43344391df24f60eab7d96e53827528edbb9ec39
[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         {
190                 .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
191                 .trunk_size = 64,
192                 .grow_trunk = 3,
193                 .grow_shift = 2,
194                 .need_lock = 1,
195                 .release_mem_en = 1,
196                 .malloc = mlx5_malloc,
197                 .free = mlx5_free,
198                 .type = "mlx5_encap_decap_ipool",
199         },
200         {
201                 .size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
202                 .trunk_size = 64,
203                 .grow_trunk = 3,
204                 .grow_shift = 2,
205                 .need_lock = 1,
206                 .release_mem_en = 1,
207                 .malloc = mlx5_malloc,
208                 .free = mlx5_free,
209                 .type = "mlx5_push_vlan_ipool",
210         },
211         {
212                 .size = sizeof(struct mlx5_flow_dv_tag_resource),
213                 .trunk_size = 64,
214                 .grow_trunk = 3,
215                 .grow_shift = 2,
216                 .need_lock = 1,
217                 .release_mem_en = 1,
218                 .malloc = mlx5_malloc,
219                 .free = mlx5_free,
220                 .type = "mlx5_tag_ipool",
221         },
222         {
223                 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
224                 .trunk_size = 64,
225                 .grow_trunk = 3,
226                 .grow_shift = 2,
227                 .need_lock = 1,
228                 .release_mem_en = 1,
229                 .malloc = mlx5_malloc,
230                 .free = mlx5_free,
231                 .type = "mlx5_port_id_ipool",
232         },
233         {
234                 .size = sizeof(struct mlx5_flow_tbl_data_entry),
235                 .trunk_size = 64,
236                 .grow_trunk = 3,
237                 .grow_shift = 2,
238                 .need_lock = 1,
239                 .release_mem_en = 1,
240                 .malloc = mlx5_malloc,
241                 .free = mlx5_free,
242                 .type = "mlx5_jump_ipool",
243         },
244         {
245                 .size = sizeof(struct mlx5_flow_dv_sample_resource),
246                 .trunk_size = 64,
247                 .grow_trunk = 3,
248                 .grow_shift = 2,
249                 .need_lock = 1,
250                 .release_mem_en = 1,
251                 .malloc = mlx5_malloc,
252                 .free = mlx5_free,
253                 .type = "mlx5_sample_ipool",
254         },
255         {
256                 .size = sizeof(struct mlx5_flow_dv_dest_array_resource),
257                 .trunk_size = 64,
258                 .grow_trunk = 3,
259                 .grow_shift = 2,
260                 .need_lock = 1,
261                 .release_mem_en = 1,
262                 .malloc = mlx5_malloc,
263                 .free = mlx5_free,
264                 .type = "mlx5_dest_array_ipool",
265         },
266 #endif
267         {
268                 .size = sizeof(struct mlx5_flow_meter),
269                 .trunk_size = 64,
270                 .grow_trunk = 3,
271                 .grow_shift = 2,
272                 .need_lock = 1,
273                 .release_mem_en = 1,
274                 .malloc = mlx5_malloc,
275                 .free = mlx5_free,
276                 .type = "mlx5_meter_ipool",
277         },
278         {
279                 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
280                 .trunk_size = 64,
281                 .grow_trunk = 3,
282                 .grow_shift = 2,
283                 .need_lock = 1,
284                 .release_mem_en = 1,
285                 .malloc = mlx5_malloc,
286                 .free = mlx5_free,
287                 .type = "mlx5_mcp_ipool",
288         },
289         {
290                 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
291                 .trunk_size = 64,
292                 .grow_trunk = 3,
293                 .grow_shift = 2,
294                 .need_lock = 1,
295                 .release_mem_en = 1,
296                 .malloc = mlx5_malloc,
297                 .free = mlx5_free,
298                 .type = "mlx5_hrxq_ipool",
299         },
300         {
301                 /*
302                  * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
303                  * It set in run time according to PCI function configuration.
304                  */
305                 .size = 0,
306                 .trunk_size = 64,
307                 .grow_trunk = 3,
308                 .grow_shift = 2,
309                 .need_lock = 1,
310                 .release_mem_en = 1,
311                 .malloc = mlx5_malloc,
312                 .free = mlx5_free,
313                 .type = "mlx5_flow_handle_ipool",
314         },
315         {
316                 .size = sizeof(struct rte_flow),
317                 .trunk_size = 4096,
318                 .need_lock = 1,
319                 .release_mem_en = 1,
320                 .malloc = mlx5_malloc,
321                 .free = mlx5_free,
322                 .type = "rte_flow_ipool",
323         },
324         {
325                 .size = 0,
326                 .need_lock = 1,
327                 .type = "mlx5_flow_rss_id_ipool",
328         },
329         {
330                 .size = 0,
331                 .need_lock = 1,
332                 .type = "mlx5_flow_tnl_flow_ipool",
333         },
334         {
335                 .size = 0,
336                 .need_lock = 1,
337                 .type = "mlx5_flow_tnl_tbl_ipool",
338         },
339         {
340                 .size = sizeof(struct mlx5_shared_action_rss),
341                 .trunk_size = 64,
342                 .grow_trunk = 3,
343                 .grow_shift = 2,
344                 .need_lock = 1,
345                 .release_mem_en = 1,
346                 .malloc = mlx5_malloc,
347                 .free = mlx5_free,
348                 .type = "mlx5_shared_action_rss",
349         },
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         memset(&sh->aso_age_mng, 0, sizeof(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->max_port = spawn->max_port;
919         strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
920                 sizeof(sh->ibdev_name) - 1);
921         strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
922                 sizeof(sh->ibdev_path) - 1);
923         /*
924          * Setting port_id to max unallowed value means
925          * there is no interrupt subhandler installed for
926          * the given port index i.
