2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2005 Cisco Systems, Inc. All rights reserved.
5 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
8 * This software is available to you under a choice of one of two
9 * licenses. You may choose to be licensed under the terms of the GNU
10 * General Public License (GPL) Version 2, available from the file
11 * COPYING in the main directory of this source tree, or the
12 * OpenIB.org BSD license below:
14 * Redistribution and use in source and binary forms, with or
15 * without modification, are permitted provided that the following
18 * - Redistributions of source code must retain the above
19 * copyright notice, this list of conditions and the following
22 * - Redistributions in binary form must reproduce the above
23 * copyright notice, this list of conditions and the following
24 * disclaimer in the documentation and/or other materials
25 * provided with the distribution.
27 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 * $Id: mthca_cq.c 1369 2004-12-20 16:17:07Z roland $
39 #include <linux/init.h>
40 #include <linux/hardirq.h>
42 #include <rdma/ib_pack.h>
44 #include "mthca_dev.h"
45 #include "mthca_cmd.h"
46 #include "mthca_memfree.h"
49 MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
53 MTHCA_CQ_ENTRY_SIZE = 0x20
57 * Must be packed because start is 64 bits but only aligned to 32 bits.
59 struct mthca_cq_context {
62 __be32 logsize_usrpage;
63 __be32 error_eqn; /* Tavor only */
67 __be32 last_notified_index;
68 __be32 solicit_producer_index;
69 __be32 consumer_index;
70 __be32 producer_index;
72 __be32 ci_db; /* Arbel only */
73 __be32 state_db; /* Arbel only */
75 } __attribute__((packed));
77 #define MTHCA_CQ_STATUS_OK ( 0 << 28)
78 #define MTHCA_CQ_STATUS_OVERFLOW ( 9 << 28)
79 #define MTHCA_CQ_STATUS_WRITE_FAIL (10 << 28)
80 #define MTHCA_CQ_FLAG_TR ( 1 << 18)
81 #define MTHCA_CQ_FLAG_OI ( 1 << 17)
82 #define MTHCA_CQ_STATE_DISARMED ( 0 << 8)
83 #define MTHCA_CQ_STATE_ARMED ( 1 << 8)
84 #define MTHCA_CQ_STATE_ARMED_SOL ( 4 << 8)
85 #define MTHCA_EQ_STATE_FIRED (10 << 8)
88 MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
92 SYNDROME_LOCAL_LENGTH_ERR = 0x01,
93 SYNDROME_LOCAL_QP_OP_ERR = 0x02,
94 SYNDROME_LOCAL_EEC_OP_ERR = 0x03,
95 SYNDROME_LOCAL_PROT_ERR = 0x04,
96 SYNDROME_WR_FLUSH_ERR = 0x05,
97 SYNDROME_MW_BIND_ERR = 0x06,
98 SYNDROME_BAD_RESP_ERR = 0x10,
99 SYNDROME_LOCAL_ACCESS_ERR = 0x11,
100 SYNDROME_REMOTE_INVAL_REQ_ERR = 0x12,
101 SYNDROME_REMOTE_ACCESS_ERR = 0x13,
102 SYNDROME_REMOTE_OP_ERR = 0x14,
103 SYNDROME_RETRY_EXC_ERR = 0x15,
104 SYNDROME_RNR_RETRY_EXC_ERR = 0x16,
105 SYNDROME_LOCAL_RDD_VIOL_ERR = 0x20,
106 SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
107 SYNDROME_REMOTE_ABORTED_ERR = 0x22,
108 SYNDROME_INVAL_EECN_ERR = 0x23,
109 SYNDROME_INVAL_EEC_STATE_ERR = 0x24
118 __be32 imm_etype_pkey_eec;
127 struct mthca_err_cqe {
140 #define MTHCA_CQ_ENTRY_OWNER_SW (0 << 7)
141 #define MTHCA_CQ_ENTRY_OWNER_HW (1 << 7)
143 #define MTHCA_TAVOR_CQ_DB_INC_CI (1 << 24)
144 #define MTHCA_TAVOR_CQ_DB_REQ_NOT (2 << 24)
145 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL (3 << 24)
146 #define MTHCA_TAVOR_CQ_DB_SET_CI (4 << 24)
147 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
149 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL (1 << 24)
150 #define MTHCA_ARBEL_CQ_DB_REQ_NOT (2 << 24)
151 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
153 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
156 return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
158 return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
159 + (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
162 static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
164 struct mthca_cqe *cqe = get_cqe(cq, i);
165 return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
168 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
170 return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
173 static inline void set_cqe_hw(struct mthca_cqe *cqe)
175 cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
178 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
180 __be32 *cqe = cqe_ptr;
182 (void) cqe; /* avoid warning if mthca_dbg compiled away... */
183 mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
184 be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
185 be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
186 be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
190 * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
191 * should be correct before calling update_cons_index().
