xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.h (revision 1edccebcccdbe600dc0a3a418fae68336648a87e)
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 
32 #include "mlx5_utils.h"
33 #include "mlx5.h"
34 #include "mlx5_mr.h"
35 #include "mlx5_autoconf.h"
36 #include "mlx5_defs.h"
37 #include "mlx5_prm.h"
38 
39 /* Support tunnel matching. */
40 #define MLX5_FLOW_TUNNEL 5
41 
42 struct mlx5_rxq_stats {
43 	unsigned int idx; /**< Mapping index. */
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 	unsigned int idx; /**< Mapping index. */
54 #ifdef MLX5_PMD_SOFT_COUNTERS
55 	uint64_t opackets; /**< Total of successfully sent packets. */
56 	uint64_t obytes; /**< Total of successfully sent bytes. */
57 #endif
58 	uint64_t oerrors; /**< Total number of failed transmitted packets. */
59 };
60 
61 struct priv;
62 
63 /* Compressed CQE context. */
64 struct rxq_zip {
65 	uint16_t ai; /* Array index. */
66 	uint16_t ca; /* Current array index. */
67 	uint16_t na; /* Next array index. */
68 	uint16_t cq_ci; /* The next CQE. */
69 	uint32_t cqe_cnt; /* Number of CQEs. */
70 };
71 
72 /* Multi-Packet RQ buffer header. */
73 struct mlx5_mprq_buf {
74 	struct rte_mempool *mp;
75 	rte_atomic16_t refcnt; /* Atomically accessed refcnt. */
76 	uint8_t pad[RTE_PKTMBUF_HEADROOM]; /* Headroom for the first packet. */
77 } __rte_cache_aligned;
78 
79 /* Get pointer to the first stride. */
80 #define mlx5_mprq_buf_addr(ptr) ((ptr) + 1)
81 
82 /* RX queue descriptor. */
83 struct mlx5_rxq_data {
84 	unsigned int csum:1; /* Enable checksum offloading. */
85 	unsigned int hw_timestamp:1; /* Enable HW timestamp. */
86 	unsigned int vlan_strip:1; /* Enable VLAN stripping. */
87 	unsigned int crc_present:1; /* CRC must be subtracted. */
88 	unsigned int sges_n:2; /* Log 2 of SGEs (max buffers per packet). */
89 	unsigned int cqe_n:4; /* Log 2 of CQ elements. */
90 	unsigned int elts_n:4; /* Log 2 of Mbufs. */
91 	unsigned int rss_hash:1; /* RSS hash result is enabled. */
92 	unsigned int mark:1; /* Marked flow available on the queue. */
93 	unsigned int strd_num_n:5; /* Log 2 of the number of stride. */
94 	unsigned int strd_sz_n:4; /* Log 2 of stride size. */
95 	unsigned int strd_shift_en:1; /* Enable 2bytes shift on a stride. */
96 	unsigned int :6; /* Remaining bits. */
97 	volatile uint32_t *rq_db;
98 	volatile uint32_t *cq_db;
99 	uint16_t port_id;
100 	uint32_t rq_ci;
101 	uint16_t consumed_strd; /* Number of consumed strides in WQE. */
102 	uint32_t rq_pi;
103 	uint32_t cq_ci;
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 	struct mlx5_rxq_stats stats;
118 	uint64_t mbuf_initializer; /* Default rearm_data for vectorized Rx. */
119 	struct rte_mbuf fake_mbuf; /* elts padding for vectorized Rx. */
120 	void *cq_uar; /* CQ user access region. */
121 	uint32_t cqn; /* CQ number. */
122 	uint8_t cq_arm_sn; /* CQ arm seq number. */
123 #ifndef RTE_ARCH_64
124 	rte_spinlock_t *uar_lock_cq;
125 	/* CQ (UAR) access lock required for 32bit implementations */
126 #endif
127 	uint32_t tunnel; /* Tunnel information. */
128 } __rte_cache_aligned;
129 
130 /* Verbs Rx queue elements. */
131 struct mlx5_rxq_ibv {
132 	LIST_ENTRY(mlx5_rxq_ibv) next; /* Pointer to the next element. */
133 	rte_atomic32_t refcnt; /* Reference counter. */
134 	struct mlx5_rxq_ctrl *rxq_ctrl; /* Back pointer to parent. */
135 	struct ibv_cq *cq; /* Completion Queue. */
136 	struct ibv_wq *wq; /* Work Queue. */
137 	struct ibv_comp_channel *channel;
138 };
139 
140 /* RX queue control descriptor. */
141 struct mlx5_rxq_ctrl {
142 	LIST_ENTRY(mlx5_rxq_ctrl) next; /* Pointer to the next element. */
