xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.c (revision b79e4c00af0e7cfb8601ab0208659d226b82bd10)
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 "-Wpedantic"
43 #endif
44 #include <infiniband/verbs.h>
45 #include <infiniband/mlx5_hw.h>
46 #include <infiniband/arch.h>
47 #ifdef PEDANTIC
48 #pragma GCC diagnostic error "-Wpedantic"
49 #endif
50 
51 /* DPDK headers don't like -pedantic. */
52 #ifdef PEDANTIC
53 #pragma GCC diagnostic ignored "-Wpedantic"
54 #endif
55 #include <rte_mbuf.h>
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>
61 #ifdef PEDANTIC
62 #pragma GCC diagnostic error "-Wpedantic"
63 #endif
64 
65 #include "mlx5.h"
66 #include "mlx5_utils.h"
67 #include "mlx5_rxtx.h"
68 #include "mlx5_autoconf.h"
69 #include "mlx5_defs.h"
70 #include "mlx5_prm.h"
71 
72 static __rte_always_inline uint32_t
73 rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe);
74 
75 static __rte_always_inline int
76 mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe,
77 		 uint16_t cqe_cnt, uint32_t *rss_hash);
78 
79 static __rte_always_inline uint32_t
80 rxq_cq_to_ol_flags(struct rxq *rxq, volatile struct mlx5_cqe *cqe);
81 
82 /*
83  * The index to the array should have:
84  * bit[1:0] = l3_hdr_type, bit[2] = tunneled, bit[3] = outer_l3_type
85  */
86 const uint32_t mlx5_ptype_table[] = {
87 	RTE_PTYPE_UNKNOWN,
88 	RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,               /* b0001 */
89 	RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,               /* b0010 */
90 	RTE_PTYPE_UNKNOWN, RTE_PTYPE_UNKNOWN,
91 	RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
92 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN, /* b0101 */
93 	RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
94 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN, /* b0110 */
95 	RTE_PTYPE_UNKNOWN, RTE_PTYPE_UNKNOWN,
96 	RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,               /* b1001 */
97 	RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,               /* b1010 */
98 	RTE_PTYPE_UNKNOWN, RTE_PTYPE_UNKNOWN,
99 	RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
100 		RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN, /* b1101 */
101 	RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
102 		RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN, /* b1110 */
103 	RTE_PTYPE_ALL_MASK			     /* b1111 */
104 };
105 
106 /**
107  * Return the size of tailroom of WQ.
108  *
109  * @param txq
110  *   Pointer to TX queue structure.
111  * @param addr
112  *   Pointer to tail of WQ.
113  *
114  * @return
115  *   Size of tailroom.
116  */
117 static inline size_t
118 tx_mlx5_wq_tailroom(struct txq *txq, void *addr)
119 {
120 	size_t tailroom;
121 	tailroom = (uintptr_t)(txq->wqes) +
122 		   (1 << txq->wqe_n) * MLX5_WQE_SIZE -
123 		   (uintptr_t)addr;
124 	return tailroom;
125 }
126 
127 /**
128  * Copy data to tailroom of circular queue.
129  *
130  * @param dst
131  *   Pointer to destination.
132  * @param src
133  *   Pointer to source.
134  * @param n
135  *   Number of bytes to copy.
136  * @param base
137  *   Pointer to head of queue.
138  * @param tailroom
139  *   Size of tailroom from dst.
140  *
141  * @return
142  *   Pointer after copied data.
143  */
144 static inline void *
145 mlx5_copy_to_wq(void *dst, const void *src, size_t n,
146 		void *base, size_t tailroom)
147 {
148 	void *ret;
149 
150 	if (n > tailroom) {
151 		rte_memcpy(dst, src, tailroom);
152 		rte_memcpy(base, (void *)((uintptr_t)src + tailroom),
153 			   n - tailroom);
154 		ret = (uint8_t *)base + n - tailroom;
155 	} else {
156 		rte_memcpy(dst, src, n);
157 		ret = (n == tailroom) ? base : (uint8_t *)dst + n;
158 	}
159 	return ret;
160 }
161 
162 /**
163  * DPDK callback to check the status of a tx descriptor.
164  *
165  * @param tx_queue
166  *   The tx queue.
167  * @param[in] offset
168  *   The index of the descriptor in the ring.
169  *
170  * @return
171  *   The status of the tx descriptor.
172  */
173 int
174 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
175 {
176 	struct txq *txq = tx_queue;
177 	uint16_t used;
178 
179 	mlx5_tx_complete(txq);
180 	used = txq->elts_head - txq->elts_tail;
181 	if (offset < used)
182 		return RTE_ETH_TX_DESC_FULL;
183 	return RTE_ETH_TX_DESC_DONE;
184 }
185 
186 /**
187  * DPDK callback to check the status of a rx descriptor.
188  *
189  * @param rx_queue
190  *   The rx queue.
191  * @param[in] offset
192  *   The index of the descriptor in the ring.
193  *
194  * @return
195  *   The status of the tx descriptor.
196  */
197 int
198 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
199 {
200 	struct rxq *rxq = rx_queue;
201 	struct rxq_zip *zip = &rxq->zip;
202 	volatile struct mlx5_cqe *cqe;
203 	const unsigned int cqe_n = (1 << rxq->cqe_n);
204 	const unsigned int cqe_cnt = cqe_n - 1;
205 	unsigned int cq_ci;
206 	unsigned int used;
207 
208 	/* if we are processing a compressed cqe */
209 	if (zip->ai) {
210 		used = zip->cqe_cnt - zip->ca;
211 		cq_ci = zip->cq_ci;
212 	} else {
213 		used = 0;
214 		cq_ci = rxq->cq_ci;
215 	}
216 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
217 	while (check_cqe(cqe, cqe_n, cq_ci) == 0) {
218 		int8_t op_own;
219 		unsigned int n;
220 
221 		op_own = cqe->op_own;
222 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
223 			n = ntohl(cqe->byte_cnt);
224 		else
225 			n = 1;
226 		cq_ci += n;
227 		used += n;
228 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
229 	}
230 	used = RTE_MIN(used, (1U << rxq->elts_n) - 1);
231 	if (offset < used)
232 		return RTE_ETH_RX_DESC_DONE;
233 	return RTE_ETH_RX_DESC_AVAIL;
234 }
235 
236 /**
237  * DPDK callback for TX.
238  *
239  * @param dpdk_txq
240  *   Generic pointer to TX queue structure.
241  * @param[in] pkts
242  *   Packets to transmit.
243  * @param pkts_n
244  *   Number of packets in array.
245  *
246  * @return
247  *   Number of packets successfully transmitted (<= pkts_n).
248  */
249 uint16_t
250 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
251 {
252 	struct txq *txq = (struct txq *)dpdk_txq;
253 	uint16_t elts_head = txq->elts_head;
254 	const uint16_t elts_n = 1 << txq->elts_n;
255 	const uint16_t elts_m = elts_n - 1;
256 	unsigned int i = 0;
257 	unsigned int j = 0;
258 	unsigned int k = 0;
259 	uint16_t max_elts;
260 	unsigned int max_inline = txq->max_inline;
261 	const unsigned int inline_en = !!max_inline && txq->inline_en;
262 	uint16_t max_wqe;
263 	unsigned int comp;
264 	volatile struct mlx5_wqe_v *wqe = NULL;
265 	volatile struct mlx5_wqe_ctrl *last_wqe = NULL;
266 	unsigned int segs_n = 0;
267 	struct rte_mbuf *buf = NULL;
268 	uint8_t *raw;
269 
270 	if (unlikely(!pkts_n))
271 		return 0;
272 	/* Prefetch first packet cacheline. */
273 	rte_prefetch0(*pkts);
274 	/* Start processing. */
275 	mlx5_tx_complete(txq);
276 	max_elts = (elts_n - (elts_head - txq->elts_tail));
277 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
278 	if (unlikely(!max_wqe))
279 		return 0;
280 	do {
281 		volatile rte_v128u32_t *dseg = NULL;
282 		uint32_t length;
283 		unsigned int ds = 0;
284 		unsigned int sg = 0; /* counter of additional segs attached. */
285 		uintptr_t addr;
286 		uint64_t naddr;
287 		uint16_t pkt_inline_sz = MLX5_WQE_DWORD_SIZE + 2;
288 		uint16_t tso_header_sz = 0;
289 		uint16_t ehdr;
290 		uint8_t cs_flags = 0;
291 		uint64_t tso = 0;
292 		uint16_t tso_segsz = 0;
293 #ifdef MLX5_PMD_SOFT_COUNTERS
294 		uint32_t total_length = 0;
295 #endif
296 
297 		/* first_seg */
298 		buf = *pkts;
299 		segs_n = buf->nb_segs;
300 		/*
301 		 * Make sure there is enough room to store this packet and
302 		 * that one ring entry remains unused.
