xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.h (revision 1cde1b9a9b4dbf31cb5e5ccdfc5da3cb079f43a2)
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 #include <rte_malloc.h>
33 
34 #include "mlx5_utils.h"
35 #include "mlx5.h"
36 #include "mlx5_mr.h"
37 #include "mlx5_autoconf.h"
38 #include "mlx5_defs.h"
39 #include "mlx5_prm.h"
40 #include "mlx5_glue.h"
41 
42 /* Support tunnel matching. */
43 #define MLX5_FLOW_TUNNEL 8
44 
45 struct mlx5_rxq_stats {
46 #ifdef MLX5_PMD_SOFT_COUNTERS
47 	uint64_t ipackets; /**< Total of successfully received packets. */
48 	uint64_t ibytes; /**< Total of successfully received bytes. */
49 #endif
50 	uint64_t idropped; /**< Total of packets dropped when RX ring full. */
51 	uint64_t rx_nombuf; /**< Total of RX mbuf allocation failures. */
52 };
53 
54 struct mlx5_txq_stats {
55 #ifdef MLX5_PMD_SOFT_COUNTERS
56 	uint64_t opackets; /**< Total of successfully sent packets. */
57 	uint64_t obytes; /**< Total of successfully sent bytes. */
58 #endif
59 	uint64_t oerrors; /**< Total number of failed transmitted packets. */
60 };
61 
62 struct mlx5_priv;
63 
64 /* Compressed CQE context. */
65 struct rxq_zip {
66 	uint16_t ai; /* Array index. */
67 	uint16_t ca; /* Current array index. */
68 	uint16_t na; /* Next array index. */
69 	uint16_t cq_ci; /* The next CQE. */
70 	uint32_t cqe_cnt; /* Number of CQEs. */
71 };
72 
73 /* Multi-Packet RQ buffer header. */
74 struct mlx5_mprq_buf {
75 	struct rte_mempool *mp;
76 	rte_atomic16_t refcnt; /* Atomically accessed refcnt. */
77 	uint8_t pad[RTE_PKTMBUF_HEADROOM]; /* Headroom for the first packet. */
78 	struct rte_mbuf_ext_shared_info shinfos[];
79 	/*
80 	 * Shared information per stride.
81 	 * More memory will be allocated for the first stride head-room and for
82 	 * the strides data.
83 	 */
84 } __rte_cache_aligned;
85 
86 /* Get pointer to the first stride. */
87 #define mlx5_mprq_buf_addr(ptr, strd_n) (RTE_PTR_ADD((ptr), \
88 				sizeof(struct mlx5_mprq_buf) + \
89 				(strd_n) * \
90 				sizeof(struct rte_mbuf_ext_shared_info) + \
91 				RTE_PKTMBUF_HEADROOM))
92 
93 #define MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES 6
94 #define MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES 9
95 
96 enum mlx5_rxq_err_state {
97 	MLX5_RXQ_ERR_STATE_NO_ERROR = 0,
98 	MLX5_RXQ_ERR_STATE_NEED_RESET,
99 	MLX5_RXQ_ERR_STATE_NEED_READY,
100 };
101 
102 /* RX queue descriptor. */
103 struct mlx5_rxq_data {
104 	unsigned int csum:1; /* Enable checksum offloading. */
105 	unsigned int hw_timestamp:1; /* Enable HW timestamp. */
106 	unsigned int vlan_strip:1; /* Enable VLAN stripping. */
107 	unsigned int crc_present:1; /* CRC must be subtracted. */
108 	unsigned int sges_n:3; /* Log 2 of SGEs (max buffers per packet). */
109 	unsigned int cqe_n:4; /* Log 2 of CQ elements. */
110 	unsigned int elts_n:4; /* Log 2 of Mbufs. */
111 	unsigned int rss_hash:1; /* RSS hash result is enabled. */
112 	unsigned int mark:1; /* Marked flow available on the queue. */
113 	unsigned int strd_num_n:5; /* Log 2 of the number of stride. */
114 	unsigned int strd_sz_n:4; /* Log 2 of stride size. */
115 	unsigned int strd_shift_en:1; /* Enable 2bytes shift on a stride. */
116 	unsigned int err_state:2; /* enum mlx5_rxq_err_state. */
117 	unsigned int strd_headroom_en:1; /* Enable mbuf headroom in MPRQ. */
118 	unsigned int lro:1; /* Enable LRO. */
119 	unsigned int :1; /* Remaining bits. */
120 	volatile uint32_t *rq_db;
121 	volatile uint32_t *cq_db;
122 	uint16_t port_id;
123 	uint32_t rq_ci;
124 	uint16_t consumed_strd; /* Number of consumed strides in WQE. */
125 	uint32_t rq_pi;
126 	uint32_t cq_ci;
