net/mlx5: split multi-thread flow handling per OS
[dpdk.git] / drivers / net / mlx5 / mlx5.c
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 #include <rte_cycles.h>
25
26 #include <mlx5_glue.h>
27 #include <mlx5_devx_cmds.h>
28 #include <mlx5_common.h>
29 #include <mlx5_common_os.h>
30 #include <mlx5_common_mp.h>
31 #include <mlx5_common_pci.h>
32 #include <mlx5_malloc.h>
33
34 #include "mlx5_defs.h"
35 #include "mlx5.h"
36 #include "mlx5_utils.h"
37 #include "mlx5_rxtx.h"
38 #include "mlx5_autoconf.h"
39 #include "mlx5_mr.h"
40 #include "mlx5_flow.h"
41 #include "mlx5_flow_os.h"
42 #include "rte_pmd_mlx5.h"
43
44 /* Device parameter to enable RX completion queue compression. */
45 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
46
47 /* Device parameter to enable RX completion entry padding to 128B. */
48 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
49
50 /* Device parameter to enable padding Rx packet to cacheline size. */
51 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
52
53 /* Device parameter to enable Multi-Packet Rx queue. */
54 #define MLX5_RX_MPRQ_EN "mprq_en"
55
56 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
57 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
58
59 /* Device parameter to configure log 2 of the stride size for MPRQ. */
60 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
61
62 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
63 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
64
65 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
66 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
67
68 /* Device parameter to configure inline send. Deprecated, ignored.*/
69 #define MLX5_TXQ_INLINE "txq_inline"
70
71 /* Device parameter to limit packet size to inline with ordinary SEND. */
72 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
73
74 /* Device parameter to configure minimal data size to inline. */
75 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
76
77 /* Device parameter to limit packet size to inline with Enhanced MPW. */
78 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
79
80 /*
81  * Device parameter to configure the number of TX queues threshold for
82  * enabling inline send.
83  */
84 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
85
86 /*
87  * Device parameter to configure the number of TX queues threshold for
88  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
89  */
90 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
91
92 /* Device parameter to enable multi-packet send WQEs. */
93 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
94
95 /*
96  * Device parameter to force doorbell register mapping
97  * to non-cahed region eliminating the extra write memory barrier.
98  */
99 #define MLX5_TX_DB_NC "tx_db_nc"
100
101 /*
102  * Device parameter to include 2 dsegs in the title WQEBB.
103  * Deprecated, ignored.
104  */
105 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
106
107 /*
108  * Device parameter to limit the size of inlining packet.
109  * Deprecated, ignored.
110  */
111 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
112
113 /*
114  * Device parameter to enable Tx scheduling on timestamps
115  * and specify the packet pacing granularity in nanoseconds.
116  */
117 #define MLX5_TX_PP "tx_pp"
118
119 /*
120  * Device parameter to specify skew in nanoseconds on Tx datapath,
121  * it represents the time between SQ start WQE processing and
122  * appearing actual packet data on the wire.
123  */
124 #define MLX5_TX_SKEW "tx_skew"
125
126 /*
127  * Device parameter to enable hardware Tx vector.
128  * Deprecated, ignored (no vectorized Tx routines anymore).
129  */
130 #define MLX5_TX_VEC_EN "tx_vec_en"
131
132 /* Device parameter to enable hardware Rx vector. */
133 #define MLX5_RX_VEC_EN "rx_vec_en"
134
135 /* Allow L3 VXLAN flow creation. */
136 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
137
138 /* Activate DV E-Switch flow steering. */
139 #define MLX5_DV_ESW_EN "dv_esw_en"
140
141 /* Activate DV flow steering. */
142 #define MLX5_DV_FLOW_EN "dv_flow_en"
143
144 /* Enable extensive flow metadata support. */
145 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
146
147 /* Device parameter to let the user manage the lacp traffic of bonded device */
148 #define MLX5_LACP_BY_USER "lacp_by_user"
149
150 /* Activate Netlink support in VF mode. */
151 #define MLX5_VF_NL_EN "vf_nl_en"
152
153 /* Enable extending memsegs when creating a MR. */
154 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
155
156 /* Select port representors to instantiate. */
157 #define MLX5_REPRESENTOR "representor"
158
159 /* Device parameter to configure the maximum number of dump files per queue. */
160 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
161
162 /* Configure timeout of LRO session (in microseconds). */
163 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
164
165 /*
166  * Device parameter to configure the total data buffer size for a single
167  * hairpin queue (logarithm value).
168  */
169 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
170
171 /* Flow memory reclaim mode. */
172 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
173
174 /* The default memory allocator used in PMD. */
175 #define MLX5_SYS_MEM_EN "sys_mem_en"
176 /* Decap will be used or not. */
177 #define MLX5_DECAP_EN "decap_en"
178
179 /* Shared memory between primary and secondary processes. */
180 struct mlx5_shared_data *mlx5_shared_data;
181
182 /** Driver-specific log messages type. */
183 int mlx5_logtype;
184
185 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
186                                                 LIST_HEAD_INITIALIZER();
187 static pthread_mutex_t mlx5_dev_ctx_list_mutex;
188 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
189 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
190         [MLX5_IPOOL_DECAP_ENCAP] = {
191                 .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
192                 .trunk_size = 64,
193                 .grow_trunk = 3,
194                 .grow_shift = 2,
195                 .need_lock = 1,
196                 .release_mem_en = 1,
197                 .malloc = mlx5_malloc,
198                 .free = mlx5_free,
199                 .type = "mlx5_encap_decap_ipool",
200         },
201         [MLX5_IPOOL_PUSH_VLAN] = {
202                 .size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
203                 .trunk_size = 64,
204                 .grow_trunk = 3,
205                 .grow_shift = 2,
206                 .need_lock = 1,
207                 .release_mem_en = 1,
208                 .malloc = mlx5_malloc,
209                 .free = mlx5_free,
210                 .type = "mlx5_push_vlan_ipool",
211         },
212         [MLX5_IPOOL_TAG] = {
213                 .size = sizeof(struct mlx5_flow_dv_tag_resource),
214                 .trunk_size = 64,
215                 .grow_trunk = 3,
216                 .grow_shift = 2,
217                 .need_lock = 1,
218                 .release_mem_en = 1,
219                 .malloc = mlx5_malloc,
220                 .free = mlx5_free,
221                 .type = "mlx5_tag_ipool",
222         },
223         [MLX5_IPOOL_PORT_ID] = {
224                 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
225                 .trunk_size = 64,
226                 .grow_trunk = 3,
227                 .grow_shift = 2,
228                 .need_lock = 1,
229                 .release_mem_en = 1,
230                 .malloc = mlx5_malloc,
231                 .free = mlx5_free,
232                 .type = "mlx5_port_id_ipool",
233         },
234         [MLX5_IPOOL_JUMP] = {
235                 .size = sizeof(struct mlx5_flow_tbl_data_entry),
236                 .trunk_size = 64,
237                 .grow_trunk = 3,
238                 .grow_shift = 2,
239                 .need_lock = 1,
240                 .release_mem_en = 1,
241                 .malloc = mlx5_malloc,
242                 .free = mlx5_free,
243                 .type = "mlx5_jump_ipool",
244         },
245         [MLX5_IPOOL_SAMPLE] = {
246                 .size = sizeof(struct mlx5_flow_dv_sample_resource),
247                 .trunk_size = 64,
248                 .grow_trunk = 3,
249                 .grow_shift = 2,
250                 .need_lock = 1,
251                 .release_mem_en = 1,
252                 .malloc = mlx5_malloc,
253                 .free = mlx5_free,
254                 .type = "mlx5_sample_ipool",
255         },
256         [MLX5_IPOOL_DEST_ARRAY] = {
257                 .size = sizeof(struct mlx5_flow_dv_dest_array_resource),
258                 .trunk_size = 64,
259                 .grow_trunk = 3,
260                 .grow_shift = 2,
261                 .need_lock = 1,
262                 .release_mem_en = 1,
263                 .malloc = mlx5_malloc,
264                 .free = mlx5_free,
265                 .type = "mlx5_dest_array_ipool",
266         },
267         [MLX5_IPOOL_TUNNEL_ID] = {
268                 .size = sizeof(struct mlx5_flow_tunnel),
269                 .trunk_size = MLX5_MAX_TUNNELS,
270                 .need_lock = 1,
271                 .release_mem_en = 1,
272                 .type = "mlx5_tunnel_offload",
273         },
274         [MLX5_IPOOL_TNL_TBL_ID] = {
275                 .size = 0,
276                 .need_lock = 1,
277                 .type = "mlx5_flow_tnl_tbl_ipool",
278         },
279 #endif
280         [MLX5_IPOOL_MTR] = {
281                 .size = sizeof(struct mlx5_flow_meter),
282                 .trunk_size = 64,
283                 .grow_trunk = 3,
284                 .grow_shift = 2,
285                 .need_lock = 1,
286                 .release_mem_en = 1,
287                 .malloc = mlx5_malloc,
288                 .free = mlx5_free,
289                 .type = "mlx5_meter_ipool",
290         },
291         [MLX5_IPOOL_MCP] = {
292                 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
293                 .trunk_size = 64,
294                 .grow_trunk = 3,
295                 .grow_shift = 2,
296                 .need_lock = 1,
297                 .release_mem_en = 1,
298                 .malloc = mlx5_malloc,
299                 .free = mlx5_free,
300                 .type = "mlx5_mcp_ipool",
301         },
302         [MLX5_IPOOL_HRXQ] = {
303                 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
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_hrxq_ipool",
312         },
313         [MLX5_IPOOL_MLX5_FLOW] = {
314                 /*
315                  * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
316                  * It set in run time according to PCI function configuration.
