4 * Copyright 2015 6WIND S.A.
5 * Copyright 2015 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.
43 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
45 #pragma GCC diagnostic ignored "-pedantic"
47 #include <infiniband/verbs.h>
49 #pragma GCC diagnostic error "-pedantic"
52 /* DPDK headers don't like -pedantic. */
54 #pragma GCC diagnostic ignored "-pedantic"
56 #include <rte_malloc.h>
57 #include <rte_ethdev.h>
59 #include <rte_common.h>
61 #pragma GCC diagnostic error "-pedantic"
65 #include "mlx5_utils.h"
66 #include "mlx5_rxtx.h"
67 #include "mlx5_autoconf.h"
68 #include "mlx5_defs.h"
71 * DPDK callback to close the device.
73 * Destroy all queues and objects, free memory.
76 * Pointer to Ethernet device structure.
79 mlx5_dev_close(struct rte_eth_dev *dev)
81 struct priv *priv = dev->data->dev_private;
86 DEBUG("%p: closing device \"%s\"",
88 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
89 /* In case mlx5_dev_stop() has not been called. */
90 priv_dev_interrupt_handler_uninstall(priv, dev);
91 priv_special_flow_disable_all(priv);
92 priv_mac_addrs_disable(priv);
93 priv_destroy_hash_rxqs(priv);
95 /* Remove flow director elements. */
96 priv_fdir_disable(priv);
97 priv_fdir_delete_filters_list(priv);
99 /* Prevent crashes when queues are still in use. */
100 dev->rx_pkt_burst = removed_rx_burst;
101 dev->tx_pkt_burst = removed_tx_burst;
102 if (priv->rxqs != NULL) {
103 /* XXX race condition if mlx5_rx_burst() is still running. */
105 for (i = 0; (i != priv->rxqs_n); ++i) {
106 tmp = (*priv->rxqs)[i];
109 (*priv->rxqs)[i] = NULL;
116 if (priv->txqs != NULL) {
117 /* XXX race condition if mlx5_tx_burst() is still running. */
119 for (i = 0; (i != priv->txqs_n); ++i) {
120 tmp = (*priv->txqs)[i];
123 (*priv->txqs)[i] = NULL;
130 if (priv->pd != NULL) {
131 assert(priv->ctx != NULL);
132 claim_zero(ibv_dealloc_pd(priv->pd));
133 claim_zero(ibv_close_device(priv->ctx));
135 assert(priv->ctx == NULL);
136 if (priv->rss_conf != NULL) {
137 for (i = 0; (i != hash_rxq_init_n); ++i)
138 rte_free((*priv->rss_conf)[i]);
139 rte_free(priv->rss_conf);
141 if (priv->reta_idx != NULL)
142 rte_free(priv->reta_idx);
144 memset(priv, 0, sizeof(*priv));
147 static const struct eth_dev_ops mlx5_dev_ops = {
148 .dev_configure = mlx5_dev_configure,
149 .dev_start = mlx5_dev_start,
150 .dev_stop = mlx5_dev_stop,
151 .dev_close = mlx5_dev_close,
152 .promiscuous_enable = mlx5_promiscuous_enable,
153 .promiscuous_disable = mlx5_promiscuous_disable,
154 .allmulticast_enable = mlx5_allmulticast_enable,
155 .allmulticast_disable = mlx5_allmulticast_disable,
156 .link_update = mlx5_link_update,
157 .stats_get = mlx5_stats_get,
158 .stats_reset = mlx5_stats_reset,
159 .dev_infos_get = mlx5_dev_infos_get,
160 .vlan_filter_set = mlx5_vlan_filter_set,
161 .rx_queue_setup = mlx5_rx_queue_setup,
162 .tx_queue_setup = mlx5_tx_queue_setup,
163 .rx_queue_release = mlx5_rx_queue_release,
164 .tx_queue_release = mlx5_tx_queue_release,
165 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
166 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
167 .mac_addr_remove = mlx5_mac_addr_remove,
168 .mac_addr_add = mlx5_mac_addr_add,
169 .mac_addr_set = mlx5_mac_addr_set,
170 .mtu_set = mlx5_dev_set_mtu,
171 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
172 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
173 .vlan_offload_set = mlx5_vlan_offload_set,
174 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
175 .reta_update = mlx5_dev_rss_reta_update,
176 .reta_query = mlx5_dev_rss_reta_query,
177 .rss_hash_update = mlx5_rss_hash_update,
178 .rss_hash_conf_get = mlx5_rss_hash_conf_get,
179 #ifdef MLX5_FDIR_SUPPORT
180 .filter_ctrl = mlx5_dev_filter_ctrl,
181 #endif /* MLX5_FDIR_SUPPORT */
185 struct rte_pci_addr pci_addr; /* associated PCI address */
186 uint32_t ports; /* physical ports bitfield. */
190 * Get device index in mlx5_dev[] from PCI bus address.
