net/mlx5: switch to the shared IB device context
[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 <assert.h>
10 #include <dlfcn.h>
11 #include <stdint.h>
12 #include <stdlib.h>
13 #include <errno.h>
14 #include <net/if.h>
15 #include <sys/mman.h>
16 #include <linux/rtnetlink.h>
17
18 /* Verbs header. */
19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
20 #ifdef PEDANTIC
21 #pragma GCC diagnostic ignored "-Wpedantic"
22 #endif
23 #include <infiniband/verbs.h>
24 #ifdef PEDANTIC
25 #pragma GCC diagnostic error "-Wpedantic"
26 #endif
27
28 #include <rte_malloc.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_ethdev_pci.h>
31 #include <rte_pci.h>
32 #include <rte_bus_pci.h>
33 #include <rte_common.h>
34 #include <rte_config.h>
35 #include <rte_eal_memconfig.h>
36 #include <rte_kvargs.h>
37 #include <rte_rwlock.h>
38 #include <rte_spinlock.h>
39 #include <rte_string_fns.h>
40
41 #include "mlx5.h"
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_autoconf.h"
45 #include "mlx5_defs.h"
46 #include "mlx5_glue.h"
47 #include "mlx5_mr.h"
48 #include "mlx5_flow.h"
49
50 /* Device parameter to enable RX completion queue compression. */
51 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
52
53 /* Device parameter to enable RX completion entry padding to 128B. */
54 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
55
56 /* Device parameter to enable padding Rx packet to cacheline size. */
57 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
58
59 /* Device parameter to enable Multi-Packet Rx queue. */
60 #define MLX5_RX_MPRQ_EN "mprq_en"
61
62 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
63 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
64
65 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
66 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
67
68 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
69 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
70
71 /* Device parameter to configure inline send. */
72 #define MLX5_TXQ_INLINE "txq_inline"
73
74 /*
75  * Device parameter to configure the number of TX queues threshold for
76  * enabling inline send.
77  */
78 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
79
80 /*
81  * Device parameter to configure the number of TX queues threshold for
82  * enabling vectorized Tx.
83  */
84 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
85
86 /* Device parameter to enable multi-packet send WQEs. */
87 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
88
89 /* Device parameter to include 2 dsegs in the title WQEBB. */
90 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
91
92 /* Device parameter to limit the size of inlining packet. */
93 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
94
95 /* Device parameter to enable hardware Tx vector. */
96 #define MLX5_TX_VEC_EN "tx_vec_en"
97
98 /* Device parameter to enable hardware Rx vector. */
99 #define MLX5_RX_VEC_EN "rx_vec_en"
100
101 /* Allow L3 VXLAN flow creation. */
102 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
103
104 /* Activate DV flow steering. */
105 #define MLX5_DV_FLOW_EN "dv_flow_en"
106
107 /* Activate Netlink support in VF mode. */
108 #define MLX5_VF_NL_EN "vf_nl_en"
109
110 /* Select port representors to instantiate. */
111 #define MLX5_REPRESENTOR "representor"
112
113 #ifndef HAVE_IBV_MLX5_MOD_MPW
114 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
115 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
116 #endif
117
118 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
119 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
120 #endif
121
122 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
123
124 /* Shared memory between primary and secondary processes. */
125 struct mlx5_shared_data *mlx5_shared_data;
126
127 /* Spinlock for mlx5_shared_data allocation. */
128 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
129
130 /** Driver-specific log messages type. */
131 int mlx5_logtype;
132
133 /** Data associated with devices to spawn. */
134 struct mlx5_dev_spawn_data {
135         uint32_t ifindex; /**< Network interface index. */
136         uint32_t max_port; /**< IB device maximal port index. */
137         uint32_t ibv_port; /**< IB device physical port index. */
138         struct mlx5_switch_info info; /**< Switch information. */
139         struct ibv_device *ibv_dev; /**< Associated IB device. */
140         struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
141 };
142
143 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
144 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
145
146 /**
147  * Allocate shared IB device context. If there is multiport device the
148  * master and representors will share this context, if there is single
149  * port dedicated IB device, the context will be used by only given
150  * port due to unification.
151  *
152  * Routine first searches the context for the spesified IB device name,
153  * if found the shared context assumed and reference counter is incremented.
154  * If no context found the new one is created and initialized with specified
155  * IB device context and parameters.
156  *
157  * @param[in] spawn
158  *   Pointer to the IB device attributes (name, port, etc).
159  *
160  * @return
161  *   Pointer to mlx5_ibv_shared object on success,
162  *   otherwise NULL and rte_errno is set.
163  */
164 static struct mlx5_ibv_shared *
165 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
166 {
167         struct mlx5_ibv_shared *sh;
168         int err = 0;
169
170         assert(spawn);
171         /* Secondary process should not create the shared context. */
172         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
173         pthread_mutex_lock(&mlx5_ibv_list_mutex);
174         /* Search for IB context by device name. */
175         LIST_FOREACH(sh, &mlx5_ibv_list, next) {
176                 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
177                         sh->refcnt++;
178                         goto exit;
179                 }
180         }
181         /* No device found, we have to create new sharted context. */
182         assert(spawn->max_port);
183         sh = rte_zmalloc("ethdev shared ib context",
184                          sizeof(struct mlx5_ibv_shared) +
185                          spawn->max_port *
186                          sizeof(struct mlx5_ibv_shared_port),
187                          RTE_CACHE_LINE_SIZE);
188         if (!sh) {
189                 DRV_LOG(ERR, "shared context allocation failure");
190                 rte_errno  = ENOMEM;
191                 goto exit;
192         }
193         /* Try to open IB device with DV first, then usual Verbs. */
194         errno = 0;
195         sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
196         if (sh->ctx) {
197                 sh->devx = 1;
198                 DRV_LOG(DEBUG, "DevX is supported");
199         } else {
200                 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
201                 if (!sh->ctx) {
202                         err = errno ? errno : ENODEV;
203                         goto error;
204                 }
205                 DRV_LOG(DEBUG, "DevX is NOT supported");
206         }
207         err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
208         if (err) {
209                 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
210                 goto error;
211         }
212         sh->refcnt = 1;
213         sh->max_port = spawn->max_port;
214         strncpy(sh->ibdev_name, sh->ctx->device->name,
215                 sizeof(sh->ibdev_name));
216         strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
217                 sizeof(sh->ibdev_path));
218         sh->pd = mlx5_glue->alloc_pd(sh->ctx);
219         if (sh->pd == NULL) {
220                 DRV_LOG(ERR, "PD allocation failure");
221                 err = ENOMEM;
222                 goto error;
223         }
224         LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
225 exit:
226         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
227         return sh;
228 error:
229         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
230         assert(sh);
231         if (sh->pd)
232                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
233         if (sh->ctx)
234                 claim_zero(mlx5_glue->close_device(sh->ctx));
235         rte_free(sh);
236         assert(err > 0);
237         rte_errno = err;
238         return NULL;
239 }
240
241 /**
242  * Free shared IB device context. Decrement counter and if zero free
243  * all allocated resources and close handles.
244  *
245  * @param[in] sh
246  *   Pointer to mlx5_ibv_shared object to free
247  */
248 static void
249 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
250 {
251         pthread_mutex_lock(&mlx5_ibv_list_mutex);
252 #ifndef NDEBUG
253         /* Check the object presence in the list. */
254         struct mlx5_ibv_shared *lctx;
255
256         LIST_FOREACH(lctx, &mlx5_ibv_list, next)
257                 if (lctx == sh)
258                         break;
259         assert(lctx);
260         if (lctx != sh) {
261                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
262                 goto exit;
263         }
264 #endif
265         assert(sh);
266         assert(sh->refcnt);
267         /* Secondary process should not free the shared context. */
268         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
269         if (--sh->refcnt)
270                 goto exit;
271         LIST_REMOVE(sh, next);
272         if (sh->pd)
273                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
274         if (sh->ctx)
275                 claim_zero(mlx5_glue->close_device(sh->ctx));
276         rte_free(sh);
277 exit:
278         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
279 }
280
281
282 /**
283  * Prepare shared data between primary and secondary process.
284  */
285 static void
286 mlx5_prepare_shared_data(void)
287 {
288         const struct rte_memzone *mz;
289
290         rte_spinlock_lock(&mlx5_shared_data_lock);
291         if (mlx5_shared_data == NULL) {
292                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
293                         /* Allocate shared memory. */
294                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
295                                                  sizeof(*mlx5_shared_data),
296                                                  SOCKET_ID_ANY, 0);
297                 } else {
298                         /* Lookup allocated shared memory. */
299                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
300                 }
301                 if (mz == NULL)
302                         rte_panic("Cannot allocate mlx5 shared data\n");
303                 mlx5_shared_data = mz->addr;
304                 /* Initialize shared data. */
305                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
306                         LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
307                         rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
308                 }
309                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
310                                                 mlx5_mr_mem_event_cb, NULL);
311         }
312         rte_spinlock_unlock(&mlx5_shared_data_lock);
313 }
314
315 /**
316  * Retrieve integer value from environment variable.
