1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
16 #include <linux/rtnetlink.h>
19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
21 #pragma GCC diagnostic ignored "-Wpedantic"
23 #include <infiniband/verbs.h>
25 #pragma GCC diagnostic error "-Wpedantic"
28 #include <rte_malloc.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_ethdev_pci.h>
32 #include <rte_bus_pci.h>
33 #include <rte_common.h>
34 #include <rte_config.h>
35 #include <rte_eal_memconfig.h>
36 #include <rte_kvargs.h>
39 #include "mlx5_utils.h"
40 #include "mlx5_rxtx.h"
41 #include "mlx5_autoconf.h"
42 #include "mlx5_defs.h"
43 #include "mlx5_glue.h"
45 /* Device parameter to enable RX completion queue compression. */
46 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
48 /* Device parameter to configure inline send. */
49 #define MLX5_TXQ_INLINE "txq_inline"
52 * Device parameter to configure the number of TX queues threshold for
53 * enabling inline send.
55 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
57 /* Device parameter to enable multi-packet send WQEs. */
58 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
60 /* Device parameter to include 2 dsegs in the title WQEBB. */
61 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
63 /* Device parameter to limit the size of inlining packet. */
64 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
66 /* Device parameter to enable hardware Tx vector. */
67 #define MLX5_TX_VEC_EN "tx_vec_en"
69 /* Device parameter to enable hardware Rx vector. */
70 #define MLX5_RX_VEC_EN "rx_vec_en"
72 /* Allow L3 VXLAN flow creation. */
73 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
75 /* Activate Netlink support in VF mode. */
76 #define MLX5_VF_NL_EN "vf_nl_en"
78 #ifndef HAVE_IBV_MLX5_MOD_MPW
79 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
80 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
83 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
84 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
87 /** Driver-specific log messages type. */
91 * Retrieve integer value from environment variable.
94 * Environment variable name.
97 * Integer value, 0 if the variable is not set.
100 mlx5_getenv_int(const char *name)
102 const char *val = getenv(name);
110 * Verbs callback to allocate a memory. This function should allocate the space
111 * according to the size provided residing inside a huge page.
112 * Please note that all allocation must respect the alignment from libmlx5
113 * (i.e. currently sysconf(_SC_PAGESIZE)).
116 * The size in bytes of the memory to allocate.
118 * A pointer to the callback data.
121 * Allocated buffer, NULL otherwise and rte_errno is set.
124 mlx5_alloc_verbs_buf(size_t size, void *data)
126 struct priv *priv = data;
128 size_t alignment = sysconf(_SC_PAGESIZE);
129 unsigned int socket = SOCKET_ID_ANY;
131 if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
132 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
134 socket = ctrl->socket;
135 } else if (priv->verbs_alloc_ctx.type ==
136 MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
137 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
139 socket = ctrl->socket;
141 assert(data != NULL);
142 ret = rte_malloc_socket(__func__, size, alignment, socket);
149 * Verbs callback to free a memory.
152 * A pointer to the memory to free.
154 * A pointer to the callback data.
157 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
159 assert(data != NULL);
164 * DPDK callback to close the device.
166 * Destroy all queues and objects, free memory.
169 * Pointer to Ethernet device structure.
