4 * Copyright 2015 6WIND S.A.
5 * Copyright 2015 Mellanox.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of 6WIND S.A. nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
42 #pragma GCC diagnostic ignored "-pedantic"
44 #include <infiniband/verbs.h>
46 #pragma GCC diagnostic error "-pedantic"
49 /* DPDK headers don't like -pedantic. */
51 #pragma GCC diagnostic ignored "-pedantic"
54 #include <rte_mempool.h>
55 #include <rte_prefetch.h>
56 #include <rte_common.h>
57 #include <rte_branch_prediction.h>
59 #pragma GCC diagnostic error "-pedantic"
63 #include "mlx5_utils.h"
64 #include "mlx5_rxtx.h"
65 #include "mlx5_defs.h"
68 * Manage TX completions.
70 * When sending a burst, mlx5_tx_burst() posts several WRs.
71 * To improve performance, a completion event is only required once every
72 * MLX5_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
73 * for other WRs, but this information would not be used anyway.
76 * Pointer to TX queue structure.
79 * 0 on success, -1 on failure.
82 txq_complete(struct txq *txq)
84 unsigned int elts_comp = txq->elts_comp;
85 unsigned int elts_tail = txq->elts_tail;
86 const unsigned int elts_n = txq->elts_n;
89 if (unlikely(elts_comp == 0))
92 DEBUG("%p: processing %u work requests completions",
93 (void *)txq, elts_comp);
95 wcs_n = txq->if_cq->poll_cnt(txq->cq, elts_comp);
96 if (unlikely(wcs_n == 0))
98 if (unlikely(wcs_n < 0)) {
99 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
104 assert(elts_comp <= txq->elts_comp);
106 * Assume WC status is successful as nothing can be done about it
109 elts_tail += wcs_n * txq->elts_comp_cd_init;
110 if (elts_tail >= elts_n)
112 txq->elts_tail = elts_tail;
113 txq->elts_comp = elts_comp;
118 * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
119 * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
120 * remove an entry first.
123 * Pointer to TX queue structure.
125 * Memory Pool for which a Memory Region lkey must be returned.
128 * mr->lkey on success, (uint32_t)-1 on failure.
131 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
136 for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
137 if (unlikely(txq->mp2mr[i].mp == NULL)) {
138 /* Unknown MP, add a new MR for it. */
141 if (txq->mp2mr[i].mp == mp) {
142 assert(txq->mp2mr[i].lkey != (uint32_t)-1);
143 assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
144 return txq->mp2mr[i].lkey;
147 /* Add a new entry, register MR first. */
148 DEBUG("%p: discovered new memory pool %p", (void *)txq, (void *)mp);
149 mr = ibv_reg_mr(txq->priv->pd,
150 (void *)mp->elt_va_start,
151 (mp->elt_va_end - mp->elt_va_start),
152 (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
153 if (unlikely(mr == NULL)) {
154 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
158 if (unlikely(i == RTE_DIM(txq->mp2mr))) {
159 /* Table is full, remove oldest entry. */
160 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
163 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
164 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
165 (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
167 /* Store the new entry. */
168 txq->mp2mr[i].mp = mp;
169 txq->mp2mr[i].mr = mr;
170 txq->mp2mr[i].lkey = mr->lkey;
171 DEBUG("%p: new MR lkey for MP %p: 0x%08" PRIu32,
172 (void *)txq, (void *)mp, txq->mp2mr[i].lkey);
173 return txq->mp2mr[i].lkey;
176 #if MLX5_PMD_SGE_WR_N > 1
179 * Copy scattered mbuf contents to a single linear buffer.
182 * Linear output buffer.
184 * Scattered input buffer.
187 * Number of bytes copied to the output buffer or 0 if not large enough.
190 linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
192 unsigned int size = 0;
196 unsigned int len = DATA_LEN(buf);
200 if (unlikely(size > sizeof(*linear)))
202 memcpy(&(*linear)[offset],
203 rte_pktmbuf_mtod(buf, uint8_t *),
206 } while (buf != NULL);
211 * Handle scattered buffers for mlx5_tx_burst().
