xref: /dpdk/drivers/net/mlx5/mlx5.c (revision aaf3b44c6618f60ec878237e3a3dcad0912bedf4)
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_flow.h"
40 #include "mlx5_flow_os.h"
41 #include "rte_pmd_mlx5.h"
42 
43 #define MLX5_ETH_DRIVER_NAME mlx5_eth
44 
45 /* Driver type key for new device global syntax. */
46 #define MLX5_DRIVER_KEY "driver"
47 
48 /* Device parameter to enable RX completion queue compression. */
49 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
50 
51 /* Device parameter to enable padding Rx packet to cacheline size. */
52 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
53 
54 /* Device parameter to enable Multi-Packet Rx queue. */
55 #define MLX5_RX_MPRQ_EN "mprq_en"
56 
57 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
58 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
59 
60 /* Device parameter to configure log 2 of the stride size for MPRQ. */
61 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
62 
63 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
64 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
65 
66 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
67 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
68 
69 /* Device parameter to configure inline send. Deprecated, ignored.*/
70 #define MLX5_TXQ_INLINE "txq_inline"
71 
72 /* Device parameter to limit packet size to inline with ordinary SEND. */
73 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
74 
75 /* Device parameter to configure minimal data size to inline. */
76 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
77 
78 /* Device parameter to limit packet size to inline with Enhanced MPW. */
79 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
80 
81 /*
82  * Device parameter to configure the number of TX queues threshold for
83  * enabling inline send.
84  */
85 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
86 
87 /*
88  * Device parameter to configure the number of TX queues threshold for
89  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
90  */
91 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
92 
93 /* Device parameter to enable multi-packet send WQEs. */
94 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
95 
96 /*
97  * Device parameter to force doorbell register mapping
98  * to non-cahed region eliminating the extra write memory barrier.
99  */
100 #define MLX5_TX_DB_NC "tx_db_nc"
101 
102 /*
103  * Device parameter to include 2 dsegs in the title WQEBB.
104  * Deprecated, ignored.
105  */
106 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
107 
108 /*
109  * Device parameter to limit the size of inlining packet.
110  * Deprecated, ignored.
111  */
112 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
113 
114 /*
115  * Device parameter to enable Tx scheduling on timestamps
116  * and specify the packet pacing granularity in nanoseconds.
117  */
118 #define MLX5_TX_PP "tx_pp"
119 
120 /*
121  * Device parameter to specify skew in nanoseconds on Tx datapath,
122  * it represents the time between SQ start WQE processing and
123  * appearing actual packet data on the wire.
124  */
125 #define MLX5_TX_SKEW "tx_skew"
126 
127 /*
128  * Device parameter to enable hardware Tx vector.
129  * Deprecated, ignored (no vectorized Tx routines anymore).
130  */
131 #define MLX5_TX_VEC_EN "tx_vec_en"
132 
133 /* Device parameter to enable hardware Rx vector. */
134 #define MLX5_RX_VEC_EN "rx_vec_en"
135 
136 /* Allow L3 VXLAN flow creation. */
137 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
138 
139 /* Activate DV E-Switch flow steering. */
140 #define MLX5_DV_ESW_EN "dv_esw_en"
141 
142 /* Activate DV flow steering. */
143 #define MLX5_DV_FLOW_EN "dv_flow_en"
144 
145 /* Enable extensive flow metadata support. */
146 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
147 
148 /* Device parameter to let the user manage the lacp traffic of bonded device */
149 #define MLX5_LACP_BY_USER "lacp_by_user"
150 
151 /* Activate Netlink support in VF mode. */
152 #define MLX5_VF_NL_EN "vf_nl_en"
153 
154 /* Enable extending memsegs when creating a MR. */
155 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
156 
157 /* Select port representors to instantiate. */
158 #define MLX5_REPRESENTOR "representor"
159 
160 /* Device parameter to configure the maximum number of dump files per queue. */
161 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
162 
163 /* Configure timeout of LRO session (in microseconds). */
164 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
165 
166 /*
167  * Device parameter to configure the total data buffer size for a single
168  * hairpin queue (logarithm value).
169  */
170 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
171 
172 /* Flow memory reclaim mode. */
173 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
174 
175 /* The default memory allocator used in PMD. */
176 #define MLX5_SYS_MEM_EN "sys_mem_en"
177 /* Decap will be used or not. */
178 #define MLX5_DECAP_EN "decap_en"
179 
180 /* Device parameter to configure allow or prevent duplicate rules pattern. */
181 #define MLX5_ALLOW_DUPLICATE_PATTERN "allow_duplicate_pattern"
182 
183 /* Device parameter to configure implicit registration of mempool memory. */
184 #define MLX5_MR_MEMPOOL_REG_EN "mr_mempool_reg_en"
185 
186 /* Shared memory between primary and secondary processes. */
187 struct mlx5_shared_data *mlx5_shared_data;
188 
189 /** Driver-specific log messages type. */
190 int mlx5_logtype;
191 
192 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
193 						LIST_HEAD_INITIALIZER();
194 static pthread_mutex_t mlx5_dev_ctx_list_mutex;
195 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
196 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
197 	[MLX5_IPOOL_DECAP_ENCAP] = {
198 		.size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
199 		.trunk_size = 64,
200 		.grow_trunk = 3,
201 		.grow_shift = 2,
202 		.need_lock = 1,
203 		.release_mem_en = 1,
204 		.malloc = mlx5_malloc,
205 		.free = mlx5_free,
206 		.type = "mlx5_encap_decap_ipool",
207 	},
208 	[MLX5_IPOOL_PUSH_VLAN] = {
209 		.size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
210 		.trunk_size = 64,
211 		.grow_trunk = 3,
212 		.grow_shift = 2,
213 		.need_lock = 1,
214 		.release_mem_en = 1,
215 		.malloc = mlx5_malloc,
216 		.free = mlx5_free,
217 		.type = "mlx5_push_vlan_ipool",
218 	},
219 	[MLX5_IPOOL_TAG] = {
220 		.size = sizeof(struct mlx5_flow_dv_tag_resource),
221 		.trunk_size = 64,
222 		.grow_trunk = 3,
223 		.grow_shift = 2,
224 		.need_lock = 1,
225 		.release_mem_en = 0,
226 		.per_core_cache = (1 << 16),
227 		.malloc = mlx5_malloc,
228 		.free = mlx5_free,
229 		.type = "mlx5_tag_ipool",
230 	},
231 	[MLX5_IPOOL_PORT_ID] = {
232 		.size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
233 		.trunk_size = 64,
234 		.grow_trunk = 3,
235 		.grow_shift = 2,
236 		.need_lock = 1,
237 		.release_mem_en = 1,
238 		.malloc = mlx5_malloc,
239 		.free = mlx5_free,
240 		.type = "mlx5_port_id_ipool",
241 	},
242 	[MLX5_IPOOL_JUMP] = {
243 		.size = sizeof(struct mlx5_flow_tbl_data_entry),
244 		.trunk_size = 64,
245 		.grow_trunk = 3,
246 		.grow_shift = 2,
247 		.need_lock = 1,
248 		.release_mem_en = 1,
249 		.malloc = mlx5_malloc,
250 		.free = mlx5_free,
251 		.type = "mlx5_jump_ipool",
252 	},
253 	[MLX5_IPOOL_SAMPLE] = {
254 		.size = sizeof(struct mlx5_flow_dv_sample_resource),
255 		.trunk_size = 64,
256 		.grow_trunk = 3,
257 		.grow_shift = 2,
258 		.need_lock = 1,
259 		.release_mem_en = 1,
260 		.malloc = mlx5_malloc,
261 		.free = mlx5_free,
262 		.type = "mlx5_sample_ipool",
263 	},
264 	[MLX5_IPOOL_DEST_ARRAY] = {
265 		.size = sizeof(struct mlx5_flow_dv_dest_array_resource),
266 		.trunk_size = 64,
267 		.grow_trunk = 3,
268 		.grow_shift = 2,
269 		.need_lock = 1,
270 		.release_mem_en = 1,
271 		.malloc = mlx5_malloc,
272 		.free = mlx5_free,
273 		.type = "mlx5_dest_array_ipool",
274 	},
275 	[MLX5_IPOOL_TUNNEL_ID] = {
276 		.size = sizeof(struct mlx5_flow_tunnel),
277 		.trunk_size = MLX5_MAX_TUNNELS,
278 		.need_lock = 1,
279 		.release_mem_en = 1,
280 		.type = "mlx5_tunnel_offload",
281 	},
282 	[MLX5_IPOOL_TNL_TBL_ID] = {
283 		.size = 0,
284 		.need_lock = 1,
285 		.type = "mlx5_flow_tnl_tbl_ipool",
286 	},
287 #endif
288 	[MLX5_IPOOL_MTR] = {
289 		/**
290 		 * The ipool index should grow continually from small to big,
291 		 * for meter idx, so not set grow_trunk to avoid meter index
292 		 * not jump continually.
293 		 */
294 		.size = sizeof(struct mlx5_legacy_flow_meter),
295 		.trunk_size = 64,
296 		.need_lock = 1,
297 		.release_mem_en = 1,
298 		.malloc = mlx5_malloc,
299 		.free = mlx5_free,
300 		.type = "mlx5_meter_ipool",
301 	},
302 	[MLX5_IPOOL_MCP] = {
303 		.size = sizeof(struct mlx5_flow_mreg_copy_resource),
304 		.trunk_size = 64,
305 		.grow_trunk = 3,
306 		.grow_shift = 2,
307 		.need_lock = 1,
308 		.release_mem_en = 1,
309 		.malloc = mlx5_malloc,
310 		.free = mlx5_free,
311 		.type = "mlx5_mcp_ipool",
312 	},
313 	[MLX5_IPOOL_HRXQ] = {
314 		.size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
315 		.trunk_size = 64,
316 		.grow_trunk = 3,
317 		.grow_shift = 2,
318 		.need_lock = 1,
319 		.release_mem_en = 1,
320 		.malloc = mlx5_malloc,
321 		.free = mlx5_free,
322 		.type = "mlx5_hrxq_ipool",
323 	},
324 	[MLX5_IPOOL_MLX5_FLOW] = {
325 		/*
326 		 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
327 		 * It set in run time according to PCI function configuration.
328 		 */
329 		.size = 0,
330 		.trunk_size = 64,
331 		.grow_trunk = 3,
332 		.grow_shift = 2,
333 		.need_lock = 1,
334 		.release_mem_en = 0,
335 		.per_core_cache = 1 << 19,
336 		.malloc = mlx5_malloc,
337 		.free = mlx5_free,
338 		.type = "mlx5_flow_handle_ipool",
339 	},
340 	[MLX5_IPOOL_RTE_FLOW] = {
341 		.size = sizeof(struct rte_flow),
342 		.trunk_size = 4096,
343 		.need_lock = 1,
344 		.release_mem_en = 1,
345 		.malloc = mlx5_malloc,
346 		.free = mlx5_free,
347 		.type = "rte_flow_ipool",
348 	},
349 	[MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
350 		.size = 0,
351 		.need_lock = 1,
352 		.type = "mlx5_flow_rss_id_ipool",
353 	},
354 	[MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
355 		.size = sizeof(struct mlx5_shared_action_rss),
356 		.trunk_size = 64,
357 		.grow_trunk = 3,
358 		.grow_shift = 2,
359 		.need_lock = 1,
360 		.release_mem_en = 1,
361 		.malloc = mlx5_malloc,
362 		.free = mlx5_free,
363 		.type = "mlx5_shared_action_rss",
364 	},
365 	[MLX5_IPOOL_MTR_POLICY] = {
366 		/**
367 		 * The ipool index should grow continually from small to big,
368 		 * for policy idx, so not set grow_trunk to avoid policy index
369 		 * not jump continually.
