xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.h (revision f10aadfd2f2fc0e43a00083dfff7ba9a64871b2d)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #ifndef RTE_PMD_MLX5_RXTX_H_
7 #define RTE_PMD_MLX5_RXTX_H_
8 
9 #include <stddef.h>
10 #include <stdint.h>
11 #include <sys/queue.h>
12 
13 /* Verbs header. */
14 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
15 #ifdef PEDANTIC
16 #pragma GCC diagnostic ignored "-Wpedantic"
17 #endif
18 #include <infiniband/verbs.h>
19 #include <infiniband/mlx5dv.h>
20 #ifdef PEDANTIC
21 #pragma GCC diagnostic error "-Wpedantic"
22 #endif
23 
24 #include <rte_mbuf.h>
25 #include <rte_mempool.h>
26 #include <rte_common.h>
27 #include <rte_hexdump.h>
28 #include <rte_atomic.h>
29 #include <rte_spinlock.h>
30 #include <rte_io.h>
31 #include <rte_bus_pci.h>
32 
33 #include "mlx5_utils.h"
34 #include "mlx5.h"
35 #include "mlx5_mr.h"
36 #include "mlx5_autoconf.h"
37 #include "mlx5_defs.h"
38 #include "mlx5_prm.h"
39 
40 /* Support tunnel matching. */
41 #define MLX5_FLOW_TUNNEL 5
42 
43 struct mlx5_rxq_stats {
44 #ifdef MLX5_PMD_SOFT_COUNTERS
45 	uint64_t ipackets; /**< Total of successfully received packets. */
46 	uint64_t ibytes; /**< Total of successfully received bytes. */
47 #endif
48 	uint64_t idropped; /**< Total of packets dropped when RX ring full. */
49 	uint64_t rx_nombuf; /**< Total of RX mbuf allocation failures. */
50 };
51 
52 struct mlx5_txq_stats {
53 #ifdef MLX5_PMD_SOFT_COUNTERS
54 	uint64_t opackets; /**< Total of successfully sent packets. */
55 	uint64_t obytes; /**< Total of successfully sent bytes. */
56 #endif
57 	uint64_t oerrors; /**< Total number of failed transmitted packets. */
58 };
59 
60 struct mlx5_priv;
61 
62 /* Compressed CQE context. */
63 struct rxq_zip {
64 	uint16_t ai; /* Array index. */
65 	uint16_t ca; /* Current array index. */
66 	uint16_t na; /* Next array index. */
67 	uint16_t cq_ci; /* The next CQE. */
68 	uint32_t cqe_cnt; /* Number of CQEs. */
69 };
70 
71 /* Multi-Packet RQ buffer header. */
72 struct mlx5_mprq_buf {
73 	struct rte_mempool *mp;
74 	rte_atomic16_t refcnt; /* Atomically accessed refcnt. */
75 	uint8_t pad[RTE_PKTMBUF_HEADROOM]; /* Headroom for the first packet. */
76 } __rte_cache_aligned;
77 
78 /* Get pointer to the first stride. */
79 #define mlx5_mprq_buf_addr(ptr) ((ptr) + 1)
80 
81 /* RX queue descriptor. */
82 struct mlx5_rxq_data {
83 	unsigned int csum:1; /* Enable checksum offloading. */
84 	unsigned int hw_timestamp:1; /* Enable HW timestamp. */
85 	unsigned int vlan_strip:1; /* Enable VLAN stripping. */
86 	unsigned int crc_present:1; /* CRC must be subtracted. */
87 	unsigned int sges_n:2; /* Log 2 of SGEs (max buffers per packet). */
88 	unsigned int cqe_n:4; /* Log 2 of CQ elements. */
89 	unsigned int elts_n:4; /* Log 2 of Mbufs. */
90 	unsigned int rss_hash:1; /* RSS hash result is enabled. */
91 	unsigned int mark:1; /* Marked flow available on the queue. */
92 	unsigned int strd_num_n:5; /* Log 2 of the number of stride. */
93 	unsigned int strd_sz_n:4; /* Log 2 of stride size. */
94 	unsigned int strd_shift_en:1; /* Enable 2bytes shift on a stride. */
95 	unsigned int :6; /* Remaining bits. */
96 	volatile uint32_t *rq_db;
97 	volatile uint32_t *cq_db;
98 	uint16_t port_id;
99 	uint32_t rq_ci;
100 	uint16_t consumed_strd; /* Number of consumed strides in WQE. */
101 	uint32_t rq_pi;
102 	uint32_t cq_ci;
103 	uint16_t rq_repl_thresh; /* Threshold for buffer replenishment. */
104 	struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
105 	uint16_t mprq_max_memcpy_len; /* Maximum size of packet to memcpy. */
106 	volatile void *wqes;
107 	volatile struct mlx5_cqe(*cqes)[];
108 	struct rxq_zip zip; /* Compressed context. */
109 	RTE_STD_C11
110 	union  {
111 		struct rte_mbuf *(*elts)[];
112 		struct mlx5_mprq_buf *(*mprq_bufs)[];
113 	};
114 	struct rte_mempool *mp;
115 	struct rte_mempool *mprq_mp; /* Mempool for Multi-Packet RQ. */
