1 files changed, 662 insertions, 0 deletions

diff --git a/drivers/crypto/ccree/cc_request_mgr.c b/drivers/crypto/ccree/cc_request_mgr.c
new file mode 100644
index 000000000..887162df5
--- /dev/null
+++ b/drivers/crypto/ccree/cc_request_mgr.c
@@ -0,0 +1,662 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (C) 2012-2019 ARM Limited (or its affiliates). */
+
+#include <linux/kernel.h>
+#include <linux/nospec.h>
+#include "cc_driver.h"
+#include "cc_buffer_mgr.h"
+#include "cc_request_mgr.h"
+#include "cc_pm.h"
+
+#define CC_MAX_POLL_ITER	10
+/* The highest descriptor count in used */
+#define CC_MAX_DESC_SEQ_LEN	23
+
+struct cc_req_mgr_handle {
+	/* Request manager resources */
+	unsigned int hw_queue_size; /* HW capability */
+	unsigned int min_free_hw_slots;
+	unsigned int max_used_sw_slots;
+	struct cc_crypto_req req_queue[MAX_REQUEST_QUEUE_SIZE];
+	u32 req_queue_head;
+	u32 req_queue_tail;
+	u32 axi_completed;
+	u32 q_free_slots;
+	/* This lock protects access to HW register
+	 * that must be single request at a time
+	 */
+	spinlock_t hw_lock;
+	struct cc_hw_desc compl_desc;
+	u8 *dummy_comp_buff;
+	dma_addr_t dummy_comp_buff_dma;
+
+	/* backlog queue */
+	struct list_head backlog;
+	unsigned int bl_len;
+	spinlock_t bl_lock; /* protect backlog queue */
+
+#ifdef COMP_IN_WQ
+	struct workqueue_struct *workq;
+	struct delayed_work compwork;
+#else
+	struct tasklet_struct comptask;
+#endif
+};
+
+struct cc_bl_item {
+	struct cc_crypto_req creq;
+	struct cc_hw_desc desc[CC_MAX_DESC_SEQ_LEN];
+	unsigned int len;
+	struct list_head list;
+	bool notif;
+};
+
+static const u32 cc_cpp_int_masks[CC_CPP_NUM_ALGS][CC_CPP_NUM_SLOTS] = {
+	{ BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_0_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_1_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_2_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_3_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_4_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_5_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_6_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_7_INT_BIT_SHIFT) },
+	{ BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_0_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_1_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_2_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_3_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_4_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_5_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_6_INT_BIT_SHIFT),
+	  BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_7_INT_BIT_SHIFT) }
+};
+
+static void comp_handler(unsigned long devarg);
+#ifdef COMP_IN_WQ
+static void comp_work_handler(struct work_struct *work);
+#endif
+
+static inline u32 cc_cpp_int_mask(enum cc_cpp_alg alg, int slot)
+{
+	alg = array_index_nospec(alg, CC_CPP_NUM_ALGS);
+	slot = array_index_nospec(slot, CC_CPP_NUM_SLOTS);
+
+	return cc_cpp_int_masks[alg][slot];
+}
+
+void cc_req_mgr_fini(struct cc_drvdata *drvdata)
+{
+	struct cc_req_mgr_handle *req_mgr_h = drvdata->request_mgr_handle;
+	struct device *dev = drvdata_to_dev(drvdata);
+
+	if (!req_mgr_h)
+		return; /* Not allocated */
+
+	if (req_mgr_h->dummy_comp_buff_dma) {
+		dma_free_coherent(dev, sizeof(u32), req_mgr_h->dummy_comp_buff,
+				  req_mgr_h->dummy_comp_buff_dma);
+	}
+