143 	rte_atomic32_t refcnt; /* Reference counter. */
144 	struct mlx5_rxq_ibv *ibv; /* Verbs elements. */
145 	struct priv *priv; /* Back pointer to private data. */
146 	struct mlx5_rxq_data rxq; /* Data path structure. */
147 	unsigned int socket; /* CPU socket ID for allocations. */
148 	unsigned int irq:1; /* Whether IRQ is enabled. */
149 	uint16_t idx; /* Queue index. */
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 	uint64_t hash_fields; /* Verbs Hash fields. */
170 	uint32_t rss_key_len; /* Hash key length in bytes. */
171 	uint8_t rss_key[]; /* Hash key. */
172 };
173 
174 /* TX queue descriptor. */
175 __extension__
176 struct mlx5_txq_data {
177 	uint16_t elts_head; /* Current counter in (*elts)[]. */
178 	uint16_t elts_tail; /* Counter of first element awaiting completion. */
179 	uint16_t elts_comp; /* Counter since last completion request. */
180 	uint16_t mpw_comp; /* WQ index since last completion request. */
181 	uint16_t cq_ci; /* Consumer index for completion queue. */
182 #ifndef NDEBUG
183 	uint16_t cq_pi; /* Producer index for completion queue. */
184 #endif
185 	uint16_t wqe_ci; /* Consumer index for work queue. */
186 	uint16_t wqe_pi; /* Producer index for work queue. */
187 	uint16_t elts_n:4; /* (*elts)[] length (in log2). */
188 	uint16_t cqe_n:4; /* Number of CQ elements (in log2). */
189 	uint16_t wqe_n:4; /* Number of of WQ elements (in log2). */
190 	uint16_t tso_en:1; /* When set hardware TSO is enabled. */
191 	uint16_t tunnel_en:1;
192 	/* When set TX offload for tunneled packets are supported. */
193 	uint16_t swp_en:1; /* Whether SW parser is enabled. */
194 	uint16_t mpw_hdr_dseg:1; /* Enable DSEGs in the title WQEBB. */
195 	uint16_t max_inline; /* Multiple of RTE_CACHE_LINE_SIZE to inline. */
196 	uint16_t inline_max_packet_sz; /* Max packet size for inlining. */
197 	uint32_t qp_num_8s; /* QP number shifted by 8. */
198 	uint64_t offloads; /* Offloads for Tx Queue. */
199 	struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
200 	volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
201 	volatile void *wqes; /* Work queue (use volatile to write into). */
202 	volatile uint32_t *qp_db; /* Work queue doorbell. */
203 	volatile uint32_t *cq_db; /* Completion queue doorbell. */
204 	volatile void *bf_reg; /* Blueflame register remapped. */
205 	struct rte_mbuf *(*elts)[]; /* TX elements. */
206 	struct mlx5_txq_stats stats; /* TX queue counters. */
207 #ifndef RTE_ARCH_64
208 	rte_spinlock_t *uar_lock;
209 	/* UAR access lock required for 32bit implementations */
210 #endif
211 } __rte_cache_aligned;
212 
213 /* Verbs Rx queue elements. */
214 struct mlx5_txq_ibv {
215 	LIST_ENTRY(mlx5_txq_ibv) next; /* Pointer to the next element. */
216 	rte_atomic32_t refcnt; /* Reference counter. */
217 	struct mlx5_txq_ctrl *txq_ctrl; /* Pointer to the control queue. */
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 	unsigned int socket; /* CPU socket ID for allocations. */
227 	unsigned int max_inline_data; /* Max inline data. */
228 	unsigned int max_tso_header; /* Max TSO header size. */
229 	struct mlx5_txq_ibv *ibv; /* Verbs queue object. */
230 	struct priv *priv; /* Back pointer to private data. */
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 	uint16_t idx; /* Queue index. */
235 };
236 
237 /* mlx5_rxq.c */
238 
239 extern uint8_t rss_hash_default_key[];
240 
241 int mlx5_check_mprq_support(struct rte_eth_dev *dev);
242 int mlx5_rxq_mprq_enabled(struct mlx5_rxq_data *rxq);
243 int mlx5_mprq_enabled(struct rte_eth_dev *dev);
244 int mlx5_mprq_free_mp(struct rte_eth_dev *dev);
245 int mlx5_mprq_alloc_mp(struct rte_eth_dev *dev);