116 	struct mlx5_mprq_buf *mprq_repl; /* Stashed mbuf for replenish. */
117 	uint16_t idx; /* Queue index. */
118 	struct mlx5_rxq_stats stats;
119 	uint64_t mbuf_initializer; /* Default rearm_data for vectorized Rx. */
120 	struct rte_mbuf fake_mbuf; /* elts padding for vectorized Rx. */
121 	void *cq_uar; /* CQ user access region. */
122 	uint32_t cqn; /* CQ number. */
123 	uint8_t cq_arm_sn; /* CQ arm seq number. */
124 #ifndef RTE_ARCH_64
125 	rte_spinlock_t *uar_lock_cq;
126 	/* CQ (UAR) access lock required for 32bit implementations */
127 #endif
128 	uint32_t tunnel; /* Tunnel information. */
129 } __rte_cache_aligned;
130 
131 /* Verbs Rx queue elements. */
132 struct mlx5_rxq_ibv {
133 	LIST_ENTRY(mlx5_rxq_ibv) next; /* Pointer to the next element. */
134 	rte_atomic32_t refcnt; /* Reference counter. */
135 	struct mlx5_rxq_ctrl *rxq_ctrl; /* Back pointer to parent. */
136 	struct ibv_cq *cq; /* Completion Queue. */
137 	struct ibv_wq *wq; /* Work Queue. */
138 	struct ibv_comp_channel *channel;
139 };
140 
141 /* RX queue control descriptor. */
142 struct mlx5_rxq_ctrl {
143 	struct mlx5_rxq_data rxq; /* Data path structure. */
144 	LIST_ENTRY(mlx5_rxq_ctrl) next; /* Pointer to the next element. */
145 	rte_atomic32_t refcnt; /* Reference counter. */
146 	struct mlx5_rxq_ibv *ibv; /* Verbs elements. */
147 	struct mlx5_priv *priv; /* Back pointer to private data. */
148 	unsigned int socket; /* CPU socket ID for allocations. */
149 	unsigned int irq:1; /* Whether IRQ is enabled. */
150 	uint32_t flow_mark_n; /* Number of Mark/Flag flows using this Queue. */
151 	uint32_t flow_tunnels_n[MLX5_FLOW_TUNNEL]; /* Tunnels counters. */
152 };
153 
154 /* Indirection table. */
155 struct mlx5_ind_table_ibv {
156 	LIST_ENTRY(mlx5_ind_table_ibv) next; /* Pointer to the next element. */
157 	rte_atomic32_t refcnt; /* Reference counter. */
158 	struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
159 	uint32_t queues_n; /**< Number of queues in the list. */
160 	uint16_t queues[]; /**< Queue list. */
161 };
162 
163 /* Hash Rx queue. */
164 struct mlx5_hrxq {
165 	LIST_ENTRY(mlx5_hrxq) next; /* Pointer to the next element. */
166 	rte_atomic32_t refcnt; /* Reference counter. */
167 	struct mlx5_ind_table_ibv *ind_table; /* Indirection table. */
168 	struct ibv_qp *qp; /* Verbs queue pair. */
169 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
170 	void *action; /* DV QP action pointer. */
171 #endif
172 	uint64_t hash_fields; /* Verbs Hash fields. */
173 	uint32_t rss_key_len; /* Hash key length in bytes. */
174 	uint8_t rss_key[]; /* Hash key. */
175 };
176 
177 /* TX queue descriptor. */
178 __extension__
179 struct mlx5_txq_data {
180 	uint16_t elts_head; /* Current counter in (*elts)[]. */
181 	uint16_t elts_tail; /* Counter of first element awaiting completion. */
182 	uint16_t elts_comp; /* Counter since last completion request. */
183 	uint16_t mpw_comp; /* WQ index since last completion request. */
184 	uint16_t cq_ci; /* Consumer index for completion queue. */
185 #ifndef NDEBUG
186 	uint16_t cq_pi; /* Producer index for completion queue. */
187 #endif
188 	uint16_t wqe_ci; /* Consumer index for work queue. */
189 	uint16_t wqe_pi; /* Producer index for work queue. */
190 	uint16_t elts_n:4; /* (*elts)[] length (in log2). */
191 	uint16_t cqe_n:4; /* Number of CQ elements (in log2). */
192 	uint16_t wqe_n:4; /* Number of of WQ elements (in log2). */
193 	uint16_t tso_en:1; /* When set hardware TSO is enabled. */
194 	uint16_t tunnel_en:1;
195 	/* When set TX offload for tunneled packets are supported. */
196 	uint16_t swp_en:1; /* Whether SW parser is enabled. */
197 	uint16_t mpw_hdr_dseg:1; /* Enable DSEGs in the title WQEBB. */
198 	uint16_t max_inline; /* Multiple of RTE_CACHE_LINE_SIZE to inline. */
199 	uint16_t inline_max_packet_sz; /* Max packet size for inlining. */
200 	uint32_t qp_num_8s; /* QP number shifted by 8. */
201 	uint64_t offloads; /* Offloads for Tx Queue. */