370 		 */
371 		.size = sizeof(struct mlx5_flow_meter_sub_policy),
372 		.trunk_size = 64,
373 		.need_lock = 1,
374 		.release_mem_en = 1,
375 		.malloc = mlx5_malloc,
376 		.free = mlx5_free,
377 		.type = "mlx5_meter_policy_ipool",
378 	},
379 };
380 
381 
382 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
383 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
384 
385 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 1024
386 
387 /**
388  * Decide whether representor ID is a HPF(host PF) port on BF2.
389  *
390  * @param dev
391  *   Pointer to Ethernet device structure.
392  *
393  * @return
394  *   Non-zero if HPF, otherwise 0.
395  */
396 bool
397 mlx5_is_hpf(struct rte_eth_dev *dev)
398 {
399 	struct mlx5_priv *priv = dev->data->dev_private;
400 	uint16_t repr = MLX5_REPRESENTOR_REPR(priv->representor_id);
401 	int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
402 
403 	return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_VF &&
404 	       MLX5_REPRESENTOR_REPR(-1) == repr;
405 }
406 
407 /**
408  * Decide whether representor ID is a SF port representor.
409  *
410  * @param dev
411  *   Pointer to Ethernet device structure.
412  *
413  * @return
414  *   Non-zero if HPF, otherwise 0.
415  */
416 bool
417 mlx5_is_sf_repr(struct rte_eth_dev *dev)
418 {
419 	struct mlx5_priv *priv = dev->data->dev_private;
420 	int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
421 
422 	return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_SF;
423 }
424 
425 /**
426  * Initialize the ASO aging management structure.
427  *
428  * @param[in] sh
429  *   Pointer to mlx5_dev_ctx_shared object to free
430  *
431  * @return
432  *   0 on success, a negative errno value otherwise and rte_errno is set.
433  */
434 int
435 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
436 {
437 	int err;
438 
439 	if (sh->aso_age_mng)
440 		return 0;
441 	sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
442 				      RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
443 	if (!sh->aso_age_mng) {
444 		DRV_LOG(ERR, "aso_age_mng allocation was failed.");
445 		rte_errno = ENOMEM;
446 		return -ENOMEM;
447 	}
448 	err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_FLOW_HIT);
449 	if (err) {
450 		mlx5_free(sh->aso_age_mng);
451 		return -1;
452 	}
453 	rte_rwlock_init(&sh->aso_age_mng->resize_rwl);
454 	rte_spinlock_init(&sh->aso_age_mng->free_sl);
455 	LIST_INIT(&sh->aso_age_mng->free);
456 	return 0;
457 }
458 
459 /**
460  * Close and release all the resources of the ASO aging management structure.
461  *
462  * @param[in] sh
463  *   Pointer to mlx5_dev_ctx_shared object to free.
464  */
465 static void
466 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
467 {
468 	int i, j;
469 
470 	mlx5_aso_flow_hit_queue_poll_stop(sh);
471 	mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_FLOW_HIT);
472 	if (sh->aso_age_mng->pools) {
473 		struct mlx5_aso_age_pool *pool;
474 
475 		for (i = 0; i < sh->aso_age_mng->next; ++i) {
476 			pool = sh->aso_age_mng->pools[i];
477 			claim_zero(mlx5_devx_cmd_destroy
478 						(pool->flow_hit_aso_obj));
479 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
480 				if (pool->actions[j].dr_action)
481 					claim_zero
482 					    (mlx5_flow_os_destroy_flow_action
483 					      (pool->actions[j].dr_action));
484 			mlx5_free(pool);
485 		}
486 		mlx5_free(sh->aso_age_mng->pools);
487 	}
488 	mlx5_free(sh->aso_age_mng);
489 }
490 
491 /**
492  * Initialize the shared aging list information per port.
493  *
494  * @param[in] sh
495  *   Pointer to mlx5_dev_ctx_shared object.
496  */
497 static void
498 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
499 {
500 	uint32_t i;
501 	struct mlx5_age_info *age_info;
502 
503 	for (i = 0; i < sh->max_port; i++) {
504 		age_info = &sh->port[i].age_info;
505 		age_info->flags = 0;
506 		TAILQ_INIT(&age_info->aged_counters);
507 		LIST_INIT(&age_info->aged_aso);
508 		rte_spinlock_init(&age_info->aged_sl);
509 		MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
510 	}
511 }
512 
513 /**
514  * Initialize the counters management structure.
515  *
516  * @param[in] sh
517  *   Pointer to mlx5_dev_ctx_shared object to free
518  */
519 static void
520 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
521 {
522 	struct mlx5_hca_attr *attr = &sh->cdev->config.hca_attr;
523 	int i;
524 
525 	memset(&sh->cmng, 0, sizeof(sh->cmng));
526 	TAILQ_INIT(&sh->cmng.flow_counters);
527 	sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
528 	sh->cmng.max_id = -1;
529 	sh->cmng.last_pool_idx = POOL_IDX_INVALID;
530 	rte_spinlock_init(&sh->cmng.pool_update_sl);
531 	for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
532 		TAILQ_INIT(&sh->cmng.counters[i]);
533 		rte_spinlock_init(&sh->cmng.csl[i]);
534 	}
535 	if (sh->devx && !haswell_broadwell_cpu) {
536 		sh->cmng.relaxed_ordering_write = attr->relaxed_ordering_write;
537 		sh->cmng.relaxed_ordering_read = attr->relaxed_ordering_read;
538 	}
539 }
540 
541 /**
542  * Destroy all the resources allocated for a counter memory management.
543  *
544  * @param[in] mng
545  *   Pointer to the memory management structure.
546  */
547 static void
548 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
549 {
550 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
551 
552 	LIST_REMOVE(mng, next);
553 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
554 	claim_zero(mlx5_os_umem_dereg(mng->umem));
555 	mlx5_free(mem);
556 }
557 
558 /**
559  * Close and release all the resources of the counters management.
560  *
561  * @param[in] sh
562  *   Pointer to mlx5_dev_ctx_shared object to free.
563  */
564 static void
565 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
566 {
567 	struct mlx5_counter_stats_mem_mng *mng;
568 	int i, j;
569 	int retries = 1024;
570 
571 	rte_errno = 0;
572 	while (--retries) {
573 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
574 		if (rte_errno != EINPROGRESS)
575 			break;
576 		rte_pause();
577 	}
578 
579 	if (sh->cmng.pools) {
580 		struct mlx5_flow_counter_pool *pool;
581 		uint16_t n_valid = sh->cmng.n_valid;
582 		bool fallback = sh->cmng.counter_fallback;
583 
584 		for (i = 0; i < n_valid; ++i) {
585 			pool = sh->cmng.pools[i];
586 			if (!fallback && pool->min_dcs)
587 				claim_zero(mlx5_devx_cmd_destroy
588 							       (pool->min_dcs));
589 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
590 				struct mlx5_flow_counter *cnt =
591 						MLX5_POOL_GET_CNT(pool, j);
592 
593 				if (cnt->action)
594 					claim_zero
595 					 (mlx5_flow_os_destroy_flow_action
596 					  (cnt->action));
597 				if (fallback && MLX5_POOL_GET_CNT
598 				    (pool, j)->dcs_when_free)
599 					claim_zero(mlx5_devx_cmd_destroy
600 						   (cnt->dcs_when_free));
601 			}
602 			mlx5_free(pool);
603 		}
604 		mlx5_free(sh->cmng.pools);
605 	}
606 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
607 	while (mng) {
608 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
609 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
610 	}
611 	memset(&sh->cmng, 0, sizeof(sh->cmng));
612 }
613 
614 /**
615  * Initialize the aso flow meters management structure.
616  *
617  * @param[in] sh
618  *   Pointer to mlx5_dev_ctx_shared object to free
619  */
620 int
621 mlx5_aso_flow_mtrs_mng_init(struct mlx5_dev_ctx_shared *sh)
622 {
623 	if (!sh->mtrmng) {
624 		sh->mtrmng = mlx5_malloc(MLX5_MEM_ZERO,
625 			sizeof(*sh->mtrmng),
626 			RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
627 		if (!sh->mtrmng) {
628 			DRV_LOG(ERR,
629 			"meter management allocation was failed.");
630 			rte_errno = ENOMEM;
631 			return -ENOMEM;
632 		}
633 		if (sh->meter_aso_en) {
634 			rte_spinlock_init(&sh->mtrmng->pools_mng.mtrsl);
635 			rte_rwlock_init(&sh->mtrmng->pools_mng.resize_mtrwl);
636 			LIST_INIT(&sh->mtrmng->pools_mng.meters);
637 		}
638 		sh->mtrmng->def_policy_id = MLX5_INVALID_POLICY_ID;
639 	}
640 	return 0;
641 }
642 
643 /**
644  * Close and release all the resources of
645  * the ASO flow meter management structure.
646  *
647  * @param[in] sh
648  *   Pointer to mlx5_dev_ctx_shared object to free.
649  */
650 static void
651 mlx5_aso_flow_mtrs_mng_close(struct mlx5_dev_ctx_shared *sh)
652 {
653 	struct mlx5_aso_mtr_pool *mtr_pool;
654 	struct mlx5_flow_mtr_mng *mtrmng = sh->mtrmng;
655 	uint32_t idx;
656 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
657 	struct mlx5_aso_mtr *aso_mtr;
658 	int i;
659 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
660 
661 	if (sh->meter_aso_en) {
662 		mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_POLICER);
663 		idx = mtrmng->pools_mng.n_valid;
664 		while (idx--) {
665 			mtr_pool = mtrmng->pools_mng.pools[idx];
666 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
667 			for (i = 0; i < MLX5_ASO_MTRS_PER_POOL; i++) {
668 				aso_mtr = &mtr_pool->mtrs[i];
669 				if (aso_mtr->fm.meter_action)
670 					claim_zero
671 					(mlx5_glue->destroy_flow_action
672 					(aso_mtr->fm.meter_action));
673 			}
674 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
675 			claim_zero(mlx5_devx_cmd_destroy
676 						(mtr_pool->devx_obj));
677 			mtrmng->pools_mng.n_valid--;
678 			mlx5_free(mtr_pool);
679 		}
680 		mlx5_free(sh->mtrmng->pools_mng.pools);
681 	}
682 	mlx5_free(sh->mtrmng);
683 	sh->mtrmng = NULL;
684 }
685 
686 /* Send FLOW_AGED event if needed. */
687 void
688 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
689 {
690 	struct mlx5_age_info *age_info;
691 	uint32_t i;
692 
693 	for (i = 0; i < sh->max_port; i++) {
694 		age_info = &sh->port[i].age_info;
695 		if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
696 			continue;
697 		MLX5_AGE_UNSET(age_info, MLX5_AGE_EVENT_NEW);
698 		if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER)) {
699 			MLX5_AGE_UNSET(age_info, MLX5_AGE_TRIGGER);
700 			rte_eth_dev_callback_process
701 				(&rte_eth_devices[sh->port[i].devx_ih_port_id],
702 				RTE_ETH_EVENT_FLOW_AGED, NULL);
703 		}
704 	}
705 }
706 
707 /*
708  * Initialize the ASO connection tracking structure.
709  *
710  * @param[in] sh
711  *   Pointer to mlx5_dev_ctx_shared object.
712  *
713  * @return
714  *   0 on success, a negative errno value otherwise and rte_errno is set.
