xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.h (revision 5dba3b9c4c131b88a78bcecfef39db23ebc47873)
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 pending_err:1; /* CQE error needs to be handled. */
118 	unsigned int :14; /* Remaining bits. */
119 	volatile uint32_t *rq_db;
120 	volatile uint32_t *cq_db;
121 	uint16_t port_id;
122 	uint16_t rq_ci;
123 	uint16_t rq_pi;
124 	uint16_t cq_ci;
125 	volatile struct mlx5_wqe_data_seg(*wqes)[];
126 	volatile struct mlx5_cqe(*cqes)[];
127 	struct rxq_zip zip; /* Compressed context. */
128 	struct rte_mbuf *(*elts)[];
129 	struct rte_mempool *mp;
130 	struct mlx5_rxq_stats stats;
131 	uint64_t mbuf_initializer; /* Default rearm_data for vectorized Rx. */
132 	struct rte_mbuf fake_mbuf; /* elts padding for vectorized Rx. */
133 	void *cq_uar; /* CQ user access region. */
134 	uint32_t cqn; /* CQ number. */
135 	uint8_t cq_arm_sn; /* CQ arm seq number. */
136 } __rte_cache_aligned;
137 
138 /* Verbs Rx queue elements. */
139 struct mlx5_rxq_ibv {
140 	LIST_ENTRY(mlx5_rxq_ibv) next; /* Pointer to the next element. */
141 	rte_atomic32_t refcnt; /* Reference counter. */
142 	struct mlx5_rxq_ctrl *rxq_ctrl; /* Back pointer to parent. */
143 	struct ibv_cq *cq; /* Completion Queue. */
144 	struct ibv_wq *wq; /* Work Queue. */
145 	struct ibv_comp_channel *channel;
146 	struct mlx5_mr *mr; /* Memory Region (for mp). */
147 };
148 
149 /* RX queue control descriptor. */
150 struct mlx5_rxq_ctrl {
151 	LIST_ENTRY(mlx5_rxq_ctrl) next; /* Pointer to the next element. */
152 	rte_atomic32_t refcnt; /* Reference counter. */
153 	struct priv *priv; /* Back pointer to private data. */
154 	struct mlx5_rxq_ibv *ibv; /* Verbs elements. */
155 	struct mlx5_rxq_data rxq; /* Data path structure. */
156 	unsigned int socket; /* CPU socket ID for allocations. */
157 	unsigned int irq:1; /* Whether IRQ is enabled. */
158 };
159 
160 /* Indirection table. */
161 struct mlx5_ind_table_ibv {
162 	LIST_ENTRY(mlx5_ind_table_ibv) next; /* Pointer to the next element. */
163 	rte_atomic32_t refcnt; /* Reference counter. */
164 	struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
165 	uint16_t queues_n; /**< Number of queues in the list. */
166 	uint16_t queues[]; /**< Queue list. */
167 };
168 
169 /* Hash Rx queue. */
170 struct mlx5_hrxq {
171 	LIST_ENTRY(mlx5_hrxq) next; /* Pointer to the next element. */
172 	rte_atomic32_t refcnt; /* Reference counter. */
173 	struct mlx5_ind_table_ibv *ind_table; /* Indirection table. */
174 	struct ibv_qp *qp; /* Verbs queue pair. */
175 	uint64_t hash_fields; /* Verbs Hash fields. */
176 	uint8_t rss_key_len; /* Hash key length in bytes. */
177 	uint8_t rss_key[]; /* Hash key. */
178 };
179 
180 /* TX queue descriptor. */
181 __extension__
182 struct mlx5_txq_data {
183 	uint16_t elts_head; /* Current counter in (*elts)[]. */
184 	uint16_t elts_tail; /* Counter of first element awaiting completion. */
185 	uint16_t elts_comp; /* Counter since last completion request. */
186 	uint16_t mpw_comp; /* WQ index since last completion request. */
187 	uint16_t cq_ci; /* Consumer index for completion queue. */
188 #ifndef NDEBUG
189 	uint16_t cq_pi; /* Producer index for completion queue. */
190 #endif
191 	uint16_t wqe_ci; /* Consumer index for work queue. */
