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.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 attaching all configured flows.
121 * Pointer to Ethernet device structure.
124 * 0 on success, negative errno value otherwise and rte_errno is set.
127 mlx4_dev_start(struct rte_eth_dev *dev)
129 struct priv *priv = dev->data->dev_private;
130 struct rte_flow_error error;
135 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
137 ret = mlx4_intr_install(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;
163 * DPDK callback to stop the device.
165 * Simulate device stop by detaching all configured flows.
168 * Pointer to Ethernet device structure.
171 mlx4_dev_stop(struct rte_eth_dev *dev)
173 struct priv *priv = dev->data->dev_private;
177 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
179 dev->tx_pkt_burst = mlx4_tx_burst_removed;
180 dev->rx_pkt_burst = mlx4_rx_burst_removed;
182 mlx4_flow_sync(priv, NULL);
183 mlx4_intr_uninstall(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 for (i = 0; i != dev->data->nb_rx_queues; ++i)
208 mlx4_rx_queue_release(dev->data->rx_queues[i]);
209 for (i = 0; i != dev->data->nb_tx_queues; ++i)
210 mlx4_tx_queue_release(dev->data->tx_queues[i]);
211 if (priv->pd != NULL) {
212 assert(priv->ctx != NULL);
213 claim_zero(ibv_dealloc_pd(priv->pd));
214 claim_zero(ibv_close_device(priv->ctx));
216 assert(priv->ctx == NULL);
217 mlx4_intr_uninstall(priv);
218 memset(priv, 0, sizeof(*priv));
221 static const struct eth_dev_ops mlx4_dev_ops = {
222 .dev_configure = mlx4_dev_configure,
223 .dev_start = mlx4_dev_start,
224 .dev_stop = mlx4_dev_stop,
225 .dev_set_link_down = mlx4_dev_set_link_down,
226 .dev_set_link_up = mlx4_dev_set_link_up,
227 .dev_close = mlx4_dev_close,
228 .link_update = mlx4_link_update,
229 .promiscuous_enable = mlx4_promiscuous_enable,
230 .promiscuous_disable = mlx4_promiscuous_disable,
231 .allmulticast_enable = mlx4_allmulticast_enable,
232 .allmulticast_disable = mlx4_allmulticast_disable,
233 .mac_addr_remove = mlx4_mac_addr_remove,
234 .mac_addr_add = mlx4_mac_addr_add,
235 .mac_addr_set = mlx4_mac_addr_set,
236 .stats_get = mlx4_stats_get,
237 .stats_reset = mlx4_stats_reset,
238 .dev_infos_get = mlx4_dev_infos_get,
239 .vlan_filter_set = mlx4_vlan_filter_set,
240 .rx_queue_setup = mlx4_rx_queue_setup,
241 .tx_queue_setup = mlx4_tx_queue_setup,
242 .rx_queue_release = mlx4_rx_queue_release,
243 .tx_queue_release = mlx4_tx_queue_release,
244 .flow_ctrl_get = mlx4_flow_ctrl_get,
245 .flow_ctrl_set = mlx4_flow_ctrl_set,
246 .mtu_set = mlx4_mtu_set,
247 .filter_ctrl = mlx4_filter_ctrl,
248 .rx_queue_intr_enable = mlx4_rx_intr_enable,
249 .rx_queue_intr_disable = mlx4_rx_intr_disable,
253 * Get PCI information from struct ibv_device.
256 * Pointer to Ethernet device structure.
257 * @param[out] pci_addr
258 * PCI bus address output buffer.
261 * 0 on success, negative errno value otherwise and rte_errno is set.
264 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
265 struct rte_pci_addr *pci_addr)
269 MKSTR(path, "%s/device/uevent", device->ibdev_path);
271 file = fopen(path, "rb");
276 while (fgets(line, sizeof(line), file) == line) {
277 size_t len = strlen(line);
280 /* Truncate long lines. */
281 if (len == (sizeof(line) - 1))
282 while (line[(len - 1)] != '\n') {
286 line[(len - 1)] = ret;
288 /* Extract information. */
291 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
295 &pci_addr->function) == 4) {
305 * Verify and store value for device argument.
308 * Key argument to verify.
310 * Value associated with key.
311 * @param[in, out] conf
312 * Shared configuration data.
315 * 0 on success, negative errno value otherwise and rte_errno is set.
