2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/mlx4/cq.h>
35 #include <linux/mlx4/qp.h>
36 #include <linux/mlx4/srq.h>
37 #include <linux/slab.h>
40 #include <rdma/mlx4-abi.h>
41 #include <rdma/uverbs_ioctl.h>
43 static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
45 struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
46 ibcq->comp_handler(ibcq, ibcq->cq_context);
49 static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
51 struct ib_event event;
54 if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
55 pr_warn("Unexpected event type %d "
56 "on CQ %06x\n", type, cq->cqn);
60 ibcq = &to_mibcq(cq)->ibcq;
61 if (ibcq->event_handler) {
62 event.device = ibcq->device;
63 event.event = IB_EVENT_CQ_ERR;
64 event.element.cq = ibcq;
65 ibcq->event_handler(&event, ibcq->cq_context);
69 static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
71 return mlx4_buf_offset(&buf->buf, n * buf->entry_size);
74 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
76 return get_cqe_from_buf(&cq->buf, n);
79 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
81 struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
82 struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
84 return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
85 !!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
88 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
90 return get_sw_cqe(cq, cq->mcq.cons_index);
93 int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
95 struct mlx4_ib_cq *mcq = to_mcq(cq);
96 struct mlx4_ib_dev *dev = to_mdev(cq->device);
98 return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
101 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
105 err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
106 PAGE_SIZE * 2, &buf->buf);
111 buf->entry_size = dev->dev->caps.cqe_size;
112 err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
117 err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf);
124 mlx4_mtt_cleanup(dev->dev, &buf->mtt);
127 mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
133 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
135 mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
138 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_udata *udata,
139 struct mlx4_ib_cq_buf *buf,
140 struct ib_umem **umem, u64 buf_addr, int cqe)
143 int cqe_size = dev->dev->caps.cqe_size;
147 *umem = ib_umem_get(udata, buf_addr, cqe * cqe_size,
148 IB_ACCESS_LOCAL_WRITE, 1);
150 return PTR_ERR(*umem);
152 n = ib_umem_page_count(*umem);
153 shift = mlx4_ib_umem_calc_optimal_mtt_size(*umem, 0, &n);
154 err = mlx4_mtt_init(dev->dev, n, shift, &buf->mtt);
159 err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
166 mlx4_mtt_cleanup(dev->dev, &buf->mtt);
169 ib_umem_release(*umem);
174 #define CQ_CREATE_FLAGS_SUPPORTED IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION
175 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev,
176 const struct ib_cq_init_attr *attr,
177 struct ib_udata *udata)
179 int entries = attr->cqe;
180 int vector = attr->comp_vector;
181 struct mlx4_ib_dev *dev = to_mdev(ibdev);
182 struct mlx4_ib_cq *cq;
183 struct mlx4_uar *uar;
186 struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
187 udata, struct mlx4_ib_ucontext, ibucontext);
189 if (entries < 1 || entries > dev->dev->caps.max_cqes)
190 return ERR_PTR(-EINVAL);
192 if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
193 return ERR_PTR(-EINVAL);
195 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
197 return ERR_PTR(-ENOMEM);
199 entries = roundup_pow_of_two(entries + 1);
200 cq->ibcq.cqe = entries - 1;
201 mutex_init(&cq->resize_mutex);
202 spin_lock_init(&cq->lock);
203 cq->resize_buf = NULL;
204 cq->resize_umem = NULL;
205 cq->create_flags = attr->flags;
206 INIT_LIST_HEAD(&cq->send_qp_list);
207 INIT_LIST_HEAD(&cq->recv_qp_list);
210 struct mlx4_ib_create_cq ucmd;
212 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
217 buf_addr = (void *)(unsigned long)ucmd.buf_addr;
218 err = mlx4_ib_get_cq_umem(dev, udata, &cq->buf, &cq->umem,
219 ucmd.buf_addr, entries);