172 mlx5_dev_close(struct rte_eth_dev *dev)
174 struct priv *priv = dev->data->dev_private;
178 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
180 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
181 /* In case mlx5_dev_stop() has not been called. */
182 mlx5_dev_interrupt_handler_uninstall(dev);
183 mlx5_traffic_disable(dev);
184 /* Prevent crashes when queues are still in use. */
185 dev->rx_pkt_burst = removed_rx_burst;
186 dev->tx_pkt_burst = removed_tx_burst;
187 if (priv->rxqs != NULL) {
188 /* XXX race condition if mlx5_rx_burst() is still running. */
190 for (i = 0; (i != priv->rxqs_n); ++i)
191 mlx5_rxq_release(dev, i);
195 if (priv->txqs != NULL) {
196 /* XXX race condition if mlx5_tx_burst() is still running. */
198 for (i = 0; (i != priv->txqs_n); ++i)
199 mlx5_txq_release(dev, i);
203 mlx5_flow_delete_drop_queue(dev);
204 if (priv->pd != NULL) {
205 assert(priv->ctx != NULL);
206 claim_zero(mlx5_glue->dealloc_pd(priv->pd));
207 claim_zero(mlx5_glue->close_device(priv->ctx));
209 assert(priv->ctx == NULL);
210 if (priv->rss_conf.rss_key != NULL)
211 rte_free(priv->rss_conf.rss_key);
212 if (priv->reta_idx != NULL)
213 rte_free(priv->reta_idx);
214 if (priv->primary_socket)
215 mlx5_socket_uninit(dev);
217 mlx5_nl_mac_addr_flush(dev);
218 if (priv->nl_socket >= 0)
219 close(priv->nl_socket);
220 ret = mlx5_hrxq_ibv_verify(dev);
222 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
224 ret = mlx5_ind_table_ibv_verify(dev);
226 DRV_LOG(WARNING, "port %u some indirection table still remain",
228 ret = mlx5_rxq_ibv_verify(dev);
230 DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
232 ret = mlx5_rxq_verify(dev);
234 DRV_LOG(WARNING, "port %u some Rx queues still remain",
236 ret = mlx5_txq_ibv_verify(dev);
238 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
240 ret = mlx5_txq_verify(dev);
242 DRV_LOG(WARNING, "port %u some Tx queues still remain",
244 ret = mlx5_flow_verify(dev);
246 DRV_LOG(WARNING, "port %u some flows still remain",
248 memset(priv, 0, sizeof(*priv));
251 const struct eth_dev_ops mlx5_dev_ops = {
252 .dev_configure = mlx5_dev_configure,
253 .dev_start = mlx5_dev_start,
254 .dev_stop = mlx5_dev_stop,
255 .dev_set_link_down = mlx5_set_link_down,
256 .dev_set_link_up = mlx5_set_link_up,
257 .dev_close = mlx5_dev_close,
258 .promiscuous_enable = mlx5_promiscuous_enable,
259 .promiscuous_disable = mlx5_promiscuous_disable,
260 .allmulticast_enable = mlx5_allmulticast_enable,
261 .allmulticast_disable = mlx5_allmulticast_disable,
262 .link_update = mlx5_link_update,
263 .stats_get = mlx5_stats_get,
264 .stats_reset = mlx5_stats_reset,
265 .xstats_get = mlx5_xstats_get,
266 .xstats_reset = mlx5_xstats_reset,
267 .xstats_get_names = mlx5_xstats_get_names,
268 .dev_infos_get = mlx5_dev_infos_get,
269 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
270 .vlan_filter_set = mlx5_vlan_filter_set,
271 .rx_queue_setup = mlx5_rx_queue_setup,
272 .tx_queue_setup = mlx5_tx_queue_setup,
273 .rx_queue_release = mlx5_rx_queue_release,
274 .tx_queue_release = mlx5_tx_queue_release,
275 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
276 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
277 .mac_addr_remove = mlx5_mac_addr_remove,
278 .mac_addr_add = mlx5_mac_addr_add,
279 .mac_addr_set = mlx5_mac_addr_set,
280 .set_mc_addr_list = mlx5_set_mc_addr_list,
281 .mtu_set = mlx5_dev_set_mtu,
282 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
283 .vlan_offload_set = mlx5_vlan_offload_set,
284 .reta_update = mlx5_dev_rss_reta_update,
285 .reta_query = mlx5_dev_rss_reta_query,
286 .rss_hash_update = mlx5_rss_hash_update,
287 .rss_hash_conf_get = mlx5_rss_hash_conf_get,
288 .filter_ctrl = mlx5_dev_filter_ctrl,
289 .rx_descriptor_status = mlx5_rx_descriptor_status,
290 .tx_descriptor_status = mlx5_tx_descriptor_status,
291 .rx_queue_intr_enable = mlx5_rx_intr_enable,
292 .rx_queue_intr_disable = mlx5_rx_intr_disable,
293 .is_removed = mlx5_is_removed,
296 static const struct eth_dev_ops mlx5_dev_sec_ops = {
297 .stats_get = mlx5_stats_get,
298 .stats_reset = mlx5_stats_reset,
299 .xstats_get = mlx5_xstats_get,
