xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 16b8e92d49b87705fffbedd9c3241204d9a584a6)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 
13 #include <rte_malloc.h>
14 #include <ethdev_driver.h>
15 #include <rte_pci.h>
16 #include <rte_bus_pci.h>
17 #include <rte_common.h>
18 #include <rte_kvargs.h>
19 #include <rte_rwlock.h>
20 #include <rte_spinlock.h>
21 #include <rte_string_fns.h>
22 #include <rte_alarm.h>
23 #include <rte_cycles.h>
24 
25 #include <mlx5_glue.h>
26 #include <mlx5_devx_cmds.h>
27 #include <mlx5_common.h>
28 #include <mlx5_common_os.h>
29 #include <mlx5_common_mp.h>
30 #include <mlx5_malloc.h>
31 
32 #include "mlx5_defs.h"
33 #include "mlx5.h"
34 #include "mlx5_utils.h"
35 #include "mlx5_rxtx.h"
36 #include "mlx5_rx.h"
37 #include "mlx5_tx.h"
38 #include "mlx5_autoconf.h"
39 #include "mlx5_mr.h"
40 #include "mlx5_flow.h"
41 #include "mlx5_flow_os.h"
42 #include "rte_pmd_mlx5.h"
43 
44 #define MLX5_ETH_DRIVER_NAME mlx5_eth
45 
46 /* Device parameter to enable RX completion queue compression. */
47 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
48 
49 /* Device parameter to enable padding Rx packet to cacheline size. */
50 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
51 
52 /* Device parameter to enable Multi-Packet Rx queue. */
53 #define MLX5_RX_MPRQ_EN "mprq_en"
54 
55 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
56 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
57 
58 /* Device parameter to configure log 2 of the stride size for MPRQ. */
59 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
60 
61 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
62 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
63 
64 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
65 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
66 
67 /* Device parameter to configure inline send. Deprecated, ignored.*/
68 #define MLX5_TXQ_INLINE "txq_inline"
69 
70 /* Device parameter to limit packet size to inline with ordinary SEND. */
71 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
72 
73 /* Device parameter to configure minimal data size to inline. */
74 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
75 
76 /* Device parameter to limit packet size to inline with Enhanced MPW. */
77 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
78 
79 /*
80  * Device parameter to configure the number of TX queues threshold for
81  * enabling inline send.
82  */
83 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
84 
85 /*
86  * Device parameter to configure the number of TX queues threshold for
87  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
88  */
89 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
90 
91 /* Device parameter to enable multi-packet send WQEs. */
92 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
93 
94 /*
95  * Device parameter to force doorbell register mapping
96  * to non-cahed region eliminating the extra write memory barrier.
97  */
98 #define MLX5_TX_DB_NC "tx_db_nc"
99 
100 /*
101  * Device parameter to include 2 dsegs in the title WQEBB.
102  * Deprecated, ignored.
103  */
104 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
105 
106 /*
107  * Device parameter to limit the size of inlining packet.
108  * Deprecated, ignored.
109  */
110 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
111 
112 /*
113  * Device parameter to enable Tx scheduling on timestamps
114  * and specify the packet pacing granularity in nanoseconds.
115  */
116 #define MLX5_TX_PP "tx_pp"
117 
118 /*
119  * Device parameter to specify skew in nanoseconds on Tx datapath,
120  * it represents the time between SQ start WQE processing and
121  * appearing actual packet data on the wire.
122  */
123 #define MLX5_TX_SKEW "tx_skew"
124 
125 /*
126  * Device parameter to enable hardware Tx vector.
127  * Deprecated, ignored (no vectorized Tx routines anymore).
128  */
129 #define MLX5_TX_VEC_EN "tx_vec_en"
130 
131 /* Device parameter to enable hardware Rx vector. */
132 #define MLX5_RX_VEC_EN "rx_vec_en"
133 
134 /* Allow L3 VXLAN flow creation. */
135 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
136 
137 /* Activate DV E-Switch flow steering. */
138 #define MLX5_DV_ESW_EN "dv_esw_en"
139 
140 /* Activate DV flow steering. */
141 #define MLX5_DV_FLOW_EN "dv_flow_en"
142 
143 /* Enable extensive flow metadata support. */
144 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
145 
146 /* Device parameter to let the user manage the lacp traffic of bonded device */
147 #define MLX5_LACP_BY_USER "lacp_by_user"
148 
149 /* Activate Netlink support in VF mode. */
150 #define MLX5_VF_NL_EN "vf_nl_en"
151 
152 /* Enable extending memsegs when creating a MR. */
153 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
154 
155 /* Select port representors to instantiate. */
156 #define MLX5_REPRESENTOR "representor"
157 
158 /* Device parameter to configure the maximum number of dump files per queue. */
159 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
160 
161 /* Configure timeout of LRO session (in microseconds). */
162 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
163 
164 /*
165  * Device parameter to configure the total data buffer size for a single
166  * hairpin queue (logarithm value).
167  */
168 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
169 
170 /* Flow memory reclaim mode. */
171 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
172 
173 /* The default memory allocator used in PMD. */
174 #define MLX5_SYS_MEM_EN "sys_mem_en"
175 /* Decap will be used or not. */
176 #define MLX5_DECAP_EN "decap_en"
177 
178 /* Device parameter to configure allow or prevent duplicate rules pattern. */
179 #define MLX5_ALLOW_DUPLICATE_PATTERN "allow_duplicate_pattern"
180 
181 /* Shared memory between primary and secondary processes. */
182 struct mlx5_shared_data *mlx5_shared_data;
183 
184 /** Driver-specific log messages type. */
185 int mlx5_logtype;
186 
187 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
188 						LIST_HEAD_INITIALIZER();
189 static pthread_mutex_t mlx5_dev_ctx_list_mutex;
190 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
191 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
192 	[MLX5_IPOOL_DECAP_ENCAP] = {
193 		.size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
194 		.trunk_size = 64,
195 		.grow_trunk = 3,
196 		.grow_shift = 2,
197 		.need_lock = 1,
198 		.release_mem_en = 1,
199 		.malloc = mlx5_malloc,
200 		.free = mlx5_free,
201 		.type = "mlx5_encap_decap_ipool",
202 	},
203 	[MLX5_IPOOL_PUSH_VLAN] = {
204 		.size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
205 		.trunk_size = 64,
206 		.grow_trunk = 3,
207 		.grow_shift = 2,
208 		.need_lock = 1,
209 		.release_mem_en = 1,
210 		.malloc = mlx5_malloc,
211 		.free = mlx5_free,
212 		.type = "mlx5_push_vlan_ipool",
213 	},
214 	[MLX5_IPOOL_TAG] = {
215 		.size = sizeof(struct mlx5_flow_dv_tag_resource),
216 		.trunk_size = 64,
217 		.grow_trunk = 3,
218 		.grow_shift = 2,
219 		.need_lock = 1,
220 		.release_mem_en = 0,
221 		.per_core_cache = (1 << 16),
222 		.malloc = mlx5_malloc,
223 		.free = mlx5_free,
224 		.type = "mlx5_tag_ipool",
225 	},
226 	[MLX5_IPOOL_PORT_ID] = {
227 		.size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
228 		.trunk_size = 64,
229 		.grow_trunk = 3,
230 		.grow_shift = 2,
231 		.need_lock = 1,
232 		.release_mem_en = 1,
233 		.malloc = mlx5_malloc,
234 		.free = mlx5_free,
235 		.type = "mlx5_port_id_ipool",
236 	},
237 	[MLX5_IPOOL_JUMP] = {
238 		.size = sizeof(struct mlx5_flow_tbl_data_entry),
239 		.trunk_size = 64,
240 		.grow_trunk = 3,
241 		.grow_shift = 2,
242 		.need_lock = 1,
243 		.release_mem_en = 1,
244 		.malloc = mlx5_malloc,
245 		.free = mlx5_free,
246 		.type = "mlx5_jump_ipool",
247 	},
248 	[MLX5_IPOOL_SAMPLE] = {
249 		.size = sizeof(struct mlx5_flow_dv_sample_resource),
250 		.trunk_size = 64,
251 		.grow_trunk = 3,
252 		.grow_shift = 2,
253 		.need_lock = 1,
254 		.release_mem_en = 1,
255 		.malloc = mlx5_malloc,
256 		.free = mlx5_free,
257 		.type = "mlx5_sample_ipool",
258 	},
259 	[MLX5_IPOOL_DEST_ARRAY] = {
260 		.size = sizeof(struct mlx5_flow_dv_dest_array_resource),
261 		.trunk_size = 64,
262 		.grow_trunk = 3,
263 		.grow_shift = 2,
264 		.need_lock = 1,
265 		.release_mem_en = 1,
266 		.malloc = mlx5_malloc,
267 		.free = mlx5_free,
268 		.type = "mlx5_dest_array_ipool",
269 	},
270 	[MLX5_IPOOL_TUNNEL_ID] = {
271 		.size = sizeof(struct mlx5_flow_tunnel),
272 		.trunk_size = MLX5_MAX_TUNNELS,
273 		.need_lock = 1,
274 		.release_mem_en = 1,
275 		.type = "mlx5_tunnel_offload",
276 	},
277 	[MLX5_IPOOL_TNL_TBL_ID] = {
278 		.size = 0,
279 		.need_lock = 1,
280 		.type = "mlx5_flow_tnl_tbl_ipool",
281 	},
282 #endif
283 	[MLX5_IPOOL_MTR] = {
284 		/**
285 		 * The ipool index should grow continually from small to big,
286 		 * for meter idx, so not set grow_trunk to avoid meter index
287 		 * not jump continually.
288 		 */
289 		.size = sizeof(struct mlx5_legacy_flow_meter),
290 		.trunk_size = 64,
291 		.need_lock = 1,
292 		.release_mem_en = 1,
293 		.malloc = mlx5_malloc,
294 		.free = mlx5_free,
295 		.type = "mlx5_meter_ipool",
296 	},
297 	[MLX5_IPOOL_MCP] = {
298 		.size = sizeof(struct mlx5_flow_mreg_copy_resource),
299 		.trunk_size = 64,
300 		.grow_trunk = 3,
301 		.grow_shift = 2,
302 		.need_lock = 1,
303 		.release_mem_en = 1,
304 		.malloc = mlx5_malloc,
305 		.free = mlx5_free,
306 		.type = "mlx5_mcp_ipool",
307 	},
308 	[MLX5_IPOOL_HRXQ] = {
309 		.size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
310 		.trunk_size = 64,
311 		.grow_trunk = 3,
312 		.grow_shift = 2,
313 		.need_lock = 1,
314 		.release_mem_en = 1,
315 		.malloc = mlx5_malloc,
316 		.free = mlx5_free,
317 		.type = "mlx5_hrxq_ipool",
318 	},
319 	[MLX5_IPOOL_MLX5_FLOW] = {
320 		/*
321 		 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
322 		 * It set in run time according to PCI function configuration.