246 void mlx5_rxq_cleanup(struct mlx5_rxq_ctrl *rxq_ctrl);
247 int mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
248 			unsigned int socket, const struct rte_eth_rxconf *conf,
249 			struct rte_mempool *mp);
250 void mlx5_rx_queue_release(void *dpdk_rxq);
251 int mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev);
252 void mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev);
253 int mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
254 int mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
255 struct mlx5_rxq_ibv *mlx5_rxq_ibv_new(struct rte_eth_dev *dev, uint16_t idx);
256 struct mlx5_rxq_ibv *mlx5_rxq_ibv_get(struct rte_eth_dev *dev, uint16_t idx);
257 int mlx5_rxq_ibv_release(struct mlx5_rxq_ibv *rxq_ibv);
258 int mlx5_rxq_ibv_releasable(struct mlx5_rxq_ibv *rxq_ibv);
259 struct mlx5_rxq_ibv *mlx5_rxq_ibv_drop_new(struct rte_eth_dev *dev);
260 void mlx5_rxq_ibv_drop_release(struct rte_eth_dev *dev);
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_releasable(struct rte_eth_dev *dev, uint16_t idx);
269 int mlx5_rxq_verify(struct rte_eth_dev *dev);
270 int rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl);
271 int rxq_alloc_mprq_buf(struct mlx5_rxq_ctrl *rxq_ctrl);
272 struct mlx5_ind_table_ibv *mlx5_ind_table_ibv_new(struct rte_eth_dev *dev,
273 						  const uint16_t *queues,
274 						  uint32_t queues_n);
275 struct mlx5_ind_table_ibv *mlx5_ind_table_ibv_get(struct rte_eth_dev *dev,
276 						  const uint16_t *queues,
277 						  uint32_t queues_n);
278 int mlx5_ind_table_ibv_release(struct rte_eth_dev *dev,
279 			       struct mlx5_ind_table_ibv *ind_tbl);
280 int mlx5_ind_table_ibv_verify(struct rte_eth_dev *dev);
281 struct mlx5_ind_table_ibv *mlx5_ind_table_ibv_drop_new(struct rte_eth_dev *dev);
282 void mlx5_ind_table_ibv_drop_release(struct rte_eth_dev *dev);
283 struct mlx5_hrxq *mlx5_hrxq_new(struct rte_eth_dev *dev,
284 				const uint8_t *rss_key, uint32_t rss_key_len,
285 				uint64_t hash_fields,
286 				const uint16_t *queues, uint32_t queues_n,
287 				int tunnel __rte_unused);
288 struct mlx5_hrxq *mlx5_hrxq_get(struct rte_eth_dev *dev,
289 				const uint8_t *rss_key, uint32_t rss_key_len,
290 				uint64_t hash_fields,
291 				const uint16_t *queues, uint32_t queues_n);
292 int mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hxrq);
293 int mlx5_hrxq_ibv_verify(struct rte_eth_dev *dev);
294 struct mlx5_hrxq *mlx5_hrxq_drop_new(struct rte_eth_dev *dev);
295 void mlx5_hrxq_drop_release(struct rte_eth_dev *dev);
296 uint64_t mlx5_get_rx_port_offloads(void);
297 uint64_t mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev);
298 
299 /* mlx5_txq.c */
300 
301 int mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
302 			unsigned int socket, const struct rte_eth_txconf *conf);
303 void mlx5_tx_queue_release(void *dpdk_txq);
304 int mlx5_tx_uar_remap(struct rte_eth_dev *dev, int fd);
305 struct mlx5_txq_ibv *mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx);
306 struct mlx5_txq_ibv *mlx5_txq_ibv_get(struct rte_eth_dev *dev, uint16_t idx);
307 int mlx5_txq_ibv_release(struct mlx5_txq_ibv *txq_ibv);
308 int mlx5_txq_ibv_releasable(struct mlx5_txq_ibv *txq_ibv);
309 int mlx5_txq_ibv_verify(struct rte_eth_dev *dev);
310 struct mlx5_txq_ctrl *mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx,
311 				   uint16_t desc, unsigned int socket,
312 				   const struct rte_eth_txconf *conf);
313 struct mlx5_txq_ctrl *mlx5_txq_get(struct rte_eth_dev *dev, uint16_t idx);
314 int mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx);
315 int mlx5_txq_releasable(struct rte_eth_dev *dev, uint16_t idx);
