xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.h (revision a103a97e7191179ad6a451ce85182df2ecb10c26)
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 #ifndef RTE_PMD_MLX5_RXTX_H_
35 #define RTE_PMD_MLX5_RXTX_H_
36 
37 #include <stddef.h>
38 #include <stdint.h>
39 
40 /* Verbs header. */
41 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
42 #ifdef PEDANTIC
43 #pragma GCC diagnostic ignored "-Wpedantic"
44 #endif
45 #include <infiniband/verbs.h>
46 #include <infiniband/mlx5dv.h>
47 #ifdef PEDANTIC
48 #pragma GCC diagnostic error "-Wpedantic"
49 #endif
50 
51 #include <rte_mbuf.h>
52 #include <rte_mempool.h>
53 #include <rte_common.h>
54 #include <rte_hexdump.h>
55 
56 #include "mlx5_utils.h"
57 #include "mlx5.h"
58 #include "mlx5_autoconf.h"
59 #include "mlx5_defs.h"
60 #include "mlx5_prm.h"
61 
62 struct mlx5_rxq_stats {
63 	unsigned int idx; /**< Mapping index. */
64 #ifdef MLX5_PMD_SOFT_COUNTERS
65 	uint64_t ipackets; /**< Total of successfully received packets. */
66 	uint64_t ibytes; /**< Total of successfully received bytes. */
67 #endif
68 	uint64_t idropped; /**< Total of packets dropped when RX ring full. */
69 	uint64_t rx_nombuf; /**< Total of RX mbuf allocation failures. */
70 };
71 
72 struct mlx5_txq_stats {
73 	unsigned int idx; /**< Mapping index. */
74 #ifdef MLX5_PMD_SOFT_COUNTERS
75 	uint64_t opackets; /**< Total of successfully sent packets. */
76 	uint64_t obytes; /**< Total of successfully sent bytes. */
77 #endif
78 	uint64_t oerrors; /**< Total number of failed transmitted packets. */
79 };
80 
81 /* Flow director queue structure. */
82 struct fdir_queue {
83 	struct ibv_qp *qp; /* Associated RX QP. */
84 	struct ibv_rwq_ind_table *ind_table; /* Indirection table. */
85 	struct ibv_wq *wq; /* Work queue. */
86 	struct ibv_cq *cq; /* Completion queue. */
87 };
88 
89 struct priv;
90 
91 /* Compressed CQE context. */
92 struct rxq_zip {
93 	uint16_t ai; /* Array index. */
94 	uint16_t ca; /* Current array index. */
95 	uint16_t na; /* Next array index. */
96 	uint16_t cq_ci; /* The next CQE. */
97 	uint32_t cqe_cnt; /* Number of CQEs. */
98 };
99 
100 /* RX queue descriptor. */
101 struct rxq {
102 	unsigned int csum:1; /* Enable checksum offloading. */
103 	unsigned int csum_l2tun:1; /* Same for L2 tunnels. */
104 	unsigned int vlan_strip:1; /* Enable VLAN stripping. */
105 	unsigned int crc_present:1; /* CRC must be subtracted. */
106 	unsigned int sges_n:2; /* Log 2 of SGEs (max buffers per packet). */
107 	unsigned int cqe_n:4; /* Log 2 of CQ elements. */
108 	unsigned int elts_n:4; /* Log 2 of Mbufs. */
109 	unsigned int rss_hash:1; /* RSS hash result is enabled. */
110 	unsigned int mark:1; /* Marked flow available on the queue. */
111 	unsigned int pending_err:1; /* CQE error needs to be handled. */
112 	unsigned int :15; /* Remaining bits. */
113 	volatile uint32_t *rq_db;
114 	volatile uint32_t *cq_db;
115 	uint16_t port_id;
116 	uint16_t rq_ci;
117 	uint16_t rq_pi;
118 	uint16_t cq_ci;
119 	volatile struct mlx5_wqe_data_seg(*wqes)[];
120 	volatile struct mlx5_cqe(*cqes)[];
121 	struct rxq_zip zip; /* Compressed context. */
122 	struct rte_mbuf *(*elts)[];
123 	struct rte_mempool *mp;
124 	struct mlx5_rxq_stats stats;
125 	uint64_t mbuf_initializer; /* Default rearm_data for vectorized Rx. */