323 		 */
324 		.size = 0,
325 		.trunk_size = 64,
326 		.grow_trunk = 3,
327 		.grow_shift = 2,
328 		.need_lock = 1,
329 		.release_mem_en = 0,
330 		.per_core_cache = 1 << 19,
331 		.malloc = mlx5_malloc,
332 		.free = mlx5_free,
333 		.type = "mlx5_flow_handle_ipool",
334 	},
335 	[MLX5_IPOOL_RTE_FLOW] = {
336 		.size = sizeof(struct rte_flow),
337 		.trunk_size = 4096,
338 		.need_lock = 1,
339 		.release_mem_en = 1,
340 		.malloc = mlx5_malloc,
341 		.free = mlx5_free,
342 		.type = "rte_flow_ipool",
343 	},
344 	[MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
345 		.size = 0,
346 		.need_lock = 1,
347 		.type = "mlx5_flow_rss_id_ipool",
348 	},
349 	[MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
350 		.size = sizeof(struct mlx5_shared_action_rss),
351 		.trunk_size = 64,
352 		.grow_trunk = 3,
353 		.grow_shift = 2,
354 		.need_lock = 1,
355 		.release_mem_en = 1,
356 		.malloc = mlx5_malloc,
357 		.free = mlx5_free,
358 		.type = "mlx5_shared_action_rss",
359 	},
360 	[MLX5_IPOOL_MTR_POLICY] = {
361 		/**
362 		 * The ipool index should grow continually from small to big,
363 		 * for policy idx, so not set grow_trunk to avoid policy index
364 		 * not jump continually.
365 		 */
366 		.size = sizeof(struct mlx5_flow_meter_sub_policy),
367 		.trunk_size = 64,
368 		.need_lock = 1,
369 		.release_mem_en = 1,
370 		.malloc = mlx5_malloc,
371 		.free = mlx5_free,
372 		.type = "mlx5_meter_policy_ipool",
373 	},
374 };
375 
376 
377 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
378 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
379 
380 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 1024
381 
382 /**
383  * Decide whether representor ID is a HPF(host PF) port on BF2.
384  *
385  * @param dev
386  *   Pointer to Ethernet device structure.
387  *
388  * @return
389  *   Non-zero if HPF, otherwise 0.
390  */
391 bool
392 mlx5_is_hpf(struct rte_eth_dev *dev)
393 {
394 	struct mlx5_priv *priv = dev->data->dev_private;
395 	uint16_t repr = MLX5_REPRESENTOR_REPR(priv->representor_id);
396 	int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
397 
398 	return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_VF &&
399 	       MLX5_REPRESENTOR_REPR(-1) == repr;
400 }
401 
402 /**
403  * Decide whether representor ID is a SF port representor.
404  *
405  * @param dev
406  *   Pointer to Ethernet device structure.
407  *
408  * @return
409  *   Non-zero if HPF, otherwise 0.
410  */
411 bool
412 mlx5_is_sf_repr(struct rte_eth_dev *dev)
413 {
414 	struct mlx5_priv *priv = dev->data->dev_private;
415 	int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
416 
417 	return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_SF;
418 }
419 
420 /**
421  * Initialize the ASO aging management structure.
422  *
423  * @param[in] sh
424  *   Pointer to mlx5_dev_ctx_shared object to free
425  *
426  * @return
427  *   0 on success, a negative errno value otherwise and rte_errno is set.
428  */
429 int
430 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
431 {
432 	int err;
433 
434 	if (sh->aso_age_mng)
435 		return 0;
436 	sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
437 				      RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
438 	if (!sh->aso_age_mng) {
439 		DRV_LOG(ERR, "aso_age_mng allocation was failed.");
440 		rte_errno = ENOMEM;
441 		return -ENOMEM;
442 	}
443 	err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_FLOW_HIT);
444 	if (err) {
445 		mlx5_free(sh->aso_age_mng);
446 		return -1;
447 	}
448 	rte_spinlock_init(&sh->aso_age_mng->resize_sl);
449 	rte_spinlock_init(&sh->aso_age_mng->free_sl);
450 	LIST_INIT(&sh->aso_age_mng->free);
451 	return 0;
452 }
453 
454 /**
455  * Close and release all the resources of the ASO aging management structure.
456  *
457  * @param[in] sh
458  *   Pointer to mlx5_dev_ctx_shared object to free.
459  */
460 static void
461 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
462 {
463 	int i, j;
464 
465 	mlx5_aso_flow_hit_queue_poll_stop(sh);
466 	mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_FLOW_HIT);
467 	if (sh->aso_age_mng->pools) {
468 		struct mlx5_aso_age_pool *pool;
469 
470 		for (i = 0; i < sh->aso_age_mng->next; ++i) {
471 			pool = sh->aso_age_mng->pools[i];
472 			claim_zero(mlx5_devx_cmd_destroy
473 						(pool->flow_hit_aso_obj));
474 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
475 				if (pool->actions[j].dr_action)
476 					claim_zero
477 					    (mlx5_flow_os_destroy_flow_action
478 					      (pool->actions[j].dr_action));
479 			mlx5_free(pool);
480 		}
481 		mlx5_free(sh->aso_age_mng->pools);
482 	}
483 	mlx5_free(sh->aso_age_mng);
484 }
485 
486 /**
487  * Initialize the shared aging list information per port.
488  *
489  * @param[in] sh
490  *   Pointer to mlx5_dev_ctx_shared object.
491  */
492 static void
493 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
494 {
495 	uint32_t i;
496 	struct mlx5_age_info *age_info;
497 
498 	for (i = 0; i < sh->max_port; i++) {
499 		age_info = &sh->port[i].age_info;
500 		age_info->flags = 0;
501 		TAILQ_INIT(&age_info->aged_counters);
502 		LIST_INIT(&age_info->aged_aso);
503 		rte_spinlock_init(&age_info->aged_sl);
504 		MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
505 	}
506 }
507 
508 /**
509  * Initialize the counters management structure.
510  *
511  * @param[in] sh
512  *   Pointer to mlx5_dev_ctx_shared object to free
513  */
514 static void
515 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
516 {
517 	int i;
518 
519 	memset(&sh->cmng, 0, sizeof(sh->cmng));
520 	TAILQ_INIT(&sh->cmng.flow_counters);
521 	sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
522 	sh->cmng.max_id = -1;
523 	sh->cmng.last_pool_idx = POOL_IDX_INVALID;
524 	rte_spinlock_init(&sh->cmng.pool_update_sl);
525 	for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
526 		TAILQ_INIT(&sh->cmng.counters[i]);
527 		rte_spinlock_init(&sh->cmng.csl[i]);
528 	}
529 }
530 
531 /**
532  * Destroy all the resources allocated for a counter memory management.
533  *
534  * @param[in] mng
535  *   Pointer to the memory management structure.
536  */
537 static void
538 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
539 {
540 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
541 
542 	LIST_REMOVE(mng, next);
543 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
544 	claim_zero(mlx5_os_umem_dereg(mng->umem));
545 	mlx5_free(mem);
546 }
547 
548 /**
549  * Close and release all the resources of the counters management.
550  *
551  * @param[in] sh
552  *   Pointer to mlx5_dev_ctx_shared object to free.
553  */
554 static void
555 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
556 {
557 	struct mlx5_counter_stats_mem_mng *mng;
558 	int i, j;
559 	int retries = 1024;
560 
561 	rte_errno = 0;
562 	while (--retries) {
563 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
564 		if (rte_errno != EINPROGRESS)
565 			break;
566 		rte_pause();
567 	}
568 
569 	if (sh->cmng.pools) {
570 		struct mlx5_flow_counter_pool *pool;
571 		uint16_t n_valid = sh->cmng.n_valid;
572 		bool fallback = sh->cmng.counter_fallback;
573 
574 		for (i = 0; i < n_valid; ++i) {
575 			pool = sh->cmng.pools[i];
576 			if (!fallback && pool->min_dcs)
577 				claim_zero(mlx5_devx_cmd_destroy
578 							       (pool->min_dcs));
579 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
580 				struct mlx5_flow_counter *cnt =
581 						MLX5_POOL_GET_CNT(pool, j);
582 
583 				if (cnt->action)
584 					claim_zero
585 					 (mlx5_flow_os_destroy_flow_action
586 					  (cnt->action));
587 				if (fallback && MLX5_POOL_GET_CNT
588 				    (pool, j)->dcs_when_free)
589 					claim_zero(mlx5_devx_cmd_destroy
590 						   (cnt->dcs_when_free));
591 			}
592 			mlx5_free(pool);
593 		}
594 		mlx5_free(sh->cmng.pools);
595 	}
596 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
597 	while (mng) {
598 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
599 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
600 	}
601 	memset(&sh->cmng, 0, sizeof(sh->cmng));
602 }
603 
604 /**
605  * Initialize the aso flow meters management structure.
606  *
607  * @param[in] sh
608  *   Pointer to mlx5_dev_ctx_shared object to free
609  */
610 int
611 mlx5_aso_flow_mtrs_mng_init(struct mlx5_dev_ctx_shared *sh)
612 {
613 	if (!sh->mtrmng) {
614 		sh->mtrmng = mlx5_malloc(MLX5_MEM_ZERO,
615 			sizeof(*sh->mtrmng),
616 			RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
617 		if (!sh->mtrmng) {
618 			DRV_LOG(ERR,
619 			"meter management allocation was failed.");
620 			rte_errno = ENOMEM;
621 			return -ENOMEM;
622 		}
623 		if (sh->meter_aso_en) {
624 			rte_spinlock_init(&sh->mtrmng->pools_mng.mtrsl);
625 			LIST_INIT(&sh->mtrmng->pools_mng.meters);
626 		}
627 		sh->mtrmng->def_policy_id = MLX5_INVALID_POLICY_ID;
628 	}
629 	return 0;
630 }
631 
632 /**
633  * Close and release all the resources of
634  * the ASO flow meter management structure.
635  *
636  * @param[in] sh
637  *   Pointer to mlx5_dev_ctx_shared object to free.
