diff options
Diffstat (limited to 'arch/mips/include/asm/octeon/cvmx-pow.h')
-rw-r--r-- | arch/mips/include/asm/octeon/cvmx-pow.h | 2215 |
1 files changed, 2215 insertions, 0 deletions
diff --git a/arch/mips/include/asm/octeon/cvmx-pow.h b/arch/mips/include/asm/octeon/cvmx-pow.h new file mode 100644 index 000000000..a3b23811e --- /dev/null +++ b/arch/mips/include/asm/octeon/cvmx-pow.h @@ -0,0 +1,2215 @@ +/***********************license start*************** + * Author: Cavium Networks + * + * Contact: support@caviumnetworks.com + * This file is part of the OCTEON SDK + * + * Copyright (c) 2003-2008 Cavium Networks + * + * This file is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License, Version 2, as + * published by the Free Software Foundation. + * + * This file is distributed in the hope that it will be useful, but + * AS-IS and WITHOUT ANY WARRANTY; without even the implied warranty + * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, TITLE, or + * NONINFRINGEMENT. See the GNU General Public License for more + * details. + * + * You should have received a copy of the GNU General Public License + * along with this file; if not, write to the Free Software + * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA + * or visit http://www.gnu.org/licenses/. + * + * This file may also be available under a different license from Cavium. + * Contact Cavium Networks for more information + ***********************license end**************************************/ + +/** + * Interface to the hardware Packet Order / Work unit. + * + * New, starting with SDK 1.7.0, cvmx-pow supports a number of + * extended consistency checks. The define + * CVMX_ENABLE_POW_CHECKS controls the runtime insertion of POW + * internal state checks to find common programming errors. If + * CVMX_ENABLE_POW_CHECKS is not defined, checks are by default + * enabled. For example, cvmx-pow will check for the following + * program errors or POW state inconsistency. + * - Requesting a POW operation with an active tag switch in + * progress. + * - Waiting for a tag switch to complete for an excessively + * long period. This is normally a sign of an error in locking + * causing deadlock. + * - Illegal tag switches from NULL_NULL. + * - Illegal tag switches from NULL. + * - Illegal deschedule request. + * - WQE pointer not matching the one attached to the core by + * the POW. + * + */ + +#ifndef __CVMX_POW_H__ +#define __CVMX_POW_H__ + +#include <asm/octeon/cvmx-pow-defs.h> + +#include <asm/octeon/cvmx-scratch.h> +#include <asm/octeon/cvmx-wqe.h> + +/* Default to having all POW constancy checks turned on */ +#ifndef CVMX_ENABLE_POW_CHECKS +#define CVMX_ENABLE_POW_CHECKS 1 +#endif + +enum cvmx_pow_tag_type { + /* Tag ordering is maintained */ + CVMX_POW_TAG_TYPE_ORDERED = 0L, + /* Tag ordering is maintained, and at most one PP has the tag */ + CVMX_POW_TAG_TYPE_ATOMIC = 1L, + /* + * The work queue entry from the order - NEVER tag switch from + * NULL to NULL + */ + CVMX_POW_TAG_TYPE_NULL = 2L, + /* A tag switch to NULL, and there is no space reserved in POW + * - NEVER tag switch to NULL_NULL + * - NEVER tag switch from NULL_NULL + * - NULL_NULL is entered at the beginning of time and on a deschedule. + * - NULL_NULL can be exited by a new work request. A NULL_SWITCH + * load can also switch the state to NULL + */ + CVMX_POW_TAG_TYPE_NULL_NULL = 3L +}; + +/** + * Wait flag values for pow functions. + */ +typedef enum { + CVMX_POW_WAIT = 1, + CVMX_POW_NO_WAIT = 0, +} cvmx_pow_wait_t; + +/** + * POW tag operations. These are used in the data stored to the POW. + */ +typedef enum { + /* + * switch the tag (only) for this PP + * - the previous tag should be non-NULL in this case + * - tag switch response required + * - fields used: op, type, tag + */ + CVMX_POW_TAG_OP_SWTAG = 0L, + /* + * switch the tag for this PP, with full information + * - this should be used when the previous tag is NULL + * - tag switch response required + * - fields used: address, op, grp, type, tag + */ + CVMX_POW_TAG_OP_SWTAG_FULL = 1L, + /* + * switch the tag (and/or group) for this PP and de-schedule + * - OK to keep the tag the same and only change the group + * - fields used: op, no_sched, grp, type, tag + */ + CVMX_POW_TAG_OP_SWTAG_DESCH = 2L, + /* + * just de-schedule + * - fields used: op, no_sched + */ + CVMX_POW_TAG_OP_DESCH = 3L, + /* + * create an entirely new work queue entry + * - fields used: address, op, qos, grp, type, tag + */ + CVMX_POW_TAG_OP_ADDWQ = 4L, + /* + * just update the work queue pointer and grp for this PP + * - fields used: address, op, grp + */ + CVMX_POW_TAG_OP_UPDATE_WQP_GRP = 5L, + /* + * set the no_sched bit on the de-schedule list + * + * - does nothing if the selected entry is not on the + * de-schedule list + * + * - does nothing if the stored work queue pointer does not + * match the address field + * + * - fields used: address, index, op + * + * Before issuing a *_NSCHED operation, SW must guarantee + * that all prior deschedules and set/clr NSCHED operations + * are complete and all prior switches are complete. The + * hardware provides the opsdone bit and swdone bit for SW + * polling. After issuing a *_NSCHED operation, SW must + * guarantee that the set/clr NSCHED is complete before any + * subsequent operations. + */ + CVMX_POW_TAG_OP_SET_NSCHED = 6L, + /* + * clears the no_sched bit on the de-schedule list + * + * - does nothing if the selected entry is not on the + * de-schedule list + * + * - does nothing if the stored work queue pointer does not + * match the address field + * + * - fields used: address, index, op + * + * Before issuing a *_NSCHED operation, SW must guarantee that + * all prior deschedules and set/clr NSCHED operations are + * complete and all prior switches are complete. The hardware + * provides the opsdone bit and swdone bit for SW + * polling. After issuing a *_NSCHED operation, SW must + * guarantee that the set/clr NSCHED is complete before any + * subsequent operations. + */ + CVMX_POW_TAG_OP_CLR_NSCHED = 7L, + /* do nothing */ + CVMX_POW_TAG_OP_NOP = 15L +} cvmx_pow_tag_op_t; + +/** + * This structure defines the store data on a store to POW + */ +typedef union { + uint64_t u64; + struct { +#ifdef __BIG_ENDIAN_BITFIELD + /* + * Don't reschedule this entry. no_sched is used for + * CVMX_POW_TAG_OP_SWTAG_DESCH and + * CVMX_POW_TAG_OP_DESCH + */ + uint64_t no_sched:1; + uint64_t unused:2; + /* Tontains index of entry for a CVMX_POW_TAG_OP_*_NSCHED */ + uint64_t index:13; + /* The operation to perform */ + cvmx_pow_tag_op_t op:4; + uint64_t unused2:2; + /* + * The QOS level for the packet. qos is only used for + * CVMX_POW_TAG_OP_ADDWQ + */ + uint64_t qos:3; + /* + * The group that the work queue entry will be + * scheduled to grp is used for CVMX_POW_TAG_OP_ADDWQ, + * CVMX_POW_TAG_OP_SWTAG_FULL, + * CVMX_POW_TAG_OP_SWTAG_DESCH, and + * CVMX_POW_TAG_OP_UPDATE_WQP_GRP + */ + uint64_t grp:4; + /* + * The type of the tag. type is used for everything + * except CVMX_POW_TAG_OP_DESCH, + * CVMX_POW_TAG_OP_UPDATE_WQP_GRP, and + * CVMX_POW_TAG_OP_*_NSCHED + */ + uint64_t type:3; + /* + * The actual tag. tag is used for everything except + * CVMX_POW_TAG_OP_DESCH, + * CVMX_POW_TAG_OP_UPDATE_WQP_GRP, and + * CVMX_POW_TAG_OP_*_NSCHED + */ + uint64_t tag:32; +#else + uint64_t tag:32; + uint64_t type:3; + uint64_t grp:4; + uint64_t qos:3; + uint64_t unused2:2; + cvmx_pow_tag_op_t op:4; + uint64_t index:13; + uint64_t unused:2; + uint64_t no_sched:1; +#endif + } s; +} cvmx_pow_tag_req_t; + +/** + * This structure describes the address to load stuff from POW + */ +typedef union { + uint64_t u64; + + /** + * Address for new work request loads (did<2:0> == 0) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + /* Mips64 address region. Should be CVMX_IO_SEG */ + uint64_t mem_region:2; + /* Must be zero */ + uint64_t reserved_49_61:13; + /* Must be one */ + uint64_t is_io:1; + /* the ID of POW -- did<2:0> == 0 in this case */ + uint64_t did:8; + /* Must be zero */ + uint64_t reserved_4_39:36; + /* + * If set, don't return load response until work is + * available. + */ + uint64_t wait:1; + /* Must be zero */ + uint64_t reserved_0_2:3; +#else + uint64_t reserved_0_2:3; + uint64_t wait:1; + uint64_t reserved_4_39:36; + uint64_t did:8; + uint64_t is_io:1; + uint64_t reserved_49_61:13; + uint64_t mem_region:2; +#endif + } swork; + + /** + * Address for loads to get POW internal status + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + /* Mips64 address region. Should be CVMX_IO_SEG */ + uint64_t mem_region:2; + /* Must be zero */ + uint64_t reserved_49_61:13; + /* Must be one */ + uint64_t is_io:1; + /* the ID of POW -- did<2:0> == 1 in this case */ + uint64_t did:8; + /* Must be zero */ + uint64_t reserved_10_39:30; + /* The core id to get status for */ + uint64_t coreid:4; + /* + * If set and get_cur is set, return reverse tag-list + * pointer rather than forward tag-list pointer. + */ + uint64_t get_rev:1; + /* + * If set, return current status rather than pending + * status. + */ + uint64_t get_cur:1; + /* + * If set, get the work-queue pointer rather than + * tag/type. + */ + uint64_t get_wqp:1; + /* Must be zero */ + uint64_t reserved_0_2:3; +#else + uint64_t reserved_0_2:3; + uint64_t get_wqp:1; + uint64_t get_cur:1; + uint64_t get_rev:1; + uint64_t coreid:4; + uint64_t reserved_10_39:30; + uint64_t did:8; + uint64_t is_io:1; + uint64_t reserved_49_61:13; + uint64_t mem_region:2; +#endif + } sstatus; + + /** + * Address for memory loads to get POW internal state + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + /* Mips64 address region. Should be CVMX_IO_SEG */ + uint64_t mem_region:2; + /* Must be zero */ + uint64_t reserved_49_61:13; + /* Must be one */ + uint64_t is_io:1; + /* the ID of POW -- did<2:0> == 2 in this case */ + uint64_t did:8; + /* Must be zero */ + uint64_t reserved_16_39:24; + /* POW memory index */ + uint64_t index:11; + /* + * If set, return deschedule information rather than + * the standard response for work-queue index (invalid + * if the work-queue entry is not on the deschedule + * list). + */ + uint64_t get_des:1; + /* + * If set, get the work-queue pointer rather than + * tag/type (no effect when get_des set). + */ + uint64_t get_wqp:1; + /* Must be zero */ + uint64_t reserved_0_2:3; +#else + uint64_t reserved_0_2:3; + uint64_t get_wqp:1; + uint64_t get_des:1; + uint64_t index:11; + uint64_t reserved_16_39:24; + uint64_t did:8; + uint64_t is_io:1; + uint64_t reserved_49_61:13; + uint64_t mem_region:2; +#endif + } smemload; + + /** + * Address for index/pointer loads + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + /* Mips64 address region. Should be CVMX_IO_SEG */ + uint64_t mem_region:2; + /* Must be zero */ + uint64_t reserved_49_61:13; + /* Must be one */ + uint64_t is_io:1; + /* the ID of POW -- did<2:0> == 3 in this case */ + uint64_t did:8; + /* Must be zero */ + uint64_t reserved_9_39:31; + /* + * when {get_rmt ==0 AND get_des_get_tail == 0}, this + * field selects one of eight POW internal-input + * queues (0-7), one per QOS level; values 8-15 are + * illegal in this case; when {get_rmt ==0 AND + * get_des_get_tail == 1}, this field selects one of + * 16 deschedule lists (per group); when get_rmt ==1, + * this field selects one of 16 memory-input queue + * lists. The two memory-input queue lists associated + * with each QOS level are: + * + * - qosgrp = 0, qosgrp = 8: QOS0 + * - qosgrp = 1, qosgrp = 9: QOS1 + * - qosgrp = 2, qosgrp = 10: QOS2 + * - qosgrp = 3, qosgrp = 11: QOS3 + * - qosgrp = 4, qosgrp = 12: QOS4 + * - qosgrp = 5, qosgrp = 13: QOS5 + * - qosgrp = 6, qosgrp = 14: QOS6 + * - qosgrp = 7, qosgrp = 15: QOS7 + */ + uint64_t qosgrp:4; + /* + * If set and get_rmt is clear, return deschedule list + * indexes rather than indexes for the specified qos + * level; if set and get_rmt is set, return the tail + * pointer rather than the head pointer for the + * specified qos level. + */ + uint64_t get_des_get_tail:1; + /* + * If set, return remote pointers rather than the + * local indexes for the specified qos level. + */ + uint64_t get_rmt:1; + /* Must be zero */ + uint64_t reserved_0_2:3; +#else + uint64_t reserved_0_2:3; + uint64_t get_rmt:1; + uint64_t get_des_get_tail:1; + uint64_t qosgrp:4; + uint64_t reserved_9_39:31; + uint64_t did:8; + uint64_t is_io:1; + uint64_t reserved_49_61:13; + uint64_t mem_region:2; +#endif + } sindexload; + + /** + * address for NULL_RD request (did<2:0> == 4) when this is read, + * HW attempts to change the state to NULL if it is NULL_NULL (the + * hardware cannot switch from NULL_NULL to NULL if a POW entry is + * not available - software may need to recover by finishing + * another piece of work before a POW entry can ever become + * available.) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + /* Mips64 address region. Should be CVMX_IO_SEG */ + uint64_t mem_region:2; + /* Must be zero */ + uint64_t reserved_49_61:13; + /* Must be one */ + uint64_t is_io:1; + /* the ID of POW -- did<2:0> == 4 in this case */ + uint64_t did:8; + /* Must be zero */ + uint64_t reserved_0_39:40; +#else + uint64_t reserved_0_39:40; + uint64_t did:8; + uint64_t is_io:1; + uint64_t reserved_49_61:13; + uint64_t mem_region:2; +#endif + } snull_rd; +} cvmx_pow_load_addr_t; + +/** + * This structure defines the response to a load/SENDSINGLE to POW + * (except CSR reads) + */ +typedef union { + uint64_t u64; + + /** + * Response to new work request loads + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + /* + * Set when no new work queue entry was returned. * + * If there was de-scheduled work, the HW will + * definitely return it. When this bit is set, it + * could mean either mean: + * + * - There was no work, or + * + * - There was no work that the HW could find. This + * case can happen, regardless of the wait bit value + * in the original request, when there is work in + * the IQ's that is too deep down the list. + */ + uint64_t no_work:1; + /* Must be zero */ + uint64_t reserved_40_62:23; + /* 36 in O1 -- the work queue pointer */ + uint64_t addr:40; +#else + uint64_t addr:40; + uint64_t reserved_40_62:23; + uint64_t no_work:1; +#endif + } s_work; + + /** + * Result for a POW Status Load (when get_cur==0 and get_wqp==0) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_62_63:2; + /* Set when there is a pending non-NULL SWTAG or + * SWTAG_FULL, and the POW entry has not left the list + * for the original tag. */ + uint64_t pend_switch:1; + /* Set when SWTAG_FULL and pend_switch is set. */ + uint64_t pend_switch_full:1; + /* + * Set when there is a pending NULL SWTAG, or an + * implicit switch to NULL. + */ + uint64_t pend_switch_null:1; + /* Set when there is a pending DESCHED or SWTAG_DESCHED. */ + uint64_t pend_desched:1; + /* + * Set when there is a pending SWTAG_DESCHED and + * pend_desched is set. + */ + uint64_t pend_desched_switch:1; + /* Set when nosched is desired and pend_desched is set. */ + uint64_t pend_nosched:1; + /* Set when there is a pending GET_WORK. */ + uint64_t pend_new_work:1; + /* + * When pend_new_work is set, this bit indicates that + * the wait bit was set. + */ + uint64_t pend_new_work_wait:1; + /* Set when there is a pending NULL_RD. */ + uint64_t pend_null_rd:1; + /* Set when there is a pending CLR_NSCHED. */ + uint64_t pend_nosched_clr:1; + uint64_t reserved_51:1; + /* This is the index when pend_nosched_clr is set. */ + uint64_t pend_index:11; + /* + * This is the new_grp when (pend_desched AND + * pend_desched_switch) is set. + */ + uint64_t pend_grp:4; + uint64_t reserved_34_35:2; + /* + * This is the tag type when pend_switch or + * (pend_desched AND pend_desched_switch) are set. + */ + uint64_t pend_type:2; + /* + * - this is the tag when pend_switch or (pend_desched + * AND pend_desched_switch) are set. + */ + uint64_t pend_tag:32; +#else + uint64_t pend_tag:32; + uint64_t pend_type:2; + uint64_t reserved_34_35:2; + uint64_t pend_grp:4; + uint64_t pend_index:11; + uint64_t reserved_51:1; + uint64_t pend_nosched_clr:1; + uint64_t pend_null_rd:1; + uint64_t pend_new_work_wait:1; + uint64_t pend_new_work:1; + uint64_t pend_nosched:1; + uint64_t pend_desched_switch:1; + uint64_t