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 <ethdev_driver.h>
37 #include <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 /** Configuration structure for device arguments. */
68 struct mlx4_conf {
69 struct {
70 uint32_t present; /**< Bit-field for existing ports. */
71 uint32_t enabled; /**< Bit-field for user-enabled ports. */
72 } ports;
73 int mr_ext_memseg_en;
74 /** Whether memseg should be extended for MR creation. */
75 };
76
77 /* Available parameters list. */
78 const char *pmd_mlx4_init_params[] = {
79 MLX4_PMD_PORT_KVARG,
80 MLX4_MR_EXT_MEMSEG_EN_KVARG,
81 NULL,
82 };
83
84 static int mlx4_dev_stop(struct rte_eth_dev *dev);
85
86 /**
87 * Initialize shared data between primary and secondary process.
88 *
89 * A memzone is reserved by primary process and secondary processes attach to
90 * the memzone.
91 *
92 * @return
93 * 0 on success, a negative errno value otherwise and rte_errno is set.
94 */
95 static int
mlx4_init_shared_data(void)96 mlx4_init_shared_data(void)
97 {
98 const struct rte_memzone *mz;
99 int ret = 0;
100
101 rte_spinlock_lock(&mlx4_shared_data_lock);
102 if (mlx4_shared_data == NULL) {
103 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
104 /* Allocate shared memory. */
105 mz = rte_memzone_reserve(MZ_MLX4_PMD_SHARED_DATA,
106 sizeof(*mlx4_shared_data),
107 SOCKET_ID_ANY, 0);
108 if (mz == NULL) {
109 ERROR("Cannot allocate mlx4 shared data\n");
110 ret = -rte_errno;
111 goto error;
112 }
113 mlx4_shared_data = mz->addr;
114 memset(mlx4_shared_data, 0, sizeof(*mlx4_shared_data));
115 rte_spinlock_init(&mlx4_shared_data->lock);
116 } else {
117 /* Lookup allocated shared memory. */
118 mz = rte_memzone_lookup(MZ_MLX4_PMD_SHARED_DATA);
119 if (mz == NULL) {
120 ERROR("Cannot attach mlx4 shared data\n");
121 ret = -rte_errno;
122 goto error;
123 }
124 mlx4_shared_data = mz->addr;
125 memset(&mlx4_local_data, 0, sizeof(mlx4_local_data));
126 }
127 }
128 error:
129 rte_spinlock_unlock(&mlx4_shared_data_lock);
130 return ret;
131 }
132
133 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
134 /**
135 * Verbs callback to allocate a memory. This function should allocate the space
136 * according to the size provided residing inside a huge page.
137 * Please note that all allocation must respect the alignment from libmlx4
138 * (i.e. currently sysconf(_SC_PAGESIZE)).
139 *
140 * @param[in] size
141 * The size in bytes of the memory to allocate.
142 * @param[in] data
143 * A pointer to the callback data.
144 *
145 * @return
146 * Allocated buffer, NULL otherwise and rte_errno is set.
147 */
148 static void *
mlx4_alloc_verbs_buf(size_t size,void * data)149 mlx4_alloc_verbs_buf(size_t size, void *data)
150 {
151 struct mlx4_priv *priv = data;
152 void *ret;
153 size_t alignment = sysconf(_SC_PAGESIZE);
154 unsigned int socket = SOCKET_ID_ANY;
155
156 if (priv->verbs_alloc_ctx.type == MLX4_VERBS_ALLOC_TYPE_TX_QUEUE) {
157 const struct txq *txq = priv->verbs_alloc_ctx.obj;
158
159 socket = txq->socket;
160 } else if (priv->verbs_alloc_ctx.type ==
161 MLX4_VERBS_ALLOC_TYPE_RX_QUEUE) {
162 const struct rxq *rxq = priv->verbs_alloc_ctx.obj;
163
164 socket = rxq->socket;
165 }
166 MLX4_ASSERT(data != NULL);
167 ret = rte_malloc_socket(__func__, size, alignment, socket);
168 if (!ret && size)
169 rte_errno = ENOMEM;
170 return ret;
171 }
172
173 /**
174 * Verbs callback to free a memory.
175 *
176 * @param[in] ptr
177 * A pointer to the memory to free.
178 * @param[in] data
179 * A pointer to the callback data.
180 */
181 static void
mlx4_free_verbs_buf(void * ptr,void * data __rte_unused)182 mlx4_free_verbs_buf(void *ptr, void *data __rte_unused)
183 {
184 MLX4_ASSERT(data != NULL);
185 rte_free(ptr);
186 }
187 #endif
188
189 /**
190 * Initialize process private data structure.
191 *
192 * @param dev
193 * Pointer to Ethernet device structure.
194 *
195 * @return
196 * 0 on success, a negative errno value otherwise and rte_errno is set.
197 */
198 int
mlx4_proc_priv_init(struct rte_eth_dev * dev)199 mlx4_proc_priv_init(struct rte_eth_dev *dev)
200 {
201 struct mlx4_proc_priv *ppriv;
202 size_t ppriv_size;
203
204 mlx4_proc_priv_uninit(dev);
205 /*
206 * UAR register table follows the process private structure. BlueFlame
207 * registers for Tx queues are stored in the table.
208 */
209 ppriv_size = sizeof(struct mlx4_proc_priv) +
210 dev->data->nb_tx_queues * sizeof(void *);
211 ppriv = rte_zmalloc_socket("mlx4_proc_priv", ppriv_size,
212 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
213 if (!ppriv) {
214 rte_errno = ENOMEM;
215 return -rte_errno;
216 }
217 ppriv->uar_table_sz = dev->data->nb_tx_queues;
218 dev->process_private = ppriv;
219 return 0;
220 }
221
222 /**
223 * Un-initialize process private data structure.
224 *
225 * @param dev
226 * Pointer to Ethernet device structure.
