4 * Copyright 2012 6WIND S.A.
5 * Copyright 2012 Mellanox
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of 6WIND S.A. nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 * mlx4 driver initialization.
48 /* Verbs headers do not support -pedantic. */
50 #pragma GCC diagnostic ignored "-Wpedantic"
52 #include <infiniband/verbs.h>
54 #pragma GCC diagnostic error "-Wpedantic"
57 #include <rte_common.h>
59 #include <rte_errno.h>
60 #include <rte_ethdev_driver.h>
61 #include <rte_ethdev_pci.h>
62 #include <rte_ether.h>
64 #include <rte_interrupts.h>
65 #include <rte_kvargs.h>
66 #include <rte_malloc.h>
70 #include "mlx4_flow.h"
71 #include "mlx4_rxtx.h"
72 #include "mlx4_utils.h"
74 /** Configuration structure for device arguments. */
77 uint32_t present; /**< Bit-field for existing ports. */
78 uint32_t enabled; /**< Bit-field for user-enabled ports. */
82 /* Available parameters list. */
83 const char *pmd_mlx4_init_params[] = {
89 * DPDK callback for Ethernet device configuration.
92 * Pointer to Ethernet device structure.
95 * 0 on success, negative errno value otherwise and rte_errno is set.
98 mlx4_dev_configure(struct rte_eth_dev *dev)
100 struct priv *priv = dev->data->dev_private;
101 struct rte_flow_error error;
104 /* Prepare internal flow rules. */
105 ret = mlx4_flow_sync(priv, &error);
107 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
108 " flow error type %d, cause %p, message: %s",
109 -ret, strerror(-ret), error.type, error.cause,
110 error.message ? error.message : "(unspecified)");
116 * DPDK callback to start the device.
118 * Simulate device start by initializing common RSS resources and attaching
119 * all configured flows.
122 * Pointer to Ethernet device structure.
125 * 0 on success, negative errno value otherwise and rte_errno is set.
128 mlx4_dev_start(struct rte_eth_dev *dev)
130 struct priv *priv = dev->data->dev_private;
131 struct rte_flow_error error;
136 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
138 ret = mlx4_rss_init(priv);
140 ERROR("%p: cannot initialize RSS resources: %s",
141 (void *)dev, strerror(-ret));
144 ret = mlx4_intr_install(priv);
146 ERROR("%p: interrupt handler installation failed",
150 ret = mlx4_flow_sync(priv, &error);
152 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
153 " flow error type %d, cause %p, message: %s",
155 -ret, strerror(-ret), error.type, error.cause,
156 error.message ? error.message : "(unspecified)");
160 dev->tx_pkt_burst = mlx4_tx_burst;
161 dev->rx_pkt_burst = mlx4_rx_burst;
170 * DPDK callback to stop the device.
172 * Simulate device stop by detaching all configured flows.
175 * Pointer to Ethernet device structure.
178 mlx4_dev_stop(struct rte_eth_dev *dev)
180 struct priv *priv = dev->data->dev_private;
184 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
186 dev->tx_pkt_burst = mlx4_tx_burst_removed;
187 dev->rx_pkt_burst = mlx4_rx_burst_removed;
189 mlx4_flow_sync(priv, NULL);
190 mlx4_intr_uninstall(priv);
191 mlx4_rss_deinit(priv);
195 * DPDK callback to close the device.
197 * Destroy all queues and objects, free memory.
200 * Pointer to Ethernet device structure.