316 int mlx5_txq_verify(struct rte_eth_dev *dev);
317 void txq_alloc_elts(struct mlx5_txq_ctrl *txq_ctrl);
318 uint64_t mlx5_get_tx_port_offloads(struct rte_eth_dev *dev);
319 
320 /* mlx5_rxtx.c */
321 
322 extern uint32_t mlx5_ptype_table[];
323 extern uint8_t mlx5_cksum_table[];
324 extern uint8_t mlx5_swp_types_table[];
325 
326 void mlx5_set_ptype_table(void);
327 void mlx5_set_cksum_table(void);
328 void mlx5_set_swp_types_table(void);
329 uint16_t mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts,
330 		       uint16_t pkts_n);
331 uint16_t mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts,
332 			   uint16_t pkts_n);
333 uint16_t mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
334 				  uint16_t pkts_n);
335 uint16_t mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts,
336 			    uint16_t pkts_n);
337 uint16_t mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
338 void mlx5_mprq_buf_free_cb(void *addr, void *opaque);
339 void mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf);
340 uint16_t mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts,
341 			    uint16_t pkts_n);
342 uint16_t removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts,
343 			  uint16_t pkts_n);
344 uint16_t removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts,
345 			  uint16_t pkts_n);
346 int mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset);
347 int mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset);
348 
349 /* Vectorized version of mlx5_rxtx.c */
350 int mlx5_check_raw_vec_tx_support(struct rte_eth_dev *dev);
351 int mlx5_check_vec_tx_support(struct rte_eth_dev *dev);
352 int mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq_data);
353 int mlx5_check_vec_rx_support(struct rte_eth_dev *dev);
354 uint16_t mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts,
355 			       uint16_t pkts_n);
356 uint16_t mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts,
357 			   uint16_t pkts_n);
358 uint16_t mlx5_rx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts,
359 			   uint16_t pkts_n);
360 
361 /* mlx5_mr.c */
362 
363 void mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl);
364 uint32_t mlx5_rx_addr2mr_bh(struct mlx5_rxq_data *rxq, uintptr_t addr);
365 uint32_t mlx5_tx_addr2mr_bh(struct mlx5_txq_data *txq, uintptr_t addr);
366 uint32_t mlx5_tx_update_ext_mp(struct mlx5_txq_data *txq, uintptr_t addr,
367 			       struct rte_mempool *mp);
368 
369 /**
370  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
371  * 64bit architectures.
372  *
373  * @param val
374  *   value to write in CPU endian format.
375  * @param addr
376  *   Address to write to.
377  * @param lock
378  *   Address of the lock to use for that UAR access.
379  */
380 static __rte_always_inline void
381 __mlx5_uar_write64_relaxed(uint64_t val, volatile void *addr,
382 			   rte_spinlock_t *lock __rte_unused)
383 {
384 #ifdef RTE_ARCH_64
385 	rte_write64_relaxed(val, addr);
386 #else /* !RTE_ARCH_64 */
387 	rte_spinlock_lock(lock);
388 	rte_write32_relaxed(val, addr);
389 	rte_io_wmb();
390 	rte_write32_relaxed(val >> 32,
391 			    (volatile void *)((volatile char *)addr + 4));
392 	rte_spinlock_unlock(lock);
393 #endif
394 }
395 
396 /**
397  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
398  * 64bit architectures while guaranteeing the order of execution with the
399  * code being executed.
400  *
401  * @param val
402  *   value to write in CPU endian format.
403  * @param addr
404  *   Address to write to.
405  * @param lock
406  *   Address of the lock to use for that UAR access.
