net/mlx5: allow multiple flow tables on same level
[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 <ethdev_driver.h>
15 #include <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_rx.h"
39 #include "mlx5_tx.h"
40 #include "mlx5_autoconf.h"
41 #include "mlx5_mr.h"
42 #include "mlx5_flow.h"
43 #include "mlx5_flow_os.h"
44 #include "rte_pmd_mlx5.h"
45
46 /* Device parameter to enable RX completion queue compression. */
47 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
48
49 /* Device parameter to enable padding Rx packet to cacheline size. */
50 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
51
52 /* Device parameter to enable Multi-Packet Rx queue. */
53 #define MLX5_RX_MPRQ_EN "mprq_en"
54
55 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
56 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
57
58 /* Device parameter to configure log 2 of the stride size for MPRQ. */
59 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
60
61 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
62 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
63
64 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
65 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
66
67 /* Device parameter to configure inline send. Deprecated, ignored.*/
68 #define MLX5_TXQ_INLINE "txq_inline"
69
70 /* Device parameter to limit packet size to inline with ordinary SEND. */
71 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
72
73 /* Device parameter to configure minimal data size to inline. */
74 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
75
76 /* Device parameter to limit packet size to inline with Enhanced MPW. */
77 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
78
79 /*
80  * Device parameter to configure the number of TX queues threshold for
81  * enabling inline send.
82  */
83 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
84
85 /*
86  * Device parameter to configure the number of TX queues threshold for
87  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
88  */
89 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
90
91 /* Device parameter to enable multi-packet send WQEs. */
92 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
93
94 /*
95  * Device parameter to force doorbell register mapping
96  * to non-cahed region eliminating the extra write memory barrier.
97  */
98 #define MLX5_TX_DB_NC "tx_db_nc"
99
100 /*
101  * Device parameter to include 2 dsegs in the title WQEBB.
102  * Deprecated, ignored.
103  */
104 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
105
106 /*
107  * Device parameter to limit the size of inlining packet.
108  * Deprecated, ignored.
109  */
110 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
111
112 /*
113  * Device parameter to enable Tx scheduling on timestamps
114  * and specify the packet pacing granularity in nanoseconds.
115  */
116 #define MLX5_TX_PP "tx_pp"
117
118 /*
119  * Device parameter to specify skew in nanoseconds on Tx datapath,
120  * it represents the time between SQ start WQE processing and
121  * appearing actual packet data on the wire.
122  */
123 #define MLX5_TX_SKEW "tx_skew"
124
125 /*
126  * Device parameter to enable hardware Tx vector.
127  * Deprecated, ignored (no vectorized Tx routines anymore).
128  */
129 #define MLX5_TX_VEC_EN "tx_vec_en"
130
131 /* Device parameter to enable hardware Rx vector. */
132 #define MLX5_RX_VEC_EN "rx_vec_en"
133
134 /* Allow L3 VXLAN flow creation. */
135 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
136
137 /* Activate DV E-Switch flow steering. */
138 #define MLX5_DV_ESW_EN "dv_esw_en"
139
140 /* Activate DV flow steering. */
141 #define MLX5_DV_FLOW_EN "dv_flow_en"
142
143 /* Enable extensive flow metadata support. */
144 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
145
146 /* Device parameter to let the user manage the lacp traffic of bonded device */
147 #define MLX5_LACP_BY_USER "lacp_by_user"
148
149 /* Activate Netlink support in VF mode. */
150 #define MLX5_VF_NL_EN "vf_nl_en"
151
152 /* Enable extending memsegs when creating a MR. */
153 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
154
155 /* Select port representors to instantiate. */
156 #define MLX5_REPRESENTOR "representor"
157
158 /* Device parameter to configure the maximum number of dump files per queue. */
159 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
160
161 /* Configure timeout of LRO session (in microseconds). */
162 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
163
164 /*
165  * Device parameter to configure the total data buffer size for a single
166  * hairpin queue (logarithm value).
167  */
168 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
169
170 /* Flow memory reclaim mode. */
171 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
172
173 /* The default memory allocator used in PMD. */
174 #define MLX5_SYS_MEM_EN "sys_mem_en"
175 /* Decap will be used or not. */
176 #define MLX5_DECAP_EN "decap_en"
177
178 /* Shared memory between primary and secondary processes. */
179 struct mlx5_shared_data *mlx5_shared_data;
180
181 /** Driver-specific log messages type. */
182 int mlx5_logtype;
183
184 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
185                                                 LIST_HEAD_INITIALIZER();
186 static pthread_mutex_t mlx5_dev_ctx_list_mutex;
187 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
188 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
189         [MLX5_IPOOL_DECAP_ENCAP] = {
190                 .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
191                 .trunk_size = 64,
192                 .grow_trunk = 3,
193                 .grow_shift = 2,
194                 .need_lock = 1,
195                 .release_mem_en = 1,
196                 .malloc = mlx5_malloc,
197                 .free = mlx5_free,
198                 .type = "mlx5_encap_decap_ipool",
199         },
200         [MLX5_IPOOL_PUSH_VLAN] = {
201                 .size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
202                 .trunk_size = 64,
203                 .grow_trunk = 3,
204                 .grow_shift = 2,
205                 .need_lock = 1,
206                 .release_mem_en = 1,
207                 .malloc = mlx5_malloc,
208                 .free = mlx5_free,
209                 .type = "mlx5_push_vlan_ipool",
210         },
211         [MLX5_IPOOL_TAG] = {
212                 .size = sizeof(struct mlx5_flow_dv_tag_resource),
213                 .trunk_size = 64,
214                 .grow_trunk = 3,
215                 .grow_shift = 2,
216                 .need_lock = 1,
217                 .release_mem_en = 1,
218                 .malloc = mlx5_malloc,
219                 .free = mlx5_free,
220                 .type = "mlx5_tag_ipool",
221         },
222         [MLX5_IPOOL_PORT_ID] = {
223                 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
224                 .trunk_size = 64,
225                 .grow_trunk = 3,
226                 .grow_shift = 2,
227                 .need_lock = 1,
228                 .release_mem_en = 1,
229                 .malloc = mlx5_malloc,
230                 .free = mlx5_free,
231                 .type = "mlx5_port_id_ipool",
232         },
233         [MLX5_IPOOL_JUMP] = {
234                 .size = sizeof(struct mlx5_flow_tbl_data_entry),
235                 .trunk_size = 64,
236                 .grow_trunk = 3,
237                 .grow_shift = 2,
238                 .need_lock = 1,
239                 .release_mem_en = 1,
240                 .malloc = mlx5_malloc,
241                 .free = mlx5_free,
242                 .type = "mlx5_jump_ipool",
243         },
244         [MLX5_IPOOL_SAMPLE] = {
245                 .size = sizeof(struct mlx5_flow_dv_sample_resource),
246                 .trunk_size = 64,
247                 .grow_trunk = 3,
248                 .grow_shift = 2,
249                 .need_lock = 1,
250                 .release_mem_en = 1,
251                 .malloc = mlx5_malloc,
252                 .free = mlx5_free,
253                 .type = "mlx5_sample_ipool",
254         },
255         [MLX5_IPOOL_DEST_ARRAY] = {
256                 .size = sizeof(struct mlx5_flow_dv_dest_array_resource),
257                 .trunk_size = 64,
258                 .grow_trunk = 3,
259                 .grow_shift = 2,
260                 .need_lock = 1,
261                 .release_mem_en = 1,
262                 .malloc = mlx5_malloc,
263                 .free = mlx5_free,
264                 .type = "mlx5_dest_array_ipool",
265         },
266         [MLX5_IPOOL_TUNNEL_ID] = {
267                 .size = sizeof(struct mlx5_flow_tunnel),
268                 .trunk_size = MLX5_MAX_TUNNELS,
269                 .need_lock = 1,
270                 .release_mem_en = 1,
271                 .type = "mlx5_tunnel_offload",
272         },
273         [MLX5_IPOOL_TNL_TBL_ID] = {
274                 .size = 0,
275                 .need_lock = 1,
276                 .type = "mlx5_flow_tnl_tbl_ipool",
277         },
278 #endif
279         [MLX5_IPOOL_MTR] = {
280                 /**
281                  * The ipool index should grow continually from small to big,
282                  * for meter idx, so not set grow_trunk to avoid meter index
283                  * not jump continually.
284                  */
285                 .size = sizeof(struct mlx5_legacy_flow_meter),
286                 .trunk_size = 64,
287                 .need_lock = 1,
288                 .release_mem_en = 1,
289                 .malloc = mlx5_malloc,
290                 .free = mlx5_free,
291                 .type = "mlx5_meter_ipool",
292         },
293         [MLX5_IPOOL_MCP] = {
294                 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
295                 .trunk_size = 64,
296                 .grow_trunk = 3,
297                 .grow_shift = 2,
298                 .need_lock = 1,
299                 .release_mem_en = 1,
300                 .malloc = mlx5_malloc,
301                 .free = mlx5_free,
302                 .type = "mlx5_mcp_ipool",
303         },
304         [MLX5_IPOOL_HRXQ] = {
305                 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
306                 .trunk_size = 64,
307                 .grow_trunk = 3,
308                 .grow_shift = 2,
309                 .need_lock = 1,
310                 .release_mem_en = 1,
311                 .malloc = mlx5_malloc,
312                 .free = mlx5_free,
313                 .type = "mlx5_hrxq_ipool",
314         },
315         [MLX5_IPOOL_MLX5_FLOW] = {
316                 /*
317                  * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
318                  * It set in run time according to PCI function configuration.
