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.
49 /* Verbs headers do not support -pedantic. */
51 #pragma GCC diagnostic ignored "-Wpedantic"
53 #include <infiniband/verbs.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>
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.
91 * Prepare the driver for a given number of TX and RX queues.
94 * Pointer to Ethernet device structure.
97 * 0 on success, negative errno value otherwise and rte_errno is set.
100 mlx4_dev_configure(struct rte_eth_dev *dev)
102 struct priv *priv = dev->data->dev_private;
103 unsigned int rxqs_n = dev->data->nb_rx_queues;
104 unsigned int txqs_n = dev->data->nb_tx_queues;
106 priv->rxqs = (void *)dev->data->rx_queues;
107 priv->txqs = (void *)dev->data->tx_queues;
108 if (txqs_n != priv->txqs_n) {
109 INFO("%p: TX queues number update: %u -> %u",
110 (void *)dev, priv->txqs_n, txqs_n);
111 priv->txqs_n = txqs_n;
113 if (rxqs_n != priv->rxqs_n) {
114 INFO("%p: Rx queues number update: %u -> %u",
115 (void *)dev, priv->rxqs_n, rxqs_n);
116 priv->rxqs_n = rxqs_n;
122 * DPDK callback to start the device.
124 * Simulate device start by attaching all configured flows.
127 * Pointer to Ethernet device structure.
130 * 0 on success, negative errno value otherwise and rte_errno is set.
133 mlx4_dev_start(struct rte_eth_dev *dev)
135 struct priv *priv = dev->data->dev_private;
140 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
142 ret = mlx4_mac_addr_add(priv);
145 ret = mlx4_intr_install(priv);
147 ERROR("%p: interrupt handler installation failed",
151 ret = mlx4_flow_start(priv);
153 ERROR("%p: flow start failed: %s",
154 (void *)dev, strerror(ret));
160 mlx4_mac_addr_del(priv);
166 * DPDK callback to stop the device.
168 * Simulate device stop by detaching all configured flows.
171 * Pointer to Ethernet device structure.
174 mlx4_dev_stop(struct rte_eth_dev *dev)
176 struct priv *priv = dev->data->dev_private;
180 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
182 mlx4_flow_stop(priv);
183 mlx4_intr_uninstall(priv);
184 mlx4_mac_addr_del(priv);
188 * DPDK callback to close the device.
190 * Destroy all queues and objects, free memory.
193 * Pointer to Ethernet device structure.
196 mlx4_dev_close(struct rte_eth_dev *dev)
198 struct priv *priv = dev->data->dev_private;
204 DEBUG("%p: closing device \"%s\"",
206 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
207 mlx4_mac_addr_del(priv);
209 * Prevent crashes when queues are still in use. This is unfortunately
210 * still required for DPDK 1.3 because some programs (such as testpmd)
211 * never release them before closing the device.
213 dev->rx_pkt_burst = mlx4_rx_burst_removed;
214 dev->tx_pkt_burst = mlx4_tx_burst_removed;
215 if (priv->rxqs != NULL) {
216 /* XXX race condition if mlx4_rx_burst() is still running. */
218 for (i = 0; (i != priv->rxqs_n); ++i) {
219 tmp = (*priv->rxqs)[i];
222 (*priv->rxqs)[i] = NULL;
223 mlx4_rxq_cleanup(tmp);
229 if (priv->txqs != NULL) {
230 /* XXX race condition if mlx4_tx_burst() is still running. */
232 for (i = 0; (i != priv->txqs_n); ++i) {
233 tmp = (*priv->txqs)[i];
236 (*priv->txqs)[i] = NULL;
237 mlx4_txq_cleanup(tmp);
243 if (priv->pd != NULL) {
244 assert(priv->ctx != NULL);
245 claim_zero(ibv_dealloc_pd(priv->pd));
246 claim_zero(ibv_close_device(priv->ctx));
248 assert(priv->ctx == NULL);
249 mlx4_intr_uninstall(priv);
250 memset(priv, 0, sizeof(*priv));
253 static const struct eth_dev_ops mlx4_dev_ops = {
254 .dev_configure = mlx4_dev_configure,
255 .dev_start = mlx4_dev_start,
256 .dev_stop = mlx4_dev_stop,
257 .dev_set_link_down = mlx4_dev_set_link_down,
258 .dev_set_link_up = mlx4_dev_set_link_up,
259 .dev_close = mlx4_dev_close,
260 .link_update = mlx4_link_update,
261 .stats_get = mlx4_stats_get,
262 .stats_reset = mlx4_stats_reset,
263 .dev_infos_get = mlx4_dev_infos_get,
264 .rx_queue_setup = mlx4_rx_queue_setup,
265 .tx_queue_setup = mlx4_tx_queue_setup,
266 .rx_queue_release = mlx4_rx_queue_release,
267 .tx_queue_release = mlx4_tx_queue_release,
268 .flow_ctrl_get = mlx4_flow_ctrl_get,
269 .flow_ctrl_set = mlx4_flow_ctrl_set,
270 .mtu_set = mlx4_mtu_set,
271 .filter_ctrl = mlx4_filter_ctrl,
272 .rx_queue_intr_enable = mlx4_rx_intr_enable,
273 .rx_queue_intr_disable = mlx4_rx_intr_disable,
277 * Get PCI information from struct ibv_device.