317  *
318  * @param[in] name
319  *   Environment variable name.
320  *
321  * @return
322  *   Integer value, 0 if the variable is not set.
323  */
324 int
325 mlx5_getenv_int(const char *name)
326 {
327         const char *val = getenv(name);
328
329         if (val == NULL)
330                 return 0;
331         return atoi(val);
332 }
333
334 /**
335  * Verbs callback to allocate a memory. This function should allocate the space
336  * according to the size provided residing inside a huge page.
337  * Please note that all allocation must respect the alignment from libmlx5
338  * (i.e. currently sysconf(_SC_PAGESIZE)).
339  *
340  * @param[in] size
341  *   The size in bytes of the memory to allocate.
342  * @param[in] data
343  *   A pointer to the callback data.
344  *
345  * @return
346  *   Allocated buffer, NULL otherwise and rte_errno is set.
347  */
348 static void *
349 mlx5_alloc_verbs_buf(size_t size, void *data)
350 {
351         struct mlx5_priv *priv = data;
352         void *ret;
353         size_t alignment = sysconf(_SC_PAGESIZE);
354         unsigned int socket = SOCKET_ID_ANY;
355
356         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
357                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
358
359                 socket = ctrl->socket;
360         } else if (priv->verbs_alloc_ctx.type ==
361                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
362                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
363
364                 socket = ctrl->socket;
365         }
366         assert(data != NULL);
367         ret = rte_malloc_socket(__func__, size, alignment, socket);
368         if (!ret && size)
369                 rte_errno = ENOMEM;
370         return ret;
371 }
372
373 /**
374  * Verbs callback to free a memory.
375  *
376  * @param[in] ptr
377  *   A pointer to the memory to free.
378  * @param[in] data
379  *   A pointer to the callback data.
380  */
381 static void
382 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
383 {
384         assert(data != NULL);
385         rte_free(ptr);
386 }
387
388 /**
389  * DPDK callback to close the device.
390  *
391  * Destroy all queues and objects, free memory.
392  *
393  * @param dev
394  *   Pointer to Ethernet device structure.
395  */
396 static void
397 mlx5_dev_close(struct rte_eth_dev *dev)
398 {
399         struct mlx5_priv *priv = dev->data->dev_private;
400         unsigned int i;
401         int ret;
402
403         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
404                 dev->data->port_id,
405                 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
406         /* In case mlx5_dev_stop() has not been called. */
407         mlx5_dev_interrupt_handler_uninstall(dev);
408         mlx5_traffic_disable(dev);
409         mlx5_flow_flush(dev, NULL);
410         /* Prevent crashes when queues are still in use. */
411         dev->rx_pkt_burst = removed_rx_burst;
412         dev->tx_pkt_burst = removed_tx_burst;
413         if (priv->rxqs != NULL) {
414                 /* XXX race condition if mlx5_rx_burst() is still running. */
415                 usleep(1000);
416                 for (i = 0; (i != priv->rxqs_n); ++i)
417                         mlx5_rxq_release(dev, i);
418                 priv->rxqs_n = 0;
419                 priv->rxqs = NULL;
420         }
421         if (priv->txqs != NULL) {
422                 /* XXX race condition if mlx5_tx_burst() is still running. */
423                 usleep(1000);
424                 for (i = 0; (i != priv->txqs_n); ++i)
425                         mlx5_txq_release(dev, i);
426                 priv->txqs_n = 0;
427                 priv->txqs = NULL;
428         }
429         mlx5_mprq_free_mp(dev);
430         mlx5_mr_release(dev);
431         assert(priv->sh);
432         if (priv->sh)
433                 mlx5_free_shared_ibctx(priv->sh);
434         priv->sh = NULL;
435         if (priv->rss_conf.rss_key != NULL)
436                 rte_free(priv->rss_conf.rss_key);
437         if (priv->reta_idx != NULL)
438                 rte_free(priv->reta_idx);
439         if (priv->primary_socket)
440                 mlx5_socket_uninit(dev);
441         if (priv->config.vf)
442                 mlx5_nl_mac_addr_flush(dev);
443         if (priv->nl_socket_route >= 0)
444                 close(priv->nl_socket_route);
445         if (priv->nl_socket_rdma >= 0)
446                 close(priv->nl_socket_rdma);
447         if (priv->tcf_context)
448                 mlx5_flow_tcf_context_destroy(priv->tcf_context);
449         ret = mlx5_hrxq_ibv_verify(dev);
450         if (ret)
451                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
452                         dev->data->port_id);
453         ret = mlx5_ind_table_ibv_verify(dev);
454         if (ret)
455                 DRV_LOG(WARNING, "port %u some indirection table still remain",
456                         dev->data->port_id);
457         ret = mlx5_rxq_ibv_verify(dev);
458         if (ret)
459                 DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
460                         dev->data->port_id);
461         ret = mlx5_rxq_verify(dev);
462         if (ret)
463                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
464                         dev->data->port_id);
465         ret = mlx5_txq_ibv_verify(dev);
466         if (ret)
467                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
468                         dev->data->port_id);
469         ret = mlx5_txq_verify(dev);
470         if (ret)
471                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
472                         dev->data->port_id);
473         ret = mlx5_flow_verify(dev);
474         if (ret)
475                 DRV_LOG(WARNING, "port %u some flows still remain",
476                         dev->data->port_id);
477         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
478                 unsigned int c = 0;
479                 unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
480                 uint16_t port_id[i];
481
482                 i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
483                 while (i--) {
484                         struct mlx5_priv *opriv =
485                                 rte_eth_devices[port_id[i]].data->dev_private;
486
487                         if (!opriv ||
488                             opriv->domain_id != priv->domain_id ||
489                             &rte_eth_devices[port_id[i]] == dev)
490                                 continue;
491                         ++c;
492                 }
493                 if (!c)
494                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
495         }
496         memset(priv, 0, sizeof(*priv));
497         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
498         /*
499          * Reset mac_addrs to NULL such that it is not freed as part of
500          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
501          * it is freed when dev_private is freed.