317                  */
318                 .size = 0,
319                 .trunk_size = 64,
320                 .grow_trunk = 3,
321                 .grow_shift = 2,
322                 .need_lock = 1,
323                 .release_mem_en = 1,
324                 .malloc = mlx5_malloc,
325                 .free = mlx5_free,
326                 .type = "mlx5_flow_handle_ipool",
327         },
328         [MLX5_IPOOL_RTE_FLOW] = {
329                 .size = sizeof(struct rte_flow),
330                 .trunk_size = 4096,
331                 .need_lock = 1,
332                 .release_mem_en = 1,
333                 .malloc = mlx5_malloc,
334                 .free = mlx5_free,
335                 .type = "rte_flow_ipool",
336         },
337         [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
338                 .size = 0,
339                 .need_lock = 1,
340                 .type = "mlx5_flow_rss_id_ipool",
341         },
342         [MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
343                 .size = sizeof(struct mlx5_shared_action_rss),
344                 .trunk_size = 64,
345                 .grow_trunk = 3,
346                 .grow_shift = 2,
347                 .need_lock = 1,
348                 .release_mem_en = 1,
349                 .malloc = mlx5_malloc,
350                 .free = mlx5_free,
351                 .type = "mlx5_shared_action_rss",
352         },
353 };
354
355
356 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
357 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
358
359 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
360
361 /**
362  * Initialize the ASO aging management structure.
363  *
364  * @param[in] sh
365  *   Pointer to mlx5_dev_ctx_shared object to free
366  *
367  * @return
368  *   0 on success, a negative errno value otherwise and rte_errno is set.
369  */
370 int
371 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
372 {
373         int err;
374
375         if (sh->aso_age_mng)
376                 return 0;
377         sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
378                                       RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
379         if (!sh->aso_age_mng) {
380                 DRV_LOG(ERR, "aso_age_mng allocation was failed.");
381                 rte_errno = ENOMEM;
382                 return -ENOMEM;
383         }
384         err = mlx5_aso_queue_init(sh);
385         if (err) {
386                 mlx5_free(sh->aso_age_mng);
387                 return -1;
388         }
389         rte_spinlock_init(&sh->aso_age_mng->resize_sl);
390         rte_spinlock_init(&sh->aso_age_mng->free_sl);
391         LIST_INIT(&sh->aso_age_mng->free);
392         return 0;
393 }
394
395 /**
396  * Close and release all the resources of the ASO aging management structure.
397  *
398  * @param[in] sh
399  *   Pointer to mlx5_dev_ctx_shared object to free.
400  */
401 static void
402 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
403 {
404         int i, j;
405
406         mlx5_aso_queue_stop(sh);
407         mlx5_aso_queue_uninit(sh);
408         if (sh->aso_age_mng->pools) {
409                 struct mlx5_aso_age_pool *pool;
410
411                 for (i = 0; i < sh->aso_age_mng->next; ++i) {
412                         pool = sh->aso_age_mng->pools[i];
413                         claim_zero(mlx5_devx_cmd_destroy
414                                                 (pool->flow_hit_aso_obj));
415                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
416                                 if (pool->actions[j].dr_action)
417                                         claim_zero
418                                             (mlx5_flow_os_destroy_flow_action
419                                               (pool->actions[j].dr_action));
420                         mlx5_free(pool);
421                 }
422                 mlx5_free(sh->aso_age_mng->pools);
423         }
424         mlx5_free(sh->aso_age_mng);
425 }
426
427 /**
428  * Initialize the shared aging list information per port.
429  *
430  * @param[in] sh
431  *   Pointer to mlx5_dev_ctx_shared object.
432  */
433 static void
434 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
435 {
436         uint32_t i;
437         struct mlx5_age_info *age_info;
438
439         for (i = 0; i < sh->max_port; i++) {
440                 age_info = &sh->port[i].age_info;
441                 age_info->flags = 0;
442                 TAILQ_INIT(&age_info->aged_counters);
443                 LIST_INIT(&age_info->aged_aso);
444                 rte_spinlock_init(&age_info->aged_sl);
445                 MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
446         }
447 }
448
449 /**
450  * Initialize the counters management structure.
451  *
452  * @param[in] sh
453  *   Pointer to mlx5_dev_ctx_shared object to free
454  */
455 static void
456 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
457 {
458         int i;
459
460         memset(&sh->cmng, 0, sizeof(sh->cmng));
461         TAILQ_INIT(&sh->cmng.flow_counters);
462         sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
463         sh->cmng.max_id = -1;
464         sh->cmng.last_pool_idx = POOL_IDX_INVALID;
465         rte_spinlock_init(&sh->cmng.pool_update_sl);
466         for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
467                 TAILQ_INIT(&sh->cmng.counters[i]);
468                 rte_spinlock_init(&sh->cmng.csl[i]);
469         }
470 }
471
472 /**
473  * Destroy all the resources allocated for a counter memory management.
474  *
475  * @param[in] mng
476  *   Pointer to the memory management structure.
477  */
478 static void
479 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
480 {
481         uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
482
483         LIST_REMOVE(mng, next);
484         claim_zero(mlx5_devx_cmd_destroy(mng->dm));
485         claim_zero(mlx5_os_umem_dereg(mng->umem));
486         mlx5_free(mem);
487 }
488
489 /**
490  * Close and release all the resources of the counters management.
491  *
492  * @param[in] sh
493  *   Pointer to mlx5_dev_ctx_shared object to free.
494  */
495 static void
496 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
497 {
498         struct mlx5_counter_stats_mem_mng *mng;
499         int i, j;
500         int retries = 1024;
501
502         rte_errno = 0;
503         while (--retries) {
504                 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
505                 if (rte_errno != EINPROGRESS)
506                         break;
507                 rte_pause();
508         }
509
510         if (sh->cmng.pools) {
511                 struct mlx5_flow_counter_pool *pool;
512                 uint16_t n_valid = sh->cmng.n_valid;
513                 bool fallback = sh->cmng.counter_fallback;
514
515                 for (i = 0; i < n_valid; ++i) {
516                         pool = sh->cmng.pools[i];
517                         if (!fallback && pool->min_dcs)
518                                 claim_zero(mlx5_devx_cmd_destroy
519                                                                (pool->min_dcs));
520                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
521                                 struct mlx5_flow_counter *cnt =
522                                                 MLX5_POOL_GET_CNT(pool, j);
523
524                                 if (cnt->action)
525                                         claim_zero
526                                          (mlx5_flow_os_destroy_flow_action
527                                           (cnt->action));
528                                 if (fallback && MLX5_POOL_GET_CNT
529                                     (pool, j)->dcs_when_free)
530                                         claim_zero(mlx5_devx_cmd_destroy
531                                                    (cnt->dcs_when_free));
532                         }
533                         mlx5_free(pool);
534                 }
535                 mlx5_free(sh->cmng.pools);
536         }
537         mng = LIST_FIRST(&sh->cmng.mem_mngs);
538         while (mng) {
539                 mlx5_flow_destroy_counter_stat_mem_mng(mng);
540                 mng = LIST_FIRST(&sh->cmng.mem_mngs);
541         }
542         memset(&sh->cmng, 0, sizeof(sh->cmng));
543 }
544
545 /* Send FLOW_AGED event if needed. */
546 void
547 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
548 {
549         struct mlx5_age_info *age_info;
550         uint32_t i;
551
552         for (i = 0; i < sh->max_port; i++) {
553                 age_info = &sh->port[i].age_info;
554                 if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
555                         continue;
556                 if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER))
557                         rte_eth_dev_callback_process
558                                 (&rte_eth_devices[sh->port[i].devx_ih_port_id],
559                                 RTE_ETH_EVENT_FLOW_AGED, NULL);
560                 age_info->flags = 0;
561         }
562 }
563
564 /**
565  * Initialize the flow resources' indexed mempool.
566  *
567  * @param[in] sh
568  *   Pointer to mlx5_dev_ctx_shared object.
569  * @param[in] sh
570  *   Pointer to user dev config.
571  */
572 static void
573 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
574                        const struct mlx5_dev_config *config)
575 {
576         uint8_t i;
577         struct mlx5_indexed_pool_config cfg;
578
579         for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
580                 cfg = mlx5_ipool_cfg[i];
581                 switch (i) {
582                 default:
583                         break;
584                 /*
585                  * Set MLX5_IPOOL_MLX5_FLOW ipool size
586                  * according to PCI function flow configuration.
587                  */
588                 case MLX5_IPOOL_MLX5_FLOW:
589                         cfg.size = config->dv_flow_en ?
590                                 sizeof(struct mlx5_flow_handle) :
591                                 MLX5_FLOW_HANDLE_VERBS_SIZE;
592                         break;
593                 }
594                 if (config->reclaim_mode)
595                         cfg.release_mem_en = 1;
596                 sh->ipool[i] = mlx5_ipool_create(&cfg);
597         }
598 }
599
600 /**
601  * Release the flow resources' indexed mempool.
