xref: /dpdk/drivers/net/mlx5/mlx5.c (revision daa02b5cddbb8e11b31d41e2bf7bb1ae64dcae2f)
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_spinlock_init(&sh->aso_age_mng->resize_sl);
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 			LIST_INIT(&sh->mtrmng->pools_mng.meters);
636 		}
637 		sh->mtrmng->def_policy_id = MLX5_INVALID_POLICY_ID;
638 	}
639 	return 0;
640 }
641 
642 /**
643  * Close and release all the resources of
644  * the ASO flow meter management structure.
645  *
646  * @param[in] sh
647  *   Pointer to mlx5_dev_ctx_shared object to free.
648  */
649 static void
650 mlx5_aso_flow_mtrs_mng_close(struct mlx5_dev_ctx_shared *sh)
651 {
652 	struct mlx5_aso_mtr_pool *mtr_pool;
653 	struct mlx5_flow_mtr_mng *mtrmng = sh->mtrmng;
654 	uint32_t idx;
655 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
656 	struct mlx5_aso_mtr *aso_mtr;
657 	int i;
658 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
659 
660 	if (sh->meter_aso_en) {
661 		mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_POLICER);
662 		idx = mtrmng->pools_mng.n_valid;
663 		while (idx--) {
664 			mtr_pool = mtrmng->pools_mng.pools[idx];
665 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
666 			for (i = 0; i < MLX5_ASO_MTRS_PER_POOL; i++) {
667 				aso_mtr = &mtr_pool->mtrs[i];
668 				if (aso_mtr->fm.meter_action)
669 					claim_zero
670 					(mlx5_glue->destroy_flow_action
671 					(aso_mtr->fm.meter_action));
672 			}
673 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
674 			claim_zero(mlx5_devx_cmd_destroy
675 						(mtr_pool->devx_obj));
676 			mtrmng->pools_mng.n_valid--;
677 			mlx5_free(mtr_pool);
678 		}
679 		mlx5_free(sh->mtrmng->pools_mng.pools);
680 	}
681 	mlx5_free(sh->mtrmng);
682 	sh->mtrmng = NULL;
683 }
684 
685 /* Send FLOW_AGED event if needed. */
686 void
687 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
688 {
689 	struct mlx5_age_info *age_info;
690 	uint32_t i;
691 
692 	for (i = 0; i < sh->max_port; i++) {
693 		age_info = &sh->port[i].age_info;
694 		if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
695 			continue;
696 		MLX5_AGE_UNSET(age_info, MLX5_AGE_EVENT_NEW);
697 		if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER)) {
698 			MLX5_AGE_UNSET(age_info, MLX5_AGE_TRIGGER);
699 			rte_eth_dev_callback_process
700 				(&rte_eth_devices[sh->port[i].devx_ih_port_id],
701 				RTE_ETH_EVENT_FLOW_AGED, NULL);
702 		}
703 	}
704 }
705 
706 /*
707  * Initialize the ASO connection tracking structure.
708  *
709  * @param[in] sh
710  *   Pointer to mlx5_dev_ctx_shared object.
711  *
712  * @return
713  *   0 on success, a negative errno value otherwise and rte_errno is set.
714  */
715 int
716 mlx5_flow_aso_ct_mng_init(struct mlx5_dev_ctx_shared *sh)
717 {
718 	int err;
719 
720 	if (sh->ct_mng)
721 		return 0;
722 	sh->ct_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->ct_mng),
723 				 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
724 	if (!sh->ct_mng) {
725 		DRV_LOG(ERR, "ASO CT management allocation failed.");
726 		rte_errno = ENOMEM;
727 		return -rte_errno;
728 	}
729 	err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
730 	if (err) {
731 		mlx5_free(sh->ct_mng);
732 		/* rte_errno should be extracted from the failure. */
733 		rte_errno = EINVAL;
734 		return -rte_errno;
735 	}
736 	rte_spinlock_init(&sh->ct_mng->ct_sl);
737 	rte_rwlock_init(&sh->ct_mng->resize_rwl);
738 	LIST_INIT(&sh->ct_mng->free_cts);
739 	return 0;
740 }
741 
742 /*
743  * Close and release all the resources of the
744  * ASO connection tracking management structure.
745  *
746  * @param[in] sh
747  *   Pointer to mlx5_dev_ctx_shared object to free.
748  */
749 static void
750 mlx5_flow_aso_ct_mng_close(struct mlx5_dev_ctx_shared *sh)
751 {
752 	struct mlx5_aso_ct_pools_mng *mng = sh->ct_mng;
753 	struct mlx5_aso_ct_pool *ct_pool;
754 	struct mlx5_aso_ct_action *ct;
755 	uint32_t idx;
756 	uint32_t val;
757 	uint32_t cnt;
758 	int i;
759 
760 	mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
761 	idx = mng->next;
762 	while (idx--) {
763 		cnt = 0;
764 		ct_pool = mng->pools[idx];
765 		for (i = 0; i < MLX5_ASO_CT_ACTIONS_PER_POOL; i++) {
766 			ct = &ct_pool->actions[i];
767 			val = __atomic_fetch_sub(&ct->refcnt, 1,
768 						 __ATOMIC_RELAXED);
769 			MLX5_ASSERT(val == 1);
770 			if (val > 1)
771 				cnt++;
772 #ifdef HAVE_MLX5_DR_ACTION_ASO_CT
773 			if (ct->dr_action_orig)
774 				claim_zero(mlx5_glue->destroy_flow_action
775 							(ct->dr_action_orig));
776 			if (ct->dr_action_rply)
777 				claim_zero(mlx5_glue->destroy_flow_action
778 							(ct->dr_action_rply));
779 #endif
780 		}
781 		claim_zero(mlx5_devx_cmd_destroy(ct_pool->devx_obj));
782 		if (cnt) {
783 			DRV_LOG(DEBUG, "%u ASO CT objects are being used in the pool %u",
784 				cnt, i);
785 		}
786 		mlx5_free(ct_pool);
787 		/* in case of failure. */
788 		mng->next--;
789 	}
790 	mlx5_free(mng->pools);
791 	mlx5_free(mng);
792 	/* Management structure must be cleared to 0s during allocation. */
793 	sh->ct_mng = NULL;
794 }
795 
796 /**
797  * Initialize the flow resources' indexed mempool.
798  *
799  * @param[in] sh
800  *   Pointer to mlx5_dev_ctx_shared object.
801  * @param[in] config
802  *   Pointer to user dev config.
803  */
804 static void
805 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
806 		       const struct mlx5_dev_config *config)
807 {
808 	uint8_t i;
809 	struct mlx5_indexed_pool_config cfg;
810 
811 	for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
812 		cfg = mlx5_ipool_cfg[i];
813 		switch (i) {
814 		default:
815 			break;
816 		/*
817 		 * Set MLX5_IPOOL_MLX5_FLOW ipool size
818 		 * according to PCI function flow configuration.
819 		 */
820 		case MLX5_IPOOL_MLX5_FLOW:
821 			cfg.size = config->dv_flow_en ?
822 				sizeof(struct mlx5_flow_handle) :
823 				MLX5_FLOW_HANDLE_VERBS_SIZE;
824 			break;
825 		}
826 		if (config->reclaim_mode) {
827 			cfg.release_mem_en = 1;
828 			cfg.per_core_cache = 0;
829 		} else {
830 			cfg.release_mem_en = 0;
831 		}
832 		sh->ipool[i] = mlx5_ipool_create(&cfg);
833 	}
834 }
835 
836 
837 /**
838  * Release the flow resources' indexed mempool.
839  *
840  * @param[in] sh
841  *   Pointer to mlx5_dev_ctx_shared object.
842  */
843 static void
844 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
845 {
846 	uint8_t i;
847 
848 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
849 		mlx5_ipool_destroy(sh->ipool[i]);
850 	for (i = 0; i < MLX5_MAX_MODIFY_NUM; ++i)
851 		if (sh->mdh_ipools[i])
852 			mlx5_ipool_destroy(sh->mdh_ipools[i]);
853 }
854 
855 /*
856  * Check if dynamic flex parser for eCPRI already exists.
857  *
858  * @param dev
859  *   Pointer to Ethernet device structure.
860  *
861  * @return
862  *   true on exists, false on not.
863  */
864 bool
865 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
866 {
867 	struct mlx5_priv *priv = dev->data->dev_private;
868 	struct mlx5_flex_parser_profiles *prf =
869 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
870 
871 	return !!prf->obj;
872 }
873 
874 /*
875  * Allocation of a flex parser for eCPRI. Once created, this parser related
876  * resources will be held until the device is closed.
877  *
878  * @param dev
879  *   Pointer to Ethernet device structure.
880  *
881  * @return
882  *   0 on success, a negative errno value otherwise and rte_errno is set.