203 mlx4_dev_close(struct rte_eth_dev *dev)
205 struct priv *priv = dev->data->dev_private;
208 DEBUG("%p: closing device \"%s\"",
210 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
211 dev->rx_pkt_burst = mlx4_rx_burst_removed;
212 dev->tx_pkt_burst = mlx4_tx_burst_removed;
214 mlx4_flow_clean(priv);
215 for (i = 0; i != dev->data->nb_rx_queues; ++i)
216 mlx4_rx_queue_release(dev->data->rx_queues[i]);
217 for (i = 0; i != dev->data->nb_tx_queues; ++i)
218 mlx4_tx_queue_release(dev->data->tx_queues[i]);
219 if (priv->pd != NULL) {
220 assert(priv->ctx != NULL);
221 claim_zero(ibv_dealloc_pd(priv->pd));
222 claim_zero(ibv_close_device(priv->ctx));
224 assert(priv->ctx == NULL);
225 mlx4_intr_uninstall(priv);
226 memset(priv, 0, sizeof(*priv));
229 static const struct eth_dev_ops mlx4_dev_ops = {
230 .dev_configure = mlx4_dev_configure,
231 .dev_start = mlx4_dev_start,
232 .dev_stop = mlx4_dev_stop,
233 .dev_set_link_down = mlx4_dev_set_link_down,
234 .dev_set_link_up = mlx4_dev_set_link_up,
235 .dev_close = mlx4_dev_close,
236 .link_update = mlx4_link_update,
237 .promiscuous_enable = mlx4_promiscuous_enable,
238 .promiscuous_disable = mlx4_promiscuous_disable,
239 .allmulticast_enable = mlx4_allmulticast_enable,
240 .allmulticast_disable = mlx4_allmulticast_disable,
241 .mac_addr_remove = mlx4_mac_addr_remove,
242 .mac_addr_add = mlx4_mac_addr_add,
243 .mac_addr_set = mlx4_mac_addr_set,
244 .stats_get = mlx4_stats_get,
245 .stats_reset = mlx4_stats_reset,
246 .dev_infos_get = mlx4_dev_infos_get,
247 .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
248 .vlan_filter_set = mlx4_vlan_filter_set,
249 .rx_queue_setup = mlx4_rx_queue_setup,
250 .tx_queue_setup = mlx4_tx_queue_setup,
251 .rx_queue_release = mlx4_rx_queue_release,
252 .tx_queue_release = mlx4_tx_queue_release,
253 .flow_ctrl_get = mlx4_flow_ctrl_get,
254 .flow_ctrl_set = mlx4_flow_ctrl_set,
255 .mtu_set = mlx4_mtu_set,
256 .filter_ctrl = mlx4_filter_ctrl,
257 .rx_queue_intr_enable = mlx4_rx_intr_enable,
258 .rx_queue_intr_disable = mlx4_rx_intr_disable,
259 .is_removed = mlx4_is_removed,
263 * Get PCI information from struct ibv_device.
266 * Pointer to Ethernet device structure.
267 * @param[out] pci_addr
268 * PCI bus address output buffer.
271 * 0 on success, negative errno value otherwise and rte_errno is set.
274 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
275 struct rte_pci_addr *pci_addr)
279 MKSTR(path, "%s/device/uevent", device->ibdev_path);
281 file = fopen(path, "rb");
286 while (fgets(line, sizeof(line), file) == line) {
287 size_t len = strlen(line);
290 /* Truncate long lines. */
291 if (len == (sizeof(line) - 1))
292 while (line[(len - 1)] != '\n') {
296 line[(len - 1)] = ret;
298 /* Extract information. */
301 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
305 &pci_addr->function) == 4) {
315 * Verify and store value for device argument.
318 * Key argument to verify.
320 * Value associated with key.
321 * @param[in, out] conf
322 * Shared configuration data.
325 * 0 on success, negative errno value otherwise and rte_errno is set.
328 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
333 tmp = strtoul(val, NULL, 0);
336 WARN("%s: \"%s\" is not a valid integer", key, val);
339 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
340 uint32_t ports = rte_log2_u32(conf->ports.present + 1);
343 ERROR("port index %lu outside range [0,%" PRIu32 ")",
347 if (!(conf->ports.present & (1 << tmp))) {
349 ERROR("invalid port index %lu", tmp);
352 conf->ports.enabled |= 1 << tmp;
355 WARN("%s: unknown parameter", key);
362 * Parse device parameters.
365 * Device arguments structure.
368 * 0 on success, negative errno value otherwise and rte_errno is set.
371 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
373 struct rte_kvargs *kvlist;
374 unsigned int arg_count;
380 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
381 if (kvlist == NULL) {
383 ERROR("failed to parse kvargs");
386 /* Process parameters. */
387 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
388 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
389 while (arg_count-- > 0) {
390 ret = rte_kvargs_process(kvlist,
392 (int (*)(const char *,
402 rte_kvargs_free(kvlist);
406 static struct rte_pci_driver mlx4_driver;
409 * DPDK callback to register a PCI device.
411 * This function creates an Ethernet device for each port of a given
415 * PCI driver structure (mlx4_driver).
417 * PCI device information.
420 * 0 on success, negative errno value otherwise and rte_errno is set.
