1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2012 6WIND S.A.
3 * Copyright 2012 Mellanox Technologies, Ltd
8 * mlx4 driver initialization.
23 /* Verbs headers do not support -pedantic. */
25 #pragma GCC diagnostic ignored "-Wpedantic"
27 #include <infiniband/verbs.h>
29 #pragma GCC diagnostic error "-Wpedantic"
32 #include <rte_common.h>
33 #include <rte_config.h>
35 #include <rte_errno.h>
36 #include <rte_ethdev_driver.h>
37 #include <rte_ethdev_pci.h>
38 #include <rte_ether.h>
40 #include <rte_interrupts.h>
41 #include <rte_kvargs.h>
42 #include <rte_malloc.h>
46 #include "mlx4_glue.h"
47 #include "mlx4_flow.h"
49 #include "mlx4_rxtx.h"
50 #include "mlx4_utils.h"
52 static const char *MZ_MLX4_PMD_SHARED_DATA = "mlx4_pmd_shared_data";
54 /* Shared memory between primary and secondary processes. */
55 struct mlx4_shared_data *mlx4_shared_data;
57 /* Spinlock for mlx4_shared_data allocation. */
58 static rte_spinlock_t mlx4_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
60 /* Process local data for secondary processes. */
61 static struct mlx4_local_data mlx4_local_data;
63 /** Configuration structure for device arguments. */
66 uint32_t present; /**< Bit-field for existing ports. */
67 uint32_t enabled; /**< Bit-field for user-enabled ports. */
70 /** Whether memseg should be extended for MR creation. */
73 /* Available parameters list. */
74 const char *pmd_mlx4_init_params[] = {
76 MLX4_MR_EXT_MEMSEG_EN_KVARG,
80 static void mlx4_dev_stop(struct rte_eth_dev *dev);
83 * Initialize shared data between primary and secondary process.
85 * A memzone is reserved by primary process and secondary processes attach to
89 * 0 on success, a negative errno value otherwise and rte_errno is set.
92 mlx4_init_shared_data(void)
94 const struct rte_memzone *mz;
97 rte_spinlock_lock(&mlx4_shared_data_lock);
98 if (mlx4_shared_data == NULL) {
99 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
100 /* Allocate shared memory. */
101 mz = rte_memzone_reserve(MZ_MLX4_PMD_SHARED_DATA,
102 sizeof(*mlx4_shared_data),
105 ERROR("Cannot allocate mlx4 shared data\n");
109 mlx4_shared_data = mz->addr;
110 memset(mlx4_shared_data, 0, sizeof(*mlx4_shared_data));
111 rte_spinlock_init(&mlx4_shared_data->lock);
113 /* Lookup allocated shared memory. */
114 mz = rte_memzone_lookup(MZ_MLX4_PMD_SHARED_DATA);
116 ERROR("Cannot attach mlx4 shared data\n");
120 mlx4_shared_data = mz->addr;
121 memset(&mlx4_local_data, 0, sizeof(mlx4_local_data));
125 rte_spinlock_unlock(&mlx4_shared_data_lock);
129 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
131 * Verbs callback to allocate a memory. This function should allocate the space
132 * according to the size provided residing inside a huge page.
133 * Please note that all allocation must respect the alignment from libmlx4
134 * (i.e. currently sysconf(_SC_PAGESIZE)).
137 * The size in bytes of the memory to allocate.
139 * A pointer to the callback data.
142 * Allocated buffer, NULL otherwise and rte_errno is set.
145 mlx4_alloc_verbs_buf(size_t size, void *data)
147 struct mlx4_priv *priv = data;
149 size_t alignment = sysconf(_SC_PAGESIZE);
150 unsigned int socket = SOCKET_ID_ANY;
152 if (priv->verbs_alloc_ctx.type == MLX4_VERBS_ALLOC_TYPE_TX_QUEUE) {
153 const struct txq *txq = priv->verbs_alloc_ctx.obj;
155 socket = txq->socket;
156 } else if (priv->verbs_alloc_ctx.type ==
157 MLX4_VERBS_ALLOC_TYPE_RX_QUEUE) {
158 const struct rxq *rxq = priv->verbs_alloc_ctx.obj;
160 socket = rxq->socket;
162 assert(data != NULL);
163 ret = rte_malloc_socket(__func__, size, alignment, socket);
170 * Verbs callback to free a memory.
