xref: /dpdk/drivers/net/mlx5/mlx5.c (revision f8dbaebbf1c9efcbb2e2354b341ed62175466a57)
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_eal_paging.h>
23 #include <rte_alarm.h>
24 #include <rte_cycles.h>
25 
26 #include <mlx5_glue.h>
27 #include <mlx5_devx_cmds.h>
28 #include <mlx5_common.h>
29 #include <mlx5_common_os.h>
30 #include <mlx5_common_mp.h>
31 #include <mlx5_malloc.h>
32 
33 #include "mlx5_defs.h"
34 #include "mlx5.h"
35 #include "mlx5_utils.h"
36 #include "mlx5_rxtx.h"
37 #include "mlx5_rx.h"
38 #include "mlx5_tx.h"
39 #include "mlx5_autoconf.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 /* Driver type key for new device global syntax. */
47 #define MLX5_DRIVER_KEY "driver"
48 
49 /* Device parameter to enable RX completion queue compression. */
50 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
51 
52 /* Device parameter to enable padding Rx packet to cacheline size. */
53 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
54 
55 /* Device parameter to enable Multi-Packet Rx queue. */
56 #define MLX5_RX_MPRQ_EN "mprq_en"
57 
58 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
59 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
60 
61 /* Device parameter to configure log 2 of the stride size for MPRQ. */
62 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
63 
64 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
65 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
66 
67 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
68 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
69 
70 /* Device parameter to configure inline send. Deprecated, ignored.*/
71 #define MLX5_TXQ_INLINE "txq_inline"
72 
73 /* Device parameter to limit packet size to inline with ordinary SEND. */
74 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
75 
76 /* Device parameter to configure minimal data size to inline. */
77 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
78 
79 /* Device parameter to limit packet size to inline with Enhanced MPW. */
80 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
81 
82 /*
83  * Device parameter to configure the number of TX queues threshold for
84  * enabling inline send.
85  */
86 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
87 
88 /*
89  * Device parameter to configure the number of TX queues threshold for
90  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
91  */
92 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
93 
94 /* Device parameter to enable multi-packet send WQEs. */
95 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
96 
97 /*
98  * Device parameter to force doorbell register mapping
99  * to non-cahed region eliminating the extra write memory barrier.
100  */
101 #define MLX5_TX_DB_NC "tx_db_nc"
102 
103 /*
104  * Device parameter to include 2 dsegs in the title WQEBB.
105  * Deprecated, ignored.
106  */
107 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
108 
109 /*
110  * Device parameter to limit the size of inlining packet.
111  * Deprecated, ignored.
112  */
113 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
114 
115 /*
116  * Device parameter to enable Tx scheduling on timestamps
117  * and specify the packet pacing granularity in nanoseconds.
118  */
119 #define MLX5_TX_PP "tx_pp"
120 
121 /*
122  * Device parameter to specify skew in nanoseconds on Tx datapath,
123  * it represents the time between SQ start WQE processing and
124  * appearing actual packet data on the wire.
125  */
126 #define MLX5_TX_SKEW "tx_skew"
127 
128 /*
129  * Device parameter to enable hardware Tx vector.
130  * Deprecated, ignored (no vectorized Tx routines anymore).
131  */
132 #define MLX5_TX_VEC_EN "tx_vec_en"
133 
134 /* Device parameter to enable hardware Rx vector. */
135 #define MLX5_RX_VEC_EN "rx_vec_en"
136 
137 /* Allow L3 VXLAN flow creation. */
138 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
139 
140 /* Activate DV E-Switch flow steering. */
141 #define MLX5_DV_ESW_EN "dv_esw_en"
142 
143 /* Activate DV flow steering. */
144 #define MLX5_DV_FLOW_EN "dv_flow_en"
145 
146 /* Enable extensive flow metadata support. */
147 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
148 
149 /* Device parameter to let the user manage the lacp traffic of bonded device */
150 #define MLX5_LACP_BY_USER "lacp_by_user"
151 
152 /* Activate Netlink support in VF mode. */
153 #define MLX5_VF_NL_EN "vf_nl_en"
154 
155 /* Enable extending memsegs when creating a MR. */
156 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
157 
158 /* Select port representors to instantiate. */
159 #define MLX5_REPRESENTOR "representor"
160 
161 /* Device parameter to configure the maximum number of dump files per queue. */
162 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
163 
164 /* Configure timeout of LRO session (in microseconds). */
165 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
166 
167 /*
168  * Device parameter to configure the total data buffer size for a single
169  * hairpin queue (logarithm value).
170  */
171 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
172 
173 /* Flow memory reclaim mode. */
174 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
175 
176 /* The default memory allocator used in PMD. */
177 #define MLX5_SYS_MEM_EN "sys_mem_en"
178 /* Decap will be used or not. */
179 #define MLX5_DECAP_EN "decap_en"
180 
181 /* Device parameter to configure allow or prevent duplicate rules pattern. */
182 #define MLX5_ALLOW_DUPLICATE_PATTERN "allow_duplicate_pattern"
183 
184 /* Device parameter to configure implicit registration of mempool memory. */
185 #define MLX5_MR_MEMPOOL_REG_EN "mr_mempool_reg_en"
186 
187 /* Device parameter to configure the delay drop when creating Rxqs. */
188 #define MLX5_DELAY_DROP "delay_drop"
189 
190 /* Shared memory between primary and secondary processes. */
191 struct mlx5_shared_data *mlx5_shared_data;
192 
193 /** Driver-specific log messages type. */
194 int mlx5_logtype;
195 
196 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
197 						LIST_HEAD_INITIALIZER();
198 static pthread_mutex_t mlx5_dev_ctx_list_mutex;
199 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
200 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
201 	[MLX5_IPOOL_DECAP_ENCAP] = {
202 		.size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
203 		.trunk_size = 64,
204 		.grow_trunk = 3,
205 		.grow_shift = 2,
206 		.need_lock = 1,
207 		.release_mem_en = 1,
208 		.malloc = mlx5_malloc,
209 		.free = mlx5_free,
210 		.type = "mlx5_encap_decap_ipool",
211 	},
212 	[MLX5_IPOOL_PUSH_VLAN] = {
213 		.size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
214 		.trunk_size = 64,
215 		.grow_trunk = 3,
216 		.grow_shift = 2,
217 		.need_lock = 1,
218 		.release_mem_en = 1,
219 		.malloc = mlx5_malloc,
220 		.free = mlx5_free,
221 		.type = "mlx5_push_vlan_ipool",
222 	},
223 	[MLX5_IPOOL_TAG] = {
224 		.size = sizeof(struct mlx5_flow_dv_tag_resource),
225 		.trunk_size = 64,
226 		.grow_trunk = 3,
227 		.grow_shift = 2,
228 		.need_lock = 1,
229 		.release_mem_en = 0,
230 		.per_core_cache = (1 << 16),
231 		.malloc = mlx5_malloc,
232 		.free = mlx5_free,
233 		.type = "mlx5_tag_ipool",
234 	},
235 	[MLX5_IPOOL_PORT_ID] = {
236 		.size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
237 		.trunk_size = 64,
238 		.grow_trunk = 3,
239 		.grow_shift = 2,
240 		.need_lock = 1,
241 		.release_mem_en = 1,
242 		.malloc = mlx5_malloc,
243 		.free = mlx5_free,
244 		.type = "mlx5_port_id_ipool",
245 	},
246 	[MLX5_IPOOL_JUMP] = {
247 		.size = sizeof(struct mlx5_flow_tbl_data_entry),
248 		.trunk_size = 64,
249 		.grow_trunk = 3,
250 		.grow_shift = 2,
251 		.need_lock = 1,
252 		.release_mem_en = 1,
253 		.malloc = mlx5_malloc,
254 		.free = mlx5_free,
255 		.type = "mlx5_jump_ipool",
256 	},
257 	[MLX5_IPOOL_SAMPLE] = {
258 		.size = sizeof(struct mlx5_flow_dv_sample_resource),
259 		.trunk_size = 64,
260 		.grow_trunk = 3,
261 		.grow_shift = 2,
262 		.need_lock = 1,
263 		.release_mem_en = 1,
264 		.malloc = mlx5_malloc,
265 		.free = mlx5_free,
266 		.type = "mlx5_sample_ipool",
267 	},
268 	[MLX5_IPOOL_DEST_ARRAY] = {
269 		.size = sizeof(struct mlx5_flow_dv_dest_array_resource),
270 		.trunk_size = 64,
271 		.grow_trunk = 3,
272 		.grow_shift = 2,
273 		.need_lock = 1,
274 		.release_mem_en = 1,
275 		.malloc = mlx5_malloc,
276 		.free = mlx5_free,
277 		.type = "mlx5_dest_array_ipool",
278 	},
279 	[MLX5_IPOOL_TUNNEL_ID] = {
280 		.size = sizeof(struct mlx5_flow_tunnel),
281 		.trunk_size = MLX5_MAX_TUNNELS,
282 		.need_lock = 1,
283 		.release_mem_en = 1,
284 		.type = "mlx5_tunnel_offload",
285 	},
286 	[MLX5_IPOOL_TNL_TBL_ID] = {
287 		.size = 0,
288 		.need_lock = 1,
289 		.type = "mlx5_flow_tnl_tbl_ipool",
290 	},
291 #endif
292 	[MLX5_IPOOL_MTR] = {
293 		/**
294 		 * The ipool index should grow continually from small to big,
295 		 * for meter idx, so not set grow_trunk to avoid meter index
296 		 * not jump continually.
297 		 */
298 		.size = sizeof(struct mlx5_legacy_flow_meter),
299 		.trunk_size = 64,
300 		.need_lock = 1,
301 		.release_mem_en = 1,
302 		.malloc = mlx5_malloc,
303 		.free = mlx5_free,
304 		.type = "mlx5_meter_ipool",
305 	},
306 	[MLX5_IPOOL_MCP] = {
307 		.size = sizeof(struct mlx5_flow_mreg_copy_resource),
308 		.trunk_size = 64,
309 		.grow_trunk = 3,
310 		.grow_shift = 2,
311 		.need_lock = 1,
312 		.release_mem_en = 1,
313 		.malloc = mlx5_malloc,
314 		.free = mlx5_free,
315 		.type = "mlx5_mcp_ipool",
316 	},
317 	[MLX5_IPOOL_HRXQ] = {
318 		.size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
319 		.trunk_size = 64,
320 		.grow_trunk = 3,
321 		.grow_shift = 2,
322 		.need_lock = 1,
323 		.release_mem_en = 1,
324 		.malloc = mlx5_malloc,
325 		.free = mlx5_free,
326 		.type = "mlx5_hrxq_ipool",
327 	},
328 	[MLX5_IPOOL_MLX5_FLOW] = {
329 		/*
330 		 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
331 		 * It set in run time according to PCI function configuration.