602  *
603  * @param[in] sh
604  *   Pointer to mlx5_dev_ctx_shared object.
605  */
606 static void
607 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
608 {
609         uint8_t i;
610
611         for (i = 0; i < MLX5_IPOOL_MAX; ++i)
612                 mlx5_ipool_destroy(sh->ipool[i]);
613 }
614
615 /*
616  * Check if dynamic flex parser for eCPRI already exists.
617  *
618  * @param dev
619  *   Pointer to Ethernet device structure.
620  *
621  * @return
622  *   true on exists, false on not.
623  */
624 bool
625 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
626 {
627         struct mlx5_priv *priv = dev->data->dev_private;
628         struct mlx5_flex_parser_profiles *prf =
629                                 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
630
631         return !!prf->obj;
632 }
633
634 /*
635  * Allocation of a flex parser for eCPRI. Once created, this parser related
636  * resources will be held until the device is closed.
637  *
638  * @param dev
639  *   Pointer to Ethernet device structure.
640  *
641  * @return
642  *   0 on success, a negative errno value otherwise and rte_errno is set.
643  */
644 int
645 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
646 {
647         struct mlx5_priv *priv = dev->data->dev_private;
648         struct mlx5_flex_parser_profiles *prf =
649                                 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
650         struct mlx5_devx_graph_node_attr node = {
651                 .modify_field_select = 0,
652         };
653         uint32_t ids[8];
654         int ret;
655
656         if (!priv->config.hca_attr.parse_graph_flex_node) {
657                 DRV_LOG(ERR, "Dynamic flex parser is not supported "
658                         "for device %s.", priv->dev_data->name);
659                 return -ENOTSUP;
660         }
661         node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
662         /* 8 bytes now: 4B common header + 4B message body header. */
663         node.header_length_base_value = 0x8;
664         /* After MAC layer: Ether / VLAN. */
665         node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
666         /* Type of compared condition should be 0xAEFE in the L2 layer. */
667         node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
668         /* Sample #0: type in common header. */
669         node.sample[0].flow_match_sample_en = 1;
670         /* Fixed offset. */
671         node.sample[0].flow_match_sample_offset_mode = 0x0;
672         /* Only the 2nd byte will be used. */
673         node.sample[0].flow_match_sample_field_base_offset = 0x0;
674         /* Sample #1: message payload. */
675         node.sample[1].flow_match_sample_en = 1;
676         /* Fixed offset. */
677         node.sample[1].flow_match_sample_offset_mode = 0x0;
678         /*
679          * Only the first two bytes will be used right now, and its offset will
680          * start after the common header that with the length of a DW(u32).
681          */
682         node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
683         prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
684         if (!prf->obj) {
685                 DRV_LOG(ERR, "Failed to create flex parser node object.");
686                 return (rte_errno == 0) ? -ENODEV : -rte_errno;
687         }
688         prf->num = 2;
689         ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
690         if (ret) {
691                 DRV_LOG(ERR, "Failed to query sample IDs.");
692                 return (rte_errno == 0) ? -ENODEV : -rte_errno;
693         }
694         prf->offset[0] = 0x0;
695         prf->offset[1] = sizeof(uint32_t);
696         prf->ids[0] = ids[0];
697         prf->ids[1] = ids[1];
698         return 0;
699 }
700
701 /*
702  * Destroy the flex parser node, including the parser itself, input / output
703  * arcs and DW samples. Resources could be reused then.
704  *
705  * @param dev
706  *   Pointer to Ethernet device structure.
707  */
708 static void
709 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
710 {
711         struct mlx5_priv *priv = dev->data->dev_private;
712         struct mlx5_flex_parser_profiles *prf =
713                                 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
714
715         if (prf->obj)
716                 mlx5_devx_cmd_destroy(prf->obj);
717         prf->obj = NULL;
718 }
719
720 /*
721  * Allocate Rx and Tx UARs in robust fashion.
722  * This routine handles the following UAR allocation issues:
723  *
724  *  - tries to allocate the UAR with the most appropriate memory
725  *    mapping type from the ones supported by the host
726  *
727  *  - tries to allocate the UAR with non-NULL base address
728  *    OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
729  *    UAR base address if UAR was not the first object in the UAR page.
730  *    It caused the PMD failure and we should try to get another UAR
731  *    till we get the first one with non-NULL base address returned.
732  */
733 static int
734 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
735                      const struct mlx5_dev_config *config)
736 {
737         uint32_t uar_mapping, retry;
738         int err = 0;
739         void *base_addr;
740
741         for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
742 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
743                 /* Control the mapping type according to the settings. */
744                 uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
745                               MLX5DV_UAR_ALLOC_TYPE_NC :
746                               MLX5DV_UAR_ALLOC_TYPE_BF;
747 #else
748                 RTE_SET_USED(config);
749                 /*
750                  * It seems we have no way to control the memory mapping type
751                  * for the UAR, the default "Write-Combining" type is supposed.
752                  * The UAR initialization on queue creation queries the
753                  * actual mapping type done by Verbs/kernel and setups the
754                  * PMD datapath accordingly.
755                  */
756                 uar_mapping = 0;
757 #endif
758                 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
759 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
760                 if (!sh->tx_uar &&
761                     uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
762                         if (config->dbnc == MLX5_TXDB_CACHED ||
763                             config->dbnc == MLX5_TXDB_HEURISTIC)
764                                 DRV_LOG(WARNING, "Devarg tx_db_nc setting "
765                                                  "is not supported by DevX");
766                         /*
767                          * In some environments like virtual machine
768                          * the Write Combining mapped might be not supported
769                          * and UAR allocation fails. We try "Non-Cached"
770                          * mapping for the case. The tx_burst routines take
771                          * the UAR mapping type into account on UAR setup
772                          * on queue creation.
773                          */
774                         DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (BF)");
775                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
776                         sh->tx_uar = mlx5_glue->devx_alloc_uar
777                                                         (sh->ctx, uar_mapping);
778                 } else if (!sh->tx_uar &&
779                            uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
780                         if (config->dbnc == MLX5_TXDB_NCACHED)
781                                 DRV_LOG(WARNING, "Devarg tx_db_nc settings "
782                                                  "is not supported by DevX");
783                         /*
784                          * If Verbs/kernel does not support "Non-Cached"
785                          * try the "Write-Combining".
786                          */
787                         DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (NC)");
788                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
789                         sh->tx_uar = mlx5_glue->devx_alloc_uar
790                                                         (sh->ctx, uar_mapping);
791                 }
792 #endif
793                 if (!sh->tx_uar) {
794                         DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
795                         err = ENOMEM;
796                         goto exit;
797                 }
798                 base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
799                 if (base_addr)
800                         break;
801                 /*
802                  * The UARs are allocated by rdma_core within the
803                  * IB device context, on context closure all UARs
804                  * will be freed, should be no memory/object leakage.
805                  */
806                 DRV_LOG(WARNING, "Retrying to allocate Tx DevX UAR");
807                 sh->tx_uar = NULL;
808         }
809         /* Check whether we finally succeeded with valid UAR allocation. */
810         if (!sh->tx_uar) {
811                 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
812                 err = ENOMEM;
813                 goto exit;
814         }
815         for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
816                 uar_mapping = 0;
817                 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
818                                                         (sh->ctx, uar_mapping);
819 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
820                 if (!sh->devx_rx_uar &&
821                     uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
822                         /*
823                          * Rx UAR is used to control interrupts only,
824                          * should be no datapath noticeable impact,
825                          * can try "Non-Cached" mapping safely.
826                          */
827                         DRV_LOG(WARNING, "Failed to allocate Rx DevX UAR (BF)");
828                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
829                         sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
830                                                         (sh->ctx, uar_mapping);
831                 }
832 #endif
833                 if (!sh->devx_rx_uar) {
834                         DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
835                         err = ENOMEM;
836                         goto exit;
837                 }
838                 base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
839                 if (base_addr)
840                         break;
841                 /*
842                  * The UARs are allocated by rdma_core within the
843                  * IB device context, on context closure all UARs
844                  * will be freed, should be no memory/object leakage.
845                  */
846                 DRV_LOG(WARNING, "Retrying to allocate Rx DevX UAR");
847                 sh->devx_rx_uar = NULL;
848         }
849         /* Check whether we finally succeeded with valid UAR allocation. */
850         if (!sh->devx_rx_uar) {
851                 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
852                 err = ENOMEM;
853         }
854 exit:
855         return err;
856 }
857
858 /**
859  * Allocate shared device context. If there is multiport device the
860  * master and representors will share this context, if there is single
861  * port dedicated device, the context will be used by only given
862  * port due to unification.
863  *
864  * Routine first searches the context for the specified device name,
865  * if found the shared context assumed and reference counter is incremented.
866  * If no context found the new one is created and initialized with specified
867  * device context and parameters.
868  *
869  * @param[in] spawn
870  *   Pointer to the device attributes (name, port, etc).
871  * @param[in] config
872  *   Pointer to device configuration structure.
873  *
874  * @return
875  *   Pointer to mlx5_dev_ctx_shared object on success,
876  *   otherwise NULL and rte_errno is set.
