xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.h (revision 89f0711f9ddfb5822da9d34f384b92f72a61c4dc)
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 #include <sys/queue.h>
40 
41 /* Verbs header. */
42 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
43 #ifdef PEDANTIC
44 #pragma GCC diagnostic ignored "-Wpedantic"
45 #endif
46 #include <infiniband/verbs.h>
47 #include <infiniband/mlx5dv.h>
48 #ifdef PEDANTIC
49 #pragma GCC diagnostic error "-Wpedantic"
50 #endif
51 
52 #include <rte_mbuf.h>
53 #include <rte_mempool.h>
54 #include <rte_common.h>
55 #include <rte_hexdump.h>
56 #include <rte_atomic.h>
57 
58 #include "mlx5_utils.h"
59 #include "mlx5.h"
60 #include "mlx5_autoconf.h"
61 #include "mlx5_defs.h"
62 #include "mlx5_prm.h"
63 
64 struct mlx5_rxq_stats {
65 	unsigned int idx; /**< Mapping index. */
66 #ifdef MLX5_PMD_SOFT_COUNTERS
67 	uint64_t ipackets; /**< Total of successfully received packets. */
68 	uint64_t ibytes; /**< Total of successfully received bytes. */
69 #endif
70 	uint64_t idropped; /**< Total of packets dropped when RX ring full. */
71 	uint64_t rx_nombuf; /**< Total of RX mbuf allocation failures. */
72 };
73 
74 struct mlx5_txq_stats {
75 	unsigned int idx; /**< Mapping index. */
76 #ifdef MLX5_PMD_SOFT_COUNTERS
77 	uint64_t opackets; /**< Total of successfully sent packets. */
78 	uint64_t obytes; /**< Total of successfully sent bytes. */
79 #endif
80 	uint64_t oerrors; /**< Total number of failed transmitted packets. */
81 };
82 
83 struct priv;
84 
85 /* Memory region queue object. */
86 struct mlx5_mr {
87 	LIST_ENTRY(mlx5_mr) next; /**< Pointer to the next element. */
88 	rte_atomic32_t refcnt; /*<< Reference counter. */
89 	uint32_t lkey; /*<< rte_cpu_to_be_32(mr->lkey) */
90 	uintptr_t start; /* Start address of MR */
91 	uintptr_t end; /* End address of MR */
92 	struct ibv_mr *mr; /*<< Memory Region. */
93 	struct rte_mempool *mp; /*<< Memory Pool. */
94 };
95 
96 /* Compressed CQE context. */
97 struct rxq_zip {
98 	uint16_t ai; /* Array index. */
99 	uint16_t ca; /* Current array index. */
100 	uint16_t na; /* Next array index. */
101 	uint16_t cq_ci; /* The next CQE. */
102 	uint32_t cqe_cnt; /* Number of CQEs. */
103 };
104 
105 /* RX queue descriptor. */
106 struct mlx5_rxq_data {
107 	unsigned int csum:1; /* Enable checksum offloading. */
108 	unsigned int csum_l2tun:1; /* Same for L2 tunnels. */
109 	unsigned int hw_timestamp:1; /* Enable HW timestamp. */
110 	unsigned int vlan_strip:1; /* Enable VLAN stripping. */
111 	unsigned int crc_present:1; /* CRC must be subtracted. */
112 	unsigned int sges_n:2; /* Log 2 of SGEs (max buffers per packet). */
113 	unsigned int cqe_n:4; /* Log 2 of CQ elements. */
114 	unsigned int elts_n:4; /* Log 2 of Mbufs. */
115 	unsigned int rss_hash:1; /* RSS hash result is enabled. */
116 	unsigned int mark:1; /* Marked flow available on the queue. */
117 	unsigned int :15; /* Remaining bits. */
118 	volatile uint32_t *rq_db;
119 	volatile uint32_t *cq_db;
120 	uint16_t port_id;
121 	uint16_t rq_ci;
122 	uint16_t rq_pi;
123 	uint16_t cq_ci;
