net/mlx4: fix build with -fno-common
[dpdk.git] / drivers / net / mlx4 / mlx4.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2012 6WIND S.A.
3  * Copyright 2012 Mellanox Technologies, Ltd
4  */
5
6 /**
7  * @file
8  * mlx4 driver initialization.
9  */
10
11 #include <errno.h>
12 #include <inttypes.h>
13 #include <stddef.h>
14 #include <stdint.h>
15 #include <stdio.h>
16 #include <stdlib.h>
17 #include <string.h>
18 #include <sys/mman.h>
19 #include <unistd.h>
20 #ifdef RTE_IBVERBS_LINK_DLOPEN
21 #include <dlfcn.h>
22 #endif
23
24 /* Verbs headers do not support -pedantic. */
25 #ifdef PEDANTIC
26 #pragma GCC diagnostic ignored "-Wpedantic"
27 #endif
28 #include <infiniband/verbs.h>
29 #ifdef PEDANTIC
30 #pragma GCC diagnostic error "-Wpedantic"
31 #endif
32
33 #include <rte_common.h>
34 #include <rte_dev.h>
35 #include <rte_errno.h>
36 #include <rte_ethdev_driver.h>
37 #include <rte_ethdev_pci.h>
38 #include <rte_ether.h>
39 #include <rte_flow.h>
40 #include <rte_interrupts.h>
41 #include <rte_kvargs.h>
42 #include <rte_malloc.h>
43 #include <rte_mbuf.h>
44
45 #include "mlx4.h"
46 #include "mlx4_glue.h"
47 #include "mlx4_flow.h"
48 #include "mlx4_mr.h"
49 #include "mlx4_rxtx.h"
50 #include "mlx4_utils.h"
51
52 #ifdef MLX4_GLUE
53 const struct mlx4_glue *mlx4_glue;
54 #endif
55
56 static const char *MZ_MLX4_PMD_SHARED_DATA = "mlx4_pmd_shared_data";
57
58 /* Shared memory between primary and secondary processes. */
59 struct mlx4_shared_data *mlx4_shared_data;
60
61 /* Spinlock for mlx4_shared_data allocation. */
62 static rte_spinlock_t mlx4_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
63
64 /* Process local data for secondary processes. */
65 static struct mlx4_local_data mlx4_local_data;
66
67 /** Driver-specific log messages type. */
68 int mlx4_logtype;
69
70 /** Configuration structure for device arguments. */
71 struct mlx4_conf {
72         struct {
73                 uint32_t present; /**< Bit-field for existing ports. */
74                 uint32_t enabled; /**< Bit-field for user-enabled ports. */
75         } ports;
76         int mr_ext_memseg_en;
77         /** Whether memseg should be extended for MR creation. */
78 };
79
80 /* Available parameters list. */
81 const char *pmd_mlx4_init_params[] = {
82         MLX4_PMD_PORT_KVARG,
83         MLX4_MR_EXT_MEMSEG_EN_KVARG,
84         NULL,
85 };
86
87 static void mlx4_dev_stop(struct rte_eth_dev *dev);
88
89 /**
90  * Initialize shared data between primary and secondary process.
91  *
92  * A memzone is reserved by primary process and secondary processes attach to
93  * the memzone.
94  *
95  * @return
96  *   0 on success, a negative errno value otherwise and rte_errno is set.
97  */
98 static int
99 mlx4_init_shared_data(void)
100 {
101         const struct rte_memzone *mz;
102         int ret = 0;
103
104         rte_spinlock_lock(&mlx4_shared_data_lock);
105         if (mlx4_shared_data == NULL) {
106                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
107                         /* Allocate shared memory. */
108                         mz = rte_memzone_reserve(MZ_MLX4_PMD_SHARED_DATA,
109                                                  sizeof(*mlx4_shared_data),
110                                                  SOCKET_ID_ANY, 0);
111                         if (mz == NULL) {
112                                 ERROR("Cannot allocate mlx4 shared data\n");
113                                 ret = -rte_errno;
114                                 goto error;
115                         }
116                         mlx4_shared_data = mz->addr;
117                         memset(mlx4_shared_data, 0, sizeof(*mlx4_shared_data));
118                         rte_spinlock_init(&mlx4_shared_data->lock);
119                 } else {
120                         /* Lookup allocated shared memory. */
121                         mz = rte_memzone_lookup(MZ_MLX4_PMD_SHARED_DATA);
122                         if (mz == NULL) {
123                                 ERROR("Cannot attach mlx4 shared data\n");
124                                 ret = -rte_errno;
125                                 goto error;
126                         }
127                         mlx4_shared_data = mz->addr;
128                         memset(&mlx4_local_data, 0, sizeof(mlx4_local_data));
129                 }
130         }
131 error:
132         rte_spinlock_unlock(&mlx4_shared_data_lock);
133         return ret;
134 }
135
136 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
137 /**
138  * Verbs callback to allocate a memory. This function should allocate the space
139  * according to the size provided residing inside a huge page.
140  * Please note that all allocation must respect the alignment from libmlx4
141  * (i.e. currently sysconf(_SC_PAGESIZE)).
142  *
143  * @param[in] size
144  *   The size in bytes of the memory to allocate.
145  * @param[in] data
146  *   A pointer to the callback data.
147  *
148  * @return
149  *   Allocated buffer, NULL otherwise and rte_errno is set.
150  */
151 static void *
152 mlx4_alloc_verbs_buf(size_t size, void *data)
153 {
154         struct mlx4_priv *priv = data;
155         void *ret;
156         size_t alignment = sysconf(_SC_PAGESIZE);
157         unsigned int socket = SOCKET_ID_ANY;
158
159         if (priv->verbs_alloc_ctx.type == MLX4_VERBS_ALLOC_TYPE_TX_QUEUE) {
160                 const struct txq *txq = priv->verbs_alloc_ctx.obj;
161
162                 socket = txq->socket;
163         } else if (priv->verbs_alloc_ctx.type ==
164                    MLX4_VERBS_ALLOC_TYPE_RX_QUEUE) {
165                 const struct rxq *rxq = priv->verbs_alloc_ctx.obj;
166
167                 socket = rxq->socket;
168         }
169         MLX4_ASSERT(data != NULL);
170         ret = rte_malloc_socket(__func__, size, alignment, socket);
171         if (!ret && size)
172                 rte_errno = ENOMEM;
173         return ret;
174 }
175
176 /**
177  * Verbs callback to free a memory.