126 	struct rte_mbuf fake_mbuf; /* elts padding for vectorized Rx. */
127 	void *cq_uar; /* CQ user access region. */
128 	uint32_t cqn; /* CQ number. */
129 	uint8_t cq_arm_sn; /* CQ arm seq number. */
130 } __rte_cache_aligned;
131 
132 /* RX queue control descriptor. */
133 struct rxq_ctrl {
134 	struct priv *priv; /* Back pointer to private data. */
135 	struct ibv_cq *cq; /* Completion Queue. */
136 	struct ibv_wq *wq; /* Work Queue. */
137 	struct fdir_queue *fdir_queue; /* Flow director queue. */
138 	struct ibv_mr *mr; /* Memory Region (for mp). */
139 	struct ibv_comp_channel *channel;
140 	unsigned int socket; /* CPU socket ID for allocations. */
141 	struct rxq rxq; /* Data path structure. */
142 };
143 
144 /* Hash RX queue types. */
145 enum hash_rxq_type {
146 	HASH_RXQ_TCPV4,
147 	HASH_RXQ_UDPV4,
148 	HASH_RXQ_IPV4,
149 	HASH_RXQ_TCPV6,
150 	HASH_RXQ_UDPV6,
151 	HASH_RXQ_IPV6,
152 	HASH_RXQ_ETH,
153 };
154 
155 /* Flow structure with Ethernet specification. It is packed to prevent padding
156  * between attr and spec as this layout is expected by libibverbs. */
157 struct flow_attr_spec_eth {
158 	struct ibv_flow_attr attr;
159 	struct ibv_flow_spec_eth spec;
160 } __attribute__((packed));
161 
162 /* Define a struct flow_attr_spec_eth object as an array of at least
163  * "size" bytes. Room after the first index is normally used to store
164  * extra flow specifications. */
165 #define FLOW_ATTR_SPEC_ETH(name, size) \
166 	struct flow_attr_spec_eth name \
167 		[((size) / sizeof(struct flow_attr_spec_eth)) + \
168 		 !!((size) % sizeof(struct flow_attr_spec_eth))]
169 
170 /* Initialization data for hash RX queue. */
171 struct hash_rxq_init {
172 	uint64_t hash_fields; /* Fields that participate in the hash. */
173 	uint64_t dpdk_rss_hf; /* Matching DPDK RSS hash fields. */
174 	unsigned int flow_priority; /* Flow priority to use. */
175 	union {
176 		struct {
177 			enum ibv_flow_spec_type type;
178 			uint16_t size;
179 		} hdr;
180 		struct ibv_flow_spec_tcp_udp tcp_udp;
181 		struct ibv_flow_spec_ipv4 ipv4;
182 		struct ibv_flow_spec_ipv6 ipv6;
183 		struct ibv_flow_spec_eth eth;
184 	} flow_spec; /* Flow specification template. */
185 	const struct hash_rxq_init *underlayer; /* Pointer to underlayer. */
186 };
187 
188 /* Initialization data for indirection table. */
189 struct ind_table_init {
190 	unsigned int max_size; /* Maximum number of WQs. */
191 	/* Hash RX queues using this table. */
192 	unsigned int hash_types;
193 	unsigned int hash_types_n;
194 };
195 
196 /* Initialization data for special flows. */
197 struct special_flow_init {
198 	uint8_t dst_mac_val[6];
199 	uint8_t dst_mac_mask[6];
200 	unsigned int hash_types;
201 	unsigned int per_vlan:1;
202 };
203 
204 enum hash_rxq_flow_type {
205 	HASH_RXQ_FLOW_TYPE_PROMISC,
206 	HASH_RXQ_FLOW_TYPE_ALLMULTI,
207 	HASH_RXQ_FLOW_TYPE_BROADCAST,
208 	HASH_RXQ_FLOW_TYPE_IPV6MULTI,
209 	HASH_RXQ_FLOW_TYPE_MAC,
210 };
211 
212 #ifndef NDEBUG
213 static inline const char *
214 hash_rxq_flow_type_str(enum hash_rxq_flow_type flow_type)
215 {
216 	switch (flow_type) {
217 	case HASH_RXQ_FLOW_TYPE_PROMISC:
218 		return "promiscuous";
219 	case HASH_RXQ_FLOW_TYPE_ALLMULTI:
220 		return "allmulticast";
221 	case HASH_RXQ_FLOW_TYPE_BROADCAST:
222 		return "broadcast";
223 	case HASH_RXQ_FLOW_TYPE_IPV6MULTI:
224 		return "IPv6 multicast";
225 	case HASH_RXQ_FLOW_TYPE_MAC:
226 		return "MAC";
227 	}
228 	return NULL;
229 }
230 #endif /* NDEBUG */
231 
232 struct hash_rxq {
233 	struct priv *priv; /* Back pointer to private data. */
234 	struct ibv_qp *qp; /* Hash RX QP. */
235 	enum hash_rxq_type type; /* Hash RX queue type. */
236 	/* MAC flow steering rules, one per VLAN ID. */
237 	struct ibv_flow *mac_flow
238 		[MLX5_MAX_MAC_ADDRESSES][MLX5_MAX_VLAN_IDS];
239 	struct ibv_flow *special_flow
240 		[MLX5_MAX_SPECIAL_FLOWS][MLX5_MAX_VLAN_IDS];
241 };
242 
243 /* TX queue descriptor. */
244 __extension__
245 struct txq {
246 	uint16_t elts_head; /* Current counter in (*elts)[]. */
247 	uint16_t elts_tail; /* Counter of first element awaiting completion. */
248 	uint16_t elts_comp; /* Counter since last completion request. */
249 	uint16_t mpw_comp; /* WQ index since last completion request. */
250 	uint16_t cq_ci; /* Consumer index for completion queue. */
251 	uint16_t cq_pi; /* Producer index for completion queue. */
252 	uint16_t wqe_ci; /* Consumer index for work queue. */
253 	uint16_t wqe_pi; /* Producer index for work queue. */
254 	uint16_t elts_n:4; /* (*elts)[] length (in log2). */
255 	uint16_t cqe_n:4; /* Number of CQ elements (in log2). */
256 	uint16_t wqe_n:4; /* Number of of WQ elements (in log2). */
257 	uint16_t inline_en:1; /* When set inline is enabled. */
258 	uint16_t tso_en:1; /* When set hardware TSO is enabled. */
259 	uint16_t tunnel_en:1;
260 	/* When set TX offload for tunneled packets are supported. */
261 	uint16_t mpw_hdr_dseg:1; /* Enable DSEGs in the title WQEBB. */
262 	uint16_t max_inline; /* Multiple of RTE_CACHE_LINE_SIZE to inline. */
263 	uint16_t inline_max_packet_sz; /* Max packet size for inlining. */
264 	uint32_t qp_num_8s; /* QP number shifted by 8. */
265 	uint32_t flags; /* Flags for Tx Queue. */
266 	volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
267 	volatile void *wqes; /* Work queue (use volatile to write into). */
268 	volatile uint32_t *qp_db; /* Work queue doorbell. */
269 	volatile uint32_t *cq_db; /* Completion queue doorbell. */
270 	volatile void *bf_reg; /* Blueflame register. */
271 	struct {
272 		uintptr_t start; /* Start address of MR */
273 		uintptr_t end; /* End address of MR */
274 		struct ibv_mr *mr; /* Memory Region (for mp). */
275 		uint32_t lkey; /* rte_cpu_to_be_32(mr->lkey) */
276 	} mp2mr[MLX5_PMD_TX_MP_CACHE]; /* MP to MR translation table. */
277 	uint16_t mr_cache_idx; /* Index of last hit entry. */
278 	struct rte_mbuf *(*elts)[]; /* TX elements. */
279 	struct mlx5_txq_stats stats; /* TX queue counters. */
280 } __rte_cache_aligned;
281 
282 /* TX queue control descriptor. */
283 struct txq_ctrl {
284 	struct priv *priv; /* Back pointer to private data. */
285 	struct ibv_cq *cq; /* Completion Queue. */
286 	struct ibv_qp *qp; /* Queue Pair. */
287 	unsigned int socket; /* CPU socket ID for allocations. */
288 	struct txq txq; /* Data path structure. */
289 };
290 
291 /* mlx5_rxq.c */
292 
293 extern const struct hash_rxq_init hash_rxq_init[];
294 extern const unsigned int hash_rxq_init_n;
295 
296 extern uint8_t rss_hash_default_key[];
297 extern const size_t rss_hash_default_key_len;
298 
299 size_t priv_flow_attr(struct priv *, struct ibv_flow_attr *,