318 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
323 tmp = strtoul(val, NULL, 0);
326 WARN("%s: \"%s\" is not a valid integer", key, val);
329 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
330 uint32_t ports = rte_log2_u32(conf->ports.present);
333 ERROR("port index %lu outside range [0,%" PRIu32 ")",
337 if (!(conf->ports.present & (1 << tmp))) {
339 ERROR("invalid port index %lu", tmp);
342 conf->ports.enabled |= 1 << tmp;
345 WARN("%s: unknown parameter", key);
352 * Parse device parameters.
355 * Device arguments structure.
358 * 0 on success, negative errno value otherwise and rte_errno is set.
361 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
363 struct rte_kvargs *kvlist;
364 unsigned int arg_count;
370 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
371 if (kvlist == NULL) {
373 ERROR("failed to parse kvargs");
376 /* Process parameters. */
377 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
378 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
379 while (arg_count-- > 0) {
380 ret = rte_kvargs_process(kvlist,
382 (int (*)(const char *,
392 rte_kvargs_free(kvlist);
396 static struct rte_pci_driver mlx4_driver;
399 * DPDK callback to register a PCI device.
401 * This function creates an Ethernet device for each port of a given
405 * PCI driver structure (mlx4_driver).
407 * PCI device information.
410 * 0 on success, negative errno value otherwise and rte_errno is set.
413 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
415 struct ibv_device **list;
416 struct ibv_device *ibv_dev;
418 struct ibv_context *attr_ctx = NULL;
419 struct ibv_device_attr device_attr;
420 struct mlx4_conf conf = {
427 assert(pci_drv == &mlx4_driver);
428 list = ibv_get_device_list(&i);
432 if (rte_errno == ENOSYS)
433 ERROR("cannot list devices, is ib_uverbs loaded?");
438 * For each listed device, check related sysfs entry against
439 * the provided PCI ID.
442 struct rte_pci_addr pci_addr;
445 DEBUG("checking device \"%s\"", list[i]->name);
446 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
448 if ((pci_dev->addr.domain != pci_addr.domain) ||
449 (pci_dev->addr.bus != pci_addr.bus) ||
450 (pci_dev->addr.devid != pci_addr.devid) ||
451 (pci_dev->addr.function != pci_addr.function))
453 vf = (pci_dev->id.device_id ==
454 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
455 INFO("PCI information matches, using device \"%s\" (VF: %s)",
456 list[i]->name, (vf ? "true" : "false"));
457 attr_ctx = ibv_open_device(list[i]);
461 if (attr_ctx == NULL) {
462 ibv_free_device_list(list);
466 ERROR("cannot access device, is mlx4_ib loaded?");
470 ERROR("cannot use device, are drivers up to date?");
478 DEBUG("device opened");
479 if (ibv_query_device(attr_ctx, &device_attr)) {
483 INFO("%u port(s) detected", device_attr.phys_port_cnt);
484 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
485 if (mlx4_args(pci_dev->device.devargs, &conf)) {
486 ERROR("failed to process device arguments");
490 /* Use all ports when none are defined */
491 if (!conf.ports.enabled)
492 conf.ports.enabled = conf.ports.present;
493 for (i = 0; i < device_attr.phys_port_cnt; i++) {
494 uint32_t port = i + 1; /* ports are indexed from one */
495 struct ibv_context *ctx = NULL;
496 struct ibv_port_attr port_attr;
497 struct ibv_pd *pd = NULL;
498 struct priv *priv = NULL;
499 struct rte_eth_dev *eth_dev = NULL;
500 struct ether_addr mac;
502 /* If port is not enabled, skip. */
503 if (!(conf.ports.enabled & (1 << i)))
505 DEBUG("using port %u", port);
506 ctx = ibv_open_device(ibv_dev);
511 /* Check port status. */
512 err = ibv_query_port(ctx, port, &port_attr);
515 ERROR("port query failed: %s", strerror(rte_errno));
518 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
520 ERROR("port %d is not configured in Ethernet mode",
524 if (port_attr.state != IBV_PORT_ACTIVE)
525 DEBUG("port %d is not active: \"%s\" (%d)",
526 port, ibv_port_state_str(port_attr.state),
528 /* Make asynchronous FD non-blocking to handle interrupts. */
529 if (mlx4_fd_set_non_blocking(ctx->async_fd) < 0) {
530 ERROR("cannot make asynchronous FD non-blocking: %s",
531 strerror(rte_errno));
534 /* Allocate protection domain. */
535 pd = ibv_alloc_pd(ctx);
538 ERROR("PD allocation failure");
541 /* from rte_ethdev.c */
542 priv = rte_zmalloc("ethdev private structure",
544 RTE_CACHE_LINE_SIZE);
547 ERROR("priv allocation failure");
551 priv->device_attr = device_attr;
554 priv->mtu = ETHER_MTU;