124 	volatile struct mlx5_wqe_data_seg(*wqes)[];
125 	volatile struct mlx5_cqe(*cqes)[];
126 	struct rxq_zip zip; /* Compressed context. */
127 	struct rte_mbuf *(*elts)[];
128 	struct rte_mempool *mp;
129 	struct mlx5_rxq_stats stats;
130 	uint64_t mbuf_initializer; /* Default rearm_data for vectorized Rx. */
131 	struct rte_mbuf fake_mbuf; /* elts padding for vectorized Rx. */
132 	void *cq_uar; /* CQ user access region. */
133 	uint32_t cqn; /* CQ number. */
134 	uint8_t cq_arm_sn; /* CQ arm seq number. */
135 } __rte_cache_aligned;
136 
137 /* Verbs Rx queue elements. */
138 struct mlx5_rxq_ibv {
139 	LIST_ENTRY(mlx5_rxq_ibv) next; /* Pointer to the next element. */
140 	rte_atomic32_t refcnt; /* Reference counter. */
141 	struct mlx5_rxq_ctrl *rxq_ctrl; /* Back pointer to parent. */
142 	struct ibv_cq *cq; /* Completion Queue. */
143 	struct ibv_wq *wq; /* Work Queue. */
144 	struct ibv_comp_channel *channel;
145 	struct mlx5_mr *mr; /* Memory Region (for mp). */
146 };
147 
148 /* RX queue control descriptor. */
149 struct mlx5_rxq_ctrl {
150 	LIST_ENTRY(mlx5_rxq_ctrl) next; /* Pointer to the next element. */
151 	rte_atomic32_t refcnt; /* Reference counter. */
152 	struct priv *priv; /* Back pointer to private data. */
153 	struct mlx5_rxq_ibv *ibv; /* Verbs elements. */
154 	struct mlx5_rxq_data rxq; /* Data path structure. */
155 	unsigned int socket; /* CPU socket ID for allocations. */
156 	unsigned int irq:1; /* Whether IRQ is enabled. */
157 };
158 
159 /* Indirection table. */
160 struct mlx5_ind_table_ibv {
161 	LIST_ENTRY(mlx5_ind_table_ibv) next; /* Pointer to the next element. */
162 	rte_atomic32_t refcnt; /* Reference counter. */
163 	struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
164 	uint16_t queues_n; /**< Number of queues in the list. */
165 	uint16_t queues[]; /**< Queue list. */
166 };
167 
168 /* Hash Rx queue. */
169 struct mlx5_hrxq {
170 	LIST_ENTRY(mlx5_hrxq) next; /* Pointer to the next element. */
171 	rte_atomic32_t refcnt; /* Reference counter. */
172 	struct mlx5_ind_table_ibv *ind_table; /* Indirection table. */
173 	struct ibv_qp *qp; /* Verbs queue pair. */
174 	uint64_t hash_fields; /* Verbs Hash fields. */
175 	uint8_t rss_key_len; /* Hash key length in bytes. */
176 	uint8_t rss_key[]; /* Hash key. */
177 };
178 
179 /* TX queue descriptor. */
180 __extension__
181 struct mlx5_txq_data {
182 	uint16_t elts_head; /* Current counter in (*elts)[]. */
183 	uint16_t elts_tail; /* Counter of first element awaiting completion. */
184 	uint16_t elts_comp; /* Counter since last completion request. */
185 	uint16_t mpw_comp; /* WQ index since last completion request. */
186 	uint16_t cq_ci; /* Consumer index for completion queue. */
187 #ifndef NDEBUG
188 	uint16_t cq_pi; /* Producer index for completion queue. */
189 #endif
190 	uint16_t wqe_ci; /* Consumer index for work queue. */
191 	uint16_t wqe_pi; /* Producer index for work queue. */
192 	uint16_t elts_n:4; /* (*elts)[] length (in log2). */
193 	uint16_t cqe_n:4; /* Number of CQ elements (in log2). */
194 	uint16_t wqe_n:4; /* Number of of WQ elements (in log2). */