193 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
198 if (mthca_is_memfree(dev)) {
199 *cq->set_ci_db = cpu_to_be32(cq->cons_index);
202 doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
203 doorbell[1] = cpu_to_be32(incr - 1);
205 mthca_write64(doorbell,
206 dev->kar + MTHCA_CQ_DOORBELL,
207 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
211 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
215 cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
218 mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
224 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
227 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
228 enum ib_event_type event_type)
231 struct ib_event event;
233 spin_lock(&dev->cq_table.lock);
235 cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
238 atomic_inc(&cq->refcount);
239 spin_unlock(&dev->cq_table.lock);
242 mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
246 event.device = &dev->ib_dev;
247 event.event = event_type;
248 event.element.cq = &cq->ibcq;
249 if (cq->ibcq.event_handler)
250 cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
252 if (atomic_dec_and_test(&cq->refcount))
256 void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
257 struct mthca_srq *srq)
260 struct mthca_cqe *cqe;
264 spin_lock_irq(&dev->cq_table.lock);
265 cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
267 atomic_inc(&cq->refcount);
268 spin_unlock_irq(&dev->cq_table.lock);
273 spin_lock_irq(&cq->lock);
276 * First we need to find the current producer index, so we
277 * know where to start cleaning from. It doesn't matter if HW
278 * adds new entries after this loop -- the QP we're worried
279 * about is already in RESET, so the new entries won't come
280 * from our QP and therefore don't need to be checked.
282 for (prod_index = cq->cons_index;
283 cqe_sw(cq, prod_index & cq->ibcq.cqe);
285 if (prod_index == cq->cons_index + cq->ibcq.cqe)
289 mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
290 qpn, cqn, cq->cons_index, prod_index);
293 * Now sweep backwards through the CQ, removing CQ entries
294 * that match our QP by copying older entries on top of them.
296 while ((int) --prod_index - (int) cq->cons_index >= 0) {
297 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
298 if (cqe->my_qpn == cpu_to_be32(qpn)) {
300 mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
303 memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
304 cqe, MTHCA_CQ_ENTRY_SIZE);
309 cq->cons_index += nfreed;
310 update_cons_index(dev, cq, nfreed);
313 spin_unlock_irq(&cq->lock);
314 if (atomic_dec_and_test(&cq->refcount))
318 static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
319 struct mthca_qp *qp, int wqe_index, int is_send,
320 struct mthca_err_cqe *cqe,
321 struct ib_wc *entry, int *free_cqe)
327 if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
328 mthca_dbg(dev, "local QP operation err "
329 "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
330 be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
331 cq->cqn, cq->cons_index);
336 * For completions in error, only work request ID, status (and
337 * freed resource count for RD) have to be set.