715  */
716 int
717 mlx5_flow_aso_ct_mng_init(struct mlx5_dev_ctx_shared *sh)
718 {
719 	int err;
720 
721 	if (sh->ct_mng)
722 		return 0;
723 	sh->ct_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->ct_mng),
724 				 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
725 	if (!sh->ct_mng) {
726 		DRV_LOG(ERR, "ASO CT management allocation failed.");
727 		rte_errno = ENOMEM;
728 		return -rte_errno;
729 	}
730 	err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
731 	if (err) {
732 		mlx5_free(sh->ct_mng);
733 		/* rte_errno should be extracted from the failure. */
734 		rte_errno = EINVAL;
735 		return -rte_errno;
736 	}
737 	rte_spinlock_init(&sh->ct_mng->ct_sl);
738 	rte_rwlock_init(&sh->ct_mng->resize_rwl);
739 	LIST_INIT(&sh->ct_mng->free_cts);
740 	return 0;
741 }
742 
743 /*
744  * Close and release all the resources of the
745  * ASO connection tracking management structure.
746  *
747  * @param[in] sh
748  *   Pointer to mlx5_dev_ctx_shared object to free.
749  */
750 static void
751 mlx5_flow_aso_ct_mng_close(struct mlx5_dev_ctx_shared *sh)
752 {
753 	struct mlx5_aso_ct_pools_mng *mng = sh->ct_mng;
754 	struct mlx5_aso_ct_pool *ct_pool;
755 	struct mlx5_aso_ct_action *ct;
756 	uint32_t idx;
757 	uint32_t val;
758 	uint32_t cnt;
759 	int i;
760 
761 	mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
762 	idx = mng->next;
763 	while (idx--) {
764 		cnt = 0;
765 		ct_pool = mng->pools[idx];
766 		for (i = 0; i < MLX5_ASO_CT_ACTIONS_PER_POOL; i++) {
767 			ct = &ct_pool->actions[i];
768 			val = __atomic_fetch_sub(&ct->refcnt, 1,
769 						 __ATOMIC_RELAXED);
770 			MLX5_ASSERT(val == 1);
771 			if (val > 1)
772 				cnt++;
773 #ifdef HAVE_MLX5_DR_ACTION_ASO_CT
774 			if (ct->dr_action_orig)
775 				claim_zero(mlx5_glue->destroy_flow_action
776 							(ct->dr_action_orig));
777 			if (ct->dr_action_rply)
778 				claim_zero(mlx5_glue->destroy_flow_action
779 							(ct->dr_action_rply));
780 #endif
781 		}
782 		claim_zero(mlx5_devx_cmd_destroy(ct_pool->devx_obj));
783 		if (cnt) {
784 			DRV_LOG(DEBUG, "%u ASO CT objects are being used in the pool %u",
785 				cnt, i);
786 		}
787 		mlx5_free(ct_pool);
788 		/* in case of failure. */
789 		mng->next--;
790 	}
791 	mlx5_free(mng->pools);
792 	mlx5_free(mng);
793 	/* Management structure must be cleared to 0s during allocation. */
794 	sh->ct_mng = NULL;
795 }
796 
797 /**
798  * Initialize the flow resources' indexed mempool.
799  *
800  * @param[in] sh
801  *   Pointer to mlx5_dev_ctx_shared object.
802  * @param[in] config
803  *   Pointer to user dev config.
804  */
805 static void
806 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
807 		       const struct mlx5_dev_config *config)
808 {
809 	uint8_t i;
810 	struct mlx5_indexed_pool_config cfg;
811 
812 	for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
813 		cfg = mlx5_ipool_cfg[i];
814 		switch (i) {
815 		default:
816 			break;
817 		/*
818 		 * Set MLX5_IPOOL_MLX5_FLOW ipool size
819 		 * according to PCI function flow configuration.
820 		 */
821 		case MLX5_IPOOL_MLX5_FLOW:
822 			cfg.size = config->dv_flow_en ?
823 				sizeof(struct mlx5_flow_handle) :
824 				MLX5_FLOW_HANDLE_VERBS_SIZE;
825 			break;
826 		}
827 		if (config->reclaim_mode) {
828 			cfg.release_mem_en = 1;
829 			cfg.per_core_cache = 0;
830 		} else {
831 			cfg.release_mem_en = 0;
832 		}
833 		sh->ipool[i] = mlx5_ipool_create(&cfg);
834 	}
835 }
836 
837 
838 /**
839  * Release the flow resources' indexed mempool.
840  *
841  * @param[in] sh
842  *   Pointer to mlx5_dev_ctx_shared object.
843  */
844 static void
845 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
846 {
847 	uint8_t i;
848 
849 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
850 		mlx5_ipool_destroy(sh->ipool[i]);
851 	for (i = 0; i < MLX5_MAX_MODIFY_NUM; ++i)
852 		if (sh->mdh_ipools[i])
853 			mlx5_ipool_destroy(sh->mdh_ipools[i]);
854 }
855 
856 /*
857  * Check if dynamic flex parser for eCPRI already exists.
858  *
859  * @param dev
860  *   Pointer to Ethernet device structure.
861  *
862  * @return
863  *   true on exists, false on not.
864  */
865 bool
866 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
867 {
868 	struct mlx5_priv *priv = dev->data->dev_private;
869 	struct mlx5_flex_parser_profiles *prf =
870 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
871 
872 	return !!prf->obj;
873 }
874 
875 /*
876  * Allocation of a flex parser for eCPRI. Once created, this parser related
877  * resources will be held until the device is closed.
878  *
879  * @param dev
880  *   Pointer to Ethernet device structure.
881  *
882  * @return
883  *   0 on success, a negative errno value otherwise and rte_errno is set.
884  */
885 int
886 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
887 {
888 	struct mlx5_priv *priv = dev->data->dev_private;
889 	struct mlx5_flex_parser_profiles *prf =
890 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
891 	struct mlx5_devx_graph_node_attr node = {
892 		.modify_field_select = 0,
893 	};
894 	uint32_t ids[8];
895 	int ret;
896 
897 	if (!priv->config.hca_attr.parse_graph_flex_node) {
898 		DRV_LOG(ERR, "Dynamic flex parser is not supported "
899 			"for device %s.", priv->dev_data->name);
900 		return -ENOTSUP;
901 	}
902 	node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
903 	/* 8 bytes now: 4B common header + 4B message body header. */
904 	node.header_length_base_value = 0x8;
905 	/* After MAC layer: Ether / VLAN. */
906 	node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
907 	/* Type of compared condition should be 0xAEFE in the L2 layer. */
908 	node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
909 	/* Sample #0: type in common header. */
910 	node.sample[0].flow_match_sample_en = 1;
911 	/* Fixed offset. */
912 	node.sample[0].flow_match_sample_offset_mode = 0x0;
913 	/* Only the 2nd byte will be used. */
914 	node.sample[0].flow_match_sample_field_base_offset = 0x0;
915 	/* Sample #1: message payload. */
916 	node.sample[1].flow_match_sample_en = 1;
917 	/* Fixed offset. */
918 	node.sample[1].flow_match_sample_offset_mode = 0x0;
919 	/*
920 	 * Only the first two bytes will be used right now, and its offset will
921 	 * start after the common header that with the length of a DW(u32).
922 	 */
923 	node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
924 	prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->cdev->ctx, &node);
925 	if (!prf->obj) {
926 		DRV_LOG(ERR, "Failed to create flex parser node object.");
927 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
928 	}
929 	prf->num = 2;
930 	ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
931 	if (ret) {
932 		DRV_LOG(ERR, "Failed to query sample IDs.");
933 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
934 	}
935 	prf->offset[0] = 0x0;
936 	prf->offset[1] = sizeof(uint32_t);
937 	prf->ids[0] = ids[0];
938 	prf->ids[1] = ids[1];
939 	return 0;
940 }
941 
942 /*
943  * Destroy the flex parser node, including the parser itself, input / output
944  * arcs and DW samples. Resources could be reused then.
945  *
946  * @param dev
947  *   Pointer to Ethernet device structure.
948  */
949 static void
950 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
951 {
952 	struct mlx5_priv *priv = dev->data->dev_private;
953 	struct mlx5_flex_parser_profiles *prf =
954 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
955 
956 	if (prf->obj)
957 		mlx5_devx_cmd_destroy(prf->obj);
958 	prf->obj = NULL;
959 }
960 
961 uint32_t
962 mlx5_get_supported_sw_parsing_offloads(const struct mlx5_hca_attr *attr)
963 {
964 	uint32_t sw_parsing_offloads = 0;
965 
966 	if (attr->swp) {
967 		sw_parsing_offloads |= MLX5_SW_PARSING_CAP;
968 		if (attr->swp_csum)
969 			sw_parsing_offloads |= MLX5_SW_PARSING_CSUM_CAP;
970 
971 		if (attr->swp_lso)
972 			sw_parsing_offloads |= MLX5_SW_PARSING_TSO_CAP;
973 	}
974 	return sw_parsing_offloads;
975 }
976 
977 uint32_t
978 mlx5_get_supported_tunneling_offloads(const struct mlx5_hca_attr *attr)
979 {
980 	uint32_t tn_offloads = 0;
981 
982 	if (attr->tunnel_stateless_vxlan)
983 		tn_offloads |= MLX5_TUNNELED_OFFLOADS_VXLAN_CAP;
984 	if (attr->tunnel_stateless_gre)
985 		tn_offloads |= MLX5_TUNNELED_OFFLOADS_GRE_CAP;
986 	if (attr->tunnel_stateless_geneve_rx)
987 		tn_offloads |= MLX5_TUNNELED_OFFLOADS_GENEVE_CAP;
988 	return tn_offloads;
989 }
990 
991 /*
992  * Allocate Rx and Tx UARs in robust fashion.
993  * This routine handles the following UAR allocation issues:
994  *
995  *  - tries to allocate the UAR with the most appropriate memory
996  *    mapping type from the ones supported by the host
997  *
998  *  - tries to allocate the UAR with non-NULL base address
999  *    OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
1000  *    UAR base address if UAR was not the first object in the UAR page.
1001  *    It caused the PMD failure and we should try to get another UAR
1002  *    till we get the first one with non-NULL base address returned.
1003  */
1004 static int
1005 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
1006 		     const struct mlx5_common_dev_config *config)
1007 {
1008 	uint32_t uar_mapping, retry;
1009 	int err = 0;
1010 	void *base_addr;
1011 
1012 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
1013 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1014 		/* Control the mapping type according to the settings. */
1015 		uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
1016 			      MLX5DV_UAR_ALLOC_TYPE_NC :
1017 			      MLX5DV_UAR_ALLOC_TYPE_BF;
1018 #else
1019 		RTE_SET_USED(config);
1020 		/*
1021 		 * It seems we have no way to control the memory mapping type
1022 		 * for the UAR, the default "Write-Combining" type is supposed.
1023 		 * The UAR initialization on queue creation queries the
1024 		 * actual mapping type done by Verbs/kernel and setups the
1025 		 * PMD datapath accordingly.
1026 		 */
1027 		uar_mapping = 0;
1028 #endif
1029 		sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1030 						       uar_mapping);
1031 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1032 		if (!sh->tx_uar &&
1033 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
1034 			if (config->dbnc == MLX5_TXDB_CACHED ||
1035 			    config->dbnc == MLX5_TXDB_HEURISTIC)
1036 				DRV_LOG(WARNING, "Devarg tx_db_nc setting "
1037 						 "is not supported by DevX");
1038 			/*
1039 			 * In some environments like virtual machine
1040 			 * the Write Combining mapped might be not supported
1041 			 * and UAR allocation fails. We try "Non-Cached"
1042 			 * mapping for the case. The tx_burst routines take
1043 			 * the UAR mapping type into account on UAR setup
1044 			 * on queue creation.
