xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.c (revision b218a1bf36d764c54ebf422b0799d5ac4f79310a)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright 2015 6WIND S.A.
5  *   Copyright 2015 Mellanox.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
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
16  *       distribution.
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.
20  *
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.
32  */
33 
34 #include <assert.h>
35 #include <stdint.h>
36 #include <string.h>
37 #include <stdlib.h>
38 
39 /* Verbs header. */
40 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
41 #ifdef PEDANTIC
42 #pragma GCC diagnostic ignored "-pedantic"
43 #endif
44 #include <infiniband/verbs.h>
45 #ifdef PEDANTIC
46 #pragma GCC diagnostic error "-pedantic"
47 #endif
48 
49 /* DPDK headers don't like -pedantic. */
50 #ifdef PEDANTIC
51 #pragma GCC diagnostic ignored "-pedantic"
52 #endif
53 #include <rte_mbuf.h>
54 #include <rte_mempool.h>
55 #include <rte_prefetch.h>
56 #include <rte_common.h>
57 #include <rte_branch_prediction.h>
58 #include <rte_memory.h>
59 #ifdef PEDANTIC
60 #pragma GCC diagnostic error "-pedantic"
61 #endif
62 
63 #include "mlx5.h"
64 #include "mlx5_utils.h"
65 #include "mlx5_rxtx.h"
66 #include "mlx5_autoconf.h"
67 #include "mlx5_defs.h"
68 
69 /**
70  * Manage TX completions.
71  *
72  * When sending a burst, mlx5_tx_burst() posts several WRs.
73  * To improve performance, a completion event is only required once every
74  * MLX5_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
75  * for other WRs, but this information would not be used anyway.
76  *
77  * @param txq
78  *   Pointer to TX queue structure.
79  *
80  * @return
81  *   0 on success, -1 on failure.
82  */
83 static int
84 txq_complete(struct txq *txq)
85 {
86 	unsigned int elts_comp = txq->elts_comp;
87 	unsigned int elts_tail = txq->elts_tail;
88 	unsigned int elts_free = txq->elts_tail;
89 	const unsigned int elts_n = txq->elts_n;
90 	int wcs_n;
91 
92 	if (unlikely(elts_comp == 0))
93 		return 0;
94 #ifdef DEBUG_SEND
95 	DEBUG("%p: processing %u work requests completions",
96 	      (void *)txq, elts_comp);
97 #endif
98 	wcs_n = txq->poll_cnt(txq->cq, elts_comp);
99 	if (unlikely(wcs_n == 0))
100 		return 0;
101 	if (unlikely(wcs_n < 0)) {
102 		DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
103 		      (void *)txq, wcs_n);
104 		return -1;
105 	}
106 	elts_comp -= wcs_n;
107 	assert(elts_comp <= txq->elts_comp);
108 	/*
109 	 * Assume WC status is successful as nothing can be done about it
110 	 * anyway.
111 	 */
112 	elts_tail += wcs_n * txq->elts_comp_cd_init;
113 	if (elts_tail >= elts_n)
114 		elts_tail -= elts_n;
115 
116 	while (elts_free != elts_tail) {
117 		struct txq_elt *elt = &(*txq->elts)[elts_free];
118 		unsigned int elts_free_next =
119 			(((elts_free + 1) == elts_n) ? 0 : elts_free + 1);
120 		struct rte_mbuf *tmp = elt->buf;
121 		struct txq_elt *elt_next = &(*txq->elts)[elts_free_next];
122 
123 #ifndef NDEBUG
124 		/* Poisoning. */
125 		memset(elt, 0x66, sizeof(*elt));
126 #endif
127 		RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
128 		/* Faster than rte_pktmbuf_free(). */
129 		do {
130 			struct rte_mbuf *next = NEXT(tmp);
131 
132 			rte_pktmbuf_free_seg(tmp);
133 			tmp = next;
134 		} while (tmp != NULL);
135 		elts_free = elts_free_next;
136 	}
137 
138 	txq->elts_tail = elts_tail;
139 	txq->elts_comp = elts_comp;
140 	return 0;
141 }
142 
143 struct mlx5_check_mempool_data {
144 	int ret;
145 	char *start;
146 	char *end;
147 };
148 
149 /* Called by mlx5_check_mempool() when iterating the memory chunks. */
150 static void mlx5_check_mempool_cb(struct rte_mempool *mp,
151 	void *opaque, struct rte_mempool_memhdr *memhdr,
152 	unsigned mem_idx)
153 {
154 	struct mlx5_check_mempool_data *data = opaque;
155 
156 	(void)mp;
157 	(void)mem_idx;
158 
159 	/* It already failed, skip the next chunks. */
160 	if (data->ret != 0)
161 		return;
162 	/* It is the first chunk. */
163 	if (data->start == NULL && data->end == NULL) {
164 		data->start = memhdr->addr;
165 		data->end = data->start + memhdr->len;
166 		return;
167 	}
168 	if (data->end == memhdr->addr) {
169 		data->end += memhdr->len;
170 		return;
171 	}
172 	if (data->start == (char *)memhdr->addr + memhdr->len) {
173 		data->start -= memhdr->len;
174 		return;
175 	}
176 	/* Error, mempool is not virtually contigous. */
177 	data->ret = -1;
178 }
179 
180 /**
181  * Check if a mempool can be used: it must be virtually contiguous.
182  *
183  * @param[in] mp
184  *   Pointer to memory pool.
185  * @param[out] start
186  *   Pointer to the start address of the mempool virtual memory area
187  * @param[out] end
188  *   Pointer to the end address of the mempool virtual memory area
189  *
190  * @return
191  *   0 on success (mempool is virtually contiguous), -1 on error.
