1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2012 6WIND S.A.
3 * Copyright 2012 Mellanox Technologies, Ltd
8 * mlx4 driver initialization.
22 /* Verbs headers do not support -pedantic. */
24 #pragma GCC diagnostic ignored "-Wpedantic"
26 #include <infiniband/verbs.h>
28 #pragma GCC diagnostic error "-Wpedantic"
31 #include <rte_common.h>
32 #include <rte_config.h>
34 #include <rte_errno.h>
35 #include <rte_ethdev_driver.h>
36 #include <rte_ethdev_pci.h>
37 #include <rte_ether.h>
39 #include <rte_interrupts.h>
40 #include <rte_kvargs.h>
41 #include <rte_malloc.h>
45 #include "mlx4_glue.h"
46 #include "mlx4_flow.h"
48 #include "mlx4_rxtx.h"
49 #include "mlx4_utils.h"
51 struct mlx4_dev_list mlx4_mem_event_cb_list =
52 LIST_HEAD_INITIALIZER(mlx4_mem_event_cb_list);
54 rte_rwlock_t mlx4_mem_event_rwlock = RTE_RWLOCK_INITIALIZER;
56 /** Configuration structure for device arguments. */
59 uint32_t present; /**< Bit-field for existing ports. */
60 uint32_t enabled; /**< Bit-field for user-enabled ports. */
64 /* Available parameters list. */
65 const char *pmd_mlx4_init_params[] = {
70 static void mlx4_dev_stop(struct rte_eth_dev *dev);
73 * DPDK callback for Ethernet device configuration.
76 * Pointer to Ethernet device structure.
79 * 0 on success, negative errno value otherwise and rte_errno is set.
82 mlx4_dev_configure(struct rte_eth_dev *dev)
84 struct priv *priv = dev->data->dev_private;
85 struct rte_flow_error error;
88 /* Prepare internal flow rules. */
89 ret = mlx4_flow_sync(priv, &error);
91 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
92 " flow error type %d, cause %p, message: %s",
93 -ret, strerror(-ret), error.type, error.cause,
94 error.message ? error.message : "(unspecified)");
97 ret = mlx4_intr_install(priv);
99 ERROR("%p: interrupt handler installation failed",
106 * DPDK callback to start the device.
108 * Simulate device start by initializing common RSS resources and attaching
109 * all configured flows.
112 * Pointer to Ethernet device structure.
115 * 0 on success, negative errno value otherwise and rte_errno is set.
118 mlx4_dev_start(struct rte_eth_dev *dev)
120 struct priv *priv = dev->data->dev_private;
121 struct rte_flow_error error;
126 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
128 ret = mlx4_rss_init(priv);
130 ERROR("%p: cannot initialize RSS resources: %s",
131 (void *)dev, strerror(-ret));
135 mlx4_mr_dump_dev(dev);
137 ret = mlx4_rxq_intr_enable(priv);
139 ERROR("%p: interrupt handler installation failed",
143 ret = mlx4_flow_sync(priv, &error);
145 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
146 " flow error type %d, cause %p, message: %s",
148 -ret, strerror(-ret), error.type, error.cause,
149 error.message ? error.message : "(unspecified)");
153 dev->tx_pkt_burst = mlx4_tx_burst;
154 dev->rx_pkt_burst = mlx4_rx_burst;
162 * DPDK callback to stop the device.
164 * Simulate device stop by detaching all configured flows.
167 * Pointer to Ethernet device structure.
170 mlx4_dev_stop(struct rte_eth_dev *dev)
172 struct priv *priv = dev->data->dev_private;
176 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
178 dev->tx_pkt_burst = mlx4_tx_burst_removed;
179 dev->rx_pkt_burst = mlx4_rx_burst_removed;
181 mlx4_flow_sync(priv, NULL);
182 mlx4_rxq_intr_disable(priv);
183 mlx4_rss_deinit(priv);
187 * DPDK callback to close the device.
189 * Destroy all queues and objects, free memory.
192 * Pointer to Ethernet device structure.