319                  */
320                 .size = 0,
321                 .trunk_size = 64,
322                 .grow_trunk = 3,
323                 .grow_shift = 2,
324                 .need_lock = 1,
325                 .release_mem_en = 1,
326                 .malloc = mlx5_malloc,
327                 .free = mlx5_free,
328                 .type = "mlx5_flow_handle_ipool",
329         },
330         [MLX5_IPOOL_RTE_FLOW] = {
331                 .size = sizeof(struct rte_flow),
332                 .trunk_size = 4096,
333                 .need_lock = 1,
334                 .release_mem_en = 1,
335                 .malloc = mlx5_malloc,
336                 .free = mlx5_free,
337                 .type = "rte_flow_ipool",
338         },
339         [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
340                 .size = 0,
341                 .need_lock = 1,
342                 .type = "mlx5_flow_rss_id_ipool",
343         },
344         [MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
345                 .size = sizeof(struct mlx5_shared_action_rss),
346                 .trunk_size = 64,
347                 .grow_trunk = 3,
348                 .grow_shift = 2,
349                 .need_lock = 1,
350                 .release_mem_en = 1,
351                 .malloc = mlx5_malloc,
352                 .free = mlx5_free,
353                 .type = "mlx5_shared_action_rss",
354         },
355 };
356
357
358 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
359 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
360
361 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
362
363 /**
364  * Decide whether representor ID is a HPF(host PF) port on BF2.
365  *
366  * @param dev
367  *   Pointer to Ethernet device structure.
368  *
369  * @return
370  *   Non-zero if HPF, otherwise 0.
371  */
372 bool
373 mlx5_is_hpf(struct rte_eth_dev *dev)
374 {
375         struct mlx5_priv *priv = dev->data->dev_private;
376         uint16_t repr = MLX5_REPRESENTOR_REPR(priv->representor_id);
377         int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
378
379         return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_VF &&
380                MLX5_REPRESENTOR_REPR(-1) == repr;
381 }
382
383 /**
384  * Initialize the ASO aging management structure.
385  *
386  * @param[in] sh
387  *   Pointer to mlx5_dev_ctx_shared object to free
388  *
389  * @return
390  *   0 on success, a negative errno value otherwise and rte_errno is set.
391  */
392 int
393 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
394 {
395         int err;
396
397         if (sh->aso_age_mng)
398                 return 0;
399         sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
400                                       RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
401         if (!sh->aso_age_mng) {
402                 DRV_LOG(ERR, "aso_age_mng allocation was failed.");
403                 rte_errno = ENOMEM;
404                 return -ENOMEM;
405         }
406         err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_FLOW_HIT);
407         if (err) {
408                 mlx5_free(sh->aso_age_mng);
409                 return -1;
410         }
411         rte_spinlock_init(&sh->aso_age_mng->resize_sl);
412         rte_spinlock_init(&sh->aso_age_mng->free_sl);
413         LIST_INIT(&sh->aso_age_mng->free);
414         return 0;
415 }
416
417 /**
418  * Close and release all the resources of the ASO aging management structure.
419  *
420  * @param[in] sh
421  *   Pointer to mlx5_dev_ctx_shared object to free.
422  */
423 static void
424 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
425 {
426         int i, j;
427
428         mlx5_aso_flow_hit_queue_poll_stop(sh);
429         mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_FLOW_HIT);
430         if (sh->aso_age_mng->pools) {
431                 struct mlx5_aso_age_pool *pool;
432
433                 for (i = 0; i < sh->aso_age_mng->next; ++i) {
434                         pool = sh->aso_age_mng->pools[i];
435                         claim_zero(mlx5_devx_cmd_destroy
436                                                 (pool->flow_hit_aso_obj));
437                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
438                                 if (pool->actions[j].dr_action)
439                                         claim_zero
440                                             (mlx5_flow_os_destroy_flow_action
441                                               (pool->actions[j].dr_action));
442                         mlx5_free(pool);
443                 }
444                 mlx5_free(sh->aso_age_mng->pools);
445         }
446         mlx5_free(sh->aso_age_mng);
447 }
448
449 /**
450  * Initialize the shared aging list information per port.
451  *
452  * @param[in] sh
453  *   Pointer to mlx5_dev_ctx_shared object.
454  */
455 static void
456 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
457 {
458         uint32_t i;
459         struct mlx5_age_info *age_info;
460
461         for (i = 0; i < sh->max_port; i++) {
462                 age_info = &sh->port[i].age_info;
463                 age_info->flags = 0;
464                 TAILQ_INIT(&age_info->aged_counters);
465                 LIST_INIT(&age_info->aged_aso);
466                 rte_spinlock_init(&age_info->aged_sl);
467                 MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
468         }
469 }
470
471 /**
472  * Initialize the counters management structure.
473  *
474  * @param[in] sh
475  *   Pointer to mlx5_dev_ctx_shared object to free
476  */
477 static void
478 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
479 {
480         int i;
481
482         memset(&sh->cmng, 0, sizeof(sh->cmng));
483         TAILQ_INIT(&sh->cmng.flow_counters);
484         sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
485         sh->cmng.max_id = -1;
486         sh->cmng.last_pool_idx = POOL_IDX_INVALID;
487         rte_spinlock_init(&sh->cmng.pool_update_sl);
488         for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
489                 TAILQ_INIT(&sh->cmng.counters[i]);
490                 rte_spinlock_init(&sh->cmng.csl[i]);
491         }
492 }
493
494 /**
495  * Destroy all the resources allocated for a counter memory management.
496  *
497  * @param[in] mng
498  *   Pointer to the memory management structure.
499  */
500 static void
501 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
502 {
503         uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
504
505         LIST_REMOVE(mng, next);
506         claim_zero(mlx5_devx_cmd_destroy(mng->dm));
507         claim_zero(mlx5_os_umem_dereg(mng->umem));
508         mlx5_free(mem);
509 }
510
511 /**
512  * Close and release all the resources of the counters management.
513  *
514  * @param[in] sh
515  *   Pointer to mlx5_dev_ctx_shared object to free.
516  */
517 static void
518 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
519 {
520         struct mlx5_counter_stats_mem_mng *mng;
521         int i, j;
522         int retries = 1024;
523
524         rte_errno = 0;
525         while (--retries) {
526                 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
527                 if (rte_errno != EINPROGRESS)
528                         break;
529                 rte_pause();
530         }
531
532         if (sh->cmng.pools) {
533                 struct mlx5_flow_counter_pool *pool;
534                 uint16_t n_valid = sh->cmng.n_valid;
535                 bool fallback = sh->cmng.counter_fallback;
536
537                 for (i = 0; i < n_valid; ++i) {
538                         pool = sh->cmng.pools[i];
539                         if (!fallback && pool->min_dcs)
540                                 claim_zero(mlx5_devx_cmd_destroy
541                                                                (pool->min_dcs));
542                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
543                                 struct mlx5_flow_counter *cnt =
544                                                 MLX5_POOL_GET_CNT(pool, j);
545
546                                 if (cnt->action)
547                                         claim_zero
548                                          (mlx5_flow_os_destroy_flow_action
549                                           (cnt->action));
550                                 if (fallback && MLX5_POOL_GET_CNT
551                                     (pool, j)->dcs_when_free)
552                                         claim_zero(mlx5_devx_cmd_destroy
553                                                    (cnt->dcs_when_free));
554                         }
555                         mlx5_free(pool);
556                 }
557                 mlx5_free(sh->cmng.pools);
558         }
559         mng = LIST_FIRST(&sh->cmng.mem_mngs);
560         while (mng) {
561                 mlx5_flow_destroy_counter_stat_mem_mng(mng);
562                 mng = LIST_FIRST(&sh->cmng.mem_mngs);
563         }
564         memset(&sh->cmng, 0, sizeof(sh->cmng));
565 }
566
567 /**
568  * Initialize the aso flow meters management structure.
569  *
570  * @param[in] sh
571  *   Pointer to mlx5_dev_ctx_shared object to free
572  */
573 int
574 mlx5_aso_flow_mtrs_mng_init(struct mlx5_priv *priv)
575 {
576         if (!priv->mtr_idx_tbl) {
577                 priv->mtr_idx_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
578                 if (!priv->mtr_idx_tbl) {
579                         DRV_LOG(ERR, "fail to create meter lookup table.");
580                         rte_errno = ENOMEM;
581                         return -ENOMEM;
582                 }
583         }
584         if (!priv->sh->mtrmng) {
585                 priv->sh->mtrmng = mlx5_malloc(MLX5_MEM_ZERO,
586                         sizeof(*priv->sh->mtrmng),
587                         RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
588                 if (!priv->sh->mtrmng) {
589                         DRV_LOG(ERR, "mlx5_aso_mtr_pools_mng allocation was failed.");
590                         rte_errno = ENOMEM;
591                         return -ENOMEM;
592                 }
593                 rte_spinlock_init(&priv->sh->mtrmng->mtrsl);
594                 LIST_INIT(&priv->sh->mtrmng->meters);
595         }
596         return 0;
597 }
598
599 /**
600  * Close and release all the resources of
601  * the ASO flow meter management structure.
602  *
603  * @param[in] sh
604  *   Pointer to mlx5_dev_ctx_shared object to free.