173 * A pointer to the memory to free.
175 * A pointer to the callback data.
178 mlx4_free_verbs_buf(void *ptr, void *data __rte_unused)
180 assert(data != NULL);
186 * Initialize process private data structure.
189 * Pointer to Ethernet device structure.
192 * 0 on success, a negative errno value otherwise and rte_errno is set.
195 mlx4_proc_priv_init(struct rte_eth_dev *dev)
197 struct mlx4_proc_priv *ppriv;
201 * UAR register table follows the process private structure. BlueFlame
202 * registers for Tx queues are stored in the table.
204 ppriv_size = sizeof(struct mlx4_proc_priv) +
205 dev->data->nb_tx_queues * sizeof(void *);
206 ppriv = rte_malloc_socket("mlx4_proc_priv", ppriv_size,
207 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
212 ppriv->uar_table_sz = ppriv_size;
213 dev->process_private = ppriv;
218 * Un-initialize process private data structure.
221 * Pointer to Ethernet device structure.
224 mlx4_proc_priv_uninit(struct rte_eth_dev *dev)
226 if (!dev->process_private)
228 rte_free(dev->process_private);
229 dev->process_private = NULL;
233 * DPDK callback for Ethernet device configuration.
236 * Pointer to Ethernet device structure.
239 * 0 on success, negative errno value otherwise and rte_errno is set.
242 mlx4_dev_configure(struct rte_eth_dev *dev)
244 struct mlx4_priv *priv = dev->data->dev_private;
245 struct rte_flow_error error;
248 /* Prepare internal flow rules. */
249 ret = mlx4_flow_sync(priv, &error);
251 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
252 " flow error type %d, cause %p, message: %s",
253 -ret, strerror(-ret), error.type, error.cause,
254 error.message ? error.message : "(unspecified)");
257 ret = mlx4_intr_install(priv);
259 ERROR("%p: interrupt handler installation failed",
263 ret = mlx4_proc_priv_init(dev);
265 ERROR("%p: process private data allocation failed",
274 * DPDK callback to start the device.
276 * Simulate device start by initializing common RSS resources and attaching
277 * all configured flows.
280 * Pointer to Ethernet device structure.
283 * 0 on success, negative errno value otherwise and rte_errno is set.
286 mlx4_dev_start(struct rte_eth_dev *dev)
288 struct mlx4_priv *priv = dev->data->dev_private;
289 struct rte_flow_error error;
294 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
296 ret = mlx4_rss_init(priv);
298 ERROR("%p: cannot initialize RSS resources: %s",
299 (void *)dev, strerror(-ret));
303 mlx4_mr_dump_dev(dev);
305 ret = mlx4_rxq_intr_enable(priv);
307 ERROR("%p: interrupt handler installation failed",
311 ret = mlx4_flow_sync(priv, &error);
313 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
314 " flow error type %d, cause %p, message: %s",
316 -ret, strerror(-ret), error.type, error.cause,
317 error.message ? error.message : "(unspecified)");
321 dev->tx_pkt_burst = mlx4_tx_burst;
322 dev->rx_pkt_burst = mlx4_rx_burst;
323 /* Enable datapath on secondary process. */
324 mlx4_mp_req_start_rxtx(dev);
332 * DPDK callback to stop the device.
334 * Simulate device stop by detaching all configured flows.
337 * Pointer to Ethernet device structure.
340 mlx4_dev_stop(struct rte_eth_dev *dev)
342 struct mlx4_priv *priv = dev->data->dev_private;
346 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
348 dev->tx_pkt_burst = mlx4_tx_burst_removed;
349 dev->rx_pkt_burst = mlx4_rx_burst_removed;
351 /* Disable datapath on secondary process. */
352 mlx4_mp_req_stop_rxtx(dev);
353 mlx4_flow_sync(priv, NULL);
354 mlx4_rxq_intr_disable(priv);
355 mlx4_rss_deinit(priv);
359 * DPDK callback to close the device.