332 		 */
333 		.size = 0,
334 		.trunk_size = 64,
335 		.grow_trunk = 3,
336 		.grow_shift = 2,
337 		.need_lock = 1,
338 		.release_mem_en = 0,
339 		.per_core_cache = 1 << 19,
340 		.malloc = mlx5_malloc,
341 		.free = mlx5_free,
342 		.type = "mlx5_flow_handle_ipool",
343 	},
344 	[MLX5_IPOOL_RTE_FLOW] = {
345 		.size = sizeof(struct rte_flow),
346 		.trunk_size = 4096,
347 		.need_lock = 1,
348 		.release_mem_en = 1,
349 		.malloc = mlx5_malloc,
350 		.free = mlx5_free,
351 		.type = "rte_flow_ipool",
352 	},
353 	[MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
354 		.size = 0,
355 		.need_lock = 1,
356 		.type = "mlx5_flow_rss_id_ipool",
357 	},
358 	[MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
359 		.size = sizeof(struct mlx5_shared_action_rss),
360 		.trunk_size = 64,
361 		.grow_trunk = 3,
362 		.grow_shift = 2,
363 		.need_lock = 1,
364 		.release_mem_en = 1,
365 		.malloc = mlx5_malloc,
366 		.free = mlx5_free,
367 		.type = "mlx5_shared_action_rss",
368 	},
369 	[MLX5_IPOOL_MTR_POLICY] = {
370 		/**
371 		 * The ipool index should grow continually from small to big,
372 		 * for policy idx, so not set grow_trunk to avoid policy index
373 		 * not jump continually.
374 		 */
375 		.size = sizeof(struct mlx5_flow_meter_sub_policy),
376 		.trunk_size = 64,
377 		.need_lock = 1,
378 		.release_mem_en = 1,
379 		.malloc = mlx5_malloc,
380 		.free = mlx5_free,
381 		.type = "mlx5_meter_policy_ipool",
382 	},
383 };
384 
385 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
386 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
387 
388 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 1024
389 
390 /**
391  * Decide whether representor ID is a HPF(host PF) port on BF2.
392  *
393  * @param dev
394  *   Pointer to Ethernet device structure.
395  *
396  * @return
397  *   Non-zero if HPF, otherwise 0.
398  */
399 bool
400 mlx5_is_hpf(struct rte_eth_dev *dev)
401 {
402 	struct mlx5_priv *priv = dev->data->dev_private;
403 	uint16_t repr = MLX5_REPRESENTOR_REPR(priv->representor_id);
404 	int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
405 
406 	return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_VF &&
407 	       MLX5_REPRESENTOR_REPR(-1) == repr;
408 }
409 
410 /**
411  * Decide whether representor ID is a SF port representor.
412  *
413  * @param dev
414  *   Pointer to Ethernet device structure.
415  *
416  * @return
417  *   Non-zero if HPF, otherwise 0.
418  */
419 bool
420 mlx5_is_sf_repr(struct rte_eth_dev *dev)
421 {
422 	struct mlx5_priv *priv = dev->data->dev_private;
423 	int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
424 
425 	return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_SF;
426 }
427 
428 /**
429  * Initialize the ASO aging management structure.
430  *
431  * @param[in] sh
432  *   Pointer to mlx5_dev_ctx_shared object to free
433  *
434  * @return
435  *   0 on success, a negative errno value otherwise and rte_errno is set.
436  */
437 int
438 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
439 {
440 	int err;
441 
442 	if (sh->aso_age_mng)
443 		return 0;
444 	sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
445 				      RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
446 	if (!sh->aso_age_mng) {
447 		DRV_LOG(ERR, "aso_age_mng allocation was failed.");
448 		rte_errno = ENOMEM;
449 		return -ENOMEM;
450 	}
451 	err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_FLOW_HIT);
452 	if (err) {
453 		mlx5_free(sh->aso_age_mng);
454 		return -1;
455 	}
456 	rte_rwlock_init(&sh->aso_age_mng->resize_rwl);
457 	rte_spinlock_init(&sh->aso_age_mng->free_sl);
458 	LIST_INIT(&sh->aso_age_mng->free);
459 	return 0;
460 }
461 
462 /**
463  * Close and release all the resources of the ASO aging management structure.
464  *
465  * @param[in] sh
466  *   Pointer to mlx5_dev_ctx_shared object to free.
467  */
468 static void
469 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
470 {
471 	int i, j;
472 
473 	mlx5_aso_flow_hit_queue_poll_stop(sh);
474 	mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_FLOW_HIT);
475 	if (sh->aso_age_mng->pools) {
476 		struct mlx5_aso_age_pool *pool;
477 
478 		for (i = 0; i < sh->aso_age_mng->next; ++i) {
479 			pool = sh->aso_age_mng->pools[i];
480 			claim_zero(mlx5_devx_cmd_destroy
481 						(pool->flow_hit_aso_obj));
482 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
483 				if (pool->actions[j].dr_action)
484 					claim_zero
485 					    (mlx5_flow_os_destroy_flow_action
486 					      (pool->actions[j].dr_action));
487 			mlx5_free(pool);
488 		}
489 		mlx5_free(sh->aso_age_mng->pools);
490 	}
491 	mlx5_free(sh->aso_age_mng);
492 }
493 
494 /**
495  * Initialize the shared aging list information per port.
496  *
497  * @param[in] sh
498  *   Pointer to mlx5_dev_ctx_shared object.
499  */
500 static void
501 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
502 {
503 	uint32_t i;
504 	struct mlx5_age_info *age_info;
505 
506 	for (i = 0; i < sh->max_port; i++) {
507 		age_info = &sh->port[i].age_info;
508 		age_info->flags = 0;
509 		TAILQ_INIT(&age_info->aged_counters);
510 		LIST_INIT(&age_info->aged_aso);
511 		rte_spinlock_init(&age_info->aged_sl);
512 		MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
513 	}
514 }
515 
516 /**
517  * Initialize the counters management structure.
518  *
519  * @param[in] sh
520  *   Pointer to mlx5_dev_ctx_shared object to free
521  */
522 static void
523 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
524 {
525 	int i;
526 
527 	memset(&sh->cmng, 0, sizeof(sh->cmng));
528 	TAILQ_INIT(&sh->cmng.flow_counters);
529 	sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
530 	sh->cmng.max_id = -1;
531 	sh->cmng.last_pool_idx = POOL_IDX_INVALID;
532 	rte_spinlock_init(&sh->cmng.pool_update_sl);
533 	for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
534 		TAILQ_INIT(&sh->cmng.counters[i]);
535 		rte_spinlock_init(&sh->cmng.csl[i]);
536 	}
537 }
538 
539 /**
540  * Destroy all the resources allocated for a counter memory management.
541  *
542  * @param[in] mng
543  *   Pointer to the memory management structure.
544  */
545 static void
546 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
547 {
548 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
549 
550 	LIST_REMOVE(mng, next);
551 	mlx5_os_wrapped_mkey_destroy(&mng->wm);
552 	mlx5_free(mem);
553 }
554 
555 /**
556  * Close and release all the resources of the counters management.
557  *
558  * @param[in] sh
559  *   Pointer to mlx5_dev_ctx_shared object to free.
560  */
561 static void
562 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
563 {
564 	struct mlx5_counter_stats_mem_mng *mng;
565 	int i, j;
566 	int retries = 1024;
567 
568 	rte_errno = 0;
569 	while (--retries) {
570 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
571 		if (rte_errno != EINPROGRESS)
572 			break;
573 		rte_pause();
574 	}
575 
576 	if (sh->cmng.pools) {
577 		struct mlx5_flow_counter_pool *pool;
578 		uint16_t n_valid = sh->cmng.n_valid;
579 		bool fallback = sh->cmng.counter_fallback;
580 
581 		for (i = 0; i < n_valid; ++i) {
582 			pool = sh->cmng.pools[i];
583 			if (!fallback && pool->min_dcs)
584 				claim_zero(mlx5_devx_cmd_destroy
585 							       (pool->min_dcs));
586 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
587 				struct mlx5_flow_counter *cnt =
588 						MLX5_POOL_GET_CNT(pool, j);
589 
590 				if (cnt->action)
591 					claim_zero
592 					 (mlx5_flow_os_destroy_flow_action
593 					  (cnt->action));
594 				if (fallback && MLX5_POOL_GET_CNT
595 				    (pool, j)->dcs_when_free)
596 					claim_zero(mlx5_devx_cmd_destroy
597 						   (cnt->dcs_when_free));
598 			}
599 			mlx5_free(pool);
600 		}
601 		mlx5_free(sh->cmng.pools);
602 	}
603 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
604 	while (mng) {
605 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
606 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
607 	}
608 	memset(&sh->cmng, 0, sizeof(sh->cmng));
609 }
610 
611 /**
612  * Initialize the aso flow meters management structure.
613  *
614  * @param[in] sh
615  *   Pointer to mlx5_dev_ctx_shared object to free
616  */
617 int
618 mlx5_aso_flow_mtrs_mng_init(struct mlx5_dev_ctx_shared *sh)
619 {
620 	if (!sh->mtrmng) {
621 		sh->mtrmng = mlx5_malloc(MLX5_MEM_ZERO,
622 			sizeof(*sh->mtrmng),
623 			RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
624 		if (!sh->mtrmng) {
625 			DRV_LOG(ERR,
626 			"meter management allocation was failed.");
627 			rte_errno = ENOMEM;
628 			return -ENOMEM;
629 		}
630 		if (sh->meter_aso_en) {
631 			rte_spinlock_init(&sh->mtrmng->pools_mng.mtrsl);
632 			rte_rwlock_init(&sh->mtrmng->pools_mng.resize_mtrwl);
633 			LIST_INIT(&sh->mtrmng->pools_mng.meters);
634 		}
635 		sh->mtrmng->def_policy_id = MLX5_INVALID_POLICY_ID;
636 	}
637 	return 0;
638 }
639 
640 /**
641  * Close and release all the resources of
642  * the ASO flow meter management structure.
643  *
644  * @param[in] sh
645  *   Pointer to mlx5_dev_ctx_shared object to free.
646  */
647 static void
648 mlx5_aso_flow_mtrs_mng_close(struct mlx5_dev_ctx_shared *sh)
649 {
650 	struct mlx5_aso_mtr_pool *mtr_pool;
651 	struct mlx5_flow_mtr_mng *mtrmng = sh->mtrmng;
652 	uint32_t idx;
653 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
654 	struct mlx5_aso_mtr *aso_mtr;
655 	int i;
656 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
657 
658 	if (sh->meter_aso_en) {
659 		mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_POLICER);
660 		idx = mtrmng->pools_mng.n_valid;
661 		while (idx--) {
662 			mtr_pool = mtrmng->pools_mng.pools[idx];
663 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
664 			for (i = 0; i < MLX5_ASO_MTRS_PER_POOL; i++) {
665 				aso_mtr = &mtr_pool->mtrs[i];
666 				if (aso_mtr->fm.meter_action)
667 					claim_zero
668 					(mlx5_glue->destroy_flow_action
669 					(aso_mtr->fm.meter_action));
670 			}
671 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
672 			claim_zero(mlx5_devx_cmd_destroy
673 						(mtr_pool->devx_obj));
674 			mtrmng->pools_mng.n_valid--;
675 			mlx5_free(mtr_pool);
676 		}
677 		mlx5_free(sh->mtrmng->pools_mng.pools);
678 	}
679 	mlx5_free(sh->mtrmng);
680 	sh->mtrmng = NULL;
681 }
682 
683 /* Send FLOW_AGED event if needed. */
684 void
685 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
686 {
687 	struct mlx5_age_info *age_info;
688 	uint32_t i;
689 
690 	for (i = 0; i < sh->max_port; i++) {
691 		age_info = &sh->port[i].age_info;
692 		if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
693 			continue;
694 		MLX5_AGE_UNSET(age_info, MLX5_AGE_EVENT_NEW);
695 		if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER)) {
696 			MLX5_AGE_UNSET(age_info, MLX5_AGE_TRIGGER);
697 			rte_eth_dev_callback_process
698 				(&rte_eth_devices[sh->port[i].devx_ih_port_id],
699 				RTE_ETH_EVENT_FLOW_AGED, NULL);
700 		}
701 	}
702 }
703 
704 /*
705  * Initialize the ASO connection tracking structure.