214 * TX queue structure.
216 * Number of segments in buf.
218 * TX queue element to fill.
222 * Index of the linear buffer to use if necessary (normally txq->elts_head).
224 * Array filled with SGEs on success.
227 * A structure containing the processed packet size in bytes and the
228 * number of SGEs. Both fields are set to (unsigned int)-1 in case of
231 static struct tx_burst_sg_ret {
235 tx_burst_sg(struct txq *txq, unsigned int segs, struct txq_elt *elt,
236 struct rte_mbuf *buf, unsigned int elts_head,
237 struct ibv_sge (*sges)[MLX5_PMD_SGE_WR_N])
239 unsigned int sent_size = 0;
243 /* When there are too many segments, extra segments are
244 * linearized in the last SGE. */
245 if (unlikely(segs > RTE_DIM(*sges))) {
246 segs = (RTE_DIM(*sges) - 1);
249 /* Update element. */
251 /* Register segments as SGEs. */
252 for (j = 0; (j != segs); ++j) {
253 struct ibv_sge *sge = &(*sges)[j];
256 /* Retrieve Memory Region key for this memory pool. */
257 lkey = txq_mp2mr(txq, buf->pool);
258 if (unlikely(lkey == (uint32_t)-1)) {
259 /* MR does not exist. */
260 DEBUG("%p: unable to get MP <-> MR association",
262 /* Clean up TX element. */
267 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
269 rte_prefetch0((volatile void *)
270 (uintptr_t)sge->addr);
271 sge->length = DATA_LEN(buf);
273 sent_size += sge->length;
276 /* If buf is not NULL here and is not going to be linearized,
277 * nb_segs is not valid. */
279 assert((buf == NULL) || (linearize));
280 /* Linearize extra segments. */
282 struct ibv_sge *sge = &(*sges)[segs];
283 linear_t *linear = &(*txq->elts_linear)[elts_head];
284 unsigned int size = linearize_mbuf(linear, buf);
286 assert(segs == (RTE_DIM(*sges) - 1));
288 /* Invalid packet. */
289 DEBUG("%p: packet too large to be linearized.",
291 /* Clean up TX element. */
295 /* If MLX5_PMD_SGE_WR_N is 1, free mbuf immediately. */
296 if (RTE_DIM(*sges) == 1) {
298 struct rte_mbuf *next = NEXT(buf);
300 rte_pktmbuf_free_seg(buf);
302 } while (buf != NULL);
306 sge->addr = (uintptr_t)&(*linear)[0];
308 sge->lkey = txq->mr_linear->lkey;
311 return (struct tx_burst_sg_ret){
316 return (struct tx_burst_sg_ret){
322 #endif /* MLX5_PMD_SGE_WR_N > 1 */
325 * DPDK callback for TX.
328 * Generic pointer to TX queue structure.
330 * Packets to transmit.
332 * Number of packets in array.
335 * Number of packets successfully transmitted (<= pkts_n).
