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, int entries, int vector,
170 struct ib_ucontext *context,
171 struct ib_udata *udata)
173 struct mlx4_ib_dev *dev = to_mdev(ibdev);
174 struct mlx4_ib_cq *cq;
175 struct mlx4_uar *uar;
178 if (entries < 1 || entries > dev->dev->caps.max_cqes)
179 return ERR_PTR(-EINVAL);
181 cq = kmalloc(sizeof *cq, GFP_KERNEL);
183 return ERR_PTR(-ENOMEM);
185 entries = roundup_pow_of_two(entries + 1);
186 cq->ibcq.cqe = entries - 1;
187 mutex_init(&cq->resize_mutex);
188 spin_lock_init(&cq->lock);
189 cq->resize_buf = NULL;
190 cq->resize_umem = NULL;
191 INIT_LIST_HEAD(&cq->send_qp_list);
192 INIT_LIST_HEAD(&cq->recv_qp_list);
195 struct mlx4_ib_create_cq ucmd;
197 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
202 err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
203 ucmd.buf_addr, entries);
207 err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
212 uar = &to_mucontext(context)->uar;
214 err = mlx4_db_alloc(dev->dev, &cq->db, 1, GFP_KERNEL);
218 cq->mcq.set_ci_db = cq->db.db;
219 cq->mcq.arm_db = cq->db.db + 1;
220 *cq->mcq.set_ci_db = 0;
223 err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
227 uar = &dev->priv_uar;
231 vector = dev->eq_table[vector % ibdev->num_comp_vectors];
233 err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
234 cq->db.dma, &cq->mcq, vector, 0, 0);
239 cq->mcq.tasklet_ctx.comp = mlx4_ib_cq_comp;
241 cq->mcq.comp = mlx4_ib_cq_comp;
242 cq->mcq.event = mlx4_ib_cq_event;
245 if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
254 mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
257 mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
260 ib_umem_release(cq->umem);
262 mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
266 mlx4_db_free(dev->dev, &cq->db);
274 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
282 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
286 err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
288 kfree(cq->resize_buf);
289 cq->resize_buf = NULL;
293 cq->resize_buf->cqe = entries - 1;
298 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
299 int entries, struct ib_udata *udata)
301 struct mlx4_ib_resize_cq ucmd;
307 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
310 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
314 err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
315 &cq->resize_umem, ucmd.buf_addr, entries);
317 kfree(cq->resize_buf);
318 cq->resize_buf = NULL;
322 cq->resize_buf->cqe = entries - 1;
327 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
331 i = cq->mcq.cons_index;
332 while (get_sw_cqe(cq, i))
335 return i - cq->mcq.cons_index;
338 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
340 struct mlx4_cqe *cqe, *new_cqe;
342 int cqe_size = cq->buf.entry_size;
343 int cqe_inc = cqe_size == 64 ? 1 : 0;
345 i = cq->mcq.cons_index;
346 cqe = get_cqe(cq, i & cq->ibcq.cqe);
349 while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
350 new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
351 (i + 1) & cq->resize_buf->cqe);
352 memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
355 new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
356 (((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
357 cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
360 ++cq->mcq.cons_index;
363 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
365 struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
366 struct mlx4_ib_cq *cq = to_mcq(ibcq);
371 mutex_lock(&cq->resize_mutex);
378 entries = roundup_pow_of_two(entries + 1);
379 if (entries == ibcq->cqe + 1) {
384 if (entries > dev->dev->caps.max_cqes) {
390 err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
394 /* Can't be smaller than the number of outstanding CQEs */
395 outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
396 if (entries < outst_cqe + 1) {
401 err = mlx4_alloc_resize_buf(dev, cq, entries);
408 err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
412 mlx4_mtt_cleanup(dev->dev, &mtt);
414 cq->buf = cq->resize_buf->buf;
415 cq->ibcq.cqe = cq->resize_buf->cqe;
416 ib_umem_release(cq->umem);
417 cq->umem = cq->resize_umem;
419 kfree(cq->resize_buf);
420 cq->resize_buf = NULL;
421 cq->resize_umem = NULL;
423 struct mlx4_ib_cq_buf tmp_buf;
426 spin_lock_irq(&cq->lock);
427 if (cq->resize_buf) {
428 mlx4_ib_cq_resize_copy_cqes(cq);