605  */
606 static void
607 mlx5_aso_flow_mtrs_mng_close(struct mlx5_dev_ctx_shared *sh)
608 {
609         struct mlx5_aso_mtr_pool *mtr_pool;
610         struct mlx5_aso_mtr_pools_mng *mtrmng = sh->mtrmng;
611         uint32_t idx;
612 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
613         struct mlx5_aso_mtr *aso_mtr;
614         int i;
615 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
616
617         mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_POLICER);
618         idx = mtrmng->n_valid;
619         while (idx--) {
620                 mtr_pool = mtrmng->pools[idx];
621 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
622                 for (i = 0; i < MLX5_ASO_MTRS_PER_POOL; i++) {
623                         aso_mtr = &mtr_pool->mtrs[i];
624                         if (aso_mtr->fm.meter_action)
625                                 claim_zero(mlx5_glue->destroy_flow_action
626                                                 (aso_mtr->fm.meter_action));
627                 }
628 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
629                 claim_zero(mlx5_devx_cmd_destroy
630                                                 (mtr_pool->devx_obj));
631                 mtrmng->n_valid--;
632                 mlx5_free(mtr_pool);
633         }
634         mlx5_free(sh->mtrmng->pools);
635         mlx5_free(sh->mtrmng);
636         sh->mtrmng = NULL;
637 }
638
639 /* Send FLOW_AGED event if needed. */
640 void
641 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
642 {
643         struct mlx5_age_info *age_info;
644         uint32_t i;
645
646         for (i = 0; i < sh->max_port; i++) {
647                 age_info = &sh->port[i].age_info;
648                 if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
649                         continue;
650                 if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER))
651                         rte_eth_dev_callback_process
652                                 (&rte_eth_devices[sh->port[i].devx_ih_port_id],
653                                 RTE_ETH_EVENT_FLOW_AGED, NULL);
654                 age_info->flags = 0;
655         }
656 }
657
658 /**
659  * Initialize the flow resources' indexed mempool.
660  *
661  * @param[in] sh
662  *   Pointer to mlx5_dev_ctx_shared object.
663  * @param[in] sh
664  *   Pointer to user dev config.
665  */
666 static void
667 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
668                        const struct mlx5_dev_config *config)
669 {
670         uint8_t i;
671         struct mlx5_indexed_pool_config cfg;
672
673         for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
674                 cfg = mlx5_ipool_cfg[i];
675                 switch (i) {
676                 default:
677                         break;
678                 /*
679                  * Set MLX5_IPOOL_MLX5_FLOW ipool size
680                  * according to PCI function flow configuration.
681                  */
682                 case MLX5_IPOOL_MLX5_FLOW:
683                         cfg.size = config->dv_flow_en ?
684                                 sizeof(struct mlx5_flow_handle) :
685                                 MLX5_FLOW_HANDLE_VERBS_SIZE;
686                         break;
687                 }
688                 if (config->reclaim_mode)
689                         cfg.release_mem_en = 1;
690                 sh->ipool[i] = mlx5_ipool_create(&cfg);
691         }
692 }
693
694 /**
695  * Release the flow resources' indexed mempool.
696  *
697  * @param[in] sh
698  *   Pointer to mlx5_dev_ctx_shared object.
699  */
700 static void
701 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
702 {
703         uint8_t i;
704
705         for (i = 0; i < MLX5_IPOOL_MAX; ++i)
706                 mlx5_ipool_destroy(sh->ipool[i]);
707 }
708
709 /*
710  * Check if dynamic flex parser for eCPRI already exists.
711  *
712  * @param dev
713  *   Pointer to Ethernet device structure.
714  *
715  * @return
716  *   true on exists, false on not.
717  */
718 bool
719 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
720 {
721         struct mlx5_priv *priv = dev->data->dev_private;
722         struct mlx5_flex_parser_profiles *prf =
723                                 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
724
725         return !!prf->obj;
726 }
727
728 /*
729  * Allocation of a flex parser for eCPRI. Once created, this parser related
730  * resources will be held until the device is closed.
731  *
732  * @param dev
733  *   Pointer to Ethernet device structure.
734  *
735  * @return
736  *   0 on success, a negative errno value otherwise and rte_errno is set.
737  */
738 int
739 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
740 {
741         struct mlx5_priv *priv = dev->data->dev_private;
742         struct mlx5_flex_parser_profiles *prf =
743                                 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
744         struct mlx5_devx_graph_node_attr node = {
745                 .modify_field_select = 0,
746         };
747         uint32_t ids[8];
748         int ret;
749
750         if (!priv->config.hca_attr.parse_graph_flex_node) {
751                 DRV_LOG(ERR, "Dynamic flex parser is not supported "
752                         "for device %s.", priv->dev_data->name);
753                 return -ENOTSUP;
754         }
755         node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
756         /* 8 bytes now: 4B common header + 4B message body header. */
757         node.header_length_base_value = 0x8;
758         /* After MAC layer: Ether / VLAN. */
759         node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
760         /* Type of compared condition should be 0xAEFE in the L2 layer. */
761         node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
762         /* Sample #0: type in common header. */
763         node.sample[0].flow_match_sample_en = 1;
764         /* Fixed offset. */
765         node.sample[0].flow_match_sample_offset_mode = 0x0;
766         /* Only the 2nd byte will be used. */
767         node.sample[0].flow_match_sample_field_base_offset = 0x0;
768         /* Sample #1: message payload. */
769         node.sample[1].flow_match_sample_en = 1;
770         /* Fixed offset. */
771         node.sample[1].flow_match_sample_offset_mode = 0x0;
772         /*
773          * Only the first two bytes will be used right now, and its offset will
774          * start after the common header that with the length of a DW(u32).
775          */
776         node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
777         prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
778         if (!prf->obj) {
779                 DRV_LOG(ERR, "Failed to create flex parser node object.");
780                 return (rte_errno == 0) ? -ENODEV : -rte_errno;
781         }
782         prf->num = 2;
783         ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
784         if (ret) {
785                 DRV_LOG(ERR, "Failed to query sample IDs.");
786                 return (rte_errno == 0) ? -ENODEV : -rte_errno;
787         }
788         prf->offset[0] = 0x0;
789         prf->offset[1] = sizeof(uint32_t);
790         prf->ids[0] = ids[0];
791         prf->ids[1] = ids[1];
792         return 0;
793 }
794
795 /*
796  * Destroy the flex parser node, including the parser itself, input / output
797  * arcs and DW samples. Resources could be reused then.
798  *
799  * @param dev
800  *   Pointer to Ethernet device structure.
801  */
802 static void
803 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
804 {
805         struct mlx5_priv *priv = dev->data->dev_private;
806         struct mlx5_flex_parser_profiles *prf =
807                                 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
808
809         if (prf->obj)
810                 mlx5_devx_cmd_destroy(prf->obj);
811         prf->obj = NULL;
812 }
813
814 /*
815  * Allocate Rx and Tx UARs in robust fashion.
816  * This routine handles the following UAR allocation issues:
817  *
818  *  - tries to allocate the UAR with the most appropriate memory
819  *    mapping type from the ones supported by the host
820  *
821  *  - tries to allocate the UAR with non-NULL base address
822  *    OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
823  *    UAR base address if UAR was not the first object in the UAR page.
824  *    It caused the PMD failure and we should try to get another UAR
825  *    till we get the first one with non-NULL base address returned.
826  */
827 static int
828 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
829                      const struct mlx5_dev_config *config)
830 {
831         uint32_t uar_mapping, retry;
832         int err = 0;
833         void *base_addr;
834
835         for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
836 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
837                 /* Control the mapping type according to the settings. */
838                 uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
839                               MLX5DV_UAR_ALLOC_TYPE_NC :
840                               MLX5DV_UAR_ALLOC_TYPE_BF;
841 #else
842                 RTE_SET_USED(config);
843                 /*
844                  * It seems we have no way to control the memory mapping type
845                  * for the UAR, the default "Write-Combining" type is supposed.
846                  * The UAR initialization on queue creation queries the
847                  * actual mapping type done by Verbs/kernel and setups the
848                  * PMD datapath accordingly.
849                  */
850                 uar_mapping = 0;
851 #endif
852                 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
853 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
854                 if (!sh->tx_uar &&
855                     uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
856                         if (config->dbnc == MLX5_TXDB_CACHED ||
857                             config->dbnc == MLX5_TXDB_HEURISTIC)
858                                 DRV_LOG(WARNING, "Devarg tx_db_nc setting "
859                                                  "is not supported by DevX");
860                         /*
861                          * In some environments like virtual machine
862                          * the Write Combining mapped might be not supported
863                          * and UAR allocation fails. We try "Non-Cached"
864                          * mapping for the case. The tx_burst routines take
865                          * the UAR mapping type into account on UAR setup
866                          * on queue creation.
867                          */
868                         DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (BF)");
869                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
870                         sh->tx_uar = mlx5_glue->devx_alloc_uar
871                                                         (sh->ctx, uar_mapping);
872                 } else if (!sh->tx_uar &&
873                            uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
874                         if (config->dbnc == MLX5_TXDB_NCACHED)
875                                 DRV_LOG(WARNING, "Devarg tx_db_nc settings "
876                                                  "is not supported by DevX");
877                         /*
878                          * If Verbs/kernel does not support "Non-Cached"
879                          * try the "Write-Combining".
880                          */
881                         DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (NC)");
882                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
883                         sh->tx_uar = mlx5_glue->devx_alloc_uar
884                                                         (sh->ctx, uar_mapping);
885                 }
886 #endif
887                 if (!sh->tx_uar) {
888                         DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
889                         err = ENOMEM;
890                         goto exit;
891                 }
892                 base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
893                 if (base_addr)
894                         break;
895                 /*
896                  * The UARs are allocated by rdma_core within the
897                  * IB device context, on context closure all UARs
898                  * will be freed, should be no memory/object leakage.
899                  */
900                 DRV_LOG(DEBUG, "Retrying to allocate Tx DevX UAR");
901                 sh->tx_uar = NULL;
902         }
903         /* Check whether we finally succeeded with valid UAR allocation. */
904         if (!sh->tx_uar) {
905                 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
906                 err = ENOMEM;
907                 goto exit;
908         }
909         for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
910                 uar_mapping = 0;
911                 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
912                                                         (sh->ctx, uar_mapping);
913 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
914                 if (!sh->devx_rx_uar &&
915                     uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
916                         /*
917                          * Rx UAR is used to control interrupts only,
918                          * should be no datapath noticeable impact,
919                          * can try "Non-Cached" mapping safely.