338 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
340 struct txq *txq = (struct txq *)dpdk_txq;
341 unsigned int elts_head = txq->elts_head;
342 const unsigned int elts_tail = txq->elts_tail;
343 const unsigned int elts_n = txq->elts_n;
344 unsigned int elts_comp_cd = txq->elts_comp_cd;
345 unsigned int elts_comp = 0;
350 assert(elts_comp_cd != 0);
352 max = (elts_n - (elts_head - elts_tail));
356 assert(max <= elts_n);
357 /* Always leave one free entry in the ring. */
363 for (i = 0; (i != max); ++i) {
364 struct rte_mbuf *buf = pkts[i];
365 unsigned int elts_head_next =
366 (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
367 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
368 struct txq_elt *elt = &(*txq->elts)[elts_head];
369 unsigned int segs = NB_SEGS(buf);
370 uint32_t send_flags = 0;
372 /* Clean up old buffer. */
373 if (likely(elt->buf != NULL)) {
374 struct rte_mbuf *tmp = elt->buf;
376 /* Faster than rte_pktmbuf_free(). */
378 struct rte_mbuf *next = NEXT(tmp);
380 rte_pktmbuf_free_seg(tmp);
382 } while (tmp != NULL);
384 /* Request TX completion. */
385 if (unlikely(--elts_comp_cd == 0)) {
386 elts_comp_cd = txq->elts_comp_cd_init;
388 send_flags |= IBV_EXP_QP_BURST_SIGNALED;
390 if (likely(segs == 1)) {
395 /* Retrieve buffer information. */
396 addr = rte_pktmbuf_mtod(buf, uintptr_t);
397 length = DATA_LEN(buf);
398 /* Retrieve Memory Region key for this memory pool. */
399 lkey = txq_mp2mr(txq, buf->pool);
400 if (unlikely(lkey == (uint32_t)-1)) {
401 /* MR does not exist. */
402 DEBUG("%p: unable to get MP <-> MR"
403 " association", (void *)txq);
404 /* Clean up TX element. */
408 /* Update element. */
411 rte_prefetch0((volatile void *)
413 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
414 /* Put packet into send queue. */
415 #if MLX5_PMD_MAX_INLINE > 0
416 if (length <= txq->max_inline)
417 err = txq->if_qp->send_pending_inline
424 err = txq->if_qp->send_pending
433 #if MLX5_PMD_SGE_WR_N > 1
434 struct ibv_sge sges[MLX5_PMD_SGE_WR_N];
435 struct tx_burst_sg_ret ret;
437 ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
439 if (ret.length == (unsigned int)-1)
441 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
442 /* Put SG list into send queue. */
443 err = txq->if_qp->send_pending_sg_list
450 #else /* MLX5_PMD_SGE_WR_N > 1 */
451 DEBUG("%p: TX scattered buffers support not"
452 " compiled in", (void *)txq);
454 #endif /* MLX5_PMD_SGE_WR_N > 1 */
456 elts_head = elts_head_next;
459 /* Take a shortcut if nothing must be sent. */
460 if (unlikely(i == 0))
462 /* Ring QP doorbell. */
463 err = txq->if_qp->send_flush(txq->qp);
465 /* A nonzero value is not supposed to be returned.
466 * Nothing can be done about it. */
467 DEBUG("%p: send_flush() failed with error %d",
470 txq->elts_head = elts_head;
471 txq->elts_comp += elts_comp;
472 txq->elts_comp_cd = elts_comp_cd;
477 * DPDK callback for RX with scattered packets support.
480 * Generic pointer to RX queue structure.
482 * Array to store received packets.
484 * Maximum number of packets in array.
487 * Number of packets successfully received (<= pkts_n).
490 mlx5_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
492 struct rxq *rxq = (struct rxq *)dpdk_rxq;
493 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
494 const unsigned int elts_n = rxq->elts_n;
495 unsigned int elts_head = rxq->elts_head;
496 struct ibv_recv_wr head;
497 struct ibv_recv_wr **next = &head.next;
498 struct ibv_recv_wr *bad_wr;
500 unsigned int pkts_ret = 0;
503 if (unlikely(!rxq->sp))
504 return mlx5_rx_burst(dpdk_rxq, pkts, pkts_n);
505 if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
507 for (i = 0; (i != pkts_n); ++i) {
508 struct rxq_elt_sp *elt = &(*elts)[elts_head];
509 struct ibv_recv_wr *wr = &elt->wr;
510 uint64_t wr_id = wr->wr_id;
512 unsigned int pkt_buf_len;
513 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
514 struct rte_mbuf **pkt_buf_next = &pkt_buf;
515 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
523 assert(wr_id < rxq->elts_n);
524 assert(wr->sg_list == elt->sges);
525 assert(wr->num_sge == RTE_DIM(elt->sges));
526 assert(elts_head < rxq->elts_n);
527 assert(rxq->elts_head < rxq->elts_n);
528 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
530 if (unlikely(ret < 0)) {
534 DEBUG("rxq=%p, poll_length() failed (ret=%d)",
536 /* ibv_poll_cq() must be used in case of failure. */
537 wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
538 if (unlikely(wcs_n == 0))
540 if (unlikely(wcs_n < 0)) {
541 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
546 if (unlikely(wc.status != IBV_WC_SUCCESS)) {
547 /* Whatever, just repost the offending WR. */
548 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
549 " completion status (%d): %s",
550 (void *)rxq, wc.wr_id, wc.status,
551 ibv_wc_status_str(wc.status));
552 /* Link completed WRs together for repost. */
563 /* Link completed WRs together for repost. */
567 * Replace spent segments with new ones, concatenate and
568 * return them as pkt_buf.
