2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2005, 2006 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
37 #include <linux/hardirq.h>
38 #include <linux/sched.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 MTHCA_ATOMIC_BYTE_LEN = 8
61 * Must be packed because start is 64 bits but only aligned to 32 bits.
63 struct mthca_cq_context {
66 __be32 logsize_usrpage;
67 __be32 error_eqn; /* Tavor only */
71 __be32 last_notified_index;
72 __be32 solicit_producer_index;
73 __be32 consumer_index;
74 __be32 producer_index;
76 __be32 ci_db; /* Arbel only */
77 __be32 state_db; /* Arbel only */
79 } __attribute__((packed));
81 #define MTHCA_CQ_STATUS_OK ( 0 << 28)
82 #define MTHCA_CQ_STATUS_OVERFLOW ( 9 << 28)
83 #define MTHCA_CQ_STATUS_WRITE_FAIL (10 << 28)
84 #define MTHCA_CQ_FLAG_TR ( 1 << 18)
85 #define MTHCA_CQ_FLAG_OI ( 1 << 17)
86 #define MTHCA_CQ_STATE_DISARMED ( 0 << 8)
87 #define MTHCA_CQ_STATE_ARMED ( 1 << 8)
88 #define MTHCA_CQ_STATE_ARMED_SOL ( 4 << 8)
89 #define MTHCA_EQ_STATE_FIRED (10 << 8)
92 MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
96 SYNDROME_LOCAL_LENGTH_ERR = 0x01,
97 SYNDROME_LOCAL_QP_OP_ERR = 0x02,
98 SYNDROME_LOCAL_EEC_OP_ERR = 0x03,
99 SYNDROME_LOCAL_PROT_ERR = 0x04,
100 SYNDROME_WR_FLUSH_ERR = 0x05,
101 SYNDROME_MW_BIND_ERR = 0x06,
102 SYNDROME_BAD_RESP_ERR = 0x10,
103 SYNDROME_LOCAL_ACCESS_ERR = 0x11,
104 SYNDROME_REMOTE_INVAL_REQ_ERR = 0x12,
105 SYNDROME_REMOTE_ACCESS_ERR = 0x13,
106 SYNDROME_REMOTE_OP_ERR = 0x14,
107 SYNDROME_RETRY_EXC_ERR = 0x15,
108 SYNDROME_RNR_RETRY_EXC_ERR = 0x16,
109 SYNDROME_LOCAL_RDD_VIOL_ERR = 0x20,
110 SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
111 SYNDROME_REMOTE_ABORTED_ERR = 0x22,
112 SYNDROME_INVAL_EECN_ERR = 0x23,
113 SYNDROME_INVAL_EEC_STATE_ERR = 0x24
123 __be32 imm_etype_pkey_eec;
132 struct mthca_err_cqe {
145 #define MTHCA_CQ_ENTRY_OWNER_SW (0 << 7)
146 #define MTHCA_CQ_ENTRY_OWNER_HW (1 << 7)
148 #define MTHCA_TAVOR_CQ_DB_INC_CI (1 << 24)
149 #define MTHCA_TAVOR_CQ_DB_REQ_NOT (2 << 24)
150 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL (3 << 24)
151 #define MTHCA_TAVOR_CQ_DB_SET_CI (4 << 24)
152 #define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
154 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL (1 << 24)
155 #define MTHCA_ARBEL_CQ_DB_REQ_NOT (2 << 24)
156 #define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
158 static inline struct mthca_cqe *get_cqe_from_buf(struct mthca_cq_buf *buf,
162 return buf->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
164 return buf->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
165 + (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
168 static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
170 return get_cqe_from_buf(&cq->buf, entry);
173 static inline struct mthca_cqe *cqe_sw(struct mthca_cqe *cqe)
175 return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
178 static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
180 return cqe_sw(get_cqe(cq, cq->cons_index & cq->ibcq.cqe));
183 static inline void set_cqe_hw(struct mthca_cqe *cqe)
185 cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
188 static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
190 __be32 *cqe = cqe_ptr;
192 (void) cqe; /* avoid warning if mthca_dbg compiled away... */
193 mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
194 be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
195 be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
196 be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
200 * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
201 * should be correct before calling update_cons_index().
