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_glue.h"
71 #include "mlx4_flow.h"
72 #include "mlx4_rxtx.h"
73 #include "mlx4_utils.h"
75 /** Configuration structure for device arguments. */
78 uint32_t present; /**< Bit-field for existing ports. */
79 uint32_t enabled; /**< Bit-field for user-enabled ports. */
83 /* Available parameters list. */
84 const char *pmd_mlx4_init_params[] = {
90 * DPDK callback for Ethernet device configuration.
93 * Pointer to Ethernet device structure.
96 * 0 on success, negative errno value otherwise and rte_errno is set.
99 mlx4_dev_configure(struct rte_eth_dev *dev)
101 struct priv *priv = dev->data->dev_private;
102 struct rte_flow_error error;
105 /* Prepare internal flow rules. */
106 ret = mlx4_flow_sync(priv, &error);
108 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
109 " flow error type %d, cause %p, message: %s",
110 -ret, strerror(-ret), error.type, error.cause,
111 error.message ? error.message : "(unspecified)");
114 ret = mlx4_intr_install(priv);
116 ERROR("%p: interrupt handler installation failed",
123 * DPDK callback to start the device.
125 * Simulate device start by initializing common RSS resources and attaching
126 * all configured flows.
129 * Pointer to Ethernet device structure.
132 * 0 on success, negative errno value otherwise and rte_errno is set.
135 mlx4_dev_start(struct rte_eth_dev *dev)
137 struct priv *priv = dev->data->dev_private;
138 struct rte_flow_error error;
143 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
145 ret = mlx4_rss_init(priv);
147 ERROR("%p: cannot initialize RSS resources: %s",
148 (void *)dev, strerror(-ret));
151 ret = mlx4_rxq_intr_enable(priv);
153 ERROR("%p: interrupt handler installation failed",
157 ret = mlx4_flow_sync(priv, &error);
159 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
160 " flow error type %d, cause %p, message: %s",
162 -ret, strerror(-ret), error.type, error.cause,
163 error.message ? error.message : "(unspecified)");
167 dev->tx_pkt_burst = mlx4_tx_burst;
168 dev->rx_pkt_burst = mlx4_rx_burst;
177 * DPDK callback to stop the device.
179 * Simulate device stop by detaching all configured flows.
182 * Pointer to Ethernet device structure.
185 mlx4_dev_stop(struct rte_eth_dev *dev)
187 struct priv *priv = dev->data->dev_private;
191 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
193 dev->tx_pkt_burst = mlx4_tx_burst_removed;
194 dev->rx_pkt_burst = mlx4_rx_burst_removed;
196 mlx4_flow_sync(priv, NULL);
197 mlx4_rxq_intr_disable(priv);
198 mlx4_rss_deinit(priv);
202 * DPDK callback to close the device.
204 * Destroy all queues and objects, free memory.
207 * Pointer to Ethernet device structure.
210 mlx4_dev_close(struct rte_eth_dev *dev)
212 struct priv *priv = dev->data->dev_private;
215 DEBUG("%p: closing device \"%s\"",
217 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
218 dev->rx_pkt_burst = mlx4_rx_burst_removed;
219 dev->tx_pkt_burst = mlx4_tx_burst_removed;
221 mlx4_flow_clean(priv);
222 for (i = 0; i != dev->data->nb_rx_queues; ++i)
223 mlx4_rx_queue_release(dev->data->rx_queues[i]);
224 for (i = 0; i != dev->data->nb_tx_queues; ++i)
225 mlx4_tx_queue_release(dev->data->tx_queues[i]);
226 if (priv->pd != NULL) {
227 assert(priv->ctx != NULL);
228 claim_zero(mlx4_glue->dealloc_pd(priv->pd));
229 claim_zero(mlx4_glue->close_device(priv->ctx));
231 assert(priv->ctx == NULL);
232 mlx4_intr_uninstall(priv);
233 memset(priv, 0, sizeof(*priv));
236 static const struct eth_dev_ops mlx4_dev_ops = {
237 .dev_configure = mlx4_dev_configure,
238 .dev_start = mlx4_dev_start,
239 .dev_stop = mlx4_dev_stop,
240 .dev_set_link_down = mlx4_dev_set_link_down,
241 .dev_set_link_up = mlx4_dev_set_link_up,
242 .dev_close = mlx4_dev_close,
243 .link_update = mlx4_link_update,
244 .promiscuous_enable = mlx4_promiscuous_enable,
245 .promiscuous_disable = mlx4_promiscuous_disable,