423 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
425 struct ibv_device **list;
426 struct ibv_device *ibv_dev;
428 struct ibv_context *attr_ctx = NULL;
429 struct ibv_device_attr device_attr;
430 struct ibv_device_attr_ex device_attr_ex;
431 struct mlx4_conf conf = {
438 assert(pci_drv == &mlx4_driver);
439 list = ibv_get_device_list(&i);
443 if (rte_errno == ENOSYS)
444 ERROR("cannot list devices, is ib_uverbs loaded?");
449 * For each listed device, check related sysfs entry against
450 * the provided PCI ID.
453 struct rte_pci_addr pci_addr;
456 DEBUG("checking device \"%s\"", list[i]->name);
457 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
459 if ((pci_dev->addr.domain != pci_addr.domain) ||
460 (pci_dev->addr.bus != pci_addr.bus) ||
461 (pci_dev->addr.devid != pci_addr.devid) ||
462 (pci_dev->addr.function != pci_addr.function))
464 vf = (pci_dev->id.device_id ==
465 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
466 INFO("PCI information matches, using device \"%s\" (VF: %s)",
467 list[i]->name, (vf ? "true" : "false"));
468 attr_ctx = ibv_open_device(list[i]);
472 if (attr_ctx == NULL) {
473 ibv_free_device_list(list);
477 ERROR("cannot access device, is mlx4_ib loaded?");
481 ERROR("cannot use device, are drivers up to date?");
489 DEBUG("device opened");
490 if (ibv_query_device(attr_ctx, &device_attr)) {
494 INFO("%u port(s) detected", device_attr.phys_port_cnt);
495 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
496 if (mlx4_args(pci_dev->device.devargs, &conf)) {
497 ERROR("failed to process device arguments");
501 /* Use all ports when none are defined */
502 if (!conf.ports.enabled)
503 conf.ports.enabled = conf.ports.present;
504 /* Retrieve extended device attributes. */
505 if (ibv_query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
509 assert(device_attr.max_sge >= MLX4_MAX_SGE);
510 for (i = 0; i < device_attr.phys_port_cnt; i++) {
511 uint32_t port = i + 1; /* ports are indexed from one */
512 struct ibv_context *ctx = NULL;
513 struct ibv_port_attr port_attr;
514 struct ibv_pd *pd = NULL;
515 struct priv *priv = NULL;
516 struct rte_eth_dev *eth_dev = NULL;
517 struct ether_addr mac;
519 /* If port is not enabled, skip. */
520 if (!(conf.ports.enabled & (1 << i)))
522 DEBUG("using port %u", port);
523 ctx = ibv_open_device(ibv_dev);
528 /* Check port status. */
529 err = ibv_query_port(ctx, port, &port_attr);
532 ERROR("port query failed: %s", strerror(rte_errno));
535 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
537 ERROR("port %d is not configured in Ethernet mode",
541 if (port_attr.state != IBV_PORT_ACTIVE)
542 DEBUG("port %d is not active: \"%s\" (%d)",
543 port, ibv_port_state_str(port_attr.state),
545 /* Make asynchronous FD non-blocking to handle interrupts. */
546 if (mlx4_fd_set_non_blocking(ctx->async_fd) < 0) {
547 ERROR("cannot make asynchronous FD non-blocking: %s",
548 strerror(rte_errno));
551 /* Allocate protection domain. */
552 pd = ibv_alloc_pd(ctx);
555 ERROR("PD allocation failure");
558 /* from rte_ethdev.c */
559 priv = rte_zmalloc("ethdev private structure",
561 RTE_CACHE_LINE_SIZE);
564 ERROR("priv allocation failure");
568 priv->device_attr = device_attr;
571 priv->mtu = ETHER_MTU;
573 priv->hw_csum = !!(device_attr.device_cap_flags &
574 IBV_DEVICE_RAW_IP_CSUM);
575 DEBUG("checksum offloading is %ssupported",
576 (priv->hw_csum ? "" : "not "));
577 /* Only ConnectX-3 Pro supports tunneling. */
578 priv->hw_csum_l2tun =
580 (device_attr.vendor_part_id ==
581 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
582 DEBUG("L2 tunnel checksum offloads are %ssupported",
583 (priv->hw_csum_l2tun ? "" : "not "));
584 priv->hw_rss_sup = device_attr_ex.rss_caps.rx_hash_fields_mask;
585 if (!priv->hw_rss_sup) {
586 WARN("no RSS capabilities reported; disabling support"
587 " for UDP RSS and inner VXLAN RSS");
588 /* Fake support for all possible RSS hash fields. */
589 priv->hw_rss_sup = ~UINT64_C(0);
590 priv->hw_rss_sup = mlx4_conv_rss_hf(priv, -1);
