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 #ifndef HAVE_IBV_MLX5_MOD_MPW
73 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
74 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
77 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
78 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
81 /** Driver-specific log messages type. */
85 * Retrieve integer value from environment variable.
88 * Environment variable name.
91 * Integer value, 0 if the variable is not set.
94 mlx5_getenv_int(const char *name)
96 const char *val = getenv(name);
104 * Verbs callback to allocate a memory. This function should allocate the space
105 * according to the size provided residing inside a huge page.
106 * Please note that all allocation must respect the alignment from libmlx5
107 * (i.e. currently sysconf(_SC_PAGESIZE)).
110 * The size in bytes of the memory to allocate.
112 * A pointer to the callback data.
115 * Allocated buffer, NULL otherwise and rte_errno is set.
118 mlx5_alloc_verbs_buf(size_t size, void *data)
120 struct priv *priv = data;
122 size_t alignment = sysconf(_SC_PAGESIZE);
123 unsigned int socket = SOCKET_ID_ANY;
125 if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
126 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
128 socket = ctrl->socket;
129 } else if (priv->verbs_alloc_ctx.type ==
130 MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
131 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
133 socket = ctrl->socket;
135 assert(data != NULL);
136 ret = rte_malloc_socket(__func__, size, alignment, socket);
143 * Verbs callback to free a memory.
146 * A pointer to the memory to free.
148 * A pointer to the callback data.
151 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
153 assert(data != NULL);
158 * DPDK callback to close the device.
160 * Destroy all queues and objects, free memory.
163 * Pointer to Ethernet device structure.
166 mlx5_dev_close(struct rte_eth_dev *dev)
168 struct priv *priv = dev->data->dev_private;
172 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
174 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
175 /* In case mlx5_dev_stop() has not been called. */
176 mlx5_dev_interrupt_handler_uninstall(dev);
177 mlx5_traffic_disable(dev);
178 /* Prevent crashes when queues are still in use. */
179 dev->rx_pkt_burst = removed_rx_burst;
180 dev->tx_pkt_burst = removed_tx_burst;
181 if (priv->rxqs != NULL) {
182 /* XXX race condition if mlx5_rx_burst() is still running. */
184 for (i = 0; (i != priv->rxqs_n); ++i)
185 mlx5_rxq_release(dev, i);
189 if (priv->txqs != NULL) {
190 /* XXX race condition if mlx5_tx_burst() is still running. */
192 for (i = 0; (i != priv->txqs_n); ++i)
193 mlx5_txq_release(dev, i);
197 if (priv->pd != NULL) {
198 assert(priv->ctx != NULL);
199 claim_zero(mlx5_glue->dealloc_pd(priv->pd));
200 claim_zero(mlx5_glue->close_device(priv->ctx));
202 assert(priv->ctx == NULL);
203 if (priv->rss_conf.rss_key != NULL)
204 rte_free(priv->rss_conf.rss_key);
205 if (priv->reta_idx != NULL)
206 rte_free(priv->reta_idx);
207 if (priv->primary_socket)
208 mlx5_socket_uninit(dev);
210 mlx5_nl_mac_addr_flush(dev);
211 if (priv->nl_socket >= 0)
212 close(priv->nl_socket);
213 ret = mlx5_hrxq_ibv_verify(dev);
215 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
217 ret = mlx5_ind_table_ibv_verify(dev);
219 DRV_LOG(WARNING, "port %u some indirection table still remain",
221 ret = mlx5_rxq_ibv_verify(dev);
223 DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
225 ret = mlx5_rxq_verify(dev);
227 DRV_LOG(WARNING, "port %u some Rx queues still remain",
229 ret = mlx5_txq_ibv_verify(dev);
231 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
233 ret = mlx5_txq_verify(dev);
235 DRV_LOG(WARNING, "port %u some Tx queues still remain",
237 ret = mlx5_flow_verify(dev);
239 DRV_LOG(WARNING, "port %u some flows still remain",
241 ret = mlx5_mr_verify(dev);
243 DRV_LOG(WARNING, "port %u some memory region still remain",
245 memset(priv, 0, sizeof(*priv));
