xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 9f3b3a96dec2f4c01cc92a132d763b8887d29e6a)
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 <rte_ethdev_driver.h>
15 #include <rte_ethdev_pci.h>
16 #include <rte_pci.h>
17 #include <rte_bus_pci.h>
18 #include <rte_common.h>
19 #include <rte_kvargs.h>
20 #include <rte_rwlock.h>
21 #include <rte_spinlock.h>
22 #include <rte_string_fns.h>
23 #include <rte_alarm.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_common_pci.h>
31 #include <mlx5_malloc.h>
32 
33 #include "mlx5_defs.h"
34 #include "mlx5.h"
35 #include "mlx5_utils.h"
36 #include "mlx5_rxtx.h"
37 #include "mlx5_autoconf.h"
38 #include "mlx5_mr.h"
39 #include "mlx5_flow.h"
40 #include "rte_pmd_mlx5.h"
41 
42 /* Device parameter to enable RX completion queue compression. */
43 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
44 
45 /* Device parameter to enable RX completion entry padding to 128B. */
46 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
47 
48 /* Device parameter to enable padding Rx packet to cacheline size. */
49 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
50 
51 /* Device parameter to enable Multi-Packet Rx queue. */
52 #define MLX5_RX_MPRQ_EN "mprq_en"
53 
54 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
55 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
56 
57 /* Device parameter to configure log 2 of the stride size for MPRQ. */
58 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
59 
60 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
61 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
62 
63 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
64 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
65 
66 /* Device parameter to configure inline send. Deprecated, ignored.*/
67 #define MLX5_TXQ_INLINE "txq_inline"
68 
69 /* Device parameter to limit packet size to inline with ordinary SEND. */
70 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
71 
72 /* Device parameter to configure minimal data size to inline. */
73 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
74 
75 /* Device parameter to limit packet size to inline with Enhanced MPW. */
76 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
77 
78 /*
79  * Device parameter to configure the number of TX queues threshold for
80  * enabling inline send.
81  */
82 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
83 
84 /*
85  * Device parameter to configure the number of TX queues threshold for
86  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
87  */
88 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
89 
90 /* Device parameter to enable multi-packet send WQEs. */
91 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
92 
93 /*
94  * Device parameter to force doorbell register mapping
95  * to non-cahed region eliminating the extra write memory barrier.
96  */
97 #define MLX5_TX_DB_NC "tx_db_nc"
98 
99 /*
100  * Device parameter to include 2 dsegs in the title WQEBB.
101  * Deprecated, ignored.
102  */
103 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
104 
105 /*
106  * Device parameter to limit the size of inlining packet.
107  * Deprecated, ignored.
108  */
109 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
110 
111 /*
112  * Device parameter to enable Tx scheduling on timestamps
113  * and specify the packet pacing granularity in nanoseconds.
114  */
115 #define MLX5_TX_PP "tx_pp"
116 
117 /*
118  * Device parameter to specify skew in nanoseconds on Tx datapath,
119  * it represents the time between SQ start WQE processing and
120  * appearing actual packet data on the wire.
121  */
122 #define MLX5_TX_SKEW "tx_skew"
123 
124 /*
125  * Device parameter to enable hardware Tx vector.
126  * Deprecated, ignored (no vectorized Tx routines anymore).
127  */
128 #define MLX5_TX_VEC_EN "tx_vec_en"
129 
130 /* Device parameter to enable hardware Rx vector. */
131 #define MLX5_RX_VEC_EN "rx_vec_en"
132 
133 /* Allow L3 VXLAN flow creation. */
134 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
135 
136 /* Activate DV E-Switch flow steering. */
137 #define MLX5_DV_ESW_EN "dv_esw_en"
138 
139 /* Activate DV flow steering. */
140 #define MLX5_DV_FLOW_EN "dv_flow_en"
141 
142 /* Enable extensive flow metadata support. */
143 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
144 
145 /* Device parameter to let the user manage the lacp traffic of bonded device */
146 #define MLX5_LACP_BY_USER "lacp_by_user"
147 
148 /* Activate Netlink support in VF mode. */
149 #define MLX5_VF_NL_EN "vf_nl_en"
150 
151 /* Enable extending memsegs when creating a MR. */
152 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
153 
154 /* Select port representors to instantiate. */
155 #define MLX5_REPRESENTOR "representor"
156 
157 /* Device parameter to configure the maximum number of dump files per queue. */
158 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
159 
160 /* Configure timeout of LRO session (in microseconds). */
161 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
162 
163 /*
164  * Device parameter to configure the total data buffer size for a single
165  * hairpin queue (logarithm value).
166  */
167 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
168 
169 /* Flow memory reclaim mode. */
170 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
171 
172 /* The default memory allocator used in PMD. */
173 #define MLX5_SYS_MEM_EN "sys_mem_en"
174 /* Decap will be used or not. */
175 #define MLX5_DECAP_EN "decap_en"
176 
177 /* Shared memory between primary and secondary processes. */
178 struct mlx5_shared_data *mlx5_shared_data;
179 
180 /** Driver-specific log messages type. */
181 int mlx5_logtype;
182 
183 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
184 						LIST_HEAD_INITIALIZER();
185 static pthread_mutex_t mlx5_dev_ctx_list_mutex = PTHREAD_MUTEX_INITIALIZER;
186 
187 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
188 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
189 	{
190 		.size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
191 		.trunk_size = 64,
192 		.grow_trunk = 3,
193 		.grow_shift = 2,
194 		.need_lock = 1,
195 		.release_mem_en = 1,
196 		.malloc = mlx5_malloc,
197 		.free = mlx5_free,
198 		.type = "mlx5_encap_decap_ipool",
199 	},
200 	{
201 		.size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
202 		.trunk_size = 64,
203 		.grow_trunk = 3,
204 		.grow_shift = 2,
205 		.need_lock = 1,
206 		.release_mem_en = 1,
207 		.malloc = mlx5_malloc,
208 		.free = mlx5_free,
209 		.type = "mlx5_push_vlan_ipool",
210 	},
211 	{
212 		.size = sizeof(struct mlx5_flow_dv_tag_resource),
213 		.trunk_size = 64,
214 		.grow_trunk = 3,
215 		.grow_shift = 2,
216 		.need_lock = 1,
217 		.release_mem_en = 1,
218 		.malloc = mlx5_malloc,
219 		.free = mlx5_free,
220 		.type = "mlx5_tag_ipool",
221 	},
222 	{
223 		.size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
224 		.trunk_size = 64,
225 		.grow_trunk = 3,
226 		.grow_shift = 2,
227 		.need_lock = 1,
228 		.release_mem_en = 1,
229 		.malloc = mlx5_malloc,
230 		.free = mlx5_free,
231 		.type = "mlx5_port_id_ipool",
232 	},
233 	{
234 		.size = sizeof(struct mlx5_flow_tbl_data_entry),
235 		.trunk_size = 64,
236 		.grow_trunk = 3,
237 		.grow_shift = 2,
238 		.need_lock = 1,
239 		.release_mem_en = 1,
240 		.malloc = mlx5_malloc,
241 		.free = mlx5_free,
242 		.type = "mlx5_jump_ipool",
243 	},
244 	{
245 		.size = sizeof(struct mlx5_flow_dv_sample_resource),
246 		.trunk_size = 64,
247 		.grow_trunk = 3,
248 		.grow_shift = 2,
249 		.need_lock = 1,
250 		.release_mem_en = 1,
251 		.malloc = mlx5_malloc,
252 		.free = mlx5_free,
253 		.type = "mlx5_sample_ipool",
254 	},
255 	{
256 		.size = sizeof(struct mlx5_flow_dv_dest_array_resource),
257 		.trunk_size = 64,
258 		.grow_trunk = 3,
259 		.grow_shift = 2,
260 		.need_lock = 1,
261 		.release_mem_en = 1,
262 		.malloc = mlx5_malloc,
263 		.free = mlx5_free,
264 		.type = "mlx5_dest_array_ipool",
265 	},
266 #endif
267 	{
268 		.size = sizeof(struct mlx5_flow_meter),
269 		.trunk_size = 64,
270 		.grow_trunk = 3,
271 		.grow_shift = 2,
272 		.need_lock = 1,
273 		.release_mem_en = 1,
274 		.malloc = mlx5_malloc,
275 		.free = mlx5_free,
276 		.type = "mlx5_meter_ipool",
277 	},
278 	{
279 		.size = sizeof(struct mlx5_flow_mreg_copy_resource),
280 		.trunk_size = 64,
281 		.grow_trunk = 3,
282 		.grow_shift = 2,
283 		.need_lock = 1,
284 		.release_mem_en = 1,
285 		.malloc = mlx5_malloc,
286 		.free = mlx5_free,
287 		.type = "mlx5_mcp_ipool",
288 	},
289 	{
290 		.size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
291 		.trunk_size = 64,
292 		.grow_trunk = 3,
293 		.grow_shift = 2,
294 		.need_lock = 1,
295 		.release_mem_en = 1,
296 		.malloc = mlx5_malloc,
297 		.free = mlx5_free,
298 		.type = "mlx5_hrxq_ipool",
299 	},
300 	{
301 		/*
302 		 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
303 		 * It set in run time according to PCI function configuration.
