Adding upstream version 6.1.76.upstream/6.1.76upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

28 files changed, 14816 insertions, 0 deletions

diff --git a/drivers/infiniband/hw/mthca/Kconfig b/drivers/infiniband/hw/mthca/Kconfig
new file mode 100644
index 000000000..faa7381f7
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/Kconfig
@@ -0,0 +1,18 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config INFINIBAND_MTHCA
+	tristate "Mellanox HCA support"
+	depends on PCI
+	help
+	  This is a low-level driver for Mellanox InfiniHost host
+	  channel adapters (HCAs), including the MT23108 PCI-X HCA
+	  ("Tavor") and the MT25208 PCI Express HCA ("Arbel").
+
+config INFINIBAND_MTHCA_DEBUG
+	bool "Verbose debugging output" if EXPERT
+	depends on INFINIBAND_MTHCA
+	default y
+	help
+	  This option causes debugging code to be compiled into the
+	  mthca driver.  The output can be turned on via the
+	  debug_level module parameter (which can also be set after
+	  the driver is loaded through sysfs).
diff --git a/drivers/infiniband/hw/mthca/Makefile b/drivers/infiniband/hw/mthca/Makefile
new file mode 100644
index 000000000..3a09e9ffd
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/Makefile
@@ -0,0 +1,8 @@
+# SPDX-License-Identifier: GPL-2.0
+obj-$(CONFIG_INFINIBAND_MTHCA) += ib_mthca.o
+
+ib_mthca-y :=	mthca_main.o mthca_cmd.o mthca_profile.o mthca_reset.o \
+		mthca_allocator.o mthca_eq.o mthca_pd.o mthca_cq.o \
+		mthca_mr.o mthca_qp.o mthca_av.o mthca_mcg.o mthca_mad.o \
+		mthca_provider.o mthca_memfree.o mthca_uar.o mthca_srq.o \
+		mthca_catas.o
diff --git a/drivers/infiniband/hw/mthca/mthca_allocator.c b/drivers/infiniband/hw/mthca/mthca_allocator.c
new file mode 100644
index 000000000..9f0f79d02
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_allocator.c
@@ -0,0 +1,298 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/errno.h>
+#include <linux/slab.h>
+#include <linux/bitmap.h>
+
+#include "mthca_dev.h"
+
+/* Trivial bitmap-based allocator */
+u32 mthca_alloc(struct mthca_alloc *alloc)
+{
+	unsigned long flags;
+	u32 obj;
+
+	spin_lock_irqsave(&alloc->lock, flags);
+
+	obj = find_next_zero_bit(alloc->table, alloc->max, alloc->last);
+	if (obj >= alloc->max) {
+		alloc->top = (alloc->top + alloc->max) & alloc->mask;
+		obj = find_first_zero_bit(alloc->table, alloc->max);
+	}
+
+	if (obj < alloc->max) {
+		__set_bit(obj, alloc->table);
+		obj |= alloc->top;
+	} else
+		obj = -1;
+
+	spin_unlock_irqrestore(&alloc->lock, flags);
+
+	return obj;
+}
+
+void mthca_free(struct mthca_alloc *alloc, u32 obj)
+{
+	unsigned long flags;
+
+	obj &= alloc->max - 1;
+
+	spin_lock_irqsave(&alloc->lock, flags);
+
+	__clear_bit(obj, alloc->table);
+	alloc->last = min(alloc->last, obj);
+	alloc->top = (alloc->top + alloc->max) & alloc->mask;
+
+	spin_unlock_irqrestore(&alloc->lock, flags);
+}
+
+int mthca_alloc_init(struct mthca_alloc *alloc, u32 num, u32 mask,
+		     u32 reserved)
+{
+	/* num must be a power of 2 */
+	if (num != 1 << (ffs(num) - 1))
+		return -EINVAL;
+
+	alloc->last = 0;
+	alloc->top  = 0;
+	alloc->max  = num;
+	alloc->mask = mask;
+	spin_lock_init(&alloc->lock);
+	alloc->table = bitmap_zalloc(num, GFP_KERNEL);
+	if (!alloc->table)
+		return -ENOMEM;
+
+	bitmap_set(alloc->table, 0, reserved);
+
+	return 0;
+}
+
+void mthca_alloc_cleanup(struct mthca_alloc *alloc)
+{
+	bitmap_free(alloc->table);
+}
+
+/*
+ * Array of pointers with lazy allocation of leaf pages.  Callers of
+ * _get, _set and _clear methods must use a lock or otherwise
+ * serialize access to the array.
+ */
+
+#define MTHCA_ARRAY_MASK (PAGE_SIZE / sizeof (void *) - 1)
+
+void *mthca_array_get(struct mthca_array *array, int index)
+{
+	int p = (index * sizeof (void *)) >> PAGE_SHIFT;
+
+	if (array->page_list[p].page)
+		return array->page_list[p].page[index & MTHCA_ARRAY_MASK];
+	else
+		return NULL;
+}
+
+int mthca_array_set(struct mthca_array *array, int index, void *value)
+{
+	int p = (index * sizeof (void *)) >> PAGE_SHIFT;
+
+	/* Allocate with GFP_ATOMIC because we'll be called with locks held. */
+	if (!array->page_list[p].page)
+		array->page_list[p].page = (void **) get_zeroed_page(GFP_ATOMIC);
+
+	if (!array->page_list[p].page)
+		return -ENOMEM;
+
+	array->page_list[p].page[index & MTHCA_ARRAY_MASK] = value;
+	++array->page_list[p].used;
+
+	return 0;
+}
+
+void mthca_array_clear(struct mthca_array *array, int index)
+{
+	int p = (index * sizeof (void *)) >> PAGE_SHIFT;
+
+	if (--array->page_list[p].used == 0) {
+		free_page((unsigned long) array->page_list[p].page);
+		array->page_list[p].page = NULL;
+	} else
+		array->page_list[p].page[index & MTHCA_ARRAY_MASK] = NULL;
+
+	if (array->page_list[p].used < 0)
+		pr_debug("Array %p index %d page %d with ref count %d < 0\n",
+			 array, index, p, array->page_list[p].used);
+}
+
+int mthca_array_init(struct mthca_array *array, int nent)
+{
+	int npage = (nent * sizeof (void *) + PAGE_SIZE - 1) / PAGE_SIZE;
+	int i;
+
+	array->page_list = kmalloc_array(npage, sizeof(*array->page_list),
+					 GFP_KERNEL);
+	if (!array->page_list)
+		return -ENOMEM;
+
+	for (i = 0; i < npage; ++i) {
+		array->page_list[i].page = NULL;
+		array->page_list[i].used = 0;
+	}
+
+	return 0;
+}
+
+void mthca_array_cleanup(struct mthca_array *array, int nent)
+{
+	int i;
+
+	for (i = 0; i < (nent * sizeof (void *) + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
+		free_page((unsigned long) array->page_list[i].page);
+
+	kfree(array->page_list);
+}
+
+/*
+ * Handling for queue buffers -- we allocate a bunch of memory and
+ * register it in a memory region at HCA virtual address 0.  If the
+ * requested size is > max_direct, we split the allocation into
+ * multiple pages, so we don't require too much contiguous memory.
+ */
+
+int mthca_buf_alloc(struct mthca_dev *dev, int size, int max_direct,
+		    union mthca_buf *buf, int *is_direct, struct mthca_pd *pd,
+		    int hca_write, struct mthca_mr *mr)
+{
+	int err = -ENOMEM;
+	int npages, shift;
+	u64 *dma_list = NULL;
+	dma_addr_t t;
+	int i;
+
+	if (size <= max_direct) {
+		*is_direct = 1;
+		npages     = 1;
+		shift      = get_order(size) + PAGE_SHIFT;
+
+		buf->direct.buf = dma_alloc_coherent(&dev->pdev->dev,
+						     size, &t, GFP_KERNEL);
+		if (!buf->direct.buf)
+			return -ENOMEM;
+
+		dma_unmap_addr_set(&buf->direct, mapping, t);
+
+		while (t & ((1 << shift) - 1)) {
+			--shift;
+			npages *= 2;
+		}
+
+		dma_list = kmalloc_array(npages, sizeof(*dma_list),
+					 GFP_KERNEL);
+		if (!dma_list)
+			goto err_free;
+
+		for (i = 0; i < npages; ++i)
+			dma_list[i] = t + i * (1 << shift);
+	} else {
+		*is_direct = 0;
+		npages     = (size + PAGE_SIZE - 1) / PAGE_SIZE;
+		shift      = PAGE_SHIFT;
+
+		dma_list = kmalloc_array(npages, sizeof(*dma_list),
+					 GFP_KERNEL);
+		if (!dma_list)
+			return -ENOMEM;
+
+		buf->page_list = kmalloc_array(npages,
+					       sizeof(*buf->page_list),
+					       GFP_KERNEL);
+		if (!buf->page_list)
+			goto err_out;
+
+		for (i = 0; i < npages; ++i)
+			buf->page_list[i].buf = NULL;
+
+		for (i = 0; i < npages; ++i) {
+			buf->page_list[i].buf =
+				dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
+						   &t, GFP_KERNEL);
+			if (!buf->page_list[i].buf)
+				goto err_free;
+
+			dma_list[i] = t;
+			dma_unmap_addr_set(&buf->page_list[i], mapping, t);
+
+			clear_page(buf->page_list[i].buf);
+		}
+	}
+
+	err = mthca_mr_alloc_phys(dev, pd->pd_num,
+				  dma_list, shift, npages,
+				  0, size,
+				  MTHCA_MPT_FLAG_LOCAL_READ |
+				  (hca_write ? MTHCA_MPT_FLAG_LOCAL_WRITE : 0),
+				  mr);
+	if (err)
+		goto err_free;
+
+	kfree(dma_list);
+
+	return 0;
+
+err_free:
+	mthca_buf_free(dev, size, buf, *is_direct, NULL);
+
+err_out:
+	kfree(dma_list);
+
+	return err;
+}
+
+void mthca_buf_free(struct mthca_dev *dev, int size, union mthca_buf *buf,
+		    int is_direct, struct mthca_mr *mr)
+{
+	int i;
+
+	if (mr)
+		mthca_free_mr(dev, mr);
+
+	if (is_direct)
+		dma_free_coherent(&dev->pdev->dev, size, buf->direct.buf,
+				  dma_unmap_addr(&buf->direct, mapping));
+	else {
+		for (i = 0; i < (size + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
+			dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+					  buf->page_list[i].buf,
+					  dma_unmap_addr(&buf->page_list[i],
+							 mapping));
+		kfree(buf->page_list);
+	}
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_av.c b/drivers/infiniband/hw/mthca/mthca_av.c
new file mode 100644
index 000000000..3df1f5ff7
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_av.c
@@ -0,0 +1,376 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/string.h>
+#include <linux/slab.h>
+
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_cache.h>
+
+#include "mthca_dev.h"
+
+enum {
+      MTHCA_RATE_TAVOR_FULL   = 0,
+      MTHCA_RATE_TAVOR_1X     = 1,
+      MTHCA_RATE_TAVOR_4X     = 2,
+      MTHCA_RATE_TAVOR_1X_DDR = 3
+};
+
+enum {
+      MTHCA_RATE_MEMFREE_FULL    = 0,
+      MTHCA_RATE_MEMFREE_QUARTER = 1,
+      MTHCA_RATE_MEMFREE_EIGHTH  = 2,
+      MTHCA_RATE_MEMFREE_HALF    = 3
+};
+
+struct mthca_av {
+	__be32 port_pd;
+	u8     reserved1;
+	u8     g_slid;
+	__be16 dlid;
+	u8     reserved2;
+	u8     gid_index;
+	u8     msg_sr;
+	u8     hop_limit;
+	__be32 sl_tclass_flowlabel;
+	__be32 dgid[4];
+};
+
+static enum ib_rate memfree_rate_to_ib(u8 mthca_rate, u8 port_rate)
+{
+	switch (mthca_rate) {
+	case MTHCA_RATE_MEMFREE_EIGHTH:
+		return mult_to_ib_rate(port_rate >> 3);
+	case MTHCA_RATE_MEMFREE_QUARTER:
+		return mult_to_ib_rate(port_rate >> 2);
+	case MTHCA_RATE_MEMFREE_HALF:
+		return mult_to_ib_rate(port_rate >> 1);
+	case MTHCA_RATE_MEMFREE_FULL:
+	default:
+		return mult_to_ib_rate(port_rate);
+	}
+}
+
+static enum ib_rate tavor_rate_to_ib(u8 mthca_rate, u8 port_rate)
+{
+	switch (mthca_rate) {
+	case MTHCA_RATE_TAVOR_1X:     return IB_RATE_2_5_GBPS;
+	case MTHCA_RATE_TAVOR_1X_DDR: return IB_RATE_5_GBPS;
+	case MTHCA_RATE_TAVOR_4X:     return IB_RATE_10_GBPS;
+	default:		      return mult_to_ib_rate(port_rate);
+	}
+}
+
+enum ib_rate mthca_rate_to_ib(struct mthca_dev *dev, u8 mthca_rate, u32 port)
+{
+	if (mthca_is_memfree(dev)) {
+		/* Handle old Arbel FW */
+		if (dev->limits.stat_rate_support == 0x3 && mthca_rate)
+			return IB_RATE_2_5_GBPS;
+
+		return memfree_rate_to_ib(mthca_rate, dev->rate[port - 1]);
+	} else
+		return tavor_rate_to_ib(mthca_rate, dev->rate[port - 1]);
+}
+
+static u8 ib_rate_to_memfree(u8 req_rate, u8 cur_rate)
+{
+	if (cur_rate <= req_rate)
+		return 0;
+
+	/*
+	 * Inter-packet delay (IPD) to get from rate X down to a rate
+	 * no more than Y is (X - 1) / Y.
+	 */
+	switch ((cur_rate - 1) / req_rate) {
+	case 0:	 return MTHCA_RATE_MEMFREE_FULL;
+	case 1:	 return MTHCA_RATE_MEMFREE_HALF;
+	case 2:
+	case 3:	 return MTHCA_RATE_MEMFREE_QUARTER;
+	default: return MTHCA_RATE_MEMFREE_EIGHTH;
+	}
+}
+
+static u8 ib_rate_to_tavor(u8 static_rate)
+{
+	switch (static_rate) {
+	case IB_RATE_2_5_GBPS: return MTHCA_RATE_TAVOR_1X;
+	case IB_RATE_5_GBPS:   return MTHCA_RATE_TAVOR_1X_DDR;
+	case IB_RATE_10_GBPS:  return MTHCA_RATE_TAVOR_4X;
+	default:	       return MTHCA_RATE_TAVOR_FULL;
+	}
+}
+
+u8 mthca_get_rate(struct mthca_dev *dev, int static_rate, u32 port)
+{
+	u8 rate;
+
+	if (!static_rate || ib_rate_to_mult(static_rate) >= dev->rate[port - 1])
+		return 0;
+
+	if (mthca_is_memfree(dev))
+		rate = ib_rate_to_memfree(ib_rate_to_mult(static_rate),
+					  dev->rate[port - 1]);
+	else
+		rate = ib_rate_to_tavor(static_rate);
+
+	if (!(dev->limits.stat_rate_support & (1 << rate)))
+		rate = 1;
+
+	return rate;
+}
+
+int mthca_create_ah(struct mthca_dev *dev,
+		    struct mthca_pd *pd,
+		    struct rdma_ah_attr *ah_attr,
+		    struct mthca_ah *ah)
+{
+	u32 index = -1;
+	struct mthca_av *av = NULL;
+
+	ah->type = MTHCA_AH_PCI_POOL;
+
+	if (mthca_is_memfree(dev)) {
+		ah->av   = kmalloc(sizeof *ah->av, GFP_ATOMIC);
+		if (!ah->av)
+			return -ENOMEM;
+
+		ah->type = MTHCA_AH_KMALLOC;
+		av       = ah->av;
+	} else if (!atomic_read(&pd->sqp_count) &&
+		 !(dev->mthca_flags & MTHCA_FLAG_DDR_HIDDEN)) {
+		index = mthca_alloc(&dev->av_table.alloc);
+
+		/* fall back to allocate in host memory */
+		if (index == -1)
+			goto on_hca_fail;
+
+		av = kmalloc(sizeof *av, GFP_ATOMIC);
+		if (!av)
+			goto on_hca_fail;
+
+		ah->type = MTHCA_AH_ON_HCA;
+		ah->avdma  = dev->av_table.ddr_av_base +
+			index * MTHCA_AV_SIZE;
+	}
+
+on_hca_fail:
+	if (ah->type == MTHCA_AH_PCI_POOL) {
+		ah->av = dma_pool_zalloc(dev->av_table.pool,
+					 GFP_ATOMIC, &ah->avdma);
+		if (!ah->av)
+			return -ENOMEM;
+
+		av = ah->av;
+	}
+
+	ah->key = pd->ntmr.ibmr.lkey;
+
+	av->port_pd = cpu_to_be32(pd->pd_num |
+				  (rdma_ah_get_port_num(ah_attr) << 24));
+	av->g_slid  = rdma_ah_get_path_bits(ah_attr);
+	av->dlid    = cpu_to_be16(rdma_ah_get_dlid(ah_attr));
+	av->msg_sr  = (3 << 4) | /* 2K message */
+		mthca_get_rate(dev, rdma_ah_get_static_rate(ah_attr),
+			       rdma_ah_get_port_num(ah_attr));
+	av->sl_tclass_flowlabel = cpu_to_be32(rdma_ah_get_sl(ah_attr) << 28);
+	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
+		const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
+
+		av->g_slid |= 0x80;
+		av->gid_index = (rdma_ah_get_port_num(ah_attr) - 1) *
+				  dev->limits.gid_table_len +
+				  grh->sgid_index;
+		av->hop_limit = grh->hop_limit;
+		av->sl_tclass_flowlabel |=
+			cpu_to_be32((grh->traffic_class << 20) |
+				    grh->flow_label);
+		memcpy(av->dgid, grh->dgid.raw, 16);
+	} else {
+		/* Arbel workaround -- low byte of GID must be 2 */
+		av->dgid[3] = cpu_to_be32(2);
+	}
+
+	if (0) {
+		int j;
+
+		mthca_dbg(dev, "Created UDAV at %p/%08lx:\n",
+			  av, (unsigned long) ah->avdma);
+		for (j = 0; j < 8; ++j)
+			printk(KERN_DEBUG "  [%2x] %08x\n",
+			       j * 4, be32_to_cpu(((__be32 *) av)[j]));
+	}
+
+	if (ah->type == MTHCA_AH_ON_HCA) {
+		memcpy_toio(dev->av_table.av_map + index * MTHCA_AV_SIZE,
+			    av, MTHCA_AV_SIZE);
+		kfree(av);
+	}
+
+	return 0;
+}
+
+int mthca_destroy_ah(struct mthca_dev *dev, struct mthca_ah *ah)
+{
+	switch (ah->type) {
+	case MTHCA_AH_ON_HCA:
+		mthca_free(&dev->av_table.alloc,
+			   (ah->avdma - dev->av_table.ddr_av_base) /
+			   MTHCA_AV_SIZE);
+		break;
+
+	case MTHCA_AH_PCI_POOL:
+		dma_pool_free(dev->av_table.pool, ah->av, ah->avdma);
+		break;
+
+	case MTHCA_AH_KMALLOC:
+		kfree(ah->av);
+		break;
+	}
+
+	return 0;
+}
+
+int mthca_ah_grh_present(struct mthca_ah *ah)
+{
+	return !!(ah->av->g_slid & 0x80);
+}
+
+int mthca_read_ah(struct mthca_dev *dev, struct mthca_ah *ah,
+		  struct ib_ud_header *header)
+{
+	if (ah->type == MTHCA_AH_ON_HCA)
+		return -EINVAL;
+
+	header->lrh.service_level   = be32_to_cpu(ah->av->sl_tclass_flowlabel) >> 28;
+	header->lrh.destination_lid = ah->av->dlid;
+	header->lrh.source_lid      = cpu_to_be16(ah->av->g_slid & 0x7f);
+	if (mthca_ah_grh_present(ah)) {
+		header->grh.traffic_class =
+			(be32_to_cpu(ah->av->sl_tclass_flowlabel) >> 20) & 0xff;
+		header->grh.flow_label    =
+			ah->av->sl_tclass_flowlabel & cpu_to_be32(0xfffff);
+		header->grh.hop_limit     = ah->av->hop_limit;
+		header->grh.source_gid = ah->ibah.sgid_attr->gid;
+		memcpy(header->grh.destination_gid.raw,
+		       ah->av->dgid, 16);
+	}
+
+	return 0;
+}
+
+int mthca_ah_query(struct ib_ah *ibah, struct rdma_ah_attr *attr)
+{
+	struct mthca_ah *ah   = to_mah(ibah);
+	struct mthca_dev *dev = to_mdev(ibah->device);
+	u32 port_num = be32_to_cpu(ah->av->port_pd) >> 24;
+
+	/* Only implement for MAD and memfree ah for now. */
+	if (ah->type == MTHCA_AH_ON_HCA)
+		return -ENOSYS;
+
+	memset(attr, 0, sizeof *attr);
+	attr->type = ibah->type;
+	rdma_ah_set_dlid(attr, be16_to_cpu(ah->av->dlid));
+	rdma_ah_set_sl(attr, be32_to_cpu(ah->av->sl_tclass_flowlabel) >> 28);
+	rdma_ah_set_port_num(attr, port_num);
+	rdma_ah_set_static_rate(attr,
+				mthca_rate_to_ib(dev, ah->av->msg_sr & 0x7,
+						 port_num));
+	rdma_ah_set_path_bits(attr, ah->av->g_slid & 0x7F);
+	if (mthca_ah_grh_present(ah)) {
+		u32 tc_fl = be32_to_cpu(ah->av->sl_tclass_flowlabel);
+
+		rdma_ah_set_grh(attr, NULL,
+				tc_fl & 0xfffff,
+				ah->av->gid_index &
+				(dev->limits.gid_table_len - 1),
+				ah->av->hop_limit,
+				(tc_fl >> 20) & 0xff);
+		rdma_ah_set_dgid_raw(attr, ah->av->dgid);
+	}
+
+	return 0;
+}
+
+int mthca_init_av_table(struct mthca_dev *dev)
+{
+	int err;
+
+	if (mthca_is_memfree(dev))
+		return 0;
+
+	err = mthca_alloc_init(&dev->av_table.alloc,
+			       dev->av_table.num_ddr_avs,
+			       dev->av_table.num_ddr_avs - 1,
+			       0);
+	if (err)
+		return err;
+
+	dev->av_table.pool = dma_pool_create("mthca_av", &dev->pdev->dev,
+					     MTHCA_AV_SIZE,
+					     MTHCA_AV_SIZE, 0);
+	if (!dev->av_table.pool)
+		goto out_free_alloc;
+
+	if (!(dev->mthca_flags & MTHCA_FLAG_DDR_HIDDEN)) {
+		dev->av_table.av_map = ioremap(pci_resource_start(dev->pdev, 4) +
+					       dev->av_table.ddr_av_base -
+					       dev->ddr_start,
+					       dev->av_table.num_ddr_avs *
+					       MTHCA_AV_SIZE);
+		if (!dev->av_table.av_map)
+			goto out_free_pool;
+	} else
+		dev->av_table.av_map = NULL;
+
+	return 0;
+
+ out_free_pool:
+	dma_pool_destroy(dev->av_table.pool);
+
+ out_free_alloc:
+	mthca_alloc_cleanup(&dev->av_table.alloc);
+	return -ENOMEM;
+}
+
+void mthca_cleanup_av_table(struct mthca_dev *dev)
+{
+	if (mthca_is_memfree(dev))
+		return;
+
+	if (dev->av_table.av_map)
+		iounmap(dev->av_table.av_map);
+	dma_pool_destroy(dev->av_table.pool);
+	mthca_alloc_cleanup(&dev->av_table.alloc);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_catas.c b/drivers/infiniband/hw/mthca/mthca_catas.c
new file mode 100644
index 000000000..ffb98eaaf
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_catas.c
@@ -0,0 +1,198 @@
+/*
+ * Copyright (c) 2005 Cisco Systems.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/jiffies.h>
+#include <linux/module.h>
+#include <linux/timer.h>
+#include <linux/workqueue.h>
+
+#include "mthca_dev.h"
+
+enum {
+	MTHCA_CATAS_POLL_INTERVAL	= 5 * HZ,
+
+	MTHCA_CATAS_TYPE_INTERNAL	= 0,
+	MTHCA_CATAS_TYPE_UPLINK		= 3,
+	MTHCA_CATAS_TYPE_DDR		= 4,
+	MTHCA_CATAS_TYPE_PARITY		= 5,
+};
+
+static DEFINE_SPINLOCK(catas_lock);
+
+static LIST_HEAD(catas_list);
+static struct workqueue_struct *catas_wq;
+static struct work_struct catas_work;
+
+static int catas_reset_disable;
+module_param_named(catas_reset_disable, catas_reset_disable, int, 0644);
+MODULE_PARM_DESC(catas_reset_disable, "disable reset on catastrophic event if nonzero");
+
+static void catas_reset(struct work_struct *work)
+{
+	struct mthca_dev *dev, *tmpdev;
+	LIST_HEAD(tlist);
+	int ret;
+
+	mutex_lock(&mthca_device_mutex);
+
+	spin_lock_irq(&catas_lock);
+	list_splice_init(&catas_list, &tlist);
+	spin_unlock_irq(&catas_lock);
+
+	list_for_each_entry_safe(dev, tmpdev, &tlist, catas_err.list) {
+		struct pci_dev *pdev = dev->pdev;
+		ret = __mthca_restart_one(dev->pdev);
+		/* 'dev' now is not valid */
+		if (ret)
+			printk(KERN_ERR "mthca %s: Reset failed (%d)\n",
+			       pci_name(pdev), ret);
+		else {
+			struct mthca_dev *d = pci_get_drvdata(pdev);
+			mthca_dbg(d, "Reset succeeded\n");
+		}
+	}
+
+	mutex_unlock(&mthca_device_mutex);
+}
+
+static void handle_catas(struct mthca_dev *dev)
+{
+	struct ib_event event;
+	unsigned long flags;
+	const char *type;
+	int i;
+
+	event.device = &dev->ib_dev;
+	event.event  = IB_EVENT_DEVICE_FATAL;
+	event.element.port_num = 0;
+	dev->active = false;
+
+	ib_dispatch_event(&event);
+
+	switch (swab32(readl(dev->catas_err.map)) >> 24) {
+	case MTHCA_CATAS_TYPE_INTERNAL:
+		type = "internal error";
+		break;
+	case MTHCA_CATAS_TYPE_UPLINK:
+		type = "uplink bus error";
+		break;
+	case MTHCA_CATAS_TYPE_DDR:
+		type = "DDR data error";
+		break;
+	case MTHCA_CATAS_TYPE_PARITY:
+		type = "internal parity error";
+		break;
+	default:
+		type = "unknown error";
+		break;
+	}
+
+	mthca_err(dev, "Catastrophic error detected: %s\n", type);
+	for (i = 0; i < dev->catas_err.size; ++i)
+		mthca_err(dev, "  buf[%02x]: %08x\n",
+			  i, swab32(readl(dev->catas_err.map + i)));
+
+	if (catas_reset_disable)
+		return;
+
+	spin_lock_irqsave(&catas_lock, flags);
+	list_add(&dev->catas_err.list, &catas_list);
+	queue_work(catas_wq, &catas_work);
+	spin_unlock_irqrestore(&catas_lock, flags);
+}
+
+static void poll_catas(struct timer_list *t)
+{
+	struct mthca_dev *dev = from_timer(dev, t, catas_err.timer);
+	int i;
+
+	for (i = 0; i < dev->catas_err.size; ++i)
+		if (readl(dev->catas_err.map + i)) {
+			handle_catas(dev);
+			return;
+		}
+
+	mod_timer(&dev->catas_err.timer,
+		  round_jiffies(jiffies + MTHCA_CATAS_POLL_INTERVAL));
+}
+
+void mthca_start_catas_poll(struct mthca_dev *dev)
+{
+	phys_addr_t addr;
+
+	timer_setup(&dev->catas_err.timer, poll_catas, 0);
+	dev->catas_err.map  = NULL;
+
+	addr = pci_resource_start(dev->pdev, 0) +
+		((pci_resource_len(dev->pdev, 0) - 1) &
+		 dev->catas_err.addr);
+
+	dev->catas_err.map = ioremap(addr, dev->catas_err.size * 4);
+	if (!dev->catas_err.map) {
+		mthca_warn(dev, "couldn't map catastrophic error region "
+			   "at 0x%llx/0x%x\n", (unsigned long long) addr,
+			   dev->catas_err.size * 4);
+		return;
+	}
+
+	dev->catas_err.timer.expires  = jiffies + MTHCA_CATAS_POLL_INTERVAL;
+	INIT_LIST_HEAD(&dev->catas_err.list);
+	add_timer(&dev->catas_err.timer);
+}
+
+void mthca_stop_catas_poll(struct mthca_dev *dev)
+{
+	del_timer_sync(&dev->catas_err.timer);
+
+	if (dev->catas_err.map)
+		iounmap(dev->catas_err.map);
+
+	spin_lock_irq(&catas_lock);
+	list_del(&dev->catas_err.list);
+	spin_unlock_irq(&catas_lock);
+}
+
+int __init mthca_catas_init(void)
+{
+	INIT_WORK(&catas_work, catas_reset);
+
+	catas_wq = alloc_ordered_workqueue("mthca_catas", WQ_MEM_RECLAIM);
+	if (!catas_wq)
+		return -ENOMEM;
+
+	return 0;
+}
+
+void mthca_catas_cleanup(void)
+{
+	destroy_workqueue(catas_wq);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_cmd.c b/drivers/infiniband/hw/mthca/mthca_cmd.c
new file mode 100644
index 000000000..8fe0cef7e
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_cmd.c
@@ -0,0 +1,1971 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/completion.h>
+#include <linux/pci.h>
+#include <linux/errno.h>
+#include <linux/sched.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <asm/io.h>
+#include <rdma/ib_mad.h>
+
+#include "mthca_dev.h"
+#include "mthca_config_reg.h"
+#include "mthca_cmd.h"
+#include "mthca_memfree.h"
+
+#define CMD_POLL_TOKEN 0xffff
+
+enum {
+	HCR_IN_PARAM_OFFSET    = 0x00,
+	HCR_IN_MODIFIER_OFFSET = 0x08,
+	HCR_OUT_PARAM_OFFSET   = 0x0c,
+	HCR_TOKEN_OFFSET       = 0x14,
+	HCR_STATUS_OFFSET      = 0x18,
+
+	HCR_OPMOD_SHIFT        = 12,
+	HCA_E_BIT              = 22,
+	HCR_GO_BIT             = 23
+};
+
+enum {
+	/* initialization and general commands */
+	CMD_SYS_EN          = 0x1,
+	CMD_SYS_DIS         = 0x2,
+	CMD_MAP_FA          = 0xfff,
+	CMD_UNMAP_FA        = 0xffe,
+	CMD_RUN_FW          = 0xff6,
+	CMD_MOD_STAT_CFG    = 0x34,
+	CMD_QUERY_DEV_LIM   = 0x3,
+	CMD_QUERY_FW        = 0x4,
+	CMD_ENABLE_LAM      = 0xff8,
+	CMD_DISABLE_LAM     = 0xff7,
+	CMD_QUERY_DDR       = 0x5,
+	CMD_QUERY_ADAPTER   = 0x6,
+	CMD_INIT_HCA        = 0x7,
+	CMD_CLOSE_HCA       = 0x8,
+	CMD_INIT_IB         = 0x9,
+	CMD_CLOSE_IB        = 0xa,
+	CMD_QUERY_HCA       = 0xb,
+	CMD_SET_IB          = 0xc,
+	CMD_ACCESS_DDR      = 0x2e,
+	CMD_MAP_ICM         = 0xffa,
+	CMD_UNMAP_ICM       = 0xff9,
+	CMD_MAP_ICM_AUX     = 0xffc,
+	CMD_UNMAP_ICM_AUX   = 0xffb,
+	CMD_SET_ICM_SIZE    = 0xffd,
+
+	/* TPT commands */
+	CMD_SW2HW_MPT 	    = 0xd,
+	CMD_QUERY_MPT 	    = 0xe,
+	CMD_HW2SW_MPT 	    = 0xf,
+	CMD_READ_MTT        = 0x10,
+	CMD_WRITE_MTT       = 0x11,
+	CMD_SYNC_TPT        = 0x2f,
+
+	/* EQ commands */
+	CMD_MAP_EQ          = 0x12,
+	CMD_SW2HW_EQ 	    = 0x13,
+	CMD_HW2SW_EQ 	    = 0x14,
+	CMD_QUERY_EQ        = 0x15,
+
+	/* CQ commands */
+	CMD_SW2HW_CQ 	    = 0x16,
+	CMD_HW2SW_CQ 	    = 0x17,
+	CMD_QUERY_CQ 	    = 0x18,
+	CMD_RESIZE_CQ       = 0x2c,
+
+	/* SRQ commands */
+	CMD_SW2HW_SRQ 	    = 0x35,
+	CMD_HW2SW_SRQ 	    = 0x36,
+	CMD_QUERY_SRQ       = 0x37,
+	CMD_ARM_SRQ         = 0x40,
+
+	/* QP/EE commands */
+	CMD_RST2INIT_QPEE   = 0x19,
+	CMD_INIT2RTR_QPEE   = 0x1a,
+	CMD_RTR2RTS_QPEE    = 0x1b,
+	CMD_RTS2RTS_QPEE    = 0x1c,
+	CMD_SQERR2RTS_QPEE  = 0x1d,
+	CMD_2ERR_QPEE       = 0x1e,
+	CMD_RTS2SQD_QPEE    = 0x1f,
+	CMD_SQD2SQD_QPEE    = 0x38,
+	CMD_SQD2RTS_QPEE    = 0x20,
+	CMD_ERR2RST_QPEE    = 0x21,
+	CMD_QUERY_QPEE      = 0x22,
+	CMD_INIT2INIT_QPEE  = 0x2d,
+	CMD_SUSPEND_QPEE    = 0x32,
+	CMD_UNSUSPEND_QPEE  = 0x33,
+	/* special QPs and management commands */
+	CMD_CONF_SPECIAL_QP = 0x23,
+	CMD_MAD_IFC         = 0x24,
+
+	/* multicast commands */
+	CMD_READ_MGM        = 0x25,
+	CMD_WRITE_MGM       = 0x26,
+	CMD_MGID_HASH       = 0x27,
+
+	/* miscellaneous commands */
+	CMD_DIAG_RPRT       = 0x30,
+	CMD_NOP             = 0x31,
+
+	/* debug commands */
+	CMD_QUERY_DEBUG_MSG = 0x2a,
+	CMD_SET_DEBUG_MSG   = 0x2b,
+};
+
+/*
+ * According to Mellanox code, FW may be starved and never complete
+ * commands.  So we can't use strict timeouts described in PRM -- we
+ * just arbitrarily select 60 seconds for now.
+ */
+#if 0
+/*
+ * Round up and add 1 to make sure we get the full wait time (since we
+ * will be starting in the middle of a jiffy)
+ */
+enum {
+	CMD_TIME_CLASS_A = (HZ + 999) / 1000 + 1,
+	CMD_TIME_CLASS_B = (HZ +  99) /  100 + 1,
+	CMD_TIME_CLASS_C = (HZ +   9) /   10 + 1,
+	CMD_TIME_CLASS_D = 60 * HZ
+};
+#else
+enum {
+	CMD_TIME_CLASS_A = 60 * HZ,
+	CMD_TIME_CLASS_B = 60 * HZ,
+	CMD_TIME_CLASS_C = 60 * HZ,
+	CMD_TIME_CLASS_D = 60 * HZ
+};
+#endif
+
+enum {
+	GO_BIT_TIMEOUT = HZ * 10
+};
+
+struct mthca_cmd_context {
+	struct completion done;
+	int               result;
+	int               next;
+	u64               out_param;
+	u16               token;
+	u8                status;
+};
+
+static int fw_cmd_doorbell = 0;
+module_param(fw_cmd_doorbell, int, 0644);
+MODULE_PARM_DESC(fw_cmd_doorbell, "post FW commands through doorbell page if nonzero "
+		 "(and supported by FW)");
+
+static inline int go_bit(struct mthca_dev *dev)
+{
+	return readl(dev->hcr + HCR_STATUS_OFFSET) &
+		swab32(1 << HCR_GO_BIT);
+}
+
+static void mthca_cmd_post_dbell(struct mthca_dev *dev,
+				 u64 in_param,
+				 u64 out_param,
+				 u32 in_modifier,
+				 u8 op_modifier,
+				 u16 op,
+				 u16 token)
+{
+	void __iomem *ptr = dev->cmd.dbell_map;
+	u16 *offs = dev->cmd.dbell_offsets;
+
+	__raw_writel((__force u32) cpu_to_be32(in_param >> 32),           ptr + offs[0]);
+	wmb();
+	__raw_writel((__force u32) cpu_to_be32(in_param & 0xfffffffful),  ptr + offs[1]);
+	wmb();
+	__raw_writel((__force u32) cpu_to_be32(in_modifier),              ptr + offs[2]);
+	wmb();
+	__raw_writel((__force u32) cpu_to_be32(out_param >> 32),          ptr + offs[3]);
+	wmb();
+	__raw_writel((__force u32) cpu_to_be32(out_param & 0xfffffffful), ptr + offs[4]);
+	wmb();
+	__raw_writel((__force u32) cpu_to_be32(token << 16),              ptr + offs[5]);
+	wmb();
+	__raw_writel((__force u32) cpu_to_be32((1 << HCR_GO_BIT)                |
+					       (1 << HCA_E_BIT)                 |
+					       (op_modifier << HCR_OPMOD_SHIFT) |
+						op),			  ptr + offs[6]);
+	wmb();
+	__raw_writel((__force u32) 0,                                     ptr + offs[7]);
+	wmb();
+}
+
+static int mthca_cmd_post_hcr(struct mthca_dev *dev,
+			      u64 in_param,
+			      u64 out_param,
+			      u32 in_modifier,
+			      u8 op_modifier,
+			      u16 op,
+			      u16 token,
+			      int event)
+{
+	if (event) {
+		unsigned long end = jiffies + GO_BIT_TIMEOUT;
+
+		while (go_bit(dev) && time_before(jiffies, end)) {
+			set_current_state(TASK_RUNNING);
+			schedule();
+		}
+	}
+
+	if (go_bit(dev))
+		return -EAGAIN;
+
+	/*
+	 * We use writel (instead of something like memcpy_toio)
+	 * because writes of less than 32 bits to the HCR don't work
+	 * (and some architectures such as ia64 implement memcpy_toio
+	 * in terms of writeb).
+	 */
+	__raw_writel((__force u32) cpu_to_be32(in_param >> 32),           dev->hcr + 0 * 4);
+	__raw_writel((__force u32) cpu_to_be32(in_param & 0xfffffffful),  dev->hcr + 1 * 4);
+	__raw_writel((__force u32) cpu_to_be32(in_modifier),              dev->hcr + 2 * 4);
+	__raw_writel((__force u32) cpu_to_be32(out_param >> 32),          dev->hcr + 3 * 4);
+	__raw_writel((__force u32) cpu_to_be32(out_param & 0xfffffffful), dev->hcr + 4 * 4);
+	__raw_writel((__force u32) cpu_to_be32(token << 16),              dev->hcr + 5 * 4);
+
+	/* __raw_writel may not order writes. */
+	wmb();
+
+	__raw_writel((__force u32) cpu_to_be32((1 << HCR_GO_BIT)                |
+					       (event ? (1 << HCA_E_BIT) : 0)   |
+					       (op_modifier << HCR_OPMOD_SHIFT) |
+					       op),                       dev->hcr + 6 * 4);
+
+	return 0;
+}
+
+static int mthca_cmd_post(struct mthca_dev *dev,
+			  u64 in_param,
+			  u64 out_param,
+			  u32 in_modifier,
+			  u8 op_modifier,
+			  u16 op,
+			  u16 token,
+			  int event)
+{
+	int err = 0;
+
+	mutex_lock(&dev->cmd.hcr_mutex);
+
+	if (event && dev->cmd.flags & MTHCA_CMD_POST_DOORBELLS && fw_cmd_doorbell)
+		mthca_cmd_post_dbell(dev, in_param, out_param, in_modifier,
+					   op_modifier, op, token);
+	else
+		err = mthca_cmd_post_hcr(dev, in_param, out_param, in_modifier,
+					 op_modifier, op, token, event);
+
+	mutex_unlock(&dev->cmd.hcr_mutex);
+	return err;
+}
+
+
+static int mthca_status_to_errno(u8 status)
+{
+	static const int trans_table[] = {
+		[MTHCA_CMD_STAT_INTERNAL_ERR]   = -EIO,
+		[MTHCA_CMD_STAT_BAD_OP]         = -EPERM,
+		[MTHCA_CMD_STAT_BAD_PARAM]      = -EINVAL,
+		[MTHCA_CMD_STAT_BAD_SYS_STATE]  = -ENXIO,
+		[MTHCA_CMD_STAT_BAD_RESOURCE]   = -EBADF,
+		[MTHCA_CMD_STAT_RESOURCE_BUSY]  = -EBUSY,
+		[MTHCA_CMD_STAT_DDR_MEM_ERR]    = -ENOMEM,
+		[MTHCA_CMD_STAT_EXCEED_LIM]     = -ENOMEM,
+		[MTHCA_CMD_STAT_BAD_RES_STATE]  = -EBADF,
+		[MTHCA_CMD_STAT_BAD_INDEX]      = -EBADF,
+		[MTHCA_CMD_STAT_BAD_NVMEM]      = -EFAULT,
+		[MTHCA_CMD_STAT_BAD_QPEE_STATE] = -EINVAL,
+		[MTHCA_CMD_STAT_BAD_SEG_PARAM]  = -EFAULT,
+		[MTHCA_CMD_STAT_REG_BOUND]      = -EBUSY,
+		[MTHCA_CMD_STAT_LAM_NOT_PRE]    = -EAGAIN,
+		[MTHCA_CMD_STAT_BAD_PKT]        = -EBADMSG,
+		[MTHCA_CMD_STAT_BAD_SIZE]       = -ENOMEM,
+	};
+
+	if (status >= ARRAY_SIZE(trans_table) ||
+			(status != MTHCA_CMD_STAT_OK
+			 && trans_table[status] == 0))
+		return -EINVAL;
+
+	return trans_table[status];
+}
+
+
+static int mthca_cmd_poll(struct mthca_dev *dev,
+			  u64 in_param,
+			  u64 *out_param,
+			  int out_is_imm,
+			  u32 in_modifier,
+			  u8 op_modifier,
+			  u16 op,
+			  unsigned long timeout)
+{
+	int err = 0;
+	unsigned long end;
+	u8 status;
+
+	down(&dev->cmd.poll_sem);
+
+	err = mthca_cmd_post(dev, in_param,
+			     out_param ? *out_param : 0,
+			     in_modifier, op_modifier,
+			     op, CMD_POLL_TOKEN, 0);
+	if (err)
+		goto out;
+
+	end = timeout + jiffies;
+	while (go_bit(dev) && time_before(jiffies, end)) {
+		set_current_state(TASK_RUNNING);
+		schedule();
+	}
+
+	if (go_bit(dev)) {
+		err = -EBUSY;
+		goto out;
+	}
+
+	if (out_is_imm && out_param) {
+		*out_param =
+			(u64) be32_to_cpu((__force __be32)
+					  __raw_readl(dev->hcr + HCR_OUT_PARAM_OFFSET)) << 32 |
+			(u64) be32_to_cpu((__force __be32)
+					  __raw_readl(dev->hcr + HCR_OUT_PARAM_OFFSET + 4));
+	} else if (out_is_imm) {
+		err = -EINVAL;
+		goto out;
+	}
+
+	status = be32_to_cpu((__force __be32) __raw_readl(dev->hcr + HCR_STATUS_OFFSET)) >> 24;
+	if (status) {
+		mthca_dbg(dev, "Command %02x completed with status %02x\n",
+			  op, status);
+		err = mthca_status_to_errno(status);
+	}
+
+out:
+	up(&dev->cmd.poll_sem);
+	return err;
+}
+
+void mthca_cmd_event(struct mthca_dev *dev,
+		     u16 token,
+		     u8  status,
+		     u64 out_param)
+{
+	struct mthca_cmd_context *context =
+		&dev->cmd.context[token & dev->cmd.token_mask];
+
+	/* previously timed out command completing at long last */
+	if (token != context->token)
+		return;
+
+	context->result    = 0;
+	context->status    = status;
+	context->out_param = out_param;
+
+	complete(&context->done);
+}
+
+static int mthca_cmd_wait(struct mthca_dev *dev,
+			  u64 in_param,
+			  u64 *out_param,
+			  int out_is_imm,
+			  u32 in_modifier,
+			  u8 op_modifier,
+			  u16 op,
+			  unsigned long timeout)
+{
+	int err = 0;
+	struct mthca_cmd_context *context;
+
+	down(&dev->cmd.event_sem);
+
+	spin_lock(&dev->cmd.context_lock);
+	BUG_ON(dev->cmd.free_head < 0);
+	context = &dev->cmd.context[dev->cmd.free_head];
+	context->token += dev->cmd.token_mask + 1;
+	dev->cmd.free_head = context->next;
+	spin_unlock(&dev->cmd.context_lock);
+
+	init_completion(&context->done);
+
+	err = mthca_cmd_post(dev, in_param,
+			     out_param ? *out_param : 0,
+			     in_modifier, op_modifier,
+			     op, context->token, 1);
+	if (err)
+		goto out;
+
+	if (!wait_for_completion_timeout(&context->done, timeout)) {
+		err = -EBUSY;
+		goto out;
+	}
+
+	err = context->result;
+	if (err)
+		goto out;
+
+	if (context->status) {
+		mthca_dbg(dev, "Command %02x completed with status %02x\n",
+			  op, context->status);
+		err = mthca_status_to_errno(context->status);
+	}
+
+	if (out_is_imm && out_param) {
+		*out_param = context->out_param;
+	} else if (out_is_imm) {
+		err = -EINVAL;
+		goto out;
+	}
+
+out:
+	spin_lock(&dev->cmd.context_lock);
+	context->next = dev->cmd.free_head;
+	dev->cmd.free_head = context - dev->cmd.context;
+	spin_unlock(&dev->cmd.context_lock);
+
+	up(&dev->cmd.event_sem);
+	return err;
+}
+
+/* Invoke a command with an output mailbox */
+static int mthca_cmd_box(struct mthca_dev *dev,
+			 u64 in_param,
+			 u64 out_param,
+			 u32 in_modifier,
+			 u8 op_modifier,
+			 u16 op,
+			 unsigned long timeout)
+{
+	if (dev->cmd.flags & MTHCA_CMD_USE_EVENTS)
+		return mthca_cmd_wait(dev, in_param, &out_param, 0,
+				      in_modifier, op_modifier, op,
+				      timeout);
+	else
+		return mthca_cmd_poll(dev, in_param, &out_param, 0,
+				      in_modifier, op_modifier, op,
+				      timeout);
+}
+
+/* Invoke a command with no output parameter */
+static int mthca_cmd(struct mthca_dev *dev,
+		     u64 in_param,
+		     u32 in_modifier,
+		     u8 op_modifier,
+		     u16 op,
+		     unsigned long timeout)
+{
+	return mthca_cmd_box(dev, in_param, 0, in_modifier,
+			     op_modifier, op, timeout);
+}
+
+/*
+ * Invoke a command with an immediate output parameter (and copy the
+ * output into the caller's out_param pointer after the command
+ * executes).
+ */
+static int mthca_cmd_imm(struct mthca_dev *dev,
+			 u64 in_param,
+			 u64 *out_param,
+			 u32 in_modifier,
+			 u8 op_modifier,
+			 u16 op,
+			 unsigned long timeout)
+{
+	if (dev->cmd.flags & MTHCA_CMD_USE_EVENTS)
+		return mthca_cmd_wait(dev, in_param, out_param, 1,
+				      in_modifier, op_modifier, op,
+				      timeout);
+	else
+		return mthca_cmd_poll(dev, in_param, out_param, 1,
+				      in_modifier, op_modifier, op,
+				      timeout);
+}
+
+int mthca_cmd_init(struct mthca_dev *dev)
+{
+	mutex_init(&dev->cmd.hcr_mutex);
+	sema_init(&dev->cmd.poll_sem, 1);
+	dev->cmd.flags = 0;
+
+	dev->hcr = ioremap(pci_resource_start(dev->pdev, 0) + MTHCA_HCR_BASE,
+			   MTHCA_HCR_SIZE);
+	if (!dev->hcr) {
+		mthca_err(dev, "Couldn't map command register.");
+		return -ENOMEM;
+	}
+
+	dev->cmd.pool = dma_pool_create("mthca_cmd", &dev->pdev->dev,
+					MTHCA_MAILBOX_SIZE,
+					MTHCA_MAILBOX_SIZE, 0);
+	if (!dev->cmd.pool) {
+		iounmap(dev->hcr);
+		return -ENOMEM;
+	}
+
+	return 0;
+}
+
+void mthca_cmd_cleanup(struct mthca_dev *dev)
+{
+	dma_pool_destroy(dev->cmd.pool);
+	iounmap(dev->hcr);
+	if (dev->cmd.flags & MTHCA_CMD_POST_DOORBELLS)
+		iounmap(dev->cmd.dbell_map);
+}
+
+/*
+ * Switch to using events to issue FW commands (should be called after
+ * event queue to command events has been initialized).
+ */
+int mthca_cmd_use_events(struct mthca_dev *dev)
+{
+	int i;
+
+	dev->cmd.context = kmalloc_array(dev->cmd.max_cmds,
+					 sizeof(struct mthca_cmd_context),
+					 GFP_KERNEL);
+	if (!dev->cmd.context)
+		return -ENOMEM;
+
+	for (i = 0; i < dev->cmd.max_cmds; ++i) {
+		dev->cmd.context[i].token = i;
+		dev->cmd.context[i].next = i + 1;
+	}
+
+	dev->cmd.context[dev->cmd.max_cmds - 1].next = -1;
+	dev->cmd.free_head = 0;
+
+	sema_init(&dev->cmd.event_sem, dev->cmd.max_cmds);
+	spin_lock_init(&dev->cmd.context_lock);
+
+	for (dev->cmd.token_mask = 1;
+	     dev->cmd.token_mask < dev->cmd.max_cmds;
+	     dev->cmd.token_mask <<= 1)
+		; /* nothing */
+	--dev->cmd.token_mask;
+
+	dev->cmd.flags |= MTHCA_CMD_USE_EVENTS;
+
+	down(&dev->cmd.poll_sem);
+
+	return 0;
+}
+
+/*
+ * Switch back to polling (used when shutting down the device)
+ */
+void mthca_cmd_use_polling(struct mthca_dev *dev)
+{
+	int i;
+
+	dev->cmd.flags &= ~MTHCA_CMD_USE_EVENTS;
+
+	for (i = 0; i < dev->cmd.max_cmds; ++i)
+		down(&dev->cmd.event_sem);
+
+	kfree(dev->cmd.context);
+
+	up(&dev->cmd.poll_sem);
+}
+
+struct mthca_mailbox *mthca_alloc_mailbox(struct mthca_dev *dev,
+					  gfp_t gfp_mask)
+{
+	struct mthca_mailbox *mailbox;
+
+	mailbox = kmalloc(sizeof *mailbox, gfp_mask);
+	if (!mailbox)
+		return ERR_PTR(-ENOMEM);
+
+	mailbox->buf = dma_pool_alloc(dev->cmd.pool, gfp_mask, &mailbox->dma);
+	if (!mailbox->buf) {
+		kfree(mailbox);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	return mailbox;
+}
+
+void mthca_free_mailbox(struct mthca_dev *dev, struct mthca_mailbox *mailbox)
+{
+	if (!mailbox)
+		return;
+
+	dma_pool_free(dev->cmd.pool, mailbox->buf, mailbox->dma);
+	kfree(mailbox);
+}
+
+int mthca_SYS_EN(struct mthca_dev *dev)
+{
+	u64 out = 0;
+	int ret;
+
+	ret = mthca_cmd_imm(dev, 0, &out, 0, 0, CMD_SYS_EN, CMD_TIME_CLASS_D);
+
+	if (ret == -ENOMEM)
+		mthca_warn(dev, "SYS_EN DDR error: syn=%x, sock=%d, "
+			   "sladdr=%d, SPD source=%s\n",
+			   (int) (out >> 6) & 0xf, (int) (out >> 4) & 3,
+			   (int) (out >> 1) & 7, (int) out & 1 ? "NVMEM" : "DIMM");
+
+	return ret;
+}
+
+int mthca_SYS_DIS(struct mthca_dev *dev)
+{
+	return mthca_cmd(dev, 0, 0, 0, CMD_SYS_DIS, CMD_TIME_CLASS_C);
+}
+
+static int mthca_map_cmd(struct mthca_dev *dev, u16 op, struct mthca_icm *icm,
+			 u64 virt)
+{
+	struct mthca_mailbox *mailbox;
+	struct mthca_icm_iter iter;
+	__be64 *pages;
+	int lg;
+	int nent = 0;
+	int i;
+	int err = 0;
+	int ts = 0, tc = 0;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	memset(mailbox->buf, 0, MTHCA_MAILBOX_SIZE);
+	pages = mailbox->buf;
+
+	for (mthca_icm_first(icm, &iter);
+	     !mthca_icm_last(&iter);
+	     mthca_icm_next(&iter)) {
+		/*
+		 * We have to pass pages that are aligned to their
+		 * size, so find the least significant 1 in the
+		 * address or size and use that as our log2 size.
+		 */
+		lg = ffs(mthca_icm_addr(&iter) | mthca_icm_size(&iter)) - 1;
+		if (lg < MTHCA_ICM_PAGE_SHIFT) {
+			mthca_warn(dev, "Got FW area not aligned to %d (%llx/%lx).\n",
+				   MTHCA_ICM_PAGE_SIZE,
+				   (unsigned long long) mthca_icm_addr(&iter),
+				   mthca_icm_size(&iter));
+			err = -EINVAL;
+			goto out;
+		}
+		for (i = 0; i < mthca_icm_size(&iter) >> lg; ++i) {
+			if (virt != -1) {
+				pages[nent * 2] = cpu_to_be64(virt);
+				virt += 1ULL << lg;
+			}
+
+			pages[nent * 2 + 1] =
+				cpu_to_be64((mthca_icm_addr(&iter) + (i << lg)) |
+					    (lg - MTHCA_ICM_PAGE_SHIFT));
+			ts += 1 << (lg - 10);
+			++tc;
+
+			if (++nent == MTHCA_MAILBOX_SIZE / 16) {
+				err = mthca_cmd(dev, mailbox->dma, nent, 0, op,
+						CMD_TIME_CLASS_B);
+				if (err)
+					goto out;
+				nent = 0;
+			}
+		}
+	}
+
+	if (nent)
+		err = mthca_cmd(dev, mailbox->dma, nent, 0, op,
+				CMD_TIME_CLASS_B);
+
+	switch (op) {
+	case CMD_MAP_FA:
+		mthca_dbg(dev, "Mapped %d chunks/%d KB for FW.\n", tc, ts);
+		break;
+	case CMD_MAP_ICM_AUX:
+		mthca_dbg(dev, "Mapped %d chunks/%d KB for ICM aux.\n", tc, ts);
+		break;
+	case CMD_MAP_ICM:
+		mthca_dbg(dev, "Mapped %d chunks/%d KB at %llx for ICM.\n",
+			  tc, ts, (unsigned long long) virt - (ts << 10));
+		break;
+	}
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_MAP_FA(struct mthca_dev *dev, struct mthca_icm *icm)
+{
+	return mthca_map_cmd(dev, CMD_MAP_FA, icm, -1);
+}
+
+int mthca_UNMAP_FA(struct mthca_dev *dev)
+{
+	return mthca_cmd(dev, 0, 0, 0, CMD_UNMAP_FA, CMD_TIME_CLASS_B);
+}
+
+int mthca_RUN_FW(struct mthca_dev *dev)
+{
+	return mthca_cmd(dev, 0, 0, 0, CMD_RUN_FW, CMD_TIME_CLASS_A);
+}
+
+static void mthca_setup_cmd_doorbells(struct mthca_dev *dev, u64 base)
+{
+	phys_addr_t addr;
+	u16 max_off = 0;
+	int i;
+
+	for (i = 0; i < 8; ++i)
+		max_off = max(max_off, dev->cmd.dbell_offsets[i]);
+
+	if ((base & PAGE_MASK) != ((base + max_off) & PAGE_MASK)) {
+		mthca_warn(dev, "Firmware doorbell region at 0x%016llx, "
+			   "length 0x%x crosses a page boundary\n",
+			   (unsigned long long) base, max_off);
+		return;
+	}
+
+	addr = pci_resource_start(dev->pdev, 2) +
+		((pci_resource_len(dev->pdev, 2) - 1) & base);
+	dev->cmd.dbell_map = ioremap(addr, max_off + sizeof(u32));
+	if (!dev->cmd.dbell_map)
+		return;
+
+	dev->cmd.flags |= MTHCA_CMD_POST_DOORBELLS;
+	mthca_dbg(dev, "Mapped doorbell page for posting FW commands\n");
+}
+
+int mthca_QUERY_FW(struct mthca_dev *dev)
+{
+	struct mthca_mailbox *mailbox;
+	u32 *outbox;
+	u64 base;
+	u32 tmp;
+	int err = 0;
+	u8 lg;
+	int i;
+
+#define QUERY_FW_OUT_SIZE             0x100
+#define QUERY_FW_VER_OFFSET            0x00
+#define QUERY_FW_MAX_CMD_OFFSET        0x0f
+#define QUERY_FW_ERR_START_OFFSET      0x30
+#define QUERY_FW_ERR_SIZE_OFFSET       0x38
+
+#define QUERY_FW_CMD_DB_EN_OFFSET      0x10
+#define QUERY_FW_CMD_DB_OFFSET         0x50
+#define QUERY_FW_CMD_DB_BASE           0x60
+
+#define QUERY_FW_START_OFFSET          0x20
+#define QUERY_FW_END_OFFSET            0x28
+
+#define QUERY_FW_SIZE_OFFSET           0x00
+#define QUERY_FW_CLR_INT_BASE_OFFSET   0x20
+#define QUERY_FW_EQ_ARM_BASE_OFFSET    0x40
+#define QUERY_FW_EQ_SET_CI_BASE_OFFSET 0x48
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
+
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_QUERY_FW,
+			    CMD_TIME_CLASS_A);
+
+	if (err)
+		goto out;
+
+	MTHCA_GET(dev->fw_ver,   outbox, QUERY_FW_VER_OFFSET);
+	/*
+	 * FW subminor version is at more significant bits than minor
+	 * version, so swap here.
+	 */
+	dev->fw_ver = (dev->fw_ver & 0xffff00000000ull) |
+		((dev->fw_ver & 0xffff0000ull) >> 16) |
+		((dev->fw_ver & 0x0000ffffull) << 16);
+
+	MTHCA_GET(lg, outbox, QUERY_FW_MAX_CMD_OFFSET);
+	dev->cmd.max_cmds = 1 << lg;
+
+	mthca_dbg(dev, "FW version %012llx, max commands %d\n",
+		  (unsigned long long) dev->fw_ver, dev->cmd.max_cmds);
+
+	MTHCA_GET(dev->catas_err.addr, outbox, QUERY_FW_ERR_START_OFFSET);
+	MTHCA_GET(dev->catas_err.size, outbox, QUERY_FW_ERR_SIZE_OFFSET);
+
+	mthca_dbg(dev, "Catastrophic error buffer at 0x%llx, size 0x%x\n",
+		  (unsigned long long) dev->catas_err.addr, dev->catas_err.size);
+
+	MTHCA_GET(tmp, outbox, QUERY_FW_CMD_DB_EN_OFFSET);
+	if (tmp & 0x1) {
+		mthca_dbg(dev, "FW supports commands through doorbells\n");
+
+		MTHCA_GET(base, outbox, QUERY_FW_CMD_DB_BASE);
+		for (i = 0; i < MTHCA_CMD_NUM_DBELL_DWORDS; ++i)
+			MTHCA_GET(dev->cmd.dbell_offsets[i], outbox,
+				  QUERY_FW_CMD_DB_OFFSET + (i << 1));
+
+		mthca_setup_cmd_doorbells(dev, base);
+	}
+
+	if (mthca_is_memfree(dev)) {
+		MTHCA_GET(dev->fw.arbel.fw_pages,       outbox, QUERY_FW_SIZE_OFFSET);
+		MTHCA_GET(dev->fw.arbel.clr_int_base,   outbox, QUERY_FW_CLR_INT_BASE_OFFSET);
+		MTHCA_GET(dev->fw.arbel.eq_arm_base,    outbox, QUERY_FW_EQ_ARM_BASE_OFFSET);
+		MTHCA_GET(dev->fw.arbel.eq_set_ci_base, outbox, QUERY_FW_EQ_SET_CI_BASE_OFFSET);
+		mthca_dbg(dev, "FW size %d KB\n", dev->fw.arbel.fw_pages << 2);
+
+		/*
+		 * Round up number of system pages needed in case
+		 * MTHCA_ICM_PAGE_SIZE < PAGE_SIZE.