+	dev_dbg(dev, "max_used_hw_slots=%d\n", (req_mgr_h->hw_queue_size -
+						req_mgr_h->min_free_hw_slots));
+	dev_dbg(dev, "max_used_sw_slots=%d\n", req_mgr_h->max_used_sw_slots);
+
+#ifdef COMP_IN_WQ
+	destroy_workqueue(req_mgr_h->workq);
+#else
+	/* Kill tasklet */
+	tasklet_kill(&req_mgr_h->comptask);
+#endif
+	kfree_sensitive(req_mgr_h);
+	drvdata->request_mgr_handle = NULL;
+}
+
+int cc_req_mgr_init(struct cc_drvdata *drvdata)
+{
+	struct cc_req_mgr_handle *req_mgr_h;
+	struct device *dev = drvdata_to_dev(drvdata);
+	int rc = 0;
+
+	req_mgr_h = kzalloc(sizeof(*req_mgr_h), GFP_KERNEL);
+	if (!req_mgr_h) {
+		rc = -ENOMEM;
+		goto req_mgr_init_err;
+	}
+
+	drvdata->request_mgr_handle = req_mgr_h;
+
+	spin_lock_init(&req_mgr_h->hw_lock);
+	spin_lock_init(&req_mgr_h->bl_lock);
+	INIT_LIST_HEAD(&req_mgr_h->backlog);
+
+#ifdef COMP_IN_WQ
+	dev_dbg(dev, "Initializing completion workqueue\n");
+	req_mgr_h->workq = create_singlethread_workqueue("ccree");
+	if (!req_mgr_h->workq) {
+		dev_err(dev, "Failed creating work queue\n");
+		rc = -ENOMEM;
+		goto req_mgr_init_err;
+	}
+	INIT_DELAYED_WORK(&req_mgr_h->compwork, comp_work_handler);
+#else
+	dev_dbg(dev, "Initializing completion tasklet\n");
+	tasklet_init(&req_mgr_h->comptask, comp_handler,
+		     (unsigned long)drvdata);
+#endif
+	req_mgr_h->hw_queue_size = cc_ioread(drvdata,
+					     CC_REG(DSCRPTR_QUEUE_SRAM_SIZE));
+	dev_dbg(dev, "hw_queue_size=0x%08X\n", req_mgr_h->hw_queue_size);
+	if (req_mgr_h->hw_queue_size < MIN_HW_QUEUE_SIZE) {
+		dev_err(dev, "Invalid HW queue size = %u (Min. required is %u)\n",
+			req_mgr_h->hw_queue_size, MIN_HW_QUEUE_SIZE);
+		rc = -ENOMEM;
+		goto req_mgr_init_err;
+	}
+	req_mgr_h->min_free_hw_slots = req_mgr_h->hw_queue_size;
+	req_mgr_h->max_used_sw_slots = 0;
+
+	/* Allocate DMA word for "dummy" completion descriptor use */
+	req_mgr_h->dummy_comp_buff =
+		dma_alloc_coherent(dev, sizeof(u32),
+				   &req_mgr_h->dummy_comp_buff_dma,
+				   GFP_KERNEL);
+	if (!req_mgr_h->dummy_comp_buff) {
+		dev_err(dev, "Not enough memory to allocate DMA (%zu) dropped buffer\n",
+			sizeof(u32));
+		rc = -ENOMEM;
+		goto req_mgr_init_err;
+	}
+
+	/* Init. "dummy" completion descriptor */
+	hw_desc_init(&req_mgr_h->compl_desc);
+	set_din_const(&req_mgr_h->compl_desc, 0, sizeof(u32));
+	set_dout_dlli(&req_mgr_h->compl_desc, req_mgr_h->dummy_comp_buff_dma,
+		      sizeof(u32), NS_BIT, 1);
+	set_flow_mode(&req_mgr_h->compl_desc, BYPASS);
+	set_queue_last_ind(drvdata, &req_mgr_h->compl_desc);
+
+	return 0;
+
+req_mgr_init_err:
+	cc_req_mgr_fini(drvdata);
+	return rc;
+}
+
+static void enqueue_seq(struct cc_drvdata *drvdata, struct cc_hw_desc seq[],
+			unsigned int seq_len)
+{
+	int i, w;
+	void __iomem *reg = drvdata->cc_base + CC_REG(DSCRPTR_QUEUE_WORD0);
+	struct device *dev = drvdata_to_dev(drvdata);
+
+	/*
+	 * We do indeed write all 6 command words to the same
+	 * register. The HW supports this.
+	 */
+
+	for (i = 0; i < seq_len; i++) {
+		for (w = 0; w <= 5; w++)
+			writel_relaxed(seq[i].word[w], reg);
+
+		if (cc_dump_desc)
+			dev_dbg(dev, "desc[%02d]: 0x%08X 0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n",
+				i, seq[i].word[0], seq[i].word[1],
+				seq[i].word[2], seq[i].word[3],
+				seq[i].word[4], seq[i].word[5]);
+	}
+}
+
+/**
+ * request_mgr_complete() - Completion will take place if and only if user
+ * requested completion by cc_send_sync_request().