280 * Pointer to Ethernet device structure.
281 * @param[out] pci_addr
282 * PCI bus address output buffer.
285 * 0 on success, negative errno value otherwise and rte_errno is set.
288 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
289 struct rte_pci_addr *pci_addr)
293 MKSTR(path, "%s/device/uevent", device->ibdev_path);
295 file = fopen(path, "rb");
300 while (fgets(line, sizeof(line), file) == line) {
301 size_t len = strlen(line);
304 /* Truncate long lines. */
305 if (len == (sizeof(line) - 1))
306 while (line[(len - 1)] != '\n') {
310 line[(len - 1)] = ret;
312 /* Extract information. */
315 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
319 &pci_addr->function) == 4) {
329 * Verify and store value for device argument.
332 * Key argument to verify.
334 * Value associated with key.
335 * @param[in, out] conf
336 * Shared configuration data.
339 * 0 on success, negative errno value otherwise and rte_errno is set.
342 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
347 tmp = strtoul(val, NULL, 0);
350 WARN("%s: \"%s\" is not a valid integer", key, val);
353 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
354 uint32_t ports = rte_log2_u32(conf->ports.present);
357 ERROR("port index %lu outside range [0,%" PRIu32 ")",
361 if (!(conf->ports.present & (1 << tmp))) {
363 ERROR("invalid port index %lu", tmp);
366 conf->ports.enabled |= 1 << tmp;
369 WARN("%s: unknown parameter", key);
376 * Parse device parameters.
379 * Device arguments structure.
382 * 0 on success, negative errno value otherwise and rte_errno is set.
385 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
387 struct rte_kvargs *kvlist;
388 unsigned int arg_count;
394 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
395 if (kvlist == NULL) {
397 ERROR("failed to parse kvargs");
400 /* Process parameters. */
401 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
402 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
403 while (arg_count-- > 0) {
404 ret = rte_kvargs_process(kvlist,
406 (int (*)(const char *,
416 rte_kvargs_free(kvlist);
420 static struct rte_pci_driver mlx4_driver;
423 * DPDK callback to register a PCI device.
425 * This function creates an Ethernet device for each port of a given
429 * PCI driver structure (mlx4_driver).
431 * PCI device information.
434 * 0 on success, negative errno value otherwise and rte_errno is set.
437 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
439 struct ibv_device **list;
440 struct ibv_device *ibv_dev;
442 struct ibv_context *attr_ctx = NULL;
443 struct ibv_device_attr device_attr;
444 struct mlx4_conf conf = {
451 assert(pci_drv == &mlx4_driver);
452 list = ibv_get_device_list(&i);
456 if (rte_errno == ENOSYS)
457 ERROR("cannot list devices, is ib_uverbs loaded?");
462 * For each listed device, check related sysfs entry against
463 * the provided PCI ID.
