xref: /dpdk/drivers/net/mlx5/mlx5_rx.c (revision b53d106d34b5c638f5a2cbdfee0da5bd42d4383f)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2021 6WIND S.A.
3  * Copyright 2021 Mellanox Technologies, Ltd
4  */
5 
6 #include <stdint.h>
7 #include <string.h>
8 #include <stdlib.h>
9 
10 #include <rte_mbuf.h>
11 #include <rte_mempool.h>
12 #include <rte_prefetch.h>
13 #include <rte_common.h>
14 #include <rte_branch_prediction.h>
15 #include <rte_ether.h>
16 #include <rte_cycles.h>
17 #include <rte_flow.h>
18 
19 #include <mlx5_prm.h>
20 #include <mlx5_common.h>
21 #include <mlx5_common_mr.h>
22 
23 #include "mlx5_autoconf.h"
24 #include "mlx5_defs.h"
25 #include "mlx5.h"
26 #include "mlx5_utils.h"
27 #include "mlx5_rxtx.h"
28 #include "mlx5_rx.h"
29 
30 
31 static __rte_always_inline uint32_t
32 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
33 		   volatile struct mlx5_mini_cqe8 *mcqe);
34 
35 static __rte_always_inline int
36 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
37 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe);
38 
39 static __rte_always_inline uint32_t
40 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe);
41 
42 static __rte_always_inline void
43 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
44 	       volatile struct mlx5_cqe *cqe,
45 	       volatile struct mlx5_mini_cqe8 *mcqe);
46 
47 static inline void
48 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *__rte_restrict tcp,
49 			volatile struct mlx5_cqe *__rte_restrict cqe,
50 			uint32_t phcsum, uint8_t l4_type);
51 
52 static inline void
53 mlx5_lro_update_hdr(uint8_t *__rte_restrict padd,
54 		    volatile struct mlx5_cqe *__rte_restrict cqe,
55 		    volatile struct mlx5_mini_cqe8 *mcqe,
56 		    struct mlx5_rxq_data *rxq, uint32_t len);
57 
58 
59 /**
60  * Internal function to compute the number of used descriptors in an RX queue.
61  *
62  * @param rxq
63  *   The Rx queue.
64  *
65  * @return
66  *   The number of used Rx descriptor.
67  */
68 static uint32_t
69 rx_queue_count(struct mlx5_rxq_data *rxq)
70 {
71 	struct rxq_zip *zip = &rxq->zip;
72 	volatile struct mlx5_cqe *cqe;
73 	const unsigned int cqe_n = (1 << rxq->cqe_n);
74 	const unsigned int sges_n = (1 << rxq->sges_n);
75 	const unsigned int elts_n = (1 << rxq->elts_n);
76 	const unsigned int strd_n = (1 << rxq->strd_num_n);
77 	const unsigned int cqe_cnt = cqe_n - 1;
78 	unsigned int cq_ci, used;
79 
80 	/* if we are processing a compressed cqe */
81 	if (zip->ai) {
82 		used = zip->cqe_cnt - zip->ai;
83 		cq_ci = zip->cq_ci;
84 	} else {
85 		used = 0;
86 		cq_ci = rxq->cq_ci;
87 	}
88 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
89 	while (check_cqe(cqe, cqe_n, cq_ci) != MLX5_CQE_STATUS_HW_OWN) {
90 		int8_t op_own;
91 		unsigned int n;
92 
93 		op_own = cqe->op_own;
94 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
95 			n = rte_be_to_cpu_32(cqe->byte_cnt);
96 		else
97 			n = 1;
98 		cq_ci += n;
99 		used += n;
100 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
101 	}
102 	used = RTE_MIN(used * sges_n, elts_n * strd_n);
103 	return used;
104 }
105 
106 /**
107  * DPDK callback to check the status of a Rx descriptor.
108  *
109  * @param rx_queue
110  *   The Rx queue.
111  * @param[in] offset
112  *   The index of the descriptor in the ring.
113  *
114  * @return
115  *   The status of the Rx descriptor.
116  */
117 int
118 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
119 {
120 	struct mlx5_rxq_data *rxq = rx_queue;
121 
122 	if (offset >= (1 << rxq->cqe_n)) {
123 		rte_errno = EINVAL;
124 		return -rte_errno;
125 	}
126 	if (offset < rx_queue_count(rxq))
127 		return RTE_ETH_RX_DESC_DONE;
128 	return RTE_ETH_RX_DESC_AVAIL;
129 }
130 
131 /**
132  * DPDK callback to get the RX queue information.
133  *
134  * @param dev
135  *   Pointer to the device structure.
136  *
137  * @param rx_queue_id
138  *   Rx queue identificator.
139  *
140  * @param qinfo
141  *   Pointer to the RX queue information structure.
142  *
143  * @return
144  *   None.
145  */
146 
147 void
148 mlx5_rxq_info_get(struct rte_eth_dev *dev, uint16_t rx_queue_id,
149 		  struct rte_eth_rxq_info *qinfo)
150 {
151 	struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, rx_queue_id);
152 	struct mlx5_rxq_data *rxq = mlx5_rxq_data_get(dev, rx_queue_id);
153 
154 	if (!rxq)
155 		return;
156 	qinfo->mp = mlx5_rxq_mprq_enabled(rxq) ?
157 					rxq->mprq_mp : rxq->mp;
158 	qinfo->conf.rx_thresh.pthresh = 0;
159 	qinfo->conf.rx_thresh.hthresh = 0;
160 	qinfo->conf.rx_thresh.wthresh = 0;
161 	qinfo->conf.rx_free_thresh = rxq->rq_repl_thresh;
162 	qinfo->conf.rx_drop_en = 1;
163 	if (rxq_ctrl == NULL || rxq_ctrl->obj == NULL)
164 		qinfo->conf.rx_deferred_start = 0;
165 	else
166 		qinfo->conf.rx_deferred_start = 1;
167 	qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
168 	qinfo->scattered_rx = dev->data->scattered_rx;
169 	qinfo->nb_desc = mlx5_rxq_mprq_enabled(rxq) ?