361 * Destroy all queues and objects, free memory.
364 * Pointer to Ethernet device structure.
367 mlx4_dev_close(struct rte_eth_dev *dev)
369 struct mlx4_priv *priv = dev->data->dev_private;
372 DEBUG("%p: closing device \"%s\"",
374 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
375 dev->rx_pkt_burst = mlx4_rx_burst_removed;
376 dev->tx_pkt_burst = mlx4_tx_burst_removed;
378 /* Disable datapath on secondary process. */
379 mlx4_mp_req_stop_rxtx(dev);
380 mlx4_flow_clean(priv);
381 mlx4_rss_deinit(priv);
382 for (i = 0; i != dev->data->nb_rx_queues; ++i)
383 mlx4_rx_queue_release(dev->data->rx_queues[i]);
384 for (i = 0; i != dev->data->nb_tx_queues; ++i)
385 mlx4_tx_queue_release(dev->data->tx_queues[i]);
386 mlx4_proc_priv_uninit(dev);
387 mlx4_mr_release(dev);
388 if (priv->pd != NULL) {
389 assert(priv->ctx != NULL);
390 claim_zero(mlx4_glue->dealloc_pd(priv->pd));
391 claim_zero(mlx4_glue->close_device(priv->ctx));
393 assert(priv->ctx == NULL);
394 mlx4_intr_uninstall(priv);
395 memset(priv, 0, sizeof(*priv));
398 static const struct eth_dev_ops mlx4_dev_ops = {
399 .dev_configure = mlx4_dev_configure,
400 .dev_start = mlx4_dev_start,
401 .dev_stop = mlx4_dev_stop,
402 .dev_set_link_down = mlx4_dev_set_link_down,
403 .dev_set_link_up = mlx4_dev_set_link_up,
404 .dev_close = mlx4_dev_close,
405 .link_update = mlx4_link_update,
406 .promiscuous_enable = mlx4_promiscuous_enable,
407 .promiscuous_disable = mlx4_promiscuous_disable,
408 .allmulticast_enable = mlx4_allmulticast_enable,
409 .allmulticast_disable = mlx4_allmulticast_disable,
410 .mac_addr_remove = mlx4_mac_addr_remove,
411 .mac_addr_add = mlx4_mac_addr_add,
412 .mac_addr_set = mlx4_mac_addr_set,
413 .stats_get = mlx4_stats_get,
414 .stats_reset = mlx4_stats_reset,
415 .fw_version_get = mlx4_fw_version_get,
416 .dev_infos_get = mlx4_dev_infos_get,
417 .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
418 .vlan_filter_set = mlx4_vlan_filter_set,
419 .rx_queue_setup = mlx4_rx_queue_setup,
420 .tx_queue_setup = mlx4_tx_queue_setup,
421 .rx_queue_release = mlx4_rx_queue_release,
422 .tx_queue_release = mlx4_tx_queue_release,
423 .flow_ctrl_get = mlx4_flow_ctrl_get,
424 .flow_ctrl_set = mlx4_flow_ctrl_set,
425 .mtu_set = mlx4_mtu_set,
426 .filter_ctrl = mlx4_filter_ctrl,
427 .rx_queue_intr_enable = mlx4_rx_intr_enable,
428 .rx_queue_intr_disable = mlx4_rx_intr_disable,
429 .is_removed = mlx4_is_removed,
432 /* Available operations from secondary process. */
433 static const struct eth_dev_ops mlx4_dev_sec_ops = {
434 .stats_get = mlx4_stats_get,
435 .stats_reset = mlx4_stats_reset,
436 .fw_version_get = mlx4_fw_version_get,
437 .dev_infos_get = mlx4_dev_infos_get,
441 * Get PCI information from struct ibv_device.
444 * Pointer to Ethernet device structure.
445 * @param[out] pci_addr
446 * PCI bus address output buffer.
449 * 0 on success, negative errno value otherwise and rte_errno is set.