300 		      size_t, enum hash_rxq_type);
301 int priv_create_hash_rxqs(struct priv *);
302 void priv_destroy_hash_rxqs(struct priv *);
303 int priv_allow_flow_type(struct priv *, enum hash_rxq_flow_type);
304 int priv_rehash_flows(struct priv *);
305 void rxq_cleanup(struct rxq_ctrl *);
306 int mlx5_rx_queue_setup(struct rte_eth_dev *, uint16_t, uint16_t, unsigned int,
307 			const struct rte_eth_rxconf *, struct rte_mempool *);
308 void mlx5_rx_queue_release(void *);
309 int priv_rx_intr_vec_enable(struct priv *priv);
310 void priv_rx_intr_vec_disable(struct priv *priv);
311 int mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
312 int mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
313 
314 /* mlx5_txq.c */
315 
316 void txq_cleanup(struct txq_ctrl *);
317 int txq_ctrl_setup(struct rte_eth_dev *, struct txq_ctrl *, uint16_t,
318 		   unsigned int, const struct rte_eth_txconf *);
319 int mlx5_tx_queue_setup(struct rte_eth_dev *, uint16_t, uint16_t, unsigned int,
320 			const struct rte_eth_txconf *);
321 void mlx5_tx_queue_release(void *);
322 
323 /* mlx5_rxtx.c */
324 
325 extern uint32_t mlx5_ptype_table[];
326 
327 void mlx5_set_ptype_table(void);
328 uint16_t mlx5_tx_burst(void *, struct rte_mbuf **, uint16_t);
329 uint16_t mlx5_tx_burst_mpw(void *, struct rte_mbuf **, uint16_t);
330 uint16_t mlx5_tx_burst_mpw_inline(void *, struct rte_mbuf **, uint16_t);
331 uint16_t mlx5_tx_burst_empw(void *, struct rte_mbuf **, uint16_t);
332 uint16_t mlx5_rx_burst(void *, struct rte_mbuf **, uint16_t);
333 uint16_t removed_tx_burst(void *, struct rte_mbuf **, uint16_t);
334 uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
335 int mlx5_rx_descriptor_status(void *, uint16_t);
336 int mlx5_tx_descriptor_status(void *, uint16_t);
337 
338 /* Vectorized version of mlx5_rxtx.c */
339 int priv_check_raw_vec_tx_support(struct priv *);
340 int priv_check_vec_tx_support(struct priv *);
341 int rxq_check_vec_support(struct rxq *);
342 int priv_check_vec_rx_support(struct priv *);
343 uint16_t mlx5_tx_burst_raw_vec(void *, struct rte_mbuf **, uint16_t);
344 uint16_t mlx5_tx_burst_vec(void *, struct rte_mbuf **, uint16_t);
345 uint16_t mlx5_rx_burst_vec(void *, struct rte_mbuf **, uint16_t);
346 
347 /* mlx5_mr.c */
348 
349 struct ibv_mr *mlx5_mp2mr(struct ibv_pd *, struct rte_mempool *);
350 void txq_mp2mr_iter(struct rte_mempool *, void *);
351 uint32_t txq_mp2mr_reg(struct txq *, struct rte_mempool *, unsigned int);
352 
353 #ifndef NDEBUG
354 /**
355  * Verify or set magic value in CQE.
356  *
357  * @param cqe
358  *   Pointer to CQE.
359  *
360  * @return
361  *   0 the first time.
362  */
363 static inline int
364 check_cqe_seen(volatile struct mlx5_cqe *cqe)
365 {
366 	static const uint8_t magic[] = "seen";
367 	volatile uint8_t (*buf)[sizeof(cqe->rsvd0)] = &cqe->rsvd0;
368 	int ret = 1;
369 	unsigned int i;
370 
371 	for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
372 		if (!ret || (*buf)[i] != magic[i]) {
373 			ret = 0;
374 			(*buf)[i] = magic[i];
375 		}
376 	return ret;
377 }
378 #endif /* NDEBUG */
379 
380 /**
381  * Check whether CQE is valid.
382  *
383  * @param cqe
384  *   Pointer to CQE.
385  * @param cqes_n
386  *   Size of completion queue.
387  * @param ci
388  *   Consumer index.
389  *
390  * @return
391  *   0 on success, 1 on failure.