203 static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
206 if (mthca_is_memfree(dev)) {
207 *cq->set_ci_db = cpu_to_be32(cq->cons_index);
210 mthca_write64(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn, incr - 1,
211 dev->kar + MTHCA_CQ_DOORBELL,
212 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
214 * Make sure doorbells don't leak out of CQ spinlock
215 * and reach the HCA out of order:
221 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn)
225 cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
228 mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
234 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
237 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
238 enum ib_event_type event_type)
241 struct ib_event event;
243 spin_lock(&dev->cq_table.lock);
245 cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
249 spin_unlock(&dev->cq_table.lock);
252 mthca_warn(dev, "Async event for bogus CQ %08x\n", cqn);
256 event.device = &dev->ib_dev;
257 event.event = event_type;
258 event.element.cq = &cq->ibcq;
259 if (cq->ibcq.event_handler)
260 cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
262 spin_lock(&dev->cq_table.lock);
265 spin_unlock(&dev->cq_table.lock);
268 static inline int is_recv_cqe(struct mthca_cqe *cqe)
270 if ((cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
271 MTHCA_ERROR_CQE_OPCODE_MASK)
272 return !(cqe->opcode & 0x01);
274 return !(cqe->is_send & 0x80);
277 void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
278 struct mthca_srq *srq)
280 struct mthca_cqe *cqe;
284 spin_lock_irq(&cq->lock);
287 * First we need to find the current producer index, so we
288 * know where to start cleaning from. It doesn't matter if HW
289 * adds new entries after this loop -- the QP we're worried
290 * about is already in RESET, so the new entries won't come
291 * from our QP and therefore don't need to be checked.
293 for (prod_index = cq->cons_index;
294 cqe_sw(get_cqe(cq, prod_index & cq->ibcq.cqe));
296 if (prod_index == cq->cons_index + cq->ibcq.cqe)
300 mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
301 qpn, cq->cqn, cq->cons_index, prod_index);
304 * Now sweep backwards through the CQ, removing CQ entries
305 * that match our QP by copying older entries on top of them.
307 while ((int) --prod_index - (int) cq->cons_index >= 0) {
308 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
309 if (cqe->my_qpn == cpu_to_be32(qpn)) {
310 if (srq && is_recv_cqe(cqe))
311 mthca_free_srq_wqe(srq, be32_to_cpu(cqe->wqe));
314 memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe),
315 cqe, MTHCA_CQ_ENTRY_SIZE);
319 for (i = 0; i < nfreed; ++i)
320 set_cqe_hw(get_cqe(cq, (cq->cons_index + i) & cq->ibcq.cqe));
322 cq->cons_index += nfreed;
323 update_cons_index(dev, cq, nfreed);
326 spin_unlock_irq(&cq->lock);
329 void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
334 * In Tavor mode, the hardware keeps the consumer and producer
335 * indices mod the CQ size. Since we might be making the CQ
336 * bigger, we need to deal with the case where the producer
337 * index wrapped around before the CQ was resized.
339 if (!mthca_is_memfree(to_mdev(cq->ibcq.device)) &&
340 cq->ibcq.cqe < cq->resize_buf->cqe) {
341 cq->cons_index &= cq->ibcq.cqe;
342 if (cqe_sw(get_cqe(cq, cq->ibcq.cqe)))
343 cq->cons_index -= cq->ibcq.cqe + 1;
346 for (i = cq->cons_index; cqe_sw(get_cqe(cq, i & cq->ibcq.cqe)); ++i)
347 memcpy(get_cqe_from_buf(&cq->resize_buf->buf,
348 i & cq->resize_buf->cqe),
349 get_cqe(cq, i & cq->ibcq.cqe), MTHCA_CQ_ENTRY_SIZE);
352 int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent)
357 ret = mthca_buf_alloc(dev, nent * MTHCA_CQ_ENTRY_SIZE,
358 MTHCA_MAX_DIRECT_CQ_SIZE,
359 &buf->queue, &buf->is_direct,
360 &dev->driver_pd, 1, &buf->mr);
364 for (i = 0; i < nent; ++i)
365 set_cqe_hw(get_cqe_from_buf(buf, i));
370 void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int cqe)
372 mthca_buf_free(dev, (cqe + 1) * MTHCA_CQ_ENTRY_SIZE, &buf->queue,
373 buf->is_direct, &buf->mr);
376 static void handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
377 struct mthca_qp *qp, int wqe_index, int is_send,
378 struct mthca_err_cqe *cqe,
379 struct ib_wc *entry, int *free_cqe)
384 if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
385 mthca_dbg(dev, "local QP operation err "
386 "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
387 be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
388 cq->cqn, cq->cons_index);
393 * For completions in error, only work request ID, status, vendor error
394 * (and freed resource count for RD) have to be set.