304 		 */
305 		.size = 0,
306 		.trunk_size = 64,
307 		.grow_trunk = 3,
308 		.grow_shift = 2,
309 		.need_lock = 1,
310 		.release_mem_en = 1,
311 		.malloc = mlx5_malloc,
312 		.free = mlx5_free,
313 		.type = "mlx5_flow_handle_ipool",
314 	},
315 	{
316 		.size = sizeof(struct rte_flow),
317 		.trunk_size = 4096,
318 		.need_lock = 1,
319 		.release_mem_en = 1,
320 		.malloc = mlx5_malloc,
321 		.free = mlx5_free,
322 		.type = "rte_flow_ipool",
323 	},
324 	{
325 		.size = 0,
326 		.need_lock = 1,
327 		.type = "mlx5_flow_rss_id_ipool",
328 	},
329 	{
330 		.size = 0,
331 		.need_lock = 1,
332 		.type = "mlx5_flow_tnl_flow_ipool",
333 	},
334 	{
335 		.size = 0,
336 		.need_lock = 1,
337 		.type = "mlx5_flow_tnl_tbl_ipool",
338 	},
339 };
340 
341 
342 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
343 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
344 
345 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
346 
347 /**
348  * Initialize the shared aging list information per port.
349  *
350  * @param[in] sh
351  *   Pointer to mlx5_dev_ctx_shared object.
352  */
353 static void
354 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
355 {
356 	uint32_t i;
357 	struct mlx5_age_info *age_info;
358 
359 	for (i = 0; i < sh->max_port; i++) {
360 		age_info = &sh->port[i].age_info;
361 		age_info->flags = 0;
362 		TAILQ_INIT(&age_info->aged_counters);
363 		rte_spinlock_init(&age_info->aged_sl);
364 		MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
365 	}
366 }
367 
368 /**
369  * Initialize the counters management structure.
370  *
371  * @param[in] sh
372  *   Pointer to mlx5_dev_ctx_shared object to free
373  */
374 static void
375 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
376 {
377 	int i;
378 
379 	memset(&sh->cmng, 0, sizeof(sh->cmng));
380 	TAILQ_INIT(&sh->cmng.flow_counters);
381 	sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
382 	sh->cmng.max_id = -1;
383 	sh->cmng.last_pool_idx = POOL_IDX_INVALID;
384 	rte_spinlock_init(&sh->cmng.pool_update_sl);
385 	for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
386 		TAILQ_INIT(&sh->cmng.counters[i]);
387 		rte_spinlock_init(&sh->cmng.csl[i]);
388 	}
389 }
390 
391 /**
392  * Destroy all the resources allocated for a counter memory management.
393  *
394  * @param[in] mng
395  *   Pointer to the memory management structure.
396  */
397 static void
398 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
399 {
400 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
401 
402 	LIST_REMOVE(mng, next);
403 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
404 	claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
405 	mlx5_free(mem);
406 }
407 
408 /**
409  * Close and release all the resources of the counters management.
410  *
411  * @param[in] sh
412  *   Pointer to mlx5_dev_ctx_shared object to free.
413  */
414 static void
415 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
416 {
417 	struct mlx5_counter_stats_mem_mng *mng;
418 	int i, j;
419 	int retries = 1024;
420 
421 	rte_errno = 0;
422 	while (--retries) {
423 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
424 		if (rte_errno != EINPROGRESS)
425 			break;
426 		rte_pause();
427 	}
428 
429 	if (sh->cmng.pools) {
430 		struct mlx5_flow_counter_pool *pool;
431 		uint16_t n_valid = sh->cmng.n_valid;
432 		bool fallback = sh->cmng.counter_fallback;
433 
434 		for (i = 0; i < n_valid; ++i) {
435 			pool = sh->cmng.pools[i];
436 			if (!fallback && pool->min_dcs)
437 				claim_zero(mlx5_devx_cmd_destroy
438 							       (pool->min_dcs));
439 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
440 				struct mlx5_flow_counter *cnt =
441 						MLX5_POOL_GET_CNT(pool, j);
442 
443 				if (cnt->action)
444 					claim_zero
445 					 (mlx5_glue->destroy_flow_action
446 					  (cnt->action));
447 				if (fallback && MLX5_POOL_GET_CNT
448 				    (pool, j)->dcs_when_free)
449 					claim_zero(mlx5_devx_cmd_destroy
450 						   (cnt->dcs_when_free));
451 			}
452 			mlx5_free(pool);
453 		}
454 		mlx5_free(sh->cmng.pools);
455 	}
456 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
457 	while (mng) {
458 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
459 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
460 	}
461 	memset(&sh->cmng, 0, sizeof(sh->cmng));
462 }
463 
464 /**
465  * Initialize the flow resources' indexed mempool.
466  *
467  * @param[in] sh
468  *   Pointer to mlx5_dev_ctx_shared object.
469  * @param[in] sh
470  *   Pointer to user dev config.
471  */
472 static void
473 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
474 		       const struct mlx5_dev_config *config)
475 {
476 	uint8_t i;
477 	struct mlx5_indexed_pool_config cfg;
478 
479 	for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
480 		cfg = mlx5_ipool_cfg[i];
481 		switch (i) {
482 		default:
483 			break;
484 		/*
485 		 * Set MLX5_IPOOL_MLX5_FLOW ipool size
486 		 * according to PCI function flow configuration.
487 		 */
488 		case MLX5_IPOOL_MLX5_FLOW:
489 			cfg.size = config->dv_flow_en ?
490 				sizeof(struct mlx5_flow_handle) :
491 				MLX5_FLOW_HANDLE_VERBS_SIZE;
492 			break;
493 		}
494 		if (config->reclaim_mode)
495 			cfg.release_mem_en = 1;
496 		sh->ipool[i] = mlx5_ipool_create(&cfg);
497 	}
498 }
499 
500 /**
501  * Release the flow resources' indexed mempool.
502  *
503  * @param[in] sh
504  *   Pointer to mlx5_dev_ctx_shared object.
505  */
506 static void
507 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
508 {
509 	uint8_t i;
510 
511 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
512 		mlx5_ipool_destroy(sh->ipool[i]);
513 }
514 
515 /*
516  * Check if dynamic flex parser for eCPRI already exists.
517  *
518  * @param dev
519  *   Pointer to Ethernet device structure.
520  *
521  * @return
522  *   true on exists, false on not.
523  */
524 bool
525 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
526 {
527 	struct mlx5_priv *priv = dev->data->dev_private;
528 	struct mlx5_flex_parser_profiles *prf =
529 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
530 
531 	return !!prf->obj;
532 }
533 
534 /*
535  * Allocation of a flex parser for eCPRI. Once created, this parser related
536  * resources will be held until the device is closed.