452 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
453 struct rte_pci_addr *pci_addr)
457 MKSTR(path, "%s/device/uevent", device->ibdev_path);
459 file = fopen(path, "rb");
464 while (fgets(line, sizeof(line), file) == line) {
465 size_t len = strlen(line);
468 /* Truncate long lines. */
469 if (len == (sizeof(line) - 1))
470 while (line[(len - 1)] != '\n') {
474 line[(len - 1)] = ret;
476 /* Extract information. */
479 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
483 &pci_addr->function) == 4) {
493 * Verify and store value for device argument.
496 * Key argument to verify.
498 * Value associated with key.
499 * @param[in, out] conf
500 * Shared configuration data.
503 * 0 on success, negative errno value otherwise and rte_errno is set.
506 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
511 tmp = strtoul(val, NULL, 0);
514 WARN("%s: \"%s\" is not a valid integer", key, val);
517 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
518 uint32_t ports = rte_log2_u32(conf->ports.present + 1);
521 ERROR("port index %lu outside range [0,%" PRIu32 ")",
525 if (!(conf->ports.present & (1 << tmp))) {
527 ERROR("invalid port index %lu", tmp);
530 conf->ports.enabled |= 1 << tmp;
531 } else if (strcmp(MLX4_MR_EXT_MEMSEG_EN_KVARG, key) == 0) {
532 conf->mr_ext_memseg_en = !!tmp;
535 WARN("%s: unknown parameter", key);
542 * Parse device parameters.
545 * Device arguments structure.
548 * 0 on success, negative errno value otherwise and rte_errno is set.
551 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
553 struct rte_kvargs *kvlist;
554 unsigned int arg_count;
560 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
561 if (kvlist == NULL) {
563 ERROR("failed to parse kvargs");
566 /* Process parameters. */
567 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
568 arg_count = rte_kvargs_count(kvlist, pmd_mlx4_init_params[i]);
569 while (arg_count-- > 0) {
570 ret = rte_kvargs_process(kvlist,
571 pmd_mlx4_init_params[i],
572 (int (*)(const char *,
582 rte_kvargs_free(kvlist);
587 * Interpret RSS capabilities reported by device.
589 * This function returns the set of usable Verbs RSS hash fields, kernel
590 * quirks taken into account.
595 * Verbs protection domain.
596 * @param device_attr_ex
597 * Extended device attributes to interpret.
600 * Usable RSS hash fields mask in Verbs format.
603 mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd,
604 struct ibv_device_attr_ex *device_attr_ex)
606 uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask;
607 struct ibv_cq *cq = NULL;
608 struct ibv_wq *wq = NULL;
609 struct ibv_rwq_ind_table *ind = NULL;
610 struct ibv_qp *qp = NULL;
613 WARN("no RSS capabilities reported; disabling support for UDP"
614 " RSS and inner VXLAN RSS");
615 return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
616 IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
617 IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
619 if (!(hw_rss_sup & IBV_RX_HASH_INNER))
622 * Although reported as supported, missing code in some Linux
623 * versions (v4.15, v4.16) prevents the creation of hash QPs with
626 * There is no choice but to attempt to instantiate a temporary RSS
627 * context in order to confirm its support.
629 cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0);
630 wq = cq ? mlx4_glue->create_wq
632 &(struct ibv_wq_init_attr){
633 .wq_type = IBV_WQT_RQ,
639 ind = wq ? mlx4_glue->create_rwq_ind_table
641 &(struct ibv_rwq_ind_table_init_attr){
642 .log_ind_tbl_size = 0,
646 qp = ind ? mlx4_glue->create_qp_ex
648 &(struct ibv_qp_init_attr_ex){
650 (IBV_QP_INIT_ATTR_PD |
651 IBV_QP_INIT_ATTR_RX_HASH |
652 IBV_QP_INIT_ATTR_IND_TABLE),
653 .qp_type = IBV_QPT_RAW_PACKET,
657 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
658 .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
659 .rx_hash_key = mlx4_rss_hash_key_default,
660 .rx_hash_fields_mask = hw_rss_sup,
664 WARN("disabling unusable inner RSS capability due to kernel"
666 hw_rss_sup &= ~IBV_RX_HASH_INNER;
668 claim_zero(mlx4_glue->destroy_qp(qp));
671 claim_zero(mlx4_glue->destroy_rwq_ind_table(ind));
673 claim_zero(mlx4_glue->destroy_wq(wq));
675 claim_zero(mlx4_glue->destroy_cq(cq));
679 static struct rte_pci_driver mlx4_driver;
682 * PMD global initialization.