920                          */
921                         DRV_LOG(DEBUG, "Failed to allocate Rx DevX UAR (BF)");
922                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
923                         sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
924                                                         (sh->ctx, uar_mapping);
925                 }
926 #endif
927                 if (!sh->devx_rx_uar) {
928                         DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
929                         err = ENOMEM;
930                         goto exit;
931                 }
932                 base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
933                 if (base_addr)
934                         break;
935                 /*
936                  * The UARs are allocated by rdma_core within the
937                  * IB device context, on context closure all UARs
938                  * will be freed, should be no memory/object leakage.
939                  */
940                 DRV_LOG(DEBUG, "Retrying to allocate Rx DevX UAR");
941                 sh->devx_rx_uar = NULL;
942         }
943         /* Check whether we finally succeeded with valid UAR allocation. */
944         if (!sh->devx_rx_uar) {
945                 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
946                 err = ENOMEM;
947         }
948 exit:
949         return err;
950 }
951
952 /**
953  * Allocate shared device context. If there is multiport device the
954  * master and representors will share this context, if there is single
955  * port dedicated device, the context will be used by only given
956  * port due to unification.
957  *
958  * Routine first searches the context for the specified device name,
959  * if found the shared context assumed and reference counter is incremented.
960  * If no context found the new one is created and initialized with specified
961  * device context and parameters.
962  *
963  * @param[in] spawn
964  *   Pointer to the device attributes (name, port, etc).
965  * @param[in] config
966  *   Pointer to device configuration structure.
967  *
968  * @return
969  *   Pointer to mlx5_dev_ctx_shared object on success,
970  *   otherwise NULL and rte_errno is set.
971  */
972 struct mlx5_dev_ctx_shared *
973 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
974                            const struct mlx5_dev_config *config)
975 {
976         struct mlx5_dev_ctx_shared *sh;
977         int err = 0;
978         uint32_t i;
979         struct mlx5_devx_tis_attr tis_attr = { 0 };
980
981         MLX5_ASSERT(spawn);
982         /* Secondary process should not create the shared context. */
983         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
984         pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
985         /* Search for IB context by device name. */
986         LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
987                 if (!strcmp(sh->ibdev_name,
988                         mlx5_os_get_dev_device_name(spawn->phys_dev))) {
989                         sh->refcnt++;
990                         goto exit;
991                 }
992         }
993         /* No device found, we have to create new shared context. */
994         MLX5_ASSERT(spawn->max_port);
995         sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
996                          sizeof(struct mlx5_dev_ctx_shared) +
997                          spawn->max_port *
998                          sizeof(struct mlx5_dev_shared_port),
999                          RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1000         if (!sh) {
1001                 DRV_LOG(ERR, "shared context allocation failure");
1002                 rte_errno  = ENOMEM;
1003                 goto exit;
1004         }
1005         if (spawn->bond_info)
1006                 sh->bond = *spawn->bond_info;
1007         err = mlx5_os_open_device(spawn, config, sh);
1008         if (!sh->ctx)
1009                 goto error;
1010         err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
1011         if (err) {
1012                 DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
1013                 goto error;
1014         }
1015         sh->refcnt = 1;
1016         sh->max_port = spawn->max_port;
1017         strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
1018                 sizeof(sh->ibdev_name) - 1);
1019         strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
1020                 sizeof(sh->ibdev_path) - 1);
1021         /*
1022          * Setting port_id to max unallowed value means
1023          * there is no interrupt subhandler installed for
1024          * the given port index i.
1025          */
1026         for (i = 0; i < sh->max_port; i++) {
1027                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
1028                 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
1029         }
1030         sh->pd = mlx5_os_alloc_pd(sh->ctx);
1031         if (sh->pd == NULL) {
1032                 DRV_LOG(ERR, "PD allocation failure");
1033                 err = ENOMEM;
1034                 goto error;
1035         }
1036         if (sh->devx) {
1037                 err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
1038                 if (err) {
1039                         DRV_LOG(ERR, "Fail to extract pdn from PD");
1040                         goto error;
1041                 }
1042                 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
1043                 if (!sh->td) {
1044                         DRV_LOG(ERR, "TD allocation failure");
1045                         err = ENOMEM;
1046                         goto error;
1047                 }
1048                 tis_attr.transport_domain = sh->td->id;
1049                 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
1050                 if (!sh->tis) {
1051                         DRV_LOG(ERR, "TIS allocation failure");
1052                         err = ENOMEM;
1053                         goto error;
1054                 }
1055                 err = mlx5_alloc_rxtx_uars(sh, config);
1056                 if (err)
1057                         goto error;
1058                 MLX5_ASSERT(sh->tx_uar);
1059                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
1060
1061                 MLX5_ASSERT(sh->devx_rx_uar);
1062                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
1063         }
1064 #ifndef RTE_ARCH_64
1065         /* Initialize UAR access locks for 32bit implementations. */
1066         rte_spinlock_init(&sh->uar_lock_cq);
1067         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1068                 rte_spinlock_init(&sh->uar_lock[i]);
1069 #endif
1070         /*
1071          * Once the device is added to the list of memory event
1072          * callback, its global MR cache table cannot be expanded
1073          * on the fly because of deadlock. If it overflows, lookup
1074          * should be done by searching MR list linearly, which is slow.
1075          *
1076          * At this point the device is not added to the memory
1077          * event list yet, context is just being created.
1078          */
1079         err = mlx5_mr_btree_init(&sh->share_cache.cache,
1080                                  MLX5_MR_BTREE_CACHE_N * 2,
1081                                  spawn->pci_dev->device.numa_node);
1082         if (err) {
1083                 err = rte_errno;
1084                 goto error;
1085         }
1086         mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
1087                               &sh->share_cache.dereg_mr_cb);
1088         mlx5_os_dev_shared_handler_install(sh);
1089         sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
1090         if (!sh->cnt_id_tbl) {
1091                 err = rte_errno;
1092                 goto error;
1093         }
1094         if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1095                 err = mlx5_flow_os_init_workspace_once();
1096                 if (err)
1097                         goto error;
1098         }
1099         mlx5_flow_aging_init(sh);
1100         mlx5_flow_counters_mng_init(sh);
1101         mlx5_flow_ipool_create(sh, config);
1102         /* Add device to memory callback list. */
1103         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1104         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1105                          sh, mem_event_cb);
1106         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1107         /* Add context to the global device list. */
1108         LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1109         rte_spinlock_init(&sh->geneve_tlv_opt_sl);
1110 exit:
1111         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1112         return sh;
1113 error:
1114         pthread_mutex_destroy(&sh->txpp.mutex);
1115         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1116         MLX5_ASSERT(sh);
1117         if (sh->cnt_id_tbl)
1118                 mlx5_l3t_destroy(sh->cnt_id_tbl);
1119         if (sh->tis)
1120                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1121         if (sh->td)
1122                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1123         if (sh->devx_rx_uar)
1124                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1125         if (sh->tx_uar)
1126                 mlx5_glue->devx_free_uar(sh->tx_uar);
1127         if (sh->pd)
1128                 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1129         if (sh->ctx)
1130                 claim_zero(mlx5_glue->close_device(sh->ctx));
1131         mlx5_free(sh);
1132         MLX5_ASSERT(err > 0);
1133         rte_errno = err;
1134         return NULL;
1135 }
1136
1137 /**
1138  * Free shared IB device context. Decrement counter and if zero free
1139  * all allocated resources and close handles.
1140  *
1141  * @param[in] sh
1142  *   Pointer to mlx5_dev_ctx_shared object to free
1143  */
1144 void
1145 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1146 {
1147         pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1148 #ifdef RTE_LIBRTE_MLX5_DEBUG
1149         /* Check the object presence in the list. */
1150         struct mlx5_dev_ctx_shared *lctx;
1151
1152         LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1153                 if (lctx == sh)
1154                         break;
1155         MLX5_ASSERT(lctx);
1156         if (lctx != sh) {
1157                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
1158                 goto exit;
1159         }
1160 #endif
1161         MLX5_ASSERT(sh);
1162         MLX5_ASSERT(sh->refcnt);
1163         /* Secondary process should not free the shared context. */
1164         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1165         if (--sh->refcnt)
1166                 goto exit;
1167         /* Remove from memory callback device list. */
1168         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1169         LIST_REMOVE(sh, mem_event_cb);
1170         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1171         /* Release created Memory Regions. */
1172         mlx5_mr_release_cache(&sh->share_cache);
1173         /* Remove context from the global device list. */
1174         LIST_REMOVE(sh, next);
1175         /* Release flow workspaces objects on the last device. */
1176         if (LIST_EMPTY(&mlx5_dev_ctx_list))
1177                 mlx5_flow_os_release_workspace();
1178         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1179         /*
1180          *  Ensure there is no async event handler installed.
1181          *  Only primary process handles async device events.
1182          **/
1183         mlx5_flow_counters_mng_close(sh);
1184         if (sh->aso_age_mng) {
1185                 mlx5_flow_aso_age_mng_close(sh);
1186                 sh->aso_age_mng = NULL;
1187         }
1188         if (sh->mtrmng)
1189                 mlx5_aso_flow_mtrs_mng_close(sh);
1190         mlx5_flow_ipool_destroy(sh);
1191         mlx5_os_dev_shared_handler_uninstall(sh);
1192         if (sh->cnt_id_tbl) {
1193                 mlx5_l3t_destroy(sh->cnt_id_tbl);
1194                 sh->cnt_id_tbl = NULL;
1195         }
1196         if (sh->tx_uar) {
1197                 mlx5_glue->devx_free_uar(sh->tx_uar);
1198                 sh->tx_uar = NULL;
1199         }
1200         if (sh->pd)
1201                 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1202         if (sh->tis)
1203                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1204         if (sh->td)
1205                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1206         if (sh->devx_rx_uar)
1207                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1208         if (sh->ctx)
1209                 claim_zero(mlx5_glue->close_device(sh->ctx));
1210         MLX5_ASSERT(sh->geneve_tlv_option_resource == NULL);
1211         pthread_mutex_destroy(&sh->txpp.mutex);
1212         mlx5_free(sh);
1213         return;
1214 exit:
1215         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1216 }
1217
1218 /**
1219  * Destroy table hash list.
