net/mlx5: preserve promiscuous flag for flow isolation mode
[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/netlink.h>
17 #include <linux/rtnetlink.h>
18
19 /* Verbs header. */
20 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
21 #ifdef PEDANTIC
22 #pragma GCC diagnostic ignored "-Wpedantic"
23 #endif
24 #include <infiniband/verbs.h>
25 #ifdef PEDANTIC
26 #pragma GCC diagnostic error "-Wpedantic"
27 #endif
28
29 #include <rte_malloc.h>
30 #include <rte_ethdev_driver.h>
31 #include <rte_ethdev_pci.h>
32 #include <rte_pci.h>
33 #include <rte_bus_pci.h>
34 #include <rte_common.h>
35 #include <rte_config.h>
36 #include <rte_eal_memconfig.h>
37 #include <rte_kvargs.h>
38 #include <rte_rwlock.h>
39 #include <rte_spinlock.h>
40 #include <rte_string_fns.h>
41
42 #include "mlx5.h"
43 #include "mlx5_utils.h"
44 #include "mlx5_rxtx.h"
45 #include "mlx5_autoconf.h"
46 #include "mlx5_defs.h"
47 #include "mlx5_glue.h"
48 #include "mlx5_mr.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 Multi-Packet Rx queue. */
54 #define MLX5_RX_MPRQ_EN "mprq_en"
55
56 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
57 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
58
59 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
60 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
61
62 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
63 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
64
65 /* Device parameter to configure inline send. */
66 #define MLX5_TXQ_INLINE "txq_inline"
67
68 /*
69  * Device parameter to configure the number of TX queues threshold for
70  * enabling inline send.
71  */
72 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
73
74 /* Device parameter to enable multi-packet send WQEs. */
75 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
76
77 /* Device parameter to include 2 dsegs in the title WQEBB. */
78 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
79
80 /* Device parameter to limit the size of inlining packet. */
81 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
82
83 /* Device parameter to enable hardware Tx vector. */
84 #define MLX5_TX_VEC_EN "tx_vec_en"
85
86 /* Device parameter to enable hardware Rx vector. */
87 #define MLX5_RX_VEC_EN "rx_vec_en"
88
89 /* Allow L3 VXLAN flow creation. */
90 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
91
92 /* Activate Netlink support in VF mode. */
93 #define MLX5_VF_NL_EN "vf_nl_en"
94
95 /* Select port representors to instantiate. */
96 #define MLX5_REPRESENTOR "representor"
97
98 #ifndef HAVE_IBV_MLX5_MOD_MPW
99 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
100 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
101 #endif
102
103 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
104 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
105 #endif
106
107 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
108
109 /* Shared memory between primary and secondary processes. */
110 struct mlx5_shared_data *mlx5_shared_data;
111
112 /* Spinlock for mlx5_shared_data allocation. */
113 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
114
115 /** Driver-specific log messages type. */
116 int mlx5_logtype;
117
118 /**
119  * Prepare shared data between primary and secondary process.
120  */
121 static void
122 mlx5_prepare_shared_data(void)
123 {
124         const struct rte_memzone *mz;
125
126         rte_spinlock_lock(&mlx5_shared_data_lock);
127         if (mlx5_shared_data == NULL) {
128                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
129                         /* Allocate shared memory. */
130                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
131                                                  sizeof(*mlx5_shared_data),
132                                                  SOCKET_ID_ANY, 0);
133                 } else {
134                         /* Lookup allocated shared memory. */
135                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
136                 }
137                 if (mz == NULL)
138                         rte_panic("Cannot allocate mlx5 shared data\n");
139                 mlx5_shared_data = mz->addr;
140                 /* Initialize shared data. */
141                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
142                         LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
143                         rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
144                 }
145                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
146                                                 mlx5_mr_mem_event_cb, NULL);
147         }
148         rte_spinlock_unlock(&mlx5_shared_data_lock);
149 }
150
151 /**
152  * Retrieve integer value from environment variable.
153  *
154  * @param[in] name
155  *   Environment variable name.
156  *
157  * @return
158  *   Integer value, 0 if the variable is not set.
159  */
160 int
161 mlx5_getenv_int(const char *name)
162 {
163         const char *val = getenv(name);
164
165         if (val == NULL)
166                 return 0;
167         return atoi(val);
168 }
169
170 /**
171  * Verbs callback to allocate a memory. This function should allocate the space
172  * according to the size provided residing inside a huge page.
173  * Please note that all allocation must respect the alignment from libmlx5
174  * (i.e. currently sysconf(_SC_PAGESIZE)).
175  *
176  * @param[in] size
177  *   The size in bytes of the memory to allocate.
178  * @param[in] data
179  *   A pointer to the callback data.
180  *
181  * @return
182  *   Allocated buffer, NULL otherwise and rte_errno is set.
183  */
184 static void *
185 mlx5_alloc_verbs_buf(size_t size, void *data)
186 {
187         struct priv *priv = data;
188         void *ret;
189         size_t alignment = sysconf(_SC_PAGESIZE);
190         unsigned int socket = SOCKET_ID_ANY;
191
192         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
193                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
194
195                 socket = ctrl->socket;
196         } else if (priv->verbs_alloc_ctx.type ==
197                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
198                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
199
200                 socket = ctrl->socket;
201         }
202         assert(data != NULL);
203         ret = rte_malloc_socket(__func__, size, alignment, socket);
204         if (!ret && size)
205                 rte_errno = ENOMEM;
206         return ret;
207 }
208
209 /**
210  * Verbs callback to free a memory.
211  *
212  * @param[in] ptr
213  *   A pointer to the memory to free.
214  * @param[in] data
215  *   A pointer to the callback data.
216  */
217 static void
218 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
219 {
220         assert(data != NULL);
221         rte_free(ptr);
222 }
223
224 /**
225  * DPDK callback to close the device.
226  *
227  * Destroy all queues and objects, free memory.
228  *
229  * @param dev
230  *   Pointer to Ethernet device structure.