684 * Independent from individual device, this function initializes global
685 * per-PMD data structures distinguishing primary and secondary processes.
686 * Hence, each initialization is called once per a process.
689 * 0 on success, a negative errno value otherwise and rte_errno is set.
694 struct mlx4_shared_data *sd;
695 struct mlx4_local_data *ld = &mlx4_local_data;
697 if (mlx4_init_shared_data())
699 sd = mlx4_shared_data;
701 rte_spinlock_lock(&sd->lock);
702 switch (rte_eal_process_type()) {
703 case RTE_PROC_PRIMARY:
706 LIST_INIT(&sd->mem_event_cb_list);
707 rte_rwlock_init(&sd->mem_event_rwlock);
708 rte_mem_event_callback_register("MLX4_MEM_EVENT_CB",
709 mlx4_mr_mem_event_cb, NULL);
710 mlx4_mp_init_primary();
711 sd->init_done = true;
713 case RTE_PROC_SECONDARY:
716 mlx4_mp_init_secondary();
718 ld->init_done = true;
723 rte_spinlock_unlock(&sd->lock);
728 * DPDK callback to register a PCI device.
730 * This function creates an Ethernet device for each port of a given
734 * PCI driver structure (mlx4_driver).
736 * PCI device information.
739 * 0 on success, negative errno value otherwise and rte_errno is set.
742 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
744 struct ibv_device **list;
745 struct ibv_device *ibv_dev;
747 struct ibv_context *attr_ctx = NULL;
748 struct ibv_device_attr device_attr;
749 struct ibv_device_attr_ex device_attr_ex;
750 struct mlx4_conf conf = {
752 .mr_ext_memseg_en = 1,
758 err = mlx4_init_once();
760 ERROR("unable to init PMD global data: %s",
761 strerror(rte_errno));
764 assert(pci_drv == &mlx4_driver);
765 list = mlx4_glue->get_device_list(&i);
769 if (rte_errno == ENOSYS)
770 ERROR("cannot list devices, is ib_uverbs loaded?");
775 * For each listed device, check related sysfs entry against
776 * the provided PCI ID.
779 struct rte_pci_addr pci_addr;
782 DEBUG("checking device \"%s\"", list[i]->name);
783 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
785 if ((pci_dev->addr.domain != pci_addr.domain) ||
786 (pci_dev->addr.bus != pci_addr.bus) ||
787 (pci_dev->addr.devid != pci_addr.devid) ||
788 (pci_dev->addr.function != pci_addr.function))
790 vf = (pci_dev->id.device_id ==
791 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
792 INFO("PCI information matches, using device \"%s\" (VF: %s)",
793 list[i]->name, (vf ? "true" : "false"));
794 attr_ctx = mlx4_glue->open_device(list[i]);
798 if (attr_ctx == NULL) {
799 mlx4_glue->free_device_list(list);
803 ERROR("cannot access device, is mlx4_ib loaded?");
807 ERROR("cannot use device, are drivers up to date?");
815 DEBUG("device opened");
816 if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
820 INFO("%u port(s) detected", device_attr.phys_port_cnt);
821 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
822 if (mlx4_args(pci_dev->device.devargs, &conf)) {
823 ERROR("failed to process device arguments");
827 /* Use all ports when none are defined */
828 if (!conf.ports.enabled)
829 conf.ports.enabled = conf.ports.present;
830 /* Retrieve extended device attributes. */
831 if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
835 assert(device_attr.max_sge >= MLX4_MAX_SGE);
836 for (i = 0; i < device_attr.phys_port_cnt; i++) {
837 uint32_t port = i + 1; /* ports are indexed from one */
838 struct ibv_context *ctx = NULL;
839 struct ibv_port_attr port_attr;
840 struct ibv_pd *pd = NULL;
841 struct mlx4_priv *priv = NULL;
842 struct rte_eth_dev *eth_dev = NULL;