227 */
228 void
mlx4_proc_priv_uninit(struct rte_eth_dev * dev)229 mlx4_proc_priv_uninit(struct rte_eth_dev *dev)
230 {
231 if (!dev->process_private)
232 return;
233 rte_free(dev->process_private);
234 dev->process_private = NULL;
235 }
236
237 /**
238 * DPDK callback for Ethernet device configuration.
239 *
240 * @param dev
241 * Pointer to Ethernet device structure.
242 *
243 * @return
244 * 0 on success, negative errno value otherwise and rte_errno is set.
245 */
246 static int
mlx4_dev_configure(struct rte_eth_dev * dev)247 mlx4_dev_configure(struct rte_eth_dev *dev)
248 {
249 struct mlx4_priv *priv = dev->data->dev_private;
250 struct rte_flow_error error;
251 int ret;
252
253 /* Prepare internal flow rules. */
254 ret = mlx4_flow_sync(priv, &error);
255 if (ret) {
256 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
257 " flow error type %d, cause %p, message: %s",
258 -ret, strerror(-ret), error.type, error.cause,
259 error.message ? error.message : "(unspecified)");
260 goto exit;
261 }
262 ret = mlx4_intr_install(priv);
263 if (ret) {
264 ERROR("%p: interrupt handler installation failed",
265 (void *)dev);
266 goto exit;
267 }
268 ret = mlx4_proc_priv_init(dev);
269 if (ret) {
270 ERROR("%p: process private data allocation failed",
271 (void *)dev);
272 goto exit;
273 }
274 exit:
275 return ret;
276 }
277
278 /**
279 * DPDK callback to start the device.
280 *
281 * Simulate device start by initializing common RSS resources and attaching
282 * all configured flows.
283 *
284 * @param dev
285 * Pointer to Ethernet device structure.
286 *
287 * @return
288 * 0 on success, negative errno value otherwise and rte_errno is set.
289 */
290 static int
mlx4_dev_start(struct rte_eth_dev * dev)291 mlx4_dev_start(struct rte_eth_dev *dev)
292 {
293 struct mlx4_priv *priv = dev->data->dev_private;
294 struct rte_flow_error error;
295 int ret;
296
297 if (priv->started)
298 return 0;
299 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
300 priv->started = 1;
301 ret = mlx4_rss_init(priv);
302 if (ret) {
303 ERROR("%p: cannot initialize RSS resources: %s",
304 (void *)dev, strerror(-ret));
305 goto err;
306 }
307 #ifdef RTE_LIBRTE_MLX4_DEBUG
308 mlx4_mr_dump_dev(dev);
309 #endif
310 ret = mlx4_rxq_intr_enable(priv);
311 if (ret) {
312 ERROR("%p: interrupt handler installation failed",
313 (void *)dev);
314 goto err;
315 }
316 ret = mlx4_flow_sync(priv, &error);
317 if (ret) {
318 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
319 " flow error type %d, cause %p, message: %s",
320 (void *)dev,
321 -ret, strerror(-ret), error.type, error.cause,
322 error.message ? error.message : "(unspecified)");
323 goto err;
324 }
325 rte_wmb();
326 dev->tx_pkt_burst = mlx4_tx_burst;
327 dev->rx_pkt_burst = mlx4_rx_burst;
328 /* Enable datapath on secondary process. */
329 mlx4_mp_req_start_rxtx(dev);
330 return 0;
331 err:
332 mlx4_dev_stop(dev);
333 return ret;
334 }
335
336 /**
337 * DPDK callback to stop the device.
338 *
339 * Simulate device stop by detaching all configured flows.
340 *
341 * @param dev
342 * Pointer to Ethernet device structure.
343 */
344 static int
mlx4_dev_stop(struct rte_eth_dev * dev)345 mlx4_dev_stop(struct rte_eth_dev *dev)
346 {
347 struct mlx4_priv *priv = dev->data->dev_private;
348
349 if (!priv->started)
350 return 0;
351 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
352 priv->started = 0;
353 dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
354 dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
355 rte_wmb();
356 /* Disable datapath on secondary process. */
357 mlx4_mp_req_stop_rxtx(dev);
358 mlx4_flow_sync(priv, NULL);
359 mlx4_rxq_intr_disable(priv);
360 mlx4_rss_deinit(priv);
361
362 return 0;
363 }
364
365 /**
366 * DPDK callback to close the device.
367 *
368 * Destroy all queues and objects, free memory.
369 *
370 * @param dev
371 * Pointer to Ethernet device structure.