178  *
179  * @param[in] ptr
180  *   A pointer to the memory to free.
181  * @param[in] data
182  *   A pointer to the callback data.
183  */
184 static void
185 mlx4_free_verbs_buf(void *ptr, void *data __rte_unused)
186 {
187         MLX4_ASSERT(data != NULL);
188         rte_free(ptr);
189 }
190 #endif
191
192 /**
193  * Initialize process private data structure.
194  *
195  * @param dev
196  *   Pointer to Ethernet device structure.
197  *
198  * @return
199  *   0 on success, a negative errno value otherwise and rte_errno is set.
200  */
201 static int
202 mlx4_proc_priv_init(struct rte_eth_dev *dev)
203 {
204         struct mlx4_proc_priv *ppriv;
205         size_t ppriv_size;
206
207         /*
208          * UAR register table follows the process private structure. BlueFlame
209          * registers for Tx queues are stored in the table.
210          */
211         ppriv_size = sizeof(struct mlx4_proc_priv) +
212                      dev->data->nb_tx_queues * sizeof(void *);
213         ppriv = rte_malloc_socket("mlx4_proc_priv", ppriv_size,
214                                   RTE_CACHE_LINE_SIZE, dev->device->numa_node);
215         if (!ppriv) {
216                 rte_errno = ENOMEM;
217                 return -rte_errno;
218         }
219         ppriv->uar_table_sz = ppriv_size;
220         dev->process_private = ppriv;
221         return 0;
222 }
223
224 /**
225  * Un-initialize process private data structure.
226  *
227  * @param dev
228  *   Pointer to Ethernet device structure.
229  */
230 static void
231 mlx4_proc_priv_uninit(struct rte_eth_dev *dev)
232 {
233         if (!dev->process_private)
234                 return;
235         rte_free(dev->process_private);
236         dev->process_private = NULL;
237 }
238
239 /**
240  * DPDK callback for Ethernet device configuration.
241  *
242  * @param dev
243  *   Pointer to Ethernet device structure.
244  *
245  * @return
246  *   0 on success, negative errno value otherwise and rte_errno is set.
247  */
248 static int
249 mlx4_dev_configure(struct rte_eth_dev *dev)
250 {
251         struct mlx4_priv *priv = dev->data->dev_private;
252         struct rte_flow_error error;
253         int ret;
254
255         if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
256                 dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
257
258         /* Prepare internal flow rules. */
259         ret = mlx4_flow_sync(priv, &error);
260         if (ret) {
261                 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
262                       " flow error type %d, cause %p, message: %s",
263                       -ret, strerror(-ret), error.type, error.cause,
264                       error.message ? error.message : "(unspecified)");
265                 goto exit;
266         }
267         ret = mlx4_intr_install(priv);
268         if (ret) {
269                 ERROR("%p: interrupt handler installation failed",
270                       (void *)dev);
271                 goto exit;
272         }
273         ret = mlx4_proc_priv_init(dev);
274         if (ret) {
275                 ERROR("%p: process private data allocation failed",
276                       (void *)dev);
277                 goto exit;
278         }
279 exit:
280         return ret;
281 }
282
283 /**
284  * DPDK callback to start the device.
285  *
286  * Simulate device start by initializing common RSS resources and attaching
287  * all configured flows.
288  *
289  * @param dev
290  *   Pointer to Ethernet device structure.
291  *
292  * @return
293  *   0 on success, negative errno value otherwise and rte_errno is set.
294  */
295 static int
296 mlx4_dev_start(struct rte_eth_dev *dev)
297 {
298         struct mlx4_priv *priv = dev->data->dev_private;
299         struct rte_flow_error error;
300         int ret;
301
302         if (priv->started)
303                 return 0;
304         DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
305         priv->started = 1;
306         ret = mlx4_rss_init(priv);
307         if (ret) {
308                 ERROR("%p: cannot initialize RSS resources: %s",
309                       (void *)dev, strerror(-ret));
310                 goto err;
311         }
312 #ifdef RTE_LIBRTE_MLX4_DEBUG
313         mlx4_mr_dump_dev(dev);
314 #endif
315         ret = mlx4_rxq_intr_enable(priv);
316         if (ret) {
317                 ERROR("%p: interrupt handler installation failed",
318                      (void *)dev);
319                 goto err;
320         }
321         ret = mlx4_flow_sync(priv, &error);
322         if (ret) {
323                 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
324                       " flow error type %d, cause %p, message: %s",
325                       (void *)dev,
326                       -ret, strerror(-ret), error.type, error.cause,
327                       error.message ? error.message : "(unspecified)");
328                 goto err;
329         }
330         rte_wmb();
331         dev->tx_pkt_burst = mlx4_tx_burst;
332         dev->rx_pkt_burst = mlx4_rx_burst;
333         /* Enable datapath on secondary process. */
334         mlx4_mp_req_start_rxtx(dev);
335         return 0;
336 err:
337         mlx4_dev_stop(dev);
338         return ret;
339 }
340
341 /**
342  * DPDK callback to stop the device.
343  *
344  * Simulate device stop by detaching all configured flows.
345  *
346  * @param dev
347  *   Pointer to Ethernet device structure.
348  */
349 static void
350 mlx4_dev_stop(struct rte_eth_dev *dev)
351 {
352         struct mlx4_priv *priv = dev->data->dev_private;
353
354         if (!priv->started)
355                 return;
356         DEBUG("%p: detaching flows from all RX queues", (void *)dev);
357         priv->started = 0;
358         dev->tx_pkt_burst = mlx4_tx_burst_removed;
359         dev->rx_pkt_burst = mlx4_rx_burst_removed;
360         rte_wmb();
361         /* Disable datapath on secondary process. */
362         mlx4_mp_req_stop_rxtx(dev);
363         mlx4_flow_sync(priv, NULL);
364         mlx4_rxq_intr_disable(priv);
365         mlx4_rss_deinit(priv);
366 }
367
368 /**
369  * DPDK callback to close the device.