+		 */
+		dev->fw.arbel.fw_pages =
+			ALIGN(dev->fw.arbel.fw_pages, PAGE_SIZE / MTHCA_ICM_PAGE_SIZE) >>
+				(PAGE_SHIFT - MTHCA_ICM_PAGE_SHIFT);
+
+		mthca_dbg(dev, "Clear int @ %llx, EQ arm @ %llx, EQ set CI @ %llx\n",
+			  (unsigned long long) dev->fw.arbel.clr_int_base,
+			  (unsigned long long) dev->fw.arbel.eq_arm_base,
+			  (unsigned long long) dev->fw.arbel.eq_set_ci_base);
+	} else {
+		MTHCA_GET(dev->fw.tavor.fw_start, outbox, QUERY_FW_START_OFFSET);
+		MTHCA_GET(dev->fw.tavor.fw_end,   outbox, QUERY_FW_END_OFFSET);
+
+		mthca_dbg(dev, "FW size %d KB (start %llx, end %llx)\n",
+			  (int) ((dev->fw.tavor.fw_end - dev->fw.tavor.fw_start) >> 10),
+			  (unsigned long long) dev->fw.tavor.fw_start,
+			  (unsigned long long) dev->fw.tavor.fw_end);
+	}
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_ENABLE_LAM(struct mthca_dev *dev)
+{
+	struct mthca_mailbox *mailbox;
+	u8 info;
+	u32 *outbox;
+	int err = 0;
+
+#define ENABLE_LAM_OUT_SIZE         0x100
+#define ENABLE_LAM_START_OFFSET     0x00
+#define ENABLE_LAM_END_OFFSET       0x08
+#define ENABLE_LAM_INFO_OFFSET      0x13
+
+#define ENABLE_LAM_INFO_HIDDEN_FLAG (1 << 4)
+#define ENABLE_LAM_INFO_ECC_MASK    0x3
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
+
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_ENABLE_LAM,
+			    CMD_TIME_CLASS_C);
+
+	if (err)
+		goto out;
+
+	MTHCA_GET(dev->ddr_start, outbox, ENABLE_LAM_START_OFFSET);
+	MTHCA_GET(dev->ddr_end,   outbox, ENABLE_LAM_END_OFFSET);
+	MTHCA_GET(info,           outbox, ENABLE_LAM_INFO_OFFSET);
+
+	if (!!(info & ENABLE_LAM_INFO_HIDDEN_FLAG) !=
+	    !!(dev->mthca_flags & MTHCA_FLAG_DDR_HIDDEN)) {
+		mthca_info(dev, "FW reports that HCA-attached memory "
+			   "is %s hidden; does not match PCI config\n",
+			   (info & ENABLE_LAM_INFO_HIDDEN_FLAG) ?
+			   "" : "not");
+	}
+	if (info & ENABLE_LAM_INFO_HIDDEN_FLAG)
+		mthca_dbg(dev, "HCA-attached memory is hidden.\n");
+
+	mthca_dbg(dev, "HCA memory size %d KB (start %llx, end %llx)\n",
+		  (int) ((dev->ddr_end - dev->ddr_start) >> 10),
+		  (unsigned long long) dev->ddr_start,
+		  (unsigned long long) dev->ddr_end);
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_DISABLE_LAM(struct mthca_dev *dev)
+{
+	return mthca_cmd(dev, 0, 0, 0, CMD_SYS_DIS, CMD_TIME_CLASS_C);
+}
+
+int mthca_QUERY_DDR(struct mthca_dev *dev)
+{
+	struct mthca_mailbox *mailbox;
+	u8 info;
+	u32 *outbox;
+	int err = 0;
+
+#define QUERY_DDR_OUT_SIZE         0x100
+#define QUERY_DDR_START_OFFSET     0x00
+#define QUERY_DDR_END_OFFSET       0x08
+#define QUERY_DDR_INFO_OFFSET      0x13
+
+#define QUERY_DDR_INFO_HIDDEN_FLAG (1 << 4)
+#define QUERY_DDR_INFO_ECC_MASK    0x3
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
+
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_QUERY_DDR,
+			    CMD_TIME_CLASS_A);
+
+	if (err)
+		goto out;
+
+	MTHCA_GET(dev->ddr_start, outbox, QUERY_DDR_START_OFFSET);
+	MTHCA_GET(dev->ddr_end,   outbox, QUERY_DDR_END_OFFSET);
+	MTHCA_GET(info,           outbox, QUERY_DDR_INFO_OFFSET);
+
+	if (!!(info & QUERY_DDR_INFO_HIDDEN_FLAG) !=
+	    !!(dev->mthca_flags & MTHCA_FLAG_DDR_HIDDEN)) {
+		mthca_info(dev, "FW reports that HCA-attached memory "
+			   "is %s hidden; does not match PCI config\n",
+			   (info & QUERY_DDR_INFO_HIDDEN_FLAG) ?
+			   "" : "not");
+	}
+	if (info & QUERY_DDR_INFO_HIDDEN_FLAG)
+		mthca_dbg(dev, "HCA-attached memory is hidden.\n");
+
+	mthca_dbg(dev, "HCA memory size %d KB (start %llx, end %llx)\n",
+		  (int) ((dev->ddr_end - dev->ddr_start) >> 10),
+		  (unsigned long long) dev->ddr_start,
+		  (unsigned long long) dev->ddr_end);
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_QUERY_DEV_LIM(struct mthca_dev *dev,
+			struct mthca_dev_lim *dev_lim)
+{
+	struct mthca_mailbox *mailbox;
+	u32 *outbox;
+	u8 field;
+	u16 size;
+	u16 stat_rate;
+	int err;
+
+#define QUERY_DEV_LIM_OUT_SIZE             0x100
+#define QUERY_DEV_LIM_MAX_SRQ_SZ_OFFSET     0x10
+#define QUERY_DEV_LIM_MAX_QP_SZ_OFFSET      0x11
+#define QUERY_DEV_LIM_RSVD_QP_OFFSET        0x12
+#define QUERY_DEV_LIM_MAX_QP_OFFSET         0x13
+#define QUERY_DEV_LIM_RSVD_SRQ_OFFSET       0x14
+#define QUERY_DEV_LIM_MAX_SRQ_OFFSET        0x15
+#define QUERY_DEV_LIM_RSVD_EEC_OFFSET       0x16
+#define QUERY_DEV_LIM_MAX_EEC_OFFSET        0x17
+#define QUERY_DEV_LIM_MAX_CQ_SZ_OFFSET      0x19
+#define QUERY_DEV_LIM_RSVD_CQ_OFFSET        0x1a
+#define QUERY_DEV_LIM_MAX_CQ_OFFSET         0x1b
+#define QUERY_DEV_LIM_MAX_MPT_OFFSET        0x1d
+#define QUERY_DEV_LIM_RSVD_EQ_OFFSET        0x1e
+#define QUERY_DEV_LIM_MAX_EQ_OFFSET         0x1f
+#define QUERY_DEV_LIM_RSVD_MTT_OFFSET       0x20
+#define QUERY_DEV_LIM_MAX_MRW_SZ_OFFSET     0x21
+#define QUERY_DEV_LIM_RSVD_MRW_OFFSET       0x22
+#define QUERY_DEV_LIM_MAX_MTT_SEG_OFFSET    0x23
+#define QUERY_DEV_LIM_MAX_AV_OFFSET         0x27
+#define QUERY_DEV_LIM_MAX_REQ_QP_OFFSET     0x29
+#define QUERY_DEV_LIM_MAX_RES_QP_OFFSET     0x2b
+#define QUERY_DEV_LIM_MAX_RDMA_OFFSET       0x2f
+#define QUERY_DEV_LIM_RSZ_SRQ_OFFSET        0x33
+#define QUERY_DEV_LIM_ACK_DELAY_OFFSET      0x35
+#define QUERY_DEV_LIM_MTU_WIDTH_OFFSET      0x36
+#define QUERY_DEV_LIM_VL_PORT_OFFSET        0x37
+#define QUERY_DEV_LIM_MAX_GID_OFFSET        0x3b
+#define QUERY_DEV_LIM_RATE_SUPPORT_OFFSET   0x3c
+#define QUERY_DEV_LIM_MAX_PKEY_OFFSET       0x3f
+#define QUERY_DEV_LIM_FLAGS_OFFSET          0x44
+#define QUERY_DEV_LIM_RSVD_UAR_OFFSET       0x48
+#define QUERY_DEV_LIM_UAR_SZ_OFFSET         0x49
+#define QUERY_DEV_LIM_PAGE_SZ_OFFSET        0x4b
+#define QUERY_DEV_LIM_MAX_SG_OFFSET         0x51
+#define QUERY_DEV_LIM_MAX_DESC_SZ_OFFSET    0x52
+#define QUERY_DEV_LIM_MAX_SG_RQ_OFFSET      0x55
+#define QUERY_DEV_LIM_MAX_DESC_SZ_RQ_OFFSET 0x56
+#define QUERY_DEV_LIM_MAX_QP_MCG_OFFSET     0x61
+#define QUERY_DEV_LIM_RSVD_MCG_OFFSET       0x62
+#define QUERY_DEV_LIM_MAX_MCG_OFFSET        0x63
+#define QUERY_DEV_LIM_RSVD_PD_OFFSET        0x64
+#define QUERY_DEV_LIM_MAX_PD_OFFSET         0x65
+#define QUERY_DEV_LIM_RSVD_RDD_OFFSET       0x66
+#define QUERY_DEV_LIM_MAX_RDD_OFFSET        0x67
+#define QUERY_DEV_LIM_EEC_ENTRY_SZ_OFFSET   0x80
+#define QUERY_DEV_LIM_QPC_ENTRY_SZ_OFFSET   0x82
+#define QUERY_DEV_LIM_EEEC_ENTRY_SZ_OFFSET  0x84
+#define QUERY_DEV_LIM_EQPC_ENTRY_SZ_OFFSET  0x86
+#define QUERY_DEV_LIM_EQC_ENTRY_SZ_OFFSET   0x88
+#define QUERY_DEV_LIM_CQC_ENTRY_SZ_OFFSET   0x8a
+#define QUERY_DEV_LIM_SRQ_ENTRY_SZ_OFFSET   0x8c
+#define QUERY_DEV_LIM_UAR_ENTRY_SZ_OFFSET   0x8e
+#define QUERY_DEV_LIM_MTT_ENTRY_SZ_OFFSET   0x90
+#define QUERY_DEV_LIM_MPT_ENTRY_SZ_OFFSET   0x92
+#define QUERY_DEV_LIM_PBL_SZ_OFFSET         0x96
+#define QUERY_DEV_LIM_BMME_FLAGS_OFFSET     0x97
+#define QUERY_DEV_LIM_RSVD_LKEY_OFFSET      0x98
+#define QUERY_DEV_LIM_LAMR_OFFSET           0x9f
+#define QUERY_DEV_LIM_MAX_ICM_SZ_OFFSET     0xa0
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
+
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_QUERY_DEV_LIM,
+			    CMD_TIME_CLASS_A);
+
+	if (err)
+		goto out;
+
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_QP_OFFSET);
+	dev_lim->reserved_qps = 1 << (field & 0xf);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_QP_OFFSET);
+	dev_lim->max_qps = 1 << (field & 0x1f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_SRQ_OFFSET);
+	dev_lim->reserved_srqs = 1 << (field >> 4);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_SRQ_OFFSET);
+	dev_lim->max_srqs = 1 << (field & 0x1f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_EEC_OFFSET);
+	dev_lim->reserved_eecs = 1 << (field & 0xf);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_EEC_OFFSET);
+	dev_lim->max_eecs = 1 << (field & 0x1f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_CQ_SZ_OFFSET);
+	dev_lim->max_cq_sz = 1 << field;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_CQ_OFFSET);
+	dev_lim->reserved_cqs = 1 << (field & 0xf);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_CQ_OFFSET);
+	dev_lim->max_cqs = 1 << (field & 0x1f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_MPT_OFFSET);
+	dev_lim->max_mpts = 1 << (field & 0x3f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_EQ_OFFSET);
+	dev_lim->reserved_eqs = 1 << (field & 0xf);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_EQ_OFFSET);
+	dev_lim->max_eqs = 1 << (field & 0x7);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_MTT_OFFSET);
+	if (mthca_is_memfree(dev))
+		dev_lim->reserved_mtts = ALIGN((1 << (field >> 4)) * sizeof(u64),
+					       dev->limits.mtt_seg_size) / dev->limits.mtt_seg_size;
+	else
+		dev_lim->reserved_mtts = 1 << (field >> 4);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_MRW_SZ_OFFSET);
+	dev_lim->max_mrw_sz = 1 << field;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_MRW_OFFSET);
+	dev_lim->reserved_mrws = 1 << (field & 0xf);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_MTT_SEG_OFFSET);
+	dev_lim->max_mtt_seg = 1 << (field & 0x3f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_REQ_QP_OFFSET);
+	dev_lim->max_requester_per_qp = 1 << (field & 0x3f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_RES_QP_OFFSET);
+	dev_lim->max_responder_per_qp = 1 << (field & 0x3f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_RDMA_OFFSET);
+	dev_lim->max_rdma_global = 1 << (field & 0x3f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_ACK_DELAY_OFFSET);
+	dev_lim->local_ca_ack_delay = field & 0x1f;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MTU_WIDTH_OFFSET);
+	dev_lim->max_mtu        = field >> 4;
+	dev_lim->max_port_width = field & 0xf;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_VL_PORT_OFFSET);
+	dev_lim->max_vl    = field >> 4;
+	dev_lim->num_ports = field & 0xf;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_GID_OFFSET);
+	dev_lim->max_gids = 1 << (field & 0xf);
+	MTHCA_GET(stat_rate, outbox, QUERY_DEV_LIM_RATE_SUPPORT_OFFSET);
+	dev_lim->stat_rate_support = stat_rate;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_PKEY_OFFSET);
+	dev_lim->max_pkeys = 1 << (field & 0xf);
+	MTHCA_GET(dev_lim->flags, outbox, QUERY_DEV_LIM_FLAGS_OFFSET);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_UAR_OFFSET);
+	dev_lim->reserved_uars = field >> 4;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_UAR_SZ_OFFSET);
+	dev_lim->uar_size = 1 << ((field & 0x3f) + 20);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_PAGE_SZ_OFFSET);
+	dev_lim->min_page_sz = 1 << field;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_SG_OFFSET);
+	dev_lim->max_sg = field;
+
+	MTHCA_GET(size, outbox, QUERY_DEV_LIM_MAX_DESC_SZ_OFFSET);
+	dev_lim->max_desc_sz = size;
+
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_QP_MCG_OFFSET);
+	dev_lim->max_qp_per_mcg = 1 << field;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_MCG_OFFSET);
+	dev_lim->reserved_mgms = field & 0xf;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_MCG_OFFSET);
+	dev_lim->max_mcgs = 1 << field;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_PD_OFFSET);
+	dev_lim->reserved_pds = field >> 4;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_PD_OFFSET);
+	dev_lim->max_pds = 1 << (field & 0x3f);
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSVD_RDD_OFFSET);
+	dev_lim->reserved_rdds = field >> 4;
+	MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_RDD_OFFSET);
+	dev_lim->max_rdds = 1 << (field & 0x3f);
+
+	MTHCA_GET(size, outbox, QUERY_DEV_LIM_EEC_ENTRY_SZ_OFFSET);
+	dev_lim->eec_entry_sz = size;
+	MTHCA_GET(size, outbox, QUERY_DEV_LIM_QPC_ENTRY_SZ_OFFSET);
+	dev_lim->qpc_entry_sz = size;
+	MTHCA_GET(size, outbox, QUERY_DEV_LIM_EEEC_ENTRY_SZ_OFFSET);
+	dev_lim->eeec_entry_sz = size;
+	MTHCA_GET(size, outbox, QUERY_DEV_LIM_EQPC_ENTRY_SZ_OFFSET);
+	dev_lim->eqpc_entry_sz = size;
+	MTHCA_GET(size, outbox, QUERY_DEV_LIM_EQC_ENTRY_SZ_OFFSET);
+	dev_lim->eqc_entry_sz = size;
+	MTHCA_GET(size, outbox, QUERY_DEV_LIM_CQC_ENTRY_SZ_OFFSET);
+	dev_lim->cqc_entry_sz = size;
+	MTHCA_GET(size, outbox, QUERY_DEV_LIM_SRQ_ENTRY_SZ_OFFSET);
+	dev_lim->srq_entry_sz = size;
+	MTHCA_GET(size, outbox, QUERY_DEV_LIM_UAR_ENTRY_SZ_OFFSET);
+	dev_lim->uar_scratch_entry_sz = size;
+
+	if (mthca_is_memfree(dev)) {
+		MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_SRQ_SZ_OFFSET);
+		dev_lim->max_srq_sz = 1 << field;
+		MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_QP_SZ_OFFSET);
+		dev_lim->max_qp_sz = 1 << field;
+		MTHCA_GET(field, outbox, QUERY_DEV_LIM_RSZ_SRQ_OFFSET);
+		dev_lim->hca.arbel.resize_srq = field & 1;
+		MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_SG_RQ_OFFSET);
+		dev_lim->max_sg = min_t(int, field, dev_lim->max_sg);
+		MTHCA_GET(size, outbox, QUERY_DEV_LIM_MAX_DESC_SZ_RQ_OFFSET);
+		dev_lim->max_desc_sz = min_t(int, size, dev_lim->max_desc_sz);
+		MTHCA_GET(size, outbox, QUERY_DEV_LIM_MPT_ENTRY_SZ_OFFSET);
+		dev_lim->mpt_entry_sz = size;
+		MTHCA_GET(field, outbox, QUERY_DEV_LIM_PBL_SZ_OFFSET);
+		dev_lim->hca.arbel.max_pbl_sz = 1 << (field & 0x3f);
+		MTHCA_GET(dev_lim->hca.arbel.bmme_flags, outbox,
+			  QUERY_DEV_LIM_BMME_FLAGS_OFFSET);
+		MTHCA_GET(dev_lim->hca.arbel.reserved_lkey, outbox,
+			  QUERY_DEV_LIM_RSVD_LKEY_OFFSET);
+		MTHCA_GET(field, outbox, QUERY_DEV_LIM_LAMR_OFFSET);
+		dev_lim->hca.arbel.lam_required = field & 1;
+		MTHCA_GET(dev_lim->hca.arbel.max_icm_sz, outbox,
+			  QUERY_DEV_LIM_MAX_ICM_SZ_OFFSET);
+
+		if (dev_lim->hca.arbel.bmme_flags & 1)
+			mthca_dbg(dev, "Base MM extensions: yes "
+				  "(flags %d, max PBL %d, rsvd L_Key %08x)\n",
+				  dev_lim->hca.arbel.bmme_flags,
+				  dev_lim->hca.arbel.max_pbl_sz,
+				  dev_lim->hca.arbel.reserved_lkey);
+		else
+			mthca_dbg(dev, "Base MM extensions: no\n");
+
+		mthca_dbg(dev, "Max ICM size %lld MB\n",
+			  (unsigned long long) dev_lim->hca.arbel.max_icm_sz >> 20);
+	} else {
+		MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_SRQ_SZ_OFFSET);
+		dev_lim->max_srq_sz = (1 << field) - 1;
+		MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_QP_SZ_OFFSET);
+		dev_lim->max_qp_sz = (1 << field) - 1;
+		MTHCA_GET(field, outbox, QUERY_DEV_LIM_MAX_AV_OFFSET);
+		dev_lim->hca.tavor.max_avs = 1 << (field & 0x3f);
+		dev_lim->mpt_entry_sz = MTHCA_MPT_ENTRY_SIZE;
+	}
+
+	mthca_dbg(dev, "Max QPs: %d, reserved QPs: %d, entry size: %d\n",
+		  dev_lim->max_qps, dev_lim->reserved_qps, dev_lim->qpc_entry_sz);
+	mthca_dbg(dev, "Max SRQs: %d, reserved SRQs: %d, entry size: %d\n",
+		  dev_lim->max_srqs, dev_lim->reserved_srqs, dev_lim->srq_entry_sz);
+	mthca_dbg(dev, "Max CQs: %d, reserved CQs: %d, entry size: %d\n",
+		  dev_lim->max_cqs, dev_lim->reserved_cqs, dev_lim->cqc_entry_sz);
+	mthca_dbg(dev, "Max EQs: %d, reserved EQs: %d, entry size: %d\n",
+		  dev_lim->max_eqs, dev_lim->reserved_eqs, dev_lim->eqc_entry_sz);
+	mthca_dbg(dev, "reserved MPTs: %d, reserved MTTs: %d\n",
+		  dev_lim->reserved_mrws, dev_lim->reserved_mtts);
+	mthca_dbg(dev, "Max PDs: %d, reserved PDs: %d, reserved UARs: %d\n",
+		  dev_lim->max_pds, dev_lim->reserved_pds, dev_lim->reserved_uars);
+	mthca_dbg(dev, "Max QP/MCG: %d, reserved MGMs: %d\n",
+		  dev_lim->max_pds, dev_lim->reserved_mgms);
+	mthca_dbg(dev, "Max CQEs: %d, max WQEs: %d, max SRQ WQEs: %d\n",
+		  dev_lim->max_cq_sz, dev_lim->max_qp_sz, dev_lim->max_srq_sz);
+
+	mthca_dbg(dev, "Flags: %08x\n", dev_lim->flags);
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+static void get_board_id(void *vsd, char *board_id)
+{
+	int i;
+
+#define VSD_OFFSET_SIG1		0x00
+#define VSD_OFFSET_SIG2		0xde
+#define VSD_OFFSET_MLX_BOARD_ID	0xd0
+#define VSD_OFFSET_TS_BOARD_ID	0x20
+
+#define VSD_SIGNATURE_TOPSPIN	0x5ad
+
+	memset(board_id, 0, MTHCA_BOARD_ID_LEN);
+
+	if (be16_to_cpup(vsd + VSD_OFFSET_SIG1) == VSD_SIGNATURE_TOPSPIN &&
+	    be16_to_cpup(vsd + VSD_OFFSET_SIG2) == VSD_SIGNATURE_TOPSPIN) {
+		strscpy(board_id, vsd + VSD_OFFSET_TS_BOARD_ID, MTHCA_BOARD_ID_LEN);
+	} else {
+		/*
+		 * The board ID is a string but the firmware byte
+		 * swaps each 4-byte word before passing it back to
+		 * us.  Therefore we need to swab it before printing.
+		 */
+		for (i = 0; i < 4; ++i)
+			((u32 *) board_id)[i] =
+				swab32(*(u32 *) (vsd + VSD_OFFSET_MLX_BOARD_ID + i * 4));
+	}
+}
+
+int mthca_QUERY_ADAPTER(struct mthca_dev *dev,
+			struct mthca_adapter *adapter)
+{
+	struct mthca_mailbox *mailbox;
+	u32 *outbox;
+	int err;
+
+#define QUERY_ADAPTER_OUT_SIZE             0x100
+#define QUERY_ADAPTER_VENDOR_ID_OFFSET     0x00
+#define QUERY_ADAPTER_DEVICE_ID_OFFSET     0x04
+#define QUERY_ADAPTER_REVISION_ID_OFFSET   0x08
+#define QUERY_ADAPTER_INTA_PIN_OFFSET      0x10
+#define QUERY_ADAPTER_VSD_OFFSET           0x20
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
+
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_QUERY_ADAPTER,
+			    CMD_TIME_CLASS_A);
+
+	if (err)
+		goto out;
+
+	if (!mthca_is_memfree(dev)) {
+		MTHCA_GET(adapter->vendor_id, outbox,
+			  QUERY_ADAPTER_VENDOR_ID_OFFSET);
+		MTHCA_GET(adapter->device_id, outbox,
+			  QUERY_ADAPTER_DEVICE_ID_OFFSET);
+		MTHCA_GET(adapter->revision_id, outbox,
+			  QUERY_ADAPTER_REVISION_ID_OFFSET);
+	}
+	MTHCA_GET(adapter->inta_pin, outbox,    QUERY_ADAPTER_INTA_PIN_OFFSET);
+
+	get_board_id(outbox + QUERY_ADAPTER_VSD_OFFSET / 4,
+		     adapter->board_id);
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_INIT_HCA(struct mthca_dev *dev,
+		   struct mthca_init_hca_param *param)
+{
+	struct mthca_mailbox *mailbox;
+	__be32 *inbox;
+	int err;
+
+#define INIT_HCA_IN_SIZE             	 0x200
+#define INIT_HCA_FLAGS1_OFFSET           0x00c
+#define INIT_HCA_FLAGS2_OFFSET           0x014
+#define INIT_HCA_QPC_OFFSET          	 0x020
+#define  INIT_HCA_QPC_BASE_OFFSET    	 (INIT_HCA_QPC_OFFSET + 0x10)
+#define  INIT_HCA_LOG_QP_OFFSET      	 (INIT_HCA_QPC_OFFSET + 0x17)
+#define  INIT_HCA_EEC_BASE_OFFSET    	 (INIT_HCA_QPC_OFFSET + 0x20)
+#define  INIT_HCA_LOG_EEC_OFFSET     	 (INIT_HCA_QPC_OFFSET + 0x27)
+#define  INIT_HCA_SRQC_BASE_OFFSET   	 (INIT_HCA_QPC_OFFSET + 0x28)
+#define  INIT_HCA_LOG_SRQ_OFFSET     	 (INIT_HCA_QPC_OFFSET + 0x2f)
+#define  INIT_HCA_CQC_BASE_OFFSET    	 (INIT_HCA_QPC_OFFSET + 0x30)
+#define  INIT_HCA_LOG_CQ_OFFSET      	 (INIT_HCA_QPC_OFFSET + 0x37)
+#define  INIT_HCA_EQPC_BASE_OFFSET   	 (INIT_HCA_QPC_OFFSET + 0x40)
+#define  INIT_HCA_EEEC_BASE_OFFSET   	 (INIT_HCA_QPC_OFFSET + 0x50)
+#define  INIT_HCA_EQC_BASE_OFFSET    	 (INIT_HCA_QPC_OFFSET + 0x60)
+#define  INIT_HCA_LOG_EQ_OFFSET      	 (INIT_HCA_QPC_OFFSET + 0x67)
+#define  INIT_HCA_RDB_BASE_OFFSET    	 (INIT_HCA_QPC_OFFSET + 0x70)
+#define INIT_HCA_UDAV_OFFSET         	 0x0b0
+#define  INIT_HCA_UDAV_LKEY_OFFSET   	 (INIT_HCA_UDAV_OFFSET + 0x0)
+#define  INIT_HCA_UDAV_PD_OFFSET     	 (INIT_HCA_UDAV_OFFSET + 0x4)
+#define INIT_HCA_MCAST_OFFSET        	 0x0c0
+#define  INIT_HCA_MC_BASE_OFFSET         (INIT_HCA_MCAST_OFFSET + 0x00)
+#define  INIT_HCA_LOG_MC_ENTRY_SZ_OFFSET (INIT_HCA_MCAST_OFFSET + 0x12)
+#define  INIT_HCA_MC_HASH_SZ_OFFSET      (INIT_HCA_MCAST_OFFSET + 0x16)
+#define  INIT_HCA_LOG_MC_TABLE_SZ_OFFSET (INIT_HCA_MCAST_OFFSET + 0x1b)
+#define INIT_HCA_TPT_OFFSET              0x0f0
+#define  INIT_HCA_MPT_BASE_OFFSET        (INIT_HCA_TPT_OFFSET + 0x00)
+#define  INIT_HCA_MTT_SEG_SZ_OFFSET      (INIT_HCA_TPT_OFFSET + 0x09)
+#define  INIT_HCA_LOG_MPT_SZ_OFFSET      (INIT_HCA_TPT_OFFSET + 0x0b)
+#define  INIT_HCA_MTT_BASE_OFFSET        (INIT_HCA_TPT_OFFSET + 0x10)
+#define INIT_HCA_UAR_OFFSET              0x120
+#define  INIT_HCA_UAR_BASE_OFFSET        (INIT_HCA_UAR_OFFSET + 0x00)
+#define  INIT_HCA_UARC_SZ_OFFSET         (INIT_HCA_UAR_OFFSET + 0x09)
+#define  INIT_HCA_LOG_UAR_SZ_OFFSET      (INIT_HCA_UAR_OFFSET + 0x0a)
+#define  INIT_HCA_UAR_PAGE_SZ_OFFSET     (INIT_HCA_UAR_OFFSET + 0x0b)
+#define  INIT_HCA_UAR_SCATCH_BASE_OFFSET (INIT_HCA_UAR_OFFSET + 0x10)
+#define  INIT_HCA_UAR_CTX_BASE_OFFSET    (INIT_HCA_UAR_OFFSET + 0x18)
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	inbox = mailbox->buf;
+
+	memset(inbox, 0, INIT_HCA_IN_SIZE);
+
+	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
+		MTHCA_PUT(inbox, 0x1, INIT_HCA_FLAGS1_OFFSET);
+
+#if defined(__LITTLE_ENDIAN)
+	*(inbox + INIT_HCA_FLAGS2_OFFSET / 4) &= ~cpu_to_be32(1 << 1);
+#elif defined(__BIG_ENDIAN)
+	*(inbox + INIT_HCA_FLAGS2_OFFSET / 4) |= cpu_to_be32(1 << 1);
+#else
+#error Host endianness not defined
+#endif
+	/* Check port for UD address vector: */
+	*(inbox + INIT_HCA_FLAGS2_OFFSET / 4) |= cpu_to_be32(1);
+
+	/* Enable IPoIB checksumming if we can: */
+	if (dev->device_cap_flags & IB_DEVICE_UD_IP_CSUM)
+		*(inbox + INIT_HCA_FLAGS2_OFFSET / 4) |= cpu_to_be32(7 << 3);
+
+	/* We leave wqe_quota, responder_exu, etc as 0 (default) */
+
+	/* QPC/EEC/CQC/EQC/RDB attributes */
+
+	MTHCA_PUT(inbox, param->qpc_base,     INIT_HCA_QPC_BASE_OFFSET);
+	MTHCA_PUT(inbox, param->log_num_qps,  INIT_HCA_LOG_QP_OFFSET);
+	MTHCA_PUT(inbox, param->eec_base,     INIT_HCA_EEC_BASE_OFFSET);
+	MTHCA_PUT(inbox, param->log_num_eecs, INIT_HCA_LOG_EEC_OFFSET);
+	MTHCA_PUT(inbox, param->srqc_base,    INIT_HCA_SRQC_BASE_OFFSET);
+	MTHCA_PUT(inbox, param->log_num_srqs, INIT_HCA_LOG_SRQ_OFFSET);
+	MTHCA_PUT(inbox, param->cqc_base,     INIT_HCA_CQC_BASE_OFFSET);
+	MTHCA_PUT(inbox, param->log_num_cqs,  INIT_HCA_LOG_CQ_OFFSET);
+	MTHCA_PUT(inbox, param->eqpc_base,    INIT_HCA_EQPC_BASE_OFFSET);
+	MTHCA_PUT(inbox, param->eeec_base,    INIT_HCA_EEEC_BASE_OFFSET);
+	MTHCA_PUT(inbox, param->eqc_base,     INIT_HCA_EQC_BASE_OFFSET);
+	MTHCA_PUT(inbox, param->log_num_eqs,  INIT_HCA_LOG_EQ_OFFSET);
+	MTHCA_PUT(inbox, param->rdb_base,     INIT_HCA_RDB_BASE_OFFSET);
+
+	/* UD AV attributes */
+
+	/* multicast attributes */
+
+	MTHCA_PUT(inbox, param->mc_base,         INIT_HCA_MC_BASE_OFFSET);
+	MTHCA_PUT(inbox, param->log_mc_entry_sz, INIT_HCA_LOG_MC_ENTRY_SZ_OFFSET);
+	MTHCA_PUT(inbox, param->mc_hash_sz,      INIT_HCA_MC_HASH_SZ_OFFSET);
+	MTHCA_PUT(inbox, param->log_mc_table_sz, INIT_HCA_LOG_MC_TABLE_SZ_OFFSET);
+
+	/* TPT attributes */
+
+	MTHCA_PUT(inbox, param->mpt_base,   INIT_HCA_MPT_BASE_OFFSET);
+	if (!mthca_is_memfree(dev))
+		MTHCA_PUT(inbox, param->mtt_seg_sz, INIT_HCA_MTT_SEG_SZ_OFFSET);
+	MTHCA_PUT(inbox, param->log_mpt_sz, INIT_HCA_LOG_MPT_SZ_OFFSET);
+	MTHCA_PUT(inbox, param->mtt_base,   INIT_HCA_MTT_BASE_OFFSET);
+
+	/* UAR attributes */
+	{
+		u8 uar_page_sz = PAGE_SHIFT - 12;
+		MTHCA_PUT(inbox, uar_page_sz, INIT_HCA_UAR_PAGE_SZ_OFFSET);
+	}
+
+	MTHCA_PUT(inbox, param->uar_scratch_base, INIT_HCA_UAR_SCATCH_BASE_OFFSET);
+
+	if (mthca_is_memfree(dev)) {
+		MTHCA_PUT(inbox, param->log_uarc_sz, INIT_HCA_UARC_SZ_OFFSET);
+		MTHCA_PUT(inbox, param->log_uar_sz,  INIT_HCA_LOG_UAR_SZ_OFFSET);
+		MTHCA_PUT(inbox, param->uarc_base,   INIT_HCA_UAR_CTX_BASE_OFFSET);
+	}
+
+	err = mthca_cmd(dev, mailbox->dma, 0, 0,
+			CMD_INIT_HCA, CMD_TIME_CLASS_D);
+
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_INIT_IB(struct mthca_dev *dev,
+		  struct mthca_init_ib_param *param,
+		  int port)
+{
+	struct mthca_mailbox *mailbox;
+	u32 *inbox;
+	int err;
+	u32 flags;
+
+#define INIT_IB_IN_SIZE          56
+#define INIT_IB_FLAGS_OFFSET     0x00
+#define INIT_IB_FLAG_SIG         (1 << 18)
+#define INIT_IB_FLAG_NG          (1 << 17)
+#define INIT_IB_FLAG_G0          (1 << 16)
+#define INIT_IB_VL_SHIFT         4
+#define INIT_IB_PORT_WIDTH_SHIFT 8
+#define INIT_IB_MTU_SHIFT        12
+#define INIT_IB_MAX_GID_OFFSET   0x06
+#define INIT_IB_MAX_PKEY_OFFSET  0x0a
+#define INIT_IB_GUID0_OFFSET     0x10
+#define INIT_IB_NODE_GUID_OFFSET 0x18
+#define INIT_IB_SI_GUID_OFFSET   0x20
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	inbox = mailbox->buf;
+
+	memset(inbox, 0, INIT_IB_IN_SIZE);
+
+	flags = 0;
+	flags |= param->set_guid0     ? INIT_IB_FLAG_G0  : 0;
+	flags |= param->set_node_guid ? INIT_IB_FLAG_NG  : 0;
+	flags |= param->set_si_guid   ? INIT_IB_FLAG_SIG : 0;
+	flags |= param->vl_cap << INIT_IB_VL_SHIFT;
+	flags |= param->port_width << INIT_IB_PORT_WIDTH_SHIFT;
+	flags |= param->mtu_cap << INIT_IB_MTU_SHIFT;
+	MTHCA_PUT(inbox, flags, INIT_IB_FLAGS_OFFSET);
+
+	MTHCA_PUT(inbox, param->gid_cap,   INIT_IB_MAX_GID_OFFSET);
+	MTHCA_PUT(inbox, param->pkey_cap,  INIT_IB_MAX_PKEY_OFFSET);
+	MTHCA_PUT(inbox, param->guid0,     INIT_IB_GUID0_OFFSET);
+	MTHCA_PUT(inbox, param->node_guid, INIT_IB_NODE_GUID_OFFSET);
+	MTHCA_PUT(inbox, param->si_guid,   INIT_IB_SI_GUID_OFFSET);
+
+	err = mthca_cmd(dev, mailbox->dma, port, 0, CMD_INIT_IB,
+			CMD_TIME_CLASS_A);
+
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_CLOSE_IB(struct mthca_dev *dev, int port)
+{
+	return mthca_cmd(dev, 0, port, 0, CMD_CLOSE_IB, CMD_TIME_CLASS_A);
+}
+
+int mthca_CLOSE_HCA(struct mthca_dev *dev, int panic)
+{
+	return mthca_cmd(dev, 0, 0, panic, CMD_CLOSE_HCA, CMD_TIME_CLASS_C);
+}
+
+int mthca_SET_IB(struct mthca_dev *dev, struct mthca_set_ib_param *param,
+		 int port)
+{
+	struct mthca_mailbox *mailbox;
+	u32 *inbox;
+	int err;
+	u32 flags = 0;
+
+#define SET_IB_IN_SIZE         0x40
+#define SET_IB_FLAGS_OFFSET    0x00
+#define SET_IB_FLAG_SIG        (1 << 18)
+#define SET_IB_FLAG_RQK        (1 <<  0)
+#define SET_IB_CAP_MASK_OFFSET 0x04
+#define SET_IB_SI_GUID_OFFSET  0x08
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	inbox = mailbox->buf;
+
+	memset(inbox, 0, SET_IB_IN_SIZE);
+
+	flags |= param->set_si_guid     ? SET_IB_FLAG_SIG : 0;
+	flags |= param->reset_qkey_viol ? SET_IB_FLAG_RQK : 0;
+	MTHCA_PUT(inbox, flags, SET_IB_FLAGS_OFFSET);
+
+	MTHCA_PUT(inbox, param->cap_mask, SET_IB_CAP_MASK_OFFSET);
+	MTHCA_PUT(inbox, param->si_guid,  SET_IB_SI_GUID_OFFSET);
+
+	err = mthca_cmd(dev, mailbox->dma, port, 0, CMD_SET_IB,
+			CMD_TIME_CLASS_B);
+
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_MAP_ICM(struct mthca_dev *dev, struct mthca_icm *icm, u64 virt)
+{
+	return mthca_map_cmd(dev, CMD_MAP_ICM, icm, virt);
+}
+
+int mthca_MAP_ICM_page(struct mthca_dev *dev, u64 dma_addr, u64 virt)
+{
+	struct mthca_mailbox *mailbox;
+	__be64 *inbox;
+	int err;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	inbox = mailbox->buf;
+
+	inbox[0] = cpu_to_be64(virt);
+	inbox[1] = cpu_to_be64(dma_addr);
+
+	err = mthca_cmd(dev, mailbox->dma, 1, 0, CMD_MAP_ICM,
+			CMD_TIME_CLASS_B);
+
+	mthca_free_mailbox(dev, mailbox);
+
+	if (!err)
+		mthca_dbg(dev, "Mapped page at %llx to %llx for ICM.\n",
+			  (unsigned long long) dma_addr, (unsigned long long) virt);
+
+	return err;
+}
+
+int mthca_UNMAP_ICM(struct mthca_dev *dev, u64 virt, u32 page_count)
+{
+	mthca_dbg(dev, "Unmapping %d pages at %llx from ICM.\n",
+		  page_count, (unsigned long long) virt);
+
+	return mthca_cmd(dev, virt, page_count, 0,
+			CMD_UNMAP_ICM, CMD_TIME_CLASS_B);
+}
+
+int mthca_MAP_ICM_AUX(struct mthca_dev *dev, struct mthca_icm *icm)
+{
+	return mthca_map_cmd(dev, CMD_MAP_ICM_AUX, icm, -1);
+}
+
+int mthca_UNMAP_ICM_AUX(struct mthca_dev *dev)
+{
+	return mthca_cmd(dev, 0, 0, 0, CMD_UNMAP_ICM_AUX, CMD_TIME_CLASS_B);
+}
+
+int mthca_SET_ICM_SIZE(struct mthca_dev *dev, u64 icm_size, u64 *aux_pages)
+{
+	int ret = mthca_cmd_imm(dev, icm_size, aux_pages, 0,
+			0, CMD_SET_ICM_SIZE, CMD_TIME_CLASS_A);
+
+	if (ret)
+		return ret;
+
+	/*
+	 * Round up number of system pages needed in case
+	 * MTHCA_ICM_PAGE_SIZE < PAGE_SIZE.
+	 */
+	*aux_pages = ALIGN(*aux_pages, PAGE_SIZE / MTHCA_ICM_PAGE_SIZE) >>
+		(PAGE_SHIFT - MTHCA_ICM_PAGE_SHIFT);
+
+	return 0;
+}
+
+int mthca_SW2HW_MPT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int mpt_index)
+{
+	return mthca_cmd(dev, mailbox->dma, mpt_index, 0, CMD_SW2HW_MPT,
+			 CMD_TIME_CLASS_B);
+}
+
+int mthca_HW2SW_MPT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int mpt_index)
+{
+	return mthca_cmd_box(dev, 0, mailbox ? mailbox->dma : 0, mpt_index,
+			     !mailbox, CMD_HW2SW_MPT,
+			     CMD_TIME_CLASS_B);
+}
+
+int mthca_WRITE_MTT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int num_mtt)
+{
+	return mthca_cmd(dev, mailbox->dma, num_mtt, 0, CMD_WRITE_MTT,
+			 CMD_TIME_CLASS_B);
+}
+
+int mthca_SYNC_TPT(struct mthca_dev *dev)
+{
+	return mthca_cmd(dev, 0, 0, 0, CMD_SYNC_TPT, CMD_TIME_CLASS_B);
+}
+
+int mthca_MAP_EQ(struct mthca_dev *dev, u64 event_mask, int unmap,
+		 int eq_num)
+{
+	mthca_dbg(dev, "%s mask %016llx for eqn %d\n",
+		  unmap ? "Clearing" : "Setting",
+		  (unsigned long long) event_mask, eq_num);
+	return mthca_cmd(dev, event_mask, (unmap << 31) | eq_num,
+			 0, CMD_MAP_EQ, CMD_TIME_CLASS_B);
+}
+
+int mthca_SW2HW_EQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		   int eq_num)
+{
+	return mthca_cmd(dev, mailbox->dma, eq_num, 0, CMD_SW2HW_EQ,
+			 CMD_TIME_CLASS_A);
+}
+
+int mthca_HW2SW_EQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		   int eq_num)
+{
+	return mthca_cmd_box(dev, 0, mailbox->dma, eq_num, 0,
+			     CMD_HW2SW_EQ,
+			     CMD_TIME_CLASS_A);
+}
+
+int mthca_SW2HW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		   int cq_num)
+{
+	return mthca_cmd(dev, mailbox->dma, cq_num, 0, CMD_SW2HW_CQ,
+			CMD_TIME_CLASS_A);
+}
+
+int mthca_HW2SW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		   int cq_num)
+{
+	return mthca_cmd_box(dev, 0, mailbox->dma, cq_num, 0,
+			     CMD_HW2SW_CQ,
+			     CMD_TIME_CLASS_A);
+}
+
+int mthca_RESIZE_CQ(struct mthca_dev *dev, int cq_num, u32 lkey, u8 log_size)
+{
+	struct mthca_mailbox *mailbox;
+	__be32 *inbox;
+	int err;
+
+#define RESIZE_CQ_IN_SIZE		0x40
+#define RESIZE_CQ_LOG_SIZE_OFFSET	0x0c
+#define RESIZE_CQ_LKEY_OFFSET		0x1c
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	inbox = mailbox->buf;
+
+	memset(inbox, 0, RESIZE_CQ_IN_SIZE);
+	/*
+	 * Leave start address fields zeroed out -- mthca assumes that
+	 * MRs for CQs always start at virtual address 0.
+	 */
+	MTHCA_PUT(inbox, log_size, RESIZE_CQ_LOG_SIZE_OFFSET);
+	MTHCA_PUT(inbox, lkey,     RESIZE_CQ_LKEY_OFFSET);
+
+	err = mthca_cmd(dev, mailbox->dma, cq_num, 1, CMD_RESIZE_CQ,
+			CMD_TIME_CLASS_B);
+
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_SW2HW_SRQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int srq_num)
+{
+	return mthca_cmd(dev, mailbox->dma, srq_num, 0, CMD_SW2HW_SRQ,
+			CMD_TIME_CLASS_A);
+}
+
+int mthca_HW2SW_SRQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int srq_num)
+{
+	return mthca_cmd_box(dev, 0, mailbox->dma, srq_num, 0,
+			     CMD_HW2SW_SRQ,
+			     CMD_TIME_CLASS_A);
+}
+
+int mthca_QUERY_SRQ(struct mthca_dev *dev, u32 num,
+		    struct mthca_mailbox *mailbox)
+{
+	return mthca_cmd_box(dev, 0, mailbox->dma, num, 0,
+			     CMD_QUERY_SRQ, CMD_TIME_CLASS_A);
+}
+
+int mthca_ARM_SRQ(struct mthca_dev *dev, int srq_num, int limit)
+{
+	return mthca_cmd(dev, limit, srq_num, 0, CMD_ARM_SRQ,
+			 CMD_TIME_CLASS_B);
+}
+
+int mthca_MODIFY_QP(struct mthca_dev *dev, enum ib_qp_state cur,
+		    enum ib_qp_state next, u32 num, int is_ee,
+		    struct mthca_mailbox *mailbox, u32 optmask)
+{
+	static const u16 op[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
+		[IB_QPS_RESET] = {
+			[IB_QPS_RESET]	= CMD_ERR2RST_QPEE,
+			[IB_QPS_ERR]	= CMD_2ERR_QPEE,
+			[IB_QPS_INIT]	= CMD_RST2INIT_QPEE,
+		},
+		[IB_QPS_INIT]  = {
+			[IB_QPS_RESET]	= CMD_ERR2RST_QPEE,
+			[IB_QPS_ERR]	= CMD_2ERR_QPEE,
+			[IB_QPS_INIT]	= CMD_INIT2INIT_QPEE,
+			[IB_QPS_RTR]	= CMD_INIT2RTR_QPEE,
+		},
+		[IB_QPS_RTR]   = {
+			[IB_QPS_RESET]	= CMD_ERR2RST_QPEE,
+			[IB_QPS_ERR]	= CMD_2ERR_QPEE,
+			[IB_QPS_RTS]	= CMD_RTR2RTS_QPEE,
+		},
+		[IB_QPS_RTS]   = {
+			[IB_QPS_RESET]	= CMD_ERR2RST_QPEE,
+			[IB_QPS_ERR]	= CMD_2ERR_QPEE,
+			[IB_QPS_RTS]	= CMD_RTS2RTS_QPEE,
+			[IB_QPS_SQD]	= CMD_RTS2SQD_QPEE,
+		},
+		[IB_QPS_SQD] = {
+			[IB_QPS_RESET]	= CMD_ERR2RST_QPEE,
+			[IB_QPS_ERR]	= CMD_2ERR_QPEE,
+			[IB_QPS_RTS]	= CMD_SQD2RTS_QPEE,
+			[IB_QPS_SQD]	= CMD_SQD2SQD_QPEE,
+		},
+		[IB_QPS_SQE] = {
+			[IB_QPS_RESET]	= CMD_ERR2RST_QPEE,
+			[IB_QPS_ERR]	= CMD_2ERR_QPEE,
+			[IB_QPS_RTS]	= CMD_SQERR2RTS_QPEE,
+		},
+		[IB_QPS_ERR] = {
+			[IB_QPS_RESET]	= CMD_ERR2RST_QPEE,
+			[IB_QPS_ERR]	= CMD_2ERR_QPEE,
+		}
+	};
+
+	u8 op_mod = 0;
+	int my_mailbox = 0;
+	int err;
+
+	if (op[cur][next] == CMD_ERR2RST_QPEE) {
+		op_mod = 3;	/* don't write outbox, any->reset */
+
+		/* For debugging */
+		if (!mailbox) {
+			mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+			if (!IS_ERR(mailbox)) {
+				my_mailbox = 1;
+				op_mod     = 2;	/* write outbox, any->reset */
+			} else
+				mailbox = NULL;
+		}
+
+		err = mthca_cmd_box(dev, 0, mailbox ? mailbox->dma : 0,
+				    (!!is_ee << 24) | num, op_mod,
+				    op[cur][next], CMD_TIME_CLASS_C);
+
+		if (0 && mailbox) {
+			int i;
+			mthca_dbg(dev, "Dumping QP context:\n");
+			printk(" %08x\n", be32_to_cpup(mailbox->buf));
+			for (i = 0; i < 0x100 / 4; ++i) {
+				if (i % 8 == 0)
+					printk("[%02x] ", i * 4);
+				printk(" %08x",
+				       be32_to_cpu(((__be32 *) mailbox->buf)[i + 2]));
+				if ((i + 1) % 8 == 0)
+					printk("\n");
+			}
+		}
+
+		if (my_mailbox)
+			mthca_free_mailbox(dev, mailbox);
+	} else {
+		if (0) {
+			int i;
+			mthca_dbg(dev, "Dumping QP context:\n");
+			printk("  opt param mask: %08x\n", be32_to_cpup(mailbox->buf));
+			for (i = 0; i < 0x100 / 4; ++i) {
+				if (i % 8 == 0)
+					printk("  [%02x] ", i * 4);
+				printk(" %08x",
+				       be32_to_cpu(((__be32 *) mailbox->buf)[i + 2]));
+				if ((i + 1) % 8 == 0)
+					printk("\n");
+			}
+		}
+
+		err = mthca_cmd(dev, mailbox->dma, optmask | (!!is_ee << 24) | num,
+				op_mod, op[cur][next], CMD_TIME_CLASS_C);
+	}
+
+	return err;
+}
+
+int mthca_QUERY_QP(struct mthca_dev *dev, u32 num, int is_ee,
+		   struct mthca_mailbox *mailbox)
+{
+	return mthca_cmd_box(dev, 0, mailbox->dma, (!!is_ee << 24) | num, 0,
+			     CMD_QUERY_QPEE, CMD_TIME_CLASS_A);
+}
+
+int mthca_CONF_SPECIAL_QP(struct mthca_dev *dev, int type, u32 qpn)
+{
+	u8 op_mod;
+
+	switch (type) {
+	case IB_QPT_SMI:
+		op_mod = 0;
+		break;
+	case IB_QPT_GSI:
+		op_mod = 1;
+		break;
+	case IB_QPT_RAW_IPV6:
+		op_mod = 2;
+		break;
+	case IB_QPT_RAW_ETHERTYPE:
+		op_mod = 3;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	return mthca_cmd(dev, 0, qpn, op_mod, CMD_CONF_SPECIAL_QP,
+			 CMD_TIME_CLASS_B);
+}
+
+int mthca_MAD_IFC(struct mthca_dev *dev, int ignore_mkey, int ignore_bkey,
+		  int port, const struct ib_wc *in_wc, const struct ib_grh *in_grh,
+		  const void *in_mad, void *response_mad)
+{
+	struct mthca_mailbox *inmailbox, *outmailbox;
+	void *inbox;
+	int err;
+	u32 in_modifier = port;
+	u8 op_modifier = 0;
+
+#define MAD_IFC_BOX_SIZE      0x400
+#define MAD_IFC_MY_QPN_OFFSET 0x100
+#define MAD_IFC_RQPN_OFFSET   0x108
+#define MAD_IFC_SL_OFFSET     0x10c
+#define MAD_IFC_G_PATH_OFFSET 0x10d
+#define MAD_IFC_RLID_OFFSET   0x10e
+#define MAD_IFC_PKEY_OFFSET   0x112
+#define MAD_IFC_GRH_OFFSET    0x140
+
+	inmailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(inmailbox))
+		return PTR_ERR(inmailbox);
+	inbox = inmailbox->buf;
+
+	outmailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(outmailbox)) {
+		mthca_free_mailbox(dev, inmailbox);
+		return PTR_ERR(outmailbox);
+	}
+
+	memcpy(inbox, in_mad, 256);
+
+	/*
+	 * Key check traps can't be generated unless we have in_wc to
+	 * tell us where to send the trap.
