net/mlx5: optimize hash list entry memory
[dpdk.git] / drivers / net / mlx5 / mlx5.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12
13 #include <rte_malloc.h>
14 #include <rte_ethdev_driver.h>
15 #include <rte_ethdev_pci.h>
16 #include <rte_pci.h>
17 #include <rte_bus_pci.h>
18 #include <rte_common.h>
19 #include <rte_kvargs.h>
20 #include <rte_rwlock.h>
21 #include <rte_spinlock.h>
22 #include <rte_string_fns.h>
23 #include <rte_alarm.h>
24
25 #include <mlx5_glue.h>
26 #include <mlx5_devx_cmds.h>
27 #include <mlx5_common.h>
28 #include <mlx5_common_os.h>
29 #include <mlx5_common_mp.h>
30 #include <mlx5_common_pci.h>
31 #include <mlx5_malloc.h>
32
33 #include "mlx5_defs.h"
34 #include "mlx5.h"
35 #include "mlx5_utils.h"
36 #include "mlx5_rxtx.h"
37 #include "mlx5_autoconf.h"
38 #include "mlx5_mr.h"
39 #include "mlx5_flow.h"
40 #include "rte_pmd_mlx5.h"
41
42 /* Device parameter to enable RX completion queue compression. */
43 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
44
45 /* Device parameter to enable RX completion entry padding to 128B. */
46 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
47
48 /* Device parameter to enable padding Rx packet to cacheline size. */
49 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
50
51 /* Device parameter to enable Multi-Packet Rx queue. */
52 #define MLX5_RX_MPRQ_EN "mprq_en"
53
54 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
55 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
56
57 /* Device parameter to configure log 2 of the stride size for MPRQ. */
58 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
59
60 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
61 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
62
63 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
64 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
65
66 /* Device parameter to configure inline send. Deprecated, ignored.*/
67 #define MLX5_TXQ_INLINE "txq_inline"
68
69 /* Device parameter to limit packet size to inline with ordinary SEND. */
70 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
71
72 /* Device parameter to configure minimal data size to inline. */
73 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
74
75 /* Device parameter to limit packet size to inline with Enhanced MPW. */
76 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
77
78 /*
79  * Device parameter to configure the number of TX queues threshold for
80  * enabling inline send.
81  */
82 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
83
84 /*
85  * Device parameter to configure the number of TX queues threshold for
86  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
87  */
88 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
89
90 /* Device parameter to enable multi-packet send WQEs. */
91 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
92
93 /*
94  * Device parameter to force doorbell register mapping
95  * to non-cahed region eliminating the extra write memory barrier.
96  */
97 #define MLX5_TX_DB_NC "tx_db_nc"
98
99 /*
100  * Device parameter to include 2 dsegs in the title WQEBB.
101  * Deprecated, ignored.
102  */
103 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
104
105 /*
106  * Device parameter to limit the size of inlining packet.
107  * Deprecated, ignored.
108  */
109 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
110
111 /*
112  * Device parameter to enable Tx scheduling on timestamps
113  * and specify the packet pacing granularity in nanoseconds.
114  */
115 #define MLX5_TX_PP "tx_pp"
116
117 /*
118  * Device parameter to specify skew in nanoseconds on Tx datapath,
119  * it represents the time between SQ start WQE processing and
120  * appearing actual packet data on the wire.
121  */
122 #define MLX5_TX_SKEW "tx_skew"
123
124 /*
125  * Device parameter to enable hardware Tx vector.
126  * Deprecated, ignored (no vectorized Tx routines anymore).
127  */
128 #define MLX5_TX_VEC_EN "tx_vec_en"
129
130 /* Device parameter to enable hardware Rx vector. */
131 #define MLX5_RX_VEC_EN "rx_vec_en"
132
133 /* Allow L3 VXLAN flow creation. */
134 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
135
136 /* Activate DV E-Switch flow steering. */
137 #define MLX5_DV_ESW_EN "dv_esw_en"
138
139 /* Activate DV flow steering. */
140 #define MLX5_DV_FLOW_EN "dv_flow_en"
141
142 /* Enable extensive flow metadata support. */
143 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
144
145 /* Device parameter to let the user manage the lacp traffic of bonded device */
146 #define MLX5_LACP_BY_USER "lacp_by_user"
147
148 /* Activate Netlink support in VF mode. */
149 #define MLX5_VF_NL_EN "vf_nl_en"
150
151 /* Enable extending memsegs when creating a MR. */
152 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
153
154 /* Select port representors to instantiate. */
155 #define MLX5_REPRESENTOR "representor"
156
157 /* Device parameter to configure the maximum number of dump files per queue. */
158 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
159
160 /* Configure timeout of LRO session (in microseconds). */
161 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
162
163 /*
164  * Device parameter to configure the total data buffer size for a single
165  * hairpin queue (logarithm value).
166  */
167 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
168
169 /* Flow memory reclaim mode. */
170 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
171
172 /* The default memory allocator used in PMD. */
173 #define MLX5_SYS_MEM_EN "sys_mem_en"
174 /* Decap will be used or not. */
175 #define MLX5_DECAP_EN "decap_en"
176
177 /* Shared memory between primary and secondary processes. */
178 struct mlx5_shared_data *mlx5_shared_data;
179
180 /** Driver-specific log messages type. */
181 int mlx5_logtype;
182
183 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
184                                                 LIST_HEAD_INITIALIZER();
185 static pthread_mutex_t mlx5_dev_ctx_list_mutex = PTHREAD_MUTEX_INITIALIZER;
186
187 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
188 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
189         [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                 .need_lock = 1,
269                 .release_mem_en = 1,
270                 .type = "mlx5_tunnel_offload",
271         },
272         [MLX5_IPOOL_TNL_TBL_ID] = {
273                 .size = 0,
274                 .need_lock = 1,
275                 .type = "mlx5_flow_tnl_tbl_ipool",
276         },
277 #endif
278         [MLX5_IPOOL_MTR] = {
279                 .size = sizeof(struct mlx5_flow_meter),
280                 .trunk_size = 64,
281                 .grow_trunk = 3,
282                 .grow_shift = 2,
283                 .need_lock = 1,
284                 .release_mem_en = 1,
285                 .malloc = mlx5_malloc,
286                 .free = mlx5_free,
287                 .type = "mlx5_meter_ipool",
288         },
289         [MLX5_IPOOL_MCP] = {
290                 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
291                 .trunk_size = 64,
292                 .grow_trunk = 3,
293                 .grow_shift = 2,
294                 .need_lock = 1,
295                 .release_mem_en = 1,
296                 .malloc = mlx5_malloc,
297                 .free = mlx5_free,
298                 .type = "mlx5_mcp_ipool",
299         },
300         [MLX5_IPOOL_HRXQ] = {
301                 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
302                 .trunk_size = 64,
303                 .grow_trunk = 3,
304                 .grow_shift = 2,
305                 .need_lock = 1,
306                 .release_mem_en = 1,
307                 .malloc = mlx5_malloc,
308                 .free = mlx5_free,
309                 .type = "mlx5_hrxq_ipool",
310         },
311         [MLX5_IPOOL_MLX5_FLOW] = {
312                 /*
313                  * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
314                  * It set in run time according to PCI function configuration.
315                  */
316                 .size = 0,
317                 .trunk_size = 64,
318                 .grow_trunk = 3,
319                 .grow_shift = 2,
320                 .need_lock = 1,
321                 .release_mem_en = 1,
322                 .malloc = mlx5_malloc,
323                 .free = mlx5_free,
324                 .type = "mlx5_flow_handle_ipool",
325         },
326         [MLX5_IPOOL_RTE_FLOW] = {
327                 .size = sizeof(struct rte_flow),
328                 .trunk_size = 4096,
329                 .need_lock = 1,
330                 .release_mem_en = 1,
331                 .malloc = mlx5_malloc,
332                 .free = mlx5_free,
333                 .type = "rte_flow_ipool",
334         },
335         [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
336                 .size = 0,
337                 .need_lock = 1,
338                 .type = "mlx5_flow_rss_id_ipool",
339         },
340         [MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
341                 .size = sizeof(struct mlx5_shared_action_rss),
342                 .trunk_size = 64,
343                 .grow_trunk = 3,
344                 .grow_shift = 2,
345                 .need_lock = 1,
346                 .release_mem_en = 1,
347                 .malloc = mlx5_malloc,
348                 .free = mlx5_free,
349                 .type = "mlx5_shared_action_rss",
350         },
351 };
352
353
354 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
355 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
356
357 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
358
359 /**
360  * Initialize the ASO aging management structure.