927          */
928         for (i = 0; i < sh->max_port; i++) {
929                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
930                 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
931         }
932         sh->pd = mlx5_glue->alloc_pd(sh->ctx);
933         if (sh->pd == NULL) {
934                 DRV_LOG(ERR, "PD allocation failure");
935                 err = ENOMEM;
936                 goto error;
937         }
938         if (sh->devx) {
939                 /* Query the EQN for this core. */
940                 err = mlx5_glue->devx_query_eqn(sh->ctx, 0, &sh->eqn);
941                 if (err) {
942                         rte_errno = errno;
943                         DRV_LOG(ERR, "Failed to query event queue number %d.",
944                                 rte_errno);
945                         goto error;
946                 }
947                 err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
948                 if (err) {
949                         DRV_LOG(ERR, "Fail to extract pdn from PD");
950                         goto error;
951                 }
952                 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
953                 if (!sh->td) {
954                         DRV_LOG(ERR, "TD allocation failure");
955                         err = ENOMEM;
956                         goto error;
957                 }
958                 tis_attr.transport_domain = sh->td->id;
959                 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
960                 if (!sh->tis) {
961                         DRV_LOG(ERR, "TIS allocation failure");
962                         err = ENOMEM;
963                         goto error;
964                 }
965                 err = mlx5_alloc_rxtx_uars(sh, config);
966                 if (err)
967                         goto error;
968                 MLX5_ASSERT(sh->tx_uar);
969                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
970
971                 MLX5_ASSERT(sh->devx_rx_uar);
972                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
973         }
974 #ifndef RTE_ARCH_64
975         /* Initialize UAR access locks for 32bit implementations. */
976         rte_spinlock_init(&sh->uar_lock_cq);
977         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
978                 rte_spinlock_init(&sh->uar_lock[i]);
979 #endif
980         /*
981          * Once the device is added to the list of memory event
982          * callback, its global MR cache table cannot be expanded
983          * on the fly because of deadlock. If it overflows, lookup
984          * should be done by searching MR list linearly, which is slow.
985          *
986          * At this point the device is not added to the memory
987          * event list yet, context is just being created.
988          */
989         err = mlx5_mr_btree_init(&sh->share_cache.cache,
990                                  MLX5_MR_BTREE_CACHE_N * 2,
991                                  spawn->pci_dev->device.numa_node);
992         if (err) {
993                 err = rte_errno;
994                 goto error;
995         }
996         mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
997                               &sh->share_cache.dereg_mr_cb);
998         mlx5_os_dev_shared_handler_install(sh);
999         sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
1000         if (!sh->cnt_id_tbl) {
1001                 err = rte_errno;
1002                 goto error;
1003         }
1004         mlx5_flow_aging_init(sh);
1005         mlx5_flow_counters_mng_init(sh);
1006         mlx5_flow_ipool_create(sh, config);
1007         /* Add device to memory callback list. */
1008         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1009         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1010                          sh, mem_event_cb);
1011         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1012         /* Add context to the global device list. */
1013         LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1014 exit:
1015         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1016         return sh;
1017 error:
1018         pthread_mutex_destroy(&sh->txpp.mutex);
1019         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1020         MLX5_ASSERT(sh);
1021         if (sh->cnt_id_tbl)
1022                 mlx5_l3t_destroy(sh->cnt_id_tbl);
1023         if (sh->tis)
1024                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1025         if (sh->td)
1026                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1027         if (sh->devx_rx_uar)
1028                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1029         if (sh->tx_uar)
1030                 mlx5_glue->devx_free_uar(sh->tx_uar);
1031         if (sh->pd)
1032                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
1033         if (sh->ctx)
1034                 claim_zero(mlx5_glue->close_device(sh->ctx));
1035         mlx5_free(sh);
1036         MLX5_ASSERT(err > 0);
1037         rte_errno = err;
1038         return NULL;
1039 }
1040
1041 /**
1042  * Free shared IB device context. Decrement counter and if zero free
1043  * all allocated resources and close handles.
1044  *
1045  * @param[in] sh
1046  *   Pointer to mlx5_dev_ctx_shared object to free
1047  */
1048 void
1049 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1050 {
1051         pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1052 #ifdef RTE_LIBRTE_MLX5_DEBUG
1053         /* Check the object presence in the list. */
1054         struct mlx5_dev_ctx_shared *lctx;
1055
1056         LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1057                 if (lctx == sh)
1058                         break;
1059         MLX5_ASSERT(lctx);
1060         if (lctx != sh) {
1061                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
1062                 goto exit;
1063         }
1064 #endif
1065         MLX5_ASSERT(sh);
1066         MLX5_ASSERT(sh->refcnt);
1067         /* Secondary process should not free the shared context. */
1068         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1069         if (--sh->refcnt)
1070                 goto exit;
1071         /* Remove from memory callback device list. */
1072         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1073         LIST_REMOVE(sh, mem_event_cb);
1074         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1075         /* Release created Memory Regions. */
1076         mlx5_mr_release_cache(&sh->share_cache);
1077         /* Remove context from the global device list. */
1078         LIST_REMOVE(sh, next);
1079         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1080         /*
1081          *  Ensure there is no async event handler installed.