+	 */
+	if (ignore_mkey || !in_wc)
+		op_modifier |= 0x1;
+	if (ignore_bkey || !in_wc)
+		op_modifier |= 0x2;
+
+	if (in_wc) {
+		u8 val;
+
+		memset(inbox + 256, 0, 256);
+
+		MTHCA_PUT(inbox, in_wc->qp->qp_num, MAD_IFC_MY_QPN_OFFSET);
+		MTHCA_PUT(inbox, in_wc->src_qp,     MAD_IFC_RQPN_OFFSET);
+
+		val = in_wc->sl << 4;
+		MTHCA_PUT(inbox, val,               MAD_IFC_SL_OFFSET);
+
+		val = in_wc->dlid_path_bits |
+			(in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
+		MTHCA_PUT(inbox, val,               MAD_IFC_G_PATH_OFFSET);
+
+		MTHCA_PUT(inbox, ib_lid_cpu16(in_wc->slid), MAD_IFC_RLID_OFFSET);
+		MTHCA_PUT(inbox, in_wc->pkey_index, MAD_IFC_PKEY_OFFSET);
+
+		if (in_grh)
+			memcpy(inbox + MAD_IFC_GRH_OFFSET, in_grh, 40);
+
+		op_modifier |= 0x4;
+
+		in_modifier |= ib_lid_cpu16(in_wc->slid) << 16;
+	}
+
+	err = mthca_cmd_box(dev, inmailbox->dma, outmailbox->dma,
+			    in_modifier, op_modifier,
+			    CMD_MAD_IFC, CMD_TIME_CLASS_C);
+
+	if (!err)
+		memcpy(response_mad, outmailbox->buf, 256);
+
+	mthca_free_mailbox(dev, inmailbox);
+	mthca_free_mailbox(dev, outmailbox);
+	return err;
+}
+
+int mthca_READ_MGM(struct mthca_dev *dev, int index,
+		   struct mthca_mailbox *mailbox)
+{
+	return mthca_cmd_box(dev, 0, mailbox->dma, index, 0,
+			     CMD_READ_MGM, CMD_TIME_CLASS_A);
+}
+
+int mthca_WRITE_MGM(struct mthca_dev *dev, int index,
+		    struct mthca_mailbox *mailbox)
+{
+	return mthca_cmd(dev, mailbox->dma, index, 0, CMD_WRITE_MGM,
+			 CMD_TIME_CLASS_A);
+}
+
+int mthca_MGID_HASH(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    u16 *hash)
+{
+	u64 imm = 0;
+	int err;
+
+	err = mthca_cmd_imm(dev, mailbox->dma, &imm, 0, 0, CMD_MGID_HASH,
+			    CMD_TIME_CLASS_A);
+
+	*hash = imm;
+	return err;
+}
+
+int mthca_NOP(struct mthca_dev *dev)
+{
+	return mthca_cmd(dev, 0, 0x1f, 0, CMD_NOP, msecs_to_jiffies(100));
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_cmd.h b/drivers/infiniband/hw/mthca/mthca_cmd.h
new file mode 100644
index 000000000..d2e5b194b
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_cmd.h
@@ -0,0 +1,325 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2006 Cisco Systems.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef MTHCA_CMD_H
+#define MTHCA_CMD_H
+
+#include <rdma/ib_verbs.h>
+
+#define MTHCA_MAILBOX_SIZE 4096
+
+enum {
+	/* command completed successfully: */
+	MTHCA_CMD_STAT_OK 	      = 0x00,
+	/* Internal error (such as a bus error) occurred while processing command: */
+	MTHCA_CMD_STAT_INTERNAL_ERR   = 0x01,
+	/* Operation/command not supported or opcode modifier not supported: */
+	MTHCA_CMD_STAT_BAD_OP 	      = 0x02,
+	/* Parameter not supported or parameter out of range: */
+	MTHCA_CMD_STAT_BAD_PARAM      = 0x03,
+	/* System not enabled or bad system state: */
+	MTHCA_CMD_STAT_BAD_SYS_STATE  = 0x04,
+	/* Attempt to access reserved or unallocaterd resource: */
+	MTHCA_CMD_STAT_BAD_RESOURCE   = 0x05,
+	/* Requested resource is currently executing a command, or is otherwise busy: */
+	MTHCA_CMD_STAT_RESOURCE_BUSY  = 0x06,
+	/* memory error: */
+	MTHCA_CMD_STAT_DDR_MEM_ERR    = 0x07,
+	/* Required capability exceeds device limits: */
+	MTHCA_CMD_STAT_EXCEED_LIM     = 0x08,
+	/* Resource is not in the appropriate state or ownership: */
+	MTHCA_CMD_STAT_BAD_RES_STATE  = 0x09,
+	/* Index out of range: */
+	MTHCA_CMD_STAT_BAD_INDEX      = 0x0a,
+	/* FW image corrupted: */
+	MTHCA_CMD_STAT_BAD_NVMEM      = 0x0b,
+	/* Attempt to modify a QP/EE which is not in the presumed state: */
+	MTHCA_CMD_STAT_BAD_QPEE_STATE = 0x10,
+	/* Bad segment parameters (Address/Size): */
+	MTHCA_CMD_STAT_BAD_SEG_PARAM  = 0x20,
+	/* Memory Region has Memory Windows bound to: */
+	MTHCA_CMD_STAT_REG_BOUND      = 0x21,
+	/* HCA local attached memory not present: */
+	MTHCA_CMD_STAT_LAM_NOT_PRE    = 0x22,
+	/* Bad management packet (silently discarded): */
+	MTHCA_CMD_STAT_BAD_PKT 	      = 0x30,
+	/* More outstanding CQEs in CQ than new CQ size: */
+	MTHCA_CMD_STAT_BAD_SIZE       = 0x40
+};
+
+enum {
+	MTHCA_TRANS_INVALID = 0,
+	MTHCA_TRANS_RST2INIT,
+	MTHCA_TRANS_INIT2INIT,
+	MTHCA_TRANS_INIT2RTR,
+	MTHCA_TRANS_RTR2RTS,
+	MTHCA_TRANS_RTS2RTS,
+	MTHCA_TRANS_SQERR2RTS,
+	MTHCA_TRANS_ANY2ERR,
+	MTHCA_TRANS_RTS2SQD,
+	MTHCA_TRANS_SQD2SQD,
+	MTHCA_TRANS_SQD2RTS,
+	MTHCA_TRANS_ANY2RST,
+};
+
+enum {
+	DEV_LIM_FLAG_RC                 = 1 << 0,
+	DEV_LIM_FLAG_UC                 = 1 << 1,
+	DEV_LIM_FLAG_UD                 = 1 << 2,
+	DEV_LIM_FLAG_RD                 = 1 << 3,
+	DEV_LIM_FLAG_RAW_IPV6           = 1 << 4,
+	DEV_LIM_FLAG_RAW_ETHER          = 1 << 5,
+	DEV_LIM_FLAG_SRQ                = 1 << 6,
+	DEV_LIM_FLAG_IPOIB_CSUM		= 1 << 7,
+	DEV_LIM_FLAG_BAD_PKEY_CNTR      = 1 << 8,
+	DEV_LIM_FLAG_BAD_QKEY_CNTR      = 1 << 9,
+	DEV_LIM_FLAG_MW                 = 1 << 16,
+	DEV_LIM_FLAG_AUTO_PATH_MIG      = 1 << 17,
+	DEV_LIM_FLAG_ATOMIC             = 1 << 18,
+	DEV_LIM_FLAG_RAW_MULTI          = 1 << 19,
+	DEV_LIM_FLAG_UD_AV_PORT_ENFORCE = 1 << 20,
+	DEV_LIM_FLAG_UD_MULTI           = 1 << 21,
+};
+
+struct mthca_mailbox {
+	dma_addr_t dma;
+	void      *buf;
+};
+
+struct mthca_dev_lim {
+	int max_srq_sz;
+	int max_qp_sz;
+	int reserved_qps;
+	int max_qps;
+	int reserved_srqs;
+	int max_srqs;
+	int reserved_eecs;
+	int max_eecs;
+	int max_cq_sz;
+	int reserved_cqs;
+	int max_cqs;
+	int max_mpts;
+	int reserved_eqs;
+	int max_eqs;
+	int reserved_mtts;
+	int max_mrw_sz;
+	int reserved_mrws;
+	int max_mtt_seg;
+	int max_requester_per_qp;
+	int max_responder_per_qp;
+	int max_rdma_global;
+	int local_ca_ack_delay;
+	int max_mtu;
+	int max_port_width;
+	int max_vl;
+	int num_ports;
+	int max_gids;
+	u16 stat_rate_support;
+	int max_pkeys;
+	u32 flags;
+	int reserved_uars;
+	int uar_size;
+	int min_page_sz;
+	int max_sg;
+	int max_desc_sz;
+	int max_qp_per_mcg;
+	int reserved_mgms;
+	int max_mcgs;
+	int reserved_pds;
+	int max_pds;
+	int reserved_rdds;
+	int max_rdds;
+	int eec_entry_sz;
+	int qpc_entry_sz;
+	int eeec_entry_sz;
+	int eqpc_entry_sz;
+	int eqc_entry_sz;
+	int cqc_entry_sz;
+	int srq_entry_sz;
+	int uar_scratch_entry_sz;
+	int mpt_entry_sz;
+	union {
+		struct {
+			int max_avs;
+		} tavor;
+		struct {
+			int resize_srq;
+			int max_pbl_sz;
+			u8  bmme_flags;
+			u32 reserved_lkey;
+			int lam_required;
+			u64 max_icm_sz;
+		} arbel;
+	} hca;
+};
+
+struct mthca_adapter {
+	u32  vendor_id;
+	u32  device_id;
+	u32  revision_id;
+	char board_id[MTHCA_BOARD_ID_LEN];
+	u8   inta_pin;
+};
+
+struct mthca_init_hca_param {
+	u64 qpc_base;
+	u64 eec_base;
+	u64 srqc_base;
+	u64 cqc_base;
+	u64 eqpc_base;
+	u64 eeec_base;
+	u64 eqc_base;
+	u64 rdb_base;
+	u64 mc_base;
+	u64 mpt_base;
+	u64 mtt_base;
+	u64 uar_scratch_base;
+	u64 uarc_base;
+	u16 log_mc_entry_sz;
+	u16 mc_hash_sz;
+	u8  log_num_qps;
+	u8  log_num_eecs;
+	u8  log_num_srqs;
+	u8  log_num_cqs;
+	u8  log_num_eqs;
+	u8  log_mc_table_sz;
+	u8  mtt_seg_sz;
+	u8  log_mpt_sz;
+	u8  log_uar_sz;
+	u8  log_uarc_sz;
+};
+
+struct mthca_init_ib_param {
+	int port_width;
+	int vl_cap;
+	int mtu_cap;
+	u16 gid_cap;
+	u16 pkey_cap;
+	int set_guid0;
+	u64 guid0;
+	int set_node_guid;
+	u64 node_guid;
+	int set_si_guid;
+	u64 si_guid;
+};
+
+struct mthca_set_ib_param {
+	int set_si_guid;
+	int reset_qkey_viol;
+	u64 si_guid;
+	u32 cap_mask;
+};
+
+int mthca_cmd_init(struct mthca_dev *dev);
+void mthca_cmd_cleanup(struct mthca_dev *dev);
+int mthca_cmd_use_events(struct mthca_dev *dev);
+void mthca_cmd_use_polling(struct mthca_dev *dev);
+void mthca_cmd_event(struct mthca_dev *dev, u16 token,
+		     u8  status, u64 out_param);
+
+struct mthca_mailbox *mthca_alloc_mailbox(struct mthca_dev *dev,
+					  gfp_t gfp_mask);
+void mthca_free_mailbox(struct mthca_dev *dev, struct mthca_mailbox *mailbox);
+
+int mthca_SYS_EN(struct mthca_dev *dev);
+int mthca_SYS_DIS(struct mthca_dev *dev);
+int mthca_MAP_FA(struct mthca_dev *dev, struct mthca_icm *icm);
+int mthca_UNMAP_FA(struct mthca_dev *dev);
+int mthca_RUN_FW(struct mthca_dev *dev);
+int mthca_QUERY_FW(struct mthca_dev *dev);
+int mthca_ENABLE_LAM(struct mthca_dev *dev);
+int mthca_DISABLE_LAM(struct mthca_dev *dev);
+int mthca_QUERY_DDR(struct mthca_dev *dev);
+int mthca_QUERY_DEV_LIM(struct mthca_dev *dev,
+			struct mthca_dev_lim *dev_lim);
+int mthca_QUERY_ADAPTER(struct mthca_dev *dev,
+			struct mthca_adapter *adapter);
+int mthca_INIT_HCA(struct mthca_dev *dev,
+		   struct mthca_init_hca_param *param);
+int mthca_INIT_IB(struct mthca_dev *dev,
+		  struct mthca_init_ib_param *param,
+		  int port);
+int mthca_CLOSE_IB(struct mthca_dev *dev, int port);
+int mthca_CLOSE_HCA(struct mthca_dev *dev, int panic);
+int mthca_SET_IB(struct mthca_dev *dev, struct mthca_set_ib_param *param,
+		 int port);
+int mthca_MAP_ICM(struct mthca_dev *dev, struct mthca_icm *icm, u64 virt);
+int mthca_MAP_ICM_page(struct mthca_dev *dev, u64 dma_addr, u64 virt);
+int mthca_UNMAP_ICM(struct mthca_dev *dev, u64 virt, u32 page_count);
+int mthca_MAP_ICM_AUX(struct mthca_dev *dev, struct mthca_icm *icm);
+int mthca_UNMAP_ICM_AUX(struct mthca_dev *dev);
+int mthca_SET_ICM_SIZE(struct mthca_dev *dev, u64 icm_size, u64 *aux_pages);
+int mthca_SW2HW_MPT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int mpt_index);
+int mthca_HW2SW_MPT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int mpt_index);
+int mthca_WRITE_MTT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int num_mtt);
+int mthca_SYNC_TPT(struct mthca_dev *dev);
+int mthca_MAP_EQ(struct mthca_dev *dev, u64 event_mask, int unmap,
+		 int eq_num);
+int mthca_SW2HW_EQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		   int eq_num);
+int mthca_HW2SW_EQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		   int eq_num);
+int mthca_SW2HW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		   int cq_num);
+int mthca_HW2SW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		   int cq_num);
+int mthca_RESIZE_CQ(struct mthca_dev *dev, int cq_num, u32 lkey, u8 log_size);
+int mthca_SW2HW_SRQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int srq_num);
+int mthca_HW2SW_SRQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    int srq_num);
+int mthca_QUERY_SRQ(struct mthca_dev *dev, u32 num,
+		    struct mthca_mailbox *mailbox);
+int mthca_ARM_SRQ(struct mthca_dev *dev, int srq_num, int limit);
+int mthca_MODIFY_QP(struct mthca_dev *dev, enum ib_qp_state cur,
+		    enum ib_qp_state next, u32 num, int is_ee,
+		    struct mthca_mailbox *mailbox, u32 optmask);
+int mthca_QUERY_QP(struct mthca_dev *dev, u32 num, int is_ee,
+		   struct mthca_mailbox *mailbox);
+int mthca_CONF_SPECIAL_QP(struct mthca_dev *dev, int type, u32 qpn);
+int mthca_MAD_IFC(struct mthca_dev *dev, int ignore_mkey, int ignore_bkey,
+		  int port, const struct ib_wc *in_wc, const struct ib_grh *in_grh,
+		  const void *in_mad, void *response_mad);
+int mthca_READ_MGM(struct mthca_dev *dev, int index,
+		   struct mthca_mailbox *mailbox);
+int mthca_WRITE_MGM(struct mthca_dev *dev, int index,
+		    struct mthca_mailbox *mailbox);
+int mthca_MGID_HASH(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    u16 *hash);
+int mthca_NOP(struct mthca_dev *dev);
+
+#endif /* MTHCA_CMD_H */
diff --git a/drivers/infiniband/hw/mthca/mthca_config_reg.h b/drivers/infiniband/hw/mthca/mthca_config_reg.h
new file mode 100644
index 000000000..155bc6639
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_config_reg.h
@@ -0,0 +1,48 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef MTHCA_CONFIG_REG_H
+#define MTHCA_CONFIG_REG_H
+
+#define MTHCA_HCR_BASE         0x80680
+#define MTHCA_HCR_SIZE         0x0001c
+#define MTHCA_ECR_BASE         0x80700
+#define MTHCA_ECR_SIZE         0x00008
+#define MTHCA_ECR_CLR_BASE     0x80708
+#define MTHCA_ECR_CLR_SIZE     0x00008
+#define MTHCA_MAP_ECR_SIZE     (MTHCA_ECR_SIZE + MTHCA_ECR_CLR_SIZE)
+#define MTHCA_CLR_INT_BASE     0xf00d8
+#define MTHCA_CLR_INT_SIZE     0x00008
+#define MTHCA_EQ_SET_CI_SIZE   (8 * 32)
+
+#endif /* MTHCA_CONFIG_REG_H */
diff --git a/drivers/infiniband/hw/mthca/mthca_cq.c b/drivers/infiniband/hw/mthca/mthca_cq.c
new file mode 100644
index 000000000..26c3408dc
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_cq.c
@@ -0,0 +1,978 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems, Inc. All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/gfp.h>
+#include <linux/hardirq.h>
+#include <linux/sched.h>
+
+#include <asm/io.h>
+
+#include <rdma/ib_pack.h>
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+#include "mthca_memfree.h"
+
+enum {
+	MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
+};
+
+enum {
+	MTHCA_CQ_ENTRY_SIZE = 0x20
+};
+
+enum {
+	MTHCA_ATOMIC_BYTE_LEN = 8
+};
+
+/*
+ * Must be packed because start is 64 bits but only aligned to 32 bits.
+ */
+struct mthca_cq_context {
+	__be32 flags;
+	__be64 start;
+	__be32 logsize_usrpage;
+	__be32 error_eqn;	/* Tavor only */
+	__be32 comp_eqn;
+	__be32 pd;
+	__be32 lkey;
+	__be32 last_notified_index;
+	__be32 solicit_producer_index;
+	__be32 consumer_index;
+	__be32 producer_index;
+	__be32 cqn;
+	__be32 ci_db;		/* Arbel only */
+	__be32 state_db;	/* Arbel only */
+	u32    reserved;
+} __packed;
+
+#define MTHCA_CQ_STATUS_OK          ( 0 << 28)
+#define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
+#define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
+#define MTHCA_CQ_FLAG_TR            ( 1 << 18)
+#define MTHCA_CQ_FLAG_OI            ( 1 << 17)
+#define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
+#define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
+#define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
+#define MTHCA_EQ_STATE_FIRED        (10 <<  8)
+
+enum {
+	MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
+};
+
+enum {
+	SYNDROME_LOCAL_LENGTH_ERR 	 = 0x01,
+	SYNDROME_LOCAL_QP_OP_ERR  	 = 0x02,
+	SYNDROME_LOCAL_EEC_OP_ERR 	 = 0x03,
+	SYNDROME_LOCAL_PROT_ERR   	 = 0x04,
+	SYNDROME_WR_FLUSH_ERR     	 = 0x05,
+	SYNDROME_MW_BIND_ERR      	 = 0x06,
+	SYNDROME_BAD_RESP_ERR     	 = 0x10,
+	SYNDROME_LOCAL_ACCESS_ERR 	 = 0x11,
+	SYNDROME_REMOTE_INVAL_REQ_ERR 	 = 0x12,
+	SYNDROME_REMOTE_ACCESS_ERR 	 = 0x13,
+	SYNDROME_REMOTE_OP_ERR     	 = 0x14,
+	SYNDROME_RETRY_EXC_ERR 		 = 0x15,
+	SYNDROME_RNR_RETRY_EXC_ERR 	 = 0x16,
+	SYNDROME_LOCAL_RDD_VIOL_ERR 	 = 0x20,
+	SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
+	SYNDROME_REMOTE_ABORTED_ERR 	 = 0x22,
+	SYNDROME_INVAL_EECN_ERR 	 = 0x23,
+	SYNDROME_INVAL_EEC_STATE_ERR 	 = 0x24
+};
+
+struct mthca_cqe {
+	__be32 my_qpn;
+	__be32 my_ee;
+	__be32 rqpn;
+	u8     sl_ipok;
+	u8     g_mlpath;
+	__be16 rlid;
+	__be32 imm_etype_pkey_eec;
+	__be32 byte_cnt;
+	__be32 wqe;
+	u8     opcode;
+	u8     is_send;
+	u8     reserved;
+	u8     owner;
+};
+
+struct mthca_err_cqe {
+	__be32 my_qpn;
+	u32    reserved1[3];
+	u8     syndrome;
+	u8     vendor_err;
+	__be16 db_cnt;
+	u32    reserved2;
+	__be32 wqe;
+	u8     opcode;
+	u8     reserved3[2];
+	u8     owner;
+};
+
+#define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
+#define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
+
+#define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
+#define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
+#define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
+#define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
+#define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
+
+#define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
+#define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
+#define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
+
+static inline struct mthca_cqe *get_cqe_from_buf(struct mthca_cq_buf *buf,
+						 int entry)
+{
+	if (buf->is_direct)
+		return buf->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
+	else
+		return buf->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
+			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
+}
+
+static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
+{
+	return get_cqe_from_buf(&cq->buf, entry);
+}
+
+static inline struct mthca_cqe *cqe_sw(struct mthca_cqe *cqe)
+{
+	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
+}
+
+static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
+{
+	return cqe_sw(get_cqe(cq, cq->cons_index & cq->ibcq.cqe));
+}
+
+static inline void set_cqe_hw(struct mthca_cqe *cqe)
+{
+	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
+}
+
+static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
+{
+	__be32 *cqe = cqe_ptr;
+
+	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
+	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
+		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
+		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
+		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
+}
+
+/*
+ * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
+ * should be correct before calling update_cons_index().
+ */
+static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
+				     int incr)
+{
+	if (mthca_is_memfree(dev)) {
+		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
+		wmb();
+	} else {
+		mthca_write64(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn, incr - 1,
+			      dev->kar + MTHCA_CQ_DOORBELL,
+			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+	}
+}
+
+void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
+{
+	struct mthca_cq *cq;
+
+	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
+
+	if (!cq) {
+		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
+		return;
+	}
+
+	++cq->arm_sn;
+
+	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
+}
+
+void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
+		    enum ib_event_type event_type)
+{
+	struct mthca_cq *cq;
+	struct ib_event event;
+
+	spin_lock(&dev->cq_table.lock);
+
+	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
+	if (cq)
+		++cq->refcount;
+
+	spin_unlock(&dev->cq_table.lock);
+
+	if (!cq) {
+		mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
+		return;
+	}
+
+	event.device      = &dev->ib_dev;
+	event.event       = event_type;
+	event.element.cq  = &cq->ibcq;
+	if (cq->ibcq.event_handler)
+		cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
+
+	spin_lock(&dev->cq_table.lock);
+	if (!--cq->refcount)
+		wake_up(&cq->wait);
+	spin_unlock(&dev->cq_table.lock);
+}
+
+static inline int is_recv_cqe(struct mthca_cqe *cqe)
+{
+	if ((cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
+	    MTHCA_ERROR_CQE_OPCODE_MASK)
+		return !(cqe->opcode & 0x01);
+	else
+		return !(cqe->is_send & 0x80);
+}
+
+void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
+		    struct mthca_srq *srq)
+{
+	struct mthca_cqe *cqe;
+	u32 prod_index;
+	int i, nfreed = 0;
+
+	spin_lock_irq(&cq->lock);
+
+	/*
+	 * First we need to find the current producer index, so we
+	 * know where to start cleaning from.  It doesn't matter if HW
+	 * adds new entries after this loop -- the QP we're worried
+	 * about is already in RESET, so the new entries won't come
+	 * from our QP and therefore don't need to be checked.
+	 */
+	for (prod_index = cq->cons_index;
+	     cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
+	     ++prod_index)
+		if (prod_index == cq->cons_index + cq->ibcq.cqe)
+			break;
+
+	if (0)
+		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
+			  qpn, cq->cqn, cq->cons_index, prod_index);
+
+	/*
+	 * Now sweep backwards through the CQ, removing CQ entries
+	 * that match our QP by copying older entries on top of them.
+	 */
+	while ((int) --prod_index - (int) cq->cons_index >= 0) {
+		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
+		if (cqe->my_qpn == cpu_to_be32(qpn)) {
+			if (srq && is_recv_cqe(cqe))
+				mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
+			++nfreed;
+		} else if (nfreed)
+			memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
+			       cqe, MTHCA_CQ_ENTRY_SIZE);
+	}
+
+	if (nfreed) {
+		for (i = 0; i < nfreed; ++i)
+			set_cqe_hw(get_cqe(cq, (cq->cons_index + i) & cq->ibcq.cqe));
+		wmb();
+		cq->cons_index += nfreed;
+		update_cons_index(dev, cq, nfreed);
+	}
+
+	spin_unlock_irq(&cq->lock);
+}
+
+void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
+{
+	int i;
+
+	/*
+	 * In Tavor mode, the hardware keeps the consumer and producer
+	 * indices mod the CQ size.  Since we might be making the CQ
+	 * bigger, we need to deal with the case where the producer
+	 * index wrapped around before the CQ was resized.
+	 */
+	if (!mthca_is_memfree(to_mdev(cq->ibcq.device)) &&
+	    cq->ibcq.cqe < cq->resize_buf->cqe) {
+		cq->cons_index &= cq->ibcq.cqe;
+		if (cqe_sw(get_cqe(cq, cq->ibcq.cqe)))
+			cq->cons_index -= cq->ibcq.cqe + 1;
+	}
+
+	for (i = cq->cons_index; cqe_sw(get_cqe(cq, i & cq->ibcq.cqe)); ++i)
+		memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
+					i & cq->resize_buf->cqe),
+		       get_cqe(cq, i & cq->ibcq.cqe), MTHCA_CQ_ENTRY_SIZE);
+}
+
+int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent)
+{
+	int ret;
+	int i;
+
+	ret = mthca_buf_alloc(dev, nent * MTHCA_CQ_ENTRY_SIZE,
+			      MTHCA_MAX_DIRECT_CQ_SIZE,
+			      &buf->queue, &buf->is_direct,
+			      &dev->driver_pd, 1, &buf->mr);
+	if (ret)
+		return ret;
+
+	for (i = 0; i < nent; ++i)
+		set_cqe_hw(get_cqe_from_buf(buf, i));
+
+	return 0;
+}
+
+void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe)
+{
+	mthca_buf_free(dev, (cqe + 1) * MTHCA_CQ_ENTRY_SIZE, &buf->queue,
+		       buf->is_direct, &buf->mr);
+}
+
+static void handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
+			     struct mthca_qp *qp, int wqe_index, int is_send,
+			     struct mthca_err_cqe *cqe,
+			     struct ib_wc *entry, int *free_cqe)
+{
+	int dbd;
+	__be32 new_wqe;
+
+	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
+		mthca_dbg(dev, "local QP operation err "
+			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
+			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
+			  cq->cqn, cq->cons_index);
+		dump_cqe(dev, cqe);
+	}
+
+	/*
+	 * For completions in error, only work request ID, status, vendor error
+	 * (and freed resource count for RD) have to be set.
+	 */
+	switch (cqe->syndrome) {
+	case SYNDROME_LOCAL_LENGTH_ERR:
+		entry->status = IB_WC_LOC_LEN_ERR;
+		break;
+	case SYNDROME_LOCAL_QP_OP_ERR:
+		entry->status = IB_WC_LOC_QP_OP_ERR;
+		break;
+	case SYNDROME_LOCAL_EEC_OP_ERR:
+		entry->status = IB_WC_LOC_EEC_OP_ERR;
+		break;
+	case SYNDROME_LOCAL_PROT_ERR:
+		entry->status = IB_WC_LOC_PROT_ERR;
+		break;
+	case SYNDROME_WR_FLUSH_ERR:
+		entry->status = IB_WC_WR_FLUSH_ERR;
+		break;
+	case SYNDROME_MW_BIND_ERR:
+		entry->status = IB_WC_MW_BIND_ERR;
+		break;
+	case SYNDROME_BAD_RESP_ERR:
+		entry->status = IB_WC_BAD_RESP_ERR;
+		break;
+	case SYNDROME_LOCAL_ACCESS_ERR:
+		entry->status = IB_WC_LOC_ACCESS_ERR;
+		break;
+	case SYNDROME_REMOTE_INVAL_REQ_ERR:
+		entry->status = IB_WC_REM_INV_REQ_ERR;
+		break;
+	case SYNDROME_REMOTE_ACCESS_ERR:
+		entry->status = IB_WC_REM_ACCESS_ERR;
+		break;
+	case SYNDROME_REMOTE_OP_ERR:
+		entry->status = IB_WC_REM_OP_ERR;
+		break;
+	case SYNDROME_RETRY_EXC_ERR:
+		entry->status = IB_WC_RETRY_EXC_ERR;
+		break;
+	case SYNDROME_RNR_RETRY_EXC_ERR:
+		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
+		break;
+	case SYNDROME_LOCAL_RDD_VIOL_ERR:
+		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
+		break;
+	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
+		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
+		break;
+	case SYNDROME_REMOTE_ABORTED_ERR:
+		entry->status = IB_WC_REM_ABORT_ERR;
+		break;
+	case SYNDROME_INVAL_EECN_ERR:
+		entry->status = IB_WC_INV_EECN_ERR;
+		break;
+	case SYNDROME_INVAL_EEC_STATE_ERR:
+		entry->status = IB_WC_INV_EEC_STATE_ERR;
+		break;
+	default:
+		entry->status = IB_WC_GENERAL_ERR;
+		break;
+	}
+
+	entry->vendor_err = cqe->vendor_err;
+
+	/*
+	 * Mem-free HCAs always generate one CQE per WQE, even in the
+	 * error case, so we don't have to check the doorbell count, etc.
+	 */
+	if (mthca_is_memfree(dev))
+		return;
+
+	mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
+
+	/*
+	 * If we're at the end of the WQE chain, or we've used up our
+	 * doorbell count, free the CQE.  Otherwise just update it for
+	 * the next poll operation.
+	 */
+	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
+		return;
+
+	be16_add_cpu(&cqe->db_cnt, -dbd);
+	cqe->wqe      = new_wqe;
+	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
+
+	*free_cqe = 0;
+}
+
+static inline int mthca_poll_one(struct mthca_dev *dev,
+				 struct mthca_cq *cq,
+				 struct mthca_qp **cur_qp,
+				 int *freed,
+				 struct ib_wc *entry)
+{
+	struct mthca_wq *wq;
+	struct mthca_cqe *cqe;
+	int wqe_index;
+	int is_error;
+	int is_send;
+	int free_cqe = 1;
+	int err = 0;
+	u16 checksum;
+
+	cqe = next_cqe_sw(cq);
+	if (!cqe)
+		return -EAGAIN;
+
+	/*
+	 * Make sure we read CQ entry contents after we've checked the
+	 * ownership bit.
+	 */
+	rmb();
+
+	if (0) {
+		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
+			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
+			  be32_to_cpu(cqe->wqe));
+		dump_cqe(dev, cqe);
+	}
+
+	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
+		MTHCA_ERROR_CQE_OPCODE_MASK;
+	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
+
+	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
+		/*
+		 * We do not have to take the QP table lock here,
+		 * because CQs will be locked while QPs are removed
+		 * from the table.
+		 */
+		*cur_qp = mthca_array_get(&dev->qp_table.qp,
+					  be32_to_cpu(cqe->my_qpn) &
+					  (dev->limits.num_qps - 1));
+		if (!*cur_qp) {
+			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
+				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
+			err = -EINVAL;
+			goto out;
+		}
+	}
+
+	entry->qp = &(*cur_qp)->ibqp;
+
+	if (is_send) {
+		wq = &(*cur_qp)->sq;
+		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
+			     >> wq->wqe_shift);
+		entry->wr_id = (*cur_qp)->wrid[wqe_index +
+					       (*cur_qp)->rq.max];
+	} else if ((*cur_qp)->ibqp.srq) {
+		struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
+		u32 wqe = be32_to_cpu(cqe->wqe);
+		wq = NULL;
+		wqe_index = wqe >> srq->wqe_shift;
+		entry->wr_id = srq->wrid[wqe_index];
+		mthca_free_srq_wqe(srq, wqe);
+	} else {
+		s32 wqe;
+		wq = &(*cur_qp)->rq;
+		wqe = be32_to_cpu(cqe->wqe);
+		wqe_index = wqe >> wq->wqe_shift;
+		/*
+		 * WQE addr == base - 1 might be reported in receive completion
+		 * with error instead of (rq size - 1) by Sinai FW 1.0.800 and
+		 * Arbel FW 5.1.400.  This bug should be fixed in later FW revs.
+		 */
+		if (unlikely(wqe_index < 0))
+			wqe_index = wq->max - 1;
+		entry->wr_id = (*cur_qp)->wrid[wqe_index];
+	}
+
+	if (wq) {
+		if (wq->last_comp < wqe_index)
+			wq->tail += wqe_index - wq->last_comp;
+		else
+			wq->tail += wqe_index + wq->max - wq->last_comp;
+
+		wq->last_comp = wqe_index;
+	}
+
+	if (is_error) {
+		handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
+				 (struct mthca_err_cqe *) cqe,
+				 entry, &free_cqe);
+		goto out;
+	}
+
+	if (is_send) {
+		entry->wc_flags = 0;
+		switch (cqe->opcode) {
+		case MTHCA_OPCODE_RDMA_WRITE:
+			entry->opcode    = IB_WC_RDMA_WRITE;
+			break;
+		case MTHCA_OPCODE_RDMA_WRITE_IMM:
+			entry->opcode    = IB_WC_RDMA_WRITE;
+			entry->wc_flags |= IB_WC_WITH_IMM;
+			break;
+		case MTHCA_OPCODE_SEND:
+			entry->opcode    = IB_WC_SEND;
+			break;
+		case MTHCA_OPCODE_SEND_IMM:
+			entry->opcode    = IB_WC_SEND;
+			entry->wc_flags |= IB_WC_WITH_IMM;
+			break;
+		case MTHCA_OPCODE_RDMA_READ:
+			entry->opcode    = IB_WC_RDMA_READ;
+			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
+			break;
+		case MTHCA_OPCODE_ATOMIC_CS:
+			entry->opcode    = IB_WC_COMP_SWAP;
+			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
+			break;
+		case MTHCA_OPCODE_ATOMIC_FA:
+			entry->opcode    = IB_WC_FETCH_ADD;
+			entry->byte_len  = MTHCA_ATOMIC_BYTE_LEN;
+			break;
+		default:
+			entry->opcode = 0xFF;
+			break;
+		}
+	} else {
+		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
+		switch (cqe->opcode & 0x1f) {
+		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
+		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
+			entry->wc_flags = IB_WC_WITH_IMM;
+			entry->ex.imm_data = cqe->imm_etype_pkey_eec;
+			entry->opcode = IB_WC_RECV;
+			break;
+		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
+		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
+			entry->wc_flags = IB_WC_WITH_IMM;
+			entry->ex.imm_data = cqe->imm_etype_pkey_eec;
+			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
+			break;
+		default:
+			entry->wc_flags = 0;
+			entry->opcode = IB_WC_RECV;
+			break;
+		}
+		entry->slid 	   = be16_to_cpu(cqe->rlid);
+		entry->sl   	   = cqe->sl_ipok >> 4;
+		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
+		entry->dlid_path_bits = cqe->g_mlpath & 0x7f;
+		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
+		entry->wc_flags   |= cqe->g_mlpath & 0x80 ? IB_WC_GRH : 0;
+		checksum = (be32_to_cpu(cqe->rqpn) >> 24) |
+				((be32_to_cpu(cqe->my_ee) >> 16) & 0xff00);
+		entry->wc_flags	  |=  (cqe->sl_ipok & 1 && checksum == 0xffff) ?
+							IB_WC_IP_CSUM_OK : 0;
+	}
+
+	entry->status = IB_WC_SUCCESS;
+
+ out:
+	if (likely(free_cqe)) {
+		set_cqe_hw(cqe);
+		++(*freed);
+		++cq->cons_index;
+	}
+
+	return err;
+}
+
+int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
+		  struct ib_wc *entry)
+{
+	struct mthca_dev *dev = to_mdev(ibcq->device);
+	struct mthca_cq *cq = to_mcq(ibcq);
+	struct mthca_qp *qp = NULL;
+	unsigned long flags;
+	int err = 0;
+	int freed = 0;
+	int npolled;
+
+	spin_lock_irqsave(&cq->lock, flags);
+
+	npolled = 0;
+repoll:
+	while (npolled < num_entries) {
+		err = mthca_poll_one(dev, cq, &qp,
+				     &freed, entry + npolled);
+		if (err)
+			break;
+		++npolled;
+	}
+
+	if (freed) {
+		wmb();
+		update_cons_index(dev, cq, freed);
+	}
+
+	/*
+	 * If a CQ resize is in progress and we discovered that the
+	 * old buffer is empty, then peek in the new buffer, and if
+	 * it's not empty, switch to the new buffer and continue
+	 * polling there.
+	 */
+	if (unlikely(err == -EAGAIN && cq->resize_buf &&
+		     cq->resize_buf->state == CQ_RESIZE_READY)) {
+		/*
+		 * In Tavor mode, the hardware keeps the producer
+		 * index modulo the CQ size.  Since we might be making
+		 * the CQ bigger, we need to mask our consumer index
+		 * using the size of the old CQ buffer before looking
+		 * in the new CQ buffer.
+		 */
+		if (!mthca_is_memfree(dev))
+			cq->cons_index &= cq->ibcq.cqe;
+
+		if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
+					    cq->cons_index & cq->resize_buf->cqe))) {
+			struct mthca_cq_buf tbuf;
+			int tcqe;
+
+			tbuf         = cq->buf;
+			tcqe         = cq->ibcq.cqe;
+			cq->buf      = cq->resize_buf->buf;
+			cq->ibcq.cqe = cq->resize_buf->cqe;
+
+			cq->resize_buf->buf   = tbuf;
+			cq->resize_buf->cqe   = tcqe;
+			cq->resize_buf->state = CQ_RESIZE_SWAPPED;
+
+			goto repoll;
+		}
+	}
+
+	spin_unlock_irqrestore(&cq->lock, flags);
+
+	return err == 0 || err == -EAGAIN ? npolled : err;
+}
+
+int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags)
+{
+	u32 dbhi = ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
+		    MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
+		    MTHCA_TAVOR_CQ_DB_REQ_NOT) |
+		to_mcq(cq)->cqn;
+
+	mthca_write64(dbhi, 0xffffffff, to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
+		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
+
+	return 0;
+}
+
+int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
+{
+	struct mthca_cq *cq = to_mcq(ibcq);
+	__be32 db_rec[2];
+	u32 dbhi;
+	u32 sn = cq->arm_sn & 3;
+
+	db_rec[0] = cpu_to_be32(cq->cons_index);
+	db_rec[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
+				((flags & IB_CQ_SOLICITED_MASK) ==
+				 IB_CQ_SOLICITED ? 1 : 2));
+
+	mthca_write_db_rec(db_rec, cq->arm_db);
+
+	/*
+	 * Make sure that the doorbell record in host memory is
+	 * written before ringing the doorbell via PCI MMIO.
+	 */
+	wmb();
+
+	dbhi = (sn << 28) |
+		((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
+		 MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
+		 MTHCA_ARBEL_CQ_DB_REQ_NOT) | cq->cqn;
+
+	mthca_write64(dbhi, cq->cons_index,
+		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
+		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
+
+	return 0;
+}
+
+int mthca_init_cq(struct mthca_dev *dev, int nent,
+		  struct mthca_ucontext *ctx, u32 pdn,
+		  struct mthca_cq *cq)
+{
+	struct mthca_mailbox *mailbox;
+	struct mthca_cq_context *cq_context;
+	int err = -ENOMEM;
+
+	cq->ibcq.cqe  = nent - 1;
+	cq->is_kernel = !ctx;
+
+	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
+	if (cq->cqn == -1)
+		return -ENOMEM;
+
+	if (mthca_is_memfree(dev)) {
+		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
+		if (err)
+			goto err_out;
+
+		if (cq->is_kernel) {
+			cq->arm_sn = 1;
+
+			err = -ENOMEM;
+
+			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
+							     cq->cqn, &cq->set_ci_db);
+			if (cq->set_ci_db_index < 0)
+				goto err_out_icm;
+
+			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
+							  cq->cqn, &cq->arm_db);
+			if (cq->arm_db_index < 0)
+				goto err_out_ci;
+		}
+	}
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		err = PTR_ERR(mailbox);
+		goto err_out_arm;
+	}
+
+	cq_context = mailbox->buf;
+
+	if (cq->is_kernel) {
+		err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
+		if (err)
+			goto err_out_mailbox;
+	}
+
+	spin_lock_init(&cq->lock);
+	cq->refcount = 1;
+	init_waitqueue_head(&cq->wait);
+	mutex_init(&cq->mutex);
+
+	memset(cq_context, 0, sizeof *cq_context);
+	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
+						  MTHCA_CQ_STATE_DISARMED |
+						  MTHCA_CQ_FLAG_TR);
+	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
+	if (ctx)
+		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
+	else
+		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
+	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
+	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
+	cq_context->pd              = cpu_to_be32(pdn);
+	cq_context->lkey            = cpu_to_be32(cq->buf.mr.ibmr.lkey);
+	cq_context->cqn             = cpu_to_be32(cq->cqn);
+
+	if (mthca_is_memfree(dev)) {
+		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
+		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
+	}
+
+	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn);
+	if (err) {
+		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
+		goto err_out_free_mr;
+	}
+
+	spin_lock_irq(&dev->cq_table.lock);
+	err = mthca_array_set(&dev->cq_table.cq,
+			      cq->cqn & (dev->limits.num_cqs - 1), cq);
+	if (err) {
+		spin_unlock_irq(&dev->cq_table.lock);
+		goto err_out_free_mr;
+	}
+	spin_unlock_irq(&dev->cq_table.lock);
+
+	cq->cons_index = 0;
+
+	mthca_free_mailbox(dev, mailbox);
+
+	return 0;
+
+err_out_free_mr:
+	if (cq->is_kernel)
+		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
+
+err_out_mailbox:
+	mthca_free_mailbox(dev, mailbox);
+
+err_out_arm:
+	if (cq->is_kernel && mthca_is_memfree(dev))
+		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
+
+err_out_ci:
+	if (cq->is_kernel && mthca_is_memfree(dev))
+		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
+
+err_out_icm:
+	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
+
+err_out:
+	mthca_free(&dev->cq_table.alloc, cq->cqn);
+
+	return err;
+}
+
+static inline int get_cq_refcount(struct mthca_dev *dev, struct mthca_cq *cq)
+{
+	int c;
+
+	spin_lock_irq(&dev->cq_table.lock);
+	c = cq->refcount;
+	spin_unlock_irq(&dev->cq_table.lock);
+
+	return c;
+}
+
+void mthca_free_cq(struct mthca_dev *dev,
+		   struct mthca_cq *cq)
+{
+	struct mthca_mailbox *mailbox;
+	int err;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
+		return;
+	}
+
+	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn);
+	if (err)
+		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
+
+	if (0) {
+		__be32 *ctx = mailbox->buf;
+		int j;
+
+		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
+		       cq->cqn, cq->cons_index,
+		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
+		for (j = 0; j < 16; ++j)
+			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
+	}
+
+	spin_lock_irq(&dev->cq_table.lock);
+	mthca_array_clear(&dev->cq_table.cq,
+			  cq->cqn & (dev->limits.num_cqs - 1));
+	--cq->refcount;
+	spin_unlock_irq(&dev->cq_table.lock);
+
+	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
+		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
+	else
+		synchronize_irq(dev->pdev->irq);
+
+	wait_event(cq->wait, !get_cq_refcount(dev, cq));
+
+	if (cq->is_kernel) {
+		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
+		if (mthca_is_memfree(dev)) {
+			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
+			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
+		}
+	}
+
+	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
+	mthca_free(&dev->cq_table.alloc, cq->cqn);
+	mthca_free_mailbox(dev, mailbox);
+}
+
+int mthca_init_cq_table(struct mthca_dev *dev)
+{
+	int err;
+
+	spin_lock_init(&dev->cq_table.lock);
+
+	err = mthca_alloc_init(&dev->cq_table.alloc,
+			       dev->limits.num_cqs,
+			       (1 << 24) - 1,
+			       dev->limits.reserved_cqs);
+	if (err)
+		return err;
+
+	err = mthca_array_init(&dev->cq_table.cq,
+			       dev->limits.num_cqs);
+	if (err)
+		mthca_alloc_cleanup(&dev->cq_table.alloc);
+
+	return err;
+}
+
+void mthca_cleanup_cq_table(struct mthca_dev *dev)
+{
+	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
+	mthca_alloc_cleanup(&dev->cq_table.alloc);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_dev.h b/drivers/infiniband/hw/mthca/mthca_dev.h
new file mode 100644
index 000000000..a4a9d871d
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_dev.h
@@ -0,0 +1,585 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef MTHCA_DEV_H
+#define MTHCA_DEV_H
+
+#include <linux/spinlock.h>
+#include <linux/kernel.h>
+#include <linux/pci.h>
+#include <linux/dma-mapping.h>
+#include <linux/timer.h>
+#include <linux/mutex.h>
+#include <linux/list.h>
+#include <linux/semaphore.h>
+
+#include "mthca_provider.h"
+#include "mthca_doorbell.h"
+
+#define DRV_NAME	"ib_mthca"
+#define PFX		DRV_NAME ": "
+#define DRV_VERSION	"1.0"
+#define DRV_RELDATE	"April 4, 2008"
+
+enum {
+	MTHCA_FLAG_DDR_HIDDEN = 1 << 1,
+	MTHCA_FLAG_SRQ        = 1 << 2,
+	MTHCA_FLAG_MSI_X      = 1 << 3,
+	MTHCA_FLAG_NO_LAM     = 1 << 4,
+	MTHCA_FLAG_FMR        = 1 << 5,
+	MTHCA_FLAG_MEMFREE    = 1 << 6,
+	MTHCA_FLAG_PCIE       = 1 << 7,
+	MTHCA_FLAG_SINAI_OPT  = 1 << 8
+};
+
+enum {
+	MTHCA_MAX_PORTS = 2
+};
+
+enum {
+	MTHCA_BOARD_ID_LEN = 64
+};
+
+enum {
+	MTHCA_EQ_CONTEXT_SIZE =  0x40,
+	MTHCA_CQ_CONTEXT_SIZE =  0x40,
+	MTHCA_QP_CONTEXT_SIZE = 0x200,
+	MTHCA_RDB_ENTRY_SIZE  =  0x20,
+	MTHCA_AV_SIZE         =  0x20,
+	MTHCA_MGM_ENTRY_SIZE  = 0x100,
+
+	/* Arbel FW gives us these, but we need them for Tavor */
+	MTHCA_MPT_ENTRY_SIZE  =  0x40,
+	MTHCA_MTT_SEG_SIZE    =  0x40,
+
+	MTHCA_QP_PER_MGM      = 4 * (MTHCA_MGM_ENTRY_SIZE / 16 - 2)
+};
+
+enum {
+	MTHCA_EQ_CMD,
+	MTHCA_EQ_ASYNC,
+	MTHCA_EQ_COMP,
+	MTHCA_NUM_EQ
+};
+
+enum {
+	MTHCA_OPCODE_NOP            = 0x00,
+	MTHCA_OPCODE_RDMA_WRITE     = 0x08,
+	MTHCA_OPCODE_RDMA_WRITE_IMM = 0x09,
+	MTHCA_OPCODE_SEND           = 0x0a,
+	MTHCA_OPCODE_SEND_IMM       = 0x0b,
+	MTHCA_OPCODE_RDMA_READ      = 0x10,
+	MTHCA_OPCODE_ATOMIC_CS      = 0x11,
+	MTHCA_OPCODE_ATOMIC_FA      = 0x12,
+	MTHCA_OPCODE_BIND_MW        = 0x18,
+};
+
+enum {
+	MTHCA_CMD_USE_EVENTS         = 1 << 0,
+	MTHCA_CMD_POST_DOORBELLS     = 1 << 1
+};
+
+enum {
+	MTHCA_CMD_NUM_DBELL_DWORDS = 8
+};
+
+struct mthca_cmd {
+	struct dma_pool          *pool;
+	struct mutex              hcr_mutex;
+	struct semaphore 	  poll_sem;
+	struct semaphore 	  event_sem;
+	int              	  max_cmds;
+	spinlock_t                context_lock;
+	int                       free_head;
+	struct mthca_cmd_context *context;
+	u16                       token_mask;
+	u32                       flags;
+	void __iomem             *dbell_map;
+	u16                       dbell_offsets[MTHCA_CMD_NUM_DBELL_DWORDS];
+};
+
+struct mthca_limits {
+	int      num_ports;
+	int      vl_cap;
+	int      mtu_cap;
+	int      gid_table_len;
+	int      pkey_table_len;
+	int      local_ca_ack_delay;
+	int      num_uars;
+	int      max_sg;
+	int      num_qps;
+	int      max_wqes;
+	int	 max_desc_sz;
+	int	 max_qp_init_rdma;
+	int      reserved_qps;
+	int      num_srqs;
+	int      max_srq_wqes;
+	int      max_srq_sge;
+	int      reserved_srqs;
+	int      num_eecs;
+	int      reserved_eecs;
+	int      num_cqs;
+	int      max_cqes;
+	int      reserved_cqs;
+	int      num_eqs;
+	int      reserved_eqs;
+	int      num_mpts;
+	int      num_mtt_segs;
+	int	 mtt_seg_size;
+	int      fmr_reserved_mtts;
+	int      reserved_mtts;
+	int      reserved_mrws;
+	int      reserved_uars;
+	int      num_mgms;
+	int      num_amgms;
+	int      reserved_mcgs;
+	int      num_pds;
+	int      reserved_pds;
+	u32      page_size_cap;
+	u32      flags;
+	u16      stat_rate_support;
+	u8       port_width_cap;
+};
+
+struct mthca_alloc {
+	u32            last;
+	u32            top;
+	u32            max;
+	u32            mask;
+	spinlock_t     lock;
+	unsigned long *table;
+};
+
+struct mthca_array {
+	struct {
+		void    **page;
+		int       used;
+	} *page_list;
+};
+
+struct mthca_uar_table {
+	struct mthca_alloc alloc;
+	u64                uarc_base;
+	int                uarc_size;
+};
+
+struct mthca_pd_table {
+	struct mthca_alloc alloc;
+};
+
+struct mthca_buddy {
+	unsigned long **bits;
+	int	       *num_free;
+	int             max_order;
+	spinlock_t      lock;
+};
+
+struct mthca_mr_table {
+	struct mthca_alloc      mpt_alloc;
+	struct mthca_buddy      mtt_buddy;
+	struct mthca_buddy     *fmr_mtt_buddy;
+	u64                     mtt_base;
+	u64                     mpt_base;
+	struct mthca_icm_table *mtt_table;
+	struct mthca_icm_table *mpt_table;
+	struct {
+		void __iomem   *mpt_base;
+		void __iomem   *mtt_base;
+		struct mthca_buddy mtt_buddy;
+	} tavor_fmr;
+};
+
+struct mthca_eq_table {
+	struct mthca_alloc alloc;
+	void __iomem      *clr_int;
+	u32                clr_mask;
+	u32                arm_mask;
+	struct mthca_eq    eq[MTHCA_NUM_EQ];
+	u64                icm_virt;
+	struct page       *icm_page;
+	dma_addr_t         icm_dma;
+	int                have_irq;
+	u8                 inta_pin;
+};
+
+struct mthca_cq_table {
+	struct mthca_alloc 	alloc;
+	spinlock_t         	lock;
+	struct mthca_array      cq;
+	struct mthca_icm_table *table;
+};
+
+struct mthca_srq_table {
+	struct mthca_alloc 	alloc;
+	spinlock_t         	lock;
+	struct mthca_array      srq;
+	struct mthca_icm_table *table;
+};
+
+struct mthca_qp_table {
+	struct mthca_alloc     	alloc;
+	u32                    	rdb_base;
+	int                    	rdb_shift;
+	int                    	sqp_start;
+	spinlock_t             	lock;
+	struct mthca_array     	qp;
+	struct mthca_icm_table *qp_table;
+	struct mthca_icm_table *eqp_table;
+	struct mthca_icm_table *rdb_table;
+};
+
+struct mthca_av_table {
+	struct dma_pool   *pool;
+	int                num_ddr_avs;
+	u64                ddr_av_base;
+	void __iomem      *av_map;
+	struct mthca_alloc alloc;
+};
+
+struct mthca_mcg_table {
+	struct mutex		mutex;
+	struct mthca_alloc 	alloc;
+	struct mthca_icm_table *table;
+};
+
+struct mthca_catas_err {
+	u64			addr;
+	u32 __iomem	       *map;
+	u32			size;
+	struct timer_list	timer;
+	struct list_head	list;
+};
+
+extern struct mutex mthca_device_mutex;
+
+struct mthca_dev {
+	struct ib_device  ib_dev;
+	struct pci_dev   *pdev;
+
+	int          	 hca_type;
+	unsigned long	 mthca_flags;
+	unsigned long    device_cap_flags;
+
+	u32              rev_id;
+	char             board_id[MTHCA_BOARD_ID_LEN];
+
+	/* firmware info */
+	u64              fw_ver;
+	union {
+		struct {
+			u64 fw_start;
+			u64 fw_end;
+		}        tavor;
+		struct {
+			u64 clr_int_base;
+			u64 eq_arm_base;
+			u64 eq_set_ci_base;
+			struct mthca_icm *fw_icm;
+			struct mthca_icm *aux_icm;
+			u16 fw_pages;
+		}        arbel;
+	}                fw;
+
+	u64              ddr_start;
+	u64              ddr_end;
+
+	MTHCA_DECLARE_DOORBELL_LOCK(doorbell_lock)
+	struct mutex cap_mask_mutex;
+
+	void __iomem    *hcr;
+	void __iomem    *kar;
+	void __iomem    *clr_base;
+	union {
+		struct {
+			void __iomem *ecr_base;
+		} tavor;
+		struct {
+			void __iomem *eq_arm;
+			void __iomem *eq_set_ci_base;
+		} arbel;
+	} eq_regs;
+
+	struct mthca_cmd    cmd;
+	struct mthca_limits limits;
+
+	struct mthca_uar_table uar_table;
+	struct mthca_pd_table  pd_table;
+	struct mthca_mr_table  mr_table;
+	struct mthca_eq_table  eq_table;
+	struct mthca_cq_table  cq_table;
+	struct mthca_srq_table srq_table;
+	struct mthca_qp_table  qp_table;
+	struct mthca_av_table  av_table;
+	struct mthca_mcg_table mcg_table;
+
+	struct mthca_catas_err catas_err;
+
+	struct mthca_uar       driver_uar;
+	struct mthca_db_table *db_tab;
+	struct mthca_pd        driver_pd;
+	struct mthca_mr        driver_mr;
+
+	struct ib_mad_agent  *send_agent[MTHCA_MAX_PORTS][2];
+	struct ib_ah         *sm_ah[MTHCA_MAX_PORTS];
+	spinlock_t            sm_lock;
+	u8                    rate[MTHCA_MAX_PORTS];
+	bool		      active;
+};
+
+#ifdef CONFIG_INFINIBAND_MTHCA_DEBUG
+extern int mthca_debug_level;
+
+#define mthca_dbg(mdev, format, arg...)					\
+	do {								\
+		if (mthca_debug_level)					\
+			dev_printk(KERN_DEBUG, &mdev->pdev->dev, format, ## arg); \
+	} while (0)
+
+#else /* CONFIG_INFINIBAND_MTHCA_DEBUG */
+
+#define mthca_dbg(mdev, format, arg...) do { (void) mdev; } while (0)
+
+#endif /* CONFIG_INFINIBAND_MTHCA_DEBUG */
+
+#define mthca_err(mdev, format, arg...) \
+	dev_err(&mdev->pdev->dev, format, ## arg)
+#define mthca_info(mdev, format, arg...) \
+	dev_info(&mdev->pdev->dev, format, ## arg)
+#define mthca_warn(mdev, format, arg...) \
+	dev_warn(&mdev->pdev->dev, format, ## arg)
+
+extern void __buggy_use_of_MTHCA_GET(void);
+extern void __buggy_use_of_MTHCA_PUT(void);
+
+#define MTHCA_GET(dest, source, offset)                               \
+	do {                                                          \
+		void *__p = (char *) (source) + (offset);             \
+		switch (sizeof (dest)) {                              \
+		case 1: (dest) = *(u8 *) __p;       break;	      \
+		case 2: (dest) = be16_to_cpup(__p); break;	      \
+		case 4: (dest) = be32_to_cpup(__p); break;	      \
+		case 8: (dest) = be64_to_cpup(__p); break;	      \
+		default: __buggy_use_of_MTHCA_GET();		      \
+		}                                                     \
+	} while (0)
+
+#define MTHCA_PUT(dest, source, offset)                               \
+	do {                                                          \
+		void *__d = ((char *) (dest) + (offset));	      \
+		switch (sizeof(source)) {                             \
+		case 1: *(u8 *) __d = (source);                break; \
+		case 2:	*(__be16 *) __d = cpu_to_be16(source); break; \
+		case 4:	*(__be32 *) __d = cpu_to_be32(source); break; \
+		case 8:	*(__be64 *) __d = cpu_to_be64(source); break; \
+		default: __buggy_use_of_MTHCA_PUT();		      \
+		}                                                     \
+	} while (0)
+
+int mthca_reset(struct mthca_dev *mdev);
+
+u32 mthca_alloc(struct mthca_alloc *alloc);
+void mthca_free(struct mthca_alloc *alloc, u32 obj);
+int mthca_alloc_init(struct mthca_alloc *alloc, u32 num, u32 mask,
+		     u32 reserved);
+void mthca_alloc_cleanup(struct mthca_alloc *alloc);
+void *mthca_array_get(struct mthca_array *array, int index);
+int mthca_array_set(struct mthca_array *array, int index, void *value);
+void mthca_array_clear(struct mthca_array *array, int index);
+int mthca_array_init(struct mthca_array *array, int nent);
+void mthca_array_cleanup(struct mthca_array *array, int nent);
+int mthca_buf_alloc(struct mthca_dev *dev, int size, int max_direct,
+		    union mthca_buf *buf, int *is_direct, struct mthca_pd *pd,
+		    int hca_write, struct mthca_mr *mr);
+void mthca_buf_free(struct mthca_dev *dev, int size, union mthca_buf *buf,
+		    int is_direct, struct mthca_mr *mr);
+
+int mthca_init_uar_table(struct mthca_dev *dev);
+int mthca_init_pd_table(struct mthca_dev *dev);
+int mthca_init_mr_table(struct mthca_dev *dev);
+int mthca_init_eq_table(struct mthca_dev *dev);
+int mthca_init_cq_table(struct mthca_dev *dev);
+int mthca_init_srq_table(struct mthca_dev *dev);
+int mthca_init_qp_table(struct mthca_dev *dev);
+int mthca_init_av_table(struct mthca_dev *dev);
+int mthca_init_mcg_table(struct mthca_dev *dev);
+
+void mthca_cleanup_uar_table(struct mthca_dev *dev);
+void mthca_cleanup_pd_table(struct mthca_dev *dev);
+void mthca_cleanup_mr_table(struct mthca_dev *dev);
+void mthca_cleanup_eq_table(struct mthca_dev *dev);
+void mthca_cleanup_cq_table(struct mthca_dev *dev);
+void mthca_cleanup_srq_table(struct mthca_dev *dev);
+void mthca_cleanup_qp_table(struct mthca_dev *dev);
+void mthca_cleanup_av_table(struct mthca_dev *dev);
+void mthca_cleanup_mcg_table(struct mthca_dev *dev);
+
+int mthca_register_device(struct mthca_dev *dev);
+void mthca_unregister_device(struct mthca_dev *dev);
+
+void mthca_start_catas_poll(struct mthca_dev *dev);
+void mthca_stop_catas_poll(struct mthca_dev *dev);
+int __mthca_restart_one(struct pci_dev *pdev);
+int mthca_catas_init(void);
+void mthca_catas_cleanup(void);
+
+int mthca_uar_alloc(struct mthca_dev *dev, struct mthca_uar *uar);
+void mthca_uar_free(struct mthca_dev *dev, struct mthca_uar *uar);
+
+int mthca_pd_alloc(struct mthca_dev *dev, int privileged, struct mthca_pd *pd);
+void mthca_pd_free(struct mthca_dev *dev, struct mthca_pd *pd);
+
+int mthca_write_mtt_size(struct mthca_dev *dev);
+
+struct mthca_mtt *mthca_alloc_mtt(struct mthca_dev *dev, int size);
+void mthca_free_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt);
+int mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,
+		    int start_index, u64 *buffer_list, int list_len);
+int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift,
+		   u64 iova, u64 total_size, u32 access, struct mthca_mr *mr);
+int mthca_mr_alloc_notrans(struct mthca_dev *dev, u32 pd,
+			   u32 access, struct mthca_mr *mr);
+int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
+			u64 *buffer_list, int buffer_size_shift,
+			int list_len, u64 iova, u64 total_size,
+			u32 access, struct mthca_mr *mr);
+void mthca_free_mr(struct mthca_dev *dev,  struct mthca_mr *mr);
+
+int mthca_map_eq_icm(struct mthca_dev *dev, u64 icm_virt);
+void mthca_unmap_eq_icm(struct mthca_dev *dev);
+
+int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
+		  struct ib_wc *entry);
+int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags);
+int mthca_arbel_arm_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags);
+int mthca_init_cq(struct mthca_dev *dev, int nent,
+		  struct mthca_ucontext *ctx, u32 pdn,
+		  struct mthca_cq *cq);
+void mthca_free_cq(struct mthca_dev *dev,
+		   struct mthca_cq *cq);
+void mthca_cq_completion(struct mthca_dev *dev, u32 cqn);
+void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
+		    enum ib_event_type event_type);
+void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
+		    struct mthca_srq *srq);
+void mthca_cq_resize_copy_cqes(struct mthca_cq *cq);
+int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent);
+void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe);
+
+int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
+		    struct ib_srq_attr *attr, struct mthca_srq *srq,
+		    struct ib_udata *udata);
+void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq);
+int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
+		     enum ib_srq_attr_mask attr_mask, struct ib_udata *udata);
+int mthca_query_srq(struct ib_srq *srq, struct ib_srq_attr *srq_attr);
+int mthca_max_srq_sge(struct mthca_dev *dev);
+void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
+		     enum ib_event_type event_type);
+void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr);
+int mthca_tavor_post_srq_recv(struct ib_srq *srq, const struct ib_recv_wr *wr,
+			      const struct ib_recv_wr **bad_wr);
+int mthca_arbel_post_srq_recv(struct ib_srq *srq, const struct ib_recv_wr *wr,
+			      const struct ib_recv_wr **bad_wr);
+
+void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
+		    enum ib_event_type event_type);
+int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
+		   struct ib_qp_init_attr *qp_init_attr);
+int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
+		    struct ib_udata *udata);
+int mthca_tavor_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
+			  const struct ib_send_wr **bad_wr);
+int mthca_tavor_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
+			     const struct ib_recv_wr **bad_wr);
+int mthca_arbel_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
+			  const struct ib_send_wr **bad_wr);
+int mthca_arbel_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
+			     const struct ib_recv_wr **bad_wr);
+void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
+			int index, int *dbd, __be32 *new_wqe);
+int mthca_alloc_qp(struct mthca_dev *dev,
+		   struct mthca_pd *pd,
+		   struct mthca_cq *send_cq,
+		   struct mthca_cq *recv_cq,
+		   enum ib_qp_type type,
+		   enum ib_sig_type send_policy,
+		   struct ib_qp_cap *cap,
+		   struct mthca_qp *qp,
+		   struct ib_udata *udata);
+int mthca_alloc_sqp(struct mthca_dev *dev,
+		    struct mthca_pd *pd,
+		    struct mthca_cq *send_cq,
+		    struct mthca_cq *recv_cq,
+		    enum ib_sig_type send_policy,
+		    struct ib_qp_cap *cap,
+		    int qpn,
+		    u32 port,
+		    struct mthca_qp *qp,
+		    struct ib_udata *udata);
+void mthca_free_qp(struct mthca_dev *dev, struct mthca_qp *qp);
+int mthca_create_ah(struct mthca_dev *dev,
+		    struct mthca_pd *pd,
+		    struct rdma_ah_attr *ah_attr,
+		    struct mthca_ah *ah);
+int mthca_destroy_ah(struct mthca_dev *dev, struct mthca_ah *ah);
+int mthca_read_ah(struct mthca_dev *dev, struct mthca_ah *ah,
+		  struct ib_ud_header *header);
+int mthca_ah_query(struct ib_ah *ibah, struct rdma_ah_attr *attr);
+int mthca_ah_grh_present(struct mthca_ah *ah);
+u8 mthca_get_rate(struct mthca_dev *dev, int static_rate, u32 port);
+enum ib_rate mthca_rate_to_ib(struct mthca_dev *dev, u8 mthca_rate, u32 port);
+
+int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid);
+int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid);
+
+int mthca_process_mad(struct ib_device *ibdev, int mad_flags, u32 port_num,
+		      const struct ib_wc *in_wc, const struct ib_grh *in_grh,
+		      const struct ib_mad *in, struct ib_mad *out,
+		      size_t *out_mad_size, u16 *out_mad_pkey_index);
+int mthca_create_agents(struct mthca_dev *dev);
+void mthca_free_agents(struct mthca_dev *dev);
+
+static inline struct mthca_dev *to_mdev(struct ib_device *ibdev)
+{
+	return container_of(ibdev, struct mthca_dev, ib_dev);
+}
+
+static inline int mthca_is_memfree(struct mthca_dev *dev)
+{
+	return dev->mthca_flags & MTHCA_FLAG_MEMFREE;
+}
+
+#endif /* MTHCA_DEV_H */
diff --git a/drivers/infiniband/hw/mthca/mthca_doorbell.h b/drivers/infiniband/hw/mthca/mthca_doorbell.h
new file mode 100644
index 000000000..14f51ef97
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_doorbell.h
@@ -0,0 +1,109 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/types.h>
+
+#define MTHCA_RD_DOORBELL      0x00
+#define MTHCA_SEND_DOORBELL    0x10
+#define MTHCA_RECEIVE_DOORBELL 0x18
+#define MTHCA_CQ_DOORBELL      0x20
+#define MTHCA_EQ_DOORBELL      0x28
+
+#if BITS_PER_LONG == 64
+/*
+ * Assume that we can just write a 64-bit doorbell atomically.  s390
+ * actually doesn't have writeq() but S/390 systems don't even have
+ * PCI so we won't worry about it.
+ */
+
+#define MTHCA_DECLARE_DOORBELL_LOCK(name)
+#define MTHCA_INIT_DOORBELL_LOCK(ptr)    do { } while (0)
+#define MTHCA_GET_DOORBELL_LOCK(ptr)      (NULL)
+
+static inline void mthca_write64_raw(__be64 val, void __iomem *dest)
+{
+	__raw_writeq((__force u64) val, dest);
+}
+
+static inline void mthca_write64(u32 hi, u32 lo, void __iomem *dest,
+				 spinlock_t *doorbell_lock)
+{
+	__raw_writeq((__force u64) cpu_to_be64((u64) hi << 32 | lo), dest);
+}
+
+static inline void mthca_write_db_rec(__be32 val[2], __be32 *db)
+{
+	*(u64 *) db = *(u64 *) val;
+}
+
+#else
+
+/*
+ * Just fall back to a spinlock to protect the doorbell if
+ * BITS_PER_LONG is 32 -- there's no portable way to do atomic 64-bit
+ * MMIO writes.