361  *
362  * @param[in] sh
363  *   Pointer to mlx5_dev_ctx_shared object to free
364  *
365  * @return
366  *   0 on success, a negative errno value otherwise and rte_errno is set.
367  */
368 int
369 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
370 {
371         int err;
372
373         if (sh->aso_age_mng)
374                 return 0;
375         sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
376                                       RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
377         if (!sh->aso_age_mng) {
378                 DRV_LOG(ERR, "aso_age_mng allocation was failed.");
379                 rte_errno = ENOMEM;
380                 return -ENOMEM;
381         }
382         err = mlx5_aso_queue_init(sh);
383         if (err) {
384                 mlx5_free(sh->aso_age_mng);
385                 return -1;
386         }
387         rte_spinlock_init(&sh->aso_age_mng->resize_sl);
388         rte_spinlock_init(&sh->aso_age_mng->free_sl);
389         LIST_INIT(&sh->aso_age_mng->free);
390         return 0;
391 }
392
393 /**
394  * Close and release all the resources of the ASO aging management structure.
395  *
396  * @param[in] sh
397  *   Pointer to mlx5_dev_ctx_shared object to free.
398  */
399 static void
400 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
401 {
402         int i, j;
403
404         mlx5_aso_queue_stop(sh);
405         mlx5_aso_queue_uninit(sh);
406         if (sh->aso_age_mng->pools) {
407                 struct mlx5_aso_age_pool *pool;
408
409                 for (i = 0; i < sh->aso_age_mng->next; ++i) {
410                         pool = sh->aso_age_mng->pools[i];
411                         claim_zero(mlx5_devx_cmd_destroy
412                                                 (pool->flow_hit_aso_obj));
413                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
414                                 if (pool->actions[j].dr_action)
415                                         claim_zero
416                                                 (mlx5_glue->destroy_flow_action
417                                                   (pool->actions[j].dr_action));
418                         mlx5_free(pool);
419                 }
420                 mlx5_free(sh->aso_age_mng->pools);
421         }
422         mlx5_free(sh->aso_age_mng);
423 }
424
425 /**
426  * Initialize the shared aging list information per port.
427  *
428  * @param[in] sh
429  *   Pointer to mlx5_dev_ctx_shared object.
430  */
431 static void
432 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
433 {
434         uint32_t i;
435         struct mlx5_age_info *age_info;
436
437         for (i = 0; i < sh->max_port; i++) {
438                 age_info = &sh->port[i].age_info;
439                 age_info->flags = 0;
440                 TAILQ_INIT(&age_info->aged_counters);
441                 LIST_INIT(&age_info->aged_aso);
442                 rte_spinlock_init(&age_info->aged_sl);
443                 MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
444         }
445 }
446
447 /**
448  * Initialize the counters management structure.
449  *
450  * @param[in] sh
451  *   Pointer to mlx5_dev_ctx_shared object to free
452  */
453 static void
454 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
455 {
456         int i;
457
458         memset(&sh->cmng, 0, sizeof(sh->cmng));
459         TAILQ_INIT(&sh->cmng.flow_counters);
460         sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
461         sh->cmng.max_id = -1;
462         sh->cmng.last_pool_idx = POOL_IDX_INVALID;
463         rte_spinlock_init(&sh->cmng.pool_update_sl);
464         for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
465                 TAILQ_INIT(&sh->cmng.counters[i]);
466                 rte_spinlock_init(&sh->cmng.csl[i]);
467         }
468 }
469
470 /**
471  * Destroy all the resources allocated for a counter memory management.
472  *
473  * @param[in] mng
474  *   Pointer to the memory management structure.
475  */
476 static void
477 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
478 {
479         uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
480
481         LIST_REMOVE(mng, next);
482         claim_zero(mlx5_devx_cmd_destroy(mng->dm));
483         claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
484         mlx5_free(mem);
485 }
486
487 /**
488  * Close and release all the resources of the counters management.
489  *
490  * @param[in] sh
491  *   Pointer to mlx5_dev_ctx_shared object to free.
492  */
493 static void
494 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
495 {
496         struct mlx5_counter_stats_mem_mng *mng;
497         int i, j;
498         int retries = 1024;
499
500         rte_errno = 0;
501         while (--retries) {
502                 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
503                 if (rte_errno != EINPROGRESS)
504                         break;
505                 rte_pause();
506         }
507
508         if (sh->cmng.pools) {
509                 struct mlx5_flow_counter_pool *pool;
510                 uint16_t n_valid = sh->cmng.n_valid;
511                 bool fallback = sh->cmng.counter_fallback;
512
513                 for (i = 0; i < n_valid; ++i) {
514                         pool = sh->cmng.pools[i];
515                         if (!fallback && pool->min_dcs)
516                                 claim_zero(mlx5_devx_cmd_destroy
517                                                                (pool->min_dcs));
518                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
519                                 struct mlx5_flow_counter *cnt =
520                                                 MLX5_POOL_GET_CNT(pool, j);
521
522                                 if (cnt->action)
523                                         claim_zero
524                                          (mlx5_glue->destroy_flow_action
525                                           (cnt->action));
526                                 if (fallback && MLX5_POOL_GET_CNT
527                                     (pool, j)->dcs_when_free)
528                                         claim_zero(mlx5_devx_cmd_destroy
529                                                    (cnt->dcs_when_free));
530                         }
531                         mlx5_free(pool);
532                 }
533                 mlx5_free(sh->cmng.pools);
534         }
535         mng = LIST_FIRST(&sh->cmng.mem_mngs);
536         while (mng) {
537                 mlx5_flow_destroy_counter_stat_mem_mng(mng);
538                 mng = LIST_FIRST(&sh->cmng.mem_mngs);
539         }
540         memset(&sh->cmng, 0, sizeof(sh->cmng));
541 }
542
543 /* Send FLOW_AGED event if needed. */
544 void
545 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
546 {
547         struct mlx5_age_info *age_info;
548         uint32_t i;
549
550         for (i = 0; i < sh->max_port; i++) {
551                 age_info = &sh->port[i].age_info;
552                 if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
553                         continue;
554                 if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER))
555                         rte_eth_dev_callback_process
556                                 (&rte_eth_devices[sh->port[i].devx_ih_port_id],
557                                 RTE_ETH_EVENT_FLOW_AGED, NULL);
558                 age_info->flags = 0;
559         }
560 }
561
562 /**
563  * Initialize the flow resources' indexed mempool.
564  *
565  * @param[in] sh
566  *   Pointer to mlx5_dev_ctx_shared object.
567  * @param[in] sh
568  *   Pointer to user dev config.
569  */
570 static void
571 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
572                        const struct mlx5_dev_config *config)
573 {
574         uint8_t i;
575         struct mlx5_indexed_pool_config cfg;
576
577         for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
578                 cfg = mlx5_ipool_cfg[i];
579                 switch (i) {
580                 default:
581                         break;
582                 /*
583                  * Set MLX5_IPOOL_MLX5_FLOW ipool size
584                  * according to PCI function flow configuration.
585                  */
586                 case MLX5_IPOOL_MLX5_FLOW:
587                         cfg.size = config->dv_flow_en ?
588                                 sizeof(struct mlx5_flow_handle) :
589                                 MLX5_FLOW_HANDLE_VERBS_SIZE;
590                         break;
591                 }
592                 if (config->reclaim_mode)
593                         cfg.release_mem_en = 1;
594                 sh->ipool[i] = mlx5_ipool_create(&cfg);
595         }
596 }
597
598 /**
599  * Release the flow resources' indexed mempool.