192  */
193 static int mlx5_check_mempool(struct rte_mempool *mp, uintptr_t *start,
194 	uintptr_t *end)
195 {
196 	struct mlx5_check_mempool_data data;
197 
198 	memset(&data, 0, sizeof(data));
199 	rte_mempool_mem_iter(mp, mlx5_check_mempool_cb, &data);
200 	*start = (uintptr_t)data.start;
201 	*end = (uintptr_t)data.end;
202 
203 	return data.ret;
204 }
205 
206 /* For best performance, this function should not be inlined. */
207 struct ibv_mr *mlx5_mp2mr(struct ibv_pd *, struct rte_mempool *)
208 	__attribute__((noinline));
209 
210 /**
211  * Register mempool as a memory region.
212  *
213  * @param pd
214  *   Pointer to protection domain.
215  * @param mp
216  *   Pointer to memory pool.
217  *
218  * @return
219  *   Memory region pointer, NULL in case of error.
220  */
221 struct ibv_mr *
222 mlx5_mp2mr(struct ibv_pd *pd, struct rte_mempool *mp)
223 {
224 	const struct rte_memseg *ms = rte_eal_get_physmem_layout();
225 	uintptr_t start;
226 	uintptr_t end;
227 	unsigned int i;
228 
229 	if (mlx5_check_mempool(mp, &start, &end) != 0) {
230 		ERROR("mempool %p: not virtually contiguous",
231 			(void *)mp);
232 		return NULL;
233 	}
234 
235 	DEBUG("mempool %p area start=%p end=%p size=%zu",
236 	      (void *)mp, (void *)start, (void *)end,
237 	      (size_t)(end - start));
238 	/* Round start and end to page boundary if found in memory segments. */
239 	for (i = 0; (i < RTE_MAX_MEMSEG) && (ms[i].addr != NULL); ++i) {
240 		uintptr_t addr = (uintptr_t)ms[i].addr;
241 		size_t len = ms[i].len;
242 		unsigned int align = ms[i].hugepage_sz;
243 
244 		if ((start > addr) && (start < addr + len))
245 			start = RTE_ALIGN_FLOOR(start, align);
246 		if ((end > addr) && (end < addr + len))
247 			end = RTE_ALIGN_CEIL(end, align);
248 	}
249 	DEBUG("mempool %p using start=%p end=%p size=%zu for MR",
250 	      (void *)mp, (void *)start, (void *)end,
251 	      (size_t)(end - start));
252 	return ibv_reg_mr(pd,
253 			  (void *)start,
254 			  end - start,
255 			  IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE);
256 }
257 
258 /**
259  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
260  * the cloned mbuf is allocated is returned instead.
261  *
262  * @param buf
263  *   Pointer to mbuf.
264  *
265  * @return
266  *   Memory pool where data is located for given mbuf.
267  */
268 static struct rte_mempool *
269 txq_mb2mp(struct rte_mbuf *buf)
270 {
271 	if (unlikely(RTE_MBUF_INDIRECT(buf)))
272 		return rte_mbuf_from_indirect(buf)->pool;
273 	return buf->pool;
274 }
275 
276 /**
277  * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
278  * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
279  * remove an entry first.
280  *
281  * @param txq
282  *   Pointer to TX queue structure.
283  * @param[in] mp
284  *   Memory Pool for which a Memory Region lkey must be returned.
285  *
286  * @return
287  *   mr->lkey on success, (uint32_t)-1 on failure.
288  */
289 static uint32_t
290 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
291 {
292 	unsigned int i;
293 	struct ibv_mr *mr;
294 
295 	for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
296 		if (unlikely(txq->mp2mr[i].mp == NULL)) {
297 			/* Unknown MP, add a new MR for it. */
298 			break;
299 		}
300 		if (txq->mp2mr[i].mp == mp) {
301 			assert(txq->mp2mr[i].lkey != (uint32_t)-1);
302 			assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
303 			return txq->mp2mr[i].lkey;
304 		}
305 	}
306 	/* Add a new entry, register MR first. */
307 	DEBUG("%p: discovered new memory pool \"%s\" (%p)",
308 	      (void *)txq, mp->name, (void *)mp);
309 	mr = mlx5_mp2mr(txq->priv->pd, mp);
310 	if (unlikely(mr == NULL)) {
311 		DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
312 		      (void *)txq);
313 		return (uint32_t)-1;
314 	}
315 	if (unlikely(i == RTE_DIM(txq->mp2mr))) {
316 		/* Table is full, remove oldest entry. */
317 		DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
318 		      (void *)txq);
319 		--i;
320 		claim_zero(ibv_dereg_mr(txq->mp2mr[0].mr));
321 		memmove(&txq->mp2mr[0], &txq->mp2mr[1],
322 			(sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
323 	}
324 	/* Store the new entry. */
325 	txq->mp2mr[i].mp = mp;
326 	txq->mp2mr[i].mr = mr;
327 	txq->mp2mr[i].lkey = mr->lkey;
328 	DEBUG("%p: new MR lkey for MP \"%s\" (%p): 0x%08" PRIu32,
329 	      (void *)txq, mp->name, (void *)mp, txq->mp2mr[i].lkey);
330 	return txq->mp2mr[i].lkey;
331 }
332 
333 struct txq_mp2mr_mbuf_check_data {
334 	int ret;
335 };
336 
337 /**
338  * Callback function for rte_mempool_obj_iter() to check whether a given
339  * mempool object looks like a mbuf.
340  *
341  * @param[in] mp
342  *   The mempool pointer
343  * @param[in] arg
344  *   Context data (struct txq_mp2mr_mbuf_check_data). Contains the
345  *   return value.
346  * @param[in] obj
347  *   Object address.
348  * @param index
349  *   Object index, unused.
350  */
351 static void
352 txq_mp2mr_mbuf_check(struct rte_mempool *mp, void *arg, void *obj,
353 	uint32_t index __rte_unused)
354 {
355 	struct txq_mp2mr_mbuf_check_data *data = arg;
356 	struct rte_mbuf *buf = obj;
357 
358 	/* Check whether mbuf structure fits element size and whether mempool
359 	 * pointer is valid. */
360 	if (sizeof(*buf) > mp->elt_size || buf->pool != mp)
361 		data->ret = -1;
362 }
363 
364 /**
365  * Iterator function for rte_mempool_walk() to register existing mempools and
366  * fill the MP to MR cache of a TX queue.