537  *
538  * @param dev
539  *   Pointer to Ethernet device structure.
540  *
541  * @return
542  *   0 on success, a negative errno value otherwise and rte_errno is set.
543  */
544 int
545 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
546 {
547 	struct mlx5_priv *priv = dev->data->dev_private;
548 	struct mlx5_flex_parser_profiles *prf =
549 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
550 	struct mlx5_devx_graph_node_attr node = {
551 		.modify_field_select = 0,
552 	};
553 	uint32_t ids[8];
554 	int ret;
555 
556 	if (!priv->config.hca_attr.parse_graph_flex_node) {
557 		DRV_LOG(ERR, "Dynamic flex parser is not supported "
558 			"for device %s.", priv->dev_data->name);
559 		return -ENOTSUP;
560 	}
561 	node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
562 	/* 8 bytes now: 4B common header + 4B message body header. */
563 	node.header_length_base_value = 0x8;
564 	/* After MAC layer: Ether / VLAN. */
565 	node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
566 	/* Type of compared condition should be 0xAEFE in the L2 layer. */
567 	node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
568 	/* Sample #0: type in common header. */
569 	node.sample[0].flow_match_sample_en = 1;
570 	/* Fixed offset. */
571 	node.sample[0].flow_match_sample_offset_mode = 0x0;
572 	/* Only the 2nd byte will be used. */
573 	node.sample[0].flow_match_sample_field_base_offset = 0x0;
574 	/* Sample #1: message payload. */
575 	node.sample[1].flow_match_sample_en = 1;
576 	/* Fixed offset. */
577 	node.sample[1].flow_match_sample_offset_mode = 0x0;
578 	/*
579 	 * Only the first two bytes will be used right now, and its offset will
580 	 * start after the common header that with the length of a DW(u32).
581 	 */
582 	node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
583 	prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
584 	if (!prf->obj) {
585 		DRV_LOG(ERR, "Failed to create flex parser node object.");
586 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
587 	}
588 	prf->num = 2;
589 	ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
590 	if (ret) {
591 		DRV_LOG(ERR, "Failed to query sample IDs.");
592 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
593 	}
594 	prf->offset[0] = 0x0;
595 	prf->offset[1] = sizeof(uint32_t);
596 	prf->ids[0] = ids[0];
597 	prf->ids[1] = ids[1];
598 	return 0;
599 }
600 
601 /*
602  * Destroy the flex parser node, including the parser itself, input / output
603  * arcs and DW samples. Resources could be reused then.
604  *
605  * @param dev
606  *   Pointer to Ethernet device structure.
607  */
608 static void
609 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
610 {
611 	struct mlx5_priv *priv = dev->data->dev_private;
612 	struct mlx5_flex_parser_profiles *prf =
613 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
614 
615 	if (prf->obj)
616 		mlx5_devx_cmd_destroy(prf->obj);
617 	prf->obj = NULL;
618 }
619 
620 /*
621  * Allocate Rx and Tx UARs in robust fashion.
622  * This routine handles the following UAR allocation issues:
623  *
624  *  - tries to allocate the UAR with the most appropriate memory
625  *    mapping type from the ones supported by the host
626  *
627  *  - tries to allocate the UAR with non-NULL base address
628  *    OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
629  *    UAR base address if UAR was not the first object in the UAR page.
630  *    It caused the PMD failure and we should try to get another UAR
631  *    till we get the first one with non-NULL base address returned.
632  */
633 static int
634 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
635 		     const struct mlx5_dev_config *config)
636 {
637 	uint32_t uar_mapping, retry;
638 	int err = 0;
639 	void *base_addr;
640 
641 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
642 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
643 		/* Control the mapping type according to the settings. */
644 		uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
645 			      MLX5DV_UAR_ALLOC_TYPE_NC :
646 			      MLX5DV_UAR_ALLOC_TYPE_BF;
647 #else
648 		RTE_SET_USED(config);
649 		/*
650 		 * It seems we have no way to control the memory mapping type
651 		 * for the UAR, the default "Write-Combining" type is supposed.
652 		 * The UAR initialization on queue creation queries the
653 		 * actual mapping type done by Verbs/kernel and setups the
654 		 * PMD datapath accordingly.
655 		 */
656 		uar_mapping = 0;
657 #endif
658 		sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
659 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
660 		if (!sh->tx_uar &&
661 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
662 			if (config->dbnc == MLX5_TXDB_CACHED ||
663 			    config->dbnc == MLX5_TXDB_HEURISTIC)
664 				DRV_LOG(WARNING, "Devarg tx_db_nc setting "
665 						 "is not supported by DevX");
666 			/*
667 			 * In some environments like virtual machine
668 			 * the Write Combining mapped might be not supported
669 			 * and UAR allocation fails. We try "Non-Cached"
670 			 * mapping for the case. The tx_burst routines take
671 			 * the UAR mapping type into account on UAR setup
672 			 * on queue creation.
673 			 */
674 			DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (BF)");
675 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
676 			sh->tx_uar = mlx5_glue->devx_alloc_uar
677 							(sh->ctx, uar_mapping);
678 		} else if (!sh->tx_uar &&
679 			   uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
680 			if (config->dbnc == MLX5_TXDB_NCACHED)
681 				DRV_LOG(WARNING, "Devarg tx_db_nc settings "
682 						 "is not supported by DevX");
683 			/*
684 			 * If Verbs/kernel does not support "Non-Cached"
685 			 * try the "Write-Combining".
686 			 */
687 			DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (NC)");
688 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
689 			sh->tx_uar = mlx5_glue->devx_alloc_uar
690 							(sh->ctx, uar_mapping);
691 		}
692 #endif
693 		if (!sh->tx_uar) {
694 			DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
695 			err = ENOMEM;
696 			goto exit;
697 		}
698 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
699 		if (base_addr)
700 			break;
701 		/*
702 		 * The UARs are allocated by rdma_core within the
703 		 * IB device context, on context closure all UARs
704 		 * will be freed, should be no memory/object leakage.
705 		 */
706 		DRV_LOG(WARNING, "Retrying to allocate Tx DevX UAR");
707 		sh->tx_uar = NULL;
708 	}
709 	/* Check whether we finally succeeded with valid UAR allocation. */
710 	if (!sh->tx_uar) {
711 		DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
712 		err = ENOMEM;
713 		goto exit;
714 	}
715 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
716 		uar_mapping = 0;
717 		sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
718 							(sh->ctx, uar_mapping);
719 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
720 		if (!sh->devx_rx_uar &&
721 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
722 			/*
723 			 * Rx UAR is used to control interrupts only,
724 			 * should be no datapath noticeable impact,
725 			 * can try "Non-Cached" mapping safely.
726 			 */
727 			DRV_LOG(WARNING, "Failed to allocate Rx DevX UAR (BF)");
728 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
729 			sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
730 							(sh->ctx, uar_mapping);
731 		}
732 #endif
733 		if (!sh->devx_rx_uar) {
734 			DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
735 			err = ENOMEM;
736 			goto exit;
737 		}
738 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
739 		if (base_addr)
740 			break;
741 		/*
742 		 * The UARs are allocated by rdma_core within the
743 		 * IB device context, on context closure all UARs
744 		 * will be freed, should be no memory/object leakage.
745 		 */
746 		DRV_LOG(WARNING, "Retrying to allocate Rx DevX UAR");
747 		sh->devx_rx_uar = NULL;
748 	}
749 	/* Check whether we finally succeeded with valid UAR allocation. */
750 	if (!sh->devx_rx_uar) {
751 		DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
752 		err = ENOMEM;
753 	}
754 exit:
755 	return err;
756 }
757 
758 /**
759  * Allocate shared device context. If there is multiport device the
760  * master and representors will share this context, if there is single
761  * port dedicated device, the context will be used by only given
762  * port due to unification.