303 		 */
304 		assert(segs_n);
305 		if (max_elts < segs_n)
306 			break;
307 		max_elts -= segs_n;
308 		--segs_n;
309 		if (unlikely(--max_wqe == 0))
310 			break;
311 		wqe = (volatile struct mlx5_wqe_v *)
312 			tx_mlx5_wqe(txq, txq->wqe_ci);
313 		rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
314 		if (pkts_n - i > 1)
315 			rte_prefetch0(*(pkts + 1));
316 		addr = rte_pktmbuf_mtod(buf, uintptr_t);
317 		length = DATA_LEN(buf);
318 		ehdr = (((uint8_t *)addr)[1] << 8) |
319 		       ((uint8_t *)addr)[0];
320 #ifdef MLX5_PMD_SOFT_COUNTERS
321 		total_length = length;
322 #endif
323 		if (length < (MLX5_WQE_DWORD_SIZE + 2))
324 			break;
325 		/* Update element. */
326 		(*txq->elts)[elts_head & elts_m] = buf;
327 		/* Prefetch next buffer data. */
328 		if (pkts_n - i > 1)
329 			rte_prefetch0(
330 			    rte_pktmbuf_mtod(*(pkts + 1), volatile void *));
331 		/* Should we enable HW CKSUM offload */
332 		if (buf->ol_flags &
333 		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
334 			const uint64_t is_tunneled = buf->ol_flags &
335 						     (PKT_TX_TUNNEL_GRE |
336 						      PKT_TX_TUNNEL_VXLAN);
337 
338 			if (is_tunneled && txq->tunnel_en) {
339 				cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
340 					   MLX5_ETH_WQE_L4_INNER_CSUM;
341 				if (buf->ol_flags & PKT_TX_OUTER_IP_CKSUM)
342 					cs_flags |= MLX5_ETH_WQE_L3_CSUM;
343 			} else {
344 				cs_flags = MLX5_ETH_WQE_L3_CSUM |
345 					   MLX5_ETH_WQE_L4_CSUM;
346 			}
347 		}
348 		raw = ((uint8_t *)(uintptr_t)wqe) + 2 * MLX5_WQE_DWORD_SIZE;
349 		/* Replace the Ethernet type by the VLAN if necessary. */
350 		if (buf->ol_flags & PKT_TX_VLAN_PKT) {
351 			uint32_t vlan = htonl(0x81000000 | buf->vlan_tci);
352 			unsigned int len = 2 * ETHER_ADDR_LEN - 2;
353 
354 			addr += 2;
355 			length -= 2;
356 			/* Copy Destination and source mac address. */
357 			memcpy((uint8_t *)raw, ((uint8_t *)addr), len);
358 			/* Copy VLAN. */
359 			memcpy((uint8_t *)raw + len, &vlan, sizeof(vlan));
360 			/* Copy missing two bytes to end the DSeg. */
361 			memcpy((uint8_t *)raw + len + sizeof(vlan),
362 			       ((uint8_t *)addr) + len, 2);
363 			addr += len + 2;
364 			length -= (len + 2);
365 		} else {
366 			memcpy((uint8_t *)raw, ((uint8_t *)addr) + 2,
367 			       MLX5_WQE_DWORD_SIZE);
368 			length -= pkt_inline_sz;
369 			addr += pkt_inline_sz;
370 		}
371 		if (txq->tso_en) {
372 			tso = buf->ol_flags & PKT_TX_TCP_SEG;
373 			if (tso) {
374 				uintptr_t end = (uintptr_t)
375 						(((uintptr_t)txq->wqes) +
376 						(1 << txq->wqe_n) *
377 						MLX5_WQE_SIZE);
378 				unsigned int copy_b;
379 				uint8_t vlan_sz = (buf->ol_flags &
380 						  PKT_TX_VLAN_PKT) ? 4 : 0;
381 				const uint64_t is_tunneled =
382 							buf->ol_flags &
383 							(PKT_TX_TUNNEL_GRE |
384 							 PKT_TX_TUNNEL_VXLAN);
385 
386 				tso_header_sz = buf->l2_len + vlan_sz +
387 						buf->l3_len + buf->l4_len;
388 				tso_segsz = buf->tso_segsz;
389 
390 				if (is_tunneled	&& txq->tunnel_en) {
391 					tso_header_sz += buf->outer_l2_len +
392 							 buf->outer_l3_len;
393 					cs_flags |= MLX5_ETH_WQE_L4_INNER_CSUM;
394 				} else {
395 					cs_flags |= MLX5_ETH_WQE_L4_CSUM;
396 				}
397 				if (unlikely(tso_header_sz >
398 					     MLX5_MAX_TSO_HEADER))
399 					break;
400 				copy_b = tso_header_sz - pkt_inline_sz;
401 				/* First seg must contain all headers. */
402 				assert(copy_b <= length);
403 				raw += MLX5_WQE_DWORD_SIZE;
404 				if (copy_b &&
405 				   ((end - (uintptr_t)raw) > copy_b)) {
406 					uint16_t n = (MLX5_WQE_DS(copy_b) -
407 						      1 + 3) / 4;
408 
409 					if (unlikely(max_wqe < n))
410 						break;
411 					max_wqe -= n;
412 					rte_memcpy((void *)raw,
413 						   (void *)addr, copy_b);
414 					addr += copy_b;
415 					length -= copy_b;
416 					pkt_inline_sz += copy_b;
417 					/*
418 					 * Another DWORD will be added
419 					 * in the inline part.
420 					 */
421 					raw += MLX5_WQE_DS(copy_b) *
422 					       MLX5_WQE_DWORD_SIZE -
423 					       MLX5_WQE_DWORD_SIZE;
424 				} else {
425 					/* NOP WQE. */
426 					wqe->ctrl = (rte_v128u32_t){
427 						     htonl(txq->wqe_ci << 8),
428 						     htonl(txq->qp_num_8s | 1),
429 						     0,
430 						     0,
431 					};
432 					ds = 1;
433 					total_length = 0;
434 					k++;
435 					goto next_wqe;
436 				}
437 			}
438 		}
439 		/* Inline if enough room. */
440 		if (inline_en || tso) {
441 			uintptr_t end = (uintptr_t)
442 				(((uintptr_t)txq->wqes) +
443 				 (1 << txq->wqe_n) * MLX5_WQE_SIZE);
444 			unsigned int inline_room = max_inline *
445 						   RTE_CACHE_LINE_SIZE -
446 						   (pkt_inline_sz - 2);
447 			uintptr_t addr_end = (addr + inline_room) &
448 					     ~(RTE_CACHE_LINE_SIZE - 1);
449 			unsigned int copy_b = (addr_end > addr) ?
450 				RTE_MIN((addr_end - addr), length) :
451 				0;
452 
453 			raw += MLX5_WQE_DWORD_SIZE;
454 			if (copy_b && ((end - (uintptr_t)raw) > copy_b)) {
455 				/*
456 				 * One Dseg remains in the current WQE.  To
457 				 * keep the computation positive, it is
458 				 * removed after the bytes to Dseg conversion.
459 				 */
460 				uint16_t n = (MLX5_WQE_DS(copy_b) - 1 + 3) / 4;
461 
462 				if (unlikely(max_wqe < n))
463 					break;
464 				max_wqe -= n;
465 				if (tso) {
466 					uint32_t inl =
467 						htonl(copy_b | MLX5_INLINE_SEG);
468 
469 					pkt_inline_sz =
470 						MLX5_WQE_DS(tso_header_sz) *
471 						MLX5_WQE_DWORD_SIZE;
472 					rte_memcpy((void *)raw,
473 						   (void *)&inl, sizeof(inl));
474 					raw += sizeof(inl);
475 					pkt_inline_sz += sizeof(inl);
476 				}
477 				rte_memcpy((void *)raw, (void *)addr, copy_b);
478 				addr += copy_b;
479 				length -= copy_b;
480 				pkt_inline_sz += copy_b;
481 			}
482 			/*
483 			 * 2 DWORDs consumed by the WQE header + ETH segment +
484 			 * the size of the inline part of the packet.
485 			 */
486 			ds = 2 + MLX5_WQE_DS(pkt_inline_sz - 2);
487 			if (length > 0) {
488 				if (ds % (MLX5_WQE_SIZE /
489 					  MLX5_WQE_DWORD_SIZE) == 0) {
490 					if (unlikely(--max_wqe == 0))
491 						break;
492 					dseg = (volatile rte_v128u32_t *)
493 					       tx_mlx5_wqe(txq, txq->wqe_ci +
494 							   ds / 4);
495 				} else {
496 					dseg = (volatile rte_v128u32_t *)
497 						((uintptr_t)wqe +
498 						 (ds * MLX5_WQE_DWORD_SIZE));
499 				}
500 				goto use_dseg;
501 			} else if (!segs_n) {
502 				goto next_pkt;
503 			} else {
504 				/* dseg will be advance as part of next_seg */
505 				dseg = (volatile rte_v128u32_t *)
506 					((uintptr_t)wqe +
507 					 ((ds - 1) * MLX5_WQE_DWORD_SIZE));
508 				goto next_seg;
509 			}
510 		} else {
511 			/*
512 			 * No inline has been done in the packet, only the
513 			 * Ethernet Header as been stored.
514 			 */
515 			dseg = (volatile rte_v128u32_t *)
516 				((uintptr_t)wqe + (3 * MLX5_WQE_DWORD_SIZE));
517 			ds = 3;
518 use_dseg:
519 			/* Add the remaining packet as a simple ds. */
520 			naddr = htonll(addr);
521 			*dseg = (rte_v128u32_t){
522 				htonl(length),
523 				mlx5_tx_mb2mr(txq, buf),
524 				naddr,
525 				naddr >> 32,
526 			};
527 			++ds;
528 			if (!segs_n)
529 				goto next_pkt;
530 		}
531 next_seg:
532 		assert(buf);
533 		assert(ds);
534 		assert(wqe);
535 		/*
536 		 * Spill on next WQE when the current one does not have
537 		 * enough room left. Size of WQE must a be a multiple
538 		 * of data segment size.