367  *
368  * @param[in] mp
369  *   Memory Pool to register.
370  * @param *arg
371  *   Pointer to TX queue structure.
372  */
373 void
374 txq_mp2mr_iter(struct rte_mempool *mp, void *arg)
375 {
376 	struct txq *txq = arg;
377 	struct txq_mp2mr_mbuf_check_data data = {
378 		.ret = 0,
379 	};
380 
381 	/* Register mempool only if the first element looks like a mbuf. */
382 	if (rte_mempool_obj_iter(mp, txq_mp2mr_mbuf_check, &data) == 0 ||
383 			data.ret == -1)
384 		return;
385 	txq_mp2mr(txq, mp);
386 }
387 
388 /**
389  * Insert VLAN using mbuf headroom space.
390  *
391  * @param buf
392  *   Buffer for VLAN insertion.
393  *
394  * @return
395  *   0 on success, errno value on failure.
396  */
397 static inline int
398 insert_vlan_sw(struct rte_mbuf *buf)
399 {
400 	uintptr_t addr;
401 	uint32_t vlan;
402 	uint16_t head_room_len = rte_pktmbuf_headroom(buf);
403 
404 	if (head_room_len < 4)
405 		return EINVAL;
406 
407 	addr = rte_pktmbuf_mtod(buf, uintptr_t);
408 	vlan = htonl(0x81000000 | buf->vlan_tci);
409 	memmove((void *)(addr - 4), (void *)addr, 12);
410 	memcpy((void *)(addr + 8), &vlan, sizeof(vlan));
411 
412 	SET_DATA_OFF(buf, head_room_len - 4);
413 	DATA_LEN(buf) += 4;
414 
415 	return 0;
416 }
417 
418 #if MLX5_PMD_SGE_WR_N > 1
419 
420 /**
421  * Copy scattered mbuf contents to a single linear buffer.
422  *
423  * @param[out] linear
424  *   Linear output buffer.
425  * @param[in] buf
426  *   Scattered input buffer.
427  *
428  * @return
429  *   Number of bytes copied to the output buffer or 0 if not large enough.
430  */
431 static unsigned int
432 linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
433 {
434 	unsigned int size = 0;
435 	unsigned int offset;
436 
437 	do {
438 		unsigned int len = DATA_LEN(buf);
439 
440 		offset = size;
441 		size += len;
442 		if (unlikely(size > sizeof(*linear)))
443 			return 0;
444 		memcpy(&(*linear)[offset],
445 		       rte_pktmbuf_mtod(buf, uint8_t *),
446 		       len);
447 		buf = NEXT(buf);
448 	} while (buf != NULL);
449 	return size;
450 }
451 
452 /**
453  * Handle scattered buffers for mlx5_tx_burst().
454  *
455  * @param txq
456  *   TX queue structure.
457  * @param segs
458  *   Number of segments in buf.
459  * @param elt
460  *   TX queue element to fill.
461  * @param[in] buf
462  *   Buffer to process.
463  * @param elts_head
464  *   Index of the linear buffer to use if necessary (normally txq->elts_head).
465  * @param[out] sges
466  *   Array filled with SGEs on success.
467  *
468  * @return
469  *   A structure containing the processed packet size in bytes and the
470  *   number of SGEs. Both fields are set to (unsigned int)-1 in case of
471  *   failure.
472  */
473 static struct tx_burst_sg_ret {
474 	unsigned int length;
475 	unsigned int num;
476 }
477 tx_burst_sg(struct txq *txq, unsigned int segs, struct txq_elt *elt,
478 	    struct rte_mbuf *buf, unsigned int elts_head,
479 	    struct ibv_sge (*sges)[MLX5_PMD_SGE_WR_N])
480 {
481 	unsigned int sent_size = 0;
482 	unsigned int j;
483 	int linearize = 0;
484 
485 	/* When there are too many segments, extra segments are
486 	 * linearized in the last SGE. */
487 	if (unlikely(segs > RTE_DIM(*sges))) {
488 		segs = (RTE_DIM(*sges) - 1);
489 		linearize = 1;
490 	}
491 	/* Update element. */
492 	elt->buf = buf;
493 	/* Register segments as SGEs. */
494 	for (j = 0; (j != segs); ++j) {
495 		struct ibv_sge *sge = &(*sges)[j];
496 		uint32_t lkey;
497 
498 		/* Retrieve Memory Region key for this memory pool. */
499 		lkey = txq_mp2mr(txq, txq_mb2mp(buf));
500 		if (unlikely(lkey == (uint32_t)-1)) {
501 			/* MR does not exist. */
502 			DEBUG("%p: unable to get MP <-> MR association",
503 			      (void *)txq);
504 			/* Clean up TX element. */
505 			elt->buf = NULL;
506 			goto stop;
507 		}
508 		/* Update SGE. */
509 		sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
510 		if (txq->priv->vf)
511 			rte_prefetch0((volatile void *)
512 				      (uintptr_t)sge->addr);
513 		sge->length = DATA_LEN(buf);
514 		sge->lkey = lkey;
515 		sent_size += sge->length;
516 		buf = NEXT(buf);
517 	}
518 	/* If buf is not NULL here and is not going to be linearized,
519 	 * nb_segs is not valid. */
520 	assert(j == segs);
521 	assert((buf == NULL) || (linearize));
522 	/* Linearize extra segments. */
523 	if (linearize) {
524 		struct ibv_sge *sge = &(*sges)[segs];
525 		linear_t *linear = &(*txq->elts_linear)[elts_head];
526 		unsigned int size = linearize_mbuf(linear, buf);
527 
528 		assert(segs == (RTE_DIM(*sges) - 1));
529 		if (size == 0) {
530 			/* Invalid packet. */
531 			DEBUG("%p: packet too large to be linearized.",
532 			      (void *)txq);
533 			/* Clean up TX element. */
534 			elt->buf = NULL;
535 			goto stop;
536 		}
537 		/* If MLX5_PMD_SGE_WR_N is 1, free mbuf immediately. */
538 		if (RTE_DIM(*sges) == 1) {
539 			do {
540 				struct rte_mbuf *next = NEXT(buf);
541 
542 				rte_pktmbuf_free_seg(buf);
543 				buf = next;
544 			} while (buf != NULL);
545 			elt->buf = NULL;
546 		}
547 		/* Update SGE. */
548 		sge->addr = (uintptr_t)&(*linear)[0];
549 		sge->length = size;
550 		sge->lkey = txq->mr_linear->lkey;
551 		sent_size += size;
552 		/* Include last segment. */
553 		segs++;
554 	}
555 	return (struct tx_burst_sg_ret){
556 		.length = sent_size,
557 		.num = segs,
558 	};
559 stop:
560 	return (struct tx_burst_sg_ret){
561 		.length = -1,
562 		.num = -1,
563 	};
564 }
565 
566 #endif /* MLX5_PMD_SGE_WR_N > 1 */
567 
568 /**
569  * DPDK callback for TX.