877  */
878 struct mlx5_dev_ctx_shared *
879 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
880                            const struct mlx5_dev_config *config)
881 {
882         struct mlx5_dev_ctx_shared *sh;
883         int err = 0;
884         uint32_t i;
885         struct mlx5_devx_tis_attr tis_attr = { 0 };
886
887         MLX5_ASSERT(spawn);
888         /* Secondary process should not create the shared context. */
889         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
890         pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
891         /* Search for IB context by device name. */
892         LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
893                 if (!strcmp(sh->ibdev_name,
894                         mlx5_os_get_dev_device_name(spawn->phys_dev))) {
895                         sh->refcnt++;
896                         goto exit;
897                 }
898         }
899         /* No device found, we have to create new shared context. */
900         MLX5_ASSERT(spawn->max_port);
901         sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
902                          sizeof(struct mlx5_dev_ctx_shared) +
903                          spawn->max_port *
904                          sizeof(struct mlx5_dev_shared_port),
905                          RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
906         if (!sh) {
907                 DRV_LOG(ERR, "shared context allocation failure");
908                 rte_errno  = ENOMEM;
909                 goto exit;
910         }
911         err = mlx5_os_open_device(spawn, config, sh);
912         if (!sh->ctx)
913                 goto error;
914         err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
915         if (err) {
916                 DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
917                 goto error;
918         }
919         sh->refcnt = 1;
920         sh->bond_dev = UINT16_MAX;
921         sh->max_port = spawn->max_port;
922         strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
923                 sizeof(sh->ibdev_name) - 1);
924         strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
925                 sizeof(sh->ibdev_path) - 1);
926         /*
927          * Setting port_id to max unallowed value means
928          * there is no interrupt subhandler installed for
929          * the given port index i.
930          */
931         for (i = 0; i < sh->max_port; i++) {
932                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
933                 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
934         }
935         sh->pd = mlx5_os_alloc_pd(sh->ctx);
936         if (sh->pd == NULL) {
937                 DRV_LOG(ERR, "PD allocation failure");
938                 err = ENOMEM;
939                 goto error;
940         }
941         if (sh->devx) {
942                 /* Query the EQN for this core. */
943                 err = mlx5_glue->devx_query_eqn(sh->ctx, 0, &sh->eqn);
944                 if (err) {
945                         rte_errno = errno;
946                         DRV_LOG(ERR, "Failed to query event queue number %d.",
947                                 rte_errno);
948                         goto error;
949                 }
950                 err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
951                 if (err) {
952                         DRV_LOG(ERR, "Fail to extract pdn from PD");
953                         goto error;
954                 }
955                 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
956                 if (!sh->td) {
957                         DRV_LOG(ERR, "TD allocation failure");
958                         err = ENOMEM;
959                         goto error;
960                 }
961                 tis_attr.transport_domain = sh->td->id;
962                 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
963                 if (!sh->tis) {
964                         DRV_LOG(ERR, "TIS allocation failure");
965                         err = ENOMEM;
966                         goto error;
967                 }
968                 err = mlx5_alloc_rxtx_uars(sh, config);
969                 if (err)
970                         goto error;
971                 MLX5_ASSERT(sh->tx_uar);
972                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
973
974                 MLX5_ASSERT(sh->devx_rx_uar);
975                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
976         }
977 #ifndef RTE_ARCH_64
978         /* Initialize UAR access locks for 32bit implementations. */
979         rte_spinlock_init(&sh->uar_lock_cq);
980         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
981                 rte_spinlock_init(&sh->uar_lock[i]);
982 #endif
983         /*
984          * Once the device is added to the list of memory event
985          * callback, its global MR cache table cannot be expanded
986          * on the fly because of deadlock. If it overflows, lookup
987          * should be done by searching MR list linearly, which is slow.
988          *
989          * At this point the device is not added to the memory
990          * event list yet, context is just being created.
991          */
992         err = mlx5_mr_btree_init(&sh->share_cache.cache,
993                                  MLX5_MR_BTREE_CACHE_N * 2,
994                                  spawn->pci_dev->device.numa_node);
995         if (err) {
996                 err = rte_errno;
997                 goto error;
998         }
999         mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
1000                               &sh->share_cache.dereg_mr_cb);
1001         mlx5_os_dev_shared_handler_install(sh);
1002         sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
1003         if (!sh->cnt_id_tbl) {
1004                 err = rte_errno;
1005                 goto error;
1006         }
1007         if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1008                 err = mlx5_flow_os_init_workspace_once();
1009                 if (err)
1010                         goto error;
1011         }
1012         mlx5_flow_aging_init(sh);
1013         mlx5_flow_counters_mng_init(sh);
1014         mlx5_flow_ipool_create(sh, config);
1015         /* Add device to memory callback list. */
1016         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1017         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1018                          sh, mem_event_cb);
1019         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1020         /* Add context to the global device list. */
1021         LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1022 exit:
1023         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1024         return sh;
1025 error:
1026         pthread_mutex_destroy(&sh->txpp.mutex);
1027         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1028         MLX5_ASSERT(sh);
1029         if (sh->cnt_id_tbl)
1030                 mlx5_l3t_destroy(sh->cnt_id_tbl);
1031         if (sh->tis)
1032                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1033         if (sh->td)
1034                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1035         if (sh->devx_rx_uar)
1036                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1037         if (sh->tx_uar)
1038                 mlx5_glue->devx_free_uar(sh->tx_uar);
1039         if (sh->pd)
1040                 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1041         if (sh->ctx)
1042                 claim_zero(mlx5_glue->close_device(sh->ctx));
1043         mlx5_free(sh);
1044         MLX5_ASSERT(err > 0);
1045         rte_errno = err;
1046         return NULL;
1047 }
1048
1049 /**
1050  * Free shared IB device context. Decrement counter and if zero free
1051  * all allocated resources and close handles.
1052  *
1053  * @param[in] sh
1054  *   Pointer to mlx5_dev_ctx_shared object to free
1055  */
1056 void
1057 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1058 {
1059         pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1060 #ifdef RTE_LIBRTE_MLX5_DEBUG
1061         /* Check the object presence in the list. */
1062         struct mlx5_dev_ctx_shared *lctx;
1063
1064         LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1065                 if (lctx == sh)
1066                         break;
1067         MLX5_ASSERT(lctx);
1068         if (lctx != sh) {
1069                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
1070                 goto exit;
1071         }
1072 #endif
1073         MLX5_ASSERT(sh);
1074         MLX5_ASSERT(sh->refcnt);
1075         /* Secondary process should not free the shared context. */
1076         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1077         if (--sh->refcnt)
1078                 goto exit;
1079         /* Remove from memory callback device list. */
1080         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1081         LIST_REMOVE(sh, mem_event_cb);
1082         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1083         /* Release created Memory Regions. */
1084         mlx5_mr_release_cache(&sh->share_cache);
1085         /* Remove context from the global device list. */
1086         LIST_REMOVE(sh, next);
1087         /* Release flow workspaces objects on the last device. */
1088         if (LIST_EMPTY(&mlx5_dev_ctx_list))
1089                 mlx5_flow_os_release_workspace();
1090         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1091         /*
1092          *  Ensure there is no async event handler installed.
1093          *  Only primary process handles async device events.
1094          **/
1095         mlx5_flow_counters_mng_close(sh);
1096         if (sh->aso_age_mng) {
1097                 mlx5_flow_aso_age_mng_close(sh);
1098                 sh->aso_age_mng = NULL;
1099         }
1100         mlx5_flow_ipool_destroy(sh);
1101         mlx5_os_dev_shared_handler_uninstall(sh);
1102         if (sh->cnt_id_tbl) {
1103                 mlx5_l3t_destroy(sh->cnt_id_tbl);
1104                 sh->cnt_id_tbl = NULL;
1105         }
1106         if (sh->tx_uar) {
1107                 mlx5_glue->devx_free_uar(sh->tx_uar);
1108                 sh->tx_uar = NULL;
1109         }
1110         if (sh->pd)
1111                 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1112         if (sh->tis)
1113                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1114         if (sh->td)
1115                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1116         if (sh->devx_rx_uar)
1117                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1118         if (sh->ctx)
1119                 claim_zero(mlx5_glue->close_device(sh->ctx));
1120         pthread_mutex_destroy(&sh->txpp.mutex);
1121         mlx5_free(sh);
1122         return;
1123 exit:
1124         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1125 }
1126
1127 /**
1128  * Destroy table hash list.
1129  *
1130  * @param[in] priv
1131  *   Pointer to the private device data structure.
1132  */
1133 void
1134 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1135 {
1136         struct mlx5_dev_ctx_shared *sh = priv->sh;
1137
1138         if (!sh->flow_tbls)
1139                 return;
1140         mlx5_hlist_destroy(sh->flow_tbls);
1141 }
1142
1143 /**
1144  * Initialize flow table hash list and create the root tables entry
1145  * for each domain.
1146  *
1147  * @param[in] priv
1148  *   Pointer to the private device data structure.
1149  *
1150  * @return
1151  *   Zero on success, positive error code otherwise.