539 		 */
540 		assert(!(MLX5_WQE_SIZE % MLX5_WQE_DWORD_SIZE));
541 		if (!(ds % (MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE))) {
542 			if (unlikely(--max_wqe == 0))
543 				break;
544 			dseg = (volatile rte_v128u32_t *)
545 			       tx_mlx5_wqe(txq, txq->wqe_ci + ds / 4);
546 			rte_prefetch0(tx_mlx5_wqe(txq,
547 						  txq->wqe_ci + ds / 4 + 1));
548 		} else {
549 			++dseg;
550 		}
551 		++ds;
552 		buf = buf->next;
553 		assert(buf);
554 		length = DATA_LEN(buf);
555 #ifdef MLX5_PMD_SOFT_COUNTERS
556 		total_length += length;
557 #endif
558 		/* Store segment information. */
559 		naddr = htonll(rte_pktmbuf_mtod(buf, uintptr_t));
560 		*dseg = (rte_v128u32_t){
561 			htonl(length),
562 			mlx5_tx_mb2mr(txq, buf),
563 			naddr,
564 			naddr >> 32,
565 		};
566 		(*txq->elts)[++elts_head & elts_m] = buf;
567 		++sg;
568 		/* Advance counter only if all segs are successfully posted. */
569 		if (sg < segs_n)
570 			goto next_seg;
571 		else
572 			j += sg;
573 next_pkt:
574 		++elts_head;
575 		++pkts;
576 		++i;
577 		/* Initialize known and common part of the WQE structure. */
578 		if (tso) {
579 			wqe->ctrl = (rte_v128u32_t){
580 				htonl((txq->wqe_ci << 8) | MLX5_OPCODE_TSO),
581 				htonl(txq->qp_num_8s | ds),
582 				0,
583 				0,
584 			};
585 			wqe->eseg = (rte_v128u32_t){
586 				0,
587 				cs_flags | (htons(tso_segsz) << 16),
588 				0,
589 				(ehdr << 16) | htons(tso_header_sz),
590 			};
591 		} else {
592 			wqe->ctrl = (rte_v128u32_t){
593 				htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND),
594 				htonl(txq->qp_num_8s | ds),
595 				0,
596 				0,
597 			};
598 			wqe->eseg = (rte_v128u32_t){
599 				0,
600 				cs_flags,
601 				0,
602 				(ehdr << 16) | htons(pkt_inline_sz),
603 			};
604 		}
605 next_wqe:
606 		txq->wqe_ci += (ds + 3) / 4;
607 		/* Save the last successful WQE for completion request */
608 		last_wqe = (volatile struct mlx5_wqe_ctrl *)wqe;
609 #ifdef MLX5_PMD_SOFT_COUNTERS
610 		/* Increment sent bytes counter. */
611 		txq->stats.obytes += total_length;
612 #endif
613 	} while (i < pkts_n);
614 	/* Take a shortcut if nothing must be sent. */
615 	if (unlikely((i + k) == 0))
616 		return 0;
617 	txq->elts_head += (i + j);
618 	/* Check whether completion threshold has been reached. */
619 	comp = txq->elts_comp + i + j + k;
620 	if (comp >= MLX5_TX_COMP_THRESH) {
621 		/* Request completion on last WQE. */
622 		last_wqe->ctrl2 = htonl(8);
623 		/* Save elts_head in unused "immediate" field of WQE. */
624 		last_wqe->ctrl3 = txq->elts_head;
625 		txq->elts_comp = 0;
626 	} else {
627 		txq->elts_comp = comp;
628 	}
629 #ifdef MLX5_PMD_SOFT_COUNTERS
630 	/* Increment sent packets counter. */
631 	txq->stats.opackets += i;
632 #endif
633 	/* Ring QP doorbell. */
634 	mlx5_tx_dbrec(txq, (volatile struct mlx5_wqe *)last_wqe);
635 	return i;
636 }
637 
638 /**
639  * Open a MPW session.
640  *
641  * @param txq
642  *   Pointer to TX queue structure.
643  * @param mpw
644  *   Pointer to MPW session structure.
645  * @param length
646  *   Packet length.
647  */
648 static inline void
649 mlx5_mpw_new(struct txq *txq, struct mlx5_mpw *mpw, uint32_t length)
650 {
651 	uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
652 	volatile struct mlx5_wqe_data_seg (*dseg)[MLX5_MPW_DSEG_MAX] =
653 		(volatile struct mlx5_wqe_data_seg (*)[])
654 		tx_mlx5_wqe(txq, idx + 1);
655 
656 	mpw->state = MLX5_MPW_STATE_OPENED;
657 	mpw->pkts_n = 0;
658 	mpw->len = length;
659 	mpw->total_len = 0;
660 	mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
661 	mpw->wqe->eseg.mss = htons(length);
662 	mpw->wqe->eseg.inline_hdr_sz = 0;
663 	mpw->wqe->eseg.rsvd0 = 0;
664 	mpw->wqe->eseg.rsvd1 = 0;
665 	mpw->wqe->eseg.rsvd2 = 0;
666 	mpw->wqe->ctrl[0] = htonl((MLX5_OPC_MOD_MPW << 24) |
667 				  (txq->wqe_ci << 8) | MLX5_OPCODE_TSO);
668 	mpw->wqe->ctrl[2] = 0;
669 	mpw->wqe->ctrl[3] = 0;
670 	mpw->data.dseg[0] = (volatile struct mlx5_wqe_data_seg *)
671 		(((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
672 	mpw->data.dseg[1] = (volatile struct mlx5_wqe_data_seg *)
673 		(((uintptr_t)mpw->wqe) + (3 * MLX5_WQE_DWORD_SIZE));
674 	mpw->data.dseg[2] = &(*dseg)[0];
675 	mpw->data.dseg[3] = &(*dseg)[1];
676 	mpw->data.dseg[4] = &(*dseg)[2];
677 }
678 
679 /**
680  * Close a MPW session.
681  *
682  * @param txq
683  *   Pointer to TX queue structure.
684  * @param mpw
685  *   Pointer to MPW session structure.
686  */
687 static inline void
688 mlx5_mpw_close(struct txq *txq, struct mlx5_mpw *mpw)
689 {
690 	unsigned int num = mpw->pkts_n;
691 
692 	/*
693 	 * Store size in multiple of 16 bytes. Control and Ethernet segments
694 	 * count as 2.
695 	 */
696 	mpw->wqe->ctrl[1] = htonl(txq->qp_num_8s | (2 + num));
697 	mpw->state = MLX5_MPW_STATE_CLOSED;
698 	if (num < 3)
699 		++txq->wqe_ci;
700 	else
701 		txq->wqe_ci += 2;
702 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
703 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
704 }
705 
706 /**
707  * DPDK callback for TX with MPW support.
708  *
709  * @param dpdk_txq
710  *   Generic pointer to TX queue structure.
711  * @param[in] pkts
712  *   Packets to transmit.
713  * @param pkts_n
714  *   Number of packets in array.
715  *
716  * @return
717  *   Number of packets successfully transmitted (<= pkts_n).
718  */
719 uint16_t
720 mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
721 {
722 	struct txq *txq = (struct txq *)dpdk_txq;
723 	uint16_t elts_head = txq->elts_head;
724 	const uint16_t elts_n = 1 << txq->elts_n;
725 	const uint16_t elts_m = elts_n - 1;
726 	unsigned int i = 0;
727 	unsigned int j = 0;
728 	uint16_t max_elts;
729 	uint16_t max_wqe;
730 	unsigned int comp;
731 	struct mlx5_mpw mpw = {
732 		.state = MLX5_MPW_STATE_CLOSED,
733 	};
734 
735 	if (unlikely(!pkts_n))
736 		return 0;
737 	/* Prefetch first packet cacheline. */
738 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
739 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
740 	/* Start processing. */
741 	mlx5_tx_complete(txq);
742 	max_elts = (elts_n - (elts_head - txq->elts_tail));
743 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
744 	if (unlikely(!max_wqe))
745 		return 0;
746 	do {
747 		struct rte_mbuf *buf = *(pkts++);
748 		uint32_t length;
749 		unsigned int segs_n = buf->nb_segs;
750 		uint32_t cs_flags = 0;
751 
752 		/*
753 		 * Make sure there is enough room to store this packet and
754 		 * that one ring entry remains unused.
755 		 */
756 		assert(segs_n);
757 		if (max_elts < segs_n)
758 			break;
759 		/* Do not bother with large packets MPW cannot handle. */
760 		if (segs_n > MLX5_MPW_DSEG_MAX)
761 			break;
762 		max_elts -= segs_n;
763 		--pkts_n;
764 		/* Should we enable HW CKSUM offload */
765 		if (buf->ol_flags &
766 		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
767 			cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
768 		/* Retrieve packet information. */
769 		length = PKT_LEN(buf);
770 		assert(length);
771 		/* Start new session if packet differs. */
772 		if ((mpw.state == MLX5_MPW_STATE_OPENED) &&
773 		    ((mpw.len != length) ||
774 		     (segs_n != 1) ||
775 		     (mpw.wqe->eseg.cs_flags != cs_flags)))
776 			mlx5_mpw_close(txq, &mpw);
777 		if (mpw.state == MLX5_MPW_STATE_CLOSED) {
778 			/*
779 			 * Multi-Packet WQE consumes at most two WQE.
780 			 * mlx5_mpw_new() expects to be able to use such
781 			 * resources.