127 	uint16_t rq_repl_thresh; /* Threshold for buffer replenishment. */
128 	union {
129 		struct rxq_zip zip; /* Compressed context. */
130 		uint16_t decompressed;
131 		/* Number of ready mbufs decompressed from the CQ. */
132 	};
133 	struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
134 	uint16_t mprq_max_memcpy_len; /* Maximum size of packet to memcpy. */
135 	volatile void *wqes;
136 	volatile struct mlx5_cqe(*cqes)[];
137 	RTE_STD_C11
138 	union  {
139 		struct rte_mbuf *(*elts)[];
140 		struct mlx5_mprq_buf *(*mprq_bufs)[];
141 	};
142 	struct rte_mempool *mp;
143 	struct rte_mempool *mprq_mp; /* Mempool for Multi-Packet RQ. */
144 	struct mlx5_mprq_buf *mprq_repl; /* Stashed mbuf for replenish. */
145 	uint16_t idx; /* Queue index. */
146 	struct mlx5_rxq_stats stats;
147 	uint64_t mbuf_initializer; /* Default rearm_data for vectorized Rx. */
148 	struct rte_mbuf fake_mbuf; /* elts padding for vectorized Rx. */
149 	void *cq_uar; /* CQ user access region. */
150 	uint32_t cqn; /* CQ number. */
151 	uint8_t cq_arm_sn; /* CQ arm seq number. */
152 #ifndef RTE_ARCH_64
153 	rte_spinlock_t *uar_lock_cq;
154 	/* CQ (UAR) access lock required for 32bit implementations */
155 #endif
156 	uint32_t tunnel; /* Tunnel information. */
157 } __rte_cache_aligned;
158 
159 enum mlx5_rxq_obj_type {
160 	MLX5_RXQ_OBJ_TYPE_IBV,		/* mlx5_rxq_obj with ibv_wq. */
161 	MLX5_RXQ_OBJ_TYPE_DEVX_RQ,	/* mlx5_rxq_obj with mlx5_devx_rq. */
162 };
163 
164 /* Verbs/DevX Rx queue elements. */
165 struct mlx5_rxq_obj {
166 	LIST_ENTRY(mlx5_rxq_obj) next; /* Pointer to the next element. */
167 	rte_atomic32_t refcnt; /* Reference counter. */
168 	struct mlx5_rxq_ctrl *rxq_ctrl; /* Back pointer to parent. */
169 	struct ibv_cq *cq; /* Completion Queue. */
170 	enum mlx5_rxq_obj_type type;
171 	RTE_STD_C11
172 	union {
173 		struct ibv_wq *wq; /* Work Queue. */
174 		struct mlx5_devx_obj *rq; /* DevX object for Rx Queue. */
175 	};
176 	struct ibv_comp_channel *channel;
177 };
178 
179 /* RX queue control descriptor. */
180 struct mlx5_rxq_ctrl {
181 	struct mlx5_rxq_data rxq; /* Data path structure. */
182 	LIST_ENTRY(mlx5_rxq_ctrl) next; /* Pointer to the next element. */
183 	rte_atomic32_t refcnt; /* Reference counter. */
184 	struct mlx5_rxq_obj *obj; /* Verbs/DevX elements. */
185 	struct mlx5_priv *priv; /* Back pointer to private data. */
186 	unsigned int socket; /* CPU socket ID for allocations. */
187 	unsigned int irq:1; /* Whether IRQ is enabled. */
188 	unsigned int dbr_umem_id_valid:1; /* dbr_umem_id holds a valid value. */
189 	uint32_t flow_mark_n; /* Number of Mark/Flag flows using this Queue. */
190 	uint32_t flow_tunnels_n[MLX5_FLOW_TUNNEL]; /* Tunnels counters. */
191 	uint32_t wqn; /* WQ number. */
192 	uint16_t dump_file_n; /* Number of dump files. */
193 	uint32_t dbr_umem_id; /* Storing door-bell information, */
194 	uint64_t dbr_offset;  /* needed when freeing door-bell. */
195 	struct mlx5dv_devx_umem *wq_umem; /* WQ buffer registration info. */
196 };
197 
198 enum mlx5_ind_tbl_type {
199 	MLX5_IND_TBL_TYPE_IBV,
200 	MLX5_IND_TBL_TYPE_DEVX,
201 };
202 
203 /* Indirection table. */
204 struct mlx5_ind_table_obj {
205 	LIST_ENTRY(mlx5_ind_table_obj) next; /* Pointer to the next element. */
206 	rte_atomic32_t refcnt; /* Reference counter. */
207 	enum mlx5_ind_tbl_type type;
208 	RTE_STD_C11
209 	union {
210 		struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
211 		struct mlx5_devx_obj *rqt; /* DevX RQT object. */
212 	};
213 	uint32_t queues_n; /**< Number of queues in the list. */
214 	uint16_t queues[]; /**< Queue list. */
215 };
216 
217 /* Hash Rx queue. */
218 struct mlx5_hrxq {
219 	LIST_ENTRY(mlx5_hrxq) next; /* Pointer to the next element. */