1082          *  Only primary process handles async device events.
1083          **/
1084         mlx5_flow_counters_mng_close(sh);
1085         if (sh->aso_age_mng) {
1086                 mlx5_flow_aso_age_mng_close(sh);
1087                 sh->aso_age_mng = NULL;
1088         }
1089         mlx5_flow_ipool_destroy(sh);
1090         mlx5_os_dev_shared_handler_uninstall(sh);
1091         if (sh->cnt_id_tbl) {
1092                 mlx5_l3t_destroy(sh->cnt_id_tbl);
1093                 sh->cnt_id_tbl = NULL;
1094         }
1095         if (sh->tx_uar) {
1096                 mlx5_glue->devx_free_uar(sh->tx_uar);
1097                 sh->tx_uar = NULL;
1098         }
1099         if (sh->pd)
1100                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
1101         if (sh->tis)
1102                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1103         if (sh->td)
1104                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1105         if (sh->devx_rx_uar)
1106                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1107         if (sh->ctx)
1108                 claim_zero(mlx5_glue->close_device(sh->ctx));
1109         pthread_mutex_destroy(&sh->txpp.mutex);
1110         mlx5_free(sh);
1111         return;
1112 exit:
1113         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1114 }
1115
1116 /**
1117  * Destroy table hash list.
1118  *
1119  * @param[in] priv
1120  *   Pointer to the private device data structure.
1121  */
1122 void
1123 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1124 {
1125         struct mlx5_dev_ctx_shared *sh = priv->sh;
1126
1127         if (!sh->flow_tbls)
1128                 return;
1129         mlx5_hlist_destroy(sh->flow_tbls);
1130 }
1131
1132 /**
1133  * Initialize flow table hash list and create the root tables entry
1134  * for each domain.
1135  *
1136  * @param[in] priv
1137  *   Pointer to the private device data structure.
1138  *
1139  * @return
1140  *   Zero on success, positive error code otherwise.
1141  */
1142 int
1143 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1144 {
1145         int err = 0;
1146         /* Tables are only used in DV and DR modes. */
1147 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1148         struct mlx5_dev_ctx_shared *sh = priv->sh;
1149         char s[MLX5_HLIST_NAMESIZE];
1150
1151         MLX5_ASSERT(sh);
1152         snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1153         sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1154                                           0, 0, flow_dv_tbl_create_cb, NULL,
1155                                           flow_dv_tbl_remove_cb);
1156         if (!sh->flow_tbls) {
1157                 DRV_LOG(ERR, "flow tables with hash creation failed.");
1158                 err = ENOMEM;
1159                 return err;
1160         }
1161         sh->flow_tbls->ctx = sh;
1162 #ifndef HAVE_MLX5DV_DR
1163         struct rte_flow_error error;
1164         struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1165
1166         /*
1167          * In case we have not DR support, the zero tables should be created
1168          * because DV expect to see them even if they cannot be created by
1169          * RDMA-CORE.
1170          */
1171         if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0, NULL, 0, 1, &error) ||
1172             !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0, NULL, 0, 1, &error) ||
1173             !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0, NULL, 0, 1, &error)) {
1174                 err = ENOMEM;
1175                 goto error;
1176         }
1177         return err;
1178 error:
1179         mlx5_free_table_hash_list(priv);
1180 #endif /* HAVE_MLX5DV_DR */
1181 #endif
1182         return err;
1183 }
1184
1185 /**
1186  * Retrieve integer value from environment variable.
1187  *
1188  * @param[in] name
1189  *   Environment variable name.
1190  *
1191  * @return
1192  *   Integer value, 0 if the variable is not set.
1193  */
1194 int
1195 mlx5_getenv_int(const char *name)
1196 {
1197         const char *val = getenv(name);
1198
1199         if (val == NULL)
1200                 return 0;
1201         return atoi(val);
1202 }
1203
1204 /**
1205  * DPDK callback to add udp tunnel port
1206  *
1207  * @param[in] dev
1208  *   A pointer to eth_dev
1209  * @param[in] udp_tunnel
1210  *   A pointer to udp tunnel
1211  *
1212  * @return
1213  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1214  */
1215 int
1216 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1217                          struct rte_eth_udp_tunnel *udp_tunnel)
1218 {
1219         MLX5_ASSERT(udp_tunnel != NULL);
1220         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1221             udp_tunnel->udp_port == 4789)
1222                 return 0;
1223         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1224             udp_tunnel->udp_port == 4790)
1225                 return 0;
1226         return -ENOTSUP;
1227 }
1228
1229 /**
1230  * Initialize process private data structure.
1231  *
1232  * @param dev
1233  *   Pointer to Ethernet device structure.
1234  *
1235  * @return
1236  *   0 on success, a negative errno value otherwise and rte_errno is set.
1237  */
1238 int
1239 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1240 {
1241         struct mlx5_priv *priv = dev->data->dev_private;
1242         struct mlx5_proc_priv *ppriv;
1243         size_t ppriv_size;
1244
1245         /*
1246          * UAR register table follows the process private structure. BlueFlame
1247          * registers for Tx queues are stored in the table.
1248          */
1249         ppriv_size =
1250                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1251         ppriv = mlx5_malloc(MLX5_MEM_RTE, ppriv_size, RTE_CACHE_LINE_SIZE,
1252                             dev->device->numa_node);
1253         if (!ppriv) {
1254                 rte_errno = ENOMEM;
1255                 return -rte_errno;
1256         }
1257         ppriv->uar_table_sz = ppriv_size;
1258         dev->process_private = ppriv;
1259         return 0;
1260 }
1261
1262 /**
1263  * Un-initialize process private data structure.
