xref: /dpdk/drivers/net/mlx5/mlx5_rxtx_vec.c (revision 99a2dd955fba6e4cc23b77d590a033650ced9c45)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2017 6WIND S.A.
3  * Copyright 2017 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_vect.h>
14 
15 #include <mlx5_glue.h>
16 #include <mlx5_prm.h>
17 
18 #include "mlx5_defs.h"
19 #include "mlx5.h"
20 #include "mlx5_utils.h"
21 #include "mlx5_rxtx.h"
22 #include "mlx5_rx.h"
23 #include "mlx5_rxtx_vec.h"
24 #include "mlx5_autoconf.h"
25 
26 #if defined RTE_ARCH_X86_64
27 #include "mlx5_rxtx_vec_sse.h"
28 #elif defined RTE_ARCH_ARM64
29 #include "mlx5_rxtx_vec_neon.h"
30 #elif defined RTE_ARCH_PPC_64
31 #include "mlx5_rxtx_vec_altivec.h"
32 #else
33 #error "This should not be compiled if SIMD instructions are not supported."
34 #endif
35 
36 /**
37  * Skip error packets.
38  *
39  * @param rxq
40  *   Pointer to RX queue structure.
41  * @param[out] pkts
42  *   Array to store received packets.
43  * @param pkts_n
44  *   Maximum number of packets in array.
45  *
46  * @return
47  *   Number of packets successfully received (<= pkts_n).
48  */
49 static uint16_t
50 rxq_handle_pending_error(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
51 			 uint16_t pkts_n)
52 {
53 	uint16_t n = 0;
54 	unsigned int i;
55 #ifdef MLX5_PMD_SOFT_COUNTERS
56 	uint32_t err_bytes = 0;
57 #endif
58 
59 	for (i = 0; i < pkts_n; ++i) {
60 		struct rte_mbuf *pkt = pkts[i];
61 
62 		if (pkt->packet_type == RTE_PTYPE_ALL_MASK || rxq->err_state) {
63 #ifdef MLX5_PMD_SOFT_COUNTERS
64 			err_bytes += PKT_LEN(pkt);
65 #endif
66 			rte_pktmbuf_free_seg(pkt);
67 		} else {
68 			pkts[n++] = pkt;
69 		}
70 	}
71 	rxq->stats.idropped += (pkts_n - n);
72 #ifdef MLX5_PMD_SOFT_COUNTERS
73 	/* Correct counters of errored completions. */
74 	rxq->stats.ipackets -= (pkts_n - n);
75 	rxq->stats.ibytes -= err_bytes;
76 #endif
77 	mlx5_rx_err_handle(rxq, 1);
78 	return n;
79 }
80 
81 /**
82  * Replenish buffers for RX in bulk.
83  *
84  * @param rxq
85  *   Pointer to RX queue structure.
86  */
87 static inline void
88 mlx5_rx_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq)
89 {
90 	const uint16_t q_n = 1 << rxq->elts_n;
91 	const uint16_t q_mask = q_n - 1;
92 	uint16_t n = q_n - (rxq->rq_ci - rxq->rq_pi);
93 	uint16_t elts_idx = rxq->rq_ci & q_mask;
94 	struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
95 	volatile struct mlx5_wqe_data_seg *wq =
96 		&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[elts_idx];
97 	unsigned int i;
98 
99 	if (n >= rxq->rq_repl_thresh) {
100 		MLX5_ASSERT(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n));
101 		MLX5_ASSERT(MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n) >
102 			    MLX5_VPMD_DESCS_PER_LOOP);
103 		/* Not to cross queue end. */
104 		n = RTE_MIN(n - MLX5_VPMD_DESCS_PER_LOOP, q_n - elts_idx);
105 		if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
106 			rxq->stats.rx_nombuf += n;
107 			return;
108 		}
109 		for (i = 0; i < n; ++i) {
110 			void *buf_addr;
111 
112 			/*
113 			 * In order to support the mbufs with external attached
114 			 * data buffer we should use the buf_addr pointer
115 			 * instead of rte_mbuf_buf_addr(). It touches the mbuf
116 			 * itself and may impact the performance.