195 mlx4_dev_close(struct rte_eth_dev *dev)
197 struct priv *priv = dev->data->dev_private;
200 DEBUG("%p: closing device \"%s\"",
202 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
203 dev->rx_pkt_burst = mlx4_rx_burst_removed;
204 dev->tx_pkt_burst = mlx4_tx_burst_removed;
206 mlx4_flow_clean(priv);
207 mlx4_rss_deinit(priv);
208 for (i = 0; i != dev->data->nb_rx_queues; ++i)
209 mlx4_rx_queue_release(dev->data->rx_queues[i]);
210 for (i = 0; i != dev->data->nb_tx_queues; ++i)
211 mlx4_tx_queue_release(dev->data->tx_queues[i]);
212 mlx4_mr_release(dev);
213 if (priv->pd != NULL) {
214 assert(priv->ctx != NULL);
215 claim_zero(mlx4_glue->dealloc_pd(priv->pd));
216 claim_zero(mlx4_glue->close_device(priv->ctx));
218 assert(priv->ctx == NULL);
219 mlx4_intr_uninstall(priv);
220 memset(priv, 0, sizeof(*priv));
223 static const struct eth_dev_ops mlx4_dev_ops = {
224 .dev_configure = mlx4_dev_configure,
225 .dev_start = mlx4_dev_start,
226 .dev_stop = mlx4_dev_stop,
227 .dev_set_link_down = mlx4_dev_set_link_down,
228 .dev_set_link_up = mlx4_dev_set_link_up,
229 .dev_close = mlx4_dev_close,
230 .link_update = mlx4_link_update,
231 .promiscuous_enable = mlx4_promiscuous_enable,
232 .promiscuous_disable = mlx4_promiscuous_disable,
233 .allmulticast_enable = mlx4_allmulticast_enable,
234 .allmulticast_disable = mlx4_allmulticast_disable,
235 .mac_addr_remove = mlx4_mac_addr_remove,
236 .mac_addr_add = mlx4_mac_addr_add,
237 .mac_addr_set = mlx4_mac_addr_set,
238 .stats_get = mlx4_stats_get,
239 .stats_reset = mlx4_stats_reset,
240 .dev_infos_get = mlx4_dev_infos_get,
241 .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
242 .vlan_filter_set = mlx4_vlan_filter_set,
243 .rx_queue_setup = mlx4_rx_queue_setup,
244 .tx_queue_setup = mlx4_tx_queue_setup,
245 .rx_queue_release = mlx4_rx_queue_release,
246 .tx_queue_release = mlx4_tx_queue_release,
247 .flow_ctrl_get = mlx4_flow_ctrl_get,
248 .flow_ctrl_set = mlx4_flow_ctrl_set,
249 .mtu_set = mlx4_mtu_set,
250 .filter_ctrl = mlx4_filter_ctrl,
251 .rx_queue_intr_enable = mlx4_rx_intr_enable,
252 .rx_queue_intr_disable = mlx4_rx_intr_disable,
253 .is_removed = mlx4_is_removed,
257 * Get PCI information from struct ibv_device.
260 * Pointer to Ethernet device structure.
261 * @param[out] pci_addr
262 * PCI bus address output buffer.
265 * 0 on success, negative errno value otherwise and rte_errno is set.
268 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
269 struct rte_pci_addr *pci_addr)
273 MKSTR(path, "%s/device/uevent", device->ibdev_path);
275 file = fopen(path, "rb");
280 while (fgets(line, sizeof(line), file) == line) {
281 size_t len = strlen(line);
284 /* Truncate long lines. */
285 if (len == (sizeof(line) - 1))
286 while (line[(len - 1)] != '\n') {
290 line[(len - 1)] = ret;
292 /* Extract information. */
295 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
299 &pci_addr->function) == 4) {
309 * Verify and store value for device argument.
312 * Key argument to verify.
314 * Value associated with key.
315 * @param[in, out] conf
316 * Shared configuration data.
319 * 0 on success, negative errno value otherwise and rte_errno is set.