300 .xstats_reset = mlx5_xstats_reset,
301 .xstats_get_names = mlx5_xstats_get_names,
302 .dev_infos_get = mlx5_dev_infos_get,
303 .rx_descriptor_status = mlx5_rx_descriptor_status,
304 .tx_descriptor_status = mlx5_tx_descriptor_status,
307 /* Available operators in flow isolated mode. */
308 const struct eth_dev_ops mlx5_dev_ops_isolate = {
309 .dev_configure = mlx5_dev_configure,
310 .dev_start = mlx5_dev_start,
311 .dev_stop = mlx5_dev_stop,
312 .dev_set_link_down = mlx5_set_link_down,
313 .dev_set_link_up = mlx5_set_link_up,
314 .dev_close = mlx5_dev_close,
315 .link_update = mlx5_link_update,
316 .stats_get = mlx5_stats_get,
317 .stats_reset = mlx5_stats_reset,
318 .xstats_get = mlx5_xstats_get,
319 .xstats_reset = mlx5_xstats_reset,
320 .xstats_get_names = mlx5_xstats_get_names,
321 .dev_infos_get = mlx5_dev_infos_get,
322 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
323 .vlan_filter_set = mlx5_vlan_filter_set,
324 .rx_queue_setup = mlx5_rx_queue_setup,
325 .tx_queue_setup = mlx5_tx_queue_setup,
326 .rx_queue_release = mlx5_rx_queue_release,
327 .tx_queue_release = mlx5_tx_queue_release,
328 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
329 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
330 .mac_addr_remove = mlx5_mac_addr_remove,
331 .mac_addr_add = mlx5_mac_addr_add,
332 .mac_addr_set = mlx5_mac_addr_set,
333 .set_mc_addr_list = mlx5_set_mc_addr_list,
334 .mtu_set = mlx5_dev_set_mtu,
335 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
336 .vlan_offload_set = mlx5_vlan_offload_set,
337 .filter_ctrl = mlx5_dev_filter_ctrl,
338 .rx_descriptor_status = mlx5_rx_descriptor_status,
339 .tx_descriptor_status = mlx5_tx_descriptor_status,
340 .rx_queue_intr_enable = mlx5_rx_intr_enable,
341 .rx_queue_intr_disable = mlx5_rx_intr_disable,
342 .is_removed = mlx5_is_removed,
346 struct rte_pci_addr pci_addr; /* associated PCI address */
347 uint32_t ports; /* physical ports bitfield. */
351 * Get device index in mlx5_dev[] from PCI bus address.
353 * @param[in] pci_addr
354 * PCI bus address to look for.
357 * mlx5_dev[] index on success, -1 on failure.
360 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
365 assert(pci_addr != NULL);
366 for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
367 if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
368 (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
369 (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
370 (mlx5_dev[i].pci_addr.function == pci_addr->function))
372 if ((mlx5_dev[i].ports == 0) && (ret == -1))
379 * Verify and store value for device argument.
382 * Key argument to verify.
384 * Value associated with key.
389 * 0 on success, a negative errno value otherwise and rte_errno is set.
392 mlx5_args_check(const char *key, const char *val, void *opaque)
394 struct mlx5_dev_config *config = opaque;
398 tmp = strtoul(val, NULL, 0);
401 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
404 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
405 config->cqe_comp = !!tmp;
406 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
407 config->txq_inline = tmp;
408 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
409 config->txqs_inline = tmp;
410 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
411 config->mps = !!tmp ? config->mps : 0;
412 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
413 config->mpw_hdr_dseg = !!tmp;
414 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
415 config->inline_max_packet_sz = tmp;
416 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
417 config->tx_vec_en = !!tmp;
418 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
419 config->rx_vec_en = !!tmp;
420 } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
421 config->l3_vxlan_en = !!tmp;
422 } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
423 config->vf_nl_en = !!tmp;
425 DRV_LOG(WARNING, "%s: unknown parameter", key);
433 * Parse device parameters.
436 * Pointer to device configuration structure.
438 * Device arguments structure.
441 * 0 on success, a negative errno value otherwise and rte_errno is set.