600  *
601  * @param[in] sh
602  *   Pointer to mlx5_dev_ctx_shared object.
603  */
604 static void
605 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
606 {
607         uint8_t i;
608
609         for (i = 0; i < MLX5_IPOOL_MAX; ++i)
610                 mlx5_ipool_destroy(sh->ipool[i]);
611 }
612
613 /*
614  * Check if dynamic flex parser for eCPRI already exists.
615  *
616  * @param dev
617  *   Pointer to Ethernet device structure.
618  *
619  * @return
620  *   true on exists, false on not.
621  */
622 bool
623 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
624 {
625         struct mlx5_priv *priv = dev->data->dev_private;
626         struct mlx5_flex_parser_profiles *prf =
627                                 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
628
629         return !!prf->obj;
630 }
631
632 /*
633  * Allocation of a flex parser for eCPRI. Once created, this parser related
634  * resources will be held until the device is closed.
635  *
636  * @param dev
637  *   Pointer to Ethernet device structure.
638  *
639  * @return
640  *   0 on success, a negative errno value otherwise and rte_errno is set.
641  */
642 int
643 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
644 {
645         struct mlx5_priv *priv = dev->data->dev_private;
646         struct mlx5_flex_parser_profiles *prf =
647                                 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
648         struct mlx5_devx_graph_node_attr node = {
649                 .modify_field_select = 0,
650         };
651         uint32_t ids[8];
652         int ret;
653
654         if (!priv->config.hca_attr.parse_graph_flex_node) {
655                 DRV_LOG(ERR, "Dynamic flex parser is not supported "
656                         "for device %s.", priv->dev_data->name);
657                 return -ENOTSUP;
658         }
659         node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
660         /* 8 bytes now: 4B common header + 4B message body header. */
661         node.header_length_base_value = 0x8;
662         /* After MAC layer: Ether / VLAN. */
663         node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
664         /* Type of compared condition should be 0xAEFE in the L2 layer. */
665         node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
666         /* Sample #0: type in common header. */
667         node.sample[0].flow_match_sample_en = 1;
668         /* Fixed offset. */
669         node.sample[0].flow_match_sample_offset_mode = 0x0;
670         /* Only the 2nd byte will be used. */
671         node.sample[0].flow_match_sample_field_base_offset = 0x0;
672         /* Sample #1: message payload. */
673         node.sample[1].flow_match_sample_en = 1;
674         /* Fixed offset. */
675         node.sample[1].flow_match_sample_offset_mode = 0x0;
676         /*
677          * Only the first two bytes will be used right now, and its offset will
678          * start after the common header that with the length of a DW(u32).
679          */
680         node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
681         prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
682         if (!prf->obj) {
683                 DRV_LOG(ERR, "Failed to create flex parser node object.");
684                 return (rte_errno == 0) ? -ENODEV : -rte_errno;
685         }
686         prf->num = 2;
687         ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
688         if (ret) {
689                 DRV_LOG(ERR, "Failed to query sample IDs.");
690                 return (rte_errno == 0) ? -ENODEV : -rte_errno;
691         }
692         prf->offset[0] = 0x0;
693         prf->offset[1] = sizeof(uint32_t);
694         prf->ids[0] = ids[0];
695         prf->ids[1] = ids[1];
696         return 0;
697 }
698
699 /*
700  * Destroy the flex parser node, including the parser itself, input / output
701  * arcs and DW samples. Resources could be reused then.
702  *
703  * @param dev
704  *   Pointer to Ethernet device structure.
705  */
706 static void
707 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
708 {
709         struct mlx5_priv *priv = dev->data->dev_private;
710         struct mlx5_flex_parser_profiles *prf =
711                                 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
712
713         if (prf->obj)
714                 mlx5_devx_cmd_destroy(prf->obj);
715         prf->obj = NULL;
716 }
717
718 /*
719  * Allocate Rx and Tx UARs in robust fashion.
720  * This routine handles the following UAR allocation issues:
721  *
722  *  - tries to allocate the UAR with the most appropriate memory
723  *    mapping type from the ones supported by the host
724  *
725  *  - tries to allocate the UAR with non-NULL base address
726  *    OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
727  *    UAR base address if UAR was not the first object in the UAR page.
728  *    It caused the PMD failure and we should try to get another UAR
729  *    till we get the first one with non-NULL base address returned.
730  */
731 static int
732 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
733                      const struct mlx5_dev_config *config)
734 {
735         uint32_t uar_mapping, retry;
736         int err = 0;
737         void *base_addr;
738
739         for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
740 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
741                 /* Control the mapping type according to the settings. */
742                 uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
743                               MLX5DV_UAR_ALLOC_TYPE_NC :
744                               MLX5DV_UAR_ALLOC_TYPE_BF;
745 #else
746                 RTE_SET_USED(config);
747                 /*
748                  * It seems we have no way to control the memory mapping type
749                  * for the UAR, the default "Write-Combining" type is supposed.
750                  * The UAR initialization on queue creation queries the
751                  * actual mapping type done by Verbs/kernel and setups the
752                  * PMD datapath accordingly.
753                  */
754                 uar_mapping = 0;
755 #endif
756                 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
757 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
758                 if (!sh->tx_uar &&
759                     uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
760                         if (config->dbnc == MLX5_TXDB_CACHED ||
761                             config->dbnc == MLX5_TXDB_HEURISTIC)
762                                 DRV_LOG(WARNING, "Devarg tx_db_nc setting "
763                                                  "is not supported by DevX");
764                         /*
765                          * In some environments like virtual machine
766                          * the Write Combining mapped might be not supported
767                          * and UAR allocation fails. We try "Non-Cached"
768                          * mapping for the case. The tx_burst routines take
769                          * the UAR mapping type into account on UAR setup
770                          * on queue creation.
771                          */
772                         DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (BF)");
773                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
774                         sh->tx_uar = mlx5_glue->devx_alloc_uar
775                                                         (sh->ctx, uar_mapping);
776                 } else if (!sh->tx_uar &&
777                            uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
778                         if (config->dbnc == MLX5_TXDB_NCACHED)
779                                 DRV_LOG(WARNING, "Devarg tx_db_nc settings "
780                                                  "is not supported by DevX");
781                         /*
782                          * If Verbs/kernel does not support "Non-Cached"
783                          * try the "Write-Combining".
784                          */
785                         DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (NC)");
786                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
787                         sh->tx_uar = mlx5_glue->devx_alloc_uar
788                                                         (sh->ctx, uar_mapping);
789                 }
790 #endif
791                 if (!sh->tx_uar) {
792                         DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
793                         err = ENOMEM;
794                         goto exit;
795                 }
796                 base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
797                 if (base_addr)
798                         break;
799                 /*
800                  * The UARs are allocated by rdma_core within the
801                  * IB device context, on context closure all UARs
802                  * will be freed, should be no memory/object leakage.
803                  */
804                 DRV_LOG(WARNING, "Retrying to allocate Tx DevX UAR");
805                 sh->tx_uar = NULL;
806         }
807         /* Check whether we finally succeeded with valid UAR allocation. */
808         if (!sh->tx_uar) {
809                 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
810                 err = ENOMEM;
811                 goto exit;
812         }
813         for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
814                 uar_mapping = 0;
815                 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
816                                                         (sh->ctx, uar_mapping);
817 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
818                 if (!sh->devx_rx_uar &&
819                     uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
820                         /*
821                          * Rx UAR is used to control interrupts only,
822                          * should be no datapath noticeable impact,
823                          * can try "Non-Cached" mapping safely.
824                          */
825                         DRV_LOG(WARNING, "Failed to allocate Rx DevX UAR (BF)");
826                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
827                         sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
828                                                         (sh->ctx, uar_mapping);
829                 }
830 #endif
831                 if (!sh->devx_rx_uar) {
832                         DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
833                         err = ENOMEM;
834                         goto exit;
835                 }
836                 base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
837                 if (base_addr)
838                         break;
839                 /*
840                  * The UARs are allocated by rdma_core within the
841                  * IB device context, on context closure all UARs
842                  * will be freed, should be no memory/object leakage.