339 switch (cqe->syndrome) {
340 case SYNDROME_LOCAL_LENGTH_ERR:
341 entry->status = IB_WC_LOC_LEN_ERR;
343 case SYNDROME_LOCAL_QP_OP_ERR:
344 entry->status = IB_WC_LOC_QP_OP_ERR;
346 case SYNDROME_LOCAL_EEC_OP_ERR:
347 entry->status = IB_WC_LOC_EEC_OP_ERR;
349 case SYNDROME_LOCAL_PROT_ERR:
350 entry->status = IB_WC_LOC_PROT_ERR;
352 case SYNDROME_WR_FLUSH_ERR:
353 entry->status = IB_WC_WR_FLUSH_ERR;
355 case SYNDROME_MW_BIND_ERR:
356 entry->status = IB_WC_MW_BIND_ERR;
358 case SYNDROME_BAD_RESP_ERR:
359 entry->status = IB_WC_BAD_RESP_ERR;
361 case SYNDROME_LOCAL_ACCESS_ERR:
362 entry->status = IB_WC_LOC_ACCESS_ERR;
364 case SYNDROME_REMOTE_INVAL_REQ_ERR:
365 entry->status = IB_WC_REM_INV_REQ_ERR;
367 case SYNDROME_REMOTE_ACCESS_ERR:
368 entry->status = IB_WC_REM_ACCESS_ERR;
370 case SYNDROME_REMOTE_OP_ERR:
371 entry->status = IB_WC_REM_OP_ERR;
373 case SYNDROME_RETRY_EXC_ERR:
374 entry->status = IB_WC_RETRY_EXC_ERR;
376 case SYNDROME_RNR_RETRY_EXC_ERR:
377 entry->status = IB_WC_RNR_RETRY_EXC_ERR;
379 case SYNDROME_LOCAL_RDD_VIOL_ERR:
380 entry->status = IB_WC_LOC_RDD_VIOL_ERR;
382 case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
383 entry->status = IB_WC_REM_INV_RD_REQ_ERR;
385 case SYNDROME_REMOTE_ABORTED_ERR:
386 entry->status = IB_WC_REM_ABORT_ERR;
388 case SYNDROME_INVAL_EECN_ERR:
389 entry->status = IB_WC_INV_EECN_ERR;
391 case SYNDROME_INVAL_EEC_STATE_ERR:
392 entry->status = IB_WC_INV_EEC_STATE_ERR;
395 entry->status = IB_WC_GENERAL_ERR;
400 * Mem-free HCAs always generate one CQE per WQE, even in the
401 * error case, so we don't have to check the doorbell count, etc.
403 if (mthca_is_memfree(dev))
406 err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
411 * If we're at the end of the WQE chain, or we've used up our
412 * doorbell count, free the CQE. Otherwise just update it for
413 * the next poll operation.
415 if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
418 cqe->db_cnt = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
420 cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
427 static inline int mthca_poll_one(struct mthca_dev *dev,
429 struct mthca_qp **cur_qp,
434 struct mthca_cqe *cqe;
441 cqe = next_cqe_sw(cq);
446 * Make sure we read CQ entry contents after we've checked the
452 mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
453 cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
454 be32_to_cpu(cqe->wqe));
458 is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
459 MTHCA_ERROR_CQE_OPCODE_MASK;
460 is_send = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
462 if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
464 * We do not have to take the QP table lock here,
465 * because CQs will be locked while QPs are removed
468 *cur_qp = mthca_array_get(&dev->qp_table.qp,
469 be32_to_cpu(cqe->my_qpn) &
470 (dev->limits.num_qps - 1));
472 mthca_warn(dev, "CQ entry for unknown QP %06x\n",
473 be32_to_cpu(cqe->my_qpn) & 0xffffff);
479 entry->qp_num = (*cur_qp)->qpn;
483 wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
485 entry->wr_id = (*cur_qp)->wrid[wqe_index +
487 } else if ((*cur_qp)->ibqp.srq) {
488 struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
489 u32 wqe = be32_to_cpu(cqe->wqe);
491 wqe_index = wqe >> srq->wqe_shift;
492 entry->wr_id = srq->wrid[wqe_index];
493 mthca_free_srq_wqe(srq, wqe);
496 wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
497 entry->wr_id = (*cur_qp)->wrid[wqe_index];
501 if (wq->last_comp < wqe_index)