570  *
571  * @param dpdk_txq
572  *   Generic pointer to TX queue structure.
573  * @param[in] pkts
574  *   Packets to transmit.
575  * @param pkts_n
576  *   Number of packets in array.
577  *
578  * @return
579  *   Number of packets successfully transmitted (<= pkts_n).
580  */
581 uint16_t
582 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
583 {
584 	struct txq *txq = (struct txq *)dpdk_txq;
585 	unsigned int elts_head = txq->elts_head;
586 	const unsigned int elts_n = txq->elts_n;
587 	unsigned int elts_comp_cd = txq->elts_comp_cd;
588 	unsigned int elts_comp = 0;
589 	unsigned int i;
590 	unsigned int max;
591 	int err;
592 	struct rte_mbuf *buf = pkts[0];
593 
594 	assert(elts_comp_cd != 0);
595 	/* Prefetch first packet cacheline. */
596 	rte_prefetch0(buf);
597 	txq_complete(txq);
598 	max = (elts_n - (elts_head - txq->elts_tail));
599 	if (max > elts_n)
600 		max -= elts_n;
601 	assert(max >= 1);
602 	assert(max <= elts_n);
603 	/* Always leave one free entry in the ring. */
604 	--max;
605 	if (max == 0)
606 		return 0;
607 	if (max > pkts_n)
608 		max = pkts_n;
609 	for (i = 0; (i != max); ++i) {
610 		struct rte_mbuf *buf_next = pkts[i + 1];
611 		unsigned int elts_head_next =
612 			(((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
613 		struct txq_elt *elt = &(*txq->elts)[elts_head];
614 		unsigned int segs = NB_SEGS(buf);
615 #ifdef MLX5_PMD_SOFT_COUNTERS
616 		unsigned int sent_size = 0;
617 #endif
618 		uint32_t send_flags = 0;
619 #ifdef HAVE_VERBS_VLAN_INSERTION
620 		int insert_vlan = 0;
621 #endif /* HAVE_VERBS_VLAN_INSERTION */
622 
623 		if (i + 1 < max)
624 			rte_prefetch0(buf_next);
625 		/* Request TX completion. */
626 		if (unlikely(--elts_comp_cd == 0)) {
627 			elts_comp_cd = txq->elts_comp_cd_init;
628 			++elts_comp;
629 			send_flags |= IBV_EXP_QP_BURST_SIGNALED;
630 		}
631 		/* Should we enable HW CKSUM offload */
632 		if (buf->ol_flags &
633 		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
634 			send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
635 			/* HW does not support checksum offloads at arbitrary
636 			 * offsets but automatically recognizes the packet
637 			 * type. For inner L3/L4 checksums, only VXLAN (UDP)
638 			 * tunnels are currently supported. */
639 			if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
640 				send_flags |= IBV_EXP_QP_BURST_TUNNEL;
641 		}
642 		if (buf->ol_flags & PKT_TX_VLAN_PKT) {
643 #ifdef HAVE_VERBS_VLAN_INSERTION
644 			if (!txq->priv->mps)
645 				insert_vlan = 1;
646 			else
647 #endif /* HAVE_VERBS_VLAN_INSERTION */
648 			{
649 				err = insert_vlan_sw(buf);
650 				if (unlikely(err))
651 					goto stop;
652 			}
653 		}
654 		if (likely(segs == 1)) {
655 			uintptr_t addr;
656 			uint32_t length;
657 			uint32_t lkey;
658 			uintptr_t buf_next_addr;
659 
660 			/* Retrieve buffer information. */
661 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
662 			length = DATA_LEN(buf);
663 			/* Update element. */
664 			elt->buf = buf;
665 			if (txq->priv->vf)
666 				rte_prefetch0((volatile void *)
667 					      (uintptr_t)addr);
668 			/* Prefetch next buffer data. */
669 			if (i + 1 < max) {
670 				buf_next_addr =
671 					rte_pktmbuf_mtod(buf_next, uintptr_t);
672 				rte_prefetch0((volatile void *)
673 					      (uintptr_t)buf_next_addr);
674 			}
675 			/* Put packet into send queue. */
676 #if MLX5_PMD_MAX_INLINE > 0
677 			if (length <= txq->max_inline) {
678 #ifdef HAVE_VERBS_VLAN_INSERTION
679 				if (insert_vlan)
680 					err = txq->send_pending_inline_vlan
681 						(txq->qp,
682 						 (void *)addr,
683 						 length,
684 						 send_flags,
685 						 &buf->vlan_tci);
686 				else
687 #endif /* HAVE_VERBS_VLAN_INSERTION */
688 					err = txq->send_pending_inline
689 						(txq->qp,
690 						 (void *)addr,
691 						 length,
692 						 send_flags);
693 			} else
694 #endif
695 			{
696 				/* Retrieve Memory Region key for this
697 				 * memory pool. */
698 				lkey = txq_mp2mr(txq, txq_mb2mp(buf));
699 				if (unlikely(lkey == (uint32_t)-1)) {