782 			 */
783 			if (unlikely(max_wqe < 2))
784 				break;
785 			max_wqe -= 2;
786 			mlx5_mpw_new(txq, &mpw, length);
787 			mpw.wqe->eseg.cs_flags = cs_flags;
788 		}
789 		/* Multi-segment packets must be alone in their MPW. */
790 		assert((segs_n == 1) || (mpw.pkts_n == 0));
791 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
792 		length = 0;
793 #endif
794 		do {
795 			volatile struct mlx5_wqe_data_seg *dseg;
796 			uintptr_t addr;
797 
798 			assert(buf);
799 			(*txq->elts)[elts_head++ & elts_m] = buf;
800 			dseg = mpw.data.dseg[mpw.pkts_n];
801 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
802 			*dseg = (struct mlx5_wqe_data_seg){
803 				.byte_count = htonl(DATA_LEN(buf)),
804 				.lkey = mlx5_tx_mb2mr(txq, buf),
805 				.addr = htonll(addr),
806 			};
807 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
808 			length += DATA_LEN(buf);
809 #endif
810 			buf = buf->next;
811 			++mpw.pkts_n;
812 			++j;
813 		} while (--segs_n);
814 		assert(length == mpw.len);
815 		if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
816 			mlx5_mpw_close(txq, &mpw);
817 #ifdef MLX5_PMD_SOFT_COUNTERS
818 		/* Increment sent bytes counter. */
819 		txq->stats.obytes += length;
820 #endif
821 		++i;
822 	} while (pkts_n);
823 	/* Take a shortcut if nothing must be sent. */
824 	if (unlikely(i == 0))
825 		return 0;
826 	/* Check whether completion threshold has been reached. */
827 	/* "j" includes both packets and segments. */
828 	comp = txq->elts_comp + j;
829 	if (comp >= MLX5_TX_COMP_THRESH) {
830 		volatile struct mlx5_wqe *wqe = mpw.wqe;
831 
832 		/* Request completion on last WQE. */
833 		wqe->ctrl[2] = htonl(8);
834 		/* Save elts_head in unused "immediate" field of WQE. */
835 		wqe->ctrl[3] = elts_head;
836 		txq->elts_comp = 0;
837 	} else {
838 		txq->elts_comp = comp;
839 	}
840 #ifdef MLX5_PMD_SOFT_COUNTERS
841 	/* Increment sent packets counter. */
842 	txq->stats.opackets += i;
843 #endif
844 	/* Ring QP doorbell. */
845 	if (mpw.state == MLX5_MPW_STATE_OPENED)
846 		mlx5_mpw_close(txq, &mpw);
847 	mlx5_tx_dbrec(txq, mpw.wqe);
848 	txq->elts_head = elts_head;
849 	return i;
850 }
851 
852 /**
853  * Open a MPW inline session.
854  *
855  * @param txq
856  *   Pointer to TX queue structure.
857  * @param mpw
858  *   Pointer to MPW session structure.
859  * @param length
860  *   Packet length.
861  */
862 static inline void
863 mlx5_mpw_inline_new(struct txq *txq, struct mlx5_mpw *mpw, uint32_t length)
864 {
865 	uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
866 	struct mlx5_wqe_inl_small *inl;
867 
868 	mpw->state = MLX5_MPW_INL_STATE_OPENED;
869 	mpw->pkts_n = 0;
870 	mpw->len = length;
871 	mpw->total_len = 0;
872 	mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
873 	mpw->wqe->ctrl[0] = htonl((MLX5_OPC_MOD_MPW << 24) |
874 				  (txq->wqe_ci << 8) |
875 				  MLX5_OPCODE_TSO);
876 	mpw->wqe->ctrl[2] = 0;
877 	mpw->wqe->ctrl[3] = 0;
878 	mpw->wqe->eseg.mss = htons(length);
879 	mpw->wqe->eseg.inline_hdr_sz = 0;
880 	mpw->wqe->eseg.cs_flags = 0;
881 	mpw->wqe->eseg.rsvd0 = 0;
882 	mpw->wqe->eseg.rsvd1 = 0;
883 	mpw->wqe->eseg.rsvd2 = 0;
884 	inl = (struct mlx5_wqe_inl_small *)
885 		(((uintptr_t)mpw->wqe) + 2 * MLX5_WQE_DWORD_SIZE);
886 	mpw->data.raw = (uint8_t *)&inl->raw;
887 }
888 
889 /**
890  * Close a MPW inline session.
891  *
892  * @param txq
893  *   Pointer to TX queue structure.
894  * @param mpw
895  *   Pointer to MPW session structure.
896  */
897 static inline void
898 mlx5_mpw_inline_close(struct txq *txq, struct mlx5_mpw *mpw)
899 {
900 	unsigned int size;
901 	struct mlx5_wqe_inl_small *inl = (struct mlx5_wqe_inl_small *)
902 		(((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
903 
904 	size = MLX5_WQE_SIZE - MLX5_MWQE64_INL_DATA + mpw->total_len;
905 	/*
906 	 * Store size in multiple of 16 bytes. Control and Ethernet segments
907 	 * count as 2.
908 	 */
909 	mpw->wqe->ctrl[1] = htonl(txq->qp_num_8s | MLX5_WQE_DS(size));
910 	mpw->state = MLX5_MPW_STATE_CLOSED;
911 	inl->byte_cnt = htonl(mpw->total_len | MLX5_INLINE_SEG);
912 	txq->wqe_ci += (size + (MLX5_WQE_SIZE - 1)) / MLX5_WQE_SIZE;
913 }
914 
915 /**
916  * DPDK callback for TX with MPW inline support.
917  *
918  * @param dpdk_txq
919  *   Generic pointer to TX queue structure.
920  * @param[in] pkts
921  *   Packets to transmit.
922  * @param pkts_n
923  *   Number of packets in array.
924  *
925  * @return
926  *   Number of packets successfully transmitted (<= pkts_n).
927  */
928 uint16_t
929 mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
930 			 uint16_t pkts_n)
931 {
932 	struct txq *txq = (struct txq *)dpdk_txq;
933 	uint16_t elts_head = txq->elts_head;
934 	const uint16_t elts_n = 1 << txq->elts_n;
935 	const uint16_t elts_m = elts_n - 1;
936 	unsigned int i = 0;
937 	unsigned int j = 0;
938 	uint16_t max_elts;
939 	uint16_t max_wqe;
940 	unsigned int comp;
941 	unsigned int inline_room = txq->max_inline * RTE_CACHE_LINE_SIZE;
942 	struct mlx5_mpw mpw = {
943 		.state = MLX5_MPW_STATE_CLOSED,
944 	};
945 	/*
946 	 * Compute the maximum number of WQE which can be consumed by inline
947 	 * code.
948 	 * - 2 DSEG for:
949 	 *   - 1 control segment,
950 	 *   - 1 Ethernet segment,
951 	 * - N Dseg from the inline request.
952 	 */
953 	const unsigned int wqe_inl_n =
954 		((2 * MLX5_WQE_DWORD_SIZE +
955 		  txq->max_inline * RTE_CACHE_LINE_SIZE) +
956 		 RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
957 
958 	if (unlikely(!pkts_n))
959 		return 0;
960 	/* Prefetch first packet cacheline. */
961 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
962 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
963 	/* Start processing. */
964 	mlx5_tx_complete(txq);
965 	max_elts = (elts_n - (elts_head - txq->elts_tail));
966 	do {
967 		struct rte_mbuf *buf = *(pkts++);
968 		uintptr_t addr;
969 		uint32_t length;
970 		unsigned int segs_n = buf->nb_segs;
971 		uint32_t cs_flags = 0;
972 
973 		/*
974 		 * Make sure there is enough room to store this packet and
975 		 * that one ring entry remains unused.
976 		 */
977 		assert(segs_n);
978 		if (max_elts < segs_n)
979 			break;
980 		/* Do not bother with large packets MPW cannot handle. */
981 		if (segs_n > MLX5_MPW_DSEG_MAX)
982 			break;
983 		max_elts -= segs_n;
984 		--pkts_n;
985 		/*
986 		 * Compute max_wqe in case less WQE were consumed in previous
987 		 * iteration.
988 		 */
989 		max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
990 		/* Should we enable HW CKSUM offload */
991 		if (buf->ol_flags &
992 		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
993 			cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
994 		/* Retrieve packet information. */
995 		length = PKT_LEN(buf);
996 		/* Start new session if packet differs. */
997 		if (mpw.state == MLX5_MPW_STATE_OPENED) {
998 			if ((mpw.len != length) ||
999 			    (segs_n != 1) ||
1000 			    (mpw.wqe->eseg.cs_flags != cs_flags))
1001 				mlx5_mpw_close(txq, &mpw);
1002 		} else if (mpw.state == MLX5_MPW_INL_STATE_OPENED) {
1003 			if ((mpw.len != length) ||
1004 			    (segs_n != 1) ||
1005 			    (length > inline_room) ||
1006 			    (mpw.wqe->eseg.cs_flags != cs_flags)) {
1007 				mlx5_mpw_inline_close(txq, &mpw);
1008 				inline_room =
1009 					txq->max_inline * RTE_CACHE_LINE_SIZE;
1010 			}
1011 		}
1012 		if (mpw.state == MLX5_MPW_STATE_CLOSED) {
1013 			if ((segs_n != 1) ||
1014 			    (length > inline_room)) {
1015 				/*
1016 				 * Multi-Packet WQE consumes at most two WQE.
1017 				 * mlx5_mpw_new() expects to be able to use
1018 				 * such resources.