pend_desched:1; + uint64_t pend_switch_null:1; + uint64_t pend_switch_full:1; + uint64_t pend_switch:1; + uint64_t reserved_62_63:2; +#endif + } s_sstatus0; + + /** + * Result for a POW Status Load (when get_cur==0 and get_wqp==1) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_62_63:2; + /* + * Set when there is a pending non-NULL SWTAG or + * SWTAG_FULL, and the POW entry has not left the list + * for the original tag. + */ + uint64_t pend_switch:1; + /* Set when SWTAG_FULL and pend_switch is set. */ + uint64_t pend_switch_full:1; + /* + * Set when there is a pending NULL SWTAG, or an + * implicit switch to NULL. + */ + uint64_t pend_switch_null:1; + /* + * Set when there is a pending DESCHED or + * SWTAG_DESCHED. + */ + uint64_t pend_desched:1; + /* + * Set when there is a pending SWTAG_DESCHED and + * pend_desched is set. + */ + uint64_t pend_desched_switch:1; + /* Set when nosched is desired and pend_desched is set. */ + uint64_t pend_nosched:1; + /* Set when there is a pending GET_WORK. */ + uint64_t pend_new_work:1; + /* + * When pend_new_work is set, this bit indicates that + * the wait bit was set. + */ + uint64_t pend_new_work_wait:1; + /* Set when there is a pending NULL_RD. */ + uint64_t pend_null_rd:1; + /* Set when there is a pending CLR_NSCHED. */ + uint64_t pend_nosched_clr:1; + uint64_t reserved_51:1; + /* This is the index when pend_nosched_clr is set. */ + uint64_t pend_index:11; + /* + * This is the new_grp when (pend_desched AND + * pend_desched_switch) is set. + */ + uint64_t pend_grp:4; + /* This is the wqp when pend_nosched_clr is set. */ + uint64_t pend_wqp:36; +#else + uint64_t pend_wqp:36; + uint64_t pend_grp:4; + uint64_t pend_index:11; + uint64_t reserved_51:1; + uint64_t pend_nosched_clr:1; + uint64_t pend_null_rd:1; + uint64_t pend_new_work_wait:1; + uint64_t pend_new_work:1; + uint64_t pend_nosched:1; + uint64_t pend_desched_switch:1; + uint64_t pend_desched:1; + uint64_t pend_switch_null:1; + uint64_t pend_switch_full:1; + uint64_t pend_switch:1; + uint64_t reserved_62_63:2; +#endif + } s_sstatus1; + + /** + * Result for a POW Status Load (when get_cur==1, get_wqp==0, and + * get_rev==0) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_62_63:2; + /* + * Points to the next POW entry in the tag list when + * tail == 0 (and tag_type is not NULL or NULL_NULL). + */ + uint64_t link_index:11; + /* The POW entry attached to the core. */ + uint64_t index:11; + /* + * The group attached to the core (updated when new + * tag list entered on SWTAG_FULL). + */ + uint64_t grp:4; + /* + * Set when this POW entry is at the head of its tag + * list (also set when in the NULL or NULL_NULL + * state). + */ + uint64_t head:1; + /* + * Set when this POW entry is at the tail of its tag + * list (also set when in the NULL or NULL_NULL + * state). + */ + uint64_t tail:1; + /* + * The tag type attached to the core (updated when new + * tag list entered on SWTAG, SWTAG_FULL, or + * SWTAG_DESCHED). + */ + uint64_t tag_type:2; + /* + * The tag attached to the core (updated when new tag + * list entered on SWTAG, SWTAG_FULL, or + * SWTAG_DESCHED). + */ + uint64_t tag:32; +#else + uint64_t tag:32; + uint64_t tag_type:2; + uint64_t tail:1; + uint64_t head:1; + uint64_t grp:4; + uint64_t index:11; + uint64_t link_index:11; + uint64_t reserved_62_63:2; +#endif + } s_sstatus2; + + /** + * Result for a POW Status Load (when get_cur==1, get_wqp==0, and get_rev==1) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_62_63:2; + /* + * Points to the prior POW entry in the tag list when + * head == 0 (and tag_type is not NULL or + * NULL_NULL). This field is unpredictable when the + * core's state is NULL or NULL_NULL. + */ + uint64_t revlink_index:11; + /* The POW entry attached to the core. */ + uint64_t index:11; + /* + * The group attached to the core (updated when new + * tag list entered on SWTAG_FULL). + */ + uint64_t grp:4; + /* Set when this POW entry is at the head of its tag + * list (also set when in the NULL or NULL_NULL + * state). + */ + uint64_t head:1; + /* + * Set when this POW entry is at the tail of its tag + * list (also set when in the NULL or NULL_NULL + * state). + */ + uint64_t tail:1; + /* + * The tag type attached to the core (updated when new + * tag list entered on SWTAG, SWTAG_FULL, or + * SWTAG_DESCHED). + */ + uint64_t tag_type:2; + /* + * The tag attached to the core (updated when new tag + * list entered on SWTAG, SWTAG_FULL, or + * SWTAG_DESCHED). + */ + uint64_t tag:32; +#else + uint64_t tag:32; + uint64_t tag_type:2; + uint64_t tail:1; + uint64_t head:1; + uint64_t grp:4; + uint64_t index:11; + uint64_t revlink_index:11; + uint64_t reserved_62_63:2; +#endif + } s_sstatus3; + + /** + * Result for a POW Status Load (when get_cur==1, get_wqp==1, and + * get_rev==0) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_62_63:2; + /* + * Points to the next POW entry in the tag list when + * tail == 0 (and tag_type is not NULL or NULL_NULL). + */ + uint64_t link_index:11; + /* The POW entry attached to the core. */ + uint64_t index:11; + /* + * The group attached to the core (updated when new + * tag list entered on SWTAG_FULL). + */ + uint64_t grp:4; + /* + * The wqp attached to the core (updated when new tag + * list entered on SWTAG_FULL). + */ + uint64_t wqp:36; +#else + uint64_t wqp:36; + uint64_t grp:4; + uint64_t index:11; + uint64_t link_index:11; + uint64_t reserved_62_63:2; +#endif + } s_sstatus4; + + /** + * Result for a POW Status Load (when get_cur==1, get_wqp==1, and + * get_rev==1) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_62_63:2; + /* + * Points to the prior POW entry in the tag list when + * head == 0 (and tag_type is not NULL or + * NULL_NULL). This field is unpredictable when the + * core's state is NULL or NULL_NULL. + */ + uint64_t revlink_index:11; + /* The POW entry attached to the core. */ + uint64_t index:11; + /* + * The group attached to the core (updated when new + * tag list entered on SWTAG_FULL). + */ + uint64_t grp:4; + /* + * The wqp attached to the core (updated when new tag + * list entered on SWTAG_FULL). + */ + uint64_t wqp:36; +#else + uint64_t wqp:36; + uint64_t grp:4; + uint64_t index:11; + uint64_t revlink_index:11; + uint64_t reserved_62_63:2; +#endif + } s_sstatus5; + + /** + * Result For POW Memory Load (get_des == 0 and get_wqp == 0) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_51_63:13; + /* + * The next entry in the input, free, descheduled_head + * list (unpredictable if entry is the tail of the + * list). + */ + uint64_t next_index:11; + /* The group of the POW entry. */ + uint64_t grp:4; + uint64_t reserved_35:1; + /* + * Set when this POW entry is at the tail of its tag + * list (also set when in the NULL or NULL_NULL + * state). + */ + uint64_t tail:1; + /* The tag type of the POW entry. */ + uint64_t tag_type:2; + /* The tag of the POW entry. */ + uint64_t tag:32; +#else + uint64_t tag:32; + uint64_t tag_type:2; + uint64_t tail:1; + uint64_t reserved_35:1; + uint64_t grp:4; + uint64_t next_index:11; + uint64_t reserved_51_63:13; +#endif + } s_smemload0; + + /** + * Result For POW Memory Load (get_des == 0 and get_wqp == 1) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_51_63:13; + /* + * The next entry in the input, free, descheduled_head + * list (unpredictable if entry is the tail of the + * list). + */ + uint64_t next_index:11; + /* The group of the POW entry. */ + uint64_t grp:4; + /* The WQP held in the POW entry. */ + uint64_t wqp:36; +#else + uint64_t wqp:36; + uint64_t grp:4; + uint64_t next_index:11; + uint64_t reserved_51_63:13; +#endif + } s_smemload1; + + /** + * Result For POW Memory Load (get_des == 1) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_51_63:13; + /* + * The next entry in the tag list connected to the + * descheduled head. + */ + uint64_t fwd_index:11; + /* The group of the POW entry. */ + uint64_t grp:4; + /* The nosched bit for the POW entry. */ + uint64_t nosched:1; + /* There is a pending tag switch */ + uint64_t pend_switch:1; + /* + * The next tag type for the new tag list when + * pend_switch is set. + */ + uint64_t pend_type:2; + /* + * The next tag for the new tag list when pend_switch + * is set. + */ + uint64_t pend_tag:32; +#else + uint64_t pend_tag:32; + uint64_t pend_type:2; + uint64_t pend_switch:1; + uint64_t nosched:1; + uint64_t grp:4; + uint64_t fwd_index:11; + uint64_t reserved_51_63:13; +#endif + } s_smemload2; + + /** + * Result For POW Index/Pointer Load (get_rmt == 0/get_des_get_tail == 0) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_52_63:12; + /* + * set when there is one or more POW entries on the + * free list. + */ + uint64_t free_val:1; + /* + * set when there is exactly one POW entry on the free + * list. + */ + uint64_t free_one:1; + uint64_t reserved_49:1; + /* + * when free_val is set, indicates the first entry on + * the free list. + */ + uint64_t free_head:11; + uint64_t reserved_37:1; + /* + * when free_val is set, indicates the last entry on + * the free list. + */ + uint64_t free_tail:11; + /* + * set when there is one or more POW entries on the + * input Q list selected by qosgrp. + */ + uint64_t loc_val:1; + /* + * set when there is exactly one POW entry on the + * input Q list selected by qosgrp. + */ + uint64_t loc_one:1; + uint64_t reserved_23:1; + /* + * when loc_val is set, indicates the first entry on + * the input Q list selected by qosgrp. + */ + uint64_t loc_head:11; + uint64_t reserved_11:1; + /* + * when loc_val is set, indicates the last entry on + * the input Q list selected by qosgrp. + */ + uint64_t loc_tail:11; +#else + uint64_t loc_tail:11; + uint64_t reserved_11:1; + uint64_t loc_head:11; + uint64_t reserved_23:1; + uint64_t loc_one:1; + uint64_t loc_val:1; + uint64_t free_tail:11; + uint64_t reserved_37:1; + uint64_t free_head:11; + uint64_t reserved_49:1; + uint64_t free_one:1; + uint64_t free_val:1; + uint64_t reserved_52_63:12; +#endif + } sindexload0; + + /** + * Result For POW Index/Pointer Load (get_rmt == 0/get_des_get_tail == 1) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_52_63:12; + /* + * set when there is one or more POW entries on the + * nosched list. + */ + uint64_t nosched_val:1; + /* + * set when there is exactly one POW entry on the + * nosched list. + */ + uint64_t nosched_one:1; + uint64_t reserved_49:1; + /* + * when nosched_val is set, indicates the first entry + * on the nosched list. + */ + uint64_t nosched_head:11; + uint64_t reserved_37:1; + /* + * when nosched_val is set, indicates the last entry + * on the nosched list. + */ + uint64_t nosched_tail:11; + /* + * set when there is one or more descheduled heads on + * the descheduled list selected by qosgrp. + */ + uint64_t des_val:1; + /* + * set when there is exactly one descheduled head on + * the descheduled list selected by qosgrp. + */ + uint64_t des_one:1; + uint64_t reserved_23:1; + /* + * when des_val is set, indicates the first + * descheduled head on the descheduled list selected + * by qosgrp. + */ + uint64_t des_head:11; + uint64_t reserved_11:1; + /* + * when des_val is set, indicates the last descheduled + * head on the descheduled list selected by qosgrp. + */ + uint64_t des_tail:11; +#else + uint64_t des_tail:11; + uint64_t reserved_11:1; + uint64_t des_head:11; + uint64_t reserved_23:1; + uint64_t des_one:1; + uint64_t des_val:1; + uint64_t nosched_tail:11; + uint64_t reserved_37:1; + uint64_t nosched_head:11; + uint64_t reserved_49:1; + uint64_t nosched_one:1; + uint64_t nosched_val:1; + uint64_t reserved_52_63:12; +#endif + } sindexload1; + + /** + * Result For POW Index/Pointer Load (get_rmt == 1/get_des_get_tail == 0) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_39_63:25; + /* + * Set when this DRAM list is the current head + * (i.e. is the next to be reloaded when the POW + * hardware reloads a POW entry from DRAM). The POW + * hardware alternates between the two DRAM lists + * associated with a QOS level when it reloads work + * from DRAM into the POW unit. + */ + uint64_t rmt_is_head:1; + /* + * Set when the DRAM portion of the input Q list + * selected by qosgrp contains one or more pieces of + * work. + */ + uint64_t rmt_val:1; + /* + * Set when the DRAM portion of the input Q list + * selected by qosgrp contains exactly one piece of + * work. + */ + uint64_t rmt_one:1; + /* + * When rmt_val is set, indicates the first piece of + * work on the DRAM input Q list selected by + * qosgrp. + */ + uint64_t rmt_head:36; +#else + uint64_t rmt_head:36; + uint64_t rmt_one:1; + uint64_t rmt_val:1; + uint64_t rmt_is_head:1; + uint64_t reserved_39_63:25; +#endif + } sindexload2; + + /** + * Result For POW Index/Pointer Load (get_rmt == + * 1/get_des_get_tail == 1) + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t reserved_39_63:25; + /* + * set when this DRAM list is the current head + * (i.e. is the next to be reloaded when the POW + * hardware reloads a POW entry from DRAM). The POW + * hardware alternates between the two DRAM lists + * associated with a QOS level when it reloads work + * from DRAM into the POW unit. + */ + uint64_t rmt_is_head:1; + /* + * set when the DRAM portion of the input Q list + * selected by qosgrp contains one or more pieces of + * work. + */ + uint64_t rmt_val:1; + /* + * set when the DRAM portion of the input Q list + * selected by qosgrp contains exactly one piece of + * work. + */ + uint64_t rmt_one:1; + /* + * when rmt_val is set, indicates the last piece of + * work on the DRAM input Q list selected by + * qosgrp. + */ + uint64_t rmt_tail:36; +#else + uint64_t rmt_tail:36; + uint64_t rmt_one:1; + uint64_t rmt_val:1; + uint64_t rmt_is_head:1; + uint64_t reserved_39_63:25; +#endif + } sindexload3; + + /** + * Response to NULL_RD request loads + */ + struct { +#ifdef __BIG_ENDIAN_BITFIELD + uint64_t unused:62; + /* of type cvmx_pow_tag_type_t. state is one of the + * following: + * + * - CVMX_POW_TAG_TYPE_ORDERED + * - CVMX_POW_TAG_TYPE_ATOMIC + * - CVMX_POW_TAG_TYPE_NULL + * - CVMX_POW_TAG_TYPE_NULL_NULL + */ + uint64_t state:2; +#else + uint64_t state:2; + uint64_t unused:62; +#endif + } s_null_rd; + +} cvmx_pow_tag_load_resp_t; + +/** + * This structure describes the address used for stores to the POW. + * The store address is meaningful on stores to the POW. The + * hardware assumes that an aligned 64-bit store was used for all + * these stores. Note the assumption that the work queue entry is + * aligned on an 8-byte boundary (since the low-order 3 address bits + * must be zero). Note that not all fields are used by all + * operations. + * + * NOTE: The following is the behavior of the pending switch bit at the PP + * for POW stores (i.e. when did<7:3> == 0xc) + * - did<2:0> == 0 => pending switch bit is set + * - did<2:0> == 1 => no affect on the pending switch bit + * - did<2:0> == 3 => pending switch bit is cleared + * - did<2:0> == 7 => no affect on the pending switch bit + * - did<2:0> == others => must not be used + * - No other loads/stores have an affect on the pending switch bit + * - The switch bus from POW can clear the pending switch bit + * + * NOTE: did<2:0> == 2 is used by the HW for a special single-cycle + * ADDWQ command that only contains the pointer). SW must never use + * did<2:0> == 2. + */ +typedef union { + /** + * Unsigned 64 bit integer representation of store address + */ + uint64_t u64; + + struct { +#ifdef __BIG_ENDIAN_BITFIELD + /* Memory region. Should be CVMX_IO_SEG in most cases */ + uint64_t mem_reg:2; + uint64_t reserved_49_61:13; /* Must be zero */ + uint64_t is_io:1; /* Must be one */ + /* Device ID of POW. Note that different sub-dids are used. */ + uint64_t did:8; + uint64_t reserved_36_39:4; /* Must be zero */ + /* Address field. addr<2:0> must be zero */ + uint64_t addr:36; +#else + uint64_t addr:36; + uint64_t reserved_36_39:4; + uint64_t did:8; + uint64_t is_io:1; + uint64_t reserved_49_61:13; + uint64_t mem_reg:2; +#endif + } stag; +} cvmx_pow_tag_store_addr_t; + +/** + * decode of the store data when an IOBDMA SENDSINGLE is sent to POW + */ +typedef union { + uint64_t u64; + + struct { +#ifdef __BIG_ENDIAN_BITFIELD + /* + * the (64-bit word) location in scratchpad to write + * to (if len != 0) + */ + uint64_t scraddr:8; + /* the number of words in the response (0 => no response) */ + uint64_t len:8; + /* the ID of the device on the non-coherent bus */ + uint64_t did:8; + uint64_t unused:36; + /* if set, don't return load response until work is available */ + uint64_t wait:1; + uint64_t unused2:3; +#else + uint64_t unused2:3; + uint64_t wait:1; + uint64_t unused:36; + uint64_t did:8; + uint64_t len:8; + uint64_t scraddr:8; +#endif + } s; + +} cvmx_pow_iobdma_store_t; + +/* CSR typedefs have been moved to cvmx-csr-*.h */ + +/** + * Get the POW tag for this core. This returns the current + * tag type, tag, group, and POW entry index associated with + * this core. Index is only valid if the tag type isn't NULL_NULL. + * If a tag switch is pending this routine returns the tag before + * the tag switch, not after. + * + * Returns Current tag + */ +static inline cvmx_pow_tag_req_t cvmx_pow_get_current_tag(void) +{ + cvmx_pow_load_addr_t load_addr; + cvmx_pow_tag_load_resp_t load_resp; + cvmx_pow_tag_req_t result; + + load_addr.u64 = 0; + load_addr.sstatus.mem_region = CVMX_IO_SEG; + load_addr.sstatus.is_io = 1; + load_addr.sstatus.did = CVMX_OCT_DID_TAG_TAG1; + load_addr.sstatus.coreid = cvmx_get_core_num(); + load_addr.sstatus.get_cur = 1; + load_resp.u64 = cvmx_read_csr(load_addr.u64); + result.u64 = 0; + result.s.grp = load_resp.s_sstatus2.grp; + result.s.index = load_resp.s_sstatus2.index; + result.s.type = load_resp.s_sstatus2.tag_type; + result.s.tag = load_resp.s_sstatus2.tag; + return result; +} + +/** + * Get the POW WQE for this core. This returns the work queue + * entry currently associated with this core. + * + * Returns WQE pointer + */ +static inline struct cvmx_wqe *cvmx_pow_get_current_wqp(void) +{ + cvmx_pow_load_addr_t load_addr; + cvmx_pow_tag_load_resp_t load_resp; + + load_addr.u64 = 0; + load_addr.sstatus.mem_region = CVMX_IO_SEG; + load_addr.sstatus.is_io = 1; + load_addr.sstatus.did = CVMX_OCT_DID_TAG_TAG1; + load_addr.sstatus.coreid = cvmx_get_core_num(); + load_addr.sstatus.get_cur = 1; + load_addr.sstatus.get_wqp = 1; + load_resp.u64 = cvmx_read_csr(load_addr.u64); + return (struct cvmx_wqe *) cvmx_phys_to_ptr(load_resp.s_sstatus4.wqp); +} + +#ifndef CVMX_MF_CHORD +#define CVMX_MF_CHORD(dest) CVMX_RDHWR(dest, 30) +#endif + +/** + * Print a warning if a tag switch is pending for this core + * + * @function: Function name checking for a pending tag switch + */ +static inline void __cvmx_pow_warn_if_pending_switch(const char *function) +{ + uint64_t switch_complete; + CVMX_MF_CHORD(switch_complete); + if (!switch_complete) + pr_warn("%s called with tag switch in progress\n", function); +} + +/** + * Waits for a tag switch to complete by polling the completion bit. + * Note that switches to NULL complete immediately and do not need + * to be waited for. + */ +static inline void cvmx_pow_tag_sw_wait(void) +{ + const uint64_t MAX_CYCLES = 1ull << 31; + uint64_t switch_complete; + uint64_t start_cycle = cvmx_get_cycle(); + while (1) { + CVMX_MF_CHORD(switch_complete); + if (unlikely(switch_complete)) + break; + if (unlikely(cvmx_get_cycle() > start_cycle + MAX_CYCLES)) { + pr_warn("Tag switch is taking a long time, possible deadlock\n"); + start_cycle = -MAX_CYCLES - 1; + } + } +} + +/** + * Synchronous work request. Requests work from the POW. + * This function does NOT wait for previous tag switches to complete, + * so the caller must ensure that there is not a pending tag switch. + * + * @wait: When set, call stalls until work becomes avaiable, or times out. + * If not set, returns immediately. + * + * Returns: the WQE pointer from POW. Returns NULL if no work + * was available. + */ +static inline struct cvmx_wqe *cvmx_pow_work_request_sync_nocheck(cvmx_pow_wait_t + wait) +{ + cvmx_pow_load_addr_t ptr; + cvmx_pow_tag_load_resp_t result; + + if (CVMX_ENABLE_POW_CHECKS) + __cvmx_pow_warn_if_pending_switch(__func__); + + ptr.u64 = 0; + ptr.swork.mem_region = CVMX_IO_SEG; + ptr.swork.is_io = 1; + ptr.swork.did = CVMX_OCT_DID_TAG_SWTAG; + ptr.swork.wait = wait; + + result.u64 = cvmx_read_csr(ptr.u64); + + if (result.s_work.no_work) + return NULL; + else + return (struct cvmx_wqe *) cvmx_phys_to_ptr(result.s_work.addr); +} + +/** + * Synchronous work request. Requests work from the POW. + * This function waits for any previous tag switch to complete before + * requesting the new work. + * + * @wait: When set, call stalls until work becomes avaiable, or times out. + * If not set, returns immediately. + * + * Returns: the WQE pointer from POW. Returns NULL if no work + * was available. + */ +static inline struct cvmx_wqe *cvmx_pow_work_request_sync(cvmx_pow_wait_t wait) +{ + if (CVMX_ENABLE_POW_CHECKS) + __cvmx_pow_warn_if_pending_switch(__func__); + + /* Must not have a switch pending when requesting work */ + cvmx_pow_tag_sw_wait(); + return cvmx_pow_work_request_sync_nocheck(wait); + +} + +/** + * Synchronous null_rd request. Requests a switch out of NULL_NULL POW state. + * This function waits for any previous tag switch to complete before + * requesting the null_rd. + * + * Returns: the POW state of type cvmx_pow_tag_type_t. + */ +static inline enum cvmx_pow_tag_type cvmx_pow_work_request_null_rd(void) +{ + cvmx_pow_load_addr_t ptr; + cvmx_pow_tag_load_resp_t result; + + if (CVMX_ENABLE_POW_CHECKS) + __cvmx_pow_warn_if_pending_switch(__func__); + + /* Must not have a switch pending when requesting work */ + cvmx_pow_tag_sw_wait(); + + ptr.u64 = 0; + ptr.snull_rd.mem_region = CVMX_IO_SEG; + ptr.snull_rd.is_io = 1; + ptr.snull_rd.did = CVMX_OCT_DID_TAG_NULL_RD; + + result.u64 = cvmx_read_csr(ptr.u64); + + return (enum cvmx_pow_tag_type) result.s_null_rd.state; +} + +/** + * Asynchronous work request. Work is requested from the POW unit, + * and should later be checked with function + * cvmx_pow_work_response_async. This function does NOT wait for + * previous tag switches to complete, so the caller must ensure that + * there is not a pending tag switch. + * + * @scr_addr: Scratch memory address that response will be returned + * to, which is either a valid WQE, or a response with the + * invalid bit set. Byte address, must be 8 byte aligned. + * + * @wait: 1 to cause response to wait for work to become available (or + * timeout), 0 to cause response to return immediately + */ +static inline void cvmx_pow_work_request_async_nocheck(int scr_addr, + cvmx_pow_wait_t wait) +{ + cvmx_pow_iobdma_store_t data; + + if (CVMX_ENABLE_POW_CHECKS) + __cvmx_pow_warn_if_pending_switch(__func__); + + /* scr_addr must be 8 byte aligned */ + data.s.scraddr = scr_addr >> 3; + data.s.len = 1; + data.s.did = CVMX_OCT_DID_TAG_SWTAG; + data.s.wait = wait; + cvmx_send_single(data.u64); +} + +/** + * Asynchronous work request. Work is requested from the POW unit, + * and should later be checked with function + * cvmx_pow_work_response_async. This function waits for any previous + * tag switch to complete before requesting the new work. + * + * @scr_addr: Scratch memory address that response will be returned + * to, which is either a valid WQE, or a response with the + * invalid bit set. Byte address, must be 8 byte aligned. + * + * @wait: 1 to cause response to wait for work to become available (or + * timeout), 0 to cause response to return immediately + */ +static inline void cvmx_pow_work_request_async(int scr_addr, + cvmx_pow_wait_t wait) +{ + if (CVMX_ENABLE_POW_CHECKS) + __cvmx_pow_warn_if_pending_switch(__func__); + + /* Must not have a switch pending when requesting work */ + cvmx_pow_tag_sw_wait(); + cvmx_pow_work_request_async_nocheck(scr_addr, wait); +} + +/** + * Gets result of asynchronous work request. Performs a IOBDMA sync + * to wait for the response. + * + * @scr_addr: Scratch memory address to get result from Byte address, + * must be 8 byte aligned. + * + * Returns: the WQE from the scratch register, or NULL if no + * work was available. + */ +static inline struct cvmx_wqe *cvmx_pow_work_response_async(int scr_addr) +{ + cvmx_pow_tag_load_resp_t result; + + CVMX_SYNCIOBDMA; + result.u64 = cvmx_scratch_read64(scr_addr); + + if (result.s_work.no_work) + return NULL; + else + return (struct cvmx_wqe *) cvmx_phys_to_ptr(result.s_work.addr); +} + +/** + * Checks if a work queue entry pointer returned by a work + * request is valid. It may be invalid due to no work + * being available or due to a timeout. + * + * @wqe_ptr: pointer to a work queue entry returned by the POW + * + * Returns 0 if pointer is valid + * 1 if invalid (no work was returned) + */ +static inline uint64_t cvmx_pow_work_invalid(struct cvmx_wqe *wqe_ptr) +{ + return wqe_ptr == NULL; +} + +/** + * Starts a tag switch to the provided tag value and tag type. + * Completion for the tag switch must be checked for separately. This + * function does NOT update the work queue entry in dram to match tag + * value and type, so the application must keep track of these if they + * are important to the application. This tag switch command must not + * be used for switches to NULL, as the tag switch pending bit will be + * set by the switch request, but never cleared by the hardware. + * + * NOTE: This should not be used when switching from a NULL tag. Use + * cvmx_pow_tag_sw_full() instead. + * + * This function does no checks, so the caller must ensure that any + * previous tag switch has completed. + * + * @tag: new tag value + * @tag_type: new tag type (ordered or atomic) + */ +static inline void cvmx_pow_tag_sw_nocheck(uint32_t tag, + enum cvmx_pow_tag_type tag_type) +{ + cvmx_addr_t ptr; + cvmx_pow_tag_req_t tag_req; + + if (CVMX_ENABLE_POW_CHECKS) { + cvmx_pow_tag_req_t current_tag; + __cvmx_pow_warn_if_pending_switch(__func__); + current_tag = cvmx_pow_get_current_tag(); + if (current_tag.s.type == CVMX_POW_TAG_TYPE_NULL_NULL) + pr_warn("%s called with NULL_NULL tag\n", __func__); + if (current_tag.s.type == CVMX_POW_TAG_TYPE_NULL) + pr_warn("%s called with NULL tag\n", __func__); + if ((current_tag.s.type == tag_type) + && (current_tag.s.tag == tag)) + pr_warn("%s called to perform a tag switch to the same tag\n", + __func__); + if (tag_type == CVMX_POW_TAG_TYPE_NULL) + pr_warn("%s called to perform a tag switch to NULL. Use cvmx_pow_tag_sw_null() instead\n", + __func__); + } + + /* + * Note that WQE in DRAM is not updated here, as the POW does + * not read from DRAM once the WQE is in flight. See hardware + * manual for complete details. It is the application's + * responsibility to keep track of the current tag value if + * that is important. + */ + + tag_req.u64 = 0; + tag_req.s.op = CVMX_POW_TAG_OP_SWTAG; + tag_req.s.tag = tag; + tag_req.s.type = tag_type; + + ptr.u64 = 0; + ptr.sio.mem_region = CVMX_IO_SEG; + ptr.sio.is_io = 1; + ptr.sio.did = CVMX_OCT_DID_TAG_SWTAG; + + /* once this store arrives at POW, it will attempt the switch + software must wait for the switch to complete separately */ + cvmx_write_io(ptr.u64, tag_req.u64); +} + +/** + * Starts a tag switch to the provided tag value and tag type. + * Completion for the tag switch must be checked for separately. This + * function does NOT update the work queue entry in dram to match tag + * value and type, so the application must keep track of these if they + * are important to the application. This tag switch command must not + * be used for switches to NULL, as the tag switch pending bit will be + * set by the switch request, but never cleared by the hardware. + * + * NOTE: This should not be used when switching from a NULL tag. Use + * cvmx_pow_tag_sw_full() instead. + * + * This function waits for any previous tag switch to complete, and also + * displays an error on tag switches to NULL. + * + * @tag: new tag value + * @tag_type: new tag type (ordered or atomic) + */ +static inline void cvmx_pow_tag_sw(uint32_t tag, + enum cvmx_pow_tag_type tag_type) +{ + if (CVMX_ENABLE_POW_CHECKS) + __cvmx_pow_warn_if_pending_switch(__func__); + + /* + * Note that WQE in DRAM is not updated here, as the POW does + * not read from DRAM once the WQE is in flight. See hardware + * manual for complete details. It is the application's + * responsibility to keep track of the current tag value if + * that is important. + */ + + /* + * Ensure that there is not a pending tag switch, as a tag + * switch cannot be started if a previous switch is still + * pending. + */ + cvmx_pow_tag_sw_wait(); + cvmx_pow_tag_sw_nocheck(tag, tag_type); +} + +/** + * Starts a tag switch to the provided tag value and tag type. + * Completion for the tag switch must be checked for separately. This + * function does NOT update the work queue entry in dram to match tag + * value and type, so the application must keep track of these if they + * are important to the application. This tag switch command must not + * be used for switches to NULL, as the tag switch pending bit will be + * set by the switch request, but never cleared by the hardware. + * + * This function must be used for tag switches from NULL. + * + * This function does no checks, so the caller must ensure that any + * previous tag switch has completed. + * + * @wqp: pointer to work queue entry to submit. This entry is + * updated to match the other parameters + * @tag: tag value to be assigned to work queue entry + * @tag_type: type of tag + * @group: group value for the work queue entry. + */ +static inline void cvmx_pow_tag_sw_full_nocheck(struct cvmx_wqe *wqp, uint32_t tag, + enum cvmx_pow_tag_type tag_type, + uint64_t group) +{ + cvmx_addr_t ptr; + cvmx_pow_tag_req_t tag_req; + + if (CVMX_ENABLE_POW_CHECKS) { + cvmx_pow_tag_req_t current_tag; + __cvmx_pow_warn_if_pending_switch(__func__); + current_tag = cvmx_pow_get_current_tag(); + if (current_tag.s.type == CVMX_POW_TAG_TYPE_NULL_NULL) + pr_warn("%s called with NULL_NULL tag\n", __func__); + if ((current_tag.s.type == tag_type) + && (current_tag.s.tag == tag)) + pr_warn("%s called to perform a tag switch to the same tag\n", + __func__); + if (tag_type == CVMX_POW_TAG_TYPE_NULL) + pr_warn("%s called to perform a tag switch to NULL. Use cvmx_pow_tag_sw_null() instead\n", + __func__); + if (wqp != cvmx_phys_to_ptr(0x80)) + if (wqp != cvmx_pow_get_current_wqp()) + pr_warn("%s passed WQE(%p) doesn't match the address in the POW(%p)\n", + __func__, wqp, + cvmx_pow_get_current_wqp()); + } + + /* + * Note that WQE in DRAM is not updated here, as the POW does + * not read from DRAM once the WQE is in flight. See hardware + * manual for complete details. It is the application's + * responsibility to keep track of the current tag value if + * that is important. + */ + + tag_req.u64 = 0; + tag_req.s.op = CVMX_POW_TAG_OP_SWTAG_FULL; + tag_req.s.tag = tag; + tag_req.s.type = tag_type; + tag_req.s.grp = group; + + ptr.u64 = 0; + ptr.sio.mem_region = CVMX_IO_SEG; + ptr.sio.is_io = 1; + ptr.sio.did = CVMX_OCT_DID_TAG_SWTAG; + ptr.sio.offset = CAST64(wqp); + + /* + * once this store arrives at POW, it will attempt the switch + * software must wait for the switch to complete separately. + */ + cvmx_write_io(ptr.u64, tag_req.u64); +} + +/** + * Starts a tag switch to the provided tag value and tag type. + * Completion for the tag switch must be checked for separately. This + * function does NOT update the work queue entry in dram to match tag + * value and type, so the application must keep track of these if they + * are important to the application. This tag switch command must not + * be used for switches to NULL, as the tag switch pending bit will be + * set by the switch request, but never cleared by the hardware. + * + * This function must be used for tag switches from NULL. + * + * This function waits for any pending tag switches to complete + * before requesting the tag switch. + * + * @wqp: pointer to work queue entry to submit. This entry is updated + * to match the other parameters + * @tag: tag value to be assigned to work queue entry + * @tag_type: type of tag + * @group: group value for the work queue entry. + */ +static inline void cvmx_pow_tag_sw_full(struct cvmx_wqe *wqp, uint32_t tag, + enum cvmx_pow_tag_type tag_type, + uint64_t group) +{ + if (CVMX_ENABLE_POW_CHECKS) + __cvmx_pow_warn_if_pending_switch(__func__); + + /* + * Ensure that there is not a pending tag switch, as a tag + * switch