1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2017 6WIND S.A.
3 * Copyright 2017 Mellanox Technologies, Ltd
11 #include <rte_mempool.h>
12 #include <rte_prefetch.h>
15 #include <mlx5_glue.h>
18 #include "mlx5_defs.h"
20 #include "mlx5_utils.h"
21 #include "mlx5_rxtx.h"
22 #include "mlx5_rxtx_vec.h"
23 #include "mlx5_autoconf.h"
25 #if defined RTE_ARCH_X86_64
26 #include "mlx5_rxtx_vec_sse.h"
27 #elif defined RTE_ARCH_ARM64
28 #include "mlx5_rxtx_vec_neon.h"
29 #elif defined RTE_ARCH_PPC_64
30 #include "mlx5_rxtx_vec_altivec.h"
32 #error "This should not be compiled if SIMD instructions are not supported."
39 * Pointer to RX queue structure.
41 * Array to store received packets.
43 * Maximum number of packets in array.
46 * Number of packets successfully received (<= pkts_n).
49 rxq_handle_pending_error(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
54 #ifdef MLX5_PMD_SOFT_COUNTERS
55 uint32_t err_bytes = 0;
58 for (i = 0; i < pkts_n; ++i) {
59 struct rte_mbuf *pkt = pkts[i];
61 if (pkt->packet_type == RTE_PTYPE_ALL_MASK || rxq->err_state) {
62 #ifdef MLX5_PMD_SOFT_COUNTERS
63 err_bytes += PKT_LEN(pkt);
65 rte_pktmbuf_free_seg(pkt);
70 rxq->stats.idropped += (pkts_n - n);
71 #ifdef MLX5_PMD_SOFT_COUNTERS
72 /* Correct counters of errored completions. */
73 rxq->stats.ipackets -= (pkts_n - n);
74 rxq->stats.ibytes -= err_bytes;
76 mlx5_rx_err_handle(rxq, 1);
81 * Replenish buffers for RX in bulk.
84 * Pointer to RX queue structure.
87 mlx5_rx_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq)
89 const uint16_t q_n = 1 << rxq->elts_n;
90 const uint16_t q_mask = q_n - 1;
91 uint16_t n = q_n - (rxq->rq_ci - rxq->rq_pi);
92 uint16_t elts_idx = rxq->rq_ci & q_mask;
93 struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
94 volatile struct mlx5_wqe_data_seg *wq =
95 &((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[elts_idx];
98 if (n >= rxq->rq_repl_thresh) {
99 MLX5_ASSERT(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n));
100 MLX5_ASSERT(MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n) >
101 MLX5_VPMD_DESCS_PER_LOOP);
102 /* Not to cross queue end. */
103 n = RTE_MIN(n - MLX5_VPMD_DESCS_PER_LOOP, q_n - elts_idx);
104 if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
105 rxq->stats.rx_nombuf += n;
108 for (i = 0; i < n; ++i) {
112 * In order to support the mbufs with external attached
113 * data buffer we should use the buf_addr pointer
114 * instead of rte_mbuf_buf_addr(). It touches the mbuf
115 * itself and may impact the performance.
117 buf_addr = elts[i]->buf_addr;
118 wq[i].addr = rte_cpu_to_be_64((uintptr_t)buf_addr +
119 RTE_PKTMBUF_HEADROOM);
120 /* If there's a single MR, no need to replace LKey. */
121 if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh)
123 wq[i].lkey = mlx5_rx_mb2mr(rxq, elts[i]);
126 /* Prevent overflowing into consumed mbufs. */
127 elts_idx = rxq->rq_ci & q_mask;
128 for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
129 (*rxq->elts)[elts_idx + i] = &rxq->fake_mbuf;
131 *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
136 * Replenish buffers for MPRQ RX in bulk.
139 * Pointer to RX queue structure.