502          */
503         dev->data->mac_addrs = NULL;
504 }
505
506 const struct eth_dev_ops mlx5_dev_ops = {
507         .dev_configure = mlx5_dev_configure,
508         .dev_start = mlx5_dev_start,
509         .dev_stop = mlx5_dev_stop,
510         .dev_set_link_down = mlx5_set_link_down,
511         .dev_set_link_up = mlx5_set_link_up,
512         .dev_close = mlx5_dev_close,
513         .promiscuous_enable = mlx5_promiscuous_enable,
514         .promiscuous_disable = mlx5_promiscuous_disable,
515         .allmulticast_enable = mlx5_allmulticast_enable,
516         .allmulticast_disable = mlx5_allmulticast_disable,
517         .link_update = mlx5_link_update,
518         .stats_get = mlx5_stats_get,
519         .stats_reset = mlx5_stats_reset,
520         .xstats_get = mlx5_xstats_get,
521         .xstats_reset = mlx5_xstats_reset,
522         .xstats_get_names = mlx5_xstats_get_names,
523         .fw_version_get = mlx5_fw_version_get,
524         .dev_infos_get = mlx5_dev_infos_get,
525         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
526         .vlan_filter_set = mlx5_vlan_filter_set,
527         .rx_queue_setup = mlx5_rx_queue_setup,
528         .tx_queue_setup = mlx5_tx_queue_setup,
529         .rx_queue_release = mlx5_rx_queue_release,
530         .tx_queue_release = mlx5_tx_queue_release,
531         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
532         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
533         .mac_addr_remove = mlx5_mac_addr_remove,
534         .mac_addr_add = mlx5_mac_addr_add,
535         .mac_addr_set = mlx5_mac_addr_set,
536         .set_mc_addr_list = mlx5_set_mc_addr_list,
537         .mtu_set = mlx5_dev_set_mtu,
538         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
539         .vlan_offload_set = mlx5_vlan_offload_set,
540         .reta_update = mlx5_dev_rss_reta_update,
541         .reta_query = mlx5_dev_rss_reta_query,
542         .rss_hash_update = mlx5_rss_hash_update,
543         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
544         .filter_ctrl = mlx5_dev_filter_ctrl,
545         .rx_descriptor_status = mlx5_rx_descriptor_status,
546         .tx_descriptor_status = mlx5_tx_descriptor_status,
547         .rx_queue_count = mlx5_rx_queue_count,
548         .rx_queue_intr_enable = mlx5_rx_intr_enable,
549         .rx_queue_intr_disable = mlx5_rx_intr_disable,
550         .is_removed = mlx5_is_removed,
551 };
552
553 /* Available operations from secondary process. */
554 static const struct eth_dev_ops mlx5_dev_sec_ops = {
555         .stats_get = mlx5_stats_get,
556         .stats_reset = mlx5_stats_reset,
557         .xstats_get = mlx5_xstats_get,
558         .xstats_reset = mlx5_xstats_reset,
559         .xstats_get_names = mlx5_xstats_get_names,
560         .fw_version_get = mlx5_fw_version_get,
561         .dev_infos_get = mlx5_dev_infos_get,
562         .rx_descriptor_status = mlx5_rx_descriptor_status,
563         .tx_descriptor_status = mlx5_tx_descriptor_status,
564 };
565
566 /* Available operations in flow isolated mode. */
567 const struct eth_dev_ops mlx5_dev_ops_isolate = {
568         .dev_configure = mlx5_dev_configure,
569         .dev_start = mlx5_dev_start,
570         .dev_stop = mlx5_dev_stop,
571         .dev_set_link_down = mlx5_set_link_down,
572         .dev_set_link_up = mlx5_set_link_up,
573         .dev_close = mlx5_dev_close,
574         .promiscuous_enable = mlx5_promiscuous_enable,
575         .promiscuous_disable = mlx5_promiscuous_disable,
576         .allmulticast_enable = mlx5_allmulticast_enable,
577         .allmulticast_disable = mlx5_allmulticast_disable,
578         .link_update = mlx5_link_update,
579         .stats_get = mlx5_stats_get,
580         .stats_reset = mlx5_stats_reset,
581         .xstats_get = mlx5_xstats_get,
582         .xstats_reset = mlx5_xstats_reset,
583         .xstats_get_names = mlx5_xstats_get_names,
584         .fw_version_get = mlx5_fw_version_get,
585         .dev_infos_get = mlx5_dev_infos_get,
586         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
587         .vlan_filter_set = mlx5_vlan_filter_set,
588         .rx_queue_setup = mlx5_rx_queue_setup,
589         .tx_queue_setup = mlx5_tx_queue_setup,
590         .rx_queue_release = mlx5_rx_queue_release,
591         .tx_queue_release = mlx5_tx_queue_release,
592         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
593         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
594         .mac_addr_remove = mlx5_mac_addr_remove,
595         .mac_addr_add = mlx5_mac_addr_add,
596         .mac_addr_set = mlx5_mac_addr_set,
597         .set_mc_addr_list = mlx5_set_mc_addr_list,
598         .mtu_set = mlx5_dev_set_mtu,
599         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
600         .vlan_offload_set = mlx5_vlan_offload_set,
601         .filter_ctrl = mlx5_dev_filter_ctrl,
602         .rx_descriptor_status = mlx5_rx_descriptor_status,
603         .tx_descriptor_status = mlx5_tx_descriptor_status,
604         .rx_queue_intr_enable = mlx5_rx_intr_enable,
605         .rx_queue_intr_disable = mlx5_rx_intr_disable,
606         .is_removed = mlx5_is_removed,
607 };
608
609 /**
610  * Verify and store value for device argument.
611  *
612  * @param[in] key
613  *   Key argument to verify.
614  * @param[in] val
615  *   Value associated with key.
616  * @param opaque
617  *   User data.
618  *
619  * @return
620  *   0 on success, a negative errno value otherwise and rte_errno is set.
621  */
622 static int
623 mlx5_args_check(const char *key, const char *val, void *opaque)
624 {
625         struct mlx5_dev_config *config = opaque;
626         unsigned long tmp;
627
628         /* No-op, port representors are processed in mlx5_dev_spawn(). */
629         if (!strcmp(MLX5_REPRESENTOR, key))
630                 return 0;
631         errno = 0;
632         tmp = strtoul(val, NULL, 0);
633         if (errno) {
634                 rte_errno = errno;
635                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
636                 return -rte_errno;
637         }
638         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
639                 config->cqe_comp = !!tmp;
640         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
641                 config->cqe_pad = !!tmp;
642         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
643                 config->hw_padding = !!tmp;
644         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
645                 config->mprq.enabled = !!tmp;
646         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
647                 config->mprq.stride_num_n = tmp;
648         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
649                 config->mprq.max_memcpy_len = tmp;
650         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
651                 config->mprq.min_rxqs_num = tmp;
652         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
653                 config->txq_inline = tmp;
654         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
655                 config->txqs_inline = tmp;
656         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
657                 config->txqs_vec = tmp;
658         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
659                 config->mps = !!tmp;
660         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
661                 config->mpw_hdr_dseg = !!tmp;
662         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
663                 config->inline_max_packet_sz = tmp;
664         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
665                 config->tx_vec_en = !!tmp;
666         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
667                 config->rx_vec_en = !!tmp;
668         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
669                 config->l3_vxlan_en = !!tmp;
670         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
671                 config->vf_nl_en = !!tmp;
672         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
673                 config->dv_flow_en = !!tmp;
674         } else {
675                 DRV_LOG(WARNING, "%s: unknown parameter", key);
676                 rte_errno = EINVAL;
677                 return -rte_errno;
678         }
679         return 0;
680 }
681
682 /**
683  * Parse device parameters.
684  *
685  * @param config
686  *   Pointer to device configuration structure.
687  * @param devargs
688  *   Device arguments structure.
689  *
690  * @return
691  *   0 on success, a negative errno value otherwise and rte_errno is set.
692  */
693 static int
694 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
695 {
696         const char **params = (const char *[]){
697                 MLX5_RXQ_CQE_COMP_EN,
698                 MLX5_RXQ_CQE_PAD_EN,
699                 MLX5_RXQ_PKT_PAD_EN,
700                 MLX5_RX_MPRQ_EN,
701                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
702                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
703                 MLX5_RXQS_MIN_MPRQ,
704                 MLX5_TXQ_INLINE,
705                 MLX5_TXQS_MIN_INLINE,
706                 MLX5_TXQS_MAX_VEC,
707                 MLX5_TXQ_MPW_EN,
708                 MLX5_TXQ_MPW_HDR_DSEG_EN,
709                 MLX5_TXQ_MAX_INLINE_LEN,
710                 MLX5_TX_VEC_EN,
711                 MLX5_RX_VEC_EN,
712                 MLX5_L3_VXLAN_EN,
713                 MLX5_VF_NL_EN,
714                 MLX5_DV_FLOW_EN,
715                 MLX5_REPRESENTOR,
716                 NULL,
717         };
718         struct rte_kvargs *kvlist;
719         int ret = 0;
720         int i;
721
722         if (devargs == NULL)
723                 return 0;
724         /* Following UGLY cast is done to pass checkpatch. */
725         kvlist = rte_kvargs_parse(devargs->args, params);
726         if (kvlist == NULL)
727                 return 0;
728         /* Process parameters. */
729         for (i = 0; (params[i] != NULL); ++i) {
730                 if (rte_kvargs_count(kvlist, params[i])) {
731                         ret = rte_kvargs_process(kvlist, params[i],
732                                                  mlx5_args_check, config);
733                         if (ret) {
734                                 rte_errno = EINVAL;
735                                 rte_kvargs_free(kvlist);
736                                 return -rte_errno;
737                         }
738                 }
739         }
740         rte_kvargs_free(kvlist);
741         return 0;
742 }
743
744 static struct rte_pci_driver mlx5_driver;
745
746 /*
747  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
748  * local resource used by both primary and secondary to avoid duplicate
749  * reservation.
750  * The space has to be available on both primary and secondary process,
751  * TXQ UAR maps to this area using fixed mmap w/o double check.
752  */
753 static void *uar_base;
754
755 static int
756 find_lower_va_bound(const struct rte_memseg_list *msl,
757                 const struct rte_memseg *ms, void *arg)
758 {
759         void **addr = arg;
760
761         if (msl->external)
762                 return 0;
763         if (*addr == NULL)
764                 *addr = ms->addr;
765         else
766                 *addr = RTE_MIN(*addr, ms->addr);
767
768         return 0;
769 }
770
771 /**
772  * Reserve UAR address space for primary process.
773  *
774  * @param[in] dev
775  *   Pointer to Ethernet device.
776  *
777  * @return
778  *   0 on success, a negative errno value otherwise and rte_errno is set.