1045 			 */
1046 			DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (BF)");
1047 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1048 			sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1049 							       uar_mapping);
1050 		} else if (!sh->tx_uar &&
1051 			   uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
1052 			if (config->dbnc == MLX5_TXDB_NCACHED)
1053 				DRV_LOG(WARNING, "Devarg tx_db_nc settings "
1054 						 "is not supported by DevX");
1055 			/*
1056 			 * If Verbs/kernel does not support "Non-Cached"
1057 			 * try the "Write-Combining".
1058 			 */
1059 			DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (NC)");
1060 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
1061 			sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1062 							       uar_mapping);
1063 		}
1064 #endif
1065 		if (!sh->tx_uar) {
1066 			DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
1067 			err = ENOMEM;
1068 			goto exit;
1069 		}
1070 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
1071 		if (base_addr)
1072 			break;
1073 		/*
1074 		 * The UARs are allocated by rdma_core within the
1075 		 * IB device context, on context closure all UARs
1076 		 * will be freed, should be no memory/object leakage.
1077 		 */
1078 		DRV_LOG(DEBUG, "Retrying to allocate Tx DevX UAR");
1079 		sh->tx_uar = NULL;
1080 	}
1081 	/* Check whether we finally succeeded with valid UAR allocation. */
1082 	if (!sh->tx_uar) {
1083 		DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
1084 		err = ENOMEM;
1085 		goto exit;
1086 	}
1087 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
1088 		uar_mapping = 0;
1089 		sh->devx_rx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1090 							    uar_mapping);
1091 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1092 		if (!sh->devx_rx_uar &&
1093 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
1094 			/*
1095 			 * Rx UAR is used to control interrupts only,
1096 			 * should be no datapath noticeable impact,
1097 			 * can try "Non-Cached" mapping safely.
1098 			 */
1099 			DRV_LOG(DEBUG, "Failed to allocate Rx DevX UAR (BF)");
1100 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1101 			sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
1102 						   (sh->cdev->ctx, uar_mapping);
1103 		}
1104 #endif
1105 		if (!sh->devx_rx_uar) {
1106 			DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
1107 			err = ENOMEM;
1108 			goto exit;
1109 		}
1110 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
1111 		if (base_addr)
1112 			break;
1113 		/*
1114 		 * The UARs are allocated by rdma_core within the
1115 		 * IB device context, on context closure all UARs
1116 		 * will be freed, should be no memory/object leakage.
1117 		 */
1118 		DRV_LOG(DEBUG, "Retrying to allocate Rx DevX UAR");
1119 		sh->devx_rx_uar = NULL;
1120 	}
1121 	/* Check whether we finally succeeded with valid UAR allocation. */
1122 	if (!sh->devx_rx_uar) {
1123 		DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
1124 		err = ENOMEM;
1125 	}
1126 exit:
1127 	return err;
1128 }
1129 
1130 /**
1131  * rte_mempool_walk() callback to unregister Rx mempools.
1132  * It used when implicit mempool registration is disabled.
1133  *
1134  * @param mp
1135  *   The mempool being walked.
1136  * @param arg
1137  *   Pointer to the device shared context.
1138  */
1139 static void
1140 mlx5_dev_ctx_shared_rx_mempool_unregister_cb(struct rte_mempool *mp, void *arg)
1141 {
1142 	struct mlx5_dev_ctx_shared *sh = arg;
1143 
1144 	mlx5_dev_mempool_unregister(sh->cdev, mp);
1145 }
1146 
1147 /**
1148  * Callback used when implicit mempool registration is disabled
1149  * in order to track Rx mempool destruction.
1150  *
1151  * @param event
1152  *   Mempool life cycle event.
1153  * @param mp
1154  *   An Rx mempool registered explicitly when the port is started.
1155  * @param arg
1156  *   Pointer to a device shared context.
1157  */
1158 static void
1159 mlx5_dev_ctx_shared_rx_mempool_event_cb(enum rte_mempool_event event,
1160 					struct rte_mempool *mp, void *arg)
1161 {
1162 	struct mlx5_dev_ctx_shared *sh = arg;
1163 
1164 	if (event == RTE_MEMPOOL_EVENT_DESTROY)
1165 		mlx5_dev_mempool_unregister(sh->cdev, mp);
1166 }
1167 
1168 int
1169 mlx5_dev_ctx_shared_mempool_subscribe(struct rte_eth_dev *dev)
1170 {
1171 	struct mlx5_priv *priv = dev->data->dev_private;
1172 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1173 	int ret;
1174 
1175 	/* Check if we only need to track Rx mempool destruction. */
1176 	if (!sh->cdev->config.mr_mempool_reg_en) {
1177 		ret = rte_mempool_event_callback_register
1178 				(mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1179 		return ret == 0 || rte_errno == EEXIST ? 0 : ret;
1180 	}
1181 	return mlx5_dev_mempool_subscribe(sh->cdev);
1182 }
1183 
1184 /**
1185  * Set up multiple TISs with different affinities according to
1186  * number of bonding ports
1187  *
1188  * @param priv
1189  * Pointer of shared context.
1190  *
1191  * @return
1192  * Zero on success, -1 otherwise.
1193  */
1194 static int
1195 mlx5_setup_tis(struct mlx5_dev_ctx_shared *sh)
1196 {
1197 	int i;
1198 	struct mlx5_devx_lag_context lag_ctx = { 0 };
1199 	struct mlx5_devx_tis_attr tis_attr = { 0 };
1200 
1201 	tis_attr.transport_domain = sh->td->id;
1202 	if (sh->bond.n_port) {
1203 		if (!mlx5_devx_cmd_query_lag(sh->cdev->ctx, &lag_ctx)) {
1204 			sh->lag.tx_remap_affinity[0] =
1205 				lag_ctx.tx_remap_affinity_1;
1206 			sh->lag.tx_remap_affinity[1] =
1207 				lag_ctx.tx_remap_affinity_2;
1208 			sh->lag.affinity_mode = lag_ctx.port_select_mode;
1209 		} else {
1210 			DRV_LOG(ERR, "Failed to query lag affinity.");
1211 			return -1;
1212 		}
1213 		if (sh->lag.affinity_mode == MLX5_LAG_MODE_TIS) {
1214 			for (i = 0; i < sh->bond.n_port; i++) {
1215 				tis_attr.lag_tx_port_affinity =
1216 					MLX5_IFC_LAG_MAP_TIS_AFFINITY(i,
1217 							sh->bond.n_port);
1218 				sh->tis[i] = mlx5_devx_cmd_create_tis(sh->cdev->ctx,
1219 						&tis_attr);
1220 				if (!sh->tis[i]) {
1221 					DRV_LOG(ERR, "Failed to TIS %d/%d for bonding device"
1222 						" %s.", i, sh->bond.n_port,
1223 						sh->ibdev_name);
1224 					return -1;
1225 				}
1226 			}
1227 			DRV_LOG(DEBUG, "LAG number of ports : %d, affinity_1 & 2 : pf%d & %d.\n",
1228 				sh->bond.n_port, lag_ctx.tx_remap_affinity_1,
1229 				lag_ctx.tx_remap_affinity_2);
1230 			return 0;
1231 		}
1232 		if (sh->lag.affinity_mode == MLX5_LAG_MODE_HASH)
1233 			DRV_LOG(INFO, "Device %s enabled HW hash based LAG.",
1234 					sh->ibdev_name);
1235 	}
1236 	tis_attr.lag_tx_port_affinity = 0;
1237 	sh->tis[0] = mlx5_devx_cmd_create_tis(sh->cdev->ctx, &tis_attr);
1238 	if (!sh->tis[0]) {
1239 		DRV_LOG(ERR, "Failed to TIS 0 for bonding device"
1240 			" %s.", sh->ibdev_name);
1241 		return -1;
1242 	}
1243 	return 0;
1244 }
1245 
1246 /**
1247  * Allocate shared device context. If there is multiport device the
1248  * master and representors will share this context, if there is single
1249  * port dedicated device, the context will be used by only given
1250  * port due to unification.
1251  *
1252  * Routine first searches the context for the specified device name,
1253  * if found the shared context assumed and reference counter is incremented.
1254  * If no context found the new one is created and initialized with specified
1255  * device context and parameters.
1256  *
1257  * @param[in] spawn
1258  *   Pointer to the device attributes (name, port, etc).
1259  * @param[in] config
1260  *   Pointer to device configuration structure.
1261  *
1262  * @return
1263  *   Pointer to mlx5_dev_ctx_shared object on success,
1264  *   otherwise NULL and rte_errno is set.
1265  */
1266 struct mlx5_dev_ctx_shared *
1267 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
1268 			  const struct mlx5_dev_config *config)
1269 {
1270 	struct mlx5_dev_ctx_shared *sh;
1271 	int err = 0;
1272 	uint32_t i;
1273 
1274 	MLX5_ASSERT(spawn);
1275 	/* Secondary process should not create the shared context. */
1276 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1277 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1278 	/* Search for IB context by device name. */
1279 	LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
1280 		if (!strcmp(sh->ibdev_name, spawn->phys_dev_name)) {
1281 			sh->refcnt++;
1282 			goto exit;
1283 		}
1284 	}
1285 	/* No device found, we have to create new shared context. */
1286 	MLX5_ASSERT(spawn->max_port);
1287 	sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1288 			 sizeof(struct mlx5_dev_ctx_shared) +
1289 			 spawn->max_port *
1290 			 sizeof(struct mlx5_dev_shared_port),
1291 			 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1292 	if (!sh) {
1293 		DRV_LOG(ERR, "shared context allocation failure");
1294 		rte_errno  = ENOMEM;
1295 		goto exit;
1296 	}
1297 	pthread_mutex_init(&sh->txpp.mutex, NULL);
1298 	sh->numa_node = spawn->cdev->dev->numa_node;
1299 	sh->cdev = spawn->cdev;
1300 	sh->devx = sh->cdev->config.devx;
1301 	if (spawn->bond_info)
1302 		sh->bond = *spawn->bond_info;
1303 	err = mlx5_os_get_dev_attr(sh->cdev, &sh->device_attr);
1304 	if (err) {
1305 		DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
1306 		goto error;
1307 	}
1308 	sh->refcnt = 1;
1309 	sh->max_port = spawn->max_port;
1310 	sh->reclaim_mode = config->reclaim_mode;
1311 	strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->cdev->ctx),
1312 		sizeof(sh->ibdev_name) - 1);
1313 	strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->cdev->ctx),
1314 		sizeof(sh->ibdev_path) - 1);
1315 	/*
1316 	 * Setting port_id to max unallowed value means
1317 	 * there is no interrupt subhandler installed for
1318 	 * the given port index i.
