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 "-Wpedantic"
44 #include <infiniband/verbs.h>
45 #include <infiniband/mlx5_hw.h>
46 #include <infiniband/arch.h>
48 #pragma GCC diagnostic error "-Wpedantic"
51 /* DPDK headers don't like -pedantic. */
53 #pragma GCC diagnostic ignored "-Wpedantic"
56 #include <rte_mempool.h>
57 #include <rte_prefetch.h>
58 #include <rte_common.h>
59 #include <rte_branch_prediction.h>
60 #include <rte_ether.h>
62 #pragma GCC diagnostic error "-Wpedantic"
66 #include "mlx5_utils.h"
67 #include "mlx5_rxtx.h"
68 #include "mlx5_autoconf.h"
69 #include "mlx5_defs.h"
73 check_cqe(volatile struct mlx5_cqe *cqe,
74 unsigned int cqes_n, const uint16_t ci)
75 __attribute__((always_inline));
78 txq_complete(struct txq *txq) __attribute__((always_inline));
80 static inline uint32_t
81 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
82 __attribute__((always_inline));
85 mlx5_tx_dbrec(struct txq *txq, volatile struct mlx5_wqe *wqe)
86 __attribute__((always_inline));
88 static inline uint32_t
89 rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe)
90 __attribute__((always_inline));
93 mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe,
94 uint16_t cqe_cnt, uint32_t *rss_hash)
95 __attribute__((always_inline));
97 static inline uint32_t
98 rxq_cq_to_ol_flags(struct rxq *rxq, volatile struct mlx5_cqe *cqe)
99 __attribute__((always_inline));
104 * Verify or set magic value in CQE.
113 check_cqe_seen(volatile struct mlx5_cqe *cqe)
115 static const uint8_t magic[] = "seen";
116 volatile uint8_t (*buf)[sizeof(cqe->rsvd0)] = &cqe->rsvd0;
120 for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
121 if (!ret || (*buf)[i] != magic[i]) {
123 (*buf)[i] = magic[i];
131 * Check whether CQE is valid.
136 * Size of completion queue.
141 * 0 on success, 1 on failure.
144 check_cqe(volatile struct mlx5_cqe *cqe,
145 unsigned int cqes_n, const uint16_t ci)
147 uint16_t idx = ci & cqes_n;
148 uint8_t op_own = cqe->op_own;
149 uint8_t op_owner = MLX5_CQE_OWNER(op_own);
150 uint8_t op_code = MLX5_CQE_OPCODE(op_own);
152 if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
153 return 1; /* No CQE. */
155 if ((op_code == MLX5_CQE_RESP_ERR) ||
156 (op_code == MLX5_CQE_REQ_ERR)) {
157 volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
158 uint8_t syndrome = err_cqe->syndrome;
160 if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
161 (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
163 if (!check_cqe_seen(cqe))
164 ERROR("unexpected CQE error %u (0x%02x)"
166 op_code, op_code, syndrome);
168 } else if ((op_code != MLX5_CQE_RESP_SEND) &&
169 (op_code != MLX5_CQE_REQ)) {
170 if (!check_cqe_seen(cqe))
171 ERROR("unexpected CQE opcode %u (0x%02x)",
180 * Return the address of the WQE.
183 * Pointer to TX queue structure.
185 * WQE consumer index.
190 static inline uintptr_t *
191 tx_mlx5_wqe(struct txq *txq, uint16_t ci)
193 ci &= ((1 << txq->wqe_n) - 1);
194 return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
198 * Manage TX completions.
200 * When sending a burst, mlx5_tx_burst() posts several WRs.
203 * Pointer to TX queue structure.
206 txq_complete(struct txq *txq)
208 const unsigned int elts_n = 1 << txq->elts_n;
209 const unsigned int cqe_n = 1 << txq->cqe_n;
210 const unsigned int cqe_cnt = cqe_n - 1;
211 uint16_t elts_free = txq->elts_tail;
213 uint16_t cq_ci = txq->cq_ci;
214 volatile struct mlx5_cqe *cqe = NULL;
215 volatile struct mlx5_wqe_ctrl *ctrl;
218 volatile struct mlx5_cqe *tmp;
220 tmp = &(*txq->cqes)[cq_ci & cqe_cnt];
221 if (check_cqe(tmp, cqe_n, cq_ci))
225 if (MLX5_CQE_FORMAT(cqe->op_own) == MLX5_COMPRESSED) {
226 if (!check_cqe_seen(cqe))
227 ERROR("unexpected compressed CQE, TX stopped");
230 if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
231 (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
232 if (!check_cqe_seen(cqe))
233 ERROR("unexpected error CQE, TX stopped");
239 if (unlikely(cqe == NULL))
241 txq->wqe_pi = ntohs(cqe->wqe_counter);
242 ctrl = (volatile struct mlx5_wqe_ctrl *)
243 tx_mlx5_wqe(txq, txq->wqe_pi);
244 elts_tail = ctrl->ctrl3;
245 assert(elts_tail < (1 << txq->wqe_n));
247 while (elts_free != elts_tail) {
248 struct rte_mbuf *elt = (*txq->elts)[elts_free];
249 unsigned int elts_free_next =
250 (elts_free + 1) & (elts_n - 1);
251 struct rte_mbuf *elt_next = (*txq->elts)[elts_free_next];
255 memset(&(*txq->elts)[elts_free],
257 sizeof((*txq->elts)[elts_free]));
259 RTE_MBUF_PREFETCH_TO_FREE(elt_next);
260 /* Only one segment needs to be freed. */
261 rte_pktmbuf_free_seg(elt);
262 elts_free = elts_free_next;
265 txq->elts_tail = elts_tail;
266 /* Update the consumer index. */
268 *txq->cq_db = htonl(cq_ci);
272 * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
273 * the cloned mbuf is allocated is returned instead.
