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>
53 #include <infiniband/mlx4dv.h>
55 #pragma GCC diagnostic error "-Wpedantic"
58 #include <rte_common.h>
60 #include <rte_errno.h>
61 #include <rte_ethdev.h>
62 #include <rte_ethdev_pci.h>
63 #include <rte_ether.h>
65 #include <rte_interrupts.h>
66 #include <rte_kvargs.h>
67 #include <rte_malloc.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;
103 uint8_t log2_range = rte_log2_u32(dev->data->nb_rx_queues);
106 /* Prepare range for RSS contexts before creating the first WQ. */
107 ret = mlx4dv_set_context_attr(priv->ctx,
108 MLX4DV_SET_CTX_ATTR_LOG_WQS_RANGE_SZ,
111 ERROR("cannot set up range size for RSS context to %u"
112 " (for %u Rx queues), error: %s",
113 1 << log2_range, dev->data->nb_rx_queues, strerror(ret));
117 /* Prepare internal flow rules. */
118 ret = mlx4_flow_sync(priv, &error);
120 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
121 " flow error type %d, cause %p, message: %s",
122 -ret, strerror(-ret), error.type, error.cause,
123 error.message ? error.message : "(unspecified)");
129 * DPDK callback to start the device.
131 * Simulate device start by attaching all configured flows.
134 * Pointer to Ethernet device structure.
137 * 0 on success, negative errno value otherwise and rte_errno is set.
140 mlx4_dev_start(struct rte_eth_dev *dev)
142 struct priv *priv = dev->data->dev_private;
143 struct rte_flow_error error;
148 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
150 ret = mlx4_intr_install(priv);
152 ERROR("%p: interrupt handler installation failed",
156 ret = mlx4_flow_sync(priv, &error);
158 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
159 " flow error type %d, cause %p, message: %s",
161 -ret, strerror(-ret), error.type, error.cause,
162 error.message ? error.message : "(unspecified)");
166 dev->tx_pkt_burst = mlx4_tx_burst;
167 dev->rx_pkt_burst = mlx4_rx_burst;
176 * DPDK callback to stop the device.
178 * Simulate device stop by detaching all configured flows.
181 * Pointer to Ethernet device structure.
184 mlx4_dev_stop(struct rte_eth_dev *dev)
186 struct priv *priv = dev->data->dev_private;
190 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
192 dev->tx_pkt_burst = mlx4_tx_burst_removed;
193 dev->rx_pkt_burst = mlx4_rx_burst_removed;
195 mlx4_flow_sync(priv, NULL);
196 mlx4_intr_uninstall(priv);
200 * DPDK callback to close the device.
202 * Destroy all queues and objects, free memory.
205 * Pointer to Ethernet device structure.
208 mlx4_dev_close(struct rte_eth_dev *dev)
210 struct priv *priv = dev->data->dev_private;
213 DEBUG("%p: closing device \"%s\"",
215 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
216 dev->rx_pkt_burst = mlx4_rx_burst_removed;
217 dev->tx_pkt_burst = mlx4_tx_burst_removed;
219 mlx4_flow_clean(priv);
220 for (i = 0; i != dev->data->nb_rx_queues; ++i)
221 mlx4_rx_queue_release(dev->data->rx_queues[i]);
222 for (i = 0; i != dev->data->nb_tx_queues; ++i)
223 mlx4_tx_queue_release(dev->data->tx_queues[i]);
224 if (priv->pd != NULL) {
225 assert(priv->ctx != NULL);
226 claim_zero(ibv_dealloc_pd(priv->pd));
227 claim_zero(ibv_close_device(priv->ctx));
229 assert(priv->ctx == NULL);
230 mlx4_intr_uninstall(priv);
231 memset(priv, 0, sizeof(*priv));
234 static const struct eth_dev_ops mlx4_dev_ops = {
235 .dev_configure = mlx4_dev_configure,
236 .dev_start = mlx4_dev_start,
237 .dev_stop = mlx4_dev_stop,
238 .dev_set_link_down = mlx4_dev_set_link_down,