638  */
639 static void
640 mlx5_aso_flow_mtrs_mng_close(struct mlx5_dev_ctx_shared *sh)
641 {
642 	struct mlx5_aso_mtr_pool *mtr_pool;
643 	struct mlx5_flow_mtr_mng *mtrmng = sh->mtrmng;
644 	uint32_t idx;
645 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
646 	struct mlx5_aso_mtr *aso_mtr;
647 	int i;
648 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
649 
650 	if (sh->meter_aso_en) {
651 		mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_POLICER);
652 		idx = mtrmng->pools_mng.n_valid;
653 		while (idx--) {
654 			mtr_pool = mtrmng->pools_mng.pools[idx];
655 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
656 			for (i = 0; i < MLX5_ASO_MTRS_PER_POOL; i++) {
657 				aso_mtr = &mtr_pool->mtrs[i];
658 				if (aso_mtr->fm.meter_action)
659 					claim_zero
660 					(mlx5_glue->destroy_flow_action
661 					(aso_mtr->fm.meter_action));
662 			}
663 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
664 			claim_zero(mlx5_devx_cmd_destroy
665 						(mtr_pool->devx_obj));
666 			mtrmng->pools_mng.n_valid--;
667 			mlx5_free(mtr_pool);
668 		}
669 		mlx5_free(sh->mtrmng->pools_mng.pools);
670 	}
671 	mlx5_free(sh->mtrmng);
672 	sh->mtrmng = NULL;
673 }
674 
675 /* Send FLOW_AGED event if needed. */
676 void
677 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
678 {
679 	struct mlx5_age_info *age_info;
680 	uint32_t i;
681 
682 	for (i = 0; i < sh->max_port; i++) {
683 		age_info = &sh->port[i].age_info;
684 		if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
685 			continue;
686 		MLX5_AGE_UNSET(age_info, MLX5_AGE_EVENT_NEW);
687 		if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER)) {
688 			MLX5_AGE_UNSET(age_info, MLX5_AGE_TRIGGER);
689 			rte_eth_dev_callback_process
690 				(&rte_eth_devices[sh->port[i].devx_ih_port_id],
691 				RTE_ETH_EVENT_FLOW_AGED, NULL);
692 		}
693 	}
694 }
695 
696 /*
697  * Initialize the ASO connection tracking structure.
698  *
699  * @param[in] sh
700  *   Pointer to mlx5_dev_ctx_shared object.
701  *
702  * @return
703  *   0 on success, a negative errno value otherwise and rte_errno is set.
704  */
705 int
706 mlx5_flow_aso_ct_mng_init(struct mlx5_dev_ctx_shared *sh)
707 {
708 	int err;
709 
710 	if (sh->ct_mng)
711 		return 0;
712 	sh->ct_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->ct_mng),
713 				 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
714 	if (!sh->ct_mng) {
715 		DRV_LOG(ERR, "ASO CT management allocation failed.");
716 		rte_errno = ENOMEM;
717 		return -rte_errno;
718 	}
719 	err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
720 	if (err) {
721 		mlx5_free(sh->ct_mng);
722 		/* rte_errno should be extracted from the failure. */
723 		rte_errno = EINVAL;
724 		return -rte_errno;
725 	}
726 	rte_spinlock_init(&sh->ct_mng->ct_sl);
727 	rte_rwlock_init(&sh->ct_mng->resize_rwl);
728 	LIST_INIT(&sh->ct_mng->free_cts);
729 	return 0;
730 }
731 
732 /*
733  * Close and release all the resources of the
734  * ASO connection tracking management structure.
735  *
736  * @param[in] sh
737  *   Pointer to mlx5_dev_ctx_shared object to free.
738  */
739 static void
740 mlx5_flow_aso_ct_mng_close(struct mlx5_dev_ctx_shared *sh)
741 {
742 	struct mlx5_aso_ct_pools_mng *mng = sh->ct_mng;
743 	struct mlx5_aso_ct_pool *ct_pool;
744 	struct mlx5_aso_ct_action *ct;
745 	uint32_t idx;
746 	uint32_t val;
747 	uint32_t cnt;
748 	int i;
749 
750 	mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
751 	idx = mng->next;
752 	while (idx--) {
753 		cnt = 0;
754 		ct_pool = mng->pools[idx];
755 		for (i = 0; i < MLX5_ASO_CT_ACTIONS_PER_POOL; i++) {
756 			ct = &ct_pool->actions[i];
757 			val = __atomic_fetch_sub(&ct->refcnt, 1,
758 						 __ATOMIC_RELAXED);
759 			MLX5_ASSERT(val == 1);
760 			if (val > 1)
761 				cnt++;
762 #ifdef HAVE_MLX5_DR_ACTION_ASO_CT
763 			if (ct->dr_action_orig)
764 				claim_zero(mlx5_glue->destroy_flow_action
765 							(ct->dr_action_orig));
766 			if (ct->dr_action_rply)
767 				claim_zero(mlx5_glue->destroy_flow_action
768 							(ct->dr_action_rply));
769 #endif
770 		}
771 		claim_zero(mlx5_devx_cmd_destroy(ct_pool->devx_obj));
772 		if (cnt) {
773 			DRV_LOG(DEBUG, "%u ASO CT objects are being used in the pool %u",
774 				cnt, i);
775 		}
776 		mlx5_free(ct_pool);
777 		/* in case of failure. */
778 		mng->next--;
779 	}
780 	mlx5_free(mng->pools);
781 	mlx5_free(mng);
782 	/* Management structure must be cleared to 0s during allocation. */
783 	sh->ct_mng = NULL;
784 }
785 
786 /**
787  * Initialize the flow resources' indexed mempool.
788  *
789  * @param[in] sh
790  *   Pointer to mlx5_dev_ctx_shared object.
791  * @param[in] config
792  *   Pointer to user dev config.
793  */
794 static void
795 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
796 		       const struct mlx5_dev_config *config)
797 {
798 	uint8_t i;
799 	struct mlx5_indexed_pool_config cfg;
800 
801 	for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
802 		cfg = mlx5_ipool_cfg[i];
803 		switch (i) {
804 		default:
805 			break;
806 		/*
807 		 * Set MLX5_IPOOL_MLX5_FLOW ipool size
808 		 * according to PCI function flow configuration.
809 		 */
810 		case MLX5_IPOOL_MLX5_FLOW:
811 			cfg.size = config->dv_flow_en ?
812 				sizeof(struct mlx5_flow_handle) :
813 				MLX5_FLOW_HANDLE_VERBS_SIZE;
814 			break;
815 		}
816 		if (config->reclaim_mode) {
817 			cfg.release_mem_en = 1;
818 			cfg.per_core_cache = 0;
819 		} else {
820 			cfg.release_mem_en = 0;
821 		}
822 		sh->ipool[i] = mlx5_ipool_create(&cfg);
823 	}
824 }
825 
826 
827 /**
828  * Release the flow resources' indexed mempool.
829  *
830  * @param[in] sh
831  *   Pointer to mlx5_dev_ctx_shared object.
832  */
833 static void
834 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
835 {
836 	uint8_t i;
837 
838 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
839 		mlx5_ipool_destroy(sh->ipool[i]);
840 	for (i = 0; i < MLX5_MAX_MODIFY_NUM; ++i)
841 		if (sh->mdh_ipools[i])
842 			mlx5_ipool_destroy(sh->mdh_ipools[i]);
843 }
844 
845 /*
846  * Check if dynamic flex parser for eCPRI already exists.
847  *
848  * @param dev
849  *   Pointer to Ethernet device structure.
850  *
851  * @return
852  *   true on exists, false on not.
853  */
854 bool
855 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
856 {
857 	struct mlx5_priv *priv = dev->data->dev_private;
858 	struct mlx5_flex_parser_profiles *prf =
859 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
860 
861 	return !!prf->obj;
862 }
863 
864 /*
865  * Allocation of a flex parser for eCPRI. Once created, this parser related
866  * resources will be held until the device is closed.
867  *
868  * @param dev
869  *   Pointer to Ethernet device structure.
870  *
871  * @return
872  *   0 on success, a negative errno value otherwise and rte_errno is set.
873  */
874 int
875 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
876 {
877 	struct mlx5_priv *priv = dev->data->dev_private;
878 	struct mlx5_flex_parser_profiles *prf =
879 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
880 	struct mlx5_devx_graph_node_attr node = {
881 		.modify_field_select = 0,
882 	};
883 	uint32_t ids[8];
884 	int ret;
885 
886 	if (!priv->config.hca_attr.parse_graph_flex_node) {
887 		DRV_LOG(ERR, "Dynamic flex parser is not supported "
888 			"for device %s.", priv->dev_data->name);
889 		return -ENOTSUP;
890 	}
891 	node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
892 	/* 8 bytes now: 4B common header + 4B message body header. */
893 	node.header_length_base_value = 0x8;
894 	/* After MAC layer: Ether / VLAN. */
895 	node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
896 	/* Type of compared condition should be 0xAEFE in the L2 layer. */
897 	node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
898 	/* Sample #0: type in common header. */
899 	node.sample[0].flow_match_sample_en = 1;
900 	/* Fixed offset. */
901 	node.sample[0].flow_match_sample_offset_mode = 0x0;
902 	/* Only the 2nd byte will be used. */
903 	node.sample[0].flow_match_sample_field_base_offset = 0x0;
904 	/* Sample #1: message payload. */
905 	node.sample[1].flow_match_sample_en = 1;
906 	/* Fixed offset. */
907 	node.sample[1].flow_match_sample_offset_mode = 0x0;
908 	/*
909 	 * Only the first two bytes will be used right now, and its offset will
910 	 * start after the common header that with the length of a DW(u32).
911 	 */
912 	node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
913 	prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
914 	if (!prf->obj) {
915 		DRV_LOG(ERR, "Failed to create flex parser node object.");
916 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
917 	}
918 	prf->num = 2;
919 	ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
920 	if (ret) {
921 		DRV_LOG(ERR, "Failed to query sample IDs.");
922 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
923 	}
924 	prf->offset[0] = 0x0;
925 	prf->offset[1] = sizeof(uint32_t);
926 	prf->ids[0] = ids[0];
927 	prf->ids[1] = ids[1];
928 	return 0;
929 }
930 
931 /*
932  * Destroy the flex parser node, including the parser itself, input / output
933  * arcs and DW samples. Resources could be reused then.
934  *
935  * @param dev
936  *   Pointer to Ethernet device structure.
937  */
938 static void
939 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
940 {
941 	struct mlx5_priv *priv = dev->data->dev_private;
942 	struct mlx5_flex_parser_profiles *prf =
943 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
944 
945 	if (prf->obj)
946 		mlx5_devx_cmd_destroy(prf->obj);
947 	prf->obj = NULL;
948 }
949 
950 /*
951  * Allocate Rx and Tx UARs in robust fashion.
952  * This routine handles the following UAR allocation issues:
953  *
954  *  - tries to allocate the UAR with the most appropriate memory
955  *    mapping type from the ones supported by the host
956  *
957  *  - tries to allocate the UAR with non-NULL base address
958  *    OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
959  *    UAR base address if UAR was not the first object in the UAR page.
960  *    It caused the PMD failure and we should try to get another UAR
961  *    till we get the first one with non-NULL base address returned.
962  */
963 static int
964 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
965 		     const struct mlx5_dev_config *config)
966 {
967 	uint32_t uar_mapping, retry;
968 	int err = 0;
969 	void *base_addr;
970 
971 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
972 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
973 		/* Control the mapping type according to the settings. */
974 		uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
975 			      MLX5DV_UAR_ALLOC_TYPE_NC :
976 			      MLX5DV_UAR_ALLOC_TYPE_BF;
977 #else
978 		RTE_SET_USED(config);
979 		/*
980 		 * It seems we have no way to control the memory mapping type
981 		 * for the UAR, the default "Write-Combining" type is supposed.