202 	struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
203 	volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
204 	volatile void *wqes; /* Work queue (use volatile to write into). */
205 	volatile uint32_t *qp_db; /* Work queue doorbell. */
206 	volatile uint32_t *cq_db; /* Completion queue doorbell. */
207 	struct rte_mbuf *(*elts)[]; /* TX elements. */
208 	uint16_t port_id; /* Port ID of device. */
209 	uint16_t idx; /* Queue index. */
210 	struct mlx5_txq_stats stats; /* TX queue counters. */
211 #ifndef RTE_ARCH_64
212 	rte_spinlock_t *uar_lock;
213 	/* UAR access lock required for 32bit implementations */
214 #endif
215 } __rte_cache_aligned;
216 
217 /* Verbs Rx queue elements. */
218 struct mlx5_txq_ibv {
219 	LIST_ENTRY(mlx5_txq_ibv) next; /* Pointer to the next element. */
220 	rte_atomic32_t refcnt; /* Reference counter. */
221 	struct mlx5_txq_ctrl *txq_ctrl; /* Pointer to the control queue. */
222 	struct ibv_cq *cq; /* Completion Queue. */
223 	struct ibv_qp *qp; /* Queue Pair. */
224 };
225 
226 /* TX queue control descriptor. */
227 struct mlx5_txq_ctrl {
228 	struct mlx5_txq_data txq; /* Data path structure. */
229 	LIST_ENTRY(mlx5_txq_ctrl) next; /* Pointer to the next element. */
230 	rte_atomic32_t refcnt; /* Reference counter. */
231 	unsigned int socket; /* CPU socket ID for allocations. */
232 	unsigned int max_inline_data; /* Max inline data. */
233 	unsigned int max_tso_header; /* Max TSO header size. */
234 	struct mlx5_txq_ibv *ibv; /* Verbs queue object. */
235 	struct mlx5_priv *priv; /* Back pointer to private data. */
236 	off_t uar_mmap_offset; /* UAR mmap offset for non-primary process. */
237 	void *bf_reg; /* BlueFlame register from Verbs. */
238 };
239 
240 #define MLX5_TX_BFREG(txq) \
241 		(MLX5_PROC_PRIV((txq)->port_id)->uar_table[(txq)->idx])
242 
243 /* mlx5_rxq.c */
244 
245 extern uint8_t rss_hash_default_key[];
246 
247 int mlx5_check_mprq_support(struct rte_eth_dev *dev);
248 int mlx5_rxq_mprq_enabled(struct mlx5_rxq_data *rxq);
249 int mlx5_mprq_enabled(struct rte_eth_dev *dev);
250 int mlx5_mprq_free_mp(struct rte_eth_dev *dev);
251 int mlx5_mprq_alloc_mp(struct rte_eth_dev *dev);
252 int mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
253 			unsigned int socket, const struct rte_eth_rxconf *conf,
254 			struct rte_mempool *mp);
255 void mlx5_rx_queue_release(void *dpdk_rxq);
256 int mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev);
257 void mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev);
258 int mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
259 int mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
260 struct mlx5_rxq_ibv *mlx5_rxq_ibv_new(struct rte_eth_dev *dev, uint16_t idx);
261 int mlx5_rxq_ibv_verify(struct rte_eth_dev *dev);
262 struct mlx5_rxq_ctrl *mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx,
263 				   uint16_t desc, unsigned int socket,
264 				   const struct rte_eth_rxconf *conf,
265 				   struct rte_mempool *mp);
266 struct mlx5_rxq_ctrl *mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx);
267 int mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx);
268 int mlx5_rxq_verify(struct rte_eth_dev *dev);
269 int rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl);
270 int mlx5_ind_table_ibv_verify(struct rte_eth_dev *dev);
271 struct mlx5_hrxq *mlx5_hrxq_new(struct rte_eth_dev *dev,
272 				const uint8_t *rss_key, uint32_t rss_key_len,
273 				uint64_t hash_fields,
274 				const uint16_t *queues, uint32_t queues_n,
275 				int tunnel __rte_unused);
276 struct mlx5_hrxq *mlx5_hrxq_get(struct rte_eth_dev *dev,
277 				const uint8_t *rss_key, uint32_t rss_key_len,
278 				uint64_t hash_fields,
279 				const uint16_t *queues, uint32_t queues_n);
280 int mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hxrq);
281 int mlx5_hrxq_ibv_verify(struct rte_eth_dev *dev);
282 struct mlx5_hrxq *mlx5_hrxq_drop_new(struct rte_eth_dev *dev);