239 .dev_set_link_up = mlx4_dev_set_link_up,
240 .dev_close = mlx4_dev_close,
241 .link_update = mlx4_link_update,
242 .promiscuous_enable = mlx4_promiscuous_enable,
243 .promiscuous_disable = mlx4_promiscuous_disable,
244 .allmulticast_enable = mlx4_allmulticast_enable,
245 .allmulticast_disable = mlx4_allmulticast_disable,
246 .mac_addr_remove = mlx4_mac_addr_remove,
247 .mac_addr_add = mlx4_mac_addr_add,
248 .mac_addr_set = mlx4_mac_addr_set,
249 .stats_get = mlx4_stats_get,
250 .stats_reset = mlx4_stats_reset,
251 .dev_infos_get = mlx4_dev_infos_get,
252 .vlan_filter_set = mlx4_vlan_filter_set,
253 .rx_queue_setup = mlx4_rx_queue_setup,
254 .tx_queue_setup = mlx4_tx_queue_setup,
255 .rx_queue_release = mlx4_rx_queue_release,
256 .tx_queue_release = mlx4_tx_queue_release,
257 .flow_ctrl_get = mlx4_flow_ctrl_get,
258 .flow_ctrl_set = mlx4_flow_ctrl_set,
259 .mtu_set = mlx4_mtu_set,
260 .filter_ctrl = mlx4_filter_ctrl,
261 .rx_queue_intr_enable = mlx4_rx_intr_enable,
262 .rx_queue_intr_disable = mlx4_rx_intr_disable,
266 * Get PCI information from struct ibv_device.
269 * Pointer to Ethernet device structure.
270 * @param[out] pci_addr
271 * PCI bus address output buffer.
274 * 0 on success, negative errno value otherwise and rte_errno is set.
277 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
278 struct rte_pci_addr *pci_addr)
282 MKSTR(path, "%s/device/uevent", device->ibdev_path);
284 file = fopen(path, "rb");
289 while (fgets(line, sizeof(line), file) == line) {
290 size_t len = strlen(line);
293 /* Truncate long lines. */
294 if (len == (sizeof(line) - 1))
295 while (line[(len - 1)] != '\n') {
299 line[(len - 1)] = ret;
301 /* Extract information. */
304 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
308 &pci_addr->function) == 4) {
318 * Verify and store value for device argument.
321 * Key argument to verify.
323 * Value associated with key.
324 * @param[in, out] conf
325 * Shared configuration data.
328 * 0 on success, negative errno value otherwise and rte_errno is set.
331 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
336 tmp = strtoul(val, NULL, 0);
339 WARN("%s: \"%s\" is not a valid integer", key, val);
342 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
343 uint32_t ports = rte_log2_u32(conf->ports.present);
346 ERROR("port index %lu outside range [0,%" PRIu32 ")",
350 if (!(conf->ports.present & (1 << tmp))) {
352 ERROR("invalid port index %lu", tmp);
355 conf->ports.enabled |= 1 << tmp;
358 WARN("%s: unknown parameter", key);
365 * Parse device parameters.
368 * Device arguments structure.
371 * 0 on success, negative errno value otherwise and rte_errno is set.
374 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
376 struct rte_kvargs *kvlist;
377 unsigned int arg_count;
383 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
384 if (kvlist == NULL) {
386 ERROR("failed to parse kvargs");
389 /* Process parameters. */
390 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
391 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
392 while (arg_count-- > 0) {
393 ret = rte_kvargs_process(kvlist,
395 (int (*)(const char *,
405 rte_kvargs_free(kvlist);
409 static struct rte_pci_driver mlx4_driver;
412 * DPDK callback to register a PCI device.
414 * This function creates an Ethernet device for each port of a given
418 * PCI driver structure (mlx4_driver).
420 * PCI device information.
423 * 0 on success, negative errno value otherwise and rte_errno is set.