279 * Memory pool where data is located for given mbuf.
281 static struct rte_mempool *
282 txq_mb2mp(struct rte_mbuf *buf)
284 if (unlikely(RTE_MBUF_INDIRECT(buf)))
285 return rte_mbuf_from_indirect(buf)->pool;
290 * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
291 * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
292 * remove an entry first.
295 * Pointer to TX queue structure.
297 * Memory Pool for which a Memory Region lkey must be returned.
300 * mr->lkey on success, (uint32_t)-1 on failure.
302 static inline uint32_t
303 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
306 uint32_t lkey = (uint32_t)-1;
308 for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
309 if (unlikely(txq->mp2mr[i].mp == NULL)) {
310 /* Unknown MP, add a new MR for it. */
313 if (txq->mp2mr[i].mp == mp) {
314 assert(txq->mp2mr[i].lkey != (uint32_t)-1);
315 assert(htonl(txq->mp2mr[i].mr->lkey) ==
317 lkey = txq->mp2mr[i].lkey;
321 if (unlikely(lkey == (uint32_t)-1))
322 lkey = txq_mp2mr_reg(txq, mp, i);
327 * Ring TX queue doorbell.
330 * Pointer to TX queue structure.
332 * Pointer to the last WQE posted in the NIC.
335 mlx5_tx_dbrec(struct txq *txq, volatile struct mlx5_wqe *wqe)
337 uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
338 volatile uint64_t *src = ((volatile uint64_t *)wqe);
341 *txq->qp_db = htonl(txq->wqe_ci);
342 /* Ensure ordering between DB record and BF copy. */
348 * DPDK callback to check the status of a tx descriptor.
353 * The index of the descriptor in the ring.
356 * The status of the tx descriptor.
359 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
361 struct txq *txq = tx_queue;
362 const unsigned int elts_n = 1 << txq->elts_n;
363 const unsigned int elts_cnt = elts_n - 1;
367 used = (txq->elts_head - txq->elts_tail) & elts_cnt;
369 return RTE_ETH_TX_DESC_FULL;
370 return RTE_ETH_TX_DESC_DONE;
374 * DPDK callback to check the status of a rx descriptor.
379 * The index of the descriptor in the ring.
382 * The status of the tx descriptor.
385 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
387 struct rxq *rxq = rx_queue;
388 struct rxq_zip *zip = &rxq->zip;
389 volatile struct mlx5_cqe *cqe;
390 const unsigned int cqe_n = (1 << rxq->cqe_n);
391 const unsigned int cqe_cnt = cqe_n - 1;
395 /* if we are processing a compressed cqe */
397 used = zip->cqe_cnt - zip->ca;
403 cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
404 while (check_cqe(cqe, cqe_n, cq_ci) == 0) {
408 op_own = cqe->op_own;
409 if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
410 n = ntohl(cqe->byte_cnt);
415 cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
417 used = RTE_MIN(used, (1U << rxq->elts_n) - 1);
419 return RTE_ETH_RX_DESC_DONE;
420 return RTE_ETH_RX_DESC_AVAIL;
424 * DPDK callback for TX.
427 * Generic pointer to TX queue structure.
429 * Packets to transmit.
431 * Number of packets in array.
434 * Number of packets successfully transmitted (<= pkts_n).
437 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
439 struct txq *txq = (struct txq *)dpdk_txq;
440 uint16_t elts_head = txq->elts_head;
441 const unsigned int elts_n = 1 << txq->elts_n;
448 volatile struct mlx5_wqe_v *wqe = NULL;
449 unsigned int segs_n = 0;
450 struct rte_mbuf *buf = NULL;
453 if (unlikely(!pkts_n))
455 /* Prefetch first packet cacheline. */
456 rte_prefetch0(*pkts);
457 /* Start processing. */
459 max = (elts_n - (elts_head - txq->elts_tail));
462 max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
463 if (unlikely(!max_wqe))
466 volatile rte_v128u32_t *dseg = NULL;
471 uint16_t pkt_inline_sz = MLX5_WQE_DWORD_SIZE + 2;
472 uint16_t tso_header_sz = 0;
474 uint8_t cs_flags = 0;
476 #ifdef MLX5_PMD_SOFT_COUNTERS
477 uint32_t total_length = 0;
482 segs_n = buf->nb_segs;
484 * Make sure there is enough room to store this packet and
485 * that one ring entry remains unused.