706  *
707  * @param[in] sh
708  *   Pointer to mlx5_dev_ctx_shared object.
709  *
710  * @return
711  *   0 on success, a negative errno value otherwise and rte_errno is set.
712  */
713 int
714 mlx5_flow_aso_ct_mng_init(struct mlx5_dev_ctx_shared *sh)
715 {
716 	int err;
717 
718 	if (sh->ct_mng)
719 		return 0;
720 	sh->ct_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->ct_mng),
721 				 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
722 	if (!sh->ct_mng) {
723 		DRV_LOG(ERR, "ASO CT management allocation failed.");
724 		rte_errno = ENOMEM;
725 		return -rte_errno;
726 	}
727 	err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
728 	if (err) {
729 		mlx5_free(sh->ct_mng);
730 		/* rte_errno should be extracted from the failure. */
731 		rte_errno = EINVAL;
732 		return -rte_errno;
733 	}
734 	rte_spinlock_init(&sh->ct_mng->ct_sl);
735 	rte_rwlock_init(&sh->ct_mng->resize_rwl);
736 	LIST_INIT(&sh->ct_mng->free_cts);
737 	return 0;
738 }
739 
740 /*
741  * Close and release all the resources of the
742  * ASO connection tracking management structure.
743  *
744  * @param[in] sh
745  *   Pointer to mlx5_dev_ctx_shared object to free.
746  */
747 static void
748 mlx5_flow_aso_ct_mng_close(struct mlx5_dev_ctx_shared *sh)
749 {
750 	struct mlx5_aso_ct_pools_mng *mng = sh->ct_mng;
751 	struct mlx5_aso_ct_pool *ct_pool;
752 	struct mlx5_aso_ct_action *ct;
753 	uint32_t idx;
754 	uint32_t val;
755 	uint32_t cnt;
756 	int i;
757 
758 	mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
759 	idx = mng->next;
760 	while (idx--) {
761 		cnt = 0;
762 		ct_pool = mng->pools[idx];
763 		for (i = 0; i < MLX5_ASO_CT_ACTIONS_PER_POOL; i++) {
764 			ct = &ct_pool->actions[i];
765 			val = __atomic_fetch_sub(&ct->refcnt, 1,
766 						 __ATOMIC_RELAXED);
767 			MLX5_ASSERT(val == 1);
768 			if (val > 1)
769 				cnt++;
770 #ifdef HAVE_MLX5_DR_ACTION_ASO_CT
771 			if (ct->dr_action_orig)
772 				claim_zero(mlx5_glue->destroy_flow_action
773 							(ct->dr_action_orig));
774 			if (ct->dr_action_rply)
775 				claim_zero(mlx5_glue->destroy_flow_action
776 							(ct->dr_action_rply));
777 #endif
778 		}
779 		claim_zero(mlx5_devx_cmd_destroy(ct_pool->devx_obj));
780 		if (cnt) {
781 			DRV_LOG(DEBUG, "%u ASO CT objects are being used in the pool %u",
782 				cnt, i);
783 		}
784 		mlx5_free(ct_pool);
785 		/* in case of failure. */
786 		mng->next--;
787 	}
788 	mlx5_free(mng->pools);
789 	mlx5_free(mng);
790 	/* Management structure must be cleared to 0s during allocation. */
791 	sh->ct_mng = NULL;
792 }
793 
794 /**
795  * Initialize the flow resources' indexed mempool.
796  *
797  * @param[in] sh
798  *   Pointer to mlx5_dev_ctx_shared object.
799  * @param[in] config
800  *   Pointer to user dev config.
801  */
802 static void
803 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
804 		       const struct mlx5_dev_config *config)
805 {
806 	uint8_t i;
807 	struct mlx5_indexed_pool_config cfg;
808 
809 	for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
810 		cfg = mlx5_ipool_cfg[i];
811 		switch (i) {
812 		default:
813 			break;
814 		/*
815 		 * Set MLX5_IPOOL_MLX5_FLOW ipool size
816 		 * according to PCI function flow configuration.
817 		 */
818 		case MLX5_IPOOL_MLX5_FLOW:
819 			cfg.size = config->dv_flow_en ?
820 				sizeof(struct mlx5_flow_handle) :
821 				MLX5_FLOW_HANDLE_VERBS_SIZE;
822 			break;
823 		}
824 		if (config->reclaim_mode) {
825 			cfg.release_mem_en = 1;
826 			cfg.per_core_cache = 0;
827 		} else {
828 			cfg.release_mem_en = 0;
829 		}
830 		sh->ipool[i] = mlx5_ipool_create(&cfg);
831 	}
832 }
833 
834 
835 /**
836  * Release the flow resources' indexed mempool.
837  *
838  * @param[in] sh
839  *   Pointer to mlx5_dev_ctx_shared object.
840  */
841 static void
842 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
843 {
844 	uint8_t i;
845 
846 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
847 		mlx5_ipool_destroy(sh->ipool[i]);
848 	for (i = 0; i < MLX5_MAX_MODIFY_NUM; ++i)
849 		if (sh->mdh_ipools[i])
850 			mlx5_ipool_destroy(sh->mdh_ipools[i]);
851 }
852 
853 /*
854  * Check if dynamic flex parser for eCPRI already exists.
855  *
856  * @param dev
857  *   Pointer to Ethernet device structure.
858  *
859  * @return
860  *   true on exists, false on not.
861  */
862 bool
863 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
864 {
865 	struct mlx5_priv *priv = dev->data->dev_private;
866 	struct mlx5_ecpri_parser_profile *prf = &priv->sh->ecpri_parser;
867 
868 	return !!prf->obj;
869 }
870 
871 /*
872  * Allocation of a flex parser for eCPRI. Once created, this parser related
873  * resources will be held until the device is closed.
874  *
875  * @param dev
876  *   Pointer to Ethernet device structure.
877  *
878  * @return
879  *   0 on success, a negative errno value otherwise and rte_errno is set.
880  */
881 int
882 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
883 {
884 	struct mlx5_priv *priv = dev->data->dev_private;
885 	struct mlx5_ecpri_parser_profile *prf =	&priv->sh->ecpri_parser;
886 	struct mlx5_devx_graph_node_attr node = {
887 		.modify_field_select = 0,
888 	};
889 	uint32_t ids[8];
890 	int ret;
891 
892 	if (!priv->config.hca_attr.parse_graph_flex_node) {
893 		DRV_LOG(ERR, "Dynamic flex parser is not supported "
894 			"for device %s.", priv->dev_data->name);
895 		return -ENOTSUP;
896 	}
897 	node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
898 	/* 8 bytes now: 4B common header + 4B message body header. */
899 	node.header_length_base_value = 0x8;
900 	/* After MAC layer: Ether / VLAN. */
901 	node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
902 	/* Type of compared condition should be 0xAEFE in the L2 layer. */
903 	node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
904 	/* Sample #0: type in common header. */
905 	node.sample[0].flow_match_sample_en = 1;
906 	/* Fixed offset. */
907 	node.sample[0].flow_match_sample_offset_mode = 0x0;
908 	/* Only the 2nd byte will be used. */
909 	node.sample[0].flow_match_sample_field_base_offset = 0x0;
910 	/* Sample #1: message payload. */
911 	node.sample[1].flow_match_sample_en = 1;
912 	/* Fixed offset. */
913 	node.sample[1].flow_match_sample_offset_mode = 0x0;
914 	/*
915 	 * Only the first two bytes will be used right now, and its offset will
916 	 * start after the common header that with the length of a DW(u32).
917 	 */
918 	node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
919 	prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->cdev->ctx, &node);
920 	if (!prf->obj) {
921 		DRV_LOG(ERR, "Failed to create flex parser node object.");
922 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
923 	}
924 	prf->num = 2;
925 	ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
926 	if (ret) {
927 		DRV_LOG(ERR, "Failed to query sample IDs.");
928 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
929 	}
930 	prf->offset[0] = 0x0;
931 	prf->offset[1] = sizeof(uint32_t);
932 	prf->ids[0] = ids[0];
933 	prf->ids[1] = ids[1];
934 	return 0;
935 }
936 
937 /*
938  * Destroy the flex parser node, including the parser itself, input / output
939  * arcs and DW samples. Resources could be reused then.
940  *
941  * @param dev
942  *   Pointer to Ethernet device structure.
943  */
944 static void
945 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
946 {
947 	struct mlx5_priv *priv = dev->data->dev_private;
948 	struct mlx5_ecpri_parser_profile *prf =	&priv->sh->ecpri_parser;
949 
950 	if (prf->obj)
951 		mlx5_devx_cmd_destroy(prf->obj);
952 	prf->obj = NULL;
953 }
954 
955 uint32_t
956 mlx5_get_supported_sw_parsing_offloads(const struct mlx5_hca_attr *attr)
957 {
958 	uint32_t sw_parsing_offloads = 0;
959 
960 	if (attr->swp) {
961 		sw_parsing_offloads |= MLX5_SW_PARSING_CAP;
962 		if (attr->swp_csum)
963 			sw_parsing_offloads |= MLX5_SW_PARSING_CSUM_CAP;
964 
965 		if (attr->swp_lso)
966 			sw_parsing_offloads |= MLX5_SW_PARSING_TSO_CAP;
967 	}
968 	return sw_parsing_offloads;
969 }
970 
971 uint32_t
972 mlx5_get_supported_tunneling_offloads(const struct mlx5_hca_attr *attr)
973 {
974 	uint32_t tn_offloads = 0;
975 
976 	if (attr->tunnel_stateless_vxlan)
977 		tn_offloads |= MLX5_TUNNELED_OFFLOADS_VXLAN_CAP;
978 	if (attr->tunnel_stateless_gre)
979 		tn_offloads |= MLX5_TUNNELED_OFFLOADS_GRE_CAP;
980 	if (attr->tunnel_stateless_geneve_rx)
981 		tn_offloads |= MLX5_TUNNELED_OFFLOADS_GENEVE_CAP;
982 	return tn_offloads;
983 }
984 
985 /* Fill all fields of UAR structure. */
986 static int
987 mlx5_rxtx_uars_prepare(struct mlx5_dev_ctx_shared *sh)
988 {
989 	int ret;
990 
991 	ret = mlx5_devx_uar_prepare(sh->cdev, &sh->tx_uar);
992 	if (ret) {
993 		DRV_LOG(ERR, "Failed to prepare Tx DevX UAR.");
994 		return -rte_errno;
995 	}
996 	MLX5_ASSERT(sh->tx_uar.obj);
997 	MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar.obj));
998 	ret = mlx5_devx_uar_prepare(sh->cdev, &sh->rx_uar);
999 	if (ret) {
1000 		DRV_LOG(ERR, "Failed to prepare Rx DevX UAR.");
1001 		mlx5_devx_uar_release(&sh->tx_uar);
1002 		return -rte_errno;
1003 	}
1004 	MLX5_ASSERT(sh->rx_uar.obj);
1005 	MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->rx_uar.obj));
1006 	return 0;
1007 }
1008 
1009 static void
1010 mlx5_rxtx_uars_release(struct mlx5_dev_ctx_shared *sh)
1011 {
1012 	mlx5_devx_uar_release(&sh->rx_uar);
1013 	mlx5_devx_uar_release(&sh->tx_uar);
1014 }
1015 
1016 /**
1017  * rte_mempool_walk() callback to unregister Rx mempools.