248 const struct eth_dev_ops mlx5_dev_ops = {
249 .dev_configure = mlx5_dev_configure,
250 .dev_start = mlx5_dev_start,
251 .dev_stop = mlx5_dev_stop,
252 .dev_set_link_down = mlx5_set_link_down,
253 .dev_set_link_up = mlx5_set_link_up,
254 .dev_close = mlx5_dev_close,
255 .promiscuous_enable = mlx5_promiscuous_enable,
256 .promiscuous_disable = mlx5_promiscuous_disable,
257 .allmulticast_enable = mlx5_allmulticast_enable,
258 .allmulticast_disable = mlx5_allmulticast_disable,
259 .link_update = mlx5_link_update,
260 .stats_get = mlx5_stats_get,
261 .stats_reset = mlx5_stats_reset,
262 .xstats_get = mlx5_xstats_get,
263 .xstats_reset = mlx5_xstats_reset,
264 .xstats_get_names = mlx5_xstats_get_names,
265 .dev_infos_get = mlx5_dev_infos_get,
266 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
267 .vlan_filter_set = mlx5_vlan_filter_set,
268 .rx_queue_setup = mlx5_rx_queue_setup,
269 .tx_queue_setup = mlx5_tx_queue_setup,
270 .rx_queue_release = mlx5_rx_queue_release,
271 .tx_queue_release = mlx5_tx_queue_release,
272 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
273 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
274 .mac_addr_remove = mlx5_mac_addr_remove,
275 .mac_addr_add = mlx5_mac_addr_add,
276 .mac_addr_set = mlx5_mac_addr_set,
277 .mtu_set = mlx5_dev_set_mtu,
278 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
279 .vlan_offload_set = mlx5_vlan_offload_set,
280 .reta_update = mlx5_dev_rss_reta_update,
281 .reta_query = mlx5_dev_rss_reta_query,
282 .rss_hash_update = mlx5_rss_hash_update,
283 .rss_hash_conf_get = mlx5_rss_hash_conf_get,
284 .filter_ctrl = mlx5_dev_filter_ctrl,
285 .rx_descriptor_status = mlx5_rx_descriptor_status,
286 .tx_descriptor_status = mlx5_tx_descriptor_status,
287 .rx_queue_intr_enable = mlx5_rx_intr_enable,
288 .rx_queue_intr_disable = mlx5_rx_intr_disable,
289 .is_removed = mlx5_is_removed,
292 static const struct eth_dev_ops mlx5_dev_sec_ops = {
293 .stats_get = mlx5_stats_get,
294 .stats_reset = mlx5_stats_reset,
295 .xstats_get = mlx5_xstats_get,
296 .xstats_reset = mlx5_xstats_reset,
297 .xstats_get_names = mlx5_xstats_get_names,
298 .dev_infos_get = mlx5_dev_infos_get,
299 .rx_descriptor_status = mlx5_rx_descriptor_status,
300 .tx_descriptor_status = mlx5_tx_descriptor_status,
303 /* Available operators in flow isolated mode. */
304 const struct eth_dev_ops mlx5_dev_ops_isolate = {
305 .dev_configure = mlx5_dev_configure,
306 .dev_start = mlx5_dev_start,
307 .dev_stop = mlx5_dev_stop,
308 .dev_set_link_down = mlx5_set_link_down,
309 .dev_set_link_up = mlx5_set_link_up,
310 .dev_close = mlx5_dev_close,
311 .link_update = mlx5_link_update,
312 .stats_get = mlx5_stats_get,
313 .stats_reset = mlx5_stats_reset,
314 .xstats_get = mlx5_xstats_get,
315 .xstats_reset = mlx5_xstats_reset,
316 .xstats_get_names = mlx5_xstats_get_names,
317 .dev_infos_get = mlx5_dev_infos_get,
318 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
319 .vlan_filter_set = mlx5_vlan_filter_set,
320 .rx_queue_setup = mlx5_rx_queue_setup,
321 .tx_queue_setup = mlx5_tx_queue_setup,
322 .rx_queue_release = mlx5_rx_queue_release,
323 .tx_queue_release = mlx5_tx_queue_release,
324 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
325 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
326 .mac_addr_remove = mlx5_mac_addr_remove,
327 .mac_addr_add = mlx5_mac_addr_add,
328 .mac_addr_set = mlx5_mac_addr_set,
329 .mtu_set = mlx5_dev_set_mtu,
330 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
331 .vlan_offload_set = mlx5_vlan_offload_set,
332 .filter_ctrl = mlx5_dev_filter_ctrl,
333 .rx_descriptor_status = mlx5_rx_descriptor_status,
334 .tx_descriptor_status = mlx5_tx_descriptor_status,
335 .rx_queue_intr_enable = mlx5_rx_intr_enable,
336 .rx_queue_intr_disable = mlx5_rx_intr_disable,
337 .is_removed = mlx5_is_removed,
341 struct rte_pci_addr pci_addr; /* associated PCI address */
342 uint32_t ports; /* physical ports bitfield. */
346 * Get device index in mlx5_dev[] from PCI bus address.
