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];
153 if (n >= rxq->rq_repl_thresh &&
154 rxq->elts_ci - rxq->rq_pi <= rxq->rq_repl_thresh) {
155 MLX5_ASSERT(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(elts_n));
156 MLX5_ASSERT(MLX5_VPMD_RXQ_RPLNSH_THRESH(elts_n) >
157 MLX5_VPMD_DESCS_PER_LOOP);
158 /* Not to cross queue end. */
159 n = RTE_MIN(n - MLX5_VPMD_DESCS_PER_LOOP, elts_n - elts_idx);
160 /* Limit replenish number to threshold value. */
161 n = RTE_MIN(n, rxq->rq_repl_thresh);
162 if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
163 rxq->stats.rx_nombuf += n;
167 /* Prevent overflowing into consumed mbufs. */
168 elts_idx = rxq->elts_ci & wqe_mask;
169 for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
170 (*rxq->elts)[elts_idx + i] = &rxq->fake_mbuf;
175 * Copy or attach MPRQ buffers to RX SW ring.
178 * Pointer to RX queue structure.
180 * Pointer to array of packets to be stored.
182 * Number of packets to be stored.
185 * Number of packets successfully copied/attached (<= pkts_n).
187 static inline uint16_t
188 rxq_copy_mprq_mbuf_v(struct mlx5_rxq_data *rxq,
189 struct rte_mbuf **pkts, uint16_t pkts_n)
191 const uint16_t wqe_n = 1 << rxq->elts_n;
192 const uint16_t wqe_mask = wqe_n - 1;
193 const uint16_t strd_sz = 1 << rxq->strd_sz_n;
194 const uint32_t strd_n = 1 << rxq->strd_num_n;
195 const uint32_t elts_n = wqe_n * strd_n;
196 const uint32_t elts_mask = elts_n - 1;
197 uint32_t elts_idx = rxq->rq_pi & elts_mask;
198 struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
199 uint32_t rq_ci = rxq->rq_ci;
200 struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wqe_mask];
204 for (i = 0; i < pkts_n; ++i) {
206 enum mlx5_rqx_code rxq_code;
208 if (rxq->consumed_strd == strd_n) {
209 /* Replace WQE if the buffer is still in use. */
210 mprq_buf_replace(rxq, rq_ci & wqe_mask);
211 /* Advance to the next WQE. */
212 rxq->consumed_strd = 0;
214 buf = (*rxq->mprq_bufs)[rq_ci & wqe_mask];
217 if (!elts[i]->pkt_len) {
218 rxq->consumed_strd = strd_n;
219 rte_pktmbuf_free_seg(elts[i]);
220 #ifdef MLX5_PMD_SOFT_COUNTERS
221 rxq->stats.ipackets -= 1;
225 strd_cnt = (elts[i]->pkt_len / strd_sz) +
226 ((elts[i]->pkt_len % strd_sz) ? 1 : 0);
227 rxq_code = mprq_buf_to_pkt(rxq, elts[i], elts[i]->pkt_len,
228 buf, rxq->consumed_strd, strd_cnt);
229 rxq->consumed_strd += strd_cnt;
230 if (unlikely(rxq_code != MLX5_RXQ_CODE_EXIT)) {
231 rte_pktmbuf_free_seg(elts[i]);
232 #ifdef MLX5_PMD_SOFT_COUNTERS
233 rxq->stats.ipackets -= 1;
234 rxq->stats.ibytes -= elts[i]->pkt_len;
236 if (rxq_code == MLX5_RXQ_CODE_NOMBUF) {
237 ++rxq->stats.rx_nombuf;
240 if (rxq_code == MLX5_RXQ_CODE_DROPPED) {
241 ++rxq->stats.idropped;
245 pkts[copied++] = elts[i];
250 *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
251 if (rq_ci != rxq->rq_ci) {
254 *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
260 * Receive burst of packets. An errored completion also consumes a mbuf, but the
261 * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
262 * before returning to application.
265 * Pointer to RX queue structure.
267 * Array to store received packets.
269 * Maximum number of packets in array.
271 * Pointer to a flag. Set non-zero value if pkts array has at least one error
274 * Pointer to a boolean. Set true if no new CQE seen.
277 * Number of packets received including errors (<= pkts_n).