430 tmp_cqe = cq->ibcq.cqe;
431 cq->buf = cq->resize_buf->buf;
432 cq->ibcq.cqe = cq->resize_buf->cqe;
434 kfree(cq->resize_buf);
435 cq->resize_buf = NULL;
437 spin_unlock_irq(&cq->lock);
440 mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
446 mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
448 mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
449 cq->resize_buf->cqe);
451 kfree(cq->resize_buf);
452 cq->resize_buf = NULL;
454 if (cq->resize_umem) {
455 ib_umem_release(cq->resize_umem);
456 cq->resize_umem = NULL;
460 mutex_unlock(&cq->resize_mutex);
465 int mlx4_ib_destroy_cq(struct ib_cq *cq)
467 struct mlx4_ib_dev *dev = to_mdev(cq->device);
468 struct mlx4_ib_cq *mcq = to_mcq(cq);
470 mlx4_cq_free(dev->dev, &mcq->mcq);
471 mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
474 mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
475 ib_umem_release(mcq->umem);
477 mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
478 mlx4_db_free(dev->dev, &mcq->db);
486 static void dump_cqe(void *cqe)
490 pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
491 be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
492 be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
493 be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
496 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
499 if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
500 pr_debug("local QP operation err "
501 "(QPN %06x, WQE index %x, vendor syndrome %02x, "
503 be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
504 cqe->vendor_err_syndrome,
505 cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
509 switch (cqe->syndrome) {
510 case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
511 wc->status = IB_WC_LOC_LEN_ERR;
513 case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
514 wc->status = IB_WC_LOC_QP_OP_ERR;
516 case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
517 wc->status = IB_WC_LOC_PROT_ERR;
519 case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
520 wc->status = IB_WC_WR_FLUSH_ERR;
522 case MLX4_CQE_SYNDROME_MW_BIND_ERR:
523 wc->status = IB_WC_MW_BIND_ERR;
525 case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
526 wc->status = IB_WC_BAD_RESP_ERR;
528 case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
529 wc->status = IB_WC_LOC_ACCESS_ERR;
531 case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
532 wc->status = IB_WC_REM_INV_REQ_ERR;
534 case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
535 wc->status = IB_WC_REM_ACCESS_ERR;
537 case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
538 wc->status = IB_WC_REM_OP_ERR;
540 case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
541 wc->status = IB_WC_RETRY_EXC_ERR;
543 case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
544 wc->status = IB_WC_RNR_RETRY_EXC_ERR;
546 case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
547 wc->status = IB_WC_REM_ABORT_ERR;
550 wc->status = IB_WC_GENERAL_ERR;
554 wc->vendor_err = cqe->vendor_err_syndrome;
557 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
559 return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
560 MLX4_CQE_STATUS_IPV4F |
561 MLX4_CQE_STATUS_IPV4OPT |
562 MLX4_CQE_STATUS_IPV6 |
563 MLX4_CQE_STATUS_IPOK)) ==
564 cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
565 MLX4_CQE_STATUS_IPOK)) &&
566 (status & cpu_to_be16(MLX4_CQE_STATUS_UDP |
567 MLX4_CQE_STATUS_TCP)) &&
568 checksum == cpu_to_be16(0xffff);
571 static int use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
572 unsigned tail, struct mlx4_cqe *cqe, int is_eth)
574 struct mlx4_ib_proxy_sqp_hdr *hdr;
576 ib_dma_sync_single_for_cpu(qp->ibqp.device,
577 qp->sqp_proxy_rcv[tail].map,
578 sizeof (struct mlx4_ib_proxy_sqp_hdr),
580 hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
581 wc->pkey_index = be16_to_cpu(hdr->tun.pkey_index);
582 wc->src_qp = be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
583 wc->wc_flags |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
584 wc->dlid_path_bits = 0;
587 wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
588 memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
589 memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
590 wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
592 wc->slid = be16_to_cpu(hdr->tun.slid_mac_47_32);
593 wc->sl = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