488 if (max < segs_n + 1)
494 if (unlikely(--max_wqe == 0))
496 wqe = (volatile struct mlx5_wqe_v *)
497 tx_mlx5_wqe(txq, txq->wqe_ci);
498 rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
500 rte_prefetch0(*pkts);
501 addr = rte_pktmbuf_mtod(buf, uintptr_t);
502 length = DATA_LEN(buf);
503 ehdr = (((uint8_t *)addr)[1] << 8) |
504 ((uint8_t *)addr)[0];
505 #ifdef MLX5_PMD_SOFT_COUNTERS
506 total_length = length;
508 assert(length >= MLX5_WQE_DWORD_SIZE);
509 /* Update element. */
510 (*txq->elts)[elts_head] = buf;
511 elts_head = (elts_head + 1) & (elts_n - 1);
512 /* Prefetch next buffer data. */
514 volatile void *pkt_addr;
516 pkt_addr = rte_pktmbuf_mtod(*pkts, volatile void *);
517 rte_prefetch0(pkt_addr);
519 /* Should we enable HW CKSUM offload */
521 (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
522 const uint64_t is_tunneled = buf->ol_flags &
524 PKT_TX_TUNNEL_VXLAN);
526 if (is_tunneled && txq->tunnel_en) {
527 cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
528 MLX5_ETH_WQE_L4_INNER_CSUM;
529 if (buf->ol_flags & PKT_TX_OUTER_IP_CKSUM)
530 cs_flags |= MLX5_ETH_WQE_L3_CSUM;
532 cs_flags = MLX5_ETH_WQE_L3_CSUM |
533 MLX5_ETH_WQE_L4_CSUM;
536 raw = ((uint8_t *)(uintptr_t)wqe) + 2 * MLX5_WQE_DWORD_SIZE;
537 /* Replace the Ethernet type by the VLAN if necessary. */
538 if (buf->ol_flags & PKT_TX_VLAN_PKT) {
539 uint32_t vlan = htonl(0x81000000 | buf->vlan_tci);
540 unsigned int len = 2 * ETHER_ADDR_LEN - 2;
544 /* Copy Destination and source mac address. */
545 memcpy((uint8_t *)raw, ((uint8_t *)addr), len);
547 memcpy((uint8_t *)raw + len, &vlan, sizeof(vlan));
548 /* Copy missing two bytes to end the DSeg. */
549 memcpy((uint8_t *)raw + len + sizeof(vlan),
550 ((uint8_t *)addr) + len, 2);
554 memcpy((uint8_t *)raw, ((uint8_t *)addr) + 2,
555 MLX5_WQE_DWORD_SIZE);
556 length -= pkt_inline_sz;
557 addr += pkt_inline_sz;
560 tso = buf->ol_flags & PKT_TX_TCP_SEG;
562 uintptr_t end = (uintptr_t)
563 (((uintptr_t)txq->wqes) +
567 uint8_t vlan_sz = (buf->ol_flags &
568 PKT_TX_VLAN_PKT) ? 4 : 0;
569 const uint64_t is_tunneled =
572 PKT_TX_TUNNEL_VXLAN);
574 tso_header_sz = buf->l2_len + vlan_sz +
575 buf->l3_len + buf->l4_len;
577 if (is_tunneled && txq->tunnel_en) {
578 tso_header_sz += buf->outer_l2_len +
580 cs_flags |= MLX5_ETH_WQE_L4_INNER_CSUM;
582 cs_flags |= MLX5_ETH_WQE_L4_CSUM;
584 if (unlikely(tso_header_sz >
585 MLX5_MAX_TSO_HEADER))
587 copy_b = tso_header_sz - pkt_inline_sz;
588 /* First seg must contain all headers. */
589 assert(copy_b <= length);
590 raw += MLX5_WQE_DWORD_SIZE;
592 ((end - (uintptr_t)raw) > copy_b)) {
593 uint16_t n = (MLX5_WQE_DS(copy_b) -
596 if (unlikely(max_wqe < n))
599 rte_memcpy((void *)raw,
600 (void *)addr, copy_b);
603 pkt_inline_sz += copy_b;
605 * Another DWORD will be added
606 * in the inline part.
608 raw += MLX5_WQE_DS(copy_b) *
609 MLX5_WQE_DWORD_SIZE -
613 wqe->ctrl = (rte_v128u32_t){
614 htonl(txq->wqe_ci << 8),
615 htonl(txq->qp_num_8s | 1),
623 elts_head = (elts_head - 1) &
630 /* Inline if enough room. */
631 if (txq->inline_en || tso) {
632 uintptr_t end = (uintptr_t)
633 (((uintptr_t)txq->wqes) +
634 (1 << txq->wqe_n) * MLX5_WQE_SIZE);
635 unsigned int max_inline = txq->max_inline *
636 RTE_CACHE_LINE_SIZE -
638 uintptr_t addr_end = (addr + max_inline) &
639 ~(RTE_CACHE_LINE_SIZE - 1);
640 unsigned int copy_b = (addr_end > addr) ?
641 RTE_MIN((addr_end - addr), length) :
644 raw += MLX5_WQE_DWORD_SIZE;
645 if (copy_b && ((end - (uintptr_t)raw) > copy_b)) {
647 * One Dseg remains in the current WQE. To
648 * keep the computation positive, it is
649 * removed after the bytes to Dseg conversion.
651 uint16_t n = (MLX5_WQE_DS(copy_b) - 1 + 3) / 4;
653 if (unlikely(max_wqe < n))
658 htonl(copy_b | MLX5_INLINE_SEG);
661 MLX5_WQE_DS(tso_header_sz) *
663 rte_memcpy((void *)raw,
664 (void *)&inl, sizeof(inl));
666 pkt_inline_sz += sizeof(inl);
668 rte_memcpy((void *)raw, (void *)addr, copy_b);
671 pkt_inline_sz += copy_b;
674 * 2 DWORDs consumed by the WQE header + ETH segment +
675 * the size of the inline part of the packet.
677 ds = 2 + MLX5_WQE_DS(pkt_inline_sz - 2);
679 if (ds % (MLX5_WQE_SIZE /
680 MLX5_WQE_DWORD_SIZE) == 0) {
681 if (unlikely(--max_wqe == 0))
683 dseg = (volatile rte_v128u32_t *)
684 tx_mlx5_wqe(txq, txq->wqe_ci +
687 dseg = (volatile rte_v128u32_t *)
689 (ds * MLX5_WQE_DWORD_SIZE));
692 } else if (!segs_n) {
695 /* dseg will be advance as part of next_seg */
696 dseg = (volatile rte_v128u32_t *)
698 ((ds - 1) * MLX5_WQE_DWORD_SIZE));
703 * No inline has been done in the packet, only the
704 * Ethernet Header as been stored.
706 dseg = (volatile rte_v128u32_t *)
707 ((uintptr_t)wqe + (3 * MLX5_WQE_DWORD_SIZE));
710 /* Add the remaining packet as a simple ds. */
711 naddr = htonll(addr);
712 *dseg = (rte_v128u32_t){
714 txq_mp2mr(txq, txq_mb2mp(buf)),
727 * Spill on next WQE when the current one does not have
728 * enough room left. Size of WQE must a be a multiple
729 * of data segment size.