372 */
373 static int
mlx4_dev_close(struct rte_eth_dev * dev)374 mlx4_dev_close(struct rte_eth_dev *dev)
375 {
376 struct mlx4_priv *priv = dev->data->dev_private;
377 unsigned int i;
378
379 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
380 rte_eth_dev_release_port(dev);
381 return 0;
382 }
383 DEBUG("%p: closing device \"%s\"",
384 (void *)dev,
385 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
386 dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
387 dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
388 rte_wmb();
389 /* Disable datapath on secondary process. */
390 mlx4_mp_req_stop_rxtx(dev);
391 mlx4_flow_clean(priv);
392 mlx4_rss_deinit(priv);
393 for (i = 0; i != dev->data->nb_rx_queues; ++i)
394 mlx4_rx_queue_release(dev, i);
395 for (i = 0; i != dev->data->nb_tx_queues; ++i)
396 mlx4_tx_queue_release(dev, i);
397 mlx4_proc_priv_uninit(dev);
398 mlx4_mr_release(dev);
399 if (priv->pd != NULL) {
400 MLX4_ASSERT(priv->ctx != NULL);
401 claim_zero(mlx4_glue->dealloc_pd(priv->pd));
402 claim_zero(mlx4_glue->close_device(priv->ctx));
403 } else
404 MLX4_ASSERT(priv->ctx == NULL);
405 mlx4_intr_uninstall(priv);
406 memset(priv, 0, sizeof(*priv));
407 /* mac_addrs must not be freed because part of dev_private */
408 dev->data->mac_addrs = NULL;
409 return 0;
410 }
411
412 static const struct eth_dev_ops mlx4_dev_ops = {
413 .dev_configure = mlx4_dev_configure,
414 .dev_start = mlx4_dev_start,
415 .dev_stop = mlx4_dev_stop,
416 .dev_set_link_down = mlx4_dev_set_link_down,
417 .dev_set_link_up = mlx4_dev_set_link_up,
418 .dev_close = mlx4_dev_close,
419 .link_update = mlx4_link_update,
420 .promiscuous_enable = mlx4_promiscuous_enable,
421 .promiscuous_disable = mlx4_promiscuous_disable,
422 .allmulticast_enable = mlx4_allmulticast_enable,
423 .allmulticast_disable = mlx4_allmulticast_disable,
424 .mac_addr_remove = mlx4_mac_addr_remove,
425 .mac_addr_add = mlx4_mac_addr_add,
426 .mac_addr_set = mlx4_mac_addr_set,
427 .set_mc_addr_list = mlx4_set_mc_addr_list,
428 .stats_get = mlx4_stats_get,
429 .stats_reset = mlx4_stats_reset,
430 .fw_version_get = mlx4_fw_version_get,
431 .dev_infos_get = mlx4_dev_infos_get,
432 .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
433 .vlan_filter_set = mlx4_vlan_filter_set,
434 .rx_queue_setup = mlx4_rx_queue_setup,
435 .tx_queue_setup = mlx4_tx_queue_setup,
436 .rx_queue_release = mlx4_rx_queue_release,
437 .tx_queue_release = mlx4_tx_queue_release,
438 .flow_ctrl_get = mlx4_flow_ctrl_get,
439 .flow_ctrl_set = mlx4_flow_ctrl_set,
440 .mtu_set = mlx4_mtu_set,
441 .flow_ops_get = mlx4_flow_ops_get,
442 .rx_queue_intr_enable = mlx4_rx_intr_enable,
443 .rx_queue_intr_disable = mlx4_rx_intr_disable,
444 .is_removed = mlx4_is_removed,
445 };
446
447 /* Available operations from secondary process. */
448 static const struct eth_dev_ops mlx4_dev_sec_ops = {
449 .stats_get = mlx4_stats_get,
450 .stats_reset = mlx4_stats_reset,
451 .fw_version_get = mlx4_fw_version_get,
452 .dev_infos_get = mlx4_dev_infos_get,
453 };
454
455 /**
456 * Get PCI information from struct ibv_device.
457 *
458 * @param device
459 * Pointer to Ethernet device structure.
460 * @param[out] pci_addr
461 * PCI bus address output buffer.
462 *
463 * @return
464 * 0 on success, negative errno value otherwise and rte_errno is set.
465 */
466 static int
mlx4_ibv_device_to_pci_addr(const struct ibv_device * device,struct rte_pci_addr * pci_addr)467 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
468 struct rte_pci_addr *pci_addr)
469 {
470 FILE *file;
471 char line[32];
472 MKSTR(path, "%s/device/uevent", device->ibdev_path);
473
474 file = fopen(path, "rb");
475 if (file == NULL) {
476 rte_errno = errno;
477 return -rte_errno;
478 }
479 while (fgets(line, sizeof(line), file) == line) {
480 size_t len = strlen(line);
481 int ret;
482
483 /* Truncate long lines. */
484 if (len == (sizeof(line) - 1))
485 while (line[(len - 1)] != '\n') {
486 ret = fgetc(file);
487 if (ret == EOF)
488 break;
489 line[(len - 1)] = ret;
490 }
491 /* Extract information. */
492 if (sscanf(line,
493 "PCI_SLOT_NAME="
494 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
495 &pci_addr->domain,
496 &pci_addr->bus,
497 &pci_addr->devid,
498 &pci_addr->function) == 4) {
499 break;
500 }
501 }
502 fclose(file);
503 return 0;
504 }
505
506 /**
507 * Verify and store value for device argument.
508 *
509 * @param[in] key
510 * Key argument to verify.
511 * @param[in] val
512 * Value associated with key.
513 * @param[in, out] conf
514 * Shared configuration data.
515 *
516 * @return
517 * 0 on success, negative errno value otherwise and rte_errno is set.
518 */
519 static int
mlx4_arg_parse(const char * key,const char * val,struct mlx4_conf * conf)520 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
521 {
522 unsigned long tmp;
523
524 errno = 0;
525 tmp = strtoul(val, NULL, 0);
526 if (errno) {
527 rte_errno = errno;
528 WARN("%s: \"%s\" is not a valid integer", key, val);
529 return -rte_errno;
530 }
531 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
532 uint32_t ports = rte_log2_u32(conf->ports.present + 1);
533
534 if (tmp >= ports) {
535 ERROR("port index %lu outside range [0,%" PRIu32 ")",
536 tmp, ports);
537 return -EINVAL;
538 }
539 if (!(conf->ports.present & (1 << tmp))) {
540 rte_errno = EINVAL;
541 ERROR("invalid port index %lu", tmp);
542 return -rte_errno;
543 }
544 conf->ports.enabled |= 1 << tmp;
545 } else if (strcmp(MLX4_MR_EXT_MEMSEG_EN_KVARG, key) == 0) {
546 conf->mr_ext_memseg_en = !!tmp;
547 } else {
548 rte_errno = EINVAL;
549 WARN("%s: unknown parameter", key);
550 return -rte_errno;
551 }
552 return 0;
553 }
554
555 /**
556 * Parse device parameters.
557 *
558 * @param devargs
559 * Device arguments structure.
560 *
561 * @return
562 * 0 on success, negative errno value otherwise and rte_errno is set.