142 mlx5_rx_mprq_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq)
144 const uint16_t wqe_n = 1 << rxq->elts_n;
145 const uint32_t strd_n = 1 << rxq->strd_num_n;
146 const uint32_t elts_n = wqe_n * strd_n;
147 const uint32_t wqe_mask = elts_n - 1;
148 uint32_t n = elts_n - (rxq->elts_ci - rxq->rq_pi);
149 uint32_t elts_idx = rxq->elts_ci & wqe_mask;
150 struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
152 /* Not to cross queue end. */
153 if (n >= rxq->rq_repl_thresh) {
154 MLX5_ASSERT(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(elts_n));
155 MLX5_ASSERT(MLX5_VPMD_RXQ_RPLNSH_THRESH(elts_n) >
156 MLX5_VPMD_DESCS_PER_LOOP);
157 n = RTE_MIN(n, elts_n - elts_idx);
158 if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
159 rxq->stats.rx_nombuf += n;
167 * Copy or attach MPRQ buffers to RX SW ring.
170 * Pointer to RX queue structure.
172 * Pointer to array of packets to be stored.
174 * Number of packets to be stored.
177 * Number of packets successfully copied/attached (<= pkts_n).
179 static inline uint16_t
180 rxq_copy_mprq_mbuf_v(struct mlx5_rxq_data *rxq,
181 struct rte_mbuf **pkts, uint16_t pkts_n)
183 const uint16_t wqe_n = 1 << rxq->elts_n;
184 const uint16_t wqe_mask = wqe_n - 1;
185 const uint16_t strd_sz = 1 << rxq->strd_sz_n;
186 const uint32_t strd_n = 1 << rxq->strd_num_n;
187 const uint32_t elts_n = wqe_n * strd_n;
188 const uint32_t elts_mask = elts_n - 1;
189 uint32_t elts_idx = rxq->rq_pi & elts_mask;
190 struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
191 uint32_t rq_ci = rxq->rq_ci;
192 struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wqe_mask];
196 for (i = 0; i < pkts_n; ++i) {
198 enum mlx5_rqx_code rxq_code;
200 if (rxq->consumed_strd == strd_n) {
201 /* Replace WQE if the buffer is still in use. */
202 mprq_buf_replace(rxq, rq_ci & wqe_mask);
203 /* Advance to the next WQE. */
204 rxq->consumed_strd = 0;
206 buf = (*rxq->mprq_bufs)[rq_ci & wqe_mask];
209 if (!elts[i]->pkt_len) {
210 rxq->consumed_strd = strd_n;
211 rte_pktmbuf_free_seg(elts[i]);
212 #ifdef MLX5_PMD_SOFT_COUNTERS
213 rxq->stats.ipackets -= 1;
217 strd_cnt = (elts[i]->pkt_len / strd_sz) +
218 ((elts[i]->pkt_len % strd_sz) ? 1 : 0);
219 rxq_code = mprq_buf_to_pkt(rxq, elts[i], elts[i]->pkt_len,
220 buf, rxq->consumed_strd, strd_cnt);
221 rxq->consumed_strd += strd_cnt;
222 if (unlikely(rxq_code != MLX5_RXQ_CODE_EXIT)) {
223 rte_pktmbuf_free_seg(elts[i]);
224 #ifdef MLX5_PMD_SOFT_COUNTERS
225 rxq->stats.ipackets -= 1;
226 rxq->stats.ibytes -= elts[i]->pkt_len;
228 if (rxq_code == MLX5_RXQ_CODE_NOMBUF) {
229 ++rxq->stats.rx_nombuf;
232 if (rxq_code == MLX5_RXQ_CODE_DROPPED) {
233 ++rxq->stats.idropped;
237 pkts[copied++] = elts[i];
242 *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
243 if (rq_ci != rxq->rq_ci) {
246 *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
252 * Receive burst of packets. An errored completion also consumes a mbuf, but the
253 * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
254 * before returning to application.
257 * Pointer to RX queue structure.
259 * Array to store received packets.
261 * Maximum number of packets in array.
263 * Pointer to a flag. Set non-zero value if pkts array has at least one error
266 * Pointer to a boolean. Set true if no new CQE seen.
269 * Number of packets received including errors (<= pkts_n).
