1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2012 6WIND S.A.
3 * Copyright 2012 Mellanox Technologies, Ltd
8 * mlx4 driver initialization.
20 #ifdef RTE_IBVERBS_LINK_DLOPEN
24 /* Verbs headers do not support -pedantic. */
26 #pragma GCC diagnostic ignored "-Wpedantic"
28 #include <infiniband/verbs.h>
30 #pragma GCC diagnostic error "-Wpedantic"
33 #include <rte_common.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"
53 const struct mlx4_glue *mlx4_glue;
56 static const char *MZ_MLX4_PMD_SHARED_DATA = "mlx4_pmd_shared_data";
58 /* Shared memory between primary and secondary processes. */
59 struct mlx4_shared_data *mlx4_shared_data;
61 /* Spinlock for mlx4_shared_data allocation. */
62 static rte_spinlock_t mlx4_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
64 /* Process local data for secondary processes. */
65 static struct mlx4_local_data mlx4_local_data;
67 /** Configuration structure for device arguments. */
70 uint32_t present; /**< Bit-field for existing ports. */
71 uint32_t enabled; /**< Bit-field for user-enabled ports. */
74 /** Whether memseg should be extended for MR creation. */
77 /* Available parameters list. */
78 const char *pmd_mlx4_init_params[] = {
80 MLX4_MR_EXT_MEMSEG_EN_KVARG,
84 static void mlx4_dev_stop(struct rte_eth_dev *dev);
87 * Initialize shared data between primary and secondary process.
89 * A memzone is reserved by primary process and secondary processes attach to
93 * 0 on success, a negative errno value otherwise and rte_errno is set.
96 mlx4_init_shared_data(void)
98 const struct rte_memzone *mz;
101 rte_spinlock_lock(&mlx4_shared_data_lock);
102 if (mlx4_shared_data == NULL) {
103 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
104 /* Allocate shared memory. */
105 mz = rte_memzone_reserve(MZ_MLX4_PMD_SHARED_DATA,
106 sizeof(*mlx4_shared_data),
109 ERROR("Cannot allocate mlx4 shared data\n");
113 mlx4_shared_data = mz->addr;
114 memset(mlx4_shared_data, 0, sizeof(*mlx4_shared_data));
115 rte_spinlock_init(&mlx4_shared_data->lock);
117 /* Lookup allocated shared memory. */
118 mz = rte_memzone_lookup(MZ_MLX4_PMD_SHARED_DATA);
120 ERROR("Cannot attach mlx4 shared data\n");
124 mlx4_shared_data = mz->addr;
125 memset(&mlx4_local_data, 0, sizeof(mlx4_local_data));
129 rte_spinlock_unlock(&mlx4_shared_data_lock);
133 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
135 * Verbs callback to allocate a memory. This function should allocate the space
136 * according to the size provided residing inside a huge page.
137 * Please note that all allocation must respect the alignment from libmlx4
138 * (i.e. currently sysconf(_SC_PAGESIZE)).
141 * The size in bytes of the memory to allocate.
143 * A pointer to the callback data.
146 * Allocated buffer, NULL otherwise and rte_errno is set.
149 mlx4_alloc_verbs_buf(size_t size, void *data)
151 struct mlx4_priv *priv = data;
153 size_t alignment = sysconf(_SC_PAGESIZE);
154 unsigned int socket = SOCKET_ID_ANY;
156 if (priv->verbs_alloc_ctx.type == MLX4_VERBS_ALLOC_TYPE_TX_QUEUE) {
157 const struct txq *txq = priv->verbs_alloc_ctx.obj;
159 socket = txq->socket;
160 } else if (priv->verbs_alloc_ctx.type ==
161 MLX4_VERBS_ALLOC_TYPE_RX_QUEUE) {
162 const struct rxq *rxq = priv->verbs_alloc_ctx.obj;
164 socket = rxq->socket;
166 MLX4_ASSERT(data != NULL);
167 ret = rte_malloc_socket(__func__, size, alignment, socket);
174 * Verbs callback to free a memory.
177 * A pointer to the memory to free.
179 * A pointer to the callback data.
182 mlx4_free_verbs_buf(void *ptr, void *data __rte_unused)
184 MLX4_ASSERT(data != NULL);
190 * Initialize process private data structure.