192 * @param[in] pci_addr
193 * PCI bus address to look for.
196 * mlx5_dev[] index on success, -1 on failure.
199 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
204 assert(pci_addr != NULL);
205 for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
206 if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
207 (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
208 (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
209 (mlx5_dev[i].pci_addr.function == pci_addr->function))
211 if ((mlx5_dev[i].ports == 0) && (ret == -1))
217 static struct eth_driver mlx5_driver;
220 * DPDK callback to register a PCI device.
222 * This function creates an Ethernet device for each port of a given
226 * PCI driver structure (mlx5_driver).
228 * PCI device information.
231 * 0 on success, negative errno value on failure.
234 mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
236 struct ibv_device **list;
237 struct ibv_device *ibv_dev;
239 struct ibv_context *attr_ctx = NULL;
240 struct ibv_device_attr device_attr;
246 assert(pci_drv == &mlx5_driver.pci_drv);
247 /* Get mlx5_dev[] index. */
248 idx = mlx5_dev_idx(&pci_dev->addr);
250 ERROR("this driver cannot support any more adapters");
253 DEBUG("using driver device index %d", idx);
255 /* Save PCI address. */
256 mlx5_dev[idx].pci_addr = pci_dev->addr;
257 list = ibv_get_device_list(&i);
260 if (errno == ENOSYS) {
261 WARN("cannot list devices, is ib_uverbs loaded?");
268 * For each listed device, check related sysfs entry against
269 * the provided PCI ID.
272 struct rte_pci_addr pci_addr;
275 DEBUG("checking device \"%s\"", list[i]->name);
276 if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
278 if ((pci_dev->addr.domain != pci_addr.domain) ||
279 (pci_dev->addr.bus != pci_addr.bus) ||
280 (pci_dev->addr.devid != pci_addr.devid) ||
281 (pci_dev->addr.function != pci_addr.function))
283 vf = ((pci_dev->id.device_id ==
284 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
285 (pci_dev->id.device_id ==
286 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF));
287 INFO("PCI information matches, using device \"%s\" (VF: %s)",
288 list[i]->name, (vf ? "true" : "false"));
289 attr_ctx = ibv_open_device(list[i]);
293 if (attr_ctx == NULL) {
294 ibv_free_device_list(list);
297 WARN("cannot access device, is mlx5_ib loaded?");
300 WARN("cannot use device, are drivers up to date?");
308 DEBUG("device opened");
309 if (ibv_query_device(attr_ctx, &device_attr))
311 INFO("%u port(s) detected", device_attr.phys_port_cnt);
313 for (i = 0; i < device_attr.phys_port_cnt; i++) {
314 uint32_t port = i + 1; /* ports are indexed from one */
315 uint32_t test = (1 << i);
316 struct ibv_context *ctx = NULL;
317 struct ibv_port_attr port_attr;
318 struct ibv_pd *pd = NULL;
319 struct priv *priv = NULL;
320 struct rte_eth_dev *eth_dev;
321 #ifdef HAVE_EXP_QUERY_DEVICE
