xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.h (revision 5ecb687a5698d2d8ec1f3b3b5a7a16bceca3e29c)
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 	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 	struct rte_mbuf *(*elts)[]; /* TX elements. */
205 	uint16_t port_id; /* Port ID of device. */
206 	uint16_t idx; /* Queue index. */
207 	struct mlx5_txq_stats stats; /* TX queue counters. */
208 #ifndef RTE_ARCH_64
209 	rte_spinlock_t *uar_lock;
210 	/* UAR access lock required for 32bit implementations */
211 #endif
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 mlx5_txq_ctrl *txq_ctrl; /* Pointer to the control queue. */
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 	struct mlx5_txq_data txq; /* Data path structure. */
226 	LIST_ENTRY(mlx5_txq_ctrl) next; /* Pointer to the next element. */
227 	rte_atomic32_t refcnt; /* Reference counter. */
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_priv *priv; /* Back pointer to private data. */
233 	off_t uar_mmap_offset; /* UAR mmap offset for non-primary process. */
234 	void *bf_reg; /* BlueFlame register from Verbs. */
235 };
236 
237 #define MLX5_TX_BFREG(txq) \
238 		(MLX5_PROC_PRIV((txq)->port_id)->uar_table[(txq)->idx])
239 
240 /* mlx5_rxq.c */
241 
242 extern uint8_t rss_hash_default_key[];
243 
244 int mlx5_check_mprq_support(struct rte_eth_dev *dev);
245 int mlx5_rxq_mprq_enabled(struct mlx5_rxq_data *rxq);
246 int mlx5_mprq_enabled(struct rte_eth_dev *dev);
247 int mlx5_mprq_free_mp(struct rte_eth_dev *dev);
248 int mlx5_mprq_alloc_mp(struct rte_eth_dev *dev);
249 void mlx5_rxq_cleanup(struct mlx5_rxq_ctrl *rxq_ctrl);
250 int mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
251 			unsigned int socket, const struct rte_eth_rxconf *conf,
252 			struct rte_mempool *mp);
253 void mlx5_rx_queue_release(void *dpdk_rxq);
254 int mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev);
255 void mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev);
256 int mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
257 int mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
258 struct mlx5_rxq_ibv *mlx5_rxq_ibv_new(struct rte_eth_dev *dev, uint16_t idx);
259 struct mlx5_rxq_ibv *mlx5_rxq_ibv_get(struct rte_eth_dev *dev, uint16_t idx);
260 int mlx5_rxq_ibv_release(struct mlx5_rxq_ibv *rxq_ibv);
261 int mlx5_rxq_ibv_releasable(struct mlx5_rxq_ibv *rxq_ibv);
262 struct mlx5_rxq_ibv *mlx5_rxq_ibv_drop_new(struct rte_eth_dev *dev);
263 void mlx5_rxq_ibv_drop_release(struct rte_eth_dev *dev);
264 int mlx5_rxq_ibv_verify(struct rte_eth_dev *dev);
265 struct mlx5_rxq_ctrl *mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx,
266 				   uint16_t desc, unsigned int socket,
267 				   const struct rte_eth_rxconf *conf,
268 				   struct rte_mempool *mp);
269 struct mlx5_rxq_ctrl *mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx);
270 int mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx);
271 int mlx5_rxq_releasable(struct rte_eth_dev *dev, uint16_t idx);
272 int mlx5_rxq_verify(struct rte_eth_dev *dev);
273 int rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl);
274 int rxq_alloc_mprq_buf(struct mlx5_rxq_ctrl *rxq_ctrl);
275 struct mlx5_ind_table_ibv *mlx5_ind_table_ibv_new(struct rte_eth_dev *dev,
276 						  const uint16_t *queues,
277 						  uint32_t queues_n);
278 struct mlx5_ind_table_ibv *mlx5_ind_table_ibv_get(struct rte_eth_dev *dev,
279 						  const uint16_t *queues,
280 						  uint32_t queues_n);
281 int mlx5_ind_table_ibv_release(struct rte_eth_dev *dev,
282 			       struct mlx5_ind_table_ibv *ind_tbl);