223 err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &cq->db);
228 cq->mcq.usage = MLX4_RES_USAGE_USER_VERBS;
230 err = mlx4_db_alloc(dev->dev, &cq->db, 1);
234 cq->mcq.set_ci_db = cq->db.db;
235 cq->mcq.arm_db = cq->db.db + 1;
236 *cq->mcq.set_ci_db = 0;
239 err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
243 buf_addr = &cq->buf.buf;
245 uar = &dev->priv_uar;
246 cq->mcq.usage = MLX4_RES_USAGE_DRIVER;
250 vector = dev->eq_table[vector % ibdev->num_comp_vectors];
252 err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar, cq->db.dma,
254 !!(cq->create_flags &
255 IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION),
261 cq->mcq.tasklet_ctx.comp = mlx4_ib_cq_comp;
263 cq->mcq.comp = mlx4_ib_cq_comp;
264 cq->mcq.event = mlx4_ib_cq_event;
267 if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
275 mlx4_cq_free(dev->dev, &cq->mcq);
279 mlx4_ib_db_unmap_user(context, &cq->db);
282 mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
285 ib_umem_release(cq->umem);
287 mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
291 mlx4_db_free(dev->dev, &cq->db);
299 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
307 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
311 err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
313 kfree(cq->resize_buf);
314 cq->resize_buf = NULL;
318 cq->resize_buf->cqe = entries - 1;
323 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
324 int entries, struct ib_udata *udata)
326 struct mlx4_ib_resize_cq ucmd;
332 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
335 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
339 err = mlx4_ib_get_cq_umem(dev, udata, &cq->resize_buf->buf,
340 &cq->resize_umem, ucmd.buf_addr, entries);
342 kfree(cq->resize_buf);
343 cq->resize_buf = NULL;
347 cq->resize_buf->cqe = entries - 1;
352 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
356 i = cq->mcq.cons_index;
357 while (get_sw_cqe(cq, i))
360 return i - cq->mcq.cons_index;
363 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
365 struct mlx4_cqe *cqe, *new_cqe;
367 int cqe_size = cq->buf.entry_size;
368 int cqe_inc = cqe_size == 64 ? 1 : 0;
370 i = cq->mcq.cons_index;
371 cqe = get_cqe(cq, i & cq->ibcq.cqe);
374 while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
375 new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
376 (i + 1) & cq->resize_buf->cqe);
377 memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
380 new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
381 (((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
382 cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
385 ++cq->mcq.cons_index;
388 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
390 struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
391 struct mlx4_ib_cq *cq = to_mcq(ibcq);
396 mutex_lock(&cq->resize_mutex);
397 if (entries < 1 || entries > dev->dev->caps.max_cqes) {
402 entries = roundup_pow_of_two(entries + 1);
403 if (entries == ibcq->cqe + 1) {
408 if (entries > dev->dev->caps.max_cqes + 1) {
414 err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
418 /* Can't be smaller than the number of outstanding CQEs */
419 outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
420 if (entries < outst_cqe + 1) {
425 err = mlx4_alloc_resize_buf(dev, cq, entries);
432 err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
436 mlx4_mtt_cleanup(dev->dev, &mtt);
438 cq->buf = cq->resize_buf->buf;
439 cq->ibcq.cqe = cq->resize_buf->cqe;
440 ib_umem_release(cq->umem);
441 cq->umem = cq->resize_umem;
443 kfree(cq->resize_buf);
444 cq->resize_buf = NULL;
445 cq->resize_umem = NULL;
447 struct mlx4_ib_cq_buf tmp_buf;
450 spin_lock_irq(&cq->lock);
451 if (cq->resize_buf) {
452 mlx4_ib_cq_resize_copy_cqes(cq);
454 tmp_cqe = cq->ibcq.cqe;
455 cq->buf = cq->resize_buf->buf;
456 cq->ibcq.cqe = cq->resize_buf->cqe;
458 kfree(cq->resize_buf);
459 cq->resize_buf = NULL;
461 spin_unlock_irq(&cq->lock);
464 mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
470 mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
472 mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
473 cq->resize_buf->cqe);
475 kfree(cq->resize_buf);