170 		(1 << rxq->elts_n) * (1 << rxq->strd_num_n) :
171 		(1 << rxq->elts_n);
172 }
173 
174 /**
175  * DPDK callback to get the RX packet burst mode information.
176  *
177  * @param dev
178  *   Pointer to the device structure.
179  *
180  * @param rx_queue_id
181  *   Rx queue identificatior.
182  *
183  * @param mode
184  *   Pointer to the burts mode information.
185  *
186  * @return
187  *   0 as success, -EINVAL as failure.
188  */
189 int
190 mlx5_rx_burst_mode_get(struct rte_eth_dev *dev,
191 		       uint16_t rx_queue_id __rte_unused,
192 		       struct rte_eth_burst_mode *mode)
193 {
194 	eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
195 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, rx_queue_id);
196 
197 	if (!rxq) {
198 		rte_errno = EINVAL;
199 		return -rte_errno;
200 	}
201 	if (pkt_burst == mlx5_rx_burst) {
202 		snprintf(mode->info, sizeof(mode->info), "%s", "Scalar");
203 	} else if (pkt_burst == mlx5_rx_burst_mprq) {
204 		snprintf(mode->info, sizeof(mode->info), "%s", "Multi-Packet RQ");
205 	} else if (pkt_burst == mlx5_rx_burst_vec) {
206 #if defined RTE_ARCH_X86_64
207 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector SSE");
208 #elif defined RTE_ARCH_ARM64
209 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector Neon");
210 #elif defined RTE_ARCH_PPC_64
211 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector AltiVec");
212 #else
213 		return -EINVAL;
214 #endif
215 	} else if (pkt_burst == mlx5_rx_burst_mprq_vec) {
216 #if defined RTE_ARCH_X86_64
217 		snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector SSE");
218 #elif defined RTE_ARCH_ARM64
219 		snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector Neon");
220 #elif defined RTE_ARCH_PPC_64
221 		snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector AltiVec");
222 #else
223 		return -EINVAL;
224 #endif
225 	} else {
226 		return -EINVAL;
227 	}
228 	return 0;
229 }
230 
231 /**
232  * DPDK callback to get the number of used descriptors in a RX queue.
233  *
234  * @param rx_queue
235  *   The Rx queue pointer.
236  *
237  * @return
238  *   The number of used rx descriptor.
239  *   -EINVAL if the queue is invalid
240  */
241 uint32_t
242 mlx5_rx_queue_count(void *rx_queue)
243 {
244 	struct mlx5_rxq_data *rxq = rx_queue;
245 	struct rte_eth_dev *dev;
246 
247 	if (!rxq) {
248 		rte_errno = EINVAL;
249 		return -rte_errno;
250 	}
251 
252 	dev = &rte_eth_devices[rxq->port_id];
253 
254 	if (dev->rx_pkt_burst == NULL ||
255 	    dev->rx_pkt_burst == removed_rx_burst) {
256 		rte_errno = ENOTSUP;
257 		return -rte_errno;
258 	}
259 
260 	return rx_queue_count(rxq);
261 }
262 
263 #define CLB_VAL_IDX 0
264 #define CLB_MSK_IDX 1
265 static int
266 mlx5_monitor_callback(const uint64_t value,
267 		const uint64_t opaque[RTE_POWER_MONITOR_OPAQUE_SZ])
268 {
269 	const uint64_t m = opaque[CLB_MSK_IDX];
270 	const uint64_t v = opaque[CLB_VAL_IDX];
271 
272 	return (value & m) == v ? -1 : 0;
273 }
274 
275 int mlx5_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc)
276 {
277 	struct mlx5_rxq_data *rxq = rx_queue;
278 	const unsigned int cqe_num = 1 << rxq->cqe_n;
279 	const unsigned int cqe_mask = cqe_num - 1;
280 	const uint16_t idx = rxq->cq_ci & cqe_num;
281 	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask];
282 
283 	if (unlikely(rxq->cqes == NULL)) {
284 		rte_errno = EINVAL;
285 		return -rte_errno;
286 	}
287 	pmc->addr = &cqe->op_own;
288 	pmc->opaque[CLB_VAL_IDX] = !!idx;
289 	pmc->opaque[CLB_MSK_IDX] = MLX5_CQE_OWNER_MASK;
290 	pmc->fn = mlx5_monitor_callback;
291 	pmc->size = sizeof(uint8_t);
292 	return 0;
293 }
294 
295 /**
296  * Translate RX completion flags to packet type.
297  *
298  * @param[in] rxq
299  *   Pointer to RX queue structure.
300  * @param[in] cqe
301  *   Pointer to CQE.
302  *
303  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
304  *
305  * @return
306  *   Packet type for struct rte_mbuf.
307  */
308 static inline uint32_t
309 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
310 				   volatile struct mlx5_mini_cqe8 *mcqe)
311 {
312 	uint8_t idx;
313 	uint8_t ptype;
314 	uint8_t pinfo = (cqe->pkt_info & 0x3) << 6;
315 
316 	/* Get l3/l4 header from mini-CQE in case L3/L4 format*/
317 	if (mcqe == NULL ||
318 	    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
319 		ptype = (cqe->hdr_type_etc & 0xfc00) >> 10;
320 	else
321 		ptype = mcqe->hdr_type >> 2;
322 	/*
323 	 * The index to the array should have:
324 	 * bit[1:0] = l3_hdr_type
325 	 * bit[4:2] = l4_hdr_type
326 	 * bit[5] = ip_frag
327 	 * bit[6] = tunneled
328 	 * bit[7] = outer_l3_type
329 	 */
330 	idx = pinfo | ptype;
331 	return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
332 }
333 
334 /**
335  * Initialize Rx WQ and indexes.