322 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
327 tmp = strtoul(val, NULL, 0);
330 WARN("%s: \"%s\" is not a valid integer", key, val);
333 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
334 uint32_t ports = rte_log2_u32(conf->ports.present + 1);
337 ERROR("port index %lu outside range [0,%" PRIu32 ")",
341 if (!(conf->ports.present & (1 << tmp))) {
343 ERROR("invalid port index %lu", tmp);
346 conf->ports.enabled |= 1 << tmp;
349 WARN("%s: unknown parameter", key);
356 * Parse device parameters.
359 * Device arguments structure.
362 * 0 on success, negative errno value otherwise and rte_errno is set.
365 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
367 struct rte_kvargs *kvlist;
368 unsigned int arg_count;
374 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
375 if (kvlist == NULL) {
377 ERROR("failed to parse kvargs");
380 /* Process parameters. */
381 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
382 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
383 while (arg_count-- > 0) {
384 ret = rte_kvargs_process(kvlist,
386 (int (*)(const char *,
396 rte_kvargs_free(kvlist);
401 * Interpret RSS capabilities reported by device.
403 * This function returns the set of usable Verbs RSS hash fields, kernel
404 * quirks taken into account.
409 * Verbs protection domain.
410 * @param device_attr_ex
411 * Extended device attributes to interpret.
414 * Usable RSS hash fields mask in Verbs format.
417 mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd,
418 struct ibv_device_attr_ex *device_attr_ex)
420 uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask;
421 struct ibv_cq *cq = NULL;
422 struct ibv_wq *wq = NULL;
423 struct ibv_rwq_ind_table *ind = NULL;
424 struct ibv_qp *qp = NULL;
427 WARN("no RSS capabilities reported; disabling support for UDP"
428 " RSS and inner VXLAN RSS");
429 return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
430 IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
431 IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
433 if (!(hw_rss_sup & IBV_RX_HASH_INNER))
436 * Although reported as supported, missing code in some Linux
437 * versions (v4.15, v4.16) prevents the creation of hash QPs with
440 * There is no choice but to attempt to instantiate a temporary RSS
441 * context in order to confirm its support.
443 cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0);
444 wq = cq ? mlx4_glue->create_wq
446 &(struct ibv_wq_init_attr){
447 .wq_type = IBV_WQT_RQ,
453 ind = wq ? mlx4_glue->create_rwq_ind_table
455 &(struct ibv_rwq_ind_table_init_attr){
456 .log_ind_tbl_size = 0,
460 qp = ind ? mlx4_glue->create_qp_ex
462 &(struct ibv_qp_init_attr_ex){
464 (IBV_QP_INIT_ATTR_PD |
465 IBV_QP_INIT_ATTR_RX_HASH |
466 IBV_QP_INIT_ATTR_IND_TABLE),
467 .qp_type = IBV_QPT_RAW_PACKET,
471 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
472 .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
473 .rx_hash_key = mlx4_rss_hash_key_default,
474 .rx_hash_fields_mask = hw_rss_sup,
478 WARN("disabling unusable inner RSS capability due to kernel"
480 hw_rss_sup &= ~IBV_RX_HASH_INNER;
482 claim_zero(mlx4_glue->destroy_qp(qp));
485 claim_zero(mlx4_glue->destroy_rwq_ind_table(ind));
487 claim_zero(mlx4_glue->destroy_wq(wq));
489 claim_zero(mlx4_glue->destroy_cq(cq));
493 static struct rte_pci_driver mlx4_driver;
496 * DPDK callback to register a PCI device.
498 * This function creates an Ethernet device for each port of a given
502 * PCI driver structure (mlx4_driver).
504 * PCI device information.
507 * 0 on success, negative errno value otherwise and rte_errno is set.