731 assert(!(MLX5_WQE_SIZE % MLX5_WQE_DWORD_SIZE));
732 if (!(ds % (MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE))) {
733 if (unlikely(--max_wqe == 0))
735 dseg = (volatile rte_v128u32_t *)
736 tx_mlx5_wqe(txq, txq->wqe_ci + ds / 4);
737 rte_prefetch0(tx_mlx5_wqe(txq,
738 txq->wqe_ci + ds / 4 + 1));
745 length = DATA_LEN(buf);
746 #ifdef MLX5_PMD_SOFT_COUNTERS
747 total_length += length;
749 /* Store segment information. */
750 naddr = htonll(rte_pktmbuf_mtod(buf, uintptr_t));
751 *dseg = (rte_v128u32_t){
753 txq_mp2mr(txq, txq_mb2mp(buf)),
757 (*txq->elts)[elts_head] = buf;
758 elts_head = (elts_head + 1) & (elts_n - 1);
767 /* Initialize known and common part of the WQE structure. */
769 wqe->ctrl = (rte_v128u32_t){
770 htonl((txq->wqe_ci << 8) | MLX5_OPCODE_TSO),
771 htonl(txq->qp_num_8s | ds),
775 wqe->eseg = (rte_v128u32_t){
777 cs_flags | (htons(buf->tso_segsz) << 16),
779 (ehdr << 16) | htons(tso_header_sz),
782 wqe->ctrl = (rte_v128u32_t){
783 htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND),
784 htonl(txq->qp_num_8s | ds),
788 wqe->eseg = (rte_v128u32_t){
792 (ehdr << 16) | htons(pkt_inline_sz),
796 txq->wqe_ci += (ds + 3) / 4;
797 #ifdef MLX5_PMD_SOFT_COUNTERS
798 /* Increment sent bytes counter. */
799 txq->stats.obytes += total_length;
802 /* Take a shortcut if nothing must be sent. */
803 if (unlikely((i + k) == 0))
805 /* Check whether completion threshold has been reached. */
806 comp = txq->elts_comp + i + j + k;
807 if (comp >= MLX5_TX_COMP_THRESH) {
808 volatile struct mlx5_wqe_ctrl *w =
809 (volatile struct mlx5_wqe_ctrl *)wqe;
811 /* Request completion on last WQE. */
813 /* Save elts_head in unused "immediate" field of WQE. */
814 w->ctrl3 = elts_head;
817 txq->elts_comp = comp;
819 #ifdef MLX5_PMD_SOFT_COUNTERS
820 /* Increment sent packets counter. */
821 txq->stats.opackets += i;
823 /* Ring QP doorbell. */
824 mlx5_tx_dbrec(txq, (volatile struct mlx5_wqe *)wqe);
825 txq->elts_head = elts_head;
830 * Open a MPW session.
833 * Pointer to TX queue structure.
835 * Pointer to MPW session structure.
840 mlx5_mpw_new(struct txq *txq, struct mlx5_mpw *mpw, uint32_t length)
842 uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
843 volatile struct mlx5_wqe_data_seg (*dseg)[MLX5_MPW_DSEG_MAX] =
844 (volatile struct mlx5_wqe_data_seg (*)[])
845 tx_mlx5_wqe(txq, idx + 1);
847 mpw->state = MLX5_MPW_STATE_OPENED;
851 mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
852 mpw->wqe->eseg.mss = htons(length);
853 mpw->wqe->eseg.inline_hdr_sz = 0;
854 mpw->wqe->eseg.rsvd0 = 0;
855 mpw->wqe->eseg.rsvd1 = 0;
856 mpw->wqe->eseg.rsvd2 = 0;
857 mpw->wqe->ctrl[0] = htonl((MLX5_OPC_MOD_MPW << 24) |
858 (txq->wqe_ci << 8) | MLX5_OPCODE_TSO);
859 mpw->wqe->ctrl[2] = 0;
860 mpw->wqe->ctrl[3] = 0;
861 mpw->data.dseg[0] = (volatile struct mlx5_wqe_data_seg *)
862 (((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
863 mpw->data.dseg[1] = (volatile struct mlx5_wqe_data_seg *)
864 (((uintptr_t)mpw->wqe) + (3 * MLX5_WQE_DWORD_SIZE));
865 mpw->data.dseg[2] = &(*dseg)[0];
866 mpw->data.dseg[3] = &(*dseg)[1];
867 mpw->data.dseg[4] = &(*dseg)[2];
871 * Close a MPW session.
874 * Pointer to TX queue structure.
876 * Pointer to MPW session structure.
879 mlx5_mpw_close(struct txq *txq, struct mlx5_mpw *mpw)
881 unsigned int num = mpw->pkts_n;
884 * Store size in multiple of 16 bytes. Control and Ethernet segments
887 mpw->wqe->ctrl[1] = htonl(txq->qp_num_8s | (2 + num));
888 mpw->state = MLX5_MPW_STATE_CLOSED;
893 rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
894 rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
898 * DPDK callback for TX with MPW support.
901 * Generic pointer to TX queue structure.
903 * Packets to transmit.
905 * Number of packets in array.
908 * Number of packets successfully transmitted (<= pkts_n).
911 mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
913 struct txq *txq = (struct txq *)dpdk_txq;
914 uint16_t elts_head = txq->elts_head;
915 const unsigned int elts_n = 1 << txq->elts_n;
921 struct mlx5_mpw mpw = {
922 .state = MLX5_MPW_STATE_CLOSED,
925 if (unlikely(!pkts_n))
927 /* Prefetch first packet cacheline. */
928 rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
929 rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
930 /* Start processing. */
932 max = (elts_n - (elts_head - txq->elts_tail));
935 max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
936 if (unlikely(!max_wqe))
939 struct rte_mbuf *buf = *(pkts++);
940 unsigned int elts_head_next;
942 unsigned int segs_n = buf->nb_segs;
943 uint32_t cs_flags = 0;
946 * Make sure there is enough room to store this packet and
947 * that one ring entry remains unused.