192 	uint16_t wqe_pi; /* Producer index for work queue. */
193 	uint16_t elts_n:4; /* (*elts)[] length (in log2). */
194 	uint16_t cqe_n:4; /* Number of CQ elements (in log2). */
195 	uint16_t wqe_n:4; /* Number of of WQ elements (in log2). */
196 	uint16_t tso_en:1; /* When set hardware TSO is enabled. */
197 	uint16_t tunnel_en:1;
198 	/* When set TX offload for tunneled packets are supported. */
199 	uint16_t mpw_hdr_dseg:1; /* Enable DSEGs in the title WQEBB. */
200 	uint16_t max_inline; /* Multiple of RTE_CACHE_LINE_SIZE to inline. */
201 	uint16_t inline_max_packet_sz; /* Max packet size for inlining. */
202 	uint16_t mr_cache_idx; /* Index of last hit entry. */
203 	uint32_t qp_num_8s; /* QP number shifted by 8. */
204 	uint32_t flags; /* Flags for Tx Queue. */
205 	volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
206 	volatile void *wqes; /* Work queue (use volatile to write into). */
207 	volatile uint32_t *qp_db; /* Work queue doorbell. */
208 	volatile uint32_t *cq_db; /* Completion queue doorbell. */
209 	volatile void *bf_reg; /* Blueflame register. */
210 	struct mlx5_mr *mp2mr[MLX5_PMD_TX_MP_CACHE]; /* MR translation table. */
211 	struct rte_mbuf *(*elts)[]; /* TX elements. */
212 	struct mlx5_txq_stats stats; /* TX queue counters. */
213 } __rte_cache_aligned;
214 
215 /* Verbs Rx queue elements. */
216 struct mlx5_txq_ibv {
217 	LIST_ENTRY(mlx5_txq_ibv) next; /* Pointer to the next element. */
218 	rte_atomic32_t refcnt; /* Reference counter. */
219 	struct ibv_cq *cq; /* Completion Queue. */
220 	struct ibv_qp *qp; /* Queue Pair. */
221 };
222 
223 /* TX queue control descriptor. */
224 struct mlx5_txq_ctrl {
225 	LIST_ENTRY(mlx5_txq_ctrl) next; /* Pointer to the next element. */
226 	rte_atomic32_t refcnt; /* Reference counter. */
227 	struct priv *priv; /* Back pointer to private data. */
228 	unsigned int socket; /* CPU socket ID for allocations. */
229 	unsigned int max_inline_data; /* Max inline data. */
230 	unsigned int max_tso_header; /* Max TSO header size. */
231 	struct mlx5_txq_ibv *ibv; /* Verbs queue object. */
232 	struct mlx5_txq_data txq; /* Data path structure. */
233 	off_t uar_mmap_offset; /* UAR mmap offset for non-primary process. */
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 					struct rte_mempool *);
257 struct mlx5_rxq_ctrl *mlx5_priv_rxq_get(struct priv *, uint16_t);
258 int mlx5_priv_rxq_release(struct priv *, uint16_t);
259 int mlx5_priv_rxq_releasable(struct priv *, uint16_t);
260 int mlx5_priv_rxq_verify(struct priv *);
261 int rxq_alloc_elts(struct mlx5_rxq_ctrl *);
262 struct mlx5_ind_table_ibv *mlx5_priv_ind_table_ibv_new(struct priv *,
263 						       uint16_t [],
264 						       uint16_t);
265 struct mlx5_ind_table_ibv *mlx5_priv_ind_table_ibv_get(struct priv *,
266 						       uint16_t [],
267 						       uint16_t);
268 int mlx5_priv_ind_table_ibv_release(struct priv *, struct mlx5_ind_table_ibv *);
269 int mlx5_priv_ind_table_ibv_verify(struct priv *);
270 struct mlx5_hrxq *mlx5_priv_hrxq_new(struct priv *, uint8_t *, uint8_t,
271 				     uint64_t, uint16_t [], uint16_t);
272 struct mlx5_hrxq *mlx5_priv_hrxq_get(struct priv *, uint8_t *, uint8_t,
273 				     uint64_t, uint16_t [], uint16_t);
274 int mlx5_priv_hrxq_release(struct priv *, struct mlx5_hrxq *);