444 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
446 const char **params = (const char *[]){
447 MLX5_RXQ_CQE_COMP_EN,
449 MLX5_TXQS_MIN_INLINE,
451 MLX5_TXQ_MPW_HDR_DSEG_EN,
452 MLX5_TXQ_MAX_INLINE_LEN,
459 struct rte_kvargs *kvlist;
465 /* Following UGLY cast is done to pass checkpatch. */
466 kvlist = rte_kvargs_parse(devargs->args, params);
469 /* Process parameters. */
470 for (i = 0; (params[i] != NULL); ++i) {
471 if (rte_kvargs_count(kvlist, params[i])) {
472 ret = rte_kvargs_process(kvlist, params[i],
473 mlx5_args_check, config);
476 rte_kvargs_free(kvlist);
481 rte_kvargs_free(kvlist);
485 static struct rte_pci_driver mlx5_driver;
488 * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
489 * local resource used by both primary and secondary to avoid duplicate
491 * The space has to be available on both primary and secondary process,
492 * TXQ UAR maps to this area using fixed mmap w/o double check.
494 static void *uar_base;
497 find_lower_va_bound(const struct rte_memseg_list *msl __rte_unused,
498 const struct rte_memseg *ms, void *arg)
505 *addr = RTE_MIN(*addr, ms->addr);
511 * Reserve UAR address space for primary process.
514 * Pointer to Ethernet device.
517 * 0 on success, a negative errno value otherwise and rte_errno is set.
520 mlx5_uar_init_primary(struct rte_eth_dev *dev)
522 struct priv *priv = dev->data->dev_private;
523 void *addr = (void *)0;
525 if (uar_base) { /* UAR address space mapped. */
526 priv->uar_base = uar_base;
529 /* find out lower bound of hugepage segments */
530 rte_memseg_walk(find_lower_va_bound, &addr);
532 /* keep distance to hugepages to minimize potential conflicts. */
533 addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
534 /* anonymous mmap, no real memory consumption. */
535 addr = mmap(addr, MLX5_UAR_SIZE,
536 PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
537 if (addr == MAP_FAILED) {
539 "port %u failed to reserve UAR address space, please"
540 " adjust MLX5_UAR_SIZE or try --base-virtaddr",
545 /* Accept either same addr or a new addr returned from mmap if target
548 DRV_LOG(INFO, "port %u reserved UAR address space: %p",
549 dev->data->port_id, addr);
550 priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
551 uar_base = addr; /* process local, don't reserve again. */
556 * Reserve UAR address space for secondary process, align with
560 * Pointer to Ethernet device.
563 * 0 on success, a negative errno value otherwise and rte_errno is set.
566 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
568 struct priv *priv = dev->data->dev_private;
571 assert(priv->uar_base);
572 if (uar_base) { /* already reserved. */
573 assert(uar_base == priv->uar_base);
576 /* anonymous mmap, no real memory consumption. */
577 addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
578 PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
579 if (addr == MAP_FAILED) {
580 DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
581 dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
585 if (priv->uar_base != addr) {
587 "port %u UAR address %p size %llu occupied, please"
588 " adjust MLX5_UAR_OFFSET or try EAL parameter"
590 dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
594 uar_base = addr; /* process local, don't reserve again */
595 DRV_LOG(INFO, "port %u reserved UAR address space: %p",
596 dev->data->port_id, addr);
601 * DPDK callback to register a PCI device.
603 * This function creates an Ethernet device for each port of a given
607 * PCI driver structure (mlx5_driver).
609 * PCI device information.
612 * 0 on success, a negative errno value otherwise and rte_errno is set.
615 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
616 struct rte_pci_device *pci_dev)
618 struct ibv_device **list = NULL;
619 struct ibv_device *ibv_dev;
621 struct ibv_context *attr_ctx = NULL;
622 struct ibv_device_attr_ex device_attr;
625 unsigned int cqe_comp;
626 unsigned int tunnel_en = 0;
627 unsigned int swp = 0;
628 unsigned int verb_priorities = 0;
631 struct mlx5dv_context attrs_out = {0};
632 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
633 struct ibv_counter_set_description cs_desc;
636 assert(pci_drv == &mlx5_driver);
637 /* Get mlx5_dev[] index. */
638 idx = mlx5_dev_idx(&pci_dev->addr);
640 DRV_LOG(ERR, "this driver cannot support any more adapters");
644 DRV_LOG(DEBUG, "using driver device index %d", idx);
645 /* Save PCI address. */
646 mlx5_dev[idx].pci_addr = pci_dev->addr;
647 list = mlx5_glue->get_device_list(&i);
653 "cannot list devices, is ib_uverbs loaded?");
658 * For each listed device, check related sysfs entry against
659 * the provided PCI ID.