193 * Pointer to Ethernet device structure.
196 * 0 on success, a negative errno value otherwise and rte_errno is set.
199 mlx4_proc_priv_init(struct rte_eth_dev *dev)
201 struct mlx4_proc_priv *ppriv;
205 * UAR register table follows the process private structure. BlueFlame
206 * registers for Tx queues are stored in the table.
208 ppriv_size = sizeof(struct mlx4_proc_priv) +
209 dev->data->nb_tx_queues * sizeof(void *);
210 ppriv = rte_malloc_socket("mlx4_proc_priv", ppriv_size,
211 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
216 ppriv->uar_table_sz = ppriv_size;
217 dev->process_private = ppriv;
222 * Un-initialize process private data structure.
225 * Pointer to Ethernet device structure.
228 mlx4_proc_priv_uninit(struct rte_eth_dev *dev)
230 if (!dev->process_private)
232 rte_free(dev->process_private);
233 dev->process_private = NULL;
237 * DPDK callback for Ethernet device configuration.
240 * Pointer to Ethernet device structure.
243 * 0 on success, negative errno value otherwise and rte_errno is set.
246 mlx4_dev_configure(struct rte_eth_dev *dev)
248 struct mlx4_priv *priv = dev->data->dev_private;
249 struct rte_flow_error error;
252 if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
253 dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
255 /* Prepare internal flow rules. */
256 ret = mlx4_flow_sync(priv, &error);
258 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
259 " flow error type %d, cause %p, message: %s",
260 -ret, strerror(-ret), error.type, error.cause,
261 error.message ? error.message : "(unspecified)");
264 ret = mlx4_intr_install(priv);
266 ERROR("%p: interrupt handler installation failed",
270 ret = mlx4_proc_priv_init(dev);
272 ERROR("%p: process private data allocation failed",
281 * DPDK callback to start the device.
283 * Simulate device start by initializing common RSS resources and attaching
284 * all configured flows.
287 * Pointer to Ethernet device structure.
290 * 0 on success, negative errno value otherwise and rte_errno is set.
293 mlx4_dev_start(struct rte_eth_dev *dev)
295 struct mlx4_priv *priv = dev->data->dev_private;
296 struct rte_flow_error error;
301 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
303 ret = mlx4_rss_init(priv);
305 ERROR("%p: cannot initialize RSS resources: %s",
306 (void *)dev, strerror(-ret));
309 #ifdef RTE_LIBRTE_MLX4_DEBUG
310 mlx4_mr_dump_dev(dev);
312 ret = mlx4_rxq_intr_enable(priv);
314 ERROR("%p: interrupt handler installation failed",
318 ret = mlx4_flow_sync(priv, &error);
320 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
321 " flow error type %d, cause %p, message: %s",
323 -ret, strerror(-ret), error.type, error.cause,
324 error.message ? error.message : "(unspecified)");
328 dev->tx_pkt_burst = mlx4_tx_burst;
329 dev->rx_pkt_burst = mlx4_rx_burst;
330 /* Enable datapath on secondary process. */
331 mlx4_mp_req_start_rxtx(dev);
339 * DPDK callback to stop the device.
341 * Simulate device stop by detaching all configured flows.
344 * Pointer to Ethernet device structure.
347 mlx4_dev_stop(struct rte_eth_dev *dev)
349 struct mlx4_priv *priv = dev->data->dev_private;
353 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
355 dev->tx_pkt_burst = mlx4_tx_burst_removed;
356 dev->rx_pkt_burst = mlx4_rx_burst_removed;
358 /* Disable datapath on secondary process. */
359 mlx4_mp_req_stop_rxtx(dev);
360 mlx4_flow_sync(priv, NULL);
361 mlx4_rxq_intr_disable(priv);
362 mlx4_rss_deinit(priv);
366 * DPDK callback to close the device.
368 * Destroy all queues and objects, free memory.
371 * Pointer to Ethernet device structure.