117 			 */
118 			buf_addr = elts[i]->buf_addr;
119 			wq[i].addr = rte_cpu_to_be_64((uintptr_t)buf_addr +
120 						      RTE_PKTMBUF_HEADROOM);
121 			/* If there's a single MR, no need to replace LKey. */
122 			if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh)
123 				     > 1))
124 				wq[i].lkey = mlx5_rx_mb2mr(rxq, elts[i]);
125 		}
126 		rxq->rq_ci += n;
127 		/* Prevent overflowing into consumed mbufs. */
128 		elts_idx = rxq->rq_ci & q_mask;
129 		for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
130 			(*rxq->elts)[elts_idx + i] = &rxq->fake_mbuf;
131 		rte_io_wmb();
132 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
133 	}
134 }
135 
136 /**
137  * Replenish buffers for MPRQ RX in bulk.
138  *
139  * @param rxq
140  *   Pointer to RX queue structure.
141  */
142 static inline void
143 mlx5_rx_mprq_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq)
144 {
145 	const uint16_t wqe_n = 1 << rxq->elts_n;
146 	const uint32_t strd_n = 1 << rxq->strd_num_n;
147 	const uint32_t elts_n = wqe_n * strd_n;
148 	const uint32_t wqe_mask = elts_n - 1;
149 	uint32_t n = elts_n - (rxq->elts_ci - rxq->rq_pi);
150 	uint32_t elts_idx = rxq->elts_ci & wqe_mask;
151 	struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
152 	unsigned int i;
153 
154 	if (n >= rxq->rq_repl_thresh &&
155 	    rxq->elts_ci - rxq->rq_pi <= rxq->rq_repl_thresh) {
156 		MLX5_ASSERT(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(elts_n));
157 		MLX5_ASSERT(MLX5_VPMD_RXQ_RPLNSH_THRESH(elts_n) >
158 			     MLX5_VPMD_DESCS_PER_LOOP);
159 		/* Not to cross queue end. */
160 		n = RTE_MIN(n - MLX5_VPMD_DESCS_PER_LOOP, elts_n - elts_idx);
161 		/* Limit replenish number to threshold value. */
162 		n = RTE_MIN(n, rxq->rq_repl_thresh);
163 		if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
164 			rxq->stats.rx_nombuf += n;
165 			return;
166 		}
167 		rxq->elts_ci += n;
168 		/* Prevent overflowing into consumed mbufs. */
169 		elts_idx = rxq->elts_ci & wqe_mask;
170 		for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
171 			(*rxq->elts)[elts_idx + i] = &rxq->fake_mbuf;
172 	}
173 }
174 
175 /**
176  * Copy or attach MPRQ buffers to RX SW ring.
177  *
178  * @param rxq
179  *   Pointer to RX queue structure.
180  * @param pkts
181  *   Pointer to array of packets to be stored.
182  * @param pkts_n
183  *   Number of packets to be stored.
184  *
185  * @return
186  *   Number of packets successfully copied/attached (<= pkts_n).