283 void mlx5_hrxq_drop_release(struct rte_eth_dev *dev);
284 uint64_t mlx5_get_rx_port_offloads(void);
285 uint64_t mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev);
286 
287 /* mlx5_txq.c */
288 
289 int mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
290 			unsigned int socket, const struct rte_eth_txconf *conf);
291 void mlx5_tx_queue_release(void *dpdk_txq);
292 int mlx5_tx_uar_init_secondary(struct rte_eth_dev *dev, int fd);
293 struct mlx5_txq_ibv *mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx);
294 struct mlx5_txq_ibv *mlx5_txq_ibv_get(struct rte_eth_dev *dev, uint16_t idx);
295 int mlx5_txq_ibv_release(struct mlx5_txq_ibv *txq_ibv);
296 int mlx5_txq_ibv_verify(struct rte_eth_dev *dev);
297 struct mlx5_txq_ctrl *mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx,
298 				   uint16_t desc, unsigned int socket,
299 				   const struct rte_eth_txconf *conf);
300 struct mlx5_txq_ctrl *mlx5_txq_get(struct rte_eth_dev *dev, uint16_t idx);
301 int mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx);
302 int mlx5_txq_releasable(struct rte_eth_dev *dev, uint16_t idx);
303 int mlx5_txq_verify(struct rte_eth_dev *dev);
304 void txq_alloc_elts(struct mlx5_txq_ctrl *txq_ctrl);
305 uint64_t mlx5_get_tx_port_offloads(struct rte_eth_dev *dev);
306 
307 /* mlx5_rxtx.c */
308 
309 extern uint32_t mlx5_ptype_table[];
310 extern uint8_t mlx5_cksum_table[];
311 extern uint8_t mlx5_swp_types_table[];
312 
313 void mlx5_set_ptype_table(void);
314 void mlx5_set_cksum_table(void);
315 void mlx5_set_swp_types_table(void);
316 uint16_t mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts,
317 		       uint16_t pkts_n);
318 uint16_t mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts,
319 			   uint16_t pkts_n);
320 uint16_t mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
321 				  uint16_t pkts_n);
322 uint16_t mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts,
323 			    uint16_t pkts_n);
324 uint16_t mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
325 void mlx5_mprq_buf_free_cb(void *addr, void *opaque);
326 void mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf);
327 uint16_t mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts,
328 			    uint16_t pkts_n);
329 uint16_t removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts,
330 			  uint16_t pkts_n);
331 uint16_t removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts,
332 			  uint16_t pkts_n);
333 int mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset);
334 int mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset);
335 uint32_t mlx5_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id);
336 
337 /* Vectorized version of mlx5_rxtx.c */
338 int mlx5_check_raw_vec_tx_support(struct rte_eth_dev *dev);
339 int mlx5_check_vec_tx_support(struct rte_eth_dev *dev);
340 int mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq_data);
341 int mlx5_check_vec_rx_support(struct rte_eth_dev *dev);
342 uint16_t mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts,
343 			       uint16_t pkts_n);
344 uint16_t mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts,
345 			   uint16_t pkts_n);
346 uint16_t mlx5_rx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts,
347 			   uint16_t pkts_n);
348 
349 /* mlx5_mr.c */
350 
351 void mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl);
352 uint32_t mlx5_rx_addr2mr_bh(struct mlx5_rxq_data *rxq, uintptr_t addr);
353 uint32_t mlx5_tx_mb2mr_bh(struct mlx5_txq_data *txq, struct rte_mbuf *mb);
354 uint32_t mlx5_tx_update_ext_mp(struct mlx5_txq_data *txq, uintptr_t addr,
355 			       struct rte_mempool *mp);
356 int mlx5_dma_map(struct rte_pci_device *pdev, void *addr, uint64_t iova,
357 		 size_t len);
358 int mlx5_dma_unmap(struct rte_pci_device *pdev, void *addr, uint64_t iova,
359 		   size_t len);
360 
361 /**
362  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
363  * 64bit architectures.