275 int mlx5_priv_hrxq_ibv_verify(struct priv *);
276 
277 /* mlx5_txq.c */
278 
279 int mlx5_tx_queue_setup(struct rte_eth_dev *, uint16_t, uint16_t, unsigned int,
280 			const struct rte_eth_txconf *);
281 void mlx5_tx_queue_release(void *);
282 int priv_tx_uar_remap(struct priv *priv, int fd);
283 struct mlx5_txq_ibv *mlx5_priv_txq_ibv_new(struct priv *, uint16_t);
284 struct mlx5_txq_ibv *mlx5_priv_txq_ibv_get(struct priv *, uint16_t);
285 int mlx5_priv_txq_ibv_release(struct priv *, struct mlx5_txq_ibv *);
286 int mlx5_priv_txq_ibv_releasable(struct priv *, struct mlx5_txq_ibv *);
287 int mlx5_priv_txq_ibv_verify(struct priv *);
288 struct mlx5_txq_ctrl *mlx5_priv_txq_new(struct priv *, uint16_t,
289 					uint16_t, unsigned int,
290 					const struct rte_eth_txconf *);
291 struct mlx5_txq_ctrl *mlx5_priv_txq_get(struct priv *, uint16_t);
292 int mlx5_priv_txq_release(struct priv *, uint16_t);
293 int mlx5_priv_txq_releasable(struct priv *, uint16_t);
294 int mlx5_priv_txq_verify(struct priv *);
295 void txq_alloc_elts(struct mlx5_txq_ctrl *);
296 
297 /* mlx5_rxtx.c */
298 
299 extern uint32_t mlx5_ptype_table[];
300 
301 void mlx5_set_ptype_table(void);
302 uint16_t mlx5_tx_burst(void *, struct rte_mbuf **, uint16_t);
303 uint16_t mlx5_tx_burst_mpw(void *, struct rte_mbuf **, uint16_t);
304 uint16_t mlx5_tx_burst_mpw_inline(void *, struct rte_mbuf **, uint16_t);
305 uint16_t mlx5_tx_burst_empw(void *, struct rte_mbuf **, uint16_t);
306 uint16_t mlx5_rx_burst(void *, struct rte_mbuf **, uint16_t);
307 uint16_t removed_tx_burst(void *, struct rte_mbuf **, uint16_t);
308 uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
309 int mlx5_rx_descriptor_status(void *, uint16_t);
310 int mlx5_tx_descriptor_status(void *, uint16_t);
311 
312 /* Vectorized version of mlx5_rxtx.c */
313 int priv_check_raw_vec_tx_support(struct priv *);
314 int priv_check_vec_tx_support(struct priv *);
315 int rxq_check_vec_support(struct mlx5_rxq_data *);
316 int priv_check_vec_rx_support(struct priv *);
317 uint16_t mlx5_tx_burst_raw_vec(void *, struct rte_mbuf **, uint16_t);
318 uint16_t mlx5_tx_burst_vec(void *, struct rte_mbuf **, uint16_t);
319 uint16_t mlx5_rx_burst_vec(void *, struct rte_mbuf **, uint16_t);
320 
321 /* mlx5_mr.c */
322 
323 void mlx5_mp2mr_iter(struct rte_mempool *, void *);
324 struct mlx5_mr *priv_txq_mp2mr_reg(struct priv *priv, struct mlx5_txq_data *,
325 				   struct rte_mempool *, unsigned int);
326 struct mlx5_mr *mlx5_txq_mp2mr_reg(struct mlx5_txq_data *, struct rte_mempool *,
327 				   unsigned int);
328 
329 #ifndef NDEBUG
330 /**
331  * Verify or set magic value in CQE.
332  *
333  * @param cqe
334  *   Pointer to CQE.
335  *
336  * @return
337  *   0 the first time.
338  */
339 static inline int
340 check_cqe_seen(volatile struct mlx5_cqe *cqe)
341 {
342 	static const uint8_t magic[] = "seen";
343 	volatile uint8_t (*buf)[sizeof(cqe->rsvd0)] = &cqe->rsvd0;
344 	int ret = 1;
345 	unsigned int i;
346 
347 	for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
348 		if (!ret || (*buf)[i] != magic[i]) {
349 			ret = 0;
350 			(*buf)[i] = magic[i];
351 		}
352 	return ret;
353 }
354 #endif /* NDEBUG */
355 
356 /**
357  * Check whether CQE is valid.
358  *
359  * @param cqe
360  *   Pointer to CQE.