+ *
+ * @dev: Device pointer
+ * @dx_compl_h: The completion event to signal
+ * @dummy: unused error code
+ */
+static void request_mgr_complete(struct device *dev, void *dx_compl_h,
+				 int dummy)
+{
+	struct completion *this_compl = dx_compl_h;
+
+	complete(this_compl);
+}
+
+static int cc_queues_status(struct cc_drvdata *drvdata,
+			    struct cc_req_mgr_handle *req_mgr_h,
+			    unsigned int total_seq_len)
+{
+	unsigned long poll_queue;
+	struct device *dev = drvdata_to_dev(drvdata);
+
+	/* SW queue is checked only once as it will not
+	 * be changed during the poll because the spinlock_bh
+	 * is held by the thread
+	 */
+	if (((req_mgr_h->req_queue_head + 1) & (MAX_REQUEST_QUEUE_SIZE - 1)) ==
+	    req_mgr_h->req_queue_tail) {
+		dev_err(dev, "SW FIFO is full. req_queue_head=%d sw_fifo_len=%d\n",
+			req_mgr_h->req_queue_head, MAX_REQUEST_QUEUE_SIZE);
+		return -ENOSPC;
+	}
+
+	if (req_mgr_h->q_free_slots >= total_seq_len)
+		return 0;
+
+	/* Wait for space in HW queue. Poll constant num of iterations. */
+	for (poll_queue = 0; poll_queue < CC_MAX_POLL_ITER ; poll_queue++) {
+		req_mgr_h->q_free_slots =
+			cc_ioread(drvdata, CC_REG(DSCRPTR_QUEUE_CONTENT));
+		if (req_mgr_h->q_free_slots < req_mgr_h->min_free_hw_slots)
+			req_mgr_h->min_free_hw_slots = req_mgr_h->q_free_slots;
+
+		if (req_mgr_h->q_free_slots >= total_seq_len) {
+			/* If there is enough place return */
+			return 0;
+		}
+
+		dev_dbg(dev, "HW FIFO is full. q_free_slots=%d total_seq_len=%d\n",
+			req_mgr_h->q_free_slots, total_seq_len);
+	}
+	/* No room in the HW queue try again later */
+	dev_dbg(dev, "HW FIFO full, timeout. req_queue_head=%d sw_fifo_len=%d q_free_slots=%d total_seq_len=%d\n",
+		req_mgr_h->req_queue_head, MAX_REQUEST_QUEUE_SIZE,
+		req_mgr_h->q_free_slots, total_seq_len);
+	return -ENOSPC;
+}
+
+/**
+ * cc_do_send_request() - Enqueue caller request to crypto hardware.
+ * Need to be called with HW lock held and PM running
+ *
+ * @drvdata: Associated device driver context
+ * @cc_req: The request to enqueue
+ * @desc: The crypto sequence
+ * @len: The crypto sequence length
+ * @add_comp: If "true": add an artificial dout DMA to mark completion
+ *
+ */
+static void cc_do_send_request(struct cc_drvdata *drvdata,
+			       struct cc_crypto_req *cc_req,
+			       struct cc_hw_desc *desc, unsigned int len,
+			       bool add_comp)
+{
+	struct cc_req_mgr_handle *req_mgr_h = drvdata->request_mgr_handle;
+	unsigned int used_sw_slots;
+	unsigned int total_seq_len = len; /*initial sequence length*/
+	struct device *dev = drvdata_to_dev(drvdata);
+
+	used_sw_slots = ((req_mgr_h->req_queue_head -
+			  req_mgr_h->req_queue_tail) &
+			 (MAX_REQUEST_QUEUE_SIZE - 1));
+	if (used_sw_slots > req_mgr_h->max_used_sw_slots)
+		req_mgr_h->max_used_sw_slots = used_sw_slots;
+
+	/* Enqueue request - must be locked with HW lock*/
+	req_mgr_h->req_queue[req_mgr_h->req_queue_head] = *cc_req;
+	req_mgr_h->req_queue_head = (req_mgr_h->req_queue_head + 1) &
+				    (MAX_REQUEST_QUEUE_SIZE - 1);
+
+	dev_dbg(dev, "Enqueue request head=%u\n", req_mgr_h->req_queue_head);
+
+	/*
+	 * We are about to push command to the HW via the command registers
+	 * that may reference host memory. We need to issue a memory barrier
+	 * to make sure there are no outstanding memory writes
+	 */
+	wmb();
+
+	/* STAT_PHASE_4: Push sequence */
+
+	enqueue_seq(drvdata, desc, len);
+
+	if (add_comp) {
+		enqueue_seq(drvdata, &req_mgr_h->compl_desc, 1);
+		total_seq_len++;
+	}
+
+	if (req_mgr_h->q_free_slots < total_seq_len) {
+		/* This situation should never occur. Maybe indicating problem
+		 * with resuming power. Set the free slot count to 0 and hope
+		 * for the best.