662 struct rte_pci_addr pci_addr;
665 DRV_LOG(DEBUG, "checking device \"%s\"", list[i]->name);
666 if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
668 if ((pci_dev->addr.domain != pci_addr.domain) ||
669 (pci_dev->addr.bus != pci_addr.bus) ||
670 (pci_dev->addr.devid != pci_addr.devid) ||
671 (pci_dev->addr.function != pci_addr.function))
673 DRV_LOG(INFO, "PCI information matches, using device \"%s\"",
675 vf = ((pci_dev->id.device_id ==
676 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
677 (pci_dev->id.device_id ==
678 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF) ||
679 (pci_dev->id.device_id ==
680 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF) ||
681 (pci_dev->id.device_id ==
682 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF));
683 attr_ctx = mlx5_glue->open_device(list[i]);
688 if (attr_ctx == NULL) {
692 "cannot access device, is mlx5_ib loaded?");
697 "cannot use device, are drivers up to date?");
703 DRV_LOG(DEBUG, "device opened");
704 #ifdef HAVE_IBV_MLX5_MOD_SWP
705 attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
708 * Multi-packet send is supported by ConnectX-4 Lx PF as well
709 * as all ConnectX-5 devices.
711 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
712 attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
714 mlx5_glue->dv_query_device(attr_ctx, &attrs_out);
715 if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
716 if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
717 DRV_LOG(DEBUG, "enhanced MPW is supported");
718 mps = MLX5_MPW_ENHANCED;
720 DRV_LOG(DEBUG, "MPW is supported");
724 DRV_LOG(DEBUG, "MPW isn't supported");
725 mps = MLX5_MPW_DISABLED;
727 #ifdef HAVE_IBV_MLX5_MOD_SWP
728 if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
729 swp = attrs_out.sw_parsing_caps.sw_parsing_offloads;
730 DRV_LOG(DEBUG, "SWP support: %u", swp);
732 if (RTE_CACHE_LINE_SIZE == 128 &&
733 !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
737 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
738 if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
739 tunnel_en = ((attrs_out.tunnel_offloads_caps &
740 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
741 (attrs_out.tunnel_offloads_caps &
742 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
744 DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
745 tunnel_en ? "" : "not ");
748 "tunnel offloading disabled due to old OFED/rdma-core version");
750 err = mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr);
752 DEBUG("ibv_query_device_ex() failed");
755 DRV_LOG(INFO, "%u port(s) detected",
756 device_attr.orig_attr.phys_port_cnt);
757 for (i = 0; i < device_attr.orig_attr.phys_port_cnt; i++) {
758 char name[RTE_ETH_NAME_MAX_LEN];
760 uint32_t port = i + 1; /* ports are indexed from one */
761 uint32_t test = (1 << i);
762 struct ibv_context *ctx = NULL;
763 struct ibv_port_attr port_attr;
764 struct ibv_pd *pd = NULL;
765 struct priv *priv = NULL;
766 struct rte_eth_dev *eth_dev = NULL;
767 struct ibv_device_attr_ex device_attr_ex;
768 struct ether_addr mac;
769 struct mlx5_dev_config config = {
770 .cqe_comp = cqe_comp,
772 .tunnel_en = tunnel_en,
776 .txq_inline = MLX5_ARG_UNSET,
777 .txqs_inline = MLX5_ARG_UNSET,
778 .inline_max_packet_sz = MLX5_ARG_UNSET,
783 len = snprintf(name, sizeof(name), PCI_PRI_FMT,
784 pci_dev->addr.domain, pci_dev->addr.bus,
785 pci_dev->addr.devid, pci_dev->addr.function);
786 if (device_attr.orig_attr.phys_port_cnt > 1)
787 snprintf(name + len, sizeof(name), " port %u", i);
788 mlx5_dev[idx].ports |= test;
789 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
790 eth_dev = rte_eth_dev_attach_secondary(name);
791 if (eth_dev == NULL) {
792 DRV_LOG(ERR, "can not attach rte ethdev");