883  */
884 int
885 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
886 {
887 	struct mlx5_priv *priv = dev->data->dev_private;
888 	struct mlx5_flex_parser_profiles *prf =
889 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
890 	struct mlx5_devx_graph_node_attr node = {
891 		.modify_field_select = 0,
892 	};
893 	uint32_t ids[8];
894 	int ret;
895 
896 	if (!priv->config.hca_attr.parse_graph_flex_node) {
897 		DRV_LOG(ERR, "Dynamic flex parser is not supported "
898 			"for device %s.", priv->dev_data->name);
899 		return -ENOTSUP;
900 	}
901 	node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
902 	/* 8 bytes now: 4B common header + 4B message body header. */
903 	node.header_length_base_value = 0x8;
904 	/* After MAC layer: Ether / VLAN. */
905 	node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
906 	/* Type of compared condition should be 0xAEFE in the L2 layer. */
907 	node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
908 	/* Sample #0: type in common header. */
909 	node.sample[0].flow_match_sample_en = 1;
910 	/* Fixed offset. */
911 	node.sample[0].flow_match_sample_offset_mode = 0x0;
912 	/* Only the 2nd byte will be used. */
913 	node.sample[0].flow_match_sample_field_base_offset = 0x0;
914 	/* Sample #1: message payload. */
915 	node.sample[1].flow_match_sample_en = 1;
916 	/* Fixed offset. */
917 	node.sample[1].flow_match_sample_offset_mode = 0x0;
918 	/*
919 	 * Only the first two bytes will be used right now, and its offset will
920 	 * start after the common header that with the length of a DW(u32).
921 	 */
922 	node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
923 	prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->cdev->ctx, &node);
924 	if (!prf->obj) {
925 		DRV_LOG(ERR, "Failed to create flex parser node object.");
926 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
927 	}
928 	prf->num = 2;
929 	ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
930 	if (ret) {
931 		DRV_LOG(ERR, "Failed to query sample IDs.");
932 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
933 	}
934 	prf->offset[0] = 0x0;
935 	prf->offset[1] = sizeof(uint32_t);
936 	prf->ids[0] = ids[0];
937 	prf->ids[1] = ids[1];
938 	return 0;
939 }
940 
941 /*
942  * Destroy the flex parser node, including the parser itself, input / output
943  * arcs and DW samples. Resources could be reused then.
944  *
945  * @param dev
946  *   Pointer to Ethernet device structure.
947  */
948 static void
949 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
950 {
951 	struct mlx5_priv *priv = dev->data->dev_private;
952 	struct mlx5_flex_parser_profiles *prf =
953 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
954 
955 	if (prf->obj)
956 		mlx5_devx_cmd_destroy(prf->obj);
957 	prf->obj = NULL;
958 }
959 
960 uint32_t
961 mlx5_get_supported_sw_parsing_offloads(const struct mlx5_hca_attr *attr)
962 {
963 	uint32_t sw_parsing_offloads = 0;
964 
965 	if (attr->swp) {
966 		sw_parsing_offloads |= MLX5_SW_PARSING_CAP;
967 		if (attr->swp_csum)
968 			sw_parsing_offloads |= MLX5_SW_PARSING_CSUM_CAP;
969 
970 		if (attr->swp_lso)
971 			sw_parsing_offloads |= MLX5_SW_PARSING_TSO_CAP;
972 	}
973 	return sw_parsing_offloads;
974 }
975 
976 uint32_t
977 mlx5_get_supported_tunneling_offloads(const struct mlx5_hca_attr *attr)
978 {
979 	uint32_t tn_offloads = 0;
980 
981 	if (attr->tunnel_stateless_vxlan)
982 		tn_offloads |= MLX5_TUNNELED_OFFLOADS_VXLAN_CAP;
983 	if (attr->tunnel_stateless_gre)
984 		tn_offloads |= MLX5_TUNNELED_OFFLOADS_GRE_CAP;
985 	if (attr->tunnel_stateless_geneve_rx)
986 		tn_offloads |= MLX5_TUNNELED_OFFLOADS_GENEVE_CAP;
987 	return tn_offloads;
988 }
989 
990 /*
991  * Allocate Rx and Tx UARs in robust fashion.
992  * This routine handles the following UAR allocation issues:
993  *
994  *  - tries to allocate the UAR with the most appropriate memory
995  *    mapping type from the ones supported by the host
996  *
997  *  - tries to allocate the UAR with non-NULL base address
998  *    OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
999  *    UAR base address if UAR was not the first object in the UAR page.
1000  *    It caused the PMD failure and we should try to get another UAR
1001  *    till we get the first one with non-NULL base address returned.
1002  */
1003 static int
1004 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
1005 		     const struct mlx5_common_dev_config *config)
1006 {
1007 	uint32_t uar_mapping, retry;
1008 	int err = 0;
1009 	void *base_addr;
1010 
1011 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
1012 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1013 		/* Control the mapping type according to the settings. */
1014 		uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
1015 			      MLX5DV_UAR_ALLOC_TYPE_NC :
1016 			      MLX5DV_UAR_ALLOC_TYPE_BF;
1017 #else
1018 		RTE_SET_USED(config);
1019 		/*
1020 		 * It seems we have no way to control the memory mapping type
1021 		 * for the UAR, the default "Write-Combining" type is supposed.
1022 		 * The UAR initialization on queue creation queries the
1023 		 * actual mapping type done by Verbs/kernel and setups the
1024 		 * PMD datapath accordingly.
1025 		 */
1026 		uar_mapping = 0;
1027 #endif
1028 		sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1029 						       uar_mapping);
1030 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1031 		if (!sh->tx_uar &&
1032 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
1033 			if (config->dbnc == MLX5_TXDB_CACHED ||
1034 			    config->dbnc == MLX5_TXDB_HEURISTIC)
1035 				DRV_LOG(WARNING, "Devarg tx_db_nc setting "
1036 						 "is not supported by DevX");
1037 			/*
1038 			 * In some environments like virtual machine
1039 			 * the Write Combining mapped might be not supported
1040 			 * and UAR allocation fails. We try "Non-Cached"
1041 			 * mapping for the case. The tx_burst routines take
1042 			 * the UAR mapping type into account on UAR setup
1043 			 * on queue creation.
1044 			 */
1045 			DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (BF)");
1046 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1047 			sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1048 							       uar_mapping);
1049 		} else if (!sh->tx_uar &&
1050 			   uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
1051 			if (config->dbnc == MLX5_TXDB_NCACHED)
1052 				DRV_LOG(WARNING, "Devarg tx_db_nc settings "
1053 						 "is not supported by DevX");
1054 			/*
1055 			 * If Verbs/kernel does not support "Non-Cached"
1056 			 * try the "Write-Combining".
1057 			 */
1058 			DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (NC)");
1059 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
1060 			sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1061 							       uar_mapping);
1062 		}
1063 #endif
1064 		if (!sh->tx_uar) {
1065 			DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
1066 			err = ENOMEM;
1067 			goto exit;
1068 		}
1069 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
1070 		if (base_addr)
1071 			break;
1072 		/*
1073 		 * The UARs are allocated by rdma_core within the
1074 		 * IB device context, on context closure all UARs
1075 		 * will be freed, should be no memory/object leakage.
1076 		 */
1077 		DRV_LOG(DEBUG, "Retrying to allocate Tx DevX UAR");
1078 		sh->tx_uar = NULL;
1079 	}
1080 	/* Check whether we finally succeeded with valid UAR allocation. */
1081 	if (!sh->tx_uar) {
1082 		DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
1083 		err = ENOMEM;
1084 		goto exit;
1085 	}
1086 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
1087 		uar_mapping = 0;
1088 		sh->devx_rx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1089 							    uar_mapping);
1090 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1091 		if (!sh->devx_rx_uar &&
1092 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
1093 			/*
1094 			 * Rx UAR is used to control interrupts only,
1095 			 * should be no datapath noticeable impact,
1096 			 * can try "Non-Cached" mapping safely.
1097 			 */
1098 			DRV_LOG(DEBUG, "Failed to allocate Rx DevX UAR (BF)");
1099 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1100 			sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
1101 						   (sh->cdev->ctx, uar_mapping);
1102 		}
1103 #endif
1104 		if (!sh->devx_rx_uar) {
1105 			DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
1106 			err = ENOMEM;
1107 			goto exit;
1108 		}
1109 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
1110 		if (base_addr)
1111 			break;
1112 		/*
1113 		 * The UARs are allocated by rdma_core within the
1114 		 * IB device context, on context closure all UARs
1115 		 * will be freed, should be no memory/object leakage.
1116 		 */
1117 		DRV_LOG(DEBUG, "Retrying to allocate Rx DevX UAR");
1118 		sh->devx_rx_uar = NULL;
1119 	}
1120 	/* Check whether we finally succeeded with valid UAR allocation. */
1121 	if (!sh->devx_rx_uar) {
1122 		DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
1123 		err = ENOMEM;
1124 	}
1125 exit:
1126 	return err;
1127 }
1128 
1129 /**
1130  * rte_mempool_walk() callback to unregister Rx mempools.
1131  * It used when implicit mempool registration is disabled.
1132  *
1133  * @param mp
1134  *   The mempool being walked.
1135  * @param arg
1136  *   Pointer to the device shared context.