1018  * It used when implicit mempool registration is disabled.
1019  *
1020  * @param mp
1021  *   The mempool being walked.
1022  * @param arg
1023  *   Pointer to the device shared context.
1024  */
1025 static void
1026 mlx5_dev_ctx_shared_rx_mempool_unregister_cb(struct rte_mempool *mp, void *arg)
1027 {
1028 	struct mlx5_dev_ctx_shared *sh = arg;
1029 
1030 	mlx5_dev_mempool_unregister(sh->cdev, mp);
1031 }
1032 
1033 /**
1034  * Callback used when implicit mempool registration is disabled
1035  * in order to track Rx mempool destruction.
1036  *
1037  * @param event
1038  *   Mempool life cycle event.
1039  * @param mp
1040  *   An Rx mempool registered explicitly when the port is started.
1041  * @param arg
1042  *   Pointer to a device shared context.
1043  */
1044 static void
1045 mlx5_dev_ctx_shared_rx_mempool_event_cb(enum rte_mempool_event event,
1046 					struct rte_mempool *mp, void *arg)
1047 {
1048 	struct mlx5_dev_ctx_shared *sh = arg;
1049 
1050 	if (event == RTE_MEMPOOL_EVENT_DESTROY)
1051 		mlx5_dev_mempool_unregister(sh->cdev, mp);
1052 }
1053 
1054 int
1055 mlx5_dev_ctx_shared_mempool_subscribe(struct rte_eth_dev *dev)
1056 {
1057 	struct mlx5_priv *priv = dev->data->dev_private;
1058 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1059 	int ret;
1060 
1061 	/* Check if we only need to track Rx mempool destruction. */
1062 	if (!sh->cdev->config.mr_mempool_reg_en) {
1063 		ret = rte_mempool_event_callback_register
1064 				(mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1065 		return ret == 0 || rte_errno == EEXIST ? 0 : ret;
1066 	}
1067 	return mlx5_dev_mempool_subscribe(sh->cdev);
1068 }
1069 
1070 /**
1071  * Set up multiple TISs with different affinities according to
1072  * number of bonding ports
1073  *
1074  * @param priv
1075  * Pointer of shared context.
1076  *
1077  * @return
1078  * Zero on success, -1 otherwise.
1079  */
1080 static int
1081 mlx5_setup_tis(struct mlx5_dev_ctx_shared *sh)
1082 {
1083 	int i;
1084 	struct mlx5_devx_lag_context lag_ctx = { 0 };
1085 	struct mlx5_devx_tis_attr tis_attr = { 0 };
1086 
1087 	tis_attr.transport_domain = sh->td->id;
1088 	if (sh->bond.n_port) {
1089 		if (!mlx5_devx_cmd_query_lag(sh->cdev->ctx, &lag_ctx)) {
1090 			sh->lag.tx_remap_affinity[0] =
1091 				lag_ctx.tx_remap_affinity_1;
1092 			sh->lag.tx_remap_affinity[1] =
1093 				lag_ctx.tx_remap_affinity_2;
1094 			sh->lag.affinity_mode = lag_ctx.port_select_mode;
1095 		} else {
1096 			DRV_LOG(ERR, "Failed to query lag affinity.");
1097 			return -1;
1098 		}
1099 		if (sh->lag.affinity_mode == MLX5_LAG_MODE_TIS) {
1100 			for (i = 0; i < sh->bond.n_port; i++) {
1101 				tis_attr.lag_tx_port_affinity =
1102 					MLX5_IFC_LAG_MAP_TIS_AFFINITY(i,
1103 							sh->bond.n_port);
1104 				sh->tis[i] = mlx5_devx_cmd_create_tis(sh->cdev->ctx,
1105 						&tis_attr);
1106 				if (!sh->tis[i]) {
1107 					DRV_LOG(ERR, "Failed to TIS %d/%d for bonding device"
1108 						" %s.", i, sh->bond.n_port,
1109 						sh->ibdev_name);
1110 					return -1;
1111 				}
1112 			}
1113 			DRV_LOG(DEBUG, "LAG number of ports : %d, affinity_1 & 2 : pf%d & %d.\n",
1114 				sh->bond.n_port, lag_ctx.tx_remap_affinity_1,
1115 				lag_ctx.tx_remap_affinity_2);
1116 			return 0;
1117 		}
1118 		if (sh->lag.affinity_mode == MLX5_LAG_MODE_HASH)
1119 			DRV_LOG(INFO, "Device %s enabled HW hash based LAG.",
1120 					sh->ibdev_name);
1121 	}
1122 	tis_attr.lag_tx_port_affinity = 0;
1123 	sh->tis[0] = mlx5_devx_cmd_create_tis(sh->cdev->ctx, &tis_attr);
1124 	if (!sh->tis[0]) {
1125 		DRV_LOG(ERR, "Failed to TIS 0 for bonding device"
1126 			" %s.", sh->ibdev_name);
1127 		return -1;
1128 	}
1129 	return 0;
1130 }
1131 
1132 /**
1133  * Allocate shared device context. If there is multiport device the
1134  * master and representors will share this context, if there is single
1135  * port dedicated device, the context will be used by only given
1136  * port due to unification.
1137  *
1138  * Routine first searches the context for the specified device name,
1139  * if found the shared context assumed and reference counter is incremented.
1140  * If no context found the new one is created and initialized with specified
1141  * device context and parameters.
1142  *
1143  * @param[in] spawn
1144  *   Pointer to the device attributes (name, port, etc).
1145  * @param[in] config
1146  *   Pointer to device configuration structure.
1147  *
1148  * @return
1149  *   Pointer to mlx5_dev_ctx_shared object on success,
1150  *   otherwise NULL and rte_errno is set.
1151  */
1152 struct mlx5_dev_ctx_shared *
1153 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
1154 			  const struct mlx5_dev_config *config)
1155 {
1156 	struct mlx5_dev_ctx_shared *sh;
1157 	int err = 0;
1158 	uint32_t i;
1159 
1160 	MLX5_ASSERT(spawn);
1161 	/* Secondary process should not create the shared context. */
1162 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1163 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1164 	/* Search for IB context by device name. */
1165 	LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
1166 		if (!strcmp(sh->ibdev_name, spawn->phys_dev_name)) {
1167 			sh->refcnt++;
1168 			goto exit;
1169 		}
1170 	}
1171 	/* No device found, we have to create new shared context. */
1172 	MLX5_ASSERT(spawn->max_port);
1173 	sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1174 			 sizeof(struct mlx5_dev_ctx_shared) +
1175 			 spawn->max_port *
1176 			 sizeof(struct mlx5_dev_shared_port),
1177 			 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1178 	if (!sh) {
1179 		DRV_LOG(ERR, "shared context allocation failure");
1180 		rte_errno  = ENOMEM;
1181 		goto exit;
1182 	}
1183 	pthread_mutex_init(&sh->txpp.mutex, NULL);
1184 	sh->numa_node = spawn->cdev->dev->numa_node;
1185 	sh->cdev = spawn->cdev;
1186 	sh->devx = sh->cdev->config.devx;
1187 	if (spawn->bond_info)
1188 		sh->bond = *spawn->bond_info;
1189 	err = mlx5_os_get_dev_attr(sh->cdev, &sh->device_attr);
1190 	if (err) {
1191 		DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
1192 		goto error;
1193 	}
1194 	sh->refcnt = 1;
1195 	sh->max_port = spawn->max_port;
1196 	sh->reclaim_mode = config->reclaim_mode;
1197 	strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->cdev->ctx),
1198 		sizeof(sh->ibdev_name) - 1);
1199 	strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->cdev->ctx),
1200 		sizeof(sh->ibdev_path) - 1);
1201 	/*
1202 	 * Setting port_id to max unallowed value means
1203 	 * there is no interrupt subhandler installed for
1204 	 * the given port index i.
1205 	 */
1206 	for (i = 0; i < sh->max_port; i++) {
1207 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
1208 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
1209 	}
1210 	if (sh->devx) {
1211 		sh->td = mlx5_devx_cmd_create_td(sh->cdev->ctx);
1212 		if (!sh->td) {
1213 			DRV_LOG(ERR, "TD allocation failure");
1214 			err = ENOMEM;
1215 			goto error;
1216 		}
1217 		if (mlx5_setup_tis(sh)) {
1218 			DRV_LOG(ERR, "TIS allocation failure");
1219 			err = ENOMEM;
1220 			goto error;
1221 		}
1222 		err = mlx5_rxtx_uars_prepare(sh);
1223 		if (err)
1224 			goto error;
1225 #ifndef RTE_ARCH_64
1226 	} else {
1227 		/* Initialize UAR access locks for 32bit implementations. */
1228 		rte_spinlock_init(&sh->uar_lock_cq);
1229 		for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1230 			rte_spinlock_init(&sh->uar_lock[i]);
1231 #endif
1232 	}
1233 	mlx5_os_dev_shared_handler_install(sh);
1234 	if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1235 		err = mlx5_flow_os_init_workspace_once();
1236 		if (err)
1237 			goto error;
1238 	}
1239 	mlx5_flow_aging_init(sh);
1240 	mlx5_flow_counters_mng_init(sh);
1241 	mlx5_flow_ipool_create(sh, config);
1242 	/* Add context to the global device list. */
1243 	LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1244 	rte_spinlock_init(&sh->geneve_tlv_opt_sl);
1245 exit:
1246 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1247 	return sh;
1248 error:
1249 	pthread_mutex_destroy(&sh->txpp.mutex);
1250 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1251 	MLX5_ASSERT(sh);
1252 	if (sh->td)
1253 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1254 	i = 0;
1255 	do {
1256 		if (sh->tis[i])
1257 			claim_zero(mlx5_devx_cmd_destroy(sh->tis[i]));
1258 	} while (++i < (uint32_t)sh->bond.n_port);
1259 	mlx5_rxtx_uars_release(sh);
1260 	mlx5_free(sh);
1261 	MLX5_ASSERT(err > 0);
1262 	rte_errno = err;
1263 	return NULL;
1264 }
1265 
1266 /**
1267  * Free shared IB device context. Decrement counter and if zero free
1268  * all allocated resources and close handles.