361  * @param cqes_n
362  *   Size of completion queue.
363  * @param ci
364  *   Consumer index.
365  *
366  * @return
367  *   0 on success, 1 on failure.
368  */
369 static __rte_always_inline int
370 check_cqe(volatile struct mlx5_cqe *cqe,
371 	  unsigned int cqes_n, const uint16_t ci)
372 {
373 	uint16_t idx = ci & cqes_n;
374 	uint8_t op_own = cqe->op_own;
375 	uint8_t op_owner = MLX5_CQE_OWNER(op_own);
376 	uint8_t op_code = MLX5_CQE_OPCODE(op_own);
377 
378 	if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
379 		return 1; /* No CQE. */
380 #ifndef NDEBUG
381 	if ((op_code == MLX5_CQE_RESP_ERR) ||
382 	    (op_code == MLX5_CQE_REQ_ERR)) {
383 		volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
384 		uint8_t syndrome = err_cqe->syndrome;
385 
386 		if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
387 		    (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
388 			return 0;
389 		if (!check_cqe_seen(cqe)) {
390 			ERROR("unexpected CQE error %u (0x%02x)"
391 			      " syndrome 0x%02x",
392 			      op_code, op_code, syndrome);
393 			rte_hexdump(stderr, "MLX5 Error CQE:",
394 				    (const void *)((uintptr_t)err_cqe),
395 				    sizeof(*err_cqe));
396 		}
397 		return 1;
398 	} else if ((op_code != MLX5_CQE_RESP_SEND) &&
399 		   (op_code != MLX5_CQE_REQ)) {
400 		if (!check_cqe_seen(cqe)) {
401 			ERROR("unexpected CQE opcode %u (0x%02x)",
402 			      op_code, op_code);
403 			rte_hexdump(stderr, "MLX5 CQE:",
404 				    (const void *)((uintptr_t)cqe),
405 				    sizeof(*cqe));
406 		}
407 		return 1;
408 	}
409 #endif /* NDEBUG */
410 	return 0;
411 }
412 
413 /**
414  * Return the address of the WQE.
415  *
416  * @param txq
417  *   Pointer to TX queue structure.
418  * @param  wqe_ci
419  *   WQE consumer index.
420  *
421  * @return
422  *   WQE address.
423  */
424 static inline uintptr_t *
425 tx_mlx5_wqe(struct mlx5_txq_data *txq, uint16_t ci)
426 {
427 	ci &= ((1 << txq->wqe_n) - 1);
428 	return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
429 }
430 
431 /**
432  * Manage TX completions.
433  *
434  * When sending a burst, mlx5_tx_burst() posts several WRs.
435  *
436  * @param txq
437  *   Pointer to TX queue structure.
438  */
439 static __rte_always_inline void
440 mlx5_tx_complete(struct mlx5_txq_data *txq)
441 {
442 	const uint16_t elts_n = 1 << txq->elts_n;
443 	const uint16_t elts_m = elts_n - 1;
444 	const unsigned int cqe_n = 1 << txq->cqe_n;
445 	const unsigned int cqe_cnt = cqe_n - 1;
446 	uint16_t elts_free = txq->elts_tail;
447 	uint16_t elts_tail;
448 	uint16_t cq_ci = txq->cq_ci;
449 	volatile struct mlx5_cqe *cqe = NULL;
450 	volatile struct mlx5_wqe_ctrl *ctrl;
451 	struct rte_mbuf *m, *free[elts_n];
452 	struct rte_mempool *pool = NULL;
453 	unsigned int blk_n = 0;
454 
455 	cqe = &(*txq->cqes)[cq_ci & cqe_cnt];
456 	if (unlikely(check_cqe(cqe, cqe_n, cq_ci)))
457 		return;
458 #ifndef NDEBUG
459 	if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
460 	    (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
461 		if (!check_cqe_seen(cqe)) {
462 			ERROR("unexpected error CQE, TX stopped");
463 			rte_hexdump(stderr, "MLX5 TXQ:",
464 				    (const void *)((uintptr_t)txq->wqes),
465 				    ((1 << txq->wqe_n) *