271 static inline uint16_t
272 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
273 uint16_t pkts_n, uint64_t *err, bool *no_cq)
275 const uint16_t q_n = 1 << rxq->cqe_n;
276 const uint16_t q_mask = q_n - 1;
277 const uint16_t e_n = 1 << rxq->elts_n;
278 const uint16_t e_mask = e_n - 1;
279 volatile struct mlx5_cqe *cq;
280 struct rte_mbuf **elts;
281 uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
282 uint16_t nocmp_n = 0;
283 uint16_t rcvd_pkt = 0;
284 unsigned int cq_idx = rxq->cq_ci & q_mask;
285 unsigned int elts_idx;
287 MLX5_ASSERT(rxq->sges_n == 0);
288 MLX5_ASSERT(rxq->cqe_n == rxq->elts_n);
289 cq = &(*rxq->cqes)[cq_idx];
291 rte_prefetch0(cq + 1);
292 rte_prefetch0(cq + 2);
293 rte_prefetch0(cq + 3);
294 pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
295 mlx5_rx_replenish_bulk_mbuf(rxq);
296 /* See if there're unreturned mbufs from compressed CQE. */
297 rcvd_pkt = rxq->decompressed;
299 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
300 rxq_copy_mbuf_v(&(*rxq->elts)[rxq->rq_pi & e_mask],
302 rxq->rq_pi += rcvd_pkt;
303 rxq->decompressed -= rcvd_pkt;
306 elts_idx = rxq->rq_pi & e_mask;
307 elts = &(*rxq->elts)[elts_idx];
308 /* Not to overflow pkts array. */
309 pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
310 /* Not to cross queue end. */
311 pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
312 pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
317 /* At this point, there shouldn't be any remaining packets. */
318 MLX5_ASSERT(rxq->decompressed == 0);
319 /* Process all the CQEs */
320 nocmp_n = rxq_cq_process_v(rxq, cq, elts, pkts, pkts_n, err, &comp_idx);
321 /* If no new CQE seen, return without updating cq_db. */
322 if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
326 /* Update the consumer indexes for non-compressed CQEs. */
327 MLX5_ASSERT(nocmp_n <= pkts_n);
328 rxq->cq_ci += nocmp_n;
329 rxq->rq_pi += nocmp_n;
331 /* Decompress the last CQE if compressed. */
332 if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP) {
333 MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
334 rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
336 rxq->cq_ci += rxq->decompressed;
337 /* Return more packets if needed. */
338 if (nocmp_n < pkts_n) {
339 uint16_t n = rxq->decompressed;
341 n = RTE_MIN(n, pkts_n - nocmp_n);
342 rxq_copy_mbuf_v(&(*rxq->elts)[rxq->rq_pi & e_mask],
346 rxq->decompressed -= n;
350 *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
356 * DPDK callback for vectorized RX.
359 * Generic pointer to RX queue structure.
361 * Array to store received packets.
363 * Maximum number of packets in array.
366 * Number of packets successfully received (<= pkts_n).
369 mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
371 struct mlx5_rxq_data *rxq = dpdk_rxq;
378 nb_rx = rxq_burst_v(rxq, pkts + tn, pkts_n - tn,
380 if (unlikely(err | rxq->err_state))
381 nb_rx = rxq_handle_pending_error(rxq, pkts + tn, nb_rx);
385 } while (tn != pkts_n);
390 * Receive burst of packets. An errored completion also consumes a mbuf, but the
391 * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
392 * before returning to application.
395 * Pointer to RX queue structure.
397 * Array to store received packets.
399 * Maximum number of packets in array.
401 * Pointer to a flag. Set non-zero value if pkts array has at least one error
404 * Pointer to a boolean. Set true if no new CQE seen.
407 * Number of packets received including errors (<= pkts_n).