364  *
365  * @param val
366  *   value to write in CPU endian format.
367  * @param addr
368  *   Address to write to.
369  * @param lock
370  *   Address of the lock to use for that UAR access.
371  */
372 static __rte_always_inline void
373 __mlx5_uar_write64_relaxed(uint64_t val, void *addr,
374 			   rte_spinlock_t *lock __rte_unused)
375 {
376 #ifdef RTE_ARCH_64
377 	*(uint64_t *)addr = val;
378 #else /* !RTE_ARCH_64 */
379 	rte_spinlock_lock(lock);
380 	*(uint32_t *)addr = val;
381 	rte_io_wmb();
382 	*((uint32_t *)addr + 1) = val >> 32;
383 	rte_spinlock_unlock(lock);
384 #endif
385 }
386 
387 /**
388  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
389  * 64bit architectures while guaranteeing the order of execution with the
390  * code being executed.
391  *
392  * @param val
393  *   value to write in CPU endian format.
394  * @param addr
395  *   Address to write to.
396  * @param lock
397  *   Address of the lock to use for that UAR access.
398  */
399 static __rte_always_inline void
400 __mlx5_uar_write64(uint64_t val, void *addr, rte_spinlock_t *lock)
401 {
402 	rte_io_wmb();
403 	__mlx5_uar_write64_relaxed(val, addr, lock);
404 }
405 
406 /* Assist macros, used instead of directly calling the functions they wrap. */
407 #ifdef RTE_ARCH_64
408 #define mlx5_uar_write64_relaxed(val, dst, lock) \
409 		__mlx5_uar_write64_relaxed(val, dst, NULL)
410 #define mlx5_uar_write64(val, dst, lock) __mlx5_uar_write64(val, dst, NULL)
411 #else
412 #define mlx5_uar_write64_relaxed(val, dst, lock) \
413 		__mlx5_uar_write64_relaxed(val, dst, lock)
414 #define mlx5_uar_write64(val, dst, lock) __mlx5_uar_write64(val, dst, lock)
415 #endif
416 
417 #ifndef NDEBUG
418 /**
419  * Verify or set magic value in CQE.
420  *
421  * @param cqe
422  *   Pointer to CQE.
423  *
424  * @return
425  *   0 the first time.
426  */
427 static inline int
428 check_cqe_seen(volatile struct mlx5_cqe *cqe)
429 {
430 	static const uint8_t magic[] = "seen";
431 	volatile uint8_t (*buf)[sizeof(cqe->rsvd1)] = &cqe->rsvd1;
432 	int ret = 1;
433 	unsigned int i;
434 
435 	for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
436 		if (!ret || (*buf)[i] != magic[i]) {
437 			ret = 0;
438 			(*buf)[i] = magic[i];
439 		}
440 	return ret;
441 }
442 #endif /* NDEBUG */
443 
444 /**
445  * Check whether CQE is valid.
446  *
447  * @param cqe
448  *   Pointer to CQE.
449  * @param cqes_n
450  *   Size of completion queue.
451  * @param ci
452  *   Consumer index.
453  *
454  * @return
455  *   0 on success, 1 on failure.
456  */
457 static __rte_always_inline int
458 check_cqe(volatile struct mlx5_cqe *cqe,
459 	  unsigned int cqes_n, const uint16_t ci)
460 {
461 	uint16_t idx = ci & cqes_n;
462 	uint8_t op_own = cqe->op_own;
463 	uint8_t op_owner = MLX5_CQE_OWNER(op_own);
464 	uint8_t op_code = MLX5_CQE_OPCODE(op_own);
465 
466 	if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
467 		return 1; /* No CQE. */
468 #ifndef NDEBUG
469 	if ((op_code == MLX5_CQE_RESP_ERR) ||
470 	    (op_code == MLX5_CQE_REQ_ERR)) {
471 		volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
472 		uint8_t syndrome = err_cqe->syndrome;
473 
474 		if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
475 		    (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
476 			return 0;
477 		if (!check_cqe_seen(cqe)) {
478 			DRV_LOG(ERR,
479 				"unexpected CQE error %u (0x%02x) syndrome"
480 				" 0x%02x",
481 				op_code, op_code, syndrome);
482 			rte_hexdump(stderr, "MLX5 Error CQE:",
483 				    (const void *)((uintptr_t)err_cqe),
484 				    sizeof(*cqe));
485 		}
486 		return 1;
487 	} else if ((op_code != MLX5_CQE_RESP_SEND) &&
488 		   (op_code != MLX5_CQE_REQ)) {
489 		if (!check_cqe_seen(cqe)) {
490 			DRV_LOG(ERR, "unexpected CQE opcode %u (0x%02x)",
491 				op_code, op_code);
492 			rte_hexdump(stderr, "MLX5 CQE:",
493 				    (const void *)((uintptr_t)cqe),
494 				    sizeof(*cqe));
495 		}
496 		return 1;
497 	}
498 #endif /* NDEBUG */
499 	return 0;
500 }
501 
502 /**
503  * Return the address of the WQE.