466 				     MLX5_WQE_SIZE));
467 		}
468 		return;
469 	}
470 #endif /* NDEBUG */
471 	++cq_ci;
472 	txq->wqe_pi = rte_be_to_cpu_16(cqe->wqe_counter);
473 	ctrl = (volatile struct mlx5_wqe_ctrl *)
474 		tx_mlx5_wqe(txq, txq->wqe_pi);
475 	elts_tail = ctrl->ctrl3;
476 	assert((elts_tail & elts_m) < (1 << txq->wqe_n));
477 	/* Free buffers. */
478 	while (elts_free != elts_tail) {
479 		m = rte_pktmbuf_prefree_seg((*txq->elts)[elts_free++ & elts_m]);
480 		if (likely(m != NULL)) {
481 			if (likely(m->pool == pool)) {
482 				free[blk_n++] = m;
483 			} else {
484 				if (likely(pool != NULL))
485 					rte_mempool_put_bulk(pool,
486 							     (void *)free,
487 							     blk_n);
488 				free[0] = m;
489 				pool = m->pool;
490 				blk_n = 1;
491 			}
492 		}
493 	}
494 	if (blk_n)
495 		rte_mempool_put_bulk(pool, (void *)free, blk_n);
496 #ifndef NDEBUG
497 	elts_free = txq->elts_tail;
498 	/* Poisoning. */
499 	while (elts_free != elts_tail) {
500 		memset(&(*txq->elts)[elts_free & elts_m],
501 		       0x66,
502 		       sizeof((*txq->elts)[elts_free & elts_m]));
503 		++elts_free;
504 	}
505 #endif
506 	txq->cq_ci = cq_ci;
507 	txq->elts_tail = elts_tail;
508 	/* Update the consumer index. */
509 	rte_compiler_barrier();
510 	*txq->cq_db = rte_cpu_to_be_32(cq_ci);
511 }
512 
513 /**
514  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
515  * the cloned mbuf is allocated is returned instead.
516  *
517  * @param buf
518  *   Pointer to mbuf.
519  *
520  * @return
521  *   Memory pool where data is located for given mbuf.
522  */
523 static struct rte_mempool *
524 mlx5_tx_mb2mp(struct rte_mbuf *buf)
525 {
526 	if (unlikely(RTE_MBUF_INDIRECT(buf)))
527 		return rte_mbuf_from_indirect(buf)->pool;
528 	return buf->pool;
529 }
530 
531 /**
532  * Get Memory Region (MR) <-> rte_mbuf association from txq->mp2mr[].
533  * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
534  * remove an entry first.
535  *
536  * @param txq
537  *   Pointer to TX queue structure.
538  * @param[in] mp
539  *   Memory Pool for which a Memory Region lkey must be returned.
540  *
541  * @return
542  *   mr->lkey on success, (uint32_t)-1 on failure.
543  */
544 static __rte_always_inline uint32_t
545 mlx5_tx_mb2mr(struct mlx5_txq_data *txq, struct rte_mbuf *mb)
546 {
547 	uint16_t i = txq->mr_cache_idx;
548 	uintptr_t addr = rte_pktmbuf_mtod_offset(mb, uintptr_t, DATA_LEN(mb));
549 	struct mlx5_mr *mr;
550 
551 	assert(i < RTE_DIM(txq->mp2mr));
552 	if (likely(txq->mp2mr[i]->start <= addr && txq->mp2mr[i]->end >= addr))
553 		return txq->mp2mr[i]->lkey;
554 	for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
555 		if (unlikely(txq->mp2mr[i]->mr == NULL)) {
556 			/* Unknown MP, add a new MR for it. */
557 			break;
558 		}
559 		if (txq->mp2mr[i]->start <= addr &&
560 		    txq->mp2mr[i]->end >= addr) {
561 			assert(txq->mp2mr[i]->lkey != (uint32_t)-1);
562 			assert(rte_cpu_to_be_32(txq->mp2mr[i]->mr->lkey) ==
563 			       txq->mp2mr[i]->lkey);
564 			txq->mr_cache_idx = i;
565 			return txq->mp2mr[i]->lkey;
566 		}
567 	}
568 	txq->mr_cache_idx = 0;
569 	mr = mlx5_txq_mp2mr_reg(txq, mlx5_tx_mb2mp(mb), i);
570 	/*
571 	 * Request the reference to use in this queue, the original one is
572 	 * kept by the control plane.