374 mlx4_dev_close(struct rte_eth_dev *dev)
376 struct mlx4_priv *priv = dev->data->dev_private;
379 DEBUG("%p: closing device \"%s\"",
381 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
382 dev->rx_pkt_burst = mlx4_rx_burst_removed;
383 dev->tx_pkt_burst = mlx4_tx_burst_removed;
385 /* Disable datapath on secondary process. */
386 mlx4_mp_req_stop_rxtx(dev);
387 mlx4_flow_clean(priv);
388 mlx4_rss_deinit(priv);
389 for (i = 0; i != dev->data->nb_rx_queues; ++i)
390 mlx4_rx_queue_release(dev->data->rx_queues[i]);
391 for (i = 0; i != dev->data->nb_tx_queues; ++i)
392 mlx4_tx_queue_release(dev->data->tx_queues[i]);
393 mlx4_proc_priv_uninit(dev);
394 mlx4_mr_release(dev);
395 if (priv->pd != NULL) {
396 MLX4_ASSERT(priv->ctx != NULL);
397 claim_zero(mlx4_glue->dealloc_pd(priv->pd));
398 claim_zero(mlx4_glue->close_device(priv->ctx));
400 MLX4_ASSERT(priv->ctx == NULL);
401 mlx4_intr_uninstall(priv);
402 memset(priv, 0, sizeof(*priv));
406 static const struct eth_dev_ops mlx4_dev_ops = {
407 .dev_configure = mlx4_dev_configure,
408 .dev_start = mlx4_dev_start,
409 .dev_stop = mlx4_dev_stop,
410 .dev_set_link_down = mlx4_dev_set_link_down,
411 .dev_set_link_up = mlx4_dev_set_link_up,
412 .dev_close = mlx4_dev_close,
413 .link_update = mlx4_link_update,
414 .promiscuous_enable = mlx4_promiscuous_enable,
415 .promiscuous_disable = mlx4_promiscuous_disable,
416 .allmulticast_enable = mlx4_allmulticast_enable,
417 .allmulticast_disable = mlx4_allmulticast_disable,
418 .mac_addr_remove = mlx4_mac_addr_remove,
419 .mac_addr_add = mlx4_mac_addr_add,
420 .mac_addr_set = mlx4_mac_addr_set,
421 .set_mc_addr_list = mlx4_set_mc_addr_list,
422 .stats_get = mlx4_stats_get,
423 .stats_reset = mlx4_stats_reset,
424 .fw_version_get = mlx4_fw_version_get,
425 .dev_infos_get = mlx4_dev_infos_get,
426 .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
427 .vlan_filter_set = mlx4_vlan_filter_set,
428 .rx_queue_setup = mlx4_rx_queue_setup,
429 .tx_queue_setup = mlx4_tx_queue_setup,
430 .rx_queue_release = mlx4_rx_queue_release,
431 .tx_queue_release = mlx4_tx_queue_release,
432 .flow_ctrl_get = mlx4_flow_ctrl_get,
433 .flow_ctrl_set = mlx4_flow_ctrl_set,
434 .mtu_set = mlx4_mtu_set,
435 .filter_ctrl = mlx4_filter_ctrl,
436 .rx_queue_intr_enable = mlx4_rx_intr_enable,
437 .rx_queue_intr_disable = mlx4_rx_intr_disable,
438 .is_removed = mlx4_is_removed,
441 /* Available operations from secondary process. */
442 static const struct eth_dev_ops mlx4_dev_sec_ops = {
443 .stats_get = mlx4_stats_get,
444 .stats_reset = mlx4_stats_reset,
445 .fw_version_get = mlx4_fw_version_get,
446 .dev_infos_get = mlx4_dev_infos_get,
450 * Get PCI information from struct ibv_device.
453 * Pointer to Ethernet device structure.
454 * @param[out] pci_addr
455 * PCI bus address output buffer.
458 * 0 on success, negative errno value otherwise and rte_errno is set.
461 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
462 struct rte_pci_addr *pci_addr)
466 MKSTR(path, "%s/device/uevent", device->ibdev_path);
468 file = fopen(path, "rb");
473 while (fgets(line, sizeof(line), file) == line) {
474 size_t len = strlen(line);
477 /* Truncate long lines. */
478 if (len == (sizeof(line) - 1))
479 while (line[(len - 1)] != '\n') {
483 line[(len - 1)] = ret;
485 /* Extract information. */
488 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
492 &pci_addr->function) == 4) {
501 * Verify and store value for device argument.