+		 */
+		dev_err(dev, "HW free slot count mismatch.");
+		req_mgr_h->q_free_slots = 0;
+	} else {
+		/* Update the free slots in HW queue */
+		req_mgr_h->q_free_slots -= total_seq_len;
+	}
+}
+
+static void cc_enqueue_backlog(struct cc_drvdata *drvdata,
+			       struct cc_bl_item *bli)
+{
+	struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
+	struct device *dev = drvdata_to_dev(drvdata);
+
+	spin_lock_bh(&mgr->bl_lock);
+	list_add_tail(&bli->list, &mgr->backlog);
+	++mgr->bl_len;
+	dev_dbg(dev, "+++bl len: %d\n", mgr->bl_len);
+	spin_unlock_bh(&mgr->bl_lock);
+	tasklet_schedule(&mgr->comptask);
+}
+
+static void cc_proc_backlog(struct cc_drvdata *drvdata)
+{
+	struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
+	struct cc_bl_item *bli;
+	struct cc_crypto_req *creq;
+	void *req;
+	struct device *dev = drvdata_to_dev(drvdata);
+	int rc;
+
+	spin_lock(&mgr->bl_lock);
+
+	while (mgr->bl_len) {
+		bli = list_first_entry(&mgr->backlog, struct cc_bl_item, list);
+		dev_dbg(dev, "---bl len: %d\n", mgr->bl_len);
+
+		spin_unlock(&mgr->bl_lock);
+
+
+		creq = &bli->creq;
+		req = creq->user_arg;
+
+		/*
+		 * Notify the request we're moving out of the backlog
+		 * but only if we haven't done so already.
+		 */
+		if (!bli->notif) {
+			creq->user_cb(dev, req, -EINPROGRESS);
+			bli->notif = true;
+		}
+
+		spin_lock(&mgr->hw_lock);
+
+		rc = cc_queues_status(drvdata, mgr, bli->len);
+		if (rc) {
+			/*
+			 * There is still no room in the FIFO for
+			 * this request. Bail out. We'll return here
+			 * on the next completion irq.
+			 */
+			spin_unlock(&mgr->hw_lock);
+			return;
+		}
+
+		cc_do_send_request(drvdata, &bli->creq, bli->desc, bli->len,
+				   false);
+		spin_unlock(&mgr->hw_lock);
+
+		/* Remove ourselves from the backlog list */
+		spin_lock(&mgr->bl_lock);
+		list_del(&bli->list);
+		--mgr->bl_len;
+		kfree(bli);
+	}
+
+	spin_unlock(&mgr->bl_lock);
+}
+
+int cc_send_request(struct cc_drvdata *drvdata, struct cc_crypto_req *cc_req,
+		    struct cc_hw_desc *desc, unsigned int len,
+		    struct crypto_async_request *req)
+{
+	int rc;
+	struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
+	struct device *dev = drvdata_to_dev(drvdata);
+	bool backlog_ok = req->flags & CRYPTO_TFM_REQ_MAY_BACKLOG;
+	gfp_t flags = cc_gfp_flags(req);
+	struct cc_bl_item *bli;
+
+	rc = cc_pm_get(dev);
+	if (rc) {
+		dev_err(dev, "cc_pm_get returned %x\n", rc);
+		return rc;
+	}
+
+	spin_lock_bh(&mgr->hw_lock);
+	rc = cc_queues_status(drvdata, mgr, len);
+
+#ifdef CC_DEBUG_FORCE_BACKLOG
+	if (backlog_ok)
+		rc = -ENOSPC;
+#endif /* CC_DEBUG_FORCE_BACKLOG */
+
+	if (rc == -ENOSPC && backlog_ok) {
+		spin_unlock_bh(&mgr->hw_lock);
+
+		bli = kmalloc(sizeof(*bli), flags);
+		if (!bli) {
+			cc_pm_put_suspend(dev);
+			return -ENOMEM;
+		}
+
+		memcpy(&bli->creq, cc_req, sizeof(*cc_req));
+		memcpy(&bli->desc, desc, len * sizeof(*desc));
+		bli->len = len;
+		bli->notif = false;
+		cc_enqueue_backlog(drvdata, bli);
+		return -EBUSY;
+	}
+
+	if (!rc) {
+		cc_do_send_request(drvdata, cc_req, desc, len, false);
+		rc = -EINPROGRESS;
+	}
+
+	spin_unlock_bh(&mgr->hw_lock);
+	return rc;
+}
+
+int cc_send_sync_request(struct cc_drvdata *drvdata,
+			 struct cc_crypto_req *cc_req, struct cc_hw_desc *desc,