1019 				 */
1020 				if (unlikely(max_wqe < 2))
1021 					break;
1022 				max_wqe -= 2;
1023 				mlx5_mpw_new(txq, &mpw, length);
1024 				mpw.wqe->eseg.cs_flags = cs_flags;
1025 			} else {
1026 				if (unlikely(max_wqe < wqe_inl_n))
1027 					break;
1028 				max_wqe -= wqe_inl_n;
1029 				mlx5_mpw_inline_new(txq, &mpw, length);
1030 				mpw.wqe->eseg.cs_flags = cs_flags;
1031 			}
1032 		}
1033 		/* Multi-segment packets must be alone in their MPW. */
1034 		assert((segs_n == 1) || (mpw.pkts_n == 0));
1035 		if (mpw.state == MLX5_MPW_STATE_OPENED) {
1036 			assert(inline_room ==
1037 			       txq->max_inline * RTE_CACHE_LINE_SIZE);
1038 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1039 			length = 0;
1040 #endif
1041 			do {
1042 				volatile struct mlx5_wqe_data_seg *dseg;
1043 
1044 				assert(buf);
1045 				(*txq->elts)[elts_head++ & elts_m] = buf;
1046 				dseg = mpw.data.dseg[mpw.pkts_n];
1047 				addr = rte_pktmbuf_mtod(buf, uintptr_t);
1048 				*dseg = (struct mlx5_wqe_data_seg){
1049 					.byte_count = htonl(DATA_LEN(buf)),
1050 					.lkey = mlx5_tx_mb2mr(txq, buf),
1051 					.addr = htonll(addr),
1052 				};
1053 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1054 				length += DATA_LEN(buf);
1055 #endif
1056 				buf = buf->next;
1057 				++mpw.pkts_n;
1058 				++j;
1059 			} while (--segs_n);
1060 			assert(length == mpw.len);
1061 			if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
1062 				mlx5_mpw_close(txq, &mpw);
1063 		} else {
1064 			unsigned int max;
1065 
1066 			assert(mpw.state == MLX5_MPW_INL_STATE_OPENED);
1067 			assert(length <= inline_room);
1068 			assert(length == DATA_LEN(buf));
1069 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
1070 			(*txq->elts)[elts_head++ & elts_m] = buf;
1071 			/* Maximum number of bytes before wrapping. */
1072 			max = ((((uintptr_t)(txq->wqes)) +
1073 				(1 << txq->wqe_n) *
1074 				MLX5_WQE_SIZE) -
1075 			       (uintptr_t)mpw.data.raw);
1076 			if (length > max) {
1077 				rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1078 					   (void *)addr,
1079 					   max);
1080 				mpw.data.raw = (volatile void *)txq->wqes;
1081 				rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1082 					   (void *)(addr + max),
1083 					   length - max);
1084 				mpw.data.raw += length - max;
1085 			} else {
1086 				rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1087 					   (void *)addr,
1088 					   length);
1089 
1090 				if (length == max)
1091 					mpw.data.raw =
1092 						(volatile void *)txq->wqes;
1093 				else
1094 					mpw.data.raw += length;
1095 			}
1096 			++mpw.pkts_n;
1097 			mpw.total_len += length;
1098 			++j;
1099 			if (mpw.pkts_n == MLX5_MPW_DSEG_MAX) {
1100 				mlx5_mpw_inline_close(txq, &mpw);
1101 				inline_room =
1102 					txq->max_inline * RTE_CACHE_LINE_SIZE;
1103 			} else {
1104 				inline_room -= length;
1105 			}
1106 		}
1107 #ifdef MLX5_PMD_SOFT_COUNTERS
1108 		/* Increment sent bytes counter. */
1109 		txq->stats.obytes += length;
1110 #endif
1111 		++i;
1112 	} while (pkts_n);
1113 	/* Take a shortcut if nothing must be sent. */
1114 	if (unlikely(i == 0))
1115 		return 0;
1116 	/* Check whether completion threshold has been reached. */
1117 	/* "j" includes both packets and segments. */
1118 	comp = txq->elts_comp + j;
1119 	if (comp >= MLX5_TX_COMP_THRESH) {
1120 		volatile struct mlx5_wqe *wqe = mpw.wqe;
1121 
1122 		/* Request completion on last WQE. */
1123 		wqe->ctrl[2] = htonl(8);
1124 		/* Save elts_head in unused "immediate" field of WQE. */
1125 		wqe->ctrl[3] = elts_head;
1126 		txq->elts_comp = 0;
1127 	} else {
1128 		txq->elts_comp = comp;
1129 	}
1130 #ifdef MLX5_PMD_SOFT_COUNTERS
1131 	/* Increment sent packets counter. */
1132 	txq->stats.opackets += i;
1133 #endif
1134 	/* Ring QP doorbell. */
1135 	if (mpw.state == MLX5_MPW_INL_STATE_OPENED)
1136 		mlx5_mpw_inline_close(txq, &mpw);
1137 	else if (mpw.state == MLX5_MPW_STATE_OPENED)
1138 		mlx5_mpw_close(txq, &mpw);
1139 	mlx5_tx_dbrec(txq, mpw.wqe);
1140 	txq->elts_head = elts_head;
1141 	return i;
1142 }
1143 
1144 /**
1145  * Open an Enhanced MPW session.
1146  *
1147  * @param txq
1148  *   Pointer to TX queue structure.
1149  * @param mpw
1150  *   Pointer to MPW session structure.
1151  * @param length
1152  *   Packet length.
1153  */
1154 static inline void
1155 mlx5_empw_new(struct txq *txq, struct mlx5_mpw *mpw, int padding)
1156 {
1157 	uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
1158 
1159 	mpw->state = MLX5_MPW_ENHANCED_STATE_OPENED;
1160 	mpw->pkts_n = 0;
1161 	mpw->total_len = sizeof(struct mlx5_wqe);
1162 	mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
1163 	mpw->wqe->ctrl[0] = htonl((MLX5_OPC_MOD_ENHANCED_MPSW << 24) |
1164 				  (txq->wqe_ci << 8) |
1165 				  MLX5_OPCODE_ENHANCED_MPSW);
1166 	mpw->wqe->ctrl[2] = 0;
1167 	mpw->wqe->ctrl[3] = 0;
1168 	memset((void *)(uintptr_t)&mpw->wqe->eseg, 0, MLX5_WQE_DWORD_SIZE);
1169 	if (unlikely(padding)) {
1170 		uintptr_t addr = (uintptr_t)(mpw->wqe + 1);
1171 
1172 		/* Pad the first 2 DWORDs with zero-length inline header. */
1173 		*(volatile uint32_t *)addr = htonl(MLX5_INLINE_SEG);
1174 		*(volatile uint32_t *)(addr + MLX5_WQE_DWORD_SIZE) =
1175 			htonl(MLX5_INLINE_SEG);
1176 		mpw->total_len += 2 * MLX5_WQE_DWORD_SIZE;
1177 		/* Start from the next WQEBB. */
1178 		mpw->data.raw = (volatile void *)(tx_mlx5_wqe(txq, idx + 1));
1179 	} else {
1180 		mpw->data.raw = (volatile void *)(mpw->wqe + 1);
1181 	}
1182 }
1183 
1184 /**
1185  * Close an Enhanced MPW session.
1186  *
1187  * @param txq
1188  *   Pointer to TX queue structure.
1189  * @param mpw
1190  *   Pointer to MPW session structure.
1191  *
1192  * @return
1193  *   Number of consumed WQEs.
1194  */
1195 static inline uint16_t
1196 mlx5_empw_close(struct txq *txq, struct mlx5_mpw *mpw)
1197 {
1198 	uint16_t ret;
1199 
1200 	/* Store size in multiple of 16 bytes. Control and Ethernet segments
1201 	 * count as 2.
1202 	 */
1203 	mpw->wqe->ctrl[1] = htonl(txq->qp_num_8s | MLX5_WQE_DS(mpw->total_len));
1204 	mpw->state = MLX5_MPW_STATE_CLOSED;
1205 	ret = (mpw->total_len + (MLX5_WQE_SIZE - 1)) / MLX5_WQE_SIZE;
1206 	txq->wqe_ci += ret;
1207 	return ret;
1208 }
1209 
1210 /**
1211  * DPDK callback for TX with Enhanced MPW support.
1212  *
1213  * @param dpdk_txq
1214  *   Generic pointer to TX queue structure.
1215  * @param[in] pkts
1216  *   Packets to transmit.
1217  * @param pkts_n
1218  *   Number of packets in array.
1219  *
1220  * @return
1221  *   Number of packets successfully transmitted (<= pkts_n).
1222  */
1223 uint16_t
1224 mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1225 {
1226 	struct txq *txq = (struct txq *)dpdk_txq;
1227 	uint16_t elts_head = txq->elts_head;
1228 	const uint16_t elts_n = 1 << txq->elts_n;
1229 	const uint16_t elts_m = elts_n - 1;
1230 	unsigned int i = 0;
1231 	unsigned int j = 0;
1232 	uint16_t max_elts;
1233 	uint16_t max_wqe;
1234 	unsigned int max_inline = txq->max_inline * RTE_CACHE_LINE_SIZE;
1235 	unsigned int mpw_room = 0;
1236 	unsigned int inl_pad = 0;
1237 	uint32_t inl_hdr;
1238 	struct mlx5_mpw mpw = {
1239 		.state = MLX5_MPW_STATE_CLOSED,
1240 	};
1241 
1242 	if (unlikely(!pkts_n))
1243 		return 0;
1244 	/* Start processing. */
1245 	mlx5_tx_complete(txq);
1246 	max_elts = (elts_n - (elts_head - txq->elts_tail));
1247 	/* A CQE slot must always be available. */
1248 	assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
1249 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
1250 	if (unlikely(!max_wqe))
1251 		return 0;
1252 	do {
1253 		struct rte_mbuf *buf = *(pkts++);
1254 		uintptr_t addr;
1255 		uint64_t naddr;
1256 		unsigned int n;
1257 		unsigned int do_inline = 0; /* Whether inline is possible. */
1258 		uint32_t length;
1259 		unsigned int segs_n = buf->nb_segs;
1260 		uint32_t cs_flags = 0;
1261 
1262 		/*
1263 		 * Make sure there is enough room to store this packet and
1264 		 * that one ring entry remains unused.