843 struct ether_addr mac;
844 char name[RTE_ETH_NAME_MAX_LEN];
846 /* If port is not enabled, skip. */
847 if (!(conf.ports.enabled & (1 << i)))
849 DEBUG("using port %u", port);
850 ctx = mlx4_glue->open_device(ibv_dev);
855 snprintf(name, sizeof(name), "%s port %u",
856 mlx4_glue->get_device_name(ibv_dev), port);
857 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
858 eth_dev = rte_eth_dev_attach_secondary(name);
859 if (eth_dev == NULL) {
860 ERROR("can not attach rte ethdev");
865 priv = eth_dev->data->dev_private;
866 if (!priv->verbs_alloc_ctx.enabled) {
867 ERROR("secondary process is not supported"
868 " due to lack of external allocator"
874 eth_dev->device = &pci_dev->device;
875 eth_dev->dev_ops = &mlx4_dev_sec_ops;
876 err = mlx4_proc_priv_init(eth_dev);
879 /* Receive command fd from primary process. */
880 err = mlx4_mp_req_verbs_cmd_fd(eth_dev);
885 /* Remap UAR for Tx queues. */
886 err = mlx4_tx_uar_init_secondary(eth_dev, err);
892 * Ethdev pointer is still required as input since
893 * the primary device is not accessible from the
896 eth_dev->tx_pkt_burst = mlx4_tx_burst;
897 eth_dev->rx_pkt_burst = mlx4_rx_burst;
898 claim_zero(mlx4_glue->close_device(ctx));
899 rte_eth_copy_pci_info(eth_dev, pci_dev);
900 rte_eth_dev_probing_finish(eth_dev);
903 /* Check port status. */
904 err = mlx4_glue->query_port(ctx, port, &port_attr);
907 ERROR("port query failed: %s", strerror(err));
910 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
912 ERROR("port %d is not configured in Ethernet mode",
916 if (port_attr.state != IBV_PORT_ACTIVE)
917 DEBUG("port %d is not active: \"%s\" (%d)",
918 port, mlx4_glue->port_state_str(port_attr.state),
920 /* Make asynchronous FD non-blocking to handle interrupts. */
921 err = mlx4_fd_set_non_blocking(ctx->async_fd);
923 ERROR("cannot make asynchronous FD non-blocking: %s",
927 /* Allocate protection domain. */
928 pd = mlx4_glue->alloc_pd(ctx);
931 ERROR("PD allocation failure");
934 /* from rte_ethdev.c */
935 priv = rte_zmalloc("ethdev private structure",
937 RTE_CACHE_LINE_SIZE);
940 ERROR("priv allocation failure");
944 priv->device_attr = device_attr;
947 priv->mtu = ETHER_MTU;
949 priv->hw_csum = !!(device_attr.device_cap_flags &
950 IBV_DEVICE_RAW_IP_CSUM);
951 DEBUG("checksum offloading is %ssupported",
952 (priv->hw_csum ? "" : "not "));
953 /* Only ConnectX-3 Pro supports tunneling. */
954 priv->hw_csum_l2tun =
956 (device_attr.vendor_part_id ==
957 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
958 DEBUG("L2 tunnel checksum offloads are %ssupported",
959 priv->hw_csum_l2tun ? "" : "not ");
960 priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd,
962 DEBUG("supported RSS hash fields mask: %016" PRIx64,
964 priv->hw_rss_max_qps =
965 device_attr_ex.rss_caps.max_rwq_indirection_table_size;
966 DEBUG("MAX RSS queues %d", priv->hw_rss_max_qps);
967 priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
968 IBV_RAW_PACKET_CAP_SCATTER_FCS);
969 DEBUG("FCS stripping toggling is %ssupported",
970 priv->hw_fcs_strip ? "" : "not ");
972 ((device_attr_ex.tso_caps.max_tso > 0) &&
973 (device_attr_ex.tso_caps.supported_qpts &
974 (1 << IBV_QPT_RAW_PACKET)));
976 priv->tso_max_payload_sz =
977 device_attr_ex.tso_caps.max_tso;
978 DEBUG("TSO is %ssupported",
979 priv->tso ? "" : "not ");
980 priv->mr_ext_memseg_en = conf.mr_ext_memseg_en;
981 /* Configure the first MAC address by default. */
982 err = mlx4_get_mac(priv, &mac.addr_bytes);