700 					/* MR does not exist. */
701 					DEBUG("%p: unable to get MP <-> MR"
702 					      " association", (void *)txq);
703 					/* Clean up TX element. */
704 					elt->buf = NULL;
705 					goto stop;
706 				}
707 #ifdef HAVE_VERBS_VLAN_INSERTION
708 				if (insert_vlan)
709 					err = txq->send_pending_vlan
710 						(txq->qp,
711 						 addr,
712 						 length,
713 						 lkey,
714 						 send_flags,
715 						 &buf->vlan_tci);
716 				else
717 #endif /* HAVE_VERBS_VLAN_INSERTION */
718 					err = txq->send_pending
719 						(txq->qp,
720 						 addr,
721 						 length,
722 						 lkey,
723 						 send_flags);
724 			}
725 			if (unlikely(err))
726 				goto stop;
727 #ifdef MLX5_PMD_SOFT_COUNTERS
728 			sent_size += length;
729 #endif
730 		} else {
731 #if MLX5_PMD_SGE_WR_N > 1
732 			struct ibv_sge sges[MLX5_PMD_SGE_WR_N];
733 			struct tx_burst_sg_ret ret;
734 
735 			ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
736 					  &sges);
737 			if (ret.length == (unsigned int)-1)
738 				goto stop;
739 			/* Put SG list into send queue. */
740 #ifdef HAVE_VERBS_VLAN_INSERTION
741 			if (insert_vlan)
742 				err = txq->send_pending_sg_list_vlan
743 					(txq->qp,
744 					 sges,
745 					 ret.num,
746 					 send_flags,
747 					 &buf->vlan_tci);
748 			else
749 #endif /* HAVE_VERBS_VLAN_INSERTION */
750 				err = txq->send_pending_sg_list
751 					(txq->qp,
752 					 sges,
753 					 ret.num,
754 					 send_flags);
755 			if (unlikely(err))
756 				goto stop;
757 #ifdef MLX5_PMD_SOFT_COUNTERS
758 			sent_size += ret.length;
759 #endif
760 #else /* MLX5_PMD_SGE_WR_N > 1 */
761 			DEBUG("%p: TX scattered buffers support not"
762 			      " compiled in", (void *)txq);
763 			goto stop;
764 #endif /* MLX5_PMD_SGE_WR_N > 1 */
765 		}
766 		elts_head = elts_head_next;
767 		buf = buf_next;
768 #ifdef MLX5_PMD_SOFT_COUNTERS
769 		/* Increment sent bytes counter. */
770 		txq->stats.obytes += sent_size;
771 #endif
772 	}
773 stop:
774 	/* Take a shortcut if nothing must be sent. */
775 	if (unlikely(i == 0))
776 		return 0;
777 #ifdef MLX5_PMD_SOFT_COUNTERS
778 	/* Increment sent packets counter. */
779 	txq->stats.opackets += i;
780 #endif
781 	/* Ring QP doorbell. */
782 	err = txq->send_flush(txq->qp);
783 	if (unlikely(err)) {
784 		/* A nonzero value is not supposed to be returned.
785 		 * Nothing can be done about it. */
786 		DEBUG("%p: send_flush() failed with error %d",
787 		      (void *)txq, err);
788 	}
789 	txq->elts_head = elts_head;
790 	txq->elts_comp += elts_comp;
791 	txq->elts_comp_cd = elts_comp_cd;
792 	return i;
793 }
794 
795 /**
796  * Translate RX completion flags to packet type.
797  *
798  * @param flags
799  *   RX completion flags returned by poll_length_flags().
800  *
801  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
802  *
803  * @return
804  *   Packet type for struct rte_mbuf.
805  */
806 static inline uint32_t
807 rxq_cq_to_pkt_type(uint32_t flags)
808 {
809 	uint32_t pkt_type;
810 
811 	if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
812 		pkt_type =
813 			TRANSPOSE(flags,
814 				  IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
815 				  RTE_PTYPE_L3_IPV4) |
816 			TRANSPOSE(flags,
817 				  IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
818 				  RTE_PTYPE_L3_IPV6) |
819 			TRANSPOSE(flags,
820 				  IBV_EXP_CQ_RX_IPV4_PACKET,
821 				  RTE_PTYPE_INNER_L3_IPV4) |
822 			TRANSPOSE(flags,
823 				  IBV_EXP_CQ_RX_IPV6_PACKET,
824 				  RTE_PTYPE_INNER_L3_IPV6);
825 	else
826 		pkt_type =
827 			TRANSPOSE(flags,
828 				  IBV_EXP_CQ_RX_IPV4_PACKET,
829 				  RTE_PTYPE_L3_IPV4) |
830 			TRANSPOSE(flags,
831 				  IBV_EXP_CQ_RX_IPV6_PACKET,
832 				  RTE_PTYPE_L3_IPV6);
833 	return pkt_type;
834 }
835 
836 /**
837  * Translate RX completion flags to offload flags.
838  *
839  * @param[in] rxq
840  *   Pointer to RX queue structure.
841  * @param flags
842  *   RX completion flags returned by poll_length_flags().
843  *
844  * @return
845  *   Offload flags (ol_flags) for struct rte_mbuf.