1269  *
1270  * @param[in] sh
1271  *   Pointer to mlx5_dev_ctx_shared object to free
1272  */
1273 void
1274 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1275 {
1276 	int ret;
1277 	int i = 0;
1278 
1279 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1280 #ifdef RTE_LIBRTE_MLX5_DEBUG
1281 	/* Check the object presence in the list. */
1282 	struct mlx5_dev_ctx_shared *lctx;
1283 
1284 	LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1285 		if (lctx == sh)
1286 			break;
1287 	MLX5_ASSERT(lctx);
1288 	if (lctx != sh) {
1289 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
1290 		goto exit;
1291 	}
1292 #endif
1293 	MLX5_ASSERT(sh);
1294 	MLX5_ASSERT(sh->refcnt);
1295 	/* Secondary process should not free the shared context. */
1296 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1297 	if (--sh->refcnt)
1298 		goto exit;
1299 	/* Stop watching for mempool events and unregister all mempools. */
1300 	if (!sh->cdev->config.mr_mempool_reg_en) {
1301 		ret = rte_mempool_event_callback_unregister
1302 				(mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1303 		if (ret == 0)
1304 			rte_mempool_walk
1305 			     (mlx5_dev_ctx_shared_rx_mempool_unregister_cb, sh);
1306 	}
1307 	/* Remove context from the global device list. */
1308 	LIST_REMOVE(sh, next);
1309 	/* Release resources on the last device removal. */
1310 	if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1311 		mlx5_os_net_cleanup();
1312 		mlx5_flow_os_release_workspace();
1313 	}
1314 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1315 	if (sh->flex_parsers_dv) {
1316 		mlx5_list_destroy(sh->flex_parsers_dv);
1317 		sh->flex_parsers_dv = NULL;
1318 	}
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 	mlx5_rxtx_uars_release(sh);
1333 	do {
1334 		if (sh->tis[i])
1335 			claim_zero(mlx5_devx_cmd_destroy(sh->tis[i]));
1336 	} while (++i < sh->bond.n_port);
1337 	if (sh->td)
1338 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1339 	MLX5_ASSERT(sh->geneve_tlv_option_resource == NULL);
1340 	pthread_mutex_destroy(&sh->txpp.mutex);
1341 	mlx5_free(sh);
1342 	return;
1343 exit:
1344 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1345 }
1346 
1347 /**
1348  * Destroy table hash list.
1349  *
1350  * @param[in] priv
1351  *   Pointer to the private device data structure.
1352  */
1353 void
1354 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1355 {
1356 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1357 
1358 	if (!sh->flow_tbls)
1359 		return;
1360 	mlx5_hlist_destroy(sh->flow_tbls);
1361 	sh->flow_tbls = NULL;
1362 }
1363 
1364 /**
1365  * Initialize flow table hash list and create the root tables entry
1366  * for each domain.
1367  *
1368  * @param[in] priv
1369  *   Pointer to the private device data structure.
1370  *
1371  * @return
1372  *   Zero on success, positive error code otherwise.
1373  */
1374 int
1375 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1376 {
1377 	int err = 0;
1378 	/* Tables are only used in DV and DR modes. */
1379 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1380 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1381 	char s[MLX5_NAME_SIZE];
1382 
1383 	MLX5_ASSERT(sh);
1384 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1385 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1386 					  false, true, sh,
1387 					  flow_dv_tbl_create_cb,
1388 					  flow_dv_tbl_match_cb,
1389 					  flow_dv_tbl_remove_cb,
1390 					  flow_dv_tbl_clone_cb,
1391 					  flow_dv_tbl_clone_free_cb);
1392 	if (!sh->flow_tbls) {
1393 		DRV_LOG(ERR, "flow tables with hash creation failed.");
1394 		err = ENOMEM;
1395 		return err;
1396 	}
1397 #ifndef HAVE_MLX5DV_DR
1398 	struct rte_flow_error error;
1399 	struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1400 
1401 	/*
1402 	 * In case we have not DR support, the zero tables should be created
1403 	 * because DV expect to see them even if they cannot be created by
1404 	 * RDMA-CORE.
1405 	 */
1406 	if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0,
1407 		NULL, 0, 1, 0, &error) ||
1408 	    !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0,
1409 		NULL, 0, 1, 0, &error) ||
1410 	    !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0,
1411 		NULL, 0, 1, 0, &error)) {
1412 		err = ENOMEM;
1413 		goto error;
1414 	}
1415 	return err;
1416 error:
1417 	mlx5_free_table_hash_list(priv);
1418 #endif /* HAVE_MLX5DV_DR */
1419 #endif
1420 	return err;
1421 }
1422 
1423 /**
1424  * Retrieve integer value from environment variable.
1425  *
1426  * @param[in] name
1427  *   Environment variable name.
1428  *
1429  * @return
1430  *   Integer value, 0 if the variable is not set.
1431  */
1432 int
1433 mlx5_getenv_int(const char *name)
1434 {
1435 	const char *val = getenv(name);
1436 
1437 	if (val == NULL)
1438 		return 0;
1439 	return atoi(val);
1440 }
1441 
1442 /**
1443  * DPDK callback to add udp tunnel port
1444  *
1445  * @param[in] dev
1446  *   A pointer to eth_dev
1447  * @param[in] udp_tunnel
1448  *   A pointer to udp tunnel
1449  *
1450  * @return
1451  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1452  */
1453 int
1454 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1455 			 struct rte_eth_udp_tunnel *udp_tunnel)
1456 {
1457 	MLX5_ASSERT(udp_tunnel != NULL);
1458 	if (udp_tunnel->prot_type == RTE_ETH_TUNNEL_TYPE_VXLAN &&
1459 	    udp_tunnel->udp_port == 4789)
1460 		return 0;
1461 	if (udp_tunnel->prot_type == RTE_ETH_TUNNEL_TYPE_VXLAN_GPE &&
1462 	    udp_tunnel->udp_port == 4790)
1463 		return 0;
1464 	return -ENOTSUP;
1465 }
1466 
1467 /**
1468  * Initialize process private data structure.
1469  *
1470  * @param dev
1471  *   Pointer to Ethernet device structure.
1472  *
1473  * @return
1474  *   0 on success, a negative errno value otherwise and rte_errno is set.
1475  */
1476 int
1477 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1478 {
1479 	struct mlx5_priv *priv = dev->data->dev_private;
1480 	struct mlx5_proc_priv *ppriv;
1481 	size_t ppriv_size;
1482 
1483 	mlx5_proc_priv_uninit(dev);
1484 	/*
1485 	 * UAR register table follows the process private structure. BlueFlame
1486 	 * registers for Tx queues are stored in the table.
1487 	 */
1488 	ppriv_size = sizeof(struct mlx5_proc_priv) +
1489 		     priv->txqs_n * sizeof(struct mlx5_uar_data);
1490 	ppriv = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, ppriv_size,
1491 			    RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1492 	if (!ppriv) {
1493 		rte_errno = ENOMEM;
1494 		return -rte_errno;
1495 	}
1496 	ppriv->uar_table_sz = priv->txqs_n;
1497 	dev->process_private = ppriv;
1498 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1499 		priv->sh->pppriv = ppriv;
1500 	return 0;
1501 }
1502 
1503 /**
1504  * Un-initialize process private data structure.
1505  *
1506  * @param dev
1507  *   Pointer to Ethernet device structure.
1508  */
1509 void
1510 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1511 {
1512 	if (!dev->process_private)
1513 		return;
1514 	mlx5_free(dev->process_private);
1515 	dev->process_private = NULL;
1516 }
1517 
1518 /**
1519  * DPDK callback to close the device.
1520  *
1521  * Destroy all queues and objects, free memory.
1522  *
1523  * @param dev
1524  *   Pointer to Ethernet device structure.
1525  */
1526 int
1527 mlx5_dev_close(struct rte_eth_dev *dev)
1528 {
1529 	struct mlx5_priv *priv = dev->data->dev_private;
1530 	unsigned int i;
1531 	int ret;
1532 
1533 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1534 		/* Check if process_private released. */
1535 		if (!dev->process_private)
1536 			return 0;
1537 		mlx5_tx_uar_uninit_secondary(dev);
1538 		mlx5_proc_priv_uninit(dev);
1539 		rte_eth_dev_release_port(dev);
1540 		return 0;
1541 	}
1542 	if (!priv->sh)
1543 		return 0;
1544 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1545 		dev->data->port_id,
1546 		((priv->sh->cdev->ctx != NULL) ?
1547 		mlx5_os_get_ctx_device_name(priv->sh->cdev->ctx) : ""));
1548 	/*
1549 	 * If default mreg copy action is removed at the stop stage,
1550 	 * the search will return none and nothing will be done anymore.
1551 	 */
1552 	mlx5_flow_stop_default(dev);
1553 	mlx5_traffic_disable(dev);
1554 	/*
1555 	 * If all the flows are already flushed in the device stop stage,
1556 	 * then this will return directly without any action.