950 if (max < segs_n + 1)
952 /* Do not bother with large packets MPW cannot handle. */
953 if (segs_n > MLX5_MPW_DSEG_MAX)
957 /* Should we enable HW CKSUM offload */
959 (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
960 cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
961 /* Retrieve packet information. */
962 length = PKT_LEN(buf);
964 /* Start new session if packet differs. */
965 if ((mpw.state == MLX5_MPW_STATE_OPENED) &&
966 ((mpw.len != length) ||
968 (mpw.wqe->eseg.cs_flags != cs_flags)))
969 mlx5_mpw_close(txq, &mpw);
970 if (mpw.state == MLX5_MPW_STATE_CLOSED) {
972 * Multi-Packet WQE consumes at most two WQE.
973 * mlx5_mpw_new() expects to be able to use such
976 if (unlikely(max_wqe < 2))
979 mlx5_mpw_new(txq, &mpw, length);
980 mpw.wqe->eseg.cs_flags = cs_flags;
982 /* Multi-segment packets must be alone in their MPW. */
983 assert((segs_n == 1) || (mpw.pkts_n == 0));
984 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
988 volatile struct mlx5_wqe_data_seg *dseg;
991 elts_head_next = (elts_head + 1) & (elts_n - 1);
993 (*txq->elts)[elts_head] = buf;
994 dseg = mpw.data.dseg[mpw.pkts_n];
995 addr = rte_pktmbuf_mtod(buf, uintptr_t);
996 *dseg = (struct mlx5_wqe_data_seg){
997 .byte_count = htonl(DATA_LEN(buf)),
998 .lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
999 .addr = htonll(addr),
1001 elts_head = elts_head_next;
1002 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1003 length += DATA_LEN(buf);
1009 assert(length == mpw.len);
1010 if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
1011 mlx5_mpw_close(txq, &mpw);
1012 elts_head = elts_head_next;
1013 #ifdef MLX5_PMD_SOFT_COUNTERS
1014 /* Increment sent bytes counter. */
1015 txq->stats.obytes += length;
1019 /* Take a shortcut if nothing must be sent. */
1020 if (unlikely(i == 0))
1022 /* Check whether completion threshold has been reached. */
1023 /* "j" includes both packets and segments. */
1024 comp = txq->elts_comp + j;
1025 if (comp >= MLX5_TX_COMP_THRESH) {
1026 volatile struct mlx5_wqe *wqe = mpw.wqe;
1028 /* Request completion on last WQE. */
1029 wqe->ctrl[2] = htonl(8);
1030 /* Save elts_head in unused "immediate" field of WQE. */
1031 wqe->ctrl[3] = elts_head;
1034 txq->elts_comp = comp;
1036 #ifdef MLX5_PMD_SOFT_COUNTERS
1037 /* Increment sent packets counter. */
1038 txq->stats.opackets += i;
1040 /* Ring QP doorbell. */
1041 if (mpw.state == MLX5_MPW_STATE_OPENED)
1042 mlx5_mpw_close(txq, &mpw);
1043 mlx5_tx_dbrec(txq, mpw.wqe);
1044 txq->elts_head = elts_head;
1049 * Open a MPW inline session.
1052 * Pointer to TX queue structure.
1054 * Pointer to MPW session structure.
1059 mlx5_mpw_inline_new(struct txq *txq, struct mlx5_mpw *mpw, uint32_t length)
1061 uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
1062 struct mlx5_wqe_inl_small *inl;
1064 mpw->state = MLX5_MPW_INL_STATE_OPENED;
1068 mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
1069 mpw->wqe->ctrl[0] = htonl((MLX5_OPC_MOD_MPW << 24) |
1070 (txq->wqe_ci << 8) |
1072 mpw->wqe->ctrl[2] = 0;
1073 mpw->wqe->ctrl[3] = 0;
1074 mpw->wqe->eseg.mss = htons(length);
1075 mpw->wqe->eseg.inline_hdr_sz = 0;
1076 mpw->wqe->eseg.cs_flags = 0;
1077 mpw->wqe->eseg.rsvd0 = 0;
1078 mpw->wqe->eseg.rsvd1 = 0;
1079 mpw->wqe->eseg.rsvd2 = 0;
1080 inl = (struct mlx5_wqe_inl_small *)
1081 (((uintptr_t)mpw->wqe) + 2 * MLX5_WQE_DWORD_SIZE);
1082 mpw->data.raw = (uint8_t *)&inl->raw;
1086 * Close a MPW inline session.
1089 * Pointer to TX queue structure.
1091 * Pointer to MPW session structure.
1094 mlx5_mpw_inline_close(struct txq *txq, struct mlx5_mpw *mpw)
1097 struct mlx5_wqe_inl_small *inl = (struct mlx5_wqe_inl_small *)
1098 (((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
1100 size = MLX5_WQE_SIZE - MLX5_MWQE64_INL_DATA + mpw->total_len;
1102 * Store size in multiple of 16 bytes. Control and Ethernet segments
1105 mpw->wqe->ctrl[1] = htonl(txq->qp_num_8s | MLX5_WQE_DS(size));
1106 mpw->state = MLX5_MPW_STATE_CLOSED;
1107 inl->byte_cnt = htonl(mpw->total_len | MLX5_INLINE_SEG);
1108 txq->wqe_ci += (size + (MLX5_WQE_SIZE - 1)) / MLX5_WQE_SIZE;
1112 * DPDK callback for TX with MPW inline support.