426 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
428 struct ibv_device **list;
429 struct ibv_device *ibv_dev;
431 struct ibv_context *attr_ctx = NULL;
432 struct ibv_device_attr device_attr;
433 struct mlx4_conf conf = {
440 assert(pci_drv == &mlx4_driver);
441 list = ibv_get_device_list(&i);
445 if (rte_errno == ENOSYS)
446 ERROR("cannot list devices, is ib_uverbs loaded?");
451 * For each listed device, check related sysfs entry against
452 * the provided PCI ID.
455 struct rte_pci_addr pci_addr;
458 DEBUG("checking device \"%s\"", list[i]->name);
459 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
461 if ((pci_dev->addr.domain != pci_addr.domain) ||
462 (pci_dev->addr.bus != pci_addr.bus) ||
463 (pci_dev->addr.devid != pci_addr.devid) ||
464 (pci_dev->addr.function != pci_addr.function))
466 vf = (pci_dev->id.device_id ==
467 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
468 INFO("PCI information matches, using device \"%s\" (VF: %s)",
469 list[i]->name, (vf ? "true" : "false"));
470 attr_ctx = ibv_open_device(list[i]);
474 if (attr_ctx == NULL) {
475 ibv_free_device_list(list);
479 ERROR("cannot access device, is mlx4_ib loaded?");
483 ERROR("cannot use device, are drivers up to date?");
491 DEBUG("device opened");
492 if (ibv_query_device(attr_ctx, &device_attr)) {
496 INFO("%u port(s) detected", device_attr.phys_port_cnt);
497 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
498 if (mlx4_args(pci_dev->device.devargs, &conf)) {
499 ERROR("failed to process device arguments");
503 /* Use all ports when none are defined */
504 if (!conf.ports.enabled)
505 conf.ports.enabled = conf.ports.present;
506 for (i = 0; i < device_attr.phys_port_cnt; i++) {
507 uint32_t port = i + 1; /* ports are indexed from one */
508 struct ibv_context *ctx = NULL;
509 struct ibv_port_attr port_attr;
510 struct ibv_pd *pd = NULL;
511 struct priv *priv = NULL;
512 struct rte_eth_dev *eth_dev = NULL;
513 struct ether_addr mac;
515 /* If port is not enabled, skip. */
516 if (!(conf.ports.enabled & (1 << i)))
518 DEBUG("using port %u", port);
519 ctx = ibv_open_device(ibv_dev);
524 /* Check port status. */
525 err = ibv_query_port(ctx, port, &port_attr);
528 ERROR("port query failed: %s", strerror(rte_errno));
531 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
533 ERROR("port %d is not configured in Ethernet mode",
537 if (port_attr.state != IBV_PORT_ACTIVE)
538 DEBUG("port %d is not active: \"%s\" (%d)",
539 port, ibv_port_state_str(port_attr.state),
541 /* Make asynchronous FD non-blocking to handle interrupts. */
542 if (mlx4_fd_set_non_blocking(ctx->async_fd) < 0) {
543 ERROR("cannot make asynchronous FD non-blocking: %s",
544 strerror(rte_errno));
547 /* Allocate protection domain. */
548 pd = ibv_alloc_pd(ctx);
551 ERROR("PD allocation failure");
554 /* from rte_ethdev.c */
555 priv = rte_zmalloc("ethdev private structure",
557 RTE_CACHE_LINE_SIZE);
560 ERROR("priv allocation failure");
564 priv->device_attr = device_attr;
567 priv->mtu = ETHER_MTU;
569 priv->hw_csum = !!(device_attr.device_cap_flags &
570 IBV_DEVICE_RAW_IP_CSUM);
571 DEBUG("checksum offloading is %ssupported",
572 (priv->hw_csum ? "" : "not "));
573 /* Only ConnectX-3 Pro supports tunneling. */
574 priv->hw_csum_l2tun =
576 (device_attr.vendor_part_id ==
577 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
578 DEBUG("L2 tunnel checksum offloads are %ssupported",
579 (priv->hw_csum_l2tun ? "" : "not "));