797 eth_dev->device = &pci_dev->device;
798 eth_dev->dev_ops = &mlx5_dev_sec_ops;
799 err = mlx5_uar_init_secondary(eth_dev);
804 /* Receive command fd from primary process */
805 err = mlx5_socket_connect(eth_dev);
810 /* Remap UAR for Tx queues. */
811 err = mlx5_tx_uar_remap(eth_dev, err);
817 * Ethdev pointer is still required as input since
818 * the primary device is not accessible from the
821 eth_dev->rx_pkt_burst =
822 mlx5_select_rx_function(eth_dev);
823 eth_dev->tx_pkt_burst =
824 mlx5_select_tx_function(eth_dev);
827 DRV_LOG(DEBUG, "using port %u (%08" PRIx32 ")", port, test);
828 ctx = mlx5_glue->open_device(ibv_dev);
833 /* Check port status. */
834 err = mlx5_glue->query_port(ctx, port, &port_attr);
836 DRV_LOG(ERR, "port query failed: %s", strerror(err));
839 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
841 "port %d is not configured in Ethernet mode",
846 if (port_attr.state != IBV_PORT_ACTIVE)
847 DRV_LOG(DEBUG, "port %d is not active: \"%s\" (%d)",
849 mlx5_glue->port_state_str(port_attr.state),
851 /* Allocate protection domain. */
852 pd = mlx5_glue->alloc_pd(ctx);
854 DRV_LOG(ERR, "PD allocation failure");
858 mlx5_dev[idx].ports |= test;
859 /* from rte_ethdev.c */
860 priv = rte_zmalloc("ethdev private structure",
862 RTE_CACHE_LINE_SIZE);
864 DRV_LOG(ERR, "priv allocation failure");
869 strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
870 sizeof(priv->ibdev_path));
871 priv->device_attr = device_attr;
874 priv->mtu = ETHER_MTU;
875 err = mlx5_args(&config, pci_dev->device.devargs);
877 DRV_LOG(ERR, "failed to process device arguments: %s",
882 err = mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex);
884 DRV_LOG(ERR, "ibv_query_device_ex() failed");
887 config.hw_csum = !!(device_attr_ex.device_cap_flags_ex &
888 IBV_DEVICE_RAW_IP_CSUM);
889 DRV_LOG(DEBUG, "checksum offloading is %ssupported",
890 (config.hw_csum ? "" : "not "));
891 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
892 config.flow_counter_en = !!(device_attr.max_counter_sets);
893 mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
895 "counter type = %d, num of cs = %ld, attributes = %d",
896 cs_desc.counter_type, cs_desc.num_of_cs,
899 config.ind_table_max_size =
900 device_attr_ex.rss_caps.max_rwq_indirection_table_size;
901 /* Remove this check once DPDK supports larger/variable
902 * indirection tables. */
903 if (config.ind_table_max_size >
904 (unsigned int)ETH_RSS_RETA_SIZE_512)
905 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
906 DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
907 config.ind_table_max_size);
908 config.hw_vlan_strip = !!(device_attr_ex.raw_packet_caps &
909 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
910 DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
911 (config.hw_vlan_strip ? "" : "not "));
913 config.hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
914 IBV_RAW_PACKET_CAP_SCATTER_FCS);
915 DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
916 (config.hw_fcs_strip ? "" : "not "));
918 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
919 config.hw_padding = !!device_attr_ex.rx_pad_end_addr_align;
922 "hardware Rx end alignment padding is %ssupported",
923 (config.hw_padding ? "" : "not "));
925 config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
926 (device_attr_ex.tso_caps.supported_qpts &
927 (1 << IBV_QPT_RAW_PACKET)));
929 config.tso_max_payload_sz =
930 device_attr_ex.tso_caps.max_tso;
931 if (config.mps && !mps) {
933 "multi-packet send not supported on this device"
934 " (" MLX5_TXQ_MPW_EN ")");
938 DRV_LOG(INFO, "%s MPS is %s",
939 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
940 config.mps != MLX5_MPW_DISABLED ? "enabled" :