220 	rte_atomic32_t refcnt; /* Reference counter. */
221 	struct mlx5_ind_table_obj *ind_table; /* Indirection table. */
222 	RTE_STD_C11
223 	union {
224 		struct ibv_qp *qp; /* Verbs queue pair. */
225 		struct mlx5_devx_obj *tir; /* DevX TIR object. */
226 	};
227 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
228 	void *action; /* DV QP action pointer. */
229 #endif
230 	uint64_t hash_fields; /* Verbs Hash fields. */
231 	uint32_t rss_key_len; /* Hash key length in bytes. */
232 	uint8_t rss_key[]; /* Hash key. */
233 };
234 
235 /* TX queue send local data. */
236 __extension__
237 struct mlx5_txq_local {
238 	struct mlx5_wqe *wqe_last; /* last sent WQE pointer. */
239 	struct rte_mbuf *mbuf; /* first mbuf to process. */
240 	uint16_t pkts_copy; /* packets copied to elts. */
241 	uint16_t pkts_sent; /* packets sent. */
242 	uint16_t pkts_loop; /* packets sent on loop entry. */
243 	uint16_t elts_free; /* available elts remain. */
244 	uint16_t wqe_free; /* available wqe remain. */
245 	uint16_t mbuf_off; /* data offset in current mbuf. */
246 	uint16_t mbuf_nseg; /* number of remaining mbuf. */
247 };
248 
249 /* TX queue descriptor. */
250 __extension__
251 struct mlx5_txq_data {
252 	uint16_t elts_head; /* Current counter in (*elts)[]. */
253 	uint16_t elts_tail; /* Counter of first element awaiting completion. */
254 	uint16_t elts_comp; /* elts index since last completion request. */
255 	uint16_t elts_s; /* Number of mbuf elements. */
256 	uint16_t elts_m; /* Mask for mbuf elements indices. */
257 	/* Fields related to elts mbuf storage. */
258 	uint16_t wqe_ci; /* Consumer index for work queue. */
259 	uint16_t wqe_pi; /* Producer index for work queue. */
260 	uint16_t wqe_s; /* Number of WQ elements. */
261 	uint16_t wqe_m; /* Mask Number for WQ elements. */
262 	uint16_t wqe_comp; /* WQE index since last completion request. */
263 	uint16_t wqe_thres; /* WQE threshold to request completion in CQ. */
264 	/* WQ related fields. */
265 	uint16_t cq_ci; /* Consumer index for completion queue. */
266 #ifndef NDEBUG
267 	uint16_t cq_pi; /* Counter of issued CQE "always" requests. */
268 #endif
269 	uint16_t cqe_s; /* Number of CQ elements. */
270 	uint16_t cqe_m; /* Mask for CQ indices. */
271 	/* CQ related fields. */
272 	uint16_t elts_n:4; /* elts[] length (in log2). */
273 	uint16_t cqe_n:4; /* Number of CQ elements (in log2). */
274 	uint16_t wqe_n:4; /* Number of WQ elements (in log2). */
275 	uint16_t tso_en:1; /* When set hardware TSO is enabled. */
276 	uint16_t tunnel_en:1;
277 	/* When set TX offload for tunneled packets are supported. */
278 	uint16_t swp_en:1; /* Whether SW parser is enabled. */
279 	uint16_t vlan_en:1; /* VLAN insertion in WQE is supported. */
280 	uint16_t inlen_send; /* Ordinary send data inline size. */
281 	uint16_t inlen_empw; /* eMPW max packet size to inline. */
282 	uint16_t inlen_mode; /* Minimal data length to inline. */
283 	uint32_t qp_num_8s; /* QP number shifted by 8. */
284 	uint64_t offloads; /* Offloads for Tx Queue. */
285 	struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
286 	struct mlx5_wqe *wqes; /* Work queue. */
287 	struct mlx5_wqe *wqes_end; /* Work queue array limit. */
288 	volatile struct mlx5_cqe *cqes; /* Completion queue. */
289 	volatile uint32_t *qp_db; /* Work queue doorbell. */
290 	volatile uint32_t *cq_db; /* Completion queue doorbell. */
291 	uint16_t port_id; /* Port ID of device. */
292 	uint16_t idx; /* Queue index. */
293 	struct mlx5_txq_stats stats; /* TX queue counters. */
294 #ifndef RTE_ARCH_64
295 	rte_spinlock_t *uar_lock;
296 	/* UAR access lock required for 32bit implementations */