+ */
+
+#define MTHCA_DECLARE_DOORBELL_LOCK(name) spinlock_t name;
+#define MTHCA_INIT_DOORBELL_LOCK(ptr)     spin_lock_init(ptr)
+#define MTHCA_GET_DOORBELL_LOCK(ptr)      (ptr)
+
+static inline void mthca_write64_raw(__be64 val, void __iomem *dest)
+{
+	__raw_writel(((__force u32 *) &val)[0], dest);
+	__raw_writel(((__force u32 *) &val)[1], dest + 4);
+}
+
+static inline void mthca_write64(u32 hi, u32 lo, void __iomem *dest,
+				 spinlock_t *doorbell_lock)
+{
+	unsigned long flags;
+
+	hi = (__force u32) cpu_to_be32(hi);
+	lo = (__force u32) cpu_to_be32(lo);
+
+	spin_lock_irqsave(doorbell_lock, flags);
+	__raw_writel(hi, dest);
+	__raw_writel(lo, dest + 4);
+	spin_unlock_irqrestore(doorbell_lock, flags);
+}
+
+static inline void mthca_write_db_rec(__be32 val[2], __be32 *db)
+{
+	db[0] = val[0];
+	wmb();
+	db[1] = val[1];
+}
+
+#endif
diff --git a/drivers/infiniband/hw/mthca/mthca_eq.c b/drivers/infiniband/hw/mthca/mthca_eq.c
new file mode 100644
index 000000000..97287c544
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_eq.c
@@ -0,0 +1,906 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/errno.h>
+#include <linux/interrupt.h>
+#include <linux/pci.h>
+#include <linux/slab.h>
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+#include "mthca_config_reg.h"
+
+enum {
+	MTHCA_NUM_ASYNC_EQE = 0x80,
+	MTHCA_NUM_CMD_EQE   = 0x80,
+	MTHCA_NUM_SPARE_EQE = 0x80,
+	MTHCA_EQ_ENTRY_SIZE = 0x20
+};
+
+/*
+ * Must be packed because start is 64 bits but only aligned to 32 bits.
+ */
+struct mthca_eq_context {
+	__be32 flags;
+	__be64 start;
+	__be32 logsize_usrpage;
+	__be32 tavor_pd;	/* reserved for Arbel */
+	u8     reserved1[3];
+	u8     intr;
+	__be32 arbel_pd;	/* lost_count for Tavor */
+	__be32 lkey;
+	u32    reserved2[2];
+	__be32 consumer_index;
+	__be32 producer_index;
+	u32    reserved3[4];
+} __packed;
+
+#define MTHCA_EQ_STATUS_OK          ( 0 << 28)
+#define MTHCA_EQ_STATUS_OVERFLOW    ( 9 << 28)
+#define MTHCA_EQ_STATUS_WRITE_FAIL  (10 << 28)
+#define MTHCA_EQ_OWNER_SW           ( 0 << 24)
+#define MTHCA_EQ_OWNER_HW           ( 1 << 24)
+#define MTHCA_EQ_FLAG_TR            ( 1 << 18)
+#define MTHCA_EQ_FLAG_OI            ( 1 << 17)
+#define MTHCA_EQ_STATE_ARMED        ( 1 <<  8)
+#define MTHCA_EQ_STATE_FIRED        ( 2 <<  8)
+#define MTHCA_EQ_STATE_ALWAYS_ARMED ( 3 <<  8)
+#define MTHCA_EQ_STATE_ARBEL        ( 8 <<  8)
+
+enum {
+	MTHCA_EVENT_TYPE_COMP       	    = 0x00,
+	MTHCA_EVENT_TYPE_PATH_MIG   	    = 0x01,
+	MTHCA_EVENT_TYPE_COMM_EST   	    = 0x02,
+	MTHCA_EVENT_TYPE_SQ_DRAINED 	    = 0x03,
+	MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE    = 0x13,
+	MTHCA_EVENT_TYPE_SRQ_LIMIT	    = 0x14,
+	MTHCA_EVENT_TYPE_CQ_ERROR   	    = 0x04,
+	MTHCA_EVENT_TYPE_WQ_CATAS_ERROR     = 0x05,
+	MTHCA_EVENT_TYPE_EEC_CATAS_ERROR    = 0x06,
+	MTHCA_EVENT_TYPE_PATH_MIG_FAILED    = 0x07,
+	MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR = 0x10,
+	MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR    = 0x11,
+	MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR    = 0x12,
+	MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR  = 0x08,
+	MTHCA_EVENT_TYPE_PORT_CHANGE        = 0x09,
+	MTHCA_EVENT_TYPE_EQ_OVERFLOW        = 0x0f,
+	MTHCA_EVENT_TYPE_ECC_DETECT         = 0x0e,
+	MTHCA_EVENT_TYPE_CMD                = 0x0a
+};
+
+#define MTHCA_ASYNC_EVENT_MASK ((1ULL << MTHCA_EVENT_TYPE_PATH_MIG)           | \
+				(1ULL << MTHCA_EVENT_TYPE_COMM_EST)           | \
+				(1ULL << MTHCA_EVENT_TYPE_SQ_DRAINED)         | \
+				(1ULL << MTHCA_EVENT_TYPE_CQ_ERROR)           | \
+				(1ULL << MTHCA_EVENT_TYPE_WQ_CATAS_ERROR)     | \
+				(1ULL << MTHCA_EVENT_TYPE_EEC_CATAS_ERROR)    | \
+				(1ULL << MTHCA_EVENT_TYPE_PATH_MIG_FAILED)    | \
+				(1ULL << MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR) | \
+				(1ULL << MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR)    | \
+				(1ULL << MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR)  | \
+				(1ULL << MTHCA_EVENT_TYPE_PORT_CHANGE)        | \
+				(1ULL << MTHCA_EVENT_TYPE_ECC_DETECT))
+#define MTHCA_SRQ_EVENT_MASK   ((1ULL << MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR)    | \
+				(1ULL << MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE)    | \
+				(1ULL << MTHCA_EVENT_TYPE_SRQ_LIMIT))
+#define MTHCA_CMD_EVENT_MASK    (1ULL << MTHCA_EVENT_TYPE_CMD)
+
+#define MTHCA_EQ_DB_INC_CI     (1 << 24)
+#define MTHCA_EQ_DB_REQ_NOT    (2 << 24)
+#define MTHCA_EQ_DB_DISARM_CQ  (3 << 24)
+#define MTHCA_EQ_DB_SET_CI     (4 << 24)
+#define MTHCA_EQ_DB_ALWAYS_ARM (5 << 24)
+
+struct mthca_eqe {
+	u8 reserved1;
+	u8 type;
+	u8 reserved2;
+	u8 subtype;
+	union {
+		u32 raw[6];
+		struct {
+			__be32 cqn;
+		} __packed comp;
+		struct {
+			u16    reserved1;
+			__be16 token;
+			u32    reserved2;
+			u8     reserved3[3];
+			u8     status;
+			__be64 out_param;
+		} __packed cmd;
+		struct {
+			__be32 qpn;
+		} __packed qp;
+		struct {
+			__be32 srqn;
+		} __packed srq;
+		struct {
+			__be32 cqn;
+			u32    reserved1;
+			u8     reserved2[3];
+			u8     syndrome;
+		} __packed cq_err;
+		struct {
+			u32    reserved1[2];
+			__be32 port;
+		} __packed port_change;
+	} event;
+	u8 reserved3[3];
+	u8 owner;
+} __packed;
+
+#define  MTHCA_EQ_ENTRY_OWNER_SW      (0 << 7)
+#define  MTHCA_EQ_ENTRY_OWNER_HW      (1 << 7)
+
+static inline u64 async_mask(struct mthca_dev *dev)
+{
+	return dev->mthca_flags & MTHCA_FLAG_SRQ ?
+		MTHCA_ASYNC_EVENT_MASK | MTHCA_SRQ_EVENT_MASK :
+		MTHCA_ASYNC_EVENT_MASK;
+}
+
+static inline void tavor_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
+{
+	/*
+	 * This barrier makes sure that all updates to ownership bits
+	 * done by set_eqe_hw() hit memory before the consumer index
+	 * is updated.  set_eq_ci() allows the HCA to possibly write
+	 * more EQ entries, and we want to avoid the exceedingly
+	 * unlikely possibility of the HCA writing an entry and then
+	 * having set_eqe_hw() overwrite the owner field.
+	 */
+	wmb();
+	mthca_write64(MTHCA_EQ_DB_SET_CI | eq->eqn, ci & (eq->nent - 1),
+		      dev->kar + MTHCA_EQ_DOORBELL,
+		      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+}
+
+static inline void arbel_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
+{
+	/* See comment in tavor_set_eq_ci() above. */
+	wmb();
+	__raw_writel((__force u32) cpu_to_be32(ci),
+		     dev->eq_regs.arbel.eq_set_ci_base + eq->eqn * 8);
+	/* We still want ordering, just not swabbing, so add a barrier */
+	mb();
+}
+
+static inline void set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
+{
+	if (mthca_is_memfree(dev))
+		arbel_set_eq_ci(dev, eq, ci);
+	else
+		tavor_set_eq_ci(dev, eq, ci);
+}
+
+static inline void tavor_eq_req_not(struct mthca_dev *dev, int eqn)
+{
+	mthca_write64(MTHCA_EQ_DB_REQ_NOT | eqn, 0,
+		      dev->kar + MTHCA_EQ_DOORBELL,
+		      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+}
+
+static inline void arbel_eq_req_not(struct mthca_dev *dev, u32 eqn_mask)
+{
+	writel(eqn_mask, dev->eq_regs.arbel.eq_arm);
+}
+
+static inline void disarm_cq(struct mthca_dev *dev, int eqn, int cqn)
+{
+	if (!mthca_is_memfree(dev)) {
+		mthca_write64(MTHCA_EQ_DB_DISARM_CQ | eqn, cqn,
+			      dev->kar + MTHCA_EQ_DOORBELL,
+			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+	}
+}
+
+static inline struct mthca_eqe *get_eqe(struct mthca_eq *eq, u32 entry)
+{
+	unsigned long off = (entry & (eq->nent - 1)) * MTHCA_EQ_ENTRY_SIZE;
+	return eq->page_list[off / PAGE_SIZE].buf + off % PAGE_SIZE;
+}
+
+static inline struct mthca_eqe *next_eqe_sw(struct mthca_eq *eq)
+{
+	struct mthca_eqe *eqe;
+	eqe = get_eqe(eq, eq->cons_index);
+	return (MTHCA_EQ_ENTRY_OWNER_HW & eqe->owner) ? NULL : eqe;
+}
+
+static inline void set_eqe_hw(struct mthca_eqe *eqe)
+{
+	eqe->owner =  MTHCA_EQ_ENTRY_OWNER_HW;
+}
+
+static void port_change(struct mthca_dev *dev, int port, int active)
+{
+	struct ib_event record;
+
+	mthca_dbg(dev, "Port change to %s for port %d\n",
+		  active ? "active" : "down", port);
+
+	record.device = &dev->ib_dev;
+	record.event  = active ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
+	record.element.port_num = port;
+
+	ib_dispatch_event(&record);
+}
+
+static int mthca_eq_int(struct mthca_dev *dev, struct mthca_eq *eq)
+{
+	struct mthca_eqe *eqe;
+	int disarm_cqn;
+	int eqes_found = 0;
+	int set_ci = 0;
+
+	while ((eqe = next_eqe_sw(eq))) {
+		/*
+		 * Make sure we read EQ entry contents after we've
+		 * checked the ownership bit.
+		 */
+		rmb();
+
+		switch (eqe->type) {
+		case MTHCA_EVENT_TYPE_COMP:
+			disarm_cqn = be32_to_cpu(eqe->event.comp.cqn) & 0xffffff;
+			disarm_cq(dev, eq->eqn, disarm_cqn);
+			mthca_cq_completion(dev, disarm_cqn);
+			break;
+
+		case MTHCA_EVENT_TYPE_PATH_MIG:
+			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
+				       IB_EVENT_PATH_MIG);
+			break;
+
+		case MTHCA_EVENT_TYPE_COMM_EST:
+			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
+				       IB_EVENT_COMM_EST);
+			break;
+
+		case MTHCA_EVENT_TYPE_SQ_DRAINED:
+			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
+				       IB_EVENT_SQ_DRAINED);
+			break;
+
+		case MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE:
+			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
+				       IB_EVENT_QP_LAST_WQE_REACHED);
+			break;
+
+		case MTHCA_EVENT_TYPE_SRQ_LIMIT:
+			mthca_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) & 0xffffff,
+					IB_EVENT_SRQ_LIMIT_REACHED);
+			break;
+
+		case MTHCA_EVENT_TYPE_WQ_CATAS_ERROR:
+			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
+				       IB_EVENT_QP_FATAL);
+			break;
+
+		case MTHCA_EVENT_TYPE_PATH_MIG_FAILED:
+			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
+				       IB_EVENT_PATH_MIG_ERR);
+			break;
+
+		case MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
+			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
+				       IB_EVENT_QP_REQ_ERR);
+			break;
+
+		case MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR:
+			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
+				       IB_EVENT_QP_ACCESS_ERR);
+			break;
+
+		case MTHCA_EVENT_TYPE_CMD:
+			mthca_cmd_event(dev,
+					be16_to_cpu(eqe->event.cmd.token),
+					eqe->event.cmd.status,
+					be64_to_cpu(eqe->event.cmd.out_param));
+			break;
+
+		case MTHCA_EVENT_TYPE_PORT_CHANGE:
+			port_change(dev,
+				    (be32_to_cpu(eqe->event.port_change.port) >> 28) & 3,
+				    eqe->subtype == 0x4);
+			break;
+
+		case MTHCA_EVENT_TYPE_CQ_ERROR:
+			mthca_warn(dev, "CQ %s on CQN %06x\n",
+				   eqe->event.cq_err.syndrome == 1 ?
+				   "overrun" : "access violation",
+				   be32_to_cpu(eqe->event.cq_err.cqn) & 0xffffff);
+			mthca_cq_event(dev, be32_to_cpu(eqe->event.cq_err.cqn),
+				       IB_EVENT_CQ_ERR);
+			break;
+
+		case MTHCA_EVENT_TYPE_EQ_OVERFLOW:
+			mthca_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
+			break;
+
+		case MTHCA_EVENT_TYPE_EEC_CATAS_ERROR:
+		case MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR:
+		case MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR:
+		case MTHCA_EVENT_TYPE_ECC_DETECT:
+		default:
+			mthca_warn(dev, "Unhandled event %02x(%02x) on EQ %d\n",
+				   eqe->type, eqe->subtype, eq->eqn);
+			break;
+		}
+
+		set_eqe_hw(eqe);
+		++eq->cons_index;
+		eqes_found = 1;
+		++set_ci;
+
+		/*
+		 * The HCA will think the queue has overflowed if we
+		 * don't tell it we've been processing events.  We
+		 * create our EQs with MTHCA_NUM_SPARE_EQE extra
+		 * entries, so we must update our consumer index at
+		 * least that often.
+		 */
+		if (unlikely(set_ci >= MTHCA_NUM_SPARE_EQE)) {
+			/*
+			 * Conditional on hca_type is OK here because
+			 * this is a rare case, not the fast path.
+			 */
+			set_eq_ci(dev, eq, eq->cons_index);
+			set_ci = 0;
+		}
+	}
+
+	/*
+	 * Rely on caller to set consumer index so that we don't have
+	 * to test hca_type in our interrupt handling fast path.
+	 */
+	return eqes_found;
+}
+
+static irqreturn_t mthca_tavor_interrupt(int irq, void *dev_ptr)
+{
+	struct mthca_dev *dev = dev_ptr;
+	u32 ecr;
+	int i;
+
+	if (dev->eq_table.clr_mask)
+		writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
+
+	ecr = readl(dev->eq_regs.tavor.ecr_base + 4);
+	if (!ecr)
+		return IRQ_NONE;
+
+	writel(ecr, dev->eq_regs.tavor.ecr_base +
+	       MTHCA_ECR_CLR_BASE - MTHCA_ECR_BASE + 4);
+
+	for (i = 0; i < MTHCA_NUM_EQ; ++i)
+		if (ecr & dev->eq_table.eq[i].eqn_mask) {
+			if (mthca_eq_int(dev, &dev->eq_table.eq[i]))
+				tavor_set_eq_ci(dev, &dev->eq_table.eq[i],
+						dev->eq_table.eq[i].cons_index);
+			tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
+		}
+
+	return IRQ_HANDLED;
+}
+
+static irqreturn_t mthca_tavor_msi_x_interrupt(int irq, void *eq_ptr)
+{
+	struct mthca_eq  *eq  = eq_ptr;
+	struct mthca_dev *dev = eq->dev;
+
+	mthca_eq_int(dev, eq);
+	tavor_set_eq_ci(dev, eq, eq->cons_index);
+	tavor_eq_req_not(dev, eq->eqn);
+
+	/* MSI-X vectors always belong to us */
+	return IRQ_HANDLED;
+}
+
+static irqreturn_t mthca_arbel_interrupt(int irq, void *dev_ptr)
+{
+	struct mthca_dev *dev = dev_ptr;
+	int work = 0;
+	int i;
+
+	if (dev->eq_table.clr_mask)
+		writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
+
+	for (i = 0; i < MTHCA_NUM_EQ; ++i)
+		if (mthca_eq_int(dev, &dev->eq_table.eq[i])) {
+			work = 1;
+			arbel_set_eq_ci(dev, &dev->eq_table.eq[i],
+					dev->eq_table.eq[i].cons_index);
+		}
+
+	arbel_eq_req_not(dev, dev->eq_table.arm_mask);
+
+	return IRQ_RETVAL(work);
+}
+
+static irqreturn_t mthca_arbel_msi_x_interrupt(int irq, void *eq_ptr)
+{
+	struct mthca_eq  *eq  = eq_ptr;
+	struct mthca_dev *dev = eq->dev;
+
+	mthca_eq_int(dev, eq);
+	arbel_set_eq_ci(dev, eq, eq->cons_index);
+	arbel_eq_req_not(dev, eq->eqn_mask);
+
+	/* MSI-X vectors always belong to us */
+	return IRQ_HANDLED;
+}
+
+static int mthca_create_eq(struct mthca_dev *dev,
+			   int nent,
+			   u8 intr,
+			   struct mthca_eq *eq)
+{
+	int npages;
+	u64 *dma_list = NULL;
+	dma_addr_t t;
+	struct mthca_mailbox *mailbox;
+	struct mthca_eq_context *eq_context;
+	int err = -ENOMEM;
+	int i;
+
+	eq->dev  = dev;
+	eq->nent = roundup_pow_of_two(max(nent, 2));
+	npages = ALIGN(eq->nent * MTHCA_EQ_ENTRY_SIZE, PAGE_SIZE) / PAGE_SIZE;
+
+	eq->page_list = kmalloc_array(npages, sizeof(*eq->page_list),
+				      GFP_KERNEL);
+	if (!eq->page_list)
+		goto err_out;
+
+	for (i = 0; i < npages; ++i)
+		eq->page_list[i].buf = NULL;
+
+	dma_list = kmalloc_array(npages, sizeof(*dma_list), GFP_KERNEL);
+	if (!dma_list)
+		goto err_out_free;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		goto err_out_free;
+	eq_context = mailbox->buf;
+
+	for (i = 0; i < npages; ++i) {
+		eq->page_list[i].buf = dma_alloc_coherent(&dev->pdev->dev,
+							  PAGE_SIZE, &t, GFP_KERNEL);
+		if (!eq->page_list[i].buf)
+			goto err_out_free_pages;
+
+		dma_list[i] = t;
+		dma_unmap_addr_set(&eq->page_list[i], mapping, t);
+
+		clear_page(eq->page_list[i].buf);
+	}
+
+	for (i = 0; i < eq->nent; ++i)
+		set_eqe_hw(get_eqe(eq, i));
+
+	eq->eqn = mthca_alloc(&dev->eq_table.alloc);
+	if (eq->eqn == -1)
+		goto err_out_free_pages;
+
+	err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
+				  dma_list, PAGE_SHIFT, npages,
+				  0, npages * PAGE_SIZE,
+				  MTHCA_MPT_FLAG_LOCAL_WRITE |
+				  MTHCA_MPT_FLAG_LOCAL_READ,
+				  &eq->mr);
+	if (err)
+		goto err_out_free_eq;
+
+	memset(eq_context, 0, sizeof *eq_context);
+	eq_context->flags           = cpu_to_be32(MTHCA_EQ_STATUS_OK   |
+						  MTHCA_EQ_OWNER_HW    |
+						  MTHCA_EQ_STATE_ARMED |
+						  MTHCA_EQ_FLAG_TR);
+	if (mthca_is_memfree(dev))
+		eq_context->flags  |= cpu_to_be32(MTHCA_EQ_STATE_ARBEL);
+
+	eq_context->logsize_usrpage = cpu_to_be32((ffs(eq->nent) - 1) << 24);
+	if (mthca_is_memfree(dev)) {
+		eq_context->arbel_pd = cpu_to_be32(dev->driver_pd.pd_num);
+	} else {
+		eq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
+		eq_context->tavor_pd         = cpu_to_be32(dev->driver_pd.pd_num);
+	}
+	eq_context->intr            = intr;
+	eq_context->lkey            = cpu_to_be32(eq->mr.ibmr.lkey);
+
+	err = mthca_SW2HW_EQ(dev, mailbox, eq->eqn);
+	if (err) {
+		mthca_warn(dev, "SW2HW_EQ returned %d\n", err);
+		goto err_out_free_mr;
+	}
+
+	kfree(dma_list);
+	mthca_free_mailbox(dev, mailbox);
+
+	eq->eqn_mask   = swab32(1 << eq->eqn);
+	eq->cons_index = 0;
+
+	dev->eq_table.arm_mask |= eq->eqn_mask;
+
+	mthca_dbg(dev, "Allocated EQ %d with %d entries\n",
+		  eq->eqn, eq->nent);
+
+	return err;
+
+ err_out_free_mr:
+	mthca_free_mr(dev, &eq->mr);
+
+ err_out_free_eq:
+	mthca_free(&dev->eq_table.alloc, eq->eqn);
+
+ err_out_free_pages:
+	for (i = 0; i < npages; ++i)
+		if (eq->page_list[i].buf)
+			dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+					  eq->page_list[i].buf,
+					  dma_unmap_addr(&eq->page_list[i],
+							 mapping));
+
+	mthca_free_mailbox(dev, mailbox);
+
+ err_out_free:
+	kfree(eq->page_list);
+	kfree(dma_list);
+
+ err_out:
+	return err;
+}
+
+static void mthca_free_eq(struct mthca_dev *dev,
+			  struct mthca_eq *eq)
+{
+	struct mthca_mailbox *mailbox;
+	int err;
+	int npages = (eq->nent * MTHCA_EQ_ENTRY_SIZE + PAGE_SIZE - 1) /
+		PAGE_SIZE;
+	int i;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return;
+
+	err = mthca_HW2SW_EQ(dev, mailbox, eq->eqn);
+	if (err)
+		mthca_warn(dev, "HW2SW_EQ returned %d\n", err);
+
+	dev->eq_table.arm_mask &= ~eq->eqn_mask;
+
+	if (0) {
+		mthca_dbg(dev, "Dumping EQ context %02x:\n", eq->eqn);
+		for (i = 0; i < sizeof (struct mthca_eq_context) / 4; ++i) {
+			if (i % 4 == 0)
+				printk("[%02x] ", i * 4);
+			printk(" %08x", be32_to_cpup(mailbox->buf + i * 4));
+			if ((i + 1) % 4 == 0)
+				printk("\n");
+		}
+	}
+
+	mthca_free_mr(dev, &eq->mr);
+	for (i = 0; i < npages; ++i)
+		dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+				  eq->page_list[i].buf,
+				  dma_unmap_addr(&eq->page_list[i], mapping));
+
+	kfree(eq->page_list);
+	mthca_free_mailbox(dev, mailbox);
+}
+
+static void mthca_free_irqs(struct mthca_dev *dev)
+{
+	int i;
+
+	if (dev->eq_table.have_irq)
+		free_irq(dev->pdev->irq, dev);
+	for (i = 0; i < MTHCA_NUM_EQ; ++i)
+		if (dev->eq_table.eq[i].have_irq) {
+			free_irq(dev->eq_table.eq[i].msi_x_vector,
+				 dev->eq_table.eq + i);
+			dev->eq_table.eq[i].have_irq = 0;
+		}
+}
+
+static int mthca_map_reg(struct mthca_dev *dev,
+			 unsigned long offset, unsigned long size,
+			 void __iomem **map)
+{
+	phys_addr_t base = pci_resource_start(dev->pdev, 0);
+
+	*map = ioremap(base + offset, size);
+	if (!*map)
+		return -ENOMEM;
+
+	return 0;
+}
+
+static int mthca_map_eq_regs(struct mthca_dev *dev)
+{
+	if (mthca_is_memfree(dev)) {
+		/*
+		 * We assume that the EQ arm and EQ set CI registers
+		 * fall within the first BAR.  We can't trust the
+		 * values firmware gives us, since those addresses are
+		 * valid on the HCA's side of the PCI bus but not
+		 * necessarily the host side.
+		 */
+		if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
+				  dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
+				  &dev->clr_base)) {
+			mthca_err(dev, "Couldn't map interrupt clear register, "
+				  "aborting.\n");
+			return -ENOMEM;
+		}
+
+		/*
+		 * Add 4 because we limit ourselves to EQs 0 ... 31,
+		 * so we only need the low word of the register.
+		 */
+		if (mthca_map_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
+					dev->fw.arbel.eq_arm_base) + 4, 4,
+				  &dev->eq_regs.arbel.eq_arm)) {
+			mthca_err(dev, "Couldn't map EQ arm register, aborting.\n");
+			iounmap(dev->clr_base);
+			return -ENOMEM;
+		}
+
+		if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
+				  dev->fw.arbel.eq_set_ci_base,
+				  MTHCA_EQ_SET_CI_SIZE,
+				  &dev->eq_regs.arbel.eq_set_ci_base)) {
+			mthca_err(dev, "Couldn't map EQ CI register, aborting.\n");
+			iounmap(dev->eq_regs.arbel.eq_arm);
+			iounmap(dev->clr_base);
+			return -ENOMEM;
+		}
+	} else {
+		if (mthca_map_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
+				  &dev->clr_base)) {
+			mthca_err(dev, "Couldn't map interrupt clear register, "
+				  "aborting.\n");
+			return -ENOMEM;
+		}
+
+		if (mthca_map_reg(dev, MTHCA_ECR_BASE,
+				  MTHCA_ECR_SIZE + MTHCA_ECR_CLR_SIZE,
+				  &dev->eq_regs.tavor.ecr_base)) {
+			mthca_err(dev, "Couldn't map ecr register, "
+				  "aborting.\n");
+			iounmap(dev->clr_base);
+			return -ENOMEM;
+		}
+	}
+
+	return 0;
+
+}
+
+static void mthca_unmap_eq_regs(struct mthca_dev *dev)
+{
+	if (mthca_is_memfree(dev)) {
+		iounmap(dev->eq_regs.arbel.eq_set_ci_base);
+		iounmap(dev->eq_regs.arbel.eq_arm);
+		iounmap(dev->clr_base);
+	} else {
+		iounmap(dev->eq_regs.tavor.ecr_base);
+		iounmap(dev->clr_base);
+	}
+}
+
+int mthca_map_eq_icm(struct mthca_dev *dev, u64 icm_virt)
+{
+	int ret;
+
+	/*
+	 * We assume that mapping one page is enough for the whole EQ
+	 * context table.  This is fine with all current HCAs, because
+	 * we only use 32 EQs and each EQ uses 32 bytes of context
+	 * memory, or 1 KB total.
+	 */
+	dev->eq_table.icm_virt = icm_virt;
+	dev->eq_table.icm_page = alloc_page(GFP_HIGHUSER);
+	if (!dev->eq_table.icm_page)
+		return -ENOMEM;
+	dev->eq_table.icm_dma =
+		dma_map_page(&dev->pdev->dev, dev->eq_table.icm_page, 0,
+			     PAGE_SIZE, DMA_BIDIRECTIONAL);
+	if (dma_mapping_error(&dev->pdev->dev, dev->eq_table.icm_dma)) {
+		__free_page(dev->eq_table.icm_page);
+		return -ENOMEM;
+	}
+
+	ret = mthca_MAP_ICM_page(dev, dev->eq_table.icm_dma, icm_virt);
+	if (ret) {
+		dma_unmap_page(&dev->pdev->dev, dev->eq_table.icm_dma,
+			       PAGE_SIZE, DMA_BIDIRECTIONAL);
+		__free_page(dev->eq_table.icm_page);
+	}
+
+	return ret;
+}
+
+void mthca_unmap_eq_icm(struct mthca_dev *dev)
+{
+	mthca_UNMAP_ICM(dev, dev->eq_table.icm_virt, 1);
+	dma_unmap_page(&dev->pdev->dev, dev->eq_table.icm_dma, PAGE_SIZE,
+		       DMA_BIDIRECTIONAL);
+	__free_page(dev->eq_table.icm_page);
+}
+
+int mthca_init_eq_table(struct mthca_dev *dev)
+{
+	int err;
+	u8 intr;
+	int i;
+
+	err = mthca_alloc_init(&dev->eq_table.alloc,
+			       dev->limits.num_eqs,
+			       dev->limits.num_eqs - 1,
+			       dev->limits.reserved_eqs);
+	if (err)
+		return err;
+
+	err = mthca_map_eq_regs(dev);
+	if (err)
+		goto err_out_free;
+
+	if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
+		dev->eq_table.clr_mask = 0;
+	} else {
+		dev->eq_table.clr_mask =
+			swab32(1 << (dev->eq_table.inta_pin & 31));
+		dev->eq_table.clr_int  = dev->clr_base +
+			(dev->eq_table.inta_pin < 32 ? 4 : 0);
+	}
+
+	dev->eq_table.arm_mask = 0;
+
+	intr = dev->eq_table.inta_pin;
+
+	err = mthca_create_eq(dev, dev->limits.num_cqs + MTHCA_NUM_SPARE_EQE,
+			      (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 128 : intr,
+			      &dev->eq_table.eq[MTHCA_EQ_COMP]);
+	if (err)
+		goto err_out_unmap;
+
+	err = mthca_create_eq(dev, MTHCA_NUM_ASYNC_EQE + MTHCA_NUM_SPARE_EQE,
+			      (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 129 : intr,
+			      &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
+	if (err)
+		goto err_out_comp;
+
+	err = mthca_create_eq(dev, MTHCA_NUM_CMD_EQE + MTHCA_NUM_SPARE_EQE,
+			      (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 130 : intr,
+			      &dev->eq_table.eq[MTHCA_EQ_CMD]);
+	if (err)
+		goto err_out_async;
+
+	if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
+		static const char *eq_name[] = {
+			[MTHCA_EQ_COMP]  = DRV_NAME "-comp",
+			[MTHCA_EQ_ASYNC] = DRV_NAME "-async",
+			[MTHCA_EQ_CMD]   = DRV_NAME "-cmd"
+		};
+
+		for (i = 0; i < MTHCA_NUM_EQ; ++i) {
+			snprintf(dev->eq_table.eq[i].irq_name,
+				 IB_DEVICE_NAME_MAX,
+				 "%s@pci:%s", eq_name[i],
+				 pci_name(dev->pdev));
+			err = request_irq(dev->eq_table.eq[i].msi_x_vector,
+					  mthca_is_memfree(dev) ?
+					  mthca_arbel_msi_x_interrupt :
+					  mthca_tavor_msi_x_interrupt,
+					  0, dev->eq_table.eq[i].irq_name,
+					  dev->eq_table.eq + i);
+			if (err)
+				goto err_out_cmd;
+			dev->eq_table.eq[i].have_irq = 1;
+		}
+	} else {
+		snprintf(dev->eq_table.eq[0].irq_name, IB_DEVICE_NAME_MAX,
+			 DRV_NAME "@pci:%s", pci_name(dev->pdev));
+		err = request_irq(dev->pdev->irq,
+				  mthca_is_memfree(dev) ?
+				  mthca_arbel_interrupt :
+				  mthca_tavor_interrupt,
+				  IRQF_SHARED, dev->eq_table.eq[0].irq_name, dev);
+		if (err)
+			goto err_out_cmd;
+		dev->eq_table.have_irq = 1;
+	}
+
+	err = mthca_MAP_EQ(dev, async_mask(dev),
+			   0, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
+	if (err)
+		mthca_warn(dev, "MAP_EQ for async EQ %d failed (%d)\n",
+			   dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, err);
+
+	err = mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
+			   0, dev->eq_table.eq[MTHCA_EQ_CMD].eqn);
+	if (err)
+		mthca_warn(dev, "MAP_EQ for cmd EQ %d failed (%d)\n",
+			   dev->eq_table.eq[MTHCA_EQ_CMD].eqn, err);
+
+	for (i = 0; i < MTHCA_NUM_EQ; ++i)
+		if (mthca_is_memfree(dev))
+			arbel_eq_req_not(dev, dev->eq_table.eq[i].eqn_mask);
+		else
+			tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
+
+	return 0;
+
+err_out_cmd:
+	mthca_free_irqs(dev);
+	mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_CMD]);
+
+err_out_async:
+	mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
+
+err_out_comp:
+	mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_COMP]);
+
+err_out_unmap:
+	mthca_unmap_eq_regs(dev);
+
+err_out_free:
+	mthca_alloc_cleanup(&dev->eq_table.alloc);
+	return err;
+}
+
+void mthca_cleanup_eq_table(struct mthca_dev *dev)
+{
+	int i;
+
+	mthca_free_irqs(dev);
+
+	mthca_MAP_EQ(dev, async_mask(dev),
+		     1, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
+	mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
+		     1, dev->eq_table.eq[MTHCA_EQ_CMD].eqn);
+
+	for (i = 0; i < MTHCA_NUM_EQ; ++i)
+		mthca_free_eq(dev, &dev->eq_table.eq[i]);
+
+	mthca_unmap_eq_regs(dev);
+
+	mthca_alloc_cleanup(&dev->eq_table.alloc);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_mad.c b/drivers/infiniband/hw/mthca/mthca_mad.c
new file mode 100644
index 000000000..042527007
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_mad.c
@@ -0,0 +1,341 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/string.h>
+#include <linux/slab.h>
+
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_mad.h>
+#include <rdma/ib_smi.h>
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+
+enum {
+	MTHCA_VENDOR_CLASS1 = 0x9,
+	MTHCA_VENDOR_CLASS2 = 0xa
+};
+
+static int mthca_update_rate(struct mthca_dev *dev, u8 port_num)
+{
+	struct ib_port_attr *tprops = NULL;
+	int                  ret;
+
+	tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
+	if (!tprops)
+		return -ENOMEM;
+
+	ret = ib_query_port(&dev->ib_dev, port_num, tprops);
+	if (ret) {
+		dev_warn(&dev->ib_dev.dev,
+			 "ib_query_port failed (%d) forport %d\n", ret,
+			 port_num);
+		goto out;
+	}
+
+	dev->rate[port_num - 1] = tprops->active_speed *
+				  ib_width_enum_to_int(tprops->active_width);
+
+out:
+	kfree(tprops);
+	return ret;
+}
+
+static void update_sm_ah(struct mthca_dev *dev,
+			 u8 port_num, u16 lid, u8 sl)
+{
+	struct ib_ah *new_ah;
+	struct rdma_ah_attr ah_attr;
+	unsigned long flags;
+
+	if (!dev->send_agent[port_num - 1][0])
+		return;
+
+	memset(&ah_attr, 0, sizeof ah_attr);
+	ah_attr.type = rdma_ah_find_type(&dev->ib_dev, port_num);
+	rdma_ah_set_dlid(&ah_attr, lid);
+	rdma_ah_set_sl(&ah_attr, sl);
+	rdma_ah_set_port_num(&ah_attr, port_num);
+
+	new_ah = rdma_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
+				&ah_attr, 0);
+	if (IS_ERR(new_ah))
+		return;
+
+	spin_lock_irqsave(&dev->sm_lock, flags);
+	if (dev->sm_ah[port_num - 1])
+		rdma_destroy_ah(dev->sm_ah[port_num - 1], 0);
+	dev->sm_ah[port_num - 1] = new_ah;
+	spin_unlock_irqrestore(&dev->sm_lock, flags);
+}
+
+/*
+ * Snoop SM MADs for port info and P_Key table sets, so we can
+ * synthesize LID change and P_Key change events.
+ */
+static void smp_snoop(struct ib_device *ibdev,
+		      u8 port_num,
+		      const struct ib_mad *mad,
+		      u16 prev_lid)
+{
+	struct ib_event event;
+
+	if ((mad->mad_hdr.mgmt_class  == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
+	     mad->mad_hdr.mgmt_class  == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
+	    mad->mad_hdr.method     == IB_MGMT_METHOD_SET) {
+		if (mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO) {
+			struct ib_port_info *pinfo =
+				(struct ib_port_info *) ((struct ib_smp *) mad)->data;
+			u16 lid = be16_to_cpu(pinfo->lid);
+
+			mthca_update_rate(to_mdev(ibdev), port_num);
+			update_sm_ah(to_mdev(ibdev), port_num,
+				     be16_to_cpu(pinfo->sm_lid),
+				     pinfo->neighbormtu_mastersmsl & 0xf);
+
+			event.device           = ibdev;
+			event.element.port_num = port_num;
+
+			if (pinfo->clientrereg_resv_subnetto & 0x80) {
+				event.event    = IB_EVENT_CLIENT_REREGISTER;
+				ib_dispatch_event(&event);
+			}
+
+			if (prev_lid != lid) {
+				event.event    = IB_EVENT_LID_CHANGE;
+				ib_dispatch_event(&event);
+			}
+		}
+
+		if (mad->mad_hdr.attr_id == IB_SMP_ATTR_PKEY_TABLE) {
+			event.device           = ibdev;
+			event.event            = IB_EVENT_PKEY_CHANGE;
+			event.element.port_num = port_num;
+			ib_dispatch_event(&event);
+		}
+	}
+}
+
+static void node_desc_override(struct ib_device *dev,
+			       struct ib_mad *mad)
+{
+	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
+	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
+	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
+	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
+		mutex_lock(&to_mdev(dev)->cap_mask_mutex);
+		memcpy(((struct ib_smp *) mad)->data, dev->node_desc,
+		       IB_DEVICE_NODE_DESC_MAX);
+		mutex_unlock(&to_mdev(dev)->cap_mask_mutex);
+	}
+}
+
+static void forward_trap(struct mthca_dev *dev,
+			 u32 port_num,
+			 const struct ib_mad *mad)
+{
+	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
+	struct ib_mad_send_buf *send_buf;
+	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
+	int ret;
+	unsigned long flags;
+
+	if (agent) {
+		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
+					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
+					      IB_MGMT_BASE_VERSION);
+		if (IS_ERR(send_buf))
+			return;
+		/*
+		 * We rely here on the fact that MLX QPs don't use the
+		 * address handle after the send is posted (this is
+		 * wrong following the IB spec strictly, but we know
+		 * it's OK for our devices).
+		 */
+		spin_lock_irqsave(&dev->sm_lock, flags);
+		memcpy(send_buf->mad, mad, sizeof *mad);
+		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
+			ret = ib_post_send_mad(send_buf, NULL);
+		else
+			ret = -EINVAL;
+		spin_unlock_irqrestore(&dev->sm_lock, flags);
+
+		if (ret)
+			ib_free_send_mad(send_buf);
+	}
+}
+
+int mthca_process_mad(struct ib_device *ibdev, int mad_flags, u32 port_num,
+		      const struct ib_wc *in_wc, const struct ib_grh *in_grh,
+		      const struct ib_mad *in, struct ib_mad *out,
+		      size_t *out_mad_size, u16 *out_mad_pkey_index)
+{
+	int err;
+	u16 slid = in_wc ? ib_lid_cpu16(in_wc->slid) : be16_to_cpu(IB_LID_PERMISSIVE);
+	u16 prev_lid = 0;
+	struct ib_port_attr pattr;
+
+	/* Forward locally generated traps to the SM */
+	if (in->mad_hdr.method == IB_MGMT_METHOD_TRAP && !slid) {
+		forward_trap(to_mdev(ibdev), port_num, in);
+		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
+	}
+
+	/*
+	 * Only handle SM gets, sets and trap represses for SM class
+	 *
+	 * Only handle PMA and Mellanox vendor-specific class gets and
+	 * sets for other classes.
+	 */
+	if (in->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
+	    in->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
+		if (in->mad_hdr.method != IB_MGMT_METHOD_GET &&
+		    in->mad_hdr.method != IB_MGMT_METHOD_SET &&
+		    in->mad_hdr.method != IB_MGMT_METHOD_TRAP_REPRESS)
+			return IB_MAD_RESULT_SUCCESS;
+
+		/*
+		 * Don't process SMInfo queries or vendor-specific
+		 * MADs -- the SMA can't handle them.
+		 */
+		if (in->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO ||
+		    ((in->mad_hdr.attr_id & IB_SMP_ATTR_VENDOR_MASK) ==
+		     IB_SMP_ATTR_VENDOR_MASK))
+			return IB_MAD_RESULT_SUCCESS;
+	} else if (in->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
+		   in->mad_hdr.mgmt_class == MTHCA_VENDOR_CLASS1     ||
+		   in->mad_hdr.mgmt_class == MTHCA_VENDOR_CLASS2) {
+		if (in->mad_hdr.method != IB_MGMT_METHOD_GET &&
+		    in->mad_hdr.method != IB_MGMT_METHOD_SET)
+			return IB_MAD_RESULT_SUCCESS;
+	} else
+		return IB_MAD_RESULT_SUCCESS;
+	if ((in->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
+	     in->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
+	    in->mad_hdr.method == IB_MGMT_METHOD_SET &&
+	    in->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
+	    !ib_query_port(ibdev, port_num, &pattr))
+		prev_lid = ib_lid_cpu16(pattr.lid);
+
+	err = mthca_MAD_IFC(to_mdev(ibdev), mad_flags & IB_MAD_IGNORE_MKEY,
+			    mad_flags & IB_MAD_IGNORE_BKEY, port_num, in_wc,
+			    in_grh, in, out);
+	if (err == -EBADMSG)
+		return IB_MAD_RESULT_SUCCESS;
+	else if (err) {
+		mthca_err(to_mdev(ibdev), "MAD_IFC returned %d\n", err);
+		return IB_MAD_RESULT_FAILURE;
+	}
+
+	if (!out->mad_hdr.status) {
+		smp_snoop(ibdev, port_num, in, prev_lid);
+		node_desc_override(ibdev, out);
+	}
+
+	/* set return bit in status of directed route responses */
+	if (in->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
+		out->mad_hdr.status |= cpu_to_be16(1 << 15);
+
+	if (in->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
+		/* no response for trap repress */
+		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
+
+	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
+}
+
+static void send_handler(struct ib_mad_agent *agent,
+			 struct ib_mad_send_wc *mad_send_wc)
+{
+	ib_free_send_mad(mad_send_wc->send_buf);
+}
+
+int mthca_create_agents(struct mthca_dev *dev)
+{
+	struct ib_mad_agent *agent;
+	int p, q;
+	int ret;
+
+	spin_lock_init(&dev->sm_lock);
+
+	for (p = 0; p < dev->limits.num_ports; ++p)
+		for (q = 0; q <= 1; ++q) {
+			agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
+						      q ? IB_QPT_GSI : IB_QPT_SMI,
+						      NULL, 0, send_handler,
+						      NULL, NULL, 0);
+			if (IS_ERR(agent)) {
+				ret = PTR_ERR(agent);
+				goto err;
+			}
+			dev->send_agent[p][q] = agent;
+		}
+
+
+	for (p = 1; p <= dev->limits.num_ports; ++p) {
+		ret = mthca_update_rate(dev, p);
+		if (ret) {
+			mthca_err(dev, "Failed to obtain port %d rate."
+				  " aborting.\n", p);
+			goto err;
+		}
+	}
+
+	return 0;
+
+err:
+	for (p = 0; p < dev->limits.num_ports; ++p)
+		for (q = 0; q <= 1; ++q)
+			if (dev->send_agent[p][q])
+				ib_unregister_mad_agent(dev->send_agent[p][q]);
+
+	return ret;
+}
+
+void mthca_free_agents(struct mthca_dev *dev)
+{
+	struct ib_mad_agent *agent;
+	int p, q;
+
+	for (p = 0; p < dev->limits.num_ports; ++p) {
+		for (q = 0; q <= 1; ++q) {
+			agent = dev->send_agent[p][q];
+			dev->send_agent[p][q] = NULL;
+			ib_unregister_mad_agent(agent);
+		}
+
+		if (dev->sm_ah[p])
+			rdma_destroy_ah(dev->sm_ah[p],
+					RDMA_DESTROY_AH_SLEEPABLE);
+	}
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_main.c b/drivers/infiniband/hw/mthca/mthca_main.c
new file mode 100644
index 000000000..1ab268b77
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_main.c
@@ -0,0 +1,1257 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/errno.h>
+#include <linux/pci.h>
+#include <linux/interrupt.h>
+#include <linux/gfp.h>
+
+#include "mthca_dev.h"
+#include "mthca_config_reg.h"
+#include "mthca_cmd.h"
+#include "mthca_profile.h"
+#include "mthca_memfree.h"
+#include "mthca_wqe.h"
+
+MODULE_AUTHOR("Roland Dreier");
+MODULE_DESCRIPTION("Mellanox InfiniBand HCA low-level driver");
+MODULE_LICENSE("Dual BSD/GPL");
+
+#ifdef CONFIG_INFINIBAND_MTHCA_DEBUG
+
+int mthca_debug_level = 0;
+module_param_named(debug_level, mthca_debug_level, int, 0644);
+MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
+
+#endif /* CONFIG_INFINIBAND_MTHCA_DEBUG */
+
+#ifdef CONFIG_PCI_MSI
+
+static int msi_x = 1;
+module_param(msi_x, int, 0444);
+MODULE_PARM_DESC(msi_x, "attempt to use MSI-X if nonzero");
+
+#else /* CONFIG_PCI_MSI */
+
+#define msi_x (0)
+
+#endif /* CONFIG_PCI_MSI */
+
+static int tune_pci = 0;
+module_param(tune_pci, int, 0444);
+MODULE_PARM_DESC(tune_pci, "increase PCI burst from the default set by BIOS if nonzero");
+
+DEFINE_MUTEX(mthca_device_mutex);
+
+#define MTHCA_DEFAULT_NUM_QP            (1 << 16)
+#define MTHCA_DEFAULT_RDB_PER_QP        (1 << 2)
+#define MTHCA_DEFAULT_NUM_CQ            (1 << 16)
+#define MTHCA_DEFAULT_NUM_MCG           (1 << 13)
+#define MTHCA_DEFAULT_NUM_MPT           (1 << 17)
+#define MTHCA_DEFAULT_NUM_MTT           (1 << 20)
+#define MTHCA_DEFAULT_NUM_UDAV          (1 << 15)
+#define MTHCA_DEFAULT_NUM_RESERVED_MTTS (1 << 18)
+#define MTHCA_DEFAULT_NUM_UARC_SIZE     (1 << 18)
+
+static struct mthca_profile hca_profile = {
+	.num_qp             = MTHCA_DEFAULT_NUM_QP,
+	.rdb_per_qp         = MTHCA_DEFAULT_RDB_PER_QP,
+	.num_cq             = MTHCA_DEFAULT_NUM_CQ,
+	.num_mcg            = MTHCA_DEFAULT_NUM_MCG,
+	.num_mpt            = MTHCA_DEFAULT_NUM_MPT,
+	.num_mtt            = MTHCA_DEFAULT_NUM_MTT,
+	.num_udav           = MTHCA_DEFAULT_NUM_UDAV,          /* Tavor only */
+	.fmr_reserved_mtts  = MTHCA_DEFAULT_NUM_RESERVED_MTTS, /* Tavor only */
+	.uarc_size          = MTHCA_DEFAULT_NUM_UARC_SIZE,     /* Arbel only */
+};
+
+module_param_named(num_qp, hca_profile.num_qp, int, 0444);
+MODULE_PARM_DESC(num_qp, "maximum number of QPs per HCA");
+
+module_param_named(rdb_per_qp, hca_profile.rdb_per_qp, int, 0444);
+MODULE_PARM_DESC(rdb_per_qp, "number of RDB buffers per QP");
+
+module_param_named(num_cq, hca_profile.num_cq, int, 0444);
+MODULE_PARM_DESC(num_cq, "maximum number of CQs per HCA");
+
+module_param_named(num_mcg, hca_profile.num_mcg, int, 0444);
+MODULE_PARM_DESC(num_mcg, "maximum number of multicast groups per HCA");
+
+module_param_named(num_mpt, hca_profile.num_mpt, int, 0444);
+MODULE_PARM_DESC(num_mpt,
+		"maximum number of memory protection table entries per HCA");
+
+module_param_named(num_mtt, hca_profile.num_mtt, int, 0444);
+MODULE_PARM_DESC(num_mtt,
+		 "maximum number of memory translation table segments per HCA");
+
+module_param_named(num_udav, hca_profile.num_udav, int, 0444);
+MODULE_PARM_DESC(num_udav, "maximum number of UD address vectors per HCA");
+
+module_param_named(fmr_reserved_mtts, hca_profile.fmr_reserved_mtts, int, 0444);
+MODULE_PARM_DESC(fmr_reserved_mtts,
+		 "number of memory translation table segments reserved for FMR");
+
+static int log_mtts_per_seg = ilog2(MTHCA_MTT_SEG_SIZE / 8);
+module_param_named(log_mtts_per_seg, log_mtts_per_seg, int, 0444);
+MODULE_PARM_DESC(log_mtts_per_seg, "Log2 number of MTT entries per segment (1-5)");
+
+static char mthca_version[] =
+	DRV_NAME ": Mellanox InfiniBand HCA driver v"
+	DRV_VERSION " (" DRV_RELDATE ")\n";
+
+static int mthca_tune_pci(struct mthca_dev *mdev)
+{
+	if (!tune_pci)
+		return 0;
+
+	/* First try to max out Read Byte Count */
+	if (pci_find_capability(mdev->pdev, PCI_CAP_ID_PCIX)) {
+		if (pcix_set_mmrbc(mdev->pdev, pcix_get_max_mmrbc(mdev->pdev))) {
+			mthca_err(mdev, "Couldn't set PCI-X max read count, "
+				"aborting.\n");
+			return -ENODEV;
+		}
+	} else if (!(mdev->mthca_flags & MTHCA_FLAG_PCIE))
+		mthca_info(mdev, "No PCI-X capability, not setting RBC.\n");
+
+	if (pci_is_pcie(mdev->pdev)) {
+		if (pcie_set_readrq(mdev->pdev, 4096)) {
+			mthca_err(mdev, "Couldn't write PCI Express read request, "
+				"aborting.\n");
+			return -ENODEV;
+		}
+	} else if (mdev->mthca_flags & MTHCA_FLAG_PCIE)
+		mthca_info(mdev, "No PCI Express capability, "
+			   "not setting Max Read Request Size.\n");
+
+	return 0;
+}
+
+static int mthca_dev_lim(struct mthca_dev *mdev, struct mthca_dev_lim *dev_lim)
+{
+	int err;
+
+	mdev->limits.mtt_seg_size = (1 << log_mtts_per_seg) * 8;
+	err = mthca_QUERY_DEV_LIM(mdev, dev_lim);
+	if (err) {
+		mthca_err(mdev, "QUERY_DEV_LIM command returned %d"
+				", aborting.\n", err);
+		return err;
+	}
+	if (dev_lim->min_page_sz > PAGE_SIZE) {
+		mthca_err(mdev, "HCA minimum page size of %d bigger than "
+			  "kernel PAGE_SIZE of %ld, aborting.\n",
+			  dev_lim->min_page_sz, PAGE_SIZE);
+		return -ENODEV;
+	}
+	if (dev_lim->num_ports > MTHCA_MAX_PORTS) {
+		mthca_err(mdev, "HCA has %d ports, but we only support %d, "
+			  "aborting.\n",
+			  dev_lim->num_ports, MTHCA_MAX_PORTS);
+		return -ENODEV;
+	}
+
+	if (dev_lim->uar_size > pci_resource_len(mdev->pdev, 2)) {
+		mthca_err(mdev, "HCA reported UAR size of 0x%x bigger than "
+			  "PCI resource 2 size of 0x%llx, aborting.\n",
+			  dev_lim->uar_size,
+			  (unsigned long long)pci_resource_len(mdev->pdev, 2));
+		return -ENODEV;
+	}
+
+	mdev->limits.num_ports      	= dev_lim->num_ports;
+	mdev->limits.vl_cap             = dev_lim->max_vl;
+	mdev->limits.mtu_cap            = dev_lim->max_mtu;
+	mdev->limits.gid_table_len  	= dev_lim->max_gids;
+	mdev->limits.pkey_table_len 	= dev_lim->max_pkeys;
+	mdev->limits.local_ca_ack_delay = dev_lim->local_ca_ack_delay;
+	/*
+	 * Need to allow for worst case send WQE overhead and check
+	 * whether max_desc_sz imposes a lower limit than max_sg; UD
+	 * send has the biggest overhead.
+	 */
+	mdev->limits.max_sg		= min_t(int, dev_lim->max_sg,
+					      (dev_lim->max_desc_sz -
+					       sizeof (struct mthca_next_seg) -
+					       (mthca_is_memfree(mdev) ?
+						sizeof (struct mthca_arbel_ud_seg) :
+						sizeof (struct mthca_tavor_ud_seg))) /
+						sizeof (struct mthca_data_seg));
+	mdev->limits.max_wqes           = dev_lim->max_qp_sz;
+	mdev->limits.max_qp_init_rdma   = dev_lim->max_requester_per_qp;
+	mdev->limits.reserved_qps       = dev_lim->reserved_qps;
+	mdev->limits.max_srq_wqes       = dev_lim->max_srq_sz;
+	mdev->limits.reserved_srqs      = dev_lim->reserved_srqs;
+	mdev->limits.reserved_eecs      = dev_lim->reserved_eecs;
+	mdev->limits.max_desc_sz        = dev_lim->max_desc_sz;
+	mdev->limits.max_srq_sge	= mthca_max_srq_sge(mdev);
+	/*
+	 * Subtract 1 from the limit because we need to allocate a
+	 * spare CQE so the HCA HW can tell the difference between an
+	 * empty CQ and a full CQ.
+	 */
+	mdev->limits.max_cqes           = dev_lim->max_cq_sz - 1;
+	mdev->limits.reserved_cqs       = dev_lim->reserved_cqs;
+	mdev->limits.reserved_eqs       = dev_lim->reserved_eqs;
+	mdev->limits.reserved_mtts      = dev_lim->reserved_mtts;
+	mdev->limits.reserved_mrws      = dev_lim->reserved_mrws;
+	mdev->limits.reserved_uars      = dev_lim->reserved_uars;
+	mdev->limits.reserved_pds       = dev_lim->reserved_pds;
+	mdev->limits.port_width_cap     = dev_lim->max_port_width;
+	mdev->limits.page_size_cap      = ~(u32) (dev_lim->min_page_sz - 1);
+	mdev->limits.flags              = dev_lim->flags;
+	/*
+	 * For old FW that doesn't return static rate support, use a
+	 * value of 0x3 (only static rate values of 0 or 1 are handled),
+	 * except on Sinai, where even old FW can handle static rate
+	 * values of 2 and 3.
+	 */
+	if (dev_lim->stat_rate_support)
+		mdev->limits.stat_rate_support = dev_lim->stat_rate_support;
+	else if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
+		mdev->limits.stat_rate_support = 0xf;
+	else
+		mdev->limits.stat_rate_support = 0x3;
+
+	/* IB_DEVICE_RESIZE_MAX_WR not supported by driver.
+	   May be doable since hardware supports it for SRQ.
+
+	   IB_DEVICE_N_NOTIFY_CQ is supported by hardware but not by driver.