187  */
188 static inline uint16_t
189 rxq_copy_mprq_mbuf_v(struct mlx5_rxq_data *rxq,
190 		     struct rte_mbuf **pkts, uint16_t pkts_n)
191 {
192 	const uint16_t wqe_n = 1 << rxq->elts_n;
193 	const uint16_t wqe_mask = wqe_n - 1;
194 	const uint16_t strd_sz = 1 << rxq->strd_sz_n;
195 	const uint32_t strd_n = 1 << rxq->strd_num_n;
196 	const uint32_t elts_n = wqe_n * strd_n;
197 	const uint32_t elts_mask = elts_n - 1;
198 	uint32_t elts_idx = rxq->rq_pi & elts_mask;
199 	struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
200 	uint32_t rq_ci = rxq->rq_ci;
201 	struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wqe_mask];
202 	uint16_t copied = 0;
203 	uint16_t i = 0;
204 
205 	for (i = 0; i < pkts_n; ++i) {
206 		uint16_t strd_cnt;
207 		enum mlx5_rqx_code rxq_code;
208 
209 		if (rxq->consumed_strd == strd_n) {
210 			/* Replace WQE if the buffer is still in use. */
211 			mprq_buf_replace(rxq, rq_ci & wqe_mask);
212 			/* Advance to the next WQE. */
213 			rxq->consumed_strd = 0;
214 			rq_ci++;
215 			buf = (*rxq->mprq_bufs)[rq_ci & wqe_mask];
216 		}
217 
218 		if (!elts[i]->pkt_len) {
219 			rxq->consumed_strd = strd_n;
220 			rte_pktmbuf_free_seg(elts[i]);
221 #ifdef MLX5_PMD_SOFT_COUNTERS
222 			rxq->stats.ipackets -= 1;
223 #endif
224 			continue;
225 		}
226 		strd_cnt = (elts[i]->pkt_len / strd_sz) +
227 			   ((elts[i]->pkt_len % strd_sz) ? 1 : 0);
228 		rxq_code = mprq_buf_to_pkt(rxq, elts[i], elts[i]->pkt_len,
229 					   buf, rxq->consumed_strd, strd_cnt);
230 		rxq->consumed_strd += strd_cnt;
231 		if (unlikely(rxq_code != MLX5_RXQ_CODE_EXIT)) {
232 			rte_pktmbuf_free_seg(elts[i]);
233 #ifdef MLX5_PMD_SOFT_COUNTERS
234 			rxq->stats.ipackets -= 1;
235 			rxq->stats.ibytes -= elts[i]->pkt_len;
236 #endif
237 			if (rxq_code == MLX5_RXQ_CODE_NOMBUF) {
238 				++rxq->stats.rx_nombuf;
239 				break;
240 			}
241 			if (rxq_code == MLX5_RXQ_CODE_DROPPED) {
242 				++rxq->stats.idropped;
243 				continue;
244 			}
245 		}
246 		pkts[copied++] = elts[i];
247 	}
248 	rxq->rq_pi += i;
249 	rxq->cq_ci += i;
250 	rte_io_wmb();
251 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
252 	if (rq_ci != rxq->rq_ci) {
253 		rxq->rq_ci = rq_ci;
254 		rte_io_wmb();
255 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
256 	}
257 	return copied;
258 }
259 
260 /**
261  * Receive burst of packets. An errored completion also consumes a mbuf, but the
262  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
263  * before returning to application.
264  *
265  * @param rxq
266  *   Pointer to RX queue structure.
267  * @param[out] pkts
268  *   Array to store received packets.
269  * @param pkts_n
270  *   Maximum number of packets in array.
271  * @param[out] err
272  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
273  *   packet to handle.
274  * @param[out] no_cq
275  *   Pointer to a boolean. Set true if no new CQE seen.
276  *
277  * @return
278  *   Number of packets received including errors (<= pkts_n).