370  *
371  * Destroy all queues and objects, free memory.
372  *
373  * @param dev
374  *   Pointer to Ethernet device structure.
375  */
376 static void
377 mlx4_dev_close(struct rte_eth_dev *dev)
378 {
379         struct mlx4_priv *priv = dev->data->dev_private;
380         unsigned int i;
381
382         DEBUG("%p: closing device \"%s\"",
383               (void *)dev,
384               ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
385         dev->rx_pkt_burst = mlx4_rx_burst_removed;
386         dev->tx_pkt_burst = mlx4_tx_burst_removed;
387         rte_wmb();
388         /* Disable datapath on secondary process. */
389         mlx4_mp_req_stop_rxtx(dev);
390         mlx4_flow_clean(priv);
391         mlx4_rss_deinit(priv);
392         for (i = 0; i != dev->data->nb_rx_queues; ++i)
393                 mlx4_rx_queue_release(dev->data->rx_queues[i]);
394         for (i = 0; i != dev->data->nb_tx_queues; ++i)
395                 mlx4_tx_queue_release(dev->data->tx_queues[i]);
396         mlx4_proc_priv_uninit(dev);
397         mlx4_mr_release(dev);
398         if (priv->pd != NULL) {
399                 MLX4_ASSERT(priv->ctx != NULL);
400                 claim_zero(mlx4_glue->dealloc_pd(priv->pd));
401                 claim_zero(mlx4_glue->close_device(priv->ctx));
402         } else
403                 MLX4_ASSERT(priv->ctx == NULL);
404         mlx4_intr_uninstall(priv);
405         memset(priv, 0, sizeof(*priv));
406 }
407
408 static const struct eth_dev_ops mlx4_dev_ops = {
409         .dev_configure = mlx4_dev_configure,
410         .dev_start = mlx4_dev_start,
411         .dev_stop = mlx4_dev_stop,
412         .dev_set_link_down = mlx4_dev_set_link_down,
413         .dev_set_link_up = mlx4_dev_set_link_up,
414         .dev_close = mlx4_dev_close,
415         .link_update = mlx4_link_update,
416         .promiscuous_enable = mlx4_promiscuous_enable,
417         .promiscuous_disable = mlx4_promiscuous_disable,
418         .allmulticast_enable = mlx4_allmulticast_enable,
419         .allmulticast_disable = mlx4_allmulticast_disable,
420         .mac_addr_remove = mlx4_mac_addr_remove,
421         .mac_addr_add = mlx4_mac_addr_add,
422         .mac_addr_set = mlx4_mac_addr_set,
423         .set_mc_addr_list = mlx4_set_mc_addr_list,
424         .stats_get = mlx4_stats_get,
425         .stats_reset = mlx4_stats_reset,
426         .fw_version_get = mlx4_fw_version_get,
427         .dev_infos_get = mlx4_dev_infos_get,
428         .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
429         .vlan_filter_set = mlx4_vlan_filter_set,
430         .rx_queue_setup = mlx4_rx_queue_setup,
431         .tx_queue_setup = mlx4_tx_queue_setup,
432         .rx_queue_release = mlx4_rx_queue_release,
433         .tx_queue_release = mlx4_tx_queue_release,
434         .flow_ctrl_get = mlx4_flow_ctrl_get,
435         .flow_ctrl_set = mlx4_flow_ctrl_set,
436         .mtu_set = mlx4_mtu_set,
437         .filter_ctrl = mlx4_filter_ctrl,
438         .rx_queue_intr_enable = mlx4_rx_intr_enable,
439         .rx_queue_intr_disable = mlx4_rx_intr_disable,
440         .is_removed = mlx4_is_removed,
441 };
442
443 /* Available operations from secondary process. */
444 static const struct eth_dev_ops mlx4_dev_sec_ops = {
445         .stats_get = mlx4_stats_get,
446         .stats_reset = mlx4_stats_reset,
447         .fw_version_get = mlx4_fw_version_get,
448         .dev_infos_get = mlx4_dev_infos_get,
449 };
450
451 /**
452  * Get PCI information from struct ibv_device.
453  *
454  * @param device
455  *   Pointer to Ethernet device structure.
456  * @param[out] pci_addr
457  *   PCI bus address output buffer.
458  *
459  * @return
460  *   0 on success, negative errno value otherwise and rte_errno is set.
461  */
462 static int
463 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
464                             struct rte_pci_addr *pci_addr)
465 {
466         FILE *file;
467         char line[32];
468         MKSTR(path, "%s/device/uevent", device->ibdev_path);
469
470         file = fopen(path, "rb");
471         if (file == NULL) {
472                 rte_errno = errno;
473                 return -rte_errno;
474         }
475         while (fgets(line, sizeof(line), file) == line) {
476                 size_t len = strlen(line);
477                 int ret;
478
479                 /* Truncate long lines. */
480                 if (len == (sizeof(line) - 1))
481                         while (line[(len - 1)] != '\n') {
482                                 ret = fgetc(file);
483                                 if (ret == EOF)
484                                         break;
485                                 line[(len - 1)] = ret;
486                         }
487                 /* Extract information. */
488                 if (sscanf(line,
489                            "PCI_SLOT_NAME="
490                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
491                            &pci_addr->domain,
492                            &pci_addr->bus,
493                            &pci_addr->devid,
494                            &pci_addr->function) == 4) {
495                         ret = 0;
496                         break;
497                 }
498         }
499         fclose(file);
500         return 0;
501 }
502
503 /**
504  * Verify and store value for device argument.
505  *
506  * @param[in] key
507  *   Key argument to verify.
508  * @param[in] val
509  *   Value associated with key.