476 cq->resize_buf = NULL;
478 if (cq->resize_umem) {
479 ib_umem_release(cq->resize_umem);
480 cq->resize_umem = NULL;
484 mutex_unlock(&cq->resize_mutex);
489 int mlx4_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
491 struct mlx4_ib_dev *dev = to_mdev(cq->device);
492 struct mlx4_ib_cq *mcq = to_mcq(cq);
494 mlx4_cq_free(dev->dev, &mcq->mcq);
495 mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
498 mlx4_ib_db_unmap_user(
499 rdma_udata_to_drv_context(
501 struct mlx4_ib_ucontext,
504 ib_umem_release(mcq->umem);
506 mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
507 mlx4_db_free(dev->dev, &mcq->db);
515 static void dump_cqe(void *cqe)
519 pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
520 be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
521 be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
522 be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
525 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
528 if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
529 pr_debug("local QP operation err "
530 "(QPN %06x, WQE index %x, vendor syndrome %02x, "
532 be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
533 cqe->vendor_err_syndrome,
534 cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
538 switch (cqe->syndrome) {
539 case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
540 wc->status = IB_WC_LOC_LEN_ERR;
542 case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
543 wc->status = IB_WC_LOC_QP_OP_ERR;
545 case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
546 wc->status = IB_WC_LOC_PROT_ERR;
548 case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
549 wc->status = IB_WC_WR_FLUSH_ERR;
551 case MLX4_CQE_SYNDROME_MW_BIND_ERR:
552 wc->status = IB_WC_MW_BIND_ERR;
554 case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
555 wc->status = IB_WC_BAD_RESP_ERR;
557 case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
558 wc->status = IB_WC_LOC_ACCESS_ERR;
560 case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
561 wc->status = IB_WC_REM_INV_REQ_ERR;
563 case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
564 wc->status = IB_WC_REM_ACCESS_ERR;
566 case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
567 wc->status = IB_WC_REM_OP_ERR;
569 case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
570 wc->status = IB_WC_RETRY_EXC_ERR;
572 case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
573 wc->status = IB_WC_RNR_RETRY_EXC_ERR;
575 case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
576 wc->status = IB_WC_REM_ABORT_ERR;
579 wc->status = IB_WC_GENERAL_ERR;
583 wc->vendor_err = cqe->vendor_err_syndrome;
586 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
588 return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
589 MLX4_CQE_STATUS_IPV4F |
590 MLX4_CQE_STATUS_IPV4OPT |
591 MLX4_CQE_STATUS_IPV6 |
592 MLX4_CQE_STATUS_IPOK)) ==
593 cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
594 MLX4_CQE_STATUS_IPOK)) &&
595 (status & cpu_to_be16(MLX4_CQE_STATUS_UDP |
596 MLX4_CQE_STATUS_TCP)) &&
597 checksum == cpu_to_be16(0xffff);
600 static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
601 unsigned tail, struct mlx4_cqe *cqe, int is_eth)
603 struct mlx4_ib_proxy_sqp_hdr *hdr;
605 ib_dma_sync_single_for_cpu(qp->ibqp.device,
606 qp->sqp_proxy_rcv[tail].map,
607 sizeof (struct mlx4_ib_proxy_sqp_hdr),
609 hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
610 wc->pkey_index = be16_to_cpu(hdr->tun.pkey_index);
611 wc->src_qp = be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
612 wc->wc_flags |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
613 wc->dlid_path_bits = 0;
617 wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
618 memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
619 memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
620 wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
622 wc->slid = be16_to_cpu(hdr->tun.slid_mac_47_32);
623 wc->sl = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
627 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
628 struct ib_wc *wc, int *npolled, int is_send)