297 #endif
298 	struct rte_mbuf *elts[0];
299 	/* Storage for queued packets, must be the last field. */
300 } __rte_cache_aligned;
301 
302 /* Verbs Rx queue elements. */
303 struct mlx5_txq_ibv {
304 	LIST_ENTRY(mlx5_txq_ibv) next; /* Pointer to the next element. */
305 	rte_atomic32_t refcnt; /* Reference counter. */
306 	struct mlx5_txq_ctrl *txq_ctrl; /* Pointer to the control queue. */
307 	struct ibv_cq *cq; /* Completion Queue. */
308 	struct ibv_qp *qp; /* Queue Pair. */
309 };
310 
311 /* TX queue control descriptor. */
312 struct mlx5_txq_ctrl {
313 	LIST_ENTRY(mlx5_txq_ctrl) next; /* Pointer to the next element. */
314 	rte_atomic32_t refcnt; /* Reference counter. */
315 	unsigned int socket; /* CPU socket ID for allocations. */
316 	unsigned int max_inline_data; /* Max inline data. */
317 	unsigned int max_tso_header; /* Max TSO header size. */
318 	struct mlx5_txq_ibv *ibv; /* Verbs queue object. */
319 	struct mlx5_priv *priv; /* Back pointer to private data. */
320 	off_t uar_mmap_offset; /* UAR mmap offset for non-primary process. */
321 	void *bf_reg; /* BlueFlame register from Verbs. */
322 	uint16_t dump_file_n; /* Number of dump files. */
323 	struct mlx5_txq_data txq; /* Data path structure. */
324 	/* Must be the last field in the structure, contains elts[]. */
325 };
326 
327 #define MLX5_TX_BFREG(txq) \
328 		(MLX5_PROC_PRIV((txq)->port_id)->uar_table[(txq)->idx])
329 
330 /* mlx5_rxq.c */
331 
332 extern uint8_t rss_hash_default_key[];
333 
334 int mlx5_check_mprq_support(struct rte_eth_dev *dev);
335 int mlx5_rxq_mprq_enabled(struct mlx5_rxq_data *rxq);
336 int mlx5_mprq_enabled(struct rte_eth_dev *dev);
337 int mlx5_mprq_free_mp(struct rte_eth_dev *dev);
338 int mlx5_mprq_alloc_mp(struct rte_eth_dev *dev);
339 int mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
340 			unsigned int socket, const struct rte_eth_rxconf *conf,
341 			struct rte_mempool *mp);
342 void mlx5_rx_queue_release(void *dpdk_rxq);
343 int mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev);
344 void mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev);
345 int mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
346 int mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
347 struct mlx5_rxq_obj *mlx5_rxq_obj_new(struct rte_eth_dev *dev, uint16_t idx,
348 				      enum mlx5_rxq_obj_type type);
349 int mlx5_rxq_obj_verify(struct rte_eth_dev *dev);
350 struct mlx5_rxq_ctrl *mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx,
351 				   uint16_t desc, unsigned int socket,
352 				   const struct rte_eth_rxconf *conf,
353 				   struct rte_mempool *mp);
354 struct mlx5_rxq_ctrl *mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx);
355 int mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx);
356 int mlx5_rxq_verify(struct rte_eth_dev *dev);
357 int rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl);
358 int mlx5_ind_table_obj_verify(struct rte_eth_dev *dev);
359 struct mlx5_hrxq *mlx5_hrxq_new(struct rte_eth_dev *dev,
360 				const uint8_t *rss_key, uint32_t rss_key_len,
361 				uint64_t hash_fields,
362 				const uint16_t *queues, uint32_t queues_n,
363 				int tunnel __rte_unused);
364 struct mlx5_hrxq *mlx5_hrxq_get(struct rte_eth_dev *dev,
365 				const uint8_t *rss_key, uint32_t rss_key_len,
366 				uint64_t hash_fields,
367 				const uint16_t *queues, uint32_t queues_n);
368 int mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hxrq);
369 int mlx5_hrxq_verify(struct rte_eth_dev *dev);
370 struct mlx5_hrxq *mlx5_hrxq_drop_new(struct rte_eth_dev *dev);
371 void mlx5_hrxq_drop_release(struct rte_eth_dev *dev);
372 uint64_t mlx5_get_rx_port_offloads(void);