1220  *
1221  * @param[in] priv
1222  *   Pointer to the private device data structure.
1223  */
1224 void
1225 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1226 {
1227         struct mlx5_dev_ctx_shared *sh = priv->sh;
1228
1229         if (!sh->flow_tbls)
1230                 return;
1231         mlx5_hlist_destroy(sh->flow_tbls);
1232 }
1233
1234 /**
1235  * Initialize flow table hash list and create the root tables entry
1236  * for each domain.
1237  *
1238  * @param[in] priv
1239  *   Pointer to the private device data structure.
1240  *
1241  * @return
1242  *   Zero on success, positive error code otherwise.
1243  */
1244 int
1245 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1246 {
1247         int err = 0;
1248         /* Tables are only used in DV and DR modes. */
1249 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1250         struct mlx5_dev_ctx_shared *sh = priv->sh;
1251         char s[MLX5_HLIST_NAMESIZE];
1252
1253         MLX5_ASSERT(sh);
1254         snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1255         sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1256                                           0, 0, flow_dv_tbl_create_cb,
1257                                           flow_dv_tbl_match_cb,
1258                                           flow_dv_tbl_remove_cb);
1259         if (!sh->flow_tbls) {
1260                 DRV_LOG(ERR, "flow tables with hash creation failed.");
1261                 err = ENOMEM;
1262                 return err;
1263         }
1264         sh->flow_tbls->ctx = sh;
1265 #ifndef HAVE_MLX5DV_DR
1266         struct rte_flow_error error;
1267         struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1268
1269         /*
1270          * In case we have not DR support, the zero tables should be created
1271          * because DV expect to see them even if they cannot be created by
1272          * RDMA-CORE.
1273          */
1274         if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0,
1275                 NULL, 0, 1, 0, &error) ||
1276             !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0,
1277                 NULL, 0, 1, 0, &error) ||
1278             !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0,
1279                 NULL, 0, 1, 0, &error)) {
1280                 err = ENOMEM;
1281                 goto error;
1282         }
1283         return err;
1284 error:
1285         mlx5_free_table_hash_list(priv);
1286 #endif /* HAVE_MLX5DV_DR */
1287 #endif
1288         return err;
1289 }
1290
1291 /**
1292  * Retrieve integer value from environment variable.
1293  *
1294  * @param[in] name
1295  *   Environment variable name.
1296  *
1297  * @return
1298  *   Integer value, 0 if the variable is not set.
1299  */
1300 int
1301 mlx5_getenv_int(const char *name)
1302 {
1303         const char *val = getenv(name);
1304
1305         if (val == NULL)
1306                 return 0;
1307         return atoi(val);
1308 }
1309
1310 /**
1311  * DPDK callback to add udp tunnel port
1312  *
1313  * @param[in] dev
1314  *   A pointer to eth_dev
1315  * @param[in] udp_tunnel
1316  *   A pointer to udp tunnel
1317  *
1318  * @return
1319  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1320  */
1321 int
1322 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1323                          struct rte_eth_udp_tunnel *udp_tunnel)
1324 {
1325         MLX5_ASSERT(udp_tunnel != NULL);
1326         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1327             udp_tunnel->udp_port == 4789)
1328                 return 0;
1329         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1330             udp_tunnel->udp_port == 4790)
1331                 return 0;
1332         return -ENOTSUP;
1333 }
1334
1335 /**
1336  * Initialize process private data structure.
1337  *
1338  * @param dev
1339  *   Pointer to Ethernet device structure.
1340  *
1341  * @return
1342  *   0 on success, a negative errno value otherwise and rte_errno is set.
1343  */
1344 int
1345 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1346 {
1347         struct mlx5_priv *priv = dev->data->dev_private;
1348         struct mlx5_proc_priv *ppriv;
1349         size_t ppriv_size;
1350
1351         /*
1352          * UAR register table follows the process private structure. BlueFlame
1353          * registers for Tx queues are stored in the table.
1354          */
1355         ppriv_size =
1356                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1357         ppriv = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, ppriv_size,
1358                             RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1359         if (!ppriv) {
1360                 rte_errno = ENOMEM;
1361                 return -rte_errno;
1362         }
1363         ppriv->uar_table_sz = priv->txqs_n;
1364         dev->process_private = ppriv;
1365         return 0;
1366 }
1367
1368 /**
1369  * Un-initialize process private data structure.
1370  *
1371  * @param dev
1372  *   Pointer to Ethernet device structure.
1373  */
1374 void
1375 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1376 {
1377         if (!dev->process_private)
1378                 return;
1379         mlx5_free(dev->process_private);
1380         dev->process_private = NULL;
1381 }
1382
1383 /**
1384  * DPDK callback to close the device.
1385  *
1386  * Destroy all queues and objects, free memory.
1387  *
1388  * @param dev
1389  *   Pointer to Ethernet device structure.
1390  */
1391 int
1392 mlx5_dev_close(struct rte_eth_dev *dev)
1393 {
1394         struct mlx5_priv *priv = dev->data->dev_private;
1395         unsigned int i;
1396         int ret;
1397
1398         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1399                 /* Check if process_private released. */
1400                 if (!dev->process_private)
1401                         return 0;
1402                 mlx5_tx_uar_uninit_secondary(dev);
1403                 mlx5_proc_priv_uninit(dev);
1404                 rte_eth_dev_release_port(dev);
1405                 return 0;
1406         }
1407         if (!priv->sh)
1408                 return 0;
1409         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1410                 dev->data->port_id,
1411                 ((priv->sh->ctx != NULL) ?
1412                 mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1413         /*
1414          * If default mreg copy action is removed at the stop stage,
1415          * the search will return none and nothing will be done anymore.
1416          */
1417         mlx5_flow_stop_default(dev);
1418         mlx5_traffic_disable(dev);
1419         /*
1420          * If all the flows are already flushed in the device stop stage,
1421          * then this will return directly without any action.
1422          */
1423         mlx5_flow_list_flush(dev, &priv->flows, true);
1424         mlx5_action_handle_flush(dev);
1425         mlx5_flow_meter_flush(dev, NULL);
1426         /* Prevent crashes when queues are still in use. */
1427         dev->rx_pkt_burst = removed_rx_burst;
1428         dev->tx_pkt_burst = removed_tx_burst;
1429         rte_wmb();
1430         /* Disable datapath on secondary process. */
1431         mlx5_mp_os_req_stop_rxtx(dev);
1432         /* Free the eCPRI flex parser resource. */
1433         mlx5_flex_parser_ecpri_release(dev);
1434         if (priv->rxqs != NULL) {
1435                 /* XXX race condition if mlx5_rx_burst() is still running. */
1436                 rte_delay_us_sleep(1000);
1437                 for (i = 0; (i != priv->rxqs_n); ++i)
1438                         mlx5_rxq_release(dev, i);
1439                 priv->rxqs_n = 0;
1440                 priv->rxqs = NULL;
1441         }
1442         if (priv->txqs != NULL) {
1443                 /* XXX race condition if mlx5_tx_burst() is still running. */
1444                 rte_delay_us_sleep(1000);
1445                 for (i = 0; (i != priv->txqs_n); ++i)
1446                         mlx5_txq_release(dev, i);
1447                 priv->txqs_n = 0;
1448                 priv->txqs = NULL;
1449         }
1450         mlx5_proc_priv_uninit(dev);
1451         if (priv->q_counters) {
1452                 mlx5_devx_cmd_destroy(priv->q_counters);
1453                 priv->q_counters = NULL;
1454         }
1455         if (priv->drop_queue.hrxq)
1456                 mlx5_drop_action_destroy(dev);
1457         if (priv->mreg_cp_tbl)
1458                 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1459         mlx5_mprq_free_mp(dev);
1460         mlx5_os_free_shared_dr(priv);
1461         if (priv->rss_conf.rss_key != NULL)
1462                 mlx5_free(priv->rss_conf.rss_key);
1463         if (priv->reta_idx != NULL)
1464                 mlx5_free(priv->reta_idx);
1465         if (priv->config.vf)
1466                 mlx5_os_mac_addr_flush(dev);
1467         if (priv->nl_socket_route >= 0)
1468                 close(priv->nl_socket_route);
1469         if (priv->nl_socket_rdma >= 0)
1470                 close(priv->nl_socket_rdma);
1471         if (priv->vmwa_context)
1472                 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1473         ret = mlx5_hrxq_verify(dev);
1474         if (ret)
1475                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1476                         dev->data->port_id);
1477         ret = mlx5_ind_table_obj_verify(dev);
1478         if (ret)
1479                 DRV_LOG(WARNING, "port %u some indirection table still remain",
1480                         dev->data->port_id);
1481         ret = mlx5_rxq_obj_verify(dev);
1482         if (ret)
1483                 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1484                         dev->data->port_id);
1485         ret = mlx5_rxq_verify(dev);
1486         if (ret)
1487                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1488                         dev->data->port_id);
1489         ret = mlx5_txq_obj_verify(dev);
1490         if (ret)
1491                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1492                         dev->data->port_id);
1493         ret = mlx5_txq_verify(dev);
1494         if (ret)
1495                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1496                         dev->data->port_id);
1497         ret = mlx5_flow_verify(dev);
1498         if (ret)
1499                 DRV_LOG(WARNING, "port %u some flows still remain",
1500                         dev->data->port_id);
1501         mlx5_cache_list_destroy(&priv->hrxqs);
1502         /*
1503          * Free the shared context in last turn, because the cleanup
1504          * routines above may use some shared fields, like
1505          * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1506          * ifindex if Netlink fails.