336  *
337  * @param[in] rxq
338  *   Pointer to RX queue structure.
339  */
340 void
341 mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
342 {
343 	const unsigned int wqe_n = 1 << rxq->elts_n;
344 	unsigned int i;
345 
346 	for (i = 0; (i != wqe_n); ++i) {
347 		volatile struct mlx5_wqe_data_seg *scat;
348 		uintptr_t addr;
349 		uint32_t byte_count;
350 		uint32_t lkey;
351 
352 		if (mlx5_rxq_mprq_enabled(rxq)) {
353 			struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[i];
354 
355 			scat = &((volatile struct mlx5_wqe_mprq *)
356 				rxq->wqes)[i].dseg;
357 			addr = (uintptr_t)mlx5_mprq_buf_addr(buf,
358 							 1 << rxq->strd_num_n);
359 			byte_count = (1 << rxq->strd_sz_n) *
360 					(1 << rxq->strd_num_n);
361 			lkey = mlx5_rx_addr2mr(rxq, addr);
362 		} else {
363 			struct rte_mbuf *buf = (*rxq->elts)[i];
364 
365 			scat = &((volatile struct mlx5_wqe_data_seg *)
366 					rxq->wqes)[i];
367 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
368 			byte_count = DATA_LEN(buf);
369 			lkey = mlx5_rx_mb2mr(rxq, buf);
370 		}
371 		/* scat->addr must be able to store a pointer. */
372 		MLX5_ASSERT(sizeof(scat->addr) >= sizeof(uintptr_t));
373 		*scat = (struct mlx5_wqe_data_seg){
374 			.addr = rte_cpu_to_be_64(addr),
375 			.byte_count = rte_cpu_to_be_32(byte_count),
376 			.lkey = lkey,
377 		};
378 	}
379 	rxq->consumed_strd = 0;
380 	rxq->decompressed = 0;
381 	rxq->rq_pi = 0;
382 	rxq->zip = (struct rxq_zip){
383 		.ai = 0,
384 	};
385 	rxq->elts_ci = mlx5_rxq_mprq_enabled(rxq) ?
386 		(wqe_n >> rxq->sges_n) * (1 << rxq->strd_num_n) : 0;
387 	/* Update doorbell counter. */
388 	rxq->rq_ci = wqe_n >> rxq->sges_n;
389 	rte_io_wmb();
390 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
391 }
392 
393 /**
394  * Handle a Rx error.
395  * The function inserts the RQ state to reset when the first error CQE is
396  * shown, then drains the CQ by the caller function loop. When the CQ is empty,
397  * it moves the RQ state to ready and initializes the RQ.
398  * Next CQE identification and error counting are in the caller responsibility.
399  *
400  * @param[in] rxq
401  *   Pointer to RX queue structure.
402  * @param[in] vec
403  *   1 when called from vectorized Rx burst, need to prepare mbufs for the RQ.
404  *   0 when called from non-vectorized Rx burst.
405  *
406  * @return
407  *   -1 in case of recovery error, otherwise the CQE status.
408  */
409 int
410 mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
411 {
412 	const uint16_t cqe_n = 1 << rxq->cqe_n;
413 	const uint16_t cqe_mask = cqe_n - 1;
414 	const uint16_t wqe_n = 1 << rxq->elts_n;
415 	const uint16_t strd_n = 1 << rxq->strd_num_n;
416 	struct mlx5_rxq_ctrl *rxq_ctrl =
417 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
418 	union {
419 		volatile struct mlx5_cqe *cqe;
420 		volatile struct mlx5_err_cqe *err_cqe;
421 	} u = {
422 		.cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask],
423 	};
424 	struct mlx5_mp_arg_queue_state_modify sm;
425 	int ret;
426 
427 	switch (rxq->err_state) {
428 	case MLX5_RXQ_ERR_STATE_NO_ERROR:
429 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_RESET;
430 		/* Fall-through */
431 	case MLX5_RXQ_ERR_STATE_NEED_RESET:
432 		sm.is_wq = 1;
433 		sm.queue_id = rxq->idx;
434 		sm.state = IBV_WQS_RESET;
435 		if (mlx5_queue_state_modify(RXQ_DEV(rxq_ctrl), &sm))
436 			return -1;
437 		if (rxq_ctrl->dump_file_n <
438 		    RXQ_PORT(rxq_ctrl)->config.max_dump_files_num) {
439 			MKSTR(err_str, "Unexpected CQE error syndrome "
440 			      "0x%02x CQN = %u RQN = %u wqe_counter = %u"
441 			      " rq_ci = %u cq_ci = %u", u.err_cqe->syndrome,
442 			      rxq->cqn, rxq_ctrl->wqn,
443 			      rte_be_to_cpu_16(u.err_cqe->wqe_counter),
444 			      rxq->rq_ci << rxq->sges_n, rxq->cq_ci);
445 			MKSTR(name, "dpdk_mlx5_port_%u_rxq_%u_%u",
446 			      rxq->port_id, rxq->idx, (uint32_t)rte_rdtsc());
447 			mlx5_dump_debug_information(name, NULL, err_str, 0);
448 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
449 						    (const void *)((uintptr_t)
450 								    rxq->cqes),
451 						    sizeof(*u.cqe) * cqe_n);
452 			mlx5_dump_debug_information(name, "MLX5 Error RQ:",
453 						    (const void *)((uintptr_t)
454 								    rxq->wqes),
455 						    16 * wqe_n);
456 			rxq_ctrl->dump_file_n++;
457 		}
458 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_READY;
459 		/* Fall-through */
460 	case MLX5_RXQ_ERR_STATE_NEED_READY:
461 		ret = check_cqe(u.cqe, cqe_n, rxq->cq_ci);
462 		if (ret == MLX5_CQE_STATUS_HW_OWN) {
463 			rte_io_wmb();
464 			*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
465 			rte_io_wmb();
466 			/*
467 			 * The RQ consumer index must be zeroed while moving
468 			 * from RESET state to RDY state.