843                  */
844                 DRV_LOG(WARNING, "Retrying to allocate Rx DevX UAR");
845                 sh->devx_rx_uar = NULL;
846         }
847         /* Check whether we finally succeeded with valid UAR allocation. */
848         if (!sh->devx_rx_uar) {
849                 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
850                 err = ENOMEM;
851         }
852 exit:
853         return err;
854 }
855
856 /**
857  * Allocate shared device context. If there is multiport device the
858  * master and representors will share this context, if there is single
859  * port dedicated device, the context will be used by only given
860  * port due to unification.
861  *
862  * Routine first searches the context for the specified device name,
863  * if found the shared context assumed and reference counter is incremented.
864  * If no context found the new one is created and initialized with specified
865  * device context and parameters.
866  *
867  * @param[in] spawn
868  *   Pointer to the device attributes (name, port, etc).
869  * @param[in] config
870  *   Pointer to device configuration structure.
871  *
872  * @return
873  *   Pointer to mlx5_dev_ctx_shared object on success,
874  *   otherwise NULL and rte_errno is set.
875  */
876 struct mlx5_dev_ctx_shared *
877 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
878                            const struct mlx5_dev_config *config)
879 {
880         struct mlx5_dev_ctx_shared *sh;
881         int err = 0;
882         uint32_t i;
883         struct mlx5_devx_tis_attr tis_attr = { 0 };
884
885         MLX5_ASSERT(spawn);
886         /* Secondary process should not create the shared context. */
887         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
888         pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
889         /* Search for IB context by device name. */
890         LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
891                 if (!strcmp(sh->ibdev_name,
892                         mlx5_os_get_dev_device_name(spawn->phys_dev))) {
893                         sh->refcnt++;
894                         goto exit;
895                 }
896         }
897         /* No device found, we have to create new shared context. */
898         MLX5_ASSERT(spawn->max_port);
899         sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
900                          sizeof(struct mlx5_dev_ctx_shared) +
901                          spawn->max_port *
902                          sizeof(struct mlx5_dev_shared_port),
903                          RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
904         if (!sh) {
905                 DRV_LOG(ERR, "shared context allocation failure");
906                 rte_errno  = ENOMEM;
907                 goto exit;
908         }
909         err = mlx5_os_open_device(spawn, config, sh);
910         if (!sh->ctx)
911                 goto error;
912         err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
913         if (err) {
914                 DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
915                 goto error;
916         }
917         sh->refcnt = 1;
918         sh->bond_dev = UINT16_MAX;
919         sh->max_port = spawn->max_port;
920         strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
921                 sizeof(sh->ibdev_name) - 1);
922         strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
923                 sizeof(sh->ibdev_path) - 1);
924         /*
925          * Setting port_id to max unallowed value means
926          * there is no interrupt subhandler installed for
927          * the given port index i.
928          */
929         for (i = 0; i < sh->max_port; i++) {
930                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
931                 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
932         }
933         sh->pd = mlx5_glue->alloc_pd(sh->ctx);
934         if (sh->pd == NULL) {
935                 DRV_LOG(ERR, "PD allocation failure");
936                 err = ENOMEM;
937                 goto error;
938         }
939         if (sh->devx) {
940                 /* Query the EQN for this core. */
941                 err = mlx5_glue->devx_query_eqn(sh->ctx, 0, &sh->eqn);
942                 if (err) {
943                         rte_errno = errno;
944                         DRV_LOG(ERR, "Failed to query event queue number %d.",
945                                 rte_errno);
946                         goto error;
947                 }
948                 err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
949                 if (err) {
950                         DRV_LOG(ERR, "Fail to extract pdn from PD");
951                         goto error;
952                 }
953                 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
954                 if (!sh->td) {
955                         DRV_LOG(ERR, "TD allocation failure");
956                         err = ENOMEM;
957                         goto error;
958                 }
959                 tis_attr.transport_domain = sh->td->id;
960                 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
961                 if (!sh->tis) {
962                         DRV_LOG(ERR, "TIS allocation failure");
963                         err = ENOMEM;
964                         goto error;
965                 }
966                 err = mlx5_alloc_rxtx_uars(sh, config);
967                 if (err)
968                         goto error;
969                 MLX5_ASSERT(sh->tx_uar);
970                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
971
972                 MLX5_ASSERT(sh->devx_rx_uar);
973                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
974         }
975 #ifndef RTE_ARCH_64
976         /* Initialize UAR access locks for 32bit implementations. */
977         rte_spinlock_init(&sh->uar_lock_cq);
978         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
979                 rte_spinlock_init(&sh->uar_lock[i]);
980 #endif
981         /*
982          * Once the device is added to the list of memory event
983          * callback, its global MR cache table cannot be expanded
984          * on the fly because of deadlock. If it overflows, lookup
985          * should be done by searching MR list linearly, which is slow.
986          *
987          * At this point the device is not added to the memory
988          * event list yet, context is just being created.
989          */
990         err = mlx5_mr_btree_init(&sh->share_cache.cache,
991                                  MLX5_MR_BTREE_CACHE_N * 2,
992                                  spawn->pci_dev->device.numa_node);
993         if (err) {
994                 err = rte_errno;
995                 goto error;
996         }
997         mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
998                               &sh->share_cache.dereg_mr_cb);
999         mlx5_os_dev_shared_handler_install(sh);
1000         sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
1001         if (!sh->cnt_id_tbl) {
1002                 err = rte_errno;
1003                 goto error;
1004         }
1005         mlx5_flow_aging_init(sh);
1006         mlx5_flow_counters_mng_init(sh);
1007         mlx5_flow_ipool_create(sh, config);
1008         /* Add device to memory callback list. */
1009         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1010         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1011                          sh, mem_event_cb);
1012         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1013         /* Add context to the global device list. */
1014         LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1015 exit:
1016         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1017         return sh;
1018 error:
1019         pthread_mutex_destroy(&sh->txpp.mutex);
1020         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1021         MLX5_ASSERT(sh);
1022         if (sh->cnt_id_tbl)
1023                 mlx5_l3t_destroy(sh->cnt_id_tbl);
1024         if (sh->tis)
1025                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1026         if (sh->td)
1027                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1028         if (sh->devx_rx_uar)
1029                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1030         if (sh->tx_uar)
1031                 mlx5_glue->devx_free_uar(sh->tx_uar);
1032         if (sh->pd)
1033                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
1034         if (sh->ctx)
1035                 claim_zero(mlx5_glue->close_device(sh->ctx));
1036         mlx5_free(sh);
1037         MLX5_ASSERT(err > 0);
1038         rte_errno = err;
1039         return NULL;
1040 }
1041
1042 /**
1043  * Free shared IB device context. Decrement counter and if zero free
1044  * all allocated resources and close handles.
1045  *
1046  * @param[in] sh
1047  *   Pointer to mlx5_dev_ctx_shared object to free
1048  */
1049 void
1050 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1051 {
1052         pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1053 #ifdef RTE_LIBRTE_MLX5_DEBUG
1054         /* Check the object presence in the list. */
1055         struct mlx5_dev_ctx_shared *lctx;
1056
1057         LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1058                 if (lctx == sh)
1059                         break;
1060         MLX5_ASSERT(lctx);
1061         if (lctx != sh) {
1062                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
1063                 goto exit;
1064         }
1065 #endif
1066         MLX5_ASSERT(sh);
1067         MLX5_ASSERT(sh->refcnt);
1068         /* Secondary process should not free the shared context. */
1069         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1070         if (--sh->refcnt)
1071                 goto exit;
1072         /* Remove from memory callback device list. */
1073         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1074         LIST_REMOVE(sh, mem_event_cb);
1075         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1076         /* Release created Memory Regions. */
1077         mlx5_mr_release_cache(&sh->share_cache);
1078         /* Remove context from the global device list. */
1079         LIST_REMOVE(sh, next);
1080         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1081         /*
1082          *  Ensure there is no async event handler installed.
1083          *  Only primary process handles async device events.