763  *
764  * Routine first searches the context for the specified device name,
765  * if found the shared context assumed and reference counter is incremented.
766  * If no context found the new one is created and initialized with specified
767  * device context and parameters.
768  *
769  * @param[in] spawn
770  *   Pointer to the device attributes (name, port, etc).
771  * @param[in] config
772  *   Pointer to device configuration structure.
773  *
774  * @return
775  *   Pointer to mlx5_dev_ctx_shared object on success,
776  *   otherwise NULL and rte_errno is set.
777  */
778 struct mlx5_dev_ctx_shared *
779 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
780 			   const struct mlx5_dev_config *config)
781 {
782 	struct mlx5_dev_ctx_shared *sh;
783 	int err = 0;
784 	uint32_t i;
785 	struct mlx5_devx_tis_attr tis_attr = { 0 };
786 
787 	MLX5_ASSERT(spawn);
788 	/* Secondary process should not create the shared context. */
789 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
790 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
791 	/* Search for IB context by device name. */
792 	LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
793 		if (!strcmp(sh->ibdev_name,
794 			mlx5_os_get_dev_device_name(spawn->phys_dev))) {
795 			sh->refcnt++;
796 			goto exit;
797 		}
798 	}
799 	/* No device found, we have to create new shared context. */
800 	MLX5_ASSERT(spawn->max_port);
801 	sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
802 			 sizeof(struct mlx5_dev_ctx_shared) +
803 			 spawn->max_port *
804 			 sizeof(struct mlx5_dev_shared_port),
805 			 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
806 	if (!sh) {
807 		DRV_LOG(ERR, "shared context allocation failure");
808 		rte_errno  = ENOMEM;
809 		goto exit;
810 	}
811 	err = mlx5_os_open_device(spawn, config, sh);
812 	if (!sh->ctx)
813 		goto error;
814 	err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
815 	if (err) {
816 		DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
817 		goto error;
818 	}
819 	sh->refcnt = 1;
820 	sh->max_port = spawn->max_port;
821 	strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
822 		sizeof(sh->ibdev_name) - 1);
823 	strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
824 		sizeof(sh->ibdev_path) - 1);
825 	/*
826 	 * Setting port_id to max unallowed value means
827 	 * there is no interrupt subhandler installed for
828 	 * the given port index i.
829 	 */
830 	for (i = 0; i < sh->max_port; i++) {
831 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
832 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
833 	}
834 	sh->pd = mlx5_glue->alloc_pd(sh->ctx);
835 	if (sh->pd == NULL) {
836 		DRV_LOG(ERR, "PD allocation failure");
837 		err = ENOMEM;
838 		goto error;
839 	}
840 	if (sh->devx) {
841 		/* Query the EQN for this core. */
842 		err = mlx5_glue->devx_query_eqn(sh->ctx, 0, &sh->eqn);
843 		if (err) {
844 			rte_errno = errno;
845 			DRV_LOG(ERR, "Failed to query event queue number %d.",
846 				rte_errno);
847 			goto error;
848 		}
849 		err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
850 		if (err) {
851 			DRV_LOG(ERR, "Fail to extract pdn from PD");
852 			goto error;
853 		}
854 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
855 		if (!sh->td) {
856 			DRV_LOG(ERR, "TD allocation failure");
857 			err = ENOMEM;
858 			goto error;
859 		}
860 		tis_attr.transport_domain = sh->td->id;
861 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
862 		if (!sh->tis) {
863 			DRV_LOG(ERR, "TIS allocation failure");
864 			err = ENOMEM;
865 			goto error;
866 		}
867 		err = mlx5_alloc_rxtx_uars(sh, config);
868 		if (err)
869 			goto error;
870 		MLX5_ASSERT(sh->tx_uar);
871 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
872 
873 		MLX5_ASSERT(sh->devx_rx_uar);
874 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
875 	}
876 #ifndef RTE_ARCH_64
877 	/* Initialize UAR access locks for 32bit implementations. */
878 	rte_spinlock_init(&sh->uar_lock_cq);
879 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
880 		rte_spinlock_init(&sh->uar_lock[i]);
881 #endif
882 	/*
883 	 * Once the device is added to the list of memory event
884 	 * callback, its global MR cache table cannot be expanded
885 	 * on the fly because of deadlock. If it overflows, lookup
886 	 * should be done by searching MR list linearly, which is slow.
887 	 *
888 	 * At this point the device is not added to the memory
889 	 * event list yet, context is just being created.
890 	 */
891 	err = mlx5_mr_btree_init(&sh->share_cache.cache,
892 				 MLX5_MR_BTREE_CACHE_N * 2,
893 				 spawn->pci_dev->device.numa_node);
894 	if (err) {
895 		err = rte_errno;
896 		goto error;
897 	}
898 	mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
899 			      &sh->share_cache.dereg_mr_cb);
900 	mlx5_os_dev_shared_handler_install(sh);
901 	sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
902 	if (!sh->cnt_id_tbl) {
903 		err = rte_errno;
904 		goto error;
905 	}
906 	mlx5_flow_aging_init(sh);
907 	mlx5_flow_counters_mng_init(sh);
908 	mlx5_flow_ipool_create(sh, config);
909 	/* Add device to memory callback list. */
910 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
911 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
912 			 sh, mem_event_cb);
913 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
914 	/* Add context to the global device list. */
915 	LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
916 exit:
917 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
918 	return sh;
919 error:
920 	pthread_mutex_destroy(&sh->txpp.mutex);
921 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
922 	MLX5_ASSERT(sh);
923 	if (sh->cnt_id_tbl)
924 		mlx5_l3t_destroy(sh->cnt_id_tbl);
925 	if (sh->tis)
926 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
927 	if (sh->td)
928 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
929 	if (sh->devx_rx_uar)
930 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
931 	if (sh->tx_uar)
932 		mlx5_glue->devx_free_uar(sh->tx_uar);
933 	if (sh->pd)
934 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
935 	if (sh->ctx)
936 		claim_zero(mlx5_glue->close_device(sh->ctx));
937 	mlx5_free(sh);
938 	MLX5_ASSERT(err > 0);
939 	rte_errno = err;
940 	return NULL;
941 }
942 
943 /**
944  * Free shared IB device context. Decrement counter and if zero free
945  * all allocated resources and close handles.
946  *
947  * @param[in] sh
948  *   Pointer to mlx5_dev_ctx_shared object to free
949  */
950 void
951 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
952 {
953 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
954 #ifdef RTE_LIBRTE_MLX5_DEBUG
955 	/* Check the object presence in the list. */
956 	struct mlx5_dev_ctx_shared *lctx;
957 
958 	LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
959 		if (lctx == sh)
960 			break;
961 	MLX5_ASSERT(lctx);
962 	if (lctx != sh) {
963 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
964 		goto exit;
965 	}
966 #endif
967 	MLX5_ASSERT(sh);
968 	MLX5_ASSERT(sh->refcnt);
969 	/* Secondary process should not free the shared context. */
970 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
971 	if (--sh->refcnt)
972 		goto exit;
973 	/* Remove from memory callback device list. */
974 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
975 	LIST_REMOVE(sh, mem_event_cb);
976 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
977 	/* Release created Memory Regions. */
978 	mlx5_mr_release_cache(&sh->share_cache);
979 	/* Remove context from the global device list. */
980 	LIST_REMOVE(sh, next);
981 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
982 	/*
983 	 *  Ensure there is no async event handler installed.
984 	 *  Only primary process handles async device events.