466 struct rte_pci_addr pci_addr;
469 DEBUG("checking device \"%s\"", list[i]->name);
470 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
472 if ((pci_dev->addr.domain != pci_addr.domain) ||
473 (pci_dev->addr.bus != pci_addr.bus) ||
474 (pci_dev->addr.devid != pci_addr.devid) ||
475 (pci_dev->addr.function != pci_addr.function))
477 vf = (pci_dev->id.device_id ==
478 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
479 INFO("PCI information matches, using device \"%s\" (VF: %s)",
480 list[i]->name, (vf ? "true" : "false"));
481 attr_ctx = ibv_open_device(list[i]);
485 if (attr_ctx == NULL) {
486 ibv_free_device_list(list);
490 ERROR("cannot access device, is mlx4_ib loaded?");
494 ERROR("cannot use device, are drivers up to date?");
502 DEBUG("device opened");
503 if (ibv_query_device(attr_ctx, &device_attr)) {
507 INFO("%u port(s) detected", device_attr.phys_port_cnt);
508 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
509 if (mlx4_args(pci_dev->device.devargs, &conf)) {
510 ERROR("failed to process device arguments");
514 /* Use all ports when none are defined */
515 if (!conf.ports.enabled)
516 conf.ports.enabled = conf.ports.present;
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 = ibv_open_device(ibv_dev);
535 /* Check port status. */
536 err = ibv_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, ibv_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 = ibv_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 /* 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. */
593 if (mlx4_mac_addr_add(priv))
597 char ifname[IF_NAMESIZE];
599 if (mlx4_get_ifname(priv, &ifname) == 0)
600 DEBUG("port %u ifname is \"%s\"",
603 DEBUG("port %u ifname is unknown", priv->port);
606 /* Get actual MTU if possible. */
607 mlx4_mtu_get(priv, &priv->mtu);
608 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
609 /* from rte_ethdev.c */
611 char name[RTE_ETH_NAME_MAX_LEN];
613 snprintf(name, sizeof(name), "%s port %u",
614 ibv_get_device_name(ibv_dev), port);
615 eth_dev = rte_eth_dev_allocate(name);
617 if (eth_dev == NULL) {
618 ERROR("can not allocate rte ethdev");
622 eth_dev->data->dev_private = priv;
623 eth_dev->data->mac_addrs = &priv->mac;
624 eth_dev->device = &pci_dev->device;
625 rte_eth_copy_pci_info(eth_dev, pci_dev);
626 eth_dev->device->driver = &mlx4_driver.driver;
627 /* Initialize local interrupt handle for current port. */
628 priv->intr_handle = (struct rte_intr_handle){
630 .type = RTE_INTR_HANDLE_EXT,
633 * Override ethdev interrupt handle pointer with private
634 * handle instead of that of the parent PCI device used by
635 * default. This prevents it from being shared between all
636 * ports of the same PCI device since each of them is
637 * associated its own Verbs context.
639 * Rx interrupts in particular require this as the PMD has
640 * no control over the registration of queue interrupts
641 * besides setting up eth_dev->intr_handle, the rest is
642 * handled by rte_intr_rx_ctl().
644 eth_dev->intr_handle = &priv->intr_handle;
646 eth_dev->dev_ops = &mlx4_dev_ops;
647 eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
648 /* Bring Ethernet device up. */
649 DEBUG("forcing Ethernet interface up");
650 mlx4_dev_set_link_up(priv->dev);
651 /* Update link status once if waiting for LSC. */
652 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
653 mlx4_link_update(eth_dev, 0);
658 claim_zero(ibv_dealloc_pd(pd));
660 claim_zero(ibv_close_device(ctx));
662 rte_eth_dev_release_port(eth_dev);
665 if (i == device_attr.phys_port_cnt)
668 * XXX if something went wrong in the loop above, there is a resource
669 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
670 * long as the dpdk does not provide a way to deallocate a ethdev and a
671 * way to enumerate the registered ethdevs to free the previous ones.
675 claim_zero(ibv_close_device(attr_ctx));
677 ibv_free_device_list(list);
678 assert(rte_errno >= 0);
682 static const struct rte_pci_id mlx4_pci_id_map[] = {
684 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
685 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
688 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
689 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
692 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
693 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
700 static struct rte_pci_driver mlx4_driver = {
702 .name = MLX4_DRIVER_NAME
704 .id_table = mlx4_pci_id_map,
705 .probe = mlx4_pci_probe,
706 .drv_flags = RTE_PCI_DRV_INTR_LSC |
707 RTE_PCI_DRV_INTR_RMV,
711 * Driver initialization routine.
713 RTE_INIT(rte_mlx4_pmd_init);
715 rte_mlx4_pmd_init(void)
718 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
719 * huge pages. Calling ibv_fork_init() during init allows
720 * applications to use fork() safely for purposes other than
721 * using this PMD, which is not supported in forked processes.
723 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
725 rte_pci_register(&mlx4_driver);
728 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
729 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
730 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
731 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");