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. */
71 /* Available parameters list. */
72 const char *pmd_mlx4_init_params[] = {
77 static void mlx4_dev_stop(struct rte_eth_dev *dev);
80 * Initialize shared data between primary and secondary process.
82 * A memzone is reserved by primary process and secondary processes attach to
86 * 0 on success, a negative errno value otherwise and rte_errno is set.
89 mlx4_init_shared_data(void)
91 const struct rte_memzone *mz;
94 rte_spinlock_lock(&mlx4_shared_data_lock);
95 if (mlx4_shared_data == NULL) {
96 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
97 /* Allocate shared memory. */
98 mz = rte_memzone_reserve(MZ_MLX4_PMD_SHARED_DATA,
99 sizeof(*mlx4_shared_data),
102 ERROR("Cannot allocate mlx4 shared data\n");
106 mlx4_shared_data = mz->addr;
107 memset(mlx4_shared_data, 0, sizeof(*mlx4_shared_data));
108 rte_spinlock_init(&mlx4_shared_data->lock);
110 /* Lookup allocated shared memory. */
111 mz = rte_memzone_lookup(MZ_MLX4_PMD_SHARED_DATA);
113 ERROR("Cannot attach mlx4 shared data\n");
117 mlx4_shared_data = mz->addr;
118 memset(&mlx4_local_data, 0, sizeof(mlx4_local_data));
122 rte_spinlock_unlock(&mlx4_shared_data_lock);
127 * Uninitialize shared data between primary and secondary process.
129 * The pointer of secondary process is dereferenced and primary process frees
133 mlx4_uninit_shared_data(void)
135 const struct rte_memzone *mz;
137 rte_spinlock_lock(&mlx4_shared_data_lock);
138 if (mlx4_shared_data) {
139 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
140 mz = rte_memzone_lookup(MZ_MLX4_PMD_SHARED_DATA);
141 rte_memzone_free(mz);
143 memset(&mlx4_local_data, 0, sizeof(mlx4_local_data));
145 mlx4_shared_data = NULL;
147 rte_spinlock_unlock(&mlx4_shared_data_lock);
150 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
152 * Verbs callback to allocate a memory. This function should allocate the space
153 * according to the size provided residing inside a huge page.
154 * Please note that all allocation must respect the alignment from libmlx4
155 * (i.e. currently sysconf(_SC_PAGESIZE)).
158 * The size in bytes of the memory to allocate.
160 * A pointer to the callback data.
163 * Allocated buffer, NULL otherwise and rte_errno is set.
166 mlx4_alloc_verbs_buf(size_t size, void *data)
168 struct mlx4_priv *priv = data;
170 size_t alignment = sysconf(_SC_PAGESIZE);
171 unsigned int socket = SOCKET_ID_ANY;
173 if (priv->verbs_alloc_ctx.type == MLX4_VERBS_ALLOC_TYPE_TX_QUEUE) {
174 const struct txq *txq = priv->verbs_alloc_ctx.obj;
176 socket = txq->socket;
177 } else if (priv->verbs_alloc_ctx.type ==
178 MLX4_VERBS_ALLOC_TYPE_RX_QUEUE) {
179 const struct rxq *rxq = priv->verbs_alloc_ctx.obj;
181 socket = rxq->socket;
183 assert(data != NULL);
184 ret = rte_malloc_socket(__func__, size, alignment, socket);
191 * Verbs callback to free a memory.
194 * A pointer to the memory to free.
196 * A pointer to the callback data.
199 mlx4_free_verbs_buf(void *ptr, void *data __rte_unused)
201 assert(data != NULL);
207 * DPDK callback for Ethernet device configuration.
210 * Pointer to Ethernet device structure.
213 * 0 on success, negative errno value otherwise and rte_errno is set.
216 mlx4_dev_configure(struct rte_eth_dev *dev)
218 struct mlx4_priv *priv = dev->data->dev_private;
219 struct rte_flow_error error;
222 /* Prepare internal flow rules. */
223 ret = mlx4_flow_sync(priv, &error);
225 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
226 " flow error type %d, cause %p, message: %s",
227 -ret, strerror(-ret), error.type, error.cause,
228 error.message ? error.message : "(unspecified)");
231 ret = mlx4_intr_install(priv);
233 ERROR("%p: interrupt handler installation failed",
240 * DPDK callback to start the device.
242 * Simulate device start by initializing common RSS resources and attaching
243 * all configured flows.