985 	 **/
986 	mlx5_flow_counters_mng_close(sh);
987 	mlx5_flow_ipool_destroy(sh);
988 	mlx5_os_dev_shared_handler_uninstall(sh);
989 	if (sh->cnt_id_tbl) {
990 		mlx5_l3t_destroy(sh->cnt_id_tbl);
991 		sh->cnt_id_tbl = NULL;
992 	}
993 	if (sh->tx_uar) {
994 		mlx5_glue->devx_free_uar(sh->tx_uar);
995 		sh->tx_uar = NULL;
996 	}
997 	if (sh->pd)
998 		claim_zero(mlx5_glue->dealloc_pd(sh->pd));
999 	if (sh->tis)
1000 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1001 	if (sh->td)
1002 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1003 	if (sh->devx_rx_uar)
1004 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1005 	if (sh->ctx)
1006 		claim_zero(mlx5_glue->close_device(sh->ctx));
1007 	pthread_mutex_destroy(&sh->txpp.mutex);
1008 	mlx5_free(sh);
1009 	return;
1010 exit:
1011 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1012 }
1013 
1014 /**
1015  * Destroy table hash list.
1016  *
1017  * @param[in] priv
1018  *   Pointer to the private device data structure.
1019  */
1020 void
1021 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1022 {
1023 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1024 
1025 	if (!sh->flow_tbls)
1026 		return;
1027 	mlx5_hlist_destroy(sh->flow_tbls);
1028 }
1029 
1030 /**
1031  * Initialize flow table hash list and create the root tables entry
1032  * for each domain.
1033  *
1034  * @param[in] priv
1035  *   Pointer to the private device data structure.
1036  *
1037  * @return
1038  *   Zero on success, positive error code otherwise.
1039  */
1040 int
1041 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1042 {
1043 	int err = 0;
1044 	/* Tables are only used in DV and DR modes. */
1045 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1046 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1047 	char s[MLX5_HLIST_NAMESIZE];
1048 
1049 	MLX5_ASSERT(sh);
1050 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1051 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1052 					  0, 0, flow_dv_tbl_create_cb, NULL,
1053 					  flow_dv_tbl_remove_cb);
1054 	if (!sh->flow_tbls) {
1055 		DRV_LOG(ERR, "flow tables with hash creation failed.");
1056 		err = ENOMEM;
1057 		return err;
1058 	}
1059 	sh->flow_tbls->ctx = sh;
1060 #ifndef HAVE_MLX5DV_DR
1061 	struct rte_flow_error error;
1062 	struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1063 
1064 	/*
1065 	 * In case we have not DR support, the zero tables should be created
1066 	 * because DV expect to see them even if they cannot be created by
1067 	 * RDMA-CORE.
1068 	 */
1069 	if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0, NULL, 0, 1, &error) ||
1070 	    !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0, NULL, 0, 1, &error) ||
1071 	    !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0, NULL, 0, 1, &error)) {
1072 		err = ENOMEM;
1073 		goto error;
1074 	}
1075 	return err;
1076 error:
1077 	mlx5_free_table_hash_list(priv);
1078 #endif /* HAVE_MLX5DV_DR */
1079 #endif
1080 	return err;
1081 }
1082 
1083 /**
1084  * Retrieve integer value from environment variable.
1085  *
1086  * @param[in] name
1087  *   Environment variable name.
1088  *
1089  * @return
1090  *   Integer value, 0 if the variable is not set.
1091  */
1092 int
1093 mlx5_getenv_int(const char *name)
1094 {
1095 	const char *val = getenv(name);
1096 
1097 	if (val == NULL)
1098 		return 0;
1099 	return atoi(val);
1100 }
1101 
1102 /**
1103  * DPDK callback to add udp tunnel port
1104  *
1105  * @param[in] dev
1106  *   A pointer to eth_dev
1107  * @param[in] udp_tunnel
1108  *   A pointer to udp tunnel
1109  *
1110  * @return
1111  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1112  */
1113 int
1114 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1115 			 struct rte_eth_udp_tunnel *udp_tunnel)
1116 {
1117 	MLX5_ASSERT(udp_tunnel != NULL);
1118 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1119 	    udp_tunnel->udp_port == 4789)
1120 		return 0;
1121 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1122 	    udp_tunnel->udp_port == 4790)
1123 		return 0;
1124 	return -ENOTSUP;
1125 }
1126 
1127 /**
1128  * Initialize process private data structure.
1129  *
1130  * @param dev
1131  *   Pointer to Ethernet device structure.
1132  *
1133  * @return
1134  *   0 on success, a negative errno value otherwise and rte_errno is set.
1135  */
1136 int
1137 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1138 {
1139 	struct mlx5_priv *priv = dev->data->dev_private;
1140 	struct mlx5_proc_priv *ppriv;
1141 	size_t ppriv_size;
1142 
1143 	/*
1144 	 * UAR register table follows the process private structure. BlueFlame
1145 	 * registers for Tx queues are stored in the table.
1146 	 */
1147 	ppriv_size =
1148 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1149 	ppriv = mlx5_malloc(MLX5_MEM_RTE, ppriv_size, RTE_CACHE_LINE_SIZE,
1150 			    dev->device->numa_node);
1151 	if (!ppriv) {
1152 		rte_errno = ENOMEM;
1153 		return -rte_errno;
1154 	}
1155 	ppriv->uar_table_sz = ppriv_size;
1156 	dev->process_private = ppriv;
1157 	return 0;
1158 }
1159 
1160 /**
1161  * Un-initialize process private data structure.
1162  *
1163  * @param dev
1164  *   Pointer to Ethernet device structure.
1165  */
1166 static void
1167 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1168 {
1169 	if (!dev->process_private)
1170 		return;
1171 	mlx5_free(dev->process_private);
1172 	dev->process_private = NULL;
1173 }
1174 
1175 /**
1176  * DPDK callback to close the device.
1177  *
1178  * Destroy all queues and objects, free memory.
1179  *
1180  * @param dev
1181  *   Pointer to Ethernet device structure.
1182  */
1183 int
1184 mlx5_dev_close(struct rte_eth_dev *dev)
1185 {
1186 	struct mlx5_priv *priv = dev->data->dev_private;
1187 	unsigned int i;
1188 	int ret;
1189 
1190 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1191 		/* Check if process_private released. */
1192 		if (!dev->process_private)
1193 			return 0;
1194 		mlx5_tx_uar_uninit_secondary(dev);
1195 		mlx5_proc_priv_uninit(dev);
1196 		rte_eth_dev_release_port(dev);
1197 		return 0;
1198 	}
1199 	if (!priv->sh)
1200 		return 0;
1201 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1202 		dev->data->port_id,
1203 		((priv->sh->ctx != NULL) ?
1204 		mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1205 	/*
1206 	 * If default mreg copy action is removed at the stop stage,
1207 	 * the search will return none and nothing will be done anymore.
1208 	 */
1209 	mlx5_flow_stop_default(dev);
1210 	mlx5_traffic_disable(dev);
1211 	/*
1212 	 * If all the flows are already flushed in the device stop stage,
1213 	 * then this will return directly without any action.