1084          **/
1085         mlx5_flow_counters_mng_close(sh);
1086         if (sh->aso_age_mng) {
1087                 mlx5_flow_aso_age_mng_close(sh);
1088                 sh->aso_age_mng = NULL;
1089         }
1090         mlx5_flow_ipool_destroy(sh);
1091         mlx5_os_dev_shared_handler_uninstall(sh);
1092         if (sh->cnt_id_tbl) {
1093                 mlx5_l3t_destroy(sh->cnt_id_tbl);
1094                 sh->cnt_id_tbl = NULL;
1095         }
1096         if (sh->tx_uar) {
1097                 mlx5_glue->devx_free_uar(sh->tx_uar);
1098                 sh->tx_uar = NULL;
1099         }
1100         if (sh->pd)
1101                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
1102         if (sh->tis)
1103                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1104         if (sh->td)
1105                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1106         if (sh->devx_rx_uar)
1107                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1108         if (sh->ctx)
1109                 claim_zero(mlx5_glue->close_device(sh->ctx));
1110         pthread_mutex_destroy(&sh->txpp.mutex);
1111         mlx5_free(sh);
1112         return;
1113 exit:
1114         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1115 }
1116
1117 /**
1118  * Destroy table hash list.
1119  *
1120  * @param[in] priv
1121  *   Pointer to the private device data structure.
1122  */
1123 void
1124 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1125 {
1126         struct mlx5_dev_ctx_shared *sh = priv->sh;
1127
1128         if (!sh->flow_tbls)
1129                 return;
1130         mlx5_hlist_destroy(sh->flow_tbls);
1131 }
1132
1133 /**
1134  * Initialize flow table hash list and create the root tables entry
1135  * for each domain.
1136  *
1137  * @param[in] priv
1138  *   Pointer to the private device data structure.
1139  *
1140  * @return
1141  *   Zero on success, positive error code otherwise.
1142  */
1143 int
1144 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1145 {
1146         int err = 0;
1147         /* Tables are only used in DV and DR modes. */
1148 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1149         struct mlx5_dev_ctx_shared *sh = priv->sh;
1150         char s[MLX5_HLIST_NAMESIZE];
1151
1152         MLX5_ASSERT(sh);
1153         snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1154         sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1155                                           0, 0, flow_dv_tbl_create_cb,
1156                                           flow_dv_tbl_match_cb,
1157                                           flow_dv_tbl_remove_cb);
1158         if (!sh->flow_tbls) {
1159                 DRV_LOG(ERR, "flow tables with hash creation failed.");
1160                 err = ENOMEM;
1161                 return err;
1162         }
1163         sh->flow_tbls->ctx = sh;
1164 #ifndef HAVE_MLX5DV_DR
1165         struct rte_flow_error error;
1166         struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1167
1168         /*
1169          * In case we have not DR support, the zero tables should be created
1170          * because DV expect to see them even if they cannot be created by
1171          * RDMA-CORE.
1172          */
1173         if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0, NULL, 0, 1, &error) ||
1174             !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0, NULL, 0, 1, &error) ||
1175             !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0, NULL, 0, 1, &error)) {
1176                 err = ENOMEM;
1177                 goto error;
1178         }
1179         return err;
1180 error:
1181         mlx5_free_table_hash_list(priv);
1182 #endif /* HAVE_MLX5DV_DR */
1183 #endif
1184         return err;
1185 }
1186
1187 /**
1188  * Retrieve integer value from environment variable.
1189  *
1190  * @param[in] name
1191  *   Environment variable name.
1192  *
1193  * @return
1194  *   Integer value, 0 if the variable is not set.
1195  */
1196 int
1197 mlx5_getenv_int(const char *name)
1198 {
1199         const char *val = getenv(name);
1200
1201         if (val == NULL)
1202                 return 0;
1203         return atoi(val);
1204 }
1205
1206 /**
1207  * DPDK callback to add udp tunnel port
1208  *
1209  * @param[in] dev
1210  *   A pointer to eth_dev
1211  * @param[in] udp_tunnel
1212  *   A pointer to udp tunnel
1213  *
1214  * @return
1215  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1216  */
1217 int
1218 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1219                          struct rte_eth_udp_tunnel *udp_tunnel)
1220 {
1221         MLX5_ASSERT(udp_tunnel != NULL);
1222         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1223             udp_tunnel->udp_port == 4789)
1224                 return 0;
1225         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1226             udp_tunnel->udp_port == 4790)
1227                 return 0;
1228         return -ENOTSUP;
1229 }
1230
1231 /**
1232  * Initialize process private data structure.
1233  *
1234  * @param dev
1235  *   Pointer to Ethernet device structure.
1236  *
1237  * @return
1238  *   0 on success, a negative errno value otherwise and rte_errno is set.
1239  */
1240 int
1241 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1242 {
1243         struct mlx5_priv *priv = dev->data->dev_private;
1244         struct mlx5_proc_priv *ppriv;
1245         size_t ppriv_size;
1246
1247         /*
1248          * UAR register table follows the process private structure. BlueFlame
1249          * registers for Tx queues are stored in the table.
1250          */
1251         ppriv_size =
1252                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1253         ppriv = mlx5_malloc(MLX5_MEM_RTE, ppriv_size, RTE_CACHE_LINE_SIZE,
1254                             dev->device->numa_node);
1255         if (!ppriv) {
1256                 rte_errno = ENOMEM;
1257                 return -rte_errno;
1258         }
1259         ppriv->uar_table_sz = ppriv_size;
1260         dev->process_private = ppriv;
1261         return 0;
1262 }
1263
1264 /**
1265  * Un-initialize process private data structure.
1266  *
1267  * @param dev
1268  *   Pointer to Ethernet device structure.
1269  */
1270 static void
1271 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1272 {
1273         if (!dev->process_private)
1274                 return;
1275         mlx5_free(dev->process_private);
1276         dev->process_private = NULL;
1277 }
1278
1279 /**
1280  * DPDK callback to close the device.
1281  *
1282  * Destroy all queues and objects, free memory.
1283  *
1284  * @param dev
1285  *   Pointer to Ethernet device structure.
1286  */
1287 int
1288 mlx5_dev_close(struct rte_eth_dev *dev)
1289 {
1290         struct mlx5_priv *priv = dev->data->dev_private;
1291         unsigned int i;
1292         int ret;
1293
1294         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1295                 /* Check if process_private released. */
1296                 if (!dev->process_private)
1297                         return 0;
1298                 mlx5_tx_uar_uninit_secondary(dev);
1299                 mlx5_proc_priv_uninit(dev);
1300                 rte_eth_dev_release_port(dev);
1301                 return 0;
1302         }
1303         if (!priv->sh)
1304                 return 0;
1305         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1306                 dev->data->port_id,
1307                 ((priv->sh->ctx != NULL) ?
1308                 mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1309         /*
1310          * If default mreg copy action is removed at the stop stage,
1311          * the search will return none and nothing will be done anymore.
1312          */
1313         mlx5_flow_stop_default(dev);
1314         mlx5_traffic_disable(dev);
1315         /*
1316          * If all the flows are already flushed in the device stop stage,
1317          * then this will return directly without any action.