246 .allmulticast_enable = mlx4_allmulticast_enable,
247 .allmulticast_disable = mlx4_allmulticast_disable,
248 .mac_addr_remove = mlx4_mac_addr_remove,
249 .mac_addr_add = mlx4_mac_addr_add,
250 .mac_addr_set = mlx4_mac_addr_set,
251 .stats_get = mlx4_stats_get,
252 .stats_reset = mlx4_stats_reset,
253 .dev_infos_get = mlx4_dev_infos_get,
254 .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
255 .vlan_filter_set = mlx4_vlan_filter_set,
256 .rx_queue_setup = mlx4_rx_queue_setup,
257 .tx_queue_setup = mlx4_tx_queue_setup,
258 .rx_queue_release = mlx4_rx_queue_release,
259 .tx_queue_release = mlx4_tx_queue_release,
260 .flow_ctrl_get = mlx4_flow_ctrl_get,
261 .flow_ctrl_set = mlx4_flow_ctrl_set,
262 .mtu_set = mlx4_mtu_set,
263 .filter_ctrl = mlx4_filter_ctrl,
264 .rx_queue_intr_enable = mlx4_rx_intr_enable,
265 .rx_queue_intr_disable = mlx4_rx_intr_disable,
266 .is_removed = mlx4_is_removed,
270 * Get PCI information from struct ibv_device.
273 * Pointer to Ethernet device structure.
274 * @param[out] pci_addr
275 * PCI bus address output buffer.
278 * 0 on success, negative errno value otherwise and rte_errno is set.
281 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
282 struct rte_pci_addr *pci_addr)
286 MKSTR(path, "%s/device/uevent", device->ibdev_path);
288 file = fopen(path, "rb");
293 while (fgets(line, sizeof(line), file) == line) {
294 size_t len = strlen(line);
297 /* Truncate long lines. */
298 if (len == (sizeof(line) - 1))
299 while (line[(len - 1)] != '\n') {
303 line[(len - 1)] = ret;
305 /* Extract information. */
308 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
312 &pci_addr->function) == 4) {
322 * Verify and store value for device argument.
325 * Key argument to verify.
327 * Value associated with key.
328 * @param[in, out] conf
329 * Shared configuration data.
332 * 0 on success, negative errno value otherwise and rte_errno is set.
335 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
340 tmp = strtoul(val, NULL, 0);
343 WARN("%s: \"%s\" is not a valid integer", key, val);
346 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
347 uint32_t ports = rte_log2_u32(conf->ports.present + 1);
350 ERROR("port index %lu outside range [0,%" PRIu32 ")",
354 if (!(conf->ports.present & (1 << tmp))) {
356 ERROR("invalid port index %lu", tmp);
359 conf->ports.enabled |= 1 << tmp;
362 WARN("%s: unknown parameter", key);
369 * Parse device parameters.
372 * Device arguments structure.
375 * 0 on success, negative errno value otherwise and rte_errno is set.
378 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
380 struct rte_kvargs *kvlist;
381 unsigned int arg_count;
387 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
388 if (kvlist == NULL) {
390 ERROR("failed to parse kvargs");
393 /* Process parameters. */
394 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
395 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
396 while (arg_count-- > 0) {
397 ret = rte_kvargs_process(kvlist,
399 (int (*)(const char *,
409 rte_kvargs_free(kvlist);
413 static struct rte_pci_driver mlx4_driver;
416 * DPDK callback to register a PCI device.
418 * This function creates an Ethernet device for each port of a given
422 * PCI driver structure (mlx4_driver).
424 * PCI device information.
427 * 0 on success, negative errno value otherwise and rte_errno is set.
430 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
432 struct ibv_device **list;
433 struct ibv_device *ibv_dev;
435 struct ibv_context *attr_ctx = NULL;
436 struct ibv_device_attr device_attr;
437 struct ibv_device_attr_ex device_attr_ex;
438 struct mlx4_conf conf = {
445 assert(pci_drv == &mlx4_driver);
446 list = mlx4_glue->get_device_list(&i);
450 if (rte_errno == ENOSYS)
451 ERROR("cannot list devices, is ib_uverbs loaded?");
456 * For each listed device, check related sysfs entry against
457 * the provided PCI ID.