591 /* Filter out known unsupported fields. */
593 ~(uint64_t)(IBV_RX_HASH_SRC_PORT_UDP |
594 IBV_RX_HASH_DST_PORT_UDP |
597 DEBUG("supported RSS hash fields mask: %016" PRIx64,
599 /* Configure the first MAC address by default. */
600 if (mlx4_get_mac(priv, &mac.addr_bytes)) {
601 ERROR("cannot get MAC address, is mlx4_en loaded?"
602 " (rte_errno: %s)", strerror(rte_errno));
605 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
607 mac.addr_bytes[0], mac.addr_bytes[1],
608 mac.addr_bytes[2], mac.addr_bytes[3],
609 mac.addr_bytes[4], mac.addr_bytes[5]);
610 /* Register MAC address. */
614 char ifname[IF_NAMESIZE];
616 if (mlx4_get_ifname(priv, &ifname) == 0)
617 DEBUG("port %u ifname is \"%s\"",
620 DEBUG("port %u ifname is unknown", priv->port);
623 /* Get actual MTU if possible. */
624 mlx4_mtu_get(priv, &priv->mtu);
625 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
626 /* from rte_ethdev.c */
628 char name[RTE_ETH_NAME_MAX_LEN];
630 snprintf(name, sizeof(name), "%s port %u",
631 ibv_get_device_name(ibv_dev), port);
632 eth_dev = rte_eth_dev_allocate(name);
634 if (eth_dev == NULL) {
635 ERROR("can not allocate rte ethdev");
639 eth_dev->data->dev_private = priv;
640 eth_dev->data->mac_addrs = priv->mac;
641 eth_dev->device = &pci_dev->device;
642 rte_eth_copy_pci_info(eth_dev, pci_dev);
643 eth_dev->device->driver = &mlx4_driver.driver;
644 /* Initialize local interrupt handle for current port. */
645 priv->intr_handle = (struct rte_intr_handle){
647 .type = RTE_INTR_HANDLE_EXT,
650 * Override ethdev interrupt handle pointer with private
651 * handle instead of that of the parent PCI device used by
652 * default. This prevents it from being shared between all
653 * ports of the same PCI device since each of them is
654 * associated its own Verbs context.
656 * Rx interrupts in particular require this as the PMD has
657 * no control over the registration of queue interrupts
658 * besides setting up eth_dev->intr_handle, the rest is
659 * handled by rte_intr_rx_ctl().
661 eth_dev->intr_handle = &priv->intr_handle;
663 eth_dev->dev_ops = &mlx4_dev_ops;
664 /* Bring Ethernet device up. */
665 DEBUG("forcing Ethernet interface up");
666 mlx4_dev_set_link_up(priv->dev);
667 /* Update link status once if waiting for LSC. */
668 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
669 mlx4_link_update(eth_dev, 0);
674 claim_zero(ibv_dealloc_pd(pd));
676 claim_zero(ibv_close_device(ctx));
678 rte_eth_dev_release_port(eth_dev);
681 if (i == device_attr.phys_port_cnt)
684 * XXX if something went wrong in the loop above, there is a resource
685 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
686 * long as the dpdk does not provide a way to deallocate a ethdev and a
687 * way to enumerate the registered ethdevs to free the previous ones.
691 claim_zero(ibv_close_device(attr_ctx));
693 ibv_free_device_list(list);
694 assert(rte_errno >= 0);
698 static const struct rte_pci_id mlx4_pci_id_map[] = {
700 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
701 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
704 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
705 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
708 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
709 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
716 static struct rte_pci_driver mlx4_driver = {
718 .name = MLX4_DRIVER_NAME
720 .id_table = mlx4_pci_id_map,
721 .probe = mlx4_pci_probe,
722 .drv_flags = RTE_PCI_DRV_INTR_LSC |
723 RTE_PCI_DRV_INTR_RMV,
727 * Driver initialization routine.
729 RTE_INIT(rte_mlx4_pmd_init);
731 rte_mlx4_pmd_init(void)
734 * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
735 * want to get success errno value in case of calling them
736 * when the device was removed.
738 setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
740 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
741 * huge pages. Calling ibv_fork_init() during init allows
742 * applications to use fork() safely for purposes other than
743 * using this PMD, which is not supported in forked processes.
745 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
747 rte_pci_register(&mlx4_driver);
750 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
751 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
752 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
753 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");