396 switch (cqe->syndrome) {
397 case SYNDROME_LOCAL_LENGTH_ERR:
398 entry->status = IB_WC_LOC_LEN_ERR;
400 case SYNDROME_LOCAL_QP_OP_ERR:
401 entry->status = IB_WC_LOC_QP_OP_ERR;
403 case SYNDROME_LOCAL_EEC_OP_ERR:
404 entry->status = IB_WC_LOC_EEC_OP_ERR;
406 case SYNDROME_LOCAL_PROT_ERR:
407 entry->status = IB_WC_LOC_PROT_ERR;
409 case SYNDROME_WR_FLUSH_ERR:
410 entry->status = IB_WC_WR_FLUSH_ERR;
412 case SYNDROME_MW_BIND_ERR:
413 entry->status = IB_WC_MW_BIND_ERR;
415 case SYNDROME_BAD_RESP_ERR:
416 entry->status = IB_WC_BAD_RESP_ERR;
418 case SYNDROME_LOCAL_ACCESS_ERR:
419 entry->status = IB_WC_LOC_ACCESS_ERR;
421 case SYNDROME_REMOTE_INVAL_REQ_ERR:
422 entry->status = IB_WC_REM_INV_REQ_ERR;
424 case SYNDROME_REMOTE_ACCESS_ERR:
425 entry->status = IB_WC_REM_ACCESS_ERR;
427 case SYNDROME_REMOTE_OP_ERR:
428 entry->status = IB_WC_REM_OP_ERR;
430 case SYNDROME_RETRY_EXC_ERR:
431 entry->status = IB_WC_RETRY_EXC_ERR;
433 case SYNDROME_RNR_RETRY_EXC_ERR:
434 entry->status = IB_WC_RNR_RETRY_EXC_ERR;
436 case SYNDROME_LOCAL_RDD_VIOL_ERR:
437 entry->status = IB_WC_LOC_RDD_VIOL_ERR;
439 case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
440 entry->status = IB_WC_REM_INV_RD_REQ_ERR;
442 case SYNDROME_REMOTE_ABORTED_ERR:
443 entry->status = IB_WC_REM_ABORT_ERR;
445 case SYNDROME_INVAL_EECN_ERR:
446 entry->status = IB_WC_INV_EECN_ERR;
448 case SYNDROME_INVAL_EEC_STATE_ERR:
449 entry->status = IB_WC_INV_EEC_STATE_ERR;
452 entry->status = IB_WC_GENERAL_ERR;
456 entry->vendor_err = cqe->vendor_err;
459 * Mem-free HCAs always generate one CQE per WQE, even in the
460 * error case, so we don't have to check the doorbell count, etc.
462 if (mthca_is_memfree(dev))
465 mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
468 * If we're at the end of the WQE chain, or we've used up our
469 * doorbell count, free the CQE. Otherwise just update it for
470 * the next poll operation.