779  */
780 static int
781 mlx5_uar_init_primary(struct rte_eth_dev *dev)
782 {
783         struct mlx5_priv *priv = dev->data->dev_private;
784         void *addr = (void *)0;
785
786         if (uar_base) { /* UAR address space mapped. */
787                 priv->uar_base = uar_base;
788                 return 0;
789         }
790         /* find out lower bound of hugepage segments */
791         rte_memseg_walk(find_lower_va_bound, &addr);
792
793         /* keep distance to hugepages to minimize potential conflicts. */
794         addr = RTE_PTR_SUB(addr, (uintptr_t)(MLX5_UAR_OFFSET + MLX5_UAR_SIZE));
795         /* anonymous mmap, no real memory consumption. */
796         addr = mmap(addr, MLX5_UAR_SIZE,
797                     PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
798         if (addr == MAP_FAILED) {
799                 DRV_LOG(ERR,
800                         "port %u failed to reserve UAR address space, please"
801                         " adjust MLX5_UAR_SIZE or try --base-virtaddr",
802                         dev->data->port_id);
803                 rte_errno = ENOMEM;
804                 return -rte_errno;
805         }
806         /* Accept either same addr or a new addr returned from mmap if target
807          * range occupied.
808          */
809         DRV_LOG(INFO, "port %u reserved UAR address space: %p",
810                 dev->data->port_id, addr);
811         priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
812         uar_base = addr; /* process local, don't reserve again. */
813         return 0;
814 }
815
816 /**
817  * Reserve UAR address space for secondary process, align with
818  * primary process.
819  *
820  * @param[in] dev
821  *   Pointer to Ethernet device.
822  *
823  * @return
824  *   0 on success, a negative errno value otherwise and rte_errno is set.
825  */
826 static int
827 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
828 {
829         struct mlx5_priv *priv = dev->data->dev_private;
830         void *addr;
831
832         assert(priv->uar_base);
833         if (uar_base) { /* already reserved. */
834                 assert(uar_base == priv->uar_base);
835                 return 0;
836         }
837         /* anonymous mmap, no real memory consumption. */
838         addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
839                     PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
840         if (addr == MAP_FAILED) {
841                 DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
842                         dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
843                 rte_errno = ENXIO;
844                 return -rte_errno;
845         }
846         if (priv->uar_base != addr) {
847                 DRV_LOG(ERR,
848                         "port %u UAR address %p size %llu occupied, please"
849                         " adjust MLX5_UAR_OFFSET or try EAL parameter"
850                         " --base-virtaddr",
851                         dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
852                 rte_errno = ENXIO;
853                 return -rte_errno;
854         }
855         uar_base = addr; /* process local, don't reserve again */
856         DRV_LOG(INFO, "port %u reserved UAR address space: %p",
857                 dev->data->port_id, addr);
858         return 0;
859 }
860
861 /**
862  * Spawn an Ethernet device from Verbs information.
863  *
864  * @param dpdk_dev
865  *   Backing DPDK device.
866  * @param spawn
867  *   Verbs device parameters (name, port, switch_info) to spawn.
868  * @param config
869  *   Device configuration parameters.
870  *
871  * @return
872  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
873  *   is set. The following errors are defined:
874  *
875  *   EBUSY: device is not supposed to be spawned.
876  *   EEXIST: device is already spawned
877  */
878 static struct rte_eth_dev *
879 mlx5_dev_spawn(struct rte_device *dpdk_dev,
880                struct mlx5_dev_spawn_data *spawn,
881                struct mlx5_dev_config config)
882 {
883         const struct mlx5_switch_info *switch_info = &spawn->info;
884         struct mlx5_ibv_shared *sh = NULL;
885         struct ibv_port_attr port_attr;
886         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
887         struct rte_eth_dev *eth_dev = NULL;
888         struct mlx5_priv *priv = NULL;
889         int err = 0;
890         unsigned int hw_padding = 0;
891         unsigned int mps;
892         unsigned int cqe_comp;
893         unsigned int cqe_pad = 0;
894         unsigned int tunnel_en = 0;
895         unsigned int mpls_en = 0;
896         unsigned int swp = 0;
897         unsigned int mprq = 0;
898         unsigned int mprq_min_stride_size_n = 0;
899         unsigned int mprq_max_stride_size_n = 0;
900         unsigned int mprq_min_stride_num_n = 0;
901         unsigned int mprq_max_stride_num_n = 0;
902         struct ether_addr mac;
903         char name[RTE_ETH_NAME_MAX_LEN];
904         int own_domain_id = 0;
905         uint16_t port_id;
906         unsigned int i;
907
908         /* Determine if this port representor is supposed to be spawned. */
909         if (switch_info->representor && dpdk_dev->devargs) {
910                 struct rte_eth_devargs eth_da;
911
912                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
913                 if (err) {
914                         rte_errno = -err;
915                         DRV_LOG(ERR, "failed to process device arguments: %s",
916                                 strerror(rte_errno));
917                         return NULL;
918                 }
919                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
920                         if (eth_da.representor_ports[i] ==
921                             (uint16_t)switch_info->port_name)
922                                 break;
923                 if (i == eth_da.nb_representor_ports) {
924                         rte_errno = EBUSY;
925                         return NULL;
926                 }
927         }
928         /* Build device name. */
929         if (!switch_info->representor)
930                 strlcpy(name, dpdk_dev->name, sizeof(name));
931         else
932                 snprintf(name, sizeof(name), "%s_representor_%u",
933                          dpdk_dev->name, switch_info->port_name);
934         /* check if the device is already spawned */
935         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
936                 rte_errno = EEXIST;
937                 return NULL;
938         }
939         /* Prepare shared data between primary and secondary process. */
940         mlx5_prepare_shared_data();
941         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
942         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
943                 eth_dev = rte_eth_dev_attach_secondary(name);
944                 if (eth_dev == NULL) {
945                         DRV_LOG(ERR, "can not attach rte ethdev");
946                         rte_errno = ENOMEM;
947                         return NULL;
948                 }
949                 eth_dev->device = dpdk_dev;
950                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
951                 err = mlx5_uar_init_secondary(eth_dev);
952                 if (err)
953                         return NULL;
954                 /* Receive command fd from primary process */
955                 err = mlx5_socket_connect(eth_dev);
956                 if (err < 0)
957                         return NULL;
958                 /* Remap UAR for Tx queues. */
959                 err = mlx5_tx_uar_remap(eth_dev, err);
960                 if (err)
961                         return NULL;
962                 /*
963                  * Ethdev pointer is still required as input since
964                  * the primary device is not accessible from the
965                  * secondary process.
966                  */
967                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
968                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
969                 return eth_dev;
970         }
971         sh = mlx5_alloc_shared_ibctx(spawn);
972         if (!sh)
973                 return NULL;
974         config.devx = sh->devx;
975 #ifdef HAVE_IBV_MLX5_MOD_SWP
976         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
977 #endif
978         /*
979          * Multi-packet send is supported by ConnectX-4 Lx PF as well
980          * as all ConnectX-5 devices.