510 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
512 struct ibv_device **list;
513 struct ibv_device *ibv_dev;
515 struct ibv_context *attr_ctx = NULL;
516 struct ibv_device_attr device_attr;
517 struct ibv_device_attr_ex device_attr_ex;
518 struct mlx4_conf conf = {
525 assert(pci_drv == &mlx4_driver);
526 list = mlx4_glue->get_device_list(&i);
530 if (rte_errno == ENOSYS)
531 ERROR("cannot list devices, is ib_uverbs loaded?");
536 * For each listed device, check related sysfs entry against
537 * the provided PCI ID.
540 struct rte_pci_addr pci_addr;
543 DEBUG("checking device \"%s\"", list[i]->name);
544 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
546 if ((pci_dev->addr.domain != pci_addr.domain) ||
547 (pci_dev->addr.bus != pci_addr.bus) ||
548 (pci_dev->addr.devid != pci_addr.devid) ||
549 (pci_dev->addr.function != pci_addr.function))
551 vf = (pci_dev->id.device_id ==
552 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
553 INFO("PCI information matches, using device \"%s\" (VF: %s)",
554 list[i]->name, (vf ? "true" : "false"));
555 attr_ctx = mlx4_glue->open_device(list[i]);
559 if (attr_ctx == NULL) {
560 mlx4_glue->free_device_list(list);
564 ERROR("cannot access device, is mlx4_ib loaded?");
568 ERROR("cannot use device, are drivers up to date?");
576 DEBUG("device opened");
577 if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
581 INFO("%u port(s) detected", device_attr.phys_port_cnt);
582 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
583 if (mlx4_args(pci_dev->device.devargs, &conf)) {
584 ERROR("failed to process device arguments");
588 /* Use all ports when none are defined */
589 if (!conf.ports.enabled)
590 conf.ports.enabled = conf.ports.present;
591 /* Retrieve extended device attributes. */
592 if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
596 assert(device_attr.max_sge >= MLX4_MAX_SGE);
597 for (i = 0; i < device_attr.phys_port_cnt; i++) {
598 uint32_t port = i + 1; /* ports are indexed from one */
599 struct ibv_context *ctx = NULL;
600 struct ibv_port_attr port_attr;
601 struct ibv_pd *pd = NULL;
602 struct priv *priv = NULL;
603 struct rte_eth_dev *eth_dev = NULL;
604 struct ether_addr mac;
606 /* If port is not enabled, skip. */
607 if (!(conf.ports.enabled & (1 << i)))
609 DEBUG("using port %u", port);
610 ctx = mlx4_glue->open_device(ibv_dev);
615 /* Check port status. */
616 err = mlx4_glue->query_port(ctx, port, &port_attr);
619 ERROR("port query failed: %s", strerror(rte_errno));
622 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
624 ERROR("port %d is not configured in Ethernet mode",
628 if (port_attr.state != IBV_PORT_ACTIVE)
629 DEBUG("port %d is not active: \"%s\" (%d)",
630 port, mlx4_glue->port_state_str(port_attr.state),
632 /* Make asynchronous FD non-blocking to handle interrupts. */
633 if (mlx4_fd_set_non_blocking(ctx->async_fd) < 0) {
634 ERROR("cannot make asynchronous FD non-blocking: %s",
635 strerror(rte_errno));
638 /* Allocate protection domain. */
639 pd = mlx4_glue->alloc_pd(ctx);
642 ERROR("PD allocation failure");
645 /* from rte_ethdev.c */
646 priv = rte_zmalloc("ethdev private structure",
648 RTE_CACHE_LINE_SIZE);
651 ERROR("priv allocation failure");
655 priv->device_attr = device_attr;
658 priv->mtu = ETHER_MTU;
660 priv->hw_csum = !!(device_attr.device_cap_flags &
661 IBV_DEVICE_RAW_IP_CSUM);
662 DEBUG("checksum offloading is %ssupported",
663 (priv->hw_csum ? "" : "not "));
664 /* Only ConnectX-3 Pro supports tunneling. */
665 priv->hw_csum_l2tun =
667 (device_attr.vendor_part_id ==
668 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
669 DEBUG("L2 tunnel checksum offloads are %ssupported",
670 priv->hw_csum_l2tun ? "" : "not ");
671 priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd,
673 DEBUG("supported RSS hash fields mask: %016" PRIx64,
675 priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
676 IBV_RAW_PACKET_CAP_SCATTER_FCS);
677 DEBUG("FCS stripping toggling is %ssupported",
678 priv->hw_fcs_strip ? "" : "not ");