195 	uint16_t tso_en:1; /* When set hardware TSO is enabled. */
196 	uint16_t tunnel_en:1;
197 	/* When set TX offload for tunneled packets are supported. */
198 	uint16_t mpw_hdr_dseg:1; /* Enable DSEGs in the title WQEBB. */
199 	uint16_t max_inline; /* Multiple of RTE_CACHE_LINE_SIZE to inline. */
200 	uint16_t inline_max_packet_sz; /* Max packet size for inlining. */
201 	uint16_t mr_cache_idx; /* Index of last hit entry. */
202 	uint32_t qp_num_8s; /* QP number shifted by 8. */
203 	uint64_t offloads; /* Offloads for Tx Queue. */
204 	volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
205 	volatile void *wqes; /* Work queue (use volatile to write into). */
206 	volatile uint32_t *qp_db; /* Work queue doorbell. */
207 	volatile uint32_t *cq_db; /* Completion queue doorbell. */
208 	volatile void *bf_reg; /* Blueflame register remapped. */
209 	struct mlx5_mr *mp2mr[MLX5_PMD_TX_MP_CACHE]; /* MR translation table. */
210 	struct rte_mbuf *(*elts)[]; /* TX elements. */
211 	struct mlx5_txq_stats stats; /* TX queue counters. */
212 } __rte_cache_aligned;
213 
214 /* Verbs Rx queue elements. */
215 struct mlx5_txq_ibv {
216 	LIST_ENTRY(mlx5_txq_ibv) next; /* Pointer to the next element. */
217 	rte_atomic32_t refcnt; /* Reference counter. */
218 	struct ibv_cq *cq; /* Completion Queue. */
219 	struct ibv_qp *qp; /* Queue Pair. */
220 };
221 
222 /* TX queue control descriptor. */
223 struct mlx5_txq_ctrl {
224 	LIST_ENTRY(mlx5_txq_ctrl) next; /* Pointer to the next element. */
225 	rte_atomic32_t refcnt; /* Reference counter. */
226 	struct priv *priv; /* Back pointer to private data. */
227 	unsigned int socket; /* CPU socket ID for allocations. */
228 	unsigned int max_inline_data; /* Max inline data. */
229 	unsigned int max_tso_header; /* Max TSO header size. */
230 	struct mlx5_txq_ibv *ibv; /* Verbs queue object. */
231 	struct mlx5_txq_data txq; /* Data path structure. */
232 	off_t uar_mmap_offset; /* UAR mmap offset for non-primary process. */
233 	volatile void *bf_reg_orig; /* Blueflame register from verbs. */
234 };
235 
236 /* mlx5_rxq.c */
237 
238 extern uint8_t rss_hash_default_key[];
239 extern const size_t rss_hash_default_key_len;
240 
241 void mlx5_rxq_cleanup(struct mlx5_rxq_ctrl *);
242 int mlx5_rx_queue_setup(struct rte_eth_dev *, uint16_t, uint16_t, unsigned int,
243 			const struct rte_eth_rxconf *, struct rte_mempool *);
244 void mlx5_rx_queue_release(void *);
245 int priv_rx_intr_vec_enable(struct priv *priv);
246 void priv_rx_intr_vec_disable(struct priv *priv);
247 int mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
248 int mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
249 struct mlx5_rxq_ibv *mlx5_priv_rxq_ibv_new(struct priv *, uint16_t);
250 struct mlx5_rxq_ibv *mlx5_priv_rxq_ibv_get(struct priv *, uint16_t);
251 int mlx5_priv_rxq_ibv_release(struct priv *, struct mlx5_rxq_ibv *);
252 int mlx5_priv_rxq_ibv_releasable(struct priv *, struct mlx5_rxq_ibv *);
253 int mlx5_priv_rxq_ibv_verify(struct priv *);
254 struct mlx5_rxq_ctrl *mlx5_priv_rxq_new(struct priv *, uint16_t,
255 					uint16_t, unsigned int,