573 	 */
574 	if (mr) {
575 		rte_atomic32_inc(&mr->refcnt);
576 		return mr->lkey;
577 	}
578 	return (uint32_t)-1;
579 }
580 
581 /**
582  * Ring TX queue doorbell and flush the update if requested.
583  *
584  * @param txq
585  *   Pointer to TX queue structure.
586  * @param wqe
587  *   Pointer to the last WQE posted in the NIC.
588  * @param cond
589  *   Request for write memory barrier after BlueFlame update.
590  */
591 static __rte_always_inline void
592 mlx5_tx_dbrec_cond_wmb(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe,
593 		       int cond)
594 {
595 	uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
596 	volatile uint64_t *src = ((volatile uint64_t *)wqe);
597 
598 	rte_io_wmb();
599 	*txq->qp_db = rte_cpu_to_be_32(txq->wqe_ci);
600 	/* Ensure ordering between DB record and BF copy. */
601 	rte_wmb();
602 	*dst = *src;
603 	if (cond)
604 		rte_wmb();
605 }
606 
607 /**
608  * Ring TX queue doorbell and flush the update by write memory barrier.
609  *
610  * @param txq
611  *   Pointer to TX queue structure.
612  * @param wqe
613  *   Pointer to the last WQE posted in the NIC.
614  */
615 static __rte_always_inline void
616 mlx5_tx_dbrec(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe)
617 {
618 	mlx5_tx_dbrec_cond_wmb(txq, wqe, 1);
619 }
620 
621 /**
622  * Convert the Checksum offloads to Verbs.
623  *
624  * @param txq_data
625  *   Pointer to the Tx queue.
626  * @param buf
627  *   Pointer to the mbuf.
628  *
629  * @return
630  *   the converted cs_flags.
631  */
632 static __rte_always_inline uint8_t
633 txq_ol_cksum_to_cs(struct mlx5_txq_data *txq_data, struct rte_mbuf *buf)
634 {
635 	uint8_t cs_flags = 0;
636 
637 	/* Should we enable HW CKSUM offload */
638 	if (buf->ol_flags &
639 	    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM |
640 	     PKT_TX_OUTER_IP_CKSUM)) {
641 		if (txq_data->tunnel_en &&
642 		    (buf->ol_flags &
643 		     (PKT_TX_TUNNEL_GRE | PKT_TX_TUNNEL_VXLAN))) {
644 			cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
645 				   MLX5_ETH_WQE_L4_INNER_CSUM;
646 			if (buf->ol_flags & PKT_TX_OUTER_IP_CKSUM)
647 				cs_flags |= MLX5_ETH_WQE_L3_CSUM;
648 		} else {
649 			cs_flags = MLX5_ETH_WQE_L3_CSUM |
650 				   MLX5_ETH_WQE_L4_CSUM;
651 		}
652 	}
653 	return cs_flags;
654 }
655 
656 /**
657  * Count the number of contiguous single segment packets.
658  *
659  * @param pkts
660  *   Pointer to array of packets.
661  * @param pkts_n
662  *   Number of packets.
663  *
664  * @return
665  *   Number of contiguous single segment packets.
666  */
667 static __rte_always_inline unsigned int
668 txq_count_contig_single_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
669 {
670 	unsigned int pos;
671 
672 	if (!pkts_n)
673 		return 0;
674 	/* Count the number of contiguous single segment packets. */
675 	for (pos = 0; pos < pkts_n; ++pos)
676 		if (NB_SEGS(pkts[pos]) > 1)
677 			break;
678 	return pos;
679 }
680 
681 /**
682  * Count the number of contiguous multi-segment packets.
683  *
684  * @param pkts
685  *   Pointer to array of packets.
686  * @param pkts_n
687  *   Number of packets.
688  *
689  * @return
690  *   Number of contiguous multi-segment packets.
691  */
692 static __rte_always_inline unsigned int
693 txq_count_contig_multi_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
694 {
695 	unsigned int pos;
696 
697 	if (!pkts_n)
698 		return 0;
699 	/* Count the number of contiguous multi-segment packets. */
700 	for (pos = 0; pos < pkts_n; ++pos)
701 		if (NB_SEGS(pkts[pos]) == 1)
702 			break;
703 	return pos;
704 }
705 
706 #endif /* RTE_PMD_MLX5_RXTX_H_ */
707