504  *
505  * @param txq
506  *   Pointer to TX queue structure.
507  * @param  wqe_ci
508  *   WQE consumer index.
509  *
510  * @return
511  *   WQE address.
512  */
513 static inline uintptr_t *
514 tx_mlx5_wqe(struct mlx5_txq_data *txq, uint16_t ci)
515 {
516 	ci &= ((1 << txq->wqe_n) - 1);
517 	return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
518 }
519 
520 /**
521  * Manage TX completions.
522  *
523  * When sending a burst, mlx5_tx_burst() posts several WRs.
524  *
525  * @param txq
526  *   Pointer to TX queue structure.
527  */
528 static __rte_always_inline void
529 mlx5_tx_complete(struct mlx5_txq_data *txq)
530 {
531 	const uint16_t elts_n = 1 << txq->elts_n;
532 	const uint16_t elts_m = elts_n - 1;
533 	const unsigned int cqe_n = 1 << txq->cqe_n;
534 	const unsigned int cqe_cnt = cqe_n - 1;
535 	uint16_t elts_free = txq->elts_tail;
536 	uint16_t elts_tail;
537 	uint16_t cq_ci = txq->cq_ci;
538 	volatile struct mlx5_cqe *cqe = NULL;
539 	volatile struct mlx5_wqe_ctrl *ctrl;
540 	struct rte_mbuf *m, *free[elts_n];
541 	struct rte_mempool *pool = NULL;
542 	unsigned int blk_n = 0;
543 
544 	cqe = &(*txq->cqes)[cq_ci & cqe_cnt];
545 	if (unlikely(check_cqe(cqe, cqe_n, cq_ci)))
546 		return;
547 #ifndef NDEBUG
548 	if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
549 	    (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
550 		if (!check_cqe_seen(cqe)) {
551 			DRV_LOG(ERR, "unexpected error CQE, Tx stopped");
552 			rte_hexdump(stderr, "MLX5 TXQ:",
553 				    (const void *)((uintptr_t)txq->wqes),
554 				    ((1 << txq->wqe_n) *
555 				     MLX5_WQE_SIZE));
556 		}
557 		return;
558 	}
559 #endif /* NDEBUG */
560 	++cq_ci;
561 	rte_cio_rmb();
562 	txq->wqe_pi = rte_be_to_cpu_16(cqe->wqe_counter);
563 	ctrl = (volatile struct mlx5_wqe_ctrl *)
564 		tx_mlx5_wqe(txq, txq->wqe_pi);
565 	elts_tail = ctrl->ctrl3;
566 	assert((elts_tail & elts_m) < (1 << txq->wqe_n));
567 	/* Free buffers. */
568 	while (elts_free != elts_tail) {
569 		m = rte_pktmbuf_prefree_seg((*txq->elts)[elts_free++ & elts_m]);
570 		if (likely(m != NULL)) {
571 			if (likely(m->pool == pool)) {
572 				free[blk_n++] = m;
573 			} else {
574 				if (likely(pool != NULL))
575 					rte_mempool_put_bulk(pool,
576 							     (void *)free,
577 							     blk_n);
578 				free[0] = m;
579 				pool = m->pool;
580 				blk_n = 1;
581 			}
582 		}
583 	}
584 	if (blk_n)
585 		rte_mempool_put_bulk(pool, (void *)free, blk_n);
586 #ifndef NDEBUG
587 	elts_free = txq->elts_tail;
588 	/* Poisoning. */
589 	while (elts_free != elts_tail) {
590 		memset(&(*txq->elts)[elts_free & elts_m],
591 		       0x66,
592 		       sizeof((*txq->elts)[elts_free & elts_m]));
593 		++elts_free;
594 	}
595 #endif
596 	txq->cq_ci = cq_ci;
597 	txq->elts_tail = elts_tail;
598 	/* Update the consumer index. */
599 	rte_compiler_barrier();
600 	*txq->cq_db = rte_cpu_to_be_32(cq_ci);
601 }
602 
603 /**
604  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which the
605  * cloned mbuf is allocated is returned instead.
606  *
607  * @param buf
608  *   Pointer to mbuf.