1319 	 */
1320 	for (i = 0; i < sh->max_port; i++) {
1321 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
1322 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
1323 	}
1324 	if (sh->devx) {
1325 		sh->td = mlx5_devx_cmd_create_td(sh->cdev->ctx);
1326 		if (!sh->td) {
1327 			DRV_LOG(ERR, "TD allocation failure");
1328 			err = ENOMEM;
1329 			goto error;
1330 		}
1331 		if (mlx5_setup_tis(sh)) {
1332 			DRV_LOG(ERR, "TIS allocation failure");
1333 			err = ENOMEM;
1334 			goto error;
1335 		}
1336 		err = mlx5_alloc_rxtx_uars(sh, &sh->cdev->config);
1337 		if (err)
1338 			goto error;
1339 		MLX5_ASSERT(sh->tx_uar);
1340 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
1341 
1342 		MLX5_ASSERT(sh->devx_rx_uar);
1343 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
1344 	}
1345 #ifndef RTE_ARCH_64
1346 	/* Initialize UAR access locks for 32bit implementations. */
1347 	rte_spinlock_init(&sh->uar_lock_cq);
1348 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1349 		rte_spinlock_init(&sh->uar_lock[i]);
1350 #endif
1351 	mlx5_os_dev_shared_handler_install(sh);
1352 	if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1353 		err = mlx5_flow_os_init_workspace_once();
1354 		if (err)
1355 			goto error;
1356 	}
1357 	mlx5_flow_aging_init(sh);
1358 	mlx5_flow_counters_mng_init(sh);
1359 	mlx5_flow_ipool_create(sh, config);
1360 	/* Add context to the global device list. */
1361 	LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1362 	rte_spinlock_init(&sh->geneve_tlv_opt_sl);
1363 exit:
1364 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1365 	return sh;
1366 error:
1367 	pthread_mutex_destroy(&sh->txpp.mutex);
1368 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1369 	MLX5_ASSERT(sh);
1370 	if (sh->td)
1371 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1372 	i = 0;
1373 	do {
1374 		if (sh->tis[i])
1375 			claim_zero(mlx5_devx_cmd_destroy(sh->tis[i]));
1376 	} while (++i < (uint32_t)sh->bond.n_port);
1377 	if (sh->devx_rx_uar)
1378 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1379 	if (sh->tx_uar)
1380 		mlx5_glue->devx_free_uar(sh->tx_uar);
1381 	mlx5_free(sh);
1382 	MLX5_ASSERT(err > 0);
1383 	rte_errno = err;
1384 	return NULL;
1385 }
1386 
1387 /**
1388  * Free shared IB device context. Decrement counter and if zero free
1389  * all allocated resources and close handles.
1390  *
1391  * @param[in] sh
1392  *   Pointer to mlx5_dev_ctx_shared object to free
1393  */
1394 void
1395 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1396 {
1397 	int ret;
1398 	int i = 0;
1399 
1400 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1401 #ifdef RTE_LIBRTE_MLX5_DEBUG
1402 	/* Check the object presence in the list. */
1403 	struct mlx5_dev_ctx_shared *lctx;
1404 
1405 	LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1406 		if (lctx == sh)
1407 			break;
1408 	MLX5_ASSERT(lctx);
1409 	if (lctx != sh) {
1410 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
1411 		goto exit;
1412 	}
1413 #endif
1414 	MLX5_ASSERT(sh);
1415 	MLX5_ASSERT(sh->refcnt);
1416 	/* Secondary process should not free the shared context. */
1417 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1418 	if (--sh->refcnt)
1419 		goto exit;
1420 	/* Stop watching for mempool events and unregister all mempools. */
1421 	if (!sh->cdev->config.mr_mempool_reg_en) {
1422 		ret = rte_mempool_event_callback_unregister
1423 				(mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1424 		if (ret == 0)
1425 			rte_mempool_walk
1426 			     (mlx5_dev_ctx_shared_rx_mempool_unregister_cb, sh);
1427 	}
1428 	/* Remove context from the global device list. */
1429 	LIST_REMOVE(sh, next);
1430 	/* Release resources on the last device removal. */
1431 	if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1432 		mlx5_os_net_cleanup();
1433 		mlx5_flow_os_release_workspace();
1434 	}
1435 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1436 	/*
1437 	 *  Ensure there is no async event handler installed.
1438 	 *  Only primary process handles async device events.
1439 	 **/
1440 	mlx5_flow_counters_mng_close(sh);
1441 	if (sh->aso_age_mng) {
1442 		mlx5_flow_aso_age_mng_close(sh);
1443 		sh->aso_age_mng = NULL;
1444 	}
1445 	if (sh->mtrmng)
1446 		mlx5_aso_flow_mtrs_mng_close(sh);
1447 	mlx5_flow_ipool_destroy(sh);
1448 	mlx5_os_dev_shared_handler_uninstall(sh);
1449 	if (sh->tx_uar) {
1450 		mlx5_glue->devx_free_uar(sh->tx_uar);
1451 		sh->tx_uar = NULL;
1452 	}
1453 	do {
1454 		if (sh->tis[i])
1455 			claim_zero(mlx5_devx_cmd_destroy(sh->tis[i]));
1456 	} while (++i < sh->bond.n_port);
1457 	if (sh->td)
1458 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1459 	if (sh->devx_rx_uar)
1460 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1461 	MLX5_ASSERT(sh->geneve_tlv_option_resource == NULL);
1462 	pthread_mutex_destroy(&sh->txpp.mutex);
1463 	mlx5_free(sh);
1464 	return;
1465 exit:
1466 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1467 }
1468 
1469 /**
1470  * Destroy table hash list.
1471  *
1472  * @param[in] priv
1473  *   Pointer to the private device data structure.
1474  */
1475 void
1476 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1477 {
1478 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1479 
1480 	if (!sh->flow_tbls)
1481 		return;
1482 	mlx5_hlist_destroy(sh->flow_tbls);
1483 	sh->flow_tbls = NULL;
1484 }
1485 
1486 /**
1487  * Initialize flow table hash list and create the root tables entry
1488  * for each domain.
1489  *
1490  * @param[in] priv
1491  *   Pointer to the private device data structure.
1492  *
1493  * @return
1494  *   Zero on success, positive error code otherwise.
1495  */
1496 int
1497 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1498 {
1499 	int err = 0;
1500 	/* Tables are only used in DV and DR modes. */
1501 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1502 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1503 	char s[MLX5_NAME_SIZE];
1504 
1505 	MLX5_ASSERT(sh);
1506 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1507 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1508 					  false, true, sh,
1509 					  flow_dv_tbl_create_cb,
1510 					  flow_dv_tbl_match_cb,
1511 					  flow_dv_tbl_remove_cb,
1512 					  flow_dv_tbl_clone_cb,
1513 					  flow_dv_tbl_clone_free_cb);
1514 	if (!sh->flow_tbls) {
1515 		DRV_LOG(ERR, "flow tables with hash creation failed.");
1516 		err = ENOMEM;
1517 		return err;
1518 	}
1519 #ifndef HAVE_MLX5DV_DR
1520 	struct rte_flow_error error;
1521 	struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1522 
1523 	/*
1524 	 * In case we have not DR support, the zero tables should be created
1525 	 * because DV expect to see them even if they cannot be created by
1526 	 * RDMA-CORE.
1527 	 */
1528 	if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0,
1529 		NULL, 0, 1, 0, &error) ||
1530 	    !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0,
1531 		NULL, 0, 1, 0, &error) ||
1532 	    !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0,
1533 		NULL, 0, 1, 0, &error)) {
1534 		err = ENOMEM;
1535 		goto error;
1536 	}
1537 	return err;
1538 error:
1539 	mlx5_free_table_hash_list(priv);
1540 #endif /* HAVE_MLX5DV_DR */
1541 #endif
1542 	return err;
1543 }
1544 
1545 /**
1546  * Retrieve integer value from environment variable.
1547  *
1548  * @param[in] name
1549  *   Environment variable name.
1550  *
1551  * @return
1552  *   Integer value, 0 if the variable is not set.
1553  */
1554 int
1555 mlx5_getenv_int(const char *name)
1556 {
1557 	const char *val = getenv(name);
1558 
1559 	if (val == NULL)
1560 		return 0;
1561 	return atoi(val);
1562 }
1563 
1564 /**
1565  * DPDK callback to add udp tunnel port
1566  *
1567  * @param[in] dev
1568  *   A pointer to eth_dev
1569  * @param[in] udp_tunnel
1570  *   A pointer to udp tunnel
1571  *
1572  * @return
1573  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1574  */
1575 int
1576 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1577 			 struct rte_eth_udp_tunnel *udp_tunnel)
1578 {
1579 	MLX5_ASSERT(udp_tunnel != NULL);
1580 	if (udp_tunnel->prot_type == RTE_ETH_TUNNEL_TYPE_VXLAN &&
1581 	    udp_tunnel->udp_port == 4789)
1582 		return 0;
1583 	if (udp_tunnel->prot_type == RTE_ETH_TUNNEL_TYPE_VXLAN_GPE &&
1584 	    udp_tunnel->udp_port == 4790)
1585 		return 0;
1586 	return -ENOTSUP;
1587 }
1588 
1589 /**
1590  * Initialize process private data structure.
1591  *
1592  * @param dev
1593  *   Pointer to Ethernet device structure.
1594  *
1595  * @return
1596  *   0 on success, a negative errno value otherwise and rte_errno is set.
1597  */
1598 int
1599 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1600 {
1601 	struct mlx5_priv *priv = dev->data->dev_private;
1602 	struct mlx5_proc_priv *ppriv;
1603 	size_t ppriv_size;
1604 
1605 	mlx5_proc_priv_uninit(dev);
1606 	/*
1607 	 * UAR register table follows the process private structure. BlueFlame
1608 	 * registers for Tx queues are stored in the table.
1609 	 */
1610 	ppriv_size =
1611 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1612 	ppriv = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, ppriv_size,
1613 			    RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1614 	if (!ppriv) {
1615 		rte_errno = ENOMEM;
1616 		return -rte_errno;
1617 	}
1618 	ppriv->uar_table_sz = priv->txqs_n;
1619 	dev->process_private = ppriv;
1620 	return 0;
1621 }
1622 
1623 /**
1624  * Un-initialize process private data structure.
1625  *
1626  * @param dev
1627  *   Pointer to Ethernet device structure.
1628  */
1629 void
1630 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1631 {
1632 	if (!dev->process_private)
1633 		return;
1634 	mlx5_free(dev->process_private);
1635 	dev->process_private = NULL;
1636 }
1637 
1638 /**
1639  * DPDK callback to close the device.
1640  *
1641  * Destroy all queues and objects, free memory.
1642  *
1643  * @param dev
1644  *   Pointer to Ethernet device structure.
1645  */
1646 int
1647 mlx5_dev_close(struct rte_eth_dev *dev)
1648 {
1649 	struct mlx5_priv *priv = dev->data->dev_private;
1650 	unsigned int i;
1651 	int ret;
1652 
1653 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1654 		/* Check if process_private released. */
1655 		if (!dev->process_private)
1656 			return 0;
1657 		mlx5_tx_uar_uninit_secondary(dev);
1658 		mlx5_proc_priv_uninit(dev);
1659 		rte_eth_dev_release_port(dev);
1660 		return 0;
1661 	}
1662 	if (!priv->sh)
1663 		return 0;
1664 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1665 		dev->data->port_id,
1666 		((priv->sh->cdev->ctx != NULL) ?
1667 		mlx5_os_get_ctx_device_name(priv->sh->cdev->ctx) : ""));
1668 	/*
1669 	 * If default mreg copy action is removed at the stop stage,
1670 	 * the search will return none and nothing will be done anymore.
1671 	 */
1672 	mlx5_flow_stop_default(dev);
1673 	mlx5_traffic_disable(dev);
1674 	/*
1675 	 * If all the flows are already flushed in the device stop stage,
1676 	 * then this will return directly without any action.