348 * @param[in] pci_addr
349 * PCI bus address to look for.
352 * mlx5_dev[] index on success, -1 on failure.
355 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
360 assert(pci_addr != NULL);
361 for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
362 if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
363 (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
364 (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
365 (mlx5_dev[i].pci_addr.function == pci_addr->function))
367 if ((mlx5_dev[i].ports == 0) && (ret == -1))
374 * Verify and store value for device argument.
377 * Key argument to verify.
379 * Value associated with key.
384 * 0 on success, a negative errno value otherwise and rte_errno is set.
387 mlx5_args_check(const char *key, const char *val, void *opaque)
389 struct mlx5_dev_config *config = opaque;
393 tmp = strtoul(val, NULL, 0);
396 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
399 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
400 config->cqe_comp = !!tmp;
401 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
402 config->txq_inline = tmp;
403 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
404 config->txqs_inline = tmp;
405 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
406 config->mps = !!tmp ? config->mps : 0;
407 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
408 config->mpw_hdr_dseg = !!tmp;
409 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
410 config->inline_max_packet_sz = tmp;
411 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
412 config->tx_vec_en = !!tmp;
413 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
414 config->rx_vec_en = !!tmp;
416 DRV_LOG(WARNING, "%s: unknown parameter", key);
424 * Parse device parameters.
427 * Pointer to device configuration structure.
429 * Device arguments structure.
432 * 0 on success, a negative errno value otherwise and rte_errno is set.
435 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
437 const char **params = (const char *[]){
438 MLX5_RXQ_CQE_COMP_EN,
440 MLX5_TXQS_MIN_INLINE,
442 MLX5_TXQ_MPW_HDR_DSEG_EN,
443 MLX5_TXQ_MAX_INLINE_LEN,
448 struct rte_kvargs *kvlist;
454 /* Following UGLY cast is done to pass checkpatch. */
455 kvlist = rte_kvargs_parse(devargs->args, params);
458 /* Process parameters. */
459 for (i = 0; (params[i] != NULL); ++i) {
460 if (rte_kvargs_count(kvlist, params[i])) {
461 ret = rte_kvargs_process(kvlist, params[i],
462 mlx5_args_check, config);
465 rte_kvargs_free(kvlist);
470 rte_kvargs_free(kvlist);
474 static struct rte_pci_driver mlx5_driver;
477 * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
478 * local resource used by both primary and secondary to avoid duplicate
480 * The space has to be available on both primary and secondary process,
481 * TXQ UAR maps to this area using fixed mmap w/o double check.
483 static void *uar_base;
486 find_lower_va_bound(const struct rte_memseg_list *msl __rte_unused,
487 const struct rte_memseg *ms, void *arg)
494 *addr = RTE_MIN(*addr, ms->addr);
500 * Reserve UAR address space for primary process.
503 * Pointer to Ethernet device.
506 * 0 on success, a negative errno value otherwise and rte_errno is set.
509 mlx5_uar_init_primary(struct rte_eth_dev *dev)
511 struct priv *priv = dev->data->dev_private;
512 void *addr = (void *)0;
514 if (uar_base) { /* UAR address space mapped. */
515 priv->uar_base = uar_base;
518 /* find out lower bound of hugepage segments */
519 rte_memseg_walk(find_lower_va_bound, &addr);
521 /* keep distance to hugepages to minimize potential conflicts. */
522 addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
523 /* anonymous mmap, no real memory consumption. */
524 addr = mmap(addr, MLX5_UAR_SIZE,
525 PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
526 if (addr == MAP_FAILED) {
528 "port %u failed to reserve UAR address space, please"
529 " adjust MLX5_UAR_SIZE or try --base-virtaddr",
534 /* Accept either same addr or a new addr returned from mmap if target
537 DRV_LOG(INFO, "port %u reserved UAR address space: %p",
538 dev->data->port_id, addr);
539 priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
540 uar_base = addr; /* process local, don't reserve again. */
545 * Reserve UAR address space for secondary process, align with
549 * Pointer to Ethernet device.