283 int mlx5_ind_table_ibv_verify(struct rte_eth_dev *dev);
284 struct mlx5_ind_table_ibv *mlx5_ind_table_ibv_drop_new(struct rte_eth_dev *dev);
285 void mlx5_ind_table_ibv_drop_release(struct rte_eth_dev *dev);
286 struct mlx5_hrxq *mlx5_hrxq_new(struct rte_eth_dev *dev,
287 				const uint8_t *rss_key, uint32_t rss_key_len,
288 				uint64_t hash_fields,
289 				const uint16_t *queues, uint32_t queues_n,
290 				int tunnel __rte_unused);
291 struct mlx5_hrxq *mlx5_hrxq_get(struct rte_eth_dev *dev,
292 				const uint8_t *rss_key, uint32_t rss_key_len,
293 				uint64_t hash_fields,
294 				const uint16_t *queues, uint32_t queues_n);
295 int mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hxrq);
296 int mlx5_hrxq_ibv_verify(struct rte_eth_dev *dev);
297 struct mlx5_hrxq *mlx5_hrxq_drop_new(struct rte_eth_dev *dev);
298 void mlx5_hrxq_drop_release(struct rte_eth_dev *dev);
299 uint64_t mlx5_get_rx_port_offloads(void);
300 uint64_t mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev);
301 
302 /* mlx5_txq.c */
303 
304 int mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
305 			unsigned int socket, const struct rte_eth_txconf *conf);
306 void mlx5_tx_queue_release(void *dpdk_txq);
307 int mlx5_tx_uar_init_secondary(struct rte_eth_dev *dev, int fd);
308 struct mlx5_txq_ibv *mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx);
309 struct mlx5_txq_ibv *mlx5_txq_ibv_get(struct rte_eth_dev *dev, uint16_t idx);
310 int mlx5_txq_ibv_release(struct mlx5_txq_ibv *txq_ibv);
311 int mlx5_txq_ibv_releasable(struct mlx5_txq_ibv *txq_ibv);
312 int mlx5_txq_ibv_verify(struct rte_eth_dev *dev);
313 struct mlx5_txq_ctrl *mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx,
314 				   uint16_t desc, unsigned int socket,
315 				   const struct rte_eth_txconf *conf);
316 struct mlx5_txq_ctrl *mlx5_txq_get(struct rte_eth_dev *dev, uint16_t idx);
317 int mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx);
318 int mlx5_txq_releasable(struct rte_eth_dev *dev, uint16_t idx);
319 int mlx5_txq_verify(struct rte_eth_dev *dev);
320 void txq_alloc_elts(struct mlx5_txq_ctrl *txq_ctrl);
321 uint64_t mlx5_get_tx_port_offloads(struct rte_eth_dev *dev);
322 
323 /* mlx5_rxtx.c */
324 
325 extern uint32_t mlx5_ptype_table[];
326 extern uint8_t mlx5_cksum_table[];
327 extern uint8_t mlx5_swp_types_table[];
328 
329 void mlx5_set_ptype_table(void);
330 void mlx5_set_cksum_table(void);
331 void mlx5_set_swp_types_table(void);
332 uint16_t mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts,
333 		       uint16_t pkts_n);
334 uint16_t mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts,
335 			   uint16_t pkts_n);
336 uint16_t mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
337 				  uint16_t pkts_n);
338 uint16_t mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts,
339 			    uint16_t pkts_n);
340 uint16_t mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
341 void mlx5_mprq_buf_free_cb(void *addr, void *opaque);
342 void mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf);
343 uint16_t mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts,
344 			    uint16_t pkts_n);
345 uint16_t removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts,
346 			  uint16_t pkts_n);
347 uint16_t removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts,
348 			  uint16_t pkts_n);
349 int mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset);
350 int mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset);
351 uint32_t mlx5_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id);
352 