1677 	 */
1678 	mlx5_flow_list_flush(dev, MLX5_FLOW_TYPE_GEN, true);
1679 	mlx5_action_handle_flush(dev);
1680 	mlx5_flow_meter_flush(dev, NULL);
1681 	/* Prevent crashes when queues are still in use. */
1682 	dev->rx_pkt_burst = removed_rx_burst;
1683 	dev->tx_pkt_burst = removed_tx_burst;
1684 	rte_wmb();
1685 	/* Disable datapath on secondary process. */
1686 	mlx5_mp_os_req_stop_rxtx(dev);
1687 	/* Free the eCPRI flex parser resource. */
1688 	mlx5_flex_parser_ecpri_release(dev);
1689 	if (priv->rxqs != NULL) {
1690 		/* XXX race condition if mlx5_rx_burst() is still running. */
1691 		rte_delay_us_sleep(1000);
1692 		for (i = 0; (i != priv->rxqs_n); ++i)
1693 			mlx5_rxq_release(dev, i);
1694 		priv->rxqs_n = 0;
1695 		priv->rxqs = NULL;
1696 	}
1697 	if (priv->representor) {
1698 		/* Each representor has a dedicated interrupts handler */
1699 		mlx5_free(dev->intr_handle);
1700 		dev->intr_handle = NULL;
1701 	}
1702 	if (priv->txqs != NULL) {
1703 		/* XXX race condition if mlx5_tx_burst() is still running. */
1704 		rte_delay_us_sleep(1000);
1705 		for (i = 0; (i != priv->txqs_n); ++i)
1706 			mlx5_txq_release(dev, i);
1707 		priv->txqs_n = 0;
1708 		priv->txqs = NULL;
1709 	}
1710 	mlx5_proc_priv_uninit(dev);
1711 	if (priv->q_counters) {
1712 		mlx5_devx_cmd_destroy(priv->q_counters);
1713 		priv->q_counters = NULL;
1714 	}
1715 	if (priv->drop_queue.hrxq)
1716 		mlx5_drop_action_destroy(dev);
1717 	if (priv->mreg_cp_tbl)
1718 		mlx5_hlist_destroy(priv->mreg_cp_tbl);
1719 	mlx5_mprq_free_mp(dev);
1720 	if (priv->sh->ct_mng)
1721 		mlx5_flow_aso_ct_mng_close(priv->sh);
1722 	mlx5_os_free_shared_dr(priv);
1723 	if (priv->rss_conf.rss_key != NULL)
1724 		mlx5_free(priv->rss_conf.rss_key);
1725 	if (priv->reta_idx != NULL)
1726 		mlx5_free(priv->reta_idx);
1727 	if (priv->config.vf)
1728 		mlx5_os_mac_addr_flush(dev);
1729 	if (priv->nl_socket_route >= 0)
1730 		close(priv->nl_socket_route);
1731 	if (priv->nl_socket_rdma >= 0)
1732 		close(priv->nl_socket_rdma);
1733 	if (priv->vmwa_context)
1734 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1735 	ret = mlx5_hrxq_verify(dev);
1736 	if (ret)
1737 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1738 			dev->data->port_id);
1739 	ret = mlx5_ind_table_obj_verify(dev);
1740 	if (ret)
1741 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1742 			dev->data->port_id);
1743 	ret = mlx5_rxq_obj_verify(dev);
1744 	if (ret)
1745 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1746 			dev->data->port_id);
1747 	ret = mlx5_rxq_verify(dev);
1748 	if (ret)
1749 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1750 			dev->data->port_id);
1751 	ret = mlx5_txq_obj_verify(dev);
1752 	if (ret)
1753 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1754 			dev->data->port_id);
1755 	ret = mlx5_txq_verify(dev);
1756 	if (ret)
1757 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1758 			dev->data->port_id);
1759 	ret = mlx5_flow_verify(dev);
1760 	if (ret)
1761 		DRV_LOG(WARNING, "port %u some flows still remain",
1762 			dev->data->port_id);
1763 	if (priv->hrxqs)
1764 		mlx5_list_destroy(priv->hrxqs);
1765 	/*
1766 	 * Free the shared context in last turn, because the cleanup
1767 	 * routines above may use some shared fields, like
1768 	 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1769 	 * ifindex if Netlink fails.
1770 	 */
1771 	mlx5_free_shared_dev_ctx(priv->sh);
1772 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1773 		unsigned int c = 0;
1774 		uint16_t port_id;
1775 
1776 		MLX5_ETH_FOREACH_DEV(port_id, dev->device) {
1777 			struct mlx5_priv *opriv =
1778 				rte_eth_devices[port_id].data->dev_private;
1779 
1780 			if (!opriv ||
1781 			    opriv->domain_id != priv->domain_id ||
1782 			    &rte_eth_devices[port_id] == dev)
1783 				continue;
1784 			++c;
1785 			break;
1786 		}
1787 		if (!c)
1788 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1789 	}
1790 	memset(priv, 0, sizeof(*priv));
1791 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1792 	/*
1793 	 * Reset mac_addrs to NULL such that it is not freed as part of
1794 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1795 	 * it is freed when dev_private is freed.
1796 	 */
1797 	dev->data->mac_addrs = NULL;
1798 	return 0;
1799 }
1800 
1801 const struct eth_dev_ops mlx5_dev_ops = {
1802 	.dev_configure = mlx5_dev_configure,
1803 	.dev_start = mlx5_dev_start,
1804 	.dev_stop = mlx5_dev_stop,
1805 	.dev_set_link_down = mlx5_set_link_down,
1806 	.dev_set_link_up = mlx5_set_link_up,
1807 	.dev_close = mlx5_dev_close,
1808 	.promiscuous_enable = mlx5_promiscuous_enable,
1809 	.promiscuous_disable = mlx5_promiscuous_disable,
1810 	.allmulticast_enable = mlx5_allmulticast_enable,
1811 	.allmulticast_disable = mlx5_allmulticast_disable,
1812 	.link_update = mlx5_link_update,
1813 	.stats_get = mlx5_stats_get,
1814 	.stats_reset = mlx5_stats_reset,
1815 	.xstats_get = mlx5_xstats_get,
1816 	.xstats_reset = mlx5_xstats_reset,
1817 	.xstats_get_names = mlx5_xstats_get_names,
1818 	.fw_version_get = mlx5_fw_version_get,
1819 	.dev_infos_get = mlx5_dev_infos_get,
1820 	.representor_info_get = mlx5_representor_info_get,
1821 	.read_clock = mlx5_txpp_read_clock,
1822 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1823 	.vlan_filter_set = mlx5_vlan_filter_set,
1824 	.rx_queue_setup = mlx5_rx_queue_setup,
1825 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1826 	.tx_queue_setup = mlx5_tx_queue_setup,
1827 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1828 	.rx_queue_release = mlx5_rx_queue_release,
1829 	.tx_queue_release = mlx5_tx_queue_release,
1830 	.rx_queue_start = mlx5_rx_queue_start,
1831 	.rx_queue_stop = mlx5_rx_queue_stop,
1832 	.tx_queue_start = mlx5_tx_queue_start,
1833 	.tx_queue_stop = mlx5_tx_queue_stop,
1834 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1835 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1836 	.mac_addr_remove = mlx5_mac_addr_remove,
1837 	.mac_addr_add = mlx5_mac_addr_add,
1838 	.mac_addr_set = mlx5_mac_addr_set,
1839 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1840 	.mtu_set = mlx5_dev_set_mtu,
1841 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1842 	.vlan_offload_set = mlx5_vlan_offload_set,
1843 	.reta_update = mlx5_dev_rss_reta_update,
1844 	.reta_query = mlx5_dev_rss_reta_query,
1845 	.rss_hash_update = mlx5_rss_hash_update,
1846 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1847 	.flow_ops_get = mlx5_flow_ops_get,
1848 	.rxq_info_get = mlx5_rxq_info_get,
1849 	.txq_info_get = mlx5_txq_info_get,
1850 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1851 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1852 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1853 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1854 	.is_removed = mlx5_is_removed,
1855 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1856 	.get_module_info = mlx5_get_module_info,
1857 	.get_module_eeprom = mlx5_get_module_eeprom,
1858 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1859 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1860 	.hairpin_bind = mlx5_hairpin_bind,
1861 	.hairpin_unbind = mlx5_hairpin_unbind,
1862 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1863 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1864 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1865 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1866 	.get_monitor_addr = mlx5_get_monitor_addr,
1867 };
1868 
1869 /* Available operations from secondary process. */
1870 const struct eth_dev_ops mlx5_dev_sec_ops = {
1871 	.stats_get = mlx5_stats_get,
1872 	.stats_reset = mlx5_stats_reset,
1873 	.xstats_get = mlx5_xstats_get,
1874 	.xstats_reset = mlx5_xstats_reset,
1875 	.xstats_get_names = mlx5_xstats_get_names,
1876 	.fw_version_get = mlx5_fw_version_get,
1877 	.dev_infos_get = mlx5_dev_infos_get,
1878 	.representor_info_get = mlx5_representor_info_get,
1879 	.read_clock = mlx5_txpp_read_clock,
1880 	.rx_queue_start = mlx5_rx_queue_start,
1881 	.rx_queue_stop = mlx5_rx_queue_stop,
1882 	.tx_queue_start = mlx5_tx_queue_start,
1883 	.tx_queue_stop = mlx5_tx_queue_stop,
1884 	.rxq_info_get = mlx5_rxq_info_get,
1885 	.txq_info_get = mlx5_txq_info_get,
1886 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1887 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1888 	.get_module_info = mlx5_get_module_info,
1889 	.get_module_eeprom = mlx5_get_module_eeprom,
1890 };
1891 
1892 /* Available operations in flow isolated mode. */
1893 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1894 	.dev_configure = mlx5_dev_configure,
1895 	.dev_start = mlx5_dev_start,
1896 	.dev_stop = mlx5_dev_stop,
1897 	.dev_set_link_down = mlx5_set_link_down,
1898 	.dev_set_link_up = mlx5_set_link_up,
1899 	.dev_close = mlx5_dev_close,
1900 	.promiscuous_enable = mlx5_promiscuous_enable,
1901 	.promiscuous_disable = mlx5_promiscuous_disable,
1902 	.allmulticast_enable = mlx5_allmulticast_enable,
1903 	.allmulticast_disable = mlx5_allmulticast_disable,
1904 	.link_update = mlx5_link_update,
1905 	.stats_get = mlx5_stats_get,
1906 	.stats_reset = mlx5_stats_reset,
1907 	.xstats_get = mlx5_xstats_get,
1908 	.xstats_reset = mlx5_xstats_reset,
1909 	.xstats_get_names = mlx5_xstats_get_names,
1910 	.fw_version_get = mlx5_fw_version_get,
1911 	.dev_infos_get = mlx5_dev_infos_get,
1912 	.representor_info_get = mlx5_representor_info_get,
1913 	.read_clock = mlx5_txpp_read_clock,
1914 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1915 	.vlan_filter_set = mlx5_vlan_filter_set,
1916 	.rx_queue_setup = mlx5_rx_queue_setup,
1917 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1918 	.tx_queue_setup = mlx5_tx_queue_setup,
1919 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1920 	.rx_queue_release = mlx5_rx_queue_release,
1921 	.tx_queue_release = mlx5_tx_queue_release,
1922 	.rx_queue_start = mlx5_rx_queue_start,
1923 	.rx_queue_stop = mlx5_rx_queue_stop,
1924 	.tx_queue_start = mlx5_tx_queue_start,
1925 	.tx_queue_stop = mlx5_tx_queue_stop,
1926 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1927 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1928 	.mac_addr_remove = mlx5_mac_addr_remove,
1929 	.mac_addr_add = mlx5_mac_addr_add,
1930 	.mac_addr_set = mlx5_mac_addr_set,
1931 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1932 	.mtu_set = mlx5_dev_set_mtu,
1933 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1934 	.vlan_offload_set = mlx5_vlan_offload_set,
1935 	.flow_ops_get = mlx5_flow_ops_get,
1936 	.rxq_info_get = mlx5_rxq_info_get,
1937 	.txq_info_get = mlx5_txq_info_get,
1938 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1939 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1940 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1941 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1942 	.is_removed = mlx5_is_removed,
1943 	.get_module_info = mlx5_get_module_info,
1944 	.get_module_eeprom = mlx5_get_module_eeprom,
1945 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1946 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1947 	.hairpin_bind = mlx5_hairpin_bind,
1948 	.hairpin_unbind = mlx5_hairpin_unbind,
1949 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1950 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1951 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1952 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1953 	.get_monitor_addr = mlx5_get_monitor_addr,
1954 };
1955 
1956 /**
1957  * Verify and store value for device argument.