1152  */
1153 int
1154 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1155 {
1156         int err = 0;
1157         /* Tables are only used in DV and DR modes. */
1158 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1159         struct mlx5_dev_ctx_shared *sh = priv->sh;
1160         char s[MLX5_HLIST_NAMESIZE];
1161
1162         MLX5_ASSERT(sh);
1163         snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1164         sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1165                                           0, 0, flow_dv_tbl_create_cb,
1166                                           flow_dv_tbl_match_cb,
1167                                           flow_dv_tbl_remove_cb);
1168         if (!sh->flow_tbls) {
1169                 DRV_LOG(ERR, "flow tables with hash creation failed.");
1170                 err = ENOMEM;
1171                 return err;
1172         }
1173         sh->flow_tbls->ctx = sh;
1174 #ifndef HAVE_MLX5DV_DR
1175         struct rte_flow_error error;
1176         struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1177
1178         /*
1179          * In case we have not DR support, the zero tables should be created
1180          * because DV expect to see them even if they cannot be created by
1181          * RDMA-CORE.
1182          */
1183         if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0, NULL, 0, 1, &error) ||
1184             !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0, NULL, 0, 1, &error) ||
1185             !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0, NULL, 0, 1, &error)) {
1186                 err = ENOMEM;
1187                 goto error;
1188         }
1189         return err;
1190 error:
1191         mlx5_free_table_hash_list(priv);
1192 #endif /* HAVE_MLX5DV_DR */
1193 #endif
1194         return err;
1195 }
1196
1197 /**
1198  * Retrieve integer value from environment variable.
1199  *
1200  * @param[in] name
1201  *   Environment variable name.
1202  *
1203  * @return
1204  *   Integer value, 0 if the variable is not set.
1205  */
1206 int
1207 mlx5_getenv_int(const char *name)
1208 {
1209         const char *val = getenv(name);
1210
1211         if (val == NULL)
1212                 return 0;
1213         return atoi(val);
1214 }
1215
1216 /**
1217  * DPDK callback to add udp tunnel port
1218  *
1219  * @param[in] dev
1220  *   A pointer to eth_dev
1221  * @param[in] udp_tunnel
1222  *   A pointer to udp tunnel
1223  *
1224  * @return
1225  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1226  */
1227 int
1228 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1229                          struct rte_eth_udp_tunnel *udp_tunnel)
1230 {
1231         MLX5_ASSERT(udp_tunnel != NULL);
1232         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1233             udp_tunnel->udp_port == 4789)
1234                 return 0;
1235         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1236             udp_tunnel->udp_port == 4790)
1237                 return 0;
1238         return -ENOTSUP;
1239 }
1240
1241 /**
1242  * Initialize process private data structure.
1243  *
1244  * @param dev
1245  *   Pointer to Ethernet device structure.
1246  *
1247  * @return
1248  *   0 on success, a negative errno value otherwise and rte_errno is set.
1249  */
1250 int
1251 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1252 {
1253         struct mlx5_priv *priv = dev->data->dev_private;
1254         struct mlx5_proc_priv *ppriv;
1255         size_t ppriv_size;
1256
1257         /*
1258          * UAR register table follows the process private structure. BlueFlame
1259          * registers for Tx queues are stored in the table.
1260          */
1261         ppriv_size =
1262                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1263         ppriv = mlx5_malloc(MLX5_MEM_RTE, ppriv_size, RTE_CACHE_LINE_SIZE,
1264                             dev->device->numa_node);
1265         if (!ppriv) {
1266                 rte_errno = ENOMEM;
1267                 return -rte_errno;
1268         }
1269         ppriv->uar_table_sz = ppriv_size;
1270         dev->process_private = ppriv;
1271         return 0;
1272 }
1273
1274 /**
1275  * Un-initialize process private data structure.
1276  *
1277  * @param dev
1278  *   Pointer to Ethernet device structure.
1279  */
1280 static void
1281 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1282 {
1283         if (!dev->process_private)
1284                 return;
1285         mlx5_free(dev->process_private);
1286         dev->process_private = NULL;
1287 }
1288
1289 /**
1290  * DPDK callback to close the device.
1291  *
1292  * Destroy all queues and objects, free memory.
1293  *
1294  * @param dev
1295  *   Pointer to Ethernet device structure.
1296  */
1297 int
1298 mlx5_dev_close(struct rte_eth_dev *dev)
1299 {
1300         struct mlx5_priv *priv = dev->data->dev_private;
1301         unsigned int i;
1302         int ret;
1303
1304         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1305                 /* Check if process_private released. */
1306                 if (!dev->process_private)
1307                         return 0;
1308                 mlx5_tx_uar_uninit_secondary(dev);
1309                 mlx5_proc_priv_uninit(dev);
1310                 rte_eth_dev_release_port(dev);
1311                 return 0;
1312         }
1313         if (!priv->sh)
1314                 return 0;
1315         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1316                 dev->data->port_id,
1317                 ((priv->sh->ctx != NULL) ?
1318                 mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1319         /*
1320          * If default mreg copy action is removed at the stop stage,
1321          * the search will return none and nothing will be done anymore.
1322          */
1323         mlx5_flow_stop_default(dev);
1324         mlx5_traffic_disable(dev);
1325         /*
1326          * If all the flows are already flushed in the device stop stage,
1327          * then this will return directly without any action.
1328          */
1329         mlx5_flow_list_flush(dev, &priv->flows, true);
1330         mlx5_shared_action_flush(dev);
1331         mlx5_flow_meter_flush(dev, NULL);
1332         /* Prevent crashes when queues are still in use. */
1333         dev->rx_pkt_burst = removed_rx_burst;
1334         dev->tx_pkt_burst = removed_tx_burst;
1335         rte_wmb();
1336         /* Disable datapath on secondary process. */
1337         mlx5_mp_os_req_stop_rxtx(dev);
1338         /* Free the eCPRI flex parser resource. */
1339         mlx5_flex_parser_ecpri_release(dev);
1340         if (priv->rxqs != NULL) {
1341                 /* XXX race condition if mlx5_rx_burst() is still running. */
1342                 rte_delay_us_sleep(1000);
1343                 for (i = 0; (i != priv->rxqs_n); ++i)
1344                         mlx5_rxq_release(dev, i);
1345                 priv->rxqs_n = 0;
1346                 priv->rxqs = NULL;
1347         }
1348         if (priv->txqs != NULL) {
1349                 /* XXX race condition if mlx5_tx_burst() is still running. */
1350                 rte_delay_us_sleep(1000);
1351                 for (i = 0; (i != priv->txqs_n); ++i)
1352                         mlx5_txq_release(dev, i);
1353                 priv->txqs_n = 0;
1354                 priv->txqs = NULL;
1355         }
1356         mlx5_proc_priv_uninit(dev);
1357         if (priv->drop_queue.hrxq)
1358                 mlx5_drop_action_destroy(dev);
1359         if (priv->mreg_cp_tbl)
1360                 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1361         mlx5_mprq_free_mp(dev);
1362         mlx5_os_free_shared_dr(priv);
1363         if (priv->rss_conf.rss_key != NULL)
1364                 mlx5_free(priv->rss_conf.rss_key);
1365         if (priv->reta_idx != NULL)
1366                 mlx5_free(priv->reta_idx);
1367         if (priv->config.vf)
1368                 mlx5_os_mac_addr_flush(dev);
1369         if (priv->nl_socket_route >= 0)
1370                 close(priv->nl_socket_route);
1371         if (priv->nl_socket_rdma >= 0)
1372                 close(priv->nl_socket_rdma);
1373         if (priv->vmwa_context)
1374                 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1375         ret = mlx5_hrxq_verify(dev);
1376         if (ret)
1377                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1378                         dev->data->port_id);
1379         ret = mlx5_ind_table_obj_verify(dev);
1380         if (ret)
1381                 DRV_LOG(WARNING, "port %u some indirection table still remain",
1382                         dev->data->port_id);
1383         ret = mlx5_rxq_obj_verify(dev);
1384         if (ret)
1385                 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1386                         dev->data->port_id);
1387         ret = mlx5_rxq_verify(dev);
1388         if (ret)
1389                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1390                         dev->data->port_id);
1391         ret = mlx5_txq_obj_verify(dev);
1392         if (ret)
1393                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1394                         dev->data->port_id);
1395         ret = mlx5_txq_verify(dev);
1396         if (ret)
1397                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1398                         dev->data->port_id);
1399         ret = mlx5_flow_verify(dev);
1400         if (ret)
1401                 DRV_LOG(WARNING, "port %u some flows still remain",
1402                         dev->data->port_id);
1403         mlx5_cache_list_destroy(&priv->hrxqs);
1404         /*
1405          * Free the shared context in last turn, because the cleanup
1406          * routines above may use some shared fields, like
1407          * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1408          * ifindex if Netlink fails.
1409          */
1410         mlx5_free_shared_dev_ctx(priv->sh);
1411         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1412                 unsigned int c = 0;
1413                 uint16_t port_id;
1414
1415                 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1416                         struct mlx5_priv *opriv =
1417                                 rte_eth_devices[port_id].data->dev_private;
1418
1419                         if (!opriv ||
1420                             opriv->domain_id != priv->domain_id ||
1421                             &rte_eth_devices[port_id] == dev)
1422                                 continue;
1423                         ++c;
1424                         break;
1425                 }
1426                 if (!c)
1427                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1428         }
1429         memset(priv, 0, sizeof(*priv));
1430         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1431         /*
1432          * Reset mac_addrs to NULL such that it is not freed as part of
1433          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1434          * it is freed when dev_private is freed.