552 * 0 on success, a negative errno value otherwise and rte_errno is set.
555 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
557 struct priv *priv = dev->data->dev_private;
560 assert(priv->uar_base);
561 if (uar_base) { /* already reserved. */
562 assert(uar_base == priv->uar_base);
565 /* anonymous mmap, no real memory consumption. */
566 addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
567 PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
568 if (addr == MAP_FAILED) {
569 DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
570 dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
574 if (priv->uar_base != addr) {
576 "port %u UAR address %p size %llu occupied, please"
577 " adjust MLX5_UAR_OFFSET or try EAL parameter"
579 dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
583 uar_base = addr; /* process local, don't reserve again */
584 DRV_LOG(INFO, "port %u reserved UAR address space: %p",
585 dev->data->port_id, addr);
590 * DPDK callback to register a PCI device.
592 * This function creates an Ethernet device for each port of a given
596 * PCI driver structure (mlx5_driver).
598 * PCI device information.
601 * 0 on success, a negative errno value otherwise and rte_errno is set.
604 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
605 struct rte_pci_device *pci_dev)
607 struct ibv_device **list = NULL;
608 struct ibv_device *ibv_dev;
610 struct ibv_context *attr_ctx = NULL;
611 struct ibv_device_attr_ex device_attr;
614 unsigned int cqe_comp;
615 unsigned int tunnel_en = 0;
618 struct mlx5dv_context attrs_out = {0};
619 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
620 struct ibv_counter_set_description cs_desc;
623 assert(pci_drv == &mlx5_driver);
624 /* Get mlx5_dev[] index. */
625 idx = mlx5_dev_idx(&pci_dev->addr);
627 DRV_LOG(ERR, "this driver cannot support any more adapters");
631 DRV_LOG(DEBUG, "using driver device index %d", idx);
632 /* Save PCI address. */
633 mlx5_dev[idx].pci_addr = pci_dev->addr;
634 list = mlx5_glue->get_device_list(&i);
640 "cannot list devices, is ib_uverbs loaded?");
645 * For each listed device, check related sysfs entry against
646 * the provided PCI ID.
649 struct rte_pci_addr pci_addr;
652 DRV_LOG(DEBUG, "checking device \"%s\"", list[i]->name);
653 if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
655 if ((pci_dev->addr.domain != pci_addr.domain) ||
656 (pci_dev->addr.bus != pci_addr.bus) ||
657 (pci_dev->addr.devid != pci_addr.devid) ||
658 (pci_dev->addr.function != pci_addr.function))
660 DRV_LOG(INFO, "PCI information matches, using device \"%s\"",
662 vf = ((pci_dev->id.device_id ==
663 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
664 (pci_dev->id.device_id ==
665 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF) ||
666 (pci_dev->id.device_id ==
667 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF) ||
668 (pci_dev->id.device_id ==
669 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF));
670 attr_ctx = mlx5_glue->open_device(list[i]);
675 if (attr_ctx == NULL) {
676 mlx5_glue->free_device_list(list);
680 "cannot access device, is mlx5_ib loaded?");
685 "cannot use device, are drivers up to date?");
690 DRV_LOG(DEBUG, "device opened");
692 * Multi-packet send is supported by ConnectX-4 Lx PF as well
693 * as all ConnectX-5 devices.