1264  *
1265  * @param dev
1266  *   Pointer to Ethernet device structure.
1267  */
1268 static void
1269 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1270 {
1271         if (!dev->process_private)
1272                 return;
1273         mlx5_free(dev->process_private);
1274         dev->process_private = NULL;
1275 }
1276
1277 /**
1278  * DPDK callback to close the device.
1279  *
1280  * Destroy all queues and objects, free memory.
1281  *
1282  * @param dev
1283  *   Pointer to Ethernet device structure.
1284  */
1285 int
1286 mlx5_dev_close(struct rte_eth_dev *dev)
1287 {
1288         struct mlx5_priv *priv = dev->data->dev_private;
1289         unsigned int i;
1290         int ret;
1291
1292         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1293                 /* Check if process_private released. */
1294                 if (!dev->process_private)
1295                         return 0;
1296                 mlx5_tx_uar_uninit_secondary(dev);
1297                 mlx5_proc_priv_uninit(dev);
1298                 rte_eth_dev_release_port(dev);
1299                 return 0;
1300         }
1301         if (!priv->sh)
1302                 return 0;
1303         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1304                 dev->data->port_id,
1305                 ((priv->sh->ctx != NULL) ?
1306                 mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1307         /*
1308          * If default mreg copy action is removed at the stop stage,
1309          * the search will return none and nothing will be done anymore.
1310          */
1311         mlx5_flow_stop_default(dev);
1312         mlx5_traffic_disable(dev);
1313         /*
1314          * If all the flows are already flushed in the device stop stage,
1315          * then this will return directly without any action.
1316          */
1317         mlx5_flow_list_flush(dev, &priv->flows, true);
1318         mlx5_shared_action_flush(dev);
1319         mlx5_flow_meter_flush(dev, NULL);
1320         /* Prevent crashes when queues are still in use. */
1321         dev->rx_pkt_burst = removed_rx_burst;
1322         dev->tx_pkt_burst = removed_tx_burst;
1323         rte_wmb();
1324         /* Disable datapath on secondary process. */
1325         mlx5_mp_os_req_stop_rxtx(dev);
1326         /* Free the eCPRI flex parser resource. */
1327         mlx5_flex_parser_ecpri_release(dev);
1328         if (priv->rxqs != NULL) {
1329                 /* XXX race condition if mlx5_rx_burst() is still running. */
1330                 usleep(1000);
1331                 for (i = 0; (i != priv->rxqs_n); ++i)
1332                         mlx5_rxq_release(dev, i);
1333                 priv->rxqs_n = 0;
1334                 priv->rxqs = NULL;
1335         }
1336         if (priv->txqs != NULL) {
1337                 /* XXX race condition if mlx5_tx_burst() is still running. */
1338                 usleep(1000);
1339                 for (i = 0; (i != priv->txqs_n); ++i)
1340                         mlx5_txq_release(dev, i);
1341                 priv->txqs_n = 0;
1342                 priv->txqs = NULL;
1343         }
1344         mlx5_proc_priv_uninit(dev);
1345         if (priv->drop_queue.hrxq)
1346                 mlx5_drop_action_destroy(dev);
1347         if (priv->mreg_cp_tbl)
1348                 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1349         mlx5_mprq_free_mp(dev);
1350         mlx5_os_free_shared_dr(priv);
1351         if (priv->rss_conf.rss_key != NULL)
1352                 mlx5_free(priv->rss_conf.rss_key);
1353         if (priv->reta_idx != NULL)
1354                 mlx5_free(priv->reta_idx);
1355         if (priv->config.vf)
1356                 mlx5_os_mac_addr_flush(dev);
1357         if (priv->nl_socket_route >= 0)
1358                 close(priv->nl_socket_route);
1359         if (priv->nl_socket_rdma >= 0)
1360                 close(priv->nl_socket_rdma);
1361         if (priv->vmwa_context)
1362                 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1363         ret = mlx5_hrxq_verify(dev);
1364         if (ret)
1365                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1366                         dev->data->port_id);
1367         ret = mlx5_ind_table_obj_verify(dev);
1368         if (ret)
1369                 DRV_LOG(WARNING, "port %u some indirection table still remain",
1370                         dev->data->port_id);
1371         ret = mlx5_rxq_obj_verify(dev);
1372         if (ret)
1373                 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1374                         dev->data->port_id);
1375         ret = mlx5_rxq_verify(dev);
1376         if (ret)
1377                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1378                         dev->data->port_id);
1379         ret = mlx5_txq_obj_verify(dev);
1380         if (ret)
1381                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1382                         dev->data->port_id);
1383         ret = mlx5_txq_verify(dev);
1384         if (ret)
1385                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1386                         dev->data->port_id);
1387         ret = mlx5_flow_verify(dev);
1388         if (ret)
1389                 DRV_LOG(WARNING, "port %u some flows still remain",
1390                         dev->data->port_id);
1391         mlx5_cache_list_destroy(&priv->hrxqs);
1392         /*
1393          * Free the shared context in last turn, because the cleanup
1394          * routines above may use some shared fields, like
1395          * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1396          * ifindex if Netlink fails.