510  * @param[in, out] conf
511  *   Shared configuration data.
512  *
513  * @return
514  *   0 on success, negative errno value otherwise and rte_errno is set.
515  */
516 static int
517 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
518 {
519         unsigned long tmp;
520
521         errno = 0;
522         tmp = strtoul(val, NULL, 0);
523         if (errno) {
524                 rte_errno = errno;
525                 WARN("%s: \"%s\" is not a valid integer", key, val);
526                 return -rte_errno;
527         }
528         if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
529                 uint32_t ports = rte_log2_u32(conf->ports.present + 1);
530
531                 if (tmp >= ports) {
532                         ERROR("port index %lu outside range [0,%" PRIu32 ")",
533                               tmp, ports);
534                         return -EINVAL;
535                 }
536                 if (!(conf->ports.present & (1 << tmp))) {
537                         rte_errno = EINVAL;
538                         ERROR("invalid port index %lu", tmp);
539                         return -rte_errno;
540                 }
541                 conf->ports.enabled |= 1 << tmp;
542         } else if (strcmp(MLX4_MR_EXT_MEMSEG_EN_KVARG, key) == 0) {
543                 conf->mr_ext_memseg_en = !!tmp;
544         } else {
545                 rte_errno = EINVAL;
546                 WARN("%s: unknown parameter", key);
547                 return -rte_errno;
548         }
549         return 0;
550 }
551
552 /**
553  * Parse device parameters.
554  *
555  * @param devargs
556  *   Device arguments structure.
557  *
558  * @return
559  *   0 on success, negative errno value otherwise and rte_errno is set.
560  */
561 static int
562 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
563 {
564         struct rte_kvargs *kvlist;
565         unsigned int arg_count;
566         int ret = 0;
567         int i;
568
569         if (devargs == NULL)
570                 return 0;
571         kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
572         if (kvlist == NULL) {
573                 rte_errno = EINVAL;
574                 ERROR("failed to parse kvargs");
575                 return -rte_errno;
576         }
577         /* Process parameters. */
578         for (i = 0; pmd_mlx4_init_params[i]; ++i) {
579                 arg_count = rte_kvargs_count(kvlist, pmd_mlx4_init_params[i]);
580                 while (arg_count-- > 0) {
581                         ret = rte_kvargs_process(kvlist,
582                                                  pmd_mlx4_init_params[i],
583                                                  (int (*)(const char *,
584                                                           const char *,
585                                                           void *))
586                                                  mlx4_arg_parse,
587                                                  conf);
588                         if (ret != 0)
589                                 goto free_kvlist;
590                 }
591         }
592 free_kvlist:
593         rte_kvargs_free(kvlist);
594         return ret;
595 }
596
597 /**
598  * Interpret RSS capabilities reported by device.
599  *
600  * This function returns the set of usable Verbs RSS hash fields, kernel
601  * quirks taken into account.
602  *
603  * @param ctx
604  *   Verbs context.
605  * @param pd
606  *   Verbs protection domain.
607  * @param device_attr_ex
608  *   Extended device attributes to interpret.
609  *
610  * @return
611  *   Usable RSS hash fields mask in Verbs format.
612  */
613 static uint64_t
614 mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd,
615                 struct ibv_device_attr_ex *device_attr_ex)
616 {
617         uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask;
618         struct ibv_cq *cq = NULL;
619         struct ibv_wq *wq = NULL;
620         struct ibv_rwq_ind_table *ind = NULL;
621         struct ibv_qp *qp = NULL;
622
623         if (!hw_rss_sup) {
624                 WARN("no RSS capabilities reported; disabling support for UDP"
625                      " RSS and inner VXLAN RSS");
626                 return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
627                         IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
628                         IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
629         }
630         if (!(hw_rss_sup & IBV_RX_HASH_INNER))
631                 return hw_rss_sup;
632         /*
633          * Although reported as supported, missing code in some Linux
634          * versions (v4.15, v4.16) prevents the creation of hash QPs with
635          * inner capability.
636          *
637          * There is no choice but to attempt to instantiate a temporary RSS
638          * context in order to confirm its support.
639          */
640         cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0);
641         wq = cq ? mlx4_glue->create_wq
642                 (ctx,
643                  &(struct ibv_wq_init_attr){
644                         .wq_type = IBV_WQT_RQ,
645                         .max_wr = 1,
646                         .max_sge = 1,
647                         .pd = pd,
648                         .cq = cq,
649                  }) : NULL;
650         ind = wq ? mlx4_glue->create_rwq_ind_table
651                 (ctx,
652                  &(struct ibv_rwq_ind_table_init_attr){
653                         .log_ind_tbl_size = 0,
654                         .ind_tbl = &wq,
655                         .comp_mask = 0,
656                  }) : NULL;
657         qp = ind ? mlx4_glue->create_qp_ex
658                 (ctx,
659                  &(struct ibv_qp_init_attr_ex){
660                         .comp_mask =
661                                 (IBV_QP_INIT_ATTR_PD |
662                                  IBV_QP_INIT_ATTR_RX_HASH |
663                                  IBV_QP_INIT_ATTR_IND_TABLE),
664                         .qp_type = IBV_QPT_RAW_PACKET,
665                         .pd = pd,
666                         .rwq_ind_tbl = ind,
667                         .rx_hash_conf = {
668                                 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
669                                 .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
670                                 .rx_hash_key = mlx4_rss_hash_key_default,
671                                 .rx_hash_fields_mask = hw_rss_sup,
672                         },
673                  }) : NULL;
674         if (!qp) {
675                 WARN("disabling unusable inner RSS capability due to kernel"
676                      " quirk");
677                 hw_rss_sup &= ~IBV_RX_HASH_INNER;
678         } else {
679                 claim_zero(mlx4_glue->destroy_qp(qp));
680         }
681         if (ind)
682                 claim_zero(mlx4_glue->destroy_rwq_ind_table(ind));
683         if (wq)
684                 claim_zero(mlx4_glue->destroy_wq(wq));
685         if (cq)
686                 claim_zero(mlx4_glue->destroy_cq(cq));
687         return hw_rss_sup;
688 }
689
690 static struct rte_pci_driver mlx4_driver;
691
692 /**
693  * PMD global initialization.