1137  */
1138 static void
1139 mlx5_dev_ctx_shared_rx_mempool_unregister_cb(struct rte_mempool *mp, void *arg)
1140 {
1141 	struct mlx5_dev_ctx_shared *sh = arg;
1142 
1143 	mlx5_dev_mempool_unregister(sh->cdev, mp);
1144 }
1145 
1146 /**
1147  * Callback used when implicit mempool registration is disabled
1148  * in order to track Rx mempool destruction.
1149  *
1150  * @param event
1151  *   Mempool life cycle event.
1152  * @param mp
1153  *   An Rx mempool registered explicitly when the port is started.
1154  * @param arg
1155  *   Pointer to a device shared context.
1156  */
1157 static void
1158 mlx5_dev_ctx_shared_rx_mempool_event_cb(enum rte_mempool_event event,
1159 					struct rte_mempool *mp, void *arg)
1160 {
1161 	struct mlx5_dev_ctx_shared *sh = arg;
1162 
1163 	if (event == RTE_MEMPOOL_EVENT_DESTROY)
1164 		mlx5_dev_mempool_unregister(sh->cdev, mp);
1165 }
1166 
1167 int
1168 mlx5_dev_ctx_shared_mempool_subscribe(struct rte_eth_dev *dev)
1169 {
1170 	struct mlx5_priv *priv = dev->data->dev_private;
1171 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1172 	int ret;
1173 
1174 	/* Check if we only need to track Rx mempool destruction. */
1175 	if (!sh->cdev->config.mr_mempool_reg_en) {
1176 		ret = rte_mempool_event_callback_register
1177 				(mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1178 		return ret == 0 || rte_errno == EEXIST ? 0 : ret;
1179 	}
1180 	return mlx5_dev_mempool_subscribe(sh->cdev);
1181 }
1182 
1183 /**
1184  * Set up multiple TISs with different affinities according to
1185  * number of bonding ports
1186  *
1187  * @param priv
1188  * Pointer of shared context.
1189  *
1190  * @return
1191  * Zero on success, -1 otherwise.
1192  */
1193 static int
1194 mlx5_setup_tis(struct mlx5_dev_ctx_shared *sh)
1195 {
1196 	int i;
1197 	struct mlx5_devx_lag_context lag_ctx = { 0 };
1198 	struct mlx5_devx_tis_attr tis_attr = { 0 };
1199 
1200 	tis_attr.transport_domain = sh->td->id;
1201 	if (sh->bond.n_port) {
1202 		if (!mlx5_devx_cmd_query_lag(sh->cdev->ctx, &lag_ctx)) {
1203 			sh->lag.tx_remap_affinity[0] =
1204 				lag_ctx.tx_remap_affinity_1;
1205 			sh->lag.tx_remap_affinity[1] =
1206 				lag_ctx.tx_remap_affinity_2;
1207 			sh->lag.affinity_mode = lag_ctx.port_select_mode;
1208 		} else {
1209 			DRV_LOG(ERR, "Failed to query lag affinity.");
1210 			return -1;
1211 		}
1212 		if (sh->lag.affinity_mode == MLX5_LAG_MODE_TIS) {
1213 			for (i = 0; i < sh->bond.n_port; i++) {
1214 				tis_attr.lag_tx_port_affinity =
1215 					MLX5_IFC_LAG_MAP_TIS_AFFINITY(i,
1216 							sh->bond.n_port);
1217 				sh->tis[i] = mlx5_devx_cmd_create_tis(sh->cdev->ctx,
1218 						&tis_attr);
1219 				if (!sh->tis[i]) {
1220 					DRV_LOG(ERR, "Failed to TIS %d/%d for bonding device"
1221 						" %s.", i, sh->bond.n_port,
1222 						sh->ibdev_name);
1223 					return -1;
1224 				}
1225 			}
1226 			DRV_LOG(DEBUG, "LAG number of ports : %d, affinity_1 & 2 : pf%d & %d.\n",
1227 				sh->bond.n_port, lag_ctx.tx_remap_affinity_1,
1228 				lag_ctx.tx_remap_affinity_2);
1229 			return 0;
1230 		}
1231 		if (sh->lag.affinity_mode == MLX5_LAG_MODE_HASH)
1232 			DRV_LOG(INFO, "Device %s enabled HW hash based LAG.",
1233 					sh->ibdev_name);
1234 	}
1235 	tis_attr.lag_tx_port_affinity = 0;
1236 	sh->tis[0] = mlx5_devx_cmd_create_tis(sh->cdev->ctx, &tis_attr);
1237 	if (!sh->tis[0]) {
1238 		DRV_LOG(ERR, "Failed to TIS 0 for bonding device"
1239 			" %s.", sh->ibdev_name);
1240 		return -1;
1241 	}
1242 	return 0;
1243 }
1244 
1245 /**
1246  * Allocate shared device context. If there is multiport device the
1247  * master and representors will share this context, if there is single
1248  * port dedicated device, the context will be used by only given
1249  * port due to unification.
1250  *
1251  * Routine first searches the context for the specified device name,
1252  * if found the shared context assumed and reference counter is incremented.
1253  * If no context found the new one is created and initialized with specified
1254  * device context and parameters.
1255  *
1256  * @param[in] spawn
1257  *   Pointer to the device attributes (name, port, etc).
1258  * @param[in] config
1259  *   Pointer to device configuration structure.
1260  *
1261  * @return
1262  *   Pointer to mlx5_dev_ctx_shared object on success,
1263  *   otherwise NULL and rte_errno is set.
1264  */
1265 struct mlx5_dev_ctx_shared *
1266 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
1267 			  const struct mlx5_dev_config *config)
1268 {
1269 	struct mlx5_dev_ctx_shared *sh;
1270 	int err = 0;
1271 	uint32_t i;
1272 
1273 	MLX5_ASSERT(spawn);
1274 	/* Secondary process should not create the shared context. */
1275 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1276 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1277 	/* Search for IB context by device name. */
1278 	LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
1279 		if (!strcmp(sh->ibdev_name, spawn->phys_dev_name)) {
1280 			sh->refcnt++;
1281 			goto exit;
1282 		}
1283 	}
1284 	/* No device found, we have to create new shared context. */
1285 	MLX5_ASSERT(spawn->max_port);
1286 	sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1287 			 sizeof(struct mlx5_dev_ctx_shared) +
1288 			 spawn->max_port *
1289 			 sizeof(struct mlx5_dev_shared_port),
1290 			 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1291 	if (!sh) {
1292 		DRV_LOG(ERR, "shared context allocation failure");
1293 		rte_errno  = ENOMEM;
1294 		goto exit;
1295 	}
1296 	pthread_mutex_init(&sh->txpp.mutex, NULL);
1297 	sh->numa_node = spawn->cdev->dev->numa_node;
1298 	sh->cdev = spawn->cdev;
1299 	sh->devx = sh->cdev->config.devx;
1300 	if (spawn->bond_info)
1301 		sh->bond = *spawn->bond_info;
1302 	err = mlx5_os_get_dev_attr(sh->cdev, &sh->device_attr);
1303 	if (err) {
1304 		DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
1305 		goto error;
1306 	}
1307 	sh->refcnt = 1;
1308 	sh->max_port = spawn->max_port;
1309 	sh->reclaim_mode = config->reclaim_mode;
1310 	strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->cdev->ctx),
1311 		sizeof(sh->ibdev_name) - 1);
1312 	strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->cdev->ctx),
1313 		sizeof(sh->ibdev_path) - 1);
1314 	/*
1315 	 * Setting port_id to max unallowed value means
1316 	 * there is no interrupt subhandler installed for
1317 	 * the given port index i.
1318 	 */
1319 	for (i = 0; i < sh->max_port; i++) {
1320 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
1321 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
1322 	}
1323 	if (sh->devx) {
1324 		sh->td = mlx5_devx_cmd_create_td(sh->cdev->ctx);
1325 		if (!sh->td) {
1326 			DRV_LOG(ERR, "TD allocation failure");
1327 			err = ENOMEM;
1328 			goto error;
1329 		}
1330 		if (mlx5_setup_tis(sh)) {
1331 			DRV_LOG(ERR, "TIS allocation failure");
1332 			err = ENOMEM;
1333 			goto error;
1334 		}
1335 		err = mlx5_alloc_rxtx_uars(sh, &sh->cdev->config);
1336 		if (err)
1337 			goto error;
1338 		MLX5_ASSERT(sh->tx_uar);
1339 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
1340 
1341 		MLX5_ASSERT(sh->devx_rx_uar);
1342 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
1343 	}
1344 #ifndef RTE_ARCH_64
1345 	/* Initialize UAR access locks for 32bit implementations. */
1346 	rte_spinlock_init(&sh->uar_lock_cq);
1347 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1348 		rte_spinlock_init(&sh->uar_lock[i]);
1349 #endif
1350 	mlx5_os_dev_shared_handler_install(sh);
1351 	if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1352 		err = mlx5_flow_os_init_workspace_once();
1353 		if (err)
1354 			goto error;
1355 	}
1356 	mlx5_flow_aging_init(sh);
1357 	mlx5_flow_counters_mng_init(sh);
1358 	mlx5_flow_ipool_create(sh, config);
1359 	/* Add context to the global device list. */
1360 	LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1361 	rte_spinlock_init(&sh->geneve_tlv_opt_sl);
1362 exit:
1363 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1364 	return sh;
1365 error:
1366 	pthread_mutex_destroy(&sh->txpp.mutex);
1367 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1368 	MLX5_ASSERT(sh);
1369 	if (sh->td)
1370 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1371 	i = 0;
1372 	do {
1373 		if (sh->tis[i])
1374 			claim_zero(mlx5_devx_cmd_destroy(sh->tis[i]));
1375 	} while (++i < (uint32_t)sh->bond.n_port);
1376 	if (sh->devx_rx_uar)
1377 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1378 	if (sh->tx_uar)
1379 		mlx5_glue->devx_free_uar(sh->tx_uar);
1380 	mlx5_free(sh);
1381 	MLX5_ASSERT(err > 0);
1382 	rte_errno = err;
1383 	return NULL;
1384 }
1385 
1386 /**
1387  * Free shared IB device context. Decrement counter and if zero free
1388  * all allocated resources and close handles.