502 wq->tail += wqe_index - wq->last_comp;
504 wq->tail += wqe_index + wq->max - wq->last_comp;
506 wq->last_comp = wqe_index;
510 err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
511 (struct mthca_err_cqe *) cqe,
518 switch (cqe->opcode) {
519 case MTHCA_OPCODE_RDMA_WRITE:
520 entry->opcode = IB_WC_RDMA_WRITE;
522 case MTHCA_OPCODE_RDMA_WRITE_IMM:
523 entry->opcode = IB_WC_RDMA_WRITE;
524 entry->wc_flags |= IB_WC_WITH_IMM;
526 case MTHCA_OPCODE_SEND:
527 entry->opcode = IB_WC_SEND;
529 case MTHCA_OPCODE_SEND_IMM:
530 entry->opcode = IB_WC_SEND;
531 entry->wc_flags |= IB_WC_WITH_IMM;
533 case MTHCA_OPCODE_RDMA_READ:
534 entry->opcode = IB_WC_RDMA_READ;
535 entry->byte_len = be32_to_cpu(cqe->byte_cnt);
537 case MTHCA_OPCODE_ATOMIC_CS:
538 entry->opcode = IB_WC_COMP_SWAP;
539 entry->byte_len = be32_to_cpu(cqe->byte_cnt);
541 case MTHCA_OPCODE_ATOMIC_FA:
542 entry->opcode = IB_WC_FETCH_ADD;
543 entry->byte_len = be32_to_cpu(cqe->byte_cnt);
545 case MTHCA_OPCODE_BIND_MW:
546 entry->opcode = IB_WC_BIND_MW;
549 entry->opcode = MTHCA_OPCODE_INVALID;
553 entry->byte_len = be32_to_cpu(cqe->byte_cnt);
554 switch (cqe->opcode & 0x1f) {
555 case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
556 case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
557 entry->wc_flags = IB_WC_WITH_IMM;
558 entry->imm_data = cqe->imm_etype_pkey_eec;
559 entry->opcode = IB_WC_RECV;
561 case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
562 case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
563 entry->wc_flags = IB_WC_WITH_IMM;
564 entry->imm_data = cqe->imm_etype_pkey_eec;
565 entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
569 entry->opcode = IB_WC_RECV;
572 entry->slid = be16_to_cpu(cqe->rlid);
573 entry->sl = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
574 entry->src_qp = be32_to_cpu(cqe->rqpn) & 0xffffff;
575 entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
576 entry->pkey_index = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
577 entry->wc_flags |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
581 entry->status = IB_WC_SUCCESS;
584 if (likely(free_cqe)) {
593 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
596 struct mthca_dev *dev = to_mdev(ibcq->device);
597 struct mthca_cq *cq = to_mcq(ibcq);
598 struct mthca_qp *qp = NULL;
604 spin_lock_irqsave(&cq->lock, flags);
606 for (npolled = 0; npolled < num_entries; ++npolled) {
607 err = mthca_poll_one(dev, cq, &qp,
608 &freed, entry + npolled);
615 update_cons_index(dev, cq, freed);
618 spin_unlock_irqrestore(&cq->lock, flags);
620 return err == 0 || err == -EAGAIN ? npolled : err;
623 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
627 doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
628 MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
629 MTHCA_TAVOR_CQ_DB_REQ_NOT) |
631 doorbell[1] = (__force __be32) 0xffffffff;
633 mthca_write64(doorbell,
634 to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
635 MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
640 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
642 struct mthca_cq *cq = to_mcq(ibcq);
648 ci = cpu_to_be32(cq->cons_index);
651 doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
652 (notify == IB_CQ_SOLICITED ? 1 : 2));
654 mthca_write_db_rec(doorbell, cq->arm_db);
657 * Make sure that the doorbell record in host memory is
658 * written before ringing the doorbell via PCI MMIO.
662 doorbell[0] = cpu_to_be32((sn << 28) |
663 (notify == IB_CQ_SOLICITED ?