982 		 * The UAR initialization on queue creation queries the
983 		 * actual mapping type done by Verbs/kernel and setups the
984 		 * PMD datapath accordingly.
985 		 */
986 		uar_mapping = 0;
987 #endif
988 		sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
989 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
990 		if (!sh->tx_uar &&
991 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
992 			if (config->dbnc == MLX5_TXDB_CACHED ||
993 			    config->dbnc == MLX5_TXDB_HEURISTIC)
994 				DRV_LOG(WARNING, "Devarg tx_db_nc setting "
995 						 "is not supported by DevX");
996 			/*
997 			 * In some environments like virtual machine
998 			 * the Write Combining mapped might be not supported
999 			 * and UAR allocation fails. We try "Non-Cached"
1000 			 * mapping for the case. The tx_burst routines take
1001 			 * the UAR mapping type into account on UAR setup
1002 			 * on queue creation.
1003 			 */
1004 			DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (BF)");
1005 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1006 			sh->tx_uar = mlx5_glue->devx_alloc_uar
1007 							(sh->ctx, uar_mapping);
1008 		} else if (!sh->tx_uar &&
1009 			   uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
1010 			if (config->dbnc == MLX5_TXDB_NCACHED)
1011 				DRV_LOG(WARNING, "Devarg tx_db_nc settings "
1012 						 "is not supported by DevX");
1013 			/*
1014 			 * If Verbs/kernel does not support "Non-Cached"
1015 			 * try the "Write-Combining".
1016 			 */
1017 			DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (NC)");
1018 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
1019 			sh->tx_uar = mlx5_glue->devx_alloc_uar
1020 							(sh->ctx, uar_mapping);
1021 		}
1022 #endif
1023 		if (!sh->tx_uar) {
1024 			DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
1025 			err = ENOMEM;
1026 			goto exit;
1027 		}
1028 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
1029 		if (base_addr)
1030 			break;
1031 		/*
1032 		 * The UARs are allocated by rdma_core within the
1033 		 * IB device context, on context closure all UARs
1034 		 * will be freed, should be no memory/object leakage.
1035 		 */
1036 		DRV_LOG(DEBUG, "Retrying to allocate Tx DevX UAR");
1037 		sh->tx_uar = NULL;
1038 	}
1039 	/* Check whether we finally succeeded with valid UAR allocation. */
1040 	if (!sh->tx_uar) {
1041 		DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
1042 		err = ENOMEM;
1043 		goto exit;
1044 	}
1045 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
1046 		uar_mapping = 0;
1047 		sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
1048 							(sh->ctx, uar_mapping);
1049 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1050 		if (!sh->devx_rx_uar &&
1051 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
1052 			/*
1053 			 * Rx UAR is used to control interrupts only,
1054 			 * should be no datapath noticeable impact,
1055 			 * can try "Non-Cached" mapping safely.
1056 			 */
1057 			DRV_LOG(DEBUG, "Failed to allocate Rx DevX UAR (BF)");
1058 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1059 			sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
1060 							(sh->ctx, uar_mapping);
1061 		}
1062 #endif
1063 		if (!sh->devx_rx_uar) {
1064 			DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
1065 			err = ENOMEM;
1066 			goto exit;
1067 		}
1068 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
1069 		if (base_addr)
1070 			break;
1071 		/*
1072 		 * The UARs are allocated by rdma_core within the
1073 		 * IB device context, on context closure all UARs
1074 		 * will be freed, should be no memory/object leakage.
1075 		 */
1076 		DRV_LOG(DEBUG, "Retrying to allocate Rx DevX UAR");
1077 		sh->devx_rx_uar = NULL;
1078 	}
1079 	/* Check whether we finally succeeded with valid UAR allocation. */
1080 	if (!sh->devx_rx_uar) {
1081 		DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
1082 		err = ENOMEM;
1083 	}
1084 exit:
1085 	return err;
1086 }
1087 
1088 /**
1089  * Allocate shared device context. If there is multiport device the
1090  * master and representors will share this context, if there is single
1091  * port dedicated device, the context will be used by only given
1092  * port due to unification.
1093  *
1094  * Routine first searches the context for the specified device name,
1095  * if found the shared context assumed and reference counter is incremented.
1096  * If no context found the new one is created and initialized with specified
1097  * device context and parameters.
1098  *
1099  * @param[in] spawn
1100  *   Pointer to the device attributes (name, port, etc).
1101  * @param[in] config
1102  *   Pointer to device configuration structure.
1103  *
1104  * @return
1105  *   Pointer to mlx5_dev_ctx_shared object on success,
1106  *   otherwise NULL and rte_errno is set.
1107  */
1108 struct mlx5_dev_ctx_shared *
1109 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
1110 			   const struct mlx5_dev_config *config)
1111 {
1112 	struct mlx5_dev_ctx_shared *sh;
1113 	int err = 0;
1114 	uint32_t i;
1115 	struct mlx5_devx_tis_attr tis_attr = { 0 };
1116 
1117 	MLX5_ASSERT(spawn);
1118 	/* Secondary process should not create the shared context. */
1119 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1120 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1121 	/* Search for IB context by device name. */
1122 	LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
1123 		if (!strcmp(sh->ibdev_name,
1124 			mlx5_os_get_dev_device_name(spawn->phys_dev))) {
1125 			sh->refcnt++;
1126 			goto exit;
1127 		}
1128 	}
1129 	/* No device found, we have to create new shared context. */
1130 	MLX5_ASSERT(spawn->max_port);
1131 	sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1132 			 sizeof(struct mlx5_dev_ctx_shared) +
1133 			 spawn->max_port *
1134 			 sizeof(struct mlx5_dev_shared_port),
1135 			 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1136 	if (!sh) {
1137 		DRV_LOG(ERR, "shared context allocation failure");
1138 		rte_errno  = ENOMEM;
1139 		goto exit;
1140 	}
1141 	sh->numa_node = spawn->numa_node;
1142 	if (spawn->bond_info)
1143 		sh->bond = *spawn->bond_info;
1144 	err = mlx5_os_open_device(spawn, config, sh);
1145 	if (!sh->ctx)
1146 		goto error;
1147 	err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
1148 	if (err) {
1149 		DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
1150 		goto error;
1151 	}
1152 	sh->refcnt = 1;
1153 	sh->max_port = spawn->max_port;
1154 	sh->reclaim_mode = config->reclaim_mode;
1155 	strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
1156 		sizeof(sh->ibdev_name) - 1);
1157 	strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
1158 		sizeof(sh->ibdev_path) - 1);
1159 	/*
1160 	 * Setting port_id to max unallowed value means
1161 	 * there is no interrupt subhandler installed for
1162 	 * the given port index i.
1163 	 */
1164 	for (i = 0; i < sh->max_port; i++) {
1165 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
1166 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
1167 	}
1168 	sh->pd = mlx5_os_alloc_pd(sh->ctx);
1169 	if (sh->pd == NULL) {
1170 		DRV_LOG(ERR, "PD allocation failure");
1171 		err = ENOMEM;
1172 		goto error;
1173 	}
1174 	if (sh->devx) {
1175 		err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
1176 		if (err) {
1177 			DRV_LOG(ERR, "Fail to extract pdn from PD");
1178 			goto error;
1179 		}
1180 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
1181 		if (!sh->td) {
1182 			DRV_LOG(ERR, "TD allocation failure");
1183 			err = ENOMEM;
1184 			goto error;
1185 		}
1186 		tis_attr.transport_domain = sh->td->id;
1187 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
1188 		if (!sh->tis) {
1189 			DRV_LOG(ERR, "TIS allocation failure");
1190 			err = ENOMEM;
1191 			goto error;
1192 		}
1193 		err = mlx5_alloc_rxtx_uars(sh, config);
1194 		if (err)
1195 			goto error;
1196 		MLX5_ASSERT(sh->tx_uar);
1197 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
1198 
1199 		MLX5_ASSERT(sh->devx_rx_uar);
1200 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
1201 	}
1202 #ifndef RTE_ARCH_64
1203 	/* Initialize UAR access locks for 32bit implementations. */
1204 	rte_spinlock_init(&sh->uar_lock_cq);
1205 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1206 		rte_spinlock_init(&sh->uar_lock[i]);
1207 #endif
1208 	/*
1209 	 * Once the device is added to the list of memory event
1210 	 * callback, its global MR cache table cannot be expanded
1211 	 * on the fly because of deadlock. If it overflows, lookup
1212 	 * should be done by searching MR list linearly, which is slow.
1213 	 *
1214 	 * At this point the device is not added to the memory
1215 	 * event list yet, context is just being created.
1216 	 */
1217 	err = mlx5_mr_btree_init(&sh->share_cache.cache,
1218 				 MLX5_MR_BTREE_CACHE_N * 2,
1219 				 sh->numa_node);
1220 	if (err) {
1221 		err = rte_errno;
1222 		goto error;
1223 	}
1224 	mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
1225 			      &sh->share_cache.dereg_mr_cb);
1226 	mlx5_os_dev_shared_handler_install(sh);
1227 	sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
1228 	if (!sh->cnt_id_tbl) {
1229 		err = rte_errno;
1230 		goto error;
1231 	}
1232 	if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1233 		err = mlx5_flow_os_init_workspace_once();
1234 		if (err)
1235 			goto error;
1236 	}
1237 	mlx5_flow_aging_init(sh);
1238 	mlx5_flow_counters_mng_init(sh);
1239 	mlx5_flow_ipool_create(sh, config);
1240 	/* Add device to memory callback list. */
1241 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1242 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1243 			 sh, mem_event_cb);
1244 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1245 	/* Add context to the global device list. */
1246 	LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1247 	rte_spinlock_init(&sh->geneve_tlv_opt_sl);
1248 exit:
1249 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1250 	return sh;
1251 error:
1252 	pthread_mutex_destroy(&sh->txpp.mutex);
1253 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1254 	MLX5_ASSERT(sh);
1255 	if (sh->cnt_id_tbl)
1256 		mlx5_l3t_destroy(sh->cnt_id_tbl);
1257 	if (sh->share_cache.cache.table)
1258 		mlx5_mr_btree_free(&sh->share_cache.cache);
1259 	if (sh->tis)
1260 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1261 	if (sh->td)
1262 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1263 	if (sh->devx_rx_uar)
1264 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1265 	if (sh->tx_uar)
1266 		mlx5_glue->devx_free_uar(sh->tx_uar);
1267 	if (sh->pd)
1268 		claim_zero(mlx5_os_dealloc_pd(sh->pd));
1269 	if (sh->ctx)
1270 		claim_zero(mlx5_glue->close_device(sh->ctx));
1271 	mlx5_free(sh);
1272 	MLX5_ASSERT(err > 0);
1273 	rte_errno = err;
1274 	return NULL;
1275 }
1276 
1277 /**
1278  * Free shared IB device context. Decrement counter and if zero free
1279  * all allocated resources and close handles.