1389  *
1390  * @param[in] sh
1391  *   Pointer to mlx5_dev_ctx_shared object to free
1392  */
1393 void
1394 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1395 {
1396 	int ret;
1397 	int i = 0;
1398 
1399 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1400 #ifdef RTE_LIBRTE_MLX5_DEBUG
1401 	/* Check the object presence in the list. */
1402 	struct mlx5_dev_ctx_shared *lctx;
1403 
1404 	LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1405 		if (lctx == sh)
1406 			break;
1407 	MLX5_ASSERT(lctx);
1408 	if (lctx != sh) {
1409 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
1410 		goto exit;
1411 	}
1412 #endif
1413 	MLX5_ASSERT(sh);
1414 	MLX5_ASSERT(sh->refcnt);
1415 	/* Secondary process should not free the shared context. */
1416 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1417 	if (--sh->refcnt)
1418 		goto exit;
1419 	/* Stop watching for mempool events and unregister all mempools. */
1420 	if (!sh->cdev->config.mr_mempool_reg_en) {
1421 		ret = rte_mempool_event_callback_unregister
1422 				(mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1423 		if (ret == 0)
1424 			rte_mempool_walk
1425 			     (mlx5_dev_ctx_shared_rx_mempool_unregister_cb, sh);
1426 	}
1427 	/* Remove context from the global device list. */
1428 	LIST_REMOVE(sh, next);
1429 	/* Release resources on the last device removal. */
1430 	if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1431 		mlx5_os_net_cleanup();
1432 		mlx5_flow_os_release_workspace();
1433 	}
1434 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1435 	/*
1436 	 *  Ensure there is no async event handler installed.
1437 	 *  Only primary process handles async device events.
1438 	 **/
1439 	mlx5_flow_counters_mng_close(sh);
1440 	if (sh->aso_age_mng) {
1441 		mlx5_flow_aso_age_mng_close(sh);
1442 		sh->aso_age_mng = NULL;
1443 	}
1444 	if (sh->mtrmng)
1445 		mlx5_aso_flow_mtrs_mng_close(sh);
1446 	mlx5_flow_ipool_destroy(sh);
1447 	mlx5_os_dev_shared_handler_uninstall(sh);
1448 	if (sh->tx_uar) {
1449 		mlx5_glue->devx_free_uar(sh->tx_uar);
1450 		sh->tx_uar = NULL;
1451 	}
1452 	do {
1453 		if (sh->tis[i])
1454 			claim_zero(mlx5_devx_cmd_destroy(sh->tis[i]));
1455 	} while (++i < sh->bond.n_port);
1456 	if (sh->td)
1457 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1458 	if (sh->devx_rx_uar)
1459 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1460 	MLX5_ASSERT(sh->geneve_tlv_option_resource == NULL);
1461 	pthread_mutex_destroy(&sh->txpp.mutex);
1462 	mlx5_free(sh);
1463 	return;
1464 exit:
1465 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1466 }
1467 
1468 /**
1469  * Destroy table hash list.
1470  *
1471  * @param[in] priv
1472  *   Pointer to the private device data structure.
1473  */
1474 void
1475 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1476 {
1477 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1478 
1479 	if (!sh->flow_tbls)
1480 		return;
1481 	mlx5_hlist_destroy(sh->flow_tbls);
1482 	sh->flow_tbls = NULL;
1483 }
1484 
1485 /**
1486  * Initialize flow table hash list and create the root tables entry
1487  * for each domain.
1488  *
1489  * @param[in] priv
1490  *   Pointer to the private device data structure.
1491  *
1492  * @return
1493  *   Zero on success, positive error code otherwise.
1494  */
1495 int
1496 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1497 {
1498 	int err = 0;
1499 	/* Tables are only used in DV and DR modes. */
1500 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1501 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1502 	char s[MLX5_NAME_SIZE];
1503 
1504 	MLX5_ASSERT(sh);
1505 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1506 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1507 					  false, true, sh,
1508 					  flow_dv_tbl_create_cb,
1509 					  flow_dv_tbl_match_cb,
1510 					  flow_dv_tbl_remove_cb,
1511 					  flow_dv_tbl_clone_cb,
1512 					  flow_dv_tbl_clone_free_cb);
1513 	if (!sh->flow_tbls) {
1514 		DRV_LOG(ERR, "flow tables with hash creation failed.");
1515 		err = ENOMEM;
1516 		return err;
1517 	}
1518 #ifndef HAVE_MLX5DV_DR
1519 	struct rte_flow_error error;
1520 	struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1521 
1522 	/*
1523 	 * In case we have not DR support, the zero tables should be created
1524 	 * because DV expect to see them even if they cannot be created by
1525 	 * RDMA-CORE.
1526 	 */
1527 	if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0,
1528 		NULL, 0, 1, 0, &error) ||
1529 	    !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0,
1530 		NULL, 0, 1, 0, &error) ||
1531 	    !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0,
1532 		NULL, 0, 1, 0, &error)) {
1533 		err = ENOMEM;
1534 		goto error;
1535 	}
1536 	return err;
1537 error:
1538 	mlx5_free_table_hash_list(priv);
1539 #endif /* HAVE_MLX5DV_DR */
1540 #endif
1541 	return err;
1542 }
1543 
1544 /**
1545  * Retrieve integer value from environment variable.
1546  *
1547  * @param[in] name
1548  *   Environment variable name.
1549  *
1550  * @return
1551  *   Integer value, 0 if the variable is not set.
1552  */
1553 int
1554 mlx5_getenv_int(const char *name)
1555 {
1556 	const char *val = getenv(name);
1557 
1558 	if (val == NULL)
1559 		return 0;
1560 	return atoi(val);
1561 }
1562 
1563 /**
1564  * DPDK callback to add udp tunnel port
1565  *
1566  * @param[in] dev
1567  *   A pointer to eth_dev
1568  * @param[in] udp_tunnel
1569  *   A pointer to udp tunnel
1570  *
1571  * @return
1572  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1573  */
1574 int
1575 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1576 			 struct rte_eth_udp_tunnel *udp_tunnel)
1577 {
1578 	MLX5_ASSERT(udp_tunnel != NULL);
1579 	if (udp_tunnel->prot_type == RTE_ETH_TUNNEL_TYPE_VXLAN &&
1580 	    udp_tunnel->udp_port == 4789)
1581 		return 0;
1582 	if (udp_tunnel->prot_type == RTE_ETH_TUNNEL_TYPE_VXLAN_GPE &&
1583 	    udp_tunnel->udp_port == 4790)
1584 		return 0;
1585 	return -ENOTSUP;
1586 }
1587 
1588 /**
1589  * Initialize process private data structure.
1590  *
1591  * @param dev
1592  *   Pointer to Ethernet device structure.
1593  *
1594  * @return
1595  *   0 on success, a negative errno value otherwise and rte_errno is set.
1596  */
1597 int
1598 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1599 {
1600 	struct mlx5_priv *priv = dev->data->dev_private;
1601 	struct mlx5_proc_priv *ppriv;
1602 	size_t ppriv_size;
1603 
1604 	mlx5_proc_priv_uninit(dev);
1605 	/*
1606 	 * UAR register table follows the process private structure. BlueFlame
1607 	 * registers for Tx queues are stored in the table.
1608 	 */
1609 	ppriv_size =
1610 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1611 	ppriv = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, ppriv_size,
1612 			    RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1613 	if (!ppriv) {
1614 		rte_errno = ENOMEM;
1615 		return -rte_errno;
1616 	}
1617 	ppriv->uar_table_sz = priv->txqs_n;
1618 	dev->process_private = ppriv;
1619 	return 0;
1620 }
1621 
1622 /**
1623  * Un-initialize process private data structure.
1624  *
1625  * @param dev
1626  *   Pointer to Ethernet device structure.
1627  */
1628 void
1629 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1630 {
1631 	if (!dev->process_private)
1632 		return;
1633 	mlx5_free(dev->process_private);
1634 	dev->process_private = NULL;
1635 }
1636 
1637 /**
1638  * DPDK callback to close the device.
1639  *
1640  * Destroy all queues and objects, free memory.