+
+	   IB_DEVICE_SRQ_RESIZE is supported by hardware but SRQ is not
+	   supported by driver. */
+	mdev->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
+		IB_DEVICE_PORT_ACTIVE_EVENT |
+		IB_DEVICE_SYS_IMAGE_GUID |
+		IB_DEVICE_RC_RNR_NAK_GEN;
+
+	if (dev_lim->flags & DEV_LIM_FLAG_BAD_PKEY_CNTR)
+		mdev->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
+
+	if (dev_lim->flags & DEV_LIM_FLAG_BAD_QKEY_CNTR)
+		mdev->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
+
+	if (dev_lim->flags & DEV_LIM_FLAG_RAW_MULTI)
+		mdev->device_cap_flags |= IB_DEVICE_RAW_MULTI;
+
+	if (dev_lim->flags & DEV_LIM_FLAG_AUTO_PATH_MIG)
+		mdev->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
+
+	if (dev_lim->flags & DEV_LIM_FLAG_UD_AV_PORT_ENFORCE)
+		mdev->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
+
+	if (dev_lim->flags & DEV_LIM_FLAG_SRQ)
+		mdev->mthca_flags |= MTHCA_FLAG_SRQ;
+
+	if (mthca_is_memfree(mdev))
+		if (dev_lim->flags & DEV_LIM_FLAG_IPOIB_CSUM)
+			mdev->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
+
+	return 0;
+}
+
+static int mthca_init_tavor(struct mthca_dev *mdev)
+{
+	s64 size;
+	int err;
+	struct mthca_dev_lim        dev_lim;
+	struct mthca_profile        profile;
+	struct mthca_init_hca_param init_hca;
+
+	err = mthca_SYS_EN(mdev);
+	if (err) {
+		mthca_err(mdev, "SYS_EN command returned %d, aborting.\n", err);
+		return err;
+	}
+
+	err = mthca_QUERY_FW(mdev);
+	if (err) {
+		mthca_err(mdev, "QUERY_FW command returned %d,"
+				" aborting.\n", err);
+		goto err_disable;
+	}
+	err = mthca_QUERY_DDR(mdev);
+	if (err) {
+		mthca_err(mdev, "QUERY_DDR command returned %d, aborting.\n", err);
+		goto err_disable;
+	}
+
+	err = mthca_dev_lim(mdev, &dev_lim);
+	if (err) {
+		mthca_err(mdev, "QUERY_DEV_LIM command returned %d, aborting.\n", err);
+		goto err_disable;
+	}
+
+	profile = hca_profile;
+	profile.num_uar   = dev_lim.uar_size / PAGE_SIZE;
+	profile.uarc_size = 0;
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
+		profile.num_srq = dev_lim.max_srqs;
+
+	size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
+	if (size < 0) {
+		err = size;
+		goto err_disable;
+	}
+
+	err = mthca_INIT_HCA(mdev, &init_hca);
+	if (err) {
+		mthca_err(mdev, "INIT_HCA command returned %d, aborting.\n", err);
+		goto err_disable;
+	}
+
+	return 0;
+
+err_disable:
+	mthca_SYS_DIS(mdev);
+
+	return err;
+}
+
+static int mthca_load_fw(struct mthca_dev *mdev)
+{
+	int err;
+
+	/* FIXME: use HCA-attached memory for FW if present */
+
+	mdev->fw.arbel.fw_icm =
+		mthca_alloc_icm(mdev, mdev->fw.arbel.fw_pages,
+				GFP_HIGHUSER | __GFP_NOWARN, 0);
+	if (!mdev->fw.arbel.fw_icm) {
+		mthca_err(mdev, "Couldn't allocate FW area, aborting.\n");
+		return -ENOMEM;
+	}
+
+	err = mthca_MAP_FA(mdev, mdev->fw.arbel.fw_icm);
+	if (err) {
+		mthca_err(mdev, "MAP_FA command returned %d, aborting.\n", err);
+		goto err_free;
+	}
+	err = mthca_RUN_FW(mdev);
+	if (err) {
+		mthca_err(mdev, "RUN_FW command returned %d, aborting.\n", err);
+		goto err_unmap_fa;
+	}
+
+	return 0;
+
+err_unmap_fa:
+	mthca_UNMAP_FA(mdev);
+
+err_free:
+	mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
+	return err;
+}
+
+static int mthca_init_icm(struct mthca_dev *mdev,
+			  struct mthca_dev_lim *dev_lim,
+			  struct mthca_init_hca_param *init_hca,
+			  u64 icm_size)
+{
+	u64 aux_pages = 0;
+	int err;
+
+	err = mthca_SET_ICM_SIZE(mdev, icm_size, &aux_pages);
+	if (err) {
+		mthca_err(mdev, "SET_ICM_SIZE command returned %d, aborting.\n", err);
+		return err;
+	}
+
+	mthca_dbg(mdev, "%lld KB of HCA context requires %lld KB aux memory.\n",
+		  (unsigned long long) icm_size >> 10,
+		  (unsigned long long) aux_pages << 2);
+
+	mdev->fw.arbel.aux_icm = mthca_alloc_icm(mdev, aux_pages,
+						 GFP_HIGHUSER | __GFP_NOWARN, 0);
+	if (!mdev->fw.arbel.aux_icm) {
+		mthca_err(mdev, "Couldn't allocate aux memory, aborting.\n");
+		return -ENOMEM;
+	}
+
+	err = mthca_MAP_ICM_AUX(mdev, mdev->fw.arbel.aux_icm);
+	if (err) {
+		mthca_err(mdev, "MAP_ICM_AUX returned %d, aborting.\n", err);
+		goto err_free_aux;
+	}
+
+	err = mthca_map_eq_icm(mdev, init_hca->eqc_base);
+	if (err) {
+		mthca_err(mdev, "Failed to map EQ context memory, aborting.\n");
+		goto err_unmap_aux;
+	}
+
+	/* CPU writes to non-reserved MTTs, while HCA might DMA to reserved mtts */
+	mdev->limits.reserved_mtts = ALIGN(mdev->limits.reserved_mtts * mdev->limits.mtt_seg_size,
+					   dma_get_cache_alignment()) / mdev->limits.mtt_seg_size;
+
+	mdev->mr_table.mtt_table = mthca_alloc_icm_table(mdev, init_hca->mtt_base,
+							 mdev->limits.mtt_seg_size,
+							 mdev->limits.num_mtt_segs,
+							 mdev->limits.reserved_mtts,
+							 1, 0);
+	if (!mdev->mr_table.mtt_table) {
+		mthca_err(mdev, "Failed to map MTT context memory, aborting.\n");
+		err = -ENOMEM;
+		goto err_unmap_eq;
+	}
+
+	mdev->mr_table.mpt_table = mthca_alloc_icm_table(mdev, init_hca->mpt_base,
+							 dev_lim->mpt_entry_sz,
+							 mdev->limits.num_mpts,
+							 mdev->limits.reserved_mrws,
+							 1, 1);
+	if (!mdev->mr_table.mpt_table) {
+		mthca_err(mdev, "Failed to map MPT context memory, aborting.\n");
+		err = -ENOMEM;
+		goto err_unmap_mtt;
+	}
+
+	mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base,
+							dev_lim->qpc_entry_sz,
+							mdev->limits.num_qps,
+							mdev->limits.reserved_qps,
+							0, 0);
+	if (!mdev->qp_table.qp_table) {
+		mthca_err(mdev, "Failed to map QP context memory, aborting.\n");
+		err = -ENOMEM;
+		goto err_unmap_mpt;
+	}
+
+	mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base,
+							 dev_lim->eqpc_entry_sz,
+							 mdev->limits.num_qps,
+							 mdev->limits.reserved_qps,
+							 0, 0);
+	if (!mdev->qp_table.eqp_table) {
+		mthca_err(mdev, "Failed to map EQP context memory, aborting.\n");
+		err = -ENOMEM;
+		goto err_unmap_qp;
+	}
+
+	mdev->qp_table.rdb_table = mthca_alloc_icm_table(mdev, init_hca->rdb_base,
+							 MTHCA_RDB_ENTRY_SIZE,
+							 mdev->limits.num_qps <<
+							 mdev->qp_table.rdb_shift, 0,
+							 0, 0);
+	if (!mdev->qp_table.rdb_table) {
+		mthca_err(mdev, "Failed to map RDB context memory, aborting\n");
+		err = -ENOMEM;
+		goto err_unmap_eqp;
+	}
+
+	mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base,
+						     dev_lim->cqc_entry_sz,
+						     mdev->limits.num_cqs,
+						     mdev->limits.reserved_cqs,
+						     0, 0);
+	if (!mdev->cq_table.table) {
+		mthca_err(mdev, "Failed to map CQ context memory, aborting.\n");
+		err = -ENOMEM;
+		goto err_unmap_rdb;
+	}
+
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ) {
+		mdev->srq_table.table =
+			mthca_alloc_icm_table(mdev, init_hca->srqc_base,
+					      dev_lim->srq_entry_sz,
+					      mdev->limits.num_srqs,
+					      mdev->limits.reserved_srqs,
+					      0, 0);
+		if (!mdev->srq_table.table) {
+			mthca_err(mdev, "Failed to map SRQ context memory, "
+				  "aborting.\n");
+			err = -ENOMEM;
+			goto err_unmap_cq;
+		}
+	}
+
+	/*
+	 * It's not strictly required, but for simplicity just map the
+	 * whole multicast group table now.  The table isn't very big
+	 * and it's a lot easier than trying to track ref counts.
+	 */
+	mdev->mcg_table.table = mthca_alloc_icm_table(mdev, init_hca->mc_base,
+						      MTHCA_MGM_ENTRY_SIZE,
+						      mdev->limits.num_mgms +
+						      mdev->limits.num_amgms,
+						      mdev->limits.num_mgms +
+						      mdev->limits.num_amgms,
+						      0, 0);
+	if (!mdev->mcg_table.table) {
+		mthca_err(mdev, "Failed to map MCG context memory, aborting.\n");
+		err = -ENOMEM;
+		goto err_unmap_srq;
+	}
+
+	return 0;
+
+err_unmap_srq:
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
+		mthca_free_icm_table(mdev, mdev->srq_table.table);
+
+err_unmap_cq:
+	mthca_free_icm_table(mdev, mdev->cq_table.table);
+
+err_unmap_rdb:
+	mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
+
+err_unmap_eqp:
+	mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
+
+err_unmap_qp:
+	mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
+
+err_unmap_mpt:
+	mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
+
+err_unmap_mtt:
+	mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
+
+err_unmap_eq:
+	mthca_unmap_eq_icm(mdev);
+
+err_unmap_aux:
+	mthca_UNMAP_ICM_AUX(mdev);
+
+err_free_aux:
+	mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
+
+	return err;
+}
+
+static void mthca_free_icms(struct mthca_dev *mdev)
+{
+
+	mthca_free_icm_table(mdev, mdev->mcg_table.table);
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
+		mthca_free_icm_table(mdev, mdev->srq_table.table);
+	mthca_free_icm_table(mdev, mdev->cq_table.table);
+	mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
+	mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
+	mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
+	mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
+	mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
+	mthca_unmap_eq_icm(mdev);
+
+	mthca_UNMAP_ICM_AUX(mdev);
+	mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
+}
+
+static int mthca_init_arbel(struct mthca_dev *mdev)
+{
+	struct mthca_dev_lim        dev_lim;
+	struct mthca_profile        profile;
+	struct mthca_init_hca_param init_hca;
+	s64 icm_size;
+	int err;
+
+	err = mthca_QUERY_FW(mdev);
+	if (err) {
+		mthca_err(mdev, "QUERY_FW command failed %d, aborting.\n", err);
+		return err;
+	}
+
+	err = mthca_ENABLE_LAM(mdev);
+	if (err == -EAGAIN) {
+		mthca_dbg(mdev, "No HCA-attached memory (running in MemFree mode)\n");
+		mdev->mthca_flags |= MTHCA_FLAG_NO_LAM;
+	} else if (err) {
+		mthca_err(mdev, "ENABLE_LAM returned %d, aborting.\n", err);
+		return err;
+	}
+
+	err = mthca_load_fw(mdev);
+	if (err) {
+		mthca_err(mdev, "Loading FW returned %d, aborting.\n", err);
+		goto err_disable;
+	}
+
+	err = mthca_dev_lim(mdev, &dev_lim);
+	if (err) {
+		mthca_err(mdev, "QUERY_DEV_LIM returned %d, aborting.\n", err);
+		goto err_stop_fw;
+	}
+
+	profile = hca_profile;
+	profile.num_uar  = dev_lim.uar_size / PAGE_SIZE;
+	profile.num_udav = 0;
+	if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
+		profile.num_srq = dev_lim.max_srqs;
+
+	icm_size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
+	if (icm_size < 0) {
+		err = icm_size;
+		goto err_stop_fw;
+	}
+
+	err = mthca_init_icm(mdev, &dev_lim, &init_hca, icm_size);
+	if (err)
+		goto err_stop_fw;
+
+	err = mthca_INIT_HCA(mdev, &init_hca);
+	if (err) {
+		mthca_err(mdev, "INIT_HCA command returned %d, aborting.\n", err);
+		goto err_free_icm;
+	}
+
+	return 0;
+
+err_free_icm:
+	mthca_free_icms(mdev);
+
+err_stop_fw:
+	mthca_UNMAP_FA(mdev);
+	mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
+
+err_disable:
+	if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
+		mthca_DISABLE_LAM(mdev);
+
+	return err;
+}
+
+static void mthca_close_hca(struct mthca_dev *mdev)
+{
+	mthca_CLOSE_HCA(mdev, 0);
+
+	if (mthca_is_memfree(mdev)) {
+		mthca_free_icms(mdev);
+
+		mthca_UNMAP_FA(mdev);
+		mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
+
+		if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
+			mthca_DISABLE_LAM(mdev);
+	} else
+		mthca_SYS_DIS(mdev);
+}
+
+static int mthca_init_hca(struct mthca_dev *mdev)
+{
+	int err;
+	struct mthca_adapter adapter;
+
+	if (mthca_is_memfree(mdev))
+		err = mthca_init_arbel(mdev);
+	else
+		err = mthca_init_tavor(mdev);
+
+	if (err)
+		return err;
+
+	err = mthca_QUERY_ADAPTER(mdev, &adapter);
+	if (err) {
+		mthca_err(mdev, "QUERY_ADAPTER command returned %d, aborting.\n", err);
+		goto err_close;
+	}
+
+	mdev->eq_table.inta_pin = adapter.inta_pin;
+	if (!mthca_is_memfree(mdev))
+		mdev->rev_id = adapter.revision_id;
+	memcpy(mdev->board_id, adapter.board_id, sizeof mdev->board_id);
+
+	return 0;
+
+err_close:
+	mthca_close_hca(mdev);
+	return err;
+}
+
+static int mthca_setup_hca(struct mthca_dev *dev)
+{
+	int err;
+
+	MTHCA_INIT_DOORBELL_LOCK(&dev->doorbell_lock);
+
+	err = mthca_init_uar_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "user access region table, aborting.\n");
+		return err;
+	}
+
+	err = mthca_uar_alloc(dev, &dev->driver_uar);
+	if (err) {
+		mthca_err(dev, "Failed to allocate driver access region, "
+			  "aborting.\n");
+		goto err_uar_table_free;
+	}
+
+	dev->kar = ioremap((phys_addr_t) dev->driver_uar.pfn << PAGE_SHIFT, PAGE_SIZE);
+	if (!dev->kar) {
+		mthca_err(dev, "Couldn't map kernel access region, "
+			  "aborting.\n");
+		err = -ENOMEM;
+		goto err_uar_free;
+	}
+
+	err = mthca_init_pd_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "protection domain table, aborting.\n");
+		goto err_kar_unmap;
+	}
+
+	err = mthca_init_mr_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "memory region table, aborting.\n");
+		goto err_pd_table_free;
+	}
+
+	err = mthca_pd_alloc(dev, 1, &dev->driver_pd);
+	if (err) {
+		mthca_err(dev, "Failed to create driver PD, "
+			  "aborting.\n");
+		goto err_mr_table_free;
+	}
+
+	err = mthca_init_eq_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "event queue table, aborting.\n");
+		goto err_pd_free;
+	}
+
+	err = mthca_cmd_use_events(dev);
+	if (err) {
+		mthca_err(dev, "Failed to switch to event-driven "
+			  "firmware commands, aborting.\n");
+		goto err_eq_table_free;
+	}
+
+	err = mthca_NOP(dev);
+	if (err) {
+		if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
+			mthca_warn(dev, "NOP command failed to generate interrupt "
+				   "(IRQ %d).\n",
+				   dev->eq_table.eq[MTHCA_EQ_CMD].msi_x_vector);
+			mthca_warn(dev, "Trying again with MSI-X disabled.\n");
+		} else {
+			mthca_err(dev, "NOP command failed to generate interrupt "
+				  "(IRQ %d), aborting.\n",
+				  dev->pdev->irq);
+			mthca_err(dev, "BIOS or ACPI interrupt routing problem?\n");
+		}
+
+		goto err_cmd_poll;
+	}
+
+	mthca_dbg(dev, "NOP command IRQ test passed\n");
+
+	err = mthca_init_cq_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "completion queue table, aborting.\n");
+		goto err_cmd_poll;
+	}
+
+	err = mthca_init_srq_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "shared receive queue table, aborting.\n");
+		goto err_cq_table_free;
+	}
+
+	err = mthca_init_qp_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "queue pair table, aborting.\n");
+		goto err_srq_table_free;
+	}
+
+	err = mthca_init_av_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "address vector table, aborting.\n");
+		goto err_qp_table_free;
+	}
+
+	err = mthca_init_mcg_table(dev);
+	if (err) {
+		mthca_err(dev, "Failed to initialize "
+			  "multicast group table, aborting.\n");
+		goto err_av_table_free;
+	}
+
+	return 0;
+
+err_av_table_free:
+	mthca_cleanup_av_table(dev);
+
+err_qp_table_free:
+	mthca_cleanup_qp_table(dev);
+
+err_srq_table_free:
+	mthca_cleanup_srq_table(dev);
+
+err_cq_table_free:
+	mthca_cleanup_cq_table(dev);
+
+err_cmd_poll:
+	mthca_cmd_use_polling(dev);
+
+err_eq_table_free:
+	mthca_cleanup_eq_table(dev);
+
+err_pd_free:
+	mthca_pd_free(dev, &dev->driver_pd);
+
+err_mr_table_free:
+	mthca_cleanup_mr_table(dev);
+
+err_pd_table_free:
+	mthca_cleanup_pd_table(dev);
+
+err_kar_unmap:
+	iounmap(dev->kar);
+
+err_uar_free:
+	mthca_uar_free(dev, &dev->driver_uar);
+
+err_uar_table_free:
+	mthca_cleanup_uar_table(dev);
+	return err;
+}
+
+static int mthca_enable_msi_x(struct mthca_dev *mdev)
+{
+	int err;
+
+	err = pci_alloc_irq_vectors(mdev->pdev, 3, 3, PCI_IRQ_MSIX);
+	if (err < 0)
+		return err;
+
+	mdev->eq_table.eq[MTHCA_EQ_COMP ].msi_x_vector =
+			pci_irq_vector(mdev->pdev, 0);
+	mdev->eq_table.eq[MTHCA_EQ_ASYNC].msi_x_vector =
+			pci_irq_vector(mdev->pdev, 1);
+	mdev->eq_table.eq[MTHCA_EQ_CMD  ].msi_x_vector =
+			pci_irq_vector(mdev->pdev, 2);
+
+	return 0;
+}
+
+/* Types of supported HCA */
+enum {
+	TAVOR,			/* MT23108                        */
+	ARBEL_COMPAT,		/* MT25208 in Tavor compat mode   */
+	ARBEL_NATIVE,		/* MT25208 with extended features */
+	SINAI			/* MT25204 */
+};
+
+#define MTHCA_FW_VER(major, minor, subminor) \
+	(((u64) (major) << 32) | ((u64) (minor) << 16) | (u64) (subminor))
+
+static struct {
+	u64 latest_fw;
+	u32 flags;
+} mthca_hca_table[] = {
+	[TAVOR]        = { .latest_fw = MTHCA_FW_VER(3, 5, 0),
+			   .flags     = 0 },
+	[ARBEL_COMPAT] = { .latest_fw = MTHCA_FW_VER(4, 8, 200),
+			   .flags     = MTHCA_FLAG_PCIE },
+	[ARBEL_NATIVE] = { .latest_fw = MTHCA_FW_VER(5, 3, 0),
+			   .flags     = MTHCA_FLAG_MEMFREE |
+					MTHCA_FLAG_PCIE },
+	[SINAI]        = { .latest_fw = MTHCA_FW_VER(1, 2, 0),
+			   .flags     = MTHCA_FLAG_MEMFREE |
+					MTHCA_FLAG_PCIE    |
+					MTHCA_FLAG_SINAI_OPT }
+};
+
+static int __mthca_init_one(struct pci_dev *pdev, int hca_type)
+{
+	int ddr_hidden = 0;
+	int err;
+	struct mthca_dev *mdev;
+
+	printk(KERN_INFO PFX "Initializing %s\n",
+	       pci_name(pdev));
+
+	err = pci_enable_device(pdev);
+	if (err) {
+		dev_err(&pdev->dev, "Cannot enable PCI device, "
+			"aborting.\n");
+		return err;
+	}
+
+	/*
+	 * Check for BARs.  We expect 0: 1MB, 2: 8MB, 4: DDR (may not
+	 * be present)
+	 */
+	if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
+	    pci_resource_len(pdev, 0) != 1 << 20) {
+		dev_err(&pdev->dev, "Missing DCS, aborting.\n");
+		err = -ENODEV;
+		goto err_disable_pdev;
+	}
+	if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM)) {
+		dev_err(&pdev->dev, "Missing UAR, aborting.\n");
+		err = -ENODEV;
+		goto err_disable_pdev;
+	}
+	if (!(pci_resource_flags(pdev, 4) & IORESOURCE_MEM))
+		ddr_hidden = 1;
+
+	err = pci_request_regions(pdev, DRV_NAME);
+	if (err) {
+		dev_err(&pdev->dev, "Cannot obtain PCI resources, "
+			"aborting.\n");
+		goto err_disable_pdev;
+	}
+
+	pci_set_master(pdev);
+
+	err = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
+	if (err) {
+		dev_err(&pdev->dev, "Can't set PCI DMA mask, aborting.\n");
+		goto err_free_res;
+	}
+
+	/* We can handle large RDMA requests, so allow larger segments. */
+	dma_set_max_seg_size(&pdev->dev, 1024 * 1024 * 1024);
+
+	mdev = ib_alloc_device(mthca_dev, ib_dev);
+	if (!mdev) {
+		dev_err(&pdev->dev, "Device struct alloc failed, "
+			"aborting.\n");
+		err = -ENOMEM;
+		goto err_free_res;
+	}
+
+	mdev->pdev = pdev;
+
+	mdev->mthca_flags = mthca_hca_table[hca_type].flags;
+	if (ddr_hidden)
+		mdev->mthca_flags |= MTHCA_FLAG_DDR_HIDDEN;
+
+	/*
+	 * Now reset the HCA before we touch the PCI capabilities or
+	 * attempt a firmware command, since a boot ROM may have left
+	 * the HCA in an undefined state.
+	 */
+	err = mthca_reset(mdev);
+	if (err) {
+		mthca_err(mdev, "Failed to reset HCA, aborting.\n");
+		goto err_free_dev;
+	}
+
+	err = mthca_cmd_init(mdev);
+	if (err) {
+		mthca_err(mdev, "Failed to init command interface, aborting.\n");
+		goto err_free_dev;
+	}
+
+	err = mthca_tune_pci(mdev);
+	if (err)
+		goto err_cmd;
+
+	err = mthca_init_hca(mdev);
+	if (err)
+		goto err_cmd;
+
+	if (mdev->fw_ver < mthca_hca_table[hca_type].latest_fw) {
+		mthca_warn(mdev, "HCA FW version %d.%d.%03d is old (%d.%d.%03d is current).\n",
+			   (int) (mdev->fw_ver >> 32), (int) (mdev->fw_ver >> 16) & 0xffff,
+			   (int) (mdev->fw_ver & 0xffff),
+			   (int) (mthca_hca_table[hca_type].latest_fw >> 32),
+			   (int) (mthca_hca_table[hca_type].latest_fw >> 16) & 0xffff,
+			   (int) (mthca_hca_table[hca_type].latest_fw & 0xffff));
+		mthca_warn(mdev, "If you have problems, try updating your HCA FW.\n");
+	}
+
+	if (msi_x && !mthca_enable_msi_x(mdev))
+		mdev->mthca_flags |= MTHCA_FLAG_MSI_X;
+
+	err = mthca_setup_hca(mdev);
+	if (err == -EBUSY && (mdev->mthca_flags & MTHCA_FLAG_MSI_X)) {
+		pci_free_irq_vectors(pdev);
+		mdev->mthca_flags &= ~MTHCA_FLAG_MSI_X;
+
+		err = mthca_setup_hca(mdev);
+	}
+
+	if (err)
+		goto err_close;
+
+	err = mthca_register_device(mdev);
+	if (err)
+		goto err_cleanup;
+
+	err = mthca_create_agents(mdev);
+	if (err)
+		goto err_unregister;
+
+	pci_set_drvdata(pdev, mdev);
+	mdev->hca_type = hca_type;
+
+	mdev->active = true;
+
+	return 0;
+
+err_unregister:
+	mthca_unregister_device(mdev);
+
+err_cleanup:
+	mthca_cleanup_mcg_table(mdev);
+	mthca_cleanup_av_table(mdev);
+	mthca_cleanup_qp_table(mdev);
+	mthca_cleanup_srq_table(mdev);
+	mthca_cleanup_cq_table(mdev);
+	mthca_cmd_use_polling(mdev);
+	mthca_cleanup_eq_table(mdev);
+
+	mthca_pd_free(mdev, &mdev->driver_pd);
+
+	mthca_cleanup_mr_table(mdev);
+	mthca_cleanup_pd_table(mdev);
+	mthca_cleanup_uar_table(mdev);
+
+err_close:
+	if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
+		pci_free_irq_vectors(pdev);
+
+	mthca_close_hca(mdev);
+
+err_cmd:
+	mthca_cmd_cleanup(mdev);
+
+err_free_dev:
+	ib_dealloc_device(&mdev->ib_dev);
+
+err_free_res:
+	pci_release_regions(pdev);
+
+err_disable_pdev:
+	pci_disable_device(pdev);
+	pci_set_drvdata(pdev, NULL);
+	return err;
+}
+
+static void __mthca_remove_one(struct pci_dev *pdev)
+{
+	struct mthca_dev *mdev = pci_get_drvdata(pdev);
+	int p;
+
+	if (mdev) {
+		mthca_free_agents(mdev);
+		mthca_unregister_device(mdev);
+
+		for (p = 1; p <= mdev->limits.num_ports; ++p)
+			mthca_CLOSE_IB(mdev, p);
+
+		mthca_cleanup_mcg_table(mdev);
+		mthca_cleanup_av_table(mdev);
+		mthca_cleanup_qp_table(mdev);
+		mthca_cleanup_srq_table(mdev);
+		mthca_cleanup_cq_table(mdev);
+		mthca_cmd_use_polling(mdev);
+		mthca_cleanup_eq_table(mdev);
+
+		mthca_pd_free(mdev, &mdev->driver_pd);
+
+		mthca_cleanup_mr_table(mdev);
+		mthca_cleanup_pd_table(mdev);
+
+		iounmap(mdev->kar);
+		mthca_uar_free(mdev, &mdev->driver_uar);
+		mthca_cleanup_uar_table(mdev);
+		mthca_close_hca(mdev);
+		mthca_cmd_cleanup(mdev);
+
+		if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
+			pci_free_irq_vectors(pdev);
+
+		ib_dealloc_device(&mdev->ib_dev);
+		pci_release_regions(pdev);
+		pci_disable_device(pdev);
+		pci_set_drvdata(pdev, NULL);
+	}
+}
+
+int __mthca_restart_one(struct pci_dev *pdev)
+{
+	struct mthca_dev *mdev;
+	int hca_type;
+
+	mdev = pci_get_drvdata(pdev);
+	if (!mdev)
+		return -ENODEV;
+	hca_type = mdev->hca_type;
+	__mthca_remove_one(pdev);
+	return __mthca_init_one(pdev, hca_type);
+}
+
+static int mthca_init_one(struct pci_dev *pdev, const struct pci_device_id *id)
+{
+	int ret;
+
+	mutex_lock(&mthca_device_mutex);
+
+	printk_once(KERN_INFO "%s", mthca_version);
+
+	if (id->driver_data >= ARRAY_SIZE(mthca_hca_table)) {
+		printk(KERN_ERR PFX "%s has invalid driver data %lx\n",
+		       pci_name(pdev), id->driver_data);
+		mutex_unlock(&mthca_device_mutex);
+		return -ENODEV;
+	}
+
+	ret = __mthca_init_one(pdev, id->driver_data);
+
+	mutex_unlock(&mthca_device_mutex);
+
+	return ret;
+}
+
+static void mthca_remove_one(struct pci_dev *pdev)
+{
+	mutex_lock(&mthca_device_mutex);
+	__mthca_remove_one(pdev);
+	mutex_unlock(&mthca_device_mutex);
+}
+
+static const struct pci_device_id mthca_pci_table[] = {
+	{ PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_TAVOR),
+	  .driver_data = TAVOR },
+	{ PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_TAVOR),
+	  .driver_data = TAVOR },
+	{ PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
+	  .driver_data = ARBEL_COMPAT },
+	{ PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
+	  .driver_data = ARBEL_COMPAT },
+	{ PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL),
+	  .driver_data = ARBEL_NATIVE },
+	{ PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL),
+	  .driver_data = ARBEL_NATIVE },
+	{ PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI),
+	  .driver_data = SINAI },
+	{ PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI),
+	  .driver_data = SINAI },
+	{ PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
+	  .driver_data = SINAI },
+	{ PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
+	  .driver_data = SINAI },
+	{ 0, }
+};
+
+MODULE_DEVICE_TABLE(pci, mthca_pci_table);
+
+static struct pci_driver mthca_driver = {
+	.name		= DRV_NAME,
+	.id_table	= mthca_pci_table,
+	.probe		= mthca_init_one,
+	.remove		= mthca_remove_one,
+};
+
+static void __init __mthca_check_profile_val(const char *name, int *pval,
+					     int pval_default)
+{
+	/* value must be positive and power of 2 */
+	int old_pval = *pval;
+
+	if (old_pval <= 0)
+		*pval = pval_default;
+	else
+		*pval = roundup_pow_of_two(old_pval);
+
+	if (old_pval != *pval) {
+		printk(KERN_WARNING PFX "Invalid value %d for %s in module parameter.\n",
+		       old_pval, name);
+		printk(KERN_WARNING PFX "Corrected %s to %d.\n", name, *pval);
+	}
+}
+
+#define mthca_check_profile_val(name, default)				\
+	__mthca_check_profile_val(#name, &hca_profile.name, default)
+
+static void __init mthca_validate_profile(void)
+{
+	mthca_check_profile_val(num_qp,            MTHCA_DEFAULT_NUM_QP);
+	mthca_check_profile_val(rdb_per_qp,        MTHCA_DEFAULT_RDB_PER_QP);
+	mthca_check_profile_val(num_cq,            MTHCA_DEFAULT_NUM_CQ);
+	mthca_check_profile_val(num_mcg, 	   MTHCA_DEFAULT_NUM_MCG);
+	mthca_check_profile_val(num_mpt, 	   MTHCA_DEFAULT_NUM_MPT);
+	mthca_check_profile_val(num_mtt, 	   MTHCA_DEFAULT_NUM_MTT);
+	mthca_check_profile_val(num_udav,          MTHCA_DEFAULT_NUM_UDAV);
+	mthca_check_profile_val(fmr_reserved_mtts, MTHCA_DEFAULT_NUM_RESERVED_MTTS);
+
+	if (hca_profile.fmr_reserved_mtts >= hca_profile.num_mtt) {
+		printk(KERN_WARNING PFX "Invalid fmr_reserved_mtts module parameter %d.\n",
+		       hca_profile.fmr_reserved_mtts);
+		printk(KERN_WARNING PFX "(Must be smaller than num_mtt %d)\n",
+		       hca_profile.num_mtt);
+		hca_profile.fmr_reserved_mtts = hca_profile.num_mtt / 2;
+		printk(KERN_WARNING PFX "Corrected fmr_reserved_mtts to %d.\n",
+		       hca_profile.fmr_reserved_mtts);
+	}
+
+	if ((log_mtts_per_seg < 1) || (log_mtts_per_seg > 5)) {
+		printk(KERN_WARNING PFX "bad log_mtts_per_seg (%d). Using default - %d\n",
+		       log_mtts_per_seg, ilog2(MTHCA_MTT_SEG_SIZE / 8));
+		log_mtts_per_seg = ilog2(MTHCA_MTT_SEG_SIZE / 8);
+	}
+}
+
+static int __init mthca_init(void)
+{
+	int ret;
+
+	mthca_validate_profile();
+
+	ret = mthca_catas_init();
+	if (ret)
+		return ret;
+
+	ret = pci_register_driver(&mthca_driver);
+	if (ret < 0) {
+		mthca_catas_cleanup();
+		return ret;
+	}
+
+	return 0;
+}
+
+static void __exit mthca_cleanup(void)
+{
+	pci_unregister_driver(&mthca_driver);
+	mthca_catas_cleanup();
+}
+
+module_init(mthca_init);
+module_exit(mthca_cleanup);
diff --git a/drivers/infiniband/hw/mthca/mthca_mcg.c b/drivers/infiniband/hw/mthca/mthca_mcg.c
new file mode 100644
index 000000000..6304ae8f4
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_mcg.c
@@ -0,0 +1,335 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/string.h>
+#include <linux/gfp.h>
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+
+struct mthca_mgm {
+	__be32 next_gid_index;
+	u32    reserved[3];
+	u8     gid[16];
+	__be32 qp[MTHCA_QP_PER_MGM];
+};
+
+static const u8 zero_gid[16];	/* automatically initialized to 0 */
+
+/*
+ * Caller must hold MCG table semaphore.  gid and mgm parameters must
+ * be properly aligned for command interface.
+ *
+ *  Returns 0 unless a firmware command error occurs.
+ *
+ * If GID is found in MGM or MGM is empty, *index = *hash, *prev = -1
+ * and *mgm holds MGM entry.
+ *
+ * if GID is found in AMGM, *index = index in AMGM, *prev = index of
+ * previous entry in hash chain and *mgm holds AMGM entry.
+ *
+ * If no AMGM exists for given gid, *index = -1, *prev = index of last
+ * entry in hash chain and *mgm holds end of hash chain.
+ */
+static int find_mgm(struct mthca_dev *dev,
+		    u8 *gid, struct mthca_mailbox *mgm_mailbox,
+		    u16 *hash, int *prev, int *index)
+{
+	struct mthca_mailbox *mailbox;
+	struct mthca_mgm *mgm = mgm_mailbox->buf;
+	u8 *mgid;
+	int err;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return -ENOMEM;
+	mgid = mailbox->buf;
+
+	memcpy(mgid, gid, 16);
+
+	err = mthca_MGID_HASH(dev, mailbox, hash);
+	if (err) {
+		mthca_err(dev, "MGID_HASH failed (%d)\n", err);
+		goto out;
+	}
+
+	if (0)
+		mthca_dbg(dev, "Hash for %pI6 is %04x\n", gid, *hash);
+
+	*index = *hash;
+	*prev  = -1;
+
+	do {
+		err = mthca_READ_MGM(dev, *index, mgm_mailbox);
+		if (err) {
+			mthca_err(dev, "READ_MGM failed (%d)\n", err);
+			goto out;
+		}
+
+		if (!memcmp(mgm->gid, zero_gid, 16)) {
+			if (*index != *hash) {
+				mthca_err(dev, "Found zero MGID in AMGM.\n");
+				err = -EINVAL;
+			}
+			goto out;
+		}
+
+		if (!memcmp(mgm->gid, gid, 16))
+			goto out;
+
+		*prev = *index;
+		*index = be32_to_cpu(mgm->next_gid_index) >> 6;
+	} while (*index);
+
+	*index = -1;
+
+ out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
+{
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	struct mthca_mailbox *mailbox;
+	struct mthca_mgm *mgm;
+	u16 hash;
+	int index, prev;
+	int link = 0;
+	int i;
+	int err;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	mgm = mailbox->buf;
+
+	mutex_lock(&dev->mcg_table.mutex);
+
+	err = find_mgm(dev, gid->raw, mailbox, &hash, &prev, &index);
+	if (err)
+		goto out;
+
+	if (index != -1) {
+		if (!memcmp(mgm->gid, zero_gid, 16))
+			memcpy(mgm->gid, gid->raw, 16);
+	} else {
+		link = 1;
+
+		index = mthca_alloc(&dev->mcg_table.alloc);
+		if (index == -1) {
+			mthca_err(dev, "No AMGM entries left\n");
+			err = -ENOMEM;
+			goto out;
+		}
+
+		err = mthca_READ_MGM(dev, index, mailbox);
+		if (err) {
+			mthca_err(dev, "READ_MGM failed (%d)\n", err);
+			goto out;
+		}
+		memset(mgm, 0, sizeof *mgm);
+		memcpy(mgm->gid, gid->raw, 16);
+	}
+
+	for (i = 0; i < MTHCA_QP_PER_MGM; ++i)
+		if (mgm->qp[i] == cpu_to_be32(ibqp->qp_num | (1 << 31))) {
+			mthca_dbg(dev, "QP %06x already a member of MGM\n",
+				  ibqp->qp_num);
+			err = 0;
+			goto out;
+		} else if (!(mgm->qp[i] & cpu_to_be32(1 << 31))) {
+			mgm->qp[i] = cpu_to_be32(ibqp->qp_num | (1 << 31));
+			break;
+		}
+
+	if (i == MTHCA_QP_PER_MGM) {
+		mthca_err(dev, "MGM at index %x is full.\n", index);
+		err = -ENOMEM;
+		goto out;
+	}
+
+	err = mthca_WRITE_MGM(dev, index, mailbox);
+	if (err) {
+		mthca_err(dev, "WRITE_MGM failed %d\n", err);
+		err = -EINVAL;
+		goto out;
+	}
+
+	if (!link)
+		goto out;
+
+	err = mthca_READ_MGM(dev, prev, mailbox);
+	if (err) {
+		mthca_err(dev, "READ_MGM failed %d\n", err);
+		goto out;
+	}
+
+	mgm->next_gid_index = cpu_to_be32(index << 6);
+
+	err = mthca_WRITE_MGM(dev, prev, mailbox);
+	if (err)
+		mthca_err(dev, "WRITE_MGM returned %d\n", err);
+
+ out:
+	if (err && link && index != -1) {
+		BUG_ON(index < dev->limits.num_mgms);
+		mthca_free(&dev->mcg_table.alloc, index);
+	}
+	mutex_unlock(&dev->mcg_table.mutex);
+
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
+{
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	struct mthca_mailbox *mailbox;
+	struct mthca_mgm *mgm;
+	u16 hash;
+	int prev, index;
+	int i, loc;
+	int err;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	mgm = mailbox->buf;
+
+	mutex_lock(&dev->mcg_table.mutex);
+
+	err = find_mgm(dev, gid->raw, mailbox, &hash, &prev, &index);
+	if (err)
+		goto out;
+
+	if (index == -1) {
+		mthca_err(dev, "MGID %pI6 not found\n", gid->raw);
+		err = -EINVAL;
+		goto out;
+	}
+
+	for (loc = -1, i = 0; i < MTHCA_QP_PER_MGM; ++i) {
+		if (mgm->qp[i] == cpu_to_be32(ibqp->qp_num | (1 << 31)))
+			loc = i;
+		if (!(mgm->qp[i] & cpu_to_be32(1 << 31)))
+			break;
+	}
+
+	if (loc == -1) {
+		mthca_err(dev, "QP %06x not found in MGM\n", ibqp->qp_num);
+		err = -EINVAL;
+		goto out;
+	}
+
+	mgm->qp[loc]   = mgm->qp[i - 1];
+	mgm->qp[i - 1] = 0;
+
+	err = mthca_WRITE_MGM(dev, index, mailbox);
+	if (err) {
+		mthca_err(dev, "WRITE_MGM returned %d\n", err);
+		goto out;
+	}
+
+	if (i != 1)
+		goto out;
+
+	if (prev == -1) {
+		/* Remove entry from MGM */
+		int amgm_index_to_free = be32_to_cpu(mgm->next_gid_index) >> 6;
+		if (amgm_index_to_free) {
+			err = mthca_READ_MGM(dev, amgm_index_to_free,
+					     mailbox);
+			if (err) {
+				mthca_err(dev, "READ_MGM returned %d\n", err);
+				goto out;
+			}
+		} else
+			memset(mgm->gid, 0, 16);
+
+		err = mthca_WRITE_MGM(dev, index, mailbox);
+		if (err) {
+			mthca_err(dev, "WRITE_MGM returned %d\n", err);
+			goto out;
+		}
+		if (amgm_index_to_free) {
+			BUG_ON(amgm_index_to_free < dev->limits.num_mgms);
+			mthca_free(&dev->mcg_table.alloc, amgm_index_to_free);
+		}
+	} else {
+		/* Remove entry from AMGM */
+		int curr_next_index = be32_to_cpu(mgm->next_gid_index) >> 6;
+		err = mthca_READ_MGM(dev, prev, mailbox);
+		if (err) {
+			mthca_err(dev, "READ_MGM returned %d\n", err);
+			goto out;
+		}
+
+		mgm->next_gid_index = cpu_to_be32(curr_next_index << 6);
+
+		err = mthca_WRITE_MGM(dev, prev, mailbox);
+		if (err) {
+			mthca_err(dev, "WRITE_MGM returned %d\n", err);
+			goto out;
+		}
+		BUG_ON(index < dev->limits.num_mgms);
+		mthca_free(&dev->mcg_table.alloc, index);
+	}
+
+ out:
+	mutex_unlock(&dev->mcg_table.mutex);
+
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_init_mcg_table(struct mthca_dev *dev)
+{
+	int err;
+	int table_size = dev->limits.num_mgms + dev->limits.num_amgms;
+
+	err = mthca_alloc_init(&dev->mcg_table.alloc,
+			       table_size,
+			       table_size - 1,
+			       dev->limits.num_mgms);
+	if (err)
+		return err;
+
+	mutex_init(&dev->mcg_table.mutex);
+
+	return 0;
+}
+
+void mthca_cleanup_mcg_table(struct mthca_dev *dev)
+{
+	mthca_alloc_cleanup(&dev->mcg_table.alloc);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_memfree.c b/drivers/infiniband/hw/mthca/mthca_memfree.c
new file mode 100644
index 000000000..f2734a5c5
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_memfree.c
@@ -0,0 +1,763 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/mm.h>
+#include <linux/scatterlist.h>
+#include <linux/sched.h>
+#include <linux/slab.h>
+
+#include <asm/page.h>
+
+#include "mthca_memfree.h"
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+
+/*
+ * We allocate in as big chunks as we can, up to a maximum of 256 KB
+ * per chunk.
+ */
+enum {
+	MTHCA_ICM_ALLOC_SIZE   = 1 << 18,
+	MTHCA_TABLE_CHUNK_SIZE = 1 << 18
+};
+
+struct mthca_user_db_table {
+	struct mutex mutex;
+	struct {
+		u64                uvirt;
+		struct scatterlist mem;
+		int                refcount;
+	} page[];
+};
+
+static void mthca_free_icm_pages(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
+{
+	int i;
+
+	if (chunk->nsg > 0)
+		dma_unmap_sg(&dev->pdev->dev, chunk->mem, chunk->npages,
+			     DMA_BIDIRECTIONAL);
+
+	for (i = 0; i < chunk->npages; ++i)
+		__free_pages(sg_page(&chunk->mem[i]),
+			     get_order(chunk->mem[i].length));
+}
+
+static void mthca_free_icm_coherent(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
+{
+	int i;
+
+	for (i = 0; i < chunk->npages; ++i) {
+		dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
+				  lowmem_page_address(sg_page(&chunk->mem[i])),
+				  sg_dma_address(&chunk->mem[i]));
+	}
+}
+
+void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent)
+{
+	struct mthca_icm_chunk *chunk, *tmp;
+
+	if (!icm)
+		return;
+
+	list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
+		if (coherent)
+			mthca_free_icm_coherent(dev, chunk);
+		else
+			mthca_free_icm_pages(dev, chunk);
+
+		kfree(chunk);
+	}
+
+	kfree(icm);
+}
+
+static int mthca_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask)
+{
+	struct page *page;
+
+	/*
+	 * Use __GFP_ZERO because buggy firmware assumes ICM pages are
+	 * cleared, and subtle failures are seen if they aren't.
+	 */
+	page = alloc_pages(gfp_mask | __GFP_ZERO, order);
+	if (!page)
+		return -ENOMEM;
+
+	sg_set_page(mem, page, PAGE_SIZE << order, 0);
+	return 0;
+}
+
+static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
+				    int order, gfp_t gfp_mask)
+{
+	void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem),
+				       gfp_mask);
+	if (!buf)
+		return -ENOMEM;
+
+	sg_set_buf(mem, buf, PAGE_SIZE << order);
+	BUG_ON(mem->offset);
+	sg_dma_len(mem) = PAGE_SIZE << order;
+	return 0;
+}
+
+struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
+				  gfp_t gfp_mask, int coherent)
+{
+	struct mthca_icm *icm;
+	struct mthca_icm_chunk *chunk = NULL;
+	int cur_order;
+	int ret;
+
+	/* We use sg_set_buf for coherent allocs, which assumes low memory */
+	BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
+
+	icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
+	if (!icm)
+		return icm;
+
+	icm->refcount = 0;
+	INIT_LIST_HEAD(&icm->chunk_list);
+
+	cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
+
+	while (npages > 0) {
+		if (!chunk) {
+			chunk = kmalloc(sizeof *chunk,
+					gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
+			if (!chunk)
+				goto fail;
+
+			sg_init_table(chunk->mem, MTHCA_ICM_CHUNK_LEN);
+			chunk->npages = 0;
+			chunk->nsg    = 0;
+			list_add_tail(&chunk->list, &icm->chunk_list);
+		}
+
+		while (1 << cur_order > npages)
+			--cur_order;
+
+		if (coherent)
+			ret = mthca_alloc_icm_coherent(&dev->pdev->dev,
+						       &chunk->mem[chunk->npages],
+						       cur_order, gfp_mask);
+		else
+			ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
+						    cur_order, gfp_mask);
+
+		if (!ret) {
+			++chunk->npages;
+
+			if (coherent)
+				++chunk->nsg;
+			else if (chunk->npages == MTHCA_ICM_CHUNK_LEN) {
+				chunk->nsg =
+					dma_map_sg(&dev->pdev->dev, chunk->mem,
+						   chunk->npages,
+						   DMA_BIDIRECTIONAL);
+
+				if (chunk->nsg <= 0)
+					goto fail;
+			}
+
+			if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
+				chunk = NULL;
+
+			npages -= 1 << cur_order;
+		} else {
+			--cur_order;
+			if (cur_order < 0)
+				goto fail;
+		}
+	}
+
+	if (!coherent && chunk) {
+		chunk->nsg = dma_map_sg(&dev->pdev->dev, chunk->mem,
+					chunk->npages, DMA_BIDIRECTIONAL);
+
+		if (chunk->nsg <= 0)
+			goto fail;
+	}
+
+	return icm;
+
+fail:
+	mthca_free_icm(dev, icm, coherent);
+	return NULL;
+}
+
+int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
+{
+	int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
+	int ret = 0;
+
+	mutex_lock(&table->mutex);
+
+	if (table->icm[i]) {
+		++table->icm[i]->refcount;
+		goto out;
+	}
+
+	table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
+					(table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
+					__GFP_NOWARN, table->coherent);
+	if (!table->icm[i]) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	if (mthca_MAP_ICM(dev, table->icm[i],
+			  table->virt + i * MTHCA_TABLE_CHUNK_SIZE)) {
+		mthca_free_icm(dev, table->icm[i], table->coherent);
+		table->icm[i] = NULL;
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	++table->icm[i]->refcount;
+
+out:
+	mutex_unlock(&table->mutex);
+	return ret;
+}
+
+void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
+{
+	int i;
+
+	if (!mthca_is_memfree(dev))
+		return;
+
+	i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
+
+	mutex_lock(&table->mutex);
+
+	if (--table->icm[i]->refcount == 0) {
+		mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
+				MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE);
+		mthca_free_icm(dev, table->icm[i], table->coherent);
+		table->icm[i] = NULL;
+	}
+
+	mutex_unlock(&table->mutex);
+}
+
+void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
+{
+	int idx, offset, dma_offset, i;
+	struct mthca_icm_chunk *chunk;
+	struct mthca_icm *icm;
+	struct page *page = NULL;
+
+	if (!table->lowmem)
+		return NULL;
+
+	mutex_lock(&table->mutex);
+
+	idx = (obj & (table->num_obj - 1)) * table->obj_size;
+	icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
+	dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE;
+
+	if (!icm)
+		goto out;
+
+	list_for_each_entry(chunk, &icm->chunk_list, list) {
+		for (i = 0; i < chunk->npages; ++i) {
+			if (dma_handle && dma_offset >= 0) {
+				if (sg_dma_len(&chunk->mem[i]) > dma_offset)
+					*dma_handle = sg_dma_address(&chunk->mem[i]) +
+						dma_offset;
+				dma_offset -= sg_dma_len(&chunk->mem[i]);
+			}
+			/* DMA mapping can merge pages but not split them,
+			 * so if we found the page, dma_handle has already
+			 * been assigned to. */
+			if (chunk->mem[i].length > offset) {
+				page = sg_page(&chunk->mem[i]);
+				goto out;
+			}
+			offset -= chunk->mem[i].length;
+		}
+	}
+
+out:
+	mutex_unlock(&table->mutex);
+	return page ? lowmem_page_address(page) + offset : NULL;
+}
+
+int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
+			  int start, int end)
+{
+	int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
+	int i, err;
+
+	for (i = start; i <= end; i += inc) {
+		err = mthca_table_get(dev, table, i);
+		if (err)
+			goto fail;
+	}
+
+	return 0;
+
+fail:
+	while (i > start) {
+		i -= inc;
+		mthca_table_put(dev, table, i);
+	}
+
+	return err;
+}
+
+void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
+			   int start, int end)
+{
+	int i;
+
+	if (!mthca_is_memfree(dev))
+		return;
+
+	for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
+		mthca_table_put(dev, table, i);
+}
+
+struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
+					      u64 virt, int obj_size,
+					      int nobj, int reserved,
+					      int use_lowmem, int use_coherent)
+{
+	struct mthca_icm_table *table;
+	int obj_per_chunk;
+	int num_icm;
+	unsigned chunk_size;
+	int i;
+
+	obj_per_chunk = MTHCA_TABLE_CHUNK_SIZE / obj_size;
+	num_icm = DIV_ROUND_UP(nobj, obj_per_chunk);
+
+	table = kmalloc(struct_size(table, icm, num_icm), GFP_KERNEL);
+	if (!table)
+		return NULL;
+
+	table->virt     = virt;
+	table->num_icm  = num_icm;
+	table->num_obj  = nobj;
+	table->obj_size = obj_size;
+	table->lowmem   = use_lowmem;
+	table->coherent = use_coherent;
+	mutex_init(&table->mutex);
+
+	for (i = 0; i < num_icm; ++i)
+		table->icm[i] = NULL;
+
+	for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
+		chunk_size = MTHCA_TABLE_CHUNK_SIZE;
+		if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
+			chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
+
+		table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
+						(use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
+						__GFP_NOWARN, use_coherent);
+		if (!table->icm[i])
+			goto err;
+		if (mthca_MAP_ICM(dev, table->icm[i],
+				  virt + i * MTHCA_TABLE_CHUNK_SIZE)) {
+			mthca_free_icm(dev, table->icm[i], table->coherent);
+			table->icm[i] = NULL;
+			goto err;
+		}
+
+		/*
+		 * Add a reference to this ICM chunk so that it never
+		 * gets freed (since it contains reserved firmware objects).
+		 */
+		++table->icm[i]->refcount;
+	}
+
+	return table;
+
+err:
+	for (i = 0; i < num_icm; ++i)
+		if (table->icm[i]) {
+			mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
+					MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE);
+			mthca_free_icm(dev, table->icm[i], table->coherent);
+		}
+
+	kfree(table);
+
+	return NULL;
+}
+
+void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
+{
+	int i;
+
+	for (i = 0; i < table->num_icm; ++i)
+		if (table->icm[i]) {
+			mthca_UNMAP_ICM(dev,
+					table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
+					MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE);
+			mthca_free_icm(dev, table->icm[i], table->coherent);
+		}
+
+	kfree(table);
+}
+
+static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
+{
+	return dev->uar_table.uarc_base +
+		uar->index * dev->uar_table.uarc_size +
+		page * MTHCA_ICM_PAGE_SIZE;
+}
+
+int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+		      struct mthca_user_db_table *db_tab, int index, u64 uaddr)
+{
+	struct page *pages[1];
+	int ret = 0;
+	int i;
+
+	if (!mthca_is_memfree(dev))
+		return 0;
+
+	if (index < 0 || index > dev->uar_table.uarc_size / 8)
+		return -EINVAL;
+
+	mutex_lock(&db_tab->mutex);
+
+	i = index / MTHCA_DB_REC_PER_PAGE;
+
+	if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE)       ||
+	    (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
+	    (uaddr & 4095)) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	if (db_tab->page[i].refcount) {
+		++db_tab->page[i].refcount;
+		goto out;
+	}
+
+	ret = pin_user_pages_fast(uaddr & PAGE_MASK, 1,
+				  FOLL_WRITE | FOLL_LONGTERM, pages);
+	if (ret < 0)
+		goto out;
+
+	sg_set_page(&db_tab->page[i].mem, pages[0], MTHCA_ICM_PAGE_SIZE,
+			uaddr & ~PAGE_MASK);
+
+	ret = dma_map_sg(&dev->pdev->dev, &db_tab->page[i].mem, 1,
+			 DMA_TO_DEVICE);
+	if (ret < 0) {
+		unpin_user_page(pages[0]);
+		goto out;
+	}
+
+	ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
+				 mthca_uarc_virt(dev, uar, i));
+	if (ret) {
+		dma_unmap_sg(&dev->pdev->dev, &db_tab->page[i].mem, 1,
+			     DMA_TO_DEVICE);
+		unpin_user_page(sg_page(&db_tab->page[i].mem));
+		goto out;
+	}
+
+	db_tab->page[i].uvirt    = uaddr;
+	db_tab->page[i].refcount = 1;
+
+out:
+	mutex_unlock(&db_tab->mutex);
+	return ret;
+}
+
+void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+			 struct mthca_user_db_table *db_tab, int index)
+{
+	if (!mthca_is_memfree(dev))
+		return;
+
+	/*
+	 * To make our bookkeeping simpler, we don't unmap DB
+	 * pages until we clean up the whole db table.
+	 */
+
+	mutex_lock(&db_tab->mutex);
+
+	--db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
+
+	mutex_unlock(&db_tab->mutex);
+}
+
+struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
+{
+	struct mthca_user_db_table *db_tab;
+	int npages;
+	int i;
+
+	if (!mthca_is_memfree(dev))
+		return NULL;
+
+	npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
+	db_tab = kmalloc(struct_size(db_tab, page, npages), GFP_KERNEL);
+	if (!db_tab)
+		return ERR_PTR(-ENOMEM);
+
+	mutex_init(&db_tab->mutex);
+	for (i = 0; i < npages; ++i) {
+		db_tab->page[i].refcount = 0;
+		db_tab->page[i].uvirt    = 0;
+		sg_init_table(&db_tab->page[i].mem, 1);
+	}
+
+	return db_tab;
+}
+
+void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
+			       struct mthca_user_db_table *db_tab)
+{
+	int i;
+
+	if (!mthca_is_memfree(dev))
+		return;
+
+	for (i = 0; i < dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE; ++i) {
+		if (db_tab->page[i].uvirt) {
+			mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1);
+			dma_unmap_sg(&dev->pdev->dev, &db_tab->page[i].mem, 1,
+				     DMA_TO_DEVICE);
+			unpin_user_page(sg_page(&db_tab->page[i].mem));
+		}
+	}
+
+	kfree(db_tab);
+}
+
+int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
+		   u32 qn, __be32 **db)
+{
+	int group;
+	int start, end, dir;
+	int i, j;
+	struct mthca_db_page *page;
+	int ret = 0;
+
+	mutex_lock(&dev->db_tab->mutex);
+
+	switch (type) {
+	case MTHCA_DB_TYPE_CQ_ARM:
+	case MTHCA_DB_TYPE_SQ:
+		group = 0;
+		start = 0;
+		end   = dev->db_tab->max_group1;
+		dir   = 1;
+		break;
+
+	case MTHCA_DB_TYPE_CQ_SET_CI:
+	case MTHCA_DB_TYPE_RQ:
+	case MTHCA_DB_TYPE_SRQ:
+		group = 1;
+		start = dev->db_tab->npages - 1;
+		end   = dev->db_tab->min_group2;
+		dir   = -1;
+		break;
+
+	default:
+		ret = -EINVAL;
+		goto out;
+	}
+
+	for (i = start; i != end; i += dir)
+		if (dev->db_tab->page[i].db_rec &&
+		    !bitmap_full(dev->db_tab->page[i].used,
+				 MTHCA_DB_REC_PER_PAGE)) {
+			page = dev->db_tab->page + i;
+			goto found;
+		}
+
+	for (i = start; i != end; i += dir)
+		if (!dev->db_tab->page[i].db_rec) {
+			page = dev->db_tab->page + i;
+			goto alloc;
+		}
+
+	if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	if (group == 0)
+		++dev->db_tab->max_group1;
+	else
+		--dev->db_tab->min_group2;
+
+	page = dev->db_tab->page + end;
+
+alloc:
+	page->db_rec = dma_alloc_coherent(&dev->pdev->dev,
+					  MTHCA_ICM_PAGE_SIZE, &page->mapping,
+					  GFP_KERNEL);
+	if (!page->db_rec) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	ret = mthca_MAP_ICM_page(dev, page->mapping,
+				 mthca_uarc_virt(dev, &dev->driver_uar, i));
+	if (ret) {
+		dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
+				  page->db_rec, page->mapping);
+		goto out;
+	}
+
+	bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
+
+found:
+	j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
+	set_bit(j, page->used);
+
+	if (group == 1)
+		j = MTHCA_DB_REC_PER_PAGE - 1 - j;
+
+	ret = i * MTHCA_DB_REC_PER_PAGE + j;
+
+	page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
+
+	*db = (__be32 *) &page->db_rec[j];
+
+out:
+	mutex_unlock(&dev->db_tab->mutex);
+
+	return ret;
+}
+
+void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
+{
+	int i, j;
+	struct mthca_db_page *page;
+
+	i = db_index / MTHCA_DB_REC_PER_PAGE;
+	j = db_index % MTHCA_DB_REC_PER_PAGE;
+
+	page = dev->db_tab->page + i;
+
+	mutex_lock(&dev->db_tab->mutex);
+
+	page->db_rec[j] = 0;
+	if (i >= dev->db_tab->min_group2)
+		j = MTHCA_DB_REC_PER_PAGE - 1 - j;
+	clear_bit(j, page->used);
+
+	if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
+	    i >= dev->db_tab->max_group1 - 1) {
+		mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1);
+
+		dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
+				  page->db_rec, page->mapping);
+		page->db_rec = NULL;
+
+		if (i == dev->db_tab->max_group1) {
+			--dev->db_tab->max_group1;
+			/* XXX may be able to unmap more pages now */
+		}
+		if (i == dev->db_tab->min_group2)
+			++dev->db_tab->min_group2;
+	}
+
+	mutex_unlock(&dev->db_tab->mutex);
+}
+
+int mthca_init_db_tab(struct mthca_dev *dev)
+{
+	int i;
+
+	if (!mthca_is_memfree(dev))
+		return 0;
+
+	dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
+	if (!dev->db_tab)
+		return -ENOMEM;
+
+	mutex_init(&dev->db_tab->mutex);
+
+	dev->db_tab->npages     = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
+	dev->db_tab->max_group1 = 0;
+	dev->db_tab->min_group2 = dev->db_tab->npages - 1;
+
+	dev->db_tab->page = kmalloc_array(dev->db_tab->npages,
+					  sizeof(*dev->db_tab->page),
+					  GFP_KERNEL);
+	if (!dev->db_tab->page) {
+		kfree(dev->db_tab);
+		return -ENOMEM;
+	}
+
+	for (i = 0; i < dev->db_tab->npages; ++i)
+		dev->db_tab->page[i].db_rec = NULL;
+
+	return 0;
+}
+
+void mthca_cleanup_db_tab(struct mthca_dev *dev)
+{
+	int i;
+
+	if (!mthca_is_memfree(dev))
+		return;
+
+	/*
+	 * Because we don't always free our UARC pages when they
+	 * become empty to make mthca_free_db() simpler we need to
+	 * make a sweep through the doorbell pages and free any
+	 * leftover pages now.
+	 */
+	for (i = 0; i < dev->db_tab->npages; ++i) {
+		if (!dev->db_tab->page[i].db_rec)
+			continue;
+
+		if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
+			mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
+
+		mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1);
+
+		dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
+				  dev->db_tab->page[i].db_rec,
+				  dev->db_tab->page[i].mapping);
+	}
+
+	kfree(dev->db_tab->page);
+	kfree(dev->db_tab);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_memfree.h b/drivers/infiniband/hw/mthca/mthca_memfree.h
new file mode 100644
index 000000000..f9a2e65e2
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_memfree.h
@@ -0,0 +1,179 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef MTHCA_MEMFREE_H
+#define MTHCA_MEMFREE_H
+
+#include <linux/list.h>
+#include <linux/mutex.h>
+
+#define MTHCA_ICM_CHUNK_LEN \
+	((256 - sizeof (struct list_head) - 2 * sizeof (int)) /		\
+	 (sizeof (struct scatterlist)))
+
+enum {
+	MTHCA_ICM_PAGE_SHIFT	= 12,
+	MTHCA_ICM_PAGE_SIZE	= 1 << MTHCA_ICM_PAGE_SHIFT,
+	MTHCA_DB_REC_PER_PAGE	= MTHCA_ICM_PAGE_SIZE / 8
+};
+
+struct mthca_icm_chunk {
+	struct list_head   list;
+	int                npages;
+	int                nsg;
+	struct scatterlist mem[MTHCA_ICM_CHUNK_LEN];
+};
+
+struct mthca_icm {
+	struct list_head chunk_list;
+	int              refcount;
+};
+
+struct mthca_icm_table {
+	u64               virt;
+	int               num_icm;
+	int               num_obj;
+	int               obj_size;
+	int               lowmem;
+	int               coherent;
+	struct mutex      mutex;
+	struct mthca_icm *icm[];
+};
+
+struct mthca_icm_iter {
+	struct mthca_icm       *icm;
+	struct mthca_icm_chunk *chunk;
+	int                     page_idx;
+};
+
+struct mthca_dev;
+
+struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
+				  gfp_t gfp_mask, int coherent);
+void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent);
+
+struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
+					      u64 virt, int obj_size,
+					      int nobj, int reserved,
+					      int use_lowmem, int use_coherent);
+void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table);
+int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj);
+void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj);
+void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle);
+int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
+			  int start, int end);
+void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
+			   int start, int end);
+
+static inline void mthca_icm_first(struct mthca_icm *icm,
+				   struct mthca_icm_iter *iter)
+{
+	iter->icm      = icm;
+	iter->chunk    = list_empty(&icm->chunk_list) ?