695 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
696 attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
698 mlx5_glue->dv_query_device(attr_ctx, &attrs_out);
699 if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
700 if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
701 DRV_LOG(DEBUG, "enhanced MPW is supported");
702 mps = MLX5_MPW_ENHANCED;
704 DRV_LOG(DEBUG, "MPW is supported");
708 DRV_LOG(DEBUG, "MPW isn't supported");
709 mps = MLX5_MPW_DISABLED;
711 if (RTE_CACHE_LINE_SIZE == 128 &&
712 !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
716 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
717 if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
718 tunnel_en = ((attrs_out.tunnel_offloads_caps &
719 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
720 (attrs_out.tunnel_offloads_caps &
721 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
723 DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
724 tunnel_en ? "" : "not ");
727 "tunnel offloading disabled due to old OFED/rdma-core version");
729 if (mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr)) {
733 DRV_LOG(INFO, "%u port(s) detected",
734 device_attr.orig_attr.phys_port_cnt);
735 for (i = 0; i < device_attr.orig_attr.phys_port_cnt; i++) {
736 char name[RTE_ETH_NAME_MAX_LEN];
738 uint32_t port = i + 1; /* ports are indexed from one */
739 uint32_t test = (1 << i);
740 struct ibv_context *ctx = NULL;
741 struct ibv_port_attr port_attr;
742 struct ibv_pd *pd = NULL;
743 struct priv *priv = NULL;
744 struct rte_eth_dev *eth_dev = NULL;
745 struct ibv_device_attr_ex device_attr_ex;
746 struct ether_addr mac;
747 struct mlx5_dev_config config = {
748 .cqe_comp = cqe_comp,
750 .tunnel_en = tunnel_en,
754 .txq_inline = MLX5_ARG_UNSET,
755 .txqs_inline = MLX5_ARG_UNSET,
756 .inline_max_packet_sz = MLX5_ARG_UNSET,
759 len = snprintf(name, sizeof(name), PCI_PRI_FMT,
760 pci_dev->addr.domain, pci_dev->addr.bus,
761 pci_dev->addr.devid, pci_dev->addr.function);
762 if (device_attr.orig_attr.phys_port_cnt > 1)
763 snprintf(name + len, sizeof(name), " port %u", i);
764 mlx5_dev[idx].ports |= test;
765 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
766 eth_dev = rte_eth_dev_attach_secondary(name);
767 if (eth_dev == NULL) {
768 DRV_LOG(ERR, "can not attach rte ethdev");
773 eth_dev->device = &pci_dev->device;
774 eth_dev->dev_ops = &mlx5_dev_sec_ops;
775 err = mlx5_uar_init_secondary(eth_dev);
778 /* Receive command fd from primary process */
779 err = mlx5_socket_connect(eth_dev);
782 /* Remap UAR for Tx queues. */
783 err = mlx5_tx_uar_remap(eth_dev, err);
787 * Ethdev pointer is still required as input since
788 * the primary device is not accessible from the
791 eth_dev->rx_pkt_burst =
792 mlx5_select_rx_function(eth_dev);
793 eth_dev->tx_pkt_burst =
794 mlx5_select_tx_function(eth_dev);
797 DRV_LOG(DEBUG, "using port %u (%08" PRIx32 ")", port, test);
798 ctx = mlx5_glue->open_device(ibv_dev);
803 /* Check port status. */
804 err = mlx5_glue->query_port(ctx, port, &port_attr);
806 DRV_LOG(ERR, "port query failed: %s", strerror(err));
809 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
811 "port %d is not configured in Ethernet mode",
816 if (port_attr.state != IBV_PORT_ACTIVE)
817 DRV_LOG(DEBUG, "port %d is not active: \"%s\" (%d)",
819 mlx5_glue->port_state_str(port_attr.state),
821 /* Allocate protection domain. */
822 pd = mlx5_glue->alloc_pd(ctx);
824 DRV_LOG(ERR, "PD allocation failure");
828 mlx5_dev[idx].ports |= test;
829 /* from rte_ethdev.c */
830 priv = rte_zmalloc("ethdev private structure",
832 RTE_CACHE_LINE_SIZE);
834 DRV_LOG(ERR, "priv allocation failure");
839 strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
840 sizeof(priv->ibdev_path));
841 priv->device_attr = device_attr;
844 priv->mtu = ETHER_MTU;
845 err = mlx5_args(&config, pci_dev->device.devargs);
847 DRV_LOG(ERR, "failed to process device arguments: %s",
851 if (mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex)) {
852 DRV_LOG(ERR, "ibv_query_device_ex() failed");
856 config.hw_csum = !!(device_attr_ex.device_cap_flags_ex &
857 IBV_DEVICE_RAW_IP_CSUM);
858 DRV_LOG(DEBUG, "checksum offloading is %ssupported",
859 (config.hw_csum ? "" : "not "));
860 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
861 config.flow_counter_en = !!(device_attr.max_counter_sets);
862 mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