373 uint64_t mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev);
374 
375 /* mlx5_txq.c */
376 
377 int mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
378 			unsigned int socket, const struct rte_eth_txconf *conf);
379 void mlx5_tx_queue_release(void *dpdk_txq);
380 int mlx5_tx_uar_init_secondary(struct rte_eth_dev *dev, int fd);
381 struct mlx5_txq_ibv *mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx);
382 struct mlx5_txq_ibv *mlx5_txq_ibv_get(struct rte_eth_dev *dev, uint16_t idx);
383 int mlx5_txq_ibv_release(struct mlx5_txq_ibv *txq_ibv);
384 int mlx5_txq_ibv_verify(struct rte_eth_dev *dev);
385 struct mlx5_txq_ctrl *mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx,
386 				   uint16_t desc, unsigned int socket,
387 				   const struct rte_eth_txconf *conf);
388 struct mlx5_txq_ctrl *mlx5_txq_get(struct rte_eth_dev *dev, uint16_t idx);
389 int mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx);
390 int mlx5_txq_releasable(struct rte_eth_dev *dev, uint16_t idx);
391 int mlx5_txq_verify(struct rte_eth_dev *dev);
392 void txq_alloc_elts(struct mlx5_txq_ctrl *txq_ctrl);
393 uint64_t mlx5_get_tx_port_offloads(struct rte_eth_dev *dev);
394 
395 /* mlx5_rxtx.c */
396 
397 extern uint32_t mlx5_ptype_table[];
398 extern uint8_t mlx5_cksum_table[];
399 extern uint8_t mlx5_swp_types_table[];
400 
401 void mlx5_set_ptype_table(void);
402 void mlx5_set_cksum_table(void);
403 void mlx5_set_swp_types_table(void);
404 __rte_noinline int mlx5_tx_error_cqe_handle
405 				(struct mlx5_txq_data *restrict txq,
406 				 volatile struct mlx5_err_cqe *err_cqe);
407 uint16_t mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
408 void mlx5_rxq_initialize(struct mlx5_rxq_data *rxq);
409 __rte_noinline int mlx5_rx_err_handle(struct mlx5_rxq_data *rxq,
410 				      uint8_t mbuf_prepare);
411 void mlx5_mprq_buf_free_cb(void *addr, void *opaque);
412 void mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf);
413 uint16_t mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts,
414 			    uint16_t pkts_n);
415 uint16_t removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts,
416 			  uint16_t pkts_n);
417 uint16_t removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts,
418 			  uint16_t pkts_n);
419 int mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset);
420 int mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset);
421 uint32_t mlx5_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id);
422 void mlx5_dump_debug_information(const char *path, const char *title,
423 				 const void *buf, unsigned int len);
424 int mlx5_queue_state_modify_primary(struct rte_eth_dev *dev,
425 			const struct mlx5_mp_arg_queue_state_modify *sm);
426 
427 /* Vectorized version of mlx5_rxtx.c */
428 int mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq_data);
429 int mlx5_check_vec_rx_support(struct rte_eth_dev *dev);
430 uint16_t mlx5_rx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts,
431 			   uint16_t pkts_n);
432 
433 /* mlx5_mr.c */
434 
435 void mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl);
436 uint32_t mlx5_rx_addr2mr_bh(struct mlx5_rxq_data *rxq, uintptr_t addr);
437 uint32_t mlx5_tx_mb2mr_bh(struct mlx5_txq_data *txq, struct rte_mbuf *mb);
438 uint32_t mlx5_tx_update_ext_mp(struct mlx5_txq_data *txq, uintptr_t addr,
439 			       struct rte_mempool *mp);
440 int mlx5_dma_map(struct rte_pci_device *pdev, void *addr, uint64_t iova,
441 		 size_t len);
442 int mlx5_dma_unmap(struct rte_pci_device *pdev, void *addr, uint64_t iova,
443 		   size_t len);
444 
445 /**
446  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
447  * 64bit architectures.