353 /* Vectorized version of mlx5_rxtx.c */
354 int mlx5_check_raw_vec_tx_support(struct rte_eth_dev *dev);
355 int mlx5_check_vec_tx_support(struct rte_eth_dev *dev);
356 int mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq_data);
357 int mlx5_check_vec_rx_support(struct rte_eth_dev *dev);
358 uint16_t mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts,
359 			       uint16_t pkts_n);
360 uint16_t mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts,
361 			   uint16_t pkts_n);
362 uint16_t mlx5_rx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts,
363 			   uint16_t pkts_n);
364 
365 /* mlx5_mr.c */
366 
367 void mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl);
368 uint32_t mlx5_rx_addr2mr_bh(struct mlx5_rxq_data *rxq, uintptr_t addr);
369 uint32_t mlx5_tx_mb2mr_bh(struct mlx5_txq_data *txq, struct rte_mbuf *mb);
370 uint32_t mlx5_tx_update_ext_mp(struct mlx5_txq_data *txq, uintptr_t addr,
371 			       struct rte_mempool *mp);
372 int mlx5_dma_map(struct rte_pci_device *pdev, void *addr, uint64_t iova,
373 		 size_t len);
374 int mlx5_dma_unmap(struct rte_pci_device *pdev, void *addr, uint64_t iova,
375 		   size_t len);
376 
377 /**
378  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
379  * 64bit architectures.
380  *
381  * @param val
382  *   value to write in CPU endian format.
383  * @param addr
384  *   Address to write to.
385  * @param lock
386  *   Address of the lock to use for that UAR access.
387  */
388 static __rte_always_inline void
389 __mlx5_uar_write64_relaxed(uint64_t val, void *addr,
390 			   rte_spinlock_t *lock __rte_unused)
391 {
392 #ifdef RTE_ARCH_64
393 	*(uint64_t *)addr = val;
394 #else /* !RTE_ARCH_64 */
395 	rte_spinlock_lock(lock);
396 	*(uint32_t *)addr = val;
397 	rte_io_wmb();
398 	*((uint32_t *)addr + 1) = val >> 32;
399 	rte_spinlock_unlock(lock);
400 #endif
401 }
402 
403 /**
404  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
405  * 64bit architectures while guaranteeing the order of execution with the
406  * code being executed.
407  *
408  * @param val
409  *   value to write in CPU endian format.
410  * @param addr
411  *   Address to write to.
412  * @param lock
413  *   Address of the lock to use for that UAR access.
414  */
415 static __rte_always_inline void
416 __mlx5_uar_write64(uint64_t val, void *addr, rte_spinlock_t *lock)
417 {
418 	rte_io_wmb();
419 	__mlx5_uar_write64_relaxed(val, addr, lock);
420 }
421 
422 /* Assist macros, used instead of directly calling the functions they wrap. */
423 #ifdef RTE_ARCH_64
424 #define mlx5_uar_write64_relaxed(val, dst, lock) \
425 		__mlx5_uar_write64_relaxed(val, dst, NULL)
426 #define mlx5_uar_write64(val, dst, lock) __mlx5_uar_write64(val, dst, NULL)
427 #else
428 #define mlx5_uar_write64_relaxed(val, dst, lock) \
429 		__mlx5_uar_write64_relaxed(val, dst, lock)
430 #define mlx5_uar_write64(val, dst, lock) __mlx5_uar_write64(val, dst, lock)
431 #endif
432 
433 #ifndef NDEBUG
434 /**
435  * Verify or set magic value in CQE.
436  *
437  * @param cqe
438  *   Pointer to CQE.
439  *
440  * @return
441  *   0 the first time.
442  */
443 static inline int
444 check_cqe_seen(volatile struct mlx5_cqe *cqe)
445 {
446 	static const uint8_t magic[] = "seen";
447 	volatile uint8_t (*buf)[sizeof(cqe->rsvd1)] = &cqe->rsvd1;
448 	int ret = 1;
449 	unsigned int i;
450 
451 	for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
452 		if (!ret || (*buf)[i] != magic[i]) {
453 			ret = 0;
454 			(*buf)[i] = magic[i];
455 		}
456 	return ret;
457 }
458 #endif /* NDEBUG */
459 
460 /**
461  * Check whether CQE is valid.