460 struct rte_pci_addr pci_addr;
463 DEBUG("checking device \"%s\"", list[i]->name);
464 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
466 if ((pci_dev->addr.domain != pci_addr.domain) ||
467 (pci_dev->addr.bus != pci_addr.bus) ||
468 (pci_dev->addr.devid != pci_addr.devid) ||
469 (pci_dev->addr.function != pci_addr.function))
471 vf = (pci_dev->id.device_id ==
472 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
473 INFO("PCI information matches, using device \"%s\" (VF: %s)",
474 list[i]->name, (vf ? "true" : "false"));
475 attr_ctx = mlx4_glue->open_device(list[i]);
479 if (attr_ctx == NULL) {
480 mlx4_glue->free_device_list(list);
484 ERROR("cannot access device, is mlx4_ib loaded?");
488 ERROR("cannot use device, are drivers up to date?");
496 DEBUG("device opened");
497 if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
501 INFO("%u port(s) detected", device_attr.phys_port_cnt);
502 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
503 if (mlx4_args(pci_dev->device.devargs, &conf)) {
504 ERROR("failed to process device arguments");
508 /* Use all ports when none are defined */
509 if (!conf.ports.enabled)
510 conf.ports.enabled = conf.ports.present;
511 /* Retrieve extended device attributes. */
512 if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
516 assert(device_attr.max_sge >= MLX4_MAX_SGE);
517 for (i = 0; i < device_attr.phys_port_cnt; i++) {
518 uint32_t port = i + 1; /* ports are indexed from one */
519 struct ibv_context *ctx = NULL;
520 struct ibv_port_attr port_attr;
521 struct ibv_pd *pd = NULL;
522 struct priv *priv = NULL;
523 struct rte_eth_dev *eth_dev = NULL;
524 struct ether_addr mac;
526 /* If port is not enabled, skip. */
527 if (!(conf.ports.enabled & (1 << i)))
529 DEBUG("using port %u", port);
530 ctx = mlx4_glue->open_device(ibv_dev);
535 /* Check port status. */
536 err = mlx4_glue->query_port(ctx, port, &port_attr);
539 ERROR("port query failed: %s", strerror(rte_errno));
542 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
544 ERROR("port %d is not configured in Ethernet mode",
548 if (port_attr.state != IBV_PORT_ACTIVE)
549 DEBUG("port %d is not active: \"%s\" (%d)",
550 port, mlx4_glue->port_state_str(port_attr.state),
552 /* Make asynchronous FD non-blocking to handle interrupts. */
553 if (mlx4_fd_set_non_blocking(ctx->async_fd) < 0) {
554 ERROR("cannot make asynchronous FD non-blocking: %s",
555 strerror(rte_errno));
558 /* Allocate protection domain. */
559 pd = mlx4_glue->alloc_pd(ctx);
562 ERROR("PD allocation failure");
565 /* from rte_ethdev.c */
566 priv = rte_zmalloc("ethdev private structure",
568 RTE_CACHE_LINE_SIZE);
571 ERROR("priv allocation failure");
575 priv->device_attr = device_attr;
578 priv->mtu = ETHER_MTU;
580 priv->hw_csum = !!(device_attr.device_cap_flags &
581 IBV_DEVICE_RAW_IP_CSUM);
582 DEBUG("checksum offloading is %ssupported",
583 (priv->hw_csum ? "" : "not "));
584 /* Only ConnectX-3 Pro supports tunneling. */
585 priv->hw_csum_l2tun =
587 (device_attr.vendor_part_id ==
588 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
589 DEBUG("L2 tunnel checksum offloads are %ssupported",
590 (priv->hw_csum_l2tun ? "" : "not "));
591 priv->hw_rss_sup = device_attr_ex.rss_caps.rx_hash_fields_mask;
592 if (!priv->hw_rss_sup) {
593 WARN("no RSS capabilities reported; disabling support"
594 " for UDP RSS and inner VXLAN RSS");
595 /* Fake support for all possible RSS hash fields. */
596 priv->hw_rss_sup = ~UINT64_C(0);
597 priv->hw_rss_sup = mlx4_conv_rss_hf(priv, -1);