472 if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
475 be16_add_cpu(&cqe->db_cnt, -dbd);
477 cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
482 static inline int mthca_poll_one(struct mthca_dev *dev,
484 struct mthca_qp **cur_qp,
489 struct mthca_cqe *cqe;
497 cqe = next_cqe_sw(cq);
502 * Make sure we read CQ entry contents after we've checked the
508 mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
509 cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
510 be32_to_cpu(cqe->wqe));
514 is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
515 MTHCA_ERROR_CQE_OPCODE_MASK;
516 is_send = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
518 if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
520 * We do not have to take the QP table lock here,
521 * because CQs will be locked while QPs are removed
524 *cur_qp = mthca_array_get(&dev->qp_table.qp,
525 be32_to_cpu(cqe->my_qpn) &
526 (dev->limits.num_qps - 1));
528 mthca_warn(dev, "CQ entry for unknown QP %06x\n",
529 be32_to_cpu(cqe->my_qpn) & 0xffffff);
535 entry->qp = &(*cur_qp)->ibqp;
539 wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
541 entry->wr_id = (*cur_qp)->wrid[wqe_index +
543 } else if ((*cur_qp)->ibqp.srq) {
544 struct mthca_srq *srq = to_msrq((*cur_qp)->ibqp.srq);
545 u32 wqe = be32_to_cpu(cqe->wqe);
547 wqe_index = wqe >> srq->wqe_shift;
548 entry->wr_id = srq->wrid[wqe_index];
549 mthca_free_srq_wqe(srq, wqe);
553 wqe = be32_to_cpu(cqe->wqe);
554 wqe_index = wqe >> wq->wqe_shift;
556 * WQE addr == base - 1 might be reported in receive completion
557 * with error instead of (rq size - 1) by Sinai FW 1.0.800 and
558 * Arbel FW 5.1.400. This bug should be fixed in later FW revs.
560 if (unlikely(wqe_index < 0))
561 wqe_index = wq->max - 1;
562 entry->wr_id = (*cur_qp)->wrid[wqe_index];
566 if (wq->last_comp < wqe_index)
567 wq->tail += wqe_index - wq->last_comp;
569 wq->tail += wqe_index + wq->max - wq->last_comp;
571 wq->last_comp = wqe_index;
575 handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
576 (struct mthca_err_cqe *) cqe,
583 switch (cqe->opcode) {
584 case MTHCA_OPCODE_RDMA_WRITE:
585 entry->opcode = IB_WC_RDMA_WRITE;
587 case MTHCA_OPCODE_RDMA_WRITE_IMM:
588 entry->opcode = IB_WC_RDMA_WRITE;
589 entry->wc_flags |= IB_WC_WITH_IMM;
591 case MTHCA_OPCODE_SEND:
592 entry->opcode = IB_WC_SEND;
594 case MTHCA_OPCODE_SEND_IMM:
595 entry->opcode = IB_WC_SEND;
596 entry->wc_flags |= IB_WC_WITH_IMM;
598 case MTHCA_OPCODE_RDMA_READ:
599 entry->opcode = IB_WC_RDMA_READ;
600 entry->byte_len = be32_to_cpu(cqe->byte_cnt);
602 case MTHCA_OPCODE_ATOMIC_CS:
603 entry->opcode = IB_WC_COMP_SWAP;
604 entry->byte_len = MTHCA_ATOMIC_BYTE_LEN;
606 case MTHCA_OPCODE_ATOMIC_FA:
607 entry->opcode = IB_WC_FETCH_ADD;
608 entry->byte_len = MTHCA_ATOMIC_BYTE_LEN;
610 case MTHCA_OPCODE_BIND_MW:
611 entry->opcode = IB_WC_BIND_MW;
614 entry->opcode = MTHCA_OPCODE_INVALID;
618 entry->byte_len = be32_to_cpu(cqe->byte_cnt);
619 switch (cqe->opcode & 0x1f) {
620 case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
621 case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
622 entry->wc_flags = IB_WC_WITH_IMM;