1318          */
1319         mlx5_flow_list_flush(dev, &priv->flows, true);
1320         mlx5_shared_action_flush(dev);
1321         mlx5_flow_meter_flush(dev, NULL);
1322         /* Prevent crashes when queues are still in use. */
1323         dev->rx_pkt_burst = removed_rx_burst;
1324         dev->tx_pkt_burst = removed_tx_burst;
1325         rte_wmb();
1326         /* Disable datapath on secondary process. */
1327         mlx5_mp_os_req_stop_rxtx(dev);
1328         /* Free the eCPRI flex parser resource. */
1329         mlx5_flex_parser_ecpri_release(dev);
1330         if (priv->rxqs != NULL) {
1331                 /* XXX race condition if mlx5_rx_burst() is still running. */
1332                 usleep(1000);
1333                 for (i = 0; (i != priv->rxqs_n); ++i)
1334                         mlx5_rxq_release(dev, i);
1335                 priv->rxqs_n = 0;
1336                 priv->rxqs = NULL;
1337         }
1338         if (priv->txqs != NULL) {
1339                 /* XXX race condition if mlx5_tx_burst() is still running. */
1340                 usleep(1000);
1341                 for (i = 0; (i != priv->txqs_n); ++i)
1342                         mlx5_txq_release(dev, i);
1343                 priv->txqs_n = 0;
1344                 priv->txqs = NULL;
1345         }
1346         mlx5_proc_priv_uninit(dev);
1347         if (priv->drop_queue.hrxq)
1348                 mlx5_drop_action_destroy(dev);
1349         if (priv->mreg_cp_tbl)
1350                 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1351         mlx5_mprq_free_mp(dev);
1352         mlx5_os_free_shared_dr(priv);
1353         if (priv->rss_conf.rss_key != NULL)
1354                 mlx5_free(priv->rss_conf.rss_key);
1355         if (priv->reta_idx != NULL)
1356                 mlx5_free(priv->reta_idx);
1357         if (priv->config.vf)
1358                 mlx5_os_mac_addr_flush(dev);
1359         if (priv->nl_socket_route >= 0)
1360                 close(priv->nl_socket_route);
1361         if (priv->nl_socket_rdma >= 0)
1362                 close(priv->nl_socket_rdma);
1363         if (priv->vmwa_context)
1364                 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1365         ret = mlx5_hrxq_verify(dev);
1366         if (ret)
1367                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1368                         dev->data->port_id);
1369         ret = mlx5_ind_table_obj_verify(dev);
1370         if (ret)
1371                 DRV_LOG(WARNING, "port %u some indirection table still remain",
1372                         dev->data->port_id);
1373         ret = mlx5_rxq_obj_verify(dev);
1374         if (ret)
1375                 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1376                         dev->data->port_id);
1377         ret = mlx5_rxq_verify(dev);
1378         if (ret)
1379                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1380                         dev->data->port_id);
1381         ret = mlx5_txq_obj_verify(dev);
1382         if (ret)
1383                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1384                         dev->data->port_id);
1385         ret = mlx5_txq_verify(dev);
1386         if (ret)
1387                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1388                         dev->data->port_id);
1389         ret = mlx5_flow_verify(dev);
1390         if (ret)
1391                 DRV_LOG(WARNING, "port %u some flows still remain",
1392                         dev->data->port_id);
1393         mlx5_cache_list_destroy(&priv->hrxqs);
1394         /*
1395          * Free the shared context in last turn, because the cleanup
1396          * routines above may use some shared fields, like
1397          * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1398          * ifindex if Netlink fails.
1399          */
1400         mlx5_free_shared_dev_ctx(priv->sh);
1401         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1402                 unsigned int c = 0;
1403                 uint16_t port_id;
1404
1405                 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1406                         struct mlx5_priv *opriv =
1407                                 rte_eth_devices[port_id].data->dev_private;
1408
1409                         if (!opriv ||
1410                             opriv->domain_id != priv->domain_id ||
1411                             &rte_eth_devices[port_id] == dev)
1412                                 continue;
1413                         ++c;
1414                         break;
1415                 }
1416                 if (!c)
1417                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1418         }
1419         memset(priv, 0, sizeof(*priv));
1420         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1421         /*
1422          * Reset mac_addrs to NULL such that it is not freed as part of
1423          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1424          * it is freed when dev_private is freed.
1425          */
1426         dev->data->mac_addrs = NULL;
1427         return 0;
1428 }
1429
1430 /**
1431  * Verify and store value for device argument.
1432  *
1433  * @param[in] key
1434  *   Key argument to verify.
1435  * @param[in] val
1436  *   Value associated with key.
1437  * @param opaque
1438  *   User data.
1439  *
1440  * @return
1441  *   0 on success, a negative errno value otherwise and rte_errno is set.
1442  */
1443 static int
1444 mlx5_args_check(const char *key, const char *val, void *opaque)
1445 {
1446         struct mlx5_dev_config *config = opaque;
1447         unsigned long mod;
1448         signed long tmp;
1449
1450         /* No-op, port representors are processed in mlx5_dev_spawn(). */
1451         if (!strcmp(MLX5_REPRESENTOR, key))
1452                 return 0;
1453         errno = 0;
1454         tmp = strtol(val, NULL, 0);
1455         if (errno) {
1456                 rte_errno = errno;
1457                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1458                 return -rte_errno;
1459         }
1460         if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1461                 /* Negative values are acceptable for some keys only. */
1462                 rte_errno = EINVAL;
1463                 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1464                 return -rte_errno;
1465         }
1466         mod = tmp >= 0 ? tmp : -tmp;
1467         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1468                 if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1469                         DRV_LOG(ERR, "invalid CQE compression "
1470                                      "format parameter");
1471                         rte_errno = EINVAL;
1472                         return -rte_errno;
1473                 }
1474                 config->cqe_comp = !!tmp;
1475                 config->cqe_comp_fmt = tmp;
1476         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1477                 config->cqe_pad = !!tmp;
1478         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1479                 config->hw_padding = !!tmp;
1480         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1481                 config->mprq.enabled = !!tmp;
1482         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1483                 config->mprq.stride_num_n = tmp;
1484         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1485                 config->mprq.stride_size_n = tmp;
1486         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1487                 config->mprq.max_memcpy_len = tmp;
1488         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1489                 config->mprq.min_rxqs_num = tmp;
1490         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1491                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1492                                  " converted to txq_inline_max", key);
1493                 config->txq_inline_max = tmp;
1494         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1495                 config->txq_inline_max = tmp;
1496         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1497                 config->txq_inline_min = tmp;
1498         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1499                 config->txq_inline_mpw = tmp;
1500         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1501                 config->txqs_inline = tmp;
1502         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1503                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1504         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1505                 config->mps = !!tmp;
1506         } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1507                 if (tmp != MLX5_TXDB_CACHED &&
1508                     tmp != MLX5_TXDB_NCACHED &&
1509                     tmp != MLX5_TXDB_HEURISTIC) {
1510                         DRV_LOG(ERR, "invalid Tx doorbell "
1511                                      "mapping parameter");
1512                         rte_errno = EINVAL;
1513                         return -rte_errno;
1514                 }
1515                 config->dbnc = tmp;
1516         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1517                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1518         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1519                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1520                                  " converted to txq_inline_mpw", key);
1521                 config->txq_inline_mpw = tmp;
1522         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1523                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1524         } else if (strcmp(MLX5_TX_PP, key) == 0) {
1525                 if (!mod) {
1526                         DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1527                         rte_errno = EINVAL;
1528                         return -rte_errno;
1529                 }
1530                 config->tx_pp = tmp;
1531         } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1532                 config->tx_skew = tmp;
1533         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1534                 config->rx_vec_en = !!tmp;
1535         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1536                 config->l3_vxlan_en = !!tmp;
1537         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1538                 config->vf_nl_en = !!tmp;
1539         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1540                 config->dv_esw_en = !!tmp;
1541         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1542                 config->dv_flow_en = !!tmp;
1543         } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1544                 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1545                     tmp != MLX5_XMETA_MODE_META16 &&
1546                     tmp != MLX5_XMETA_MODE_META32 &&
1547                     tmp != MLX5_XMETA_MODE_MISS_INFO) {
1548                         DRV_LOG(ERR, "invalid extensive "
1549                                      "metadata parameter");
1550                         rte_errno = EINVAL;
1551                         return -rte_errno;
1552                 }
1553                 if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1554                         config->dv_xmeta_en = tmp;
1555                 else
1556                         config->dv_miss_info = 1;
1557         } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1558                 config->lacp_by_user = !!tmp;
1559         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1560                 config->mr_ext_memseg_en = !!tmp;
1561         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1562                 config->max_dump_files_num = tmp;
1563         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1564                 config->lro.timeout = tmp;
1565         } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1566                 DRV_LOG(DEBUG, "class argument is %s.", val);
1567         } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1568                 config->log_hp_size = tmp;
1569         } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1570                 if (tmp != MLX5_RCM_NONE &&
1571                     tmp != MLX5_RCM_LIGHT &&
1572                     tmp != MLX5_RCM_AGGR) {
1573                         DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1574                         rte_errno = EINVAL;
1575                         return -rte_errno;
1576                 }
1577                 config->reclaim_mode = tmp;
1578         } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1579                 config->sys_mem_en = !!tmp;
1580         } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1581                 config->decap_en = !!tmp;
1582         } else {
1583                 DRV_LOG(WARNING, "%s: unknown parameter", key);
1584                 rte_errno = EINVAL;
1585                 return -rte_errno;
1586         }
1587         return 0;
1588 }
1589
1590 /**
1591  * Parse device parameters.