609  *
610  * @return
611  *   Memory pool where data is located for given mbuf.
612  */
613 static inline struct rte_mempool *
614 mlx5_mb2mp(struct rte_mbuf *buf)
615 {
616 	if (unlikely(RTE_MBUF_CLONED(buf)))
617 		return rte_mbuf_from_indirect(buf)->pool;
618 	return buf->pool;
619 }
620 
621 /**
622  * Query LKey from a packet buffer for Rx. No need to flush local caches for Rx
623  * as mempool is pre-configured and static.
624  *
625  * @param rxq
626  *   Pointer to Rx queue structure.
627  * @param addr
628  *   Address to search.
629  *
630  * @return
631  *   Searched LKey on success, UINT32_MAX on no match.
632  */
633 static __rte_always_inline uint32_t
634 mlx5_rx_addr2mr(struct mlx5_rxq_data *rxq, uintptr_t addr)
635 {
636 	struct mlx5_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
637 	uint32_t lkey;
638 
639 	/* Linear search on MR cache array. */
640 	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
641 				    MLX5_MR_CACHE_N, addr);
642 	if (likely(lkey != UINT32_MAX))
643 		return lkey;
644 	/* Take slower bottom-half (Binary Search) on miss. */
645 	return mlx5_rx_addr2mr_bh(rxq, addr);
646 }
647 
648 #define mlx5_rx_mb2mr(rxq, mb) mlx5_rx_addr2mr(rxq, (uintptr_t)((mb)->buf_addr))
649 
650 /**
651  * Query LKey from a packet buffer for Tx. If not found, add the mempool.
652  *
653  * @param txq
654  *   Pointer to Tx queue structure.
655  * @param addr
656  *   Address to search.
657  *
658  * @return
659  *   Searched LKey on success, UINT32_MAX on no match.
660  */
661 static __rte_always_inline uint32_t
662 mlx5_tx_mb2mr(struct mlx5_txq_data *txq, struct rte_mbuf *mb)
663 {
664 	struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
665 	uintptr_t addr = (uintptr_t)mb->buf_addr;
666 	uint32_t lkey;
667 
668 	/* Check generation bit to see if there's any change on existing MRs. */
669 	if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen))
670 		mlx5_mr_flush_local_cache(mr_ctrl);
671 	/* Linear search on MR cache array. */
672 	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
673 				    MLX5_MR_CACHE_N, addr);
674 	if (likely(lkey != UINT32_MAX))
675 		return lkey;
676 	/* Take slower bottom-half on miss. */
677 	return mlx5_tx_mb2mr_bh(txq, mb);
678 }
679 
680 /**
681  * Ring TX queue doorbell and flush the update if requested.
682  *
683  * @param txq
684  *   Pointer to TX queue structure.
685  * @param wqe
686  *   Pointer to the last WQE posted in the NIC.
687  * @param cond
688  *   Request for write memory barrier after BlueFlame update.
689  */
690 static __rte_always_inline void
691 mlx5_tx_dbrec_cond_wmb(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe,
692 		       int cond)
693 {
694 	uint64_t *dst = MLX5_TX_BFREG(txq);
695 	volatile uint64_t *src = ((volatile uint64_t *)wqe);
696 
697 	rte_cio_wmb();
698 	*txq->qp_db = rte_cpu_to_be_32(txq->wqe_ci);
699 	/* Ensure ordering between DB record and BF copy. */
700 	rte_wmb();
701 	mlx5_uar_write64_relaxed(*src, dst, txq->uar_lock);
702 	if (cond)
703 		rte_wmb();
704 }
705 
706 /**
707  * Ring TX queue doorbell and flush the update by write memory barrier.
708  *
709  * @param txq
710  *   Pointer to TX queue structure.
711  * @param wqe
712  *   Pointer to the last WQE posted in the NIC.
713  */
714 static __rte_always_inline void
715 mlx5_tx_dbrec(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe)
716 {
717 	mlx5_tx_dbrec_cond_wmb(txq, wqe, 1);
718 }
719 
720 /**
721  * Convert mbuf to Verb SWP.
722  *
723  * @param txq_data
724  *   Pointer to the Tx queue.
725  * @param buf
726  *   Pointer to the mbuf.
727  * @param offsets
728  *   Pointer to the SWP header offsets.
729  * @param swp_types
730  *   Pointer to the SWP header types.