1557 	 */
1558 	mlx5_flow_list_flush(dev, MLX5_FLOW_TYPE_GEN, true);
1559 	mlx5_action_handle_flush(dev);
1560 	mlx5_flow_meter_flush(dev, NULL);
1561 	/* Prevent crashes when queues are still in use. */
1562 	dev->rx_pkt_burst = removed_rx_burst;
1563 	dev->tx_pkt_burst = removed_tx_burst;
1564 	rte_wmb();
1565 	/* Disable datapath on secondary process. */
1566 	mlx5_mp_os_req_stop_rxtx(dev);
1567 	/* Free the eCPRI flex parser resource. */
1568 	mlx5_flex_parser_ecpri_release(dev);
1569 	mlx5_flex_item_port_cleanup(dev);
1570 	if (priv->rxq_privs != NULL) {
1571 		/* XXX race condition if mlx5_rx_burst() is still running. */
1572 		rte_delay_us_sleep(1000);
1573 		for (i = 0; (i != priv->rxqs_n); ++i)
1574 			mlx5_rxq_release(dev, i);
1575 		priv->rxqs_n = 0;
1576 		mlx5_free(priv->rxq_privs);
1577 		priv->rxq_privs = NULL;
1578 	}
1579 	if (priv->txqs != NULL) {
1580 		/* XXX race condition if mlx5_tx_burst() is still running. */
1581 		rte_delay_us_sleep(1000);
1582 		for (i = 0; (i != priv->txqs_n); ++i)
1583 			mlx5_txq_release(dev, i);
1584 		priv->txqs_n = 0;
1585 		priv->txqs = NULL;
1586 	}
1587 	mlx5_proc_priv_uninit(dev);
1588 	if (priv->q_counters) {
1589 		mlx5_devx_cmd_destroy(priv->q_counters);
1590 		priv->q_counters = NULL;
1591 	}
1592 	if (priv->drop_queue.hrxq)
1593 		mlx5_drop_action_destroy(dev);
1594 	if (priv->mreg_cp_tbl)
1595 		mlx5_hlist_destroy(priv->mreg_cp_tbl);
1596 	mlx5_mprq_free_mp(dev);
1597 	if (priv->sh->ct_mng)
1598 		mlx5_flow_aso_ct_mng_close(priv->sh);
1599 	mlx5_os_free_shared_dr(priv);
1600 	if (priv->rss_conf.rss_key != NULL)
1601 		mlx5_free(priv->rss_conf.rss_key);
1602 	if (priv->reta_idx != NULL)
1603 		mlx5_free(priv->reta_idx);
1604 	if (priv->config.vf)
1605 		mlx5_os_mac_addr_flush(dev);
1606 	if (priv->nl_socket_route >= 0)
1607 		close(priv->nl_socket_route);
1608 	if (priv->nl_socket_rdma >= 0)
1609 		close(priv->nl_socket_rdma);
1610 	if (priv->vmwa_context)
1611 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1612 	ret = mlx5_hrxq_verify(dev);
1613 	if (ret)
1614 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1615 			dev->data->port_id);
1616 	ret = mlx5_ind_table_obj_verify(dev);
1617 	if (ret)
1618 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1619 			dev->data->port_id);
1620 	ret = mlx5_rxq_obj_verify(dev);
1621 	if (ret)
1622 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1623 			dev->data->port_id);
1624 	ret = mlx5_rxq_verify(dev);
1625 	if (ret)
1626 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1627 			dev->data->port_id);
1628 	ret = mlx5_txq_obj_verify(dev);
1629 	if (ret)
1630 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1631 			dev->data->port_id);
1632 	ret = mlx5_txq_verify(dev);
1633 	if (ret)
1634 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1635 			dev->data->port_id);
1636 	ret = mlx5_flow_verify(dev);
1637 	if (ret)
1638 		DRV_LOG(WARNING, "port %u some flows still remain",
1639 			dev->data->port_id);
1640 	if (priv->hrxqs)
1641 		mlx5_list_destroy(priv->hrxqs);
1642 	/*
1643 	 * Free the shared context in last turn, because the cleanup
1644 	 * routines above may use some shared fields, like
1645 	 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1646 	 * ifindex if Netlink fails.
1647 	 */
1648 	mlx5_free_shared_dev_ctx(priv->sh);
1649 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1650 		unsigned int c = 0;
1651 		uint16_t port_id;
1652 
1653 		MLX5_ETH_FOREACH_DEV(port_id, dev->device) {
1654 			struct mlx5_priv *opriv =
1655 				rte_eth_devices[port_id].data->dev_private;
1656 
1657 			if (!opriv ||
1658 			    opriv->domain_id != priv->domain_id ||
1659 			    &rte_eth_devices[port_id] == dev)
1660 				continue;
1661 			++c;
1662 			break;
1663 		}
1664 		if (!c)
1665 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1666 	}
1667 	memset(priv, 0, sizeof(*priv));
1668 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1669 	/*
1670 	 * Reset mac_addrs to NULL such that it is not freed as part of
1671 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1672 	 * it is freed when dev_private is freed.
1673 	 */
1674 	dev->data->mac_addrs = NULL;
1675 	return 0;
1676 }
1677 
1678 const struct eth_dev_ops mlx5_dev_ops = {
1679 	.dev_configure = mlx5_dev_configure,
1680 	.dev_start = mlx5_dev_start,
1681 	.dev_stop = mlx5_dev_stop,
1682 	.dev_set_link_down = mlx5_set_link_down,
1683 	.dev_set_link_up = mlx5_set_link_up,
1684 	.dev_close = mlx5_dev_close,
1685 	.promiscuous_enable = mlx5_promiscuous_enable,
1686 	.promiscuous_disable = mlx5_promiscuous_disable,
1687 	.allmulticast_enable = mlx5_allmulticast_enable,
1688 	.allmulticast_disable = mlx5_allmulticast_disable,
1689 	.link_update = mlx5_link_update,
1690 	.stats_get = mlx5_stats_get,
1691 	.stats_reset = mlx5_stats_reset,
1692 	.xstats_get = mlx5_xstats_get,
1693 	.xstats_reset = mlx5_xstats_reset,
1694 	.xstats_get_names = mlx5_xstats_get_names,
1695 	.fw_version_get = mlx5_fw_version_get,
1696 	.dev_infos_get = mlx5_dev_infos_get,
1697 	.representor_info_get = mlx5_representor_info_get,
1698 	.read_clock = mlx5_txpp_read_clock,
1699 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1700 	.vlan_filter_set = mlx5_vlan_filter_set,
1701 	.rx_queue_setup = mlx5_rx_queue_setup,
1702 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1703 	.tx_queue_setup = mlx5_tx_queue_setup,
1704 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1705 	.rx_queue_release = mlx5_rx_queue_release,
1706 	.tx_queue_release = mlx5_tx_queue_release,
1707 	.rx_queue_start = mlx5_rx_queue_start,
1708 	.rx_queue_stop = mlx5_rx_queue_stop,
1709 	.tx_queue_start = mlx5_tx_queue_start,
1710 	.tx_queue_stop = mlx5_tx_queue_stop,
1711 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1712 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1713 	.mac_addr_remove = mlx5_mac_addr_remove,
1714 	.mac_addr_add = mlx5_mac_addr_add,
1715 	.mac_addr_set = mlx5_mac_addr_set,
1716 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1717 	.mtu_set = mlx5_dev_set_mtu,
1718 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1719 	.vlan_offload_set = mlx5_vlan_offload_set,
1720 	.reta_update = mlx5_dev_rss_reta_update,
1721 	.reta_query = mlx5_dev_rss_reta_query,
1722 	.rss_hash_update = mlx5_rss_hash_update,
1723 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1724 	.flow_ops_get = mlx5_flow_ops_get,
1725 	.rxq_info_get = mlx5_rxq_info_get,
1726 	.txq_info_get = mlx5_txq_info_get,
1727 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1728 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1729 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1730 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1731 	.is_removed = mlx5_is_removed,
1732 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1733 	.get_module_info = mlx5_get_module_info,
1734 	.get_module_eeprom = mlx5_get_module_eeprom,
1735 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1736 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1737 	.hairpin_bind = mlx5_hairpin_bind,
1738 	.hairpin_unbind = mlx5_hairpin_unbind,
1739 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1740 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1741 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1742 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1743 	.get_monitor_addr = mlx5_get_monitor_addr,
1744 };
1745 
1746 /* Available operations from secondary process. */
1747 const struct eth_dev_ops mlx5_dev_sec_ops = {
1748 	.stats_get = mlx5_stats_get,
1749 	.stats_reset = mlx5_stats_reset,
1750 	.xstats_get = mlx5_xstats_get,
1751 	.xstats_reset = mlx5_xstats_reset,
1752 	.xstats_get_names = mlx5_xstats_get_names,
1753 	.fw_version_get = mlx5_fw_version_get,
1754 	.dev_infos_get = mlx5_dev_infos_get,
1755 	.representor_info_get = mlx5_representor_info_get,
1756 	.read_clock = mlx5_txpp_read_clock,
1757 	.rx_queue_start = mlx5_rx_queue_start,
1758 	.rx_queue_stop = mlx5_rx_queue_stop,
1759 	.tx_queue_start = mlx5_tx_queue_start,
1760 	.tx_queue_stop = mlx5_tx_queue_stop,
1761 	.rxq_info_get = mlx5_rxq_info_get,
1762 	.txq_info_get = mlx5_txq_info_get,
1763 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1764 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1765 	.get_module_info = mlx5_get_module_info,
1766 	.get_module_eeprom = mlx5_get_module_eeprom,
1767 };
1768 
1769 /* Available operations in flow isolated mode. */
1770 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1771 	.dev_configure = mlx5_dev_configure,
1772 	.dev_start = mlx5_dev_start,
1773 	.dev_stop = mlx5_dev_stop,
1774 	.dev_set_link_down = mlx5_set_link_down,
1775 	.dev_set_link_up = mlx5_set_link_up,
1776 	.dev_close = mlx5_dev_close,
1777 	.promiscuous_enable = mlx5_promiscuous_enable,
1778 	.promiscuous_disable = mlx5_promiscuous_disable,
1779 	.allmulticast_enable = mlx5_allmulticast_enable,
1780 	.allmulticast_disable = mlx5_allmulticast_disable,
1781 	.link_update = mlx5_link_update,
1782 	.stats_get = mlx5_stats_get,
1783 	.stats_reset = mlx5_stats_reset,
1784 	.xstats_get = mlx5_xstats_get,
1785 	.xstats_reset = mlx5_xstats_reset,
1786 	.xstats_get_names = mlx5_xstats_get_names,
1787 	.fw_version_get = mlx5_fw_version_get,
1788 	.dev_infos_get = mlx5_dev_infos_get,
1789 	.representor_info_get = mlx5_representor_info_get,
1790 	.read_clock = mlx5_txpp_read_clock,
1791 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1792 	.vlan_filter_set = mlx5_vlan_filter_set,
1793 	.rx_queue_setup = mlx5_rx_queue_setup,
1794 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1795 	.tx_queue_setup = mlx5_tx_queue_setup,
1796 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1797 	.rx_queue_release = mlx5_rx_queue_release,
1798 	.tx_queue_release = mlx5_tx_queue_release,
1799 	.rx_queue_start = mlx5_rx_queue_start,
1800 	.rx_queue_stop = mlx5_rx_queue_stop,
1801 	.tx_queue_start = mlx5_tx_queue_start,
1802 	.tx_queue_stop = mlx5_tx_queue_stop,
1803 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1804 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1805 	.mac_addr_remove = mlx5_mac_addr_remove,
1806 	.mac_addr_add = mlx5_mac_addr_add,
1807 	.mac_addr_set = mlx5_mac_addr_set,
1808 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1809 	.mtu_set = mlx5_dev_set_mtu,
1810 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1811 	.vlan_offload_set = mlx5_vlan_offload_set,
1812 	.flow_ops_get = mlx5_flow_ops_get,
1813 	.rxq_info_get = mlx5_rxq_info_get,
1814 	.txq_info_get = mlx5_txq_info_get,
1815 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1816 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1817 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1818 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1819 	.is_removed = mlx5_is_removed,
1820 	.get_module_info = mlx5_get_module_info,
1821 	.get_module_eeprom = mlx5_get_module_eeprom,
1822 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1823 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1824 	.hairpin_bind = mlx5_hairpin_bind,
1825 	.hairpin_unbind = mlx5_hairpin_unbind,
1826 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1827 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1828 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1829 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1830 	.get_monitor_addr = mlx5_get_monitor_addr,
1831 };
1832 
1833 /**
1834  * Verify and store value for device argument.