1280  *
1281  * @param[in] sh
1282  *   Pointer to mlx5_dev_ctx_shared object to free
1283  */
1284 void
1285 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1286 {
1287 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1288 #ifdef RTE_LIBRTE_MLX5_DEBUG
1289 	/* Check the object presence in the list. */
1290 	struct mlx5_dev_ctx_shared *lctx;
1291 
1292 	LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1293 		if (lctx == sh)
1294 			break;
1295 	MLX5_ASSERT(lctx);
1296 	if (lctx != sh) {
1297 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
1298 		goto exit;
1299 	}
1300 #endif
1301 	MLX5_ASSERT(sh);
1302 	MLX5_ASSERT(sh->refcnt);
1303 	/* Secondary process should not free the shared context. */
1304 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1305 	if (--sh->refcnt)
1306 		goto exit;
1307 	/* Remove from memory callback device list. */
1308 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1309 	LIST_REMOVE(sh, mem_event_cb);
1310 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1311 	/* Release created Memory Regions. */
1312 	mlx5_mr_release_cache(&sh->share_cache);
1313 	/* Remove context from the global device list. */
1314 	LIST_REMOVE(sh, next);
1315 	/* Release flow workspaces objects on the last device. */
1316 	if (LIST_EMPTY(&mlx5_dev_ctx_list))
1317 		mlx5_flow_os_release_workspace();
1318 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1319 	/*
1320 	 *  Ensure there is no async event handler installed.
1321 	 *  Only primary process handles async device events.
1322 	 **/
1323 	mlx5_flow_counters_mng_close(sh);
1324 	if (sh->aso_age_mng) {
1325 		mlx5_flow_aso_age_mng_close(sh);
1326 		sh->aso_age_mng = NULL;
1327 	}
1328 	if (sh->mtrmng)
1329 		mlx5_aso_flow_mtrs_mng_close(sh);
1330 	mlx5_flow_ipool_destroy(sh);
1331 	mlx5_os_dev_shared_handler_uninstall(sh);
1332 	if (sh->cnt_id_tbl) {
1333 		mlx5_l3t_destroy(sh->cnt_id_tbl);
1334 		sh->cnt_id_tbl = NULL;
1335 	}
1336 	if (sh->tx_uar) {
1337 		mlx5_glue->devx_free_uar(sh->tx_uar);
1338 		sh->tx_uar = NULL;
1339 	}
1340 	if (sh->pd)
1341 		claim_zero(mlx5_os_dealloc_pd(sh->pd));
1342 	if (sh->tis)
1343 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1344 	if (sh->td)
1345 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1346 	if (sh->devx_rx_uar)
1347 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1348 	if (sh->ctx)
1349 		claim_zero(mlx5_glue->close_device(sh->ctx));
1350 	MLX5_ASSERT(sh->geneve_tlv_option_resource == NULL);
1351 	pthread_mutex_destroy(&sh->txpp.mutex);
1352 	mlx5_free(sh);
1353 	return;
1354 exit:
1355 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1356 }
1357 
1358 /**
1359  * Destroy table hash list.
1360  *
1361  * @param[in] priv
1362  *   Pointer to the private device data structure.
1363  */
1364 void
1365 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1366 {
1367 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1368 
1369 	if (!sh->flow_tbls)
1370 		return;
1371 	mlx5_hlist_destroy(sh->flow_tbls);
1372 }
1373 
1374 /**
1375  * Initialize flow table hash list and create the root tables entry
1376  * for each domain.
1377  *
1378  * @param[in] priv
1379  *   Pointer to the private device data structure.
1380  *
1381  * @return
1382  *   Zero on success, positive error code otherwise.
1383  */
1384 int
1385 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1386 {
1387 	int err = 0;
1388 	/* Tables are only used in DV and DR modes. */
1389 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1390 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1391 	char s[MLX5_NAME_SIZE];
1392 
1393 	MLX5_ASSERT(sh);
1394 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1395 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1396 					  false, true, sh,
1397 					  flow_dv_tbl_create_cb,
1398 					  flow_dv_tbl_match_cb,
1399 					  flow_dv_tbl_remove_cb,
1400 					  flow_dv_tbl_clone_cb,
1401 					  flow_dv_tbl_clone_free_cb);
1402 	if (!sh->flow_tbls) {
1403 		DRV_LOG(ERR, "flow tables with hash creation failed.");
1404 		err = ENOMEM;
1405 		return err;
1406 	}
1407 #ifndef HAVE_MLX5DV_DR
1408 	struct rte_flow_error error;
1409 	struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1410 
1411 	/*
1412 	 * In case we have not DR support, the zero tables should be created
1413 	 * because DV expect to see them even if they cannot be created by
1414 	 * RDMA-CORE.
1415 	 */
1416 	if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0,
1417 		NULL, 0, 1, 0, &error) ||
1418 	    !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0,
1419 		NULL, 0, 1, 0, &error) ||
1420 	    !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0,
1421 		NULL, 0, 1, 0, &error)) {
1422 		err = ENOMEM;
1423 		goto error;
1424 	}
1425 	return err;
1426 error:
1427 	mlx5_free_table_hash_list(priv);
1428 #endif /* HAVE_MLX5DV_DR */
1429 #endif
1430 	return err;
1431 }
1432 
1433 /**
1434  * Retrieve integer value from environment variable.
1435  *
1436  * @param[in] name
1437  *   Environment variable name.
1438  *
1439  * @return
1440  *   Integer value, 0 if the variable is not set.
1441  */
1442 int
1443 mlx5_getenv_int(const char *name)
1444 {
1445 	const char *val = getenv(name);
1446 
1447 	if (val == NULL)
1448 		return 0;
1449 	return atoi(val);
1450 }
1451 
1452 /**
1453  * DPDK callback to add udp tunnel port
1454  *
1455  * @param[in] dev
1456  *   A pointer to eth_dev
1457  * @param[in] udp_tunnel
1458  *   A pointer to udp tunnel
1459  *
1460  * @return
1461  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1462  */
1463 int
1464 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1465 			 struct rte_eth_udp_tunnel *udp_tunnel)
1466 {
1467 	MLX5_ASSERT(udp_tunnel != NULL);
1468 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1469 	    udp_tunnel->udp_port == 4789)
1470 		return 0;
1471 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1472 	    udp_tunnel->udp_port == 4790)
1473 		return 0;
1474 	return -ENOTSUP;
1475 }
1476 
1477 /**
1478  * Initialize process private data structure.
1479  *
1480  * @param dev
1481  *   Pointer to Ethernet device structure.
1482  *
1483  * @return
1484  *   0 on success, a negative errno value otherwise and rte_errno is set.
1485  */
1486 int
1487 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1488 {
1489 	struct mlx5_priv *priv = dev->data->dev_private;
1490 	struct mlx5_proc_priv *ppriv;
1491 	size_t ppriv_size;
1492 
1493 	mlx5_proc_priv_uninit(dev);
1494 	/*
1495 	 * UAR register table follows the process private structure. BlueFlame
1496 	 * registers for Tx queues are stored in the table.
1497 	 */
1498 	ppriv_size =
1499 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1500 	ppriv = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, ppriv_size,
1501 			    RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1502 	if (!ppriv) {
1503 		rte_errno = ENOMEM;
1504 		return -rte_errno;
1505 	}
1506 	ppriv->uar_table_sz = priv->txqs_n;
1507 	dev->process_private = ppriv;
1508 	return 0;
1509 }
1510 
1511 /**
1512  * Un-initialize process private data structure.
1513  *
1514  * @param dev
1515  *   Pointer to Ethernet device structure.
1516  */
1517 void
1518 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1519 {
1520 	if (!dev->process_private)
1521 		return;
1522 	mlx5_free(dev->process_private);
1523 	dev->process_private = NULL;
1524 }
1525 
1526 /**
1527  * DPDK callback to close the device.
1528  *
1529  * Destroy all queues and objects, free memory.
1530  *
1531  * @param dev
1532  *   Pointer to Ethernet device structure.
1533  */
1534 int
1535 mlx5_dev_close(struct rte_eth_dev *dev)
1536 {
1537 	struct mlx5_priv *priv = dev->data->dev_private;
1538 	unsigned int i;
1539 	int ret;
1540 
1541 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1542 		/* Check if process_private released. */
1543 		if (!dev->process_private)
1544 			return 0;
1545 		mlx5_tx_uar_uninit_secondary(dev);
1546 		mlx5_proc_priv_uninit(dev);
1547 		rte_eth_dev_release_port(dev);
1548 		return 0;
1549 	}
1550 	if (!priv->sh)
1551 		return 0;
1552 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1553 		dev->data->port_id,
1554 		((priv->sh->ctx != NULL) ?
1555 		mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1556 	/*
1557 	 * If default mreg copy action is removed at the stop stage,
1558 	 * the search will return none and nothing will be done anymore.
1559 	 */
1560 	mlx5_flow_stop_default(dev);
1561 	mlx5_traffic_disable(dev);
1562 	/*
1563 	 * If all the flows are already flushed in the device stop stage,
1564 	 * then this will return directly without any action.