1214 	 */
1215 	mlx5_flow_list_flush(dev, &priv->flows, true);
1216 	mlx5_shared_action_flush(dev);
1217 	mlx5_flow_meter_flush(dev, NULL);
1218 	/* Prevent crashes when queues are still in use. */
1219 	dev->rx_pkt_burst = removed_rx_burst;
1220 	dev->tx_pkt_burst = removed_tx_burst;
1221 	rte_wmb();
1222 	/* Disable datapath on secondary process. */
1223 	mlx5_mp_os_req_stop_rxtx(dev);
1224 	/* Free the eCPRI flex parser resource. */
1225 	mlx5_flex_parser_ecpri_release(dev);
1226 	if (priv->rxqs != NULL) {
1227 		/* XXX race condition if mlx5_rx_burst() is still running. */
1228 		usleep(1000);
1229 		for (i = 0; (i != priv->rxqs_n); ++i)
1230 			mlx5_rxq_release(dev, i);
1231 		priv->rxqs_n = 0;
1232 		priv->rxqs = NULL;
1233 	}
1234 	if (priv->txqs != NULL) {
1235 		/* XXX race condition if mlx5_tx_burst() is still running. */
1236 		usleep(1000);
1237 		for (i = 0; (i != priv->txqs_n); ++i)
1238 			mlx5_txq_release(dev, i);
1239 		priv->txqs_n = 0;
1240 		priv->txqs = NULL;
1241 	}
1242 	mlx5_proc_priv_uninit(dev);
1243 	if (priv->drop_queue.hrxq)
1244 		mlx5_drop_action_destroy(dev);
1245 	if (priv->mreg_cp_tbl)
1246 		mlx5_hlist_destroy(priv->mreg_cp_tbl);
1247 	mlx5_mprq_free_mp(dev);
1248 	mlx5_os_free_shared_dr(priv);
1249 	if (priv->rss_conf.rss_key != NULL)
1250 		mlx5_free(priv->rss_conf.rss_key);
1251 	if (priv->reta_idx != NULL)
1252 		mlx5_free(priv->reta_idx);
1253 	if (priv->config.vf)
1254 		mlx5_os_mac_addr_flush(dev);
1255 	if (priv->nl_socket_route >= 0)
1256 		close(priv->nl_socket_route);
1257 	if (priv->nl_socket_rdma >= 0)
1258 		close(priv->nl_socket_rdma);
1259 	if (priv->vmwa_context)
1260 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1261 	ret = mlx5_hrxq_verify(dev);
1262 	if (ret)
1263 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1264 			dev->data->port_id);
1265 	ret = mlx5_ind_table_obj_verify(dev);
1266 	if (ret)
1267 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1268 			dev->data->port_id);
1269 	ret = mlx5_rxq_obj_verify(dev);
1270 	if (ret)
1271 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1272 			dev->data->port_id);
1273 	ret = mlx5_rxq_verify(dev);
1274 	if (ret)
1275 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1276 			dev->data->port_id);
1277 	ret = mlx5_txq_obj_verify(dev);
1278 	if (ret)
1279 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1280 			dev->data->port_id);
1281 	ret = mlx5_txq_verify(dev);
1282 	if (ret)
1283 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1284 			dev->data->port_id);
1285 	ret = mlx5_flow_verify(dev);
1286 	if (ret)
1287 		DRV_LOG(WARNING, "port %u some flows still remain",
1288 			dev->data->port_id);
1289 	mlx5_cache_list_destroy(&priv->hrxqs);
1290 	/*
1291 	 * Free the shared context in last turn, because the cleanup
1292 	 * routines above may use some shared fields, like
1293 	 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1294 	 * ifindex if Netlink fails.
1295 	 */
1296 	mlx5_free_shared_dev_ctx(priv->sh);
1297 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1298 		unsigned int c = 0;
1299 		uint16_t port_id;
1300 
1301 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1302 			struct mlx5_priv *opriv =
1303 				rte_eth_devices[port_id].data->dev_private;
1304 
1305 			if (!opriv ||
1306 			    opriv->domain_id != priv->domain_id ||
1307 			    &rte_eth_devices[port_id] == dev)
1308 				continue;
1309 			++c;
1310 			break;
1311 		}
1312 		if (!c)
1313 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1314 	}
1315 	memset(priv, 0, sizeof(*priv));
1316 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1317 	/*
1318 	 * Reset mac_addrs to NULL such that it is not freed as part of
1319 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1320 	 * it is freed when dev_private is freed.
1321 	 */
1322 	dev->data->mac_addrs = NULL;
1323 	return 0;
1324 }
1325 
1326 /**
1327  * Verify and store value for device argument.
1328  *
1329  * @param[in] key
1330  *   Key argument to verify.
1331  * @param[in] val
1332  *   Value associated with key.
1333  * @param opaque
1334  *   User data.
1335  *
1336  * @return
1337  *   0 on success, a negative errno value otherwise and rte_errno is set.
1338  */
1339 static int
1340 mlx5_args_check(const char *key, const char *val, void *opaque)
1341 {
1342 	struct mlx5_dev_config *config = opaque;
1343 	unsigned long mod;
1344 	signed long tmp;
1345 
1346 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1347 	if (!strcmp(MLX5_REPRESENTOR, key))
1348 		return 0;
1349 	errno = 0;
1350 	tmp = strtol(val, NULL, 0);
1351 	if (errno) {
1352 		rte_errno = errno;
1353 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1354 		return -rte_errno;
1355 	}
1356 	if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1357 		/* Negative values are acceptable for some keys only. */
1358 		rte_errno = EINVAL;
1359 		DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1360 		return -rte_errno;
1361 	}
1362 	mod = tmp >= 0 ? tmp : -tmp;
1363 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1364 		config->cqe_comp = !!tmp;
1365 	} else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1366 		config->cqe_pad = !!tmp;
1367 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1368 		config->hw_padding = !!tmp;
1369 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1370 		config->mprq.enabled = !!tmp;
1371 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1372 		config->mprq.stride_num_n = tmp;
1373 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1374 		config->mprq.stride_size_n = tmp;
1375 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1376 		config->mprq.max_memcpy_len = tmp;
1377 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1378 		config->mprq.min_rxqs_num = tmp;
1379 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1380 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1381 				 " converted to txq_inline_max", key);
1382 		config->txq_inline_max = tmp;
1383 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1384 		config->txq_inline_max = tmp;
1385 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1386 		config->txq_inline_min = tmp;
1387 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1388 		config->txq_inline_mpw = tmp;
1389 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1390 		config->txqs_inline = tmp;
1391 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1392 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1393 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1394 		config->mps = !!tmp;
1395 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1396 		if (tmp != MLX5_TXDB_CACHED &&
1397 		    tmp != MLX5_TXDB_NCACHED &&
1398 		    tmp != MLX5_TXDB_HEURISTIC) {
1399 			DRV_LOG(ERR, "invalid Tx doorbell "
1400 				     "mapping parameter");
1401 			rte_errno = EINVAL;
1402 			return -rte_errno;
1403 		}
1404 		config->dbnc = tmp;
1405 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1406 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1407 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1408 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1409 				 " converted to txq_inline_mpw", key);
1410 		config->txq_inline_mpw = tmp;
1411 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1412 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1413 	} else if (strcmp(MLX5_TX_PP, key) == 0) {
1414 		if (!mod) {
1415 			DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1416 			rte_errno = EINVAL;
1417 			return -rte_errno;
1418 		}
1419 		config->tx_pp = tmp;
1420 	} else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1421 		config->tx_skew = tmp;
1422 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1423 		config->rx_vec_en = !!tmp;
1424 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1425 		config->l3_vxlan_en = !!tmp;
1426 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1427 		config->vf_nl_en = !!tmp;
1428 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1429 		config->dv_esw_en = !!tmp;
1430 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1431 		config->dv_flow_en = !!tmp;
1432 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1433 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1434 		    tmp != MLX5_XMETA_MODE_META16 &&
1435 		    tmp != MLX5_XMETA_MODE_META32 &&
1436 		    tmp != MLX5_XMETA_MODE_MISS_INFO) {
1437 			DRV_LOG(ERR, "invalid extensive "
1438 				     "metadata parameter");
1439 			rte_errno = EINVAL;
1440 			return -rte_errno;
1441 		}
1442 		if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1443 			config->dv_xmeta_en = tmp;
1444 		else
1445 			config->dv_miss_info = 1;
1446 	} else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1447 		config->lacp_by_user = !!tmp;
1448 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1449 		config->mr_ext_memseg_en = !!tmp;
1450 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1451 		config->max_dump_files_num = tmp;
1452 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1453 		config->lro.timeout = tmp;
1454 	} else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1455 		DRV_LOG(DEBUG, "class argument is %s.", val);
1456 	} else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1457 		config->log_hp_size = tmp;
1458 	} else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1459 		if (tmp != MLX5_RCM_NONE &&
1460 		    tmp != MLX5_RCM_LIGHT &&
1461 		    tmp != MLX5_RCM_AGGR) {
1462 			DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1463 			rte_errno = EINVAL;
1464 			return -rte_errno;
1465 		}
1466 		config->reclaim_mode = tmp;
1467 	} else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1468 		config->sys_mem_en = !!tmp;
1469 	} else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1470 		config->decap_en = !!tmp;
1471 	} else {
1472 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1473 		rte_errno = EINVAL;
1474 		return -rte_errno;
1475 	}
1476 	return 0;
1477 }
1478 
1479 /**
1480  * Parse device parameters.