580 /* Configure the first MAC address by default. */
581 if (mlx4_get_mac(priv, &mac.addr_bytes)) {
582 ERROR("cannot get MAC address, is mlx4_en loaded?"
583 " (rte_errno: %s)", strerror(rte_errno));
586 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
588 mac.addr_bytes[0], mac.addr_bytes[1],
589 mac.addr_bytes[2], mac.addr_bytes[3],
590 mac.addr_bytes[4], mac.addr_bytes[5]);
591 /* Register MAC address. */
595 char ifname[IF_NAMESIZE];
597 if (mlx4_get_ifname(priv, &ifname) == 0)
598 DEBUG("port %u ifname is \"%s\"",
601 DEBUG("port %u ifname is unknown", priv->port);
604 /* Get actual MTU if possible. */
605 mlx4_mtu_get(priv, &priv->mtu);
606 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
607 /* from rte_ethdev.c */
609 char name[RTE_ETH_NAME_MAX_LEN];
611 snprintf(name, sizeof(name), "%s port %u",
612 ibv_get_device_name(ibv_dev), port);
613 eth_dev = rte_eth_dev_allocate(name);
615 if (eth_dev == NULL) {
616 ERROR("can not allocate rte ethdev");
620 eth_dev->data->dev_private = priv;
621 eth_dev->data->mac_addrs = priv->mac;
622 eth_dev->device = &pci_dev->device;
623 rte_eth_copy_pci_info(eth_dev, pci_dev);
624 eth_dev->device->driver = &mlx4_driver.driver;
625 /* Initialize local interrupt handle for current port. */
626 priv->intr_handle = (struct rte_intr_handle){
628 .type = RTE_INTR_HANDLE_EXT,
631 * Override ethdev interrupt handle pointer with private
632 * handle instead of that of the parent PCI device used by
633 * default. This prevents it from being shared between all
634 * ports of the same PCI device since each of them is
635 * associated its own Verbs context.
637 * Rx interrupts in particular require this as the PMD has
638 * no control over the registration of queue interrupts
639 * besides setting up eth_dev->intr_handle, the rest is
640 * handled by rte_intr_rx_ctl().
642 eth_dev->intr_handle = &priv->intr_handle;
644 eth_dev->dev_ops = &mlx4_dev_ops;
645 eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
646 /* Bring Ethernet device up. */
647 DEBUG("forcing Ethernet interface up");
648 mlx4_dev_set_link_up(priv->dev);
649 /* Update link status once if waiting for LSC. */
650 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
651 mlx4_link_update(eth_dev, 0);
656 claim_zero(ibv_dealloc_pd(pd));
658 claim_zero(ibv_close_device(ctx));
660 rte_eth_dev_release_port(eth_dev);
663 if (i == device_attr.phys_port_cnt)
666 * XXX if something went wrong in the loop above, there is a resource
667 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
668 * long as the dpdk does not provide a way to deallocate a ethdev and a
669 * way to enumerate the registered ethdevs to free the previous ones.
673 claim_zero(ibv_close_device(attr_ctx));
675 ibv_free_device_list(list);
676 assert(rte_errno >= 0);
680 static const struct rte_pci_id mlx4_pci_id_map[] = {
682 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
683 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
686 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
687 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
690 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
691 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
698 static struct rte_pci_driver mlx4_driver = {
700 .name = MLX4_DRIVER_NAME
702 .id_table = mlx4_pci_id_map,
703 .probe = mlx4_pci_probe,
704 .drv_flags = RTE_PCI_DRV_INTR_LSC |
705 RTE_PCI_DRV_INTR_RMV,
709 * Driver initialization routine.
711 RTE_INIT(rte_mlx4_pmd_init);
713 rte_mlx4_pmd_init(void)
716 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
717 * huge pages. Calling ibv_fork_init() during init allows
718 * applications to use fork() safely for purposes other than
719 * using this PMD, which is not supported in forked processes.
721 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
723 rte_pci_register(&mlx4_driver);
726 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
727 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
728 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
729 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");