1565 	 */
1566 	mlx5_flow_list_flush(dev, MLX5_FLOW_TYPE_GEN, true);
1567 	mlx5_action_handle_flush(dev);
1568 	mlx5_flow_meter_flush(dev, NULL);
1569 	/* Prevent crashes when queues are still in use. */
1570 	dev->rx_pkt_burst = removed_rx_burst;
1571 	dev->tx_pkt_burst = removed_tx_burst;
1572 	rte_wmb();
1573 	/* Disable datapath on secondary process. */
1574 	mlx5_mp_os_req_stop_rxtx(dev);
1575 	/* Free the eCPRI flex parser resource. */
1576 	mlx5_flex_parser_ecpri_release(dev);
1577 	if (priv->rxqs != NULL) {
1578 		/* XXX race condition if mlx5_rx_burst() is still running. */
1579 		rte_delay_us_sleep(1000);
1580 		for (i = 0; (i != priv->rxqs_n); ++i)
1581 			mlx5_rxq_release(dev, i);
1582 		priv->rxqs_n = 0;
1583 		priv->rxqs = NULL;
1584 	}
1585 	if (priv->representor) {
1586 		/* Each representor has a dedicated interrupts handler */
1587 		mlx5_free(dev->intr_handle);
1588 		dev->intr_handle = NULL;
1589 	}
1590 	if (priv->txqs != NULL) {
1591 		/* XXX race condition if mlx5_tx_burst() is still running. */
1592 		rte_delay_us_sleep(1000);
1593 		for (i = 0; (i != priv->txqs_n); ++i)
1594 			mlx5_txq_release(dev, i);
1595 		priv->txqs_n = 0;
1596 		priv->txqs = NULL;
1597 	}
1598 	mlx5_proc_priv_uninit(dev);
1599 	if (priv->q_counters) {
1600 		mlx5_devx_cmd_destroy(priv->q_counters);
1601 		priv->q_counters = NULL;
1602 	}
1603 	if (priv->drop_queue.hrxq)
1604 		mlx5_drop_action_destroy(dev);
1605 	if (priv->mreg_cp_tbl)
1606 		mlx5_hlist_destroy(priv->mreg_cp_tbl);
1607 	mlx5_mprq_free_mp(dev);
1608 	if (priv->sh->ct_mng)
1609 		mlx5_flow_aso_ct_mng_close(priv->sh);
1610 	mlx5_os_free_shared_dr(priv);
1611 	if (priv->rss_conf.rss_key != NULL)
1612 		mlx5_free(priv->rss_conf.rss_key);
1613 	if (priv->reta_idx != NULL)
1614 		mlx5_free(priv->reta_idx);
1615 	if (priv->config.vf)
1616 		mlx5_os_mac_addr_flush(dev);
1617 	if (priv->nl_socket_route >= 0)
1618 		close(priv->nl_socket_route);
1619 	if (priv->nl_socket_rdma >= 0)
1620 		close(priv->nl_socket_rdma);
1621 	if (priv->vmwa_context)
1622 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1623 	ret = mlx5_hrxq_verify(dev);
1624 	if (ret)
1625 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1626 			dev->data->port_id);
1627 	ret = mlx5_ind_table_obj_verify(dev);
1628 	if (ret)
1629 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1630 			dev->data->port_id);
1631 	ret = mlx5_rxq_obj_verify(dev);
1632 	if (ret)
1633 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1634 			dev->data->port_id);
1635 	ret = mlx5_rxq_verify(dev);
1636 	if (ret)
1637 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1638 			dev->data->port_id);
1639 	ret = mlx5_txq_obj_verify(dev);
1640 	if (ret)
1641 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1642 			dev->data->port_id);
1643 	ret = mlx5_txq_verify(dev);
1644 	if (ret)
1645 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1646 			dev->data->port_id);
1647 	ret = mlx5_flow_verify(dev);
1648 	if (ret)
1649 		DRV_LOG(WARNING, "port %u some flows still remain",
1650 			dev->data->port_id);
1651 	if (priv->hrxqs)
1652 		mlx5_list_destroy(priv->hrxqs);
1653 	/*
1654 	 * Free the shared context in last turn, because the cleanup
1655 	 * routines above may use some shared fields, like
1656 	 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1657 	 * ifindex if Netlink fails.
1658 	 */
1659 	mlx5_free_shared_dev_ctx(priv->sh);
1660 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1661 		unsigned int c = 0;
1662 		uint16_t port_id;
1663 
1664 		MLX5_ETH_FOREACH_DEV(port_id, dev->device) {
1665 			struct mlx5_priv *opriv =
1666 				rte_eth_devices[port_id].data->dev_private;
1667 
1668 			if (!opriv ||
1669 			    opriv->domain_id != priv->domain_id ||
1670 			    &rte_eth_devices[port_id] == dev)
1671 				continue;
1672 			++c;
1673 			break;
1674 		}
1675 		if (!c)
1676 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1677 	}
1678 	memset(priv, 0, sizeof(*priv));
1679 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1680 	/*
1681 	 * Reset mac_addrs to NULL such that it is not freed as part of
1682 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1683 	 * it is freed when dev_private is freed.
1684 	 */
1685 	dev->data->mac_addrs = NULL;
1686 	return 0;
1687 }
1688 
1689 const struct eth_dev_ops mlx5_dev_ops = {
1690 	.dev_configure = mlx5_dev_configure,
1691 	.dev_start = mlx5_dev_start,
1692 	.dev_stop = mlx5_dev_stop,
1693 	.dev_set_link_down = mlx5_set_link_down,
1694 	.dev_set_link_up = mlx5_set_link_up,
1695 	.dev_close = mlx5_dev_close,
1696 	.promiscuous_enable = mlx5_promiscuous_enable,
1697 	.promiscuous_disable = mlx5_promiscuous_disable,
1698 	.allmulticast_enable = mlx5_allmulticast_enable,
1699 	.allmulticast_disable = mlx5_allmulticast_disable,
1700 	.link_update = mlx5_link_update,
1701 	.stats_get = mlx5_stats_get,
1702 	.stats_reset = mlx5_stats_reset,
1703 	.xstats_get = mlx5_xstats_get,
1704 	.xstats_reset = mlx5_xstats_reset,
1705 	.xstats_get_names = mlx5_xstats_get_names,
1706 	.fw_version_get = mlx5_fw_version_get,
1707 	.dev_infos_get = mlx5_dev_infos_get,
1708 	.representor_info_get = mlx5_representor_info_get,
1709 	.read_clock = mlx5_txpp_read_clock,
1710 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1711 	.vlan_filter_set = mlx5_vlan_filter_set,
1712 	.rx_queue_setup = mlx5_rx_queue_setup,
1713 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1714 	.tx_queue_setup = mlx5_tx_queue_setup,
1715 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1716 	.rx_queue_release = mlx5_rx_queue_release,
1717 	.tx_queue_release = mlx5_tx_queue_release,
1718 	.rx_queue_start = mlx5_rx_queue_start,
1719 	.rx_queue_stop = mlx5_rx_queue_stop,
1720 	.tx_queue_start = mlx5_tx_queue_start,
1721 	.tx_queue_stop = mlx5_tx_queue_stop,
1722 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1723 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1724 	.mac_addr_remove = mlx5_mac_addr_remove,
1725 	.mac_addr_add = mlx5_mac_addr_add,
1726 	.mac_addr_set = mlx5_mac_addr_set,
1727 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1728 	.mtu_set = mlx5_dev_set_mtu,
1729 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1730 	.vlan_offload_set = mlx5_vlan_offload_set,
1731 	.reta_update = mlx5_dev_rss_reta_update,
1732 	.reta_query = mlx5_dev_rss_reta_query,
1733 	.rss_hash_update = mlx5_rss_hash_update,
1734 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1735 	.flow_ops_get = mlx5_flow_ops_get,
1736 	.rxq_info_get = mlx5_rxq_info_get,
1737 	.txq_info_get = mlx5_txq_info_get,
1738 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1739 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1740 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1741 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1742 	.is_removed = mlx5_is_removed,
1743 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1744 	.get_module_info = mlx5_get_module_info,
1745 	.get_module_eeprom = mlx5_get_module_eeprom,
1746 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1747 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1748 	.hairpin_bind = mlx5_hairpin_bind,
1749 	.hairpin_unbind = mlx5_hairpin_unbind,
1750 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1751 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1752 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1753 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1754 	.get_monitor_addr = mlx5_get_monitor_addr,
1755 };
1756 
1757 /* Available operations from secondary process. */
1758 const struct eth_dev_ops mlx5_dev_sec_ops = {
1759 	.stats_get = mlx5_stats_get,
1760 	.stats_reset = mlx5_stats_reset,
1761 	.xstats_get = mlx5_xstats_get,
1762 	.xstats_reset = mlx5_xstats_reset,
1763 	.xstats_get_names = mlx5_xstats_get_names,
1764 	.fw_version_get = mlx5_fw_version_get,
1765 	.dev_infos_get = mlx5_dev_infos_get,
1766 	.representor_info_get = mlx5_representor_info_get,
1767 	.read_clock = mlx5_txpp_read_clock,
1768 	.rx_queue_start = mlx5_rx_queue_start,
1769 	.rx_queue_stop = mlx5_rx_queue_stop,
1770 	.tx_queue_start = mlx5_tx_queue_start,
1771 	.tx_queue_stop = mlx5_tx_queue_stop,
1772 	.rxq_info_get = mlx5_rxq_info_get,
1773 	.txq_info_get = mlx5_txq_info_get,
1774 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1775 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1776 	.get_module_info = mlx5_get_module_info,
1777 	.get_module_eeprom = mlx5_get_module_eeprom,
1778 };
1779 
1780 /* Available operations in flow isolated mode. */
1781 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1782 	.dev_configure = mlx5_dev_configure,
1783 	.dev_start = mlx5_dev_start,
1784 	.dev_stop = mlx5_dev_stop,
1785 	.dev_set_link_down = mlx5_set_link_down,
1786 	.dev_set_link_up = mlx5_set_link_up,
1787 	.dev_close = mlx5_dev_close,
1788 	.promiscuous_enable = mlx5_promiscuous_enable,
1789 	.promiscuous_disable = mlx5_promiscuous_disable,
1790 	.allmulticast_enable = mlx5_allmulticast_enable,
1791 	.allmulticast_disable = mlx5_allmulticast_disable,
1792 	.link_update = mlx5_link_update,
1793 	.stats_get = mlx5_stats_get,
1794 	.stats_reset = mlx5_stats_reset,
1795 	.xstats_get = mlx5_xstats_get,
1796 	.xstats_reset = mlx5_xstats_reset,
1797 	.xstats_get_names = mlx5_xstats_get_names,
1798 	.fw_version_get = mlx5_fw_version_get,
1799 	.dev_infos_get = mlx5_dev_infos_get,
1800 	.representor_info_get = mlx5_representor_info_get,
1801 	.read_clock = mlx5_txpp_read_clock,
1802 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1803 	.vlan_filter_set = mlx5_vlan_filter_set,
1804 	.rx_queue_setup = mlx5_rx_queue_setup,
1805 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1806 	.tx_queue_setup = mlx5_tx_queue_setup,
1807 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1808 	.rx_queue_release = mlx5_rx_queue_release,
1809 	.tx_queue_release = mlx5_tx_queue_release,
1810 	.rx_queue_start = mlx5_rx_queue_start,
1811 	.rx_queue_stop = mlx5_rx_queue_stop,
1812 	.tx_queue_start = mlx5_tx_queue_start,
1813 	.tx_queue_stop = mlx5_tx_queue_stop,
1814 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1815 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1816 	.mac_addr_remove = mlx5_mac_addr_remove,
1817 	.mac_addr_add = mlx5_mac_addr_add,
1818 	.mac_addr_set = mlx5_mac_addr_set,
1819 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1820 	.mtu_set = mlx5_dev_set_mtu,
1821 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1822 	.vlan_offload_set = mlx5_vlan_offload_set,
1823 	.flow_ops_get = mlx5_flow_ops_get,
1824 	.rxq_info_get = mlx5_rxq_info_get,
1825 	.txq_info_get = mlx5_txq_info_get,
1826 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1827 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1828 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1829 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1830 	.is_removed = mlx5_is_removed,
1831 	.get_module_info = mlx5_get_module_info,
1832 	.get_module_eeprom = mlx5_get_module_eeprom,
1833 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1834 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1835 	.hairpin_bind = mlx5_hairpin_bind,
1836 	.hairpin_unbind = mlx5_hairpin_unbind,
1837 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1838 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1839 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1840 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1841 	.get_monitor_addr = mlx5_get_monitor_addr,
1842 };
1843 
1844 /**
1845  * Verify and store value for device argument.
1846  *
1847  * @param[in] key
1848  *   Key argument to verify.