1641  *
1642  * @param dev
1643  *   Pointer to Ethernet device structure.
1644  */
1645 int
1646 mlx5_dev_close(struct rte_eth_dev *dev)
1647 {
1648 	struct mlx5_priv *priv = dev->data->dev_private;
1649 	unsigned int i;
1650 	int ret;
1651 
1652 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1653 		/* Check if process_private released. */
1654 		if (!dev->process_private)
1655 			return 0;
1656 		mlx5_tx_uar_uninit_secondary(dev);
1657 		mlx5_proc_priv_uninit(dev);
1658 		rte_eth_dev_release_port(dev);
1659 		return 0;
1660 	}
1661 	if (!priv->sh)
1662 		return 0;
1663 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1664 		dev->data->port_id,
1665 		((priv->sh->cdev->ctx != NULL) ?
1666 		mlx5_os_get_ctx_device_name(priv->sh->cdev->ctx) : ""));
1667 	/*
1668 	 * If default mreg copy action is removed at the stop stage,
1669 	 * the search will return none and nothing will be done anymore.
1670 	 */
1671 	mlx5_flow_stop_default(dev);
1672 	mlx5_traffic_disable(dev);
1673 	/*
1674 	 * If all the flows are already flushed in the device stop stage,
1675 	 * then this will return directly without any action.
1676 	 */
1677 	mlx5_flow_list_flush(dev, MLX5_FLOW_TYPE_GEN, true);
1678 	mlx5_action_handle_flush(dev);
1679 	mlx5_flow_meter_flush(dev, NULL);
1680 	/* Prevent crashes when queues are still in use. */
1681 	dev->rx_pkt_burst = removed_rx_burst;
1682 	dev->tx_pkt_burst = removed_tx_burst;
1683 	rte_wmb();
1684 	/* Disable datapath on secondary process. */
1685 	mlx5_mp_os_req_stop_rxtx(dev);
1686 	/* Free the eCPRI flex parser resource. */
1687 	mlx5_flex_parser_ecpri_release(dev);
1688 	if (priv->rxqs != NULL) {
1689 		/* XXX race condition if mlx5_rx_burst() is still running. */
1690 		rte_delay_us_sleep(1000);
1691 		for (i = 0; (i != priv->rxqs_n); ++i)
1692 			mlx5_rxq_release(dev, i);
1693 		priv->rxqs_n = 0;
1694 		priv->rxqs = NULL;
1695 	}
1696 	if (priv->representor) {
1697 		/* Each representor has a dedicated interrupts handler */
1698 		mlx5_free(dev->intr_handle);
1699 		dev->intr_handle = NULL;
1700 	}
1701 	if (priv->txqs != NULL) {
1702 		/* XXX race condition if mlx5_tx_burst() is still running. */
1703 		rte_delay_us_sleep(1000);
1704 		for (i = 0; (i != priv->txqs_n); ++i)
1705 			mlx5_txq_release(dev, i);
1706 		priv->txqs_n = 0;
1707 		priv->txqs = NULL;
1708 	}
1709 	mlx5_proc_priv_uninit(dev);
1710 	if (priv->q_counters) {
1711 		mlx5_devx_cmd_destroy(priv->q_counters);
1712 		priv->q_counters = NULL;
1713 	}
1714 	if (priv->drop_queue.hrxq)
1715 		mlx5_drop_action_destroy(dev);
1716 	if (priv->mreg_cp_tbl)
1717 		mlx5_hlist_destroy(priv->mreg_cp_tbl);
1718 	mlx5_mprq_free_mp(dev);
1719 	if (priv->sh->ct_mng)
1720 		mlx5_flow_aso_ct_mng_close(priv->sh);
1721 	mlx5_os_free_shared_dr(priv);
1722 	if (priv->rss_conf.rss_key != NULL)
1723 		mlx5_free(priv->rss_conf.rss_key);
1724 	if (priv->reta_idx != NULL)
1725 		mlx5_free(priv->reta_idx);
1726 	if (priv->config.vf)
1727 		mlx5_os_mac_addr_flush(dev);
1728 	if (priv->nl_socket_route >= 0)
1729 		close(priv->nl_socket_route);
1730 	if (priv->nl_socket_rdma >= 0)
1731 		close(priv->nl_socket_rdma);
1732 	if (priv->vmwa_context)
1733 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1734 	ret = mlx5_hrxq_verify(dev);
1735 	if (ret)
1736 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1737 			dev->data->port_id);
1738 	ret = mlx5_ind_table_obj_verify(dev);
1739 	if (ret)
1740 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1741 			dev->data->port_id);
1742 	ret = mlx5_rxq_obj_verify(dev);
1743 	if (ret)
1744 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1745 			dev->data->port_id);
1746 	ret = mlx5_rxq_verify(dev);
1747 	if (ret)
1748 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1749 			dev->data->port_id);
1750 	ret = mlx5_txq_obj_verify(dev);
1751 	if (ret)
1752 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1753 			dev->data->port_id);
1754 	ret = mlx5_txq_verify(dev);
1755 	if (ret)
1756 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1757 			dev->data->port_id);
1758 	ret = mlx5_flow_verify(dev);
1759 	if (ret)
1760 		DRV_LOG(WARNING, "port %u some flows still remain",
1761 			dev->data->port_id);
1762 	if (priv->hrxqs)
1763 		mlx5_list_destroy(priv->hrxqs);
1764 	/*
1765 	 * Free the shared context in last turn, because the cleanup
1766 	 * routines above may use some shared fields, like
1767 	 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1768 	 * ifindex if Netlink fails.
1769 	 */
1770 	mlx5_free_shared_dev_ctx(priv->sh);
1771 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1772 		unsigned int c = 0;
1773 		uint16_t port_id;
1774 
1775 		MLX5_ETH_FOREACH_DEV(port_id, dev->device) {
1776 			struct mlx5_priv *opriv =
1777 				rte_eth_devices[port_id].data->dev_private;
1778 
1779 			if (!opriv ||
1780 			    opriv->domain_id != priv->domain_id ||
1781 			    &rte_eth_devices[port_id] == dev)
1782 				continue;
1783 			++c;
1784 			break;
1785 		}
1786 		if (!c)
1787 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1788 	}
1789 	memset(priv, 0, sizeof(*priv));
1790 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1791 	/*
1792 	 * Reset mac_addrs to NULL such that it is not freed as part of
1793 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1794 	 * it is freed when dev_private is freed.
1795 	 */
1796 	dev->data->mac_addrs = NULL;
1797 	return 0;
1798 }
1799 
1800 const struct eth_dev_ops mlx5_dev_ops = {
1801 	.dev_configure = mlx5_dev_configure,
1802 	.dev_start = mlx5_dev_start,
1803 	.dev_stop = mlx5_dev_stop,
1804 	.dev_set_link_down = mlx5_set_link_down,
1805 	.dev_set_link_up = mlx5_set_link_up,
1806 	.dev_close = mlx5_dev_close,
1807 	.promiscuous_enable = mlx5_promiscuous_enable,
1808 	.promiscuous_disable = mlx5_promiscuous_disable,
1809 	.allmulticast_enable = mlx5_allmulticast_enable,
1810 	.allmulticast_disable = mlx5_allmulticast_disable,
1811 	.link_update = mlx5_link_update,
1812 	.stats_get = mlx5_stats_get,
1813 	.stats_reset = mlx5_stats_reset,
1814 	.xstats_get = mlx5_xstats_get,
1815 	.xstats_reset = mlx5_xstats_reset,
1816 	.xstats_get_names = mlx5_xstats_get_names,
1817 	.fw_version_get = mlx5_fw_version_get,
1818 	.dev_infos_get = mlx5_dev_infos_get,
1819 	.representor_info_get = mlx5_representor_info_get,
1820 	.read_clock = mlx5_txpp_read_clock,
1821 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1822 	.vlan_filter_set = mlx5_vlan_filter_set,
1823 	.rx_queue_setup = mlx5_rx_queue_setup,
1824 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1825 	.tx_queue_setup = mlx5_tx_queue_setup,
1826 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1827 	.rx_queue_release = mlx5_rx_queue_release,
1828 	.tx_queue_release = mlx5_tx_queue_release,
1829 	.rx_queue_start = mlx5_rx_queue_start,
1830 	.rx_queue_stop = mlx5_rx_queue_stop,
1831 	.tx_queue_start = mlx5_tx_queue_start,
1832 	.tx_queue_stop = mlx5_tx_queue_stop,
1833 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1834 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1835 	.mac_addr_remove = mlx5_mac_addr_remove,
1836 	.mac_addr_add = mlx5_mac_addr_add,
1837 	.mac_addr_set = mlx5_mac_addr_set,
1838 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1839 	.mtu_set = mlx5_dev_set_mtu,
1840 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1841 	.vlan_offload_set = mlx5_vlan_offload_set,
1842 	.reta_update = mlx5_dev_rss_reta_update,
1843 	.reta_query = mlx5_dev_rss_reta_query,
1844 	.rss_hash_update = mlx5_rss_hash_update,
1845 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1846 	.flow_ops_get = mlx5_flow_ops_get,
1847 	.rxq_info_get = mlx5_rxq_info_get,
1848 	.txq_info_get = mlx5_txq_info_get,
1849 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1850 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1851 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1852 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1853 	.is_removed = mlx5_is_removed,
1854 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1855 	.get_module_info = mlx5_get_module_info,
1856 	.get_module_eeprom = mlx5_get_module_eeprom,
1857 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1858 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1859 	.hairpin_bind = mlx5_hairpin_bind,
1860 	.hairpin_unbind = mlx5_hairpin_unbind,
1861 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1862 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1863 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1864 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1865 	.get_monitor_addr = mlx5_get_monitor_addr,
1866 };
1867 
1868 /* Available operations from secondary process. */
1869 const struct eth_dev_ops mlx5_dev_sec_ops = {
1870 	.stats_get = mlx5_stats_get,
1871 	.stats_reset = mlx5_stats_reset,
1872 	.xstats_get = mlx5_xstats_get,
1873 	.xstats_reset = mlx5_xstats_reset,
1874 	.xstats_get_names = mlx5_xstats_get_names,
1875 	.fw_version_get = mlx5_fw_version_get,
1876 	.dev_infos_get = mlx5_dev_infos_get,
1877 	.representor_info_get = mlx5_representor_info_get,
1878 	.read_clock = mlx5_txpp_read_clock,
1879 	.rx_queue_start = mlx5_rx_queue_start,
1880 	.rx_queue_stop = mlx5_rx_queue_stop,
1881 	.tx_queue_start = mlx5_tx_queue_start,
1882 	.tx_queue_stop = mlx5_tx_queue_stop,