469 			 */
470 			*rxq->rq_db = rte_cpu_to_be_32(0);
471 			rte_io_wmb();
472 			sm.is_wq = 1;
473 			sm.queue_id = rxq->idx;
474 			sm.state = IBV_WQS_RDY;
475 			if (mlx5_queue_state_modify(RXQ_DEV(rxq_ctrl), &sm))
476 				return -1;
477 			if (vec) {
478 				const uint32_t elts_n =
479 					mlx5_rxq_mprq_enabled(rxq) ?
480 					wqe_n * strd_n : wqe_n;
481 				const uint32_t e_mask = elts_n - 1;
482 				uint32_t elts_ci =
483 					mlx5_rxq_mprq_enabled(rxq) ?
484 					rxq->elts_ci : rxq->rq_ci;
485 				uint32_t elt_idx;
486 				struct rte_mbuf **elt;
487 				int i;
488 				unsigned int n = elts_n - (elts_ci -
489 							  rxq->rq_pi);
490 
491 				for (i = 0; i < (int)n; ++i) {
492 					elt_idx = (elts_ci + i) & e_mask;
493 					elt = &(*rxq->elts)[elt_idx];
494 					*elt = rte_mbuf_raw_alloc(rxq->mp);
495 					if (!*elt) {
496 						for (i--; i >= 0; --i) {
497 							elt_idx = (elts_ci +
498 								   i) & elts_n;
499 							elt = &(*rxq->elts)
500 								[elt_idx];
501 							rte_pktmbuf_free_seg
502 								(*elt);
503 						}
504 						return -1;
505 					}
506 				}
507 				for (i = 0; i < (int)elts_n; ++i) {
508 					elt = &(*rxq->elts)[i];
509 					DATA_LEN(*elt) =
510 						(uint16_t)((*elt)->buf_len -
511 						rte_pktmbuf_headroom(*elt));
512 				}
513 				/* Padding with a fake mbuf for vec Rx. */
514 				for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
515 					(*rxq->elts)[elts_n + i] =
516 								&rxq->fake_mbuf;
517 			}
518 			mlx5_rxq_initialize(rxq);
519 			rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
520 		}
521 		return ret;
522 	default:
523 		return -1;
524 	}
525 }
526 
527 /**
528  * Get size of the next packet for a given CQE. For compressed CQEs, the
529  * consumer index is updated only once all packets of the current one have
530  * been processed.
531  *
532  * @param rxq
533  *   Pointer to RX queue.
534  * @param cqe
535  *   CQE to process.
536  * @param[out] mcqe
537  *   Store pointer to mini-CQE if compressed. Otherwise, the pointer is not
538  *   written.
539  *
540  * @return
541  *   0 in case of empty CQE, otherwise the packet size in bytes.
542  */
543 static inline int
544 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
545 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe)
546 {
547 	struct rxq_zip *zip = &rxq->zip;
548 	uint16_t cqe_n = cqe_cnt + 1;
549 	int len;
550 	uint16_t idx, end;
551 
552 	do {
553 		len = 0;
554 		/* Process compressed data in the CQE and mini arrays. */
555 		if (zip->ai) {
556 			volatile struct mlx5_mini_cqe8 (*mc)[8] =
557 				(volatile struct mlx5_mini_cqe8 (*)[8])
558 				(uintptr_t)(&(*rxq->cqes)[zip->ca &
559 							  cqe_cnt].pkt_info);
560 			len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt &
561 					       rxq->byte_mask);
562 			*mcqe = &(*mc)[zip->ai & 7];
563 			if ((++zip->ai & 7) == 0) {
564 				/* Invalidate consumed CQEs */
565 				idx = zip->ca;
566 				end = zip->na;
567 				while (idx != end) {
568 					(*rxq->cqes)[idx & cqe_cnt].op_own =
569 						MLX5_CQE_INVALIDATE;
570 					++idx;
571 				}
572 				/*
573 				 * Increment consumer index to skip the number
574 				 * of CQEs consumed. Hardware leaves holes in
575 				 * the CQ ring for software use.
576 				 */
577 				zip->ca = zip->na;
578 				zip->na += 8;
579 			}
580 			if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
581 				/* Invalidate the rest */
582 				idx = zip->ca;
583 				end = zip->cq_ci;
584 
585 				while (idx != end) {
586 					(*rxq->cqes)[idx & cqe_cnt].op_own =
587 						MLX5_CQE_INVALIDATE;
588 					++idx;
589 				}
590 				rxq->cq_ci = zip->cq_ci;
591 				zip->ai = 0;
592 			}
593 		/*
594 		 * No compressed data, get next CQE and verify if it is
595 		 * compressed.
596 		 */
597 		} else {
598 			int ret;
599 			int8_t op_own;
600 			uint32_t cq_ci;
601 
602 			ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
603 			if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
604 				if (unlikely(ret == MLX5_CQE_STATUS_ERR ||
605 					     rxq->err_state)) {
606 					ret = mlx5_rx_err_handle(rxq, 0);
607 					if (ret == MLX5_CQE_STATUS_HW_OWN ||
608 					    ret == -1)
609 						return 0;
610 				} else {
611 					return 0;
612 				}
613 			}
614 			/*
615 			 * Introduce the local variable to have queue cq_ci
616 			 * index in queue structure always consistent with
617 			 * actual CQE boundary (not pointing to the middle
618 			 * of compressed CQE session).
619 			 */
620 			cq_ci = rxq->cq_ci + 1;
621 			op_own = cqe->op_own;
622 			if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
623 				volatile struct mlx5_mini_cqe8 (*mc)[8] =
624 					(volatile struct mlx5_mini_cqe8 (*)[8])
625 					(uintptr_t)(&(*rxq->cqes)
626 						[cq_ci & cqe_cnt].pkt_info);
627 
628 				/* Fix endianness. */
629 				zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
630 				/*
631 				 * Current mini array position is the one
632 				 * returned by check_cqe64().