1507          */
1508         mlx5_free_shared_dev_ctx(priv->sh);
1509         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1510                 unsigned int c = 0;
1511                 uint16_t port_id;
1512
1513                 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1514                         struct mlx5_priv *opriv =
1515                                 rte_eth_devices[port_id].data->dev_private;
1516
1517                         if (!opriv ||
1518                             opriv->domain_id != priv->domain_id ||
1519                             &rte_eth_devices[port_id] == dev)
1520                                 continue;
1521                         ++c;
1522                         break;
1523                 }
1524                 if (!c)
1525                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1526         }
1527         memset(priv, 0, sizeof(*priv));
1528         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1529         /*
1530          * Reset mac_addrs to NULL such that it is not freed as part of
1531          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1532          * it is freed when dev_private is freed.
1533          */
1534         dev->data->mac_addrs = NULL;
1535         return 0;
1536 }
1537
1538 const struct eth_dev_ops mlx5_dev_ops = {
1539         .dev_configure = mlx5_dev_configure,
1540         .dev_start = mlx5_dev_start,
1541         .dev_stop = mlx5_dev_stop,
1542         .dev_set_link_down = mlx5_set_link_down,
1543         .dev_set_link_up = mlx5_set_link_up,
1544         .dev_close = mlx5_dev_close,
1545         .promiscuous_enable = mlx5_promiscuous_enable,
1546         .promiscuous_disable = mlx5_promiscuous_disable,
1547         .allmulticast_enable = mlx5_allmulticast_enable,
1548         .allmulticast_disable = mlx5_allmulticast_disable,
1549         .link_update = mlx5_link_update,
1550         .stats_get = mlx5_stats_get,
1551         .stats_reset = mlx5_stats_reset,
1552         .xstats_get = mlx5_xstats_get,
1553         .xstats_reset = mlx5_xstats_reset,
1554         .xstats_get_names = mlx5_xstats_get_names,
1555         .fw_version_get = mlx5_fw_version_get,
1556         .dev_infos_get = mlx5_dev_infos_get,
1557         .representor_info_get = mlx5_representor_info_get,
1558         .read_clock = mlx5_txpp_read_clock,
1559         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1560         .vlan_filter_set = mlx5_vlan_filter_set,
1561         .rx_queue_setup = mlx5_rx_queue_setup,
1562         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1563         .tx_queue_setup = mlx5_tx_queue_setup,
1564         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1565         .rx_queue_release = mlx5_rx_queue_release,
1566         .tx_queue_release = mlx5_tx_queue_release,
1567         .rx_queue_start = mlx5_rx_queue_start,
1568         .rx_queue_stop = mlx5_rx_queue_stop,
1569         .tx_queue_start = mlx5_tx_queue_start,
1570         .tx_queue_stop = mlx5_tx_queue_stop,
1571         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1572         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1573         .mac_addr_remove = mlx5_mac_addr_remove,
1574         .mac_addr_add = mlx5_mac_addr_add,
1575         .mac_addr_set = mlx5_mac_addr_set,
1576         .set_mc_addr_list = mlx5_set_mc_addr_list,
1577         .mtu_set = mlx5_dev_set_mtu,
1578         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1579         .vlan_offload_set = mlx5_vlan_offload_set,
1580         .reta_update = mlx5_dev_rss_reta_update,
1581         .reta_query = mlx5_dev_rss_reta_query,
1582         .rss_hash_update = mlx5_rss_hash_update,
1583         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1584         .flow_ops_get = mlx5_flow_ops_get,
1585         .rxq_info_get = mlx5_rxq_info_get,
1586         .txq_info_get = mlx5_txq_info_get,
1587         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1588         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1589         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1590         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1591         .is_removed = mlx5_is_removed,
1592         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1593         .get_module_info = mlx5_get_module_info,
1594         .get_module_eeprom = mlx5_get_module_eeprom,
1595         .hairpin_cap_get = mlx5_hairpin_cap_get,
1596         .mtr_ops_get = mlx5_flow_meter_ops_get,
1597         .hairpin_bind = mlx5_hairpin_bind,
1598         .hairpin_unbind = mlx5_hairpin_unbind,
1599         .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1600         .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1601         .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1602         .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1603 };
1604
1605 /* Available operations from secondary process. */
1606 const struct eth_dev_ops mlx5_dev_sec_ops = {
1607         .stats_get = mlx5_stats_get,
1608         .stats_reset = mlx5_stats_reset,
1609         .xstats_get = mlx5_xstats_get,
1610         .xstats_reset = mlx5_xstats_reset,
1611         .xstats_get_names = mlx5_xstats_get_names,
1612         .fw_version_get = mlx5_fw_version_get,
1613         .dev_infos_get = mlx5_dev_infos_get,
1614         .read_clock = mlx5_txpp_read_clock,
1615         .rx_queue_start = mlx5_rx_queue_start,
1616         .rx_queue_stop = mlx5_rx_queue_stop,
1617         .tx_queue_start = mlx5_tx_queue_start,
1618         .tx_queue_stop = mlx5_tx_queue_stop,
1619         .rxq_info_get = mlx5_rxq_info_get,
1620         .txq_info_get = mlx5_txq_info_get,
1621         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1622         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1623         .get_module_info = mlx5_get_module_info,
1624         .get_module_eeprom = mlx5_get_module_eeprom,
1625 };
1626
1627 /* Available operations in flow isolated mode. */
1628 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1629         .dev_configure = mlx5_dev_configure,
1630         .dev_start = mlx5_dev_start,
1631         .dev_stop = mlx5_dev_stop,
1632         .dev_set_link_down = mlx5_set_link_down,
1633         .dev_set_link_up = mlx5_set_link_up,
1634         .dev_close = mlx5_dev_close,
1635         .promiscuous_enable = mlx5_promiscuous_enable,
1636         .promiscuous_disable = mlx5_promiscuous_disable,
1637         .allmulticast_enable = mlx5_allmulticast_enable,
1638         .allmulticast_disable = mlx5_allmulticast_disable,
1639         .link_update = mlx5_link_update,
1640         .stats_get = mlx5_stats_get,
1641         .stats_reset = mlx5_stats_reset,
1642         .xstats_get = mlx5_xstats_get,
1643         .xstats_reset = mlx5_xstats_reset,
1644         .xstats_get_names = mlx5_xstats_get_names,
1645         .fw_version_get = mlx5_fw_version_get,
1646         .dev_infos_get = mlx5_dev_infos_get,
1647         .read_clock = mlx5_txpp_read_clock,
1648         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1649         .vlan_filter_set = mlx5_vlan_filter_set,
1650         .rx_queue_setup = mlx5_rx_queue_setup,
1651         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1652         .tx_queue_setup = mlx5_tx_queue_setup,
1653         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1654         .rx_queue_release = mlx5_rx_queue_release,
1655         .tx_queue_release = mlx5_tx_queue_release,
1656         .rx_queue_start = mlx5_rx_queue_start,
1657         .rx_queue_stop = mlx5_rx_queue_stop,
1658         .tx_queue_start = mlx5_tx_queue_start,
1659         .tx_queue_stop = mlx5_tx_queue_stop,
1660         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1661         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1662         .mac_addr_remove = mlx5_mac_addr_remove,
1663         .mac_addr_add = mlx5_mac_addr_add,
1664         .mac_addr_set = mlx5_mac_addr_set,
1665         .set_mc_addr_list = mlx5_set_mc_addr_list,
1666         .mtu_set = mlx5_dev_set_mtu,
1667         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1668         .vlan_offload_set = mlx5_vlan_offload_set,
1669         .flow_ops_get = mlx5_flow_ops_get,
1670         .rxq_info_get = mlx5_rxq_info_get,
1671         .txq_info_get = mlx5_txq_info_get,
1672         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1673         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1674         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1675         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1676         .is_removed = mlx5_is_removed,
1677         .get_module_info = mlx5_get_module_info,
1678         .get_module_eeprom = mlx5_get_module_eeprom,
1679         .hairpin_cap_get = mlx5_hairpin_cap_get,
1680         .mtr_ops_get = mlx5_flow_meter_ops_get,
1681         .hairpin_bind = mlx5_hairpin_bind,
1682         .hairpin_unbind = mlx5_hairpin_unbind,
1683         .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1684         .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1685         .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1686         .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1687 };
1688
1689 /**
1690  * Verify and store value for device argument.
1691  *
1692  * @param[in] key
1693  *   Key argument to verify.
1694  * @param[in] val
1695  *   Value associated with key.
1696  * @param opaque
1697  *   User data.
1698  *
1699  * @return
1700  *   0 on success, a negative errno value otherwise and rte_errno is set.