981          */
982 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
983         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
984 #endif
985 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
986         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
987 #endif
988         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
989         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
990                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
991                         DRV_LOG(DEBUG, "enhanced MPW is supported");
992                         mps = MLX5_MPW_ENHANCED;
993                 } else {
994                         DRV_LOG(DEBUG, "MPW is supported");
995                         mps = MLX5_MPW;
996                 }
997         } else {
998                 DRV_LOG(DEBUG, "MPW isn't supported");
999                 mps = MLX5_MPW_DISABLED;
1000         }
1001 #ifdef HAVE_IBV_MLX5_MOD_SWP
1002         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1003                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1004         DRV_LOG(DEBUG, "SWP support: %u", swp);
1005 #endif
1006         config.swp = !!swp;
1007 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1008         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1009                 struct mlx5dv_striding_rq_caps mprq_caps =
1010                         dv_attr.striding_rq_caps;
1011
1012                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1013                         mprq_caps.min_single_stride_log_num_of_bytes);
1014                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1015                         mprq_caps.max_single_stride_log_num_of_bytes);
1016                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1017                         mprq_caps.min_single_wqe_log_num_of_strides);
1018                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1019                         mprq_caps.max_single_wqe_log_num_of_strides);
1020                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1021                         mprq_caps.supported_qpts);
1022                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1023                 mprq = 1;
1024                 mprq_min_stride_size_n =
1025                         mprq_caps.min_single_stride_log_num_of_bytes;
1026                 mprq_max_stride_size_n =
1027                         mprq_caps.max_single_stride_log_num_of_bytes;
1028                 mprq_min_stride_num_n =
1029                         mprq_caps.min_single_wqe_log_num_of_strides;
1030                 mprq_max_stride_num_n =
1031                         mprq_caps.max_single_wqe_log_num_of_strides;
1032                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1033                                                    mprq_min_stride_num_n);
1034         }
1035 #endif
1036         if (RTE_CACHE_LINE_SIZE == 128 &&
1037             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1038                 cqe_comp = 0;
1039         else
1040                 cqe_comp = 1;
1041         config.cqe_comp = cqe_comp;
1042 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1043         /* Whether device supports 128B Rx CQE padding. */
1044         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1045                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1046 #endif
1047 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1048         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1049                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
1050                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1051                              (dv_attr.tunnel_offloads_caps &
1052                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1053         }
1054         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1055                 tunnel_en ? "" : "not ");
1056 #else
1057         DRV_LOG(WARNING,
1058                 "tunnel offloading disabled due to old OFED/rdma-core version");
1059 #endif
1060         config.tunnel_en = tunnel_en;
1061 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1062         mpls_en = ((dv_attr.tunnel_offloads_caps &
1063                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1064                    (dv_attr.tunnel_offloads_caps &
1065                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1066         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1067                 mpls_en ? "" : "not ");
1068 #else
1069         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1070                 " old OFED/rdma-core version or firmware configuration");
1071 #endif
1072         config.mpls_en = mpls_en;
1073         /* Check port status. */
1074         err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1075         if (err) {
1076                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1077                 goto error;
1078         }
1079         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1080                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1081                 err = EINVAL;
1082                 goto error;
1083         }
1084         if (port_attr.state != IBV_PORT_ACTIVE)
1085                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1086                         mlx5_glue->port_state_str(port_attr.state),
1087                         port_attr.state);
1088         /* Allocate private eth device data. */
1089         priv = rte_zmalloc("ethdev private structure",
1090                            sizeof(*priv),
1091                            RTE_CACHE_LINE_SIZE);
1092         if (priv == NULL) {
1093                 DRV_LOG(ERR, "priv allocation failure");
1094                 err = ENOMEM;
1095                 goto error;
1096         }
1097         priv->sh = sh;
1098         priv->ibv_port = spawn->ibv_port;
1099         priv->mtu = ETHER_MTU;
1100 #ifndef RTE_ARCH_64
1101         /* Initialize UAR access locks for 32bit implementations. */
1102         rte_spinlock_init(&priv->uar_lock_cq);
1103         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1104                 rte_spinlock_init(&priv->uar_lock[i]);
1105 #endif
1106         /* Some internal functions rely on Netlink sockets, open them now. */
1107         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1108         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1109         priv->nl_sn = 0;
1110         priv->representor = !!switch_info->representor;
1111         priv->master = !!switch_info->master;
1112         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1113         /*
1114          * Currently we support single E-Switch per PF configurations
1115          * only and vport_id field contains the vport index for
1116          * associated VF, which is deduced from representor port name.
1117          * For exapmple, let's have the IB device port 10, it has
1118          * attached network device eth0, which has port name attribute
1119          * pf0vf2, we can deduce the VF number as 2, and set vport index
1120          * as 3 (2+1). This assigning schema should be changed if the
1121          * multiple E-Switch instances per PF configurations or/and PCI
1122          * subfunctions are added.
1123          */
1124         priv->vport_id = switch_info->representor ?
1125                          switch_info->port_name + 1 : -1;
1126         /* representor_id field keeps the unmodified port/VF index. */
1127         priv->representor_id = switch_info->representor ?
1128                                switch_info->port_name : -1;
1129         /*
1130          * Look for sibling devices in order to reuse their switch domain
1131          * if any, otherwise allocate one.
1132          */
1133         i = mlx5_dev_to_port_id(dpdk_dev, NULL, 0);
1134         if (i > 0) {
1135                 uint16_t port_id[i];
1136
1137                 i = RTE_MIN(mlx5_dev_to_port_id(dpdk_dev, port_id, i), i);
1138                 while (i--) {
1139                         const struct mlx5_priv *opriv =
1140                                 rte_eth_devices[port_id[i]].data->dev_private;
1141
1142                         if (!opriv ||
1143                             opriv->domain_id ==
1144                             RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1145                                 continue;
1146                         priv->domain_id = opriv->domain_id;
1147                         break;
1148                 }
1149         }
1150         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1151                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1152                 if (err) {
1153                         err = rte_errno;
1154                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
1155                                 strerror(rte_errno));
1156                         goto error;
1157                 }
1158                 own_domain_id = 1;
1159         }
1160         err = mlx5_args(&config, dpdk_dev->devargs);
1161         if (err) {
1162                 err = rte_errno;
1163                 DRV_LOG(ERR, "failed to process device arguments: %s",
1164                         strerror(rte_errno));
1165                 goto error;
1166         }
1167         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1168                             IBV_DEVICE_RAW_IP_CSUM);
1169         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1170                 (config.hw_csum ? "" : "not "));
1171 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1172         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1173         DRV_LOG(DEBUG, "counters are not supported");
1174 #endif
1175 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1176         if (config.dv_flow_en) {
1177                 DRV_LOG(WARNING, "DV flow is not supported");
1178                 config.dv_flow_en = 0;
1179         }
1180 #endif
1181         config.ind_table_max_size =
1182                 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1183         /*
1184          * Remove this check once DPDK supports larger/variable
1185          * indirection tables.
1186          */
1187         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1188                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1189         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1190                 config.ind_table_max_size);
1191         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1192                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1193         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1194                 (config.hw_vlan_strip ? "" : "not "));
1195         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1196                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1197         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1198                 (config.hw_fcs_strip ? "" : "not "));
1199 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1200         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1201 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1202         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1203                         IBV_DEVICE_PCI_WRITE_END_PADDING);
1204 #endif
1205         if (config.hw_padding && !hw_padding) {
1206                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1207                 config.hw_padding = 0;
1208         } else if (config.hw_padding) {
1209                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1210         }
1211         config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1212                       (sh->device_attr.tso_caps.supported_qpts &
1213                        (1 << IBV_QPT_RAW_PACKET)));
1214         if (config.tso)
1215                 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1216         /*
1217          * MPW is disabled by default, while the Enhanced MPW is enabled
1218          * by default.
1219          */
1220         if (config.mps == MLX5_ARG_UNSET)
1221                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1222                                                           MLX5_MPW_DISABLED;
1223         else
1224                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1225         DRV_LOG(INFO, "%sMPS is %s",
1226                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1227                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1228         if (config.cqe_comp && !cqe_comp) {
1229                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1230                 config.cqe_comp = 0;
1231         }
1232         if (config.cqe_pad && !cqe_pad) {
1233                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1234                 config.cqe_pad = 0;
1235         } else if (config.cqe_pad) {
1236                 DRV_LOG(INFO, "Rx CQE padding is enabled");
1237         }
1238         if (config.mprq.enabled && mprq) {
1239                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1240                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
1241                         config.mprq.stride_num_n =
1242                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1243                                         mprq_min_stride_num_n);
1244                         DRV_LOG(WARNING,