231  */
232 static void
233 mlx5_dev_close(struct rte_eth_dev *dev)
234 {
235         struct priv *priv = dev->data->dev_private;
236         unsigned int i;
237         int ret;
238
239         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
240                 dev->data->port_id,
241                 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
242         /* In case mlx5_dev_stop() has not been called. */
243         mlx5_dev_interrupt_handler_uninstall(dev);
244         mlx5_traffic_disable(dev);
245         mlx5_flow_flush(dev, NULL);
246         /* Prevent crashes when queues are still in use. */
247         dev->rx_pkt_burst = removed_rx_burst;
248         dev->tx_pkt_burst = removed_tx_burst;
249         if (priv->rxqs != NULL) {
250                 /* XXX race condition if mlx5_rx_burst() is still running. */
251                 usleep(1000);
252                 for (i = 0; (i != priv->rxqs_n); ++i)
253                         mlx5_rxq_release(dev, i);
254                 priv->rxqs_n = 0;
255                 priv->rxqs = NULL;
256         }
257         if (priv->txqs != NULL) {
258                 /* XXX race condition if mlx5_tx_burst() is still running. */
259                 usleep(1000);
260                 for (i = 0; (i != priv->txqs_n); ++i)
261                         mlx5_txq_release(dev, i);
262                 priv->txqs_n = 0;
263                 priv->txqs = NULL;
264         }
265         mlx5_mprq_free_mp(dev);
266         mlx5_mr_release(dev);
267         if (priv->pd != NULL) {
268                 assert(priv->ctx != NULL);
269                 claim_zero(mlx5_glue->dealloc_pd(priv->pd));
270                 claim_zero(mlx5_glue->close_device(priv->ctx));
271         } else
272                 assert(priv->ctx == NULL);
273         if (priv->rss_conf.rss_key != NULL)
274                 rte_free(priv->rss_conf.rss_key);
275         if (priv->reta_idx != NULL)
276                 rte_free(priv->reta_idx);
277         if (priv->primary_socket)
278                 mlx5_socket_uninit(dev);
279         if (priv->config.vf)
280                 mlx5_nl_mac_addr_flush(dev);
281         if (priv->nl_socket_route >= 0)
282                 close(priv->nl_socket_route);
283         if (priv->nl_socket_rdma >= 0)
284                 close(priv->nl_socket_rdma);
285         if (priv->mnl_socket)
286                 mlx5_nl_flow_socket_destroy(priv->mnl_socket);
287         ret = mlx5_hrxq_ibv_verify(dev);
288         if (ret)
289                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
290                         dev->data->port_id);
291         ret = mlx5_ind_table_ibv_verify(dev);
292         if (ret)
293                 DRV_LOG(WARNING, "port %u some indirection table still remain",
294                         dev->data->port_id);
295         ret = mlx5_rxq_ibv_verify(dev);
296         if (ret)
297                 DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
298                         dev->data->port_id);
299         ret = mlx5_rxq_verify(dev);
300         if (ret)
301                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
302                         dev->data->port_id);
303         ret = mlx5_txq_ibv_verify(dev);
304         if (ret)
305                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
306                         dev->data->port_id);
307         ret = mlx5_txq_verify(dev);
308         if (ret)
309                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
310                         dev->data->port_id);
311         ret = mlx5_flow_verify(dev);
312         if (ret)
313                 DRV_LOG(WARNING, "port %u some flows still remain",
314                         dev->data->port_id);
315         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
316                 unsigned int c = 0;
317                 unsigned int i = mlx5_dev_to_port_id(dev->device, NULL, 0);
318                 uint16_t port_id[i];
319
320                 i = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, i), i);
321                 while (i--) {
322                         struct priv *opriv =
323                                 rte_eth_devices[port_id[i]].data->dev_private;
324
325                         if (!opriv ||
326                             opriv->domain_id != priv->domain_id ||
327                             &rte_eth_devices[port_id[i]] == dev)
328                                 continue;
329                         ++c;
330                 }
331                 if (!c)
332                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
333         }
334         memset(priv, 0, sizeof(*priv));
335         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
336 }
337
338 const struct eth_dev_ops mlx5_dev_ops = {
339         .dev_configure = mlx5_dev_configure,
340         .dev_start = mlx5_dev_start,
341         .dev_stop = mlx5_dev_stop,
342         .dev_set_link_down = mlx5_set_link_down,
343         .dev_set_link_up = mlx5_set_link_up,
344         .dev_close = mlx5_dev_close,
345         .promiscuous_enable = mlx5_promiscuous_enable,
346         .promiscuous_disable = mlx5_promiscuous_disable,
347         .allmulticast_enable = mlx5_allmulticast_enable,
348         .allmulticast_disable = mlx5_allmulticast_disable,
349         .link_update = mlx5_link_update,
350         .stats_get = mlx5_stats_get,
351         .stats_reset = mlx5_stats_reset,
352         .xstats_get = mlx5_xstats_get,
353         .xstats_reset = mlx5_xstats_reset,
354         .xstats_get_names = mlx5_xstats_get_names,
355         .dev_infos_get = mlx5_dev_infos_get,
356         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
357         .vlan_filter_set = mlx5_vlan_filter_set,
358         .rx_queue_setup = mlx5_rx_queue_setup,
359         .tx_queue_setup = mlx5_tx_queue_setup,
360         .rx_queue_release = mlx5_rx_queue_release,
361         .tx_queue_release = mlx5_tx_queue_release,
362         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
363         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
364         .mac_addr_remove = mlx5_mac_addr_remove,
365         .mac_addr_add = mlx5_mac_addr_add,
366         .mac_addr_set = mlx5_mac_addr_set,
367         .set_mc_addr_list = mlx5_set_mc_addr_list,
368         .mtu_set = mlx5_dev_set_mtu,
369         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
370         .vlan_offload_set = mlx5_vlan_offload_set,
371         .reta_update = mlx5_dev_rss_reta_update,
372         .reta_query = mlx5_dev_rss_reta_query,
373         .rss_hash_update = mlx5_rss_hash_update,
374         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
375         .filter_ctrl = mlx5_dev_filter_ctrl,
376         .rx_descriptor_status = mlx5_rx_descriptor_status,
377         .tx_descriptor_status = mlx5_tx_descriptor_status,
378         .rx_queue_intr_enable = mlx5_rx_intr_enable,
379         .rx_queue_intr_disable = mlx5_rx_intr_disable,
380         .is_removed = mlx5_is_removed,
381 };
382
383 static const struct eth_dev_ops mlx5_dev_sec_ops = {
384         .stats_get = mlx5_stats_get,
385         .stats_reset = mlx5_stats_reset,
386         .xstats_get = mlx5_xstats_get,
387         .xstats_reset = mlx5_xstats_reset,
388         .xstats_get_names = mlx5_xstats_get_names,
389         .dev_infos_get = mlx5_dev_infos_get,
390         .rx_descriptor_status = mlx5_rx_descriptor_status,
391         .tx_descriptor_status = mlx5_tx_descriptor_status,
392 };
393
394 /* Available operators in flow isolated mode. */
395 const struct eth_dev_ops mlx5_dev_ops_isolate = {
396         .dev_configure = mlx5_dev_configure,
397         .dev_start = mlx5_dev_start,
398         .dev_stop = mlx5_dev_stop,
399         .dev_set_link_down = mlx5_set_link_down,
400         .dev_set_link_up = mlx5_set_link_up,
401         .dev_close = mlx5_dev_close,
402         .promiscuous_enable = mlx5_promiscuous_enable,
403         .promiscuous_disable = mlx5_promiscuous_disable,
404         .link_update = mlx5_link_update,
405         .stats_get = mlx5_stats_get,
406         .stats_reset = mlx5_stats_reset,
407         .xstats_get = mlx5_xstats_get,
408         .xstats_reset = mlx5_xstats_reset,
409         .xstats_get_names = mlx5_xstats_get_names,
410         .dev_infos_get = mlx5_dev_infos_get,
411         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
412         .vlan_filter_set = mlx5_vlan_filter_set,
413         .rx_queue_setup = mlx5_rx_queue_setup,
414         .tx_queue_setup = mlx5_tx_queue_setup,
415         .rx_queue_release = mlx5_rx_queue_release,
416         .tx_queue_release = mlx5_tx_queue_release,
417         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
418         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
419         .mac_addr_remove = mlx5_mac_addr_remove,
420         .mac_addr_add = mlx5_mac_addr_add,
421         .mac_addr_set = mlx5_mac_addr_set,
422         .set_mc_addr_list = mlx5_set_mc_addr_list,
423         .mtu_set = mlx5_dev_set_mtu,
424         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
425         .vlan_offload_set = mlx5_vlan_offload_set,
426         .filter_ctrl = mlx5_dev_filter_ctrl,
427         .rx_descriptor_status = mlx5_rx_descriptor_status,
428         .tx_descriptor_status = mlx5_tx_descriptor_status,
429         .rx_queue_intr_enable = mlx5_rx_intr_enable,
430         .rx_queue_intr_disable = mlx5_rx_intr_disable,
431         .is_removed = mlx5_is_removed,
432 };
433
434 /**
435  * Verify and store value for device argument.