392  */
393 static __rte_always_inline int
394 check_cqe(volatile struct mlx5_cqe *cqe,
395 	  unsigned int cqes_n, const uint16_t ci)
396 {
397 	uint16_t idx = ci & cqes_n;
398 	uint8_t op_own = cqe->op_own;
399 	uint8_t op_owner = MLX5_CQE_OWNER(op_own);
400 	uint8_t op_code = MLX5_CQE_OPCODE(op_own);
401 
402 	if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
403 		return 1; /* No CQE. */
404 #ifndef NDEBUG
405 	if ((op_code == MLX5_CQE_RESP_ERR) ||
406 	    (op_code == MLX5_CQE_REQ_ERR)) {
407 		volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
408 		uint8_t syndrome = err_cqe->syndrome;
409 
410 		if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
411 		    (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
412 			return 0;
413 		if (!check_cqe_seen(cqe)) {
414 			ERROR("unexpected CQE error %u (0x%02x)"
415 			      " syndrome 0x%02x",
416 			      op_code, op_code, syndrome);
417 			rte_hexdump(stderr, "MLX5 Error CQE:",
418 				    (const void *)((uintptr_t)err_cqe),
419 				    sizeof(*err_cqe));
420 		}
421 		return 1;
422 	} else if ((op_code != MLX5_CQE_RESP_SEND) &&
423 		   (op_code != MLX5_CQE_REQ)) {
424 		if (!check_cqe_seen(cqe)) {
425 			ERROR("unexpected CQE opcode %u (0x%02x)",
426 			      op_code, op_code);
427 			rte_hexdump(stderr, "MLX5 CQE:",
428 				    (const void *)((uintptr_t)cqe),
429 				    sizeof(*cqe));
430 		}
431 		return 1;
432 	}
433 #endif /* NDEBUG */
434 	return 0;
435 }
436 
437 /**
438  * Return the address of the WQE.
439  *
440  * @param txq
441  *   Pointer to TX queue structure.
442  * @param  wqe_ci
443  *   WQE consumer index.
444  *
445  * @return
446  *   WQE address.
447  */
448 static inline uintptr_t *
449 tx_mlx5_wqe(struct txq *txq, uint16_t ci)
450 {
451 	ci &= ((1 << txq->wqe_n) - 1);
452 	return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
453 }
454 
455 /**
456  * Manage TX completions.
457  *
458  * When sending a burst, mlx5_tx_burst() posts several WRs.
459  *
460  * @param txq
461  *   Pointer to TX queue structure.
462  */
463 static __rte_always_inline void
464 mlx5_tx_complete(struct txq *txq)
465 {
466 	const uint16_t elts_n = 1 << txq->elts_n;
467 	const uint16_t elts_m = elts_n - 1;
468 	const unsigned int cqe_n = 1 << txq->cqe_n;
469 	const unsigned int cqe_cnt = cqe_n - 1;
470 	uint16_t elts_free = txq->elts_tail;
471 	uint16_t elts_tail;
472 	uint16_t cq_ci = txq->cq_ci;
473 	volatile struct mlx5_cqe *cqe = NULL;
474 	volatile struct mlx5_wqe_ctrl *ctrl;
475 	struct rte_mbuf *m, *free[elts_n];
476 	struct rte_mempool *pool = NULL;
477 	unsigned int blk_n = 0;
478 
479 	cqe = &(*txq->cqes)[cq_ci & cqe_cnt];
480 	if (unlikely(check_cqe(cqe, cqe_n, cq_ci)))
481 		return;
482 #ifndef NDEBUG
483 	if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
484 	    (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
485 		if (!check_cqe_seen(cqe)) {
486 			ERROR("unexpected error CQE, TX stopped");
487 			rte_hexdump(stderr, "MLX5 TXQ:",
488 				    (const void *)((uintptr_t)txq->wqes),
489 				    ((1 << txq->wqe_n) *
490 				     MLX5_WQE_SIZE));
491 		}
492 		return;
493 	}
494 #endif /* NDEBUG */
495 	++cq_ci;
496 	txq->wqe_pi = rte_be_to_cpu_16(cqe->wqe_counter);
497 	ctrl = (volatile struct mlx5_wqe_ctrl *)
498 		tx_mlx5_wqe(txq, txq->wqe_pi);
499 	elts_tail = ctrl->ctrl3;
500 	assert((elts_tail & elts_m) < (1 << txq->wqe_n));