984 ERROR("cannot get MAC address, is mlx4_en loaded?"
985 " (error: %s)", strerror(err));
988 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
990 mac.addr_bytes[0], mac.addr_bytes[1],
991 mac.addr_bytes[2], mac.addr_bytes[3],
992 mac.addr_bytes[4], mac.addr_bytes[5]);
993 /* Register MAC address. */
997 char ifname[IF_NAMESIZE];
999 if (mlx4_get_ifname(priv, &ifname) == 0)
1000 DEBUG("port %u ifname is \"%s\"",
1001 priv->port, ifname);
1003 DEBUG("port %u ifname is unknown", priv->port);
1006 /* Get actual MTU if possible. */
1007 mlx4_mtu_get(priv, &priv->mtu);
1008 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
1009 eth_dev = rte_eth_dev_allocate(name);
1010 if (eth_dev == NULL) {
1012 ERROR("can not allocate rte ethdev");
1015 eth_dev->data->dev_private = priv;
1016 eth_dev->data->mac_addrs = priv->mac;
1017 eth_dev->device = &pci_dev->device;
1018 rte_eth_copy_pci_info(eth_dev, pci_dev);
1019 /* Initialize local interrupt handle for current port. */
1020 priv->intr_handle = (struct rte_intr_handle){
1022 .type = RTE_INTR_HANDLE_EXT,
1025 * Override ethdev interrupt handle pointer with private
1026 * handle instead of that of the parent PCI device used by
1027 * default. This prevents it from being shared between all
1028 * ports of the same PCI device since each of them is
1029 * associated its own Verbs context.
1031 * Rx interrupts in particular require this as the PMD has
1032 * no control over the registration of queue interrupts
1033 * besides setting up eth_dev->intr_handle, the rest is
1034 * handled by rte_intr_rx_ctl().
1036 eth_dev->intr_handle = &priv->intr_handle;
1037 priv->dev_data = eth_dev->data;
1038 eth_dev->dev_ops = &mlx4_dev_ops;
1039 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
1040 /* Hint libmlx4 to use PMD allocator for data plane resources */
1041 struct mlx4dv_ctx_allocators alctr = {
1042 .alloc = &mlx4_alloc_verbs_buf,
1043 .free = &mlx4_free_verbs_buf,
1046 err = mlx4_glue->dv_set_context_attr
1047 (ctx, MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS,
1048 (void *)((uintptr_t)&alctr));
1050 WARN("Verbs external allocator is not supported");
1052 priv->verbs_alloc_ctx.enabled = 1;
1054 /* Bring Ethernet device up. */
1055 DEBUG("forcing Ethernet interface up");
1056 mlx4_dev_set_link_up(eth_dev);
1057 /* Update link status once if waiting for LSC. */
1058 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
1059 mlx4_link_update(eth_dev, 0);
1061 * Once the device is added to the list of memory event
1062 * callback, its global MR cache table cannot be expanded
1063 * on the fly because of deadlock. If it overflows, lookup
1064 * should be done by searching MR list linearly, which is slow.
1066 err = mlx4_mr_btree_init(&priv->mr.cache,
1067 MLX4_MR_BTREE_CACHE_N * 2,
1068 eth_dev->device->numa_node);
1070 /* rte_errno is already set. */
1073 /* Add device to memory callback list. */
1074 rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock);
1075 LIST_INSERT_HEAD(&mlx4_shared_data->mem_event_cb_list,
1076 priv, mem_event_cb);
1077 rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock);
1078 rte_eth_dev_probing_finish(eth_dev);
1082 if (eth_dev != NULL)
1083 eth_dev->data->dev_private = NULL;
1085 claim_zero(mlx4_glue->dealloc_pd(pd));
1087 claim_zero(mlx4_glue->close_device(ctx));
1088 if (eth_dev != NULL) {
1089 /* mac_addrs must not be freed because part of dev_private */
1090 eth_dev->data->mac_addrs = NULL;
1091 rte_eth_dev_release_port(eth_dev);
1096 * XXX if something went wrong in the loop above, there is a resource
1097 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1098 * long as the dpdk does not provide a way to deallocate a ethdev and a
1099 * way to enumerate the registered ethdevs to free the previous ones.