436  *
437  * @param[in] key
438  *   Key argument to verify.
439  * @param[in] val
440  *   Value associated with key.
441  * @param opaque
442  *   User data.
443  *
444  * @return
445  *   0 on success, a negative errno value otherwise and rte_errno is set.
446  */
447 static int
448 mlx5_args_check(const char *key, const char *val, void *opaque)
449 {
450         struct mlx5_dev_config *config = opaque;
451         unsigned long tmp;
452
453         /* No-op, port representors are processed in mlx5_dev_spawn(). */
454         if (!strcmp(MLX5_REPRESENTOR, key))
455                 return 0;
456         errno = 0;
457         tmp = strtoul(val, NULL, 0);
458         if (errno) {
459                 rte_errno = errno;
460                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
461                 return -rte_errno;
462         }
463         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
464                 config->cqe_comp = !!tmp;
465         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
466                 config->mprq.enabled = !!tmp;
467         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
468                 config->mprq.stride_num_n = tmp;
469         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
470                 config->mprq.max_memcpy_len = tmp;
471         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
472                 config->mprq.min_rxqs_num = tmp;
473         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
474                 config->txq_inline = tmp;
475         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
476                 config->txqs_inline = tmp;
477         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
478                 config->mps = !!tmp ? config->mps : 0;
479         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
480                 config->mpw_hdr_dseg = !!tmp;
481         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
482                 config->inline_max_packet_sz = tmp;
483         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
484                 config->tx_vec_en = !!tmp;
485         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
486                 config->rx_vec_en = !!tmp;
487         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
488                 config->l3_vxlan_en = !!tmp;
489         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
490                 config->vf_nl_en = !!tmp;
491         } else {
492                 DRV_LOG(WARNING, "%s: unknown parameter", key);
493                 rte_errno = EINVAL;
494                 return -rte_errno;
495         }
496         return 0;
497 }
498
499 /**
500  * Parse device parameters.
501  *
502  * @param config
503  *   Pointer to device configuration structure.
504  * @param devargs
505  *   Device arguments structure.
506  *
507  * @return
508  *   0 on success, a negative errno value otherwise and rte_errno is set.
509  */
510 static int
511 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
512 {
513         const char **params = (const char *[]){
514                 MLX5_RXQ_CQE_COMP_EN,
515                 MLX5_RX_MPRQ_EN,
516                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
517                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
518                 MLX5_RXQS_MIN_MPRQ,
519                 MLX5_TXQ_INLINE,
520                 MLX5_TXQS_MIN_INLINE,
521                 MLX5_TXQ_MPW_EN,
522                 MLX5_TXQ_MPW_HDR_DSEG_EN,
523                 MLX5_TXQ_MAX_INLINE_LEN,
524                 MLX5_TX_VEC_EN,
525                 MLX5_RX_VEC_EN,
526                 MLX5_L3_VXLAN_EN,
527                 MLX5_VF_NL_EN,
528                 MLX5_REPRESENTOR,
529                 NULL,
530         };
531         struct rte_kvargs *kvlist;
532         int ret = 0;
533         int i;
534
535         if (devargs == NULL)
536                 return 0;
537         /* Following UGLY cast is done to pass checkpatch. */
538         kvlist = rte_kvargs_parse(devargs->args, params);
539         if (kvlist == NULL)
540                 return 0;
541         /* Process parameters. */
542         for (i = 0; (params[i] != NULL); ++i) {
543                 if (rte_kvargs_count(kvlist, params[i])) {
544                         ret = rte_kvargs_process(kvlist, params[i],
545                                                  mlx5_args_check, config);
546                         if (ret) {
547                                 rte_errno = EINVAL;
548                                 rte_kvargs_free(kvlist);
549                                 return -rte_errno;
550                         }
551                 }
552         }
553         rte_kvargs_free(kvlist);
554         return 0;
555 }
556
557 static struct rte_pci_driver mlx5_driver;
558
559 /*
560  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
561  * local resource used by both primary and secondary to avoid duplicate
562  * reservation.
563  * The space has to be available on both primary and secondary process,
564  * TXQ UAR maps to this area using fixed mmap w/o double check.
565  */
566 static void *uar_base;
567
568 static int
569 find_lower_va_bound(const struct rte_memseg_list *msl __rte_unused,
570                 const struct rte_memseg *ms, void *arg)
571 {
572         void **addr = arg;
573
574         if (*addr == NULL)
575                 *addr = ms->addr;
576         else
577                 *addr = RTE_MIN(*addr, ms->addr);
578
579         return 0;
580 }
581
582 /**
583  * Reserve UAR address space for primary process.
584  *
585  * @param[in] dev
586  *   Pointer to Ethernet device.
587  *
588  * @return
589  *   0 on success, a negative errno value otherwise and rte_errno is set.
590  */
591 static int
592 mlx5_uar_init_primary(struct rte_eth_dev *dev)
593 {
594         struct priv *priv = dev->data->dev_private;
595         void *addr = (void *)0;
596
597         if (uar_base) { /* UAR address space mapped. */
598                 priv->uar_base = uar_base;
599                 return 0;
600         }
601         /* find out lower bound of hugepage segments */
602         rte_memseg_walk(find_lower_va_bound, &addr);
603
604         /* keep distance to hugepages to minimize potential conflicts. */
605         addr = RTE_PTR_SUB(addr, (uintptr_t)(MLX5_UAR_OFFSET + MLX5_UAR_SIZE));
606         /* anonymous mmap, no real memory consumption. */
607         addr = mmap(addr, MLX5_UAR_SIZE,
608                     PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
609         if (addr == MAP_FAILED) {
610                 DRV_LOG(ERR,
611                         "port %u failed to reserve UAR address space, please"
612                         " adjust MLX5_UAR_SIZE or try --base-virtaddr",
613                         dev->data->port_id);
614                 rte_errno = ENOMEM;
615                 return -rte_errno;
616         }
617         /* Accept either same addr or a new addr returned from mmap if target
618          * range occupied.
619          */
620         DRV_LOG(INFO, "port %u reserved UAR address space: %p",
621                 dev->data->port_id, addr);
622         priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
623         uar_base = addr; /* process local, don't reserve again. */
624         return 0;
625 }
626
627 /**
628  * Reserve UAR address space for secondary process, align with
629  * primary process.
630  *
631  * @param[in] dev
632  *   Pointer to Ethernet device.