694  *
695  * Independent from individual device, this function initializes global
696  * per-PMD data structures distinguishing primary and secondary processes.
697  * Hence, each initialization is called once per a process.
698  *
699  * @return
700  *   0 on success, a negative errno value otherwise and rte_errno is set.
701  */
702 static int
703 mlx4_init_once(void)
704 {
705         struct mlx4_shared_data *sd;
706         struct mlx4_local_data *ld = &mlx4_local_data;
707         int ret = 0;
708
709         if (mlx4_init_shared_data())
710                 return -rte_errno;
711         sd = mlx4_shared_data;
712         MLX4_ASSERT(sd);
713         rte_spinlock_lock(&sd->lock);
714         switch (rte_eal_process_type()) {
715         case RTE_PROC_PRIMARY:
716                 if (sd->init_done)
717                         break;
718                 LIST_INIT(&sd->mem_event_cb_list);
719                 rte_rwlock_init(&sd->mem_event_rwlock);
720                 rte_mem_event_callback_register("MLX4_MEM_EVENT_CB",
721                                                 mlx4_mr_mem_event_cb, NULL);
722                 ret = mlx4_mp_init_primary();
723                 if (ret)
724                         goto out;
725                 sd->init_done = 1;
726                 break;
727         case RTE_PROC_SECONDARY:
728                 if (ld->init_done)
729                         break;
730                 ret = mlx4_mp_init_secondary();
731                 if (ret)
732                         goto out;
733                 ++sd->secondary_cnt;
734                 ld->init_done = 1;
735                 break;
736         default:
737                 break;
738         }
739 out:
740         rte_spinlock_unlock(&sd->lock);
741         return ret;
742 }
743
744 /**
745  * DPDK callback to register a PCI device.
746  *
747  * This function creates an Ethernet device for each port of a given
748  * PCI device.
749  *
750  * @param[in] pci_drv
751  *   PCI driver structure (mlx4_driver).
752  * @param[in] pci_dev
753  *   PCI device information.
754  *
755  * @return
756  *   0 on success, negative errno value otherwise and rte_errno is set.
757  */
758 static int
759 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
760 {
761         struct ibv_device **list;
762         struct ibv_device *ibv_dev;
763         int err = 0;
764         struct ibv_context *attr_ctx = NULL;
765         struct ibv_device_attr device_attr;
766         struct ibv_device_attr_ex device_attr_ex;
767         struct mlx4_conf conf = {
768                 .ports.present = 0,
769                 .mr_ext_memseg_en = 1,
770         };
771         unsigned int vf;
772         int i;
773         char ifname[IF_NAMESIZE];
774
775         (void)pci_drv;
776         err = mlx4_init_once();
777         if (err) {
778                 ERROR("unable to init PMD global data: %s",
779                       strerror(rte_errno));
780                 return -rte_errno;
781         }
782         MLX4_ASSERT(pci_drv == &mlx4_driver);
783         list = mlx4_glue->get_device_list(&i);
784         if (list == NULL) {
785                 rte_errno = errno;
786                 MLX4_ASSERT(rte_errno);
787                 if (rte_errno == ENOSYS)
788                         ERROR("cannot list devices, is ib_uverbs loaded?");
789                 return -rte_errno;
790         }
791         MLX4_ASSERT(i >= 0);
792         /*
793          * For each listed device, check related sysfs entry against
794          * the provided PCI ID.
795          */
796         while (i != 0) {
797                 struct rte_pci_addr pci_addr;
798
799                 --i;
800                 DEBUG("checking device \"%s\"", list[i]->name);
801                 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
802                         continue;
803                 if ((pci_dev->addr.domain != pci_addr.domain) ||
804                     (pci_dev->addr.bus != pci_addr.bus) ||
805                     (pci_dev->addr.devid != pci_addr.devid) ||
806                     (pci_dev->addr.function != pci_addr.function))
807                         continue;
808                 vf = (pci_dev->id.device_id ==
809                       PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
810                 INFO("PCI information matches, using device \"%s\" (VF: %s)",
811                      list[i]->name, (vf ? "true" : "false"));
812                 attr_ctx = mlx4_glue->open_device(list[i]);
813                 err = errno;
814                 break;
815         }
816         if (attr_ctx == NULL) {
817                 mlx4_glue->free_device_list(list);
818                 switch (err) {
819                 case 0:
820                         rte_errno = ENODEV;
821                         ERROR("cannot access device, is mlx4_ib loaded?");
822                         return -rte_errno;
823                 case EINVAL:
824                         rte_errno = EINVAL;
825                         ERROR("cannot use device, are drivers up to date?");
826                         return -rte_errno;
827                 }
828                 MLX4_ASSERT(err > 0);
829                 rte_errno = err;
830                 return -rte_errno;
831         }
832         ibv_dev = list[i];
833         DEBUG("device opened");
834         if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
835                 err = ENODEV;
836                 goto error;
837         }
838         INFO("%u port(s) detected", device_attr.phys_port_cnt);
839         conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
840         if (mlx4_args(pci_dev->device.devargs, &conf)) {
841                 ERROR("failed to process device arguments");
842                 err = EINVAL;
843                 goto error;
844         }
845         /* Use all ports when none are defined */
846         if (!conf.ports.enabled)
847                 conf.ports.enabled = conf.ports.present;
848         /* Retrieve extended device attributes. */
849         if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
850                 err = ENODEV;
851                 goto error;
852         }
853         MLX4_ASSERT(device_attr.max_sge >= MLX4_MAX_SGE);
854         for (i = 0; i < device_attr.phys_port_cnt; i++) {
855                 uint32_t port = i + 1; /* ports are indexed from one */
856                 struct ibv_context *ctx = NULL;
857                 struct ibv_port_attr port_attr;
858                 struct ibv_pd *pd = NULL;
859                 struct mlx4_priv *priv = NULL;
860                 struct rte_eth_dev *eth_dev = NULL;
861                 struct rte_ether_addr mac;
862                 char name[RTE_ETH_NAME_MAX_LEN];
863
864                 /* If port is not enabled, skip. */
865                 if (!(conf.ports.enabled & (1 << i)))
866                         continue;
867                 DEBUG("using port %u", port);
868                 ctx = mlx4_glue->open_device(ibv_dev);
869                 if (ctx == NULL) {
870                         err = ENODEV;
871                         goto port_error;