501 	/* Free buffers. */
502 	while (elts_free != elts_tail) {
503 		m = rte_pktmbuf_prefree_seg((*txq->elts)[elts_free++ & elts_m]);
504 		if (likely(m != NULL)) {
505 			if (likely(m->pool == pool)) {
506 				free[blk_n++] = m;
507 			} else {
508 				if (likely(pool != NULL))
509 					rte_mempool_put_bulk(pool,
510 							     (void *)free,
511 							     blk_n);
512 				free[0] = m;
513 				pool = m->pool;
514 				blk_n = 1;
515 			}
516 		}
517 	}
518 	if (blk_n)
519 		rte_mempool_put_bulk(pool, (void *)free, blk_n);
520 #ifndef NDEBUG
521 	elts_free = txq->elts_tail;
522 	/* Poisoning. */
523 	while (elts_free != elts_tail) {
524 		memset(&(*txq->elts)[elts_free & elts_m],
525 		       0x66,
526 		       sizeof((*txq->elts)[elts_free & elts_m]));
527 		++elts_free;
528 	}
529 #endif
530 	txq->cq_ci = cq_ci;
531 	txq->elts_tail = elts_tail;
532 	/* Update the consumer index. */
533 	rte_wmb();
534 	*txq->cq_db = rte_cpu_to_be_32(cq_ci);
535 }
536 
537 /**
538  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
539  * the cloned mbuf is allocated is returned instead.
540  *
541  * @param buf
542  *   Pointer to mbuf.
543  *
544  * @return
545  *   Memory pool where data is located for given mbuf.
546  */
547 static struct rte_mempool *
548 mlx5_tx_mb2mp(struct rte_mbuf *buf)
549 {
550 	if (unlikely(RTE_MBUF_INDIRECT(buf)))
551 		return rte_mbuf_from_indirect(buf)->pool;
552 	return buf->pool;
553 }
554 
555 /**
556  * Get Memory Region (MR) <-> rte_mbuf association from txq->mp2mr[].
557  * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
558  * remove an entry first.
559  *
560  * @param txq
561  *   Pointer to TX queue structure.
562  * @param[in] mp
563  *   Memory Pool for which a Memory Region lkey must be returned.
564  *
565  * @return
566  *   mr->lkey on success, (uint32_t)-1 on failure.
567  */
568 static __rte_always_inline uint32_t
569 mlx5_tx_mb2mr(struct txq *txq, struct rte_mbuf *mb)
570 {
571 	uint16_t i = txq->mr_cache_idx;
572 	uintptr_t addr = rte_pktmbuf_mtod(mb, uintptr_t);
573 
574 	assert(i < RTE_DIM(txq->mp2mr));
575 	if (likely(txq->mp2mr[i].start <= addr && txq->mp2mr[i].end >= addr))
576 		return txq->mp2mr[i].lkey;
577 	for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
578 		if (unlikely(txq->mp2mr[i].mr == NULL)) {
579 			/* Unknown MP, add a new MR for it. */
580 			break;
581 		}
582 		if (txq->mp2mr[i].start <= addr &&
583 		    txq->mp2mr[i].end >= addr) {
584 			assert(txq->mp2mr[i].lkey != (uint32_t)-1);
585 			assert(rte_cpu_to_be_32(txq->mp2mr[i].mr->lkey) ==
586 			       txq->mp2mr[i].lkey);
587 			txq->mr_cache_idx = i;
588 			return txq->mp2mr[i].lkey;
589 		}
590 	}
591 	txq->mr_cache_idx = 0;
592 	return txq_mp2mr_reg(txq, mlx5_tx_mb2mp(mb), i);
593 }
594 
595 /**
596  * Ring TX queue doorbell.
597  *
598  * @param txq
599  *   Pointer to TX queue structure.
600  * @param wqe
601  *   Pointer to the last WQE posted in the NIC.
602  */
603 static __rte_always_inline void
604 mlx5_tx_dbrec(struct txq *txq, volatile struct mlx5_wqe *wqe)
605 {
606 	uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
607 	volatile uint64_t *src = ((volatile uint64_t *)wqe);
608 
609 	rte_io_wmb();
610 	*txq->qp_db = rte_cpu_to_be_32(txq->wqe_ci);
611 	/* Ensure ordering between DB record and BF copy. */
612 	rte_wmb();
613 	*dst = *src;
614 }
615 
616 #endif /* RTE_PMD_MLX5_RXTX_H_ */
617