1435          */
1436         dev->data->mac_addrs = NULL;
1437         return 0;
1438 }
1439
1440 const struct eth_dev_ops mlx5_dev_ops = {
1441         .dev_configure = mlx5_dev_configure,
1442         .dev_start = mlx5_dev_start,
1443         .dev_stop = mlx5_dev_stop,
1444         .dev_set_link_down = mlx5_set_link_down,
1445         .dev_set_link_up = mlx5_set_link_up,
1446         .dev_close = mlx5_dev_close,
1447         .promiscuous_enable = mlx5_promiscuous_enable,
1448         .promiscuous_disable = mlx5_promiscuous_disable,
1449         .allmulticast_enable = mlx5_allmulticast_enable,
1450         .allmulticast_disable = mlx5_allmulticast_disable,
1451         .link_update = mlx5_link_update,
1452         .stats_get = mlx5_stats_get,
1453         .stats_reset = mlx5_stats_reset,
1454         .xstats_get = mlx5_xstats_get,
1455         .xstats_reset = mlx5_xstats_reset,
1456         .xstats_get_names = mlx5_xstats_get_names,
1457         .fw_version_get = mlx5_fw_version_get,
1458         .dev_infos_get = mlx5_dev_infos_get,
1459         .read_clock = mlx5_txpp_read_clock,
1460         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1461         .vlan_filter_set = mlx5_vlan_filter_set,
1462         .rx_queue_setup = mlx5_rx_queue_setup,
1463         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1464         .tx_queue_setup = mlx5_tx_queue_setup,
1465         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1466         .rx_queue_release = mlx5_rx_queue_release,
1467         .tx_queue_release = mlx5_tx_queue_release,
1468         .rx_queue_start = mlx5_rx_queue_start,
1469         .rx_queue_stop = mlx5_rx_queue_stop,
1470         .tx_queue_start = mlx5_tx_queue_start,
1471         .tx_queue_stop = mlx5_tx_queue_stop,
1472         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1473         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1474         .mac_addr_remove = mlx5_mac_addr_remove,
1475         .mac_addr_add = mlx5_mac_addr_add,
1476         .mac_addr_set = mlx5_mac_addr_set,
1477         .set_mc_addr_list = mlx5_set_mc_addr_list,
1478         .mtu_set = mlx5_dev_set_mtu,
1479         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1480         .vlan_offload_set = mlx5_vlan_offload_set,
1481         .reta_update = mlx5_dev_rss_reta_update,
1482         .reta_query = mlx5_dev_rss_reta_query,
1483         .rss_hash_update = mlx5_rss_hash_update,
1484         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1485         .filter_ctrl = mlx5_dev_filter_ctrl,
1486         .rxq_info_get = mlx5_rxq_info_get,
1487         .txq_info_get = mlx5_txq_info_get,
1488         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1489         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1490         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1491         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1492         .is_removed = mlx5_is_removed,
1493         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1494         .get_module_info = mlx5_get_module_info,
1495         .get_module_eeprom = mlx5_get_module_eeprom,
1496         .hairpin_cap_get = mlx5_hairpin_cap_get,
1497         .mtr_ops_get = mlx5_flow_meter_ops_get,
1498         .hairpin_bind = mlx5_hairpin_bind,
1499         .hairpin_unbind = mlx5_hairpin_unbind,
1500         .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1501         .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1502         .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1503         .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1504 };
1505
1506 /* Available operations from secondary process. */
1507 const struct eth_dev_ops mlx5_dev_sec_ops = {
1508         .stats_get = mlx5_stats_get,
1509         .stats_reset = mlx5_stats_reset,
1510         .xstats_get = mlx5_xstats_get,
1511         .xstats_reset = mlx5_xstats_reset,
1512         .xstats_get_names = mlx5_xstats_get_names,
1513         .fw_version_get = mlx5_fw_version_get,
1514         .dev_infos_get = mlx5_dev_infos_get,
1515         .read_clock = mlx5_txpp_read_clock,
1516         .rx_queue_start = mlx5_rx_queue_start,
1517         .rx_queue_stop = mlx5_rx_queue_stop,
1518         .tx_queue_start = mlx5_tx_queue_start,
1519         .tx_queue_stop = mlx5_tx_queue_stop,
1520         .rxq_info_get = mlx5_rxq_info_get,
1521         .txq_info_get = mlx5_txq_info_get,
1522         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1523         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1524         .get_module_info = mlx5_get_module_info,
1525         .get_module_eeprom = mlx5_get_module_eeprom,
1526 };
1527
1528 /* Available operations in flow isolated mode. */
1529 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1530         .dev_configure = mlx5_dev_configure,
1531         .dev_start = mlx5_dev_start,
1532         .dev_stop = mlx5_dev_stop,
1533         .dev_set_link_down = mlx5_set_link_down,
1534         .dev_set_link_up = mlx5_set_link_up,
1535         .dev_close = mlx5_dev_close,
1536         .promiscuous_enable = mlx5_promiscuous_enable,
1537         .promiscuous_disable = mlx5_promiscuous_disable,
1538         .allmulticast_enable = mlx5_allmulticast_enable,
1539         .allmulticast_disable = mlx5_allmulticast_disable,
1540         .link_update = mlx5_link_update,
1541         .stats_get = mlx5_stats_get,
1542         .stats_reset = mlx5_stats_reset,
1543         .xstats_get = mlx5_xstats_get,
1544         .xstats_reset = mlx5_xstats_reset,
1545         .xstats_get_names = mlx5_xstats_get_names,
1546         .fw_version_get = mlx5_fw_version_get,
1547         .dev_infos_get = mlx5_dev_infos_get,
1548         .read_clock = mlx5_txpp_read_clock,
1549         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1550         .vlan_filter_set = mlx5_vlan_filter_set,
1551         .rx_queue_setup = mlx5_rx_queue_setup,
1552         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1553         .tx_queue_setup = mlx5_tx_queue_setup,
1554         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1555         .rx_queue_release = mlx5_rx_queue_release,
1556         .tx_queue_release = mlx5_tx_queue_release,
1557         .rx_queue_start = mlx5_rx_queue_start,
1558         .rx_queue_stop = mlx5_rx_queue_stop,
1559         .tx_queue_start = mlx5_tx_queue_start,
1560         .tx_queue_stop = mlx5_tx_queue_stop,
1561         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1562         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1563         .mac_addr_remove = mlx5_mac_addr_remove,
1564         .mac_addr_add = mlx5_mac_addr_add,
1565         .mac_addr_set = mlx5_mac_addr_set,
1566         .set_mc_addr_list = mlx5_set_mc_addr_list,
1567         .mtu_set = mlx5_dev_set_mtu,
1568         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1569         .vlan_offload_set = mlx5_vlan_offload_set,
1570         .filter_ctrl = mlx5_dev_filter_ctrl,
1571         .rxq_info_get = mlx5_rxq_info_get,
1572         .txq_info_get = mlx5_txq_info_get,
1573         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1574         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1575         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1576         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1577         .is_removed = mlx5_is_removed,
1578         .get_module_info = mlx5_get_module_info,
1579         .get_module_eeprom = mlx5_get_module_eeprom,
1580         .hairpin_cap_get = mlx5_hairpin_cap_get,
1581         .mtr_ops_get = mlx5_flow_meter_ops_get,
1582         .hairpin_bind = mlx5_hairpin_bind,
1583         .hairpin_unbind = mlx5_hairpin_unbind,
1584         .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1585         .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1586         .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1587         .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1588 };
1589
1590 /**
1591  * Verify and store value for device argument.
1592  *
1593  * @param[in] key
1594  *   Key argument to verify.
1595  * @param[in] val
1596  *   Value associated with key.
1597  * @param opaque
1598  *   User data.
1599  *
1600  * @return
1601  *   0 on success, a negative errno value otherwise and rte_errno is set.