407  */
408 static __rte_always_inline void
409 __mlx5_uar_write64(uint64_t val, volatile void *addr, rte_spinlock_t *lock)
410 {
411 	rte_io_wmb();
412 	__mlx5_uar_write64_relaxed(val, addr, lock);
413 }
414 
415 /* Assist macros, used instead of directly calling the functions they wrap. */
416 #ifdef RTE_ARCH_64
417 #define mlx5_uar_write64_relaxed(val, dst, lock) \
418 		__mlx5_uar_write64_relaxed(val, dst, NULL)
419 #define mlx5_uar_write64(val, dst, lock) __mlx5_uar_write64(val, dst, NULL)
420 #else
421 #define mlx5_uar_write64_relaxed(val, dst, lock) \
422 		__mlx5_uar_write64_relaxed(val, dst, lock)
423 #define mlx5_uar_write64(val, dst, lock) __mlx5_uar_write64(val, dst, lock)
424 #endif
425 
426 #ifndef NDEBUG
427 /**
428  * Verify or set magic value in CQE.
429  *
430  * @param cqe
431  *   Pointer to CQE.
432  *
433  * @return
434  *   0 the first time.
435  */
436 static inline int
437 check_cqe_seen(volatile struct mlx5_cqe *cqe)
438 {
439 	static const uint8_t magic[] = "seen";
440 	volatile uint8_t (*buf)[sizeof(cqe->rsvd1)] = &cqe->rsvd1;
441 	int ret = 1;
442 	unsigned int i;
443 
444 	for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
445 		if (!ret || (*buf)[i] != magic[i]) {
446 			ret = 0;
447 			(*buf)[i] = magic[i];
448 		}
449 	return ret;
450 }
451 #endif /* NDEBUG */
452 
453 /**
454  * Check whether CQE is valid.
455  *
456  * @param cqe
457  *   Pointer to CQE.
458  * @param cqes_n
459  *   Size of completion queue.
460  * @param ci
461  *   Consumer index.
462  *
463  * @return
464  *   0 on success, 1 on failure.
465  */
466 static __rte_always_inline int
467 check_cqe(volatile struct mlx5_cqe *cqe,
468 	  unsigned int cqes_n, const uint16_t ci)
469 {
470 	uint16_t idx = ci & cqes_n;
471 	uint8_t op_own = cqe->op_own;
472 	uint8_t op_owner = MLX5_CQE_OWNER(op_own);
473 	uint8_t op_code = MLX5_CQE_OPCODE(op_own);
474 
475 	if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
476 		return 1; /* No CQE. */
477 #ifndef NDEBUG
478 	if ((op_code == MLX5_CQE_RESP_ERR) ||
479 	    (op_code == MLX5_CQE_REQ_ERR)) {
480 		volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
481 		uint8_t syndrome = err_cqe->syndrome;
482 
483 		if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
484 		    (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
485 			return 0;
486 		if (!check_cqe_seen(cqe)) {
487 			DRV_LOG(ERR,
488 				"unexpected CQE error %u (0x%02x) syndrome"
489 				" 0x%02x",
490 				op_code, op_code, syndrome);
491 			rte_hexdump(stderr, "MLX5 Error CQE:",
492 				    (const void *)((uintptr_t)err_cqe),
493 				    sizeof(*err_cqe));
494 		}
495 		return 1;
496 	} else if ((op_code != MLX5_CQE_RESP_SEND) &&
497 		   (op_code != MLX5_CQE_REQ)) {
498 		if (!check_cqe_seen(cqe)) {
499 			DRV_LOG(ERR, "unexpected CQE opcode %u (0x%02x)",
500 				op_code, op_code);
501 			rte_hexdump(stderr, "MLX5 CQE:",
502 				    (const void *)((uintptr_t)cqe),
503 				    sizeof(*cqe));
504 		}
505 		return 1;
506 	}
507 #endif /* NDEBUG */
508 	return 0;
509 }
510 
511 /**
512  * Return the address of the WQE.
513  *
514  * @param txq
515  *   Pointer to TX queue structure.
516  * @param  wqe_ci
517  *   WQE consumer index.
518  *
519  * @return
520  *   WQE address.
521  */
522 static inline uintptr_t *
523 tx_mlx5_wqe(struct mlx5_txq_data *txq, uint16_t ci)
524 {
525 	ci &= ((1 << txq->wqe_n) - 1);
526 	return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
527 }
528 
529 /**
530  * Manage TX completions.
531  *
532  * When sending a burst, mlx5_tx_burst() posts several WRs.
533  *
534  * @param txq
535  *   Pointer to TX queue structure.