322 struct ibv_exp_device_attr exp_device_attr;
323 #endif /* HAVE_EXP_QUERY_DEVICE */
324 struct ether_addr mac;
326 #ifdef HAVE_EXP_QUERY_DEVICE
327 exp_device_attr.comp_mask =
328 IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
329 IBV_EXP_DEVICE_ATTR_RX_HASH |
330 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
331 IBV_EXP_DEVICE_ATTR_VLAN_OFFLOADS |
332 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
334 #endif /* HAVE_EXP_QUERY_DEVICE */
336 DEBUG("using port %u (%08" PRIx32 ")", port, test);
338 ctx = ibv_open_device(ibv_dev);
342 /* Check port status. */
343 err = ibv_query_port(ctx, port, &port_attr);
345 ERROR("port query failed: %s", strerror(err));
349 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
350 ERROR("port %d is not configured in Ethernet mode",
355 if (port_attr.state != IBV_PORT_ACTIVE)
356 DEBUG("port %d is not active: \"%s\" (%d)",
357 port, ibv_port_state_str(port_attr.state),
360 /* Allocate protection domain. */
361 pd = ibv_alloc_pd(ctx);
363 ERROR("PD allocation failure");
368 mlx5_dev[idx].ports |= test;
370 /* from rte_ethdev.c */
371 priv = rte_zmalloc("ethdev private structure",
373 RTE_CACHE_LINE_SIZE);
375 ERROR("priv allocation failure");
381 priv->device_attr = device_attr;
384 priv->mtu = ETHER_MTU;
385 #ifdef HAVE_EXP_QUERY_DEVICE
386 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
387 ERROR("ibv_exp_query_device() failed");
392 ((exp_device_attr.exp_device_cap_flags &
393 IBV_EXP_DEVICE_RX_CSUM_TCP_UDP_PKT) &&
394 (exp_device_attr.exp_device_cap_flags &
395 IBV_EXP_DEVICE_RX_CSUM_IP_PKT));
396 DEBUG("checksum offloading is %ssupported",
397 (priv->hw_csum ? "" : "not "));
399 priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
400 IBV_EXP_DEVICE_VXLAN_SUPPORT);
401 DEBUG("L2 tunnel checksum offloads are %ssupported",
402 (priv->hw_csum_l2tun ? "" : "not "));
404 priv->ind_table_max_size = exp_device_attr.rx_hash_caps.max_rwq_indirection_table_size;
405 /* Remove this check once DPDK supports larger/variable
406 * indirection tables. */
407 if (priv->ind_table_max_size > (unsigned int)RSS_INDIRECTION_TABLE_SIZE)
408 priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
409 DEBUG("maximum RX indirection table size is %u",
410 priv->ind_table_max_size);
411 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
412 priv->hw_vlan_strip = !!(exp_device_attr.wq_vlan_offloads_cap &
413 IBV_EXP_RECEIVE_WQ_CVLAN_STRIP);
414 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
415 DEBUG("VLAN stripping is %ssupported",
416 (priv->hw_vlan_strip ? "" : "not "));
418 #else /* HAVE_EXP_QUERY_DEVICE */
419 priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
420 #endif /* HAVE_EXP_QUERY_DEVICE */
423 /* Allocate and register default RSS hash keys. */
424 priv->rss_conf = rte_calloc(__func__, hash_rxq_init_n,
425 sizeof((*priv->rss_conf)[0]), 0);