1602  */
1603 static int
1604 mlx5_args_check(const char *key, const char *val, void *opaque)
1605 {
1606         struct mlx5_dev_config *config = opaque;
1607         unsigned long mod;
1608         signed long tmp;
1609
1610         /* No-op, port representors are processed in mlx5_dev_spawn(). */
1611         if (!strcmp(MLX5_REPRESENTOR, key))
1612                 return 0;
1613         errno = 0;
1614         tmp = strtol(val, NULL, 0);
1615         if (errno) {
1616                 rte_errno = errno;
1617                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1618                 return -rte_errno;
1619         }
1620         if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1621                 /* Negative values are acceptable for some keys only. */
1622                 rte_errno = EINVAL;
1623                 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1624                 return -rte_errno;
1625         }
1626         mod = tmp >= 0 ? tmp : -tmp;
1627         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1628                 if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1629                         DRV_LOG(ERR, "invalid CQE compression "
1630                                      "format parameter");
1631                         rte_errno = EINVAL;
1632                         return -rte_errno;
1633                 }
1634                 config->cqe_comp = !!tmp;
1635                 config->cqe_comp_fmt = tmp;
1636         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1637                 config->cqe_pad = !!tmp;
1638         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1639                 config->hw_padding = !!tmp;
1640         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1641                 config->mprq.enabled = !!tmp;
1642         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1643                 config->mprq.stride_num_n = tmp;
1644         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1645                 config->mprq.stride_size_n = tmp;
1646         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1647                 config->mprq.max_memcpy_len = tmp;
1648         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1649                 config->mprq.min_rxqs_num = tmp;
1650         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1651                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1652                                  " converted to txq_inline_max", key);
1653                 config->txq_inline_max = tmp;
1654         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1655                 config->txq_inline_max = tmp;
1656         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1657                 config->txq_inline_min = tmp;
1658         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1659                 config->txq_inline_mpw = tmp;
1660         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1661                 config->txqs_inline = tmp;
1662         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1663                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1664         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1665                 config->mps = !!tmp;
1666         } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1667                 if (tmp != MLX5_TXDB_CACHED &&
1668                     tmp != MLX5_TXDB_NCACHED &&
1669                     tmp != MLX5_TXDB_HEURISTIC) {
1670                         DRV_LOG(ERR, "invalid Tx doorbell "
1671                                      "mapping parameter");
1672                         rte_errno = EINVAL;
1673                         return -rte_errno;
1674                 }
1675                 config->dbnc = tmp;
1676         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1677                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1678         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1679                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1680                                  " converted to txq_inline_mpw", key);
1681                 config->txq_inline_mpw = tmp;
1682         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1683                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1684         } else if (strcmp(MLX5_TX_PP, key) == 0) {
1685                 if (!mod) {
1686                         DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1687                         rte_errno = EINVAL;
1688                         return -rte_errno;
1689                 }
1690                 config->tx_pp = tmp;
1691         } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1692                 config->tx_skew = tmp;
1693         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1694                 config->rx_vec_en = !!tmp;
1695         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1696                 config->l3_vxlan_en = !!tmp;
1697         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1698                 config->vf_nl_en = !!tmp;
1699         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1700                 config->dv_esw_en = !!tmp;
1701         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1702                 config->dv_flow_en = !!tmp;
1703         } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1704                 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1705                     tmp != MLX5_XMETA_MODE_META16 &&
1706                     tmp != MLX5_XMETA_MODE_META32 &&
1707                     tmp != MLX5_XMETA_MODE_MISS_INFO) {
1708                         DRV_LOG(ERR, "invalid extensive "
1709                                      "metadata parameter");
1710                         rte_errno = EINVAL;
1711                         return -rte_errno;
1712                 }
1713                 if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1714                         config->dv_xmeta_en = tmp;
1715                 else
1716                         config->dv_miss_info = 1;
1717         } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1718                 config->lacp_by_user = !!tmp;
1719         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1720                 config->mr_ext_memseg_en = !!tmp;
1721         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1722                 config->max_dump_files_num = tmp;
1723         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1724                 config->lro.timeout = tmp;
1725         } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1726                 DRV_LOG(DEBUG, "class argument is %s.", val);
1727         } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1728                 config->log_hp_size = tmp;
1729         } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1730                 if (tmp != MLX5_RCM_NONE &&
1731                     tmp != MLX5_RCM_LIGHT &&
1732                     tmp != MLX5_RCM_AGGR) {
1733                         DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1734                         rte_errno = EINVAL;
1735                         return -rte_errno;
1736                 }
1737                 config->reclaim_mode = tmp;
1738         } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1739                 config->sys_mem_en = !!tmp;
1740         } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1741                 config->decap_en = !!tmp;
1742         } else {
1743                 DRV_LOG(WARNING, "%s: unknown parameter", key);
1744                 rte_errno = EINVAL;
1745                 return -rte_errno;
1746         }
1747         return 0;
1748 }
1749
1750 /**
1751  * Parse device parameters.
1752  *
1753  * @param config
1754  *   Pointer to device configuration structure.
1755  * @param devargs
1756  *   Device arguments structure.
1757  *
1758  * @return
1759  *   0 on success, a negative errno value otherwise and rte_errno is set.
1760  */
1761 int
1762 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1763 {
1764         const char **params = (const char *[]){
1765                 MLX5_RXQ_CQE_COMP_EN,
1766                 MLX5_RXQ_CQE_PAD_EN,
1767                 MLX5_RXQ_PKT_PAD_EN,
1768                 MLX5_RX_MPRQ_EN,
1769                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1770                 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1771                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1772                 MLX5_RXQS_MIN_MPRQ,
1773                 MLX5_TXQ_INLINE,
1774                 MLX5_TXQ_INLINE_MIN,
1775                 MLX5_TXQ_INLINE_MAX,
1776                 MLX5_TXQ_INLINE_MPW,
1777                 MLX5_TXQS_MIN_INLINE,
1778                 MLX5_TXQS_MAX_VEC,
1779                 MLX5_TXQ_MPW_EN,
1780                 MLX5_TXQ_MPW_HDR_DSEG_EN,
1781                 MLX5_TXQ_MAX_INLINE_LEN,
1782                 MLX5_TX_DB_NC,
1783                 MLX5_TX_PP,
1784                 MLX5_TX_SKEW,
1785                 MLX5_TX_VEC_EN,
1786                 MLX5_RX_VEC_EN,
1787                 MLX5_L3_VXLAN_EN,
1788                 MLX5_VF_NL_EN,
1789                 MLX5_DV_ESW_EN,
1790                 MLX5_DV_FLOW_EN,
1791                 MLX5_DV_XMETA_EN,
1792                 MLX5_LACP_BY_USER,
1793                 MLX5_MR_EXT_MEMSEG_EN,
1794                 MLX5_REPRESENTOR,
1795                 MLX5_MAX_DUMP_FILES_NUM,
1796                 MLX5_LRO_TIMEOUT_USEC,
1797                 MLX5_CLASS_ARG_NAME,
1798                 MLX5_HP_BUF_SIZE,
1799                 MLX5_RECLAIM_MEM,
1800                 MLX5_SYS_MEM_EN,
1801                 MLX5_DECAP_EN,
1802                 NULL,
1803         };
1804         struct rte_kvargs *kvlist;
1805         int ret = 0;
1806         int i;
1807
1808         if (devargs == NULL)
1809                 return 0;
1810         /* Following UGLY cast is done to pass checkpatch. */
1811         kvlist = rte_kvargs_parse(devargs->args, params);
1812         if (kvlist == NULL) {
1813                 rte_errno = EINVAL;
1814                 return -rte_errno;
1815         }
1816         /* Process parameters. */
1817         for (i = 0; (params[i] != NULL); ++i) {
1818                 if (rte_kvargs_count(kvlist, params[i])) {
1819                         ret = rte_kvargs_process(kvlist, params[i],
1820                                                  mlx5_args_check, config);
1821                         if (ret) {
1822                                 rte_errno = EINVAL;
1823                                 rte_kvargs_free(kvlist);
1824                                 return -rte_errno;
1825                         }
1826                 }
1827         }
1828         rte_kvargs_free(kvlist);
1829         return 0;
1830 }
1831
1832 /**
1833  * Configures the minimal amount of data to inline into WQE
1834  * while sending packets.
1835  *
1836  * - the txq_inline_min has the maximal priority, if this
1837  *   key is specified in devargs
1838  * - if DevX is enabled the inline mode is queried from the
1839  *   device (HCA attributes and NIC vport context if needed).
1840  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1841  *   and none (0 bytes) for other NICs
1842  *
1843  * @param spawn
1844  *   Verbs device parameters (name, port, switch_info) to spawn.
1845  * @param config
1846  *   Device configuration parameters.
1847  */
1848 void
1849 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1850                     struct mlx5_dev_config *config)
1851 {
1852         if (config->txq_inline_min != MLX5_ARG_UNSET) {
1853                 /* Application defines size of inlined data explicitly. */
1854                 switch (spawn->pci_dev->id.device_id) {
1855                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1856                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1857                         if (config->txq_inline_min <
1858                                        (int)MLX5_INLINE_HSIZE_L2) {
1859                                 DRV_LOG(DEBUG,
1860                                         "txq_inline_mix aligned to minimal"
1861                                         " ConnectX-4 required value %d",
1862                                         (int)MLX5_INLINE_HSIZE_L2);
1863                                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1864                         }
1865                         break;
1866                 }
1867                 goto exit;
1868         }
1869         if (config->hca_attr.eth_net_offloads) {
1870                 /* We have DevX enabled, inline mode queried successfully. */
1871                 switch (config->hca_attr.wqe_inline_mode) {
1872                 case MLX5_CAP_INLINE_MODE_L2:
1873                         /* outer L2 header must be inlined. */
1874                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1875                         goto exit;
1876                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1877                         /* No inline data are required by NIC. */
1878                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1879                         config->hw_vlan_insert =
1880                                 config->hca_attr.wqe_vlan_insert;
1881                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1882                         goto exit;
1883                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1884                         /* inline mode is defined by NIC vport context. */
1885                         if (!config->hca_attr.eth_virt)
1886                                 break;
1887                         switch (config->hca_attr.vport_inline_mode) {
1888                         case MLX5_INLINE_MODE_NONE:
1889                                 config->txq_inline_min =
1890                                         MLX5_INLINE_HSIZE_NONE;
1891                                 goto exit;
1892                         case MLX5_INLINE_MODE_L2:
1893                                 config->txq_inline_min =
1894                                         MLX5_INLINE_HSIZE_L2;
1895                                 goto exit;
1896                         case MLX5_INLINE_MODE_IP:
1897                                 config->txq_inline_min =
1898                                         MLX5_INLINE_HSIZE_L3;
1899                                 goto exit;
1900                         case MLX5_INLINE_MODE_TCP_UDP:
1901                                 config->txq_inline_min =
1902                                         MLX5_INLINE_HSIZE_L4;
1903                                 goto exit;
1904                         case MLX5_INLINE_MODE_INNER_L2:
1905                                 config->txq_inline_min =
1906                                         MLX5_INLINE_HSIZE_INNER_L2;
1907                                 goto exit;
1908                         case MLX5_INLINE_MODE_INNER_IP:
1909                                 config->txq_inline_min =
1910                                         MLX5_INLINE_HSIZE_INNER_L3;
1911                                 goto exit;
1912                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
1913                                 config->txq_inline_min =
1914                                         MLX5_INLINE_HSIZE_INNER_L4;
1915                                 goto exit;
1916                         }
1917                 }
1918         }
1919         /*
1920          * We get here if we are unable to deduce
1921          * inline data size with DevX. Try PCI ID
1922          * to determine old NICs.