536  */
537 static __rte_always_inline void
538 mlx5_tx_complete(struct mlx5_txq_data *txq)
539 {
540 	const uint16_t elts_n = 1 << txq->elts_n;
541 	const uint16_t elts_m = elts_n - 1;
542 	const unsigned int cqe_n = 1 << txq->cqe_n;
543 	const unsigned int cqe_cnt = cqe_n - 1;
544 	uint16_t elts_free = txq->elts_tail;
545 	uint16_t elts_tail;
546 	uint16_t cq_ci = txq->cq_ci;
547 	volatile struct mlx5_cqe *cqe = NULL;
548 	volatile struct mlx5_wqe_ctrl *ctrl;
549 	struct rte_mbuf *m, *free[elts_n];
550 	struct rte_mempool *pool = NULL;
551 	unsigned int blk_n = 0;
552 
553 	cqe = &(*txq->cqes)[cq_ci & cqe_cnt];
554 	if (unlikely(check_cqe(cqe, cqe_n, cq_ci)))
555 		return;
556 #ifndef NDEBUG
557 	if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
558 	    (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
559 		if (!check_cqe_seen(cqe)) {
560 			DRV_LOG(ERR, "unexpected error CQE, Tx stopped");
561 			rte_hexdump(stderr, "MLX5 TXQ:",
562 				    (const void *)((uintptr_t)txq->wqes),
563 				    ((1 << txq->wqe_n) *
564 				     MLX5_WQE_SIZE));
565 		}
566 		return;
567 	}
568 #endif /* NDEBUG */
569 	++cq_ci;
570 	txq->wqe_pi = rte_be_to_cpu_16(cqe->wqe_counter);
571 	ctrl = (volatile struct mlx5_wqe_ctrl *)
572 		tx_mlx5_wqe(txq, txq->wqe_pi);
573 	elts_tail = ctrl->ctrl3;
574 	assert((elts_tail & elts_m) < (1 << txq->wqe_n));
575 	/* Free buffers. */
576 	while (elts_free != elts_tail) {
577 		m = rte_pktmbuf_prefree_seg((*txq->elts)[elts_free++ & elts_m]);
578 		if (likely(m != NULL)) {
579 			if (likely(m->pool == pool)) {
580 				free[blk_n++] = m;
581 			} else {
582 				if (likely(pool != NULL))
583 					rte_mempool_put_bulk(pool,
584 							     (void *)free,
585 							     blk_n);
586 				free[0] = m;
587 				pool = m->pool;
588 				blk_n = 1;
589 			}
590 		}
591 	}
592 	if (blk_n)
593 		rte_mempool_put_bulk(pool, (void *)free, blk_n);
594 #ifndef NDEBUG
595 	elts_free = txq->elts_tail;
596 	/* Poisoning. */
597 	while (elts_free != elts_tail) {
598 		memset(&(*txq->elts)[elts_free & elts_m],
599 		       0x66,
600 		       sizeof((*txq->elts)[elts_free & elts_m]));
601 		++elts_free;
602 	}
603 #endif
604 	txq->cq_ci = cq_ci;
605 	txq->elts_tail = elts_tail;
606 	/* Update the consumer index. */
607 	rte_compiler_barrier();
608 	*txq->cq_db = rte_cpu_to_be_32(cq_ci);
609 }
610 
611 /**
612  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which the
613  * cloned mbuf is allocated is returned instead.
614  *
615  * @param buf
616  *   Pointer to mbuf.
617  *
618  * @return
619  *   Memory pool where data is located for given mbuf.
620  */
621 static struct rte_mempool *
622 mlx5_mb2mp(struct rte_mbuf *buf)
623 {
624 	if (unlikely(RTE_MBUF_INDIRECT(buf)))
625 		return rte_mbuf_from_indirect(buf)->pool;
626 	return buf->pool;
627 }
628 
629 /**
630  * Query LKey from a packet buffer for Rx. No need to flush local caches for Rx
631  * as mempool is pre-configured and static.
632  *
633  * @param rxq
634  *   Pointer to Rx queue structure.
635  * @param addr
636  *   Address to search.
637  *
638  * @return
639  *   Searched LKey on success, UINT32_MAX on no match.
640  */
641 static __rte_always_inline uint32_t
642 mlx5_rx_addr2mr(struct mlx5_rxq_data *rxq, uintptr_t addr)
643 {
644 	struct mlx5_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
645 	uint32_t lkey;
646 
647 	/* Linear search on MR cache array. */
648 	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
649 				    MLX5_MR_CACHE_N, addr);
650 	if (likely(lkey != UINT32_MAX))
651 		return lkey;
652 	/* Take slower bottom-half (Binary Search) on miss. */
653 	return mlx5_rx_addr2mr_bh(rxq, addr);
654 }
655 
656 #define mlx5_rx_mb2mr(rxq, mb) mlx5_rx_addr2mr(rxq, (uintptr_t)((mb)->buf_addr))
657 
658 /**
659  * Query LKey from a packet buffer for Tx. If not found, add the mempool.