571 struct ibv_sge *sge = &elt->sges[j];
572 struct rte_mbuf *seg = elt->bufs[j];
573 struct rte_mbuf *rep;
574 unsigned int seg_tailroom;
577 * Fetch initial bytes of packet descriptor into a
578 * cacheline while allocating rep.
581 rep = __rte_mbuf_raw_alloc(rxq->mp);
582 if (unlikely(rep == NULL)) {
584 * Unable to allocate a replacement mbuf,
587 DEBUG("rxq=%p, wr_id=%" PRIu64 ":"
588 " can't allocate a new mbuf",
590 if (pkt_buf != NULL) {
591 *pkt_buf_next = NULL;
592 rte_pktmbuf_free(pkt_buf);
594 /* Increment out of memory counters. */
595 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
599 /* Poison user-modifiable fields in rep. */
600 NEXT(rep) = (void *)((uintptr_t)-1);
601 SET_DATA_OFF(rep, 0xdead);
602 DATA_LEN(rep) = 0xd00d;
603 PKT_LEN(rep) = 0xdeadd00d;
608 assert(rep->buf_len == seg->buf_len);
609 assert(rep->buf_len == rxq->mb_len);
610 /* Reconfigure sge to use rep instead of seg. */
611 assert(sge->lkey == rxq->mr->lkey);
612 sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
615 /* Update pkt_buf if it's the first segment, or link
616 * seg to the previous one and update pkt_buf_next. */
618 pkt_buf_next = &NEXT(seg);
619 /* Update seg information. */
620 seg_tailroom = (seg->buf_len - seg_headroom);
621 assert(sge->length == seg_tailroom);
622 SET_DATA_OFF(seg, seg_headroom);
623 if (likely(len <= seg_tailroom)) {
628 assert(rte_pktmbuf_headroom(seg) ==
630 assert(rte_pktmbuf_tailroom(seg) ==
631 (seg_tailroom - len));
634 DATA_LEN(seg) = seg_tailroom;
635 PKT_LEN(seg) = seg_tailroom;
637 assert(rte_pktmbuf_headroom(seg) == seg_headroom);
638 assert(rte_pktmbuf_tailroom(seg) == 0);
639 /* Fix len and clear headroom for next segments. */
643 /* Update head and tail segments. */
644 *pkt_buf_next = NULL;
645 assert(pkt_buf != NULL);
647 NB_SEGS(pkt_buf) = j;
648 PORT(pkt_buf) = rxq->port_id;
649 PKT_LEN(pkt_buf) = pkt_buf_len;
655 if (++elts_head >= elts_n)
659 if (unlikely(i == 0))
664 DEBUG("%p: reposting %d WRs", (void *)rxq, i);
666 ret = ibv_post_recv(rxq->qp, head.next, &bad_wr);
668 /* Inability to repost WRs is fatal. */
669 DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
675 rxq->elts_head = elts_head;
680 * DPDK callback for RX.
682 * The following function is the same as mlx5_rx_burst_sp(), except it doesn't
683 * manage scattered packets. Improves performance when MRU is lower than the
684 * size of the first segment.