1397          */
1398         mlx5_free_shared_dev_ctx(priv->sh);
1399         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1400                 unsigned int c = 0;
1401                 uint16_t port_id;
1402
1403                 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1404                         struct mlx5_priv *opriv =
1405                                 rte_eth_devices[port_id].data->dev_private;
1406
1407                         if (!opriv ||
1408                             opriv->domain_id != priv->domain_id ||
1409                             &rte_eth_devices[port_id] == dev)
1410                                 continue;
1411                         ++c;
1412                         break;
1413                 }
1414                 if (!c)
1415                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1416         }
1417         memset(priv, 0, sizeof(*priv));
1418         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1419         /*
1420          * Reset mac_addrs to NULL such that it is not freed as part of
1421          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1422          * it is freed when dev_private is freed.
1423          */
1424         dev->data->mac_addrs = NULL;
1425         return 0;
1426 }
1427
1428 /**
1429  * Verify and store value for device argument.
1430  *
1431  * @param[in] key
1432  *   Key argument to verify.
1433  * @param[in] val
1434  *   Value associated with key.
1435  * @param opaque
1436  *   User data.
1437  *
1438  * @return
1439  *   0 on success, a negative errno value otherwise and rte_errno is set.
1440  */
1441 static int
1442 mlx5_args_check(const char *key, const char *val, void *opaque)
1443 {
1444         struct mlx5_dev_config *config = opaque;
1445         unsigned long mod;
1446         signed long tmp;
1447
1448         /* No-op, port representors are processed in mlx5_dev_spawn(). */
1449         if (!strcmp(MLX5_REPRESENTOR, key))
1450                 return 0;
1451         errno = 0;
1452         tmp = strtol(val, NULL, 0);
1453         if (errno) {
1454                 rte_errno = errno;
1455                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1456                 return -rte_errno;
1457         }
1458         if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1459                 /* Negative values are acceptable for some keys only. */
1460                 rte_errno = EINVAL;
1461                 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1462                 return -rte_errno;
1463         }
1464         mod = tmp >= 0 ? tmp : -tmp;
1465         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1466                 if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1467                         DRV_LOG(ERR, "invalid CQE compression "
1468                                      "format parameter");
1469                         rte_errno = EINVAL;
1470                         return -rte_errno;
1471                 }
1472                 config->cqe_comp = !!tmp;
1473                 config->cqe_comp_fmt = tmp;
1474         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1475                 config->cqe_pad = !!tmp;
1476         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1477                 config->hw_padding = !!tmp;
1478         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1479                 config->mprq.enabled = !!tmp;
1480         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1481                 config->mprq.stride_num_n = tmp;
1482         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1483                 config->mprq.stride_size_n = tmp;
1484         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1485                 config->mprq.max_memcpy_len = tmp;
1486         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1487                 config->mprq.min_rxqs_num = tmp;
1488         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1489                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1490                                  " converted to txq_inline_max", key);
1491                 config->txq_inline_max = tmp;
1492         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1493                 config->txq_inline_max = tmp;
1494         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1495                 config->txq_inline_min = tmp;
1496         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1497                 config->txq_inline_mpw = tmp;
1498         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1499                 config->txqs_inline = tmp;
1500         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1501                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1502         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1503                 config->mps = !!tmp;
1504         } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1505                 if (tmp != MLX5_TXDB_CACHED &&
1506                     tmp != MLX5_TXDB_NCACHED &&
1507                     tmp != MLX5_TXDB_HEURISTIC) {
1508                         DRV_LOG(ERR, "invalid Tx doorbell "
1509                                      "mapping parameter");
1510                         rte_errno = EINVAL;
1511                         return -rte_errno;
1512                 }
1513                 config->dbnc = tmp;
1514         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1515                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1516         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1517                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1518                                  " converted to txq_inline_mpw", key);
1519                 config->txq_inline_mpw = tmp;
1520         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1521                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1522         } else if (strcmp(MLX5_TX_PP, key) == 0) {
1523                 if (!mod) {
1524                         DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1525                         rte_errno = EINVAL;
1526                         return -rte_errno;
1527                 }
1528                 config->tx_pp = tmp;
1529         } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1530                 config->tx_skew = tmp;
1531         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1532                 config->rx_vec_en = !!tmp;
1533         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1534                 config->l3_vxlan_en = !!tmp;
1535         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1536                 config->vf_nl_en = !!tmp;
1537         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1538                 config->dv_esw_en = !!tmp;
1539         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1540                 config->dv_flow_en = !!tmp;
1541         } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1542                 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1543                     tmp != MLX5_XMETA_MODE_META16 &&
1544                     tmp != MLX5_XMETA_MODE_META32 &&
1545                     tmp != MLX5_XMETA_MODE_MISS_INFO) {
1546                         DRV_LOG(ERR, "invalid extensive "
1547                                      "metadata parameter");
1548                         rte_errno = EINVAL;
1549                         return -rte_errno;
1550                 }
1551                 if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1552                         config->dv_xmeta_en = tmp;
1553                 else
1554                         config->dv_miss_info = 1;
1555         } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1556                 config->lacp_by_user = !!tmp;
1557         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1558                 config->mr_ext_memseg_en = !!tmp;
1559         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1560                 config->max_dump_files_num = tmp;
1561         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1562                 config->lro.timeout = tmp;
1563         } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1564                 DRV_LOG(DEBUG, "class argument is %s.", val);
1565         } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1566                 config->log_hp_size = tmp;
1567         } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1568                 if (tmp != MLX5_RCM_NONE &&
1569                     tmp != MLX5_RCM_LIGHT &&
1570                     tmp != MLX5_RCM_AGGR) {
1571                         DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1572                         rte_errno = EINVAL;
1573                         return -rte_errno;
1574                 }
1575                 config->reclaim_mode = tmp;
1576         } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1577                 config->sys_mem_en = !!tmp;
1578         } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1579                 config->decap_en = !!tmp;
1580         } else {
1581                 DRV_LOG(WARNING, "%s: unknown parameter", key);
1582                 rte_errno = EINVAL;
1583                 return -rte_errno;
1584         }
1585         return 0;
1586 }
1587
1588 /**
1589  * Parse device parameters.