246 * Pointer to Ethernet device structure.
249 * 0 on success, negative errno value otherwise and rte_errno is set.
252 mlx4_dev_start(struct rte_eth_dev *dev)
254 struct mlx4_priv *priv = dev->data->dev_private;
255 struct rte_flow_error error;
260 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
262 ret = mlx4_tx_uar_remap(dev, priv->ctx->cmd_fd);
264 ERROR("%p: cannot remap UAR", (void *)dev);
267 ret = mlx4_rss_init(priv);
269 ERROR("%p: cannot initialize RSS resources: %s",
270 (void *)dev, strerror(-ret));
274 mlx4_mr_dump_dev(dev);
276 ret = mlx4_rxq_intr_enable(priv);
278 ERROR("%p: interrupt handler installation failed",
282 ret = mlx4_flow_sync(priv, &error);
284 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
285 " flow error type %d, cause %p, message: %s",
287 -ret, strerror(-ret), error.type, error.cause,
288 error.message ? error.message : "(unspecified)");
292 dev->tx_pkt_burst = mlx4_tx_burst;
293 dev->rx_pkt_burst = mlx4_rx_burst;
294 /* Enable datapath on secondary process. */
295 mlx4_mp_req_start_rxtx(dev);
303 * DPDK callback to stop the device.
305 * Simulate device stop by detaching all configured flows.
308 * Pointer to Ethernet device structure.
311 mlx4_dev_stop(struct rte_eth_dev *dev)
313 struct mlx4_priv *priv = dev->data->dev_private;
314 const size_t page_size = sysconf(_SC_PAGESIZE);
319 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
321 dev->tx_pkt_burst = mlx4_tx_burst_removed;
322 dev->rx_pkt_burst = mlx4_rx_burst_removed;
324 /* Disable datapath on secondary process. */
325 mlx4_mp_req_stop_rxtx(dev);
326 mlx4_flow_sync(priv, NULL);
327 mlx4_rxq_intr_disable(priv);
328 mlx4_rss_deinit(priv);
329 for (i = 0; i != dev->data->nb_tx_queues; ++i) {
332 txq = dev->data->tx_queues[i];
335 munmap((void *)RTE_ALIGN_FLOOR((uintptr_t)txq->msq.db,
336 page_size), page_size);
341 * DPDK callback to close the device.
343 * Destroy all queues and objects, free memory.
346 * Pointer to Ethernet device structure.
349 mlx4_dev_close(struct rte_eth_dev *dev)
351 struct mlx4_priv *priv = dev->data->dev_private;
354 DEBUG("%p: closing device \"%s\"",
356 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
357 dev->rx_pkt_burst = mlx4_rx_burst_removed;
358 dev->tx_pkt_burst = mlx4_tx_burst_removed;
360 /* Disable datapath on secondary process. */
361 mlx4_mp_req_stop_rxtx(dev);
362 mlx4_flow_clean(priv);
363 mlx4_rss_deinit(priv);
364 for (i = 0; i != dev->data->nb_rx_queues; ++i)
365 mlx4_rx_queue_release(dev->data->rx_queues[i]);
366 for (i = 0; i != dev->data->nb_tx_queues; ++i)
367 mlx4_tx_queue_release(dev->data->tx_queues[i]);
368 mlx4_mr_release(dev);
369 if (priv->pd != NULL) {
370 assert(priv->ctx != NULL);
371 claim_zero(mlx4_glue->dealloc_pd(priv->pd));
372 claim_zero(mlx4_glue->close_device(priv->ctx));
374 assert(priv->ctx == NULL);
375 mlx4_intr_uninstall(priv);
376 memset(priv, 0, sizeof(*priv));
379 static const struct eth_dev_ops mlx4_dev_ops = {
380 .dev_configure = mlx4_dev_configure,
381 .dev_start = mlx4_dev_start,
382 .dev_stop = mlx4_dev_stop,
383 .dev_set_link_down = mlx4_dev_set_link_down,
384 .dev_set_link_up = mlx4_dev_set_link_up,
385 .dev_close = mlx4_dev_close,
386 .link_update = mlx4_link_update,
387 .promiscuous_enable = mlx4_promiscuous_enable,
388 .promiscuous_disable = mlx4_promiscuous_disable,
389 .allmulticast_enable = mlx4_allmulticast_enable,
390 .allmulticast_disable = mlx4_allmulticast_disable,
391 .mac_addr_remove = mlx4_mac_addr_remove,
392 .mac_addr_add = mlx4_mac_addr_add,
393 .mac_addr_set = mlx4_mac_addr_set,
394 .stats_get = mlx4_stats_get,
395 .stats_reset = mlx4_stats_reset,
396 .fw_version_get = mlx4_fw_version_get,
397 .dev_infos_get = mlx4_dev_infos_get,
398 .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
399 .vlan_filter_set = mlx4_vlan_filter_set,
400 .rx_queue_setup = mlx4_rx_queue_setup,
401 .tx_queue_setup = mlx4_tx_queue_setup,
402 .rx_queue_release = mlx4_rx_queue_release,