462  *
463  * @param cqe
464  *   Pointer to CQE.
465  * @param cqes_n
466  *   Size of completion queue.
467  * @param ci
468  *   Consumer index.
469  *
470  * @return
471  *   0 on success, 1 on failure.
472  */
473 static __rte_always_inline int
474 check_cqe(volatile struct mlx5_cqe *cqe,
475 	  unsigned int cqes_n, const uint16_t ci)
476 {
477 	uint16_t idx = ci & cqes_n;
478 	uint8_t op_own = cqe->op_own;
479 	uint8_t op_owner = MLX5_CQE_OWNER(op_own);
480 	uint8_t op_code = MLX5_CQE_OPCODE(op_own);
481 
482 	if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
483 		return 1; /* No CQE. */
484 #ifndef NDEBUG
485 	if ((op_code == MLX5_CQE_RESP_ERR) ||
486 	    (op_code == MLX5_CQE_REQ_ERR)) {
487 		volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
488 		uint8_t syndrome = err_cqe->syndrome;
489 
490 		if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
491 		    (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
492 			return 0;
493 		if (!check_cqe_seen(cqe)) {
494 			DRV_LOG(ERR,
495 				"unexpected CQE error %u (0x%02x) syndrome"
496 				" 0x%02x",
497 				op_code, op_code, syndrome);
498 			rte_hexdump(stderr, "MLX5 Error CQE:",
499 				    (const void *)((uintptr_t)err_cqe),
500 				    sizeof(*cqe));
501 		}
502 		return 1;
503 	} else if ((op_code != MLX5_CQE_RESP_SEND) &&
504 		   (op_code != MLX5_CQE_REQ)) {
505 		if (!check_cqe_seen(cqe)) {
506 			DRV_LOG(ERR, "unexpected CQE opcode %u (0x%02x)",
507 				op_code, op_code);
508 			rte_hexdump(stderr, "MLX5 CQE:",
509 				    (const void *)((uintptr_t)cqe),
510 				    sizeof(*cqe));
511 		}
512 		return 1;
513 	}
514 #endif /* NDEBUG */
515 	return 0;
516 }
517 
518 /**
519  * Return the address of the WQE.
520  *
521  * @param txq
522  *   Pointer to TX queue structure.
523  * @param  wqe_ci
524  *   WQE consumer index.
525  *
526  * @return
527  *   WQE address.
528  */
529 static inline uintptr_t *
530 tx_mlx5_wqe(struct mlx5_txq_data *txq, uint16_t ci)
531 {
532 	ci &= ((1 << txq->wqe_n) - 1);
533 	return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
534 }
535 
536 /**
537  * Manage TX completions.
538  *
539  * When sending a burst, mlx5_tx_burst() posts several WRs.
540  *
541  * @param txq
542  *   Pointer to TX queue structure.