630 struct mlx4_ib_wq *wq;
634 wq = is_send ? &qp->sq : &qp->rq;
635 cur = wq->head - wq->tail;
640 for (i = 0; i < cur && *npolled < num_entries; i++) {
641 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
642 wc->status = IB_WC_WR_FLUSH_ERR;
643 wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
651 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
652 struct ib_wc *wc, int *npolled)
654 struct mlx4_ib_qp *qp;
657 /* Find uncompleted WQEs belonging to that cq and return
658 * simulated FLUSH_ERR completions
660 list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
661 mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
662 if (*npolled >= num_entries)
666 list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
667 mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
668 if (*npolled >= num_entries)
676 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
677 struct mlx4_ib_qp **cur_qp,
680 struct mlx4_cqe *cqe;
682 struct mlx4_ib_wq *wq;
683 struct mlx4_ib_srq *srq;
684 struct mlx4_srq *msrq = NULL;
693 cqe = next_cqe_sw(cq);
697 if (cq->buf.entry_size == 64)
700 ++cq->mcq.cons_index;
703 * Make sure we read CQ entry contents after we've checked the
708 is_send = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
709 is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
710 MLX4_CQE_OPCODE_ERROR;
712 /* Resize CQ in progress */
713 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
714 if (cq->resize_buf) {
715 struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
717 mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
718 cq->buf = cq->resize_buf->buf;
719 cq->ibcq.cqe = cq->resize_buf->cqe;
721 kfree(cq->resize_buf);
722 cq->resize_buf = NULL;
729 (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
731 * We do not have to take the QP table lock here,
732 * because CQs will be locked while QPs are removed
735 mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
736 be32_to_cpu(cqe->vlan_my_qpn));
737 *cur_qp = to_mibqp(mqp);
740 wc->qp = &(*cur_qp)->ibqp;
742 if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
744 g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
745 srq_num = g_mlpath_rqpn & 0xffffff;
746 /* SRQ is also in the radix tree */
747 msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
753 if (!(*cur_qp)->sq_signal_bits) {
754 wqe_ctr = be16_to_cpu(cqe->wqe_index);
755 wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
757 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
759 } else if ((*cur_qp)->ibqp.srq) {
760 srq = to_msrq((*cur_qp)->ibqp.srq);
761 wqe_ctr = be16_to_cpu(cqe->wqe_index);
762 wc->wr_id = srq->wrid[wqe_ctr];
763 mlx4_ib_free_srq_wqe(srq, wqe_ctr);
765 srq = to_mibsrq(msrq);
766 wqe_ctr = be16_to_cpu(cqe->wqe_index);
767 wc->wr_id = srq->wrid[wqe_ctr];
768 mlx4_ib_free_srq_wqe(srq, wqe_ctr);
771 tail = wq->tail & (wq->wqe_cnt - 1);
772 wc->wr_id = wq->wrid[tail];
776 if (unlikely(is_error)) {
777 mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
781 wc->status = IB_WC_SUCCESS;
785 switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
786 case MLX4_OPCODE_RDMA_WRITE_IMM:
787 wc->wc_flags |= IB_WC_WITH_IMM;
789 case MLX4_OPCODE_RDMA_WRITE:
790 wc->opcode = IB_WC_RDMA_WRITE;
792 case MLX4_OPCODE_SEND_IMM:
793 wc->wc_flags |= IB_WC_WITH_IMM;
795 case MLX4_OPCODE_SEND:
796 case MLX4_OPCODE_SEND_INVAL:
797 wc->opcode = IB_WC_SEND;
799 case MLX4_OPCODE_RDMA_READ:
800 wc->opcode = IB_WC_RDMA_READ;
801 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
803 case MLX4_OPCODE_ATOMIC_CS:
804 wc->opcode = IB_WC_COMP_SWAP;
807 case MLX4_OPCODE_ATOMIC_FA:
808 wc->opcode = IB_WC_FETCH_ADD;
811 case MLX4_OPCODE_MASKED_ATOMIC_CS:
812 wc->opcode = IB_WC_MASKED_COMP_SWAP;
815 case MLX4_OPCODE_MASKED_ATOMIC_FA:
816 wc->opcode = IB_WC_MASKED_FETCH_ADD;
819 case MLX4_OPCODE_LSO:
820 wc->opcode = IB_WC_LSO;
822 case MLX4_OPCODE_FMR:
823 wc->opcode = IB_WC_REG_MR;
825 case MLX4_OPCODE_LOCAL_INVAL:
826 wc->opcode = IB_WC_LOCAL_INV;
830 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
832 switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
833 case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