731  */
732 static __rte_always_inline void
733 txq_mbuf_to_swp(struct mlx5_txq_data *txq, struct rte_mbuf *buf,
734 		uint8_t *offsets, uint8_t *swp_types)
735 {
736 	const uint64_t vlan = buf->ol_flags & PKT_TX_VLAN_PKT;
737 	const uint64_t tunnel = buf->ol_flags & PKT_TX_TUNNEL_MASK;
738 	const uint64_t tso = buf->ol_flags & PKT_TX_TCP_SEG;
739 	const uint64_t csum_flags = buf->ol_flags & PKT_TX_L4_MASK;
740 	const uint64_t inner_ip =
741 		buf->ol_flags & (PKT_TX_IPV4 | PKT_TX_IPV6);
742 	const uint64_t ol_flags_mask = PKT_TX_L4_MASK | PKT_TX_IPV6 |
743 				       PKT_TX_OUTER_IPV6;
744 	uint16_t idx;
745 	uint16_t off;
746 
747 	if (likely(!txq->swp_en || (tunnel != PKT_TX_TUNNEL_UDP &&
748 				    tunnel != PKT_TX_TUNNEL_IP)))
749 		return;
750 	/*
751 	 * The index should have:
752 	 * bit[0:1] = PKT_TX_L4_MASK
753 	 * bit[4] = PKT_TX_IPV6
754 	 * bit[8] = PKT_TX_OUTER_IPV6
755 	 * bit[9] = PKT_TX_OUTER_UDP
756 	 */
757 	idx = (buf->ol_flags & ol_flags_mask) >> 52;
758 	if (tunnel == PKT_TX_TUNNEL_UDP)
759 		idx |= 1 << 9;
760 	*swp_types = mlx5_swp_types_table[idx];
761 	/*
762 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
763 	 * complements HW parser. SW parser starts to engage only if HW parser
764 	 * can't reach a header. For the older devices, HW parser will not kick
765 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
766 	 * should be set regardless of HW offload.
767 	 */
768 	off = buf->outer_l2_len + (vlan ? sizeof(struct rte_vlan_hdr) : 0);
769 	offsets[1] = off >> 1; /* Outer L3 offset. */
770 	off += buf->outer_l3_len;
771 	if (tunnel == PKT_TX_TUNNEL_UDP)
772 		offsets[0] = off >> 1; /* Outer L4 offset. */
773 	if (inner_ip) {
774 		off += buf->l2_len;
775 		offsets[3] = off >> 1; /* Inner L3 offset. */
776 		if (csum_flags == PKT_TX_TCP_CKSUM || tso ||
777 		    csum_flags == PKT_TX_UDP_CKSUM) {
778 			off += buf->l3_len;
779 			offsets[2] = off >> 1; /* Inner L4 offset. */
780 		}
781 	}
782 }
783 
784 /**
785  * Convert the Checksum offloads to Verbs.
786  *
787  * @param buf
788  *   Pointer to the mbuf.
789  *
790  * @return
791  *   Converted checksum flags.
792  */
793 static __rte_always_inline uint8_t
794 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
795 {
796 	uint32_t idx;
797 	uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK);
798 	const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK |
799 				       PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM;
800 
801 	/*
802 	 * The index should have:
803 	 * bit[0] = PKT_TX_TCP_SEG
804 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
805 	 * bit[4] = PKT_TX_IP_CKSUM
806 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
807 	 * bit[9] = tunnel
808 	 */
809 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
810 	return mlx5_cksum_table[idx];
811 }
812 
813 /**
814  * Count the number of contiguous single segment packets.
815  *
816  * @param pkts
817  *   Pointer to array of packets.
818  * @param pkts_n
819  *   Number of packets.
820  *
821  * @return
822  *   Number of contiguous single segment packets.
823  */
824 static __rte_always_inline unsigned int
825 txq_count_contig_single_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
826 {
827 	unsigned int pos;
828 
829 	if (!pkts_n)
830 		return 0;
831 	/* Count the number of contiguous single segment packets. */
832 	for (pos = 0; pos < pkts_n; ++pos)
833 		if (NB_SEGS(pkts[pos]) > 1)
834 			break;
835 	return pos;
836 }
837 
838 /**
839  * Count the number of contiguous multi-segment packets.
840  *
841  * @param pkts
842  *   Pointer to array of packets.
843  * @param pkts_n
844  *   Number of packets.
845  *
846  * @return
847  *   Number of contiguous multi-segment packets.
848  */
849 static __rte_always_inline unsigned int
850 txq_count_contig_multi_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
851 {
852 	unsigned int pos;
853 
854 	if (!pkts_n)
855 		return 0;
856 	/* Count the number of contiguous multi-segment packets. */
857 	for (pos = 0; pos < pkts_n; ++pos)
858 		if (NB_SEGS(pkts[pos]) == 1)
859 			break;
860 	return pos;
861 }
862 
863 #endif /* RTE_PMD_MLX5_RXTX_H_ */
864