942 if (config.cqe_comp && !cqe_comp) {
943 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
946 eth_dev = rte_eth_dev_allocate(name);
947 if (eth_dev == NULL) {
948 DRV_LOG(ERR, "can not allocate rte ethdev");
952 eth_dev->data->dev_private = priv;
953 priv->dev_data = eth_dev->data;
954 eth_dev->data->mac_addrs = priv->mac;
955 eth_dev->device = &pci_dev->device;
956 rte_eth_copy_pci_info(eth_dev, pci_dev);
957 eth_dev->device->driver = &mlx5_driver.driver;
958 err = mlx5_uar_init_primary(eth_dev);
963 /* Configure the first MAC address by default. */
964 if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
966 "port %u cannot get MAC address, is mlx5_en"
967 " loaded? (errno: %s)",
968 eth_dev->data->port_id, strerror(errno));
973 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
974 eth_dev->data->port_id,
975 mac.addr_bytes[0], mac.addr_bytes[1],
976 mac.addr_bytes[2], mac.addr_bytes[3],
977 mac.addr_bytes[4], mac.addr_bytes[5]);
980 char ifname[IF_NAMESIZE];
982 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
983 DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
984 eth_dev->data->port_id, ifname);
986 DRV_LOG(DEBUG, "port %u ifname is unknown",
987 eth_dev->data->port_id);
990 /* Get actual MTU if possible. */
991 err = mlx5_get_mtu(eth_dev, &priv->mtu);
996 DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
999 * Initialize burst functions to prevent crashes before link-up.
1001 eth_dev->rx_pkt_burst = removed_rx_burst;
1002 eth_dev->tx_pkt_burst = removed_tx_burst;
1003 eth_dev->dev_ops = &mlx5_dev_ops;
1004 /* Register MAC address. */
1005 claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1006 priv->nl_socket = -1;
1008 if (vf && config.vf_nl_en) {
1009 priv->nl_socket = mlx5_nl_init(RTMGRP_LINK);
1010 if (priv->nl_socket < 0)
1011 priv->nl_socket = -1;
1012 mlx5_nl_mac_addr_sync(eth_dev);
1014 TAILQ_INIT(&priv->flows);
1015 TAILQ_INIT(&priv->ctrl_flows);
1016 /* Hint libmlx5 to use PMD allocator for data plane resources */
1017 struct mlx5dv_ctx_allocators alctr = {
1018 .alloc = &mlx5_alloc_verbs_buf,
1019 .free = &mlx5_free_verbs_buf,
1022 mlx5_glue->dv_set_context_attr(ctx,
1023 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1024 (void *)((uintptr_t)&alctr));
1025 /* Bring Ethernet device up. */
1026 DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1027 eth_dev->data->port_id);
1028 mlx5_set_link_up(eth_dev);
1030 * Even though the interrupt handler is not installed yet,
1031 * interrupts will still trigger on the asyn_fd from
1032 * Verbs context returned by ibv_open_device().
1034 mlx5_link_update(eth_dev, 0);
1035 /* Store device configuration on private structure. */
1036 priv->config = config;
1037 /* Create drop queue. */
1038 err = mlx5_flow_create_drop_queue(eth_dev);
1040 DRV_LOG(ERR, "port %u drop queue allocation failed: %s",
1041 eth_dev->data->port_id, strerror(rte_errno));
1045 /* Supported Verbs flow priority number detection. */
1046 if (verb_priorities == 0)
1047 verb_priorities = mlx5_get_max_verbs_prio(eth_dev);
1048 if (verb_priorities < MLX5_VERBS_FLOW_PRIO_8) {
1049 DRV_LOG(ERR, "port %u wrong Verbs flow priorities: %u",
1050 eth_dev->data->port_id, verb_priorities);
1053 priv->config.max_verbs_prio = verb_priorities;
1059 claim_zero(mlx5_glue->dealloc_pd(pd));
1061 claim_zero(mlx5_glue->close_device(ctx));
1062 if (eth_dev && rte_eal_process_type() == RTE_PROC_PRIMARY)
1063 rte_eth_dev_release_port(eth_dev);
1067 * XXX if something went wrong in the loop above, there is a resource
1068 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1069 * long as the dpdk does not provide a way to deallocate a ethdev and a
1070 * way to enumerate the registered ethdevs to free the previous ones.