+			 unsigned int len)
+{
+	int rc;
+	struct device *dev = drvdata_to_dev(drvdata);
+	struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
+
+	init_completion(&cc_req->seq_compl);
+	cc_req->user_cb = request_mgr_complete;
+	cc_req->user_arg = &cc_req->seq_compl;
+
+	rc = cc_pm_get(dev);
+	if (rc) {
+		dev_err(dev, "cc_pm_get returned %x\n", rc);
+		return rc;
+	}
+
+	while (true) {
+		spin_lock_bh(&mgr->hw_lock);
+		rc = cc_queues_status(drvdata, mgr, len + 1);
+
+		if (!rc)
+			break;
+
+		spin_unlock_bh(&mgr->hw_lock);
+		wait_for_completion_interruptible(&drvdata->hw_queue_avail);
+		reinit_completion(&drvdata->hw_queue_avail);
+	}
+
+	cc_do_send_request(drvdata, cc_req, desc, len, true);
+	spin_unlock_bh(&mgr->hw_lock);
+	wait_for_completion(&cc_req->seq_compl);
+	return 0;
+}
+
+/**
+ * send_request_init() - Enqueue caller request to crypto hardware during init
+ * process.
+ * Assume this function is not called in the middle of a flow,
+ * since we set QUEUE_LAST_IND flag in the last descriptor.
+ *
+ * @drvdata: Associated device driver context
+ * @desc: The crypto sequence
+ * @len: The crypto sequence length
+ *
+ * Return:
+ * Returns "0" upon success
+ */
+int send_request_init(struct cc_drvdata *drvdata, struct cc_hw_desc *desc,
+		      unsigned int len)
+{
+	struct cc_req_mgr_handle *req_mgr_h = drvdata->request_mgr_handle;
+	unsigned int total_seq_len = len; /*initial sequence length*/
+	int rc = 0;
+
+	/* Wait for space in HW and SW FIFO. Poll for as much as FIFO_TIMEOUT.
+	 */
+	rc = cc_queues_status(drvdata, req_mgr_h, total_seq_len);
+	if (rc)
+		return rc;
+
+	set_queue_last_ind(drvdata, &desc[(len - 1)]);
+
+	/*
+	 * We are about to push command to the HW via the command registers
+	 * that may reference host memory. We need to issue a memory barrier
+	 * to make sure there are no outstanding memory writes
+	 */
+	wmb();
+	enqueue_seq(drvdata, desc, len);
+
+	/* Update the free slots in HW queue */
+	req_mgr_h->q_free_slots =
+		cc_ioread(drvdata, CC_REG(DSCRPTR_QUEUE_CONTENT));
+
+	return 0;
+}
+
+void complete_request(struct cc_drvdata *drvdata)
+{
+	struct cc_req_mgr_handle *request_mgr_handle =
+						drvdata->request_mgr_handle;
+
+	complete(&drvdata->hw_queue_avail);
+#ifdef COMP_IN_WQ
+	queue_delayed_work(request_mgr_handle->workq,
+			   &request_mgr_handle->compwork, 0);
+#else
+	tasklet_schedule(&request_mgr_handle->comptask);
+#endif
+}
+
+#ifdef COMP_IN_WQ
+static void comp_work_handler(struct work_struct *work)
+{
+	struct cc_drvdata *drvdata =
+		container_of(work, struct cc_drvdata, compwork.work);
+
+	comp_handler((unsigned long)drvdata);
+}
+#endif
+
+static void proc_completions(struct cc_drvdata *drvdata)
+{
+	struct cc_crypto_req *cc_req;
+	struct device *dev = drvdata_to_dev(drvdata);
+	struct cc_req_mgr_handle *request_mgr_handle =
+						drvdata->request_mgr_handle;
+	unsigned int *tail = &request_mgr_handle->req_queue_tail;
+	unsigned int *head = &request_mgr_handle->req_queue_head;
+	int rc;
+	u32 mask;
+
+	while (request_mgr_handle->axi_completed) {
+		request_mgr_handle->axi_completed--;
+
+		/* Dequeue request */
+		if (*head == *tail) {
+			/* We are supposed to handle a completion but our
+			 * queue is empty. This is not normal. Return and
+			 * hope for the best.