cannot be started if a previous switch is still + * pending. + */ + cvmx_pow_tag_sw_wait(); + cvmx_pow_tag_sw_full_nocheck(wqp, tag, tag_type, group); +} + +/** + * Switch to a NULL tag, which ends any ordering or + * synchronization provided by the POW for the current + * work queue entry. This operation completes immediately, + * so completion should not be waited for. + * This function does NOT wait for previous tag switches to complete, + * so the caller must ensure that any previous tag switches have completed. + */ +static inline void cvmx_pow_tag_sw_null_nocheck(void) +{ + cvmx_addr_t ptr; + cvmx_pow_tag_req_t tag_req; + + if (CVMX_ENABLE_POW_CHECKS) { + cvmx_pow_tag_req_t current_tag; + __cvmx_pow_warn_if_pending_switch(__func__); + current_tag = cvmx_pow_get_current_tag(); + if (current_tag.s.type == CVMX_POW_TAG_TYPE_NULL_NULL) + pr_warn("%s called with NULL_NULL tag\n", __func__); + if (current_tag.s.type == CVMX_POW_TAG_TYPE_NULL) + pr_warn("%s called when we already have a NULL tag\n", + __func__); + } + + tag_req.u64 = 0; + tag_req.s.op = CVMX_POW_TAG_OP_SWTAG; + tag_req.s.type = CVMX_POW_TAG_TYPE_NULL; + + ptr.u64 = 0; + ptr.sio.mem_region = CVMX_IO_SEG; + ptr.sio.is_io = 1; + ptr.sio.did = CVMX_OCT_DID_TAG_TAG1; + + cvmx_write_io(ptr.u64, tag_req.u64); + + /* switch to NULL completes immediately */ +} + +/** + * Switch to a NULL tag, which ends any ordering or + * synchronization provided by the POW for the current + * work queue entry. This operation completes immediately, + * so completion should not be waited for. + * This function waits for any pending tag switches to complete + * before requesting the switch to NULL. + */ +static inline void cvmx_pow_tag_sw_null(void) +{ + if (CVMX_ENABLE_POW_CHECKS) + __cvmx_pow_warn_if_pending_switch(__func__); + + /* + * Ensure that there is not a pending tag switch, as a tag + * switch cannot be started if a previous switch is still + * pending. + */ + cvmx_pow_tag_sw_wait(); + cvmx_pow_tag_sw_null_nocheck(); + + /* switch to NULL completes immediately */ +} + +/** + * Submits work to an input queue. This function updates the work + * queue entry in DRAM to match the arguments given. Note that the + * tag provided is for the work queue entry submitted, and is + * unrelated to the tag that the core currently holds. + * + * @wqp: pointer to work queue entry to submit. This entry is + * updated to match the other parameters + * @tag: tag value to be assigned to work queue entry + * @tag_type: type of tag + * @qos: Input queue to add to. + * @grp: group value for the work queue entry. + */ +static inline void cvmx_pow_work_submit(struct cvmx_wqe *wqp, uint32_t tag, + enum cvmx_pow_tag_type tag_type, + uint64_t qos, uint64_t grp) +{ + cvmx_addr_t ptr; + cvmx_pow_tag_req_t tag_req; + + wqp->word1.tag = tag; + wqp->word1.tag_type = tag_type; + + cvmx_wqe_set_qos(wqp, qos); + cvmx_wqe_set_grp(wqp, grp); + + tag_req.u64 = 0; + tag_req.s.op = CVMX_POW_TAG_OP_ADDWQ; + tag_req.s.type = tag_type; + tag_req.s.tag = tag; + tag_req.s.qos = qos; + tag_req.s.grp = grp; + + ptr.u64 = 0; + ptr.sio.mem_region = CVMX_IO_SEG; + ptr.sio.is_io = 1; + ptr.sio.did = CVMX_OCT_DID_TAG_TAG1; + ptr.sio.offset = cvmx_ptr_to_phys(wqp); + + /* + * SYNC write to memory before the work submit. This is + * necessary as POW may read values from DRAM at this time. + */ + CVMX_SYNCWS; + cvmx_write_io(ptr.u64, tag_req.u64); +} + +/** + * This function sets the group mask for a core. The group mask + * indicates which groups each core will accept work from. There are + * 16 groups. + * + * @core_num: core to apply mask to + * @mask: Group mask. There are 16 groups, so only bits 0-15 are valid, + * representing groups 0-15. + * Each 1 bit in the mask enables the core to accept work from + * the corresponding group. + */ +static inline void cvmx_pow_set_group_mask(uint64_t core_num, uint64_t mask) +{ + union cvmx_pow_pp_grp_mskx grp_msk; + + grp_msk.u64 = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(core_num)); + grp_msk.s.grp_msk = mask; + cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(core_num), grp_msk.u64); +} + +/** + * This function sets POW static priorities for a core. Each input queue has + * an associated priority value. + * + * @core_num: core to apply priorities to + * @priority: Vector of 8 priorities, one per POW Input Queue (0-7). + * Highest priority is 0 and lowest is 7. A priority value + * of 0xF instructs POW to skip the Input Queue when + * scheduling to this specific core. + * NOTE: priorities should not have gaps in values, meaning + * {0,1,1,1,1,1,1,1} is a valid configuration while + * {0,2,2,2,2,2,2,2} is not. + */ +static inline void cvmx_pow_set_priority(uint64_t core_num, + const uint8_t priority[]) +{ + /* POW priorities are supported on CN5xxx and later */ + if (!OCTEON_IS_MODEL(OCTEON_CN3XXX)) { + union cvmx_pow_pp_grp_mskx grp_msk; + + grp_msk.u64 = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(core_num)); + grp_msk.s.qos0_pri = priority[0]; + grp_msk.s.qos1_pri = priority[1]; + grp_msk.s.qos2_pri = priority[2]; + grp_msk.s.qos3_pri = priority[3]; + grp_msk.s.qos4_pri = priority[4]; + grp_msk.s.qos5_pri = priority[5]; + grp_msk.s.qos6_pri = priority[6]; + grp_msk.s.qos7_pri = priority[7]; + + /* Detect gaps between priorities and flag error */ + { + int i; + uint32_t prio_mask = 0; + + for (i = 0; i < 8; i++) + if (priority[i] != 0xF) + prio_mask |= 1 << priority[i]; + + if (prio_mask ^ ((1 << cvmx_pop(prio_mask)) - 1)) { + pr_err("POW static priorities should be " + "contiguous (0x%llx)\n", + (unsigned long long)prio_mask); + return; + } + } + + cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(core_num), grp_msk.u64); + } +} + +/** + * Performs a tag switch and then an immediate deschedule. This completes + * immediately, so completion must not be waited for. This function does NOT + * update the wqe in DRAM to match arguments. + * + * This function does NOT wait for any prior tag switches to complete, so the + * calling code must do this. + * + * Note the following CAVEAT of the Octeon HW behavior when + * re-scheduling DE-SCHEDULEd items whose (next) state is + * ORDERED: + * - If there are no switches pending at the time that the + * HW executes the de-schedule, the HW will only re-schedule + * the head of the FIFO associated with the given tag. This + * means that in many respects, the HW treats this ORDERED + * tag as an ATOMIC tag. Note that in the SWTAG_DESCH + * case (to an ORDERED tag), the HW will do the switch + * before the deschedule whenever it is possible to do + * the switch immediately, so it may often look like + * this case. + * - If there is a pending switch to ORDERED at the time + * the HW executes the de-schedule, the HW will perform + * the switch at the time it re-schedules, and will be + * able to reschedule any/all of the entries with the + * same tag. + * Due to this behavior, the RECOMMENDATION to software is + * that they have a (next) state of ATOMIC when they + * DE-SCHEDULE. If an ORDERED tag is what was really desired, + * SW can choose to immediately switch to an ORDERED tag + * after the work (that has an ATOMIC tag) is re-scheduled. + * Note that since there are never any tag switches pending + * when the HW re-schedules, this switch can be IMMEDIATE upon + * the reception of the pointer during the re-schedule. + * + * @tag: New tag value + * @tag_type: New tag type + * @group: New group value + * @no_sched: Control whether this work queue entry will be rescheduled. + * - 1 : don't schedule this work + * - 0 : allow this work to be scheduled. + */ +static inline void cvmx_pow_tag_sw_desched_nocheck( + uint32_t tag, + enum cvmx_pow_tag_type tag_type, + uint64_t group, + uint64_t no_sched) +{ + cvmx_addr_t ptr; + cvmx_pow_tag_req_t tag_req; + + if (CVMX_ENABLE_POW_CHECKS) { + cvmx_pow_tag_req_t current_tag; + __cvmx_pow_warn_if_pending_switch(__func__); + current_tag = cvmx_pow_get_current_tag(); + if (current_tag.s.type == CVMX_POW_TAG_TYPE_NULL_NULL) + pr_warn("%s called with NULL_NULL tag\n", __func__); + if (current_tag.s.type == CVMX_POW_TAG_TYPE_NULL) + pr_warn("%s called with NULL tag. Deschedule not allowed from NULL state\n", + __func__); + if ((current_tag.s.type != CVMX_POW_TAG_TYPE_ATOMIC) + && (tag_type != CVMX_POW_TAG_TYPE_ATOMIC)) + pr_warn("%s called where neither the before or after tag is ATOMIC\n", + __func__); + } + + tag_req.u64 = 0; + tag_req.s.op = CVMX_POW_TAG_OP_SWTAG_DESCH; + tag_req.s.tag = tag; + tag_req.s.type = tag_type; + tag_req.s.grp = group; + tag_req.s.no_sched = no_sched; + + ptr.u64 = 0; + ptr.sio.mem_region = CVMX_IO_SEG; + ptr.sio.is_io = 1; + ptr.sio.did = CVMX_OCT_DID_TAG_TAG3; + /* + * since TAG3 is used, this store will clear the local pending + * switch bit. + */ + cvmx_write_io(ptr.u64, tag_req.u64); +} + +/** + * Performs a tag switch and then an immediate deschedule. This completes + * immediately, so completion must not be waited for. This function does NOT + * update the wqe in DRAM to match arguments. + * + * This function waits for any prior tag switches to complete, so the + * calling code may call this function with a pending tag switch. + * + * Note the following CAVEAT of the Octeon HW behavior when + * re-scheduling DE-SCHEDULEd items whose (next) state is + * ORDERED: + * - If there are no switches pending at the time that the + * HW executes the de-schedule, the HW will only re-schedule + * the head of the FIFO associated with the given tag. This + * means that in many respects, the HW treats this ORDERED + * tag as an ATOMIC tag. Note that in the SWTAG_DESCH + * case (to an ORDERED tag), the HW will do the switch + * before the deschedule whenever it is possible to do + * the switch immediately, so it may often look like + * this case. + * - If there is a pending switch to ORDERED at the time + * the HW executes the de-schedule, the HW will perform + * the switch at the time it re-schedules, and will be + * able to reschedule any/all of the entries with the + * same tag. + * Due to this behavior, the RECOMMENDATION to software is + * that they have a (next) state of ATOMIC when they + * DE-SCHEDULE. If an ORDERED tag is what was really desired, + * SW can choose to immediately switch to an ORDERED tag + * after the work (that has an ATOMIC tag) is re-scheduled. + * Note that since there are never any tag switches pending + * when the HW re-schedules, this switch can be IMMEDIATE upon + * the reception of the pointer during the re-schedule. + * + * @tag: New tag value + * @tag_type: New tag type + * @group: New group value + * @no_sched: Control whether this work queue entry will be rescheduled. + * - 1 : don't schedule this work + * - 0 : allow this work to be scheduled. + */ +static inline void cvmx_pow_tag_sw_desched(uint32_t tag, + enum cvmx_pow_tag_type tag_type, + uint64_t group, uint64_t no_sched) +{ + if (CVMX_ENABLE_POW_CHECKS) + __cvmx_pow_warn_if_pending_switch(__func__); + + /* Need to make sure any writes to the work queue entry are complete */ + CVMX_SYNCWS; + /* + * Ensure that there is not a pending tag switch, as a tag + * switch cannot be started if a previous switch is still + * pending. + */ + cvmx_pow_tag_sw_wait(); + cvmx_pow_tag_sw_desched_nocheck(tag, tag_type, group, no_sched); +} + +/** + * Deschedules the current work queue entry. + * + * @no_sched: no schedule flag value to be set on the work queue + * entry. If this is set the entry will not be + * rescheduled. + */ +static inline void cvmx_pow_desched(uint64_t no_sched) +{ + cvmx_addr_t ptr; + cvmx_pow_tag_req_t tag_req; + + if (CVMX_ENABLE_POW_CHECKS) { + cvmx_pow_tag_req_t current_tag; + __cvmx_pow_warn_if_pending_switch(__func__); + current_tag = cvmx_pow_get_current_tag(); + if (current_tag.s.type == CVMX_POW_TAG_TYPE_NULL_NULL) + pr_warn("%s called with NULL_NULL tag\n", __func__); + if (current_tag.s.type == CVMX_POW_TAG_TYPE_NULL) + pr_warn("%s called with NULL tag. Deschedule not expected from NULL state\n", + __func__); + } + + /* Need to make sure any writes to the work queue entry are complete */ + CVMX_SYNCWS; + + tag_req.u64 = 0; + tag_req.s.op = CVMX_POW_TAG_OP_DESCH; + tag_req.s.no_sched = no_sched; + + ptr.u64 = 0; + ptr.sio.mem_region = CVMX_IO_SEG; + ptr.sio.is_io = 1; + ptr.sio.did = CVMX_OCT_DID_TAG_TAG3; + /* + * since TAG3 is used, this store will clear the local pending + * switch bit. + */ + cvmx_write_io(ptr.u64, tag_req.u64); +} + +/**************************************************** +* Define usage of bits within the 32 bit tag values. +*****************************************************/ + +/* + * Number of bits of the tag used by software. The SW bits are always + * a contiguous block of the high starting at bit 31. The hardware + * bits are always the low bits. By default, the top 8 bits of the + * tag are reserved for software, and the low 24 are set by the IPD + * unit. + */ +#define CVMX_TAG_SW_BITS (8) +#define CVMX_TAG_SW_SHIFT (32 - CVMX_TAG_SW_BITS) + +/* Below is the list of values for the top 8 bits of the tag. */ +/* + * Tag values with top byte of this value are reserved for internal + * executive uses. + */ +#define CVMX_TAG_SW_BITS_INTERNAL 0x1 +/* The executive divides the remaining 24 bits as follows: + * - the upper 8 bits (bits 23 - 16 of the tag) define a subgroup + * + * - the lower 16 bits (bits 15 - 0 of the tag) define are the value + * with the subgroup + * + * Note that this section describes the format of tags generated by + * software - refer to the hardware documentation for a description of + * the tags values generated by the packet input hardware. Subgroups + * are defined here. + */ +/* Mask for the value portion of the tag */ +#define CVMX_TAG_SUBGROUP_MASK 0xFFFF +#define CVMX_TAG_SUBGROUP_SHIFT 16 +#define CVMX_TAG_SUBGROUP_PKO 0x1 + +/* End of executive tag subgroup definitions */ + +/* + * The remaining values software bit values 0x2 - 0xff are available + * for application use. + */ + +/** + * This function creates a 32 bit tag value from the two values provided. + * + * @sw_bits: The upper bits (number depends on configuration) are set + * to this value. The remainder of bits are set by the + * hw_bits parameter. + * + * @hw_bits: The lower bits (number depends on configuration) are set + * to this value. The remainder of bits are set by the + * sw_bits parameter. + * + * Returns 32 bit value of the combined hw and sw bits. + */ +static inline uint32_t cvmx_pow_tag_compose(uint64_t sw_bits, uint64_t hw_bits) +{ + return ((sw_bits & cvmx_build_mask(CVMX_TAG_SW_BITS)) << + CVMX_TAG_SW_SHIFT) | + (hw_bits & cvmx_build_mask(32 - CVMX_TAG_SW_BITS)); +} + +/** + * Extracts the bits allocated for software use from the tag + * + * @tag: 32 bit tag value + * + * Returns N bit software tag value, where N is configurable with the + * CVMX_TAG_SW_BITS define + */ +static inline uint32_t cvmx_pow_tag_get_sw_bits(uint64_t tag) +{ + return (tag >> (32 - CVMX_TAG_SW_BITS)) & + cvmx_build_mask(CVMX_TAG_SW_BITS); +} + +/** + * + * Extracts the bits allocated for hardware use from the tag + * + * @tag: 32 bit tag value + * + * Returns (32 - N) bit software tag value, where N is configurable + * with the CVMX_TAG_SW_BITS define + */ +static inline uint32_t cvmx_pow_tag_get_hw_bits(uint64_t tag) +{ + return tag & cvmx_build_mask(32 - CVMX_TAG_SW_BITS); +} + +/** + * Store the current POW internal state into the supplied + * buffer. It is recommended that you pass a buffer of at least + * 128KB. The format of the capture may change based on SDK + * version and Octeon chip. + * + * @buffer: Buffer to store capture into + * @buffer_size: + * The size of the supplied buffer + * + * Returns Zero on success, negative on failure + */ +extern int cvmx_pow_capture(void *buffer, int buffer_size); + +/** + * Dump a POW capture to the console in a human readable format. + * + * @buffer: POW capture from cvmx_pow_capture() + * @buffer_size: + * Size of the buffer + */ +extern void cvmx_pow_display(void *buffer, int buffer_size); + +/** + * Return the number of POW entries supported by this chip + * + * Returns Number of POW entries + */ +extern int cvmx_pow_get_num_entries(void); + +#endif /* __CVMX_POW_H__ */ |