1701  */
1702 static int
1703 mlx5_args_check(const char *key, const char *val, void *opaque)
1704 {
1705         struct mlx5_dev_config *config = opaque;
1706         unsigned long mod;
1707         signed long tmp;
1708
1709         /* No-op, port representors are processed in mlx5_dev_spawn(). */
1710         if (!strcmp(MLX5_REPRESENTOR, key))
1711                 return 0;
1712         errno = 0;
1713         tmp = strtol(val, NULL, 0);
1714         if (errno) {
1715                 rte_errno = errno;
1716                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1717                 return -rte_errno;
1718         }
1719         if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1720                 /* Negative values are acceptable for some keys only. */
1721                 rte_errno = EINVAL;
1722                 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1723                 return -rte_errno;
1724         }
1725         mod = tmp >= 0 ? tmp : -tmp;
1726         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1727                 if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1728                         DRV_LOG(ERR, "invalid CQE compression "
1729                                      "format parameter");
1730                         rte_errno = EINVAL;
1731                         return -rte_errno;
1732                 }
1733                 config->cqe_comp = !!tmp;
1734                 config->cqe_comp_fmt = tmp;
1735         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1736                 config->hw_padding = !!tmp;
1737         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1738                 config->mprq.enabled = !!tmp;
1739         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1740                 config->mprq.stride_num_n = tmp;
1741         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1742                 config->mprq.stride_size_n = tmp;
1743         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1744                 config->mprq.max_memcpy_len = tmp;
1745         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1746                 config->mprq.min_rxqs_num = tmp;
1747         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1748                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1749                                  " converted to txq_inline_max", key);
1750                 config->txq_inline_max = tmp;
1751         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1752                 config->txq_inline_max = tmp;
1753         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1754                 config->txq_inline_min = tmp;
1755         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1756                 config->txq_inline_mpw = tmp;
1757         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1758                 config->txqs_inline = tmp;
1759         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1760                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1761         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1762                 config->mps = !!tmp;
1763         } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1764                 if (tmp != MLX5_TXDB_CACHED &&
1765                     tmp != MLX5_TXDB_NCACHED &&
1766                     tmp != MLX5_TXDB_HEURISTIC) {
1767                         DRV_LOG(ERR, "invalid Tx doorbell "
1768                                      "mapping parameter");
1769                         rte_errno = EINVAL;
1770                         return -rte_errno;
1771                 }
1772                 config->dbnc = tmp;
1773         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1774                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1775         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1776                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1777                                  " converted to txq_inline_mpw", key);
1778                 config->txq_inline_mpw = tmp;
1779         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1780                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1781         } else if (strcmp(MLX5_TX_PP, key) == 0) {
1782                 if (!mod) {
1783                         DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1784                         rte_errno = EINVAL;
1785                         return -rte_errno;
1786                 }
1787                 config->tx_pp = tmp;
1788         } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1789                 config->tx_skew = tmp;
1790         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1791                 config->rx_vec_en = !!tmp;
1792         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1793                 config->l3_vxlan_en = !!tmp;
1794         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1795                 config->vf_nl_en = !!tmp;
1796         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1797                 config->dv_esw_en = !!tmp;
1798         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1799                 config->dv_flow_en = !!tmp;
1800         } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1801                 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1802                     tmp != MLX5_XMETA_MODE_META16 &&
1803                     tmp != MLX5_XMETA_MODE_META32 &&
1804                     tmp != MLX5_XMETA_MODE_MISS_INFO) {
1805                         DRV_LOG(ERR, "invalid extensive "
1806                                      "metadata parameter");
1807                         rte_errno = EINVAL;
1808                         return -rte_errno;
1809                 }
1810                 if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1811                         config->dv_xmeta_en = tmp;
1812                 else
1813                         config->dv_miss_info = 1;
1814         } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1815                 config->lacp_by_user = !!tmp;
1816         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1817                 config->mr_ext_memseg_en = !!tmp;
1818         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1819                 config->max_dump_files_num = tmp;
1820         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1821                 config->lro.timeout = tmp;
1822         } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1823                 DRV_LOG(DEBUG, "class argument is %s.", val);
1824         } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1825                 config->log_hp_size = tmp;
1826         } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1827                 if (tmp != MLX5_RCM_NONE &&
1828                     tmp != MLX5_RCM_LIGHT &&
1829                     tmp != MLX5_RCM_AGGR) {
1830                         DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1831                         rte_errno = EINVAL;
1832                         return -rte_errno;
1833                 }
1834                 config->reclaim_mode = tmp;
1835         } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1836                 config->sys_mem_en = !!tmp;
1837         } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1838                 config->decap_en = !!tmp;
1839         } else {
1840                 DRV_LOG(WARNING, "%s: unknown parameter", key);
1841                 rte_errno = EINVAL;
1842                 return -rte_errno;
1843         }
1844         return 0;
1845 }
1846
1847 /**
1848  * Parse device parameters.
1849  *
1850  * @param config
1851  *   Pointer to device configuration structure.
1852  * @param devargs
1853  *   Device arguments structure.
1854  *
1855  * @return
1856  *   0 on success, a negative errno value otherwise and rte_errno is set.
1857  */
1858 int
1859 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1860 {
1861         const char **params = (const char *[]){
1862                 MLX5_RXQ_CQE_COMP_EN,
1863                 MLX5_RXQ_PKT_PAD_EN,
1864                 MLX5_RX_MPRQ_EN,
1865                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1866                 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1867                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1868                 MLX5_RXQS_MIN_MPRQ,
1869                 MLX5_TXQ_INLINE,
1870                 MLX5_TXQ_INLINE_MIN,
1871                 MLX5_TXQ_INLINE_MAX,
1872                 MLX5_TXQ_INLINE_MPW,
1873                 MLX5_TXQS_MIN_INLINE,
1874                 MLX5_TXQS_MAX_VEC,
1875                 MLX5_TXQ_MPW_EN,
1876                 MLX5_TXQ_MPW_HDR_DSEG_EN,
1877                 MLX5_TXQ_MAX_INLINE_LEN,
1878                 MLX5_TX_DB_NC,
1879                 MLX5_TX_PP,
1880                 MLX5_TX_SKEW,
1881                 MLX5_TX_VEC_EN,
1882                 MLX5_RX_VEC_EN,
1883                 MLX5_L3_VXLAN_EN,
1884                 MLX5_VF_NL_EN,
1885                 MLX5_DV_ESW_EN,
1886                 MLX5_DV_FLOW_EN,
1887                 MLX5_DV_XMETA_EN,
1888                 MLX5_LACP_BY_USER,
1889                 MLX5_MR_EXT_MEMSEG_EN,
1890                 MLX5_REPRESENTOR,
1891                 MLX5_MAX_DUMP_FILES_NUM,
1892                 MLX5_LRO_TIMEOUT_USEC,
1893                 MLX5_CLASS_ARG_NAME,
1894                 MLX5_HP_BUF_SIZE,
1895                 MLX5_RECLAIM_MEM,
1896                 MLX5_SYS_MEM_EN,
1897                 MLX5_DECAP_EN,
1898                 NULL,
1899         };
1900         struct rte_kvargs *kvlist;
1901         int ret = 0;
1902         int i;
1903
1904         if (devargs == NULL)
1905                 return 0;
1906         /* Following UGLY cast is done to pass checkpatch. */
1907         kvlist = rte_kvargs_parse(devargs->args, params);
1908         if (kvlist == NULL) {
1909                 rte_errno = EINVAL;
1910                 return -rte_errno;
1911         }
1912         /* Process parameters. */
1913         for (i = 0; (params[i] != NULL); ++i) {
1914                 if (rte_kvargs_count(kvlist, params[i])) {
1915                         ret = rte_kvargs_process(kvlist, params[i],
1916                                                  mlx5_args_check, config);
1917                         if (ret) {
1918                                 rte_errno = EINVAL;
1919                                 rte_kvargs_free(kvlist);
1920                                 return -rte_errno;
1921                         }
1922                 }
1923         }
1924         rte_kvargs_free(kvlist);
1925         return 0;
1926 }
1927
1928 /**
1929  * Configures the minimal amount of data to inline into WQE
1930  * while sending packets.
1931  *
1932  * - the txq_inline_min has the maximal priority, if this
1933  *   key is specified in devargs
1934  * - if DevX is enabled the inline mode is queried from the
1935  *   device (HCA attributes and NIC vport context if needed).
1936  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1937  *   and none (0 bytes) for other NICs
1938  *
1939  * @param spawn
1940  *   Verbs device parameters (name, port, switch_info) to spawn.
1941  * @param config
1942  *   Device configuration parameters.
1943  */
1944 void
1945 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1946                     struct mlx5_dev_config *config)
1947 {
1948         if (config->txq_inline_min != MLX5_ARG_UNSET) {
1949                 /* Application defines size of inlined data explicitly. */
1950                 switch (spawn->pci_dev->id.device_id) {
1951                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1952                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1953                         if (config->txq_inline_min <
1954                                        (int)MLX5_INLINE_HSIZE_L2) {
1955                                 DRV_LOG(DEBUG,
1956                                         "txq_inline_mix aligned to minimal"
1957                                         " ConnectX-4 required value %d",
1958                                         (int)MLX5_INLINE_HSIZE_L2);
1959                                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1960                         }
1961                         break;
1962                 }
1963                 goto exit;
1964         }
1965         if (config->hca_attr.eth_net_offloads) {
1966                 /* We have DevX enabled, inline mode queried successfully. */
1967                 switch (config->hca_attr.wqe_inline_mode) {
1968                 case MLX5_CAP_INLINE_MODE_L2:
1969                         /* outer L2 header must be inlined. */
1970                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1971                         goto exit;
1972                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1973                         /* No inline data are required by NIC. */
1974                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1975                         config->hw_vlan_insert =
1976                                 config->hca_attr.wqe_vlan_insert;
1977                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1978                         goto exit;
1979                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1980                         /* inline mode is defined by NIC vport context. */
1981                         if (!config->hca_attr.eth_virt)
1982                                 break;
1983                         switch (config->hca_attr.vport_inline_mode) {
1984                         case MLX5_INLINE_MODE_NONE:
1985                                 config->txq_inline_min =
1986                                         MLX5_INLINE_HSIZE_NONE;
1987                                 goto exit;
1988                         case MLX5_INLINE_MODE_L2:
1989                                 config->txq_inline_min =
1990                                         MLX5_INLINE_HSIZE_L2;
1991                                 goto exit;
1992                         case MLX5_INLINE_MODE_IP:
1993                                 config->txq_inline_min =
1994                                         MLX5_INLINE_HSIZE_L3;
1995                                 goto exit;
1996                         case MLX5_INLINE_MODE_TCP_UDP:
1997                                 config->txq_inline_min =
1998                                         MLX5_INLINE_HSIZE_L4;
1999                                 goto exit;
2000                         case MLX5_INLINE_MODE_INNER_L2:
2001                                 config->txq_inline_min =
2002                                         MLX5_INLINE_HSIZE_INNER_L2;
2003                                 goto exit;
2004                         case MLX5_INLINE_MODE_INNER_IP:
2005                                 config->txq_inline_min =
2006                                         MLX5_INLINE_HSIZE_INNER_L3;
2007                                 goto exit;
2008                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
2009                                 config->txq_inline_min =
2010                                         MLX5_INLINE_HSIZE_INNER_L4;
2011                                 goto exit;
2012                         }
2013                 }
2014         }
2015         /*
2016          * We get here if we are unable to deduce
2017          * inline data size with DevX. Try PCI ID
2018          * to determine old NICs.