633 				 *
634 				 * If completion comprises several mini arrays,
635 				 * as a special case the second one is located
636 				 * 7 CQEs after the initial CQE instead of 8
637 				 * for subsequent ones.
638 				 */
639 				zip->ca = cq_ci;
640 				zip->na = zip->ca + 7;
641 				/* Compute the next non compressed CQE. */
642 				zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
643 				/* Get packet size to return. */
644 				len = rte_be_to_cpu_32((*mc)[0].byte_cnt &
645 						       rxq->byte_mask);
646 				*mcqe = &(*mc)[0];
647 				zip->ai = 1;
648 				/* Prefetch all to be invalidated */
649 				idx = zip->ca;
650 				end = zip->cq_ci;
651 				while (idx != end) {
652 					rte_prefetch0(&(*rxq->cqes)[(idx) &
653 								    cqe_cnt]);
654 					++idx;
655 				}
656 			} else {
657 				rxq->cq_ci = cq_ci;
658 				len = rte_be_to_cpu_32(cqe->byte_cnt);
659 			}
660 		}
661 		if (unlikely(rxq->err_state)) {
662 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
663 			++rxq->stats.idropped;
664 		} else {
665 			return len;
666 		}
667 	} while (1);
668 }
669 
670 /**
671  * Translate RX completion flags to offload flags.
672  *
673  * @param[in] cqe
674  *   Pointer to CQE.
675  *
676  * @return
677  *   Offload flags (ol_flags) for struct rte_mbuf.
678  */
679 static inline uint32_t
680 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe)
681 {
682 	uint32_t ol_flags = 0;
683 	uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
684 
685 	ol_flags =
686 		TRANSPOSE(flags,
687 			  MLX5_CQE_RX_L3_HDR_VALID,
688 			  RTE_MBUF_F_RX_IP_CKSUM_GOOD) |
689 		TRANSPOSE(flags,
690 			  MLX5_CQE_RX_L4_HDR_VALID,
691 			  RTE_MBUF_F_RX_L4_CKSUM_GOOD);
692 	return ol_flags;
693 }
694 
695 /**
696  * Fill in mbuf fields from RX completion flags.
697  * Note that pkt->ol_flags should be initialized outside of this function.
698  *
699  * @param rxq
700  *   Pointer to RX queue.
701  * @param pkt
702  *   mbuf to fill.
703  * @param cqe
704  *   CQE to process.
705  * @param rss_hash_res
706  *   Packet RSS Hash result.
707  */
708 static inline void
709 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
710 	       volatile struct mlx5_cqe *cqe,
711 	       volatile struct mlx5_mini_cqe8 *mcqe)
712 {
713 	/* Update packet information. */
714 	pkt->packet_type = rxq_cq_to_pkt_type(rxq, cqe, mcqe);
715 	pkt->port = unlikely(rxq->shared) ? cqe->user_index_low : rxq->port_id;
716 
717 	if (rxq->rss_hash) {
718 		uint32_t rss_hash_res = 0;
719 
720 		/* If compressed, take hash result from mini-CQE. */
721 		if (mcqe == NULL ||
722 		    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_HASH)
723 			rss_hash_res = rte_be_to_cpu_32(cqe->rx_hash_res);
724 		else
725 			rss_hash_res = rte_be_to_cpu_32(mcqe->rx_hash_result);
726 		if (rss_hash_res) {
727 			pkt->hash.rss = rss_hash_res;
728 			pkt->ol_flags |= RTE_MBUF_F_RX_RSS_HASH;
729 		}
730 	}
731 	if (rxq->mark) {
732 		uint32_t mark = 0;
733 
734 		/* If compressed, take flow tag from mini-CQE. */
735 		if (mcqe == NULL ||
736 		    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_FTAG_STRIDX)
737 			mark = cqe->sop_drop_qpn;
738 		else
739 			mark = ((mcqe->byte_cnt_flow & 0xff) << 8) |
740 				(mcqe->flow_tag_high << 16);
741 		if (MLX5_FLOW_MARK_IS_VALID(mark)) {
742 			pkt->ol_flags |= RTE_MBUF_F_RX_FDIR;
743 			if (mark != RTE_BE32(MLX5_FLOW_MARK_DEFAULT)) {
744 				pkt->ol_flags |= RTE_MBUF_F_RX_FDIR_ID;
745 				pkt->hash.fdir.hi = mlx5_flow_mark_get(mark);
746 			}
747 		}
748 	}
749 	if (rxq->dynf_meta) {
750 		uint32_t meta = rte_be_to_cpu_32(cqe->flow_table_metadata) &
751 			rxq->flow_meta_port_mask;
752 
753 		if (meta) {
754 			pkt->ol_flags |= rxq->flow_meta_mask;
755 			*RTE_MBUF_DYNFIELD(pkt, rxq->flow_meta_offset,
756 						uint32_t *) = meta;
757 		}
758 	}
759 	if (rxq->csum)
760 		pkt->ol_flags |= rxq_cq_to_ol_flags(cqe);
761 	if (rxq->vlan_strip) {
762 		bool vlan_strip;
763 
764 		if (mcqe == NULL ||
765 		    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
766 			vlan_strip = cqe->hdr_type_etc &
767 				     RTE_BE16(MLX5_CQE_VLAN_STRIPPED);
768 		else
769 			vlan_strip = mcqe->hdr_type &
770 				     RTE_BE16(MLX5_CQE_VLAN_STRIPPED);
771 		if (vlan_strip) {
772 			pkt->ol_flags |= RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED;
773 			pkt->vlan_tci = rte_be_to_cpu_16(cqe->vlan_info);
774 		}
775 	}
776 	if (rxq->hw_timestamp) {
777 		uint64_t ts = rte_be_to_cpu_64(cqe->timestamp);
778 
779 		if (rxq->rt_timestamp)
780 			ts = mlx5_txpp_convert_rx_ts(rxq->sh, ts);
781 		mlx5_timestamp_set(pkt, rxq->timestamp_offset, ts);
782 		pkt->ol_flags |= rxq->timestamp_rx_flag;
783 	}
784 }
785 
786 /**
787  * DPDK callback for RX.