660  *
661  * @param txq
662  *   Pointer to Tx queue structure.
663  * @param addr
664  *   Address to search.
665  *
666  * @return
667  *   Searched LKey on success, UINT32_MAX on no match.
668  */
669 static __rte_always_inline uint32_t
670 mlx5_tx_addr2mr(struct mlx5_txq_data *txq, uintptr_t addr)
671 {
672 	struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
673 	uint32_t lkey;
674 
675 	/* Check generation bit to see if there's any change on existing MRs. */
676 	if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen))
677 		mlx5_mr_flush_local_cache(mr_ctrl);
678 	/* Linear search on MR cache array. */
679 	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
680 				    MLX5_MR_CACHE_N, addr);
681 	if (likely(lkey != UINT32_MAX))
682 		return lkey;
683 	/* Take slower bottom-half (binary search) on miss. */
684 	return mlx5_tx_addr2mr_bh(txq, addr);
685 }
686 
687 static __rte_always_inline uint32_t
688 mlx5_tx_mb2mr(struct mlx5_txq_data *txq, struct rte_mbuf *mb)
689 {
690 	uintptr_t addr = (uintptr_t)mb->buf_addr;
691 	uint32_t lkey = mlx5_tx_addr2mr(txq, addr);
692 
693 	if (likely(lkey != UINT32_MAX))
694 		return lkey;
695 	if (rte_errno == ENXIO) {
696 		/* Mempool may have externally allocated memory. */
697 		lkey = mlx5_tx_update_ext_mp(txq, addr, mlx5_mb2mp(mb));
698 	}
699 	return lkey;
700 }
701 
702 /**
703  * Ring TX queue doorbell and flush the update if requested.
704  *
705  * @param txq
706  *   Pointer to TX queue structure.
707  * @param wqe
708  *   Pointer to the last WQE posted in the NIC.
709  * @param cond
710  *   Request for write memory barrier after BlueFlame update.
711  */
712 static __rte_always_inline void
713 mlx5_tx_dbrec_cond_wmb(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe,
714 		       int cond)
715 {
716 	uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
717 	volatile uint64_t *src = ((volatile uint64_t *)wqe);
718 
719 	rte_cio_wmb();
720 	*txq->qp_db = rte_cpu_to_be_32(txq->wqe_ci);
721 	/* Ensure ordering between DB record and BF copy. */
722 	rte_wmb();
723 	mlx5_uar_write64_relaxed(*src, dst, txq->uar_lock);
724 	if (cond)
725 		rte_wmb();
726 }
727 
728 /**
729  * Ring TX queue doorbell and flush the update by write memory barrier.
730  *
731  * @param txq
732  *   Pointer to TX queue structure.
733  * @param wqe
734  *   Pointer to the last WQE posted in the NIC.
735  */
736 static __rte_always_inline void
737 mlx5_tx_dbrec(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe)
738 {
739 	mlx5_tx_dbrec_cond_wmb(txq, wqe, 1);
740 }
741 
742 /**
743  * Convert mbuf to Verb SWP.
744  *
745  * @param txq_data
746  *   Pointer to the Tx queue.
747  * @param buf
748  *   Pointer to the mbuf.
749  * @param tso
750  *   TSO offloads enabled.
751  * @param vlan
752  *   VLAN offloads enabled
753  * @param offsets
754  *   Pointer to the SWP header offsets.
755  * @param swp_types
756  *   Pointer to the SWP header types.