599 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
600 struct ib_wc *wc, int *npolled, int is_send)
602 struct mlx4_ib_wq *wq;
606 wq = is_send ? &qp->sq : &qp->rq;
607 cur = wq->head - wq->tail;
612 for (i = 0; i < cur && *npolled < num_entries; i++) {
613 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
614 wc->status = IB_WC_WR_FLUSH_ERR;
615 wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
623 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
624 struct ib_wc *wc, int *npolled)
626 struct mlx4_ib_qp *qp;
629 /* Find uncompleted WQEs belonging to that cq and retrun
630 * simulated FLUSH_ERR completions
632 list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
633 mlx4_ib_qp_sw_comp(qp, num_entries, wc, npolled, 1);
634 if (*npolled >= num_entries)
638 list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
639 mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
640 if (*npolled >= num_entries)
648 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
649 struct mlx4_ib_qp **cur_qp,
652 struct mlx4_cqe *cqe;
654 struct mlx4_ib_wq *wq;
655 struct mlx4_ib_srq *srq;
656 struct mlx4_srq *msrq = NULL;
665 cqe = next_cqe_sw(cq);
669 if (cq->buf.entry_size == 64)
672 ++cq->mcq.cons_index;
675 * Make sure we read CQ entry contents after we've checked the
680 is_send = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
681 is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
682 MLX4_CQE_OPCODE_ERROR;
684 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
686 pr_warn("Completion for NOP opcode detected!\n");
690 /* Resize CQ in progress */
691 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
692 if (cq->resize_buf) {
693 struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
695 mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
696 cq->buf = cq->resize_buf->buf;
697 cq->ibcq.cqe = cq->resize_buf->cqe;
699 kfree(cq->resize_buf);
700 cq->resize_buf = NULL;
707 (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
709 * We do not have to take the QP table lock here,
710 * because CQs will be locked while QPs are removed
713 mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
714 be32_to_cpu(cqe->vlan_my_qpn));
715 if (unlikely(!mqp)) {
716 pr_warn("CQ %06x with entry for unknown QPN %06x\n",
717 cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
721 *cur_qp = to_mibqp(mqp);
724 wc->qp = &(*cur_qp)->ibqp;
726 if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
728 g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
729 srq_num = g_mlpath_rqpn & 0xffffff;
730 /* SRQ is also in the radix tree */
731 msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
733 if (unlikely(!msrq)) {
734 pr_warn("CQ %06x with entry for unknown SRQN %06x\n",
735 cq->mcq.cqn, srq_num);
742 if (!(*cur_qp)->sq_signal_bits) {
743 wqe_ctr = be16_to_cpu(cqe->wqe_index);
744 wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
746 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
748 } else if ((*cur_qp)->ibqp.srq) {
749 srq = to_msrq((*cur_qp)->ibqp.srq);
750 wqe_ctr = be16_to_cpu(cqe->wqe_index);
751 wc->wr_id = srq->wrid[wqe_ctr];
752 mlx4_ib_free_srq_wqe(srq, wqe_ctr);
754 srq = to_mibsrq(msrq);
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);
760 tail = wq->tail & (wq->wqe_cnt - 1);
761 wc->wr_id = wq->wrid[tail];
765 if (unlikely(is_error)) {
766 mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
770 wc->status = IB_WC_SUCCESS;
774 switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
775 case MLX4_OPCODE_RDMA_WRITE_IMM:
776 wc->wc_flags |= IB_WC_WITH_IMM;
777 case MLX4_OPCODE_RDMA_WRITE:
778 wc->opcode = IB_WC_RDMA_WRITE;
780 case MLX4_OPCODE_SEND_IMM:
781 wc->wc_flags |= IB_WC_WITH_IMM;
782 case MLX4_OPCODE_SEND:
783 case MLX4_OPCODE_SEND_INVAL:
784 wc->opcode = IB_WC_SEND;
786 case MLX4_OPCODE_RDMA_READ:
787 wc->opcode = IB_WC_RDMA_READ;
788 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
790 case MLX4_OPCODE_ATOMIC_CS:
791 wc->opcode = IB_WC_COMP_SWAP;
794 case MLX4_OPCODE_ATOMIC_FA:
795 wc->opcode = IB_WC_FETCH_ADD;
798 case MLX4_OPCODE_MASKED_ATOMIC_CS:
799 wc->opcode = IB_WC_MASKED_COMP_SWAP;
802 case MLX4_OPCODE_MASKED_ATOMIC_FA:
803 wc->opcode = IB_WC_MASKED_FETCH_ADD;
806 case MLX4_OPCODE_BIND_MW:
807 wc->opcode = IB_WC_BIND_MW;
809 case MLX4_OPCODE_LSO:
810 wc->opcode = IB_WC_LSO;
812 case MLX4_OPCODE_FMR:
813 wc->opcode = IB_WC_FAST_REG_MR;