1592  *
1593  * @param config
1594  *   Pointer to device configuration structure.
1595  * @param devargs
1596  *   Device arguments structure.
1597  *
1598  * @return
1599  *   0 on success, a negative errno value otherwise and rte_errno is set.
1600  */
1601 int
1602 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1603 {
1604         const char **params = (const char *[]){
1605                 MLX5_RXQ_CQE_COMP_EN,
1606                 MLX5_RXQ_CQE_PAD_EN,
1607                 MLX5_RXQ_PKT_PAD_EN,
1608                 MLX5_RX_MPRQ_EN,
1609                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1610                 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1611                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1612                 MLX5_RXQS_MIN_MPRQ,
1613                 MLX5_TXQ_INLINE,
1614                 MLX5_TXQ_INLINE_MIN,
1615                 MLX5_TXQ_INLINE_MAX,
1616                 MLX5_TXQ_INLINE_MPW,
1617                 MLX5_TXQS_MIN_INLINE,
1618                 MLX5_TXQS_MAX_VEC,
1619                 MLX5_TXQ_MPW_EN,
1620                 MLX5_TXQ_MPW_HDR_DSEG_EN,
1621                 MLX5_TXQ_MAX_INLINE_LEN,
1622                 MLX5_TX_DB_NC,
1623                 MLX5_TX_PP,
1624                 MLX5_TX_SKEW,
1625                 MLX5_TX_VEC_EN,
1626                 MLX5_RX_VEC_EN,
1627                 MLX5_L3_VXLAN_EN,
1628                 MLX5_VF_NL_EN,
1629                 MLX5_DV_ESW_EN,
1630                 MLX5_DV_FLOW_EN,
1631                 MLX5_DV_XMETA_EN,
1632                 MLX5_LACP_BY_USER,
1633                 MLX5_MR_EXT_MEMSEG_EN,
1634                 MLX5_REPRESENTOR,
1635                 MLX5_MAX_DUMP_FILES_NUM,
1636                 MLX5_LRO_TIMEOUT_USEC,
1637                 MLX5_CLASS_ARG_NAME,
1638                 MLX5_HP_BUF_SIZE,
1639                 MLX5_RECLAIM_MEM,
1640                 MLX5_SYS_MEM_EN,
1641                 MLX5_DECAP_EN,
1642                 NULL,
1643         };
1644         struct rte_kvargs *kvlist;
1645         int ret = 0;
1646         int i;
1647
1648         if (devargs == NULL)
1649                 return 0;
1650         /* Following UGLY cast is done to pass checkpatch. */
1651         kvlist = rte_kvargs_parse(devargs->args, params);
1652         if (kvlist == NULL) {
1653                 rte_errno = EINVAL;
1654                 return -rte_errno;
1655         }
1656         /* Process parameters. */
1657         for (i = 0; (params[i] != NULL); ++i) {
1658                 if (rte_kvargs_count(kvlist, params[i])) {
1659                         ret = rte_kvargs_process(kvlist, params[i],
1660                                                  mlx5_args_check, config);
1661                         if (ret) {
1662                                 rte_errno = EINVAL;
1663                                 rte_kvargs_free(kvlist);
1664                                 return -rte_errno;
1665                         }
1666                 }
1667         }
1668         rte_kvargs_free(kvlist);
1669         return 0;
1670 }
1671
1672 /**
1673  * Configures the minimal amount of data to inline into WQE
1674  * while sending packets.
1675  *
1676  * - the txq_inline_min has the maximal priority, if this
1677  *   key is specified in devargs
1678  * - if DevX is enabled the inline mode is queried from the
1679  *   device (HCA attributes and NIC vport context if needed).
1680  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1681  *   and none (0 bytes) for other NICs
1682  *
1683  * @param spawn
1684  *   Verbs device parameters (name, port, switch_info) to spawn.
1685  * @param config
1686  *   Device configuration parameters.
1687  */
1688 void
1689 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1690                     struct mlx5_dev_config *config)
1691 {
1692         if (config->txq_inline_min != MLX5_ARG_UNSET) {
1693                 /* Application defines size of inlined data explicitly. */
1694                 switch (spawn->pci_dev->id.device_id) {
1695                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1696                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1697                         if (config->txq_inline_min <
1698                                        (int)MLX5_INLINE_HSIZE_L2) {
1699                                 DRV_LOG(DEBUG,
1700                                         "txq_inline_mix aligned to minimal"
1701                                         " ConnectX-4 required value %d",
1702                                         (int)MLX5_INLINE_HSIZE_L2);
1703                                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1704                         }
1705                         break;
1706                 }
1707                 goto exit;
1708         }
1709         if (config->hca_attr.eth_net_offloads) {
1710                 /* We have DevX enabled, inline mode queried successfully. */
1711                 switch (config->hca_attr.wqe_inline_mode) {
1712                 case MLX5_CAP_INLINE_MODE_L2:
1713                         /* outer L2 header must be inlined. */
1714                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1715                         goto exit;
1716                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1717                         /* No inline data are required by NIC. */
1718                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1719                         config->hw_vlan_insert =
1720                                 config->hca_attr.wqe_vlan_insert;
1721                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1722                         goto exit;
1723                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1724                         /* inline mode is defined by NIC vport context. */
1725                         if (!config->hca_attr.eth_virt)
1726                                 break;
1727                         switch (config->hca_attr.vport_inline_mode) {
1728                         case MLX5_INLINE_MODE_NONE:
1729                                 config->txq_inline_min =
1730                                         MLX5_INLINE_HSIZE_NONE;
1731                                 goto exit;
1732                         case MLX5_INLINE_MODE_L2:
1733                                 config->txq_inline_min =
1734                                         MLX5_INLINE_HSIZE_L2;
1735                                 goto exit;
1736                         case MLX5_INLINE_MODE_IP:
1737                                 config->txq_inline_min =
1738                                         MLX5_INLINE_HSIZE_L3;
1739                                 goto exit;
1740                         case MLX5_INLINE_MODE_TCP_UDP:
1741                                 config->txq_inline_min =
1742                                         MLX5_INLINE_HSIZE_L4;
1743                                 goto exit;
1744                         case MLX5_INLINE_MODE_INNER_L2:
1745                                 config->txq_inline_min =
1746                                         MLX5_INLINE_HSIZE_INNER_L2;
1747                                 goto exit;
1748                         case MLX5_INLINE_MODE_INNER_IP:
1749                                 config->txq_inline_min =
1750                                         MLX5_INLINE_HSIZE_INNER_L3;
1751                                 goto exit;
1752                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
1753                                 config->txq_inline_min =
1754                                         MLX5_INLINE_HSIZE_INNER_L4;
1755                                 goto exit;
1756                         }
1757                 }
1758         }
1759         /*
1760          * We get here if we are unable to deduce
1761          * inline data size with DevX. Try PCI ID
1762          * to determine old NICs.
1763          */
1764         switch (spawn->pci_dev->id.device_id) {
1765         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1766         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1767         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1768         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1769                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1770                 config->hw_vlan_insert = 0;
1771                 break;
1772         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1773         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1774         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1775         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1776                 /*
1777                  * These NICs support VLAN insertion from WQE and
1778                  * report the wqe_vlan_insert flag. But there is the bug
1779                  * and PFC control may be broken, so disable feature.
1780                  */
1781                 config->hw_vlan_insert = 0;
1782                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1783                 break;
1784         default:
1785                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1786                 break;
1787         }
1788 exit:
1789         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1790 }
1791
1792 /**
1793  * Configures the metadata mask fields in the shared context.