834 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
835 wc->wc_flags = IB_WC_WITH_IMM;
836 wc->ex.imm_data = cqe->immed_rss_invalid;
838 case MLX4_RECV_OPCODE_SEND_INVAL:
839 wc->opcode = IB_WC_RECV;
840 wc->wc_flags = IB_WC_WITH_INVALIDATE;
841 wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
843 case MLX4_RECV_OPCODE_SEND:
844 wc->opcode = IB_WC_RECV;
847 case MLX4_RECV_OPCODE_SEND_IMM:
848 wc->opcode = IB_WC_RECV;
849 wc->wc_flags = IB_WC_WITH_IMM;
850 wc->ex.imm_data = cqe->immed_rss_invalid;
854 is_eth = (rdma_port_get_link_layer(wc->qp->device,
856 IB_LINK_LAYER_ETHERNET);
857 if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
858 if ((*cur_qp)->mlx4_ib_qp_type &
859 (MLX4_IB_QPT_PROXY_SMI_OWNER |
860 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
861 use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
867 g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
868 wc->src_qp = g_mlpath_rqpn & 0xffffff;
869 wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
870 wc->wc_flags |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
871 wc->pkey_index = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
872 wc->wc_flags |= mlx4_ib_ipoib_csum_ok(cqe->status,
873 cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
876 wc->sl = be16_to_cpu(cqe->sl_vid) >> 13;
877 if (be32_to_cpu(cqe->vlan_my_qpn) &
878 MLX4_CQE_CVLAN_PRESENT_MASK) {
879 wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
882 wc->vlan_id = 0xffff;
884 memcpy(wc->smac, cqe->smac, ETH_ALEN);
885 wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
887 wc->slid = be16_to_cpu(cqe->rlid);
888 wc->sl = be16_to_cpu(cqe->sl_vid) >> 12;
889 wc->vlan_id = 0xffff;
896 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
898 struct mlx4_ib_cq *cq = to_mcq(ibcq);
899 struct mlx4_ib_qp *cur_qp = NULL;
902 struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
904 spin_lock_irqsave(&cq->lock, flags);
905 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
906 mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
910 for (npolled = 0; npolled < num_entries; ++npolled) {
911 if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
915 mlx4_cq_set_ci(&cq->mcq);
918 spin_unlock_irqrestore(&cq->lock, flags);
923 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
925 mlx4_cq_arm(&to_mcq(ibcq)->mcq,
926 (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
927 MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
928 to_mdev(ibcq->device)->uar_map,
929 MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
934 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
938 struct mlx4_cqe *cqe, *dest;
940 int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
943 * First we need to find the current producer index, so we
944 * know where to start cleaning from. It doesn't matter if HW
945 * adds new entries after this loop -- the QP we're worried
946 * about is already in RESET, so the new entries won't come
947 * from our QP and therefore don't need to be checked.
949 for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
950 if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
954 * Now sweep backwards through the CQ, removing CQ entries
955 * that match our QP by copying older entries on top of them.
957 while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
958 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
961 if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
962 if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
963 mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
966 dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
969 owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
970 memcpy(dest, cqe, sizeof *cqe);
971 dest->owner_sr_opcode = owner_bit |
972 (dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
977 cq->mcq.cons_index += nfreed;
979 * Make sure update of buffer contents is done before
980 * updating consumer index.
983 mlx4_cq_set_ci(&cq->mcq);
987 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
989 spin_lock_irq(&cq->lock);
990 __mlx4_ib_cq_clean(cq, qpn, srq);
991 spin_unlock_irq(&cq->lock);