1835  *
1836  * @param[in] key
1837  *   Key argument to verify.
1838  * @param[in] val
1839  *   Value associated with key.
1840  * @param opaque
1841  *   User data.
1842  *
1843  * @return
1844  *   0 on success, a negative errno value otherwise and rte_errno is set.
1845  */
1846 static int
1847 mlx5_args_check(const char *key, const char *val, void *opaque)
1848 {
1849 	struct mlx5_dev_config *config = opaque;
1850 	unsigned long mod;
1851 	signed long tmp;
1852 
1853 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1854 	if (!strcmp(MLX5_DRIVER_KEY, key) || !strcmp(MLX5_REPRESENTOR, key) ||
1855 	    !strcmp(MLX5_SYS_MEM_EN, key) || !strcmp(MLX5_TX_DB_NC, key) ||
1856 	    !strcmp(MLX5_MR_MEMPOOL_REG_EN, key) ||
1857 	    !strcmp(MLX5_MR_EXT_MEMSEG_EN, key))
1858 		return 0;
1859 	errno = 0;
1860 	tmp = strtol(val, NULL, 0);
1861 	if (errno) {
1862 		rte_errno = errno;
1863 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1864 		return -rte_errno;
1865 	}
1866 	if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1867 		/* Negative values are acceptable for some keys only. */
1868 		rte_errno = EINVAL;
1869 		DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1870 		return -rte_errno;
1871 	}
1872 	mod = tmp >= 0 ? tmp : -tmp;
1873 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1874 		if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1875 			DRV_LOG(ERR, "invalid CQE compression "
1876 				     "format parameter");
1877 			rte_errno = EINVAL;
1878 			return -rte_errno;
1879 		}
1880 		config->cqe_comp = !!tmp;
1881 		config->cqe_comp_fmt = tmp;
1882 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1883 		config->hw_padding = !!tmp;
1884 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1885 		config->mprq.enabled = !!tmp;
1886 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1887 		config->mprq.stride_num_n = tmp;
1888 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1889 		config->mprq.stride_size_n = tmp;
1890 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1891 		config->mprq.max_memcpy_len = tmp;
1892 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1893 		config->mprq.min_rxqs_num = tmp;
1894 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1895 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1896 				 " converted to txq_inline_max", key);
1897 		config->txq_inline_max = tmp;
1898 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1899 		config->txq_inline_max = tmp;
1900 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1901 		config->txq_inline_min = tmp;
1902 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1903 		config->txq_inline_mpw = tmp;
1904 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1905 		config->txqs_inline = tmp;
1906 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1907 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1908 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1909 		config->mps = !!tmp;
1910 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1911 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1912 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1913 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1914 				 " converted to txq_inline_mpw", key);
1915 		config->txq_inline_mpw = tmp;
1916 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1917 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1918 	} else if (strcmp(MLX5_TX_PP, key) == 0) {
1919 		if (!mod) {
1920 			DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1921 			rte_errno = EINVAL;
1922 			return -rte_errno;
1923 		}
1924 		config->tx_pp = tmp;
1925 	} else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1926 		config->tx_skew = tmp;
1927 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1928 		config->rx_vec_en = !!tmp;
1929 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1930 		config->l3_vxlan_en = !!tmp;
1931 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1932 		config->vf_nl_en = !!tmp;
1933 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1934 		config->dv_esw_en = !!tmp;
1935 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1936 		config->dv_flow_en = !!tmp;
1937 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1938 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1939 		    tmp != MLX5_XMETA_MODE_META16 &&
1940 		    tmp != MLX5_XMETA_MODE_META32 &&
1941 		    tmp != MLX5_XMETA_MODE_MISS_INFO) {
1942 			DRV_LOG(ERR, "invalid extensive "
1943 				     "metadata parameter");
1944 			rte_errno = EINVAL;
1945 			return -rte_errno;
1946 		}
1947 		if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1948 			config->dv_xmeta_en = tmp;
1949 		else
1950 			config->dv_miss_info = 1;
1951 	} else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1952 		config->lacp_by_user = !!tmp;
1953 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1954 		config->max_dump_files_num = tmp;
1955 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1956 		config->lro.timeout = tmp;
1957 	} else if (strcmp(RTE_DEVARGS_KEY_CLASS, key) == 0) {
1958 		DRV_LOG(DEBUG, "class argument is %s.", val);
1959 	} else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1960 		config->log_hp_size = tmp;
1961 	} else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1962 		if (tmp != MLX5_RCM_NONE &&
1963 		    tmp != MLX5_RCM_LIGHT &&
1964 		    tmp != MLX5_RCM_AGGR) {
1965 			DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1966 			rte_errno = EINVAL;
1967 			return -rte_errno;
1968 		}
1969 		config->reclaim_mode = tmp;
1970 	} else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1971 		config->decap_en = !!tmp;
1972 	} else if (strcmp(MLX5_ALLOW_DUPLICATE_PATTERN, key) == 0) {
1973 		config->allow_duplicate_pattern = !!tmp;
1974 	} else if (strcmp(MLX5_DELAY_DROP, key) == 0) {
1975 		config->std_delay_drop = !!(tmp & MLX5_DELAY_DROP_STANDARD);
1976 		config->hp_delay_drop = !!(tmp & MLX5_DELAY_DROP_HAIRPIN);
1977 	} else {
1978 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1979 		rte_errno = EINVAL;
1980 		return -rte_errno;
1981 	}
1982 	return 0;
1983 }
1984 
1985 /**
1986  * Parse device parameters.
1987  *
1988  * @param config
1989  *   Pointer to device configuration structure.
1990  * @param devargs
1991  *   Device arguments structure.
1992  *
1993  * @return
1994  *   0 on success, a negative errno value otherwise and rte_errno is set.
1995  */
1996 int
1997 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1998 {
1999 	const char **params = (const char *[]){
2000 		MLX5_DRIVER_KEY,
2001 		MLX5_RXQ_CQE_COMP_EN,
2002 		MLX5_RXQ_PKT_PAD_EN,
2003 		MLX5_RX_MPRQ_EN,
2004 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
2005 		MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
2006 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
2007 		MLX5_RXQS_MIN_MPRQ,
2008 		MLX5_TXQ_INLINE,
2009 		MLX5_TXQ_INLINE_MIN,
2010 		MLX5_TXQ_INLINE_MAX,
2011 		MLX5_TXQ_INLINE_MPW,
2012 		MLX5_TXQS_MIN_INLINE,
2013 		MLX5_TXQS_MAX_VEC,
2014 		MLX5_TXQ_MPW_EN,
2015 		MLX5_TXQ_MPW_HDR_DSEG_EN,
2016 		MLX5_TXQ_MAX_INLINE_LEN,
2017 		MLX5_TX_DB_NC,
2018 		MLX5_TX_PP,
2019 		MLX5_TX_SKEW,
2020 		MLX5_TX_VEC_EN,
2021 		MLX5_RX_VEC_EN,
2022 		MLX5_L3_VXLAN_EN,
2023 		MLX5_VF_NL_EN,
2024 		MLX5_DV_ESW_EN,
2025 		MLX5_DV_FLOW_EN,
2026 		MLX5_DV_XMETA_EN,
2027 		MLX5_LACP_BY_USER,
2028 		MLX5_MR_EXT_MEMSEG_EN,
2029 		MLX5_REPRESENTOR,
2030 		MLX5_MAX_DUMP_FILES_NUM,
2031 		MLX5_LRO_TIMEOUT_USEC,
2032 		RTE_DEVARGS_KEY_CLASS,
2033 		MLX5_HP_BUF_SIZE,
2034 		MLX5_RECLAIM_MEM,
2035 		MLX5_SYS_MEM_EN,
2036 		MLX5_DECAP_EN,
2037 		MLX5_ALLOW_DUPLICATE_PATTERN,
2038 		MLX5_MR_MEMPOOL_REG_EN,
2039 		MLX5_DELAY_DROP,
2040 		NULL,
2041 	};
2042 	struct rte_kvargs *kvlist;
2043 	int ret = 0;
2044 	int i;
2045 
2046 	if (devargs == NULL)
2047 		return 0;
2048 	/* Following UGLY cast is done to pass checkpatch. */
2049 	kvlist = rte_kvargs_parse(devargs->args, params);
2050 	if (kvlist == NULL) {
2051 		rte_errno = EINVAL;
2052 		return -rte_errno;
2053 	}
2054 	/* Process parameters. */
2055 	for (i = 0; (params[i] != NULL); ++i) {
2056 		if (rte_kvargs_count(kvlist, params[i])) {
2057 			ret = rte_kvargs_process(kvlist, params[i],
2058 						 mlx5_args_check, config);
2059 			if (ret) {
2060 				rte_errno = EINVAL;
2061 				rte_kvargs_free(kvlist);
2062 				return -rte_errno;
2063 			}
2064 		}
2065 	}
2066 	rte_kvargs_free(kvlist);
2067 	return 0;
2068 }
2069 
2070 /**
2071  * Configures the minimal amount of data to inline into WQE
2072  * while sending packets.
2073  *
2074  * - the txq_inline_min has the maximal priority, if this
2075  *   key is specified in devargs
2076  * - if DevX is enabled the inline mode is queried from the
2077  *   device (HCA attributes and NIC vport context if needed).
2078  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
2079  *   and none (0 bytes) for other NICs
2080  *
2081  * @param spawn
2082  *   Verbs device parameters (name, port, switch_info) to spawn.
2083  * @param config
2084  *   Device configuration parameters.