+		NULL : list_entry(icm->chunk_list.next,
+				  struct mthca_icm_chunk, list);
+	iter->page_idx = 0;
+}
+
+static inline int mthca_icm_last(struct mthca_icm_iter *iter)
+{
+	return !iter->chunk;
+}
+
+static inline void mthca_icm_next(struct mthca_icm_iter *iter)
+{
+	if (++iter->page_idx >= iter->chunk->nsg) {
+		if (iter->chunk->list.next == &iter->icm->chunk_list) {
+			iter->chunk = NULL;
+			return;
+		}
+
+		iter->chunk = list_entry(iter->chunk->list.next,
+					 struct mthca_icm_chunk, list);
+		iter->page_idx = 0;
+	}
+}
+
+static inline dma_addr_t mthca_icm_addr(struct mthca_icm_iter *iter)
+{
+	return sg_dma_address(&iter->chunk->mem[iter->page_idx]);
+}
+
+static inline unsigned long mthca_icm_size(struct mthca_icm_iter *iter)
+{
+	return sg_dma_len(&iter->chunk->mem[iter->page_idx]);
+}
+
+struct mthca_db_page {
+	DECLARE_BITMAP(used, MTHCA_DB_REC_PER_PAGE);
+	__be64    *db_rec;
+	dma_addr_t mapping;
+};
+
+struct mthca_db_table {
+	int 	       	      npages;
+	int 	       	      max_group1;
+	int 	       	      min_group2;
+	struct mthca_db_page *page;
+	struct mutex          mutex;
+};
+
+enum mthca_db_type {
+	MTHCA_DB_TYPE_INVALID   = 0x0,
+	MTHCA_DB_TYPE_CQ_SET_CI = 0x1,
+	MTHCA_DB_TYPE_CQ_ARM    = 0x2,
+	MTHCA_DB_TYPE_SQ        = 0x3,
+	MTHCA_DB_TYPE_RQ        = 0x4,
+	MTHCA_DB_TYPE_SRQ       = 0x5,
+	MTHCA_DB_TYPE_GROUP_SEP = 0x7
+};
+
+struct mthca_user_db_table;
+struct mthca_uar;
+
+int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+		      struct mthca_user_db_table *db_tab, int index, u64 uaddr);
+void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+			 struct mthca_user_db_table *db_tab, int index);
+struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev);
+void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
+			       struct mthca_user_db_table *db_tab);
+
+int mthca_init_db_tab(struct mthca_dev *dev);
+void mthca_cleanup_db_tab(struct mthca_dev *dev);
+int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
+		   u32 qn, __be32 **db);
+void mthca_free_db(struct mthca_dev *dev, int type, int db_index);
+
+#endif /* MTHCA_MEMFREE_H */
diff --git a/drivers/infiniband/hw/mthca/mthca_mr.c b/drivers/infiniband/hw/mthca/mthca_mr.c
new file mode 100644
index 000000000..192f83fd7
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_mr.c
@@ -0,0 +1,702 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/slab.h>
+#include <linux/errno.h>
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+#include "mthca_memfree.h"
+
+struct mthca_mtt {
+	struct mthca_buddy *buddy;
+	int                 order;
+	u32                 first_seg;
+};
+
+/*
+ * Must be packed because mtt_seg is 64 bits but only aligned to 32 bits.
+ */
+struct mthca_mpt_entry {
+	__be32 flags;
+	__be32 page_size;
+	__be32 key;
+	__be32 pd;
+	__be64 start;
+	__be64 length;
+	__be32 lkey;
+	__be32 window_count;
+	__be32 window_count_limit;
+	__be64 mtt_seg;
+	__be32 mtt_sz;		/* Arbel only */
+	u32    reserved[2];
+} __packed;
+
+#define MTHCA_MPT_FLAG_SW_OWNS       (0xfUL << 28)
+#define MTHCA_MPT_FLAG_MIO           (1 << 17)
+#define MTHCA_MPT_FLAG_BIND_ENABLE   (1 << 15)
+#define MTHCA_MPT_FLAG_PHYSICAL      (1 <<  9)
+#define MTHCA_MPT_FLAG_REGION        (1 <<  8)
+
+#define MTHCA_MTT_FLAG_PRESENT       1
+
+#define MTHCA_MPT_STATUS_SW 0xF0
+#define MTHCA_MPT_STATUS_HW 0x00
+
+#define SINAI_FMR_KEY_INC 0x1000000
+
+/*
+ * Buddy allocator for MTT segments (currently not very efficient
+ * since it doesn't keep a free list and just searches linearly
+ * through the bitmaps)
+ */
+
+static u32 mthca_buddy_alloc(struct mthca_buddy *buddy, int order)
+{
+	int o;
+	int m;
+	u32 seg;
+
+	spin_lock(&buddy->lock);
+
+	for (o = order; o <= buddy->max_order; ++o)
+		if (buddy->num_free[o]) {
+			m = 1 << (buddy->max_order - o);
+			seg = find_first_bit(buddy->bits[o], m);
+			if (seg < m)
+				goto found;
+		}
+
+	spin_unlock(&buddy->lock);
+	return -1;
+
+ found:
+	__clear_bit(seg, buddy->bits[o]);
+	--buddy->num_free[o];
+
+	while (o > order) {
+		--o;
+		seg <<= 1;
+		__set_bit(seg ^ 1, buddy->bits[o]);
+		++buddy->num_free[o];
+	}
+
+	spin_unlock(&buddy->lock);
+
+	seg <<= order;
+
+	return seg;
+}
+
+static void mthca_buddy_free(struct mthca_buddy *buddy, u32 seg, int order)
+{
+	seg >>= order;
+
+	spin_lock(&buddy->lock);
+
+	while (test_bit(seg ^ 1, buddy->bits[order])) {
+		__clear_bit(seg ^ 1, buddy->bits[order]);
+		--buddy->num_free[order];
+		seg >>= 1;
+		++order;
+	}
+
+	__set_bit(seg, buddy->bits[order]);
+	++buddy->num_free[order];
+
+	spin_unlock(&buddy->lock);
+}
+
+static int mthca_buddy_init(struct mthca_buddy *buddy, int max_order)
+{
+	int i;
+
+	buddy->max_order = max_order;
+	spin_lock_init(&buddy->lock);
+
+	buddy->bits = kcalloc(buddy->max_order + 1, sizeof(long *),
+			      GFP_KERNEL);
+	buddy->num_free = kcalloc((buddy->max_order + 1), sizeof *buddy->num_free,
+				  GFP_KERNEL);
+	if (!buddy->bits || !buddy->num_free)
+		goto err_out;
+
+	for (i = 0; i <= buddy->max_order; ++i) {
+		buddy->bits[i] = bitmap_zalloc(1 << (buddy->max_order - i),
+					       GFP_KERNEL);
+		if (!buddy->bits[i])
+			goto err_out_free;
+	}
+
+	__set_bit(0, buddy->bits[buddy->max_order]);
+	buddy->num_free[buddy->max_order] = 1;
+
+	return 0;
+
+err_out_free:
+	for (i = 0; i <= buddy->max_order; ++i)
+		bitmap_free(buddy->bits[i]);
+
+err_out:
+	kfree(buddy->bits);
+	kfree(buddy->num_free);
+
+	return -ENOMEM;
+}
+
+static void mthca_buddy_cleanup(struct mthca_buddy *buddy)
+{
+	int i;
+
+	for (i = 0; i <= buddy->max_order; ++i)
+		bitmap_free(buddy->bits[i]);
+
+	kfree(buddy->bits);
+	kfree(buddy->num_free);
+}
+
+static u32 mthca_alloc_mtt_range(struct mthca_dev *dev, int order,
+				 struct mthca_buddy *buddy)
+{
+	u32 seg = mthca_buddy_alloc(buddy, order);
+
+	if (seg == -1)
+		return -1;
+
+	if (mthca_is_memfree(dev))
+		if (mthca_table_get_range(dev, dev->mr_table.mtt_table, seg,
+					  seg + (1 << order) - 1)) {
+			mthca_buddy_free(buddy, seg, order);
+			seg = -1;
+		}
+
+	return seg;
+}
+
+static struct mthca_mtt *__mthca_alloc_mtt(struct mthca_dev *dev, int size,
+					   struct mthca_buddy *buddy)
+{
+	struct mthca_mtt *mtt;
+	int i;
+
+	if (size <= 0)
+		return ERR_PTR(-EINVAL);
+
+	mtt = kmalloc(sizeof *mtt, GFP_KERNEL);
+	if (!mtt)
+		return ERR_PTR(-ENOMEM);
+
+	mtt->buddy = buddy;
+	mtt->order = 0;
+	for (i = dev->limits.mtt_seg_size / 8; i < size; i <<= 1)
+		++mtt->order;
+
+	mtt->first_seg = mthca_alloc_mtt_range(dev, mtt->order, buddy);
+	if (mtt->first_seg == -1) {
+		kfree(mtt);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	return mtt;
+}
+
+struct mthca_mtt *mthca_alloc_mtt(struct mthca_dev *dev, int size)
+{
+	return __mthca_alloc_mtt(dev, size, &dev->mr_table.mtt_buddy);
+}
+
+void mthca_free_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt)
+{
+	if (!mtt)
+		return;
+
+	mthca_buddy_free(mtt->buddy, mtt->first_seg, mtt->order);
+
+	mthca_table_put_range(dev, dev->mr_table.mtt_table,
+			      mtt->first_seg,
+			      mtt->first_seg + (1 << mtt->order) - 1);
+
+	kfree(mtt);
+}
+
+static int __mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,
+			     int start_index, u64 *buffer_list, int list_len)
+{
+	struct mthca_mailbox *mailbox;
+	__be64 *mtt_entry;
+	int err = 0;
+	int i;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	mtt_entry = mailbox->buf;
+
+	while (list_len > 0) {
+		mtt_entry[0] = cpu_to_be64(dev->mr_table.mtt_base +
+					   mtt->first_seg * dev->limits.mtt_seg_size +
+					   start_index * 8);
+		mtt_entry[1] = 0;
+		for (i = 0; i < list_len && i < MTHCA_MAILBOX_SIZE / 8 - 2; ++i)
+			mtt_entry[i + 2] = cpu_to_be64(buffer_list[i] |
+						       MTHCA_MTT_FLAG_PRESENT);
+
+		/*
+		 * If we have an odd number of entries to write, add
+		 * one more dummy entry for firmware efficiency.
+		 */
+		if (i & 1)
+			mtt_entry[i + 2] = 0;
+
+		err = mthca_WRITE_MTT(dev, mailbox, (i + 1) & ~1);
+		if (err) {
+			mthca_warn(dev, "WRITE_MTT failed (%d)\n", err);
+			goto out;
+		}
+
+		list_len    -= i;
+		start_index += i;
+		buffer_list += i;
+	}
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_write_mtt_size(struct mthca_dev *dev)
+{
+	if (dev->mr_table.fmr_mtt_buddy != &dev->mr_table.mtt_buddy ||
+	    !(dev->mthca_flags & MTHCA_FLAG_FMR))
+		/*
+		 * Be friendly to WRITE_MTT command
+		 * and leave two empty slots for the
+		 * index and reserved fields of the
+		 * mailbox.
+		 */
+		return PAGE_SIZE / sizeof (u64) - 2;
+
+	/* For Arbel, all MTTs must fit in the same page. */
+	return mthca_is_memfree(dev) ? (PAGE_SIZE / sizeof (u64)) : 0x7ffffff;
+}
+
+static void mthca_tavor_write_mtt_seg(struct mthca_dev *dev,
+				      struct mthca_mtt *mtt, int start_index,
+				      u64 *buffer_list, int list_len)
+{
+	u64 __iomem *mtts;
+	int i;
+
+	mtts = dev->mr_table.tavor_fmr.mtt_base + mtt->first_seg * dev->limits.mtt_seg_size +
+		start_index * sizeof (u64);
+	for (i = 0; i < list_len; ++i)
+		mthca_write64_raw(cpu_to_be64(buffer_list[i] | MTHCA_MTT_FLAG_PRESENT),
+				  mtts + i);
+}
+
+static void mthca_arbel_write_mtt_seg(struct mthca_dev *dev,
+				      struct mthca_mtt *mtt, int start_index,
+				      u64 *buffer_list, int list_len)
+{
+	__be64 *mtts;
+	dma_addr_t dma_handle;
+	int i;
+	int s = start_index * sizeof (u64);
+
+	/* For Arbel, all MTTs must fit in the same page. */
+	BUG_ON(s / PAGE_SIZE != (s + list_len * sizeof(u64) - 1) / PAGE_SIZE);
+	/* Require full segments */
+	BUG_ON(s % dev->limits.mtt_seg_size);
+
+	mtts = mthca_table_find(dev->mr_table.mtt_table, mtt->first_seg +
+				s / dev->limits.mtt_seg_size, &dma_handle);
+
+	BUG_ON(!mtts);
+
+	dma_sync_single_for_cpu(&dev->pdev->dev, dma_handle,
+				list_len * sizeof (u64), DMA_TO_DEVICE);
+
+	for (i = 0; i < list_len; ++i)
+		mtts[i] = cpu_to_be64(buffer_list[i] | MTHCA_MTT_FLAG_PRESENT);
+
+	dma_sync_single_for_device(&dev->pdev->dev, dma_handle,
+				   list_len * sizeof (u64), DMA_TO_DEVICE);
+}
+
+int mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,
+		    int start_index, u64 *buffer_list, int list_len)
+{
+	int size = mthca_write_mtt_size(dev);
+	int chunk;
+
+	if (dev->mr_table.fmr_mtt_buddy != &dev->mr_table.mtt_buddy ||
+	    !(dev->mthca_flags & MTHCA_FLAG_FMR))
+		return __mthca_write_mtt(dev, mtt, start_index, buffer_list, list_len);
+
+	while (list_len > 0) {
+		chunk = min(size, list_len);
+		if (mthca_is_memfree(dev))
+			mthca_arbel_write_mtt_seg(dev, mtt, start_index,
+						  buffer_list, chunk);
+		else
+			mthca_tavor_write_mtt_seg(dev, mtt, start_index,
+						  buffer_list, chunk);
+
+		list_len    -= chunk;
+		start_index += chunk;
+		buffer_list += chunk;
+	}
+
+	return 0;
+}
+
+static inline u32 tavor_hw_index_to_key(u32 ind)
+{
+	return ind;
+}
+
+static inline u32 tavor_key_to_hw_index(u32 key)
+{
+	return key;
+}
+
+static inline u32 arbel_hw_index_to_key(u32 ind)
+{
+	return (ind >> 24) | (ind << 8);
+}
+
+static inline u32 arbel_key_to_hw_index(u32 key)
+{
+	return (key << 24) | (key >> 8);
+}
+
+static inline u32 hw_index_to_key(struct mthca_dev *dev, u32 ind)
+{
+	if (mthca_is_memfree(dev))
+		return arbel_hw_index_to_key(ind);
+	else
+		return tavor_hw_index_to_key(ind);
+}
+
+static inline u32 key_to_hw_index(struct mthca_dev *dev, u32 key)
+{
+	if (mthca_is_memfree(dev))
+		return arbel_key_to_hw_index(key);
+	else
+		return tavor_key_to_hw_index(key);
+}
+
+static inline u32 adjust_key(struct mthca_dev *dev, u32 key)
+{
+	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
+		return ((key << 20) & 0x800000) | (key & 0x7fffff);
+	else
+		return key;
+}
+
+int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift,
+		   u64 iova, u64 total_size, u32 access, struct mthca_mr *mr)
+{
+	struct mthca_mailbox *mailbox;
+	struct mthca_mpt_entry *mpt_entry;
+	u32 key;
+	int i;
+	int err;
+
+	WARN_ON(buffer_size_shift >= 32);
+
+	key = mthca_alloc(&dev->mr_table.mpt_alloc);
+	if (key == -1)
+		return -ENOMEM;
+	key = adjust_key(dev, key);
+	mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key);
+
+	if (mthca_is_memfree(dev)) {
+		err = mthca_table_get(dev, dev->mr_table.mpt_table, key);
+		if (err)
+			goto err_out_mpt_free;
+	}
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		err = PTR_ERR(mailbox);
+		goto err_out_table;
+	}
+	mpt_entry = mailbox->buf;
+
+	mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS     |
+				       MTHCA_MPT_FLAG_MIO         |
+				       MTHCA_MPT_FLAG_REGION      |
+				       access);
+	if (!mr->mtt)
+		mpt_entry->flags |= cpu_to_be32(MTHCA_MPT_FLAG_PHYSICAL);
+
+	mpt_entry->page_size = cpu_to_be32(buffer_size_shift - 12);
+	mpt_entry->key       = cpu_to_be32(key);
+	mpt_entry->pd        = cpu_to_be32(pd);
+	mpt_entry->start     = cpu_to_be64(iova);
+	mpt_entry->length    = cpu_to_be64(total_size);
+
+	memset_startat(mpt_entry, 0, lkey);
+
+	if (mr->mtt)
+		mpt_entry->mtt_seg =
+			cpu_to_be64(dev->mr_table.mtt_base +
+				    mr->mtt->first_seg * dev->limits.mtt_seg_size);
+
+	if (0) {
+		mthca_dbg(dev, "Dumping MPT entry %08x:\n", mr->ibmr.lkey);
+		for (i = 0; i < sizeof (struct mthca_mpt_entry) / 4; ++i) {
+			if (i % 4 == 0)
+				printk("[%02x] ", i * 4);
+			printk(" %08x", be32_to_cpu(((__be32 *) mpt_entry)[i]));
+			if ((i + 1) % 4 == 0)
+				printk("\n");
+		}
+	}
+
+	err = mthca_SW2HW_MPT(dev, mailbox,
+			      key & (dev->limits.num_mpts - 1));
+	if (err) {
+		mthca_warn(dev, "SW2HW_MPT failed (%d)\n", err);
+		goto err_out_mailbox;
+	}
+
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+
+err_out_mailbox:
+	mthca_free_mailbox(dev, mailbox);
+
+err_out_table:
+	mthca_table_put(dev, dev->mr_table.mpt_table, key);
+
+err_out_mpt_free:
+	mthca_free(&dev->mr_table.mpt_alloc, key);
+	return err;
+}
+
+int mthca_mr_alloc_notrans(struct mthca_dev *dev, u32 pd,
+			   u32 access, struct mthca_mr *mr)
+{
+	mr->mtt = NULL;
+	return mthca_mr_alloc(dev, pd, 12, 0, ~0ULL, access, mr);
+}
+
+int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
+			u64 *buffer_list, int buffer_size_shift,
+			int list_len, u64 iova, u64 total_size,
+			u32 access, struct mthca_mr *mr)
+{
+	int err;
+
+	mr->mtt = mthca_alloc_mtt(dev, list_len);
+	if (IS_ERR(mr->mtt))
+		return PTR_ERR(mr->mtt);
+
+	err = mthca_write_mtt(dev, mr->mtt, 0, buffer_list, list_len);
+	if (err) {
+		mthca_free_mtt(dev, mr->mtt);
+		return err;
+	}
+
+	err = mthca_mr_alloc(dev, pd, buffer_size_shift, iova,
+			     total_size, access, mr);
+	if (err)
+		mthca_free_mtt(dev, mr->mtt);
+
+	return err;
+}
+
+/* Free mr */
+static void mthca_free_region(struct mthca_dev *dev, u32 lkey)
+{
+	mthca_table_put(dev, dev->mr_table.mpt_table,
+			key_to_hw_index(dev, lkey));
+
+	mthca_free(&dev->mr_table.mpt_alloc, key_to_hw_index(dev, lkey));
+}
+
+void mthca_free_mr(struct mthca_dev *dev, struct mthca_mr *mr)
+{
+	int err;
+
+	err = mthca_HW2SW_MPT(dev, NULL,
+			      key_to_hw_index(dev, mr->ibmr.lkey) &
+			      (dev->limits.num_mpts - 1));
+	if (err)
+		mthca_warn(dev, "HW2SW_MPT failed (%d)\n", err);
+
+	mthca_free_region(dev, mr->ibmr.lkey);
+	mthca_free_mtt(dev, mr->mtt);
+}
+
+int mthca_init_mr_table(struct mthca_dev *dev)
+{
+	phys_addr_t addr;
+	int mpts, mtts, err, i;
+
+	err = mthca_alloc_init(&dev->mr_table.mpt_alloc,
+			       dev->limits.num_mpts,
+			       ~0, dev->limits.reserved_mrws);
+	if (err)
+		return err;
+
+	if (!mthca_is_memfree(dev) &&
+	    (dev->mthca_flags & MTHCA_FLAG_DDR_HIDDEN))
+		dev->limits.fmr_reserved_mtts = 0;
+	else
+		dev->mthca_flags |= MTHCA_FLAG_FMR;
+
+	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
+		mthca_dbg(dev, "Memory key throughput optimization activated.\n");
+
+	err = mthca_buddy_init(&dev->mr_table.mtt_buddy,
+			       fls(dev->limits.num_mtt_segs - 1));
+
+	if (err)
+		goto err_mtt_buddy;
+
+	dev->mr_table.tavor_fmr.mpt_base = NULL;
+	dev->mr_table.tavor_fmr.mtt_base = NULL;
+
+	if (dev->limits.fmr_reserved_mtts) {
+		i = fls(dev->limits.fmr_reserved_mtts - 1);
+
+		if (i >= 31) {
+			mthca_warn(dev, "Unable to reserve 2^31 FMR MTTs.\n");
+			err = -EINVAL;
+			goto err_fmr_mpt;
+		}
+		mpts = mtts = 1 << i;
+	} else {
+		mtts = dev->limits.num_mtt_segs;
+		mpts = dev->limits.num_mpts;
+	}
+
+	if (!mthca_is_memfree(dev) &&
+	    (dev->mthca_flags & MTHCA_FLAG_FMR)) {
+
+		addr = pci_resource_start(dev->pdev, 4) +
+			((pci_resource_len(dev->pdev, 4) - 1) &
+			 dev->mr_table.mpt_base);
+
+		dev->mr_table.tavor_fmr.mpt_base =
+			ioremap(addr, mpts * sizeof(struct mthca_mpt_entry));
+
+		if (!dev->mr_table.tavor_fmr.mpt_base) {
+			mthca_warn(dev, "MPT ioremap for FMR failed.\n");
+			err = -ENOMEM;
+			goto err_fmr_mpt;
+		}
+
+		addr = pci_resource_start(dev->pdev, 4) +
+			((pci_resource_len(dev->pdev, 4) - 1) &
+			 dev->mr_table.mtt_base);
+
+		dev->mr_table.tavor_fmr.mtt_base =
+			ioremap(addr, mtts * dev->limits.mtt_seg_size);
+		if (!dev->mr_table.tavor_fmr.mtt_base) {
+			mthca_warn(dev, "MTT ioremap for FMR failed.\n");
+			err = -ENOMEM;
+			goto err_fmr_mtt;
+		}
+	}
+
+	if (dev->limits.fmr_reserved_mtts) {
+		err = mthca_buddy_init(&dev->mr_table.tavor_fmr.mtt_buddy, fls(mtts - 1));
+		if (err)
+			goto err_fmr_mtt_buddy;
+
+		/* Prevent regular MRs from using FMR keys */
+		err = mthca_buddy_alloc(&dev->mr_table.mtt_buddy, fls(mtts - 1));
+		if (err)
+			goto err_reserve_fmr;
+
+		dev->mr_table.fmr_mtt_buddy =
+			&dev->mr_table.tavor_fmr.mtt_buddy;
+	} else
+		dev->mr_table.fmr_mtt_buddy = &dev->mr_table.mtt_buddy;
+
+	/* FMR table is always the first, take reserved MTTs out of there */
+	if (dev->limits.reserved_mtts) {
+		i = fls(dev->limits.reserved_mtts - 1);
+
+		if (mthca_alloc_mtt_range(dev, i,
+					  dev->mr_table.fmr_mtt_buddy) == -1) {
+			mthca_warn(dev, "MTT table of order %d is too small.\n",
+				  dev->mr_table.fmr_mtt_buddy->max_order);
+			err = -ENOMEM;
+			goto err_reserve_mtts;
+		}
+	}
+
+	return 0;
+
+err_reserve_mtts:
+err_reserve_fmr:
+	if (dev->limits.fmr_reserved_mtts)
+		mthca_buddy_cleanup(&dev->mr_table.tavor_fmr.mtt_buddy);
+
+err_fmr_mtt_buddy:
+	if (dev->mr_table.tavor_fmr.mtt_base)
+		iounmap(dev->mr_table.tavor_fmr.mtt_base);
+
+err_fmr_mtt:
+	if (dev->mr_table.tavor_fmr.mpt_base)
+		iounmap(dev->mr_table.tavor_fmr.mpt_base);
+
+err_fmr_mpt:
+	mthca_buddy_cleanup(&dev->mr_table.mtt_buddy);
+
+err_mtt_buddy:
+	mthca_alloc_cleanup(&dev->mr_table.mpt_alloc);
+
+	return err;
+}
+
+void mthca_cleanup_mr_table(struct mthca_dev *dev)
+{
+	/* XXX check if any MRs are still allocated? */
+	if (dev->limits.fmr_reserved_mtts)
+		mthca_buddy_cleanup(&dev->mr_table.tavor_fmr.mtt_buddy);
+
+	mthca_buddy_cleanup(&dev->mr_table.mtt_buddy);
+
+	if (dev->mr_table.tavor_fmr.mtt_base)
+		iounmap(dev->mr_table.tavor_fmr.mtt_base);
+	if (dev->mr_table.tavor_fmr.mpt_base)
+		iounmap(dev->mr_table.tavor_fmr.mpt_base);
+
+	mthca_alloc_cleanup(&dev->mr_table.mpt_alloc);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_pd.c b/drivers/infiniband/hw/mthca/mthca_pd.c
new file mode 100644
index 000000000..266f14e47
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_pd.c
@@ -0,0 +1,81 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/errno.h>
+
+#include "mthca_dev.h"
+
+int mthca_pd_alloc(struct mthca_dev *dev, int privileged, struct mthca_pd *pd)
+{
+	int err = 0;
+
+	pd->privileged = privileged;
+
+	atomic_set(&pd->sqp_count, 0);
+	pd->pd_num = mthca_alloc(&dev->pd_table.alloc);
+	if (pd->pd_num == -1)
+		return -ENOMEM;
+
+	if (privileged) {
+		err = mthca_mr_alloc_notrans(dev, pd->pd_num,
+					     MTHCA_MPT_FLAG_LOCAL_READ |
+					     MTHCA_MPT_FLAG_LOCAL_WRITE,
+					     &pd->ntmr);
+		if (err)
+			mthca_free(&dev->pd_table.alloc, pd->pd_num);
+	}
+
+	return err;
+}
+
+void mthca_pd_free(struct mthca_dev *dev, struct mthca_pd *pd)
+{
+	if (pd->privileged)
+		mthca_free_mr(dev, &pd->ntmr);
+	mthca_free(&dev->pd_table.alloc, pd->pd_num);
+}
+
+int mthca_init_pd_table(struct mthca_dev *dev)
+{
+	return mthca_alloc_init(&dev->pd_table.alloc,
+				dev->limits.num_pds,
+				(1 << 24) - 1,
+				dev->limits.reserved_pds);
+}
+
+void mthca_cleanup_pd_table(struct mthca_dev *dev)
+{
+	/* XXX check if any PDs are still allocated? */
+	mthca_alloc_cleanup(&dev->pd_table.alloc);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_profile.c b/drivers/infiniband/hw/mthca/mthca_profile.c
new file mode 100644
index 000000000..69af65f1b
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_profile.c
@@ -0,0 +1,279 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/string.h>
+#include <linux/slab.h>
+
+#include "mthca_profile.h"
+
+enum {
+	MTHCA_RES_QP,
+	MTHCA_RES_EEC,
+	MTHCA_RES_SRQ,
+	MTHCA_RES_CQ,
+	MTHCA_RES_EQP,
+	MTHCA_RES_EEEC,
+	MTHCA_RES_EQ,
+	MTHCA_RES_RDB,
+	MTHCA_RES_MCG,
+	MTHCA_RES_MPT,
+	MTHCA_RES_MTT,
+	MTHCA_RES_UAR,
+	MTHCA_RES_UDAV,
+	MTHCA_RES_UARC,
+	MTHCA_RES_NUM
+};
+
+enum {
+	MTHCA_NUM_EQS = 32,
+	MTHCA_NUM_PDS = 1 << 15
+};
+
+s64 mthca_make_profile(struct mthca_dev *dev,
+		       struct mthca_profile *request,
+		       struct mthca_dev_lim *dev_lim,
+		       struct mthca_init_hca_param *init_hca)
+{
+	struct mthca_resource {
+		u64 size;
+		u64 start;
+		int type;
+		int num;
+		int log_num;
+	};
+
+	u64 mem_base, mem_avail;
+	s64 total_size = 0;
+	struct mthca_resource *profile;
+	int i, j;
+
+	profile = kcalloc(MTHCA_RES_NUM, sizeof(*profile), GFP_KERNEL);
+	if (!profile)
+		return -ENOMEM;
+
+	profile[MTHCA_RES_QP].size   = dev_lim->qpc_entry_sz;
+	profile[MTHCA_RES_EEC].size  = dev_lim->eec_entry_sz;
+	profile[MTHCA_RES_SRQ].size  = dev_lim->srq_entry_sz;
+	profile[MTHCA_RES_CQ].size   = dev_lim->cqc_entry_sz;
+	profile[MTHCA_RES_EQP].size  = dev_lim->eqpc_entry_sz;
+	profile[MTHCA_RES_EEEC].size = dev_lim->eeec_entry_sz;
+	profile[MTHCA_RES_EQ].size   = dev_lim->eqc_entry_sz;
+	profile[MTHCA_RES_RDB].size  = MTHCA_RDB_ENTRY_SIZE;
+	profile[MTHCA_RES_MCG].size  = MTHCA_MGM_ENTRY_SIZE;
+	profile[MTHCA_RES_MPT].size  = dev_lim->mpt_entry_sz;
+	profile[MTHCA_RES_MTT].size  = dev->limits.mtt_seg_size;
+	profile[MTHCA_RES_UAR].size  = dev_lim->uar_scratch_entry_sz;
+	profile[MTHCA_RES_UDAV].size = MTHCA_AV_SIZE;
+	profile[MTHCA_RES_UARC].size = request->uarc_size;
+
+	profile[MTHCA_RES_QP].num    = request->num_qp;
+	profile[MTHCA_RES_SRQ].num   = request->num_srq;
+	profile[MTHCA_RES_EQP].num   = request->num_qp;
+	profile[MTHCA_RES_RDB].num   = request->num_qp * request->rdb_per_qp;
+	profile[MTHCA_RES_CQ].num    = request->num_cq;
+	profile[MTHCA_RES_EQ].num    = MTHCA_NUM_EQS;
+	profile[MTHCA_RES_MCG].num   = request->num_mcg;
+	profile[MTHCA_RES_MPT].num   = request->num_mpt;
+	profile[MTHCA_RES_MTT].num   = request->num_mtt;
+	profile[MTHCA_RES_UAR].num   = request->num_uar;
+	profile[MTHCA_RES_UARC].num  = request->num_uar;
+	profile[MTHCA_RES_UDAV].num  = request->num_udav;
+
+	for (i = 0; i < MTHCA_RES_NUM; ++i) {
+		profile[i].type     = i;
+		profile[i].log_num  = max(ffs(profile[i].num) - 1, 0);
+		profile[i].size    *= profile[i].num;
+		if (mthca_is_memfree(dev))
+			profile[i].size = max(profile[i].size, (u64) PAGE_SIZE);
+	}
+
+	if (mthca_is_memfree(dev)) {
+		mem_base  = 0;
+		mem_avail = dev_lim->hca.arbel.max_icm_sz;
+	} else {
+		mem_base  = dev->ddr_start;
+		mem_avail = dev->fw.tavor.fw_start - dev->ddr_start;
+	}
+
+	/*
+	 * Sort the resources in decreasing order of size.  Since they
+	 * all have sizes that are powers of 2, we'll be able to keep
+	 * resources aligned to their size and pack them without gaps
+	 * using the sorted order.
+	 */
+	for (i = MTHCA_RES_NUM; i > 0; --i)
+		for (j = 1; j < i; ++j) {
+			if (profile[j].size > profile[j - 1].size)
+				swap(profile[j], profile[j - 1]);
+		}
+
+	for (i = 0; i < MTHCA_RES_NUM; ++i) {
+		if (profile[i].size) {
+			profile[i].start = mem_base + total_size;
+			total_size      += profile[i].size;
+		}
+		if (total_size > mem_avail) {
+			mthca_err(dev, "Profile requires 0x%llx bytes; "
+				  "won't fit in 0x%llx bytes of context memory.\n",
+				  (unsigned long long) total_size,
+				  (unsigned long long) mem_avail);
+			kfree(profile);
+			return -ENOMEM;
+		}
+
+		if (profile[i].size)
+			mthca_dbg(dev, "profile[%2d]--%2d/%2d @ 0x%16llx "
+				  "(size 0x%8llx)\n",
+				  i, profile[i].type, profile[i].log_num,
+				  (unsigned long long) profile[i].start,
+				  (unsigned long long) profile[i].size);
+	}
+
+	if (mthca_is_memfree(dev))
+		mthca_dbg(dev, "HCA context memory: reserving %d KB\n",
+			  (int) (total_size >> 10));
+	else
+		mthca_dbg(dev, "HCA memory: allocated %d KB/%d KB (%d KB free)\n",
+			  (int) (total_size >> 10), (int) (mem_avail >> 10),
+			  (int) ((mem_avail - total_size) >> 10));
+
+	for (i = 0; i < MTHCA_RES_NUM; ++i) {
+		switch (profile[i].type) {
+		case MTHCA_RES_QP:
+			dev->limits.num_qps   = profile[i].num;
+			init_hca->qpc_base    = profile[i].start;
+			init_hca->log_num_qps = profile[i].log_num;
+			break;
+		case MTHCA_RES_EEC:
+			dev->limits.num_eecs   = profile[i].num;
+			init_hca->eec_base     = profile[i].start;
+			init_hca->log_num_eecs = profile[i].log_num;
+			break;
+		case MTHCA_RES_SRQ:
+			dev->limits.num_srqs   = profile[i].num;
+			init_hca->srqc_base    = profile[i].start;
+			init_hca->log_num_srqs = profile[i].log_num;
+			break;
+		case MTHCA_RES_CQ:
+			dev->limits.num_cqs   = profile[i].num;
+			init_hca->cqc_base    = profile[i].start;
+			init_hca->log_num_cqs = profile[i].log_num;
+			break;
+		case MTHCA_RES_EQP:
+			init_hca->eqpc_base = profile[i].start;
+			break;
+		case MTHCA_RES_EEEC:
+			init_hca->eeec_base = profile[i].start;
+			break;
+		case MTHCA_RES_EQ:
+			dev->limits.num_eqs   = profile[i].num;
+			init_hca->eqc_base    = profile[i].start;
+			init_hca->log_num_eqs = profile[i].log_num;
+			break;
+		case MTHCA_RES_RDB:
+			for (dev->qp_table.rdb_shift = 0;
+			     request->num_qp << dev->qp_table.rdb_shift < profile[i].num;
+			     ++dev->qp_table.rdb_shift)
+				; /* nothing */
+			dev->qp_table.rdb_base    = (u32) profile[i].start;
+			init_hca->rdb_base        = profile[i].start;
+			break;
+		case MTHCA_RES_MCG:
+			dev->limits.num_mgms      = profile[i].num >> 1;
+			dev->limits.num_amgms     = profile[i].num >> 1;
+			init_hca->mc_base         = profile[i].start;
+			init_hca->log_mc_entry_sz = ffs(MTHCA_MGM_ENTRY_SIZE) - 1;
+			init_hca->log_mc_table_sz = profile[i].log_num;
+			init_hca->mc_hash_sz      = 1 << (profile[i].log_num - 1);
+			break;
+		case MTHCA_RES_MPT:
+			dev->limits.num_mpts   = profile[i].num;
+			dev->mr_table.mpt_base = profile[i].start;
+			init_hca->mpt_base     = profile[i].start;
+			init_hca->log_mpt_sz   = profile[i].log_num;
+			break;
+		case MTHCA_RES_MTT:
+			dev->limits.num_mtt_segs = profile[i].num;
+			dev->mr_table.mtt_base   = profile[i].start;
+			init_hca->mtt_base       = profile[i].start;
+			init_hca->mtt_seg_sz     = ffs(dev->limits.mtt_seg_size) - 7;
+			break;
+		case MTHCA_RES_UAR:
+			dev->limits.num_uars       = profile[i].num;
+			init_hca->uar_scratch_base = profile[i].start;
+			break;
+		case MTHCA_RES_UDAV:
+			dev->av_table.ddr_av_base = profile[i].start;
+			dev->av_table.num_ddr_avs = profile[i].num;
+			break;
+		case MTHCA_RES_UARC:
+			dev->uar_table.uarc_size = request->uarc_size;
+			dev->uar_table.uarc_base = profile[i].start;
+			init_hca->uarc_base   	 = profile[i].start;
+			init_hca->log_uarc_sz 	 = ffs(request->uarc_size) - 13;
+			init_hca->log_uar_sz  	 = ffs(request->num_uar) - 1;
+			break;
+		default:
+			break;
+		}
+	}
+
+	/*
+	 * PDs don't take any HCA memory, but we assign them as part
+	 * of the HCA profile anyway.
+	 */
+	dev->limits.num_pds = MTHCA_NUM_PDS;
+
+	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT &&
+	    init_hca->log_mpt_sz > 23) {
+		mthca_warn(dev, "MPT table too large (requested size 2^%d >= 2^24)\n",
+			   init_hca->log_mpt_sz);
+		mthca_warn(dev, "Disabling memory key throughput optimization.\n");
+		dev->mthca_flags &= ~MTHCA_FLAG_SINAI_OPT;
+	}
+
+	/*
+	 * For Tavor, FMRs use ioremapped PCI memory. For 32 bit
+	 * systems it may use too much vmalloc space to map all MTT
+	 * memory, so we reserve some MTTs for FMR access, taking them
+	 * out of the MR pool. They don't use additional memory, but
+	 * we assign them as part of the HCA profile anyway.
+	 */
+	if (mthca_is_memfree(dev) || BITS_PER_LONG == 64)
+		dev->limits.fmr_reserved_mtts = 0;
+	else
+		dev->limits.fmr_reserved_mtts = request->fmr_reserved_mtts;
+
+	kfree(profile);
+	return total_size;
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_profile.h b/drivers/infiniband/hw/mthca/mthca_profile.h
new file mode 100644
index 000000000..62b009cc8
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_profile.h
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef MTHCA_PROFILE_H
+#define MTHCA_PROFILE_H
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+
+struct mthca_profile {
+	int num_qp;
+	int rdb_per_qp;
+	int num_srq;
+	int num_cq;
+	int num_mcg;
+	int num_mpt;
+	int num_mtt;
+	int num_udav;
+	int num_uar;
+	int uarc_size;
+	int fmr_reserved_mtts;
+};
+
+s64 mthca_make_profile(struct mthca_dev *mdev,
+		       struct mthca_profile *request,
+		       struct mthca_dev_lim *dev_lim,
+		       struct mthca_init_hca_param *init_hca);
+
+#endif /* MTHCA_PROFILE_H */
diff --git a/drivers/infiniband/hw/mthca/mthca_provider.c b/drivers/infiniband/hw/mthca/mthca_provider.c
new file mode 100644
index 000000000..c46df53f2
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_provider.c
@@ -0,0 +1,1178 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <rdma/ib_smi.h>
+#include <rdma/ib_umem.h>
+#include <rdma/ib_user_verbs.h>
+#include <rdma/uverbs_ioctl.h>
+
+#include <linux/sched.h>
+#include <linux/slab.h>
+#include <linux/stat.h>
+#include <linux/mm.h>
+#include <linux/export.h>
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+#include <rdma/mthca-abi.h>
+#include "mthca_memfree.h"
+
+static int mthca_query_device(struct ib_device *ibdev, struct ib_device_attr *props,
+			      struct ib_udata *uhw)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int err = -ENOMEM;
+	struct mthca_dev *mdev = to_mdev(ibdev);
+
+	if (uhw->inlen || uhw->outlen)
+		return -EINVAL;
+
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	memset(props, 0, sizeof *props);
+
+	props->fw_ver              = mdev->fw_ver;
+
+	ib_init_query_mad(in_mad);
+	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
+
+	err = mthca_MAD_IFC(mdev, 1, 1,
+			    1, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	props->device_cap_flags    = mdev->device_cap_flags;
+	props->vendor_id           = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
+		0xffffff;
+	props->vendor_part_id      = be16_to_cpup((__be16 *) (out_mad->data + 30));
+	props->hw_ver              = be32_to_cpup((__be32 *) (out_mad->data + 32));
+	memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
+
+	props->max_mr_size         = ~0ull;
+	props->page_size_cap       = mdev->limits.page_size_cap;
+	props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
+	props->max_qp_wr           = mdev->limits.max_wqes;
+	props->max_send_sge        = mdev->limits.max_sg;
+	props->max_recv_sge        = mdev->limits.max_sg;
+	props->max_sge_rd          = mdev->limits.max_sg;
+	props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
+	props->max_cqe             = mdev->limits.max_cqes;
+	props->max_mr              = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
+	props->max_pd              = mdev->limits.num_pds - mdev->limits.reserved_pds;
+	props->max_qp_rd_atom      = 1 << mdev->qp_table.rdb_shift;
+	props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
+	props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
+	props->max_srq             = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
+	props->max_srq_wr          = mdev->limits.max_srq_wqes;
+	props->max_srq_sge         = mdev->limits.max_srq_sge;
+	props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
+	props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
+					IB_ATOMIC_HCA : IB_ATOMIC_NONE;
+	props->max_pkeys           = mdev->limits.pkey_table_len;
+	props->max_mcast_grp       = mdev->limits.num_mgms + mdev->limits.num_amgms;
+	props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
+	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
+					   props->max_mcast_grp;
+
+	err = 0;
+ out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static int mthca_query_port(struct ib_device *ibdev,
+			    u32 port, struct ib_port_attr *props)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int err = -ENOMEM;
+
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	/* props being zeroed by the caller, avoid zeroing it here */
+
+	ib_init_query_mad(in_mad);
+	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
+	in_mad->attr_mod = cpu_to_be32(port);
+
+	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
+			    port, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
+	props->lmc               = out_mad->data[34] & 0x7;
+	props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
+	props->sm_sl             = out_mad->data[36] & 0xf;
+	props->state             = out_mad->data[32] & 0xf;
+	props->phys_state        = out_mad->data[33] >> 4;
+	props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
+	props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
+	props->max_msg_sz        = 0x80000000;
+	props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
+	props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
+	props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
+	props->active_width      = out_mad->data[31] & 0xf;
+	props->active_speed      = out_mad->data[35] >> 4;
+	props->max_mtu           = out_mad->data[41] & 0xf;
+	props->active_mtu        = out_mad->data[36] >> 4;
+	props->subnet_timeout    = out_mad->data[51] & 0x1f;
+	props->max_vl_num        = out_mad->data[37] >> 4;
+	props->init_type_reply   = out_mad->data[41] >> 4;
+
+ out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static int mthca_modify_device(struct ib_device *ibdev,
+			       int mask,
+			       struct ib_device_modify *props)
+{
+	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
+		return -EOPNOTSUPP;
+
+	if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
+		if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
+			return -ERESTARTSYS;
+		memcpy(ibdev->node_desc, props->node_desc,
+		       IB_DEVICE_NODE_DESC_MAX);
+		mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
+	}
+
+	return 0;
+}
+
+static int mthca_modify_port(struct ib_device *ibdev,
+			     u32 port, int port_modify_mask,
+			     struct ib_port_modify *props)
+{
+	struct mthca_set_ib_param set_ib;
+	struct ib_port_attr attr;
+	int err;
+
+	if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
+		return -ERESTARTSYS;
+
+	err = ib_query_port(ibdev, port, &attr);
+	if (err)
+		goto out;
+
+	set_ib.set_si_guid     = 0;
+	set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
+
+	set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
+		~props->clr_port_cap_mask;
+
+	err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port);
+	if (err)
+		goto out;
+out:
+	mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
+	return err;
+}
+
+static int mthca_query_pkey(struct ib_device *ibdev,
+			    u32 port, u16 index, u16 *pkey)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int err = -ENOMEM;
+
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	ib_init_query_mad(in_mad);
+	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
+	in_mad->attr_mod = cpu_to_be32(index / 32);
+
+	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
+			    port, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	*pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
+
+ out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static int mthca_query_gid(struct ib_device *ibdev, u32 port,
+			   int index, union ib_gid *gid)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int err = -ENOMEM;
+
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	ib_init_query_mad(in_mad);
+	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
+	in_mad->attr_mod = cpu_to_be32(port);
+
+	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
+			    port, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	memcpy(gid->raw, out_mad->data + 8, 8);
+
+	ib_init_query_mad(in_mad);
+	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
+	in_mad->attr_mod = cpu_to_be32(index / 8);
+
+	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
+			    port, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
+
+ out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static int mthca_alloc_ucontext(struct ib_ucontext *uctx,
+				struct ib_udata *udata)
+{
+	struct ib_device *ibdev = uctx->device;
+	struct mthca_alloc_ucontext_resp uresp = {};
+	struct mthca_ucontext *context = to_mucontext(uctx);
+	int                              err;
+
+	if (!(to_mdev(ibdev)->active))
+		return -EAGAIN;
+
+	uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
+	if (mthca_is_memfree(to_mdev(ibdev)))
+		uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
+	else
+		uresp.uarc_size = 0;
+
+	err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
+	if (err)
+		return err;
+
+	context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
+	if (IS_ERR(context->db_tab)) {
+		err = PTR_ERR(context->db_tab);
+		mthca_uar_free(to_mdev(ibdev), &context->uar);
+		return err;
+	}
+
+	if (ib_copy_to_udata(udata, &uresp, sizeof(uresp))) {
+		mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
+		mthca_uar_free(to_mdev(ibdev), &context->uar);
+		return -EFAULT;
+	}
+
+	context->reg_mr_warned = 0;
+
+	return 0;
+}
+
+static void mthca_dealloc_ucontext(struct ib_ucontext *context)
+{
+	mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
+				  to_mucontext(context)->db_tab);
+	mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
+}
+
+static int mthca_mmap_uar(struct ib_ucontext *context,
+			  struct vm_area_struct *vma)
+{
+	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
+		return -EINVAL;
+
+	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+
+	if (io_remap_pfn_range(vma, vma->vm_start,
+			       to_mucontext(context)->uar.pfn,
+			       PAGE_SIZE, vma->vm_page_prot))
+		return -EAGAIN;
+
+	return 0;
+}
+
+static int mthca_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
+{
+	struct ib_device *ibdev = ibpd->device;
+	struct mthca_pd *pd = to_mpd(ibpd);
+	int err;
+
+	err = mthca_pd_alloc(to_mdev(ibdev), !udata, pd);
+	if (err)
+		return err;
+
+	if (udata) {
+		if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
+			mthca_pd_free(to_mdev(ibdev), pd);
+			return -EFAULT;
+		}
+	}
+
+	return 0;
+}
+
+static int mthca_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
+{
+	mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
+	return 0;
+}
+
+static int mthca_ah_create(struct ib_ah *ibah,
+			   struct rdma_ah_init_attr *init_attr,
+			   struct ib_udata *udata)
+
+{
+	struct mthca_ah *ah = to_mah(ibah);
+
+	return mthca_create_ah(to_mdev(ibah->device), to_mpd(ibah->pd),
+			       init_attr->ah_attr, ah);
+}
+
+static int mthca_ah_destroy(struct ib_ah *ah, u32 flags)
+{
+	mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
+	return 0;
+}
+
+static int mthca_create_srq(struct ib_srq *ibsrq,
+			    struct ib_srq_init_attr *init_attr,
+			    struct ib_udata *udata)
+{
+	struct mthca_create_srq ucmd;
+	struct mthca_ucontext *context = rdma_udata_to_drv_context(
+		udata, struct mthca_ucontext, ibucontext);
+	struct mthca_srq *srq = to_msrq(ibsrq);
+	int err;
+
+	if (init_attr->srq_type != IB_SRQT_BASIC)
+		return -EOPNOTSUPP;
+
+	if (udata) {
+		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
+			return -EFAULT;
+
+		err = mthca_map_user_db(to_mdev(ibsrq->device), &context->uar,
+					context->db_tab, ucmd.db_index,
+					ucmd.db_page);
+
+		if (err)
+			return err;
+
+		srq->mr.ibmr.lkey = ucmd.lkey;
+		srq->db_index     = ucmd.db_index;
+	}
+
+	err = mthca_alloc_srq(to_mdev(ibsrq->device), to_mpd(ibsrq->pd),
+			      &init_attr->attr, srq, udata);
+
+	if (err && udata)
+		mthca_unmap_user_db(to_mdev(ibsrq->device), &context->uar,
+				    context->db_tab, ucmd.db_index);
+
+	if (err)
+		return err;
+
+	if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof(__u32))) {
+		mthca_free_srq(to_mdev(ibsrq->device), srq);
+		return -EFAULT;
+	}
+
+	return 0;
+}
+
+static int mthca_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
+{
+	if (udata) {
+		struct mthca_ucontext *context =
+			rdma_udata_to_drv_context(
+				udata,
+				struct mthca_ucontext,
+				ibucontext);
+
+		mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
+				    context->db_tab, to_msrq(srq)->db_index);
+	}
+
+	mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
+	return 0;
+}
+
+static int mthca_create_qp(struct ib_qp *ibqp,
+			   struct ib_qp_init_attr *init_attr,
+			   struct ib_udata *udata)
+{
+	struct mthca_ucontext *context = rdma_udata_to_drv_context(
+		udata, struct mthca_ucontext, ibucontext);
+	struct mthca_create_qp ucmd;
+	struct mthca_qp *qp = to_mqp(ibqp);
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	int err;
+
+	if (init_attr->create_flags)
+		return -EOPNOTSUPP;
+
+	switch (init_attr->qp_type) {
+	case IB_QPT_RC:
+	case IB_QPT_UC:
+	case IB_QPT_UD:
+	{
+		if (udata) {
+			if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
+				return -EFAULT;
+
+			err = mthca_map_user_db(dev, &context->uar,
+						context->db_tab,
+						ucmd.sq_db_index,
+						ucmd.sq_db_page);
+			if (err)
+				return err;
+
+			err = mthca_map_user_db(dev, &context->uar,
+						context->db_tab,
+						ucmd.rq_db_index,
+						ucmd.rq_db_page);
+			if (err) {
+				mthca_unmap_user_db(dev, &context->uar,
+						    context->db_tab,
+						    ucmd.sq_db_index);
+				return err;
+			}
+
+			qp->mr.ibmr.lkey = ucmd.lkey;
+			qp->sq.db_index  = ucmd.sq_db_index;
+			qp->rq.db_index  = ucmd.rq_db_index;
+		}
+
+		err = mthca_alloc_qp(dev, to_mpd(ibqp->pd),
+				     to_mcq(init_attr->send_cq),
+				     to_mcq(init_attr->recv_cq),
+				     init_attr->qp_type, init_attr->sq_sig_type,
+				     &init_attr->cap, qp, udata);
+
+		if (err && udata) {
+			mthca_unmap_user_db(dev, &context->uar, context->db_tab,
+					    ucmd.sq_db_index);
+			mthca_unmap_user_db(dev, &context->uar, context->db_tab,
+					    ucmd.rq_db_index);
+		}
+
+		qp->ibqp.qp_num = qp->qpn;
+		break;
+	}
+	case IB_QPT_SMI:
+	case IB_QPT_GSI:
+	{
+		qp->sqp = kzalloc(sizeof(struct mthca_sqp), GFP_KERNEL);
+		if (!qp->sqp)
+			return -ENOMEM;
+
+		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
+
+		err = mthca_alloc_sqp(dev, to_mpd(ibqp->pd),
+				      to_mcq(init_attr->send_cq),
+				      to_mcq(init_attr->recv_cq),
+				      init_attr->sq_sig_type, &init_attr->cap,
+				      qp->ibqp.qp_num, init_attr->port_num, qp,
+				      udata);
+		break;
+	}
+	default:
+		/* Don't support raw QPs */
+		return -EOPNOTSUPP;
+	}
+
+	if (err) {
+		kfree(qp->sqp);
+		return err;
+	}
+
+	init_attr->cap.max_send_wr     = qp->sq.max;
+	init_attr->cap.max_recv_wr     = qp->rq.max;
+	init_attr->cap.max_send_sge    = qp->sq.max_gs;
+	init_attr->cap.max_recv_sge    = qp->rq.max_gs;
+	init_attr->cap.max_inline_data = qp->max_inline_data;
+
+	return 0;
+}
+
+static int mthca_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
+{
+	if (udata) {
+		struct mthca_ucontext *context =
+			rdma_udata_to_drv_context(
+				udata,
+				struct mthca_ucontext,
+				ibucontext);
+
+		mthca_unmap_user_db(to_mdev(qp->device),
+				    &context->uar,
+				    context->db_tab,
+				    to_mqp(qp)->sq.db_index);
+		mthca_unmap_user_db(to_mdev(qp->device),
+				    &context->uar,
+				    context->db_tab,
+				    to_mqp(qp)->rq.db_index);
+	}
+	mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
+	kfree(to_mqp(qp)->sqp);
+	return 0;
+}
+
+static int mthca_create_cq(struct ib_cq *ibcq,
+			   const struct ib_cq_init_attr *attr,
+			   struct ib_udata *udata)
+{
+	struct ib_device *ibdev = ibcq->device;
+	int entries = attr->cqe;
+	struct mthca_create_cq ucmd;
+	struct mthca_cq *cq;
+	int nent;
+	int err;
+	struct mthca_ucontext *context = rdma_udata_to_drv_context(
+		udata, struct mthca_ucontext, ibucontext);
+
+	if (attr->flags)
+		return -EOPNOTSUPP;
+
+	if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
+		return -EINVAL;
+
+	if (udata) {
+		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
+			return -EFAULT;
+
+		err = mthca_map_user_db(to_mdev(ibdev), &context->uar,
+					context->db_tab, ucmd.set_db_index,
+					ucmd.set_db_page);
+		if (err)
+			return err;
+
+		err = mthca_map_user_db(to_mdev(ibdev), &context->uar,
+					context->db_tab, ucmd.arm_db_index,
+					ucmd.arm_db_page);
+		if (err)
+			goto err_unmap_set;
+	}
+
+	cq = to_mcq(ibcq);
+
+	if (udata) {
+		cq->buf.mr.ibmr.lkey = ucmd.lkey;
+		cq->set_ci_db_index  = ucmd.set_db_index;
+		cq->arm_db_index     = ucmd.arm_db_index;
+	}
+
+	for (nent = 1; nent <= entries; nent <<= 1)
+		; /* nothing */
+
+	err = mthca_init_cq(to_mdev(ibdev), nent, context,
+			    udata ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
+			    cq);
+	if (err)
+		goto err_unmap_arm;
+
+	if (udata && ib_copy_to_udata(udata, &cq->cqn, sizeof(__u32))) {
+		mthca_free_cq(to_mdev(ibdev), cq);
+		err = -EFAULT;
+		goto err_unmap_arm;
+	}
+
+	cq->resize_buf = NULL;
+
+	return 0;
+
+err_unmap_arm:
+	if (udata)
+		mthca_unmap_user_db(to_mdev(ibdev), &context->uar,
+				    context->db_tab, ucmd.arm_db_index);
+
+err_unmap_set:
+	if (udata)
+		mthca_unmap_user_db(to_mdev(ibdev), &context->uar,
+				    context->db_tab, ucmd.set_db_index);
+
+	return err;
+}
+
+static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
+				  int entries)
+{
+	int ret;
+
+	spin_lock_irq(&cq->lock);
+	if (cq->resize_buf) {
+		ret = -EBUSY;
+		goto unlock;
+	}
+
+	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
+	if (!cq->resize_buf) {
+		ret = -ENOMEM;
+		goto unlock;
+	}
+
+	cq->resize_buf->state = CQ_RESIZE_ALLOC;
+
+	ret = 0;
+
+unlock:
+	spin_unlock_irq(&cq->lock);
+
+	if (ret)
+		return ret;
+
+	ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
+	if (ret) {
+		spin_lock_irq(&cq->lock);
+		kfree(cq->resize_buf);
+		cq->resize_buf = NULL;
+		spin_unlock_irq(&cq->lock);
+		return ret;
+	}
+
+	cq->resize_buf->cqe = entries - 1;
+
+	spin_lock_irq(&cq->lock);
+	cq->resize_buf->state = CQ_RESIZE_READY;
+	spin_unlock_irq(&cq->lock);
+
+	return 0;
+}
+
+static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
+{
+	struct mthca_dev *dev = to_mdev(ibcq->device);
+	struct mthca_cq *cq = to_mcq(ibcq);
+	struct mthca_resize_cq ucmd;
+	u32 lkey;
+	int ret;
+
+	if (entries < 1 || entries > dev->limits.max_cqes)
+		return -EINVAL;
+
+	mutex_lock(&cq->mutex);
+
+	entries = roundup_pow_of_two(entries + 1);
+	if (entries == ibcq->cqe + 1) {
+		ret = 0;
+		goto out;
+	}
+
+	if (cq->is_kernel) {
+		ret = mthca_alloc_resize_buf(dev, cq, entries);
+		if (ret)
+			goto out;
+		lkey = cq->resize_buf->buf.mr.ibmr.lkey;
+	} else {
+		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
+			ret = -EFAULT;
+			goto out;
+		}
+		lkey = ucmd.lkey;
+	}
+
+	ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries));
+
+	if (ret) {
+		if (cq->resize_buf) {
+			mthca_free_cq_buf(dev, &cq->resize_buf->buf,
+					  cq->resize_buf->cqe);
+			kfree(cq->resize_buf);
+			spin_lock_irq(&cq->lock);
+			cq->resize_buf = NULL;
+			spin_unlock_irq(&cq->lock);
+		}
+		goto out;
+	}
+
+	if (cq->is_kernel) {
+		struct mthca_cq_buf tbuf;
+		int tcqe;
+
+		spin_lock_irq(&cq->lock);
+		if (cq->resize_buf->state == CQ_RESIZE_READY) {
+			mthca_cq_resize_copy_cqes(cq);
+			tbuf         = cq->buf;
+			tcqe         = cq->ibcq.cqe;
+			cq->buf      = cq->resize_buf->buf;
+			cq->ibcq.cqe = cq->resize_buf->cqe;
+		} else {
+			tbuf = cq->resize_buf->buf;
+			tcqe = cq->resize_buf->cqe;
+		}
+
+		kfree(cq->resize_buf);
+		cq->resize_buf = NULL;
+		spin_unlock_irq(&cq->lock);
+
+		mthca_free_cq_buf(dev, &tbuf, tcqe);
+	} else
+		ibcq->cqe = entries - 1;
+
+out:
+	mutex_unlock(&cq->mutex);
+
+	return ret;
+}
+
+static int mthca_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
+{
+	if (udata) {
+		struct mthca_ucontext *context =
+			rdma_udata_to_drv_context(
+				udata,
+				struct mthca_ucontext,
+				ibucontext);
+
+		mthca_unmap_user_db(to_mdev(cq->device),
+				    &context->uar,
+				    context->db_tab,
+				    to_mcq(cq)->arm_db_index);
+		mthca_unmap_user_db(to_mdev(cq->device),
+				    &context->uar,
+				    context->db_tab,
+				    to_mcq(cq)->set_ci_db_index);
+	}
+	mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
+	return 0;
+}
+
+static inline u32 convert_access(int acc)
+{
+	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
+	       (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
+	       (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
+	       (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
+	       MTHCA_MPT_FLAG_LOCAL_READ;
+}
+
+static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
+{
+	struct mthca_mr *mr;
+	int err;
+
+	mr = kmalloc(sizeof *mr, GFP_KERNEL);
+	if (!mr)
+		return ERR_PTR(-ENOMEM);
+
+	err = mthca_mr_alloc_notrans(to_mdev(pd->device),
+				     to_mpd(pd)->pd_num,
+				     convert_access(acc), mr);
+
+	if (err) {
+		kfree(mr);
+		return ERR_PTR(err);
+	}
+
+	mr->umem = NULL;
+
+	return &mr->ibmr;
+}
+
+static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
+				       u64 virt, int acc, struct ib_udata *udata)
+{
+	struct mthca_dev *dev = to_mdev(pd->device);
+	struct ib_block_iter biter;
+	struct mthca_ucontext *context = rdma_udata_to_drv_context(
+		udata, struct mthca_ucontext, ibucontext);
+	struct mthca_mr *mr;
+	struct mthca_reg_mr ucmd;
+	u64 *pages;
+	int n, i;
+	int err = 0;
+	int write_mtt_size;
+
+	if (udata->inlen < sizeof ucmd) {
+		if (!context->reg_mr_warned) {
+			mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n",
+				   current->comm);
+			mthca_warn(dev, "  Update libmthca to fix this.\n");
+		}
+		++context->reg_mr_warned;
+		ucmd.mr_attrs = 0;
+	} else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
+		return ERR_PTR(-EFAULT);
+
+	mr = kmalloc(sizeof *mr, GFP_KERNEL);
+	if (!mr)
+		return ERR_PTR(-ENOMEM);
+
+	mr->umem = ib_umem_get(pd->device, start, length, acc);
+	if (IS_ERR(mr->umem)) {
+		err = PTR_ERR(mr->umem);
+		goto err;
+	}
+
+	n = ib_umem_num_dma_blocks(mr->umem, PAGE_SIZE);
+
+	mr->mtt = mthca_alloc_mtt(dev, n);
+	if (IS_ERR(mr->mtt)) {
+		err = PTR_ERR(mr->mtt);
+		goto err_umem;
+	}
+
+	pages = (u64 *) __get_free_page(GFP_KERNEL);
+	if (!pages) {
+		err = -ENOMEM;
+		goto err_mtt;
+	}
+
+	i = n = 0;
+
+	write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
+
+	rdma_umem_for_each_dma_block(mr->umem, &biter, PAGE_SIZE) {
+		pages[i++] = rdma_block_iter_dma_address(&biter);
+
+		/*
+		 * Be friendly to write_mtt and pass it chunks
+		 * of appropriate size.