1103 claim_zero(mlx4_glue->close_device(attr_ctx));
1105 mlx4_glue->free_device_list(list);
1111 static const struct rte_pci_id mlx4_pci_id_map[] = {
1113 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1114 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
1117 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1118 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
1121 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1122 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
1129 static struct rte_pci_driver mlx4_driver = {
1131 .name = MLX4_DRIVER_NAME
1133 .id_table = mlx4_pci_id_map,
1134 .probe = mlx4_pci_probe,
1135 .drv_flags = RTE_PCI_DRV_INTR_LSC |
1136 RTE_PCI_DRV_INTR_RMV,
1139 #ifdef RTE_IBVERBS_LINK_DLOPEN
1142 * Suffix RTE_EAL_PMD_PATH with "-glue".
1144 * This function performs a sanity check on RTE_EAL_PMD_PATH before
1145 * suffixing its last component.
1148 * Output buffer, should be large enough otherwise NULL is returned.
1153 * Pointer to @p buf or @p NULL in case suffix cannot be appended.
1156 mlx4_glue_path(char *buf, size_t size)
1158 static const char *const bad[] = { "/", ".", "..", NULL };
1159 const char *path = RTE_EAL_PMD_PATH;
1160 size_t len = strlen(path);
1164 while (len && path[len - 1] == '/')
1166 for (off = len; off && path[off - 1] != '/'; --off)
1168 for (i = 0; bad[i]; ++i)
1169 if (!strncmp(path + off, bad[i], (int)(len - off)))
1171 i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1172 if (i == -1 || (size_t)i >= size)
1176 ERROR("unable to append \"-glue\" to last component of"
1177 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1178 " please re-configure DPDK");
1183 * Initialization routine for run-time dependency on rdma-core.
1186 mlx4_glue_init(void)
1188 char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1189 const char *path[] = {
1191 * A basic security check is necessary before trusting
1192 * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1194 (geteuid() == getuid() && getegid() == getgid() ?
1195 getenv("MLX4_GLUE_PATH") : NULL),
1197 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1198 * variant, otherwise let dlopen() look up libraries on its
1201 (*RTE_EAL_PMD_PATH ?
1202 mlx4_glue_path(glue_path, sizeof(glue_path)) : ""),
1205 void *handle = NULL;
1209 while (!handle && i != RTE_DIM(path)) {
1218 end = strpbrk(path[i], ":;");
1220 end = path[i] + strlen(path[i]);
1221 len = end - path[i];
1226 ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE,
1228 (!len || *(end - 1) == '/') ? "" : "/");
1231 if (sizeof(name) != (size_t)ret + 1)
1233 DEBUG("looking for rdma-core glue as \"%s\"", name);
1234 handle = dlopen(name, RTLD_LAZY);
1245 WARN("cannot load glue library: %s", dlmsg);
1248 sym = dlsym(handle, "mlx4_glue");
1249 if (!sym || !*sym) {
1253 ERROR("cannot resolve glue symbol: %s", dlmsg);
1261 WARN("cannot initialize PMD due to missing run-time"
1262 " dependency on rdma-core libraries (libibverbs,"
1270 * Driver initialization routine.
1272 RTE_INIT(rte_mlx4_pmd_init)
1275 * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
1276 * want to get success errno value in case of calling them
1277 * when the device was removed.
1279 setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
1281 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1282 * huge pages. Calling ibv_fork_init() during init allows
1283 * applications to use fork() safely for purposes other than
1284 * using this PMD, which is not supported in forked processes.
1286 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1287 #ifdef RTE_IBVERBS_LINK_DLOPEN
1288 if (mlx4_glue_init())
1293 /* Glue structure must not contain any NULL pointers. */
1297 for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i)
1298 assert(((const void *const *)mlx4_glue)[i]);
1301 if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) {
1302 ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required",
1303 mlx4_glue->version, MLX4_GLUE_VERSION);
1306 mlx4_glue->fork_init();
1307 rte_pci_register(&mlx4_driver);
1310 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
1311 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
1312 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
1313 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");