1849  * @param[in] val
1850  *   Value associated with key.
1851  * @param opaque
1852  *   User data.
1853  *
1854  * @return
1855  *   0 on success, a negative errno value otherwise and rte_errno is set.
1856  */
1857 static int
1858 mlx5_args_check(const char *key, const char *val, void *opaque)
1859 {
1860 	struct mlx5_dev_config *config = opaque;
1861 	unsigned long mod;
1862 	signed long tmp;
1863 
1864 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1865 	if (!strcmp(MLX5_REPRESENTOR, key))
1866 		return 0;
1867 	errno = 0;
1868 	tmp = strtol(val, NULL, 0);
1869 	if (errno) {
1870 		rte_errno = errno;
1871 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1872 		return -rte_errno;
1873 	}
1874 	if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1875 		/* Negative values are acceptable for some keys only. */
1876 		rte_errno = EINVAL;
1877 		DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1878 		return -rte_errno;
1879 	}
1880 	mod = tmp >= 0 ? tmp : -tmp;
1881 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1882 		if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1883 			DRV_LOG(ERR, "invalid CQE compression "
1884 				     "format parameter");
1885 			rte_errno = EINVAL;
1886 			return -rte_errno;
1887 		}
1888 		config->cqe_comp = !!tmp;
1889 		config->cqe_comp_fmt = tmp;
1890 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1891 		config->hw_padding = !!tmp;
1892 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1893 		config->mprq.enabled = !!tmp;
1894 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1895 		config->mprq.stride_num_n = tmp;
1896 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1897 		config->mprq.stride_size_n = tmp;
1898 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1899 		config->mprq.max_memcpy_len = tmp;
1900 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1901 		config->mprq.min_rxqs_num = tmp;
1902 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1903 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1904 				 " converted to txq_inline_max", key);
1905 		config->txq_inline_max = tmp;
1906 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1907 		config->txq_inline_max = tmp;
1908 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1909 		config->txq_inline_min = tmp;
1910 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1911 		config->txq_inline_mpw = tmp;
1912 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1913 		config->txqs_inline = tmp;
1914 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1915 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1916 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1917 		config->mps = !!tmp;
1918 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1919 		if (tmp != MLX5_TXDB_CACHED &&
1920 		    tmp != MLX5_TXDB_NCACHED &&
1921 		    tmp != MLX5_TXDB_HEURISTIC) {
1922 			DRV_LOG(ERR, "invalid Tx doorbell "
1923 				     "mapping parameter");
1924 			rte_errno = EINVAL;
1925 			return -rte_errno;
1926 		}
1927 		config->dbnc = tmp;
1928 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1929 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1930 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1931 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1932 				 " converted to txq_inline_mpw", key);
1933 		config->txq_inline_mpw = tmp;
1934 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1935 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1936 	} else if (strcmp(MLX5_TX_PP, key) == 0) {
1937 		if (!mod) {
1938 			DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1939 			rte_errno = EINVAL;
1940 			return -rte_errno;
1941 		}
1942 		config->tx_pp = tmp;
1943 	} else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1944 		config->tx_skew = tmp;
1945 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1946 		config->rx_vec_en = !!tmp;
1947 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1948 		config->l3_vxlan_en = !!tmp;
1949 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1950 		config->vf_nl_en = !!tmp;
1951 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1952 		config->dv_esw_en = !!tmp;
1953 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1954 		config->dv_flow_en = !!tmp;
1955 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1956 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1957 		    tmp != MLX5_XMETA_MODE_META16 &&
1958 		    tmp != MLX5_XMETA_MODE_META32 &&
1959 		    tmp != MLX5_XMETA_MODE_MISS_INFO) {
1960 			DRV_LOG(ERR, "invalid extensive "
1961 				     "metadata parameter");
1962 			rte_errno = EINVAL;
1963 			return -rte_errno;
1964 		}
1965 		if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1966 			config->dv_xmeta_en = tmp;
1967 		else
1968 			config->dv_miss_info = 1;
1969 	} else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1970 		config->lacp_by_user = !!tmp;
1971 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1972 		config->mr_ext_memseg_en = !!tmp;
1973 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1974 		config->max_dump_files_num = tmp;
1975 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1976 		config->lro.timeout = tmp;
1977 	} else if (strcmp(RTE_DEVARGS_KEY_CLASS, key) == 0) {
1978 		DRV_LOG(DEBUG, "class argument is %s.", val);
1979 	} else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1980 		config->log_hp_size = tmp;
1981 	} else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1982 		if (tmp != MLX5_RCM_NONE &&
1983 		    tmp != MLX5_RCM_LIGHT &&
1984 		    tmp != MLX5_RCM_AGGR) {
1985 			DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1986 			rte_errno = EINVAL;
1987 			return -rte_errno;
1988 		}
1989 		config->reclaim_mode = tmp;
1990 	} else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1991 		config->sys_mem_en = !!tmp;
1992 	} else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1993 		config->decap_en = !!tmp;
1994 	} else if (strcmp(MLX5_ALLOW_DUPLICATE_PATTERN, key) == 0) {
1995 		config->allow_duplicate_pattern = !!tmp;
1996 	} else {
1997 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1998 		rte_errno = EINVAL;
1999 		return -rte_errno;
2000 	}
2001 	return 0;
2002 }
2003 
2004 /**
2005  * Parse device parameters.
2006  *
2007  * @param config
2008  *   Pointer to device configuration structure.
2009  * @param devargs
2010  *   Device arguments structure.
2011  *
2012  * @return
2013  *   0 on success, a negative errno value otherwise and rte_errno is set.
2014  */
2015 int
2016 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
2017 {
2018 	const char **params = (const char *[]){
2019 		MLX5_RXQ_CQE_COMP_EN,
2020 		MLX5_RXQ_PKT_PAD_EN,
2021 		MLX5_RX_MPRQ_EN,
2022 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
2023 		MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
2024 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
2025 		MLX5_RXQS_MIN_MPRQ,
2026 		MLX5_TXQ_INLINE,
2027 		MLX5_TXQ_INLINE_MIN,
2028 		MLX5_TXQ_INLINE_MAX,
2029 		MLX5_TXQ_INLINE_MPW,
2030 		MLX5_TXQS_MIN_INLINE,
2031 		MLX5_TXQS_MAX_VEC,
2032 		MLX5_TXQ_MPW_EN,
2033 		MLX5_TXQ_MPW_HDR_DSEG_EN,
2034 		MLX5_TXQ_MAX_INLINE_LEN,
2035 		MLX5_TX_DB_NC,
2036 		MLX5_TX_PP,
2037 		MLX5_TX_SKEW,
2038 		MLX5_TX_VEC_EN,
2039 		MLX5_RX_VEC_EN,
2040 		MLX5_L3_VXLAN_EN,
2041 		MLX5_VF_NL_EN,
2042 		MLX5_DV_ESW_EN,
2043 		MLX5_DV_FLOW_EN,
2044 		MLX5_DV_XMETA_EN,
2045 		MLX5_LACP_BY_USER,
2046 		MLX5_MR_EXT_MEMSEG_EN,
2047 		MLX5_REPRESENTOR,
2048 		MLX5_MAX_DUMP_FILES_NUM,
2049 		MLX5_LRO_TIMEOUT_USEC,
2050 		RTE_DEVARGS_KEY_CLASS,
2051 		MLX5_HP_BUF_SIZE,
2052 		MLX5_RECLAIM_MEM,
2053 		MLX5_SYS_MEM_EN,
2054 		MLX5_DECAP_EN,
2055 		MLX5_ALLOW_DUPLICATE_PATTERN,
2056 		NULL,
2057 	};
2058 	struct rte_kvargs *kvlist;
2059 	int ret = 0;
2060 	int i;
2061 
2062 	if (devargs == NULL)
2063 		return 0;
2064 	/* Following UGLY cast is done to pass checkpatch. */
2065 	kvlist = rte_kvargs_parse(devargs->args, params);
2066 	if (kvlist == NULL) {
2067 		rte_errno = EINVAL;
2068 		return -rte_errno;
2069 	}
2070 	/* Process parameters. */
2071 	for (i = 0; (params[i] != NULL); ++i) {
2072 		if (rte_kvargs_count(kvlist, params[i])) {
2073 			ret = rte_kvargs_process(kvlist, params[i],
2074 						 mlx5_args_check, config);
2075 			if (ret) {
2076 				rte_errno = EINVAL;
2077 				rte_kvargs_free(kvlist);
2078 				return -rte_errno;
2079 			}
2080 		}
2081 	}
2082 	rte_kvargs_free(kvlist);
2083 	return 0;
2084 }
2085 
2086 /**
2087  * Configures the minimal amount of data to inline into WQE
2088  * while sending packets.
2089  *
2090  * - the txq_inline_min has the maximal priority, if this
2091  *   key is specified in devargs
2092  * - if DevX is enabled the inline mode is queried from the
2093  *   device (HCA attributes and NIC vport context if needed).
2094  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
2095  *   and none (0 bytes) for other NICs
2096  *
2097  * @param spawn
2098  *   Verbs device parameters (name, port, switch_info) to spawn.
2099  * @param config
2100  *   Device configuration parameters.
2101  */
2102 void
2103 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
2104 		    struct mlx5_dev_config *config)
2105 {
2106 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
2107 		/* Application defines size of inlined data explicitly. */
2108 		if (spawn->pci_dev != NULL) {
2109 			switch (spawn->pci_dev->id.device_id) {
2110 			case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2111 			case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2112 				if (config->txq_inline_min <
2113 					       (int)MLX5_INLINE_HSIZE_L2) {
2114 					DRV_LOG(DEBUG,
2115 						"txq_inline_mix aligned to minimal ConnectX-4 required value %d",
2116 						(int)MLX5_INLINE_HSIZE_L2);
2117 					config->txq_inline_min =
2118 							MLX5_INLINE_HSIZE_L2;
2119 				}
2120 				break;
2121 			}
2122 		}
2123 		goto exit;
2124 	}
2125 	if (config->hca_attr.eth_net_offloads) {
2126 		/* We have DevX enabled, inline mode queried successfully. */
2127 		switch (config->hca_attr.wqe_inline_mode) {
2128 		case MLX5_CAP_INLINE_MODE_L2:
2129 			/* outer L2 header must be inlined. */
2130 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2131 			goto exit;
2132 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
2133 			/* No inline data are required by NIC. */
2134 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2135 			config->hw_vlan_insert =
2136 				config->hca_attr.wqe_vlan_insert;
2137 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
2138 			goto exit;
2139 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
2140 			/* inline mode is defined by NIC vport context. */
2141 			if (!config->hca_attr.eth_virt)
2142 				break;
2143 			switch (config->hca_attr.vport_inline_mode) {
2144 			case MLX5_INLINE_MODE_NONE:
2145 				config->txq_inline_min =
2146 					MLX5_INLINE_HSIZE_NONE;
2147 				goto exit;
2148 			case MLX5_INLINE_MODE_L2:
2149 				config->txq_inline_min =
2150 					MLX5_INLINE_HSIZE_L2;
2151 				goto exit;
2152 			case MLX5_INLINE_MODE_IP:
2153 				config->txq_inline_min =
2154 					MLX5_INLINE_HSIZE_L3;
2155 				goto exit;
2156 			case MLX5_INLINE_MODE_TCP_UDP:
2157 				config->txq_inline_min =
2158 					MLX5_INLINE_HSIZE_L4;
2159 				goto exit;
2160 			case MLX5_INLINE_MODE_INNER_L2:
2161 				config->txq_inline_min =
2162 					MLX5_INLINE_HSIZE_INNER_L2;
2163 				goto exit;
2164 			case MLX5_INLINE_MODE_INNER_IP:
2165 				config->txq_inline_min =
2166 					MLX5_INLINE_HSIZE_INNER_L3;
2167 				goto exit;
2168 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
2169 				config->txq_inline_min =
2170 					MLX5_INLINE_HSIZE_INNER_L4;
2171 				goto exit;
2172 			}
2173 		}
2174 	}
2175 	if (spawn->pci_dev == NULL) {
2176 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2177 		goto exit;
2178 	}
2179 	/*
2180 	 * We get here if we are unable to deduce
2181 	 * inline data size with DevX. Try PCI ID
2182 	 * to determine old NICs.