788  *
789  * @param dpdk_rxq
790  *   Generic pointer to RX queue structure.
791  * @param[out] pkts
792  *   Array to store received packets.
793  * @param pkts_n
794  *   Maximum number of packets in array.
795  *
796  * @return
797  *   Number of packets successfully received (<= pkts_n).
798  */
799 uint16_t
800 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
801 {
802 	struct mlx5_rxq_data *rxq = dpdk_rxq;
803 	const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
804 	const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
805 	const unsigned int sges_n = rxq->sges_n;
806 	struct rte_mbuf *pkt = NULL;
807 	struct rte_mbuf *seg = NULL;
808 	volatile struct mlx5_cqe *cqe =
809 		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
810 	unsigned int i = 0;
811 	unsigned int rq_ci = rxq->rq_ci << sges_n;
812 	int len = 0; /* keep its value across iterations. */
813 
814 	while (pkts_n) {
815 		unsigned int idx = rq_ci & wqe_cnt;
816 		volatile struct mlx5_wqe_data_seg *wqe =
817 			&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[idx];
818 		struct rte_mbuf *rep = (*rxq->elts)[idx];
819 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
820 
821 		if (pkt)
822 			NEXT(seg) = rep;
823 		seg = rep;
824 		rte_prefetch0(seg);
825 		rte_prefetch0(cqe);
826 		rte_prefetch0(wqe);
827 		/* Allocate the buf from the same pool. */
828 		rep = rte_mbuf_raw_alloc(seg->pool);
829 		if (unlikely(rep == NULL)) {
830 			++rxq->stats.rx_nombuf;
831 			if (!pkt) {
832 				/*
833 				 * no buffers before we even started,
834 				 * bail out silently.
835 				 */
836 				break;
837 			}
838 			while (pkt != seg) {
839 				MLX5_ASSERT(pkt != (*rxq->elts)[idx]);
840 				rep = NEXT(pkt);
841 				NEXT(pkt) = NULL;
842 				NB_SEGS(pkt) = 1;
843 				rte_mbuf_raw_free(pkt);
844 				pkt = rep;
845 			}
846 			rq_ci >>= sges_n;
847 			++rq_ci;
848 			rq_ci <<= sges_n;
849 			break;
850 		}
851 		if (!pkt) {
852 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
853 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
854 			if (!len) {
855 				rte_mbuf_raw_free(rep);
856 				break;
857 			}
858 			pkt = seg;
859 			MLX5_ASSERT(len >= (rxq->crc_present << 2));
860 			pkt->ol_flags &= RTE_MBUF_F_EXTERNAL;
861 			rxq_cq_to_mbuf(rxq, pkt, cqe, mcqe);
862 			if (rxq->crc_present)
863 				len -= RTE_ETHER_CRC_LEN;
864 			PKT_LEN(pkt) = len;
865 			if (cqe->lro_num_seg > 1) {
866 				mlx5_lro_update_hdr
867 					(rte_pktmbuf_mtod(pkt, uint8_t *), cqe,
868 					 mcqe, rxq, len);
869 				pkt->ol_flags |= RTE_MBUF_F_RX_LRO;
870 				pkt->tso_segsz = len / cqe->lro_num_seg;
871 			}
872 		}
873 		DATA_LEN(rep) = DATA_LEN(seg);
874 		PKT_LEN(rep) = PKT_LEN(seg);
875 		SET_DATA_OFF(rep, DATA_OFF(seg));
876 		PORT(rep) = PORT(seg);
877 		(*rxq->elts)[idx] = rep;
878 		/*
879 		 * Fill NIC descriptor with the new buffer. The lkey and size
880 		 * of the buffers are already known, only the buffer address
881 		 * changes.
882 		 */
883 		wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
884 		/* If there's only one MR, no need to replace LKey in WQE. */
885 		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
886 			wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
887 		if (len > DATA_LEN(seg)) {
888 			len -= DATA_LEN(seg);
889 			++NB_SEGS(pkt);
890 			++rq_ci;
891 			continue;
892 		}
893 		DATA_LEN(seg) = len;
894 #ifdef MLX5_PMD_SOFT_COUNTERS
895 		/* Increment bytes counter. */
896 		rxq->stats.ibytes += PKT_LEN(pkt);
897 #endif
898 		/* Return packet. */
899 		*(pkts++) = pkt;
900 		pkt = NULL;
901 		--pkts_n;
902 		++i;
903 		/* Align consumer index to the next stride. */
904 		rq_ci >>= sges_n;
905 		++rq_ci;
906 		rq_ci <<= sges_n;
907 	}
908 	if (unlikely(i == 0 && ((rq_ci >> sges_n) == rxq->rq_ci)))
909 		return 0;
910 	/* Update the consumer index. */
911 	rxq->rq_ci = rq_ci >> sges_n;
912 	rte_io_wmb();
913 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
914 	rte_io_wmb();
915 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
916 #ifdef MLX5_PMD_SOFT_COUNTERS
917 	/* Increment packets counter. */
918 	rxq->stats.ipackets += i;
919 #endif
920 	return i;
921 }
922 
923 /**
924  * Update LRO packet TCP header.
925  * The HW LRO feature doesn't update the TCP header after coalescing the
926  * TCP segments but supplies information in CQE to fill it by SW.
927  *
928  * @param tcp
929  *   Pointer to the TCP header.
930  * @param cqe
931  *   Pointer to the completion entry.