664 MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
665 MTHCA_ARBEL_CQ_DB_REQ_NOT) |
669 mthca_write64(doorbell,
670 to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
671 MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
676 static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
678 mthca_buf_free(dev, (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
679 &cq->queue, cq->is_direct, &cq->mr);
682 int mthca_init_cq(struct mthca_dev *dev, int nent,
683 struct mthca_ucontext *ctx, u32 pdn,
686 int size = nent * MTHCA_CQ_ENTRY_SIZE;
687 struct mthca_mailbox *mailbox;
688 struct mthca_cq_context *cq_context;
695 cq->ibcq.cqe = nent - 1;
696 cq->is_kernel = !ctx;
698 cq->cqn = mthca_alloc(&dev->cq_table.alloc);
702 if (mthca_is_memfree(dev)) {
703 err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
712 cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
713 cq->cqn, &cq->set_ci_db);
714 if (cq->set_ci_db_index < 0)
717 cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
718 cq->cqn, &cq->arm_db);
719 if (cq->arm_db_index < 0)
724 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
728 cq_context = mailbox->buf;
731 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_CQ_SIZE,
732 &cq->queue, &cq->is_direct,
733 &dev->driver_pd, 1, &cq->mr);
735 goto err_out_mailbox;
737 for (i = 0; i < nent; ++i)
738 set_cqe_hw(get_cqe(cq, i));
741 spin_lock_init(&cq->lock);
742 atomic_set(&cq->refcount, 1);
743 init_waitqueue_head(&cq->wait);
745 memset(cq_context, 0, sizeof *cq_context);
746 cq_context->flags = cpu_to_be32(MTHCA_CQ_STATUS_OK |
747 MTHCA_CQ_STATE_DISARMED |
749 cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
751 cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
753 cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
754 cq_context->error_eqn = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
755 cq_context->comp_eqn = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
756 cq_context->pd = cpu_to_be32(pdn);
757 cq_context->lkey = cpu_to_be32(cq->mr.ibmr.lkey);
758 cq_context->cqn = cpu_to_be32(cq->cqn);
760 if (mthca_is_memfree(dev)) {
761 cq_context->ci_db = cpu_to_be32(cq->set_ci_db_index);
762 cq_context->state_db = cpu_to_be32(cq->arm_db_index);
765 err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
767 mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
768 goto err_out_free_mr;
772 mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
775 goto err_out_free_mr;
778 spin_lock_irq(&dev->cq_table.lock);
779 if (mthca_array_set(&dev->cq_table.cq,
780 cq->cqn & (dev->limits.num_cqs - 1),
782 spin_unlock_irq(&dev->cq_table.lock);
783 goto err_out_free_mr;
785 spin_unlock_irq(&dev->cq_table.lock);
789 mthca_free_mailbox(dev, mailbox);
795 mthca_free_cq_buf(dev, cq);
798 mthca_free_mailbox(dev, mailbox);
801 if (cq->is_kernel && mthca_is_memfree(dev))
802 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
805 if (cq->is_kernel && mthca_is_memfree(dev))
806 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
809 mthca_table_put(dev, dev->cq_table.table, cq->cqn);
812 mthca_free(&dev->cq_table.alloc, cq->cqn);
817 void mthca_free_cq(struct mthca_dev *dev,
820 struct mthca_mailbox *mailbox;
826 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
827 if (IS_ERR(mailbox)) {
828 mthca_warn(dev, "No memory for mailbox to free CQ.\n");
832 err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
834 mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
836 mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
839 __be32 *ctx = mailbox->buf;
842 printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
843 cq->cqn, cq->cons_index,
844 cq->is_kernel ? !!next_cqe_sw(cq) : 0);
845 for (j = 0; j < 16; ++j)
846 printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
849 spin_lock_irq(&dev->cq_table.lock);
850 mthca_array_clear(&dev->cq_table.cq,
851 cq->cqn & (dev->limits.num_cqs - 1));
852 spin_unlock_irq(&dev->cq_table.lock);
854 if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
855 synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
857 synchronize_irq(dev->pdev->irq);
859 atomic_dec(&cq->refcount);
860 wait_event(cq->wait, !atomic_read(&cq->refcount));
863 mthca_free_cq_buf(dev, cq);
864 if (mthca_is_memfree(dev)) {
865 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
866 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
870 mthca_table_put(dev, dev->cq_table.table, cq->cqn);
871 mthca_free(&dev->cq_table.alloc, cq->cqn);
872 mthca_free_mailbox(dev, mailbox);
875 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
879 spin_lock_init(&dev->cq_table.lock);
881 err = mthca_alloc_init(&dev->cq_table.alloc,
884 dev->limits.reserved_cqs);
888 err = mthca_array_init(&dev->cq_table.cq,
889 dev->limits.num_cqs);
891 mthca_alloc_cleanup(&dev->cq_table.alloc);
896 void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
898 mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
899 mthca_alloc_cleanup(&dev->cq_table.alloc);