864 "counter type = %d, num of cs = %ld, attributes = %d",
865 cs_desc.counter_type, cs_desc.num_of_cs,
868 config.ind_table_max_size =
869 device_attr_ex.rss_caps.max_rwq_indirection_table_size;
870 /* Remove this check once DPDK supports larger/variable
871 * indirection tables. */
872 if (config.ind_table_max_size >
873 (unsigned int)ETH_RSS_RETA_SIZE_512)
874 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
875 DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
876 config.ind_table_max_size);
877 config.hw_vlan_strip = !!(device_attr_ex.raw_packet_caps &
878 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
879 DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
880 (config.hw_vlan_strip ? "" : "not "));
882 config.hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
883 IBV_RAW_PACKET_CAP_SCATTER_FCS);
884 DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
885 (config.hw_fcs_strip ? "" : "not "));
887 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
888 config.hw_padding = !!device_attr_ex.rx_pad_end_addr_align;
891 "hardware Rx end alignment padding is %ssupported",
892 (config.hw_padding ? "" : "not "));
894 config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
895 (device_attr_ex.tso_caps.supported_qpts &
896 (1 << IBV_QPT_RAW_PACKET)));
898 config.tso_max_payload_sz =
899 device_attr_ex.tso_caps.max_tso;
900 if (config.mps && !mps) {
902 "multi-packet send not supported on this device"
903 " (" MLX5_TXQ_MPW_EN ")");
907 DRV_LOG(INFO, "%s MPS is %s",
908 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
909 config.mps != MLX5_MPW_DISABLED ? "enabled" :
911 if (config.cqe_comp && !cqe_comp) {
912 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
915 eth_dev = rte_eth_dev_allocate(name);
916 if (eth_dev == NULL) {
917 DRV_LOG(ERR, "can not allocate rte ethdev");
921 eth_dev->data->dev_private = priv;
923 eth_dev->data->mac_addrs = priv->mac;
924 eth_dev->device = &pci_dev->device;
925 rte_eth_copy_pci_info(eth_dev, pci_dev);
926 eth_dev->device->driver = &mlx5_driver.driver;
927 err = mlx5_uar_init_primary(eth_dev);
930 /* Configure the first MAC address by default. */
931 if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
933 "port %u cannot get MAC address, is mlx5_en"
934 " loaded? (errno: %s)",
935 eth_dev->data->port_id, strerror(errno));
940 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
941 eth_dev->data->port_id,
942 mac.addr_bytes[0], mac.addr_bytes[1],
943 mac.addr_bytes[2], mac.addr_bytes[3],
944 mac.addr_bytes[4], mac.addr_bytes[5]);
947 char ifname[IF_NAMESIZE];
949 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
950 DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
951 eth_dev->data->port_id, ifname);
953 DRV_LOG(DEBUG, "port %u ifname is unknown",
954 eth_dev->data->port_id);
957 /* Get actual MTU if possible. */
958 err = mlx5_get_mtu(eth_dev, &priv->mtu);
961 DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
964 * Initialize burst functions to prevent crashes before link-up.
966 eth_dev->rx_pkt_burst = removed_rx_burst;
967 eth_dev->tx_pkt_burst = removed_tx_burst;
968 eth_dev->dev_ops = &mlx5_dev_ops;
969 /* Register MAC address. */
970 claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
971 priv->nl_socket = -1;
974 priv->nl_socket = mlx5_nl_init(RTMGRP_LINK);
975 if (priv->nl_socket < 0)
976 priv->nl_socket = -1;
977 mlx5_nl_mac_addr_sync(eth_dev);
979 TAILQ_INIT(&priv->flows);
980 TAILQ_INIT(&priv->ctrl_flows);
981 /* Hint libmlx5 to use PMD allocator for data plane resources */
982 struct mlx5dv_ctx_allocators alctr = {
983 .alloc = &mlx5_alloc_verbs_buf,
984 .free = &mlx5_free_verbs_buf,
987 mlx5_glue->dv_set_context_attr(ctx,
988 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
989 (void *)((uintptr_t)&alctr));
990 /* Bring Ethernet device up. */
991 DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
992 eth_dev->data->port_id);
993 mlx5_set_link_up(eth_dev);
994 /* Store device configuration on private structure. */
995 priv->config = config;
1001 claim_zero(mlx5_glue->dealloc_pd(pd));
1003 claim_zero(mlx5_glue->close_device(ctx));
1007 * XXX if something went wrong in the loop above, there is a resource
1008 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1009 * long as the dpdk does not provide a way to deallocate a ethdev and a
1010 * way to enumerate the registered ethdevs to free the previous ones.