1115 * Generic pointer to TX queue structure.
1117 * Packets to transmit.
1119 * Number of packets in array.
1122 * Number of packets successfully transmitted (<= pkts_n).
1125 mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
1128 struct txq *txq = (struct txq *)dpdk_txq;
1129 uint16_t elts_head = txq->elts_head;
1130 const unsigned int elts_n = 1 << txq->elts_n;
1136 unsigned int inline_room = txq->max_inline * RTE_CACHE_LINE_SIZE;
1137 struct mlx5_mpw mpw = {
1138 .state = MLX5_MPW_STATE_CLOSED,
1141 * Compute the maximum number of WQE which can be consumed by inline
1144 * - 1 control segment,
1145 * - 1 Ethernet segment,
1146 * - N Dseg from the inline request.
1148 const unsigned int wqe_inl_n =
1149 ((2 * MLX5_WQE_DWORD_SIZE +
1150 txq->max_inline * RTE_CACHE_LINE_SIZE) +
1151 RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
1153 if (unlikely(!pkts_n))
1155 /* Prefetch first packet cacheline. */
1156 rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
1157 rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
1158 /* Start processing. */
1160 max = (elts_n - (elts_head - txq->elts_tail));
1164 struct rte_mbuf *buf = *(pkts++);
1165 unsigned int elts_head_next;
1168 unsigned int segs_n = buf->nb_segs;
1169 uint32_t cs_flags = 0;
1172 * Make sure there is enough room to store this packet and
1173 * that one ring entry remains unused.
1176 if (max < segs_n + 1)
1178 /* Do not bother with large packets MPW cannot handle. */
1179 if (segs_n > MLX5_MPW_DSEG_MAX)
1184 * Compute max_wqe in case less WQE were consumed in previous
1187 max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
1188 /* Should we enable HW CKSUM offload */
1190 (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
1191 cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
1192 /* Retrieve packet information. */
1193 length = PKT_LEN(buf);
1194 /* Start new session if packet differs. */
1195 if (mpw.state == MLX5_MPW_STATE_OPENED) {
1196 if ((mpw.len != length) ||
1198 (mpw.wqe->eseg.cs_flags != cs_flags))
1199 mlx5_mpw_close(txq, &mpw);
1200 } else if (mpw.state == MLX5_MPW_INL_STATE_OPENED) {
1201 if ((mpw.len != length) ||
1203 (length > inline_room) ||
1204 (mpw.wqe->eseg.cs_flags != cs_flags)) {
1205 mlx5_mpw_inline_close(txq, &mpw);
1207 txq->max_inline * RTE_CACHE_LINE_SIZE;
1210 if (mpw.state == MLX5_MPW_STATE_CLOSED) {
1211 if ((segs_n != 1) ||
1212 (length > inline_room)) {
1214 * Multi-Packet WQE consumes at most two WQE.
1215 * mlx5_mpw_new() expects to be able to use
1218 if (unlikely(max_wqe < 2))
1221 mlx5_mpw_new(txq, &mpw, length);
1222 mpw.wqe->eseg.cs_flags = cs_flags;
1224 if (unlikely(max_wqe < wqe_inl_n))
1226 max_wqe -= wqe_inl_n;
1227 mlx5_mpw_inline_new(txq, &mpw, length);
1228 mpw.wqe->eseg.cs_flags = cs_flags;
1231 /* Multi-segment packets must be alone in their MPW. */
1232 assert((segs_n == 1) || (mpw.pkts_n == 0));
1233 if (mpw.state == MLX5_MPW_STATE_OPENED) {
1234 assert(inline_room ==
1235 txq->max_inline * RTE_CACHE_LINE_SIZE);
1236 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1240 volatile struct mlx5_wqe_data_seg *dseg;
1243 (elts_head + 1) & (elts_n - 1);
1245 (*txq->elts)[elts_head] = buf;
1246 dseg = mpw.data.dseg[mpw.pkts_n];
1247 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1248 *dseg = (struct mlx5_wqe_data_seg){
1249 .byte_count = htonl(DATA_LEN(buf)),
1250 .lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
1251 .addr = htonll(addr),
1253 elts_head = elts_head_next;
1254 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1255 length += DATA_LEN(buf);
1261 assert(length == mpw.len);
1262 if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
1263 mlx5_mpw_close(txq, &mpw);
1267 assert(mpw.state == MLX5_MPW_INL_STATE_OPENED);
1268 assert(length <= inline_room);
1269 assert(length == DATA_LEN(buf));
1270 elts_head_next = (elts_head + 1) & (elts_n - 1);
1271 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1272 (*txq->elts)[elts_head] = buf;
1273 /* Maximum number of bytes before wrapping. */
1274 max = ((((uintptr_t)(txq->wqes)) +
1277 (uintptr_t)mpw.data.raw);
1279 rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1282 mpw.data.raw = (volatile void *)txq->wqes;
1283 rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1284 (void *)(addr + max),
1286 mpw.data.raw += length - max;
1288 rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1294 (volatile void *)txq->wqes;
1296 mpw.data.raw += length;
1299 mpw.total_len += length;
1301 if (mpw.pkts_n == MLX5_MPW_DSEG_MAX) {
1302 mlx5_mpw_inline_close(txq, &mpw);
1304 txq->max_inline * RTE_CACHE_LINE_SIZE;
1306 inline_room -= length;
1309 elts_head = elts_head_next;
1310 #ifdef MLX5_PMD_SOFT_COUNTERS
1311 /* Increment sent bytes counter. */
1312 txq->stats.obytes += length;
1316 /* Take a shortcut if nothing must be sent. */
1317 if (unlikely(i == 0))
1319 /* Check whether completion threshold has been reached. */
1320 /* "j" includes both packets and segments. */
1321 comp = txq->elts_comp + j;
1322 if (comp >= MLX5_TX_COMP_THRESH) {
1323 volatile struct mlx5_wqe *wqe = mpw.wqe;
1325 /* Request completion on last WQE. */
1326 wqe->ctrl[2] = htonl(8);
1327 /* Save elts_head in unused "immediate" field of WQE. */
1328 wqe->ctrl[3] = elts_head;
1331 txq->elts_comp = comp;
1333 #ifdef MLX5_PMD_SOFT_COUNTERS
1334 /* Increment sent packets counter. */
1335 txq->stats.opackets += i;
1337 /* Ring QP doorbell. */
1338 if (mpw.state == MLX5_MPW_INL_STATE_OPENED)
1339 mlx5_mpw_inline_close(txq, &mpw);
1340 else if (mpw.state == MLX5_MPW_STATE_OPENED)
1341 mlx5_mpw_close(txq, &mpw);
1342 mlx5_tx_dbrec(txq, mpw.wqe);
1343 txq->elts_head = elts_head;
1348 * Translate RX completion flags to packet type.