1883 	.rxq_info_get = mlx5_rxq_info_get,
1884 	.txq_info_get = mlx5_txq_info_get,
1885 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1886 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1887 	.get_module_info = mlx5_get_module_info,
1888 	.get_module_eeprom = mlx5_get_module_eeprom,
1889 };
1890 
1891 /* Available operations in flow isolated mode. */
1892 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1893 	.dev_configure = mlx5_dev_configure,
1894 	.dev_start = mlx5_dev_start,
1895 	.dev_stop = mlx5_dev_stop,
1896 	.dev_set_link_down = mlx5_set_link_down,
1897 	.dev_set_link_up = mlx5_set_link_up,
1898 	.dev_close = mlx5_dev_close,
1899 	.promiscuous_enable = mlx5_promiscuous_enable,
1900 	.promiscuous_disable = mlx5_promiscuous_disable,
1901 	.allmulticast_enable = mlx5_allmulticast_enable,
1902 	.allmulticast_disable = mlx5_allmulticast_disable,
1903 	.link_update = mlx5_link_update,
1904 	.stats_get = mlx5_stats_get,
1905 	.stats_reset = mlx5_stats_reset,
1906 	.xstats_get = mlx5_xstats_get,
1907 	.xstats_reset = mlx5_xstats_reset,
1908 	.xstats_get_names = mlx5_xstats_get_names,
1909 	.fw_version_get = mlx5_fw_version_get,
1910 	.dev_infos_get = mlx5_dev_infos_get,
1911 	.representor_info_get = mlx5_representor_info_get,
1912 	.read_clock = mlx5_txpp_read_clock,
1913 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1914 	.vlan_filter_set = mlx5_vlan_filter_set,
1915 	.rx_queue_setup = mlx5_rx_queue_setup,
1916 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1917 	.tx_queue_setup = mlx5_tx_queue_setup,
1918 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1919 	.rx_queue_release = mlx5_rx_queue_release,
1920 	.tx_queue_release = mlx5_tx_queue_release,
1921 	.rx_queue_start = mlx5_rx_queue_start,
1922 	.rx_queue_stop = mlx5_rx_queue_stop,
1923 	.tx_queue_start = mlx5_tx_queue_start,
1924 	.tx_queue_stop = mlx5_tx_queue_stop,
1925 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1926 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1927 	.mac_addr_remove = mlx5_mac_addr_remove,
1928 	.mac_addr_add = mlx5_mac_addr_add,
1929 	.mac_addr_set = mlx5_mac_addr_set,
1930 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1931 	.mtu_set = mlx5_dev_set_mtu,
1932 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1933 	.vlan_offload_set = mlx5_vlan_offload_set,
1934 	.flow_ops_get = mlx5_flow_ops_get,
1935 	.rxq_info_get = mlx5_rxq_info_get,
1936 	.txq_info_get = mlx5_txq_info_get,
1937 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1938 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1939 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1940 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1941 	.is_removed = mlx5_is_removed,
1942 	.get_module_info = mlx5_get_module_info,
1943 	.get_module_eeprom = mlx5_get_module_eeprom,
1944 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1945 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1946 	.hairpin_bind = mlx5_hairpin_bind,
1947 	.hairpin_unbind = mlx5_hairpin_unbind,
1948 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1949 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1950 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1951 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1952 	.get_monitor_addr = mlx5_get_monitor_addr,
1953 };
1954 
1955 /**
1956  * Verify and store value for device argument.
1957  *
1958  * @param[in] key
1959  *   Key argument to verify.
1960  * @param[in] val
1961  *   Value associated with key.
1962  * @param opaque
1963  *   User data.
1964  *
1965  * @return
1966  *   0 on success, a negative errno value otherwise and rte_errno is set.
1967  */
1968 static int
1969 mlx5_args_check(const char *key, const char *val, void *opaque)
1970 {
1971 	struct mlx5_dev_config *config = opaque;
1972 	unsigned long mod;
1973 	signed long tmp;
1974 
1975 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1976 	if (!strcmp(MLX5_DRIVER_KEY, key) || !strcmp(MLX5_REPRESENTOR, key) ||
1977 	    !strcmp(MLX5_SYS_MEM_EN, key) || !strcmp(MLX5_TX_DB_NC, key) ||
1978 	    !strcmp(MLX5_MR_MEMPOOL_REG_EN, key) ||
1979 	    !strcmp(MLX5_MR_EXT_MEMSEG_EN, key))
1980 		return 0;
1981 	errno = 0;
1982 	tmp = strtol(val, NULL, 0);
1983 	if (errno) {
1984 		rte_errno = errno;
1985 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1986 		return -rte_errno;
1987 	}
1988 	if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1989 		/* Negative values are acceptable for some keys only. */
1990 		rte_errno = EINVAL;
1991 		DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1992 		return -rte_errno;
1993 	}
1994 	mod = tmp >= 0 ? tmp : -tmp;
1995 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1996 		if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1997 			DRV_LOG(ERR, "invalid CQE compression "
1998 				     "format parameter");
1999 			rte_errno = EINVAL;
2000 			return -rte_errno;
2001 		}
2002 		config->cqe_comp = !!tmp;
2003 		config->cqe_comp_fmt = tmp;
2004 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
2005 		config->hw_padding = !!tmp;
2006 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
2007 		config->mprq.enabled = !!tmp;
2008 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
2009 		config->mprq.stride_num_n = tmp;
2010 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
2011 		config->mprq.stride_size_n = tmp;
2012 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
2013 		config->mprq.max_memcpy_len = tmp;
2014 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
2015 		config->mprq.min_rxqs_num = tmp;
2016 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
2017 		DRV_LOG(WARNING, "%s: deprecated parameter,"
2018 				 " converted to txq_inline_max", key);
2019 		config->txq_inline_max = tmp;
2020 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
2021 		config->txq_inline_max = tmp;
2022 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
2023 		config->txq_inline_min = tmp;
2024 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
2025 		config->txq_inline_mpw = tmp;
2026 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
2027 		config->txqs_inline = tmp;
2028 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
2029 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2030 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
2031 		config->mps = !!tmp;
2032 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
2033 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2034 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
2035 		DRV_LOG(WARNING, "%s: deprecated parameter,"
2036 				 " converted to txq_inline_mpw", key);
2037 		config->txq_inline_mpw = tmp;
2038 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
2039 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2040 	} else if (strcmp(MLX5_TX_PP, key) == 0) {
2041 		if (!mod) {
2042 			DRV_LOG(ERR, "Zero Tx packet pacing parameter");
2043 			rte_errno = EINVAL;
2044 			return -rte_errno;
2045 		}
2046 		config->tx_pp = tmp;
2047 	} else if (strcmp(MLX5_TX_SKEW, key) == 0) {
2048 		config->tx_skew = tmp;
2049 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
2050 		config->rx_vec_en = !!tmp;
2051 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
2052 		config->l3_vxlan_en = !!tmp;
2053 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
2054 		config->vf_nl_en = !!tmp;
2055 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
2056 		config->dv_esw_en = !!tmp;
2057 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
2058 		config->dv_flow_en = !!tmp;
2059 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
2060 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
2061 		    tmp != MLX5_XMETA_MODE_META16 &&
2062 		    tmp != MLX5_XMETA_MODE_META32 &&
2063 		    tmp != MLX5_XMETA_MODE_MISS_INFO) {
2064 			DRV_LOG(ERR, "invalid extensive "
2065 				     "metadata parameter");
2066 			rte_errno = EINVAL;
2067 			return -rte_errno;
2068 		}
2069 		if (tmp != MLX5_XMETA_MODE_MISS_INFO)
2070 			config->dv_xmeta_en = tmp;
2071 		else
2072 			config->dv_miss_info = 1;
2073 	} else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
2074 		config->lacp_by_user = !!tmp;
2075 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
2076 		config->max_dump_files_num = tmp;
2077 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
2078 		config->lro.timeout = tmp;
2079 	} else if (strcmp(RTE_DEVARGS_KEY_CLASS, key) == 0) {
2080 		DRV_LOG(DEBUG, "class argument is %s.", val);
2081 	} else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
2082 		config->log_hp_size = tmp;
2083 	} else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
2084 		if (tmp != MLX5_RCM_NONE &&
2085 		    tmp != MLX5_RCM_LIGHT &&
2086 		    tmp != MLX5_RCM_AGGR) {
2087 			DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
2088 			rte_errno = EINVAL;
2089 			return -rte_errno;
2090 		}
2091 		config->reclaim_mode = tmp;
2092 	} else if (strcmp(MLX5_DECAP_EN, key) == 0) {
2093 		config->decap_en = !!tmp;
2094 	} else if (strcmp(MLX5_ALLOW_DUPLICATE_PATTERN, key) == 0) {
2095 		config->allow_duplicate_pattern = !!tmp;
2096 	} else {
2097 		DRV_LOG(WARNING, "%s: unknown parameter", key);
2098 		rte_errno = EINVAL;
2099 		return -rte_errno;
2100 	}
2101 	return 0;
2102 }
2103 
2104 /**
2105  * Parse device parameters.