2019          */
2020         switch (spawn->pci_dev->id.device_id) {
2021         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2022         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2023         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2024         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2025                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2026                 config->hw_vlan_insert = 0;
2027                 break;
2028         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2029         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2030         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2031         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2032                 /*
2033                  * These NICs support VLAN insertion from WQE and
2034                  * report the wqe_vlan_insert flag. But there is the bug
2035                  * and PFC control may be broken, so disable feature.
2036                  */
2037                 config->hw_vlan_insert = 0;
2038                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2039                 break;
2040         default:
2041                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2042                 break;
2043         }
2044 exit:
2045         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2046 }
2047
2048 /**
2049  * Configures the metadata mask fields in the shared context.
2050  *
2051  * @param [in] dev
2052  *   Pointer to Ethernet device.
2053  */
2054 void
2055 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2056 {
2057         struct mlx5_priv *priv = dev->data->dev_private;
2058         struct mlx5_dev_ctx_shared *sh = priv->sh;
2059         uint32_t meta, mark, reg_c0;
2060
2061         reg_c0 = ~priv->vport_meta_mask;
2062         switch (priv->config.dv_xmeta_en) {
2063         case MLX5_XMETA_MODE_LEGACY:
2064                 meta = UINT32_MAX;
2065                 mark = MLX5_FLOW_MARK_MASK;
2066                 break;
2067         case MLX5_XMETA_MODE_META16:
2068                 meta = reg_c0 >> rte_bsf32(reg_c0);
2069                 mark = MLX5_FLOW_MARK_MASK;
2070                 break;
2071         case MLX5_XMETA_MODE_META32:
2072                 meta = UINT32_MAX;
2073                 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2074                 break;
2075         default:
2076                 meta = 0;
2077                 mark = 0;
2078                 MLX5_ASSERT(false);
2079                 break;
2080         }
2081         if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2082                 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2083                                  sh->dv_mark_mask, mark);
2084         else
2085                 sh->dv_mark_mask = mark;
2086         if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2087                 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2088                                  sh->dv_meta_mask, meta);
2089         else
2090                 sh->dv_meta_mask = meta;
2091         if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2092                 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2093                                  sh->dv_meta_mask, reg_c0);
2094         else
2095                 sh->dv_regc0_mask = reg_c0;
2096         DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2097         DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2098         DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2099         DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2100 }
2101
2102 int
2103 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2104 {
2105         static const char *const dynf_names[] = {
2106                 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2107                 RTE_MBUF_DYNFLAG_METADATA_NAME,
2108                 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2109         };
2110         unsigned int i;
2111
2112         if (n < RTE_DIM(dynf_names))
2113                 return -ENOMEM;
2114         for (i = 0; i < RTE_DIM(dynf_names); i++) {
2115                 if (names[i] == NULL)
2116                         return -EINVAL;
2117                 strcpy(names[i], dynf_names[i]);
2118         }
2119         return RTE_DIM(dynf_names);
2120 }
2121
2122 /**
2123  * Comparison callback to sort device data.
2124  *
2125  * This is meant to be used with qsort().
2126  *
2127  * @param a[in]
2128  *   Pointer to pointer to first data object.
2129  * @param b[in]
2130  *   Pointer to pointer to second data object.
2131  *
2132  * @return
2133  *   0 if both objects are equal, less than 0 if the first argument is less
2134  *   than the second, greater than 0 otherwise.
2135  */
2136 int
2137 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2138                               struct mlx5_dev_config *config)
2139 {
2140         struct mlx5_dev_ctx_shared *sh = priv->sh;
2141         struct mlx5_dev_config *sh_conf = NULL;
2142         uint16_t port_id;
2143
2144         MLX5_ASSERT(sh);
2145         /* Nothing to compare for the single/first device. */
2146         if (sh->refcnt == 1)
2147                 return 0;
2148         /* Find the device with shared context. */
2149         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2150                 struct mlx5_priv *opriv =
2151                         rte_eth_devices[port_id].data->dev_private;
2152
2153                 if (opriv && opriv != priv && opriv->sh == sh) {
2154                         sh_conf = &opriv->config;
2155                         break;
2156                 }
2157         }
2158         if (!sh_conf)
2159                 return 0;
2160         if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2161                 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2162                              " for shared %s context", sh->ibdev_name);
2163                 rte_errno = EINVAL;
2164                 return rte_errno;
2165         }
2166         if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2167                 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2168                              " for shared %s context", sh->ibdev_name);
2169                 rte_errno = EINVAL;
2170                 return rte_errno;
2171         }
2172         return 0;
2173 }
2174
2175 /**
2176  * Look for the ethernet device belonging to mlx5 driver.
2177  *
2178  * @param[in] port_id
2179  *   port_id to start looking for device.
2180  * @param[in] pci_dev
2181  *   Pointer to the hint PCI device. When device is being probed
2182  *   the its siblings (master and preceding representors might
2183  *   not have assigned driver yet (because the mlx5_os_pci_probe()
2184  *   is not completed yet, for this case match on hint PCI
2185  *   device may be used to detect sibling device.
2186  *
2187  * @return
2188  *   port_id of found device, RTE_MAX_ETHPORT if not found.
2189  */
2190 uint16_t
2191 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
2192 {
2193         while (port_id < RTE_MAX_ETHPORTS) {
2194                 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2195
2196                 if (dev->state != RTE_ETH_DEV_UNUSED &&
2197                     dev->device &&
2198                     (dev->device == &pci_dev->device ||
2199                      (dev->device->driver &&
2200                      dev->device->driver->name &&
2201                      !strcmp(dev->device->driver->name, MLX5_PCI_DRIVER_NAME))))
2202                         break;
2203                 port_id++;
2204         }
2205         if (port_id >= RTE_MAX_ETHPORTS)
2206                 return RTE_MAX_ETHPORTS;
2207         return port_id;
2208 }
2209
2210 /**
2211  * DPDK callback to remove a PCI device.
2212  *
2213  * This function removes all Ethernet devices belong to a given PCI device.
2214  *
2215  * @param[in] pci_dev
2216  *   Pointer to the PCI device.
2217  *
2218  * @return
2219  *   0 on success, the function cannot fail.
2220  */
2221 static int
2222 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2223 {
2224         uint16_t port_id;
2225         int ret = 0;
2226
2227         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
2228                 /*
2229                  * mlx5_dev_close() is not registered to secondary process,
2230                  * call the close function explicitly for secondary process.
2231                  */
2232                 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2233                         ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2234                 else
2235                         ret |= rte_eth_dev_close(port_id);
2236         }
2237         return ret == 0 ? 0 : -EIO;
2238 }
2239
2240 static const struct rte_pci_id mlx5_pci_id_map[] = {
2241         {
2242                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2243                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2244         },
2245         {
2246                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2247                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2248         },
2249         {
2250                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2251                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2252         },
2253         {
2254                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2255                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2256         },
2257         {
2258                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2259                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2260         },
2261         {
2262                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2263                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2264         },
2265         {
2266                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2267                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2268         },
2269         {
2270                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2271                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2272         },
2273         {
2274                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2275                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2276         },
2277         {
2278                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2279                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2280         },
2281         {
2282                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2283                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2284         },
2285         {
2286                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2287                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2288         },
2289         {
2290                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2291                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2292         },
2293         {
2294                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2295                                 PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2296         },
2297         {
2298                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2299                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2300         },
2301         {
2302                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2303                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2304         },
2305         {
2306                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2307                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2308         },
2309         {
2310                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2311                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2312         },
2313         {
2314                 .vendor_id = 0
2315         }
2316 };
2317
2318 static struct mlx5_pci_driver mlx5_driver = {
2319         .driver_class = MLX5_CLASS_NET,
2320         .pci_driver = {
2321                 .driver = {
2322                         .name = MLX5_PCI_DRIVER_NAME,
2323                 },
2324                 .id_table = mlx5_pci_id_map,
2325                 .probe = mlx5_os_pci_probe,
2326                 .remove = mlx5_pci_remove,
2327                 .dma_map = mlx5_dma_map,
2328                 .dma_unmap = mlx5_dma_unmap,
2329                 .drv_flags = PCI_DRV_FLAGS,
2330         },
2331 };
2332
2333 /* Initialize driver log type. */
2334 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
2335
2336 /**
2337  * Driver initialization routine.
2338  */
2339 RTE_INIT(rte_mlx5_pmd_init)
2340 {
2341         pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2342         mlx5_common_init();
2343         /* Build the static tables for Verbs conversion. */
2344         mlx5_set_ptype_table();
2345         mlx5_set_cksum_table();
2346         mlx5_set_swp_types_table();
2347         if (mlx5_glue)
2348                 mlx5_pci_driver_register(&mlx5_driver);
2349 }
2350
2351 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2352 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2353 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");