598 /* Filter out known unsupported fields. */
600 ~(uint64_t)(IBV_RX_HASH_SRC_PORT_UDP |
601 IBV_RX_HASH_DST_PORT_UDP |
604 DEBUG("supported RSS hash fields mask: %016" PRIx64,
606 /* Configure the first MAC address by default. */
607 if (mlx4_get_mac(priv, &mac.addr_bytes)) {
608 ERROR("cannot get MAC address, is mlx4_en loaded?"
609 " (rte_errno: %s)", strerror(rte_errno));
612 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
614 mac.addr_bytes[0], mac.addr_bytes[1],
615 mac.addr_bytes[2], mac.addr_bytes[3],
616 mac.addr_bytes[4], mac.addr_bytes[5]);
617 /* Register MAC address. */
621 char ifname[IF_NAMESIZE];
623 if (mlx4_get_ifname(priv, &ifname) == 0)
624 DEBUG("port %u ifname is \"%s\"",
627 DEBUG("port %u ifname is unknown", priv->port);
630 /* Get actual MTU if possible. */
631 mlx4_mtu_get(priv, &priv->mtu);
632 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
633 /* from rte_ethdev.c */
635 char name[RTE_ETH_NAME_MAX_LEN];
637 snprintf(name, sizeof(name), "%s port %u",
638 mlx4_glue->get_device_name(ibv_dev), port);
639 eth_dev = rte_eth_dev_allocate(name);
641 if (eth_dev == NULL) {
642 ERROR("can not allocate rte ethdev");
646 eth_dev->data->dev_private = priv;
647 eth_dev->data->mac_addrs = priv->mac;
648 eth_dev->device = &pci_dev->device;
649 rte_eth_copy_pci_info(eth_dev, pci_dev);
650 eth_dev->device->driver = &mlx4_driver.driver;
651 /* Initialize local interrupt handle for current port. */
652 priv->intr_handle = (struct rte_intr_handle){
654 .type = RTE_INTR_HANDLE_EXT,
657 * Override ethdev interrupt handle pointer with private
658 * handle instead of that of the parent PCI device used by
659 * default. This prevents it from being shared between all
660 * ports of the same PCI device since each of them is
661 * associated its own Verbs context.
663 * Rx interrupts in particular require this as the PMD has
664 * no control over the registration of queue interrupts
665 * besides setting up eth_dev->intr_handle, the rest is
666 * handled by rte_intr_rx_ctl().
668 eth_dev->intr_handle = &priv->intr_handle;
670 eth_dev->dev_ops = &mlx4_dev_ops;
671 /* Bring Ethernet device up. */
672 DEBUG("forcing Ethernet interface up");
673 mlx4_dev_set_link_up(priv->dev);
674 /* Update link status once if waiting for LSC. */
675 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
676 mlx4_link_update(eth_dev, 0);
681 claim_zero(mlx4_glue->dealloc_pd(pd));
683 claim_zero(mlx4_glue->close_device(ctx));
685 rte_eth_dev_release_port(eth_dev);
688 if (i == device_attr.phys_port_cnt)
691 * XXX if something went wrong in the loop above, there is a resource
692 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
693 * long as the dpdk does not provide a way to deallocate a ethdev and a
694 * way to enumerate the registered ethdevs to free the previous ones.
698 claim_zero(mlx4_glue->close_device(attr_ctx));
700 mlx4_glue->free_device_list(list);
701 assert(rte_errno >= 0);
705 static const struct rte_pci_id mlx4_pci_id_map[] = {
707 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
708 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
711 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
712 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
715 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
716 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
723 static struct rte_pci_driver mlx4_driver = {
725 .name = MLX4_DRIVER_NAME
727 .id_table = mlx4_pci_id_map,
728 .probe = mlx4_pci_probe,
729 .drv_flags = RTE_PCI_DRV_INTR_LSC |
730 RTE_PCI_DRV_INTR_RMV,
734 * Driver initialization routine.
736 RTE_INIT(rte_mlx4_pmd_init);
738 rte_mlx4_pmd_init(void)
741 * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
742 * want to get success errno value in case of calling them
743 * when the device was removed.
745 setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
747 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
748 * huge pages. Calling ibv_fork_init() during init allows
749 * applications to use fork() safely for purposes other than
750 * using this PMD, which is not supported in forked processes.
752 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
753 mlx4_glue->fork_init();
754 rte_pci_register(&mlx4_driver);
757 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
758 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
759 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
760 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");