403 .tx_queue_release = mlx4_tx_queue_release,
404 .flow_ctrl_get = mlx4_flow_ctrl_get,
405 .flow_ctrl_set = mlx4_flow_ctrl_set,
406 .mtu_set = mlx4_mtu_set,
407 .filter_ctrl = mlx4_filter_ctrl,
408 .rx_queue_intr_enable = mlx4_rx_intr_enable,
409 .rx_queue_intr_disable = mlx4_rx_intr_disable,
410 .is_removed = mlx4_is_removed,
413 /* Available operations from secondary process. */
414 static const struct eth_dev_ops mlx4_dev_sec_ops = {
415 .stats_get = mlx4_stats_get,
416 .stats_reset = mlx4_stats_reset,
417 .fw_version_get = mlx4_fw_version_get,
418 .dev_infos_get = mlx4_dev_infos_get,
422 * Get PCI information from struct ibv_device.
425 * Pointer to Ethernet device structure.
426 * @param[out] pci_addr
427 * PCI bus address output buffer.
430 * 0 on success, negative errno value otherwise and rte_errno is set.
433 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
434 struct rte_pci_addr *pci_addr)
438 MKSTR(path, "%s/device/uevent", device->ibdev_path);
440 file = fopen(path, "rb");
445 while (fgets(line, sizeof(line), file) == line) {
446 size_t len = strlen(line);
449 /* Truncate long lines. */
450 if (len == (sizeof(line) - 1))
451 while (line[(len - 1)] != '\n') {
455 line[(len - 1)] = ret;
457 /* Extract information. */
460 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
464 &pci_addr->function) == 4) {
474 * Verify and store value for device argument.
477 * Key argument to verify.
479 * Value associated with key.
480 * @param[in, out] conf
481 * Shared configuration data.
484 * 0 on success, negative errno value otherwise and rte_errno is set.
487 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
492 tmp = strtoul(val, NULL, 0);
495 WARN("%s: \"%s\" is not a valid integer", key, val);
498 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
499 uint32_t ports = rte_log2_u32(conf->ports.present + 1);
502 ERROR("port index %lu outside range [0,%" PRIu32 ")",
506 if (!(conf->ports.present & (1 << tmp))) {
508 ERROR("invalid port index %lu", tmp);
511 conf->ports.enabled |= 1 << tmp;
514 WARN("%s: unknown parameter", key);
521 * Parse device parameters.
524 * Device arguments structure.
527 * 0 on success, negative errno value otherwise and rte_errno is set.
530 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
532 struct rte_kvargs *kvlist;
533 unsigned int arg_count;
539 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
540 if (kvlist == NULL) {
542 ERROR("failed to parse kvargs");
545 /* Process parameters. */
546 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
547 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
548 while (arg_count-- > 0) {
549 ret = rte_kvargs_process(kvlist,
551 (int (*)(const char *,
561 rte_kvargs_free(kvlist);
566 * Interpret RSS capabilities reported by device.
568 * This function returns the set of usable Verbs RSS hash fields, kernel
569 * quirks taken into account.
574 * Verbs protection domain.
575 * @param device_attr_ex
576 * Extended device attributes to interpret.
579 * Usable RSS hash fields mask in Verbs format.
582 mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd,
583 struct ibv_device_attr_ex *device_attr_ex)
585 uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask;
586 struct ibv_cq *cq = NULL;
587 struct ibv_wq *wq = NULL;
588 struct ibv_rwq_ind_table *ind = NULL;
589 struct ibv_qp *qp = NULL;
592 WARN("no RSS capabilities reported; disabling support for UDP"
593 " RSS and inner VXLAN RSS");
594 return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
595 IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
596 IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
598 if (!(hw_rss_sup & IBV_RX_HASH_INNER))
601 * Although reported as supported, missing code in some Linux
602 * versions (v4.15, v4.16) prevents the creation of hash QPs with
605 * There is no choice but to attempt to instantiate a temporary RSS
606 * context in order to confirm its support.