846  */
847 static inline uint32_t
848 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
849 {
850 	uint32_t ol_flags = 0;
851 
852 	if (rxq->csum) {
853 		/* Set IP checksum flag only for IPv4/IPv6 packets. */
854 		if (flags &
855 		    (IBV_EXP_CQ_RX_IPV4_PACKET | IBV_EXP_CQ_RX_IPV6_PACKET))
856 			ol_flags |=
857 				TRANSPOSE(~flags,
858 					IBV_EXP_CQ_RX_IP_CSUM_OK,
859 					PKT_RX_IP_CKSUM_BAD);
860 #ifdef HAVE_EXP_CQ_RX_TCP_PACKET
861 		/* Set L4 checksum flag only for TCP/UDP packets. */
862 		if (flags &
863 		    (IBV_EXP_CQ_RX_TCP_PACKET | IBV_EXP_CQ_RX_UDP_PACKET))
864 #endif /* HAVE_EXP_CQ_RX_TCP_PACKET */
865 			ol_flags |=
866 				TRANSPOSE(~flags,
867 					IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
868 					PKT_RX_L4_CKSUM_BAD);
869 	}
870 	/*
871 	 * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
872 	 * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
873 	 * (its value is 0).
874 	 */
875 	if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
876 		ol_flags |=
877 			TRANSPOSE(~flags,
878 				  IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
879 				  PKT_RX_IP_CKSUM_BAD) |
880 			TRANSPOSE(~flags,
881 				  IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
882 				  PKT_RX_L4_CKSUM_BAD);
883 	return ol_flags;
884 }
885 
886 /**
887  * DPDK callback for RX with scattered packets support.
888  *
889  * @param dpdk_rxq
890  *   Generic pointer to RX queue structure.
891  * @param[out] pkts
892  *   Array to store received packets.
893  * @param pkts_n
894  *   Maximum number of packets in array.
895  *
896  * @return
897  *   Number of packets successfully received (<= pkts_n).
898  */
899 uint16_t
900 mlx5_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
901 {
902 	struct rxq *rxq = (struct rxq *)dpdk_rxq;
903 	struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
904 	const unsigned int elts_n = rxq->elts_n;
905 	unsigned int elts_head = rxq->elts_head;
906 	unsigned int i;
907 	unsigned int pkts_ret = 0;
908 	int ret;
909 
910 	if (unlikely(!rxq->sp))
911 		return mlx5_rx_burst(dpdk_rxq, pkts, pkts_n);
912 	if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
913 		return 0;
914 	for (i = 0; (i != pkts_n); ++i) {
915 		struct rxq_elt_sp *elt = &(*elts)[elts_head];
916 		unsigned int len;
917 		unsigned int pkt_buf_len;
918 		struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
919 		struct rte_mbuf **pkt_buf_next = &pkt_buf;
920 		unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
921 		unsigned int j = 0;
922 		uint32_t flags;
923 		uint16_t vlan_tci;
924 
925 		/* Sanity checks. */
926 		assert(elts_head < rxq->elts_n);
927 		assert(rxq->elts_head < rxq->elts_n);
928 		ret = rxq->poll(rxq->cq, NULL, NULL, &flags, &vlan_tci);
929 		if (unlikely(ret < 0)) {
930 			struct ibv_wc wc;
931 			int wcs_n;
932 
933 			DEBUG("rxq=%p, poll_length() failed (ret=%d)",
934 			      (void *)rxq, ret);
935 			/* ibv_poll_cq() must be used in case of failure. */
936 			wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
937 			if (unlikely(wcs_n == 0))
938 				break;
939 			if (unlikely(wcs_n < 0)) {
940 				DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
941 				      (void *)rxq, wcs_n);
942 				break;
943 			}
944 			assert(wcs_n == 1);
945 			if (unlikely(wc.status != IBV_WC_SUCCESS)) {
946 				/* Whatever, just repost the offending WR. */
947 				DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
948 				      " completion status (%d): %s",
949 				      (void *)rxq, wc.wr_id, wc.status,
950 				      ibv_wc_status_str(wc.status));
951 #ifdef MLX5_PMD_SOFT_COUNTERS
952 				/* Increment dropped packets counter. */
953 				++rxq->stats.idropped;
954 #endif
955 				goto repost;
956 			}
957 			ret = wc.byte_len;
958 		}
959 		if (ret == 0)
960 			break;
961 		assert(ret >= (rxq->crc_present << 2));
962 		len = ret - (rxq->crc_present << 2);
963 		pkt_buf_len = len;
964 		/*
965 		 * Replace spent segments with new ones, concatenate and
966 		 * return them as pkt_buf.
967 		 */
968 		while (1) {
969 			struct ibv_sge *sge = &elt->sges[j];
970 			struct rte_mbuf *seg = elt->bufs[j];
971 			struct rte_mbuf *rep;
972 			unsigned int seg_tailroom;
973 
974 			assert(seg != NULL);
975 			/*
976 			 * Fetch initial bytes of packet descriptor into a
977 			 * cacheline while allocating rep.
978 			 */
979 			rte_prefetch0(seg);
980 			rep = rte_mbuf_raw_alloc(rxq->mp);
981 			if (unlikely(rep == NULL)) {
982 				/*
983 				 * Unable to allocate a replacement mbuf,
984 				 * repost WR.