1958  *
1959  * @param[in] key
1960  *   Key argument to verify.
1961  * @param[in] val
1962  *   Value associated with key.
1963  * @param opaque
1964  *   User data.
1965  *
1966  * @return
1967  *   0 on success, a negative errno value otherwise and rte_errno is set.
1968  */
1969 static int
1970 mlx5_args_check(const char *key, const char *val, void *opaque)
1971 {
1972 	struct mlx5_dev_config *config = opaque;
1973 	unsigned long mod;
1974 	signed long tmp;
1975 
1976 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1977 	if (!strcmp(MLX5_DRIVER_KEY, key) || !strcmp(MLX5_REPRESENTOR, key) ||
1978 	    !strcmp(MLX5_SYS_MEM_EN, key) || !strcmp(MLX5_TX_DB_NC, key) ||
1979 	    !strcmp(MLX5_MR_MEMPOOL_REG_EN, key) ||
1980 	    !strcmp(MLX5_MR_EXT_MEMSEG_EN, key))
1981 		return 0;
1982 	errno = 0;
1983 	tmp = strtol(val, NULL, 0);
1984 	if (errno) {
1985 		rte_errno = errno;
1986 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1987 		return -rte_errno;
1988 	}
1989 	if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1990 		/* Negative values are acceptable for some keys only. */
1991 		rte_errno = EINVAL;
1992 		DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1993 		return -rte_errno;
1994 	}
1995 	mod = tmp >= 0 ? tmp : -tmp;
1996 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1997 		if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1998 			DRV_LOG(ERR, "invalid CQE compression "
1999 				     "format parameter");
2000 			rte_errno = EINVAL;
2001 			return -rte_errno;
2002 		}
2003 		config->cqe_comp = !!tmp;
2004 		config->cqe_comp_fmt = tmp;
2005 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
2006 		config->hw_padding = !!tmp;
2007 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
2008 		config->mprq.enabled = !!tmp;
2009 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
2010 		config->mprq.stride_num_n = tmp;
2011 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
2012 		config->mprq.stride_size_n = tmp;
2013 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
2014 		config->mprq.max_memcpy_len = tmp;
2015 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
2016 		config->mprq.min_rxqs_num = tmp;
2017 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
2018 		DRV_LOG(WARNING, "%s: deprecated parameter,"
2019 				 " converted to txq_inline_max", key);
2020 		config->txq_inline_max = tmp;
2021 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
2022 		config->txq_inline_max = tmp;
2023 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
2024 		config->txq_inline_min = tmp;
2025 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
2026 		config->txq_inline_mpw = tmp;
2027 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
2028 		config->txqs_inline = tmp;
2029 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
2030 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2031 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
2032 		config->mps = !!tmp;
2033 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
2034 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2035 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
2036 		DRV_LOG(WARNING, "%s: deprecated parameter,"
2037 				 " converted to txq_inline_mpw", key);
2038 		config->txq_inline_mpw = tmp;
2039 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
2040 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2041 	} else if (strcmp(MLX5_TX_PP, key) == 0) {
2042 		if (!mod) {
2043 			DRV_LOG(ERR, "Zero Tx packet pacing parameter");
2044 			rte_errno = EINVAL;
2045 			return -rte_errno;
2046 		}
2047 		config->tx_pp = tmp;
2048 	} else if (strcmp(MLX5_TX_SKEW, key) == 0) {
2049 		config->tx_skew = tmp;
2050 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
2051 		config->rx_vec_en = !!tmp;
2052 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
2053 		config->l3_vxlan_en = !!tmp;
2054 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
2055 		config->vf_nl_en = !!tmp;
2056 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
2057 		config->dv_esw_en = !!tmp;
2058 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
2059 		config->dv_flow_en = !!tmp;
2060 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
2061 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
2062 		    tmp != MLX5_XMETA_MODE_META16 &&
2063 		    tmp != MLX5_XMETA_MODE_META32 &&
2064 		    tmp != MLX5_XMETA_MODE_MISS_INFO) {
2065 			DRV_LOG(ERR, "invalid extensive "
2066 				     "metadata parameter");
2067 			rte_errno = EINVAL;
2068 			return -rte_errno;
2069 		}
2070 		if (tmp != MLX5_XMETA_MODE_MISS_INFO)
2071 			config->dv_xmeta_en = tmp;
2072 		else
2073 			config->dv_miss_info = 1;
2074 	} else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
2075 		config->lacp_by_user = !!tmp;
2076 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
2077 		config->max_dump_files_num = tmp;
2078 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
2079 		config->lro.timeout = tmp;
2080 	} else if (strcmp(RTE_DEVARGS_KEY_CLASS, key) == 0) {
2081 		DRV_LOG(DEBUG, "class argument is %s.", val);
2082 	} else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
2083 		config->log_hp_size = tmp;
2084 	} else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
2085 		if (tmp != MLX5_RCM_NONE &&
2086 		    tmp != MLX5_RCM_LIGHT &&
2087 		    tmp != MLX5_RCM_AGGR) {
2088 			DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
2089 			rte_errno = EINVAL;
2090 			return -rte_errno;
2091 		}
2092 		config->reclaim_mode = tmp;
2093 	} else if (strcmp(MLX5_DECAP_EN, key) == 0) {
2094 		config->decap_en = !!tmp;
2095 	} else if (strcmp(MLX5_ALLOW_DUPLICATE_PATTERN, key) == 0) {
2096 		config->allow_duplicate_pattern = !!tmp;
2097 	} else {
2098 		DRV_LOG(WARNING, "%s: unknown parameter", key);
2099 		rte_errno = EINVAL;
2100 		return -rte_errno;
2101 	}
2102 	return 0;
2103 }
2104 
2105 /**
2106  * Parse device parameters.
2107  *
2108  * @param config
2109  *   Pointer to device configuration structure.
2110  * @param devargs
2111  *   Device arguments structure.
2112  *
2113  * @return
2114  *   0 on success, a negative errno value otherwise and rte_errno is set.
2115  */
2116 int
2117 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
2118 {
2119 	const char **params = (const char *[]){
2120 		MLX5_DRIVER_KEY,
2121 		MLX5_RXQ_CQE_COMP_EN,
2122 		MLX5_RXQ_PKT_PAD_EN,
2123 		MLX5_RX_MPRQ_EN,
2124 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
2125 		MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
2126 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
2127 		MLX5_RXQS_MIN_MPRQ,
2128 		MLX5_TXQ_INLINE,
2129 		MLX5_TXQ_INLINE_MIN,
2130 		MLX5_TXQ_INLINE_MAX,
2131 		MLX5_TXQ_INLINE_MPW,
2132 		MLX5_TXQS_MIN_INLINE,
2133 		MLX5_TXQS_MAX_VEC,
2134 		MLX5_TXQ_MPW_EN,
2135 		MLX5_TXQ_MPW_HDR_DSEG_EN,
2136 		MLX5_TXQ_MAX_INLINE_LEN,
2137 		MLX5_TX_DB_NC,
2138 		MLX5_TX_PP,
2139 		MLX5_TX_SKEW,
2140 		MLX5_TX_VEC_EN,
2141 		MLX5_RX_VEC_EN,
2142 		MLX5_L3_VXLAN_EN,
2143 		MLX5_VF_NL_EN,
2144 		MLX5_DV_ESW_EN,
2145 		MLX5_DV_FLOW_EN,
2146 		MLX5_DV_XMETA_EN,
2147 		MLX5_LACP_BY_USER,
2148 		MLX5_MR_EXT_MEMSEG_EN,
2149 		MLX5_REPRESENTOR,
2150 		MLX5_MAX_DUMP_FILES_NUM,
2151 		MLX5_LRO_TIMEOUT_USEC,
2152 		RTE_DEVARGS_KEY_CLASS,
2153 		MLX5_HP_BUF_SIZE,
2154 		MLX5_RECLAIM_MEM,
2155 		MLX5_SYS_MEM_EN,
2156 		MLX5_DECAP_EN,
2157 		MLX5_ALLOW_DUPLICATE_PATTERN,
2158 		MLX5_MR_MEMPOOL_REG_EN,
2159 		NULL,
2160 	};
2161 	struct rte_kvargs *kvlist;
2162 	int ret = 0;
2163 	int i;
2164 
2165 	if (devargs == NULL)
2166 		return 0;
2167 	/* Following UGLY cast is done to pass checkpatch. */
2168 	kvlist = rte_kvargs_parse(devargs->args, params);
2169 	if (kvlist == NULL) {
2170 		rte_errno = EINVAL;
2171 		return -rte_errno;
2172 	}
2173 	/* Process parameters. */
2174 	for (i = 0; (params[i] != NULL); ++i) {
2175 		if (rte_kvargs_count(kvlist, params[i])) {
2176 			ret = rte_kvargs_process(kvlist, params[i],
2177 						 mlx5_args_check, config);
2178 			if (ret) {
2179 				rte_errno = EINVAL;
2180 				rte_kvargs_free(kvlist);
2181 				return -rte_errno;
2182 			}
2183 		}
2184 	}
2185 	rte_kvargs_free(kvlist);
2186 	return 0;
2187 }
2188 
2189 /**
2190  * Configures the minimal amount of data to inline into WQE
2191  * while sending packets.
2192  *
2193  * - the txq_inline_min has the maximal priority, if this
2194  *   key is specified in devargs
2195  * - if DevX is enabled the inline mode is queried from the
2196  *   device (HCA attributes and NIC vport context if needed).
2197  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
2198  *   and none (0 bytes) for other NICs
2199  *
2200  * @param spawn
2201  *   Verbs device parameters (name, port, switch_info) to spawn.
2202  * @param config
2203  *   Device configuration parameters.