279 static inline uint16_t
280 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
281 uint16_t pkts_n, uint64_t *err, bool *no_cq)
283 const uint16_t q_n = 1 << rxq->cqe_n;
284 const uint16_t q_mask = q_n - 1;
285 const uint16_t e_n = 1 << rxq->elts_n;
286 const uint16_t e_mask = e_n - 1;
287 volatile struct mlx5_cqe *cq;
288 struct rte_mbuf **elts;
289 uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
290 uint16_t nocmp_n = 0;
291 uint16_t rcvd_pkt = 0;
292 unsigned int cq_idx = rxq->cq_ci & q_mask;
293 unsigned int elts_idx;
295 MLX5_ASSERT(rxq->sges_n == 0);
296 MLX5_ASSERT(rxq->cqe_n == rxq->elts_n);
297 cq = &(*rxq->cqes)[cq_idx];
299 rte_prefetch0(cq + 1);
300 rte_prefetch0(cq + 2);
301 rte_prefetch0(cq + 3);
302 pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
303 mlx5_rx_replenish_bulk_mbuf(rxq);
304 /* See if there're unreturned mbufs from compressed CQE. */
305 rcvd_pkt = rxq->decompressed;
307 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
308 rxq_copy_mbuf_v(&(*rxq->elts)[rxq->rq_pi & e_mask],
310 rxq->rq_pi += rcvd_pkt;
311 rxq->decompressed -= rcvd_pkt;
314 elts_idx = rxq->rq_pi & e_mask;
315 elts = &(*rxq->elts)[elts_idx];
316 /* Not to overflow pkts array. */
317 pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
318 /* Not to cross queue end. */
319 pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
320 pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
325 /* At this point, there shouldn't be any remaining packets. */
326 MLX5_ASSERT(rxq->decompressed == 0);
327 /* Process all the CQEs */
328 nocmp_n = rxq_cq_process_v(rxq, cq, elts, pkts, pkts_n, err, &comp_idx);
329 /* If no new CQE seen, return without updating cq_db. */
330 if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
334 /* Update the consumer indexes for non-compressed CQEs. */
335 MLX5_ASSERT(nocmp_n <= pkts_n);
336 rxq->cq_ci += nocmp_n;
337 rxq->rq_pi += nocmp_n;
339 /* Decompress the last CQE if compressed. */
340 if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP) {
341 MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
342 rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
344 rxq->cq_ci += rxq->decompressed;
345 /* Return more packets if needed. */
346 if (nocmp_n < pkts_n) {
347 uint16_t n = rxq->decompressed;
349 n = RTE_MIN(n, pkts_n - nocmp_n);
350 rxq_copy_mbuf_v(&(*rxq->elts)[rxq->rq_pi & e_mask],
354 rxq->decompressed -= n;
358 *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
364 * DPDK callback for vectorized RX.
367 * Generic pointer to RX queue structure.
369 * Array to store received packets.
371 * Maximum number of packets in array.
374 * Number of packets successfully received (<= pkts_n).
377 mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
379 struct mlx5_rxq_data *rxq = dpdk_rxq;
386 nb_rx = rxq_burst_v(rxq, pkts + tn, pkts_n - tn,
388 if (unlikely(err | rxq->err_state))
389 nb_rx = rxq_handle_pending_error(rxq, pkts + tn, nb_rx);
393 } while (tn != pkts_n);
398 * Receive burst of packets. An errored completion also consumes a mbuf, but the
399 * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
400 * before returning to application.
403 * Pointer to RX queue structure.
405 * Array to store received packets.
407 * Maximum number of packets in array.
409 * Pointer to a flag. Set non-zero value if pkts array has at least one error
412 * Pointer to a boolean. Set true if no new CQE seen.
415 * Number of packets received including errors (<= pkts_n).