+		 */
+		if (i == write_mtt_size) {
+			err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
+			if (err)
+				goto mtt_done;
+			n += i;
+			i = 0;
+		}
+	}
+
+	if (i)
+		err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
+mtt_done:
+	free_page((unsigned long) pages);
+	if (err)
+		goto err_mtt;
+
+	err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, PAGE_SHIFT, virt, length,
+			     convert_access(acc), mr);
+
+	if (err)
+		goto err_mtt;
+
+	return &mr->ibmr;
+
+err_mtt:
+	mthca_free_mtt(dev, mr->mtt);
+
+err_umem:
+	ib_umem_release(mr->umem);
+
+err:
+	kfree(mr);
+	return ERR_PTR(err);
+}
+
+static int mthca_dereg_mr(struct ib_mr *mr, struct ib_udata *udata)
+{
+	struct mthca_mr *mmr = to_mmr(mr);
+
+	mthca_free_mr(to_mdev(mr->device), mmr);
+	ib_umem_release(mmr->umem);
+	kfree(mmr);
+
+	return 0;
+}
+
+static ssize_t hw_rev_show(struct device *device,
+			   struct device_attribute *attr, char *buf)
+{
+	struct mthca_dev *dev =
+		rdma_device_to_drv_device(device, struct mthca_dev, ib_dev);
+
+	return sysfs_emit(buf, "%x\n", dev->rev_id);
+}
+static DEVICE_ATTR_RO(hw_rev);
+
+static const char *hca_type_string(int hca_type)
+{
+	switch (hca_type) {
+	case PCI_DEVICE_ID_MELLANOX_TAVOR:
+		return "MT23108";
+	case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
+		return "MT25208 (MT23108 compat mode)";
+	case PCI_DEVICE_ID_MELLANOX_ARBEL:
+		return "MT25208";
+	case PCI_DEVICE_ID_MELLANOX_SINAI:
+	case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
+		return "MT25204";
+	}
+
+	return "unknown";
+}
+
+static ssize_t hca_type_show(struct device *device,
+			     struct device_attribute *attr, char *buf)
+{
+	struct mthca_dev *dev =
+		rdma_device_to_drv_device(device, struct mthca_dev, ib_dev);
+
+	return sysfs_emit(buf, "%s\n", hca_type_string(dev->pdev->device));
+}
+static DEVICE_ATTR_RO(hca_type);
+
+static ssize_t board_id_show(struct device *device,
+			     struct device_attribute *attr, char *buf)
+{
+	struct mthca_dev *dev =
+		rdma_device_to_drv_device(device, struct mthca_dev, ib_dev);
+
+	return sysfs_emit(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
+}
+static DEVICE_ATTR_RO(board_id);
+
+static struct attribute *mthca_dev_attributes[] = {
+	&dev_attr_hw_rev.attr,
+	&dev_attr_hca_type.attr,
+	&dev_attr_board_id.attr,
+	NULL
+};
+
+static const struct attribute_group mthca_attr_group = {
+	.attrs = mthca_dev_attributes,
+};
+
+static int mthca_init_node_data(struct mthca_dev *dev)
+{
+	struct ib_smp *in_mad  = NULL;
+	struct ib_smp *out_mad = NULL;
+	int err = -ENOMEM;
+
+	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
+	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
+	if (!in_mad || !out_mad)
+		goto out;
+
+	ib_init_query_mad(in_mad);
+	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
+
+	err = mthca_MAD_IFC(dev, 1, 1,
+			    1, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	memcpy(dev->ib_dev.node_desc, out_mad->data, IB_DEVICE_NODE_DESC_MAX);
+
+	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
+
+	err = mthca_MAD_IFC(dev, 1, 1,
+			    1, NULL, NULL, in_mad, out_mad);
+	if (err)
+		goto out;
+
+	if (mthca_is_memfree(dev))
+		dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
+	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
+
+out:
+	kfree(in_mad);
+	kfree(out_mad);
+	return err;
+}
+
+static int mthca_port_immutable(struct ib_device *ibdev, u32 port_num,
+			        struct ib_port_immutable *immutable)
+{
+	struct ib_port_attr attr;
+	int err;
+
+	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB;
+
+	err = ib_query_port(ibdev, port_num, &attr);
+	if (err)
+		return err;
+
+	immutable->pkey_tbl_len = attr.pkey_tbl_len;
+	immutable->gid_tbl_len = attr.gid_tbl_len;
+	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
+
+	return 0;
+}
+
+static void get_dev_fw_str(struct ib_device *device, char *str)
+{
+	struct mthca_dev *dev =
+		container_of(device, struct mthca_dev, ib_dev);
+	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d",
+		 (int) (dev->fw_ver >> 32),
+		 (int) (dev->fw_ver >> 16) & 0xffff,
+		 (int) dev->fw_ver & 0xffff);
+}
+
+static const struct ib_device_ops mthca_dev_ops = {
+	.owner = THIS_MODULE,
+	.driver_id = RDMA_DRIVER_MTHCA,
+	.uverbs_abi_ver = MTHCA_UVERBS_ABI_VERSION,
+	.uverbs_no_driver_id_binding = 1,
+
+	.alloc_pd = mthca_alloc_pd,
+	.alloc_ucontext = mthca_alloc_ucontext,
+	.attach_mcast = mthca_multicast_attach,
+	.create_ah = mthca_ah_create,
+	.create_cq = mthca_create_cq,
+	.create_qp = mthca_create_qp,
+	.dealloc_pd = mthca_dealloc_pd,
+	.dealloc_ucontext = mthca_dealloc_ucontext,
+	.dereg_mr = mthca_dereg_mr,
+	.destroy_ah = mthca_ah_destroy,
+	.destroy_cq = mthca_destroy_cq,
+	.destroy_qp = mthca_destroy_qp,
+	.detach_mcast = mthca_multicast_detach,
+	.device_group = &mthca_attr_group,
+	.get_dev_fw_str = get_dev_fw_str,
+	.get_dma_mr = mthca_get_dma_mr,
+	.get_port_immutable = mthca_port_immutable,
+	.mmap = mthca_mmap_uar,
+	.modify_device = mthca_modify_device,
+	.modify_port = mthca_modify_port,
+	.modify_qp = mthca_modify_qp,
+	.poll_cq = mthca_poll_cq,
+	.process_mad = mthca_process_mad,
+	.query_ah = mthca_ah_query,
+	.query_device = mthca_query_device,
+	.query_gid = mthca_query_gid,
+	.query_pkey = mthca_query_pkey,
+	.query_port = mthca_query_port,
+	.query_qp = mthca_query_qp,
+	.reg_user_mr = mthca_reg_user_mr,
+	.resize_cq = mthca_resize_cq,
+
+	INIT_RDMA_OBJ_SIZE(ib_ah, mthca_ah, ibah),
+	INIT_RDMA_OBJ_SIZE(ib_cq, mthca_cq, ibcq),
+	INIT_RDMA_OBJ_SIZE(ib_pd, mthca_pd, ibpd),
+	INIT_RDMA_OBJ_SIZE(ib_qp, mthca_qp, ibqp),
+	INIT_RDMA_OBJ_SIZE(ib_ucontext, mthca_ucontext, ibucontext),
+};
+
+static const struct ib_device_ops mthca_dev_arbel_srq_ops = {
+	.create_srq = mthca_create_srq,
+	.destroy_srq = mthca_destroy_srq,
+	.modify_srq = mthca_modify_srq,
+	.post_srq_recv = mthca_arbel_post_srq_recv,
+	.query_srq = mthca_query_srq,
+
+	INIT_RDMA_OBJ_SIZE(ib_srq, mthca_srq, ibsrq),
+};
+
+static const struct ib_device_ops mthca_dev_tavor_srq_ops = {
+	.create_srq = mthca_create_srq,
+	.destroy_srq = mthca_destroy_srq,
+	.modify_srq = mthca_modify_srq,
+	.post_srq_recv = mthca_tavor_post_srq_recv,
+	.query_srq = mthca_query_srq,
+
+	INIT_RDMA_OBJ_SIZE(ib_srq, mthca_srq, ibsrq),
+};
+
+static const struct ib_device_ops mthca_dev_arbel_ops = {
+	.post_recv = mthca_arbel_post_receive,
+	.post_send = mthca_arbel_post_send,
+	.req_notify_cq = mthca_arbel_arm_cq,
+};
+
+static const struct ib_device_ops mthca_dev_tavor_ops = {
+	.post_recv = mthca_tavor_post_receive,
+	.post_send = mthca_tavor_post_send,
+	.req_notify_cq = mthca_tavor_arm_cq,
+};
+
+int mthca_register_device(struct mthca_dev *dev)
+{
+	int ret;
+
+	ret = mthca_init_node_data(dev);
+	if (ret)
+		return ret;
+
+	dev->ib_dev.node_type            = RDMA_NODE_IB_CA;
+	dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
+	dev->ib_dev.num_comp_vectors     = 1;
+	dev->ib_dev.dev.parent           = &dev->pdev->dev;
+
+	if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
+		if (mthca_is_memfree(dev))
+			ib_set_device_ops(&dev->ib_dev,
+					  &mthca_dev_arbel_srq_ops);
+		else
+			ib_set_device_ops(&dev->ib_dev,
+					  &mthca_dev_tavor_srq_ops);
+	}
+
+	ib_set_device_ops(&dev->ib_dev, &mthca_dev_ops);
+
+	if (mthca_is_memfree(dev))
+		ib_set_device_ops(&dev->ib_dev, &mthca_dev_arbel_ops);
+	else
+		ib_set_device_ops(&dev->ib_dev, &mthca_dev_tavor_ops);
+
+	mutex_init(&dev->cap_mask_mutex);
+
+	ret = ib_register_device(&dev->ib_dev, "mthca%d", &dev->pdev->dev);
+	if (ret)
+		return ret;
+
+	mthca_start_catas_poll(dev);
+
+	return 0;
+}
+
+void mthca_unregister_device(struct mthca_dev *dev)
+{
+	mthca_stop_catas_poll(dev);
+	ib_unregister_device(&dev->ib_dev);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_provider.h b/drivers/infiniband/hw/mthca/mthca_provider.h
new file mode 100644
index 000000000..8a77483bb
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_provider.h
@@ -0,0 +1,316 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef MTHCA_PROVIDER_H
+#define MTHCA_PROVIDER_H
+
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_pack.h>
+
+#define MTHCA_MPT_FLAG_ATOMIC        (1 << 14)
+#define MTHCA_MPT_FLAG_REMOTE_WRITE  (1 << 13)
+#define MTHCA_MPT_FLAG_REMOTE_READ   (1 << 12)
+#define MTHCA_MPT_FLAG_LOCAL_WRITE   (1 << 11)
+#define MTHCA_MPT_FLAG_LOCAL_READ    (1 << 10)
+
+struct mthca_buf_list {
+	void *buf;
+	DEFINE_DMA_UNMAP_ADDR(mapping);
+};
+
+union mthca_buf {
+	struct mthca_buf_list direct;
+	struct mthca_buf_list *page_list;
+};
+
+struct mthca_uar {
+	unsigned long pfn;
+	int           index;
+};
+
+struct mthca_user_db_table;
+
+struct mthca_ucontext {
+	struct ib_ucontext          ibucontext;
+	struct mthca_uar            uar;
+	struct mthca_user_db_table *db_tab;
+	int			    reg_mr_warned;
+};
+
+struct mthca_mtt;
+
+struct mthca_mr {
+	struct ib_mr      ibmr;
+	struct ib_umem   *umem;
+	struct mthca_mtt *mtt;
+};
+
+struct mthca_pd {
+	struct ib_pd    ibpd;
+	u32             pd_num;
+	atomic_t        sqp_count;
+	struct mthca_mr ntmr;
+	int             privileged;
+};
+
+struct mthca_eq {
+	struct mthca_dev      *dev;
+	int                    eqn;
+	u32                    eqn_mask;
+	u32                    cons_index;
+	u16                    msi_x_vector;
+	u16                    msi_x_entry;
+	int                    have_irq;
+	int                    nent;
+	struct mthca_buf_list *page_list;
+	struct mthca_mr        mr;
+	char		       irq_name[IB_DEVICE_NAME_MAX];
+};
+
+struct mthca_av;
+
+enum mthca_ah_type {
+	MTHCA_AH_ON_HCA,
+	MTHCA_AH_PCI_POOL,
+	MTHCA_AH_KMALLOC
+};
+
+struct mthca_ah {
+	struct ib_ah       ibah;
+	enum mthca_ah_type type;
+	u32                key;
+	struct mthca_av   *av;
+	dma_addr_t         avdma;
+};
+
+/*
+ * Quick description of our CQ/QP locking scheme:
+ *
+ * We have one global lock that protects dev->cq/qp_table.  Each
+ * struct mthca_cq/qp also has its own lock.  An individual qp lock
+ * may be taken inside of an individual cq lock.  Both cqs attached to
+ * a qp may be locked, with the cq with the lower cqn locked first.
+ * No other nesting should be done.
+ *
+ * Each struct mthca_cq/qp also has an ref count, protected by the
+ * corresponding table lock.  The pointer from the cq/qp_table to the
+ * struct counts as one reference.  This reference also is good for
+ * access through the consumer API, so modifying the CQ/QP etc doesn't
+ * need to take another reference.  Access to a QP because of a
+ * completion being polled does not need a reference either.
+ *
+ * Finally, each struct mthca_cq/qp has a wait_queue_head_t for the
+ * destroy function to sleep on.
+ *
+ * This means that access from the consumer API requires nothing but
+ * taking the struct's lock.
+ *
+ * Access because of a completion event should go as follows:
+ * - lock cq/qp_table and look up struct
+ * - increment ref count in struct
+ * - drop cq/qp_table lock
+ * - lock struct, do your thing, and unlock struct
+ * - decrement ref count; if zero, wake up waiters
+ *
+ * To destroy a CQ/QP, we can do the following:
+ * - lock cq/qp_table
+ * - remove pointer and decrement ref count
+ * - unlock cq/qp_table lock
+ * - wait_event until ref count is zero
+ *
+ * It is the consumer's responsibilty to make sure that no QP
+ * operations (WQE posting or state modification) are pending when a
+ * QP is destroyed.  Also, the consumer must make sure that calls to
+ * qp_modify are serialized.  Similarly, the consumer is responsible
+ * for ensuring that no CQ resize operations are pending when a CQ
+ * is destroyed.
+ *
+ * Possible optimizations (wait for profile data to see if/where we
+ * have locks bouncing between CPUs):
+ * - split cq/qp table lock into n separate (cache-aligned) locks,
+ *   indexed (say) by the page in the table
+ * - split QP struct lock into three (one for common info, one for the
+ *   send queue and one for the receive queue)
+ */
+
+struct mthca_cq_buf {
+	union mthca_buf		queue;
+	struct mthca_mr		mr;
+	int			is_direct;
+};
+
+struct mthca_cq_resize {
+	struct mthca_cq_buf	buf;
+	int			cqe;
+	enum {
+		CQ_RESIZE_ALLOC,
+		CQ_RESIZE_READY,
+		CQ_RESIZE_SWAPPED
+	}			state;
+};
+
+struct mthca_cq {
+	struct ib_cq		ibcq;
+	spinlock_t		lock;
+	int			refcount;
+	int			cqn;
+	u32			cons_index;
+	struct mthca_cq_buf	buf;
+	struct mthca_cq_resize *resize_buf;
+	int			is_kernel;
+
+	/* Next fields are Arbel only */
+	int			set_ci_db_index;
+	__be32		       *set_ci_db;
+	int			arm_db_index;
+	__be32		       *arm_db;
+	int			arm_sn;
+
+	wait_queue_head_t	wait;
+	struct mutex		mutex;
+};
+
+struct mthca_srq {
+	struct ib_srq		ibsrq;
+	spinlock_t		lock;
+	int			refcount;
+	int			srqn;
+	int			max;
+	int			max_gs;
+	int			wqe_shift;
+	int			first_free;
+	int			last_free;
+	u16			counter;  /* Arbel only */
+	int			db_index; /* Arbel only */
+	__be32		       *db;       /* Arbel only */
+	void		       *last;
+
+	int			is_direct;
+	u64		       *wrid;
+	union mthca_buf		queue;
+	struct mthca_mr		mr;
+
+	wait_queue_head_t	wait;
+	struct mutex		mutex;
+};
+
+struct mthca_wq {
+	spinlock_t lock;
+	int        max;
+	unsigned   next_ind;
+	unsigned   last_comp;
+	unsigned   head;
+	unsigned   tail;
+	void      *last;
+	int        max_gs;
+	int        wqe_shift;
+
+	int        db_index;	/* Arbel only */
+	__be32    *db;
+};
+
+struct mthca_sqp {
+	int             pkey_index;
+	u32             qkey;
+	u32             send_psn;
+	struct ib_ud_header ud_header;
+	int             header_buf_size;
+	void           *header_buf;
+	dma_addr_t      header_dma;
+};
+
+struct mthca_qp {
+	struct ib_qp           ibqp;
+	int                    refcount;
+	u32                    qpn;
+	int                    is_direct;
+	u8                     port; /* for SQP and memfree use only */
+	u8                     alt_port; /* for memfree use only */
+	u8                     transport;
+	u8                     state;
+	u8                     atomic_rd_en;
+	u8                     resp_depth;
+
+	struct mthca_mr        mr;
+
+	struct mthca_wq        rq;
+	struct mthca_wq        sq;
+	enum ib_sig_type       sq_policy;
+	int                    send_wqe_offset;
+	int                    max_inline_data;
+
+	u64                   *wrid;
+	union mthca_buf	       queue;
+
+	wait_queue_head_t      wait;
+	struct mutex	       mutex;
+	struct mthca_sqp *sqp;
+};
+
+static inline struct mthca_ucontext *to_mucontext(struct ib_ucontext *ibucontext)
+{
+	return container_of(ibucontext, struct mthca_ucontext, ibucontext);
+}
+
+static inline struct mthca_mr *to_mmr(struct ib_mr *ibmr)
+{
+	return container_of(ibmr, struct mthca_mr, ibmr);
+}
+
+static inline struct mthca_pd *to_mpd(struct ib_pd *ibpd)
+{
+	return container_of(ibpd, struct mthca_pd, ibpd);
+}
+
+static inline struct mthca_ah *to_mah(struct ib_ah *ibah)
+{
+	return container_of(ibah, struct mthca_ah, ibah);
+}
+
+static inline struct mthca_cq *to_mcq(struct ib_cq *ibcq)
+{
+	return container_of(ibcq, struct mthca_cq, ibcq);
+}
+
+static inline struct mthca_srq *to_msrq(struct ib_srq *ibsrq)
+{
+	return container_of(ibsrq, struct mthca_srq, ibsrq);
+}
+
+static inline struct mthca_qp *to_mqp(struct ib_qp *ibqp)
+{
+	return container_of(ibqp, struct mthca_qp, ibqp);
+}
+
+#endif /* MTHCA_PROVIDER_H */
diff --git a/drivers/infiniband/hw/mthca/mthca_qp.c b/drivers/infiniband/hw/mthca/mthca_qp.c
new file mode 100644
index 000000000..53f43649f
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_qp.c
@@ -0,0 +1,2317 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems. All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/string.h>
+#include <linux/slab.h>
+#include <linux/sched.h>
+
+#include <asm/io.h>
+
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_cache.h>
+#include <rdma/ib_pack.h>
+#include <rdma/uverbs_ioctl.h>
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+#include "mthca_memfree.h"
+#include "mthca_wqe.h"
+
+enum {
+	MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
+	MTHCA_ACK_REQ_FREQ       = 10,
+	MTHCA_FLIGHT_LIMIT       = 9,
+	MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
+	MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
+	MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
+};
+
+enum {
+	MTHCA_QP_STATE_RST  = 0,
+	MTHCA_QP_STATE_INIT = 1,
+	MTHCA_QP_STATE_RTR  = 2,
+	MTHCA_QP_STATE_RTS  = 3,
+	MTHCA_QP_STATE_SQE  = 4,
+	MTHCA_QP_STATE_SQD  = 5,
+	MTHCA_QP_STATE_ERR  = 6,
+	MTHCA_QP_STATE_DRAINING = 7
+};
+
+enum {
+	MTHCA_QP_ST_RC 	= 0x0,
+	MTHCA_QP_ST_UC 	= 0x1,
+	MTHCA_QP_ST_RD 	= 0x2,
+	MTHCA_QP_ST_UD 	= 0x3,
+	MTHCA_QP_ST_MLX = 0x7
+};
+
+enum {
+	MTHCA_QP_PM_MIGRATED = 0x3,
+	MTHCA_QP_PM_ARMED    = 0x0,
+	MTHCA_QP_PM_REARM    = 0x1
+};
+
+enum {
+	/* qp_context flags */
+	MTHCA_QP_BIT_DE  = 1 <<  8,
+	/* params1 */
+	MTHCA_QP_BIT_SRE = 1 << 15,
+	MTHCA_QP_BIT_SWE = 1 << 14,
+	MTHCA_QP_BIT_SAE = 1 << 13,
+	MTHCA_QP_BIT_SIC = 1 <<  4,
+	MTHCA_QP_BIT_SSC = 1 <<  3,
+	/* params2 */
+	MTHCA_QP_BIT_RRE = 1 << 15,
+	MTHCA_QP_BIT_RWE = 1 << 14,
+	MTHCA_QP_BIT_RAE = 1 << 13,
+	MTHCA_QP_BIT_RIC = 1 <<  4,
+	MTHCA_QP_BIT_RSC = 1 <<  3
+};
+
+enum {
+	MTHCA_SEND_DOORBELL_FENCE = 1 << 5
+};
+
+struct mthca_qp_path {
+	__be32 port_pkey;
+	u8     rnr_retry;
+	u8     g_mylmc;
+	__be16 rlid;
+	u8     ackto;
+	u8     mgid_index;
+	u8     static_rate;
+	u8     hop_limit;
+	__be32 sl_tclass_flowlabel;
+	u8     rgid[16];
+} __packed;
+
+struct mthca_qp_context {
+	__be32 flags;
+	__be32 tavor_sched_queue; /* Reserved on Arbel */
+	u8     mtu_msgmax;
+	u8     rq_size_stride;	/* Reserved on Tavor */
+	u8     sq_size_stride;	/* Reserved on Tavor */
+	u8     rlkey_arbel_sched_queue;	/* Reserved on Tavor */
+	__be32 usr_page;
+	__be32 local_qpn;
+	__be32 remote_qpn;
+	u32    reserved1[2];
+	struct mthca_qp_path pri_path;
+	struct mthca_qp_path alt_path;
+	__be32 rdd;
+	__be32 pd;
+	__be32 wqe_base;
+	__be32 wqe_lkey;
+	__be32 params1;
+	__be32 reserved2;
+	__be32 next_send_psn;
+	__be32 cqn_snd;
+	__be32 snd_wqe_base_l;	/* Next send WQE on Tavor */
+	__be32 snd_db_index;	/* (debugging only entries) */
+	__be32 last_acked_psn;
+	__be32 ssn;
+	__be32 params2;
+	__be32 rnr_nextrecvpsn;
+	__be32 ra_buff_indx;
+	__be32 cqn_rcv;
+	__be32 rcv_wqe_base_l;	/* Next recv WQE on Tavor */
+	__be32 rcv_db_index;	/* (debugging only entries) */
+	__be32 qkey;
+	__be32 srqn;
+	__be32 rmsn;
+	__be16 rq_wqe_counter;	/* reserved on Tavor */
+	__be16 sq_wqe_counter;	/* reserved on Tavor */
+	u32    reserved3[18];
+} __packed;
+
+struct mthca_qp_param {
+	__be32 opt_param_mask;
+	u32    reserved1;
+	struct mthca_qp_context context;
+	u32    reserved2[62];
+} __packed;
+
+enum {
+	MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
+	MTHCA_QP_OPTPAR_RRE               = 1 << 1,
+	MTHCA_QP_OPTPAR_RAE               = 1 << 2,
+	MTHCA_QP_OPTPAR_RWE               = 1 << 3,
+	MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
+	MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
+	MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
+	MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
+	MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
+	MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
+	MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
+	MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
+	MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
+	MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
+	MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
+	MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
+	MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
+};
+
+static const u8 mthca_opcode[] = {
+	[IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
+	[IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
+	[IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
+	[IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
+	[IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
+	[IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
+	[IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
+};
+
+static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
+{
+	return qp->qpn >= dev->qp_table.sqp_start &&
+		qp->qpn <= dev->qp_table.sqp_start + 3;
+}
+
+static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
+{
+	return qp->qpn >= dev->qp_table.sqp_start &&
+		qp->qpn <= dev->qp_table.sqp_start + 1;
+}
+
+static void *get_recv_wqe(struct mthca_qp *qp, int n)
+{
+	if (qp->is_direct)
+		return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
+	else
+		return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
+			((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
+}
+
+static void *get_send_wqe(struct mthca_qp *qp, int n)
+{
+	if (qp->is_direct)
+		return qp->queue.direct.buf + qp->send_wqe_offset +
+			(n << qp->sq.wqe_shift);
+	else
+		return qp->queue.page_list[(qp->send_wqe_offset +
+					    (n << qp->sq.wqe_shift)) >>
+					   PAGE_SHIFT].buf +
+			((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
+			 (PAGE_SIZE - 1));
+}
+
+static void mthca_wq_reset(struct mthca_wq *wq)
+{
+	wq->next_ind  = 0;
+	wq->last_comp = wq->max - 1;
+	wq->head      = 0;
+	wq->tail      = 0;
+}
+
+void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
+		    enum ib_event_type event_type)
+{
+	struct mthca_qp *qp;
+	struct ib_event event;
+
+	spin_lock(&dev->qp_table.lock);
+	qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
+	if (qp)
+		++qp->refcount;
+	spin_unlock(&dev->qp_table.lock);
+
+	if (!qp) {
+		mthca_warn(dev, "Async event %d for bogus QP %08x\n",
+			   event_type, qpn);
+		return;
+	}
+
+	if (event_type == IB_EVENT_PATH_MIG)
+		qp->port = qp->alt_port;
+
+	event.device      = &dev->ib_dev;
+	event.event       = event_type;
+	event.element.qp  = &qp->ibqp;
+	if (qp->ibqp.event_handler)
+		qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
+
+	spin_lock(&dev->qp_table.lock);
+	if (!--qp->refcount)
+		wake_up(&qp->wait);
+	spin_unlock(&dev->qp_table.lock);
+}
+
+static int to_mthca_state(enum ib_qp_state ib_state)
+{
+	switch (ib_state) {
+	case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
+	case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
+	case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
+	case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
+	case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
+	case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
+	case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
+	default:                return -1;
+	}
+}
+
+enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
+
+static int to_mthca_st(int transport)
+{
+	switch (transport) {
+	case RC:  return MTHCA_QP_ST_RC;
+	case UC:  return MTHCA_QP_ST_UC;
+	case UD:  return MTHCA_QP_ST_UD;
+	case RD:  return MTHCA_QP_ST_RD;
+	case MLX: return MTHCA_QP_ST_MLX;
+	default:  return -1;
+	}
+}
+
+static void store_attrs(struct mthca_sqp *sqp, const struct ib_qp_attr *attr,
+			int attr_mask)
+{
+	if (attr_mask & IB_QP_PKEY_INDEX)
+		sqp->pkey_index = attr->pkey_index;
+	if (attr_mask & IB_QP_QKEY)
+		sqp->qkey = attr->qkey;
+	if (attr_mask & IB_QP_SQ_PSN)
+		sqp->send_psn = attr->sq_psn;
+}
+
+static void init_port(struct mthca_dev *dev, int port)
+{
+	int err;
+	struct mthca_init_ib_param param;
+
+	memset(&param, 0, sizeof param);
+
+	param.port_width = dev->limits.port_width_cap;
+	param.vl_cap     = dev->limits.vl_cap;
+	param.mtu_cap    = dev->limits.mtu_cap;
+	param.gid_cap    = dev->limits.gid_table_len;
+	param.pkey_cap   = dev->limits.pkey_table_len;
+
+	err = mthca_INIT_IB(dev, &param, port);
+	if (err)
+		mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
+}
+
+static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr,
+				  int attr_mask)
+{
+	u8 dest_rd_atomic;
+	u32 access_flags;
+	u32 hw_access_flags = 0;
+
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
+		dest_rd_atomic = attr->max_dest_rd_atomic;
+	else
+		dest_rd_atomic = qp->resp_depth;
+
+	if (attr_mask & IB_QP_ACCESS_FLAGS)
+		access_flags = attr->qp_access_flags;
+	else
+		access_flags = qp->atomic_rd_en;
+
+	if (!dest_rd_atomic)
+		access_flags &= IB_ACCESS_REMOTE_WRITE;
+
+	if (access_flags & IB_ACCESS_REMOTE_READ)
+		hw_access_flags |= MTHCA_QP_BIT_RRE;
+	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
+		hw_access_flags |= MTHCA_QP_BIT_RAE;
+	if (access_flags & IB_ACCESS_REMOTE_WRITE)
+		hw_access_flags |= MTHCA_QP_BIT_RWE;
+
+	return cpu_to_be32(hw_access_flags);
+}
+
+static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
+{
+	switch (mthca_state) {
+	case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
+	case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
+	case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
+	case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
+	case MTHCA_QP_STATE_DRAINING:
+	case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
+	case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
+	case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
+	default:                      return -1;
+	}
+}
+
+static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
+{
+	switch (mthca_mig_state) {
+	case 0:  return IB_MIG_ARMED;
+	case 1:  return IB_MIG_REARM;
+	case 3:  return IB_MIG_MIGRATED;
+	default: return -1;
+	}
+}
+
+static int to_ib_qp_access_flags(int mthca_flags)
+{
+	int ib_flags = 0;
+
+	if (mthca_flags & MTHCA_QP_BIT_RRE)
+		ib_flags |= IB_ACCESS_REMOTE_READ;
+	if (mthca_flags & MTHCA_QP_BIT_RWE)
+		ib_flags |= IB_ACCESS_REMOTE_WRITE;
+	if (mthca_flags & MTHCA_QP_BIT_RAE)
+		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
+
+	return ib_flags;
+}
+
+static void to_rdma_ah_attr(struct mthca_dev *dev,
+			    struct rdma_ah_attr *ah_attr,
+			    struct mthca_qp_path *path)
+{
+	u8 port_num = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
+
+	memset(ah_attr, 0, sizeof(*ah_attr));
+
+	if (port_num == 0 || port_num > dev->limits.num_ports)
+		return;
+	ah_attr->type = rdma_ah_find_type(&dev->ib_dev, port_num);
+	rdma_ah_set_port_num(ah_attr, port_num);
+
+	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
+	rdma_ah_set_sl(ah_attr, be32_to_cpu(path->sl_tclass_flowlabel) >> 28);
+	rdma_ah_set_path_bits(ah_attr, path->g_mylmc & 0x7f);
+	rdma_ah_set_static_rate(ah_attr,
+				mthca_rate_to_ib(dev,
+						 path->static_rate & 0xf,
+						 port_num));
+	if (path->g_mylmc & (1 << 7)) {
+		u32 tc_fl = be32_to_cpu(path->sl_tclass_flowlabel);
+
+		rdma_ah_set_grh(ah_attr, NULL,
+				tc_fl & 0xfffff,
+				path->mgid_index &
+				(dev->limits.gid_table_len - 1),
+				path->hop_limit,
+				(tc_fl >> 20) & 0xff);
+		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
+	}
+}
+
+int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
+		   struct ib_qp_init_attr *qp_init_attr)
+{
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	struct mthca_qp *qp = to_mqp(ibqp);
+	int err = 0;
+	struct mthca_mailbox *mailbox = NULL;
+	struct mthca_qp_param *qp_param;
+	struct mthca_qp_context *context;
+	int mthca_state;
+
+	mutex_lock(&qp->mutex);
+
+	if (qp->state == IB_QPS_RESET) {
+		qp_attr->qp_state = IB_QPS_RESET;
+		goto done;
+	}
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		err = PTR_ERR(mailbox);
+		goto out;
+	}
+
+	err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox);
+	if (err) {
+		mthca_warn(dev, "QUERY_QP failed (%d)\n", err);
+		goto out_mailbox;
+	}
+
+	qp_param    = mailbox->buf;
+	context     = &qp_param->context;
+	mthca_state = be32_to_cpu(context->flags) >> 28;
+
+	qp->state		     = to_ib_qp_state(mthca_state);
+	qp_attr->qp_state	     = qp->state;
+	qp_attr->path_mtu 	     = context->mtu_msgmax >> 5;
+	qp_attr->path_mig_state      =
+		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
+	qp_attr->qkey 		     = be32_to_cpu(context->qkey);
+	qp_attr->rq_psn 	     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
+	qp_attr->sq_psn 	     = be32_to_cpu(context->next_send_psn) & 0xffffff;
+	qp_attr->dest_qp_num 	     = be32_to_cpu(context->remote_qpn) & 0xffffff;
+	qp_attr->qp_access_flags     =
+		to_ib_qp_access_flags(be32_to_cpu(context->params2));
+
+	if (qp->transport == RC || qp->transport == UC) {
+		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
+		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
+		qp_attr->alt_pkey_index =
+			be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
+		qp_attr->alt_port_num	=
+			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
+	}
+
+	qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
+	qp_attr->port_num   =
+		(be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
+
+	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
+	qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
+
+	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
+
+	qp_attr->max_dest_rd_atomic =
+		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
+	qp_attr->min_rnr_timer 	    =
+		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
+	qp_attr->timeout 	    = context->pri_path.ackto >> 3;
+	qp_attr->retry_cnt 	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
+	qp_attr->rnr_retry 	    = context->pri_path.rnr_retry >> 5;
+	qp_attr->alt_timeout 	    = context->alt_path.ackto >> 3;
+
+done:
+	qp_attr->cur_qp_state	     = qp_attr->qp_state;
+	qp_attr->cap.max_send_wr     = qp->sq.max;
+	qp_attr->cap.max_recv_wr     = qp->rq.max;
+	qp_attr->cap.max_send_sge    = qp->sq.max_gs;
+	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
+	qp_attr->cap.max_inline_data = qp->max_inline_data;
+
+	qp_init_attr->cap	     = qp_attr->cap;
+	qp_init_attr->sq_sig_type    = qp->sq_policy;
+
+out_mailbox:
+	mthca_free_mailbox(dev, mailbox);
+
+out:
+	mutex_unlock(&qp->mutex);
+	return err;
+}
+
+static int mthca_path_set(struct mthca_dev *dev, const struct rdma_ah_attr *ah,
+			  struct mthca_qp_path *path, u8 port)
+{
+	path->g_mylmc     = rdma_ah_get_path_bits(ah) & 0x7f;
+	path->rlid        = cpu_to_be16(rdma_ah_get_dlid(ah));
+	path->static_rate = mthca_get_rate(dev, rdma_ah_get_static_rate(ah),
+					   port);
+
+	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
+		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
+
+		if (grh->sgid_index >= dev->limits.gid_table_len) {
+			mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
+				  grh->sgid_index,
+				  dev->limits.gid_table_len - 1);
+			return -1;
+		}
+
+		path->g_mylmc   |= 1 << 7;
+		path->mgid_index = grh->sgid_index;
+		path->hop_limit  = grh->hop_limit;
+		path->sl_tclass_flowlabel =
+			cpu_to_be32((rdma_ah_get_sl(ah) << 28) |
+				    (grh->traffic_class << 20) |
+				    (grh->flow_label));
+		memcpy(path->rgid, grh->dgid.raw, 16);
+	} else {
+		path->sl_tclass_flowlabel = cpu_to_be32(rdma_ah_get_sl(ah) <<
+							28);
+	}
+
+	return 0;
+}
+
+static int __mthca_modify_qp(struct ib_qp *ibqp,
+			     const struct ib_qp_attr *attr, int attr_mask,
+			     enum ib_qp_state cur_state,
+			     enum ib_qp_state new_state,
+			     struct ib_udata *udata)
+{
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	struct mthca_qp *qp = to_mqp(ibqp);
+	struct mthca_ucontext *context = rdma_udata_to_drv_context(
+		udata, struct mthca_ucontext, ibucontext);
+	struct mthca_mailbox *mailbox;
+	struct mthca_qp_param *qp_param;
+	struct mthca_qp_context *qp_context;
+	u32 sqd_event = 0;
+	int err = -EINVAL;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		err = PTR_ERR(mailbox);
+		goto out;
+	}
+	qp_param = mailbox->buf;
+	qp_context = &qp_param->context;
+	memset(qp_param, 0, sizeof *qp_param);
+
+	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
+					     (to_mthca_st(qp->transport) << 16));
+	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
+	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
+		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
+	else {
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
+		switch (attr->path_mig_state) {
+		case IB_MIG_MIGRATED:
+			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
+			break;
+		case IB_MIG_REARM:
+			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
+			break;
+		case IB_MIG_ARMED:
+			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
+			break;
+		}
+	}
+
+	/* leave tavor_sched_queue as 0 */
+
+	if (qp->transport == MLX || qp->transport == UD)
+		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
+	else if (attr_mask & IB_QP_PATH_MTU) {
+		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
+			mthca_dbg(dev, "path MTU (%u) is invalid\n",
+				  attr->path_mtu);
+			goto out_mailbox;
+		}
+		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
+	}
+
+	if (mthca_is_memfree(dev)) {
+		if (qp->rq.max)
+			qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
+		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
+
+		if (qp->sq.max)
+			qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
+		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
+	}
+
+	/* leave arbel_sched_queue as 0 */
+
+	if (qp->ibqp.uobject)
+		qp_context->usr_page = cpu_to_be32(context->uar.index);
+	else
+		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
+	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
+	if (attr_mask & IB_QP_DEST_QPN) {
+		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
+	}
+
+	if (qp->transport == MLX)
+		qp_context->pri_path.port_pkey |=
+			cpu_to_be32(qp->port << 24);
+	else {
+		if (attr_mask & IB_QP_PORT) {
+			qp_context->pri_path.port_pkey |=
+				cpu_to_be32(attr->port_num << 24);
+			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
+		}
+	}
+
+	if (attr_mask & IB_QP_PKEY_INDEX) {
+		qp_context->pri_path.port_pkey |=
+			cpu_to_be32(attr->pkey_index);
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
+	}
+
+	if (attr_mask & IB_QP_RNR_RETRY) {
+		qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
+			attr->rnr_retry << 5;
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
+							MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
+	}
+
+	if (attr_mask & IB_QP_AV) {
+		if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
+				   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
+			goto out_mailbox;
+
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
+	}
+
+	if (ibqp->qp_type == IB_QPT_RC &&
+	    cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
+		u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
+
+		if (mthca_is_memfree(dev))
+			qp_context->rlkey_arbel_sched_queue |= sched_queue;
+		else
+			qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
+
+		qp_param->opt_param_mask |=
+			cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
+	}
+
+	if (attr_mask & IB_QP_TIMEOUT) {
+		qp_context->pri_path.ackto = attr->timeout << 3;
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
+	}
+
+	if (attr_mask & IB_QP_ALT_PATH) {
+		if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
+			mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
+				  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
+			goto out_mailbox;
+		}
+
+		if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
+			mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
+				attr->alt_port_num);
+			goto out_mailbox;
+		}
+
+		if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
+				   rdma_ah_get_port_num(&attr->alt_ah_attr)))
+			goto out_mailbox;
+
+		qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
+							      attr->alt_port_num << 24);
+		qp_context->alt_path.ackto = attr->alt_timeout << 3;
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
+	}
+
+	/* leave rdd as 0 */
+	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
+	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
+	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
+	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
+					     (MTHCA_FLIGHT_LIMIT << 24) |
+					     MTHCA_QP_BIT_SWE);
+	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
+		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
+	if (attr_mask & IB_QP_RETRY_CNT) {
+		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
+	}
+
+	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
+		if (attr->max_rd_atomic) {
+			qp_context->params1 |=
+				cpu_to_be32(MTHCA_QP_BIT_SRE |
+					    MTHCA_QP_BIT_SAE);
+			qp_context->params1 |=
+				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
+		}
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
+	}
+
+	if (attr_mask & IB_QP_SQ_PSN)
+		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
+	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
+
+	if (mthca_is_memfree(dev)) {
+		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
+		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
+	}
+
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
+		if (attr->max_dest_rd_atomic)
+			qp_context->params2 |=
+				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
+
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
+	}
+
+	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
+		qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
+							MTHCA_QP_OPTPAR_RRE |
+							MTHCA_QP_OPTPAR_RAE);
+	}
+
+	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
+
+	if (ibqp->srq)
+		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
+
+	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
+		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
+	}
+	if (attr_mask & IB_QP_RQ_PSN)
+		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
+
+	qp_context->ra_buff_indx =
+		cpu_to_be32(dev->qp_table.rdb_base +
+			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
+			     dev->qp_table.rdb_shift));
+
+	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
+
+	if (mthca_is_memfree(dev))
+		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
+
+	if (attr_mask & IB_QP_QKEY) {
+		qp_context->qkey = cpu_to_be32(attr->qkey);
+		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
+	}
+
+	if (ibqp->srq)
+		qp_context->srqn = cpu_to_be32(1 << 24 |
+					       to_msrq(ibqp->srq)->srqn);
+
+	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
+	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY		&&
+	    attr->en_sqd_async_notify)
+		sqd_event = 1 << 31;
+
+	err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
+			      mailbox, sqd_event);
+	if (err) {
+		mthca_warn(dev, "modify QP %d->%d returned %d.\n",
+			   cur_state, new_state, err);
+		goto out_mailbox;
+	}
+
+	qp->state = new_state;
+	if (attr_mask & IB_QP_ACCESS_FLAGS)
+		qp->atomic_rd_en = attr->qp_access_flags;
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
+		qp->resp_depth = attr->max_dest_rd_atomic;
+	if (attr_mask & IB_QP_PORT)
+		qp->port = attr->port_num;
+	if (attr_mask & IB_QP_ALT_PATH)
+		qp->alt_port = attr->alt_port_num;
+
+	if (is_sqp(dev, qp))
+		store_attrs(qp->sqp, attr, attr_mask);
+
+	/*
+	 * If we moved QP0 to RTR, bring the IB link up; if we moved
+	 * QP0 to RESET or ERROR, bring the link back down.
+	 */
+	if (is_qp0(dev, qp)) {
+		if (cur_state != IB_QPS_RTR &&
+		    new_state == IB_QPS_RTR)
+			init_port(dev, qp->port);
+
+		if (cur_state != IB_QPS_RESET &&
+		    cur_state != IB_QPS_ERR &&
+		    (new_state == IB_QPS_RESET ||
+		     new_state == IB_QPS_ERR))
+			mthca_CLOSE_IB(dev, qp->port);
+	}
+
+	/*
+	 * If we moved a kernel QP to RESET, clean up all old CQ
+	 * entries and reinitialize the QP.
+	 */
+	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
+		mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
+			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
+		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
+			mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
+
+		mthca_wq_reset(&qp->sq);
+		qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
+
+		mthca_wq_reset(&qp->rq);
+		qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
+
+		if (mthca_is_memfree(dev)) {
+			*qp->sq.db = 0;
+			*qp->rq.db = 0;
+		}
+	}
+
+out_mailbox:
+	mthca_free_mailbox(dev, mailbox);
+out:
+	return err;
+}
+
+int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
+		    struct ib_udata *udata)
+{
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	struct mthca_qp *qp = to_mqp(ibqp);
+	enum ib_qp_state cur_state, new_state;
+	int err = -EINVAL;
+
+	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
+		return -EOPNOTSUPP;
+
+	mutex_lock(&qp->mutex);
+	if (attr_mask & IB_QP_CUR_STATE) {
+		cur_state = attr->cur_qp_state;
+	} else {
+		spin_lock_irq(&qp->sq.lock);
+		spin_lock(&qp->rq.lock);
+		cur_state = qp->state;
+		spin_unlock(&qp->rq.lock);
+		spin_unlock_irq(&qp->sq.lock);
+	}
+
+	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
+
+	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
+				attr_mask)) {
+		mthca_dbg(dev, "Bad QP transition (transport %d) "
+			  "%d->%d with attr 0x%08x\n",
+			  qp->transport, cur_state, new_state,
+			  attr_mask);
+		goto out;
+	}
+
+	if ((attr_mask & IB_QP_PKEY_INDEX) &&
+	     attr->pkey_index >= dev->limits.pkey_table_len) {
+		mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
+			  attr->pkey_index, dev->limits.pkey_table_len-1);
+		goto out;
+	}
+
+	if ((attr_mask & IB_QP_PORT) &&
+	    (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
+		mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
+		goto out;
+	}
+
+	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
+	    attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
+		mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
+			  attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
+		goto out;
+	}
+
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
+	    attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
+		mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
+			  attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
+		goto out;
+	}
+
+	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
+		err = 0;
+		goto out;
+	}
+
+	err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state,
+				udata);
+
+out:
+	mutex_unlock(&qp->mutex);
+	return err;
+}
+
+static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
+{
+	/*
+	 * Calculate the maximum size of WQE s/g segments, excluding
+	 * the next segment and other non-data segments.
+	 */
+	int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
+
+	switch (qp->transport) {
+	case MLX:
+		max_data_size -= 2 * sizeof (struct mthca_data_seg);
+		break;
+
+	case UD:
+		if (mthca_is_memfree(dev))
+			max_data_size -= sizeof (struct mthca_arbel_ud_seg);
+		else
+			max_data_size -= sizeof (struct mthca_tavor_ud_seg);
+		break;
+
+	default:
+		max_data_size -= sizeof (struct mthca_raddr_seg);
+		break;
+	}
+
+	return max_data_size;
+}
+
+static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
+{
+	/* We don't support inline data for kernel QPs (yet). */
+	return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
+}
+
+static void mthca_adjust_qp_caps(struct mthca_dev *dev,
+				 struct mthca_pd *pd,
+				 struct mthca_qp *qp)
+{
+	int max_data_size = mthca_max_data_size(dev, qp,
+						min(dev->limits.max_desc_sz,
+						    1 << qp->sq.wqe_shift));
+
+	qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
+
+	qp->sq.max_gs = min_t(int, dev->limits.max_sg,
+			      max_data_size / sizeof (struct mthca_data_seg));
+	qp->rq.max_gs = min_t(int, dev->limits.max_sg,
+			       (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
+				sizeof (struct mthca_next_seg)) /
+			       sizeof (struct mthca_data_seg));
+}
+
+/*
+ * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
+ * rq.max_gs and sq.max_gs must all be assigned.
+ * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
+ * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
+ * queue)
+ */
+static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
+			       struct mthca_pd *pd,
+			       struct mthca_qp *qp,
+			       struct ib_udata *udata)
+{
+	int size;
+	int err = -ENOMEM;
+
+	size = sizeof (struct mthca_next_seg) +
+		qp->rq.max_gs * sizeof (struct mthca_data_seg);
+
+	if (size > dev->limits.max_desc_sz)
+		return -EINVAL;
+
+	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
+	     qp->rq.wqe_shift++)
+		; /* nothing */
+
+	size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
+	switch (qp->transport) {
+	case MLX:
+		size += 2 * sizeof (struct mthca_data_seg);
+		break;
+
+	case UD:
+		size += mthca_is_memfree(dev) ?
+			sizeof (struct mthca_arbel_ud_seg) :
+			sizeof (struct mthca_tavor_ud_seg);
+		break;
+
+	case UC:
+		size += sizeof (struct mthca_raddr_seg);
+		break;
+
+	case RC:
+		size += sizeof (struct mthca_raddr_seg);
+		/*
+		 * An atomic op will require an atomic segment, a
+		 * remote address segment and one scatter entry.
+		 */
+		size = max_t(int, size,
+			     sizeof (struct mthca_atomic_seg) +
+			     sizeof (struct mthca_raddr_seg) +
+			     sizeof (struct mthca_data_seg));
+		break;
+
+	default:
+		break;
+	}
+
+	/* Make sure that we have enough space for a bind request */
+	size = max_t(int, size, sizeof (struct mthca_bind_seg));
+
+	size += sizeof (struct mthca_next_seg);
+
+	if (size > dev->limits.max_desc_sz)
+		return -EINVAL;
+
+	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
+	     qp->sq.wqe_shift++)
+		; /* nothing */
+
+	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
+				    1 << qp->sq.wqe_shift);
+
+	/*
+	 * If this is a userspace QP, we don't actually have to
+	 * allocate anything.  All we need is to calculate the WQE
+	 * sizes and the send_wqe_offset, so we're done now.
+	 */
+	if (udata)
+		return 0;
+
+	size = PAGE_ALIGN(qp->send_wqe_offset +
+			  (qp->sq.max << qp->sq.wqe_shift));
+
+	qp->wrid = kmalloc_array(qp->rq.max + qp->sq.max, sizeof(u64),
+				 GFP_KERNEL);
+	if (!qp->wrid)
+		goto err_out;
+
+	err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
+			      &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
+	if (err)
+		goto err_out;
+
+	return 0;
+
+err_out:
+	kfree(qp->wrid);
+	return err;
+}
+
+static void mthca_free_wqe_buf(struct mthca_dev *dev,
+			       struct mthca_qp *qp)
+{
+	mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
+				       (qp->sq.max << qp->sq.wqe_shift)),
+		       &qp->queue, qp->is_direct, &qp->mr);
+	kfree(qp->wrid);
+}
+
+static int mthca_map_memfree(struct mthca_dev *dev,
+			     struct mthca_qp *qp)
+{
+	int ret;
+
+	if (mthca_is_memfree(dev)) {
+		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
+		if (ret)
+			return ret;
+
+		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
+		if (ret)
+			goto err_qpc;
+
+		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
+				      qp->qpn << dev->qp_table.rdb_shift);
+		if (ret)
+			goto err_eqpc;
+
+	}
+
+	return 0;
+
+err_eqpc:
+	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
+
+err_qpc:
+	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
+
+	return ret;
+}
+
+static void mthca_unmap_memfree(struct mthca_dev *dev,
+				struct mthca_qp *qp)
+{
+	mthca_table_put(dev, dev->qp_table.rdb_table,
+			qp->qpn << dev->qp_table.rdb_shift);
+	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
+	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
+}
+
+static int mthca_alloc_memfree(struct mthca_dev *dev,
+			       struct mthca_qp *qp)
+{
+	if (mthca_is_memfree(dev)) {
+		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
+						 qp->qpn, &qp->rq.db);
+		if (qp->rq.db_index < 0)
+			return -ENOMEM;
+
+		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
+						 qp->qpn, &qp->sq.db);
+		if (qp->sq.db_index < 0) {
+			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
+			return -ENOMEM;
+		}
+	}
+
+	return 0;
+}
+
+static void mthca_free_memfree(struct mthca_dev *dev,
+			       struct mthca_qp *qp)
+{
+	if (mthca_is_memfree(dev)) {
+		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
+		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
+	}
+}
+
+static int mthca_alloc_qp_common(struct mthca_dev *dev,
+				 struct mthca_pd *pd,
+				 struct mthca_cq *send_cq,
+				 struct mthca_cq *recv_cq,
+				 enum ib_sig_type send_policy,
+				 struct mthca_qp *qp,
+				 struct ib_udata *udata)
+{
+	int ret;
+	int i;
+	struct mthca_next_seg *next;
+
+	qp->refcount = 1;
+	init_waitqueue_head(&qp->wait);
+	mutex_init(&qp->mutex);
+	qp->state    	 = IB_QPS_RESET;
+	qp->atomic_rd_en = 0;
+	qp->resp_depth   = 0;
+	qp->sq_policy    = send_policy;
+	mthca_wq_reset(&qp->sq);
+	mthca_wq_reset(&qp->rq);
+
+	spin_lock_init(&qp->sq.lock);
+	spin_lock_init(&qp->rq.lock);
+
+	ret = mthca_map_memfree(dev, qp);
+	if (ret)
+		return ret;
+
+	ret = mthca_alloc_wqe_buf(dev, pd, qp, udata);
+	if (ret) {
+		mthca_unmap_memfree(dev, qp);
+		return ret;
+	}
+
+	mthca_adjust_qp_caps(dev, pd, qp);
+
+	/*
+	 * If this is a userspace QP, we're done now.  The doorbells
+	 * will be allocated and buffers will be initialized in
+	 * userspace.
+	 */
+	if (udata)
+		return 0;
+
+	ret = mthca_alloc_memfree(dev, qp);
+	if (ret) {
+		mthca_free_wqe_buf(dev, qp);
+		mthca_unmap_memfree(dev, qp);
+		return ret;
+	}
+
+	if (mthca_is_memfree(dev)) {
+		struct mthca_data_seg *scatter;
+		int size = (sizeof (struct mthca_next_seg) +
+			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
+
+		for (i = 0; i < qp->rq.max; ++i) {
+			next = get_recv_wqe(qp, i);
+			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
+						   qp->rq.wqe_shift);
+			next->ee_nds = cpu_to_be32(size);
+
+			for (scatter = (void *) (next + 1);
+			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
+			     ++scatter)
+				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
+		}
+
+		for (i = 0; i < qp->sq.max; ++i) {
+			next = get_send_wqe(qp, i);
+			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
+						    qp->sq.wqe_shift) +
+						   qp->send_wqe_offset);
+		}
+	} else {
+		for (i = 0; i < qp->rq.max; ++i) {
+			next = get_recv_wqe(qp, i);
+			next->nda_op = htonl((((i + 1) % qp->rq.max) <<
+					      qp->rq.wqe_shift) | 1);
+		}
+
+	}
+
+	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
+	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
+
+	return 0;
+}
+
+static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
+			     struct mthca_pd *pd, struct mthca_qp *qp)
+{
+	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
+
+	/* Sanity check QP size before proceeding */
+	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
+	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
+	    cap->max_send_sge 	 > dev->limits.max_sg   ||
+	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
+	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
+		return -EINVAL;
+
+	/*
+	 * For MLX transport we need 2 extra send gather entries:
+	 * one for the header and one for the checksum at the end
+	 */
+	if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg)
+		return -EINVAL;
+
+	if (mthca_is_memfree(dev)) {
+		qp->rq.max = cap->max_recv_wr ?
+			roundup_pow_of_two(cap->max_recv_wr) : 0;
+		qp->sq.max = cap->max_send_wr ?
+			roundup_pow_of_two(cap->max_send_wr) : 0;
+	} else {
+		qp->rq.max = cap->max_recv_wr;
+		qp->sq.max = cap->max_send_wr;
+	}
+
+	qp->rq.max_gs = cap->max_recv_sge;
+	qp->sq.max_gs = max_t(int, cap->max_send_sge,
+			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
+				    MTHCA_INLINE_CHUNK_SIZE) /
+			      sizeof (struct mthca_data_seg));
+
+	return 0;
+}
+
+int mthca_alloc_qp(struct mthca_dev *dev,
+		   struct mthca_pd *pd,
+		   struct mthca_cq *send_cq,
+		   struct mthca_cq *recv_cq,
+		   enum ib_qp_type type,
+		   enum ib_sig_type send_policy,
+		   struct ib_qp_cap *cap,
+		   struct mthca_qp *qp,
+		   struct ib_udata *udata)
+{
+	int err;
+
+	switch (type) {
+	case IB_QPT_RC: qp->transport = RC; break;
+	case IB_QPT_UC: qp->transport = UC; break;
+	case IB_QPT_UD: qp->transport = UD; break;
+	default: return -EINVAL;
+	}
+
+	err = mthca_set_qp_size(dev, cap, pd, qp);
+	if (err)
+		return err;
+
+	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
+	if (qp->qpn == -1)
+		return -ENOMEM;
+
+	/* initialize port to zero for error-catching. */
+	qp->port = 0;
+
+	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
+				    send_policy, qp, udata);
+	if (err) {
+		mthca_free(&dev->qp_table.alloc, qp->qpn);
+		return err;
+	}
+
+	spin_lock_irq(&dev->qp_table.lock);
+	mthca_array_set(&dev->qp_table.qp,
+			qp->qpn & (dev->limits.num_qps - 1), qp);
+	spin_unlock_irq(&dev->qp_table.lock);
+
+	return 0;
+}
+
+static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
+	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
+{
+	if (send_cq == recv_cq) {
+		spin_lock_irq(&send_cq->lock);
+		__acquire(&recv_cq->lock);
+	} else if (send_cq->cqn < recv_cq->cqn) {
+		spin_lock_irq(&send_cq->lock);
+		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
+	} else {
+		spin_lock_irq(&recv_cq->lock);
+		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
+	}
+}
+
+static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
+	__releases(&send_cq->lock) __releases(&recv_cq->lock)
+{
+	if (send_cq == recv_cq) {
+		__release(&recv_cq->lock);
+		spin_unlock_irq(&send_cq->lock);
+	} else if (send_cq->cqn < recv_cq->cqn) {
+		spin_unlock(&recv_cq->lock);
+		spin_unlock_irq(&send_cq->lock);
+	} else {
+		spin_unlock(&send_cq->lock);
+		spin_unlock_irq(&recv_cq->lock);
+	}
+}
+
+int mthca_alloc_sqp(struct mthca_dev *dev,
+		    struct mthca_pd *pd,
+		    struct mthca_cq *send_cq,
+		    struct mthca_cq *recv_cq,
+		    enum ib_sig_type send_policy,
+		    struct ib_qp_cap *cap,
+		    int qpn,
+		    u32 port,
+		    struct mthca_qp *qp,
+		    struct ib_udata *udata)
+{
+	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
+	int err;
+
+	qp->transport = MLX;
+	err = mthca_set_qp_size(dev, cap, pd, qp);
+	if (err)
+		return err;
+
+	qp->sqp->header_buf_size = qp->sq.max * MTHCA_UD_HEADER_SIZE;
+	qp->sqp->header_buf =
+		dma_alloc_coherent(&dev->pdev->dev, qp->sqp->header_buf_size,
+				   &qp->sqp->header_dma, GFP_KERNEL);
+	if (!qp->sqp->header_buf)
+		return -ENOMEM;
+
+	spin_lock_irq(&dev->qp_table.lock);
+	if (mthca_array_get(&dev->qp_table.qp, mqpn))
+		err = -EBUSY;
+	else
+		mthca_array_set(&dev->qp_table.qp, mqpn, qp);
+	spin_unlock_irq(&dev->qp_table.lock);
+
+	if (err)
+		goto err_out;
+
+	qp->port      = port;
+	qp->qpn       = mqpn;
+	qp->transport = MLX;
+
+	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
+				    send_policy, qp, udata);
+	if (err)
+		goto err_out_free;
+
+	atomic_inc(&pd->sqp_count);
+
+	return 0;
+
+ err_out_free:
+	/*
+	 * Lock CQs here, so that CQ polling code can do QP lookup
+	 * without taking a lock.