1072 /* no port found, complain */
1073 if (!mlx5_dev[idx].ports) {
1079 claim_zero(mlx5_glue->close_device(attr_ctx));
1081 mlx5_glue->free_device_list(list);
1089 static const struct rte_pci_id mlx5_pci_id_map[] = {
1091 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1092 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1095 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1096 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1099 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1100 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1103 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1104 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1107 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1108 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1111 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1112 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1115 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1116 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1119 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1120 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1127 static struct rte_pci_driver mlx5_driver = {
1129 .name = MLX5_DRIVER_NAME
1131 .id_table = mlx5_pci_id_map,
1132 .probe = mlx5_pci_probe,
1133 .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1136 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1139 * Suffix RTE_EAL_PMD_PATH with "-glue".
1141 * This function performs a sanity check on RTE_EAL_PMD_PATH before
1142 * suffixing its last component.
1145 * Output buffer, should be large enough otherwise NULL is returned.
1150 * Pointer to @p buf or @p NULL in case suffix cannot be appended.
1153 mlx5_glue_path(char *buf, size_t size)
1155 static const char *const bad[] = { "/", ".", "..", NULL };
1156 const char *path = RTE_EAL_PMD_PATH;
1157 size_t len = strlen(path);
1161 while (len && path[len - 1] == '/')
1163 for (off = len; off && path[off - 1] != '/'; --off)
1165 for (i = 0; bad[i]; ++i)
1166 if (!strncmp(path + off, bad[i], (int)(len - off)))
1168 i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1169 if (i == -1 || (size_t)i >= size)
1174 "unable to append \"-glue\" to last component of"
1175 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1176 " please re-configure DPDK");
1181 * Initialization routine for run-time dependency on rdma-core.
1184 mlx5_glue_init(void)
1186 char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1187 const char *path[] = {
1189 * A basic security check is necessary before trusting
1190 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1192 (geteuid() == getuid() && getegid() == getgid() ?
1193 getenv("MLX5_GLUE_PATH") : NULL),
1195 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1196 * variant, otherwise let dlopen() look up libraries on its
1199 (*RTE_EAL_PMD_PATH ?
1200 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1203 void *handle = NULL;
1207 while (!handle && i != RTE_DIM(path)) {
1216 end = strpbrk(path[i], ":;");
1218 end = path[i] + strlen(path[i]);
1219 len = end - path[i];
1224 ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1226 (!len || *(end - 1) == '/') ? "" : "/");
1229 if (sizeof(name) != (size_t)ret + 1)
1231 DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1233 handle = dlopen(name, RTLD_LAZY);
1244 DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1247 sym = dlsym(handle, "mlx5_glue");
1248 if (!sym || !*sym) {
1252 DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1261 "cannot initialize PMD due to missing run-time dependency on"
1262 " rdma-core libraries (libibverbs, libmlx5)");
1269 * Driver initialization routine.
1271 RTE_INIT(rte_mlx5_pmd_init);
1273 rte_mlx5_pmd_init(void)
1275 /* Build the static tables for Verbs conversion. */
1276 mlx5_set_ptype_table();
1277 mlx5_set_cksum_table();
1278 mlx5_set_swp_types_table();
1280 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1281 * huge pages. Calling ibv_fork_init() during init allows
1282 * applications to use fork() safely for purposes other than
1283 * using this PMD, which is not supported in forked processes.
1285 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1286 /* Match the size of Rx completion entry to the size of a cacheline. */
1287 if (RTE_CACHE_LINE_SIZE == 128)
1288 setenv("MLX5_CQE_SIZE", "128", 0);
1289 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1290 if (mlx5_glue_init())
1295 /* Glue structure must not contain any NULL pointers. */
1299 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1300 assert(((const void *const *)mlx5_glue)[i]);
1303 if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1305 "rdma-core glue \"%s\" mismatch: \"%s\" is required",
1306 mlx5_glue->version, MLX5_GLUE_VERSION);
1309 mlx5_glue->fork_init();
1310 rte_pci_register(&mlx5_driver);
1313 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1314 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1315 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1317 /** Initialize driver log type. */
1318 RTE_INIT(vdev_netvsc_init_log)
1320 mlx5_logtype = rte_log_register("pmd.net.mlx5");
1321 if (mlx5_logtype >= 0)
1322 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);