417 static inline uint16_t
418 rxq_burst_mprq_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
419 uint16_t pkts_n, uint64_t *err, bool *no_cq)
421 const uint16_t q_n = 1 << rxq->cqe_n;
422 const uint16_t q_mask = q_n - 1;
423 const uint16_t wqe_n = 1 << rxq->elts_n;
424 const uint32_t strd_n = 1 << rxq->strd_num_n;
425 const uint32_t elts_n = wqe_n * strd_n;
426 const uint32_t elts_mask = elts_n - 1;
427 volatile struct mlx5_cqe *cq;
428 struct rte_mbuf **elts;
429 uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
430 uint16_t nocmp_n = 0;
431 uint16_t rcvd_pkt = 0;
433 unsigned int cq_idx = rxq->cq_ci & q_mask;
434 unsigned int elts_idx;
436 MLX5_ASSERT(rxq->sges_n == 0);
437 cq = &(*rxq->cqes)[cq_idx];
439 rte_prefetch0(cq + 1);
440 rte_prefetch0(cq + 2);
441 rte_prefetch0(cq + 3);
442 pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
443 mlx5_rx_mprq_replenish_bulk_mbuf(rxq);
444 /* See if there're unreturned mbufs from compressed CQE. */
445 rcvd_pkt = rxq->decompressed;
447 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
448 cp_pkt = rxq_copy_mprq_mbuf_v(rxq, pkts, rcvd_pkt);
449 rxq->decompressed -= rcvd_pkt;
452 elts_idx = rxq->rq_pi & elts_mask;
453 elts = &(*rxq->elts)[elts_idx];
454 /* Not to overflow pkts array. */
455 pkts_n = RTE_ALIGN_FLOOR(pkts_n - cp_pkt, MLX5_VPMD_DESCS_PER_LOOP);
456 /* Not to cross queue end. */
457 pkts_n = RTE_MIN(pkts_n, elts_n - elts_idx);
458 pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
459 /* Not to move past the allocated mbufs. */
460 pkts_n = RTE_MIN(pkts_n, rxq->elts_ci - rxq->rq_pi);
465 /* At this point, there shouldn't be any remaining packets. */
466 MLX5_ASSERT(rxq->decompressed == 0);
467 /* Process all the CQEs */
468 nocmp_n = rxq_cq_process_v(rxq, cq, elts, pkts, pkts_n, err, &comp_idx);
469 /* If no new CQE seen, return without updating cq_db. */
470 if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
474 /* Update the consumer indexes for non-compressed CQEs. */
475 MLX5_ASSERT(nocmp_n <= pkts_n);
476 cp_pkt = rxq_copy_mprq_mbuf_v(rxq, pkts, nocmp_n);
478 /* Decompress the last CQE if compressed. */
479 if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP) {
480 MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
481 rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
483 /* Return more packets if needed. */
484 if (nocmp_n < pkts_n) {
485 uint16_t n = rxq->decompressed;
487 n = RTE_MIN(n, pkts_n - nocmp_n);
488 cp_pkt = rxq_copy_mprq_mbuf_v(rxq, &pkts[cp_pkt], n);
490 rxq->decompressed -= n;
498 * DPDK callback for vectorized MPRQ RX.
501 * Generic pointer to RX queue structure.
503 * Array to store received packets.
505 * Maximum number of packets in array.
508 * Number of packets successfully received (<= pkts_n).
511 mlx5_rx_burst_mprq_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
513 struct mlx5_rxq_data *rxq = dpdk_rxq;
520 nb_rx = rxq_burst_mprq_v(rxq, pkts + tn, pkts_n - tn,
522 if (unlikely(err | rxq->err_state))
523 nb_rx = rxq_handle_pending_error(rxq, pkts + tn, nb_rx);
527 } while (tn != pkts_n);
532 * Check a RX queue can support vectorized RX.
535 * Pointer to RX queue.
538 * 1 if supported, negative errno value if not.
541 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq)
543 struct mlx5_rxq_ctrl *ctrl =
544 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
546 if (!ctrl->priv->config.rx_vec_en || rxq->sges_n != 0)
554 * Check a device can support vectorized RX.
557 * Pointer to Ethernet device.
560 * 1 if supported, negative errno value if not.
563 mlx5_check_vec_rx_support(struct rte_eth_dev *dev)
565 struct mlx5_priv *priv = dev->data->dev_private;
568 if (rte_vect_get_max_simd_bitwidth() < RTE_VECT_SIMD_128)
570 if (!priv->config.rx_vec_en)
572 /* All the configured queues should support. */
573 for (i = 0; i < priv->rxqs_n; ++i) {
574 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
578 if (mlx5_rxq_check_vec_support(rxq) < 0)
581 if (i != priv->rxqs_n)