279  */
280 static inline uint16_t
281 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
282 	    uint16_t pkts_n, uint64_t *err, bool *no_cq)
283 {
284 	const uint16_t q_n = 1 << rxq->cqe_n;
285 	const uint16_t q_mask = q_n - 1;
286 	const uint16_t e_n = 1 << rxq->elts_n;
287 	const uint16_t e_mask = e_n - 1;
288 	volatile struct mlx5_cqe *cq;
289 	struct rte_mbuf **elts;
290 	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
291 	uint16_t nocmp_n = 0;
292 	uint16_t rcvd_pkt = 0;
293 	unsigned int cq_idx = rxq->cq_ci & q_mask;
294 	unsigned int elts_idx;
295 
296 	MLX5_ASSERT(rxq->sges_n == 0);
297 	MLX5_ASSERT(rxq->cqe_n == rxq->elts_n);
298 	cq = &(*rxq->cqes)[cq_idx];
299 	rte_prefetch0(cq);
300 	rte_prefetch0(cq + 1);
301 	rte_prefetch0(cq + 2);
302 	rte_prefetch0(cq + 3);
303 	pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
304 	mlx5_rx_replenish_bulk_mbuf(rxq);
305 	/* See if there're unreturned mbufs from compressed CQE. */
306 	rcvd_pkt = rxq->decompressed;
307 	if (rcvd_pkt > 0) {
308 		rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
309 		rxq_copy_mbuf_v(&(*rxq->elts)[rxq->rq_pi & e_mask],
310 				pkts, rcvd_pkt);
311 		rxq->rq_pi += rcvd_pkt;
312 		rxq->decompressed -= rcvd_pkt;
313 		pkts += rcvd_pkt;
314 	}
315 	elts_idx = rxq->rq_pi & e_mask;
316 	elts = &(*rxq->elts)[elts_idx];
317 	/* Not to overflow pkts array. */
318 	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
319 	/* Not to cross queue end. */
320 	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
321 	pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
322 	if (!pkts_n) {
323 		*no_cq = !rcvd_pkt;
324 		return rcvd_pkt;
325 	}
326 	/* At this point, there shouldn't be any remaining packets. */
327 	MLX5_ASSERT(rxq->decompressed == 0);
328 	/* Process all the CQEs */
329 	nocmp_n = rxq_cq_process_v(rxq, cq, elts, pkts, pkts_n, err, &comp_idx);
330 	/* If no new CQE seen, return without updating cq_db. */
331 	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
332 		*no_cq = true;
333 		return rcvd_pkt;
334 	}
335 	/* Update the consumer indexes for non-compressed CQEs. */
336 	MLX5_ASSERT(nocmp_n <= pkts_n);
337 	rxq->cq_ci += nocmp_n;
338 	rxq->rq_pi += nocmp_n;
339 	rcvd_pkt += nocmp_n;
340 	/* Decompress the last CQE if compressed. */
341 	if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP) {
342 		MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
343 		rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
344 							&elts[nocmp_n]);
345 		rxq->cq_ci += rxq->decompressed;
346 		/* Return more packets if needed. */
347 		if (nocmp_n < pkts_n) {
348 			uint16_t n = rxq->decompressed;
349 
350 			n = RTE_MIN(n, pkts_n - nocmp_n);
351 			rxq_copy_mbuf_v(&(*rxq->elts)[rxq->rq_pi & e_mask],
352 					&pkts[nocmp_n], n);
353 			rxq->rq_pi += n;
354 			rcvd_pkt += n;
355 			rxq->decompressed -= n;
356 		}
357 	}
358 	rte_io_wmb();
359 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
360 	*no_cq = !rcvd_pkt;
361 	return rcvd_pkt;
362 }
363 
364 /**
365  * DPDK callback for vectorized RX.
366  *
367  * @param dpdk_rxq
368  *   Generic pointer to RX queue structure.
369  * @param[out] pkts
370  *   Array to store received packets.
371  * @param pkts_n
372  *   Maximum number of packets in array.
373  *
374  * @return
375  *   Number of packets successfully received (<= pkts_n).