1923          */
1924         switch (spawn->pci_dev->id.device_id) {
1925         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1926         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1927         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1928         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1929                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1930                 config->hw_vlan_insert = 0;
1931                 break;
1932         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1933         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1934         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1935         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1936                 /*
1937                  * These NICs support VLAN insertion from WQE and
1938                  * report the wqe_vlan_insert flag. But there is the bug
1939                  * and PFC control may be broken, so disable feature.
1940                  */
1941                 config->hw_vlan_insert = 0;
1942                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1943                 break;
1944         default:
1945                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1946                 break;
1947         }
1948 exit:
1949         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1950 }
1951
1952 /**
1953  * Configures the metadata mask fields in the shared context.
1954  *
1955  * @param [in] dev
1956  *   Pointer to Ethernet device.
1957  */
1958 void
1959 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1960 {
1961         struct mlx5_priv *priv = dev->data->dev_private;
1962         struct mlx5_dev_ctx_shared *sh = priv->sh;
1963         uint32_t meta, mark, reg_c0;
1964
1965         reg_c0 = ~priv->vport_meta_mask;
1966         switch (priv->config.dv_xmeta_en) {
1967         case MLX5_XMETA_MODE_LEGACY:
1968                 meta = UINT32_MAX;
1969                 mark = MLX5_FLOW_MARK_MASK;
1970                 break;
1971         case MLX5_XMETA_MODE_META16:
1972                 meta = reg_c0 >> rte_bsf32(reg_c0);
1973                 mark = MLX5_FLOW_MARK_MASK;
1974                 break;
1975         case MLX5_XMETA_MODE_META32:
1976                 meta = UINT32_MAX;
1977                 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1978                 break;
1979         default:
1980                 meta = 0;
1981                 mark = 0;
1982                 MLX5_ASSERT(false);
1983                 break;
1984         }
1985         if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1986                 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1987                                  sh->dv_mark_mask, mark);
1988         else
1989                 sh->dv_mark_mask = mark;
1990         if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1991                 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1992                                  sh->dv_meta_mask, meta);
1993         else
1994                 sh->dv_meta_mask = meta;
1995         if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1996                 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1997                                  sh->dv_meta_mask, reg_c0);
1998         else
1999                 sh->dv_regc0_mask = reg_c0;
2000         DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2001         DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2002         DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2003         DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2004 }
2005
2006 int
2007 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2008 {
2009         static const char *const dynf_names[] = {
2010                 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2011                 RTE_MBUF_DYNFLAG_METADATA_NAME,
2012                 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2013         };
2014         unsigned int i;
2015
2016         if (n < RTE_DIM(dynf_names))
2017                 return -ENOMEM;
2018         for (i = 0; i < RTE_DIM(dynf_names); i++) {
2019                 if (names[i] == NULL)
2020                         return -EINVAL;
2021                 strcpy(names[i], dynf_names[i]);
2022         }
2023         return RTE_DIM(dynf_names);
2024 }
2025
2026 /**
2027  * Comparison callback to sort device data.
2028  *
2029  * This is meant to be used with qsort().
2030  *
2031  * @param a[in]
2032  *   Pointer to pointer to first data object.
2033  * @param b[in]
2034  *   Pointer to pointer to second data object.
2035  *
2036  * @return
2037  *   0 if both objects are equal, less than 0 if the first argument is less
2038  *   than the second, greater than 0 otherwise.
2039  */
2040 int
2041 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2042                               struct mlx5_dev_config *config)
2043 {
2044         struct mlx5_dev_ctx_shared *sh = priv->sh;
2045         struct mlx5_dev_config *sh_conf = NULL;
2046         uint16_t port_id;
2047
2048         MLX5_ASSERT(sh);
2049         /* Nothing to compare for the single/first device. */
2050         if (sh->refcnt == 1)
2051                 return 0;
2052         /* Find the device with shared context. */
2053         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2054                 struct mlx5_priv *opriv =
2055                         rte_eth_devices[port_id].data->dev_private;
2056
2057                 if (opriv && opriv != priv && opriv->sh == sh) {
2058                         sh_conf = &opriv->config;
2059                         break;
2060                 }
2061         }
2062         if (!sh_conf)
2063                 return 0;
2064         if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2065                 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2066                              " for shared %s context", sh->ibdev_name);
2067                 rte_errno = EINVAL;
2068                 return rte_errno;
2069         }
2070         if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2071                 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2072                              " for shared %s context", sh->ibdev_name);
2073                 rte_errno = EINVAL;
2074                 return rte_errno;
2075         }
2076         return 0;
2077 }
2078
2079 /**
2080  * Look for the ethernet device belonging to mlx5 driver.
2081  *
2082  * @param[in] port_id
2083  *   port_id to start looking for device.
2084  * @param[in] pci_dev
2085  *   Pointer to the hint PCI device. When device is being probed
2086  *   the its siblings (master and preceding representors might
2087  *   not have assigned driver yet (because the mlx5_os_pci_probe()
2088  *   is not completed yet, for this case match on hint PCI
2089  *   device may be used to detect sibling device.
2090  *
2091  * @return
2092  *   port_id of found device, RTE_MAX_ETHPORT if not found.
2093  */
2094 uint16_t
2095 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
2096 {
2097         while (port_id < RTE_MAX_ETHPORTS) {
2098                 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2099
2100                 if (dev->state != RTE_ETH_DEV_UNUSED &&
2101                     dev->device &&
2102                     (dev->device == &pci_dev->device ||
2103                      (dev->device->driver &&
2104                      dev->device->driver->name &&
2105                      !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
2106                         break;
2107                 port_id++;
2108         }
2109         if (port_id >= RTE_MAX_ETHPORTS)
2110                 return RTE_MAX_ETHPORTS;
2111         return port_id;
2112 }
2113
2114 /**
2115  * DPDK callback to remove a PCI device.
2116  *
2117  * This function removes all Ethernet devices belong to a given PCI device.
2118  *
2119  * @param[in] pci_dev
2120  *   Pointer to the PCI device.
2121  *
2122  * @return
2123  *   0 on success, the function cannot fail.
2124  */
2125 static int
2126 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2127 {
2128         uint16_t port_id;
2129         int ret = 0;
2130
2131         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
2132                 /*
2133                  * mlx5_dev_close() is not registered to secondary process,
2134                  * call the close function explicitly for secondary process.
2135                  */
2136                 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2137                         ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2138                 else
2139                         ret |= rte_eth_dev_close(port_id);
2140         }
2141         return ret == 0 ? 0 : -EIO;
2142 }
2143
2144 static const struct rte_pci_id mlx5_pci_id_map[] = {
2145         {
2146                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2147                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2148         },
2149         {
2150                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2151                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2152         },
2153         {
2154                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2155                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2156         },
2157         {
2158                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2159                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2160         },
2161         {
2162                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2163                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2164         },
2165         {
2166                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2167                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2168         },
2169         {
2170                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2171                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2172         },
2173         {
2174                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2175                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2176         },
2177         {
2178                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2179                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2180         },
2181         {
2182                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2183                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2184         },
2185         {
2186                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2187                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2188         },
2189         {
2190                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2191                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2192         },
2193         {
2194                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2195                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2196         },
2197         {
2198                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2199                                 PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2200         },
2201         {
2202                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2203                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2204         },
2205         {
2206                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2207                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2208         },
2209         {
2210                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2211                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2212         },
2213         {
2214                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2215                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2216         },
2217         {
2218                 .vendor_id = 0
2219         }
2220 };
2221
2222 static struct mlx5_pci_driver mlx5_driver = {
2223         .driver_class = MLX5_CLASS_NET,
2224         .pci_driver = {
2225                 .driver = {
2226                         .name = MLX5_DRIVER_NAME,
2227                 },
2228                 .id_table = mlx5_pci_id_map,
2229                 .probe = mlx5_os_pci_probe,
2230                 .remove = mlx5_pci_remove,
2231                 .dma_map = mlx5_dma_map,
2232                 .dma_unmap = mlx5_dma_unmap,
2233                 .drv_flags = PCI_DRV_FLAGS,
2234         },
2235 };
2236
2237 /* Initialize driver log type. */
2238 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
2239
2240 /**
2241  * Driver initialization routine.
2242  */
2243 RTE_INIT(rte_mlx5_pmd_init)
2244 {
2245         pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2246         mlx5_common_init();
2247         /* Build the static tables for Verbs conversion. */
2248         mlx5_set_ptype_table();
2249         mlx5_set_cksum_table();
2250         mlx5_set_swp_types_table();
2251         if (mlx5_glue)
2252                 mlx5_pci_driver_register(&mlx5_driver);
2253 }
2254
2255 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2256 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2257 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");