426 if (priv->rss_conf == NULL) {
430 err = rss_hash_rss_conf_new_key(priv,
431 rss_hash_default_key,
432 rss_hash_default_key_len,
436 /* Configure the first MAC address by default. */
437 if (priv_get_mac(priv, &mac.addr_bytes)) {
438 ERROR("cannot get MAC address, is mlx5_en loaded?"
439 " (errno: %s)", strerror(errno));
442 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
444 mac.addr_bytes[0], mac.addr_bytes[1],
445 mac.addr_bytes[2], mac.addr_bytes[3],
446 mac.addr_bytes[4], mac.addr_bytes[5]);
447 /* Register MAC address. */
448 claim_zero(priv_mac_addr_add(priv, 0,
449 (const uint8_t (*)[ETHER_ADDR_LEN])
451 /* Initialize FD filters list. */
452 err = fdir_init_filters_list(priv);
457 char ifname[IF_NAMESIZE];
459 if (priv_get_ifname(priv, &ifname) == 0)
460 DEBUG("port %u ifname is \"%s\"",
463 DEBUG("port %u ifname is unknown", priv->port);
466 /* Get actual MTU if possible. */
467 priv_get_mtu(priv, &priv->mtu);
468 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
470 /* from rte_ethdev.c */
472 char name[RTE_ETH_NAME_MAX_LEN];
474 snprintf(name, sizeof(name), "%s port %u",
475 ibv_get_device_name(ibv_dev), port);
476 eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
478 if (eth_dev == NULL) {
479 ERROR("can not allocate rte ethdev");
484 eth_dev->data->dev_private = priv;
485 eth_dev->pci_dev = pci_dev;
487 rte_eth_copy_pci_info(eth_dev, pci_dev);
489 eth_dev->driver = &mlx5_driver;
490 eth_dev->data->rx_mbuf_alloc_failed = 0;
491 eth_dev->data->mtu = ETHER_MTU;
494 eth_dev->dev_ops = &mlx5_dev_ops;
495 eth_dev->data->mac_addrs = priv->mac;
496 TAILQ_INIT(ð_dev->link_intr_cbs);
498 /* Bring Ethernet device up. */
499 DEBUG("forcing Ethernet interface up");
500 priv_set_flags(priv, ~IFF_UP, IFF_UP);
505 rte_free(priv->rss_conf);
509 claim_zero(ibv_dealloc_pd(pd));
511 claim_zero(ibv_close_device(ctx));
516 * XXX if something went wrong in the loop above, there is a resource
517 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
518 * long as the dpdk does not provide a way to deallocate a ethdev and a
519 * way to enumerate the registered ethdevs to free the previous ones.
522 /* no port found, complain */
523 if (!mlx5_dev[idx].ports) {
530 claim_zero(ibv_close_device(attr_ctx));
532 ibv_free_device_list(list);
537 static const struct rte_pci_id mlx5_pci_id_map[] = {
539 .vendor_id = PCI_VENDOR_ID_MELLANOX,
540 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4,
541 .subsystem_vendor_id = PCI_ANY_ID,
542 .subsystem_device_id = PCI_ANY_ID
545 .vendor_id = PCI_VENDOR_ID_MELLANOX,
546 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4VF,
547 .subsystem_vendor_id = PCI_ANY_ID,
548 .subsystem_device_id = PCI_ANY_ID
551 .vendor_id = PCI_VENDOR_ID_MELLANOX,
552 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4LX,
553 .subsystem_vendor_id = PCI_ANY_ID,
554 .subsystem_device_id = PCI_ANY_ID
557 .vendor_id = PCI_VENDOR_ID_MELLANOX,
558 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF,
559 .subsystem_vendor_id = PCI_ANY_ID,
560 .subsystem_device_id = PCI_ANY_ID
567 static struct eth_driver mlx5_driver = {
569 .name = MLX5_DRIVER_NAME,
570 .id_table = mlx5_pci_id_map,
571 .devinit = mlx5_pci_devinit,
572 .drv_flags = RTE_PCI_DRV_INTR_LSC,
574 .dev_private_size = sizeof(struct priv)
578 * Driver initialization routine.
581 rte_mlx5_pmd_init(const char *name, const char *args)
586 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
587 * huge pages. Calling ibv_fork_init() during init allows
588 * applications to use fork() safely for purposes other than
589 * using this PMD, which is not supported in forked processes.
591 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
593 rte_eal_pci_register(&mlx5_driver.pci_drv);
597 static struct rte_driver rte_mlx5_driver = {
599 .name = MLX5_DRIVER_NAME,
600 .init = rte_mlx5_pmd_init,
603 PMD_REGISTER_DRIVER(rte_mlx5_driver)