1245                                 "the number of strides"
1246                                 " for Multi-Packet RQ is out of range,"
1247                                 " setting default value (%u)",
1248                                 1 << config.mprq.stride_num_n);
1249                 }
1250                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1251                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1252         } else if (config.mprq.enabled && !mprq) {
1253                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1254                 config.mprq.enabled = 0;
1255         }
1256         eth_dev = rte_eth_dev_allocate(name);
1257         if (eth_dev == NULL) {
1258                 DRV_LOG(ERR, "can not allocate rte ethdev");
1259                 err = ENOMEM;
1260                 goto error;
1261         }
1262         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1263         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1264         if (priv->representor) {
1265                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1266                 eth_dev->data->representor_id = priv->representor_id;
1267         }
1268         eth_dev->data->dev_private = priv;
1269         priv->dev_data = eth_dev->data;
1270         eth_dev->data->mac_addrs = priv->mac;
1271         eth_dev->device = dpdk_dev;
1272         err = mlx5_uar_init_primary(eth_dev);
1273         if (err) {
1274                 err = rte_errno;
1275                 goto error;
1276         }
1277         /* Configure the first MAC address by default. */
1278         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1279                 DRV_LOG(ERR,
1280                         "port %u cannot get MAC address, is mlx5_en"
1281                         " loaded? (errno: %s)",
1282                         eth_dev->data->port_id, strerror(rte_errno));
1283                 err = ENODEV;
1284                 goto error;
1285         }
1286         DRV_LOG(INFO,
1287                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1288                 eth_dev->data->port_id,
1289                 mac.addr_bytes[0], mac.addr_bytes[1],
1290                 mac.addr_bytes[2], mac.addr_bytes[3],
1291                 mac.addr_bytes[4], mac.addr_bytes[5]);
1292 #ifndef NDEBUG
1293         {
1294                 char ifname[IF_NAMESIZE];
1295
1296                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1297                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1298                                 eth_dev->data->port_id, ifname);
1299                 else
1300                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1301                                 eth_dev->data->port_id);
1302         }
1303 #endif
1304         /* Get actual MTU if possible. */
1305         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1306         if (err) {
1307                 err = rte_errno;
1308                 goto error;
1309         }
1310         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1311                 priv->mtu);
1312         /* Initialize burst functions to prevent crashes before link-up. */
1313         eth_dev->rx_pkt_burst = removed_rx_burst;
1314         eth_dev->tx_pkt_burst = removed_tx_burst;
1315         eth_dev->dev_ops = &mlx5_dev_ops;
1316         /* Register MAC address. */
1317         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1318         if (config.vf && config.vf_nl_en)
1319                 mlx5_nl_mac_addr_sync(eth_dev);
1320         priv->tcf_context = mlx5_flow_tcf_context_create();
1321         if (!priv->tcf_context) {
1322                 err = -rte_errno;
1323                 DRV_LOG(WARNING,
1324                         "flow rules relying on switch offloads will not be"
1325                         " supported: cannot open libmnl socket: %s",
1326                         strerror(rte_errno));
1327         } else {
1328                 struct rte_flow_error error;
1329                 unsigned int ifindex = mlx5_ifindex(eth_dev);
1330
1331                 if (!ifindex) {
1332                         err = -rte_errno;
1333                         error.message =
1334                                 "cannot retrieve network interface index";
1335                 } else {
1336                         err = mlx5_flow_tcf_init(priv->tcf_context,
1337                                                  ifindex, &error);
1338                 }
1339                 if (err) {
1340                         DRV_LOG(WARNING,
1341                                 "flow rules relying on switch offloads will"
1342                                 " not be supported: %s: %s",
1343                                 error.message, strerror(rte_errno));
1344                         mlx5_flow_tcf_context_destroy(priv->tcf_context);
1345                         priv->tcf_context = NULL;
1346                 }
1347         }
1348         TAILQ_INIT(&priv->flows);
1349         TAILQ_INIT(&priv->ctrl_flows);
1350         /* Hint libmlx5 to use PMD allocator for data plane resources */
1351         struct mlx5dv_ctx_allocators alctr = {
1352                 .alloc = &mlx5_alloc_verbs_buf,
1353                 .free = &mlx5_free_verbs_buf,
1354                 .data = priv,
1355         };
1356         mlx5_glue->dv_set_context_attr(sh->ctx,
1357                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1358                                        (void *)((uintptr_t)&alctr));
1359         /* Bring Ethernet device up. */
1360         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1361                 eth_dev->data->port_id);
1362         mlx5_set_link_up(eth_dev);
1363         /*
1364          * Even though the interrupt handler is not installed yet,
1365          * interrupts will still trigger on the asyn_fd from
1366          * Verbs context returned by ibv_open_device().
1367          */
1368         mlx5_link_update(eth_dev, 0);
1369         /* Store device configuration on private structure. */
1370         priv->config = config;
1371         /* Supported Verbs flow priority number detection. */
1372         err = mlx5_flow_discover_priorities(eth_dev);
1373         if (err < 0) {
1374                 err = -err;
1375                 goto error;
1376         }
1377         priv->config.flow_prio = err;
1378         /*
1379          * Once the device is added to the list of memory event
1380          * callback, its global MR cache table cannot be expanded
1381          * on the fly because of deadlock. If it overflows, lookup
1382          * should be done by searching MR list linearly, which is slow.
1383          */
1384         err = mlx5_mr_btree_init(&priv->mr.cache,
1385                                  MLX5_MR_BTREE_CACHE_N * 2,
1386                                  eth_dev->device->numa_node);
1387         if (err) {
1388                 err = rte_errno;
1389                 goto error;
1390         }
1391         /* Add device to memory callback list. */
1392         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1393         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1394                          priv, mem_event_cb);
1395         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1396         return eth_dev;
1397 error:
1398         if (priv) {
1399                 if (priv->nl_socket_route >= 0)
1400                         close(priv->nl_socket_route);
1401                 if (priv->nl_socket_rdma >= 0)
1402                         close(priv->nl_socket_rdma);
1403                 if (priv->tcf_context)
1404                         mlx5_flow_tcf_context_destroy(priv->tcf_context);
1405                 if (own_domain_id)
1406                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1407                 rte_free(priv);
1408                 if (eth_dev != NULL)
1409                         eth_dev->data->dev_private = NULL;
1410         }
1411         if (eth_dev != NULL) {
1412                 /* mac_addrs must not be freed alone because part of dev_private */
1413                 eth_dev->data->mac_addrs = NULL;
1414                 rte_eth_dev_release_port(eth_dev);
1415         }
1416         if (sh)
1417                 mlx5_free_shared_ibctx(sh);
1418         assert(err > 0);
1419         rte_errno = err;
1420         return NULL;
1421 }
1422
1423 /**
1424  * Comparison callback to sort device data.
1425  *
1426  * This is meant to be used with qsort().
1427  *
1428  * @param a[in]
1429  *   Pointer to pointer to first data object.
1430  * @param b[in]
1431  *   Pointer to pointer to second data object.
1432  *
1433  * @return
1434  *   0 if both objects are equal, less than 0 if the first argument is less
1435  *   than the second, greater than 0 otherwise.
1436  */
1437 static int
1438 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1439 {
1440         const struct mlx5_switch_info *si_a =
1441                 &((const struct mlx5_dev_spawn_data *)a)->info;
1442         const struct mlx5_switch_info *si_b =
1443                 &((const struct mlx5_dev_spawn_data *)b)->info;
1444         int ret;
1445
1446         /* Master device first. */
1447         ret = si_b->master - si_a->master;
1448         if (ret)
1449                 return ret;
1450         /* Then representor devices. */
1451         ret = si_b->representor - si_a->representor;
1452         if (ret)
1453                 return ret;
1454         /* Unidentified devices come last in no specific order. */
1455         if (!si_a->representor)
1456                 return 0;
1457         /* Order representors by name. */
1458         return si_a->port_name - si_b->port_name;
1459 }
1460
1461 /**
1462  * DPDK callback to register a PCI device.
1463  *
1464  * This function spawns Ethernet devices out of a given PCI device.
1465  *
1466  * @param[in] pci_drv
1467  *   PCI driver structure (mlx5_driver).
1468  * @param[in] pci_dev
1469  *   PCI device information.
1470  *
1471  * @return
1472  *   0 on success, a negative errno value otherwise and rte_errno is set.
1473  */
1474 static int
1475 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1476                struct rte_pci_device *pci_dev)
1477 {
1478         struct ibv_device **ibv_list;
1479         /*
1480          * Number of found IB Devices matching with requested PCI BDF.
1481          * nd != 1 means there are multiple IB devices over the same
1482          * PCI device and we have representors and master.
1483          */
1484         unsigned int nd = 0;
1485         /*
1486          * Number of found IB device Ports. nd = 1 and np = 1..n means
1487          * we have the single multiport IB device, and there may be
1488          * representors attached to some of found ports.
1489          */
1490         unsigned int np = 0;
1491         /*
1492          * Number of DPDK ethernet devices to Spawn - either over
1493          * multiple IB devices or multiple ports of single IB device.
1494          * Actually this is the number of iterations to spawn.
1495          */
1496         unsigned int ns = 0;
1497         struct mlx5_dev_config dev_config;
1498         int ret;
1499
1500         assert(pci_drv == &mlx5_driver);
1501         errno = 0;
1502         ibv_list = mlx5_glue->get_device_list(&ret);
1503         if (!ibv_list) {
1504                 rte_errno = errno ? errno : ENOSYS;
1505                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1506                 return -rte_errno;
1507         }
1508         /*
1509          * First scan the list of all Infiniband devices to find
1510          * matching ones, gathering into the list.
1511          */
1512         struct ibv_device *ibv_match[ret + 1];
1513         int nl_route = -1;
1514         int nl_rdma = -1;
1515         unsigned int i;
1516
1517         while (ret-- > 0) {
1518                 struct rte_pci_addr pci_addr;
1519
1520                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1521                 if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1522                         continue;
1523                 if (pci_dev->addr.domain != pci_addr.domain ||
1524                     pci_dev->addr.bus != pci_addr.bus ||
1525                     pci_dev->addr.devid != pci_addr.devid ||
1526                     pci_dev->addr.function != pci_addr.function)
1527                         continue;
1528                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1529                         ibv_list[ret]->name);
1530                 ibv_match[nd++] = ibv_list[ret];
1531         }
1532         ibv_match[nd] = NULL;
1533         if (!nd) {
1534                 /* No device macthes, just complain and bail out. */
1535                 mlx5_glue->free_device_list(ibv_list);
1536                 DRV_LOG(WARNING,
1537                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1538                         " are kernel drivers loaded?",
1539                         pci_dev->addr.domain, pci_dev->addr.bus,
1540                         pci_dev->addr.devid, pci_dev->addr.function);
1541                 rte_errno = ENOENT;
1542                 ret = -rte_errno;
1543                 return ret;
1544         }
1545         nl_route = mlx5_nl_init(NETLINK_ROUTE);
1546         nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1547         if (nd == 1) {
1548                 /*
1549                  * Found single matching device may have multiple ports.