687 * Generic pointer to RX queue structure.
689 * Array to store received packets.
691 * Maximum number of packets in array.
694 * Number of packets successfully received (<= pkts_n).
697 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
699 struct rxq *rxq = (struct rxq *)dpdk_rxq;
700 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
701 const unsigned int elts_n = rxq->elts_n;
702 unsigned int elts_head = rxq->elts_head;
703 struct ibv_sge sges[pkts_n];
705 unsigned int pkts_ret = 0;
708 if (unlikely(rxq->sp))
709 return mlx5_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
710 for (i = 0; (i != pkts_n); ++i) {
711 struct rxq_elt *elt = &(*elts)[elts_head];
712 struct ibv_recv_wr *wr = &elt->wr;
713 uint64_t wr_id = wr->wr_id;
715 struct rte_mbuf *seg = (void *)((uintptr_t)elt->sge.addr -
716 WR_ID(wr_id).offset);
717 struct rte_mbuf *rep;
721 assert(WR_ID(wr_id).id < rxq->elts_n);
722 assert(wr->sg_list == &elt->sge);
723 assert(wr->num_sge == 1);
724 assert(elts_head < rxq->elts_n);
725 assert(rxq->elts_head < rxq->elts_n);
727 * Fetch initial bytes of packet descriptor into a
728 * cacheline while allocating rep.
731 rte_prefetch0(&seg->cacheline1);
732 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
734 if (unlikely(ret < 0)) {
738 DEBUG("rxq=%p, poll_length() failed (ret=%d)",
740 /* ibv_poll_cq() must be used in case of failure. */
741 wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
742 if (unlikely(wcs_n == 0))
744 if (unlikely(wcs_n < 0)) {
745 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
750 if (unlikely(wc.status != IBV_WC_SUCCESS)) {
751 /* Whatever, just repost the offending WR. */
752 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
753 " completion status (%d): %s",
754 (void *)rxq, wc.wr_id, wc.status,
755 ibv_wc_status_str(wc.status));
756 /* Add SGE to array for repost. */
765 rep = __rte_mbuf_raw_alloc(rxq->mp);
766 if (unlikely(rep == NULL)) {
768 * Unable to allocate a replacement mbuf,
771 DEBUG("rxq=%p, wr_id=%" PRIu32 ":"
772 " can't allocate a new mbuf",
773 (void *)rxq, WR_ID(wr_id).id);
774 /* Increment out of memory counters. */
775 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
779 /* Reconfigure sge to use rep instead of seg. */
780 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
781 assert(elt->sge.lkey == rxq->mr->lkey);
782 WR_ID(wr->wr_id).offset =
783 (((uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM) -
785 assert(WR_ID(wr->wr_id).id == WR_ID(wr_id).id);
787 /* Add SGE to array for repost. */
790 /* Update seg information. */
791 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
793 PORT(seg) = rxq->port_id;
802 if (++elts_head >= elts_n)
806 if (unlikely(i == 0))
810 DEBUG("%p: reposting %u WRs", (void *)rxq, i);
812 ret = rxq->if_qp->recv_burst(rxq->qp, sges, i);
814 /* Inability to repost WRs is fatal. */
815 DEBUG("%p: recv_burst(): failed (ret=%d)",
820 rxq->elts_head = elts_head;
825 * Dummy DPDK callback for TX.
827 * This function is used to temporarily replace the real callback during
828 * unsafe control operations on the queue, or in case of error.
831 * Generic pointer to TX queue structure.
833 * Packets to transmit.
835 * Number of packets in array.
838 * Number of packets successfully transmitted (<= pkts_n).
841 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
850 * Dummy DPDK callback for RX.
852 * This function is used to temporarily replace the real callback during
853 * unsafe control operations on the queue, or in case of error.
856 * Generic pointer to RX queue structure.
858 * Array to store received packets.
860 * Maximum number of packets in array.
863 * Number of packets successfully received (<= pkts_n).
866 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)