563 */
564 static int
mlx4_args(struct rte_devargs * devargs,struct mlx4_conf * conf)565 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
566 {
567 struct rte_kvargs *kvlist;
568 unsigned int arg_count;
569 int ret = 0;
570 int i;
571
572 if (devargs == NULL)
573 return 0;
574 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
575 if (kvlist == NULL) {
576 rte_errno = EINVAL;
577 ERROR("failed to parse kvargs");
578 return -rte_errno;
579 }
580 /* Process parameters. */
581 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
582 arg_count = rte_kvargs_count(kvlist, pmd_mlx4_init_params[i]);
583 while (arg_count-- > 0) {
584 ret = rte_kvargs_process(kvlist,
585 pmd_mlx4_init_params[i],
586 (int (*)(const char *,
587 const char *,
588 void *))
589 mlx4_arg_parse,
590 conf);
591 if (ret != 0)
592 goto free_kvlist;
593 }
594 }
595 free_kvlist:
596 rte_kvargs_free(kvlist);
597 return ret;
598 }
599
600 /**
601 * Interpret RSS capabilities reported by device.
602 *
603 * This function returns the set of usable Verbs RSS hash fields, kernel
604 * quirks taken into account.
605 *
606 * @param ctx
607 * Verbs context.
608 * @param pd
609 * Verbs protection domain.
610 * @param device_attr_ex
611 * Extended device attributes to interpret.
612 *
613 * @return
614 * Usable RSS hash fields mask in Verbs format.
615 */
616 static uint64_t
mlx4_hw_rss_sup(struct ibv_context * ctx,struct ibv_pd * pd,struct ibv_device_attr_ex * device_attr_ex)617 mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd,
618 struct ibv_device_attr_ex *device_attr_ex)
619 {
620 uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask;
621 struct ibv_cq *cq = NULL;
622 struct ibv_wq *wq = NULL;
623 struct ibv_rwq_ind_table *ind = NULL;
624 struct ibv_qp *qp = NULL;
625
626 if (!hw_rss_sup) {
627 WARN("no RSS capabilities reported; disabling support for UDP"
628 " RSS and inner VXLAN RSS");
629 return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
630 IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
631 IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
632 }
633 if (!(hw_rss_sup & IBV_RX_HASH_INNER))
634 return hw_rss_sup;
635 /*
636 * Although reported as supported, missing code in some Linux
637 * versions (v4.15, v4.16) prevents the creation of hash QPs with
638 * inner capability.
639 *
640 * There is no choice but to attempt to instantiate a temporary RSS
641 * context in order to confirm its support.
642 */
643 cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0);
644 wq = cq ? mlx4_glue->create_wq
645 (ctx,
646 &(struct ibv_wq_init_attr){
647 .wq_type = IBV_WQT_RQ,
648 .max_wr = 1,
649 .max_sge = 1,
650 .pd = pd,
651 .cq = cq,
652 }) : NULL;
653 ind = wq ? mlx4_glue->create_rwq_ind_table
654 (ctx,
655 &(struct ibv_rwq_ind_table_init_attr){
656 .log_ind_tbl_size = 0,
657 .ind_tbl = &wq,
658 .comp_mask = 0,
659 }) : NULL;
660 qp = ind ? mlx4_glue->create_qp_ex
661 (ctx,
662 &(struct ibv_qp_init_attr_ex){
663 .comp_mask =
664 (IBV_QP_INIT_ATTR_PD |
665 IBV_QP_INIT_ATTR_RX_HASH |
666 IBV_QP_INIT_ATTR_IND_TABLE),
667 .qp_type = IBV_QPT_RAW_PACKET,
668 .pd = pd,
669 .rwq_ind_tbl = ind,
670 .rx_hash_conf = {
671 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
672 .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
673 .rx_hash_key = mlx4_rss_hash_key_default,
674 .rx_hash_fields_mask = hw_rss_sup,
675 },
676 }) : NULL;
677 if (!qp) {
678 WARN("disabling unusable inner RSS capability due to kernel"
679 " quirk");
680 hw_rss_sup &= ~IBV_RX_HASH_INNER;
681 } else {
682 claim_zero(mlx4_glue->destroy_qp(qp));
683 }
684 if (ind)
685 claim_zero(mlx4_glue->destroy_rwq_ind_table(ind));
686 if (wq)
687 claim_zero(mlx4_glue->destroy_wq(wq));
688 if (cq)
689 claim_zero(mlx4_glue->destroy_cq(cq));
690 return hw_rss_sup;
691 }
692
693 static struct rte_pci_driver mlx4_driver;
694
695 /**
696 * PMD global initialization.
697 *
698 * Independent from individual device, this function initializes global
699 * per-PMD data structures distinguishing primary and secondary processes.
700 * Hence, each initialization is called once per a process.
701 *
702 * @return
703 * 0 on success, a negative errno value otherwise and rte_errno is set.
704 */
705 static int
mlx4_init_once(void)706 mlx4_init_once(void)
707 {
708 struct mlx4_shared_data *sd;
709 struct mlx4_local_data *ld = &mlx4_local_data;
710 int ret = 0;
711
712 if (mlx4_init_shared_data())
713 return -rte_errno;
714 sd = mlx4_shared_data;
715 MLX4_ASSERT(sd);
716 rte_spinlock_lock(&sd->lock);
717 switch (rte_eal_process_type()) {
718 case RTE_PROC_PRIMARY:
719 if (sd->init_done)
720 break;
721 LIST_INIT(&sd->mem_event_cb_list);
722 rte_rwlock_init(&sd->mem_event_rwlock);
723 rte_mem_event_callback_register("MLX4_MEM_EVENT_CB",
724 mlx4_mr_mem_event_cb, NULL);
725 ret = mlx4_mp_init_primary();
726 if (ret)
727 goto out;
728 sd->init_done = 1;
729 break;
730 case RTE_PROC_SECONDARY:
731 if (ld->init_done)
732 break;
733 ret = mlx4_mp_init_secondary();
734 if (ret)
735 goto out;
736 ++sd->secondary_cnt;
737 ld->init_done = 1;
738 break;
739 default:
740 break;
741 }
742 out:
743 rte_spinlock_unlock(&sd->lock);
744 return ret;
745 }
746
747 /**
748 * DPDK callback to register a PCI device.