1550                  * Each port may be representor, we have to check the port
1551                  * number and check the representors existence.
1552                  */
1553                 if (nl_rdma >= 0)
1554                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1555                 if (!np)
1556                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1557                                          " ports number", ibv_match[0]->name);
1558         }
1559         /*
1560          * Now we can determine the maximal
1561          * amount of devices to be spawned.
1562          */
1563         struct mlx5_dev_spawn_data list[np ? np : nd];
1564
1565         if (np > 1) {
1566                 /*
1567                  * Signle IB device with multiple ports found,
1568                  * it may be E-Switch master device and representors.
1569                  * We have to perform identification trough the ports.
1570                  */
1571                 assert(nl_rdma >= 0);
1572                 assert(ns == 0);
1573                 assert(nd == 1);
1574                 for (i = 1; i <= np; ++i) {
1575                         list[ns].max_port = np;
1576                         list[ns].ibv_port = i;
1577                         list[ns].ibv_dev = ibv_match[0];
1578                         list[ns].eth_dev = NULL;
1579                         list[ns].ifindex = mlx5_nl_ifindex
1580                                         (nl_rdma, list[ns].ibv_dev->name, i);
1581                         if (!list[ns].ifindex) {
1582                                 /*
1583                                  * No network interface index found for the
1584                                  * specified port, it means there is no
1585                                  * representor on this port. It's OK,
1586                                  * there can be disabled ports, for example
1587                                  * if sriov_numvfs < sriov_totalvfs.
1588                                  */
1589                                 continue;
1590                         }
1591                         ret = -1;
1592                         if (nl_route >= 0)
1593                                 ret = mlx5_nl_switch_info
1594                                                (nl_route,
1595                                                 list[ns].ifindex,
1596                                                 &list[ns].info);
1597                         if (ret || (!list[ns].info.representor &&
1598                                     !list[ns].info.master)) {
1599                                 /*
1600                                  * We failed to recognize representors with
1601                                  * Netlink, let's try to perform the task
1602                                  * with sysfs.
1603                                  */
1604                                 ret =  mlx5_sysfs_switch_info
1605                                                 (list[ns].ifindex,
1606                                                  &list[ns].info);
1607                         }
1608                         if (!ret && (list[ns].info.representor ^
1609                                      list[ns].info.master))
1610                                 ns++;
1611                 }
1612                 if (!ns) {
1613                         DRV_LOG(ERR,
1614                                 "unable to recognize master/representors"
1615                                 " on the IB device with multiple ports");
1616                         rte_errno = ENOENT;
1617                         ret = -rte_errno;
1618                         goto exit;
1619                 }
1620         } else {
1621                 /*
1622                  * The existence of several matching entries (nd > 1) means
1623                  * port representors have been instantiated. No existing Verbs
1624                  * call nor sysfs entries can tell them apart, this can only
1625                  * be done through Netlink calls assuming kernel drivers are
1626                  * recent enough to support them.
1627                  *
1628                  * In the event of identification failure through Netlink,
1629                  * try again through sysfs, then:
1630                  *
1631                  * 1. A single IB device matches (nd == 1) with single
1632                  *    port (np=0/1) and is not a representor, assume
1633                  *    no switch support.
1634                  *
1635                  * 2. Otherwise no safe assumptions can be made;
1636                  *    complain louder and bail out.
1637                  */
1638                 np = 1;
1639                 for (i = 0; i != nd; ++i) {
1640                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1641                         list[ns].max_port = 1;
1642                         list[ns].ibv_port = 1;
1643                         list[ns].ibv_dev = ibv_match[i];
1644                         list[ns].eth_dev = NULL;
1645                         list[ns].ifindex = 0;
1646                         if (nl_rdma >= 0)
1647                                 list[ns].ifindex = mlx5_nl_ifindex
1648                                         (nl_rdma, list[ns].ibv_dev->name, 1);
1649                         if (!list[ns].ifindex) {
1650                                 /*
1651                                  * No network interface index found for the
1652                                  * specified device, it means there it is not
1653                                  * a representor/master.
1654                                  */
1655                                 continue;
1656                         }
1657                         ret = -1;
1658                         if (nl_route >= 0)
1659                                 ret = mlx5_nl_switch_info
1660                                                (nl_route,
1661                                                 list[ns].ifindex,
1662                                                 &list[ns].info);
1663                         if (ret || (!list[ns].info.representor &&
1664                                     !list[ns].info.master)) {
1665                                 /*
1666                                  * We failed to recognize representors with
1667                                  * Netlink, let's try to perform the task
1668                                  * with sysfs.
1669                                  */
1670                                 ret =  mlx5_sysfs_switch_info
1671                                                 (list[ns].ifindex,
1672                                                  &list[ns].info);
1673                         }
1674                         if (!ret && (list[ns].info.representor ^
1675                                      list[ns].info.master)) {
1676                                 ns++;
1677                         } else if ((nd == 1) &&
1678                                    !list[ns].info.representor &&
1679                                    !list[ns].info.master) {
1680                                 /*
1681                                  * Single IB device with
1682                                  * one physical port and
1683                                  * attached network device.
1684                                  * May be SRIOV is not enabled
1685                                  * or there is no representors.
1686                                  */
1687                                 DRV_LOG(INFO, "no E-Switch support detected");
1688                                 ns++;
1689                                 break;
1690                         }
1691                 }
1692                 if (!ns) {
1693                         DRV_LOG(ERR,
1694                                 "unable to recognize master/representors"
1695                                 " on the multiple IB devices");
1696                         rte_errno = ENOENT;
1697                         ret = -rte_errno;
1698                         goto exit;
1699                 }
1700         }
1701         assert(ns);
1702         /*
1703          * Sort list to probe devices in natural order for users convenience
1704          * (i.e. master first, then representors from lowest to highest ID).
1705          */
1706         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1707         /* Default configuration. */
1708         dev_config = (struct mlx5_dev_config){
1709                 .hw_padding = 0,
1710                 .mps = MLX5_ARG_UNSET,
1711                 .tx_vec_en = 1,
1712                 .rx_vec_en = 1,
1713                 .txq_inline = MLX5_ARG_UNSET,
1714                 .txqs_inline = MLX5_ARG_UNSET,
1715                 .txqs_vec = MLX5_ARG_UNSET,
1716                 .inline_max_packet_sz = MLX5_ARG_UNSET,
1717                 .vf_nl_en = 1,
1718                 .mprq = {
1719                         .enabled = 0, /* Disabled by default. */
1720                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
1721                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1722                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1723                 },
1724         };
1725         /* Device specific configuration. */
1726         switch (pci_dev->id.device_id) {
1727         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BF:
1728                 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS_BLUEFIELD;
1729                 break;
1730         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1731         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1732         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1733         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1734                 dev_config.vf = 1;
1735                 break;
1736         default:
1737                 break;
1738         }
1739         /* Set architecture-dependent default value if unset. */
1740         if (dev_config.txqs_vec == MLX5_ARG_UNSET)
1741                 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS;
1742         for (i = 0; i != ns; ++i) {
1743                 uint32_t restore;
1744
1745                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1746                                                  &list[i],
1747                                                  dev_config);
1748                 if (!list[i].eth_dev) {
1749                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1750                                 break;
1751                         /* Device is disabled or already spawned. Ignore it. */
1752                         continue;
1753                 }
1754                 restore = list[i].eth_dev->data->dev_flags;
1755                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1756                 /* Restore non-PCI flags cleared by the above call. */
1757                 list[i].eth_dev->data->dev_flags |= restore;
1758                 rte_eth_dev_probing_finish(list[i].eth_dev);
1759         }
1760         if (i != ns) {
1761                 DRV_LOG(ERR,
1762                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1763                         " encountering an error: %s",
1764                         pci_dev->addr.domain, pci_dev->addr.bus,
1765                         pci_dev->addr.devid, pci_dev->addr.function,
1766                         strerror(rte_errno));
1767                 ret = -rte_errno;
1768                 /* Roll back. */
1769                 while (i--) {
1770                         if (!list[i].eth_dev)
1771                                 continue;
1772                         mlx5_dev_close(list[i].eth_dev);
1773                         /* mac_addrs must not be freed because in dev_private */
1774                         list[i].eth_dev->data->mac_addrs = NULL;
1775                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1776                 }
1777                 /* Restore original error. */
1778                 rte_errno = -ret;
1779         } else {
1780                 ret = 0;
1781         }
1782 exit:
1783         /*
1784          * Do the routine cleanup:
1785          * - close opened Netlink sockets
1786          * - free the Infiniband device list
1787          */
1788         if (nl_rdma >= 0)
1789                 close(nl_rdma);
1790         if (nl_route >= 0)
1791                 close(nl_route);
1792         assert(ibv_list);
1793         mlx5_glue->free_device_list(ibv_list);
1794         return ret;
1795 }
1796
1797 /**
1798  * DPDK callback to remove a PCI device.