985 				 */
986 				DEBUG("rxq=%p: can't allocate a new mbuf",
987 				      (void *)rxq);
988 				if (pkt_buf != NULL) {
989 					*pkt_buf_next = NULL;
990 					rte_pktmbuf_free(pkt_buf);
991 				}
992 				/* Increment out of memory counters. */
993 				++rxq->stats.rx_nombuf;
994 				++rxq->priv->dev->data->rx_mbuf_alloc_failed;
995 				goto repost;
996 			}
997 #ifndef NDEBUG
998 			/* Poison user-modifiable fields in rep. */
999 			NEXT(rep) = (void *)((uintptr_t)-1);
1000 			SET_DATA_OFF(rep, 0xdead);
1001 			DATA_LEN(rep) = 0xd00d;
1002 			PKT_LEN(rep) = 0xdeadd00d;
1003 			NB_SEGS(rep) = 0x2a;
1004 			PORT(rep) = 0x2a;
1005 			rep->ol_flags = -1;
1006 #endif
1007 			assert(rep->buf_len == seg->buf_len);
1008 			assert(rep->buf_len == rxq->mb_len);
1009 			/* Reconfigure sge to use rep instead of seg. */
1010 			assert(sge->lkey == rxq->mr->lkey);
1011 			sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
1012 			elt->bufs[j] = rep;
1013 			++j;
1014 			/* Update pkt_buf if it's the first segment, or link
1015 			 * seg to the previous one and update pkt_buf_next. */
1016 			*pkt_buf_next = seg;
1017 			pkt_buf_next = &NEXT(seg);
1018 			/* Update seg information. */
1019 			seg_tailroom = (seg->buf_len - seg_headroom);
1020 			assert(sge->length == seg_tailroom);
1021 			SET_DATA_OFF(seg, seg_headroom);
1022 			if (likely(len <= seg_tailroom)) {
1023 				/* Last segment. */
1024 				DATA_LEN(seg) = len;
1025 				PKT_LEN(seg) = len;
1026 				/* Sanity check. */
1027 				assert(rte_pktmbuf_headroom(seg) ==
1028 				       seg_headroom);
1029 				assert(rte_pktmbuf_tailroom(seg) ==
1030 				       (seg_tailroom - len));
1031 				break;
1032 			}
1033 			DATA_LEN(seg) = seg_tailroom;
1034 			PKT_LEN(seg) = seg_tailroom;
1035 			/* Sanity check. */
1036 			assert(rte_pktmbuf_headroom(seg) == seg_headroom);
1037 			assert(rte_pktmbuf_tailroom(seg) == 0);
1038 			/* Fix len and clear headroom for next segments. */
1039 			len -= seg_tailroom;
1040 			seg_headroom = 0;
1041 		}
1042 		/* Update head and tail segments. */
1043 		*pkt_buf_next = NULL;
1044 		assert(pkt_buf != NULL);
1045 		assert(j != 0);
1046 		NB_SEGS(pkt_buf) = j;
1047 		PORT(pkt_buf) = rxq->port_id;
1048 		PKT_LEN(pkt_buf) = pkt_buf_len;
1049 		if (rxq->csum | rxq->csum_l2tun | rxq->vlan_strip) {
1050 			pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
1051 			pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
1052 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
1053 			if (flags & IBV_EXP_CQ_RX_CVLAN_STRIPPED_V1) {
1054 				pkt_buf->ol_flags |= PKT_RX_VLAN_PKT;
1055 				pkt_buf->vlan_tci = vlan_tci;
1056 			}
1057 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
1058 		}
1059 
1060 		/* Return packet. */
1061 		*(pkts++) = pkt_buf;
1062 		++pkts_ret;
1063 #ifdef MLX5_PMD_SOFT_COUNTERS
1064 		/* Increment bytes counter. */
1065 		rxq->stats.ibytes += pkt_buf_len;
1066 #endif
1067 repost:
1068 		ret = rxq->recv(rxq->wq, elt->sges, RTE_DIM(elt->sges));
1069 		if (unlikely(ret)) {
1070 			/* Inability to repost WRs is fatal. */
1071 			DEBUG("%p: recv_sg_list(): failed (ret=%d)",
1072 			      (void *)rxq->priv,
1073 			      ret);
1074 			abort();
1075 		}
1076 		if (++elts_head >= elts_n)
1077 			elts_head = 0;
1078 		continue;
1079 	}
1080 	if (unlikely(i == 0))
1081 		return 0;
1082 	rxq->elts_head = elts_head;
1083 #ifdef MLX5_PMD_SOFT_COUNTERS
1084 	/* Increment packets counter. */
1085 	rxq->stats.ipackets += pkts_ret;
1086 #endif
1087 	return pkts_ret;
1088 }
1089 
1090 /**
1091  * DPDK callback for RX.
1092  *
1093  * The following function is the same as mlx5_rx_burst_sp(), except it doesn't
1094  * manage scattered packets. Improves performance when MRU is lower than the
1095  * size of the first segment.
1096  *
1097  * @param dpdk_rxq
1098  *   Generic pointer to RX queue structure.
1099  * @param[out] pkts
1100  *   Array to store received packets.
1101  * @param pkts_n
1102  *   Maximum number of packets in array.
1103  *
1104  * @return
1105  *   Number of packets successfully received (<= pkts_n).
1106  */
1107 uint16_t
1108 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1109 {
1110 	struct rxq *rxq = (struct rxq *)dpdk_rxq;
1111 	struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
1112 	const unsigned int elts_n = rxq->elts_n;
1113 	unsigned int elts_head = rxq->elts_head;
1114 	struct ibv_sge sges[pkts_n];
1115 	unsigned int i;
1116 	unsigned int pkts_ret = 0;
1117 	int ret;
1118 
1119 	if (unlikely(rxq->sp))
1120 		return mlx5_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
1121 	for (i = 0; (i != pkts_n); ++i) {
1122 		struct rxq_elt *elt = &(*elts)[elts_head];
1123 		unsigned int len;
1124 		struct rte_mbuf *seg = elt->buf;
1125 		struct rte_mbuf *rep;
1126 		uint32_t flags;
1127 		uint16_t vlan_tci;
1128 
1129 		/* Sanity checks. */
1130 		assert(seg != NULL);
1131 		assert(elts_head < rxq->elts_n);
1132 		assert(rxq->elts_head < rxq->elts_n);
1133 		/*
1134 		 * Fetch initial bytes of packet descriptor into a
1135 		 * cacheline while allocating rep.