1481  *
1482  * @param config
1483  *   Pointer to device configuration structure.
1484  * @param devargs
1485  *   Device arguments structure.
1486  *
1487  * @return
1488  *   0 on success, a negative errno value otherwise and rte_errno is set.
1489  */
1490 int
1491 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1492 {
1493 	const char **params = (const char *[]){
1494 		MLX5_RXQ_CQE_COMP_EN,
1495 		MLX5_RXQ_CQE_PAD_EN,
1496 		MLX5_RXQ_PKT_PAD_EN,
1497 		MLX5_RX_MPRQ_EN,
1498 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1499 		MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1500 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1501 		MLX5_RXQS_MIN_MPRQ,
1502 		MLX5_TXQ_INLINE,
1503 		MLX5_TXQ_INLINE_MIN,
1504 		MLX5_TXQ_INLINE_MAX,
1505 		MLX5_TXQ_INLINE_MPW,
1506 		MLX5_TXQS_MIN_INLINE,
1507 		MLX5_TXQS_MAX_VEC,
1508 		MLX5_TXQ_MPW_EN,
1509 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1510 		MLX5_TXQ_MAX_INLINE_LEN,
1511 		MLX5_TX_DB_NC,
1512 		MLX5_TX_PP,
1513 		MLX5_TX_SKEW,
1514 		MLX5_TX_VEC_EN,
1515 		MLX5_RX_VEC_EN,
1516 		MLX5_L3_VXLAN_EN,
1517 		MLX5_VF_NL_EN,
1518 		MLX5_DV_ESW_EN,
1519 		MLX5_DV_FLOW_EN,
1520 		MLX5_DV_XMETA_EN,
1521 		MLX5_LACP_BY_USER,
1522 		MLX5_MR_EXT_MEMSEG_EN,
1523 		MLX5_REPRESENTOR,
1524 		MLX5_MAX_DUMP_FILES_NUM,
1525 		MLX5_LRO_TIMEOUT_USEC,
1526 		MLX5_CLASS_ARG_NAME,
1527 		MLX5_HP_BUF_SIZE,
1528 		MLX5_RECLAIM_MEM,
1529 		MLX5_SYS_MEM_EN,
1530 		MLX5_DECAP_EN,
1531 		NULL,
1532 	};
1533 	struct rte_kvargs *kvlist;
1534 	int ret = 0;
1535 	int i;
1536 
1537 	if (devargs == NULL)
1538 		return 0;
1539 	/* Following UGLY cast is done to pass checkpatch. */
1540 	kvlist = rte_kvargs_parse(devargs->args, params);
1541 	if (kvlist == NULL) {
1542 		rte_errno = EINVAL;
1543 		return -rte_errno;
1544 	}
1545 	/* Process parameters. */
1546 	for (i = 0; (params[i] != NULL); ++i) {
1547 		if (rte_kvargs_count(kvlist, params[i])) {
1548 			ret = rte_kvargs_process(kvlist, params[i],
1549 						 mlx5_args_check, config);
1550 			if (ret) {
1551 				rte_errno = EINVAL;
1552 				rte_kvargs_free(kvlist);
1553 				return -rte_errno;
1554 			}
1555 		}
1556 	}
1557 	rte_kvargs_free(kvlist);
1558 	return 0;
1559 }
1560 
1561 /**
1562  * Configures the minimal amount of data to inline into WQE
1563  * while sending packets.
1564  *
1565  * - the txq_inline_min has the maximal priority, if this
1566  *   key is specified in devargs
1567  * - if DevX is enabled the inline mode is queried from the
1568  *   device (HCA attributes and NIC vport context if needed).
1569  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1570  *   and none (0 bytes) for other NICs
1571  *
1572  * @param spawn
1573  *   Verbs device parameters (name, port, switch_info) to spawn.
1574  * @param config
1575  *   Device configuration parameters.
1576  */
1577 void
1578 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1579 		    struct mlx5_dev_config *config)
1580 {
1581 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1582 		/* Application defines size of inlined data explicitly. */
1583 		switch (spawn->pci_dev->id.device_id) {
1584 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1585 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1586 			if (config->txq_inline_min <
1587 				       (int)MLX5_INLINE_HSIZE_L2) {
1588 				DRV_LOG(DEBUG,
1589 					"txq_inline_mix aligned to minimal"
1590 					" ConnectX-4 required value %d",
1591 					(int)MLX5_INLINE_HSIZE_L2);
1592 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1593 			}
1594 			break;
1595 		}
1596 		goto exit;
1597 	}
1598 	if (config->hca_attr.eth_net_offloads) {
1599 		/* We have DevX enabled, inline mode queried successfully. */
1600 		switch (config->hca_attr.wqe_inline_mode) {
1601 		case MLX5_CAP_INLINE_MODE_L2:
1602 			/* outer L2 header must be inlined. */
1603 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1604 			goto exit;
1605 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1606 			/* No inline data are required by NIC. */
1607 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1608 			config->hw_vlan_insert =
1609 				config->hca_attr.wqe_vlan_insert;
1610 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1611 			goto exit;
1612 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1613 			/* inline mode is defined by NIC vport context. */
1614 			if (!config->hca_attr.eth_virt)
1615 				break;
1616 			switch (config->hca_attr.vport_inline_mode) {
1617 			case MLX5_INLINE_MODE_NONE:
1618 				config->txq_inline_min =
1619 					MLX5_INLINE_HSIZE_NONE;
1620 				goto exit;
1621 			case MLX5_INLINE_MODE_L2:
1622 				config->txq_inline_min =
1623 					MLX5_INLINE_HSIZE_L2;
1624 				goto exit;
1625 			case MLX5_INLINE_MODE_IP:
1626 				config->txq_inline_min =
1627 					MLX5_INLINE_HSIZE_L3;
1628 				goto exit;
1629 			case MLX5_INLINE_MODE_TCP_UDP:
1630 				config->txq_inline_min =
1631 					MLX5_INLINE_HSIZE_L4;
1632 				goto exit;
1633 			case MLX5_INLINE_MODE_INNER_L2:
1634 				config->txq_inline_min =
1635 					MLX5_INLINE_HSIZE_INNER_L2;
1636 				goto exit;
1637 			case MLX5_INLINE_MODE_INNER_IP:
1638 				config->txq_inline_min =
1639 					MLX5_INLINE_HSIZE_INNER_L3;
1640 				goto exit;
1641 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1642 				config->txq_inline_min =
1643 					MLX5_INLINE_HSIZE_INNER_L4;
1644 				goto exit;
1645 			}
1646 		}
1647 	}
1648 	/*
1649 	 * We get here if we are unable to deduce
1650 	 * inline data size with DevX. Try PCI ID
1651 	 * to determine old NICs.
1652 	 */
1653 	switch (spawn->pci_dev->id.device_id) {
1654 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1655 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1656 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1657 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1658 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1659 		config->hw_vlan_insert = 0;
1660 		break;
1661 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1662 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1663 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1664 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1665 		/*
1666 		 * These NICs support VLAN insertion from WQE and
1667 		 * report the wqe_vlan_insert flag. But there is the bug
1668 		 * and PFC control may be broken, so disable feature.
1669 		 */
1670 		config->hw_vlan_insert = 0;
1671 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1672 		break;
1673 	default:
1674 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1675 		break;
1676 	}
1677 exit:
1678 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1679 }
1680 
1681 /**
1682  * Configures the metadata mask fields in the shared context.
1683  *
1684  * @param [in] dev
1685  *   Pointer to Ethernet device.