757  */
758 static __rte_always_inline void
759 txq_mbuf_to_swp(struct mlx5_txq_data *txq, struct rte_mbuf *buf,
760 		uint8_t *offsets, uint8_t *swp_types)
761 {
762 	const uint64_t vlan = buf->ol_flags & PKT_TX_VLAN_PKT;
763 	const uint64_t tunnel = buf->ol_flags & PKT_TX_TUNNEL_MASK;
764 	const uint64_t tso = buf->ol_flags & PKT_TX_TCP_SEG;
765 	const uint64_t csum_flags = buf->ol_flags & PKT_TX_L4_MASK;
766 	const uint64_t inner_ip =
767 		buf->ol_flags & (PKT_TX_IPV4 | PKT_TX_IPV6);
768 	const uint64_t ol_flags_mask = PKT_TX_L4_MASK | PKT_TX_IPV6 |
769 				       PKT_TX_OUTER_IPV6;
770 	uint16_t idx;
771 	uint16_t off;
772 
773 	if (likely(!txq->swp_en || (tunnel != PKT_TX_TUNNEL_UDP &&
774 				    tunnel != PKT_TX_TUNNEL_IP)))
775 		return;
776 	/*
777 	 * The index should have:
778 	 * bit[0:1] = PKT_TX_L4_MASK
779 	 * bit[4] = PKT_TX_IPV6
780 	 * bit[8] = PKT_TX_OUTER_IPV6
781 	 * bit[9] = PKT_TX_OUTER_UDP
782 	 */
783 	idx = (buf->ol_flags & ol_flags_mask) >> 52;
784 	if (tunnel == PKT_TX_TUNNEL_UDP)
785 		idx |= 1 << 9;
786 	*swp_types = mlx5_swp_types_table[idx];
787 	/*
788 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
789 	 * complements HW parser. SW parser starts to engage only if HW parser
790 	 * can't reach a header. For the older devices, HW parser will not kick
791 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
792 	 * should be set regardless of HW offload.
793 	 */
794 	off = buf->outer_l2_len + (vlan ? sizeof(struct vlan_hdr) : 0);
795 	offsets[1] = off >> 1; /* Outer L3 offset. */
796 	off += buf->outer_l3_len;
797 	if (tunnel == PKT_TX_TUNNEL_UDP)
798 		offsets[0] = off >> 1; /* Outer L4 offset. */
799 	if (inner_ip) {
800 		off += buf->l2_len;
801 		offsets[3] = off >> 1; /* Inner L3 offset. */
802 		if (csum_flags == PKT_TX_TCP_CKSUM || tso ||
803 		    csum_flags == PKT_TX_UDP_CKSUM) {
804 			off += buf->l3_len;
805 			offsets[2] = off >> 1; /* Inner L4 offset. */
806 		}
807 	}
808 }
809 
810 /**
811  * Convert the Checksum offloads to Verbs.
812  *
813  * @param buf
814  *   Pointer to the mbuf.
815  *
816  * @return
817  *   Converted checksum flags.
818  */
819 static __rte_always_inline uint8_t
820 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
821 {
822 	uint32_t idx;
823 	uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK);
824 	const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK |
825 				       PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM;
826 
827 	/*
828 	 * The index should have:
829 	 * bit[0] = PKT_TX_TCP_SEG
830 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
831 	 * bit[4] = PKT_TX_IP_CKSUM
832 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
833 	 * bit[9] = tunnel
834 	 */
835 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
836 	return mlx5_cksum_table[idx];
837 }
838 
839 /**
840  * Count the number of contiguous single segment packets.
841  *
842  * @param pkts
843  *   Pointer to array of packets.
844  * @param pkts_n
845  *   Number of packets.
846  *
847  * @return
848  *   Number of contiguous single segment packets.
849  */
850 static __rte_always_inline unsigned int
851 txq_count_contig_single_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
852 {
853 	unsigned int pos;
854 
855 	if (!pkts_n)
856 		return 0;
857 	/* Count the number of contiguous single segment packets. */
858 	for (pos = 0; pos < pkts_n; ++pos)
859 		if (NB_SEGS(pkts[pos]) > 1)
860 			break;
861 	return pos;
862 }
863 
864 /**
865  * Count the number of contiguous multi-segment packets.
866  *
867  * @param pkts
868  *   Pointer to array of packets.
869  * @param pkts_n
870  *   Number of packets.
871  *
872  * @return
873  *   Number of contiguous multi-segment packets.
874  */
875 static __rte_always_inline unsigned int
876 txq_count_contig_multi_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
877 {
878 	unsigned int pos;
879 
880 	if (!pkts_n)
881 		return 0;
882 	/* Count the number of contiguous multi-segment packets. */
883 	for (pos = 0; pos < pkts_n; ++pos)
884 		if (NB_SEGS(pkts[pos]) == 1)
885 			break;
886 	return pos;
887 }
888 
889 #endif /* RTE_PMD_MLX5_RXTX_H_ */
890