543  */
544 static __rte_always_inline void
545 mlx5_tx_complete(struct mlx5_txq_data *txq)
546 {
547 	const uint16_t elts_n = 1 << txq->elts_n;
548 	const uint16_t elts_m = elts_n - 1;
549 	const unsigned int cqe_n = 1 << txq->cqe_n;
550 	const unsigned int cqe_cnt = cqe_n - 1;
551 	uint16_t elts_free = txq->elts_tail;
552 	uint16_t elts_tail;
553 	uint16_t cq_ci = txq->cq_ci;
554 	volatile struct mlx5_cqe *cqe = NULL;
555 	volatile struct mlx5_wqe_ctrl *ctrl;
556 	struct rte_mbuf *m, *free[elts_n];
557 	struct rte_mempool *pool = NULL;
558 	unsigned int blk_n = 0;
559 
560 	cqe = &(*txq->cqes)[cq_ci & cqe_cnt];
561 	if (unlikely(check_cqe(cqe, cqe_n, cq_ci)))
562 		return;
563 #ifndef NDEBUG
564 	if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
565 	    (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
566 		if (!check_cqe_seen(cqe)) {
567 			DRV_LOG(ERR, "unexpected error CQE, Tx stopped");
568 			rte_hexdump(stderr, "MLX5 TXQ:",
569 				    (const void *)((uintptr_t)txq->wqes),
570 				    ((1 << txq->wqe_n) *
571 				     MLX5_WQE_SIZE));
572 		}
573 		return;
574 	}
575 #endif /* NDEBUG */
576 	++cq_ci;
577 	rte_cio_rmb();
578 	txq->wqe_pi = rte_be_to_cpu_16(cqe->wqe_counter);
579 	ctrl = (volatile struct mlx5_wqe_ctrl *)
580 		tx_mlx5_wqe(txq, txq->wqe_pi);
581 	elts_tail = ctrl->ctrl3;
582 	assert((elts_tail & elts_m) < (1 << txq->wqe_n));
583 	/* Free buffers. */
584 	while (elts_free != elts_tail) {
585 		m = rte_pktmbuf_prefree_seg((*txq->elts)[elts_free++ & elts_m]);
586 		if (likely(m != NULL)) {
587 			if (likely(m->pool == pool)) {
588 				free[blk_n++] = m;
589 			} else {
590 				if (likely(pool != NULL))
591 					rte_mempool_put_bulk(pool,
592 							     (void *)free,
593 							     blk_n);
594 				free[0] = m;
595 				pool = m->pool;
596 				blk_n = 1;
597 			}
598 		}
599 	}
600 	if (blk_n)
601 		rte_mempool_put_bulk(pool, (void *)free, blk_n);
602 #ifndef NDEBUG
603 	elts_free = txq->elts_tail;
604 	/* Poisoning. */
605 	while (elts_free != elts_tail) {
606 		memset(&(*txq->elts)[elts_free & elts_m],
607 		       0x66,
608 		       sizeof((*txq->elts)[elts_free & elts_m]));
609 		++elts_free;
610 	}
611 #endif
612 	txq->cq_ci = cq_ci;
613 	txq->elts_tail = elts_tail;
614 	/* Update the consumer index. */
615 	rte_compiler_barrier();
616 	*txq->cq_db = rte_cpu_to_be_32(cq_ci);
617 }
618 
619 /**
620  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which the
621  * cloned mbuf is allocated is returned instead.
622  *
623  * @param buf
624  *   Pointer to mbuf.
625  *
626  * @return
627  *   Memory pool where data is located for given mbuf.
628  */
629 static inline struct rte_mempool *
630 mlx5_mb2mp(struct rte_mbuf *buf)
631 {
632 	if (unlikely(RTE_MBUF_CLONED(buf)))
633 		return rte_mbuf_from_indirect(buf)->pool;
634 	return buf->pool;
635 }
636 
637 /**
638  * Query LKey from a packet buffer for Rx. No need to flush local caches for Rx
639  * as mempool is pre-configured and static.
640  *
641  * @param rxq
642  *   Pointer to Rx queue structure.
643  * @param addr
644  *   Address to search.
645  *
646  * @return
647  *   Searched LKey on success, UINT32_MAX on no match.
648  */
649 static __rte_always_inline uint32_t
650 mlx5_rx_addr2mr(struct mlx5_rxq_data *rxq, uintptr_t addr)
651 {
652 	struct mlx5_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
653 	uint32_t lkey;
654 
655 	/* Linear search on MR cache array. */
656 	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
657 				    MLX5_MR_CACHE_N, addr);
658 	if (likely(lkey != UINT32_MAX))
659 		return lkey;
660 	/* Take slower bottom-half (Binary Search) on miss. */
661 	return mlx5_rx_addr2mr_bh(rxq, addr);
662 }
663 
664 #define mlx5_rx_mb2mr(rxq, mb) mlx5_rx_addr2mr(rxq, (uintptr_t)((mb)->buf_addr))
665 
666 /**
667  * Query LKey from a packet buffer for Tx. If not found, add the mempool.
