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>
42 static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
44 struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
45 ibcq->comp_handler(ibcq, ibcq->cq_context);
48 static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
50 struct ib_event event;
53 if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
54 pr_warn("Unexpected event type %d "
55 "on CQ %06x\n", type, cq->cqn);
59 ibcq = &to_mibcq(cq)->ibcq;
60 if (ibcq->event_handler) {
61 event.device = ibcq->device;
62 event.event = IB_EVENT_CQ_ERR;
63 event.element.cq = ibcq;
64 ibcq->event_handler(&event, ibcq->cq_context);
68 static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
70 return mlx4_buf_offset(&buf->buf, n * buf->entry_size);
73 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
75 return get_cqe_from_buf(&cq->buf, n);
78 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
80 struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
81 struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
83 return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
84 !!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
87 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
89 return get_sw_cqe(cq, cq->mcq.cons_index);
92 int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
94 struct mlx4_ib_cq *mcq = to_mcq(cq);
95 struct mlx4_ib_dev *dev = to_mdev(cq->device);
97 return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
100 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
104 err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
105 PAGE_SIZE * 2, &buf->buf, GFP_KERNEL);
110 buf->entry_size = dev->dev->caps.cqe_size;
111 err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
116 err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf, GFP_KERNEL);
123 mlx4_mtt_cleanup(dev->dev, &buf->mtt);
126 mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
132 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
134 mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
137 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *context,
138 struct mlx4_ib_cq_buf *buf, struct ib_umem **umem,
139 u64 buf_addr, int cqe)
142 int cqe_size = dev->dev->caps.cqe_size;
144 *umem = ib_umem_get(context, buf_addr, cqe * cqe_size,
145 IB_ACCESS_LOCAL_WRITE, 1);
147 return PTR_ERR(*umem);
149 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(*umem),
150 ilog2((*umem)->page_size), &buf->mtt);
154 err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
161 mlx4_mtt_cleanup(dev->dev, &buf->mtt);
164 ib_umem_release(*umem);
169 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev,
170 const struct ib_cq_init_attr *attr,
171 struct ib_ucontext *context,
172 struct ib_udata *udata)
174 int entries = attr->cqe;
175 int vector = attr->comp_vector;
176 struct mlx4_ib_dev *dev = to_mdev(ibdev);
177 struct mlx4_ib_cq *cq;
178 struct mlx4_uar *uar;
182 return ERR_PTR(-EINVAL);
184 if (entries < 1 || entries > dev->dev->caps.max_cqes)
185 return ERR_PTR(-EINVAL);
187 cq = kmalloc(sizeof *cq, GFP_KERNEL);
189 return ERR_PTR(-ENOMEM);
191 entries = roundup_pow_of_two(entries + 1);
192 cq->ibcq.cqe = entries - 1;
193 mutex_init(&cq->resize_mutex);
194 spin_lock_init(&cq->lock);
195 cq->resize_buf = NULL;
196 cq->resize_umem = NULL;
197 INIT_LIST_HEAD(&cq->send_qp_list);
198 INIT_LIST_HEAD(&cq->recv_qp_list);
201 struct mlx4_ib_create_cq ucmd;
203 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
208 err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
209 ucmd.buf_addr, entries);
213 err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
218 uar = &to_mucontext(context)->uar;
220 err = mlx4_db_alloc(dev->dev, &cq->db, 1, GFP_KERNEL);
224 cq->mcq.set_ci_db = cq->db.db;
225 cq->mcq.arm_db = cq->db.db + 1;
226 *cq->mcq.set_ci_db = 0;
229 err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
233 uar = &dev->priv_uar;
237 vector = dev->eq_table[vector % ibdev->num_comp_vectors];
239 err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
240 cq->db.dma, &cq->mcq, vector, 0, 0);