409 static inline uint16_t
410 rxq_burst_mprq_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
411 uint16_t pkts_n, uint64_t *err, bool *no_cq)
413 const uint16_t q_n = 1 << rxq->cqe_n;
414 const uint16_t q_mask = q_n - 1;
415 const uint16_t wqe_n = 1 << rxq->elts_n;
416 const uint32_t strd_n = 1 << rxq->strd_num_n;
417 const uint32_t elts_n = wqe_n * strd_n;
418 const uint32_t elts_mask = elts_n - 1;
419 volatile struct mlx5_cqe *cq;
420 struct rte_mbuf **elts;
421 uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
422 uint16_t nocmp_n = 0;
423 uint16_t rcvd_pkt = 0;
425 unsigned int cq_idx = rxq->cq_ci & q_mask;
426 unsigned int elts_idx;
428 MLX5_ASSERT(rxq->sges_n == 0);
429 cq = &(*rxq->cqes)[cq_idx];
431 rte_prefetch0(cq + 1);
432 rte_prefetch0(cq + 2);
433 rte_prefetch0(cq + 3);
434 pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
435 mlx5_rx_mprq_replenish_bulk_mbuf(rxq);
436 /* See if there're unreturned mbufs from compressed CQE. */
437 rcvd_pkt = rxq->decompressed;
439 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
440 cp_pkt = rxq_copy_mprq_mbuf_v(rxq, pkts, rcvd_pkt);
441 rxq->decompressed -= rcvd_pkt;
444 elts_idx = rxq->rq_pi & elts_mask;
445 elts = &(*rxq->elts)[elts_idx];
446 /* Not to overflow pkts array. */
447 pkts_n = RTE_ALIGN_FLOOR(pkts_n - cp_pkt, MLX5_VPMD_DESCS_PER_LOOP);
448 /* Not to cross queue end. */
449 pkts_n = RTE_MIN(pkts_n, elts_n - elts_idx);
450 pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
451 /* Not to move past the allocated mbufs. */
452 pkts_n = RTE_MIN(pkts_n, rxq->elts_ci - rxq->rq_pi);
457 /* At this point, there shouldn't be any remaining packets. */
458 MLX5_ASSERT(rxq->decompressed == 0);
459 /* Process all the CQEs */
460 nocmp_n = rxq_cq_process_v(rxq, cq, elts, pkts, pkts_n, err, &comp_idx);
461 /* If no new CQE seen, return without updating cq_db. */
462 if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
466 /* Update the consumer indexes for non-compressed CQEs. */
467 MLX5_ASSERT(nocmp_n <= pkts_n);
468 cp_pkt = rxq_copy_mprq_mbuf_v(rxq, pkts, nocmp_n);
470 /* Decompress the last CQE if compressed. */
471 if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP) {
472 MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
473 rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
475 /* Return more packets if needed. */
476 if (nocmp_n < pkts_n) {
477 uint16_t n = rxq->decompressed;
479 n = RTE_MIN(n, pkts_n - nocmp_n);
480 cp_pkt = rxq_copy_mprq_mbuf_v(rxq, &pkts[cp_pkt], n);
482 rxq->decompressed -= n;
490 * DPDK callback for vectorized MPRQ RX.
493 * Generic pointer to RX queue structure.
495 * Array to store received packets.
497 * Maximum number of packets in array.
500 * Number of packets successfully received (<= pkts_n).
503 mlx5_rx_burst_mprq_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
505 struct mlx5_rxq_data *rxq = dpdk_rxq;
512 nb_rx = rxq_burst_mprq_v(rxq, pkts + tn, pkts_n - tn,
514 if (unlikely(err | rxq->err_state))
515 nb_rx = rxq_handle_pending_error(rxq, pkts + tn, nb_rx);
519 } while (tn != pkts_n);
524 * Check a RX queue can support vectorized RX.
527 * Pointer to RX queue.
530 * 1 if supported, negative errno value if not.
533 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq)
535 struct mlx5_rxq_ctrl *ctrl =
536 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
538 if (!ctrl->priv->config.rx_vec_en || rxq->sges_n != 0)
546 * Check a device can support vectorized RX.
549 * Pointer to Ethernet device.
552 * 1 if supported, negative errno value if not.
555 mlx5_check_vec_rx_support(struct rte_eth_dev *dev)
557 struct mlx5_priv *priv = dev->data->dev_private;
560 if (rte_vect_get_max_simd_bitwidth() < RTE_VECT_SIMD_128)
562 if (!priv->config.rx_vec_en)
564 /* All the configured queues should support. */
565 for (i = 0; i < priv->rxqs_n; ++i) {
566 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
570 if (mlx5_rxq_check_vec_support(rxq) < 0)
573 if (i != priv->rxqs_n)