256 					const struct rte_eth_rxconf *,
257 					struct rte_mempool *);
258 struct mlx5_rxq_ctrl *mlx5_priv_rxq_get(struct priv *, uint16_t);
259 int mlx5_priv_rxq_release(struct priv *, uint16_t);
260 int mlx5_priv_rxq_releasable(struct priv *, uint16_t);
261 int mlx5_priv_rxq_verify(struct priv *);
262 int rxq_alloc_elts(struct mlx5_rxq_ctrl *);
263 struct mlx5_ind_table_ibv *mlx5_priv_ind_table_ibv_new(struct priv *,
264 						       uint16_t [],
265 						       uint16_t);
266 struct mlx5_ind_table_ibv *mlx5_priv_ind_table_ibv_get(struct priv *,
267 						       uint16_t [],
268 						       uint16_t);
269 int mlx5_priv_ind_table_ibv_release(struct priv *, struct mlx5_ind_table_ibv *);
270 int mlx5_priv_ind_table_ibv_verify(struct priv *);
271 struct mlx5_hrxq *mlx5_priv_hrxq_new(struct priv *, uint8_t *, uint8_t,
272 				     uint64_t, uint16_t [], uint16_t);
273 struct mlx5_hrxq *mlx5_priv_hrxq_get(struct priv *, uint8_t *, uint8_t,
274 				     uint64_t, uint16_t [], uint16_t);
275 int mlx5_priv_hrxq_release(struct priv *, struct mlx5_hrxq *);
276 int mlx5_priv_hrxq_ibv_verify(struct priv *);
277 uint64_t mlx5_priv_get_rx_port_offloads(struct priv *);
278 uint64_t mlx5_priv_get_rx_queue_offloads(struct priv *);
279 
280 /* mlx5_txq.c */
281 
282 int mlx5_tx_queue_setup(struct rte_eth_dev *, uint16_t, uint16_t, unsigned int,
283 			const struct rte_eth_txconf *);
284 void mlx5_tx_queue_release(void *);
285 int priv_tx_uar_remap(struct priv *priv, int fd);
286 struct mlx5_txq_ibv *mlx5_priv_txq_ibv_new(struct priv *, uint16_t);
287 struct mlx5_txq_ibv *mlx5_priv_txq_ibv_get(struct priv *, uint16_t);
288 int mlx5_priv_txq_ibv_release(struct priv *, struct mlx5_txq_ibv *);
289 int mlx5_priv_txq_ibv_releasable(struct priv *, struct mlx5_txq_ibv *);
290 int mlx5_priv_txq_ibv_verify(struct priv *);
291 struct mlx5_txq_ctrl *mlx5_priv_txq_new(struct priv *, uint16_t,
292 					uint16_t, unsigned int,
293 					const struct rte_eth_txconf *);
294 struct mlx5_txq_ctrl *mlx5_priv_txq_get(struct priv *, uint16_t);
295 int mlx5_priv_txq_release(struct priv *, uint16_t);
296 int mlx5_priv_txq_releasable(struct priv *, uint16_t);
297 int mlx5_priv_txq_verify(struct priv *);
298 void txq_alloc_elts(struct mlx5_txq_ctrl *);
299 uint64_t mlx5_priv_get_tx_port_offloads(struct priv *);
300 
301 /* mlx5_rxtx.c */
302 
303 extern uint32_t mlx5_ptype_table[];
304 
305 void mlx5_set_ptype_table(void);
306 uint16_t mlx5_tx_burst(void *, struct rte_mbuf **, uint16_t);
307 uint16_t mlx5_tx_burst_mpw(void *, struct rte_mbuf **, uint16_t);
308 uint16_t mlx5_tx_burst_mpw_inline(void *, struct rte_mbuf **, uint16_t);
309 uint16_t mlx5_tx_burst_empw(void *, struct rte_mbuf **, uint16_t);
310 uint16_t mlx5_rx_burst(void *, struct rte_mbuf **, uint16_t);
311 uint16_t removed_tx_burst(void *, struct rte_mbuf **, uint16_t);
312 uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
313 int mlx5_rx_descriptor_status(void *, uint16_t);
314 int mlx5_tx_descriptor_status(void *, uint16_t);
315 
316 /* Vectorized version of mlx5_rxtx.c */
317 int priv_check_raw_vec_tx_support(struct priv *, struct rte_eth_dev *);