1353 * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
1356 * Packet type for struct rte_mbuf.
1358 static inline uint32_t
1359 rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe)
1362 uint16_t flags = ntohs(cqe->hdr_type_etc);
1364 if (cqe->pkt_info & MLX5_CQE_RX_TUNNEL_PACKET) {
1367 MLX5_CQE_RX_IPV4_PACKET,
1368 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN) |
1370 MLX5_CQE_RX_IPV6_PACKET,
1371 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN);
1372 pkt_type |= ((cqe->pkt_info & MLX5_CQE_RX_OUTER_PACKET) ?
1373 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN :
1374 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN);
1378 MLX5_CQE_L3_HDR_TYPE_IPV6,
1379 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN) |
1381 MLX5_CQE_L3_HDR_TYPE_IPV4,
1382 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN);
1388 * Get size of the next packet for a given CQE. For compressed CQEs, the
1389 * consumer index is updated only once all packets of the current one have
1393 * Pointer to RX queue.
1396 * @param[out] rss_hash
1397 * Packet RSS Hash result.
1400 * Packet size in bytes (0 if there is none), -1 in case of completion
1404 mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe,
1405 uint16_t cqe_cnt, uint32_t *rss_hash)
1407 struct rxq_zip *zip = &rxq->zip;
1408 uint16_t cqe_n = cqe_cnt + 1;
1412 /* Process compressed data in the CQE and mini arrays. */
1414 volatile struct mlx5_mini_cqe8 (*mc)[8] =
1415 (volatile struct mlx5_mini_cqe8 (*)[8])
1416 (uintptr_t)(&(*rxq->cqes)[zip->ca & cqe_cnt]);
1418 len = ntohl((*mc)[zip->ai & 7].byte_cnt);
1419 *rss_hash = ntohl((*mc)[zip->ai & 7].rx_hash_result);
1420 if ((++zip->ai & 7) == 0) {
1421 /* Invalidate consumed CQEs */
1424 while (idx != end) {
1425 (*rxq->cqes)[idx & cqe_cnt].op_own =
1426 MLX5_CQE_INVALIDATE;
1430 * Increment consumer index to skip the number of
1431 * CQEs consumed. Hardware leaves holes in the CQ
1432 * ring for software use.
1437 if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1438 /* Invalidate the rest */
1442 while (idx != end) {
1443 (*rxq->cqes)[idx & cqe_cnt].op_own =
1444 MLX5_CQE_INVALIDATE;
1447 rxq->cq_ci = zip->cq_ci;
1450 /* No compressed data, get next CQE and verify if it is compressed. */
1455 ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1456 if (unlikely(ret == 1))
1459 op_own = cqe->op_own;
1460 if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1461 volatile struct mlx5_mini_cqe8 (*mc)[8] =
1462 (volatile struct mlx5_mini_cqe8 (*)[8])
1463 (uintptr_t)(&(*rxq->cqes)[rxq->cq_ci &
1466 /* Fix endianness. */
1467 zip->cqe_cnt = ntohl(cqe->byte_cnt);
1469 * Current mini array position is the one returned by
1472 * If completion comprises several mini arrays, as a
1473 * special case the second one is located 7 CQEs after
1474 * the initial CQE instead of 8 for subsequent ones.
1476 zip->ca = rxq->cq_ci;
1477 zip->na = zip->ca + 7;
1478 /* Compute the next non compressed CQE. */
1480 zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1481 /* Get packet size to return. */
1482 len = ntohl((*mc)[0].byte_cnt);
1483 *rss_hash = ntohl((*mc)[0].rx_hash_result);
1485 /* Prefetch all the entries to be invalidated */
1488 while (idx != end) {
1489 rte_prefetch0(&(*rxq->cqes)[(idx) & cqe_cnt]);
1493 len = ntohl(cqe->byte_cnt);
1494 *rss_hash = ntohl(cqe->rx_hash_res);
1496 /* Error while receiving packet. */
1497 if (unlikely(MLX5_CQE_OPCODE(op_own) == MLX5_CQE_RESP_ERR))
1504 * Translate RX completion flags to offload flags.
1507 * Pointer to RX queue structure.
1512 * Offload flags (ol_flags) for struct rte_mbuf.
1514 static inline uint32_t
1515 rxq_cq_to_ol_flags(struct rxq *rxq, volatile struct mlx5_cqe *cqe)
1517 uint32_t ol_flags = 0;
1518 uint16_t flags = ntohs(cqe->hdr_type_etc);
1522 MLX5_CQE_RX_L3_HDR_VALID,
1523 PKT_RX_IP_CKSUM_GOOD) |
1525 MLX5_CQE_RX_L4_HDR_VALID,
1526 PKT_RX_L4_CKSUM_GOOD);
1527 if ((cqe->pkt_info & MLX5_CQE_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
1530 MLX5_CQE_RX_L3_HDR_VALID,
1531 PKT_RX_IP_CKSUM_GOOD) |
1533 MLX5_CQE_RX_L4_HDR_VALID,
1534 PKT_RX_L4_CKSUM_GOOD);
1539 * DPDK callback for RX.