633  *
634  * @return
635  *   0 on success, a negative errno value otherwise and rte_errno is set.
636  */
637 static int
638 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
639 {
640         struct priv *priv = dev->data->dev_private;
641         void *addr;
642
643         assert(priv->uar_base);
644         if (uar_base) { /* already reserved. */
645                 assert(uar_base == priv->uar_base);
646                 return 0;
647         }
648         /* anonymous mmap, no real memory consumption. */
649         addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
650                     PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
651         if (addr == MAP_FAILED) {
652                 DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
653                         dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
654                 rte_errno = ENXIO;
655                 return -rte_errno;
656         }
657         if (priv->uar_base != addr) {
658                 DRV_LOG(ERR,
659                         "port %u UAR address %p size %llu occupied, please"
660                         " adjust MLX5_UAR_OFFSET or try EAL parameter"
661                         " --base-virtaddr",
662                         dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
663                 rte_errno = ENXIO;
664                 return -rte_errno;
665         }
666         uar_base = addr; /* process local, don't reserve again */
667         DRV_LOG(INFO, "port %u reserved UAR address space: %p",
668                 dev->data->port_id, addr);
669         return 0;
670 }
671
672 /**
673  * Spawn an Ethernet device from Verbs information.
674  *
675  * @param dpdk_dev
676  *   Backing DPDK device.
677  * @param ibv_dev
678  *   Verbs device.
679  * @param vf
680  *   If nonzero, enable VF-specific features.
681  * @param[in] switch_info
682  *   Switch properties of Ethernet device.
683  *
684  * @return
685  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
686  *   is set. The following error is defined:
687  *
688  *   EBUSY: device is not supposed to be spawned.
689  */
690 static struct rte_eth_dev *
691 mlx5_dev_spawn(struct rte_device *dpdk_dev,
692                struct ibv_device *ibv_dev,
693                int vf,
694                const struct mlx5_switch_info *switch_info)
695 {
696         struct ibv_context *ctx;
697         struct ibv_device_attr_ex attr;
698         struct ibv_port_attr port_attr;
699         struct ibv_pd *pd = NULL;
700         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
701         struct mlx5_dev_config config = {
702                 .vf = !!vf,
703                 .tx_vec_en = 1,
704                 .rx_vec_en = 1,
705                 .mpw_hdr_dseg = 0,
706                 .txq_inline = MLX5_ARG_UNSET,
707                 .txqs_inline = MLX5_ARG_UNSET,
708                 .inline_max_packet_sz = MLX5_ARG_UNSET,
709                 .vf_nl_en = 1,
710                 .mprq = {
711                         .enabled = 0,
712                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
713                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
714                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
715                 },
716         };
717         struct rte_eth_dev *eth_dev = NULL;
718         struct priv *priv = NULL;
719         int err = 0;
720         unsigned int mps;
721         unsigned int cqe_comp;
722         unsigned int tunnel_en = 0;
723         unsigned int mpls_en = 0;
724         unsigned int swp = 0;
725         unsigned int mprq = 0;
726         unsigned int mprq_min_stride_size_n = 0;
727         unsigned int mprq_max_stride_size_n = 0;
728         unsigned int mprq_min_stride_num_n = 0;
729         unsigned int mprq_max_stride_num_n = 0;
730 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
731         struct ibv_counter_set_description cs_desc = { .counter_type = 0 };
732 #endif
733         struct ether_addr mac;
734         char name[RTE_ETH_NAME_MAX_LEN];
735         int own_domain_id = 0;
736         unsigned int i;
737
738         /* Determine if this port representor is supposed to be spawned. */
739         if (switch_info->representor && dpdk_dev->devargs) {
740                 struct rte_eth_devargs eth_da;
741
742                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
743                 if (err) {
744                         rte_errno = -err;
745                         DRV_LOG(ERR, "failed to process device arguments: %s",
746                                 strerror(rte_errno));
747                         return NULL;
748                 }
749                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
750                         if (eth_da.representor_ports[i] ==
751                             (uint16_t)switch_info->port_name)
752                                 break;
753                 if (i == eth_da.nb_representor_ports) {
754                         rte_errno = EBUSY;
755                         return NULL;
756                 }
757         }
758         /* Prepare shared data between primary and secondary process. */
759         mlx5_prepare_shared_data();
760         errno = 0;
761         ctx = mlx5_glue->open_device(ibv_dev);
762         if (!ctx) {
763                 rte_errno = errno ? errno : ENODEV;
764                 return NULL;
765         }
766 #ifdef HAVE_IBV_MLX5_MOD_SWP
767         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
768 #endif
769         /*
770          * Multi-packet send is supported by ConnectX-4 Lx PF as well
771          * as all ConnectX-5 devices.
772          */
773 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
774         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
775 #endif
776 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
777         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
778 #endif
779         mlx5_glue->dv_query_device(ctx, &dv_attr);
780         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
781                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
782                         DRV_LOG(DEBUG, "enhanced MPW is supported");
783                         mps = MLX5_MPW_ENHANCED;
784                 } else {
785                         DRV_LOG(DEBUG, "MPW is supported");
786                         mps = MLX5_MPW;
787                 }
788         } else {
789                 DRV_LOG(DEBUG, "MPW isn't supported");
790                 mps = MLX5_MPW_DISABLED;
791         }
792         config.mps = mps;
793 #ifdef HAVE_IBV_MLX5_MOD_SWP
794         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
795                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
796         DRV_LOG(DEBUG, "SWP support: %u", swp);
797 #endif
798         config.swp = !!swp;
799 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
800         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
801                 struct mlx5dv_striding_rq_caps mprq_caps =
802                         dv_attr.striding_rq_caps;
803
804                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
805                         mprq_caps.min_single_stride_log_num_of_bytes);
806                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
807                         mprq_caps.max_single_stride_log_num_of_bytes);
808                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
809                         mprq_caps.min_single_wqe_log_num_of_strides);
810                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
811                         mprq_caps.max_single_wqe_log_num_of_strides);
812                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
813                         mprq_caps.supported_qpts);
814                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
815                 mprq = 1;
816                 mprq_min_stride_size_n =
817                         mprq_caps.min_single_stride_log_num_of_bytes;
818                 mprq_max_stride_size_n =
819                         mprq_caps.max_single_stride_log_num_of_bytes;
820                 mprq_min_stride_num_n =
821                         mprq_caps.min_single_wqe_log_num_of_strides;
822                 mprq_max_stride_num_n =
823                         mprq_caps.max_single_wqe_log_num_of_strides;
824                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
825                                                    mprq_min_stride_num_n);
826         }
827 #endif
828         if (RTE_CACHE_LINE_SIZE == 128 &&
829             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
830                 cqe_comp = 0;
831         else
832                 cqe_comp = 1;
833         config.cqe_comp = cqe_comp;
834 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
835         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
836                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
837                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
838                              (dv_attr.tunnel_offloads_caps &
839                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
840         }
841         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
842                 tunnel_en ? "" : "not ");
843 #else
844         DRV_LOG(WARNING,
845                 "tunnel offloading disabled due to old OFED/rdma-core version");
846 #endif
847         config.tunnel_en = tunnel_en;
848 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
849         mpls_en = ((dv_attr.tunnel_offloads_caps &
850                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
851                    (dv_attr.tunnel_offloads_caps &
852                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
853         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
854                 mpls_en ? "" : "not ");
855 #else
856         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
857                 " old OFED/rdma-core version or firmware configuration");
858 #endif
859         config.mpls_en = mpls_en;
860         err = mlx5_glue->query_device_ex(ctx, NULL, &attr);
861         if (err) {
862                 DEBUG("ibv_query_device_ex() failed");
863                 goto error;
864         }
865         if (!switch_info->representor)
866                 rte_strlcpy(name, dpdk_dev->name, sizeof(name));
867         else
868                 snprintf(name, sizeof(name), "%s_representor_%u",
869                          dpdk_dev->name, switch_info->port_name);
870         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
871         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
872                 eth_dev = rte_eth_dev_attach_secondary(name);
873                 if (eth_dev == NULL) {
874                         DRV_LOG(ERR, "can not attach rte ethdev");
875                         rte_errno = ENOMEM;
876                         err = rte_errno;
877                         goto error;
878                 }
879                 eth_dev->device = dpdk_dev;
880                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
881                 err = mlx5_uar_init_secondary(eth_dev);
882                 if (err) {
883                         err = rte_errno;
884                         goto error;
885                 }
886                 /* Receive command fd from primary process */
887                 err = mlx5_socket_connect(eth_dev);
888                 if (err < 0) {
889                         err = rte_errno;
890                         goto error;
891                 }
892                 /* Remap UAR for Tx queues. */
893                 err = mlx5_tx_uar_remap(eth_dev, err);
894                 if (err) {
895                         err = rte_errno;
896                         goto error;
897                 }
898                 /*
899                  * Ethdev pointer is still required as input since
900                  * the primary device is not accessible from the
901                  * secondary process.