1265 		 */
1266 		assert(segs_n);
1267 		if (max_elts - j < segs_n)
1268 			break;
1269 		/* Do not bother with large packets MPW cannot handle. */
1270 		if (segs_n > MLX5_MPW_DSEG_MAX)
1271 			break;
1272 		/* Should we enable HW CKSUM offload. */
1273 		if (buf->ol_flags &
1274 		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
1275 			cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
1276 		/* Retrieve packet information. */
1277 		length = PKT_LEN(buf);
1278 		/* Start new session if:
1279 		 * - multi-segment packet
1280 		 * - no space left even for a dseg
1281 		 * - next packet can be inlined with a new WQE
1282 		 * - cs_flag differs
1283 		 * It can't be MLX5_MPW_STATE_OPENED as always have a single
1284 		 * segmented packet.
1285 		 */
1286 		if (mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED) {
1287 			if ((segs_n != 1) ||
1288 			    (inl_pad + sizeof(struct mlx5_wqe_data_seg) >
1289 			      mpw_room) ||
1290 			    (length <= txq->inline_max_packet_sz &&
1291 			     inl_pad + sizeof(inl_hdr) + length >
1292 			      mpw_room) ||
1293 			    (mpw.wqe->eseg.cs_flags != cs_flags))
1294 				max_wqe -= mlx5_empw_close(txq, &mpw);
1295 		}
1296 		if (unlikely(mpw.state == MLX5_MPW_STATE_CLOSED)) {
1297 			if (unlikely(segs_n != 1)) {
1298 				/* Fall back to legacy MPW.
1299 				 * A MPW session consumes 2 WQEs at most to
1300 				 * include MLX5_MPW_DSEG_MAX pointers.
1301 				 */
1302 				if (unlikely(max_wqe < 2))
1303 					break;
1304 				mlx5_mpw_new(txq, &mpw, length);
1305 			} else {
1306 				/* In Enhanced MPW, inline as much as the budget
1307 				 * is allowed. The remaining space is to be
1308 				 * filled with dsegs. If the title WQEBB isn't
1309 				 * padded, it will have 2 dsegs there.
1310 				 */
1311 				mpw_room = RTE_MIN(MLX5_WQE_SIZE_MAX,
1312 					    (max_inline ? max_inline :
1313 					     pkts_n * MLX5_WQE_DWORD_SIZE) +
1314 					    MLX5_WQE_SIZE);
1315 				if (unlikely(max_wqe * MLX5_WQE_SIZE <
1316 					      mpw_room))
1317 					break;
1318 				/* Don't pad the title WQEBB to not waste WQ. */
1319 				mlx5_empw_new(txq, &mpw, 0);
1320 				mpw_room -= mpw.total_len;
1321 				inl_pad = 0;
1322 				do_inline =
1323 					length <= txq->inline_max_packet_sz &&
1324 					sizeof(inl_hdr) + length <= mpw_room &&
1325 					!txq->mpw_hdr_dseg;
1326 			}
1327 			mpw.wqe->eseg.cs_flags = cs_flags;
1328 		} else {
1329 			/* Evaluate whether the next packet can be inlined.
1330 			 * Inlininig is possible when:
1331 			 * - length is less than configured value
1332 			 * - length fits for remaining space
1333 			 * - not required to fill the title WQEBB with dsegs
1334 			 */
1335 			do_inline =
1336 				length <= txq->inline_max_packet_sz &&
1337 				inl_pad + sizeof(inl_hdr) + length <=
1338 				 mpw_room &&
1339 				(!txq->mpw_hdr_dseg ||
1340 				 mpw.total_len >= MLX5_WQE_SIZE);
1341 		}
1342 		/* Multi-segment packets must be alone in their MPW. */
1343 		assert((segs_n == 1) || (mpw.pkts_n == 0));
1344 		if (unlikely(mpw.state == MLX5_MPW_STATE_OPENED)) {
1345 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1346 			length = 0;
1347 #endif
1348 			do {
1349 				volatile struct mlx5_wqe_data_seg *dseg;
1350 
1351 				assert(buf);
1352 				(*txq->elts)[elts_head++ & elts_m] = buf;
1353 				dseg = mpw.data.dseg[mpw.pkts_n];
1354 				addr = rte_pktmbuf_mtod(buf, uintptr_t);
1355 				*dseg = (struct mlx5_wqe_data_seg){
1356 					.byte_count = htonl(DATA_LEN(buf)),
1357 					.lkey = mlx5_tx_mb2mr(txq, buf),
1358 					.addr = htonll(addr),
1359 				};
1360 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1361 				length += DATA_LEN(buf);
1362 #endif
1363 				buf = buf->next;
1364 				++j;
1365 				++mpw.pkts_n;
1366 			} while (--segs_n);
1367 			/* A multi-segmented packet takes one MPW session.
1368 			 * TODO: Pack more multi-segmented packets if possible.
1369 			 */
1370 			mlx5_mpw_close(txq, &mpw);
1371 			if (mpw.pkts_n < 3)
1372 				max_wqe--;
1373 			else
1374 				max_wqe -= 2;
1375 		} else if (do_inline) {
1376 			/* Inline packet into WQE. */
1377 			unsigned int max;
1378 
1379 			assert(mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED);
1380 			assert(length == DATA_LEN(buf));
1381 			inl_hdr = htonl(length | MLX5_INLINE_SEG);
1382 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
1383 			mpw.data.raw = (volatile void *)
1384 				((uintptr_t)mpw.data.raw + inl_pad);
1385 			max = tx_mlx5_wq_tailroom(txq,
1386 					(void *)(uintptr_t)mpw.data.raw);
1387 			/* Copy inline header. */
1388 			mpw.data.raw = (volatile void *)
1389 				mlx5_copy_to_wq(
1390 					  (void *)(uintptr_t)mpw.data.raw,
1391 					  &inl_hdr,
1392 					  sizeof(inl_hdr),
1393 					  (void *)(uintptr_t)txq->wqes,
1394 					  max);
1395 			max = tx_mlx5_wq_tailroom(txq,
1396 					(void *)(uintptr_t)mpw.data.raw);
1397 			/* Copy packet data. */
1398 			mpw.data.raw = (volatile void *)
1399 				mlx5_copy_to_wq(
1400 					  (void *)(uintptr_t)mpw.data.raw,
1401 					  (void *)addr,
1402 					  length,
1403 					  (void *)(uintptr_t)txq->wqes,
1404 					  max);
1405 			++mpw.pkts_n;
1406 			mpw.total_len += (inl_pad + sizeof(inl_hdr) + length);
1407 			/* No need to get completion as the entire packet is
1408 			 * copied to WQ. Free the buf right away.
1409 			 */
1410 			rte_pktmbuf_free_seg(buf);
1411 			mpw_room -= (inl_pad + sizeof(inl_hdr) + length);
1412 			/* Add pad in the next packet if any. */
1413 			inl_pad = (((uintptr_t)mpw.data.raw +
1414 					(MLX5_WQE_DWORD_SIZE - 1)) &
1415 					~(MLX5_WQE_DWORD_SIZE - 1)) -
1416 				  (uintptr_t)mpw.data.raw;
1417 		} else {
1418 			/* No inline. Load a dseg of packet pointer. */
1419 			volatile rte_v128u32_t *dseg;
1420 
1421 			assert(mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED);
1422 			assert((inl_pad + sizeof(*dseg)) <= mpw_room);
1423 			assert(length == DATA_LEN(buf));
1424 			if (!tx_mlx5_wq_tailroom(txq,
1425 					(void *)((uintptr_t)mpw.data.raw
1426 						+ inl_pad)))
1427 				dseg = (volatile void *)txq->wqes;
1428 			else
1429 				dseg = (volatile void *)
1430 					((uintptr_t)mpw.data.raw +
1431 					 inl_pad);
1432 			(*txq->elts)[elts_head++ & elts_m] = buf;
1433 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
1434 			for (n = 0; n * RTE_CACHE_LINE_SIZE < length; n++)
1435 				rte_prefetch2((void *)(addr +
1436 						n * RTE_CACHE_LINE_SIZE));
1437 			naddr = htonll(addr);
1438 			*dseg = (rte_v128u32_t) {
1439 				htonl(length),
1440 				mlx5_tx_mb2mr(txq, buf),
1441 				naddr,
1442 				naddr >> 32,
1443 			};
1444 			mpw.data.raw = (volatile void *)(dseg + 1);
1445 			mpw.total_len += (inl_pad + sizeof(*dseg));
1446 			++j;
1447 			++mpw.pkts_n;
1448 			mpw_room -= (inl_pad + sizeof(*dseg));
1449 			inl_pad = 0;
1450 		}
1451 #ifdef MLX5_PMD_SOFT_COUNTERS
1452 		/* Increment sent bytes counter. */
1453 		txq->stats.obytes += length;
1454 #endif
1455 		++i;
1456 	} while (i < pkts_n);
1457 	/* Take a shortcut if nothing must be sent. */
1458 	if (unlikely(i == 0))
1459 		return 0;
1460 	/* Check whether completion threshold has been reached. */
1461 	if (txq->elts_comp + j >= MLX5_TX_COMP_THRESH ||
1462 			(uint16_t)(txq->wqe_ci - txq->mpw_comp) >=
1463 			 (1 << txq->wqe_n) / MLX5_TX_COMP_THRESH_INLINE_DIV) {
1464 		volatile struct mlx5_wqe *wqe = mpw.wqe;
1465 
1466 		/* Request completion on last WQE. */
1467 		wqe->ctrl[2] = htonl(8);
1468 		/* Save elts_head in unused "immediate" field of WQE. */
1469 		wqe->ctrl[3] = elts_head;
1470 		txq->elts_comp = 0;
1471 		txq->mpw_comp = txq->wqe_ci;
1472 		txq->cq_pi++;
1473 	} else {
1474 		txq->elts_comp += j;
1475 	}
1476 #ifdef MLX5_PMD_SOFT_COUNTERS
1477 	/* Increment sent packets counter. */
1478 	txq->stats.opackets += i;
1479 #endif
1480 	if (mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED)
1481 		mlx5_empw_close(txq, &mpw);
1482 	else if (mpw.state == MLX5_MPW_STATE_OPENED)
1483 		mlx5_mpw_close(txq, &mpw);
1484 	/* Ring QP doorbell. */
1485 	mlx5_tx_dbrec(txq, mpw.wqe);
1486 	txq->elts_head = elts_head;
1487 	return i;
1488 }
1489 
1490 /**
1491  * Translate RX completion flags to packet type.