245 cq->mcq.tasklet_ctx.comp = mlx4_ib_cq_comp;
247 cq->mcq.comp = mlx4_ib_cq_comp;
248 cq->mcq.event = mlx4_ib_cq_event;
251 if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
260 mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
263 mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
266 ib_umem_release(cq->umem);
268 mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
272 mlx4_db_free(dev->dev, &cq->db);
280 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
288 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
292 err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
294 kfree(cq->resize_buf);
295 cq->resize_buf = NULL;
299 cq->resize_buf->cqe = entries - 1;
304 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
305 int entries, struct ib_udata *udata)
307 struct mlx4_ib_resize_cq ucmd;
313 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
316 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
320 err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
321 &cq->resize_umem, ucmd.buf_addr, entries);
323 kfree(cq->resize_buf);
324 cq->resize_buf = NULL;
328 cq->resize_buf->cqe = entries - 1;
333 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
337 i = cq->mcq.cons_index;
338 while (get_sw_cqe(cq, i))
341 return i - cq->mcq.cons_index;
344 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
346 struct mlx4_cqe *cqe, *new_cqe;
348 int cqe_size = cq->buf.entry_size;
349 int cqe_inc = cqe_size == 64 ? 1 : 0;
351 i = cq->mcq.cons_index;
352 cqe = get_cqe(cq, i & cq->ibcq.cqe);
355 while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
356 new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
357 (i + 1) & cq->resize_buf->cqe);
358 memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
361 new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
362 (((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
363 cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
366 ++cq->mcq.cons_index;
369 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
371 struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
372 struct mlx4_ib_cq *cq = to_mcq(ibcq);
377 mutex_lock(&cq->resize_mutex);
378 if (entries < 1 || entries > dev->dev->caps.max_cqes) {
383 entries = roundup_pow_of_two(entries + 1);
384 if (entries == ibcq->cqe + 1) {
389 if (entries > dev->dev->caps.max_cqes + 1) {
395 err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
399 /* Can't be smaller than the number of outstanding CQEs */
400 outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
401 if (entries < outst_cqe + 1) {
406 err = mlx4_alloc_resize_buf(dev, cq, entries);
413 err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
417 mlx4_mtt_cleanup(dev->dev, &mtt);
419 cq->buf = cq->resize_buf->buf;
420 cq->ibcq.cqe = cq->resize_buf->cqe;
421 ib_umem_release(cq->umem);
422 cq->umem = cq->resize_umem;
424 kfree(cq->resize_buf);
425 cq->resize_buf = NULL;
426 cq->resize_umem = NULL;
428 struct mlx4_ib_cq_buf tmp_buf;
431 spin_lock_irq(&cq->lock);
432 if (cq->resize_buf) {
433 mlx4_ib_cq_resize_copy_cqes(cq);
435 tmp_cqe = cq->ibcq.cqe;
436 cq->buf = cq->resize_buf->buf;
437 cq->ibcq.cqe = cq->resize_buf->cqe;
439 kfree(cq->resize_buf);
440 cq->resize_buf = NULL;
442 spin_unlock_irq(&cq->lock);
445 mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
451 mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
453 mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
454 cq->resize_buf->cqe);
456 kfree(cq->resize_buf);
457 cq->resize_buf = NULL;
459 if (cq->resize_umem) {
460 ib_umem_release(cq->resize_umem);
461 cq->resize_umem = NULL;
465 mutex_unlock(&cq->resize_mutex);
470 int mlx4_ib_destroy_cq(struct ib_cq *cq)
472 struct mlx4_ib_dev *dev = to_mdev(cq->device);
473 struct mlx4_ib_cq *mcq = to_mcq(cq);
475 mlx4_cq_free(dev->dev, &mcq->mcq);
476 mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
479 mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
480 ib_umem_release(mcq->umem);
482 mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
483 mlx4_db_free(dev->dev, &mcq->db);