668  *
669  * @param txq
670  *   Pointer to Tx queue structure.
671  * @param addr
672  *   Address to search.
673  *
674  * @return
675  *   Searched LKey on success, UINT32_MAX on no match.
676  */
677 static __rte_always_inline uint32_t
678 mlx5_tx_mb2mr(struct mlx5_txq_data *txq, struct rte_mbuf *mb)
679 {
680 	struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
681 	uintptr_t addr = (uintptr_t)mb->buf_addr;
682 	uint32_t lkey;
683 
684 	/* Check generation bit to see if there's any change on existing MRs. */
685 	if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen))
686 		mlx5_mr_flush_local_cache(mr_ctrl);
687 	/* Linear search on MR cache array. */
688 	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
689 				    MLX5_MR_CACHE_N, addr);
690 	if (likely(lkey != UINT32_MAX))
691 		return lkey;
692 	/* Take slower bottom-half on miss. */
693 	return mlx5_tx_mb2mr_bh(txq, mb);
694 }
695 
696 /**
697  * Ring TX queue doorbell and flush the update if requested.
698  *
699  * @param txq
700  *   Pointer to TX queue structure.
701  * @param wqe
702  *   Pointer to the last WQE posted in the NIC.
703  * @param cond
704  *   Request for write memory barrier after BlueFlame update.
705  */
706 static __rte_always_inline void
707 mlx5_tx_dbrec_cond_wmb(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe,
708 		       int cond)
709 {
710 	uint64_t *dst = MLX5_TX_BFREG(txq);
711 	volatile uint64_t *src = ((volatile uint64_t *)wqe);
712 
713 	rte_cio_wmb();
714 	*txq->qp_db = rte_cpu_to_be_32(txq->wqe_ci);
715 	/* Ensure ordering between DB record and BF copy. */
716 	rte_wmb();
717 	mlx5_uar_write64_relaxed(*src, dst, txq->uar_lock);
718 	if (cond)
719 		rte_wmb();
720 }
721 
722 /**
723  * Ring TX queue doorbell and flush the update by write memory barrier.
724  *
725  * @param txq
726  *   Pointer to TX queue structure.
727  * @param wqe
728  *   Pointer to the last WQE posted in the NIC.
729  */
730 static __rte_always_inline void
731 mlx5_tx_dbrec(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe)
732 {
733 	mlx5_tx_dbrec_cond_wmb(txq, wqe, 1);
734 }
735 
736 /**
737  * Convert mbuf to Verb SWP.
738  *
739  * @param txq_data
740  *   Pointer to the Tx queue.
741  * @param buf
742  *   Pointer to the mbuf.
743  * @param offsets
744  *   Pointer to the SWP header offsets.
745  * @param swp_types
746  *   Pointer to the SWP header types.
747  */
748 static __rte_always_inline void
749 txq_mbuf_to_swp(struct mlx5_txq_data *txq, struct rte_mbuf *buf,
750 		uint8_t *offsets, uint8_t *swp_types)
751 {
752 	const uint64_t vlan = buf->ol_flags & PKT_TX_VLAN_PKT;
753 	const uint64_t tunnel = buf->ol_flags & PKT_TX_TUNNEL_MASK;
754 	const uint64_t tso = buf->ol_flags & PKT_TX_TCP_SEG;
755 	const uint64_t csum_flags = buf->ol_flags & PKT_TX_L4_MASK;
756 	const uint64_t inner_ip =
757 		buf->ol_flags & (PKT_TX_IPV4 | PKT_TX_IPV6);
758 	const uint64_t ol_flags_mask = PKT_TX_L4_MASK | PKT_TX_IPV6 |
759 				       PKT_TX_OUTER_IPV6;
760 	uint16_t idx;
761 	uint16_t off;
762 
763 	if (likely(!txq->swp_en || (tunnel != PKT_TX_TUNNEL_UDP &&
764 				    tunnel != PKT_TX_TUNNEL_IP)))
765 		return;
766 	/*
767 	 * The index should have:
768 	 * bit[0:1] = PKT_TX_L4_MASK
769 	 * bit[4] = PKT_TX_IPV6
770 	 * bit[8] = PKT_TX_OUTER_IPV6
771 	 * bit[9] = PKT_TX_OUTER_UDP
772 	 */
773 	idx = (buf->ol_flags & ol_flags_mask) >> 52;
774 	if (tunnel == PKT_TX_TUNNEL_UDP)
775 		idx |= 1 << 9;
776 	*swp_types = mlx5_swp_types_table[idx];
777 	/*
778 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
779 	 * complements HW parser. SW parser starts to engage only if HW parser
780 	 * can't reach a header. For the older devices, HW parser will not kick
781 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
782 	 * should be set regardless of HW offload.