1492  *
1493  * @param[in] cqe
1494  *   Pointer to CQE.
1495  *
1496  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
1497  *
1498  * @return
1499  *   Packet type for struct rte_mbuf.
1500  */
1501 static inline uint32_t
1502 rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe)
1503 {
1504 	uint32_t pkt_type;
1505 	uint16_t flags = ntohs(cqe->hdr_type_etc);
1506 
1507 	if (cqe->pkt_info & MLX5_CQE_RX_TUNNEL_PACKET) {
1508 		pkt_type =
1509 			TRANSPOSE(flags,
1510 				  MLX5_CQE_RX_IPV4_PACKET,
1511 				  RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN) |
1512 			TRANSPOSE(flags,
1513 				  MLX5_CQE_RX_IPV6_PACKET,
1514 				  RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN);
1515 		pkt_type |= ((cqe->pkt_info & MLX5_CQE_RX_OUTER_PACKET) ?
1516 			     RTE_PTYPE_L3_IPV6_EXT_UNKNOWN :
1517 			     RTE_PTYPE_L3_IPV4_EXT_UNKNOWN);
1518 	} else {
1519 		pkt_type =
1520 			TRANSPOSE(flags,
1521 				  MLX5_CQE_L3_HDR_TYPE_IPV6,
1522 				  RTE_PTYPE_L3_IPV6_EXT_UNKNOWN) |
1523 			TRANSPOSE(flags,
1524 				  MLX5_CQE_L3_HDR_TYPE_IPV4,
1525 				  RTE_PTYPE_L3_IPV4_EXT_UNKNOWN);
1526 	}
1527 	return pkt_type;
1528 }
1529 
1530 /**
1531  * Get size of the next packet for a given CQE. For compressed CQEs, the
1532  * consumer index is updated only once all packets of the current one have
1533  * been processed.
1534  *
1535  * @param rxq
1536  *   Pointer to RX queue.
1537  * @param cqe
1538  *   CQE to process.
1539  * @param[out] rss_hash
1540  *   Packet RSS Hash result.
1541  *
1542  * @return
1543  *   Packet size in bytes (0 if there is none), -1 in case of completion
1544  *   with error.
1545  */
1546 static inline int
1547 mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe,
1548 		 uint16_t cqe_cnt, uint32_t *rss_hash)
1549 {
1550 	struct rxq_zip *zip = &rxq->zip;
1551 	uint16_t cqe_n = cqe_cnt + 1;
1552 	int len = 0;
1553 	uint16_t idx, end;
1554 
1555 	/* Process compressed data in the CQE and mini arrays. */
1556 	if (zip->ai) {
1557 		volatile struct mlx5_mini_cqe8 (*mc)[8] =
1558 			(volatile struct mlx5_mini_cqe8 (*)[8])
1559 			(uintptr_t)(&(*rxq->cqes)[zip->ca & cqe_cnt]);
1560 
1561 		len = ntohl((*mc)[zip->ai & 7].byte_cnt);
1562 		*rss_hash = ntohl((*mc)[zip->ai & 7].rx_hash_result);
1563 		if ((++zip->ai & 7) == 0) {
1564 			/* Invalidate consumed CQEs */
1565 			idx = zip->ca;
1566 			end = zip->na;
1567 			while (idx != end) {
1568 				(*rxq->cqes)[idx & cqe_cnt].op_own =
1569 					MLX5_CQE_INVALIDATE;
1570 				++idx;
1571 			}
1572 			/*
1573 			 * Increment consumer index to skip the number of
1574 			 * CQEs consumed. Hardware leaves holes in the CQ
1575 			 * ring for software use.
1576 			 */
1577 			zip->ca = zip->na;
1578 			zip->na += 8;
1579 		}
1580 		if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1581 			/* Invalidate the rest */
1582 			idx = zip->ca;
1583 			end = zip->cq_ci;
1584 
1585 			while (idx != end) {
1586 				(*rxq->cqes)[idx & cqe_cnt].op_own =
1587 					MLX5_CQE_INVALIDATE;
1588 				++idx;
1589 			}
1590 			rxq->cq_ci = zip->cq_ci;
1591 			zip->ai = 0;
1592 		}
1593 	/* No compressed data, get next CQE and verify if it is compressed. */
1594 	} else {
1595 		int ret;
1596 		int8_t op_own;
1597 
1598 		ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1599 		if (unlikely(ret == 1))
1600 			return 0;
1601 		++rxq->cq_ci;
1602 		op_own = cqe->op_own;
1603 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1604 			volatile struct mlx5_mini_cqe8 (*mc)[8] =
1605 				(volatile struct mlx5_mini_cqe8 (*)[8])
1606 				(uintptr_t)(&(*rxq->cqes)[rxq->cq_ci &
1607 							  cqe_cnt]);
1608 
1609 			/* Fix endianness. */
1610 			zip->cqe_cnt = ntohl(cqe->byte_cnt);
1611 			/*
1612 			 * Current mini array position is the one returned by
1613 			 * check_cqe64().
1614 			 *
1615 			 * If completion comprises several mini arrays, as a
1616 			 * special case the second one is located 7 CQEs after
1617 			 * the initial CQE instead of 8 for subsequent ones.
1618 			 */
1619 			zip->ca = rxq->cq_ci;
1620 			zip->na = zip->ca + 7;
1621 			/* Compute the next non compressed CQE. */
1622 			--rxq->cq_ci;
1623 			zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1624 			/* Get packet size to return. */
1625 			len = ntohl((*mc)[0].byte_cnt);
1626 			*rss_hash = ntohl((*mc)[0].rx_hash_result);
1627 			zip->ai = 1;
1628 			/* Prefetch all the entries to be invalidated */
1629 			idx = zip->ca;
1630 			end = zip->cq_ci;
1631 			while (idx != end) {
1632 				rte_prefetch0(&(*rxq->cqes)[(idx) & cqe_cnt]);
1633 				++idx;
1634 			}
1635 		} else {
1636 			len = ntohl(cqe->byte_cnt);
1637 			*rss_hash = ntohl(cqe->rx_hash_res);
1638 		}
1639 		/* Error while receiving packet. */
1640 		if (unlikely(MLX5_CQE_OPCODE(op_own) == MLX5_CQE_RESP_ERR))
1641 			return -1;
1642 	}
1643 	return len;
1644 }
1645 
1646 /**
1647  * Translate RX completion flags to offload flags.
1648  *
1649  * @param[in] rxq
1650  *   Pointer to RX queue structure.
1651  * @param[in] cqe
1652  *   Pointer to CQE.
1653  *
1654  * @return
1655  *   Offload flags (ol_flags) for struct rte_mbuf.
1656  */
1657 static inline uint32_t
1658 rxq_cq_to_ol_flags(struct rxq *rxq, volatile struct mlx5_cqe *cqe)
1659 {
1660 	uint32_t ol_flags = 0;
1661 	uint16_t flags = ntohs(cqe->hdr_type_etc);
1662 
1663 	ol_flags =
1664 		TRANSPOSE(flags,
1665 			  MLX5_CQE_RX_L3_HDR_VALID,
1666 			  PKT_RX_IP_CKSUM_GOOD) |
1667 		TRANSPOSE(flags,
1668 			  MLX5_CQE_RX_L4_HDR_VALID,
1669 			  PKT_RX_L4_CKSUM_GOOD);
1670 	if ((cqe->pkt_info & MLX5_CQE_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
1671 		ol_flags |=
1672 			TRANSPOSE(flags,
1673 				  MLX5_CQE_RX_L3_HDR_VALID,
1674 				  PKT_RX_IP_CKSUM_GOOD) |
1675 			TRANSPOSE(flags,
1676 				  MLX5_CQE_RX_L4_HDR_VALID,
1677 				  PKT_RX_L4_CKSUM_GOOD);
1678 	return ol_flags;
1679 }
1680 
1681 /**
1682  * DPDK callback for RX.
1683  *
1684  * @param dpdk_rxq
1685  *   Generic pointer to RX queue structure.
1686  * @param[out] pkts
1687  *   Array to store received packets.
1688  * @param pkts_n
1689  *   Maximum number of packets in array.