+	 */
+	mthca_lock_cqs(send_cq, recv_cq);
+
+	spin_lock(&dev->qp_table.lock);
+	mthca_array_clear(&dev->qp_table.qp, mqpn);
+	spin_unlock(&dev->qp_table.lock);
+
+	mthca_unlock_cqs(send_cq, recv_cq);
+
+err_out:
+	dma_free_coherent(&dev->pdev->dev, qp->sqp->header_buf_size,
+			  qp->sqp->header_buf, qp->sqp->header_dma);
+	return err;
+}
+
+static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
+{
+	int c;
+
+	spin_lock_irq(&dev->qp_table.lock);
+	c = qp->refcount;
+	spin_unlock_irq(&dev->qp_table.lock);
+
+	return c;
+}
+
+void mthca_free_qp(struct mthca_dev *dev,
+		   struct mthca_qp *qp)
+{
+	struct mthca_cq *send_cq;
+	struct mthca_cq *recv_cq;
+
+	send_cq = to_mcq(qp->ibqp.send_cq);
+	recv_cq = to_mcq(qp->ibqp.recv_cq);
+
+	/*
+	 * Lock CQs here, so that CQ polling code can do QP lookup
+	 * without taking a lock.
+	 */
+	mthca_lock_cqs(send_cq, recv_cq);
+
+	spin_lock(&dev->qp_table.lock);
+	mthca_array_clear(&dev->qp_table.qp,
+			  qp->qpn & (dev->limits.num_qps - 1));
+	--qp->refcount;
+	spin_unlock(&dev->qp_table.lock);
+
+	mthca_unlock_cqs(send_cq, recv_cq);
+
+	wait_event(qp->wait, !get_qp_refcount(dev, qp));
+
+	if (qp->state != IB_QPS_RESET)
+		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
+				NULL, 0);
+
+	/*
+	 * If this is a userspace QP, the buffers, MR, CQs and so on
+	 * will be cleaned up in userspace, so all we have to do is
+	 * unref the mem-free tables and free the QPN in our table.
+	 */
+	if (!qp->ibqp.uobject) {
+		mthca_cq_clean(dev, recv_cq, qp->qpn,
+			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
+		if (send_cq != recv_cq)
+			mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
+
+		mthca_free_memfree(dev, qp);
+		mthca_free_wqe_buf(dev, qp);
+	}
+
+	mthca_unmap_memfree(dev, qp);
+
+	if (is_sqp(dev, qp)) {
+		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
+		dma_free_coherent(&dev->pdev->dev, qp->sqp->header_buf_size,
+				  qp->sqp->header_buf, qp->sqp->header_dma);
+	} else
+		mthca_free(&dev->qp_table.alloc, qp->qpn);
+}
+
+/* Create UD header for an MLX send and build a data segment for it */
+static int build_mlx_header(struct mthca_dev *dev, struct mthca_qp *qp, int ind,
+			    const struct ib_ud_wr *wr,
+			    struct mthca_mlx_seg *mlx,
+			    struct mthca_data_seg *data)
+{
+	struct mthca_sqp *sqp = qp->sqp;
+	int header_size;
+	int err;
+	u16 pkey;
+
+	ib_ud_header_init(256, /* assume a MAD */ 1, 0, 0,
+			  mthca_ah_grh_present(to_mah(wr->ah)), 0, 0, 0,
+			  &sqp->ud_header);
+
+	err = mthca_read_ah(dev, to_mah(wr->ah), &sqp->ud_header);
+	if (err)
+		return err;
+	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
+	mlx->flags |= cpu_to_be32((!qp->ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
+				  (sqp->ud_header.lrh.destination_lid ==
+				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
+				  (sqp->ud_header.lrh.service_level << 8));
+	mlx->rlid = sqp->ud_header.lrh.destination_lid;
+	mlx->vcrc = 0;
+
+	switch (wr->wr.opcode) {
+	case IB_WR_SEND:
+		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
+		sqp->ud_header.immediate_present = 0;
+		break;
+	case IB_WR_SEND_WITH_IMM:
+		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
+		sqp->ud_header.immediate_present = 1;
+		sqp->ud_header.immediate_data = wr->wr.ex.imm_data;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	sqp->ud_header.lrh.virtual_lane    = !qp->ibqp.qp_num ? 15 : 0;
+	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
+		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
+	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
+	if (!qp->ibqp.qp_num)
+		ib_get_cached_pkey(&dev->ib_dev, qp->port, sqp->pkey_index,
+				   &pkey);
+	else
+		ib_get_cached_pkey(&dev->ib_dev, qp->port, wr->pkey_index,
+				   &pkey);
+	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
+	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
+	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
+	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
+					       sqp->qkey : wr->remote_qkey);
+	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->ibqp.qp_num);
+
+	header_size = ib_ud_header_pack(&sqp->ud_header,
+					sqp->header_buf +
+					ind * MTHCA_UD_HEADER_SIZE);
+
+	data->byte_count = cpu_to_be32(header_size);
+	data->lkey       = cpu_to_be32(to_mpd(qp->ibqp.pd)->ntmr.ibmr.lkey);
+	data->addr       = cpu_to_be64(sqp->header_dma +
+				       ind * MTHCA_UD_HEADER_SIZE);
+
+	return 0;
+}
+
+static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
+				    struct ib_cq *ib_cq)
+{
+	unsigned cur;
+	struct mthca_cq *cq;
+
+	cur = wq->head - wq->tail;
+	if (likely(cur + nreq < wq->max))
+		return 0;
+
+	cq = to_mcq(ib_cq);
+	spin_lock(&cq->lock);
+	cur = wq->head - wq->tail;
+	spin_unlock(&cq->lock);
+
+	return cur + nreq >= wq->max;
+}
+
+static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
+					  u64 remote_addr, u32 rkey)
+{
+	rseg->raddr    = cpu_to_be64(remote_addr);
+	rseg->rkey     = cpu_to_be32(rkey);
+	rseg->reserved = 0;
+}
+
+static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
+					   const struct ib_atomic_wr *wr)
+{
+	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
+		aseg->swap_add = cpu_to_be64(wr->swap);
+		aseg->compare  = cpu_to_be64(wr->compare_add);
+	} else {
+		aseg->swap_add = cpu_to_be64(wr->compare_add);
+		aseg->compare  = 0;
+	}
+
+}
+
+static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
+			     const struct ib_ud_wr *wr)
+{
+	useg->lkey    = cpu_to_be32(to_mah(wr->ah)->key);
+	useg->av_addr =	cpu_to_be64(to_mah(wr->ah)->avdma);
+	useg->dqpn    =	cpu_to_be32(wr->remote_qpn);
+	useg->qkey    =	cpu_to_be32(wr->remote_qkey);
+
+}
+
+static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
+			     const struct ib_ud_wr *wr)
+{
+	memcpy(useg->av, to_mah(wr->ah)->av, MTHCA_AV_SIZE);
+	useg->dqpn = cpu_to_be32(wr->remote_qpn);
+	useg->qkey = cpu_to_be32(wr->remote_qkey);
+}
+
+int mthca_tavor_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
+			  const struct ib_send_wr **bad_wr)
+{
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	struct mthca_qp *qp = to_mqp(ibqp);
+	void *wqe;
+	void *prev_wqe;
+	unsigned long flags;
+	int err = 0;
+	int nreq;
+	int i;
+	int size;
+	/*
+	 * f0 and size0 are only used if nreq != 0, and they will
+	 * always be initialized the first time through the main loop
+	 * before nreq is incremented.  So nreq cannot become non-zero
+	 * without initializing f0 and size0, and they are in fact
+	 * never used uninitialized.
+	 */
+	int size0;
+	u32 f0;
+	int ind;
+	u8 op0 = 0;
+
+	spin_lock_irqsave(&qp->sq.lock, flags);
+
+	/* XXX check that state is OK to post send */
+
+	ind = qp->sq.next_ind;
+
+	for (nreq = 0; wr; ++nreq, wr = wr->next) {
+		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
+			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
+					" %d max, %d nreq)\n", qp->qpn,
+					qp->sq.head, qp->sq.tail,
+					qp->sq.max, nreq);
+			err = -ENOMEM;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		wqe = get_send_wqe(qp, ind);
+		prev_wqe = qp->sq.last;
+		qp->sq.last = wqe;
+
+		((struct mthca_next_seg *) wqe)->nda_op = 0;
+		((struct mthca_next_seg *) wqe)->ee_nds = 0;
+		((struct mthca_next_seg *) wqe)->flags =
+			((wr->send_flags & IB_SEND_SIGNALED) ?
+			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
+			((wr->send_flags & IB_SEND_SOLICITED) ?
+			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
+			cpu_to_be32(1);
+		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
+		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
+			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
+
+		wqe += sizeof (struct mthca_next_seg);
+		size = sizeof (struct mthca_next_seg) / 16;
+
+		switch (qp->transport) {
+		case RC:
+			switch (wr->opcode) {
+			case IB_WR_ATOMIC_CMP_AND_SWP:
+			case IB_WR_ATOMIC_FETCH_AND_ADD:
+				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
+					      atomic_wr(wr)->rkey);
+				wqe += sizeof (struct mthca_raddr_seg);
+
+				set_atomic_seg(wqe, atomic_wr(wr));
+				wqe += sizeof (struct mthca_atomic_seg);
+				size += (sizeof (struct mthca_raddr_seg) +
+					 sizeof (struct mthca_atomic_seg)) / 16;
+				break;
+
+			case IB_WR_RDMA_WRITE:
+			case IB_WR_RDMA_WRITE_WITH_IMM:
+			case IB_WR_RDMA_READ:
+				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
+					      rdma_wr(wr)->rkey);
+				wqe  += sizeof (struct mthca_raddr_seg);
+				size += sizeof (struct mthca_raddr_seg) / 16;
+				break;
+
+			default:
+				/* No extra segments required for sends */
+				break;
+			}
+
+			break;
+
+		case UC:
+			switch (wr->opcode) {
+			case IB_WR_RDMA_WRITE:
+			case IB_WR_RDMA_WRITE_WITH_IMM:
+				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
+					      rdma_wr(wr)->rkey);
+				wqe  += sizeof (struct mthca_raddr_seg);
+				size += sizeof (struct mthca_raddr_seg) / 16;
+				break;
+
+			default:
+				/* No extra segments required for sends */
+				break;
+			}
+
+			break;
+
+		case UD:
+			set_tavor_ud_seg(wqe, ud_wr(wr));
+			wqe  += sizeof (struct mthca_tavor_ud_seg);
+			size += sizeof (struct mthca_tavor_ud_seg) / 16;
+			break;
+
+		case MLX:
+			err = build_mlx_header(
+				dev, qp, ind, ud_wr(wr),
+				wqe - sizeof(struct mthca_next_seg), wqe);
+			if (err) {
+				*bad_wr = wr;
+				goto out;
+			}
+			wqe += sizeof (struct mthca_data_seg);
+			size += sizeof (struct mthca_data_seg) / 16;
+			break;
+		}
+
+		if (wr->num_sge > qp->sq.max_gs) {
+			mthca_err(dev, "too many gathers\n");
+			err = -EINVAL;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		for (i = 0; i < wr->num_sge; ++i) {
+			mthca_set_data_seg(wqe, wr->sg_list + i);
+			wqe  += sizeof (struct mthca_data_seg);
+			size += sizeof (struct mthca_data_seg) / 16;
+		}
+
+		/* Add one more inline data segment for ICRC */
+		if (qp->transport == MLX) {
+			((struct mthca_data_seg *) wqe)->byte_count =
+				cpu_to_be32((1 << 31) | 4);
+			((u32 *) wqe)[1] = 0;
+			wqe += sizeof (struct mthca_data_seg);
+			size += sizeof (struct mthca_data_seg) / 16;
+		}
+
+		qp->wrid[ind + qp->rq.max] = wr->wr_id;
+
+		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
+			mthca_err(dev, "opcode invalid\n");
+			err = -EINVAL;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		((struct mthca_next_seg *) prev_wqe)->nda_op =
+			cpu_to_be32(((ind << qp->sq.wqe_shift) +
+				     qp->send_wqe_offset) |
+				    mthca_opcode[wr->opcode]);
+		wmb();
+		((struct mthca_next_seg *) prev_wqe)->ee_nds =
+			cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
+				    ((wr->send_flags & IB_SEND_FENCE) ?
+				    MTHCA_NEXT_FENCE : 0));
+
+		if (!nreq) {
+			size0 = size;
+			op0   = mthca_opcode[wr->opcode];
+			f0    = wr->send_flags & IB_SEND_FENCE ?
+				MTHCA_SEND_DOORBELL_FENCE : 0;
+		}
+
+		++ind;
+		if (unlikely(ind >= qp->sq.max))
+			ind -= qp->sq.max;
+	}
+
+out:
+	if (likely(nreq)) {
+		wmb();
+
+		mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
+			       qp->send_wqe_offset) | f0 | op0,
+			      (qp->qpn << 8) | size0,
+			      dev->kar + MTHCA_SEND_DOORBELL,
+			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+	}
+
+	qp->sq.next_ind = ind;
+	qp->sq.head    += nreq;
+
+	spin_unlock_irqrestore(&qp->sq.lock, flags);
+	return err;
+}
+
+int mthca_tavor_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
+			     const struct ib_recv_wr **bad_wr)
+{
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	struct mthca_qp *qp = to_mqp(ibqp);
+	unsigned long flags;
+	int err = 0;
+	int nreq;
+	int i;
+	int size;
+	/*
+	 * size0 is only used if nreq != 0, and it will always be
+	 * initialized the first time through the main loop before
+	 * nreq is incremented.  So nreq cannot become non-zero
+	 * without initializing size0, and it is in fact never used
+	 * uninitialized.
+	 */
+	int size0;
+	int ind;
+	void *wqe;
+	void *prev_wqe;
+
+	spin_lock_irqsave(&qp->rq.lock, flags);
+
+	/* XXX check that state is OK to post receive */
+
+	ind = qp->rq.next_ind;
+
+	for (nreq = 0; wr; wr = wr->next) {
+		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
+			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
+					" %d max, %d nreq)\n", qp->qpn,
+					qp->rq.head, qp->rq.tail,
+					qp->rq.max, nreq);
+			err = -ENOMEM;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		wqe = get_recv_wqe(qp, ind);
+		prev_wqe = qp->rq.last;
+		qp->rq.last = wqe;
+
+		((struct mthca_next_seg *) wqe)->ee_nds =
+			cpu_to_be32(MTHCA_NEXT_DBD);
+		((struct mthca_next_seg *) wqe)->flags = 0;
+
+		wqe += sizeof (struct mthca_next_seg);
+		size = sizeof (struct mthca_next_seg) / 16;
+
+		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		for (i = 0; i < wr->num_sge; ++i) {
+			mthca_set_data_seg(wqe, wr->sg_list + i);
+			wqe  += sizeof (struct mthca_data_seg);
+			size += sizeof (struct mthca_data_seg) / 16;
+		}
+
+		qp->wrid[ind] = wr->wr_id;
+
+		((struct mthca_next_seg *) prev_wqe)->ee_nds =
+			cpu_to_be32(MTHCA_NEXT_DBD | size);
+
+		if (!nreq)
+			size0 = size;
+
+		++ind;
+		if (unlikely(ind >= qp->rq.max))
+			ind -= qp->rq.max;
+
+		++nreq;
+		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
+			nreq = 0;
+
+			wmb();
+
+			mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
+				      qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
+				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+
+			qp->rq.next_ind = ind;
+			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
+		}
+	}
+
+out:
+	if (likely(nreq)) {
+		wmb();
+
+		mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
+			      qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
+			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+	}
+
+	qp->rq.next_ind = ind;
+	qp->rq.head    += nreq;
+
+	spin_unlock_irqrestore(&qp->rq.lock, flags);
+	return err;
+}
+
+int mthca_arbel_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
+			  const struct ib_send_wr **bad_wr)
+{
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	struct mthca_qp *qp = to_mqp(ibqp);
+	u32 dbhi;
+	void *wqe;
+	void *prev_wqe;
+	unsigned long flags;
+	int err = 0;
+	int nreq;
+	int i;
+	int size;
+	/*
+	 * f0 and size0 are only used if nreq != 0, and they will
+	 * always be initialized the first time through the main loop
+	 * before nreq is incremented.  So nreq cannot become non-zero
+	 * without initializing f0 and size0, and they are in fact
+	 * never used uninitialized.
+	 */
+	int size0;
+	u32 f0;
+	int ind;
+	u8 op0 = 0;
+
+	spin_lock_irqsave(&qp->sq.lock, flags);
+
+	/* XXX check that state is OK to post send */
+
+	ind = qp->sq.head & (qp->sq.max - 1);
+
+	for (nreq = 0; wr; ++nreq, wr = wr->next) {
+		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
+			nreq = 0;
+
+			dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
+				((qp->sq.head & 0xffff) << 8) | f0 | op0;
+
+			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
+
+			/*
+			 * Make sure that descriptors are written before
+			 * doorbell record.
+			 */
+			wmb();
+			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
+
+			/*
+			 * Make sure doorbell record is written before we
+			 * write MMIO send doorbell.
+			 */
+			wmb();
+
+			mthca_write64(dbhi, (qp->qpn << 8) | size0,
+				      dev->kar + MTHCA_SEND_DOORBELL,
+				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+		}
+
+		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
+			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
+					" %d max, %d nreq)\n", qp->qpn,
+					qp->sq.head, qp->sq.tail,
+					qp->sq.max, nreq);
+			err = -ENOMEM;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		wqe = get_send_wqe(qp, ind);
+		prev_wqe = qp->sq.last;
+		qp->sq.last = wqe;
+
+		((struct mthca_next_seg *) wqe)->flags =
+			((wr->send_flags & IB_SEND_SIGNALED) ?
+			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
+			((wr->send_flags & IB_SEND_SOLICITED) ?
+			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
+			((wr->send_flags & IB_SEND_IP_CSUM) ?
+			 cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) |
+			cpu_to_be32(1);
+		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
+		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
+			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
+
+		wqe += sizeof (struct mthca_next_seg);
+		size = sizeof (struct mthca_next_seg) / 16;
+
+		switch (qp->transport) {
+		case RC:
+			switch (wr->opcode) {
+			case IB_WR_ATOMIC_CMP_AND_SWP:
+			case IB_WR_ATOMIC_FETCH_AND_ADD:
+				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
+					      atomic_wr(wr)->rkey);
+				wqe += sizeof (struct mthca_raddr_seg);
+
+				set_atomic_seg(wqe, atomic_wr(wr));
+				wqe  += sizeof (struct mthca_atomic_seg);
+				size += (sizeof (struct mthca_raddr_seg) +
+					 sizeof (struct mthca_atomic_seg)) / 16;
+				break;
+
+			case IB_WR_RDMA_READ:
+			case IB_WR_RDMA_WRITE:
+			case IB_WR_RDMA_WRITE_WITH_IMM:
+				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
+					      rdma_wr(wr)->rkey);
+				wqe  += sizeof (struct mthca_raddr_seg);
+				size += sizeof (struct mthca_raddr_seg) / 16;
+				break;
+
+			default:
+				/* No extra segments required for sends */
+				break;
+			}
+
+			break;
+
+		case UC:
+			switch (wr->opcode) {
+			case IB_WR_RDMA_WRITE:
+			case IB_WR_RDMA_WRITE_WITH_IMM:
+				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
+					      rdma_wr(wr)->rkey);
+				wqe  += sizeof (struct mthca_raddr_seg);
+				size += sizeof (struct mthca_raddr_seg) / 16;
+				break;
+
+			default:
+				/* No extra segments required for sends */
+				break;
+			}
+
+			break;
+
+		case UD:
+			set_arbel_ud_seg(wqe, ud_wr(wr));
+			wqe  += sizeof (struct mthca_arbel_ud_seg);
+			size += sizeof (struct mthca_arbel_ud_seg) / 16;
+			break;
+
+		case MLX:
+			err = build_mlx_header(
+				dev, qp, ind, ud_wr(wr),
+				wqe - sizeof(struct mthca_next_seg), wqe);
+			if (err) {
+				*bad_wr = wr;
+				goto out;
+			}
+			wqe += sizeof (struct mthca_data_seg);
+			size += sizeof (struct mthca_data_seg) / 16;
+			break;
+		}
+
+		if (wr->num_sge > qp->sq.max_gs) {
+			mthca_err(dev, "too many gathers\n");
+			err = -EINVAL;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		for (i = 0; i < wr->num_sge; ++i) {
+			mthca_set_data_seg(wqe, wr->sg_list + i);
+			wqe  += sizeof (struct mthca_data_seg);
+			size += sizeof (struct mthca_data_seg) / 16;
+		}
+
+		/* Add one more inline data segment for ICRC */
+		if (qp->transport == MLX) {
+			((struct mthca_data_seg *) wqe)->byte_count =
+				cpu_to_be32((1 << 31) | 4);
+			((u32 *) wqe)[1] = 0;
+			wqe += sizeof (struct mthca_data_seg);
+			size += sizeof (struct mthca_data_seg) / 16;
+		}
+
+		qp->wrid[ind + qp->rq.max] = wr->wr_id;
+
+		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
+			mthca_err(dev, "opcode invalid\n");
+			err = -EINVAL;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		((struct mthca_next_seg *) prev_wqe)->nda_op =
+			cpu_to_be32(((ind << qp->sq.wqe_shift) +
+				     qp->send_wqe_offset) |
+				    mthca_opcode[wr->opcode]);
+		wmb();
+		((struct mthca_next_seg *) prev_wqe)->ee_nds =
+			cpu_to_be32(MTHCA_NEXT_DBD | size |
+				    ((wr->send_flags & IB_SEND_FENCE) ?
+				     MTHCA_NEXT_FENCE : 0));
+
+		if (!nreq) {
+			size0 = size;
+			op0   = mthca_opcode[wr->opcode];
+			f0    = wr->send_flags & IB_SEND_FENCE ?
+				MTHCA_SEND_DOORBELL_FENCE : 0;
+		}
+
+		++ind;
+		if (unlikely(ind >= qp->sq.max))
+			ind -= qp->sq.max;
+	}
+
+out:
+	if (likely(nreq)) {
+		dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
+
+		qp->sq.head += nreq;
+
+		/*
+		 * Make sure that descriptors are written before
+		 * doorbell record.
+		 */
+		wmb();
+		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
+
+		/*
+		 * Make sure doorbell record is written before we
+		 * write MMIO send doorbell.
+		 */
+		wmb();
+
+		mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
+			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+	}
+
+	spin_unlock_irqrestore(&qp->sq.lock, flags);
+	return err;
+}
+
+int mthca_arbel_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
+			     const struct ib_recv_wr **bad_wr)
+{
+	struct mthca_dev *dev = to_mdev(ibqp->device);
+	struct mthca_qp *qp = to_mqp(ibqp);
+	unsigned long flags;
+	int err = 0;
+	int nreq;
+	int ind;
+	int i;
+	void *wqe;
+
+	spin_lock_irqsave(&qp->rq.lock, flags);
+
+	/* XXX check that state is OK to post receive */
+
+	ind = qp->rq.head & (qp->rq.max - 1);
+
+	for (nreq = 0; wr; ++nreq, wr = wr->next) {
+		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
+			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
+					" %d max, %d nreq)\n", qp->qpn,
+					qp->rq.head, qp->rq.tail,
+					qp->rq.max, nreq);
+			err = -ENOMEM;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		wqe = get_recv_wqe(qp, ind);
+
+		((struct mthca_next_seg *) wqe)->flags = 0;
+
+		wqe += sizeof (struct mthca_next_seg);
+
+		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			goto out;
+		}
+
+		for (i = 0; i < wr->num_sge; ++i) {
+			mthca_set_data_seg(wqe, wr->sg_list + i);
+			wqe += sizeof (struct mthca_data_seg);
+		}
+
+		if (i < qp->rq.max_gs)
+			mthca_set_data_seg_inval(wqe);
+
+		qp->wrid[ind] = wr->wr_id;
+
+		++ind;
+		if (unlikely(ind >= qp->rq.max))
+			ind -= qp->rq.max;
+	}
+out:
+	if (likely(nreq)) {
+		qp->rq.head += nreq;
+
+		/*
+		 * Make sure that descriptors are written before
+		 * doorbell record.
+		 */
+		wmb();
+		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
+	}
+
+	spin_unlock_irqrestore(&qp->rq.lock, flags);
+	return err;
+}
+
+void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
+			int index, int *dbd, __be32 *new_wqe)
+{
+	struct mthca_next_seg *next;
+
+	/*
+	 * For SRQs, all receive WQEs generate a CQE, so we're always
+	 * at the end of the doorbell chain.
+	 */
+	if (qp->ibqp.srq && !is_send) {
+		*new_wqe = 0;
+		return;
+	}
+
+	if (is_send)
+		next = get_send_wqe(qp, index);
+	else
+		next = get_recv_wqe(qp, index);
+
+	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
+	if (next->ee_nds & cpu_to_be32(0x3f))
+		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
+			(next->ee_nds & cpu_to_be32(0x3f));
+	else
+		*new_wqe = 0;
+}
+
+int mthca_init_qp_table(struct mthca_dev *dev)
+{
+	int err;
+	int i;
+
+	spin_lock_init(&dev->qp_table.lock);
+
+	/*
+	 * We reserve 2 extra QPs per port for the special QPs.  The
+	 * special QP for port 1 has to be even, so round up.
+	 */
+	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
+	err = mthca_alloc_init(&dev->qp_table.alloc,
+			       dev->limits.num_qps,
+			       (1 << 24) - 1,
+			       dev->qp_table.sqp_start +
+			       MTHCA_MAX_PORTS * 2);
+	if (err)
+		return err;
+
+	err = mthca_array_init(&dev->qp_table.qp,
+			       dev->limits.num_qps);
+	if (err) {
+		mthca_alloc_cleanup(&dev->qp_table.alloc);
+		return err;
+	}
+
+	for (i = 0; i < 2; ++i) {
+		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
+				    dev->qp_table.sqp_start + i * 2);
+		if (err) {
+			mthca_warn(dev, "CONF_SPECIAL_QP returned "
+				   "%d, aborting.\n", err);
+			goto err_out;
+		}
+	}
+	return 0;
+
+ err_out:
+	for (i = 0; i < 2; ++i)
+		mthca_CONF_SPECIAL_QP(dev, i, 0);
+
+	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
+	mthca_alloc_cleanup(&dev->qp_table.alloc);
+
+	return err;
+}
+
+void mthca_cleanup_qp_table(struct mthca_dev *dev)
+{
+	int i;
+
+	for (i = 0; i < 2; ++i)
+		mthca_CONF_SPECIAL_QP(dev, i, 0);
+
+	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
+	mthca_alloc_cleanup(&dev->qp_table.alloc);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_reset.c b/drivers/infiniband/hw/mthca/mthca_reset.c
new file mode 100644
index 000000000..2a6979e4a
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_reset.c
@@ -0,0 +1,282 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/errno.h>
+#include <linux/pci.h>
+#include <linux/delay.h>
+#include <linux/slab.h>
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+
+int mthca_reset(struct mthca_dev *mdev)
+{
+	int i;
+	int err = 0;
+	u32 *hca_header    = NULL;
+	u32 *bridge_header = NULL;
+	struct pci_dev *bridge = NULL;
+	int bridge_pcix_cap = 0;
+	int hca_pcie_cap = 0;
+	int hca_pcix_cap = 0;
+
+	u16 devctl;
+	u16 linkctl;
+
+#define MTHCA_RESET_OFFSET 0xf0010
+#define MTHCA_RESET_VALUE  swab32(1)
+
+	/*
+	 * Reset the chip.  This is somewhat ugly because we have to
+	 * save off the PCI header before reset and then restore it
+	 * after the chip reboots.  We skip config space offsets 22
+	 * and 23 since those have a special meaning.
+	 *
+	 * To make matters worse, for Tavor (PCI-X HCA) we have to
+	 * find the associated bridge device and save off its PCI
+	 * header as well.
+	 */
+
+	if (!(mdev->mthca_flags & MTHCA_FLAG_PCIE)) {
+		/* Look for the bridge -- its device ID will be 2 more
+		   than HCA's device ID. */
+		while ((bridge = pci_get_device(mdev->pdev->vendor,
+						mdev->pdev->device + 2,
+						bridge)) != NULL) {
+			if (bridge->hdr_type    == PCI_HEADER_TYPE_BRIDGE &&
+			    bridge->subordinate == mdev->pdev->bus) {
+				mthca_dbg(mdev, "Found bridge: %s\n",
+					  pci_name(bridge));
+				break;
+			}
+		}
+
+		if (!bridge) {
+			/*
+			 * Didn't find a bridge for a Tavor device --
+			 * assume we're in no-bridge mode and hope for
+			 * the best.
+			 */
+			mthca_warn(mdev, "No bridge found for %s\n",
+				  pci_name(mdev->pdev));
+		}
+
+	}
+
+	/* For Arbel do we need to save off the full 4K PCI Express header?? */
+	hca_header = kmalloc(256, GFP_KERNEL);
+	if (!hca_header) {
+		err = -ENOMEM;
+		goto put_dev;
+	}
+
+	for (i = 0; i < 64; ++i) {
+		if (i == 22 || i == 23)
+			continue;
+		if (pci_read_config_dword(mdev->pdev, i * 4, hca_header + i)) {
+			err = -ENODEV;
+			mthca_err(mdev, "Couldn't save HCA "
+				  "PCI header, aborting.\n");
+			goto free_hca;
+		}
+	}
+
+	hca_pcix_cap = pci_find_capability(mdev->pdev, PCI_CAP_ID_PCIX);
+	hca_pcie_cap = pci_pcie_cap(mdev->pdev);
+
+	if (bridge) {
+		bridge_header = kmalloc(256, GFP_KERNEL);
+		if (!bridge_header) {
+			err = -ENOMEM;
+			goto free_hca;
+		}
+
+		for (i = 0; i < 64; ++i) {
+			if (i == 22 || i == 23)
+				continue;
+			if (pci_read_config_dword(bridge, i * 4, bridge_header + i)) {
+				err = -ENODEV;
+				mthca_err(mdev, "Couldn't save HCA bridge "
+					  "PCI header, aborting.\n");
+				goto free_bh;
+			}
+		}
+		bridge_pcix_cap = pci_find_capability(bridge, PCI_CAP_ID_PCIX);
+		if (!bridge_pcix_cap) {
+				err = -ENODEV;
+				mthca_err(mdev, "Couldn't locate HCA bridge "
+					  "PCI-X capability, aborting.\n");
+				goto free_bh;
+		}
+	}
+
+	/* actually hit reset */
+	{
+		void __iomem *reset = ioremap(pci_resource_start(mdev->pdev, 0) +
+					      MTHCA_RESET_OFFSET, 4);
+
+		if (!reset) {
+			err = -ENOMEM;
+			mthca_err(mdev, "Couldn't map HCA reset register, "
+				  "aborting.\n");
+			goto free_bh;
+		}
+
+		writel(MTHCA_RESET_VALUE, reset);
+		iounmap(reset);
+	}
+
+	/* Docs say to wait one second before accessing device */
+	msleep(1000);
+
+	/* Now wait for PCI device to start responding again */
+	{
+		u32 v;
+		int c = 0;
+
+		for (c = 0; c < 100; ++c) {
+			if (pci_read_config_dword(bridge ? bridge : mdev->pdev, 0, &v)) {
+				err = -ENODEV;
+				mthca_err(mdev, "Couldn't access HCA after reset, "
+					  "aborting.\n");
+				goto free_bh;
+			}
+
+			if (v != 0xffffffff)
+				goto good;
+
+			msleep(100);
+		}
+
+		err = -ENODEV;
+		mthca_err(mdev, "PCI device did not come back after reset, "
+			  "aborting.\n");
+		goto free_bh;
+	}
+
+good:
+	/* Now restore the PCI headers */
+	if (bridge) {
+		if (pci_write_config_dword(bridge, bridge_pcix_cap + 0x8,
+				 bridge_header[(bridge_pcix_cap + 0x8) / 4])) {
+			err = -ENODEV;
+			mthca_err(mdev, "Couldn't restore HCA bridge Upstream "
+				  "split transaction control, aborting.\n");
+			goto free_bh;
+		}
+		if (pci_write_config_dword(bridge, bridge_pcix_cap + 0xc,
+				 bridge_header[(bridge_pcix_cap + 0xc) / 4])) {
+			err = -ENODEV;
+			mthca_err(mdev, "Couldn't restore HCA bridge Downstream "
+				  "split transaction control, aborting.\n");
+			goto free_bh;
+		}
+		/*
+		 * Bridge control register is at 0x3e, so we'll
+		 * naturally restore it last in this loop.
+		 */
+		for (i = 0; i < 16; ++i) {
+			if (i * 4 == PCI_COMMAND)
+				continue;
+
+			if (pci_write_config_dword(bridge, i * 4, bridge_header[i])) {
+				err = -ENODEV;
+				mthca_err(mdev, "Couldn't restore HCA bridge reg %x, "
+					  "aborting.\n", i);
+				goto free_bh;
+			}
+		}
+
+		if (pci_write_config_dword(bridge, PCI_COMMAND,
+					   bridge_header[PCI_COMMAND / 4])) {
+			err = -ENODEV;
+			mthca_err(mdev, "Couldn't restore HCA bridge COMMAND, "
+				  "aborting.\n");
+			goto free_bh;
+		}
+	}
+
+	if (hca_pcix_cap) {
+		if (pci_write_config_dword(mdev->pdev, hca_pcix_cap,
+				 hca_header[hca_pcix_cap / 4])) {
+			err = -ENODEV;
+			mthca_err(mdev, "Couldn't restore HCA PCI-X "
+				  "command register, aborting.\n");
+			goto free_bh;
+		}
+	}
+
+	if (hca_pcie_cap) {
+		devctl = hca_header[(hca_pcie_cap + PCI_EXP_DEVCTL) / 4];
+		if (pcie_capability_write_word(mdev->pdev, PCI_EXP_DEVCTL,
+					       devctl)) {
+			err = -ENODEV;
+			mthca_err(mdev, "Couldn't restore HCA PCI Express "
+				  "Device Control register, aborting.\n");
+			goto free_bh;
+		}
+		linkctl = hca_header[(hca_pcie_cap + PCI_EXP_LNKCTL) / 4];
+		if (pcie_capability_write_word(mdev->pdev, PCI_EXP_LNKCTL,
+					       linkctl)) {
+			err = -ENODEV;
+			mthca_err(mdev, "Couldn't restore HCA PCI Express "
+				  "Link control register, aborting.\n");
+			goto free_bh;
+		}
+	}
+
+	for (i = 0; i < 16; ++i) {
+		if (i * 4 == PCI_COMMAND)
+			continue;
+
+		if (pci_write_config_dword(mdev->pdev, i * 4, hca_header[i])) {
+			err = -ENODEV;
+			mthca_err(mdev, "Couldn't restore HCA reg %x, "
+				  "aborting.\n", i);
+			goto free_bh;
+		}
+	}
+
+	if (pci_write_config_dword(mdev->pdev, PCI_COMMAND,
+				   hca_header[PCI_COMMAND / 4])) {
+		err = -ENODEV;
+		mthca_err(mdev, "Couldn't restore HCA COMMAND, "
+			  "aborting.\n");
+	}
+free_bh:
+	kfree(bridge_header);
+free_hca:
+	kfree(hca_header);
+put_dev:
+	pci_dev_put(bridge);
+	return err;
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_srq.c b/drivers/infiniband/hw/mthca/mthca_srq.c
new file mode 100644
index 000000000..a85935ccc
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_srq.c
@@ -0,0 +1,698 @@
+/*
+ * Copyright (c) 2005 Cisco Systems. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/slab.h>
+#include <linux/string.h>
+#include <linux/sched.h>
+
+#include <asm/io.h>
+
+#include <rdma/uverbs_ioctl.h>
+
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+#include "mthca_memfree.h"
+#include "mthca_wqe.h"
+
+enum {
+	MTHCA_MAX_DIRECT_SRQ_SIZE = 4 * PAGE_SIZE
+};
+
+struct mthca_tavor_srq_context {
+	__be64 wqe_base_ds;	/* low 6 bits is descriptor size */
+	__be32 state_pd;
+	__be32 lkey;
+	__be32 uar;
+	__be16 limit_watermark;
+	__be16 wqe_cnt;
+	u32    reserved[2];
+};
+
+struct mthca_arbel_srq_context {
+	__be32 state_logsize_srqn;
+	__be32 lkey;
+	__be32 db_index;
+	__be32 logstride_usrpage;
+	__be64 wqe_base;
+	__be32 eq_pd;
+	__be16 limit_watermark;
+	__be16 wqe_cnt;
+	u16    reserved1;
+	__be16 wqe_counter;
+	u32    reserved2[3];
+};
+
+static void *get_wqe(struct mthca_srq *srq, int n)
+{
+	if (srq->is_direct)
+		return srq->queue.direct.buf + (n << srq->wqe_shift);
+	else
+		return srq->queue.page_list[(n << srq->wqe_shift) >> PAGE_SHIFT].buf +
+			((n << srq->wqe_shift) & (PAGE_SIZE - 1));
+}
+
+/*
+ * Return a pointer to the location within a WQE that we're using as a
+ * link when the WQE is in the free list.  We use the imm field
+ * because in the Tavor case, posting a WQE may overwrite the next
+ * segment of the previous WQE, but a receive WQE will never touch the
+ * imm field.  This avoids corrupting our free list if the previous
+ * WQE has already completed and been put on the free list when we
+ * post the next WQE.
+ */
+static inline int *wqe_to_link(void *wqe)
+{
+	return (int *) (wqe + offsetof(struct mthca_next_seg, imm));
+}
+
+static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
+					 struct mthca_pd *pd,
+					 struct mthca_srq *srq,
+					 struct mthca_tavor_srq_context *context,
+					 struct ib_udata *udata)
+{
+	struct mthca_ucontext *ucontext = rdma_udata_to_drv_context(
+		udata, struct mthca_ucontext, ibucontext);
+
+	memset(context, 0, sizeof *context);
+
+	context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
+	context->state_pd    = cpu_to_be32(pd->pd_num);
+	context->lkey        = cpu_to_be32(srq->mr.ibmr.lkey);
+
+	if (udata)
+		context->uar = cpu_to_be32(ucontext->uar.index);
+	else
+		context->uar = cpu_to_be32(dev->driver_uar.index);
+}
+
+static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
+					 struct mthca_pd *pd,
+					 struct mthca_srq *srq,
+					 struct mthca_arbel_srq_context *context,
+					 struct ib_udata *udata)
+{
+	struct mthca_ucontext *ucontext = rdma_udata_to_drv_context(
+		udata, struct mthca_ucontext, ibucontext);
+	int logsize, max;
+
+	memset(context, 0, sizeof *context);
+
+	/*
+	 * Put max in a temporary variable to work around gcc bug
+	 * triggered by ilog2() on sparc64.
+	 */
+	max = srq->max;
+	logsize = ilog2(max);
+	context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
+	context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
+	context->db_index = cpu_to_be32(srq->db_index);
+	context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
+	if (udata)
+		context->logstride_usrpage |= cpu_to_be32(ucontext->uar.index);
+	else
+		context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
+	context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
+}
+
+static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
+{
+	mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
+		       srq->is_direct, &srq->mr);
+	kfree(srq->wrid);
+}
+
+static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
+			       struct mthca_srq *srq, struct ib_udata *udata)
+{
+	struct mthca_data_seg *scatter;
+	void *wqe;
+	int err;
+	int i;
+
+	if (udata)
+		return 0;
+
+	srq->wrid = kmalloc_array(srq->max, sizeof(u64), GFP_KERNEL);
+	if (!srq->wrid)
+		return -ENOMEM;
+
+	err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
+			      MTHCA_MAX_DIRECT_SRQ_SIZE,
+			      &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
+	if (err) {
+		kfree(srq->wrid);
+		return err;
+	}
+
+	/*
+	 * Now initialize the SRQ buffer so that all of the WQEs are
+	 * linked into the list of free WQEs.  In addition, set the
+	 * scatter list L_Keys to the sentry value of 0x100.
+	 */
+	for (i = 0; i < srq->max; ++i) {
+		struct mthca_next_seg *next;
+
+		next = wqe = get_wqe(srq, i);
+
+		if (i < srq->max - 1) {
+			*wqe_to_link(wqe) = i + 1;
+			next->nda_op = htonl(((i + 1) << srq->wqe_shift) | 1);
+		} else {
+			*wqe_to_link(wqe) = -1;
+			next->nda_op = 0;
+		}
+
+		for (scatter = wqe + sizeof (struct mthca_next_seg);
+		     (void *) scatter < wqe + (1 << srq->wqe_shift);
+		     ++scatter)
+			scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
+	}
+
+	srq->last = get_wqe(srq, srq->max - 1);
+
+	return 0;
+}
+
+int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
+		    struct ib_srq_attr *attr, struct mthca_srq *srq,
+		    struct ib_udata *udata)
+{
+	struct mthca_mailbox *mailbox;
+	int ds;
+	int err;
+
+	/* Sanity check SRQ size before proceeding */
+	if (attr->max_wr  > dev->limits.max_srq_wqes ||
+	    attr->max_sge > dev->limits.max_srq_sge)
+		return -EINVAL;
+
+	srq->max      = attr->max_wr;
+	srq->max_gs   = attr->max_sge;
+	srq->counter  = 0;
+
+	if (mthca_is_memfree(dev))
+		srq->max = roundup_pow_of_two(srq->max + 1);
+	else
+		srq->max = srq->max + 1;
+
+	ds = max(64UL,
+		 roundup_pow_of_two(sizeof (struct mthca_next_seg) +
+				    srq->max_gs * sizeof (struct mthca_data_seg)));
+
+	if (!mthca_is_memfree(dev) && (ds > dev->limits.max_desc_sz))
+		return -EINVAL;
+
+	srq->wqe_shift = ilog2(ds);
+
+	srq->srqn = mthca_alloc(&dev->srq_table.alloc);
+	if (srq->srqn == -1)
+		return -ENOMEM;
+
+	if (mthca_is_memfree(dev)) {
+		err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
+		if (err)
+			goto err_out;
+
+		if (!udata) {
+			srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
+						       srq->srqn, &srq->db);
+			if (srq->db_index < 0) {
+				err = -ENOMEM;
+				goto err_out_icm;
+			}
+		}
+	}
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		err = PTR_ERR(mailbox);
+		goto err_out_db;
+	}
+
+	err = mthca_alloc_srq_buf(dev, pd, srq, udata);
+	if (err)
+		goto err_out_mailbox;
+
+	spin_lock_init(&srq->lock);
+	srq->refcount = 1;
+	init_waitqueue_head(&srq->wait);
+	mutex_init(&srq->mutex);
+
+	if (mthca_is_memfree(dev))
+		mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf, udata);
+	else
+		mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf, udata);
+
+	err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn);
+
+	if (err) {
+		mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
+		goto err_out_free_buf;
+	}
+
+	spin_lock_irq(&dev->srq_table.lock);
+	if (mthca_array_set(&dev->srq_table.srq,
+			    srq->srqn & (dev->limits.num_srqs - 1),
+			    srq)) {
+		spin_unlock_irq(&dev->srq_table.lock);
+		goto err_out_free_srq;
+	}
+	spin_unlock_irq(&dev->srq_table.lock);
+
+	mthca_free_mailbox(dev, mailbox);
+
+	srq->first_free = 0;
+	srq->last_free  = srq->max - 1;
+
+	attr->max_wr    = srq->max - 1;
+	attr->max_sge   = srq->max_gs;
+
+	return 0;
+
+err_out_free_srq:
+	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
+	if (err)
+		mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
+
+err_out_free_buf:
+	if (!udata)
+		mthca_free_srq_buf(dev, srq);
+
+err_out_mailbox:
+	mthca_free_mailbox(dev, mailbox);
+
+err_out_db:
+	if (!udata && mthca_is_memfree(dev))
+		mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
+
+err_out_icm:
+	mthca_table_put(dev, dev->srq_table.table, srq->srqn);
+
+err_out:
+	mthca_free(&dev->srq_table.alloc, srq->srqn);
+
+	return err;
+}
+
+static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
+{
+	int c;
+
+	spin_lock_irq(&dev->srq_table.lock);
+	c = srq->refcount;
+	spin_unlock_irq(&dev->srq_table.lock);
+
+	return c;
+}
+
+void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
+{
+	struct mthca_mailbox *mailbox;
+	int err;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
+		return;
+	}
+
+	err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn);
+	if (err)
+		mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
+
+	spin_lock_irq(&dev->srq_table.lock);
+	mthca_array_clear(&dev->srq_table.srq,
+			  srq->srqn & (dev->limits.num_srqs - 1));
+	--srq->refcount;
+	spin_unlock_irq(&dev->srq_table.lock);
+
+	wait_event(srq->wait, !get_srq_refcount(dev, srq));
+
+	if (!srq->ibsrq.uobject) {
+		mthca_free_srq_buf(dev, srq);
+		if (mthca_is_memfree(dev))
+			mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
+	}
+
+	mthca_table_put(dev, dev->srq_table.table, srq->srqn);
+	mthca_free(&dev->srq_table.alloc, srq->srqn);
+	mthca_free_mailbox(dev, mailbox);
+}
+
+int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
+		     enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
+{
+	struct mthca_dev *dev = to_mdev(ibsrq->device);
+	struct mthca_srq *srq = to_msrq(ibsrq);
+	int ret = 0;
+
+	/* We don't support resizing SRQs (yet?) */
+	if (attr_mask & IB_SRQ_MAX_WR)
+		return -EINVAL;
+
+	if (attr_mask & IB_SRQ_LIMIT) {
+		u32 max_wr = mthca_is_memfree(dev) ? srq->max - 1 : srq->max;
+		if (attr->srq_limit > max_wr)
+			return -EINVAL;
+
+		mutex_lock(&srq->mutex);
+		ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit);
+		mutex_unlock(&srq->mutex);
+	}
+
+	return ret;
+}
+
+int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
+{
+	struct mthca_dev *dev = to_mdev(ibsrq->device);
+	struct mthca_srq *srq = to_msrq(ibsrq);
+	struct mthca_mailbox *mailbox;
+	struct mthca_arbel_srq_context *arbel_ctx;
+	struct mthca_tavor_srq_context *tavor_ctx;
+	int err;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+
+	err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox);
+	if (err)
+		goto out;
+
+	if (mthca_is_memfree(dev)) {
+		arbel_ctx = mailbox->buf;
+		srq_attr->srq_limit = be16_to_cpu(arbel_ctx->limit_watermark);
+	} else {
+		tavor_ctx = mailbox->buf;
+		srq_attr->srq_limit = be16_to_cpu(tavor_ctx->limit_watermark);
+	}
+
+	srq_attr->max_wr  = srq->max - 1;
+	srq_attr->max_sge = srq->max_gs;
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+
+	return err;
+}
+
+void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
+		     enum ib_event_type event_type)
+{
+	struct mthca_srq *srq;
+	struct ib_event event;
+
+	spin_lock(&dev->srq_table.lock);
+	srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
+	if (srq)
+		++srq->refcount;
+	spin_unlock(&dev->srq_table.lock);
+
+	if (!srq) {
+		mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
+		return;
+	}
+
+	if (!srq->ibsrq.event_handler)
+		goto out;
+
+	event.device      = &dev->ib_dev;
+	event.event       = event_type;
+	event.element.srq = &srq->ibsrq;
+	srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
+
+out:
+	spin_lock(&dev->srq_table.lock);
+	if (!--srq->refcount)
+		wake_up(&srq->wait);
+	spin_unlock(&dev->srq_table.lock);
+}
+
+/*
+ * This function must be called with IRQs disabled.
+ */
+void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
+{
+	int ind;
+	struct mthca_next_seg *last_free;
+
+	ind = wqe_addr >> srq->wqe_shift;
+
+	spin_lock(&srq->lock);
+
+	last_free = get_wqe(srq, srq->last_free);
+	*wqe_to_link(last_free) = ind;
+	last_free->nda_op = htonl((ind << srq->wqe_shift) | 1);
+	*wqe_to_link(get_wqe(srq, ind)) = -1;
+	srq->last_free = ind;
+
+	spin_unlock(&srq->lock);
+}
+
+int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
+			      const struct ib_recv_wr **bad_wr)
+{
+	struct mthca_dev *dev = to_mdev(ibsrq->device);
+	struct mthca_srq *srq = to_msrq(ibsrq);
+	unsigned long flags;
+	int err = 0;
+	int first_ind;
+	int ind;
+	int next_ind;
+	int nreq;
+	int i;
+	void *wqe;
+	void *prev_wqe;
+
+	spin_lock_irqsave(&srq->lock, flags);
+
+	first_ind = srq->first_free;
+
+	for (nreq = 0; wr; wr = wr->next) {
+		ind       = srq->first_free;
+		wqe       = get_wqe(srq, ind);
+		next_ind  = *wqe_to_link(wqe);
+
+		if (unlikely(next_ind < 0)) {
+			mthca_err(dev, "SRQ %06x full\n", srq->srqn);
+			err = -ENOMEM;
+			*bad_wr = wr;
+			break;
+		}
+
+		prev_wqe  = srq->last;
+		srq->last = wqe;
+
+		((struct mthca_next_seg *) wqe)->ee_nds = 0;
+		/* flags field will always remain 0 */
+
+		wqe += sizeof (struct mthca_next_seg);
+
+		if (unlikely(wr->num_sge > srq->max_gs)) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			srq->last = prev_wqe;
+			break;
+		}
+
+		for (i = 0; i < wr->num_sge; ++i) {
+			mthca_set_data_seg(wqe, wr->sg_list + i);
+			wqe += sizeof (struct mthca_data_seg);
+		}
+
+		if (i < srq->max_gs)
+			mthca_set_data_seg_inval(wqe);
+
+		((struct mthca_next_seg *) prev_wqe)->ee_nds =
+			cpu_to_be32(MTHCA_NEXT_DBD);
+
+		srq->wrid[ind]  = wr->wr_id;
+		srq->first_free = next_ind;
+
+		++nreq;
+		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
+			nreq = 0;
+
+			/*
+			 * Make sure that descriptors are written
+			 * before doorbell is rung.
+			 */
+			wmb();
+
+			mthca_write64(first_ind << srq->wqe_shift, srq->srqn << 8,
+				      dev->kar + MTHCA_RECEIVE_DOORBELL,
+				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+
+			first_ind = srq->first_free;
+		}
+	}
+
+	if (likely(nreq)) {
+		/*
+		 * Make sure that descriptors are written before
+		 * doorbell is rung.
+		 */
+		wmb();
+
+		mthca_write64(first_ind << srq->wqe_shift, (srq->srqn << 8) | nreq,
+			      dev->kar + MTHCA_RECEIVE_DOORBELL,
+			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+	}
+
+	spin_unlock_irqrestore(&srq->lock, flags);
+	return err;
+}
+
+int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
+			      const struct ib_recv_wr **bad_wr)
+{
+	struct mthca_dev *dev = to_mdev(ibsrq->device);
+	struct mthca_srq *srq = to_msrq(ibsrq);
+	unsigned long flags;
+	int err = 0;
+	int ind;
+	int next_ind;
+	int nreq;
+	int i;
+	void *wqe;
+
+	spin_lock_irqsave(&srq->lock, flags);
+
+	for (nreq = 0; wr; ++nreq, wr = wr->next) {
+		ind       = srq->first_free;
+		wqe       = get_wqe(srq, ind);
+		next_ind  = *wqe_to_link(wqe);
+
+		if (unlikely(next_ind < 0)) {
+			mthca_err(dev, "SRQ %06x full\n", srq->srqn);
+			err = -ENOMEM;
+			*bad_wr = wr;
+			break;
+		}
+
+		((struct mthca_next_seg *) wqe)->ee_nds = 0;
+		/* flags field will always remain 0 */
+
+		wqe += sizeof (struct mthca_next_seg);
+
+		if (unlikely(wr->num_sge > srq->max_gs)) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			break;
+		}
+
+		for (i = 0; i < wr->num_sge; ++i) {
+			mthca_set_data_seg(wqe, wr->sg_list + i);
+			wqe += sizeof (struct mthca_data_seg);
+		}
+
+		if (i < srq->max_gs)
+			mthca_set_data_seg_inval(wqe);
+
+		srq->wrid[ind]  = wr->wr_id;
+		srq->first_free = next_ind;
+	}
+
+	if (likely(nreq)) {
+		srq->counter += nreq;
+
+		/*
+		 * Make sure that descriptors are written before
+		 * we write doorbell record.
+		 */
+		wmb();
+		*srq->db = cpu_to_be32(srq->counter);
+	}
+
+	spin_unlock_irqrestore(&srq->lock, flags);
+	return err;
+}
+
+int mthca_max_srq_sge(struct mthca_dev *dev)
+{
+	if (mthca_is_memfree(dev))
+		return dev->limits.max_sg;
+
+	/*
+	 * SRQ allocations are based on powers of 2 for Tavor,
+	 * (although they only need to be multiples of 16 bytes).
+	 *
+	 * Therefore, we need to base the max number of sg entries on
+	 * the largest power of 2 descriptor size that is <= to the
+	 * actual max WQE descriptor size, rather than return the
+	 * max_sg value given by the firmware (which is based on WQE
+	 * sizes as multiples of 16, not powers of 2).
+	 *
+	 * If SRQ implementation is changed for Tavor to be based on
+	 * multiples of 16, the calculation below can be deleted and
+	 * the FW max_sg value returned.
+	 */
+	return min_t(int, dev->limits.max_sg,
+		     ((1 << (fls(dev->limits.max_desc_sz) - 1)) -
+		      sizeof (struct mthca_next_seg)) /
+		     sizeof (struct mthca_data_seg));
+}
+
+int mthca_init_srq_table(struct mthca_dev *dev)
+{
+	int err;
+
+	if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
+		return 0;
+
+	spin_lock_init(&dev->srq_table.lock);
+
+	err = mthca_alloc_init(&dev->srq_table.alloc,
+			       dev->limits.num_srqs,
+			       dev->limits.num_srqs - 1,
+			       dev->limits.reserved_srqs);
+	if (err)
+		return err;
+
+	err = mthca_array_init(&dev->srq_table.srq,
+			       dev->limits.num_srqs);
+	if (err)
+		mthca_alloc_cleanup(&dev->srq_table.alloc);
+
+	return err;
+}
+
+void mthca_cleanup_srq_table(struct mthca_dev *dev)
+{
+	if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
+		return;
+
+	mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
+	mthca_alloc_cleanup(&dev->srq_table.alloc);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_uar.c b/drivers/infiniband/hw/mthca/mthca_uar.c
new file mode 100644
index 000000000..ca5900c96
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_uar.c
@@ -0,0 +1,78 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <asm/page.h>		/* PAGE_SHIFT */
+
+#include "mthca_dev.h"
+#include "mthca_memfree.h"
+
+int mthca_uar_alloc(struct mthca_dev *dev, struct mthca_uar *uar)
+{
+	uar->index = mthca_alloc(&dev->uar_table.alloc);
+	if (uar->index == -1)
+		return -ENOMEM;
+
+	uar->pfn = (pci_resource_start(dev->pdev, 2) >> PAGE_SHIFT) + uar->index;
+
+	return 0;
+}
+
+void mthca_uar_free(struct mthca_dev *dev, struct mthca_uar *uar)
+{
+	mthca_free(&dev->uar_table.alloc, uar->index);
+}
+
+int mthca_init_uar_table(struct mthca_dev *dev)
+{
+	int ret;
+
+	ret = mthca_alloc_init(&dev->uar_table.alloc,
+			       dev->limits.num_uars,
+			       dev->limits.num_uars - 1,
+			       dev->limits.reserved_uars + 1);
+	if (ret)
+		return ret;
+
+	ret = mthca_init_db_tab(dev);
+	if (ret)
+		mthca_alloc_cleanup(&dev->uar_table.alloc);
+
+	return ret;
+}
+
+void mthca_cleanup_uar_table(struct mthca_dev *dev)
+{
+	mthca_cleanup_db_tab(dev);
+
+	/* XXX check if any UARs are still allocated? */
+	mthca_alloc_cleanup(&dev->uar_table.alloc);
+}
diff --git a/drivers/infiniband/hw/mthca/mthca_wqe.h b/drivers/infiniband/hw/mthca/mthca_wqe.h
new file mode 100644
index 000000000..341a5ae88
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_wqe.h
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2005 Cisco Systems. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef MTHCA_WQE_H
+#define MTHCA_WQE_H
+
+#include <linux/types.h>
+
+enum {
+	MTHCA_NEXT_DBD		= 1 << 7,
+	MTHCA_NEXT_FENCE	= 1 << 6,
+	MTHCA_NEXT_CQ_UPDATE	= 1 << 3,
+	MTHCA_NEXT_EVENT_GEN	= 1 << 2,
+	MTHCA_NEXT_SOLICIT	= 1 << 1,
+	MTHCA_NEXT_IP_CSUM	= 1 << 4,
+	MTHCA_NEXT_TCP_UDP_CSUM = 1 << 5,
+
+	MTHCA_MLX_VL15		= 1 << 17,
+	MTHCA_MLX_SLR		= 1 << 16
+};
+
+enum {
+	MTHCA_INVAL_LKEY			= 0x100,
+	MTHCA_TAVOR_MAX_WQES_PER_RECV_DB	= 256,
+	MTHCA_ARBEL_MAX_WQES_PER_SEND_DB	= 255
+};
+
+struct mthca_next_seg {
+	__be32 nda_op;		/* [31:6] next WQE [4:0] next opcode */
+	__be32 ee_nds;		/* [31:8] next EE  [7] DBD [6] F [5:0] next WQE size */
+	__be32 flags;		/* [3] CQ [2] Event [1] Solicit */
+	__be32 imm;		/* immediate data */
+};
+
+struct mthca_tavor_ud_seg {
+	u32    reserved1;
+	__be32 lkey;
+	__be64 av_addr;
+	u32    reserved2[4];
+	__be32 dqpn;
+	__be32 qkey;
+	u32    reserved3[2];
+};
+
+struct mthca_arbel_ud_seg {
+	__be32 av[8];
+	__be32 dqpn;
+	__be32 qkey;
+	u32    reserved[2];
+};
+
+struct mthca_bind_seg {
+	__be32 flags;		/* [31] Atomic [30] rem write [29] rem read */
+	u32    reserved;
+	__be32 new_rkey;
+	__be32 lkey;
+	__be64 addr;
+	__be64 length;
+};
+
+struct mthca_raddr_seg {
+	__be64 raddr;
+	__be32 rkey;
+	u32    reserved;
+};
+
+struct mthca_atomic_seg {
+	__be64 swap_add;
+	__be64 compare;
+};
+
+struct mthca_data_seg {
+	__be32 byte_count;
+	__be32 lkey;
+	__be64 addr;
+};
+
+struct mthca_mlx_seg {
+	__be32 nda_op;
+	__be32 nds;
+	__be32 flags;		/* [17] VL15 [16] SLR [14:12] static rate
+				   [11:8] SL [3] C [2] E */
+	__be16 rlid;
+	__be16 vcrc;
+};
+
+static __always_inline void mthca_set_data_seg(struct mthca_data_seg *dseg,
+					       struct ib_sge *sg)
+{
+	dseg->byte_count = cpu_to_be32(sg->length);
+	dseg->lkey       = cpu_to_be32(sg->lkey);
+	dseg->addr       = cpu_to_be64(sg->addr);
+}
+
+static __always_inline void mthca_set_data_seg_inval(struct mthca_data_seg *dseg)
+{
+	dseg->byte_count = 0;
+	dseg->lkey       = cpu_to_be32(MTHCA_INVAL_LKEY);
+	dseg->addr       = 0;
+}
+
+#endif /* MTHCA_WQE_H */