749 *
750 * This function creates an Ethernet device for each port of a given
751 * PCI device.
752 *
753 * @param[in] pci_drv
754 * PCI driver structure (mlx4_driver).
755 * @param[in] pci_dev
756 * PCI device information.
757 *
758 * @return
759 * 0 on success, negative errno value otherwise and rte_errno is set.
760 */
761 static int
mlx4_pci_probe(struct rte_pci_driver * pci_drv,struct rte_pci_device * pci_dev)762 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
763 {
764 struct ibv_device **list;
765 struct ibv_device *ibv_dev;
766 int err = 0;
767 struct ibv_context *attr_ctx = NULL;
768 struct ibv_device_attr device_attr;
769 struct ibv_device_attr_ex device_attr_ex;
770 struct rte_eth_dev *prev_dev = NULL;
771 struct mlx4_conf conf = {
772 .ports.present = 0,
773 .mr_ext_memseg_en = 1,
774 };
775 unsigned int vf;
776 int i;
777 char ifname[IF_NAMESIZE];
778
779 (void)pci_drv;
780 err = mlx4_init_once();
781 if (err) {
782 ERROR("unable to init PMD global data: %s",
783 strerror(rte_errno));
784 return -rte_errno;
785 }
786 MLX4_ASSERT(pci_drv == &mlx4_driver);
787 list = mlx4_glue->get_device_list(&i);
788 if (list == NULL) {
789 rte_errno = errno;
790 MLX4_ASSERT(rte_errno);
791 if (rte_errno == ENOSYS)
792 ERROR("cannot list devices, is ib_uverbs loaded?");
793 return -rte_errno;
794 }
795 MLX4_ASSERT(i >= 0);
796 /*
797 * For each listed device, check related sysfs entry against
798 * the provided PCI ID.
799 */
800 while (i != 0) {
801 struct rte_pci_addr pci_addr;
802
803 --i;
804 DEBUG("checking device \"%s\"", list[i]->name);
805 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
806 continue;
807 if ((pci_dev->addr.domain != pci_addr.domain) ||
808 (pci_dev->addr.bus != pci_addr.bus) ||
809 (pci_dev->addr.devid != pci_addr.devid) ||
810 (pci_dev->addr.function != pci_addr.function))
811 continue;
812 vf = (pci_dev->id.device_id ==
813 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
814 INFO("PCI information matches, using device \"%s\" (VF: %s)",
815 list[i]->name, (vf ? "true" : "false"));
816 attr_ctx = mlx4_glue->open_device(list[i]);
817 err = errno;
818 break;
819 }
820 if (attr_ctx == NULL) {
821 mlx4_glue->free_device_list(list);
822 switch (err) {
823 case 0:
824 rte_errno = ENODEV;
825 ERROR("cannot access device, is mlx4_ib loaded?");
826 return -rte_errno;
827 case EINVAL:
828 rte_errno = EINVAL;
829 ERROR("cannot use device, are drivers up to date?");
830 return -rte_errno;
831 }
832 MLX4_ASSERT(err > 0);
833 rte_errno = err;
834 return -rte_errno;
835 }
836 ibv_dev = list[i];
837 DEBUG("device opened");
838 if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
839 err = ENODEV;
840 goto error;
841 }
842 INFO("%u port(s) detected", device_attr.phys_port_cnt);
843 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
844 if (mlx4_args(pci_dev->device.devargs, &conf)) {
845 ERROR("failed to process device arguments");
846 err = EINVAL;
847 goto error;
848 }
849 /* Use all ports when none are defined */
850 if (!conf.ports.enabled)
851 conf.ports.enabled = conf.ports.present;
852 /* Retrieve extended device attributes. */
853 if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
854 err = ENODEV;
855 goto error;
856 }
857 MLX4_ASSERT(device_attr.max_sge >= MLX4_MAX_SGE);
858 for (i = 0; i < device_attr.phys_port_cnt; i++) {
859 uint32_t port = i + 1; /* ports are indexed from one */
860 struct ibv_context *ctx = NULL;
861 struct ibv_port_attr port_attr;
862 struct ibv_pd *pd = NULL;
863 struct mlx4_priv *priv = NULL;
864 struct rte_eth_dev *eth_dev = NULL;
865 struct rte_ether_addr mac;
866 char name[RTE_ETH_NAME_MAX_LEN];
867
868 /* If port is not enabled, skip. */
869 if (!(conf.ports.enabled & (1 << i)))
870 continue;
871 DEBUG("using port %u", port);
872 ctx = mlx4_glue->open_device(ibv_dev);
873 if (ctx == NULL) {
874 err = ENODEV;
875 goto port_error;
876 }
877 snprintf(name, sizeof(name), "%s port %u",
878 mlx4_glue->get_device_name(ibv_dev), port);
879 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
880 eth_dev = rte_eth_dev_attach_secondary(name);
881 if (eth_dev == NULL) {
882 ERROR("can not attach rte ethdev");
883 rte_errno = ENOMEM;
884 err = rte_errno;
885 goto err_secondary;
886 }
887 priv = eth_dev->data->dev_private;
888 if (!priv->verbs_alloc_ctx.enabled) {
889 ERROR("secondary process is not supported"
890 " due to lack of external allocator"
891 " from Verbs");
892 rte_errno = ENOTSUP;
893 err = rte_errno;
894 goto err_secondary;
895 }
896 eth_dev->device = &pci_dev->device;
897 eth_dev->dev_ops = &mlx4_dev_sec_ops;
898 err = mlx4_proc_priv_init(eth_dev);
899 if (err)
900 goto err_secondary;
901 /* Receive command fd from primary process. */
902 err = mlx4_mp_req_verbs_cmd_fd(eth_dev);
903 if (err < 0) {
904 err = rte_errno;
905 goto err_secondary;
906 }
907 /* Remap UAR for Tx queues. */
908 err = mlx4_tx_uar_init_secondary(eth_dev, err);
909 if (err) {
910 err = rte_errno;
911 goto err_secondary;
912 }
913 /*
914 * Ethdev pointer is still required as input since
915 * the primary device is not accessible from the
916 * secondary process.