608 cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0);
609 wq = cq ? mlx4_glue->create_wq
611 &(struct ibv_wq_init_attr){
612 .wq_type = IBV_WQT_RQ,
618 ind = wq ? mlx4_glue->create_rwq_ind_table
620 &(struct ibv_rwq_ind_table_init_attr){
621 .log_ind_tbl_size = 0,
625 qp = ind ? mlx4_glue->create_qp_ex
627 &(struct ibv_qp_init_attr_ex){
629 (IBV_QP_INIT_ATTR_PD |
630 IBV_QP_INIT_ATTR_RX_HASH |
631 IBV_QP_INIT_ATTR_IND_TABLE),
632 .qp_type = IBV_QPT_RAW_PACKET,
636 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
637 .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
638 .rx_hash_key = mlx4_rss_hash_key_default,
639 .rx_hash_fields_mask = hw_rss_sup,
643 WARN("disabling unusable inner RSS capability due to kernel"
645 hw_rss_sup &= ~IBV_RX_HASH_INNER;
647 claim_zero(mlx4_glue->destroy_qp(qp));
650 claim_zero(mlx4_glue->destroy_rwq_ind_table(ind));
652 claim_zero(mlx4_glue->destroy_wq(wq));
654 claim_zero(mlx4_glue->destroy_cq(cq));
658 static struct rte_pci_driver mlx4_driver;
661 find_lower_va_bound(const struct rte_memseg_list *msl,
662 const struct rte_memseg *ms, void *arg)
671 *addr = RTE_MIN(*addr, ms->addr);
677 * Reserve UAR address space for primary process.
679 * Process local resource is used by both primary and secondary to avoid
680 * duplicate reservation. The space has to be available on both primary and
681 * secondary process, TXQ UAR maps to this area using fixed mmap w/o double
685 * 0 on success, a negative errno value otherwise and rte_errno is set.
688 mlx4_uar_init_primary(void)
690 struct mlx4_shared_data *sd = mlx4_shared_data;
691 void *addr = (void *)0;
695 /* find out lower bound of hugepage segments */
696 rte_memseg_walk(find_lower_va_bound, &addr);
697 /* keep distance to hugepages to minimize potential conflicts. */
698 addr = RTE_PTR_SUB(addr, (uintptr_t)(MLX4_UAR_OFFSET + MLX4_UAR_SIZE));
699 /* anonymous mmap, no real memory consumption. */
700 addr = mmap(addr, MLX4_UAR_SIZE,
701 PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
702 if (addr == MAP_FAILED) {
703 ERROR("failed to reserve UAR address space, please"
704 " adjust MLX4_UAR_SIZE or try --base-virtaddr");
708 /* Accept either same addr or a new addr returned from mmap if target
711 INFO("reserved UAR address space: %p", addr);
712 sd->uar_base = addr; /* for primary and secondary UAR re-mmap. */
717 * Unmap UAR address space reserved for primary process.
720 mlx4_uar_uninit_primary(void)
722 struct mlx4_shared_data *sd = mlx4_shared_data;
726 munmap(sd->uar_base, MLX4_UAR_SIZE);
731 * Reserve UAR address space for secondary process, align with primary process.
734 * 0 on success, a negative errno value otherwise and rte_errno is set.
737 mlx4_uar_init_secondary(void)
739 struct mlx4_shared_data *sd = mlx4_shared_data;
740 struct mlx4_local_data *ld = &mlx4_local_data;
743 if (ld->uar_base) { /* Already reserved. */
744 assert(sd->uar_base == ld->uar_base);
747 assert(sd->uar_base);
748 /* anonymous mmap, no real memory consumption. */
749 addr = mmap(sd->uar_base, MLX4_UAR_SIZE,
750 PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
751 if (addr == MAP_FAILED) {
752 ERROR("UAR mmap failed: %p size: %llu",
753 sd->uar_base, MLX4_UAR_SIZE);
757 if (sd->uar_base != addr) {
758 ERROR("UAR address %p size %llu occupied, please"
759 " adjust MLX4_UAR_OFFSET or try EAL parameter"
761 sd->uar_base, MLX4_UAR_SIZE);
766 INFO("reserved UAR address space: %p", addr);
771 * Unmap UAR address space reserved for secondary process.