491 static void dump_cqe(void *cqe)
495 pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
496 be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
497 be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
498 be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
501 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
504 if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
505 pr_debug("local QP operation err "
506 "(QPN %06x, WQE index %x, vendor syndrome %02x, "
508 be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
509 cqe->vendor_err_syndrome,
510 cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
514 switch (cqe->syndrome) {
515 case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
516 wc->status = IB_WC_LOC_LEN_ERR;
518 case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
519 wc->status = IB_WC_LOC_QP_OP_ERR;
521 case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
522 wc->status = IB_WC_LOC_PROT_ERR;
524 case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
525 wc->status = IB_WC_WR_FLUSH_ERR;
527 case MLX4_CQE_SYNDROME_MW_BIND_ERR:
528 wc->status = IB_WC_MW_BIND_ERR;
530 case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
531 wc->status = IB_WC_BAD_RESP_ERR;
533 case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
534 wc->status = IB_WC_LOC_ACCESS_ERR;
536 case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
537 wc->status = IB_WC_REM_INV_REQ_ERR;
539 case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
540 wc->status = IB_WC_REM_ACCESS_ERR;
542 case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
543 wc->status = IB_WC_REM_OP_ERR;
545 case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
546 wc->status = IB_WC_RETRY_EXC_ERR;
548 case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
549 wc->status = IB_WC_RNR_RETRY_EXC_ERR;
551 case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
552 wc->status = IB_WC_REM_ABORT_ERR;
555 wc->status = IB_WC_GENERAL_ERR;
559 wc->vendor_err = cqe->vendor_err_syndrome;
562 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
564 return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
565 MLX4_CQE_STATUS_IPV4F |
566 MLX4_CQE_STATUS_IPV4OPT |
567 MLX4_CQE_STATUS_IPV6 |
568 MLX4_CQE_STATUS_IPOK)) ==
569 cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
570 MLX4_CQE_STATUS_IPOK)) &&
571 (status & cpu_to_be16(MLX4_CQE_STATUS_UDP |
572 MLX4_CQE_STATUS_TCP)) &&
573 checksum == cpu_to_be16(0xffff);
576 static int use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
577 unsigned tail, struct mlx4_cqe *cqe, int is_eth)
579 struct mlx4_ib_proxy_sqp_hdr *hdr;
581 ib_dma_sync_single_for_cpu(qp->ibqp.device,
582 qp->sqp_proxy_rcv[tail].map,
583 sizeof (struct mlx4_ib_proxy_sqp_hdr),
585 hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
586 wc->pkey_index = be16_to_cpu(hdr->tun.pkey_index);
587 wc->src_qp = be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
588 wc->wc_flags |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
589 wc->dlid_path_bits = 0;
592 wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
593 memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
594 memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
595 wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
597 wc->slid = be16_to_cpu(hdr->tun.slid_mac_47_32);
598 wc->sl = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
604 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
605 struct ib_wc *wc, int *npolled, int is_send)
607 struct mlx4_ib_wq *wq;
611 wq = is_send ? &qp->sq : &qp->rq;
612 cur = wq->head - wq->tail;
617 for (i = 0; i < cur && *npolled < num_entries; i++) {
618 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
619 wc->status = IB_WC_WR_FLUSH_ERR;
620 wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
628 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
629 struct ib_wc *wc, int *npolled)
631 struct mlx4_ib_qp *qp;
634 /* Find uncompleted WQEs belonging to that cq and retrun
635 * simulated FLUSH_ERR completions
637 list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