932  * @param phcsum
933  *   The L3 pseudo-header checksum.
934  */
935 static inline void
936 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *__rte_restrict tcp,
937 			volatile struct mlx5_cqe *__rte_restrict cqe,
938 			uint32_t phcsum, uint8_t l4_type)
939 {
940 	/*
941 	 * The HW calculates only the TCP payload checksum, need to complete
942 	 * the TCP header checksum and the L3 pseudo-header checksum.
943 	 */
944 	uint32_t csum = phcsum + cqe->csum;
945 
946 	if (l4_type == MLX5_L4_HDR_TYPE_TCP_EMPTY_ACK ||
947 	    l4_type == MLX5_L4_HDR_TYPE_TCP_WITH_ACL) {
948 		tcp->tcp_flags |= RTE_TCP_ACK_FLAG;
949 		tcp->recv_ack = cqe->lro_ack_seq_num;
950 		tcp->rx_win = cqe->lro_tcp_win;
951 	}
952 	if (cqe->lro_tcppsh_abort_dupack & MLX5_CQE_LRO_PUSH_MASK)
953 		tcp->tcp_flags |= RTE_TCP_PSH_FLAG;
954 	tcp->cksum = 0;
955 	csum += rte_raw_cksum(tcp, (tcp->data_off >> 4) * 4);
956 	csum = ((csum & 0xffff0000) >> 16) + (csum & 0xffff);
957 	csum = (~csum) & 0xffff;
958 	if (csum == 0)
959 		csum = 0xffff;
960 	tcp->cksum = csum;
961 }
962 
963 /**
964  * Update LRO packet headers.
965  * The HW LRO feature doesn't update the L3/TCP headers after coalescing the
966  * TCP segments but supply information in CQE to fill it by SW.
967  *
968  * @param padd
969  *   The packet address.
970  * @param cqe
971  *   Pointer to the completion entry.
972  * @param len
973  *   The packet length.
974  */
975 static inline void
976 mlx5_lro_update_hdr(uint8_t *__rte_restrict padd,
977 		    volatile struct mlx5_cqe *__rte_restrict cqe,
978 		    volatile struct mlx5_mini_cqe8 *mcqe,
979 		    struct mlx5_rxq_data *rxq, uint32_t len)
980 {
981 	union {
982 		struct rte_ether_hdr *eth;
983 		struct rte_vlan_hdr *vlan;
984 		struct rte_ipv4_hdr *ipv4;
985 		struct rte_ipv6_hdr *ipv6;
986 		struct rte_tcp_hdr *tcp;
987 		uint8_t *hdr;
988 	} h = {
989 		.hdr = padd,
990 	};
991 	uint16_t proto = h.eth->ether_type;
992 	uint32_t phcsum;
993 	uint8_t l4_type;
994 
995 	h.eth++;
996 	while (proto == RTE_BE16(RTE_ETHER_TYPE_VLAN) ||
997 	       proto == RTE_BE16(RTE_ETHER_TYPE_QINQ)) {
998 		proto = h.vlan->eth_proto;
999 		h.vlan++;
1000 	}
1001 	if (proto == RTE_BE16(RTE_ETHER_TYPE_IPV4)) {
1002 		h.ipv4->time_to_live = cqe->lro_min_ttl;
1003 		h.ipv4->total_length = rte_cpu_to_be_16(len - (h.hdr - padd));
1004 		h.ipv4->hdr_checksum = 0;
1005 		h.ipv4->hdr_checksum = rte_ipv4_cksum(h.ipv4);
1006 		phcsum = rte_ipv4_phdr_cksum(h.ipv4, 0);
1007 		h.ipv4++;
1008 	} else {
1009 		h.ipv6->hop_limits = cqe->lro_min_ttl;
1010 		h.ipv6->payload_len = rte_cpu_to_be_16(len - (h.hdr - padd) -
1011 						       sizeof(*h.ipv6));
1012 		phcsum = rte_ipv6_phdr_cksum(h.ipv6, 0);
1013 		h.ipv6++;
1014 	}
1015 	if (mcqe == NULL ||
1016 	    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
1017 		l4_type = (rte_be_to_cpu_16(cqe->hdr_type_etc) &
1018 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1019 	else
1020 		l4_type = (rte_be_to_cpu_16(mcqe->hdr_type) &
1021 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1022 	mlx5_lro_update_tcp_hdr(h.tcp, cqe, phcsum, l4_type);
1023 }
1024 
1025 void
1026 mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf)
1027 {
1028 	mlx5_mprq_buf_free_cb(NULL, buf);
1029 }
1030 
1031 /**
1032  * DPDK callback for RX with Multi-Packet RQ support.
1033  *
1034  * @param dpdk_rxq
1035  *   Generic pointer to RX queue structure.
1036  * @param[out] pkts
1037  *   Array to store received packets.
1038  * @param pkts_n
1039  *   Maximum number of packets in array.
1040  *
1041  * @return
1042  *   Number of packets successfully received (<= pkts_n).