448  *
449  * @param val
450  *   value to write in CPU endian format.
451  * @param addr
452  *   Address to write to.
453  * @param lock
454  *   Address of the lock to use for that UAR access.
455  */
456 static __rte_always_inline void
457 __mlx5_uar_write64_relaxed(uint64_t val, void *addr,
458 			   rte_spinlock_t *lock __rte_unused)
459 {
460 #ifdef RTE_ARCH_64
461 	*(uint64_t *)addr = val;
462 #else /* !RTE_ARCH_64 */
463 	rte_spinlock_lock(lock);
464 	*(uint32_t *)addr = val;
465 	rte_io_wmb();
466 	*((uint32_t *)addr + 1) = val >> 32;
467 	rte_spinlock_unlock(lock);
468 #endif
469 }
470 
471 /**
472  * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
473  * 64bit architectures while guaranteeing the order of execution with the
474  * code being executed.
475  *
476  * @param val
477  *   value to write in CPU endian format.
478  * @param addr
479  *   Address to write to.
480  * @param lock
481  *   Address of the lock to use for that UAR access.
482  */
483 static __rte_always_inline void
484 __mlx5_uar_write64(uint64_t val, void *addr, rte_spinlock_t *lock)
485 {
486 	rte_io_wmb();
487 	__mlx5_uar_write64_relaxed(val, addr, lock);
488 }
489 
490 /* Assist macros, used instead of directly calling the functions they wrap. */
491 #ifdef RTE_ARCH_64
492 #define mlx5_uar_write64_relaxed(val, dst, lock) \
493 		__mlx5_uar_write64_relaxed(val, dst, NULL)
494 #define mlx5_uar_write64(val, dst, lock) __mlx5_uar_write64(val, dst, NULL)
495 #else
496 #define mlx5_uar_write64_relaxed(val, dst, lock) \
497 		__mlx5_uar_write64_relaxed(val, dst, lock)
498 #define mlx5_uar_write64(val, dst, lock) __mlx5_uar_write64(val, dst, lock)
499 #endif
500 
501 /* CQE status. */
502 enum mlx5_cqe_status {
503 	MLX5_CQE_STATUS_SW_OWN = -1,
504 	MLX5_CQE_STATUS_HW_OWN = -2,
505 	MLX5_CQE_STATUS_ERR = -3,
506 };
507 
508 /**
509  * Check whether CQE is valid.
510  *
511  * @param cqe
512  *   Pointer to CQE.
513  * @param cqes_n
514  *   Size of completion queue.
515  * @param ci
516  *   Consumer index.
517  *
518  * @return
519  *   The CQE status.
520  */
521 static __rte_always_inline enum mlx5_cqe_status
522 check_cqe(volatile struct mlx5_cqe *cqe, const uint16_t cqes_n,
523 	  const uint16_t ci)
524 {
525 	const uint16_t idx = ci & cqes_n;
526 	const uint8_t op_own = cqe->op_own;
527 	const uint8_t op_owner = MLX5_CQE_OWNER(op_own);
528 	const uint8_t op_code = MLX5_CQE_OPCODE(op_own);
529 
530 	if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
531 		return MLX5_CQE_STATUS_HW_OWN;
532 	rte_cio_rmb();
533 	if (unlikely(op_code == MLX5_CQE_RESP_ERR ||
534 		     op_code == MLX5_CQE_REQ_ERR))
535 		return MLX5_CQE_STATUS_ERR;
536 	return MLX5_CQE_STATUS_SW_OWN;
537 }
538 
539 /**
540  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which the
541  * cloned mbuf is allocated is returned instead.