1590  *
1591  * @param config
1592  *   Pointer to device configuration structure.
1593  * @param devargs
1594  *   Device arguments structure.
1595  *
1596  * @return
1597  *   0 on success, a negative errno value otherwise and rte_errno is set.
1598  */
1599 int
1600 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1601 {
1602         const char **params = (const char *[]){
1603                 MLX5_RXQ_CQE_COMP_EN,
1604                 MLX5_RXQ_CQE_PAD_EN,
1605                 MLX5_RXQ_PKT_PAD_EN,
1606                 MLX5_RX_MPRQ_EN,
1607                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1608                 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1609                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1610                 MLX5_RXQS_MIN_MPRQ,
1611                 MLX5_TXQ_INLINE,
1612                 MLX5_TXQ_INLINE_MIN,
1613                 MLX5_TXQ_INLINE_MAX,
1614                 MLX5_TXQ_INLINE_MPW,
1615                 MLX5_TXQS_MIN_INLINE,
1616                 MLX5_TXQS_MAX_VEC,
1617                 MLX5_TXQ_MPW_EN,
1618                 MLX5_TXQ_MPW_HDR_DSEG_EN,
1619                 MLX5_TXQ_MAX_INLINE_LEN,
1620                 MLX5_TX_DB_NC,
1621                 MLX5_TX_PP,
1622                 MLX5_TX_SKEW,
1623                 MLX5_TX_VEC_EN,
1624                 MLX5_RX_VEC_EN,
1625                 MLX5_L3_VXLAN_EN,
1626                 MLX5_VF_NL_EN,
1627                 MLX5_DV_ESW_EN,
1628                 MLX5_DV_FLOW_EN,
1629                 MLX5_DV_XMETA_EN,
1630                 MLX5_LACP_BY_USER,
1631                 MLX5_MR_EXT_MEMSEG_EN,
1632                 MLX5_REPRESENTOR,
1633                 MLX5_MAX_DUMP_FILES_NUM,
1634                 MLX5_LRO_TIMEOUT_USEC,
1635                 MLX5_CLASS_ARG_NAME,
1636                 MLX5_HP_BUF_SIZE,
1637                 MLX5_RECLAIM_MEM,
1638                 MLX5_SYS_MEM_EN,
1639                 MLX5_DECAP_EN,
1640                 NULL,
1641         };
1642         struct rte_kvargs *kvlist;
1643         int ret = 0;
1644         int i;
1645
1646         if (devargs == NULL)
1647                 return 0;
1648         /* Following UGLY cast is done to pass checkpatch. */
1649         kvlist = rte_kvargs_parse(devargs->args, params);
1650         if (kvlist == NULL) {
1651                 rte_errno = EINVAL;
1652                 return -rte_errno;
1653         }
1654         /* Process parameters. */
1655         for (i = 0; (params[i] != NULL); ++i) {
1656                 if (rte_kvargs_count(kvlist, params[i])) {
1657                         ret = rte_kvargs_process(kvlist, params[i],
1658                                                  mlx5_args_check, config);
1659                         if (ret) {
1660                                 rte_errno = EINVAL;
1661                                 rte_kvargs_free(kvlist);
1662                                 return -rte_errno;
1663                         }
1664                 }
1665         }
1666         rte_kvargs_free(kvlist);
1667         return 0;
1668 }
1669
1670 /**
1671  * Configures the minimal amount of data to inline into WQE
1672  * while sending packets.
1673  *
1674  * - the txq_inline_min has the maximal priority, if this
1675  *   key is specified in devargs
1676  * - if DevX is enabled the inline mode is queried from the
1677  *   device (HCA attributes and NIC vport context if needed).
1678  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1679  *   and none (0 bytes) for other NICs
1680  *
1681  * @param spawn
1682  *   Verbs device parameters (name, port, switch_info) to spawn.
1683  * @param config
1684  *   Device configuration parameters.