+			 */
+			dev_err(dev, "Request queue is empty head == tail %u\n",
+				*head);
+			break;
+		}
+
+		cc_req = &request_mgr_handle->req_queue[*tail];
+
+		if (cc_req->cpp.is_cpp) {
+
+			dev_dbg(dev, "CPP request completion slot: %d alg:%d\n",
+				cc_req->cpp.slot, cc_req->cpp.alg);
+			mask = cc_cpp_int_mask(cc_req->cpp.alg,
+					       cc_req->cpp.slot);
+			rc = (drvdata->irq & mask ? -EPERM : 0);
+			dev_dbg(dev, "Got mask: %x irq: %x rc: %d\n", mask,
+				drvdata->irq, rc);
+		} else {
+			dev_dbg(dev, "None CPP request completion\n");
+			rc = 0;
+		}
+
+		if (cc_req->user_cb)
+			cc_req->user_cb(dev, cc_req->user_arg, rc);
+		*tail = (*tail + 1) & (MAX_REQUEST_QUEUE_SIZE - 1);
+		dev_dbg(dev, "Dequeue request tail=%u\n", *tail);
+		dev_dbg(dev, "Request completed. axi_completed=%d\n",
+			request_mgr_handle->axi_completed);
+		cc_pm_put_suspend(dev);
+	}
+}
+
+static inline u32 cc_axi_comp_count(struct cc_drvdata *drvdata)
+{
+	return FIELD_GET(AXIM_MON_COMP_VALUE,
+			 cc_ioread(drvdata, drvdata->axim_mon_offset));
+}
+
+/* Deferred service handler, run as interrupt-fired tasklet */
+static void comp_handler(unsigned long devarg)
+{
+	struct cc_drvdata *drvdata = (struct cc_drvdata *)devarg;
+	struct cc_req_mgr_handle *request_mgr_handle =
+						drvdata->request_mgr_handle;
+	struct device *dev = drvdata_to_dev(drvdata);
+	u32 irq;
+
+	dev_dbg(dev, "Completion handler called!\n");
+	irq = (drvdata->irq & drvdata->comp_mask);
+
+	/* To avoid the interrupt from firing as we unmask it,
+	 * we clear it now
+	 */
+	cc_iowrite(drvdata, CC_REG(HOST_ICR), irq);
+
+	/* Avoid race with above clear: Test completion counter once more */
+
+	request_mgr_handle->axi_completed += cc_axi_comp_count(drvdata);
+
+	dev_dbg(dev, "AXI completion after updated: %d\n",
+		request_mgr_handle->axi_completed);
+
+	while (request_mgr_handle->axi_completed) {
+		do {
+			drvdata->irq |= cc_ioread(drvdata, CC_REG(HOST_IRR));
+			irq = (drvdata->irq & drvdata->comp_mask);
+			proc_completions(drvdata);
+
+			/* At this point (after proc_completions()),
+			 * request_mgr_handle->axi_completed is 0.
+			 */
+			request_mgr_handle->axi_completed +=
+						cc_axi_comp_count(drvdata);
+		} while (request_mgr_handle->axi_completed > 0);
+
+		cc_iowrite(drvdata, CC_REG(HOST_ICR), irq);
+
+		request_mgr_handle->axi_completed += cc_axi_comp_count(drvdata);
+	}
+
+	/* after verifying that there is nothing to do,
+	 * unmask AXI completion interrupt
+	 */
+	cc_iowrite(drvdata, CC_REG(HOST_IMR),
+		   cc_ioread(drvdata, CC_REG(HOST_IMR)) & ~drvdata->comp_mask);
+
+	cc_proc_backlog(drvdata);
+	dev_dbg(dev, "Comp. handler done.\n");
+}