1686  */
1687 void
1688 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1689 {
1690 	struct mlx5_priv *priv = dev->data->dev_private;
1691 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1692 	uint32_t meta, mark, reg_c0;
1693 
1694 	reg_c0 = ~priv->vport_meta_mask;
1695 	switch (priv->config.dv_xmeta_en) {
1696 	case MLX5_XMETA_MODE_LEGACY:
1697 		meta = UINT32_MAX;
1698 		mark = MLX5_FLOW_MARK_MASK;
1699 		break;
1700 	case MLX5_XMETA_MODE_META16:
1701 		meta = reg_c0 >> rte_bsf32(reg_c0);
1702 		mark = MLX5_FLOW_MARK_MASK;
1703 		break;
1704 	case MLX5_XMETA_MODE_META32:
1705 		meta = UINT32_MAX;
1706 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1707 		break;
1708 	default:
1709 		meta = 0;
1710 		mark = 0;
1711 		MLX5_ASSERT(false);
1712 		break;
1713 	}
1714 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1715 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1716 				 sh->dv_mark_mask, mark);
1717 	else
1718 		sh->dv_mark_mask = mark;
1719 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1720 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1721 				 sh->dv_meta_mask, meta);
1722 	else
1723 		sh->dv_meta_mask = meta;
1724 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1725 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1726 				 sh->dv_meta_mask, reg_c0);
1727 	else
1728 		sh->dv_regc0_mask = reg_c0;
1729 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1730 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1731 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1732 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1733 }
1734 
1735 int
1736 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
1737 {
1738 	static const char *const dynf_names[] = {
1739 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
1740 		RTE_MBUF_DYNFLAG_METADATA_NAME,
1741 		RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
1742 	};
1743 	unsigned int i;
1744 
1745 	if (n < RTE_DIM(dynf_names))
1746 		return -ENOMEM;
1747 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
1748 		if (names[i] == NULL)
1749 			return -EINVAL;
1750 		strcpy(names[i], dynf_names[i]);
1751 	}
1752 	return RTE_DIM(dynf_names);
1753 }
1754 
1755 /**
1756  * Comparison callback to sort device data.
1757  *
1758  * This is meant to be used with qsort().
1759  *
1760  * @param a[in]
1761  *   Pointer to pointer to first data object.
1762  * @param b[in]
1763  *   Pointer to pointer to second data object.
1764  *
1765  * @return
1766  *   0 if both objects are equal, less than 0 if the first argument is less
1767  *   than the second, greater than 0 otherwise.
1768  */
1769 int
1770 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1771 			      struct mlx5_dev_config *config)
1772 {
1773 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1774 	struct mlx5_dev_config *sh_conf = NULL;
1775 	uint16_t port_id;
1776 
1777 	MLX5_ASSERT(sh);
1778 	/* Nothing to compare for the single/first device. */
1779 	if (sh->refcnt == 1)
1780 		return 0;
1781 	/* Find the device with shared context. */
1782 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1783 		struct mlx5_priv *opriv =
1784 			rte_eth_devices[port_id].data->dev_private;
1785 
1786 		if (opriv && opriv != priv && opriv->sh == sh) {
1787 			sh_conf = &opriv->config;
1788 			break;
1789 		}
1790 	}
1791 	if (!sh_conf)
1792 		return 0;
1793 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1794 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1795 			     " for shared %s context", sh->ibdev_name);
1796 		rte_errno = EINVAL;
1797 		return rte_errno;
1798 	}
1799 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
1800 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
1801 			     " for shared %s context", sh->ibdev_name);
1802 		rte_errno = EINVAL;
1803 		return rte_errno;
1804 	}
1805 	return 0;
1806 }
1807 
1808 /**
1809  * Look for the ethernet device belonging to mlx5 driver.
1810  *
1811  * @param[in] port_id
1812  *   port_id to start looking for device.
1813  * @param[in] pci_dev
1814  *   Pointer to the hint PCI device. When device is being probed
1815  *   the its siblings (master and preceding representors might
1816  *   not have assigned driver yet (because the mlx5_os_pci_probe()
1817  *   is not completed yet, for this case match on hint PCI
1818  *   device may be used to detect sibling device.
1819  *
1820  * @return
1821  *   port_id of found device, RTE_MAX_ETHPORT if not found.
1822  */
1823 uint16_t
1824 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
1825 {
1826 	while (port_id < RTE_MAX_ETHPORTS) {
1827 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
1828 
1829 		if (dev->state != RTE_ETH_DEV_UNUSED &&
1830 		    dev->device &&
1831 		    (dev->device == &pci_dev->device ||
1832 		     (dev->device->driver &&
1833 		     dev->device->driver->name &&
1834 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
1835 			break;
1836 		port_id++;
1837 	}
1838 	if (port_id >= RTE_MAX_ETHPORTS)
1839 		return RTE_MAX_ETHPORTS;
1840 	return port_id;
1841 }
1842 
1843 /**
1844  * DPDK callback to remove a PCI device.
1845  *
1846  * This function removes all Ethernet devices belong to a given PCI device.
1847  *
1848  * @param[in] pci_dev
1849  *   Pointer to the PCI device.
1850  *
1851  * @return
1852  *   0 on success, the function cannot fail.
1853  */
1854 static int
1855 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1856 {
1857 	uint16_t port_id;
1858 	int ret = 0;
1859 
1860 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
1861 		/*
1862 		 * mlx5_dev_close() is not registered to secondary process,
1863 		 * call the close function explicitly for secondary process.
1864 		 */
1865 		if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1866 			ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
1867 		else
1868 			ret |= rte_eth_dev_close(port_id);
1869 	}
1870 	return ret == 0 ? 0 : -EIO;
1871 }
1872 
1873 static const struct rte_pci_id mlx5_pci_id_map[] = {
1874 	{
1875 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1876 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1877 	},
1878 	{
1879 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1880 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1881 	},
1882 	{
1883 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1884 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1885 	},
1886 	{
1887 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1888 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1889 	},
1890 	{
1891 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1892 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1893 	},
1894 	{
1895 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1896 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1897 	},
1898 	{
1899 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1900 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1901 	},
1902 	{
1903 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1904 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1905 	},
1906 	{
1907 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1908 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
1909 	},
1910 	{
1911 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1912 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
1913 	},
1914 	{
1915 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1916 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
1917 	},
1918 	{
1919 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1920 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
1921 	},
1922 	{
1923 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1924 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
1925 	},
1926 	{
1927 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1928 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
1929 	},
1930 	{
1931 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1932 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
1933 	},
1934 	{
1935 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1936 				PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
1937 	},
1938 	{
1939 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1940 				PCI_DEVICE_ID_MELLANOX_CONNECTX7)
1941 	},
1942 	{
1943 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1944 				PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
1945 	},
1946 	{
1947 		.vendor_id = 0
1948 	}
1949 };
1950 
1951 static struct mlx5_pci_driver mlx5_driver = {
1952 	.driver_class = MLX5_CLASS_NET,
1953 	.pci_driver = {
1954 		.driver = {
1955 			.name = MLX5_DRIVER_NAME,
1956 		},
1957 		.id_table = mlx5_pci_id_map,
1958 		.probe = mlx5_os_pci_probe,
1959 		.remove = mlx5_pci_remove,
1960 		.dma_map = mlx5_dma_map,
1961 		.dma_unmap = mlx5_dma_unmap,
1962 		.drv_flags = PCI_DRV_FLAGS,
1963 	},
1964 };
1965 
1966 /* Initialize driver log type. */
1967 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
1968 
1969 /**
1970  * Driver initialization routine.
1971  */
1972 RTE_INIT(rte_mlx5_pmd_init)
1973 {
1974 	mlx5_common_init();
1975 	/* Build the static tables for Verbs conversion. */
1976 	mlx5_set_ptype_table();
1977 	mlx5_set_cksum_table();
1978 	mlx5_set_swp_types_table();
1979 	if (mlx5_glue)
1980 		mlx5_pci_driver_register(&mlx5_driver);
1981 }
1982 
1983 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1984 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1985 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1986