1799  *
1800  * This function removes all Ethernet devices belong to a given PCI device.
1801  *
1802  * @param[in] pci_dev
1803  *   Pointer to the PCI device.
1804  *
1805  * @return
1806  *   0 on success, the function cannot fail.
1807  */
1808 static int
1809 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1810 {
1811         uint16_t port_id;
1812         struct rte_eth_dev *port;
1813
1814         for (port_id = 0; port_id < RTE_MAX_ETHPORTS; port_id++) {
1815                 port = &rte_eth_devices[port_id];
1816                 if (port->state != RTE_ETH_DEV_UNUSED &&
1817                                 port->device == &pci_dev->device)
1818                         rte_eth_dev_close(port_id);
1819         }
1820         return 0;
1821 }
1822
1823 static const struct rte_pci_id mlx5_pci_id_map[] = {
1824         {
1825                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1826                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1827         },
1828         {
1829                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1830                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1831         },
1832         {
1833                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1834                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1835         },
1836         {
1837                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1838                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1839         },
1840         {
1841                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1842                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1843         },
1844         {
1845                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1846                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1847         },
1848         {
1849                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1850                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1851         },
1852         {
1853                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1854                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1855         },
1856         {
1857                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1858                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
1859         },
1860         {
1861                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1862                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
1863         },
1864         {
1865                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1866                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
1867         },
1868         {
1869                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1870                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
1871         },
1872         {
1873                 .vendor_id = 0
1874         }
1875 };
1876
1877 static struct rte_pci_driver mlx5_driver = {
1878         .driver = {
1879                 .name = MLX5_DRIVER_NAME
1880         },
1881         .id_table = mlx5_pci_id_map,
1882         .probe = mlx5_pci_probe,
1883         .remove = mlx5_pci_remove,
1884         .dma_map = mlx5_dma_map,
1885         .dma_unmap = mlx5_dma_unmap,
1886         .drv_flags = (RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
1887                       RTE_PCI_DRV_PROBE_AGAIN),
1888 };
1889
1890 #ifdef RTE_IBVERBS_LINK_DLOPEN
1891
1892 /**
1893  * Suffix RTE_EAL_PMD_PATH with "-glue".
1894  *
1895  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1896  * suffixing its last component.
1897  *
1898  * @param buf[out]
1899  *   Output buffer, should be large enough otherwise NULL is returned.
1900  * @param size
1901  *   Size of @p out.
1902  *
1903  * @return
1904  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1905  */
1906 static char *
1907 mlx5_glue_path(char *buf, size_t size)
1908 {
1909         static const char *const bad[] = { "/", ".", "..", NULL };
1910         const char *path = RTE_EAL_PMD_PATH;
1911         size_t len = strlen(path);
1912         size_t off;
1913         int i;
1914
1915         while (len && path[len - 1] == '/')
1916                 --len;
1917         for (off = len; off && path[off - 1] != '/'; --off)
1918                 ;
1919         for (i = 0; bad[i]; ++i)
1920                 if (!strncmp(path + off, bad[i], (int)(len - off)))
1921                         goto error;
1922         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1923         if (i == -1 || (size_t)i >= size)
1924                 goto error;
1925         return buf;
1926 error:
1927         DRV_LOG(ERR,
1928                 "unable to append \"-glue\" to last component of"
1929                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1930                 " please re-configure DPDK");
1931         return NULL;
1932 }
1933
1934 /**
1935  * Initialization routine for run-time dependency on rdma-core.
1936  */
1937 static int
1938 mlx5_glue_init(void)
1939 {
1940         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1941         const char *path[] = {
1942                 /*
1943                  * A basic security check is necessary before trusting
1944                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1945                  */
1946                 (geteuid() == getuid() && getegid() == getgid() ?
1947                  getenv("MLX5_GLUE_PATH") : NULL),
1948                 /*
1949                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1950                  * variant, otherwise let dlopen() look up libraries on its
1951                  * own.
1952                  */
1953                 (*RTE_EAL_PMD_PATH ?
1954                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1955         };
1956         unsigned int i = 0;
1957         void *handle = NULL;
1958         void **sym;
1959         const char *dlmsg;
1960
1961         while (!handle && i != RTE_DIM(path)) {
1962                 const char *end;
1963                 size_t len;
1964                 int ret;
1965
1966                 if (!path[i]) {
1967                         ++i;
1968                         continue;
1969                 }
1970                 end = strpbrk(path[i], ":;");
1971                 if (!end)
1972                         end = path[i] + strlen(path[i]);
1973                 len = end - path[i];
1974                 ret = 0;
1975                 do {
1976                         char name[ret + 1];
1977
1978                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1979                                        (int)len, path[i],
1980                                        (!len || *(end - 1) == '/') ? "" : "/");
1981                         if (ret == -1)
1982                                 break;
1983                         if (sizeof(name) != (size_t)ret + 1)
1984                                 continue;
1985                         DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1986                                 name);
1987                         handle = dlopen(name, RTLD_LAZY);
1988                         break;
1989                 } while (1);
1990                 path[i] = end + 1;
1991                 if (!*end)
1992                         ++i;
1993         }
1994         if (!handle) {
1995                 rte_errno = EINVAL;
1996                 dlmsg = dlerror();
1997                 if (dlmsg)
1998                         DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1999                 goto glue_error;
2000         }
2001         sym = dlsym(handle, "mlx5_glue");
2002         if (!sym || !*sym) {
2003                 rte_errno = EINVAL;
2004                 dlmsg = dlerror();
2005                 if (dlmsg)
2006                         DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
2007                 goto glue_error;
2008         }
2009         mlx5_glue = *sym;
2010         return 0;
2011 glue_error:
2012         if (handle)
2013                 dlclose(handle);
2014         DRV_LOG(WARNING,
2015                 "cannot initialize PMD due to missing run-time dependency on"
2016                 " rdma-core libraries (libibverbs, libmlx5)");
2017         return -rte_errno;
2018 }
2019
2020 #endif
2021
2022 /**
2023  * Driver initialization routine.
2024  */
2025 RTE_INIT(rte_mlx5_pmd_init)
2026 {
2027         /* Initialize driver log type. */
2028         mlx5_logtype = rte_log_register("pmd.net.mlx5");
2029         if (mlx5_logtype >= 0)
2030                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
2031
2032         /* Build the static tables for Verbs conversion. */
2033         mlx5_set_ptype_table();
2034         mlx5_set_cksum_table();
2035         mlx5_set_swp_types_table();
2036         /*
2037          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
2038          * huge pages. Calling ibv_fork_init() during init allows
2039          * applications to use fork() safely for purposes other than
2040          * using this PMD, which is not supported in forked processes.
2041          */
2042         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
2043         /* Match the size of Rx completion entry to the size of a cacheline. */
2044         if (RTE_CACHE_LINE_SIZE == 128)
2045                 setenv("MLX5_CQE_SIZE", "128", 0);
2046         /*
2047          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
2048          * cleanup all the Verbs resources even when the device was removed.
2049          */
2050         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
2051 #ifdef RTE_IBVERBS_LINK_DLOPEN
2052         if (mlx5_glue_init())
2053                 return;
2054         assert(mlx5_glue);
2055 #endif
2056 #ifndef NDEBUG
2057         /* Glue structure must not contain any NULL pointers. */
2058         {
2059                 unsigned int i;
2060
2061                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
2062                         assert(((const void *const *)mlx5_glue)[i]);
2063         }
2064 #endif
2065         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
2066                 DRV_LOG(ERR,
2067                         "rdma-core glue \"%s\" mismatch: \"%s\" is required",
2068                         mlx5_glue->version, MLX5_GLUE_VERSION);
2069                 return;
2070         }
2071         mlx5_glue->fork_init();
2072         rte_pci_register(&mlx5_driver);
2073 }
2074
2075 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2076 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2077 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");