917 */
918 eth_dev->tx_pkt_burst = mlx4_tx_burst;
919 eth_dev->rx_pkt_burst = mlx4_rx_burst;
920 claim_zero(mlx4_glue->close_device(ctx));
921 rte_eth_copy_pci_info(eth_dev, pci_dev);
922 rte_eth_dev_probing_finish(eth_dev);
923 prev_dev = eth_dev;
924 continue;
925 err_secondary:
926 claim_zero(mlx4_glue->close_device(ctx));
927 rte_eth_dev_release_port(eth_dev);
928 if (prev_dev)
929 rte_eth_dev_release_port(prev_dev);
930 break;
931 }
932 /* Check port status. */
933 err = mlx4_glue->query_port(ctx, port, &port_attr);
934 if (err) {
935 err = ENODEV;
936 ERROR("port query failed: %s", strerror(err));
937 goto port_error;
938 }
939 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
940 err = ENOTSUP;
941 ERROR("port %d is not configured in Ethernet mode",
942 port);
943 goto port_error;
944 }
945 if (port_attr.state != IBV_PORT_ACTIVE)
946 DEBUG("port %d is not active: \"%s\" (%d)",
947 port, mlx4_glue->port_state_str(port_attr.state),
948 port_attr.state);
949 /* Make asynchronous FD non-blocking to handle interrupts. */
950 err = mlx4_fd_set_non_blocking(ctx->async_fd);
951 if (err) {
952 ERROR("cannot make asynchronous FD non-blocking: %s",
953 strerror(err));
954 goto port_error;
955 }
956 /* Allocate protection domain. */
957 pd = mlx4_glue->alloc_pd(ctx);
958 if (pd == NULL) {
959 err = ENOMEM;
960 ERROR("PD allocation failure");
961 goto port_error;
962 }
963 /* from rte_ethdev.c */
964 priv = rte_zmalloc("ethdev private structure",
965 sizeof(*priv),
966 RTE_CACHE_LINE_SIZE);
967 if (priv == NULL) {
968 err = ENOMEM;
969 ERROR("priv allocation failure");
970 goto port_error;
971 }
972 priv->ctx = ctx;
973 priv->device_attr = device_attr;
974 priv->port = port;
975 priv->pd = pd;
976 priv->mtu = RTE_ETHER_MTU;
977 priv->vf = vf;
978 priv->hw_csum = !!(device_attr.device_cap_flags &
979 IBV_DEVICE_RAW_IP_CSUM);
980 DEBUG("checksum offloading is %ssupported",
981 (priv->hw_csum ? "" : "not "));
982 /* Only ConnectX-3 Pro supports tunneling. */
983 priv->hw_csum_l2tun =
984 priv->hw_csum &&
985 (device_attr.vendor_part_id ==
986 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
987 DEBUG("L2 tunnel checksum offloads are %ssupported",
988 priv->hw_csum_l2tun ? "" : "not ");
989 priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd,
990 &device_attr_ex);
991 DEBUG("supported RSS hash fields mask: %016" PRIx64,
992 priv->hw_rss_sup);
993 priv->hw_rss_max_qps =
994 device_attr_ex.rss_caps.max_rwq_indirection_table_size;
995 DEBUG("MAX RSS queues %d", priv->hw_rss_max_qps);
996 priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
997 IBV_RAW_PACKET_CAP_SCATTER_FCS);
998 DEBUG("FCS stripping toggling is %ssupported",
999 priv->hw_fcs_strip ? "" : "not ");
1000 priv->tso =
1001 ((device_attr_ex.tso_caps.max_tso > 0) &&
1002 (device_attr_ex.tso_caps.supported_qpts &
1003 (1 << IBV_QPT_RAW_PACKET)));
1004 if (priv->tso)
1005 priv->tso_max_payload_sz =
1006 device_attr_ex.tso_caps.max_tso;
1007 DEBUG("TSO is %ssupported",
1008 priv->tso ? "" : "not ");
1009 priv->mr_ext_memseg_en = conf.mr_ext_memseg_en;
1010 /* Configure the first MAC address by default. */
1011 err = mlx4_get_mac(priv, &mac.addr_bytes);