542  *
543  * @param buf
544  *   Pointer to mbuf.
545  *
546  * @return
547  *   Memory pool where data is located for given mbuf.
548  */
549 static inline struct rte_mempool *
550 mlx5_mb2mp(struct rte_mbuf *buf)
551 {
552 	if (unlikely(RTE_MBUF_CLONED(buf)))
553 		return rte_mbuf_from_indirect(buf)->pool;
554 	return buf->pool;
555 }
556 
557 /**
558  * Query LKey from a packet buffer for Rx. No need to flush local caches for Rx
559  * as mempool is pre-configured and static.
560  *
561  * @param rxq
562  *   Pointer to Rx queue structure.
563  * @param addr
564  *   Address to search.
565  *
566  * @return
567  *   Searched LKey on success, UINT32_MAX on no match.
568  */
569 static __rte_always_inline uint32_t
570 mlx5_rx_addr2mr(struct mlx5_rxq_data *rxq, uintptr_t addr)
571 {
572 	struct mlx5_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
573 	uint32_t lkey;
574 
575 	/* Linear search on MR cache array. */
576 	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
577 				    MLX5_MR_CACHE_N, addr);
578 	if (likely(lkey != UINT32_MAX))
579 		return lkey;
580 	/* Take slower bottom-half (Binary Search) on miss. */
581 	return mlx5_rx_addr2mr_bh(rxq, addr);
582 }
583 
584 #define mlx5_rx_mb2mr(rxq, mb) mlx5_rx_addr2mr(rxq, (uintptr_t)((mb)->buf_addr))
585 
586 /**
587  * Query LKey from a packet buffer for Tx. If not found, add the mempool.
588  *
589  * @param txq
590  *   Pointer to Tx queue structure.
591  * @param addr
592  *   Address to search.
593  *
594  * @return
595  *   Searched LKey on success, UINT32_MAX on no match.
596  */
597 static __rte_always_inline uint32_t
598 mlx5_tx_mb2mr(struct mlx5_txq_data *txq, struct rte_mbuf *mb)
599 {
600 	struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
601 	uintptr_t addr = (uintptr_t)mb->buf_addr;
602 	uint32_t lkey;
603 
604 	/* Check generation bit to see if there's any change on existing MRs. */
605 	if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen))
606 		mlx5_mr_flush_local_cache(mr_ctrl);
607 	/* Linear search on MR cache array. */
608 	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
609 				    MLX5_MR_CACHE_N, addr);
610 	if (likely(lkey != UINT32_MAX))
611 		return lkey;
612 	/* Take slower bottom-half on miss. */
613 	return mlx5_tx_mb2mr_bh(txq, mb);
614 }
615 
616 /**
617  * Ring TX queue doorbell and flush the update if requested.
618  *
619  * @param txq
620  *   Pointer to TX queue structure.
621  * @param wqe
622  *   Pointer to the last WQE posted in the NIC.
623  * @param cond
624  *   Request for write memory barrier after BlueFlame update.
625  */
626 static __rte_always_inline void
627 mlx5_tx_dbrec_cond_wmb(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe,
628 		       int cond)
629 {
630 	uint64_t *dst = MLX5_TX_BFREG(txq);
631 	volatile uint64_t *src = ((volatile uint64_t *)wqe);
632 
633 	rte_cio_wmb();
634 	*txq->qp_db = rte_cpu_to_be_32(txq->wqe_ci);
635 	/* Ensure ordering between DB record and BF copy. */
636 	rte_wmb();
637 	mlx5_uar_write64_relaxed(*src, dst, txq->uar_lock);
638 	if (cond)
639 		rte_wmb();
640 }
641 
642 /**
643  * Ring TX queue doorbell and flush the update by write memory barrier.
644  *
645  * @param txq
646  *   Pointer to TX queue structure.
647  * @param wqe
648  *   Pointer to the last WQE posted in the NIC.
649  */
650 static __rte_always_inline void
651 mlx5_tx_dbrec(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe)
652 {
653 	mlx5_tx_dbrec_cond_wmb(txq, wqe, 1);
654 }
655 
656 #endif /* RTE_PMD_MLX5_RXTX_H_ */
657