504 * Key argument to verify.
506 * Value associated with key.
507 * @param[in, out] conf
508 * Shared configuration data.
511 * 0 on success, negative errno value otherwise and rte_errno is set.
514 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
519 tmp = strtoul(val, NULL, 0);
522 WARN("%s: \"%s\" is not a valid integer", key, val);
525 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
526 uint32_t ports = rte_log2_u32(conf->ports.present + 1);
529 ERROR("port index %lu outside range [0,%" PRIu32 ")",
533 if (!(conf->ports.present & (1 << tmp))) {
535 ERROR("invalid port index %lu", tmp);
538 conf->ports.enabled |= 1 << tmp;
539 } else if (strcmp(MLX4_MR_EXT_MEMSEG_EN_KVARG, key) == 0) {
540 conf->mr_ext_memseg_en = !!tmp;
543 WARN("%s: unknown parameter", key);
550 * Parse device parameters.
553 * Device arguments structure.
556 * 0 on success, negative errno value otherwise and rte_errno is set.
559 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
561 struct rte_kvargs *kvlist;
562 unsigned int arg_count;
568 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
569 if (kvlist == NULL) {
571 ERROR("failed to parse kvargs");
574 /* Process parameters. */
575 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
576 arg_count = rte_kvargs_count(kvlist, pmd_mlx4_init_params[i]);
577 while (arg_count-- > 0) {
578 ret = rte_kvargs_process(kvlist,
579 pmd_mlx4_init_params[i],
580 (int (*)(const char *,
590 rte_kvargs_free(kvlist);
595 * Interpret RSS capabilities reported by device.
597 * This function returns the set of usable Verbs RSS hash fields, kernel
598 * quirks taken into account.
603 * Verbs protection domain.
604 * @param device_attr_ex
605 * Extended device attributes to interpret.
608 * Usable RSS hash fields mask in Verbs format.
611 mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd,
612 struct ibv_device_attr_ex *device_attr_ex)
614 uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask;
615 struct ibv_cq *cq = NULL;
616 struct ibv_wq *wq = NULL;
617 struct ibv_rwq_ind_table *ind = NULL;
618 struct ibv_qp *qp = NULL;
621 WARN("no RSS capabilities reported; disabling support for UDP"
622 " RSS and inner VXLAN RSS");
623 return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
624 IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
625 IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
627 if (!(hw_rss_sup & IBV_RX_HASH_INNER))
630 * Although reported as supported, missing code in some Linux
631 * versions (v4.15, v4.16) prevents the creation of hash QPs with
634 * There is no choice but to attempt to instantiate a temporary RSS
635 * context in order to confirm its support.
637 cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0);
638 wq = cq ? mlx4_glue->create_wq
640 &(struct ibv_wq_init_attr){
641 .wq_type = IBV_WQT_RQ,
647 ind = wq ? mlx4_glue->create_rwq_ind_table
649 &(struct ibv_rwq_ind_table_init_attr){
650 .log_ind_tbl_size = 0,
654 qp = ind ? mlx4_glue->create_qp_ex
656 &(struct ibv_qp_init_attr_ex){
658 (IBV_QP_INIT_ATTR_PD |
659 IBV_QP_INIT_ATTR_RX_HASH |
660 IBV_QP_INIT_ATTR_IND_TABLE),
661 .qp_type = IBV_QPT_RAW_PACKET,
665 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
666 .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
667 .rx_hash_key = mlx4_rss_hash_key_default,
668 .rx_hash_fields_mask = hw_rss_sup,
672 WARN("disabling unusable inner RSS capability due to kernel"
674 hw_rss_sup &= ~IBV_RX_HASH_INNER;
676 claim_zero(mlx4_glue->destroy_qp(qp));
679 claim_zero(mlx4_glue->destroy_rwq_ind_table(ind));
681 claim_zero(mlx4_glue->destroy_wq(wq));
683 claim_zero(mlx4_glue->destroy_cq(cq));
687 static struct rte_pci_driver mlx4_driver;
690 * PMD global initialization.
692 * Independent from individual device, this function initializes global
693 * per-PMD data structures distinguishing primary and secondary processes.