1685  */
1686 void
1687 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1688                     struct mlx5_dev_config *config)
1689 {
1690         if (config->txq_inline_min != MLX5_ARG_UNSET) {
1691                 /* Application defines size of inlined data explicitly. */
1692                 switch (spawn->pci_dev->id.device_id) {
1693                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1694                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1695                         if (config->txq_inline_min <
1696                                        (int)MLX5_INLINE_HSIZE_L2) {
1697                                 DRV_LOG(DEBUG,
1698                                         "txq_inline_mix aligned to minimal"
1699                                         " ConnectX-4 required value %d",
1700                                         (int)MLX5_INLINE_HSIZE_L2);
1701                                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1702                         }
1703                         break;
1704                 }
1705                 goto exit;
1706         }
1707         if (config->hca_attr.eth_net_offloads) {
1708                 /* We have DevX enabled, inline mode queried successfully. */
1709                 switch (config->hca_attr.wqe_inline_mode) {
1710                 case MLX5_CAP_INLINE_MODE_L2:
1711                         /* outer L2 header must be inlined. */
1712                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1713                         goto exit;
1714                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1715                         /* No inline data are required by NIC. */
1716                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1717                         config->hw_vlan_insert =
1718                                 config->hca_attr.wqe_vlan_insert;
1719                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1720                         goto exit;
1721                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1722                         /* inline mode is defined by NIC vport context. */
1723                         if (!config->hca_attr.eth_virt)
1724                                 break;
1725                         switch (config->hca_attr.vport_inline_mode) {
1726                         case MLX5_INLINE_MODE_NONE:
1727                                 config->txq_inline_min =
1728                                         MLX5_INLINE_HSIZE_NONE;
1729                                 goto exit;
1730                         case MLX5_INLINE_MODE_L2:
1731                                 config->txq_inline_min =
1732                                         MLX5_INLINE_HSIZE_L2;
1733                                 goto exit;
1734                         case MLX5_INLINE_MODE_IP:
1735                                 config->txq_inline_min =
1736                                         MLX5_INLINE_HSIZE_L3;
1737                                 goto exit;
1738                         case MLX5_INLINE_MODE_TCP_UDP:
1739                                 config->txq_inline_min =
1740                                         MLX5_INLINE_HSIZE_L4;
1741                                 goto exit;
1742                         case MLX5_INLINE_MODE_INNER_L2:
1743                                 config->txq_inline_min =
1744                                         MLX5_INLINE_HSIZE_INNER_L2;
1745                                 goto exit;
1746                         case MLX5_INLINE_MODE_INNER_IP:
1747                                 config->txq_inline_min =
1748                                         MLX5_INLINE_HSIZE_INNER_L3;
1749                                 goto exit;
1750                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
1751                                 config->txq_inline_min =
1752                                         MLX5_INLINE_HSIZE_INNER_L4;
1753                                 goto exit;
1754                         }
1755                 }
1756         }
1757         /*
1758          * We get here if we are unable to deduce
1759          * inline data size with DevX. Try PCI ID
1760          * to determine old NICs.
1761          */
1762         switch (spawn->pci_dev->id.device_id) {
1763         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1764         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1765         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1766         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1767                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1768                 config->hw_vlan_insert = 0;
1769                 break;
1770         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1771         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1772         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1773         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1774                 /*
1775                  * These NICs support VLAN insertion from WQE and
1776                  * report the wqe_vlan_insert flag. But there is the bug
1777                  * and PFC control may be broken, so disable feature.
1778                  */
1779                 config->hw_vlan_insert = 0;
1780                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1781                 break;
1782         default:
1783                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1784                 break;
1785         }
1786 exit:
1787         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1788 }
1789
1790 /**
1791  * Configures the metadata mask fields in the shared context.
1792  *
1793  * @param [in] dev
1794  *   Pointer to Ethernet device.
1795  */
1796 void
1797 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1798 {
1799         struct mlx5_priv *priv = dev->data->dev_private;
1800         struct mlx5_dev_ctx_shared *sh = priv->sh;
1801         uint32_t meta, mark, reg_c0;
1802
1803         reg_c0 = ~priv->vport_meta_mask;
1804         switch (priv->config.dv_xmeta_en) {
1805         case MLX5_XMETA_MODE_LEGACY:
1806                 meta = UINT32_MAX;
1807                 mark = MLX5_FLOW_MARK_MASK;
1808                 break;
1809         case MLX5_XMETA_MODE_META16:
1810                 meta = reg_c0 >> rte_bsf32(reg_c0);
1811                 mark = MLX5_FLOW_MARK_MASK;
1812                 break;
1813         case MLX5_XMETA_MODE_META32:
1814                 meta = UINT32_MAX;
1815                 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1816                 break;
1817         default:
1818                 meta = 0;
1819                 mark = 0;
1820                 MLX5_ASSERT(false);
1821                 break;
1822         }
1823         if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1824                 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1825                                  sh->dv_mark_mask, mark);
1826         else
1827                 sh->dv_mark_mask = mark;
1828         if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1829                 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1830                                  sh->dv_meta_mask, meta);
1831         else
1832                 sh->dv_meta_mask = meta;
1833         if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1834                 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1835                                  sh->dv_meta_mask, reg_c0);
1836         else
1837                 sh->dv_regc0_mask = reg_c0;
1838         DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1839         DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1840         DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1841         DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1842 }
1843
1844 int
1845 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
1846 {
1847         static const char *const dynf_names[] = {
1848                 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
1849                 RTE_MBUF_DYNFLAG_METADATA_NAME,
1850                 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
1851         };
1852         unsigned int i;
1853
1854         if (n < RTE_DIM(dynf_names))
1855                 return -ENOMEM;
1856         for (i = 0; i < RTE_DIM(dynf_names); i++) {
1857                 if (names[i] == NULL)
1858                         return -EINVAL;
1859                 strcpy(names[i], dynf_names[i]);
1860         }
1861         return RTE_DIM(dynf_names);
1862 }
1863
1864 /**
1865  * Comparison callback to sort device data.
1866  *
1867  * This is meant to be used with qsort().
1868  *
1869  * @param a[in]
1870  *   Pointer to pointer to first data object.
1871  * @param b[in]
1872  *   Pointer to pointer to second data object.
1873  *
1874  * @return
1875  *   0 if both objects are equal, less than 0 if the first argument is less
1876  *   than the second, greater than 0 otherwise.