376  */
377 uint16_t
378 mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
379 {
380 	struct mlx5_rxq_data *rxq = dpdk_rxq;
381 	uint16_t nb_rx = 0;
382 	uint16_t tn = 0;
383 	uint64_t err = 0;
384 	bool no_cq = false;
385 
386 	do {
387 		nb_rx = rxq_burst_v(rxq, pkts + tn, pkts_n - tn,
388 				    &err, &no_cq);
389 		if (unlikely(err | rxq->err_state))
390 			nb_rx = rxq_handle_pending_error(rxq, pkts + tn, nb_rx);
391 		tn += nb_rx;
392 		if (unlikely(no_cq))
393 			break;
394 	} while (tn != pkts_n);
395 	return tn;
396 }
397 
398 /**
399  * Receive burst of packets. An errored completion also consumes a mbuf, but the
400  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
401  * before returning to application.
402  *
403  * @param rxq
404  *   Pointer to RX queue structure.
405  * @param[out] pkts
406  *   Array to store received packets.
407  * @param pkts_n
408  *   Maximum number of packets in array.
409  * @param[out] err
410  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
411  *   packet to handle.
412  * @param[out] no_cq
413  *   Pointer to a boolean. Set true if no new CQE seen.
414  *
415  * @return
416  *   Number of packets received including errors (<= pkts_n).
417  */
418 static inline uint16_t
419 rxq_burst_mprq_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
420 		 uint16_t pkts_n, uint64_t *err, bool *no_cq)
421 {
422 	const uint16_t q_n = 1 << rxq->cqe_n;
423 	const uint16_t q_mask = q_n - 1;
424 	const uint16_t wqe_n = 1 << rxq->elts_n;
425 	const uint32_t strd_n = 1 << rxq->strd_num_n;
426 	const uint32_t elts_n = wqe_n * strd_n;
427 	const uint32_t elts_mask = elts_n - 1;
428 	volatile struct mlx5_cqe *cq;
429 	struct rte_mbuf **elts;
430 	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
431 	uint16_t nocmp_n = 0;
432 	uint16_t rcvd_pkt = 0;
433 	uint16_t cp_pkt = 0;
434 	unsigned int cq_idx = rxq->cq_ci & q_mask;
435 	unsigned int elts_idx;
436 
437 	MLX5_ASSERT(rxq->sges_n == 0);
438 	cq = &(*rxq->cqes)[cq_idx];
439 	rte_prefetch0(cq);
440 	rte_prefetch0(cq + 1);
441 	rte_prefetch0(cq + 2);
442 	rte_prefetch0(cq + 3);
443 	pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
444 	mlx5_rx_mprq_replenish_bulk_mbuf(rxq);
445 	/* See if there're unreturned mbufs from compressed CQE. */
446 	rcvd_pkt = rxq->decompressed;
447 	if (rcvd_pkt > 0) {
448 		rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
449 		cp_pkt = rxq_copy_mprq_mbuf_v(rxq, pkts, rcvd_pkt);
450 		rxq->decompressed -= rcvd_pkt;
451 		pkts += cp_pkt;
452 	}
453 	elts_idx = rxq->rq_pi & elts_mask;
454 	elts = &(*rxq->elts)[elts_idx];
455 	/* Not to overflow pkts array. */
456 	pkts_n = RTE_ALIGN_FLOOR(pkts_n - cp_pkt, MLX5_VPMD_DESCS_PER_LOOP);
457 	/* Not to cross queue end. */
458 	pkts_n = RTE_MIN(pkts_n, elts_n - elts_idx);
459 	pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
460 	/* Not to move past the allocated mbufs. */
461 	pkts_n = RTE_MIN(pkts_n, rxq->elts_ci - rxq->rq_pi);
462 	if (!pkts_n) {
463 		*no_cq = !cp_pkt;
464 		return cp_pkt;
465 	}
466 	/* At this point, there shouldn't be any remaining packets. */
467 	MLX5_ASSERT(rxq->decompressed == 0);
468 	/* Process all the CQEs */
469 	nocmp_n = rxq_cq_process_v(rxq, cq, elts, pkts, pkts_n, err, &comp_idx);
470 	/* If no new CQE seen, return without updating cq_db. */
471 	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
472 		*no_cq = true;
473 		return cp_pkt;
474 	}
475 	/* Update the consumer indexes for non-compressed CQEs. */
476 	MLX5_ASSERT(nocmp_n <= pkts_n);
477 	cp_pkt = rxq_copy_mprq_mbuf_v(rxq, pkts, nocmp_n);
478 	rcvd_pkt += cp_pkt;
479 	/* Decompress the last CQE if compressed. */
480 	if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP) {
481 		MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
482 		rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
483 							&elts[nocmp_n]);
484 		/* Return more packets if needed. */
485 		if (nocmp_n < pkts_n) {
486 			uint16_t n = rxq->decompressed;
487 
488 			n = RTE_MIN(n, pkts_n - nocmp_n);
489 			cp_pkt = rxq_copy_mprq_mbuf_v(rxq, &pkts[cp_pkt], n);
490 			rcvd_pkt += cp_pkt;
491 			rxq->decompressed -= n;
492 		}
493 	}
494 	*no_cq = !rcvd_pkt;
495 	return rcvd_pkt;
496 }
497 
498 /**
499  * DPDK callback for vectorized MPRQ RX.