872                 }
873                 snprintf(name, sizeof(name), "%s port %u",
874                          mlx4_glue->get_device_name(ibv_dev), port);
875                 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
876                         eth_dev = rte_eth_dev_attach_secondary(name);
877                         if (eth_dev == NULL) {
878                                 ERROR("can not attach rte ethdev");
879                                 rte_errno = ENOMEM;
880                                 err = rte_errno;
881                                 goto error;
882                         }
883                         priv = eth_dev->data->dev_private;
884                         if (!priv->verbs_alloc_ctx.enabled) {
885                                 ERROR("secondary process is not supported"
886                                       " due to lack of external allocator"
887                                       " from Verbs");
888                                 rte_errno = ENOTSUP;
889                                 err = rte_errno;
890                                 goto error;
891                         }
892                         eth_dev->device = &pci_dev->device;
893                         eth_dev->dev_ops = &mlx4_dev_sec_ops;
894                         err = mlx4_proc_priv_init(eth_dev);
895                         if (err)
896                                 goto error;
897                         /* Receive command fd from primary process. */
898                         err = mlx4_mp_req_verbs_cmd_fd(eth_dev);
899                         if (err < 0) {
900                                 err = rte_errno;
901                                 goto error;
902                         }
903                         /* Remap UAR for Tx queues. */
904                         err = mlx4_tx_uar_init_secondary(eth_dev, err);
905                         if (err) {
906                                 err = rte_errno;
907                                 goto error;
908                         }
909                         /*
910                          * Ethdev pointer is still required as input since
911                          * the primary device is not accessible from the
912                          * secondary process.
913                          */
914                         eth_dev->tx_pkt_burst = mlx4_tx_burst;
915                         eth_dev->rx_pkt_burst = mlx4_rx_burst;
916                         claim_zero(mlx4_glue->close_device(ctx));
917                         rte_eth_copy_pci_info(eth_dev, pci_dev);
918                         rte_eth_dev_probing_finish(eth_dev);
919                         continue;
920                 }
921                 /* Check port status. */
922                 err = mlx4_glue->query_port(ctx, port, &port_attr);
923                 if (err) {
924                         err = ENODEV;
925                         ERROR("port query failed: %s", strerror(err));
926                         goto port_error;
927                 }
928                 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
929                         err = ENOTSUP;
930                         ERROR("port %d is not configured in Ethernet mode",
931                               port);
932                         goto port_error;
933                 }
934                 if (port_attr.state != IBV_PORT_ACTIVE)
935                         DEBUG("port %d is not active: \"%s\" (%d)",
936                               port, mlx4_glue->port_state_str(port_attr.state),
937                               port_attr.state);
938                 /* Make asynchronous FD non-blocking to handle interrupts. */
939                 err = mlx4_fd_set_non_blocking(ctx->async_fd);
940                 if (err) {
941                         ERROR("cannot make asynchronous FD non-blocking: %s",
942                               strerror(err));
943                         goto port_error;
944                 }
945                 /* Allocate protection domain. */
946                 pd = mlx4_glue->alloc_pd(ctx);
947                 if (pd == NULL) {
948                         err = ENOMEM;
949                         ERROR("PD allocation failure");
950                         goto port_error;
951                 }
952                 /* from rte_ethdev.c */
953                 priv = rte_zmalloc("ethdev private structure",
954                                    sizeof(*priv),
955                                    RTE_CACHE_LINE_SIZE);
956                 if (priv == NULL) {
957                         err = ENOMEM;
958                         ERROR("priv allocation failure");
959                         goto port_error;
960                 }
961                 priv->ctx = ctx;
962                 priv->device_attr = device_attr;
963                 priv->port = port;
964                 priv->pd = pd;
965                 priv->mtu = RTE_ETHER_MTU;
966                 priv->vf = vf;
967                 priv->hw_csum = !!(device_attr.device_cap_flags &
968                                    IBV_DEVICE_RAW_IP_CSUM);
969                 DEBUG("checksum offloading is %ssupported",
970                       (priv->hw_csum ? "" : "not "));
971                 /* Only ConnectX-3 Pro supports tunneling. */
972                 priv->hw_csum_l2tun =
973                         priv->hw_csum &&
974                         (device_attr.vendor_part_id ==
975                          PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
976                 DEBUG("L2 tunnel checksum offloads are %ssupported",
977                       priv->hw_csum_l2tun ? "" : "not ");
978                 priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd,
979                                                    &device_attr_ex);
980                 DEBUG("supported RSS hash fields mask: %016" PRIx64,
981                       priv->hw_rss_sup);
982                 priv->hw_rss_max_qps =
983                         device_attr_ex.rss_caps.max_rwq_indirection_table_size;
984                 DEBUG("MAX RSS queues %d", priv->hw_rss_max_qps);
985                 priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
986                                         IBV_RAW_PACKET_CAP_SCATTER_FCS);
987                 DEBUG("FCS stripping toggling is %ssupported",
988                       priv->hw_fcs_strip ? "" : "not ");
989                 priv->tso =
990                         ((device_attr_ex.tso_caps.max_tso > 0) &&
991                          (device_attr_ex.tso_caps.supported_qpts &
992                           (1 << IBV_QPT_RAW_PACKET)));
993                 if (priv->tso)
994                         priv->tso_max_payload_sz =
995                                         device_attr_ex.tso_caps.max_tso;
996                 DEBUG("TSO is %ssupported",
997                       priv->tso ? "" : "not ");
998                 priv->mr_ext_memseg_en = conf.mr_ext_memseg_en;
999                 /* Configure the first MAC address by default. */
1000                 err = mlx4_get_mac(priv, &mac.addr_bytes);