318 int priv_check_vec_tx_support(struct priv *, struct rte_eth_dev *);
319 int rxq_check_vec_support(struct mlx5_rxq_data *);
320 int priv_check_vec_rx_support(struct priv *);
321 uint16_t mlx5_tx_burst_raw_vec(void *, struct rte_mbuf **, uint16_t);
322 uint16_t mlx5_tx_burst_vec(void *, struct rte_mbuf **, uint16_t);
323 uint16_t mlx5_rx_burst_vec(void *, struct rte_mbuf **, uint16_t);
324 
325 /* mlx5_mr.c */
326 
327 void mlx5_mp2mr_iter(struct rte_mempool *, void *);
328 struct mlx5_mr *priv_txq_mp2mr_reg(struct priv *priv, struct mlx5_txq_data *,
329 				   struct rte_mempool *, unsigned int);
330 struct mlx5_mr *mlx5_txq_mp2mr_reg(struct mlx5_txq_data *, struct rte_mempool *,
331 				   unsigned int);
332 
333 #ifndef NDEBUG
334 /**
335  * Verify or set magic value in CQE.
336  *
337  * @param cqe
338  *   Pointer to CQE.
339  *
340  * @return
341  *   0 the first time.
342  */
343 static inline int
344 check_cqe_seen(volatile struct mlx5_cqe *cqe)
345 {
346 	static const uint8_t magic[] = "seen";
347 	volatile uint8_t (*buf)[sizeof(cqe->rsvd0)] = &cqe->rsvd0;
348 	int ret = 1;
349 	unsigned int i;
350 
351 	for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
352 		if (!ret || (*buf)[i] != magic[i]) {
353 			ret = 0;
354 			(*buf)[i] = magic[i];
355 		}
356 	return ret;
357 }
358 #endif /* NDEBUG */
359 
360 /**
361  * Check whether CQE is valid.
362  *
363  * @param cqe
364  *   Pointer to CQE.
365  * @param cqes_n
366  *   Size of completion queue.
367  * @param ci
368  *   Consumer index.
369  *
370  * @return
371  *   0 on success, 1 on failure.
372  */
373 static __rte_always_inline int
374 check_cqe(volatile struct mlx5_cqe *cqe,
375 	  unsigned int cqes_n, const uint16_t ci)
376 {
377 	uint16_t idx = ci & cqes_n;
378 	uint8_t op_own = cqe->op_own;
379 	uint8_t op_owner = MLX5_CQE_OWNER(op_own);
380 	uint8_t op_code = MLX5_CQE_OPCODE(op_own);
381 
382 	if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
383 		return 1; /* No CQE. */
384 #ifndef NDEBUG
385 	if ((op_code == MLX5_CQE_RESP_ERR) ||
386 	    (op_code == MLX5_CQE_REQ_ERR)) {
387 		volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
388 		uint8_t syndrome = err_cqe->syndrome;
389 
390 		if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
391 		    (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
392 			return 0;
393 		if (!check_cqe_seen(cqe)) {
394 			ERROR("unexpected CQE error %u (0x%02x)"
395 			      " syndrome 0x%02x",
396 			      op_code, op_code, syndrome);
397 			rte_hexdump(stderr, "MLX5 Error CQE:",
398 				    (const void *)((uintptr_t)err_cqe),
399 				    sizeof(*err_cqe));
400 		}
401 		return 1;
402 	} else if ((op_code != MLX5_CQE_RESP_SEND) &&
403 		   (op_code != MLX5_CQE_REQ)) {
404 		if (!check_cqe_seen(cqe)) {
405 			ERROR("unexpected CQE opcode %u (0x%02x)",
406 			      op_code, op_code);
407 			rte_hexdump(stderr, "MLX5 CQE:",
408 				    (const void *)((uintptr_t)cqe),
409 				    sizeof(*cqe));
410 		}
411 		return 1;
412 	}
413 #endif /* NDEBUG */
414 	return 0;
415 }
416 
417 /**
418  * Return the address of the WQE.