1794  *
1795  * @param [in] dev
1796  *   Pointer to Ethernet device.
1797  */
1798 void
1799 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1800 {
1801         struct mlx5_priv *priv = dev->data->dev_private;
1802         struct mlx5_dev_ctx_shared *sh = priv->sh;
1803         uint32_t meta, mark, reg_c0;
1804
1805         reg_c0 = ~priv->vport_meta_mask;
1806         switch (priv->config.dv_xmeta_en) {
1807         case MLX5_XMETA_MODE_LEGACY:
1808                 meta = UINT32_MAX;
1809                 mark = MLX5_FLOW_MARK_MASK;
1810                 break;
1811         case MLX5_XMETA_MODE_META16:
1812                 meta = reg_c0 >> rte_bsf32(reg_c0);
1813                 mark = MLX5_FLOW_MARK_MASK;
1814                 break;
1815         case MLX5_XMETA_MODE_META32:
1816                 meta = UINT32_MAX;
1817                 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1818                 break;
1819         default:
1820                 meta = 0;
1821                 mark = 0;
1822                 MLX5_ASSERT(false);
1823                 break;
1824         }
1825         if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1826                 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1827                                  sh->dv_mark_mask, mark);
1828         else
1829                 sh->dv_mark_mask = mark;
1830         if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1831                 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1832                                  sh->dv_meta_mask, meta);
1833         else
1834                 sh->dv_meta_mask = meta;
1835         if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1836                 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1837                                  sh->dv_meta_mask, reg_c0);
1838         else
1839                 sh->dv_regc0_mask = reg_c0;
1840         DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1841         DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1842         DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1843         DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1844 }
1845
1846 int
1847 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
1848 {
1849         static const char *const dynf_names[] = {
1850                 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
1851                 RTE_MBUF_DYNFLAG_METADATA_NAME,
1852                 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
1853         };
1854         unsigned int i;
1855
1856         if (n < RTE_DIM(dynf_names))
1857                 return -ENOMEM;
1858         for (i = 0; i < RTE_DIM(dynf_names); i++) {
1859                 if (names[i] == NULL)
1860                         return -EINVAL;
1861                 strcpy(names[i], dynf_names[i]);
1862         }
1863         return RTE_DIM(dynf_names);
1864 }
1865
1866 /**
1867  * Comparison callback to sort device data.
1868  *
1869  * This is meant to be used with qsort().
1870  *
1871  * @param a[in]
1872  *   Pointer to pointer to first data object.
1873  * @param b[in]
1874  *   Pointer to pointer to second data object.
1875  *
1876  * @return
1877  *   0 if both objects are equal, less than 0 if the first argument is less
1878  *   than the second, greater than 0 otherwise.
1879  */
1880 int
1881 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1882                               struct mlx5_dev_config *config)
1883 {
1884         struct mlx5_dev_ctx_shared *sh = priv->sh;
1885         struct mlx5_dev_config *sh_conf = NULL;
1886         uint16_t port_id;
1887
1888         MLX5_ASSERT(sh);
1889         /* Nothing to compare for the single/first device. */
1890         if (sh->refcnt == 1)
1891                 return 0;
1892         /* Find the device with shared context. */
1893         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1894                 struct mlx5_priv *opriv =
1895                         rte_eth_devices[port_id].data->dev_private;
1896
1897                 if (opriv && opriv != priv && opriv->sh == sh) {
1898                         sh_conf = &opriv->config;
1899                         break;
1900                 }
1901         }
1902         if (!sh_conf)
1903                 return 0;
1904         if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1905                 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1906                              " for shared %s context", sh->ibdev_name);
1907                 rte_errno = EINVAL;
1908                 return rte_errno;
1909         }
1910         if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
1911                 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
1912                              " for shared %s context", sh->ibdev_name);
1913                 rte_errno = EINVAL;
1914                 return rte_errno;
1915         }
1916         return 0;
1917 }
1918
1919 /**
1920  * Look for the ethernet device belonging to mlx5 driver.
1921  *
1922  * @param[in] port_id
1923  *   port_id to start looking for device.
1924  * @param[in] pci_dev
1925  *   Pointer to the hint PCI device. When device is being probed
1926  *   the its siblings (master and preceding representors might
1927  *   not have assigned driver yet (because the mlx5_os_pci_probe()
1928  *   is not completed yet, for this case match on hint PCI
1929  *   device may be used to detect sibling device.
1930  *
1931  * @return
1932  *   port_id of found device, RTE_MAX_ETHPORT if not found.
1933  */
1934 uint16_t
1935 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
1936 {
1937         while (port_id < RTE_MAX_ETHPORTS) {
1938                 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
1939
1940                 if (dev->state != RTE_ETH_DEV_UNUSED &&
1941                     dev->device &&
1942                     (dev->device == &pci_dev->device ||
1943                      (dev->device->driver &&
1944                      dev->device->driver->name &&
1945                      !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
1946                         break;
1947                 port_id++;
1948         }
1949         if (port_id >= RTE_MAX_ETHPORTS)
1950                 return RTE_MAX_ETHPORTS;
1951         return port_id;
1952 }
1953
1954 /**
1955  * DPDK callback to remove a PCI device.
1956  *
1957  * This function removes all Ethernet devices belong to a given PCI device.
1958  *
1959  * @param[in] pci_dev
1960  *   Pointer to the PCI device.
1961  *
1962  * @return
1963  *   0 on success, the function cannot fail.
1964  */
1965 static int
1966 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1967 {
1968         uint16_t port_id;
1969         int ret = 0;
1970
1971         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
1972                 /*
1973                  * mlx5_dev_close() is not registered to secondary process,
1974                  * call the close function explicitly for secondary process.
1975                  */
1976                 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1977                         ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
1978                 else
1979                         ret |= rte_eth_dev_close(port_id);
1980         }
1981         return ret == 0 ? 0 : -EIO;
1982 }
1983
1984 static const struct rte_pci_id mlx5_pci_id_map[] = {
1985         {
1986                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1987                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1988         },
1989         {
1990                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1991                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1992         },
1993         {
1994                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1995                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1996         },
1997         {
1998                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1999                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2000         },
2001         {
2002                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2003                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2004         },
2005         {
2006                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2007                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2008         },
2009         {
2010                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2011                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2012         },
2013         {
2014                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2015                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2016         },
2017         {
2018                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2019                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2020         },
2021         {
2022                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2023                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2024         },
2025         {
2026                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2027                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2028         },
2029         {
2030                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2031                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2032         },
2033         {
2034                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2035                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2036         },
2037         {
2038                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2039                                 PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2040         },
2041         {
2042                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2043                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2044         },
2045         {
2046                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2047                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2048         },
2049         {
2050                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2051                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2052         },
2053         {
2054                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2055                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2056         },
2057         {
2058                 .vendor_id = 0
2059         }
2060 };
2061
2062 static struct mlx5_pci_driver mlx5_driver = {
2063         .driver_class = MLX5_CLASS_NET,
2064         .pci_driver = {
2065                 .driver = {
2066                         .name = MLX5_DRIVER_NAME,
2067                 },
2068                 .id_table = mlx5_pci_id_map,
2069                 .probe = mlx5_os_pci_probe,
2070                 .remove = mlx5_pci_remove,
2071                 .dma_map = mlx5_dma_map,
2072                 .dma_unmap = mlx5_dma_unmap,
2073                 .drv_flags = PCI_DRV_FLAGS,
2074         },
2075 };
2076
2077 /* Initialize driver log type. */
2078 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
2079
2080 /**
2081  * Driver initialization routine.
2082  */
2083 RTE_INIT(rte_mlx5_pmd_init)
2084 {
2085         mlx5_common_init();
2086         /* Build the static tables for Verbs conversion. */
2087         mlx5_set_ptype_table();
2088         mlx5_set_cksum_table();
2089         mlx5_set_swp_types_table();
2090         if (mlx5_glue)
2091                 mlx5_pci_driver_register(&mlx5_driver);
2092 }
2093
2094 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2095 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2096 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");