902                  */
903                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
904                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
905                 claim_zero(mlx5_glue->close_device(ctx));
906                 return eth_dev;
907         }
908         /* Check port status. */
909         err = mlx5_glue->query_port(ctx, 1, &port_attr);
910         if (err) {
911                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
912                 goto error;
913         }
914         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
915                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
916                 err = EINVAL;
917                 goto error;
918         }
919         if (port_attr.state != IBV_PORT_ACTIVE)
920                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
921                         mlx5_glue->port_state_str(port_attr.state),
922                         port_attr.state);
923         /* Allocate protection domain. */
924         pd = mlx5_glue->alloc_pd(ctx);
925         if (pd == NULL) {
926                 DRV_LOG(ERR, "PD allocation failure");
927                 err = ENOMEM;
928                 goto error;
929         }
930         priv = rte_zmalloc("ethdev private structure",
931                            sizeof(*priv),
932                            RTE_CACHE_LINE_SIZE);
933         if (priv == NULL) {
934                 DRV_LOG(ERR, "priv allocation failure");
935                 err = ENOMEM;
936                 goto error;
937         }
938         priv->ctx = ctx;
939         strncpy(priv->ibdev_name, priv->ctx->device->name,
940                 sizeof(priv->ibdev_name));
941         strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
942                 sizeof(priv->ibdev_path));
943         priv->device_attr = attr;
944         priv->pd = pd;
945         priv->mtu = ETHER_MTU;
946 #ifndef RTE_ARCH_64
947         /* Initialize UAR access locks for 32bit implementations. */
948         rte_spinlock_init(&priv->uar_lock_cq);
949         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
950                 rte_spinlock_init(&priv->uar_lock[i]);
951 #endif
952         /* Some internal functions rely on Netlink sockets, open them now. */
953         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
954         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
955         priv->nl_sn = 0;
956         priv->representor = !!switch_info->representor;
957         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
958         priv->representor_id =
959                 switch_info->representor ? switch_info->port_name : -1;
960         /*
961          * Look for sibling devices in order to reuse their switch domain
962          * if any, otherwise allocate one.
963          */
964         i = mlx5_dev_to_port_id(dpdk_dev, NULL, 0);
965         if (i > 0) {
966                 uint16_t port_id[i];
967
968                 i = RTE_MIN(mlx5_dev_to_port_id(dpdk_dev, port_id, i), i);
969                 while (i--) {
970                         const struct priv *opriv =
971                                 rte_eth_devices[port_id[i]].data->dev_private;
972
973                         if (!opriv ||
974                             opriv->domain_id ==
975                             RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
976                                 continue;
977                         priv->domain_id = opriv->domain_id;
978                         break;
979                 }
980         }
981         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
982                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
983                 if (err) {
984                         err = rte_errno;
985                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
986                                 strerror(rte_errno));
987                         goto error;
988                 }
989                 own_domain_id = 1;
990         }
991         err = mlx5_args(&config, dpdk_dev->devargs);
992         if (err) {
993                 err = rte_errno;
994                 DRV_LOG(ERR, "failed to process device arguments: %s",
995                         strerror(rte_errno));
996                 goto error;
997         }
998         config.hw_csum = !!(attr.device_cap_flags_ex & IBV_DEVICE_RAW_IP_CSUM);
999         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1000                 (config.hw_csum ? "" : "not "));
1001 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1002         config.flow_counter_en = !!attr.max_counter_sets;
1003         mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
1004         DRV_LOG(DEBUG, "counter type = %d, num of cs = %ld, attributes = %d",
1005                 cs_desc.counter_type, cs_desc.num_of_cs,
1006                 cs_desc.attributes);
1007 #endif
1008         config.ind_table_max_size =
1009                 attr.rss_caps.max_rwq_indirection_table_size;
1010         /*
1011          * Remove this check once DPDK supports larger/variable
1012          * indirection tables.