419  *
420  * @param txq
421  *   Pointer to TX queue structure.
422  * @param  wqe_ci
423  *   WQE consumer index.
424  *
425  * @return
426  *   WQE address.
427  */
428 static inline uintptr_t *
429 tx_mlx5_wqe(struct mlx5_txq_data *txq, uint16_t ci)
430 {
431 	ci &= ((1 << txq->wqe_n) - 1);
432 	return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
433 }
434 
435 /**
436  * Manage TX completions.
437  *
438  * When sending a burst, mlx5_tx_burst() posts several WRs.
439  *
440  * @param txq
441  *   Pointer to TX queue structure.
442  */
443 static __rte_always_inline void
444 mlx5_tx_complete(struct mlx5_txq_data *txq)
445 {
446 	const uint16_t elts_n = 1 << txq->elts_n;
447 	const uint16_t elts_m = elts_n - 1;
448 	const unsigned int cqe_n = 1 << txq->cqe_n;
449 	const unsigned int cqe_cnt = cqe_n - 1;
450 	uint16_t elts_free = txq->elts_tail;
451 	uint16_t elts_tail;
452 	uint16_t cq_ci = txq->cq_ci;
453 	volatile struct mlx5_cqe *cqe = NULL;
454 	volatile struct mlx5_wqe_ctrl *ctrl;
455 	struct rte_mbuf *m, *free[elts_n];
456 	struct rte_mempool *pool = NULL;
457 	unsigned int blk_n = 0;
458 
459 	cqe = &(*txq->cqes)[cq_ci & cqe_cnt];
460 	if (unlikely(check_cqe(cqe, cqe_n, cq_ci)))
461 		return;
462 #ifndef NDEBUG
463 	if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
464 	    (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
465 		if (!check_cqe_seen(cqe)) {
466 			ERROR("unexpected error CQE, TX stopped");
467 			rte_hexdump(stderr, "MLX5 TXQ:",
468 				    (const void *)((uintptr_t)txq->wqes),
469 				    ((1 << txq->wqe_n) *
470 				     MLX5_WQE_SIZE));
471 		}
472 		return;
473 	}
474 #endif /* NDEBUG */
475 	++cq_ci;
476 	txq->wqe_pi = rte_be_to_cpu_16(cqe->wqe_counter);
477 	ctrl = (volatile struct mlx5_wqe_ctrl *)
478 		tx_mlx5_wqe(txq, txq->wqe_pi);
479 	elts_tail = ctrl->ctrl3;
480 	assert((elts_tail & elts_m) < (1 << txq->wqe_n));
481 	/* Free buffers. */
482 	while (elts_free != elts_tail) {
483 		m = rte_pktmbuf_prefree_seg((*txq->elts)[elts_free++ & elts_m]);
484 		if (likely(m != NULL)) {
485 			if (likely(m->pool == pool)) {
486 				free[blk_n++] = m;
487 			} else {
488 				if (likely(pool != NULL))
489 					rte_mempool_put_bulk(pool,
490 							     (void *)free,
491 							     blk_n);
492 				free[0] = m;
493 				pool = m->pool;
494 				blk_n = 1;
495 			}
496 		}
497 	}
498 	if (blk_n)
499 		rte_mempool_put_bulk(pool, (void *)free, blk_n);
500 #ifndef NDEBUG
501 	elts_free = txq->elts_tail;
502 	/* Poisoning. */
503 	while (elts_free != elts_tail) {
504 		memset(&(*txq->elts)[elts_free & elts_m],
505 		       0x66,
506 		       sizeof((*txq->elts)[elts_free & elts_m]));
507 		++elts_free;
508 	}
509 #endif
510 	txq->cq_ci = cq_ci;
511 	txq->elts_tail = elts_tail;
512 	/* Update the consumer index. */
513 	rte_compiler_barrier();
514 	*txq->cq_db = rte_cpu_to_be_32(cq_ci);
515 }
516 
517 /**
518  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
519  * the cloned mbuf is allocated is returned instead.