1012 /* no port found, complain */
1013 if (!mlx5_dev[idx].ports) {
1019 claim_zero(mlx5_glue->close_device(attr_ctx));
1021 mlx5_glue->free_device_list(list);
1029 static const struct rte_pci_id mlx5_pci_id_map[] = {
1031 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1032 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1035 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1036 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1039 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1040 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1043 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1044 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1047 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1048 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1051 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1052 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1055 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1056 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1059 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1060 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1067 static struct rte_pci_driver mlx5_driver = {
1069 .name = MLX5_DRIVER_NAME
1071 .id_table = mlx5_pci_id_map,
1072 .probe = mlx5_pci_probe,
1073 .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1076 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1079 * Suffix RTE_EAL_PMD_PATH with "-glue".
1081 * This function performs a sanity check on RTE_EAL_PMD_PATH before
1082 * suffixing its last component.
1085 * Output buffer, should be large enough otherwise NULL is returned.
1090 * Pointer to @p buf or @p NULL in case suffix cannot be appended.
1093 mlx5_glue_path(char *buf, size_t size)
1095 static const char *const bad[] = { "/", ".", "..", NULL };
1096 const char *path = RTE_EAL_PMD_PATH;
1097 size_t len = strlen(path);
1101 while (len && path[len - 1] == '/')
1103 for (off = len; off && path[off - 1] != '/'; --off)
1105 for (i = 0; bad[i]; ++i)
1106 if (!strncmp(path + off, bad[i], (int)(len - off)))
1108 i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1109 if (i == -1 || (size_t)i >= size)
1114 "unable to append \"-glue\" to last component of"
1115 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1116 " please re-configure DPDK");
1121 * Initialization routine for run-time dependency on rdma-core.
1124 mlx5_glue_init(void)
1126 char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1127 const char *path[] = {
1129 * A basic security check is necessary before trusting
1130 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1132 (geteuid() == getuid() && getegid() == getgid() ?
1133 getenv("MLX5_GLUE_PATH") : NULL),
1135 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1136 * variant, otherwise let dlopen() look up libraries on its
1139 (*RTE_EAL_PMD_PATH ?
1140 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1143 void *handle = NULL;
1147 while (!handle && i != RTE_DIM(path)) {
1156 end = strpbrk(path[i], ":;");
1158 end = path[i] + strlen(path[i]);
1159 len = end - path[i];
1164 ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1166 (!len || *(end - 1) == '/') ? "" : "/");
1169 if (sizeof(name) != (size_t)ret + 1)
1171 DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1173 handle = dlopen(name, RTLD_LAZY);
1184 DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1187 sym = dlsym(handle, "mlx5_glue");
1188 if (!sym || !*sym) {
1192 DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1201 "cannot initialize PMD due to missing run-time dependency on"
1202 " rdma-core libraries (libibverbs, libmlx5)");
1209 * Driver initialization routine.
1211 RTE_INIT(rte_mlx5_pmd_init);
1213 rte_mlx5_pmd_init(void)
1215 /* Build the static table for ptype conversion. */
1216 mlx5_set_ptype_table();
1218 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1219 * huge pages. Calling ibv_fork_init() during init allows
1220 * applications to use fork() safely for purposes other than
1221 * using this PMD, which is not supported in forked processes.
1223 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1224 /* Match the size of Rx completion entry to the size of a cacheline. */
1225 if (RTE_CACHE_LINE_SIZE == 128)
1226 setenv("MLX5_CQE_SIZE", "128", 0);
1227 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1228 if (mlx5_glue_init())
1233 /* Glue structure must not contain any NULL pointers. */
1237 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1238 assert(((const void *const *)mlx5_glue)[i]);
1241 if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1243 "rdma-core glue \"%s\" mismatch: \"%s\" is required",
1244 mlx5_glue->version, MLX5_GLUE_VERSION);
1247 mlx5_glue->fork_init();
1248 rte_pci_register(&mlx5_driver);
1251 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1252 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1253 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1255 /** Initialize driver log type. */
1256 RTE_INIT(vdev_netvsc_init_log)
1258 mlx5_logtype = rte_log_register("pmd.net.mlx5");
1259 if (mlx5_logtype >= 0)
1260 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);