1013          */
1014         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1015                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1016         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1017                 config.ind_table_max_size);
1018         config.hw_vlan_strip = !!(attr.raw_packet_caps &
1019                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1020         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1021                 (config.hw_vlan_strip ? "" : "not "));
1022         config.hw_fcs_strip = !!(attr.raw_packet_caps &
1023                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1024         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1025                 (config.hw_fcs_strip ? "" : "not "));
1026 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
1027         config.hw_padding = !!attr.rx_pad_end_addr_align;
1028 #endif
1029         DRV_LOG(DEBUG, "hardware Rx end alignment padding is %ssupported",
1030                 (config.hw_padding ? "" : "not "));
1031         config.tso = (attr.tso_caps.max_tso > 0 &&
1032                       (attr.tso_caps.supported_qpts &
1033                        (1 << IBV_QPT_RAW_PACKET)));
1034         if (config.tso)
1035                 config.tso_max_payload_sz = attr.tso_caps.max_tso;
1036         if (config.mps && !mps) {
1037                 DRV_LOG(ERR,
1038                         "multi-packet send not supported on this device"
1039                         " (" MLX5_TXQ_MPW_EN ")");
1040                 err = ENOTSUP;
1041                 goto error;
1042         }
1043         DRV_LOG(INFO, "%sMPS is %s",
1044                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1045                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1046         if (config.cqe_comp && !cqe_comp) {
1047                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1048                 config.cqe_comp = 0;
1049         }
1050         if (config.mprq.enabled && mprq) {
1051                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1052                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
1053                         config.mprq.stride_num_n =
1054                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1055                                         mprq_min_stride_num_n);
1056                         DRV_LOG(WARNING,
1057                                 "the number of strides"
1058                                 " for Multi-Packet RQ is out of range,"
1059                                 " setting default value (%u)",
1060                                 1 << config.mprq.stride_num_n);
1061                 }
1062                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1063                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1064         } else if (config.mprq.enabled && !mprq) {
1065                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1066                 config.mprq.enabled = 0;
1067         }
1068         eth_dev = rte_eth_dev_allocate(name);
1069         if (eth_dev == NULL) {
1070                 DRV_LOG(ERR, "can not allocate rte ethdev");
1071                 err = ENOMEM;
1072                 goto error;
1073         }
1074         if (priv->representor)
1075                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1076         eth_dev->data->dev_private = priv;
1077         priv->dev_data = eth_dev->data;
1078         eth_dev->data->mac_addrs = priv->mac;
1079         eth_dev->device = dpdk_dev;
1080         eth_dev->device->driver = &mlx5_driver.driver;
1081         err = mlx5_uar_init_primary(eth_dev);
1082         if (err) {
1083                 err = rte_errno;
1084                 goto error;
1085         }
1086         /* Configure the first MAC address by default. */
1087         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1088                 DRV_LOG(ERR,
1089                         "port %u cannot get MAC address, is mlx5_en"
1090                         " loaded? (errno: %s)",
1091                         eth_dev->data->port_id, strerror(rte_errno));
1092                 err = ENODEV;
1093                 goto error;
1094         }
1095         DRV_LOG(INFO,
1096                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1097                 eth_dev->data->port_id,
1098                 mac.addr_bytes[0], mac.addr_bytes[1],
1099                 mac.addr_bytes[2], mac.addr_bytes[3],
1100                 mac.addr_bytes[4], mac.addr_bytes[5]);
1101 #ifndef NDEBUG
1102         {
1103                 char ifname[IF_NAMESIZE];
1104
1105                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1106                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1107                                 eth_dev->data->port_id, ifname);
1108                 else
1109                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1110                                 eth_dev->data->port_id);
1111         }
1112 #endif
1113         /* Get actual MTU if possible. */
1114         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1115         if (err) {
1116                 err = rte_errno;
1117                 goto error;
1118         }
1119         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1120                 priv->mtu);
1121         /* Initialize burst functions to prevent crashes before link-up. */
1122         eth_dev->rx_pkt_burst = removed_rx_burst;
1123         eth_dev->tx_pkt_burst = removed_tx_burst;
1124         eth_dev->dev_ops = &mlx5_dev_ops;
1125         /* Register MAC address. */
1126         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1127         if (vf && config.vf_nl_en)
1128                 mlx5_nl_mac_addr_sync(eth_dev);
1129         priv->mnl_socket = mlx5_nl_flow_socket_create();
1130         if (!priv->mnl_socket) {
1131                 err = -rte_errno;
1132                 DRV_LOG(WARNING,
1133                         "flow rules relying on switch offloads will not be"
1134                         " supported: cannot open libmnl socket: %s",
1135                         strerror(rte_errno));
1136         } else {
1137                 struct rte_flow_error error;
1138                 unsigned int ifindex = mlx5_ifindex(eth_dev);
1139
1140                 if (!ifindex) {
1141                         err = -rte_errno;
1142                         error.message =
1143                                 "cannot retrieve network interface index";
1144                 } else {
1145                         err = mlx5_nl_flow_init(priv->mnl_socket, ifindex,
1146                                                 &error);
1147                 }
1148                 if (err) {
1149                         DRV_LOG(WARNING,
1150                                 "flow rules relying on switch offloads will"
1151                                 " not be supported: %s: %s",
1152                                 error.message, strerror(rte_errno));
1153                         mlx5_nl_flow_socket_destroy(priv->mnl_socket);
1154                         priv->mnl_socket = NULL;
1155                 }
1156         }
1157         TAILQ_INIT(&priv->flows);
1158         TAILQ_INIT(&priv->ctrl_flows);
1159         /* Hint libmlx5 to use PMD allocator for data plane resources */
1160         struct mlx5dv_ctx_allocators alctr = {
1161                 .alloc = &mlx5_alloc_verbs_buf,
1162                 .free = &mlx5_free_verbs_buf,
1163                 .data = priv,
1164         };
1165         mlx5_glue->dv_set_context_attr(ctx, MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1166                                        (void *)((uintptr_t)&alctr));
1167         /* Bring Ethernet device up. */
1168         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1169                 eth_dev->data->port_id);
1170         mlx5_set_link_up(eth_dev);
1171         /*
1172          * Even though the interrupt handler is not installed yet,
1173          * interrupts will still trigger on the asyn_fd from
1174          * Verbs context returned by ibv_open_device().
1175          */
1176         mlx5_link_update(eth_dev, 0);
1177         /* Store device configuration on private structure. */
1178         priv->config = config;
1179         /* Supported Verbs flow priority number detection. */
1180         err = mlx5_flow_discover_priorities(eth_dev);
1181         if (err < 0)
1182                 goto error;
1183         priv->config.flow_prio = err;
1184         /*
1185          * Once the device is added to the list of memory event
1186          * callback, its global MR cache table cannot be expanded
1187          * on the fly because of deadlock. If it overflows, lookup
1188          * should be done by searching MR list linearly, which is slow.
1189          */
1190         err = mlx5_mr_btree_init(&priv->mr.cache,
1191                                  MLX5_MR_BTREE_CACHE_N * 2,
1192                                  eth_dev->device->numa_node);
1193         if (err) {
1194                 err = rte_errno;
1195                 goto error;
1196         }
1197         /* Add device to memory callback list. */
1198         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1199         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1200                          priv, mem_event_cb);
1201         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1202         return eth_dev;
1203 error:
1204         if (priv) {
1205                 if (priv->nl_socket_route >= 0)
1206                         close(priv->nl_socket_route);
1207                 if (priv->nl_socket_rdma >= 0)
1208                         close(priv->nl_socket_rdma);
1209                 if (priv->mnl_socket)
1210                         mlx5_nl_flow_socket_destroy(priv->mnl_socket);
1211                 if (own_domain_id)
1212                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1213                 rte_free(priv);
1214         }
1215         if (pd)
1216                 claim_zero(mlx5_glue->dealloc_pd(pd));
1217         if (eth_dev)
1218                 rte_eth_dev_release_port(eth_dev);
1219         if (ctx)
1220                 claim_zero(mlx5_glue->close_device(ctx));
1221         assert(err > 0);
1222         rte_errno = err;
1223         return NULL;
1224 }
1225
1226 /** Data associated with devices to spawn. */
1227 struct mlx5_dev_spawn_data {
1228         unsigned int ifindex; /**< Network interface index. */
1229         struct mlx5_switch_info info; /**< Switch information. */
1230         struct ibv_device *ibv_dev; /**< Associated IB device. */
1231         struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
1232 };
1233
1234 /**
1235  * Comparison callback to sort device data.
1236  *
1237  * This is meant to be used with qsort().