520  *
521  * @param buf
522  *   Pointer to mbuf.
523  *
524  * @return
525  *   Memory pool where data is located for given mbuf.
526  */
527 static struct rte_mempool *
528 mlx5_tx_mb2mp(struct rte_mbuf *buf)
529 {
530 	if (unlikely(RTE_MBUF_INDIRECT(buf)))
531 		return rte_mbuf_from_indirect(buf)->pool;
532 	return buf->pool;
533 }
534 
535 /**
536  * Get Memory Region (MR) <-> rte_mbuf association from txq->mp2mr[].
537  * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
538  * remove an entry first.
539  *
540  * @param txq
541  *   Pointer to TX queue structure.
542  * @param[in] mp
543  *   Memory Pool for which a Memory Region lkey must be returned.
544  *
545  * @return
546  *   mr->lkey on success, (uint32_t)-1 on failure.
547  */
548 static __rte_always_inline uint32_t
549 mlx5_tx_mb2mr(struct mlx5_txq_data *txq, struct rte_mbuf *mb)
550 {
551 	uint16_t i = txq->mr_cache_idx;
552 	uintptr_t addr = rte_pktmbuf_mtod(mb, uintptr_t);
553 	struct mlx5_mr *mr;
554 
555 	assert(i < RTE_DIM(txq->mp2mr));
556 	if (likely(txq->mp2mr[i]->start <= addr && txq->mp2mr[i]->end > addr))
557 		return txq->mp2mr[i]->lkey;
558 	for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
559 		if (unlikely(txq->mp2mr[i] == NULL ||
560 		    txq->mp2mr[i]->mr == NULL)) {
561 			/* Unknown MP, add a new MR for it. */
562 			break;
563 		}
564 		if (txq->mp2mr[i]->start <= addr &&
565 		    txq->mp2mr[i]->end > addr) {
566 			assert(txq->mp2mr[i]->lkey != (uint32_t)-1);
567 			txq->mr_cache_idx = i;
568 			return txq->mp2mr[i]->lkey;
569 		}
570 	}
571 	mr = mlx5_txq_mp2mr_reg(txq, mlx5_tx_mb2mp(mb), i);
572 	/*
573 	 * Request the reference to use in this queue, the original one is
574 	 * kept by the control plane.
575 	 */
576 	if (mr) {
577 		rte_atomic32_inc(&mr->refcnt);
578 		txq->mr_cache_idx = i >= RTE_DIM(txq->mp2mr) ? i - 1 : i;
579 		return mr->lkey;
580 	} else {
581 		struct rte_mempool *mp = mlx5_tx_mb2mp(mb);
582 
583 		WARN("Failed to register mempool 0x%p(%s)",
584 		      (void *)mp, mp->name);
585 	}
586 	return (uint32_t)-1;
587 }
588 
589 /**
590  * Ring TX queue doorbell and flush the update if requested.
591  *
592  * @param txq
593  *   Pointer to TX queue structure.
594  * @param wqe
595  *   Pointer to the last WQE posted in the NIC.