500  *
501  * @param dpdk_rxq
502  *   Generic pointer to RX queue structure.
503  * @param[out] pkts
504  *   Array to store received packets.
505  * @param pkts_n
506  *   Maximum number of packets in array.
507  *
508  * @return
509  *   Number of packets successfully received (<= pkts_n).
510  */
511 uint16_t
512 mlx5_rx_burst_mprq_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
513 {
514 	struct mlx5_rxq_data *rxq = dpdk_rxq;
515 	uint16_t nb_rx = 0;
516 	uint16_t tn = 0;
517 	uint64_t err = 0;
518 	bool no_cq = false;
519 
520 	do {
521 		nb_rx = rxq_burst_mprq_v(rxq, pkts + tn, pkts_n - tn,
522 					 &err, &no_cq);
523 		if (unlikely(err | rxq->err_state))
524 			nb_rx = rxq_handle_pending_error(rxq, pkts + tn, nb_rx);
525 		tn += nb_rx;
526 		if (unlikely(no_cq))
527 			break;
528 	} while (tn != pkts_n);
529 	return tn;
530 }
531 
532 /**
533  * Check a RX queue can support vectorized RX.
534  *
535  * @param rxq
536  *   Pointer to RX queue.
537  *
538  * @return
539  *   1 if supported, negative errno value if not.
540  */
541 int __rte_cold
542 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq)
543 {
544 	struct mlx5_rxq_ctrl *ctrl =
545 		container_of(rxq, struct mlx5_rxq_ctrl, rxq);
546 
547 	if (!ctrl->priv->config.rx_vec_en || rxq->sges_n != 0)
548 		return -ENOTSUP;
549 	if (rxq->lro)
550 		return -ENOTSUP;
551 	return 1;
552 }
553 
554 /**
555  * Check a device can support vectorized RX.
556  *
557  * @param dev
558  *   Pointer to Ethernet device.
559  *
560  * @return
561  *   1 if supported, negative errno value if not.
562  */
563 int __rte_cold
564 mlx5_check_vec_rx_support(struct rte_eth_dev *dev)
565 {
566 	struct mlx5_priv *priv = dev->data->dev_private;
567 	uint32_t i;
568 
569 	if (rte_vect_get_max_simd_bitwidth() < RTE_VECT_SIMD_128)
570 		return -ENOTSUP;
571 	if (!priv->config.rx_vec_en)
572 		return -ENOTSUP;
573 	/* All the configured queues should support. */
574 	for (i = 0; i < priv->rxqs_n; ++i) {
575 		struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
576 
577 		if (!rxq)
578 			continue;
579 		if (mlx5_rxq_check_vec_support(rxq) < 0)
580 			break;
581 	}
582 	if (i != priv->rxqs_n)
583 		return -ENOTSUP;
584 	return 1;
585 }
586