783 	 */
784 	off = buf->outer_l2_len + (vlan ? sizeof(struct vlan_hdr) : 0);
785 	offsets[1] = off >> 1; /* Outer L3 offset. */
786 	off += buf->outer_l3_len;
787 	if (tunnel == PKT_TX_TUNNEL_UDP)
788 		offsets[0] = off >> 1; /* Outer L4 offset. */
789 	if (inner_ip) {
790 		off += buf->l2_len;
791 		offsets[3] = off >> 1; /* Inner L3 offset. */
792 		if (csum_flags == PKT_TX_TCP_CKSUM || tso ||
793 		    csum_flags == PKT_TX_UDP_CKSUM) {
794 			off += buf->l3_len;
795 			offsets[2] = off >> 1; /* Inner L4 offset. */
796 		}
797 	}
798 }
799 
800 /**
801  * Convert the Checksum offloads to Verbs.
802  *
803  * @param buf
804  *   Pointer to the mbuf.
805  *
806  * @return
807  *   Converted checksum flags.
808  */
809 static __rte_always_inline uint8_t
810 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
811 {
812 	uint32_t idx;
813 	uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK);
814 	const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK |
815 				       PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM;
816 
817 	/*
818 	 * The index should have:
819 	 * bit[0] = PKT_TX_TCP_SEG
820 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
821 	 * bit[4] = PKT_TX_IP_CKSUM
822 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
823 	 * bit[9] = tunnel
824 	 */
825 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
826 	return mlx5_cksum_table[idx];
827 }
828 
829 /**
830  * Count the number of contiguous single segment packets.
831  *
832  * @param pkts
833  *   Pointer to array of packets.
834  * @param pkts_n
835  *   Number of packets.
836  *
837  * @return
838  *   Number of contiguous single segment packets.
839  */
840 static __rte_always_inline unsigned int
841 txq_count_contig_single_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
842 {
843 	unsigned int pos;
844 
845 	if (!pkts_n)
846 		return 0;
847 	/* Count the number of contiguous single segment packets. */
848 	for (pos = 0; pos < pkts_n; ++pos)
849 		if (NB_SEGS(pkts[pos]) > 1)
850 			break;
851 	return pos;
852 }
853 
854 /**
855  * Count the number of contiguous multi-segment packets.
856  *
857  * @param pkts
858  *   Pointer to array of packets.
859  * @param pkts_n
860  *   Number of packets.
861  *
862  * @return
863  *   Number of contiguous multi-segment packets.
864  */
865 static __rte_always_inline unsigned int
866 txq_count_contig_multi_seg(struct rte_mbuf **pkts, uint16_t pkts_n)
867 {
868 	unsigned int pos;
869 
870 	if (!pkts_n)
871 		return 0;
872 	/* Count the number of contiguous multi-segment packets. */
873 	for (pos = 0; pos < pkts_n; ++pos)
874 		if (NB_SEGS(pkts[pos]) == 1)
875 			break;
876 	return pos;
877 }
878 
879 #endif /* RTE_PMD_MLX5_RXTX_H_ */
880