1238  *
1239  * @param a[in]
1240  *   Pointer to pointer to first data object.
1241  * @param b[in]
1242  *   Pointer to pointer to second data object.
1243  *
1244  * @return
1245  *   0 if both objects are equal, less than 0 if the first argument is less
1246  *   than the second, greater than 0 otherwise.
1247  */
1248 static int
1249 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1250 {
1251         const struct mlx5_switch_info *si_a =
1252                 &((const struct mlx5_dev_spawn_data *)a)->info;
1253         const struct mlx5_switch_info *si_b =
1254                 &((const struct mlx5_dev_spawn_data *)b)->info;
1255         int ret;
1256
1257         /* Master device first. */
1258         ret = si_b->master - si_a->master;
1259         if (ret)
1260                 return ret;
1261         /* Then representor devices. */
1262         ret = si_b->representor - si_a->representor;
1263         if (ret)
1264                 return ret;
1265         /* Unidentified devices come last in no specific order. */
1266         if (!si_a->representor)
1267                 return 0;
1268         /* Order representors by name. */
1269         return si_a->port_name - si_b->port_name;
1270 }
1271
1272 /**
1273  * DPDK callback to register a PCI device.
1274  *
1275  * This function spawns Ethernet devices out of a given PCI device.
1276  *
1277  * @param[in] pci_drv
1278  *   PCI driver structure (mlx5_driver).
1279  * @param[in] pci_dev
1280  *   PCI device information.
1281  *
1282  * @return
1283  *   0 on success, a negative errno value otherwise and rte_errno is set.
1284  */
1285 static int
1286 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1287                struct rte_pci_device *pci_dev)
1288 {
1289         struct ibv_device **ibv_list;
1290         unsigned int n = 0;
1291         int vf;
1292         int ret;
1293
1294         assert(pci_drv == &mlx5_driver);
1295         errno = 0;
1296         ibv_list = mlx5_glue->get_device_list(&ret);
1297         if (!ibv_list) {
1298                 rte_errno = errno ? errno : ENOSYS;
1299                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1300                 return -rte_errno;
1301         }
1302
1303         struct ibv_device *ibv_match[ret + 1];
1304
1305         while (ret-- > 0) {
1306                 struct rte_pci_addr pci_addr;
1307
1308                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1309                 if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1310                         continue;
1311                 if (pci_dev->addr.domain != pci_addr.domain ||
1312                     pci_dev->addr.bus != pci_addr.bus ||
1313                     pci_dev->addr.devid != pci_addr.devid ||
1314                     pci_dev->addr.function != pci_addr.function)
1315                         continue;
1316                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1317                         ibv_list[ret]->name);
1318                 ibv_match[n++] = ibv_list[ret];
1319         }
1320         ibv_match[n] = NULL;
1321
1322         struct mlx5_dev_spawn_data list[n];
1323         int nl_route = n ? mlx5_nl_init(NETLINK_ROUTE) : -1;
1324         int nl_rdma = n ? mlx5_nl_init(NETLINK_RDMA) : -1;
1325         unsigned int i;
1326         unsigned int u;
1327
1328         /*
1329          * The existence of several matching entries (n > 1) means port
1330          * representors have been instantiated. No existing Verbs call nor
1331          * /sys entries can tell them apart, this can only be done through
1332          * Netlink calls assuming kernel drivers are recent enough to
1333          * support them.
1334          *
1335          * In the event of identification failure through Netlink, try again
1336          * through sysfs, then either:
1337          *
1338          * 1. No device matches (n == 0), complain and bail out.
1339          * 2. A single IB device matches (n == 1) and is not a representor,
1340          *    assume no switch support.
1341          * 3. Otherwise no safe assumptions can be made; complain louder and
1342          *    bail out.
1343          */
1344         for (i = 0; i != n; ++i) {
1345                 list[i].ibv_dev = ibv_match[i];
1346                 list[i].eth_dev = NULL;
1347                 if (nl_rdma < 0)
1348                         list[i].ifindex = 0;
1349                 else
1350                         list[i].ifindex = mlx5_nl_ifindex
1351                                 (nl_rdma, list[i].ibv_dev->name);
1352                 if (nl_route < 0 ||
1353                     !list[i].ifindex ||
1354                     mlx5_nl_switch_info(nl_route, list[i].ifindex,
1355                                         &list[i].info) ||
1356                     ((!list[i].info.representor && !list[i].info.master) &&
1357                      mlx5_sysfs_switch_info(list[i].ifindex, &list[i].info))) {
1358                         list[i].ifindex = 0;
1359                         memset(&list[i].info, 0, sizeof(list[i].info));
1360                         continue;
1361                 }
1362         }
1363         if (nl_rdma >= 0)
1364                 close(nl_rdma);
1365         if (nl_route >= 0)
1366                 close(nl_route);
1367         /* Count unidentified devices. */
1368         for (u = 0, i = 0; i != n; ++i)
1369                 if (!list[i].info.master && !list[i].info.representor)
1370                         ++u;
1371         if (u) {
1372                 if (n == 1 && u == 1) {
1373                         /* Case #2. */
1374                         DRV_LOG(INFO, "no switch support detected");
1375                 } else {
1376                         /* Case #3. */
1377                         DRV_LOG(ERR,
1378                                 "unable to tell which of the matching devices"
1379                                 " is the master (lack of kernel support?)");
1380                         n = 0;
1381                 }
1382         }
1383         /*
1384          * Sort list to probe devices in natural order for users convenience
1385          * (i.e. master first, then representors from lowest to highest ID).
1386          */
1387         if (n)
1388                 qsort(list, n, sizeof(*list), mlx5_dev_spawn_data_cmp);
1389         switch (pci_dev->id.device_id) {
1390         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1391         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1392         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1393         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1394                 vf = 1;
1395                 break;
1396         default:
1397                 vf = 0;
1398         }
1399         for (i = 0; i != n; ++i) {
1400                 uint32_t restore;
1401
1402                 list[i].eth_dev = mlx5_dev_spawn
1403                         (&pci_dev->device, list[i].ibv_dev, vf, &list[i].info);
1404                 if (!list[i].eth_dev) {
1405                         if (rte_errno != EBUSY)
1406                                 break;
1407                         /* Device is disabled, ignore it. */
1408                         continue;
1409                 }
1410                 restore = list[i].eth_dev->data->dev_flags;
1411                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1412                 /* Restore non-PCI flags cleared by the above call. */
1413                 list[i].eth_dev->data->dev_flags |= restore;
1414                 rte_eth_dev_probing_finish(list[i].eth_dev);
1415         }
1416         mlx5_glue->free_device_list(ibv_list);
1417         if (!n) {
1418                 DRV_LOG(WARNING,
1419                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1420                         " are kernel drivers loaded?",
1421                         pci_dev->addr.domain, pci_dev->addr.bus,
1422                         pci_dev->addr.devid, pci_dev->addr.function);
1423                 rte_errno = ENOENT;
1424                 ret = -rte_errno;
1425         } else if (i != n) {
1426                 DRV_LOG(ERR,
1427                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1428                         " encountering an error: %s",
1429                         pci_dev->addr.domain, pci_dev->addr.bus,
1430                         pci_dev->addr.devid, pci_dev->addr.function,
1431                         strerror(rte_errno));
1432                 ret = -rte_errno;
1433                 /* Roll back. */
1434                 while (i--) {
1435                         if (!list[i].eth_dev)
1436                                 continue;
1437                         mlx5_dev_close(list[i].eth_dev);
1438                         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1439                                 rte_free(list[i].eth_dev->data->dev_private);
1440                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1441                 }
1442                 /* Restore original error. */
1443                 rte_errno = -ret;
1444         } else {
1445                 ret = 0;
1446         }
1447         return ret;
1448 }
1449
1450 static const struct rte_pci_id mlx5_pci_id_map[] = {
1451         {
1452                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1453                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1454         },
1455         {
1456                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1457                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1458         },
1459         {
1460                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1461                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1462         },
1463         {
1464                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1465                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1466         },
1467         {
1468                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1469                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1470         },
1471         {
1472                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1473                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1474         },
1475         {
1476                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1477                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1478         },
1479         {
1480                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1481                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1482         },
1483         {
1484                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1485                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
1486         },
1487         {
1488                 .vendor_id = 0
1489         }
1490 };
1491
1492 static struct rte_pci_driver mlx5_driver = {
1493         .driver = {
1494                 .name = MLX5_DRIVER_NAME
1495         },
1496         .id_table = mlx5_pci_id_map,
1497         .probe = mlx5_pci_probe,
1498         .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1499 };
1500
1501 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1502
1503 /**
1504  * Suffix RTE_EAL_PMD_PATH with "-glue".