679 /* Configure the first MAC address by default. */
680 if (mlx4_get_mac(priv, &mac.addr_bytes)) {
681 ERROR("cannot get MAC address, is mlx4_en loaded?"
682 " (rte_errno: %s)", strerror(rte_errno));
685 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
687 mac.addr_bytes[0], mac.addr_bytes[1],
688 mac.addr_bytes[2], mac.addr_bytes[3],
689 mac.addr_bytes[4], mac.addr_bytes[5]);
690 /* Register MAC address. */
694 char ifname[IF_NAMESIZE];
696 if (mlx4_get_ifname(priv, &ifname) == 0)
697 DEBUG("port %u ifname is \"%s\"",
700 DEBUG("port %u ifname is unknown", priv->port);
703 /* Get actual MTU if possible. */
704 mlx4_mtu_get(priv, &priv->mtu);
705 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
706 /* from rte_ethdev.c */
708 char name[RTE_ETH_NAME_MAX_LEN];
710 snprintf(name, sizeof(name), "%s port %u",
711 mlx4_glue->get_device_name(ibv_dev), port);
712 eth_dev = rte_eth_dev_allocate(name);
714 if (eth_dev == NULL) {
715 ERROR("can not allocate rte ethdev");
719 eth_dev->data->dev_private = priv;
720 eth_dev->data->mac_addrs = priv->mac;
721 eth_dev->device = &pci_dev->device;
722 rte_eth_copy_pci_info(eth_dev, pci_dev);
723 eth_dev->device->driver = &mlx4_driver.driver;
724 /* Initialize local interrupt handle for current port. */
725 priv->intr_handle = (struct rte_intr_handle){
727 .type = RTE_INTR_HANDLE_EXT,
730 * Override ethdev interrupt handle pointer with private
731 * handle instead of that of the parent PCI device used by
732 * default. This prevents it from being shared between all
733 * ports of the same PCI device since each of them is
734 * associated its own Verbs context.
736 * Rx interrupts in particular require this as the PMD has
737 * no control over the registration of queue interrupts
738 * besides setting up eth_dev->intr_handle, the rest is
739 * handled by rte_intr_rx_ctl().
741 eth_dev->intr_handle = &priv->intr_handle;
743 eth_dev->dev_ops = &mlx4_dev_ops;
744 /* Bring Ethernet device up. */
745 DEBUG("forcing Ethernet interface up");
746 mlx4_dev_set_link_up(priv->dev);
747 /* Update link status once if waiting for LSC. */
748 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
749 mlx4_link_update(eth_dev, 0);
751 * Once the device is added to the list of memory event
752 * callback, its global MR cache table cannot be expanded
753 * on the fly because of deadlock. If it overflows, lookup
754 * should be done by searching MR list linearly, which is slow.
756 err = mlx4_mr_btree_init(&priv->mr.cache,
757 MLX4_MR_BTREE_CACHE_N * 2,
758 eth_dev->device->numa_node);
760 /* rte_errno is already set. */
763 /* Add device to memory callback list. */
764 rte_rwlock_write_lock(&mlx4_mem_event_rwlock);
765 LIST_INSERT_HEAD(&mlx4_mem_event_cb_list, priv, mem_event_cb);
766 rte_rwlock_write_unlock(&mlx4_mem_event_rwlock);
767 rte_eth_dev_probing_finish(eth_dev);
772 claim_zero(mlx4_glue->dealloc_pd(pd));
774 claim_zero(mlx4_glue->close_device(ctx));
776 rte_eth_dev_release_port(eth_dev);
779 if (i == device_attr.phys_port_cnt)
782 * XXX if something went wrong in the loop above, there is a resource
783 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
784 * long as the dpdk does not provide a way to deallocate a ethdev and a
785 * way to enumerate the registered ethdevs to free the previous ones.