2183 	 */
2184 	switch (spawn->pci_dev->id.device_id) {
2185 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2186 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2187 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2188 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2189 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2190 		config->hw_vlan_insert = 0;
2191 		break;
2192 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2193 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2194 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2195 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2196 		/*
2197 		 * These NICs support VLAN insertion from WQE and
2198 		 * report the wqe_vlan_insert flag. But there is the bug
2199 		 * and PFC control may be broken, so disable feature.
2200 		 */
2201 		config->hw_vlan_insert = 0;
2202 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2203 		break;
2204 	default:
2205 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2206 		break;
2207 	}
2208 exit:
2209 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2210 }
2211 
2212 /**
2213  * Configures the metadata mask fields in the shared context.
2214  *
2215  * @param [in] dev
2216  *   Pointer to Ethernet device.
2217  */
2218 void
2219 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2220 {
2221 	struct mlx5_priv *priv = dev->data->dev_private;
2222 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2223 	uint32_t meta, mark, reg_c0;
2224 
2225 	reg_c0 = ~priv->vport_meta_mask;
2226 	switch (priv->config.dv_xmeta_en) {
2227 	case MLX5_XMETA_MODE_LEGACY:
2228 		meta = UINT32_MAX;
2229 		mark = MLX5_FLOW_MARK_MASK;
2230 		break;
2231 	case MLX5_XMETA_MODE_META16:
2232 		meta = reg_c0 >> rte_bsf32(reg_c0);
2233 		mark = MLX5_FLOW_MARK_MASK;
2234 		break;
2235 	case MLX5_XMETA_MODE_META32:
2236 		meta = UINT32_MAX;
2237 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2238 		break;
2239 	default:
2240 		meta = 0;
2241 		mark = 0;
2242 		MLX5_ASSERT(false);
2243 		break;
2244 	}
2245 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2246 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2247 				 sh->dv_mark_mask, mark);
2248 	else
2249 		sh->dv_mark_mask = mark;
2250 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2251 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2252 				 sh->dv_meta_mask, meta);
2253 	else
2254 		sh->dv_meta_mask = meta;
2255 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2256 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2257 				 sh->dv_meta_mask, reg_c0);
2258 	else
2259 		sh->dv_regc0_mask = reg_c0;
2260 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2261 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2262 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2263 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2264 }
2265 
2266 int
2267 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2268 {
2269 	static const char *const dynf_names[] = {
2270 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2271 		RTE_MBUF_DYNFLAG_METADATA_NAME,
2272 		RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2273 	};
2274 	unsigned int i;
2275 
2276 	if (n < RTE_DIM(dynf_names))
2277 		return -ENOMEM;
2278 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
2279 		if (names[i] == NULL)
2280 			return -EINVAL;
2281 		strcpy(names[i], dynf_names[i]);
2282 	}
2283 	return RTE_DIM(dynf_names);
2284 }
2285 
2286 /**
2287  * Comparison callback to sort device data.
2288  *
2289  * This is meant to be used with qsort().
2290  *
2291  * @param a[in]
2292  *   Pointer to pointer to first data object.
2293  * @param b[in]
2294  *   Pointer to pointer to second data object.
2295  *
2296  * @return
2297  *   0 if both objects are equal, less than 0 if the first argument is less
2298  *   than the second, greater than 0 otherwise.
2299  */
2300 int
2301 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2302 			      struct mlx5_dev_config *config,
2303 			      struct rte_device *dpdk_dev)
2304 {
2305 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2306 	struct mlx5_dev_config *sh_conf = NULL;
2307 	uint16_t port_id;
2308 
2309 	MLX5_ASSERT(sh);
2310 	/* Nothing to compare for the single/first device. */
2311 	if (sh->refcnt == 1)
2312 		return 0;
2313 	/* Find the device with shared context. */
2314 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
2315 		struct mlx5_priv *opriv =
2316 			rte_eth_devices[port_id].data->dev_private;
2317 
2318 		if (opriv && opriv != priv && opriv->sh == sh) {
2319 			sh_conf = &opriv->config;
2320 			break;
2321 		}
2322 	}
2323 	if (!sh_conf)
2324 		return 0;
2325 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2326 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2327 			     " for shared %s context", sh->ibdev_name);
2328 		rte_errno = EINVAL;
2329 		return rte_errno;
2330 	}
2331 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2332 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2333 			     " for shared %s context", sh->ibdev_name);
2334 		rte_errno = EINVAL;
2335 		return rte_errno;
2336 	}
2337 	return 0;
2338 }
2339 
2340 /**
2341  * Look for the ethernet device belonging to mlx5 driver.
2342  *
2343  * @param[in] port_id
2344  *   port_id to start looking for device.
2345  * @param[in] odev
2346  *   Pointer to the hint device. When device is being probed
2347  *   the its siblings (master and preceding representors might
2348  *   not have assigned driver yet (because the mlx5_os_pci_probe()
2349  *   is not completed yet, for this case match on hint
2350  *   device may be used to detect sibling device.
2351  *
2352  * @return
2353  *   port_id of found device, RTE_MAX_ETHPORT if not found.
2354  */
2355 uint16_t
2356 mlx5_eth_find_next(uint16_t port_id, struct rte_device *odev)
2357 {
2358 	while (port_id < RTE_MAX_ETHPORTS) {
2359 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2360 
2361 		if (dev->state != RTE_ETH_DEV_UNUSED &&
2362 		    dev->device &&
2363 		    (dev->device == odev ||
2364 		     (dev->device->driver &&
2365 		     dev->device->driver->name &&
2366 		     ((strcmp(dev->device->driver->name,
2367 			      MLX5_PCI_DRIVER_NAME) == 0) ||
2368 		      (strcmp(dev->device->driver->name,
2369 			      MLX5_AUXILIARY_DRIVER_NAME) == 0)))))
2370 			break;
2371 		port_id++;
2372 	}
2373 	if (port_id >= RTE_MAX_ETHPORTS)
2374 		return RTE_MAX_ETHPORTS;
2375 	return port_id;
2376 }
2377 
2378 /**
2379  * Callback to remove a device.
2380  *
2381  * This function removes all Ethernet devices belong to a given device.
2382  *
2383  * @param[in] dev
2384  *   Pointer to the generic device.
2385  *
2386  * @return
2387  *   0 on success, the function cannot fail.
2388  */
2389 int
2390 mlx5_net_remove(struct rte_device *dev)
2391 {
2392 	uint16_t port_id;
2393 	int ret = 0;
2394 
2395 	RTE_ETH_FOREACH_DEV_OF(port_id, dev) {
2396 		/*
2397 		 * mlx5_dev_close() is not registered to secondary process,
2398 		 * call the close function explicitly for secondary process.
2399 		 */
2400 		if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2401 			ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2402 		else
2403 			ret |= rte_eth_dev_close(port_id);
2404 	}
2405 	return ret == 0 ? 0 : -EIO;
2406 }
2407 
2408 static const struct rte_pci_id mlx5_pci_id_map[] = {
2409 	{
2410 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2411 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2412 	},
2413 	{
2414 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2415 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2416 	},
2417 	{
2418 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2419 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2420 	},
2421 	{
2422 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2423 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2424 	},
2425 	{
2426 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2427 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2428 	},
2429 	{
2430 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2431 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2432 	},
2433 	{
2434 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2435 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2436 	},
2437 	{
2438 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2439 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2440 	},
2441 	{
2442 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2443 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2444 	},
2445 	{
2446 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2447 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2448 	},
2449 	{
2450 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2451 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2452 	},
2453 	{
2454 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2455 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2456 	},
2457 	{
2458 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2459 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2460 	},
2461 	{
2462 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2463 				PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2464 	},
2465 	{
2466 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2467 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2468 	},
2469 	{
2470 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2471 				PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2472 	},
2473 	{
2474 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2475 				PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2476 	},
2477 	{
2478 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2479 				PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2480 	},
2481 	{
2482 		.vendor_id = 0
2483 	}
2484 };
2485 
2486 static struct mlx5_class_driver mlx5_net_driver = {
2487 	.drv_class = MLX5_CLASS_ETH,
2488 	.name = RTE_STR(MLX5_ETH_DRIVER_NAME),
2489 	.id_table = mlx5_pci_id_map,
2490 	.probe = mlx5_os_net_probe,
2491 	.remove = mlx5_net_remove,
2492 	.dma_map = mlx5_net_dma_map,
2493 	.dma_unmap = mlx5_net_dma_unmap,
2494 	.probe_again = 1,
2495 	.intr_lsc = 1,
2496 	.intr_rmv = 1,
2497 };
2498 
2499 /* Initialize driver log type. */
2500 RTE_LOG_REGISTER_DEFAULT(mlx5_logtype, NOTICE)
2501 
2502 /**
2503  * Driver initialization routine.
2504  */
2505 RTE_INIT(rte_mlx5_pmd_init)
2506 {
2507 	pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2508 	mlx5_common_init();
2509 	/* Build the static tables for Verbs conversion. */
2510 	mlx5_set_ptype_table();
2511 	mlx5_set_cksum_table();
2512 	mlx5_set_swp_types_table();
2513 	if (mlx5_glue)
2514 		mlx5_class_driver_register(&mlx5_net_driver);
2515 }
2516 
2517 RTE_PMD_EXPORT_NAME(MLX5_ETH_DRIVER_NAME, __COUNTER__);
2518 RTE_PMD_REGISTER_PCI_TABLE(MLX5_ETH_DRIVER_NAME, mlx5_pci_id_map);
2519 RTE_PMD_REGISTER_KMOD_DEP(MLX5_ETH_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");
2520