638 mlx4_ib_qp_sw_comp(qp, num_entries, wc, npolled, 1);
639 if (*npolled >= num_entries)
643 list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
644 mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
645 if (*npolled >= num_entries)
653 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
654 struct mlx4_ib_qp **cur_qp,
657 struct mlx4_cqe *cqe;
659 struct mlx4_ib_wq *wq;
660 struct mlx4_ib_srq *srq;
661 struct mlx4_srq *msrq = NULL;
670 cqe = next_cqe_sw(cq);
674 if (cq->buf.entry_size == 64)
677 ++cq->mcq.cons_index;
680 * Make sure we read CQ entry contents after we've checked the
685 is_send = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
686 is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
687 MLX4_CQE_OPCODE_ERROR;
689 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
691 pr_warn("Completion for NOP opcode detected!\n");
695 /* Resize CQ in progress */
696 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
697 if (cq->resize_buf) {
698 struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
700 mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
701 cq->buf = cq->resize_buf->buf;
702 cq->ibcq.cqe = cq->resize_buf->cqe;
704 kfree(cq->resize_buf);
705 cq->resize_buf = NULL;
712 (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
714 * We do not have to take the QP table lock here,
715 * because CQs will be locked while QPs are removed
718 mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
719 be32_to_cpu(cqe->vlan_my_qpn));
720 if (unlikely(!mqp)) {
721 pr_warn("CQ %06x with entry for unknown QPN %06x\n",
722 cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
726 *cur_qp = to_mibqp(mqp);
729 wc->qp = &(*cur_qp)->ibqp;
731 if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
733 g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
734 srq_num = g_mlpath_rqpn & 0xffffff;
735 /* SRQ is also in the radix tree */
736 msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
738 if (unlikely(!msrq)) {
739 pr_warn("CQ %06x with entry for unknown SRQN %06x\n",
740 cq->mcq.cqn, srq_num);
747 if (!(*cur_qp)->sq_signal_bits) {
748 wqe_ctr = be16_to_cpu(cqe->wqe_index);
749 wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
751 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
753 } else if ((*cur_qp)->ibqp.srq) {
754 srq = to_msrq((*cur_qp)->ibqp.srq);
755 wqe_ctr = be16_to_cpu(cqe->wqe_index);
756 wc->wr_id = srq->wrid[wqe_ctr];
757 mlx4_ib_free_srq_wqe(srq, wqe_ctr);
759 srq = to_mibsrq(msrq);
760 wqe_ctr = be16_to_cpu(cqe->wqe_index);
761 wc->wr_id = srq->wrid[wqe_ctr];
762 mlx4_ib_free_srq_wqe(srq, wqe_ctr);
765 tail = wq->tail & (wq->wqe_cnt - 1);
766 wc->wr_id = wq->wrid[tail];
770 if (unlikely(is_error)) {
771 mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
775 wc->status = IB_WC_SUCCESS;
779 switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
780 case MLX4_OPCODE_RDMA_WRITE_IMM:
781 wc->wc_flags |= IB_WC_WITH_IMM;
782 case MLX4_OPCODE_RDMA_WRITE:
783 wc->opcode = IB_WC_RDMA_WRITE;
785 case MLX4_OPCODE_SEND_IMM:
786 wc->wc_flags |= IB_WC_WITH_IMM;
787 case MLX4_OPCODE_SEND:
788 case MLX4_OPCODE_SEND_INVAL:
789 wc->opcode = IB_WC_SEND;
791 case MLX4_OPCODE_RDMA_READ:
792 wc->opcode = IB_WC_RDMA_READ;
793 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
795 case MLX4_OPCODE_ATOMIC_CS:
796 wc->opcode = IB_WC_COMP_SWAP;
799 case MLX4_OPCODE_ATOMIC_FA:
800 wc->opcode = IB_WC_FETCH_ADD;
803 case MLX4_OPCODE_MASKED_ATOMIC_CS:
804 wc->opcode = IB_WC_MASKED_COMP_SWAP;
807 case MLX4_OPCODE_MASKED_ATOMIC_FA:
808 wc->opcode = IB_WC_MASKED_FETCH_ADD;
811 case MLX4_OPCODE_BIND_MW:
812 wc->opcode = IB_WC_BIND_MW;
814 case MLX4_OPCODE_LSO:
815 wc->opcode = IB_WC_LSO;
817 case MLX4_OPCODE_FMR:
818 wc->opcode = IB_WC_FAST_REG_MR;
820 case MLX4_OPCODE_LOCAL_INVAL:
821 wc->opcode = IB_WC_LOCAL_INV;
825 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
827 switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
828 case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
829 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