789 claim_zero(mlx4_glue->close_device(attr_ctx));
791 mlx4_glue->free_device_list(list);
792 assert(rte_errno >= 0);
796 static const struct rte_pci_id mlx4_pci_id_map[] = {
798 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
799 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
802 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
803 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
806 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
807 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
814 static struct rte_pci_driver mlx4_driver = {
816 .name = MLX4_DRIVER_NAME
818 .id_table = mlx4_pci_id_map,
819 .probe = mlx4_pci_probe,
820 .drv_flags = RTE_PCI_DRV_INTR_LSC |
821 RTE_PCI_DRV_INTR_RMV,
824 #ifdef RTE_LIBRTE_MLX4_DLOPEN_DEPS
827 * Suffix RTE_EAL_PMD_PATH with "-glue".
829 * This function performs a sanity check on RTE_EAL_PMD_PATH before
830 * suffixing its last component.
833 * Output buffer, should be large enough otherwise NULL is returned.
838 * Pointer to @p buf or @p NULL in case suffix cannot be appended.
841 mlx4_glue_path(char *buf, size_t size)
843 static const char *const bad[] = { "/", ".", "..", NULL };
844 const char *path = RTE_EAL_PMD_PATH;
845 size_t len = strlen(path);
849 while (len && path[len - 1] == '/')
851 for (off = len; off && path[off - 1] != '/'; --off)
853 for (i = 0; bad[i]; ++i)
854 if (!strncmp(path + off, bad[i], (int)(len - off)))
856 i = snprintf(buf, size, "%.*s-glue", (int)len, path);
857 if (i == -1 || (size_t)i >= size)
861 ERROR("unable to append \"-glue\" to last component of"
862 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
863 " please re-configure DPDK");
868 * Initialization routine for run-time dependency on rdma-core.
873 char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
874 const char *path[] = {
876 * A basic security check is necessary before trusting
877 * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
879 (geteuid() == getuid() && getegid() == getgid() ?
880 getenv("MLX4_GLUE_PATH") : NULL),
882 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
883 * variant, otherwise let dlopen() look up libraries on its
887 mlx4_glue_path(glue_path, sizeof(glue_path)) : ""),
894 while (!handle && i != RTE_DIM(path)) {
903 end = strpbrk(path[i], ":;");
905 end = path[i] + strlen(path[i]);
911 ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE,
913 (!len || *(end - 1) == '/') ? "" : "/");
916 if (sizeof(name) != (size_t)ret + 1)
918 DEBUG("looking for rdma-core glue as \"%s\"", name);
919 handle = dlopen(name, RTLD_LAZY);
930 WARN("cannot load glue library: %s", dlmsg);
933 sym = dlsym(handle, "mlx4_glue");
938 ERROR("cannot resolve glue symbol: %s", dlmsg);
946 WARN("cannot initialize PMD due to missing run-time"
947 " dependency on rdma-core libraries (libibverbs,"
955 * Driver initialization routine.
957 RTE_INIT(rte_mlx4_pmd_init);
959 rte_mlx4_pmd_init(void)
962 * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
963 * want to get success errno value in case of calling them
964 * when the device was removed.
966 setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
968 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
969 * huge pages. Calling ibv_fork_init() during init allows
970 * applications to use fork() safely for purposes other than
971 * using this PMD, which is not supported in forked processes.
973 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
974 #ifdef RTE_LIBRTE_MLX4_DLOPEN_DEPS
975 if (mlx4_glue_init())
980 /* Glue structure must not contain any NULL pointers. */
984 for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i)
985 assert(((const void *const *)mlx4_glue)[i]);
988 if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) {
989 ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required",
990 mlx4_glue->version, MLX4_GLUE_VERSION);
993 mlx4_glue->fork_init();
994 rte_pci_register(&mlx4_driver);
995 rte_mem_event_callback_register("MLX4_MEM_EVENT_CB",
996 mlx4_mr_mem_event_cb, NULL);
999 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
1000 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
1001 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
1002 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");