815 case MLX4_OPCODE_LOCAL_INVAL:
816 wc->opcode = IB_WC_LOCAL_INV;
820 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
822 switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
823 case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
824 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
825 wc->wc_flags = IB_WC_WITH_IMM;
826 wc->ex.imm_data = cqe->immed_rss_invalid;
828 case MLX4_RECV_OPCODE_SEND_INVAL:
829 wc->opcode = IB_WC_RECV;
830 wc->wc_flags = IB_WC_WITH_INVALIDATE;
831 wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
833 case MLX4_RECV_OPCODE_SEND:
834 wc->opcode = IB_WC_RECV;
837 case MLX4_RECV_OPCODE_SEND_IMM:
838 wc->opcode = IB_WC_RECV;
839 wc->wc_flags = IB_WC_WITH_IMM;
840 wc->ex.imm_data = cqe->immed_rss_invalid;
844 is_eth = (rdma_port_get_link_layer(wc->qp->device,
846 IB_LINK_LAYER_ETHERNET);
847 if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
848 if ((*cur_qp)->mlx4_ib_qp_type &
849 (MLX4_IB_QPT_PROXY_SMI_OWNER |
850 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
851 return use_tunnel_data(*cur_qp, cq, wc, tail,
855 wc->slid = be16_to_cpu(cqe->rlid);
856 g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
857 wc->src_qp = g_mlpath_rqpn & 0xffffff;
858 wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
859 wc->wc_flags |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
860 wc->pkey_index = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
861 wc->wc_flags |= mlx4_ib_ipoib_csum_ok(cqe->status,
862 cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
864 wc->sl = be16_to_cpu(cqe->sl_vid) >> 13;
865 if (be32_to_cpu(cqe->vlan_my_qpn) &
866 MLX4_CQE_VLAN_PRESENT_MASK) {
867 wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
870 wc->vlan_id = 0xffff;
872 memcpy(wc->smac, cqe->smac, ETH_ALEN);
873 wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
875 wc->sl = be16_to_cpu(cqe->sl_vid) >> 12;
876 wc->vlan_id = 0xffff;
883 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
885 struct mlx4_ib_cq *cq = to_mcq(ibcq);
886 struct mlx4_ib_qp *cur_qp = NULL;
890 struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
892 spin_lock_irqsave(&cq->lock, flags);
893 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
894 mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
898 for (npolled = 0; npolled < num_entries; ++npolled) {
899 err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
904 mlx4_cq_set_ci(&cq->mcq);
907 spin_unlock_irqrestore(&cq->lock, flags);
909 if (err == 0 || err == -EAGAIN)
915 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
917 mlx4_cq_arm(&to_mcq(ibcq)->mcq,
918 (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
919 MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
920 to_mdev(ibcq->device)->uar_map,
921 MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
926 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
930 struct mlx4_cqe *cqe, *dest;
932 int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
935 * First we need to find the current producer index, so we
936 * know where to start cleaning from. It doesn't matter if HW
937 * adds new entries after this loop -- the QP we're worried
938 * about is already in RESET, so the new entries won't come
939 * from our QP and therefore don't need to be checked.
941 for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
942 if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
946 * Now sweep backwards through the CQ, removing CQ entries
947 * that match our QP by copying older entries on top of them.
949 while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
950 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
953 if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
954 if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
955 mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
958 dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
961 owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
962 memcpy(dest, cqe, sizeof *cqe);
963 dest->owner_sr_opcode = owner_bit |
964 (dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
969 cq->mcq.cons_index += nfreed;
971 * Make sure update of buffer contents is done before
972 * updating consumer index.
975 mlx4_cq_set_ci(&cq->mcq);
979 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
981 spin_lock_irq(&cq->lock);
982 __mlx4_ib_cq_clean(cq, qpn, srq);
983 spin_unlock_irq(&cq->lock);