556 /* Configure the first MAC address by default. */
557 if (mlx4_get_mac(priv, &mac.addr_bytes)) {
558 ERROR("cannot get MAC address, is mlx4_en loaded?"
559 " (rte_errno: %s)", strerror(rte_errno));
562 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
564 mac.addr_bytes[0], mac.addr_bytes[1],
565 mac.addr_bytes[2], mac.addr_bytes[3],
566 mac.addr_bytes[4], mac.addr_bytes[5]);
567 /* Register MAC address. */
571 char ifname[IF_NAMESIZE];
573 if (mlx4_get_ifname(priv, &ifname) == 0)
574 DEBUG("port %u ifname is \"%s\"",
577 DEBUG("port %u ifname is unknown", priv->port);
580 /* Get actual MTU if possible. */
581 mlx4_mtu_get(priv, &priv->mtu);
582 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
583 /* from rte_ethdev.c */
585 char name[RTE_ETH_NAME_MAX_LEN];
587 snprintf(name, sizeof(name), "%s port %u",
588 ibv_get_device_name(ibv_dev), port);
589 eth_dev = rte_eth_dev_allocate(name);
591 if (eth_dev == NULL) {
592 ERROR("can not allocate rte ethdev");
596 eth_dev->data->dev_private = priv;
597 eth_dev->data->mac_addrs = priv->mac;
598 eth_dev->device = &pci_dev->device;
599 rte_eth_copy_pci_info(eth_dev, pci_dev);
600 eth_dev->device->driver = &mlx4_driver.driver;
601 /* Initialize local interrupt handle for current port. */
602 priv->intr_handle = (struct rte_intr_handle){
604 .type = RTE_INTR_HANDLE_EXT,
607 * Override ethdev interrupt handle pointer with private
608 * handle instead of that of the parent PCI device used by
609 * default. This prevents it from being shared between all
610 * ports of the same PCI device since each of them is
611 * associated its own Verbs context.
613 * Rx interrupts in particular require this as the PMD has
614 * no control over the registration of queue interrupts
615 * besides setting up eth_dev->intr_handle, the rest is
616 * handled by rte_intr_rx_ctl().
618 eth_dev->intr_handle = &priv->intr_handle;
620 eth_dev->dev_ops = &mlx4_dev_ops;
621 eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
622 /* Bring Ethernet device up. */
623 DEBUG("forcing Ethernet interface up");
624 mlx4_dev_set_link_up(priv->dev);
625 /* Update link status once if waiting for LSC. */
626 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
627 mlx4_link_update(eth_dev, 0);
632 claim_zero(ibv_dealloc_pd(pd));
634 claim_zero(ibv_close_device(ctx));
636 rte_eth_dev_release_port(eth_dev);
639 if (i == device_attr.phys_port_cnt)
642 * XXX if something went wrong in the loop above, there is a resource
643 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
644 * long as the dpdk does not provide a way to deallocate a ethdev and a
645 * way to enumerate the registered ethdevs to free the previous ones.
649 claim_zero(ibv_close_device(attr_ctx));
651 ibv_free_device_list(list);
652 assert(rte_errno >= 0);
656 static const struct rte_pci_id mlx4_pci_id_map[] = {
658 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
659 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
662 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
663 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
666 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
667 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
674 static struct rte_pci_driver mlx4_driver = {
676 .name = MLX4_DRIVER_NAME
678 .id_table = mlx4_pci_id_map,
679 .probe = mlx4_pci_probe,
680 .drv_flags = RTE_PCI_DRV_INTR_LSC |
681 RTE_PCI_DRV_INTR_RMV,
685 * Driver initialization routine.
687 RTE_INIT(rte_mlx4_pmd_init);
689 rte_mlx4_pmd_init(void)
692 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
693 * huge pages. Calling ibv_fork_init() during init allows
694 * applications to use fork() safely for purposes other than
695 * using this PMD, which is not supported in forked processes.
697 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
699 rte_pci_register(&mlx4_driver);
702 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
703 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
704 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
705 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");