1877  */
1878 int
1879 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1880                               struct mlx5_dev_config *config)
1881 {
1882         struct mlx5_dev_ctx_shared *sh = priv->sh;
1883         struct mlx5_dev_config *sh_conf = NULL;
1884         uint16_t port_id;
1885
1886         MLX5_ASSERT(sh);
1887         /* Nothing to compare for the single/first device. */
1888         if (sh->refcnt == 1)
1889                 return 0;
1890         /* Find the device with shared context. */
1891         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1892                 struct mlx5_priv *opriv =
1893                         rte_eth_devices[port_id].data->dev_private;
1894
1895                 if (opriv && opriv != priv && opriv->sh == sh) {
1896                         sh_conf = &opriv->config;
1897                         break;
1898                 }
1899         }
1900         if (!sh_conf)
1901                 return 0;
1902         if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1903                 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1904                              " for shared %s context", sh->ibdev_name);
1905                 rte_errno = EINVAL;
1906                 return rte_errno;
1907         }
1908         if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
1909                 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
1910                              " for shared %s context", sh->ibdev_name);
1911                 rte_errno = EINVAL;
1912                 return rte_errno;
1913         }
1914         return 0;
1915 }
1916
1917 /**
1918  * Look for the ethernet device belonging to mlx5 driver.
1919  *
1920  * @param[in] port_id
1921  *   port_id to start looking for device.
1922  * @param[in] pci_dev
1923  *   Pointer to the hint PCI device. When device is being probed
1924  *   the its siblings (master and preceding representors might
1925  *   not have assigned driver yet (because the mlx5_os_pci_probe()
1926  *   is not completed yet, for this case match on hint PCI
1927  *   device may be used to detect sibling device.
1928  *
1929  * @return
1930  *   port_id of found device, RTE_MAX_ETHPORT if not found.
1931  */
1932 uint16_t
1933 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
1934 {
1935         while (port_id < RTE_MAX_ETHPORTS) {
1936                 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
1937
1938                 if (dev->state != RTE_ETH_DEV_UNUSED &&
1939                     dev->device &&
1940                     (dev->device == &pci_dev->device ||
1941                      (dev->device->driver &&
1942                      dev->device->driver->name &&
1943                      !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
1944                         break;
1945                 port_id++;
1946         }
1947         if (port_id >= RTE_MAX_ETHPORTS)
1948                 return RTE_MAX_ETHPORTS;
1949         return port_id;
1950 }
1951
1952 /**
1953  * DPDK callback to remove a PCI device.
1954  *
1955  * This function removes all Ethernet devices belong to a given PCI device.
1956  *
1957  * @param[in] pci_dev
1958  *   Pointer to the PCI device.
1959  *
1960  * @return
1961  *   0 on success, the function cannot fail.
1962  */
1963 static int
1964 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1965 {
1966         uint16_t port_id;
1967         int ret = 0;
1968
1969         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
1970                 /*
1971                  * mlx5_dev_close() is not registered to secondary process,
1972                  * call the close function explicitly for secondary process.
1973                  */
1974                 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1975                         ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
1976                 else
1977                         ret |= rte_eth_dev_close(port_id);
1978         }
1979         return ret == 0 ? 0 : -EIO;
1980 }
1981
1982 static const struct rte_pci_id mlx5_pci_id_map[] = {
1983         {
1984                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1985                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1986         },
1987         {
1988                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1989                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1990         },
1991         {
1992                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1993                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1994         },
1995         {
1996                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1997                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1998         },
1999         {
2000                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2001                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2002         },
2003         {
2004                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2005                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2006         },
2007         {
2008                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2009                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2010         },
2011         {
2012                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2013                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2014         },
2015         {
2016                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2017                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2018         },
2019         {
2020                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2021                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2022         },
2023         {
2024                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2025                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2026         },
2027         {
2028                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2029                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2030         },
2031         {
2032                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2033                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2034         },
2035         {
2036                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2037                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
2038         },
2039         {
2040                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2041                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2042         },
2043         {
2044                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2045                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2046         },
2047         {
2048                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2049                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2050         },
2051         {
2052                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2053                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2054         },
2055         {
2056                 .vendor_id = 0
2057         }
2058 };
2059
2060 static struct mlx5_pci_driver mlx5_driver = {
2061         .driver_class = MLX5_CLASS_NET,
2062         .pci_driver = {
2063                 .driver = {
2064                         .name = MLX5_DRIVER_NAME,
2065                 },
2066                 .id_table = mlx5_pci_id_map,
2067                 .probe = mlx5_os_pci_probe,
2068                 .remove = mlx5_pci_remove,
2069                 .dma_map = mlx5_dma_map,
2070                 .dma_unmap = mlx5_dma_unmap,
2071                 .drv_flags = PCI_DRV_FLAGS,
2072         },
2073 };
2074
2075 /* Initialize driver log type. */
2076 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
2077
2078 /**
2079  * Driver initialization routine.
2080  */
2081 RTE_INIT(rte_mlx5_pmd_init)
2082 {
2083         mlx5_common_init();
2084         /* Build the static tables for Verbs conversion. */
2085         mlx5_set_ptype_table();
2086         mlx5_set_cksum_table();
2087         mlx5_set_swp_types_table();
2088         if (mlx5_glue)
2089                 mlx5_pci_driver_register(&mlx5_driver);
2090 }
2091
2092 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2093 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2094 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");