623 entry->ex.imm_data = cqe->imm_etype_pkey_eec;
624 entry->opcode = IB_WC_RECV;
626 case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
627 case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
628 entry->wc_flags = IB_WC_WITH_IMM;
629 entry->ex.imm_data = cqe->imm_etype_pkey_eec;
630 entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
634 entry->opcode = IB_WC_RECV;
637 entry->slid = be16_to_cpu(cqe->rlid);
638 entry->sl = cqe->sl_ipok >> 4;
639 entry->src_qp = be32_to_cpu(cqe->rqpn) & 0xffffff;
640 entry->dlid_path_bits = cqe->g_mlpath & 0x7f;
641 entry->pkey_index = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
642 entry->wc_flags |= cqe->g_mlpath & 0x80 ? IB_WC_GRH : 0;
643 checksum = (be32_to_cpu(cqe->rqpn) >> 24) |
644 ((be32_to_cpu(cqe->my_ee) >> 16) & 0xff00);
645 entry->csum_ok = (cqe->sl_ipok & 1 && checksum == 0xffff);
648 entry->status = IB_WC_SUCCESS;
651 if (likely(free_cqe)) {
660 int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
663 struct mthca_dev *dev = to_mdev(ibcq->device);
664 struct mthca_cq *cq = to_mcq(ibcq);
665 struct mthca_qp *qp = NULL;
671 spin_lock_irqsave(&cq->lock, flags);
675 while (npolled < num_entries) {
676 err = mthca_poll_one(dev, cq, &qp,
677 &freed, entry + npolled);
685 update_cons_index(dev, cq, freed);
689 * If a CQ resize is in progress and we discovered that the
690 * old buffer is empty, then peek in the new buffer, and if
691 * it's not empty, switch to the new buffer and continue
694 if (unlikely(err == -EAGAIN && cq->resize_buf &&
695 cq->resize_buf->state == CQ_RESIZE_READY)) {
697 * In Tavor mode, the hardware keeps the producer
698 * index modulo the CQ size. Since we might be making
699 * the CQ bigger, we need to mask our consumer index
700 * using the size of the old CQ buffer before looking
701 * in the new CQ buffer.
703 if (!mthca_is_memfree(dev))
704 cq->cons_index &= cq->ibcq.cqe;
706 if (cqe_sw(get_cqe_from_buf(&cq->resize_buf->buf,
707 cq->cons_index & cq->resize_buf->cqe))) {
708 struct mthca_cq_buf tbuf;
713 cq->buf = cq->resize_buf->buf;
714 cq->ibcq.cqe = cq->resize_buf->cqe;
716 cq->resize_buf->buf = tbuf;
717 cq->resize_buf->cqe = tcqe;
718 cq->resize_buf->state = CQ_RESIZE_SWAPPED;
724 spin_unlock_irqrestore(&cq->lock, flags);
726 return err == 0 || err == -EAGAIN ? npolled : err;
729 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags)
731 u32 dbhi = ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
732 MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
733 MTHCA_TAVOR_CQ_DB_REQ_NOT) |
736 mthca_write64(dbhi, 0xffffffff, to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
737 MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
742 int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
744 struct mthca_cq *cq = to_mcq(ibcq);
747 u32 sn = cq->arm_sn & 3;
749 db_rec[0] = cpu_to_be32(cq->cons_index);
750 db_rec[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
751 ((flags & IB_CQ_SOLICITED_MASK) ==
752 IB_CQ_SOLICITED ? 1 : 2));
754 mthca_write_db_rec(db_rec, cq->arm_db);
757 * Make sure that the doorbell record in host memory is
758 * written before ringing the doorbell via PCI MMIO.