694 * Hence, each initialization is called once per a process.
697 * 0 on success, a negative errno value otherwise and rte_errno is set.
702 struct mlx4_shared_data *sd;
703 struct mlx4_local_data *ld = &mlx4_local_data;
706 if (mlx4_init_shared_data())
708 sd = mlx4_shared_data;
710 rte_spinlock_lock(&sd->lock);
711 switch (rte_eal_process_type()) {
712 case RTE_PROC_PRIMARY:
715 LIST_INIT(&sd->mem_event_cb_list);
716 rte_rwlock_init(&sd->mem_event_rwlock);
717 rte_mem_event_callback_register("MLX4_MEM_EVENT_CB",
718 mlx4_mr_mem_event_cb, NULL);
719 ret = mlx4_mp_init_primary();
724 case RTE_PROC_SECONDARY:
727 ret = mlx4_mp_init_secondary();
737 rte_spinlock_unlock(&sd->lock);
742 * DPDK callback to register a PCI device.
744 * This function creates an Ethernet device for each port of a given
748 * PCI driver structure (mlx4_driver).
750 * PCI device information.
753 * 0 on success, negative errno value otherwise and rte_errno is set.
756 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
758 struct ibv_device **list;
759 struct ibv_device *ibv_dev;
761 struct ibv_context *attr_ctx = NULL;
762 struct ibv_device_attr device_attr;
763 struct ibv_device_attr_ex device_attr_ex;
764 struct mlx4_conf conf = {
766 .mr_ext_memseg_en = 1,
770 char ifname[IF_NAMESIZE];
773 err = mlx4_init_once();
775 ERROR("unable to init PMD global data: %s",
776 strerror(rte_errno));
779 MLX4_ASSERT(pci_drv == &mlx4_driver);
780 list = mlx4_glue->get_device_list(&i);
783 MLX4_ASSERT(rte_errno);
784 if (rte_errno == ENOSYS)
785 ERROR("cannot list devices, is ib_uverbs loaded?");
790 * For each listed device, check related sysfs entry against
791 * the provided PCI ID.
794 struct rte_pci_addr pci_addr;
797 DEBUG("checking device \"%s\"", list[i]->name);
798 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
800 if ((pci_dev->addr.domain != pci_addr.domain) ||
801 (pci_dev->addr.bus != pci_addr.bus) ||
802 (pci_dev->addr.devid != pci_addr.devid) ||
803 (pci_dev->addr.function != pci_addr.function))
805 vf = (pci_dev->id.device_id ==
806 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
807 INFO("PCI information matches, using device \"%s\" (VF: %s)",
808 list[i]->name, (vf ? "true" : "false"));
809 attr_ctx = mlx4_glue->open_device(list[i]);
813 if (attr_ctx == NULL) {
814 mlx4_glue->free_device_list(list);
818 ERROR("cannot access device, is mlx4_ib loaded?");
822 ERROR("cannot use device, are drivers up to date?");
825 MLX4_ASSERT(err > 0);
830 DEBUG("device opened");
831 if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
835 INFO("%u port(s) detected", device_attr.phys_port_cnt);
836 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
837 if (mlx4_args(pci_dev->device.devargs, &conf)) {
838 ERROR("failed to process device arguments");
842 /* Use all ports when none are defined */
843 if (!conf.ports.enabled)
844 conf.ports.enabled = conf.ports.present;
845 /* Retrieve extended device attributes. */
846 if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
850 MLX4_ASSERT(device_attr.max_sge >= MLX4_MAX_SGE);
851 for (i = 0; i < device_attr.phys_port_cnt; i++) {
852 uint32_t port = i + 1; /* ports are indexed from one */
853 struct ibv_context *ctx = NULL;
854 struct ibv_port_attr port_attr;
855 struct ibv_pd *pd = NULL;
856 struct mlx4_priv *priv = NULL;
857 struct rte_eth_dev *eth_dev = NULL;
858 struct rte_ether_addr mac;
859 char name[RTE_ETH_NAME_MAX_LEN];