1136 		 */
1137 		rte_mbuf_prefetch_part1(seg);
1138 		rte_mbuf_prefetch_part2(seg);
1139 		ret = rxq->poll(rxq->cq, NULL, NULL, &flags, &vlan_tci);
1140 		if (unlikely(ret < 0)) {
1141 			struct ibv_wc wc;
1142 			int wcs_n;
1143 
1144 			DEBUG("rxq=%p, poll_length() failed (ret=%d)",
1145 			      (void *)rxq, ret);
1146 			/* ibv_poll_cq() must be used in case of failure. */
1147 			wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
1148 			if (unlikely(wcs_n == 0))
1149 				break;
1150 			if (unlikely(wcs_n < 0)) {
1151 				DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
1152 				      (void *)rxq, wcs_n);
1153 				break;
1154 			}
1155 			assert(wcs_n == 1);
1156 			if (unlikely(wc.status != IBV_WC_SUCCESS)) {
1157 				/* Whatever, just repost the offending WR. */
1158 				DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
1159 				      " completion status (%d): %s",
1160 				      (void *)rxq, wc.wr_id, wc.status,
1161 				      ibv_wc_status_str(wc.status));
1162 #ifdef MLX5_PMD_SOFT_COUNTERS
1163 				/* Increment dropped packets counter. */
1164 				++rxq->stats.idropped;
1165 #endif
1166 				/* Add SGE to array for repost. */
1167 				sges[i] = elt->sge;
1168 				goto repost;
1169 			}
1170 			ret = wc.byte_len;
1171 		}
1172 		if (ret == 0)
1173 			break;
1174 		assert(ret >= (rxq->crc_present << 2));
1175 		len = ret - (rxq->crc_present << 2);
1176 		rep = rte_mbuf_raw_alloc(rxq->mp);
1177 		if (unlikely(rep == NULL)) {
1178 			/*
1179 			 * Unable to allocate a replacement mbuf,
1180 			 * repost WR.
1181 			 */
1182 			DEBUG("rxq=%p: can't allocate a new mbuf",
1183 			      (void *)rxq);
1184 			/* Increment out of memory counters. */
1185 			++rxq->stats.rx_nombuf;
1186 			++rxq->priv->dev->data->rx_mbuf_alloc_failed;
1187 			goto repost;
1188 		}
1189 
1190 		/* Reconfigure sge to use rep instead of seg. */
1191 		elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
1192 		assert(elt->sge.lkey == rxq->mr->lkey);
1193 		elt->buf = rep;
1194 
1195 		/* Add SGE to array for repost. */
1196 		sges[i] = elt->sge;
1197 
1198 		/* Update seg information. */
1199 		SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
1200 		NB_SEGS(seg) = 1;
1201 		PORT(seg) = rxq->port_id;
1202 		NEXT(seg) = NULL;
1203 		PKT_LEN(seg) = len;
1204 		DATA_LEN(seg) = len;
1205 		if (rxq->csum | rxq->csum_l2tun | rxq->vlan_strip) {
1206 			seg->packet_type = rxq_cq_to_pkt_type(flags);
1207 			seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
1208 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
1209 			if (flags & IBV_EXP_CQ_RX_CVLAN_STRIPPED_V1) {
1210 				seg->ol_flags |= PKT_RX_VLAN_PKT;
1211 				seg->vlan_tci = vlan_tci;
1212 			}
1213 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
1214 		}
1215 		/* Return packet. */
1216 		*(pkts++) = seg;
1217 		++pkts_ret;
1218 #ifdef MLX5_PMD_SOFT_COUNTERS
1219 		/* Increment bytes counter. */
1220 		rxq->stats.ibytes += len;
1221 #endif
1222 repost:
1223 		if (++elts_head >= elts_n)
1224 			elts_head = 0;
1225 		continue;
1226 	}
1227 	if (unlikely(i == 0))
1228 		return 0;
1229 	/* Repost WRs. */
1230 #ifdef DEBUG_RECV
1231 	DEBUG("%p: reposting %u WRs", (void *)rxq, i);
1232 #endif
1233 	ret = rxq->recv(rxq->wq, sges, i);
1234 	if (unlikely(ret)) {
1235 		/* Inability to repost WRs is fatal. */
1236 		DEBUG("%p: recv_burst(): failed (ret=%d)",
1237 		      (void *)rxq->priv,
1238 		      ret);
1239 		abort();
1240 	}
1241 	rxq->elts_head = elts_head;
1242 #ifdef MLX5_PMD_SOFT_COUNTERS
1243 	/* Increment packets counter. */
1244 	rxq->stats.ipackets += pkts_ret;
1245 #endif
1246 	return pkts_ret;
1247 }
1248 
1249 /**
1250  * Dummy DPDK callback for TX.
1251  *
1252  * This function is used to temporarily replace the real callback during
1253  * unsafe control operations on the queue, or in case of error.
1254  *
1255  * @param dpdk_txq
1256  *   Generic pointer to TX queue structure.
1257  * @param[in] pkts
1258  *   Packets to transmit.
1259  * @param pkts_n
1260  *   Number of packets in array.
1261  *
1262  * @return
1263  *   Number of packets successfully transmitted (<= pkts_n).
1264  */
1265 uint16_t
1266 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1267 {
1268 	(void)dpdk_txq;
1269 	(void)pkts;
1270 	(void)pkts_n;
1271 	return 0;
1272 }
1273 
1274 /**
1275  * Dummy DPDK callback for RX.
1276  *
1277  * This function is used to temporarily replace the real callback during
1278  * unsafe control operations on the queue, or in case of error.
1279  *
1280  * @param dpdk_rxq
1281  *   Generic pointer to RX queue structure.
1282  * @param[out] pkts
1283  *   Array to store received packets.
1284  * @param pkts_n
1285  *   Maximum number of packets in array.
1286  *
1287  * @return
1288  *   Number of packets successfully received (<= pkts_n).
1289  */
1290 uint16_t
1291 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1292 {
1293 	(void)dpdk_rxq;
1294 	(void)pkts;
1295 	(void)pkts_n;
1296 	return 0;
1297 }
1298