1690  *
1691  * @return
1692  *   Number of packets successfully received (<= pkts_n).
1693  */
1694 uint16_t
1695 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1696 {
1697 	struct rxq *rxq = dpdk_rxq;
1698 	const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1699 	const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1700 	const unsigned int sges_n = rxq->sges_n;
1701 	struct rte_mbuf *pkt = NULL;
1702 	struct rte_mbuf *seg = NULL;
1703 	volatile struct mlx5_cqe *cqe =
1704 		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1705 	unsigned int i = 0;
1706 	unsigned int rq_ci = rxq->rq_ci << sges_n;
1707 	int len = 0; /* keep its value across iterations. */
1708 
1709 	while (pkts_n) {
1710 		unsigned int idx = rq_ci & wqe_cnt;
1711 		volatile struct mlx5_wqe_data_seg *wqe = &(*rxq->wqes)[idx];
1712 		struct rte_mbuf *rep = (*rxq->elts)[idx];
1713 		uint32_t rss_hash_res = 0;
1714 
1715 		if (pkt)
1716 			NEXT(seg) = rep;
1717 		seg = rep;
1718 		rte_prefetch0(seg);
1719 		rte_prefetch0(cqe);
1720 		rte_prefetch0(wqe);
1721 		rep = rte_mbuf_raw_alloc(rxq->mp);
1722 		if (unlikely(rep == NULL)) {
1723 			++rxq->stats.rx_nombuf;
1724 			if (!pkt) {
1725 				/*
1726 				 * no buffers before we even started,
1727 				 * bail out silently.
1728 				 */
1729 				break;
1730 			}
1731 			while (pkt != seg) {
1732 				assert(pkt != (*rxq->elts)[idx]);
1733 				rep = NEXT(pkt);
1734 				NEXT(pkt) = NULL;
1735 				NB_SEGS(pkt) = 1;
1736 				rte_mbuf_raw_free(pkt);
1737 				pkt = rep;
1738 			}
1739 			break;
1740 		}
1741 		if (!pkt) {
1742 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1743 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt,
1744 					       &rss_hash_res);
1745 			if (!len) {
1746 				rte_mbuf_raw_free(rep);
1747 				break;
1748 			}
1749 			if (unlikely(len == -1)) {
1750 				/* RX error, packet is likely too large. */
1751 				rte_mbuf_raw_free(rep);
1752 				++rxq->stats.idropped;
1753 				goto skip;
1754 			}
1755 			pkt = seg;
1756 			assert(len >= (rxq->crc_present << 2));
1757 			/* Update packet information. */
1758 			pkt->packet_type = 0;
1759 			pkt->ol_flags = 0;
1760 			if (rss_hash_res && rxq->rss_hash) {
1761 				pkt->hash.rss = rss_hash_res;
1762 				pkt->ol_flags = PKT_RX_RSS_HASH;
1763 			}
1764 			if (rxq->mark &&
1765 			    MLX5_FLOW_MARK_IS_VALID(cqe->sop_drop_qpn)) {
1766 				pkt->ol_flags |= PKT_RX_FDIR;
1767 				if (cqe->sop_drop_qpn !=
1768 				    htonl(MLX5_FLOW_MARK_DEFAULT)) {
1769 					uint32_t mark = cqe->sop_drop_qpn;
1770 
1771 					pkt->ol_flags |= PKT_RX_FDIR_ID;
1772 					pkt->hash.fdir.hi =
1773 						mlx5_flow_mark_get(mark);
1774 				}
1775 			}
1776 			if (rxq->csum | rxq->csum_l2tun) {
1777 				pkt->packet_type = rxq_cq_to_pkt_type(cqe);
1778 				pkt->ol_flags |= rxq_cq_to_ol_flags(rxq, cqe);
1779 			}
1780 			if (rxq->vlan_strip &&
1781 			    (cqe->hdr_type_etc &
1782 			     htons(MLX5_CQE_VLAN_STRIPPED))) {
1783 				pkt->ol_flags |= PKT_RX_VLAN_PKT |
1784 					PKT_RX_VLAN_STRIPPED;
1785 				pkt->vlan_tci = ntohs(cqe->vlan_info);
1786 			}
1787 			if (rxq->crc_present)
1788 				len -= ETHER_CRC_LEN;
1789 			PKT_LEN(pkt) = len;
1790 		}
1791 		DATA_LEN(rep) = DATA_LEN(seg);
1792 		PKT_LEN(rep) = PKT_LEN(seg);
1793 		SET_DATA_OFF(rep, DATA_OFF(seg));
1794 		PORT(rep) = PORT(seg);
1795 		(*rxq->elts)[idx] = rep;
1796 		/*
1797 		 * Fill NIC descriptor with the new buffer.  The lkey and size
1798 		 * of the buffers are already known, only the buffer address
1799 		 * changes.
1800 		 */
1801 		wqe->addr = htonll(rte_pktmbuf_mtod(rep, uintptr_t));
1802 		if (len > DATA_LEN(seg)) {
1803 			len -= DATA_LEN(seg);
1804 			++NB_SEGS(pkt);
1805 			++rq_ci;
1806 			continue;
1807 		}
1808 		DATA_LEN(seg) = len;
1809 #ifdef MLX5_PMD_SOFT_COUNTERS
1810 		/* Increment bytes counter. */
1811 		rxq->stats.ibytes += PKT_LEN(pkt);
1812 #endif
1813 		/* Return packet. */
1814 		*(pkts++) = pkt;
1815 		pkt = NULL;
1816 		--pkts_n;
1817 		++i;
1818 skip:
1819 		/* Align consumer index to the next stride. */
1820 		rq_ci >>= sges_n;
1821 		++rq_ci;
1822 		rq_ci <<= sges_n;
1823 	}
1824 	if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1825 		return 0;
1826 	/* Update the consumer index. */
1827 	rxq->rq_ci = rq_ci >> sges_n;
1828 	rte_wmb();
1829 	*rxq->cq_db = htonl(rxq->cq_ci);
1830 	rte_wmb();
1831 	*rxq->rq_db = htonl(rxq->rq_ci);
1832 #ifdef MLX5_PMD_SOFT_COUNTERS
1833 	/* Increment packets counter. */
1834 	rxq->stats.ipackets += i;
1835 #endif
1836 	return i;
1837 }
1838 
1839 /**
1840  * Dummy DPDK callback for TX.
1841  *
1842  * This function is used to temporarily replace the real callback during
1843  * unsafe control operations on the queue, or in case of error.
1844  *
1845  * @param dpdk_txq
1846  *   Generic pointer to TX queue structure.
1847  * @param[in] pkts
1848  *   Packets to transmit.
1849  * @param pkts_n
1850  *   Number of packets in array.
1851  *
1852  * @return
1853  *   Number of packets successfully transmitted (<= pkts_n).
1854  */
1855 uint16_t
1856 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1857 {
1858 	(void)dpdk_txq;
1859 	(void)pkts;
1860 	(void)pkts_n;
1861 	return 0;
1862 }
1863 
1864 /**
1865  * Dummy DPDK callback for RX.
1866  *
1867  * This function is used to temporarily replace the real callback during
1868  * unsafe control operations on the queue, or in case of error.
1869  *
1870  * @param dpdk_rxq
1871  *   Generic pointer to RX queue structure.
1872  * @param[out] pkts
1873  *   Array to store received packets.
1874  * @param pkts_n
1875  *   Maximum number of packets in array.
1876  *
1877  * @return
1878  *   Number of packets successfully received (<= pkts_n).
1879  */
1880 uint16_t
1881 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1882 {
1883 	(void)dpdk_rxq;
1884 	(void)pkts;
1885 	(void)pkts_n;
1886 	return 0;
1887 }
1888 
1889 /*
1890  * Vectorized Rx/Tx routines are not compiled in when required vector
1891  * instructions are not supported on a target architecture. The following null
1892  * stubs are needed for linkage when those are not included outside of this file
1893  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
1894  */
1895 
1896 uint16_t __attribute__((weak))
1897 mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1898 {
1899 	(void)dpdk_txq;
1900 	(void)pkts;
1901 	(void)pkts_n;
1902 	return 0;
1903 }
1904 
1905 uint16_t __attribute__((weak))
1906 mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1907 {
1908 	(void)dpdk_txq;
1909 	(void)pkts;
1910 	(void)pkts_n;
1911 	return 0;
1912 }
1913 
1914 uint16_t __attribute__((weak))
1915 mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1916 {
1917 	(void)dpdk_rxq;
1918 	(void)pkts;
1919 	(void)pkts_n;
1920 	return 0;
1921 }
1922 
1923 int __attribute__((weak))
1924 priv_check_raw_vec_tx_support(struct priv *priv)
1925 {
1926 	(void)priv;
1927 	return -ENOTSUP;
1928 }
1929 
1930 int __attribute__((weak))
1931 priv_check_vec_tx_support(struct priv *priv)
1932 {
1933 	(void)priv;
1934 	return -ENOTSUP;
1935 }
1936 
1937 int __attribute__((weak))
1938 rxq_check_vec_support(struct rxq *rxq)
1939 {
1940 	(void)rxq;
1941 	return -ENOTSUP;
1942 }
1943 
1944 int __attribute__((weak))
1945 priv_check_vec_rx_support(struct priv *priv)
1946 {
1947 	(void)priv;
1948 	return -ENOTSUP;
1949 }
1950 
1951 void __attribute__((weak))
1952 priv_prep_vec_rx_function(struct priv *priv)
1953 {
1954 	(void)priv;
1955 }
1956