2106  *
2107  * @param config
2108  *   Pointer to device configuration structure.
2109  * @param devargs
2110  *   Device arguments structure.
2111  *
2112  * @return
2113  *   0 on success, a negative errno value otherwise and rte_errno is set.
2114  */
2115 int
2116 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
2117 {
2118 	const char **params = (const char *[]){
2119 		MLX5_DRIVER_KEY,
2120 		MLX5_RXQ_CQE_COMP_EN,
2121 		MLX5_RXQ_PKT_PAD_EN,
2122 		MLX5_RX_MPRQ_EN,
2123 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
2124 		MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
2125 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
2126 		MLX5_RXQS_MIN_MPRQ,
2127 		MLX5_TXQ_INLINE,
2128 		MLX5_TXQ_INLINE_MIN,
2129 		MLX5_TXQ_INLINE_MAX,
2130 		MLX5_TXQ_INLINE_MPW,
2131 		MLX5_TXQS_MIN_INLINE,
2132 		MLX5_TXQS_MAX_VEC,
2133 		MLX5_TXQ_MPW_EN,
2134 		MLX5_TXQ_MPW_HDR_DSEG_EN,
2135 		MLX5_TXQ_MAX_INLINE_LEN,
2136 		MLX5_TX_DB_NC,
2137 		MLX5_TX_PP,
2138 		MLX5_TX_SKEW,
2139 		MLX5_TX_VEC_EN,
2140 		MLX5_RX_VEC_EN,
2141 		MLX5_L3_VXLAN_EN,
2142 		MLX5_VF_NL_EN,
2143 		MLX5_DV_ESW_EN,
2144 		MLX5_DV_FLOW_EN,
2145 		MLX5_DV_XMETA_EN,
2146 		MLX5_LACP_BY_USER,
2147 		MLX5_MR_EXT_MEMSEG_EN,
2148 		MLX5_REPRESENTOR,
2149 		MLX5_MAX_DUMP_FILES_NUM,
2150 		MLX5_LRO_TIMEOUT_USEC,
2151 		RTE_DEVARGS_KEY_CLASS,
2152 		MLX5_HP_BUF_SIZE,
2153 		MLX5_RECLAIM_MEM,
2154 		MLX5_SYS_MEM_EN,
2155 		MLX5_DECAP_EN,
2156 		MLX5_ALLOW_DUPLICATE_PATTERN,
2157 		MLX5_MR_MEMPOOL_REG_EN,
2158 		NULL,
2159 	};
2160 	struct rte_kvargs *kvlist;
2161 	int ret = 0;
2162 	int i;
2163 
2164 	if (devargs == NULL)
2165 		return 0;
2166 	/* Following UGLY cast is done to pass checkpatch. */
2167 	kvlist = rte_kvargs_parse(devargs->args, params);
2168 	if (kvlist == NULL) {
2169 		rte_errno = EINVAL;
2170 		return -rte_errno;
2171 	}
2172 	/* Process parameters. */
2173 	for (i = 0; (params[i] != NULL); ++i) {
2174 		if (rte_kvargs_count(kvlist, params[i])) {
2175 			ret = rte_kvargs_process(kvlist, params[i],
2176 						 mlx5_args_check, config);
2177 			if (ret) {
2178 				rte_errno = EINVAL;
2179 				rte_kvargs_free(kvlist);
2180 				return -rte_errno;
2181 			}
2182 		}
2183 	}
2184 	rte_kvargs_free(kvlist);
2185 	return 0;
2186 }
2187 
2188 /**
2189  * Configures the minimal amount of data to inline into WQE
2190  * while sending packets.
2191  *
2192  * - the txq_inline_min has the maximal priority, if this
2193  *   key is specified in devargs
2194  * - if DevX is enabled the inline mode is queried from the
2195  *   device (HCA attributes and NIC vport context if needed).
2196  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
2197  *   and none (0 bytes) for other NICs
2198  *
2199  * @param spawn
2200  *   Verbs device parameters (name, port, switch_info) to spawn.
2201  * @param config
2202  *   Device configuration parameters.
2203  */
2204 void
2205 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
2206 		    struct mlx5_dev_config *config)
2207 {
2208 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
2209 		/* Application defines size of inlined data explicitly. */
2210 		if (spawn->pci_dev != NULL) {
2211 			switch (spawn->pci_dev->id.device_id) {
2212 			case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2213 			case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2214 				if (config->txq_inline_min <
2215 					       (int)MLX5_INLINE_HSIZE_L2) {
2216 					DRV_LOG(DEBUG,
2217 						"txq_inline_mix aligned to minimal ConnectX-4 required value %d",
2218 						(int)MLX5_INLINE_HSIZE_L2);
2219 					config->txq_inline_min =
2220 							MLX5_INLINE_HSIZE_L2;
2221 				}
2222 				break;
2223 			}
2224 		}
2225 		goto exit;
2226 	}
2227 	if (config->hca_attr.eth_net_offloads) {
2228 		/* We have DevX enabled, inline mode queried successfully. */
2229 		switch (config->hca_attr.wqe_inline_mode) {
2230 		case MLX5_CAP_INLINE_MODE_L2:
2231 			/* outer L2 header must be inlined. */
2232 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2233 			goto exit;
2234 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
2235 			/* No inline data are required by NIC. */
2236 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2237 			config->hw_vlan_insert =
2238 				config->hca_attr.wqe_vlan_insert;
2239 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
2240 			goto exit;
2241 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
2242 			/* inline mode is defined by NIC vport context. */
2243 			if (!config->hca_attr.eth_virt)
2244 				break;
2245 			switch (config->hca_attr.vport_inline_mode) {
2246 			case MLX5_INLINE_MODE_NONE:
2247 				config->txq_inline_min =
2248 					MLX5_INLINE_HSIZE_NONE;
2249 				goto exit;
2250 			case MLX5_INLINE_MODE_L2:
2251 				config->txq_inline_min =
2252 					MLX5_INLINE_HSIZE_L2;
2253 				goto exit;
2254 			case MLX5_INLINE_MODE_IP:
2255 				config->txq_inline_min =
2256 					MLX5_INLINE_HSIZE_L3;
2257 				goto exit;
2258 			case MLX5_INLINE_MODE_TCP_UDP:
2259 				config->txq_inline_min =
2260 					MLX5_INLINE_HSIZE_L4;
2261 				goto exit;
2262 			case MLX5_INLINE_MODE_INNER_L2:
2263 				config->txq_inline_min =
2264 					MLX5_INLINE_HSIZE_INNER_L2;
2265 				goto exit;
2266 			case MLX5_INLINE_MODE_INNER_IP:
2267 				config->txq_inline_min =
2268 					MLX5_INLINE_HSIZE_INNER_L3;
2269 				goto exit;
2270 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
2271 				config->txq_inline_min =
2272 					MLX5_INLINE_HSIZE_INNER_L4;
2273 				goto exit;
2274 			}
2275 		}
2276 	}
2277 	if (spawn->pci_dev == NULL) {
2278 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2279 		goto exit;
2280 	}
2281 	/*
2282 	 * We get here if we are unable to deduce
2283 	 * inline data size with DevX. Try PCI ID
2284 	 * to determine old NICs.
2285 	 */
2286 	switch (spawn->pci_dev->id.device_id) {
2287 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2288 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2289 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2290 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2291 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2292 		config->hw_vlan_insert = 0;
2293 		break;
2294 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2295 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2296 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2297 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2298 		/*
2299 		 * These NICs support VLAN insertion from WQE and
2300 		 * report the wqe_vlan_insert flag. But there is the bug
2301 		 * and PFC control may be broken, so disable feature.