1001                 if (err) {
1002                         ERROR("cannot get MAC address, is mlx4_en loaded?"
1003                               " (error: %s)", strerror(err));
1004                         goto port_error;
1005                 }
1006                 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1007                      priv->port,
1008                      mac.addr_bytes[0], mac.addr_bytes[1],
1009                      mac.addr_bytes[2], mac.addr_bytes[3],
1010                      mac.addr_bytes[4], mac.addr_bytes[5]);
1011                 /* Register MAC address. */
1012                 priv->mac[0] = mac;
1013
1014                 if (mlx4_get_ifname(priv, &ifname) == 0) {
1015                         DEBUG("port %u ifname is \"%s\"",
1016                               priv->port, ifname);
1017                         priv->if_index = if_nametoindex(ifname);
1018                 } else {
1019                         DEBUG("port %u ifname is unknown", priv->port);
1020                 }
1021
1022                 /* Get actual MTU if possible. */
1023                 mlx4_mtu_get(priv, &priv->mtu);
1024                 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
1025                 eth_dev = rte_eth_dev_allocate(name);
1026                 if (eth_dev == NULL) {
1027                         err = ENOMEM;
1028                         ERROR("can not allocate rte ethdev");
1029                         goto port_error;
1030                 }
1031                 eth_dev->data->dev_private = priv;
1032                 eth_dev->data->mac_addrs = priv->mac;
1033                 eth_dev->device = &pci_dev->device;
1034                 rte_eth_copy_pci_info(eth_dev, pci_dev);
1035                 /* Initialize local interrupt handle for current port. */
1036                 priv->intr_handle = (struct rte_intr_handle){
1037                         .fd = -1,
1038                         .type = RTE_INTR_HANDLE_EXT,
1039                 };
1040                 /*
1041                  * Override ethdev interrupt handle pointer with private
1042                  * handle instead of that of the parent PCI device used by
1043                  * default. This prevents it from being shared between all
1044                  * ports of the same PCI device since each of them is
1045                  * associated its own Verbs context.
1046                  *
1047                  * Rx interrupts in particular require this as the PMD has
1048                  * no control over the registration of queue interrupts
1049                  * besides setting up eth_dev->intr_handle, the rest is
1050                  * handled by rte_intr_rx_ctl().
1051                  */
1052                 eth_dev->intr_handle = &priv->intr_handle;
1053                 priv->dev_data = eth_dev->data;
1054                 eth_dev->dev_ops = &mlx4_dev_ops;
1055 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
1056                 /* Hint libmlx4 to use PMD allocator for data plane resources */
1057                 struct mlx4dv_ctx_allocators alctr = {
1058                         .alloc = &mlx4_alloc_verbs_buf,
1059                         .free = &mlx4_free_verbs_buf,
1060                         .data = priv,
1061                 };
1062                 err = mlx4_glue->dv_set_context_attr
1063                         (ctx, MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS,
1064                          (void *)((uintptr_t)&alctr));
1065                 if (err)
1066                         WARN("Verbs external allocator is not supported");
1067                 else
1068                         priv->verbs_alloc_ctx.enabled = 1;
1069 #endif
1070                 /* Bring Ethernet device up. */
1071                 DEBUG("forcing Ethernet interface up");
1072                 mlx4_dev_set_link_up(eth_dev);
1073                 /* Update link status once if waiting for LSC. */
1074                 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
1075                         mlx4_link_update(eth_dev, 0);
1076                 /*
1077                  * Once the device is added to the list of memory event
1078                  * callback, its global MR cache table cannot be expanded
1079                  * on the fly because of deadlock. If it overflows, lookup
1080                  * should be done by searching MR list linearly, which is slow.
1081                  */
1082                 err = mlx4_mr_btree_init(&priv->mr.cache,
1083                                          MLX4_MR_BTREE_CACHE_N * 2,
1084                                          eth_dev->device->numa_node);
1085                 if (err) {
1086                         /* rte_errno is already set. */
1087                         goto port_error;
1088                 }
1089                 /* Add device to memory callback list. */
1090                 rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock);
1091                 LIST_INSERT_HEAD(&mlx4_shared_data->mem_event_cb_list,
1092                                  priv, mem_event_cb);
1093                 rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock);
1094                 rte_eth_dev_probing_finish(eth_dev);
1095                 continue;
1096 port_error:
1097                 rte_free(priv);
1098                 if (eth_dev != NULL)
1099                         eth_dev->data->dev_private = NULL;
1100                 if (pd)
1101                         claim_zero(mlx4_glue->dealloc_pd(pd));
1102                 if (ctx)
1103                         claim_zero(mlx4_glue->close_device(ctx));
1104                 if (eth_dev != NULL) {
1105                         /* mac_addrs must not be freed because part of dev_private */
1106                         eth_dev->data->mac_addrs = NULL;
1107                         rte_eth_dev_release_port(eth_dev);
1108                 }
1109                 break;
1110         }
1111         /*
1112          * XXX if something went wrong in the loop above, there is a resource
1113          * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1114          * long as the dpdk does not provide a way to deallocate a ethdev and a
1115          * way to enumerate the registered ethdevs to free the previous ones.
1116          */
1117 error:
1118         if (attr_ctx)
1119                 claim_zero(mlx4_glue->close_device(attr_ctx));
1120         if (list)
1121                 mlx4_glue->free_device_list(list);
1122         if (err)
1123                 rte_errno = err;
1124         return -err;
1125 }
1126
1127 static const struct rte_pci_id mlx4_pci_id_map[] = {
1128         {
1129                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1130                                PCI_DEVICE_ID_MELLANOX_CONNECTX3)
1131         },
1132         {
1133                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1134                                PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
1135         },
1136         {
1137                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1138                                PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
1139         },
1140         {
1141                 .vendor_id = 0
1142         }
1143 };
1144
1145 static struct rte_pci_driver mlx4_driver = {
1146         .driver = {
1147                 .name = MLX4_DRIVER_NAME
1148         },
1149         .id_table = mlx4_pci_id_map,
1150         .probe = mlx4_pci_probe,
1151         .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1152 };
1153
1154 #ifdef RTE_IBVERBS_LINK_DLOPEN
1155
1156 /**
1157  * Suffix RTE_EAL_PMD_PATH with "-glue".