1012 if (err) {
1013 ERROR("cannot get MAC address, is mlx4_en loaded?"
1014 " (error: %s)", strerror(err));
1015 goto port_error;
1016 }
1017 INFO("port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1018 priv->port, RTE_ETHER_ADDR_BYTES(&mac));
1019 /* Register MAC address. */
1020 priv->mac[0] = mac;
1021
1022 if (mlx4_get_ifname(priv, &ifname) == 0) {
1023 DEBUG("port %u ifname is \"%s\"",
1024 priv->port, ifname);
1025 priv->if_index = if_nametoindex(ifname);
1026 } else {
1027 DEBUG("port %u ifname is unknown", priv->port);
1028 }
1029
1030 /* Get actual MTU if possible. */
1031 mlx4_mtu_get(priv, &priv->mtu);
1032 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
1033 eth_dev = rte_eth_dev_allocate(name);
1034 if (eth_dev == NULL) {
1035 err = ENOMEM;
1036 ERROR("can not allocate rte ethdev");
1037 goto port_error;
1038 }
1039 eth_dev->data->dev_private = priv;
1040 eth_dev->data->mac_addrs = priv->mac;
1041 eth_dev->device = &pci_dev->device;
1042 rte_eth_copy_pci_info(eth_dev, pci_dev);
1043 eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1044 /* Initialize local interrupt handle for current port. */
1045 priv->intr_handle =
1046 rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
1047 if (priv->intr_handle == NULL) {
1048 RTE_LOG(ERR, EAL, "Fail to allocate intr_handle\n");
1049 goto port_error;
1050 }
1051
1052 if (rte_intr_fd_set(priv->intr_handle, -1))
1053 goto port_error;
1054
1055 if (rte_intr_type_set(priv->intr_handle, RTE_INTR_HANDLE_EXT))
1056 goto port_error;
1057
1058 /*
1059 * Override ethdev interrupt handle pointer with private
1060 * handle instead of that of the parent PCI device used by
1061 * default. This prevents it from being shared between all
1062 * ports of the same PCI device since each of them is
1063 * associated its own Verbs context.
1064 *
1065 * Rx interrupts in particular require this as the PMD has
1066 * no control over the registration of queue interrupts
1067 * besides setting up eth_dev->intr_handle, the rest is
1068 * handled by rte_intr_rx_ctl().
1069 */
1070 eth_dev->intr_handle = priv->intr_handle;
1071 priv->dev_data = eth_dev->data;
1072 eth_dev->dev_ops = &mlx4_dev_ops;
1073 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
1074 /* Hint libmlx4 to use PMD allocator for data plane resources */
1075 err = mlx4_glue->dv_set_context_attr
1076 (ctx, MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS,
1077 (void *)((uintptr_t)&(struct mlx4dv_ctx_allocators){
1078 .alloc = &mlx4_alloc_verbs_buf,
1079 .free = &mlx4_free_verbs_buf,
1080 .data = priv,
1081 }));
1082 if (err)
1083 WARN("Verbs external allocator is not supported");
1084 else
1085 priv->verbs_alloc_ctx.enabled = 1;
1086 #endif
1087 /* Bring Ethernet device up. */
1088 DEBUG("forcing Ethernet interface up");
1089 mlx4_dev_set_link_up(eth_dev);
1090 /* Update link status once if waiting for LSC. */
1091 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
1092 mlx4_link_update(eth_dev, 0);
1093 /*
1094 * Once the device is added to the list of memory event
1095 * callback, its global MR cache table cannot be expanded
1096 * on the fly because of deadlock. If it overflows, lookup
1097 * should be done by searching MR list linearly, which is slow.
1098 */
1099 err = mlx4_mr_btree_init(&priv->mr.cache,
1100 MLX4_MR_BTREE_CACHE_N * 2,
1101 eth_dev->device->numa_node);
1102 if (err) {
1103 /* rte_errno is already set. */
1104 goto port_error;
1105 }
1106 /* Add device to memory callback list. */
1107 rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock);
1108 LIST_INSERT_HEAD(&mlx4_shared_data->mem_event_cb_list,
1109 priv, mem_event_cb);
1110 rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock);
1111 rte_eth_dev_probing_finish(eth_dev);
1112 prev_dev = eth_dev;
1113 continue;
1114 port_error:
1115 if (priv != NULL)
1116 rte_intr_instance_free(priv->intr_handle);
1117 rte_free(priv);
1118 if (eth_dev != NULL)
1119 eth_dev->data->dev_private = NULL;
1120 if (pd)
1121 claim_zero(mlx4_glue->dealloc_pd(pd));
1122 if (ctx)
1123 claim_zero(mlx4_glue->close_device(ctx));
1124 if (eth_dev != NULL) {
1125 /* mac_addrs must not be freed because part of dev_private */
1126 eth_dev->data->mac_addrs = NULL;
1127 rte_eth_dev_release_port(eth_dev);
1128 }
1129 if (prev_dev)
1130 mlx4_dev_close(prev_dev);
1131 break;
1132 }
1133 error:
1134 if (attr_ctx)
1135 claim_zero(mlx4_glue->close_device(attr_ctx));
1136 if (list)
1137 mlx4_glue->free_device_list(list);
1138 if (err)
1139 rte_errno = err;
1140 return -err;
1141 }
1142
1143 /**
1144 * DPDK callback to remove a PCI device.
1145 *
1146 * This function removes all Ethernet devices belong to a given PCI device.
1147 *
1148 * @param[in] pci_dev
1149 * Pointer to the PCI device.
1150 *
1151 * @return
1152 * 0 on success, the function cannot fail.
1153 */
1154 static int
mlx4_pci_remove(struct rte_pci_device * pci_dev)1155 mlx4_pci_remove(struct rte_pci_device *pci_dev)
1156 {
1157 uint16_t port_id;
1158 int ret = 0;
1159
1160 RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
1161 /*
1162 * mlx4_dev_close() is not registered to secondary process,
1163 * call the close function explicitly for secondary process.
1164 */
1165 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1166 ret |= mlx4_dev_close(&rte_eth_devices[port_id]);
1167 else
1168 ret |= rte_eth_dev_close(port_id);
1169 }
1170 return ret == 0 ? 0 : -EIO;
1171 }
1172
1173 static const struct rte_pci_id mlx4_pci_id_map[] = {
1174 {
1175 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1176 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
1177 },
1178 {
1179 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1180 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
1181 },
1182 {
1183 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1184 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
1185 },
1186 {
1187 .vendor_id = 0
1188 }
1189 };
1190
1191 static struct rte_pci_driver mlx4_driver = {
1192 .driver = {
1193 .name = MLX4_DRIVER_NAME
1194 },
1195 .id_table = mlx4_pci_id_map,
1196 .probe = mlx4_pci_probe,
1197 .remove = mlx4_pci_remove,
1198 .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1199 };
1200
1201 #ifdef RTE_IBVERBS_LINK_DLOPEN
1202
1203 /**
1204 * Suffix RTE_EAL_PMD_PATH with "-glue".