596  * @param cond
597  *   Request for write memory barrier after BlueFlame update.
598  */
599 static __rte_always_inline void
600 mlx5_tx_dbrec_cond_wmb(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe,
601 		       int cond)
602 {
603 	uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
604 	volatile uint64_t *src = ((volatile uint64_t *)wqe);
605 
606 	rte_cio_wmb();
607 	*txq->qp_db = rte_cpu_to_be_32(txq->wqe_ci);
608 	/* Ensure ordering between DB record and BF copy. */
609 	rte_wmb();
610 	*dst = *src;
611 	if (cond)
612 		rte_wmb();
613 }
614 
615 /**
616  * Ring TX queue doorbell and flush the update by write memory barrier.
617  *
618  * @param txq
619  *   Pointer to TX queue structure.
620  * @param wqe
621  *   Pointer to the last WQE posted in the NIC.
622  */
623 static __rte_always_inline void
624 mlx5_tx_dbrec(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe)
625 {
626 	mlx5_tx_dbrec_cond_wmb(txq, wqe, 1);
627 }
628 
629 /**
630  * Convert the Checksum offloads to Verbs.
631  *
632  * @param txq_data
633  *   Pointer to the Tx queue.
634  * @param buf
635  *   Pointer to the mbuf.
636  *
637  * @return
638  *   the converted cs_flags.
639  */
640 static __rte_always_inline uint8_t
641 txq_ol_cksum_to_cs(struct mlx5_txq_data *txq_data, struct rte_mbuf *buf)
642 {
643 	uint8_t cs_flags = 0;
644 
645 	/* Should we enable HW CKSUM offload */
646 	if (buf->ol_flags &
647 	    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM |
648 	     PKT_TX_OUTER_IP_CKSUM)) {
649 		if (txq_data->tunnel_en &&
650 		    (buf->ol_flags &
651 		     (PKT_TX_TUNNEL_GRE | PKT_TX_TUNNEL_VXLAN))) {
652 			cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
653 				   MLX5_ETH_WQE_L4_INNER_CSUM;
654 			if (buf->ol_flags & PKT_TX_OUTER_IP_CKSUM)
655 				cs_flags |= MLX5_ETH_WQE_L3_CSUM;
656 		} else {
657 			cs_flags = MLX5_ETH_WQE_L3_CSUM |
658 				   MLX5_ETH_WQE_L4_CSUM;
659 		}
660 	}
661 	return cs_flags;
662 }
663 
664 /**
665  * Count the number of contiguous single segment packets.
666  *
667  * @param pkts
668  *   Pointer to array of packets.
669  * @param pkts_n
670  *   Number of packets.
671  *
672  * @return
673  *   Number of contiguous single segment packets.
674  */
675 static __rte_always_inline unsigned int
676 txq_count_contig_single_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
677 {
678 	unsigned int pos;
679 
680 	if (!pkts_n)
681 		return 0;
682 	/* Count the number of contiguous single segment packets. */
683 	for (pos = 0; pos < pkts_n; ++pos)
684 		if (NB_SEGS(pkts[pos]) > 1)
685 			break;
686 	return pos;
687 }
688 
689 /**
690  * Count the number of contiguous multi-segment packets.
691  *
692  * @param pkts
693  *   Pointer to array of packets.
694  * @param pkts_n
695  *   Number of packets.
696  *
697  * @return
698  *   Number of contiguous multi-segment packets.
699  */
700 static __rte_always_inline unsigned int
701 txq_count_contig_multi_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
702 {
703 	unsigned int pos;
704 
705 	if (!pkts_n)
706 		return 0;
707 	/* Count the number of contiguous multi-segment packets. */
708 	for (pos = 0; pos < pkts_n; ++pos)
709 		if (NB_SEGS(pkts[pos]) == 1)
710 			break;
711 	return pos;
712 }
713 
714 #endif /* RTE_PMD_MLX5_RXTX_H_ */
715