1158  *
1159  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1160  * suffixing its last component.
1161  *
1162  * @param buf[out]
1163  *   Output buffer, should be large enough otherwise NULL is returned.
1164  * @param size
1165  *   Size of @p out.
1166  *
1167  * @return
1168  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1169  */
1170 static char *
1171 mlx4_glue_path(char *buf, size_t size)
1172 {
1173         static const char *const bad[] = { "/", ".", "..", NULL };
1174         const char *path = RTE_EAL_PMD_PATH;
1175         size_t len = strlen(path);
1176         size_t off;
1177         int i;
1178
1179         while (len && path[len - 1] == '/')
1180                 --len;
1181         for (off = len; off && path[off - 1] != '/'; --off)
1182                 ;
1183         for (i = 0; bad[i]; ++i)
1184                 if (!strncmp(path + off, bad[i], (int)(len - off)))
1185                         goto error;
1186         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1187         if (i == -1 || (size_t)i >= size)
1188                 goto error;
1189         return buf;
1190 error:
1191         ERROR("unable to append \"-glue\" to last component of"
1192               " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1193               " please re-configure DPDK");
1194         return NULL;
1195 }
1196
1197 /**
1198  * Initialization routine for run-time dependency on rdma-core.
1199  */
1200 static int
1201 mlx4_glue_init(void)
1202 {
1203         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1204         const char *path[] = {
1205                 /*
1206                  * A basic security check is necessary before trusting
1207                  * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1208                  */
1209                 (geteuid() == getuid() && getegid() == getgid() ?
1210                  getenv("MLX4_GLUE_PATH") : NULL),
1211                 /*
1212                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1213                  * variant, otherwise let dlopen() look up libraries on its
1214                  * own.
1215                  */
1216                 (*RTE_EAL_PMD_PATH ?
1217                  mlx4_glue_path(glue_path, sizeof(glue_path)) : ""),
1218         };
1219         unsigned int i = 0;
1220         void *handle = NULL;
1221         void **sym;
1222         const char *dlmsg;
1223
1224         while (!handle && i != RTE_DIM(path)) {
1225                 const char *end;
1226                 size_t len;
1227                 int ret;
1228
1229                 if (!path[i]) {
1230                         ++i;
1231                         continue;
1232                 }
1233                 end = strpbrk(path[i], ":;");
1234                 if (!end)
1235                         end = path[i] + strlen(path[i]);
1236                 len = end - path[i];
1237                 ret = 0;
1238                 do {
1239                         char name[ret + 1];
1240
1241                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE,
1242                                        (int)len, path[i],
1243                                        (!len || *(end - 1) == '/') ? "" : "/");
1244                         if (ret == -1)
1245                                 break;
1246                         if (sizeof(name) != (size_t)ret + 1)
1247                                 continue;
1248                         DEBUG("looking for rdma-core glue as \"%s\"", name);
1249                         handle = dlopen(name, RTLD_LAZY);
1250                         break;
1251                 } while (1);
1252                 path[i] = end + 1;
1253                 if (!*end)
1254                         ++i;
1255         }
1256         if (!handle) {
1257                 rte_errno = EINVAL;
1258                 dlmsg = dlerror();
1259                 if (dlmsg)
1260                         WARN("cannot load glue library: %s", dlmsg);
1261                 goto glue_error;
1262         }
1263         sym = dlsym(handle, "mlx4_glue");
1264         if (!sym || !*sym) {
1265                 rte_errno = EINVAL;
1266                 dlmsg = dlerror();
1267                 if (dlmsg)
1268                         ERROR("cannot resolve glue symbol: %s", dlmsg);
1269                 goto glue_error;
1270         }
1271         mlx4_glue = *sym;
1272         return 0;
1273 glue_error:
1274         if (handle)
1275                 dlclose(handle);
1276         WARN("cannot initialize PMD due to missing run-time"
1277              " dependency on rdma-core libraries (libibverbs,"
1278              " libmlx4)");
1279         return -rte_errno;
1280 }
1281
1282 #endif
1283
1284 /**
1285  * Driver initialization routine.
1286  */
1287 RTE_INIT(rte_mlx4_pmd_init)
1288 {
1289         /* Initialize driver log type. */
1290         mlx4_logtype = rte_log_register("pmd.net.mlx4");
1291         if (mlx4_logtype >= 0)
1292                 rte_log_set_level(mlx4_logtype, RTE_LOG_NOTICE);
1293
1294         /*
1295          * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
1296          * want to get success errno value in case of calling them
1297          * when the device was removed.
1298          */
1299         setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
1300         /*
1301          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1302          * huge pages. Calling ibv_fork_init() during init allows
1303          * applications to use fork() safely for purposes other than
1304          * using this PMD, which is not supported in forked processes.
1305          */
1306         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1307 #ifdef RTE_IBVERBS_LINK_DLOPEN
1308         if (mlx4_glue_init())
1309                 return;
1310         MLX4_ASSERT(mlx4_glue);
1311 #endif
1312 #ifdef RTE_LIBRTE_MLX4_DEBUG
1313         /* Glue structure must not contain any NULL pointers. */
1314         {
1315                 unsigned int i;
1316
1317                 for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i)
1318                         MLX4_ASSERT(((const void *const *)mlx4_glue)[i]);
1319         }
1320 #endif
1321         if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) {
1322                 ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required",
1323                       mlx4_glue->version, MLX4_GLUE_VERSION);
1324                 return;
1325         }
1326         mlx4_glue->fork_init();
1327         rte_pci_register(&mlx4_driver);
1328 }
1329
1330 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
1331 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
1332 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
1333         "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");