830 wc->wc_flags = IB_WC_WITH_IMM;
831 wc->ex.imm_data = cqe->immed_rss_invalid;
833 case MLX4_RECV_OPCODE_SEND_INVAL:
834 wc->opcode = IB_WC_RECV;
835 wc->wc_flags = IB_WC_WITH_INVALIDATE;
836 wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
838 case MLX4_RECV_OPCODE_SEND:
839 wc->opcode = IB_WC_RECV;
842 case MLX4_RECV_OPCODE_SEND_IMM:
843 wc->opcode = IB_WC_RECV;
844 wc->wc_flags = IB_WC_WITH_IMM;
845 wc->ex.imm_data = cqe->immed_rss_invalid;
849 is_eth = (rdma_port_get_link_layer(wc->qp->device,
851 IB_LINK_LAYER_ETHERNET);
852 if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
853 if ((*cur_qp)->mlx4_ib_qp_type &
854 (MLX4_IB_QPT_PROXY_SMI_OWNER |
855 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
856 return use_tunnel_data(*cur_qp, cq, wc, tail,
860 wc->slid = be16_to_cpu(cqe->rlid);
861 g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
862 wc->src_qp = g_mlpath_rqpn & 0xffffff;
863 wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
864 wc->wc_flags |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
865 wc->pkey_index = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
866 wc->wc_flags |= mlx4_ib_ipoib_csum_ok(cqe->status,
867 cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
869 wc->sl = be16_to_cpu(cqe->sl_vid) >> 13;
870 if (be32_to_cpu(cqe->vlan_my_qpn) &
871 MLX4_CQE_VLAN_PRESENT_MASK) {
872 wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
875 wc->vlan_id = 0xffff;
877 memcpy(wc->smac, cqe->smac, ETH_ALEN);
878 wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
880 wc->sl = be16_to_cpu(cqe->sl_vid) >> 12;
881 wc->vlan_id = 0xffff;
888 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
890 struct mlx4_ib_cq *cq = to_mcq(ibcq);
891 struct mlx4_ib_qp *cur_qp = NULL;
895 struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
897 spin_lock_irqsave(&cq->lock, flags);
898 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
899 mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
903 for (npolled = 0; npolled < num_entries; ++npolled) {
904 err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
909 mlx4_cq_set_ci(&cq->mcq);
912 spin_unlock_irqrestore(&cq->lock, flags);
914 if (err == 0 || err == -EAGAIN)
920 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
922 mlx4_cq_arm(&to_mcq(ibcq)->mcq,
923 (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
924 MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
925 to_mdev(ibcq->device)->uar_map,
926 MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
931 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
935 struct mlx4_cqe *cqe, *dest;
937 int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
940 * First we need to find the current producer index, so we
941 * know where to start cleaning from. It doesn't matter if HW
942 * adds new entries after this loop -- the QP we're worried
943 * about is already in RESET, so the new entries won't come
944 * from our QP and therefore don't need to be checked.
946 for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
947 if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
951 * Now sweep backwards through the CQ, removing CQ entries
952 * that match our QP by copying older entries on top of them.
954 while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
955 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
958 if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
959 if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
960 mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
963 dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
966 owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
967 memcpy(dest, cqe, sizeof *cqe);
968 dest->owner_sr_opcode = owner_bit |
969 (dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
974 cq->mcq.cons_index += nfreed;
976 * Make sure update of buffer contents is done before
977 * updating consumer index.
980 mlx4_cq_set_ci(&cq->mcq);
984 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
986 spin_lock_irq(&cq->lock);
987 __mlx4_ib_cq_clean(cq, qpn, srq);
988 spin_unlock_irq(&cq->lock);