2085  */
2086 void
2087 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
2088 		    struct mlx5_dev_config *config)
2089 {
2090 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
2091 		/* Application defines size of inlined data explicitly. */
2092 		if (spawn->pci_dev != NULL) {
2093 			switch (spawn->pci_dev->id.device_id) {
2094 			case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2095 			case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2096 				if (config->txq_inline_min <
2097 					       (int)MLX5_INLINE_HSIZE_L2) {
2098 					DRV_LOG(DEBUG,
2099 						"txq_inline_mix aligned to minimal ConnectX-4 required value %d",
2100 						(int)MLX5_INLINE_HSIZE_L2);
2101 					config->txq_inline_min =
2102 							MLX5_INLINE_HSIZE_L2;
2103 				}
2104 				break;
2105 			}
2106 		}
2107 		goto exit;
2108 	}
2109 	if (config->hca_attr.eth_net_offloads) {
2110 		/* We have DevX enabled, inline mode queried successfully. */
2111 		switch (config->hca_attr.wqe_inline_mode) {
2112 		case MLX5_CAP_INLINE_MODE_L2:
2113 			/* outer L2 header must be inlined. */
2114 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2115 			goto exit;
2116 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
2117 			/* No inline data are required by NIC. */
2118 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2119 			config->hw_vlan_insert =
2120 				config->hca_attr.wqe_vlan_insert;
2121 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
2122 			goto exit;
2123 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
2124 			/* inline mode is defined by NIC vport context. */
2125 			if (!config->hca_attr.eth_virt)
2126 				break;
2127 			switch (config->hca_attr.vport_inline_mode) {
2128 			case MLX5_INLINE_MODE_NONE:
2129 				config->txq_inline_min =
2130 					MLX5_INLINE_HSIZE_NONE;
2131 				goto exit;
2132 			case MLX5_INLINE_MODE_L2:
2133 				config->txq_inline_min =
2134 					MLX5_INLINE_HSIZE_L2;
2135 				goto exit;
2136 			case MLX5_INLINE_MODE_IP:
2137 				config->txq_inline_min =
2138 					MLX5_INLINE_HSIZE_L3;
2139 				goto exit;
2140 			case MLX5_INLINE_MODE_TCP_UDP:
2141 				config->txq_inline_min =
2142 					MLX5_INLINE_HSIZE_L4;
2143 				goto exit;
2144 			case MLX5_INLINE_MODE_INNER_L2:
2145 				config->txq_inline_min =
2146 					MLX5_INLINE_HSIZE_INNER_L2;
2147 				goto exit;
2148 			case MLX5_INLINE_MODE_INNER_IP:
2149 				config->txq_inline_min =
2150 					MLX5_INLINE_HSIZE_INNER_L3;
2151 				goto exit;
2152 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
2153 				config->txq_inline_min =
2154 					MLX5_INLINE_HSIZE_INNER_L4;
2155 				goto exit;
2156 			}
2157 		}
2158 	}
2159 	if (spawn->pci_dev == NULL) {
2160 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2161 		goto exit;
2162 	}
2163 	/*
2164 	 * We get here if we are unable to deduce
2165 	 * inline data size with DevX. Try PCI ID
2166 	 * to determine old NICs.
2167 	 */
2168 	switch (spawn->pci_dev->id.device_id) {
2169 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2170 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2171 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2172 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2173 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2174 		config->hw_vlan_insert = 0;
2175 		break;
2176 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2177 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2178 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2179 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2180 		/*
2181 		 * These NICs support VLAN insertion from WQE and
2182 		 * report the wqe_vlan_insert flag. But there is the bug
2183 		 * and PFC control may be broken, so disable feature.
2184 		 */
2185 		config->hw_vlan_insert = 0;
2186 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2187 		break;
2188 	default:
2189 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2190 		break;
2191 	}
2192 exit:
2193 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2194 }
2195 
2196 /**
2197  * Configures the metadata mask fields in the shared context.
2198  *
2199  * @param [in] dev
2200  *   Pointer to Ethernet device.
2201  */
2202 void
2203 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2204 {
2205 	struct mlx5_priv *priv = dev->data->dev_private;
2206 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2207 	uint32_t meta, mark, reg_c0;
2208 
2209 	reg_c0 = ~priv->vport_meta_mask;
2210 	switch (priv->config.dv_xmeta_en) {
2211 	case MLX5_XMETA_MODE_LEGACY:
2212 		meta = UINT32_MAX;
2213 		mark = MLX5_FLOW_MARK_MASK;
2214 		break;
2215 	case MLX5_XMETA_MODE_META16:
2216 		meta = reg_c0 >> rte_bsf32(reg_c0);
2217 		mark = MLX5_FLOW_MARK_MASK;
2218 		break;
2219 	case MLX5_XMETA_MODE_META32:
2220 		meta = UINT32_MAX;
2221 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2222 		break;
2223 	default:
2224 		meta = 0;
2225 		mark = 0;
2226 		MLX5_ASSERT(false);
2227 		break;
2228 	}
2229 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2230 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2231 				 sh->dv_mark_mask, mark);
2232 	else
2233 		sh->dv_mark_mask = mark;
2234 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2235 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2236 				 sh->dv_meta_mask, meta);
2237 	else
2238 		sh->dv_meta_mask = meta;
2239 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2240 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2241 				 sh->dv_meta_mask, reg_c0);
2242 	else
2243 		sh->dv_regc0_mask = reg_c0;
2244 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2245 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2246 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2247 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2248 }
2249 
2250 int
2251 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2252 {
2253 	static const char *const dynf_names[] = {
2254 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2255 		RTE_MBUF_DYNFLAG_METADATA_NAME,
2256 		RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2257 	};
2258 	unsigned int i;
2259 
2260 	if (n < RTE_DIM(dynf_names))
2261 		return -ENOMEM;
2262 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
2263 		if (names[i] == NULL)
2264 			return -EINVAL;
2265 		strcpy(names[i], dynf_names[i]);
2266 	}
2267 	return RTE_DIM(dynf_names);
2268 }
2269 
2270 /**
2271  * Comparison callback to sort device data.
2272  *
2273  * This is meant to be used with qsort().
2274  *
2275  * @param a[in]
2276  *   Pointer to pointer to first data object.
2277  * @param b[in]
2278  *   Pointer to pointer to second data object.
2279  *
2280  * @return
2281  *   0 if both objects are equal, less than 0 if the first argument is less
2282  *   than the second, greater than 0 otherwise.
2283  */
2284 int
2285 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2286 			      struct mlx5_dev_config *config,
2287 			      struct rte_device *dpdk_dev)
2288 {
2289 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2290 	struct mlx5_dev_config *sh_conf = NULL;
2291 	uint16_t port_id;
2292 
2293 	MLX5_ASSERT(sh);
2294 	/* Nothing to compare for the single/first device. */
2295 	if (sh->refcnt == 1)
2296 		return 0;
2297 	/* Find the device with shared context. */
2298 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
2299 		struct mlx5_priv *opriv =
2300 			rte_eth_devices[port_id].data->dev_private;
2301 
2302 		if (opriv && opriv != priv && opriv->sh == sh) {
2303 			sh_conf = &opriv->config;
2304 			break;
2305 		}
2306 	}
2307 	if (!sh_conf)
2308 		return 0;
2309 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2310 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2311 			     " for shared %s context", sh->ibdev_name);
2312 		rte_errno = EINVAL;
2313 		return rte_errno;
2314 	}
2315 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2316 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2317 			     " for shared %s context", sh->ibdev_name);
2318 		rte_errno = EINVAL;
2319 		return rte_errno;
2320 	}
2321 	return 0;
2322 }
2323 
2324 /**
2325  * Look for the ethernet device belonging to mlx5 driver.
2326  *
2327  * @param[in] port_id
2328  *   port_id to start looking for device.
2329  * @param[in] odev
2330  *   Pointer to the hint device. When device is being probed
2331  *   the its siblings (master and preceding representors might
2332  *   not have assigned driver yet (because the mlx5_os_pci_probe()
2333  *   is not completed yet, for this case match on hint
2334  *   device may be used to detect sibling device.
2335  *
2336  * @return
2337  *   port_id of found device, RTE_MAX_ETHPORT if not found.
2338  */
2339 uint16_t
2340 mlx5_eth_find_next(uint16_t port_id, struct rte_device *odev)
2341 {
2342 	while (port_id < RTE_MAX_ETHPORTS) {
2343 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2344 
2345 		if (dev->state != RTE_ETH_DEV_UNUSED &&
2346 		    dev->device &&
2347 		    (dev->device == odev ||
2348 		     (dev->device->driver &&
2349 		     dev->device->driver->name &&
2350 		     ((strcmp(dev->device->driver->name,
2351 			      MLX5_PCI_DRIVER_NAME) == 0) ||
2352 		      (strcmp(dev->device->driver->name,
2353 			      MLX5_AUXILIARY_DRIVER_NAME) == 0)))))
2354 			break;
2355 		port_id++;
2356 	}
2357 	if (port_id >= RTE_MAX_ETHPORTS)
2358 		return RTE_MAX_ETHPORTS;
2359 	return port_id;
2360 }
2361 
2362 /**
2363  * Callback to remove a device.
2364  *
2365  * This function removes all Ethernet devices belong to a given device.
2366  *
2367  * @param[in] cdev
2368  *   Pointer to the generic device.
2369  *
2370  * @return
2371  *   0 on success, the function cannot fail.
2372  */
2373 int
2374 mlx5_net_remove(struct mlx5_common_device *cdev)
2375 {
2376 	uint16_t port_id;
2377 	int ret = 0;
2378 
2379 	RTE_ETH_FOREACH_DEV_OF(port_id, cdev->dev) {
2380 		/*
2381 		 * mlx5_dev_close() is not registered to secondary process,
2382 		 * call the close function explicitly for secondary process.
2383 		 */
2384 		if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2385 			ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2386 		else
2387 			ret |= rte_eth_dev_close(port_id);
2388 	}
2389 	return ret == 0 ? 0 : -EIO;
2390 }
2391 
2392 static const struct rte_pci_id mlx5_pci_id_map[] = {
2393 	{
2394 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2395 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2396 	},
2397 	{
2398 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2399 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2400 	},
2401 	{
2402 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2403 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2404 	},
2405 	{
2406 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2407 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2408 	},
2409 	{
2410 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2411 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2412 	},
2413 	{
2414 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2415 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2416 	},
2417 	{
2418 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2419 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2420 	},
2421 	{
2422 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2423 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2424 	},
2425 	{
2426 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2427 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2428 	},
2429 	{
2430 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2431 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2432 	},
2433 	{
2434 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2435 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2436 	},
2437 	{
2438 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2439 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2440 	},
2441 	{
2442 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2443 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2444 	},
2445 	{
2446 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2447 				PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2448 	},
2449 	{
2450 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2451 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2452 	},
2453 	{
2454 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2455 				PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2456 	},
2457 	{
2458 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2459 				PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2460 	},
2461 	{
2462 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2463 				PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2464 	},
2465 	{
2466 		.vendor_id = 0
2467 	}
2468 };
2469 
2470 static struct mlx5_class_driver mlx5_net_driver = {
2471 	.drv_class = MLX5_CLASS_ETH,
2472 	.name = RTE_STR(MLX5_ETH_DRIVER_NAME),
2473 	.id_table = mlx5_pci_id_map,
2474 	.probe = mlx5_os_net_probe,
2475 	.remove = mlx5_net_remove,
2476 	.probe_again = 1,
2477 	.intr_lsc = 1,
2478 	.intr_rmv = 1,
2479 };
2480 
2481 /* Initialize driver log type. */
2482 RTE_LOG_REGISTER_DEFAULT(mlx5_logtype, NOTICE)
2483 
2484 /**
2485  * Driver initialization routine.
2486  */
2487 RTE_INIT(rte_mlx5_pmd_init)
2488 {
2489 	pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2490 	mlx5_common_init();
2491 	/* Build the static tables for Verbs conversion. */
2492 	mlx5_set_ptype_table();
2493 	mlx5_set_cksum_table();
2494 	mlx5_set_swp_types_table();
2495 	if (mlx5_glue)
2496 		mlx5_class_driver_register(&mlx5_net_driver);
2497 }
2498 
2499 RTE_PMD_EXPORT_NAME(MLX5_ETH_DRIVER_NAME, __COUNTER__);
2500 RTE_PMD_REGISTER_PCI_TABLE(MLX5_ETH_DRIVER_NAME, mlx5_pci_id_map);
2501 RTE_PMD_REGISTER_KMOD_DEP(MLX5_ETH_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");
2502