1205 *
1206 * This function performs a sanity check on RTE_EAL_PMD_PATH before
1207 * suffixing its last component.
1208 *
1209 * @param buf[out]
1210 * Output buffer, should be large enough otherwise NULL is returned.
1211 * @param size
1212 * Size of @p out.
1213 *
1214 * @return
1215 * Pointer to @p buf or @p NULL in case suffix cannot be appended.
1216 */
1217 static char *
mlx4_glue_path(char * buf,size_t size)1218 mlx4_glue_path(char *buf, size_t size)
1219 {
1220 static const char *const bad[] = { "/", ".", "..", NULL };
1221 const char *path = RTE_EAL_PMD_PATH;
1222 size_t len = strlen(path);
1223 size_t off;
1224 int i;
1225
1226 while (len && path[len - 1] == '/')
1227 --len;
1228 for (off = len; off && path[off - 1] != '/'; --off)
1229 ;
1230 for (i = 0; bad[i]; ++i)
1231 if (!strncmp(path + off, bad[i], (int)(len - off)))
1232 goto error;
1233 i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1234 if (i == -1 || (size_t)i >= size)
1235 goto error;
1236 return buf;
1237 error:
1238 ERROR("unable to append \"-glue\" to last component of"
1239 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1240 " please re-configure DPDK");
1241 return NULL;
1242 }
1243
1244 /**
1245 * Initialization routine for run-time dependency on rdma-core.
1246 */
1247 static int
mlx4_glue_init(void)1248 mlx4_glue_init(void)
1249 {
1250 char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1251 const char *path[] = {
1252 /*
1253 * A basic security check is necessary before trusting
1254 * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1255 */
1256 (geteuid() == getuid() && getegid() == getgid() ?
1257 getenv("MLX4_GLUE_PATH") : NULL),
1258 /*
1259 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1260 * variant, otherwise let dlopen() look up libraries on its
1261 * own.
1262 */
1263 (*RTE_EAL_PMD_PATH ?
1264 mlx4_glue_path(glue_path, sizeof(glue_path)) : ""),
1265 };
1266 unsigned int i = 0;
1267 void *handle = NULL;
1268 void **sym;
1269 const char *dlmsg;
1270
1271 while (!handle && i != RTE_DIM(path)) {
1272 const char *end;
1273 size_t len;
1274 int ret;
1275
1276 if (!path[i]) {
1277 ++i;
1278 continue;
1279 }
1280 end = strpbrk(path[i], ":;");
1281 if (!end)
1282 end = path[i] + strlen(path[i]);
1283 len = end - path[i];
1284 ret = 0;
1285 do {
1286 char name[ret + 1];
1287
1288 ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE,
1289 (int)len, path[i],
1290 (!len || *(end - 1) == '/') ? "" : "/");
1291 if (ret == -1)
1292 break;
1293 if (sizeof(name) != (size_t)ret + 1)
1294 continue;
1295 DEBUG("looking for rdma-core glue as \"%s\"", name);
1296 handle = dlopen(name, RTLD_LAZY);
1297 break;
1298 } while (1);
1299 path[i] = end + 1;
1300 if (!*end)
1301 ++i;
1302 }
1303 if (!handle) {
1304 rte_errno = EINVAL;
1305 dlmsg = dlerror();
1306 if (dlmsg)
1307 WARN("cannot load glue library: %s", dlmsg);
1308 goto glue_error;
1309 }
1310 sym = dlsym(handle, "mlx4_glue");
1311 if (!sym || !*sym) {
1312 rte_errno = EINVAL;
1313 dlmsg = dlerror();
1314 if (dlmsg)
1315 ERROR("cannot resolve glue symbol: %s", dlmsg);
1316 goto glue_error;
1317 }
1318 mlx4_glue = *sym;
1319 return 0;
1320 glue_error:
1321 if (handle)
1322 dlclose(handle);
1323 WARN("cannot initialize PMD due to missing run-time"
1324 " dependency on rdma-core libraries (libibverbs,"
1325 " libmlx4)");
1326 return -rte_errno;
1327 }
1328
1329 #endif
1330
1331 /* Initialize driver log type. */
RTE_LOG_REGISTER_DEFAULT(mlx4_logtype,NOTICE)1332 RTE_LOG_REGISTER_DEFAULT(mlx4_logtype, NOTICE)
1333
1334 /**
1335 * Driver initialization routine.
1336 */
1337 RTE_INIT(rte_mlx4_pmd_init)
1338 {
1339 /*
1340 * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
1341 * want to get success errno value in case of calling them
1342 * when the device was removed.
1343 */
1344 setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
1345 /*
1346 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1347 * huge pages. Calling ibv_fork_init() during init allows
1348 * applications to use fork() safely for purposes other than
1349 * using this PMD, which is not supported in forked processes.
1350 */
1351 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1352 #ifdef RTE_IBVERBS_LINK_DLOPEN
1353 if (mlx4_glue_init())
1354 return;
1355 MLX4_ASSERT(mlx4_glue);
1356 #endif
1357 #ifdef RTE_LIBRTE_MLX4_DEBUG
1358 /* Glue structure must not contain any NULL pointers. */
1359 {
1360 unsigned int i;
1361
1362 for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i)
1363 MLX4_ASSERT(((const void *const *)mlx4_glue)[i]);
1364 }
1365 #endif
1366 if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) {
1367 ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required",
1368 mlx4_glue->version, MLX4_GLUE_VERSION);
1369 return;
1370 }
1371 mlx4_glue->fork_init();
1372 rte_pci_register(&mlx4_driver);
1373 }
1374
1375 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
1376 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
1377 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
1378 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");
1379