861 /* If port is not enabled, skip. */
862 if (!(conf.ports.enabled & (1 << i)))
864 DEBUG("using port %u", port);
865 ctx = mlx4_glue->open_device(ibv_dev);
870 snprintf(name, sizeof(name), "%s port %u",
871 mlx4_glue->get_device_name(ibv_dev), port);
872 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
873 eth_dev = rte_eth_dev_attach_secondary(name);
874 if (eth_dev == NULL) {
875 ERROR("can not attach rte ethdev");
880 priv = eth_dev->data->dev_private;
881 if (!priv->verbs_alloc_ctx.enabled) {
882 ERROR("secondary process is not supported"
883 " due to lack of external allocator"
889 eth_dev->device = &pci_dev->device;
890 eth_dev->dev_ops = &mlx4_dev_sec_ops;
891 err = mlx4_proc_priv_init(eth_dev);
894 /* Receive command fd from primary process. */
895 err = mlx4_mp_req_verbs_cmd_fd(eth_dev);
900 /* Remap UAR for Tx queues. */
901 err = mlx4_tx_uar_init_secondary(eth_dev, err);
907 * Ethdev pointer is still required as input since
908 * the primary device is not accessible from the
911 eth_dev->tx_pkt_burst = mlx4_tx_burst;
912 eth_dev->rx_pkt_burst = mlx4_rx_burst;
913 claim_zero(mlx4_glue->close_device(ctx));
914 rte_eth_copy_pci_info(eth_dev, pci_dev);
915 rte_eth_dev_probing_finish(eth_dev);
918 /* Check port status. */
919 err = mlx4_glue->query_port(ctx, port, &port_attr);
922 ERROR("port query failed: %s", strerror(err));
925 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
927 ERROR("port %d is not configured in Ethernet mode",
931 if (port_attr.state != IBV_PORT_ACTIVE)
932 DEBUG("port %d is not active: \"%s\" (%d)",
933 port, mlx4_glue->port_state_str(port_attr.state),
935 /* Make asynchronous FD non-blocking to handle interrupts. */
936 err = mlx4_fd_set_non_blocking(ctx->async_fd);
938 ERROR("cannot make asynchronous FD non-blocking: %s",
942 /* Allocate protection domain. */
943 pd = mlx4_glue->alloc_pd(ctx);
946 ERROR("PD allocation failure");
949 /* from rte_ethdev.c */
950 priv = rte_zmalloc("ethdev private structure",
952 RTE_CACHE_LINE_SIZE);
955 ERROR("priv allocation failure");
959 priv->device_attr = device_attr;
962 priv->mtu = RTE_ETHER_MTU;
964 priv->hw_csum = !!(device_attr.device_cap_flags &
965 IBV_DEVICE_RAW_IP_CSUM);
966 DEBUG("checksum offloading is %ssupported",
967 (priv->hw_csum ? "" : "not "));
968 /* Only ConnectX-3 Pro supports tunneling. */
969 priv->hw_csum_l2tun =
971 (device_attr.vendor_part_id ==
972 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
973 DEBUG("L2 tunnel checksum offloads are %ssupported",
974 priv->hw_csum_l2tun ? "" : "not ");
975 priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd,
977 DEBUG("supported RSS hash fields mask: %016" PRIx64,
979 priv->hw_rss_max_qps =
980 device_attr_ex.rss_caps.max_rwq_indirection_table_size;
981 DEBUG("MAX RSS queues %d", priv->hw_rss_max_qps);
982 priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
983 IBV_RAW_PACKET_CAP_SCATTER_FCS);
984 DEBUG("FCS stripping toggling is %ssupported",
985 priv->hw_fcs_strip ? "" : "not ");
987 ((device_attr_ex.tso_caps.max_tso > 0) &&
988 (device_attr_ex.tso_caps.supported_qpts &
989 (1 << IBV_QPT_RAW_PACKET)));
991 priv->tso_max_payload_sz =
992 device_attr_ex.tso_caps.max_tso;
993 DEBUG("TSO is %ssupported",
994 priv->tso ? "" : "not ");
995 priv->mr_ext_memseg_en = conf.mr_ext_memseg_en;
996 /* Configure the first MAC address by default. */
997 err = mlx4_get_mac(priv, &mac.addr_bytes);