1542 * Generic pointer to RX queue structure.
1544 * Array to store received packets.
1546 * Maximum number of packets in array.
1549 * Number of packets successfully received (<= pkts_n).
1552 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1554 struct rxq *rxq = dpdk_rxq;
1555 const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1556 const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1557 const unsigned int sges_n = rxq->sges_n;
1558 struct rte_mbuf *pkt = NULL;
1559 struct rte_mbuf *seg = NULL;
1560 volatile struct mlx5_cqe *cqe =
1561 &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1563 unsigned int rq_ci = rxq->rq_ci << sges_n;
1564 int len; /* keep its value across iterations. */
1567 unsigned int idx = rq_ci & wqe_cnt;
1568 volatile struct mlx5_wqe_data_seg *wqe = &(*rxq->wqes)[idx];
1569 struct rte_mbuf *rep = (*rxq->elts)[idx];
1570 uint32_t rss_hash_res = 0;
1578 rep = rte_mbuf_raw_alloc(rxq->mp);
1579 if (unlikely(rep == NULL)) {
1580 ++rxq->stats.rx_nombuf;
1583 * no buffers before we even started,
1584 * bail out silently.
1588 while (pkt != seg) {
1589 assert(pkt != (*rxq->elts)[idx]);
1591 rte_mbuf_refcnt_set(pkt, 0);
1592 __rte_mbuf_raw_free(pkt);
1598 cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1599 len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt,
1602 rte_mbuf_refcnt_set(rep, 0);
1603 __rte_mbuf_raw_free(rep);
1606 if (unlikely(len == -1)) {
1607 /* RX error, packet is likely too large. */
1608 rte_mbuf_refcnt_set(rep, 0);
1609 __rte_mbuf_raw_free(rep);
1610 ++rxq->stats.idropped;
1614 assert(len >= (rxq->crc_present << 2));
1615 /* Update packet information. */
1616 pkt->packet_type = 0;
1618 if (rss_hash_res && rxq->rss_hash) {
1619 pkt->hash.rss = rss_hash_res;
1620 pkt->ol_flags = PKT_RX_RSS_HASH;
1622 if (rxq->mark && (cqe->sop_drop_qpn !=
1623 htonl(MLX5_FLOW_MARK_INVALID))) {
1624 pkt->ol_flags |= PKT_RX_FDIR;
1625 if (cqe->sop_drop_qpn !=
1626 htonl(MLX5_FLOW_MARK_DEFAULT)) {
1627 uint32_t mark = cqe->sop_drop_qpn;
1629 pkt->ol_flags |= PKT_RX_FDIR_ID;
1631 mlx5_flow_mark_get(mark);
1634 if (rxq->csum | rxq->csum_l2tun | rxq->vlan_strip |
1638 rxq_cq_to_pkt_type(cqe);
1640 rxq_cq_to_ol_flags(rxq, cqe);
1642 if (ntohs(cqe->hdr_type_etc) &
1643 MLX5_CQE_VLAN_STRIPPED) {
1644 pkt->ol_flags |= PKT_RX_VLAN_PKT |
1645 PKT_RX_VLAN_STRIPPED;
1646 pkt->vlan_tci = ntohs(cqe->vlan_info);
1648 if (rxq->crc_present)
1649 len -= ETHER_CRC_LEN;
1653 DATA_LEN(rep) = DATA_LEN(seg);
1654 PKT_LEN(rep) = PKT_LEN(seg);
1655 SET_DATA_OFF(rep, DATA_OFF(seg));
1656 NB_SEGS(rep) = NB_SEGS(seg);
1657 PORT(rep) = PORT(seg);
1659 (*rxq->elts)[idx] = rep;
1661 * Fill NIC descriptor with the new buffer. The lkey and size
1662 * of the buffers are already known, only the buffer address
1665 wqe->addr = htonll(rte_pktmbuf_mtod(rep, uintptr_t));
1666 if (len > DATA_LEN(seg)) {
1667 len -= DATA_LEN(seg);
1672 DATA_LEN(seg) = len;
1673 #ifdef MLX5_PMD_SOFT_COUNTERS
1674 /* Increment bytes counter. */
1675 rxq->stats.ibytes += PKT_LEN(pkt);
1677 /* Return packet. */
1683 /* Align consumer index to the next stride. */
1688 if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1690 /* Update the consumer index. */
1691 rxq->rq_ci = rq_ci >> sges_n;
1693 *rxq->cq_db = htonl(rxq->cq_ci);
1695 *rxq->rq_db = htonl(rxq->rq_ci);
1696 #ifdef MLX5_PMD_SOFT_COUNTERS
1697 /* Increment packets counter. */
1698 rxq->stats.ipackets += i;
1704 * Dummy DPDK callback for TX.
1706 * This function is used to temporarily replace the real callback during
1707 * unsafe control operations on the queue, or in case of error.
1710 * Generic pointer to TX queue structure.
1712 * Packets to transmit.
1714 * Number of packets in array.
1717 * Number of packets successfully transmitted (<= pkts_n).
1720 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1729 * Dummy DPDK callback for RX.
1731 * This function is used to temporarily replace the real callback during
1732 * unsafe control operations on the queue, or in case of error.
1735 * Generic pointer to RX queue structure.
1737 * Array to store received packets.
1739 * Maximum number of packets in array.
1742 * Number of packets successfully received (<= pkts_n).
1745 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)