763 ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
764 MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
765 MTHCA_ARBEL_CQ_DB_REQ_NOT) | cq->cqn;
767 mthca_write64(dbhi, cq->cons_index,
768 to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
769 MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
774 int mthca_init_cq(struct mthca_dev *dev, int nent,
775 struct mthca_ucontext *ctx, u32 pdn,
778 struct mthca_mailbox *mailbox;
779 struct mthca_cq_context *cq_context;
783 cq->ibcq.cqe = nent - 1;
784 cq->is_kernel = !ctx;
786 cq->cqn = mthca_alloc(&dev->cq_table.alloc);
790 if (mthca_is_memfree(dev)) {
791 err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
800 cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
801 cq->cqn, &cq->set_ci_db);
802 if (cq->set_ci_db_index < 0)
805 cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
806 cq->cqn, &cq->arm_db);
807 if (cq->arm_db_index < 0)
812 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
816 cq_context = mailbox->buf;
819 err = mthca_alloc_cq_buf(dev, &cq->buf, nent);
821 goto err_out_mailbox;
824 spin_lock_init(&cq->lock);
826 init_waitqueue_head(&cq->wait);
827 mutex_init(&cq->mutex);
829 memset(cq_context, 0, sizeof *cq_context);
830 cq_context->flags = cpu_to_be32(MTHCA_CQ_STATUS_OK |
831 MTHCA_CQ_STATE_DISARMED |
833 cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
835 cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
837 cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
838 cq_context->error_eqn = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
839 cq_context->comp_eqn = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
840 cq_context->pd = cpu_to_be32(pdn);
841 cq_context->lkey = cpu_to_be32(cq->buf.mr.ibmr.lkey);
842 cq_context->cqn = cpu_to_be32(cq->cqn);
844 if (mthca_is_memfree(dev)) {
845 cq_context->ci_db = cpu_to_be32(cq->set_ci_db_index);
846 cq_context->state_db = cpu_to_be32(cq->arm_db_index);
849 err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
851 mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
852 goto err_out_free_mr;
856 mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
859 goto err_out_free_mr;
862 spin_lock_irq(&dev->cq_table.lock);
863 if (mthca_array_set(&dev->cq_table.cq,
864 cq->cqn & (dev->limits.num_cqs - 1),
866 spin_unlock_irq(&dev->cq_table.lock);
867 goto err_out_free_mr;
869 spin_unlock_irq(&dev->cq_table.lock);
873 mthca_free_mailbox(dev, mailbox);
879 mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
882 mthca_free_mailbox(dev, mailbox);
885 if (cq->is_kernel && mthca_is_memfree(dev))
886 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
889 if (cq->is_kernel && mthca_is_memfree(dev))
890 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
893 mthca_table_put(dev, dev->cq_table.table, cq->cqn);
896 mthca_free(&dev->cq_table.alloc, cq->cqn);
901 static inline int get_cq_refcount(struct mthca_dev *dev, struct mthca_cq *cq)
905 spin_lock_irq(&dev->cq_table.lock);
907 spin_unlock_irq(&dev->cq_table.lock);
912 void mthca_free_cq(struct mthca_dev *dev,
915 struct mthca_mailbox *mailbox;
919 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
920 if (IS_ERR(mailbox)) {
921 mthca_warn(dev, "No memory for mailbox to free CQ.\n");
925 err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
927 mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
929 mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
932 __be32 *ctx = mailbox->buf;
935 printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
936 cq->cqn, cq->cons_index,
937 cq->is_kernel ? !!next_cqe_sw(cq) : 0);
938 for (j = 0; j < 16; ++j)
939 printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
942 spin_lock_irq(&dev->cq_table.lock);
943 mthca_array_clear(&dev->cq_table.cq,
944 cq->cqn & (dev->limits.num_cqs - 1));
946 spin_unlock_irq(&dev->cq_table.lock);
948 if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
949 synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
951 synchronize_irq(dev->pdev->irq);
953 wait_event(cq->wait, !get_cq_refcount(dev, cq));
956 mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
957 if (mthca_is_memfree(dev)) {
958 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
959 mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
963 mthca_table_put(dev, dev->cq_table.table, cq->cqn);
964 mthca_free(&dev->cq_table.alloc, cq->cqn);
965 mthca_free_mailbox(dev, mailbox);
968 int mthca_init_cq_table(struct mthca_dev *dev)
972 spin_lock_init(&dev->cq_table.lock);
974 err = mthca_alloc_init(&dev->cq_table.alloc,
977 dev->limits.reserved_cqs);
981 err = mthca_array_init(&dev->cq_table.cq,
982 dev->limits.num_cqs);
984 mthca_alloc_cleanup(&dev->cq_table.alloc);
989 void mthca_cleanup_cq_table(struct mthca_dev *dev)
991 mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
992 mthca_alloc_cleanup(&dev->cq_table.alloc);