999 ERROR("cannot get MAC address, is mlx4_en loaded?"
1000 " (error: %s)", strerror(err));
1003 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1005 mac.addr_bytes[0], mac.addr_bytes[1],
1006 mac.addr_bytes[2], mac.addr_bytes[3],
1007 mac.addr_bytes[4], mac.addr_bytes[5]);
1008 /* Register MAC address. */
1011 if (mlx4_get_ifname(priv, &ifname) == 0) {
1012 DEBUG("port %u ifname is \"%s\"",
1013 priv->port, ifname);
1014 priv->if_index = if_nametoindex(ifname);
1016 DEBUG("port %u ifname is unknown", priv->port);
1019 /* Get actual MTU if possible. */
1020 mlx4_mtu_get(priv, &priv->mtu);
1021 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
1022 eth_dev = rte_eth_dev_allocate(name);
1023 if (eth_dev == NULL) {
1025 ERROR("can not allocate rte ethdev");
1028 eth_dev->data->dev_private = priv;
1029 eth_dev->data->mac_addrs = priv->mac;
1030 eth_dev->device = &pci_dev->device;
1031 rte_eth_copy_pci_info(eth_dev, pci_dev);
1032 /* Initialize local interrupt handle for current port. */
1033 memset(&priv->intr_handle, 0, sizeof(struct rte_intr_handle));
1034 priv->intr_handle.fd = -1;
1035 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1037 * Override ethdev interrupt handle pointer with private
1038 * handle instead of that of the parent PCI device used by
1039 * default. This prevents it from being shared between all
1040 * ports of the same PCI device since each of them is
1041 * associated its own Verbs context.
1043 * Rx interrupts in particular require this as the PMD has
1044 * no control over the registration of queue interrupts
1045 * besides setting up eth_dev->intr_handle, the rest is
1046 * handled by rte_intr_rx_ctl().
1048 eth_dev->intr_handle = &priv->intr_handle;
1049 priv->dev_data = eth_dev->data;
1050 eth_dev->dev_ops = &mlx4_dev_ops;
1051 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
1052 /* Hint libmlx4 to use PMD allocator for data plane resources */
1053 err = mlx4_glue->dv_set_context_attr
1054 (ctx, MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS,
1055 (void *)((uintptr_t)&(struct mlx4dv_ctx_allocators){
1056 .alloc = &mlx4_alloc_verbs_buf,
1057 .free = &mlx4_free_verbs_buf,
1061 WARN("Verbs external allocator is not supported");
1063 priv->verbs_alloc_ctx.enabled = 1;
1065 /* Bring Ethernet device up. */
1066 DEBUG("forcing Ethernet interface up");
1067 mlx4_dev_set_link_up(eth_dev);
1068 /* Update link status once if waiting for LSC. */
1069 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
1070 mlx4_link_update(eth_dev, 0);
1072 * Once the device is added to the list of memory event
1073 * callback, its global MR cache table cannot be expanded
1074 * on the fly because of deadlock. If it overflows, lookup
1075 * should be done by searching MR list linearly, which is slow.
1077 err = mlx4_mr_btree_init(&priv->mr.cache,
1078 MLX4_MR_BTREE_CACHE_N * 2,
1079 eth_dev->device->numa_node);
1081 /* rte_errno is already set. */
1084 /* Add device to memory callback list. */
1085 rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock);
1086 LIST_INSERT_HEAD(&mlx4_shared_data->mem_event_cb_list,
1087 priv, mem_event_cb);
1088 rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock);
1089 rte_eth_dev_probing_finish(eth_dev);
1093 if (eth_dev != NULL)
1094 eth_dev->data->dev_private = NULL;
1096 claim_zero(mlx4_glue->dealloc_pd(pd));
1098 claim_zero(mlx4_glue->close_device(ctx));
1099 if (eth_dev != NULL) {
1100 /* mac_addrs must not be freed because part of dev_private */
1101 eth_dev->data->mac_addrs = NULL;
1102 rte_eth_dev_release_port(eth_dev);
1107 * XXX if something went wrong in the loop above, there is a resource
1108 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1109 * long as the dpdk does not provide a way to deallocate a ethdev and a
1110 * way to enumerate the registered ethdevs to free the previous ones.