1505  *
1506  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1507  * suffixing its last component.
1508  *
1509  * @param buf[out]
1510  *   Output buffer, should be large enough otherwise NULL is returned.
1511  * @param size
1512  *   Size of @p out.
1513  *
1514  * @return
1515  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1516  */
1517 static char *
1518 mlx5_glue_path(char *buf, size_t size)
1519 {
1520         static const char *const bad[] = { "/", ".", "..", NULL };
1521         const char *path = RTE_EAL_PMD_PATH;
1522         size_t len = strlen(path);
1523         size_t off;
1524         int i;
1525
1526         while (len && path[len - 1] == '/')
1527                 --len;
1528         for (off = len; off && path[off - 1] != '/'; --off)
1529                 ;
1530         for (i = 0; bad[i]; ++i)
1531                 if (!strncmp(path + off, bad[i], (int)(len - off)))
1532                         goto error;
1533         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1534         if (i == -1 || (size_t)i >= size)
1535                 goto error;
1536         return buf;
1537 error:
1538         DRV_LOG(ERR,
1539                 "unable to append \"-glue\" to last component of"
1540                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1541                 " please re-configure DPDK");
1542         return NULL;
1543 }
1544
1545 /**
1546  * Initialization routine for run-time dependency on rdma-core.
1547  */
1548 static int
1549 mlx5_glue_init(void)
1550 {
1551         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1552         const char *path[] = {
1553                 /*
1554                  * A basic security check is necessary before trusting
1555                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1556                  */
1557                 (geteuid() == getuid() && getegid() == getgid() ?
1558                  getenv("MLX5_GLUE_PATH") : NULL),
1559                 /*
1560                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1561                  * variant, otherwise let dlopen() look up libraries on its
1562                  * own.
1563                  */
1564                 (*RTE_EAL_PMD_PATH ?
1565                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1566         };
1567         unsigned int i = 0;
1568         void *handle = NULL;
1569         void **sym;
1570         const char *dlmsg;
1571
1572         while (!handle && i != RTE_DIM(path)) {
1573                 const char *end;
1574                 size_t len;
1575                 int ret;
1576
1577                 if (!path[i]) {
1578                         ++i;
1579                         continue;
1580                 }
1581                 end = strpbrk(path[i], ":;");
1582                 if (!end)
1583                         end = path[i] + strlen(path[i]);
1584                 len = end - path[i];
1585                 ret = 0;
1586                 do {
1587                         char name[ret + 1];
1588
1589                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1590                                        (int)len, path[i],
1591                                        (!len || *(end - 1) == '/') ? "" : "/");
1592                         if (ret == -1)
1593                                 break;
1594                         if (sizeof(name) != (size_t)ret + 1)
1595                                 continue;
1596                         DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1597                                 name);
1598                         handle = dlopen(name, RTLD_LAZY);
1599                         break;
1600                 } while (1);
1601                 path[i] = end + 1;
1602                 if (!*end)
1603                         ++i;
1604         }
1605         if (!handle) {
1606                 rte_errno = EINVAL;
1607                 dlmsg = dlerror();
1608                 if (dlmsg)
1609                         DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1610                 goto glue_error;
1611         }
1612         sym = dlsym(handle, "mlx5_glue");
1613         if (!sym || !*sym) {
1614                 rte_errno = EINVAL;
1615                 dlmsg = dlerror();
1616                 if (dlmsg)
1617                         DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1618                 goto glue_error;
1619         }
1620         mlx5_glue = *sym;
1621         return 0;
1622 glue_error:
1623         if (handle)
1624                 dlclose(handle);
1625         DRV_LOG(WARNING,
1626                 "cannot initialize PMD due to missing run-time dependency on"
1627                 " rdma-core libraries (libibverbs, libmlx5)");
1628         return -rte_errno;
1629 }
1630
1631 #endif
1632
1633 /**
1634  * Driver initialization routine.
1635  */
1636 RTE_INIT(rte_mlx5_pmd_init)
1637 {
1638         /* Initialize driver log type. */
1639         mlx5_logtype = rte_log_register("pmd.net.mlx5");
1640         if (mlx5_logtype >= 0)
1641                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1642
1643         /* Build the static tables for Verbs conversion. */
1644         mlx5_set_ptype_table();
1645         mlx5_set_cksum_table();
1646         mlx5_set_swp_types_table();
1647         /*
1648          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1649          * huge pages. Calling ibv_fork_init() during init allows
1650          * applications to use fork() safely for purposes other than
1651          * using this PMD, which is not supported in forked processes.
1652          */
1653         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1654         /* Match the size of Rx completion entry to the size of a cacheline. */
1655         if (RTE_CACHE_LINE_SIZE == 128)
1656                 setenv("MLX5_CQE_SIZE", "128", 0);
1657         /*
1658          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
1659          * cleanup all the Verbs resources even when the device was removed.
1660          */
1661         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
1662 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1663         if (mlx5_glue_init())
1664                 return;
1665         assert(mlx5_glue);
1666 #endif
1667 #ifndef NDEBUG
1668         /* Glue structure must not contain any NULL pointers. */
1669         {
1670                 unsigned int i;
1671
1672                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1673                         assert(((const void *const *)mlx5_glue)[i]);
1674         }
1675 #endif
1676         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1677                 DRV_LOG(ERR,
1678                         "rdma-core glue \"%s\" mismatch: \"%s\" is required",
1679                         mlx5_glue->version, MLX5_GLUE_VERSION);
1680                 return;
1681         }
1682         mlx5_glue->fork_init();
1683         rte_pci_register(&mlx5_driver);
1684 }
1685
1686 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1687 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1688 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");