1043  */
1044 uint16_t
1045 mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1046 {
1047 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1048 	const uint32_t strd_n = 1 << rxq->strd_num_n;
1049 	const uint32_t strd_sz = 1 << rxq->strd_sz_n;
1050 	const uint32_t cq_mask = (1 << rxq->cqe_n) - 1;
1051 	const uint32_t wq_mask = (1 << rxq->elts_n) - 1;
1052 	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1053 	unsigned int i = 0;
1054 	uint32_t rq_ci = rxq->rq_ci;
1055 	uint16_t consumed_strd = rxq->consumed_strd;
1056 	struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1057 
1058 	while (i < pkts_n) {
1059 		struct rte_mbuf *pkt;
1060 		int ret;
1061 		uint32_t len;
1062 		uint16_t strd_cnt;
1063 		uint16_t strd_idx;
1064 		uint32_t byte_cnt;
1065 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1066 		enum mlx5_rqx_code rxq_code;
1067 
1068 		if (consumed_strd == strd_n) {
1069 			/* Replace WQE if the buffer is still in use. */
1070 			mprq_buf_replace(rxq, rq_ci & wq_mask);
1071 			/* Advance to the next WQE. */
1072 			consumed_strd = 0;
1073 			++rq_ci;
1074 			buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1075 		}
1076 		cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1077 		ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
1078 		if (!ret)
1079 			break;
1080 		byte_cnt = ret;
1081 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1082 		MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
1083 		if (rxq->crc_present)
1084 			len -= RTE_ETHER_CRC_LEN;
1085 		if (mcqe &&
1086 		    rxq->mcqe_format == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX)
1087 			strd_cnt = (len / strd_sz) + !!(len % strd_sz);
1088 		else
1089 			strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >>
1090 				   MLX5_MPRQ_STRIDE_NUM_SHIFT;
1091 		MLX5_ASSERT(strd_cnt);
1092 		consumed_strd += strd_cnt;
1093 		if (byte_cnt & MLX5_MPRQ_FILLER_MASK)
1094 			continue;
1095 		strd_idx = rte_be_to_cpu_16(mcqe == NULL ?
1096 					cqe->wqe_counter :
1097 					mcqe->stride_idx);
1098 		MLX5_ASSERT(strd_idx < strd_n);
1099 		MLX5_ASSERT(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) &
1100 			    wq_mask));
1101 		pkt = rte_pktmbuf_alloc(rxq->mp);
1102 		if (unlikely(pkt == NULL)) {
1103 			++rxq->stats.rx_nombuf;
1104 			break;
1105 		}
1106 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1107 		MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
1108 		if (rxq->crc_present)
1109 			len -= RTE_ETHER_CRC_LEN;
1110 		rxq_code = mprq_buf_to_pkt(rxq, pkt, len, buf,
1111 					   strd_idx, strd_cnt);
1112 		if (unlikely(rxq_code != MLX5_RXQ_CODE_EXIT)) {
1113 			rte_pktmbuf_free_seg(pkt);
1114 			if (rxq_code == MLX5_RXQ_CODE_DROPPED) {
1115 				++rxq->stats.idropped;
1116 				continue;
1117 			}
1118 			if (rxq_code == MLX5_RXQ_CODE_NOMBUF) {
1119 				++rxq->stats.rx_nombuf;
1120 				break;
1121 			}
1122 		}
1123 		rxq_cq_to_mbuf(rxq, pkt, cqe, mcqe);
1124 		if (cqe->lro_num_seg > 1) {
1125 			mlx5_lro_update_hdr(rte_pktmbuf_mtod(pkt, uint8_t *),
1126 					    cqe, mcqe, rxq, len);
1127 			pkt->ol_flags |= RTE_MBUF_F_RX_LRO;
1128 			pkt->tso_segsz = len / cqe->lro_num_seg;
1129 		}
1130 		PKT_LEN(pkt) = len;
1131 		PORT(pkt) = rxq->port_id;
1132 #ifdef MLX5_PMD_SOFT_COUNTERS
1133 		/* Increment bytes counter. */
1134 		rxq->stats.ibytes += PKT_LEN(pkt);
1135 #endif
1136 		/* Return packet. */
1137 		*(pkts++) = pkt;
1138 		++i;
1139 	}
1140 	/* Update the consumer indexes. */
1141 	rxq->consumed_strd = consumed_strd;
1142 	rte_io_wmb();
1143 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1144 	if (rq_ci != rxq->rq_ci) {
1145 		rxq->rq_ci = rq_ci;
1146 		rte_io_wmb();
1147 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1148 	}
1149 #ifdef MLX5_PMD_SOFT_COUNTERS
1150 	/* Increment packets counter. */
1151 	rxq->stats.ipackets += i;
1152 #endif
1153 	return i;
1154 }
1155 
1156 /**
1157  * Dummy DPDK callback for RX.
1158  *
1159  * This function is used to temporarily replace the real callback during
1160  * unsafe control operations on the queue, or in case of error.
1161  *
1162  * @param dpdk_rxq
1163  *   Generic pointer to RX queue structure.
1164  * @param[out] pkts
1165  *   Array to store received packets.
1166  * @param pkts_n
1167  *   Maximum number of packets in array.
1168  *
1169  * @return
1170  *   Number of packets successfully received (<= pkts_n).
1171  */
1172 uint16_t
1173 removed_rx_burst(void *dpdk_rxq __rte_unused,
1174 		 struct rte_mbuf **pkts __rte_unused,
1175 		 uint16_t pkts_n __rte_unused)
1176 {
1177 	rte_mb();
1178 	return 0;
1179 }
1180 
1181 /*
1182  * Vectorized Rx routines are not compiled in when required vector instructions
1183  * are not supported on a target architecture.
1184  * The following null stubs are needed for linkage when those are not included
1185  * outside of this file (e.g. mlx5_rxtx_vec_sse.c for x86).
1186  */
1187 
1188 __rte_weak uint16_t
1189 mlx5_rx_burst_vec(void *dpdk_rxq __rte_unused,
1190 		  struct rte_mbuf **pkts __rte_unused,
1191 		  uint16_t pkts_n __rte_unused)
1192 {
1193 	return 0;
1194 }
1195 
1196 __rte_weak uint16_t
1197 mlx5_rx_burst_mprq_vec(void *dpdk_rxq __rte_unused,
1198 		       struct rte_mbuf **pkts __rte_unused,
1199 		       uint16_t pkts_n __rte_unused)
1200 {
1201 	return 0;
1202 }
1203 
1204 __rte_weak int
1205 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq __rte_unused)
1206 {
1207 	return -ENOTSUP;
1208 }
1209 
1210 __rte_weak int
1211 mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused)
1212 {
1213 	return -ENOTSUP;
1214 }
1215 
1216