1114 claim_zero(mlx4_glue->close_device(attr_ctx));
1116 mlx4_glue->free_device_list(list);
1122 static const struct rte_pci_id mlx4_pci_id_map[] = {
1124 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1125 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
1128 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1129 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
1132 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1133 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
1140 static struct rte_pci_driver mlx4_driver = {
1142 .name = MLX4_DRIVER_NAME
1144 .id_table = mlx4_pci_id_map,
1145 .probe = mlx4_pci_probe,
1146 .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1149 #ifdef RTE_IBVERBS_LINK_DLOPEN
1152 * Suffix RTE_EAL_PMD_PATH with "-glue".
1154 * This function performs a sanity check on RTE_EAL_PMD_PATH before
1155 * suffixing its last component.
1158 * Output buffer, should be large enough otherwise NULL is returned.
1163 * Pointer to @p buf or @p NULL in case suffix cannot be appended.
1166 mlx4_glue_path(char *buf, size_t size)
1168 static const char *const bad[] = { "/", ".", "..", NULL };
1169 const char *path = RTE_EAL_PMD_PATH;
1170 size_t len = strlen(path);
1174 while (len && path[len - 1] == '/')
1176 for (off = len; off && path[off - 1] != '/'; --off)
1178 for (i = 0; bad[i]; ++i)
1179 if (!strncmp(path + off, bad[i], (int)(len - off)))
1181 i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1182 if (i == -1 || (size_t)i >= size)
1186 ERROR("unable to append \"-glue\" to last component of"
1187 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1188 " please re-configure DPDK");
1193 * Initialization routine for run-time dependency on rdma-core.
1196 mlx4_glue_init(void)
1198 char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1199 const char *path[] = {
1201 * A basic security check is necessary before trusting
1202 * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1204 (geteuid() == getuid() && getegid() == getgid() ?
1205 getenv("MLX4_GLUE_PATH") : NULL),
1207 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1208 * variant, otherwise let dlopen() look up libraries on its
1211 (*RTE_EAL_PMD_PATH ?
1212 mlx4_glue_path(glue_path, sizeof(glue_path)) : ""),
1215 void *handle = NULL;
1219 while (!handle && i != RTE_DIM(path)) {
1228 end = strpbrk(path[i], ":;");
1230 end = path[i] + strlen(path[i]);
1231 len = end - path[i];
1236 ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE,
1238 (!len || *(end - 1) == '/') ? "" : "/");
1241 if (sizeof(name) != (size_t)ret + 1)
1243 DEBUG("looking for rdma-core glue as \"%s\"", name);
1244 handle = dlopen(name, RTLD_LAZY);
1255 WARN("cannot load glue library: %s", dlmsg);
1258 sym = dlsym(handle, "mlx4_glue");
1259 if (!sym || !*sym) {
1263 ERROR("cannot resolve glue symbol: %s", dlmsg);
1271 WARN("cannot initialize PMD due to missing run-time"
1272 " dependency on rdma-core libraries (libibverbs,"
1279 /* Initialize driver log type. */
1280 RTE_LOG_REGISTER(mlx4_logtype, pmd.net.mlx4, NOTICE)
1283 * Driver initialization routine.
1285 RTE_INIT(rte_mlx4_pmd_init)
1288 * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
1289 * want to get success errno value in case of calling them
1290 * when the device was removed.
1292 setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
1294 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1295 * huge pages. Calling ibv_fork_init() during init allows
1296 * applications to use fork() safely for purposes other than
1297 * using this PMD, which is not supported in forked processes.
1299 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1300 #ifdef RTE_IBVERBS_LINK_DLOPEN
1301 if (mlx4_glue_init())
1303 MLX4_ASSERT(mlx4_glue);
1305 #ifdef RTE_LIBRTE_MLX4_DEBUG
1306 /* Glue structure must not contain any NULL pointers. */
1310 for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i)
1311 MLX4_ASSERT(((const void *const *)mlx4_glue)[i]);
1314 if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) {
1315 ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required",
1316 mlx4_glue->version, MLX4_GLUE_VERSION);
1319 mlx4_glue->fork_init();
1320 rte_pci_register(&mlx4_driver);
1323 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
1324 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
1325 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
1326 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");