2302 		 */
2303 		config->hw_vlan_insert = 0;
2304 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2305 		break;
2306 	default:
2307 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2308 		break;
2309 	}
2310 exit:
2311 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2312 }
2313 
2314 /**
2315  * Configures the metadata mask fields in the shared context.
2316  *
2317  * @param [in] dev
2318  *   Pointer to Ethernet device.
2319  */
2320 void
2321 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2322 {
2323 	struct mlx5_priv *priv = dev->data->dev_private;
2324 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2325 	uint32_t meta, mark, reg_c0;
2326 
2327 	reg_c0 = ~priv->vport_meta_mask;
2328 	switch (priv->config.dv_xmeta_en) {
2329 	case MLX5_XMETA_MODE_LEGACY:
2330 		meta = UINT32_MAX;
2331 		mark = MLX5_FLOW_MARK_MASK;
2332 		break;
2333 	case MLX5_XMETA_MODE_META16:
2334 		meta = reg_c0 >> rte_bsf32(reg_c0);
2335 		mark = MLX5_FLOW_MARK_MASK;
2336 		break;
2337 	case MLX5_XMETA_MODE_META32:
2338 		meta = UINT32_MAX;
2339 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2340 		break;
2341 	default:
2342 		meta = 0;
2343 		mark = 0;
2344 		MLX5_ASSERT(false);
2345 		break;
2346 	}
2347 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2348 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2349 				 sh->dv_mark_mask, mark);
2350 	else
2351 		sh->dv_mark_mask = mark;
2352 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2353 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2354 				 sh->dv_meta_mask, meta);
2355 	else
2356 		sh->dv_meta_mask = meta;
2357 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2358 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2359 				 sh->dv_meta_mask, reg_c0);
2360 	else
2361 		sh->dv_regc0_mask = reg_c0;
2362 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2363 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2364 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2365 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2366 }
2367 
2368 int
2369 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2370 {
2371 	static const char *const dynf_names[] = {
2372 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2373 		RTE_MBUF_DYNFLAG_METADATA_NAME,
2374 		RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2375 	};
2376 	unsigned int i;
2377 
2378 	if (n < RTE_DIM(dynf_names))
2379 		return -ENOMEM;
2380 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
2381 		if (names[i] == NULL)
2382 			return -EINVAL;
2383 		strcpy(names[i], dynf_names[i]);
2384 	}
2385 	return RTE_DIM(dynf_names);
2386 }
2387 
2388 /**
2389  * Comparison callback to sort device data.
2390  *
2391  * This is meant to be used with qsort().
2392  *
2393  * @param a[in]
2394  *   Pointer to pointer to first data object.
2395  * @param b[in]
2396  *   Pointer to pointer to second data object.
2397  *
2398  * @return
2399  *   0 if both objects are equal, less than 0 if the first argument is less
2400  *   than the second, greater than 0 otherwise.
2401  */
2402 int
2403 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2404 			      struct mlx5_dev_config *config,
2405 			      struct rte_device *dpdk_dev)
2406 {
2407 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2408 	struct mlx5_dev_config *sh_conf = NULL;
2409 	uint16_t port_id;
2410 
2411 	MLX5_ASSERT(sh);
2412 	/* Nothing to compare for the single/first device. */
2413 	if (sh->refcnt == 1)
2414 		return 0;
2415 	/* Find the device with shared context. */
2416 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
2417 		struct mlx5_priv *opriv =
2418 			rte_eth_devices[port_id].data->dev_private;
2419 
2420 		if (opriv && opriv != priv && opriv->sh == sh) {
2421 			sh_conf = &opriv->config;
2422 			break;
2423 		}
2424 	}
2425 	if (!sh_conf)
2426 		return 0;
2427 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2428 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2429 			     " for shared %s context", sh->ibdev_name);
2430 		rte_errno = EINVAL;
2431 		return rte_errno;
2432 	}
2433 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2434 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2435 			     " for shared %s context", sh->ibdev_name);
2436 		rte_errno = EINVAL;
2437 		return rte_errno;
2438 	}
2439 	return 0;
2440 }
2441 
2442 /**
2443  * Look for the ethernet device belonging to mlx5 driver.
2444  *
2445  * @param[in] port_id
2446  *   port_id to start looking for device.
2447  * @param[in] odev
2448  *   Pointer to the hint device. When device is being probed
2449  *   the its siblings (master and preceding representors might
2450  *   not have assigned driver yet (because the mlx5_os_pci_probe()
2451  *   is not completed yet, for this case match on hint
2452  *   device may be used to detect sibling device.
2453  *
2454  * @return
2455  *   port_id of found device, RTE_MAX_ETHPORT if not found.
2456  */
2457 uint16_t
2458 mlx5_eth_find_next(uint16_t port_id, struct rte_device *odev)
2459 {
2460 	while (port_id < RTE_MAX_ETHPORTS) {
2461 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2462 
2463 		if (dev->state != RTE_ETH_DEV_UNUSED &&
2464 		    dev->device &&
2465 		    (dev->device == odev ||
2466 		     (dev->device->driver &&
2467 		     dev->device->driver->name &&
2468 		     ((strcmp(dev->device->driver->name,
2469 			      MLX5_PCI_DRIVER_NAME) == 0) ||
2470 		      (strcmp(dev->device->driver->name,
2471 			      MLX5_AUXILIARY_DRIVER_NAME) == 0)))))
2472 			break;
2473 		port_id++;
2474 	}
2475 	if (port_id >= RTE_MAX_ETHPORTS)
2476 		return RTE_MAX_ETHPORTS;
2477 	return port_id;
2478 }
2479 
2480 /**
2481  * Callback to remove a device.
2482  *
2483  * This function removes all Ethernet devices belong to a given device.
2484  *
2485  * @param[in] cdev
2486  *   Pointer to the generic device.
2487  *
2488  * @return
2489  *   0 on success, the function cannot fail.
2490  */
2491 int
2492 mlx5_net_remove(struct mlx5_common_device *cdev)
2493 {
2494 	uint16_t port_id;
2495 	int ret = 0;
2496 
2497 	RTE_ETH_FOREACH_DEV_OF(port_id, cdev->dev) {
2498 		/*
2499 		 * mlx5_dev_close() is not registered to secondary process,
2500 		 * call the close function explicitly for secondary process.
2501 		 */
2502 		if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2503 			ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2504 		else
2505 			ret |= rte_eth_dev_close(port_id);
2506 	}
2507 	return ret == 0 ? 0 : -EIO;
2508 }
2509 
2510 static const struct rte_pci_id mlx5_pci_id_map[] = {
2511 	{
2512 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2513 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2514 	},
2515 	{
2516 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2517 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2518 	},
2519 	{
2520 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2521 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2522 	},
2523 	{
2524 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2525 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2526 	},
2527 	{
2528 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2529 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2530 	},
2531 	{
2532 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2533 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2534 	},
2535 	{
2536 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2537 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2538 	},
2539 	{
2540 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2541 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2542 	},
2543 	{
2544 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2545 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2546 	},
2547 	{
2548 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2549 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2550 	},
2551 	{
2552 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2553 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2554 	},
2555 	{
2556 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2557 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2558 	},
2559 	{
2560 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2561 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2562 	},
2563 	{
2564 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2565 				PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2566 	},
2567 	{
2568 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2569 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2570 	},
2571 	{
2572 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2573 				PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2574 	},
2575 	{
2576 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2577 				PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2578 	},
2579 	{
2580 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2581 				PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2582 	},
2583 	{
2584 		.vendor_id = 0
2585 	}
2586 };
2587 
2588 static struct mlx5_class_driver mlx5_net_driver = {
2589 	.drv_class = MLX5_CLASS_ETH,
2590 	.name = RTE_STR(MLX5_ETH_DRIVER_NAME),
2591 	.id_table = mlx5_pci_id_map,
2592 	.probe = mlx5_os_net_probe,
2593 	.remove = mlx5_net_remove,
2594 	.probe_again = 1,
2595 	.intr_lsc = 1,
2596 	.intr_rmv = 1,
2597 };
2598 
2599 /* Initialize driver log type. */
2600 RTE_LOG_REGISTER_DEFAULT(mlx5_logtype, NOTICE)
2601 
2602 /**
2603  * Driver initialization routine.
2604  */
2605 RTE_INIT(rte_mlx5_pmd_init)
2606 {
2607 	pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2608 	mlx5_common_init();
2609 	/* Build the static tables for Verbs conversion. */
2610 	mlx5_set_ptype_table();
2611 	mlx5_set_cksum_table();
2612 	mlx5_set_swp_types_table();
2613 	if (mlx5_glue)
2614 		mlx5_class_driver_register(&mlx5_net_driver);
2615 }
2616 
2617 RTE_PMD_EXPORT_NAME(MLX5_ETH_DRIVER_NAME, __COUNTER__);
2618 RTE_PMD_REGISTER_PCI_TABLE(MLX5_ETH_DRIVER_NAME, mlx5_pci_id_map);
2619 RTE_PMD_REGISTER_KMOD_DEP(MLX5_ETH_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");
2620