2204  */
2205 void
2206 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
2207 		    struct mlx5_dev_config *config)
2208 {
2209 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
2210 		/* Application defines size of inlined data explicitly. */
2211 		if (spawn->pci_dev != NULL) {
2212 			switch (spawn->pci_dev->id.device_id) {
2213 			case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2214 			case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2215 				if (config->txq_inline_min <
2216 					       (int)MLX5_INLINE_HSIZE_L2) {
2217 					DRV_LOG(DEBUG,
2218 						"txq_inline_mix aligned to minimal ConnectX-4 required value %d",
2219 						(int)MLX5_INLINE_HSIZE_L2);
2220 					config->txq_inline_min =
2221 							MLX5_INLINE_HSIZE_L2;
2222 				}
2223 				break;
2224 			}
2225 		}
2226 		goto exit;
2227 	}
2228 	if (config->hca_attr.eth_net_offloads) {
2229 		/* We have DevX enabled, inline mode queried successfully. */
2230 		switch (config->hca_attr.wqe_inline_mode) {
2231 		case MLX5_CAP_INLINE_MODE_L2:
2232 			/* outer L2 header must be inlined. */
2233 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2234 			goto exit;
2235 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
2236 			/* No inline data are required by NIC. */
2237 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2238 			config->hw_vlan_insert =
2239 				config->hca_attr.wqe_vlan_insert;
2240 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
2241 			goto exit;
2242 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
2243 			/* inline mode is defined by NIC vport context. */
2244 			if (!config->hca_attr.eth_virt)
2245 				break;
2246 			switch (config->hca_attr.vport_inline_mode) {
2247 			case MLX5_INLINE_MODE_NONE:
2248 				config->txq_inline_min =
2249 					MLX5_INLINE_HSIZE_NONE;
2250 				goto exit;
2251 			case MLX5_INLINE_MODE_L2:
2252 				config->txq_inline_min =
2253 					MLX5_INLINE_HSIZE_L2;
2254 				goto exit;
2255 			case MLX5_INLINE_MODE_IP:
2256 				config->txq_inline_min =
2257 					MLX5_INLINE_HSIZE_L3;
2258 				goto exit;
2259 			case MLX5_INLINE_MODE_TCP_UDP:
2260 				config->txq_inline_min =
2261 					MLX5_INLINE_HSIZE_L4;
2262 				goto exit;
2263 			case MLX5_INLINE_MODE_INNER_L2:
2264 				config->txq_inline_min =
2265 					MLX5_INLINE_HSIZE_INNER_L2;
2266 				goto exit;
2267 			case MLX5_INLINE_MODE_INNER_IP:
2268 				config->txq_inline_min =
2269 					MLX5_INLINE_HSIZE_INNER_L3;
2270 				goto exit;
2271 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
2272 				config->txq_inline_min =
2273 					MLX5_INLINE_HSIZE_INNER_L4;
2274 				goto exit;
2275 			}
2276 		}
2277 	}
2278 	if (spawn->pci_dev == NULL) {
2279 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2280 		goto exit;
2281 	}
2282 	/*
2283 	 * We get here if we are unable to deduce
2284 	 * inline data size with DevX. Try PCI ID
2285 	 * to determine old NICs.
2286 	 */
2287 	switch (spawn->pci_dev->id.device_id) {
2288 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2289 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2290 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2291 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2292 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2293 		config->hw_vlan_insert = 0;
2294 		break;
2295 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2296 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2297 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2298 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2299 		/*
2300 		 * These NICs support VLAN insertion from WQE and
2301 		 * report the wqe_vlan_insert flag. But there is the bug
2302 		 * and PFC control may be broken, so disable feature.
2303 		 */
2304 		config->hw_vlan_insert = 0;
2305 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2306 		break;
2307 	default:
2308 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2309 		break;
2310 	}
2311 exit:
2312 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2313 }
2314 
2315 /**
2316  * Configures the metadata mask fields in the shared context.
2317  *
2318  * @param [in] dev
2319  *   Pointer to Ethernet device.
2320  */
2321 void
2322 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2323 {
2324 	struct mlx5_priv *priv = dev->data->dev_private;
2325 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2326 	uint32_t meta, mark, reg_c0;
2327 
2328 	reg_c0 = ~priv->vport_meta_mask;
2329 	switch (priv->config.dv_xmeta_en) {
2330 	case MLX5_XMETA_MODE_LEGACY:
2331 		meta = UINT32_MAX;
2332 		mark = MLX5_FLOW_MARK_MASK;
2333 		break;
2334 	case MLX5_XMETA_MODE_META16:
2335 		meta = reg_c0 >> rte_bsf32(reg_c0);
2336 		mark = MLX5_FLOW_MARK_MASK;
2337 		break;
2338 	case MLX5_XMETA_MODE_META32:
2339 		meta = UINT32_MAX;
2340 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2341 		break;
2342 	default:
2343 		meta = 0;
2344 		mark = 0;
2345 		MLX5_ASSERT(false);
2346 		break;
2347 	}
2348 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2349 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2350 				 sh->dv_mark_mask, mark);
2351 	else
2352 		sh->dv_mark_mask = mark;
2353 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2354 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2355 				 sh->dv_meta_mask, meta);
2356 	else
2357 		sh->dv_meta_mask = meta;
2358 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2359 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2360 				 sh->dv_meta_mask, reg_c0);
2361 	else
2362 		sh->dv_regc0_mask = reg_c0;
2363 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2364 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2365 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2366 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2367 }
2368 
2369 int
2370 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2371 {
2372 	static const char *const dynf_names[] = {
2373 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2374 		RTE_MBUF_DYNFLAG_METADATA_NAME,
2375 		RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2376 	};
2377 	unsigned int i;
2378 
2379 	if (n < RTE_DIM(dynf_names))
2380 		return -ENOMEM;
2381 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
2382 		if (names[i] == NULL)
2383 			return -EINVAL;
2384 		strcpy(names[i], dynf_names[i]);
2385 	}
2386 	return RTE_DIM(dynf_names);
2387 }
2388 
2389 /**
2390  * Comparison callback to sort device data.
2391  *
2392  * This is meant to be used with qsort().
2393  *
2394  * @param a[in]
2395  *   Pointer to pointer to first data object.
2396  * @param b[in]
2397  *   Pointer to pointer to second data object.
2398  *
2399  * @return
2400  *   0 if both objects are equal, less than 0 if the first argument is less
2401  *   than the second, greater than 0 otherwise.
2402  */
2403 int
2404 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2405 			      struct mlx5_dev_config *config,
2406 			      struct rte_device *dpdk_dev)
2407 {
2408 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2409 	struct mlx5_dev_config *sh_conf = NULL;
2410 	uint16_t port_id;
2411 
2412 	MLX5_ASSERT(sh);
2413 	/* Nothing to compare for the single/first device. */
2414 	if (sh->refcnt == 1)
2415 		return 0;
2416 	/* Find the device with shared context. */
2417 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
2418 		struct mlx5_priv *opriv =
2419 			rte_eth_devices[port_id].data->dev_private;
2420 
2421 		if (opriv && opriv != priv && opriv->sh == sh) {
2422 			sh_conf = &opriv->config;
2423 			break;
2424 		}
2425 	}
2426 	if (!sh_conf)
2427 		return 0;
2428 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2429 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2430 			     " for shared %s context", sh->ibdev_name);
2431 		rte_errno = EINVAL;
2432 		return rte_errno;
2433 	}
2434 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2435 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2436 			     " for shared %s context", sh->ibdev_name);
2437 		rte_errno = EINVAL;
2438 		return rte_errno;
2439 	}
2440 	return 0;
2441 }
2442 
2443 /**
2444  * Look for the ethernet device belonging to mlx5 driver.
2445  *
2446  * @param[in] port_id
2447  *   port_id to start looking for device.
2448  * @param[in] odev
2449  *   Pointer to the hint device. When device is being probed
2450  *   the its siblings (master and preceding representors might
2451  *   not have assigned driver yet (because the mlx5_os_pci_probe()
2452  *   is not completed yet, for this case match on hint
2453  *   device may be used to detect sibling device.
2454  *
2455  * @return
2456  *   port_id of found device, RTE_MAX_ETHPORT if not found.
2457  */
2458 uint16_t
2459 mlx5_eth_find_next(uint16_t port_id, struct rte_device *odev)
2460 {
2461 	while (port_id < RTE_MAX_ETHPORTS) {
2462 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2463 
2464 		if (dev->state != RTE_ETH_DEV_UNUSED &&
2465 		    dev->device &&
2466 		    (dev->device == odev ||
2467 		     (dev->device->driver &&
2468 		     dev->device->driver->name &&
2469 		     ((strcmp(dev->device->driver->name,
2470 			      MLX5_PCI_DRIVER_NAME) == 0) ||
2471 		      (strcmp(dev->device->driver->name,
2472 			      MLX5_AUXILIARY_DRIVER_NAME) == 0)))))
2473 			break;
2474 		port_id++;
2475 	}
2476 	if (port_id >= RTE_MAX_ETHPORTS)
2477 		return RTE_MAX_ETHPORTS;
2478 	return port_id;
2479 }
2480 
2481 /**
2482  * Callback to remove a device.
2483  *
2484  * This function removes all Ethernet devices belong to a given device.
2485  *
2486  * @param[in] cdev
2487  *   Pointer to the generic device.
2488  *
2489  * @return
2490  *   0 on success, the function cannot fail.
2491  */
2492 int
2493 mlx5_net_remove(struct mlx5_common_device *cdev)
2494 {
2495 	uint16_t port_id;
2496 	int ret = 0;
2497 
2498 	RTE_ETH_FOREACH_DEV_OF(port_id, cdev->dev) {
2499 		/*
2500 		 * mlx5_dev_close() is not registered to secondary process,
2501 		 * call the close function explicitly for secondary process.
2502 		 */
2503 		if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2504 			ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2505 		else
2506 			ret |= rte_eth_dev_close(port_id);
2507 	}
2508 	return ret == 0 ? 0 : -EIO;
2509 }
2510 
2511 static const struct rte_pci_id mlx5_pci_id_map[] = {
2512 	{
2513 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2514 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2515 	},
2516 	{
2517 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2518 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2519 	},
2520 	{
2521 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2522 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2523 	},
2524 	{
2525 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2526 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2527 	},
2528 	{
2529 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2530 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2531 	},
2532 	{
2533 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2534 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2535 	},
2536 	{
2537 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2538 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2539 	},
2540 	{
2541 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2542 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2543 	},
2544 	{
2545 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2546 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2547 	},
2548 	{
2549 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2550 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2551 	},
2552 	{
2553 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2554 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2555 	},
2556 	{
2557 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2558 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2559 	},
2560 	{
2561 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2562 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2563 	},
2564 	{
2565 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2566 				PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2567 	},
2568 	{
2569 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2570 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2571 	},
2572 	{
2573 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2574 				PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2575 	},
2576 	{
2577 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2578 				PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2579 	},
2580 	{
2581 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2582 				PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2583 	},
2584 	{
2585 		.vendor_id = 0
2586 	}
2587 };
2588 
2589 static struct mlx5_class_driver mlx5_net_driver = {
2590 	.drv_class = MLX5_CLASS_ETH,
2591 	.name = RTE_STR(MLX5_ETH_DRIVER_NAME),
2592 	.id_table = mlx5_pci_id_map,
2593 	.probe = mlx5_os_net_probe,
2594 	.remove = mlx5_net_remove,
2595 	.probe_again = 1,
2596 	.intr_lsc = 1,
2597 	.intr_rmv = 1,
2598 };
2599 
2600 /* Initialize driver log type. */
2601 RTE_LOG_REGISTER_DEFAULT(mlx5_logtype, NOTICE)
2602 
2603 /**
2604  * Driver initialization routine.
2605  */
2606 RTE_INIT(rte_mlx5_pmd_init)
2607 {
2608 	pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2609 	mlx5_common_init();
2610 	/* Build the static tables for Verbs conversion. */
2611 	mlx5_set_ptype_table();
2612 	mlx5_set_cksum_table();
2613 	mlx5_set_swp_types_table();
2614 	if (mlx5_glue)
2615 		mlx5_class_driver_register(&mlx5_net_driver);
2616 }
2617 
2618 RTE_PMD_EXPORT_NAME(MLX5_ETH_DRIVER_NAME, __COUNTER__);
2619 RTE_PMD_REGISTER_PCI_TABLE(MLX5_ETH_DRIVER_NAME, mlx5_pci_id_map);
2620 RTE_PMD_REGISTER_KMOD_DEP(MLX5_ETH_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");
2621