774 mlx4_uar_uninit_secondary(void)
776 struct mlx4_local_data *ld = &mlx4_local_data;
780 munmap(ld->uar_base, MLX4_UAR_SIZE);
785 * PMD global initialization.
787 * Independent from individual device, this function initializes global
788 * per-PMD data structures distinguishing primary and secondary processes.
789 * Hence, each initialization is called once per a process.
792 * 0 on success, a negative errno value otherwise and rte_errno is set.
797 struct mlx4_shared_data *sd;
798 struct mlx4_local_data *ld = &mlx4_local_data;
801 if (mlx4_init_shared_data())
803 sd = mlx4_shared_data;
805 rte_spinlock_lock(&sd->lock);
806 switch (rte_eal_process_type()) {
807 case RTE_PROC_PRIMARY:
810 LIST_INIT(&sd->mem_event_cb_list);
811 rte_rwlock_init(&sd->mem_event_rwlock);
812 rte_mem_event_callback_register("MLX4_MEM_EVENT_CB",
813 mlx4_mr_mem_event_cb, NULL);
814 mlx4_mp_init_primary();
815 ret = mlx4_uar_init_primary();
818 sd->init_done = true;
820 case RTE_PROC_SECONDARY:
823 mlx4_mp_init_secondary();
824 ret = mlx4_uar_init_secondary();
828 ld->init_done = true;
833 rte_spinlock_unlock(&sd->lock);
836 switch (rte_eal_process_type()) {
837 case RTE_PROC_PRIMARY:
838 mlx4_uar_uninit_primary();
839 mlx4_mp_uninit_primary();
840 rte_mem_event_callback_unregister("MLX4_MEM_EVENT_CB", NULL);
842 case RTE_PROC_SECONDARY:
843 mlx4_uar_uninit_secondary();
844 mlx4_mp_uninit_secondary();
849 rte_spinlock_unlock(&sd->lock);
850 mlx4_uninit_shared_data();
855 * DPDK callback to register a PCI device.
857 * This function creates an Ethernet device for each port of a given
861 * PCI driver structure (mlx4_driver).
863 * PCI device information.
866 * 0 on success, negative errno value otherwise and rte_errno is set.
869 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
871 struct ibv_device **list;
872 struct ibv_device *ibv_dev;
874 struct ibv_context *attr_ctx = NULL;
875 struct ibv_device_attr device_attr;
876 struct ibv_device_attr_ex device_attr_ex;
877 struct mlx4_conf conf = {
884 err = mlx4_init_once();
886 ERROR("unable to init PMD global data: %s",
887 strerror(rte_errno));
890 assert(pci_drv == &mlx4_driver);
891 list = mlx4_glue->get_device_list(&i);
895 if (rte_errno == ENOSYS)
896 ERROR("cannot list devices, is ib_uverbs loaded?");
901 * For each listed device, check related sysfs entry against
902 * the provided PCI ID.
905 struct rte_pci_addr pci_addr;
908 DEBUG("checking device \"%s\"", list[i]->name);
909 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
911 if ((pci_dev->addr.domain != pci_addr.domain) ||
912 (pci_dev->addr.bus != pci_addr.bus) ||
913 (pci_dev->addr.devid != pci_addr.devid) ||
914 (pci_dev->addr.function != pci_addr.function))
916 vf = (pci_dev->id.device_id ==
917 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
918 INFO("PCI information matches, using device \"%s\" (VF: %s)",
919 list[i]->name, (vf ? "true" : "false"));
920 attr_ctx = mlx4_glue->open_device(list[i]);
924 if (attr_ctx == NULL) {
925 mlx4_glue->free_device_list(list);
929 ERROR("cannot access device, is mlx4_ib loaded?");
933 ERROR("cannot use device, are drivers up to date?");
941 DEBUG("device opened");
942 if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
946 INFO("%u port(s) detected", device_attr.phys_port_cnt);
947 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
948 if (mlx4_args(pci_dev->device.devargs, &conf)) {
949 ERROR("failed to process device arguments");
953 /* Use all ports when none are defined */
954 if (!conf.ports.enabled)
955 conf.ports.enabled = conf.ports.present;
956 /* Retrieve extended device attributes. */
957 if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
961 assert(device_attr.max_sge >= MLX4_MAX_SGE);
962 for (i = 0; i < device_attr.phys_port_cnt; i++) {
963 uint32_t port = i + 1; /* ports are indexed from one */
964 struct ibv_context *ctx = NULL;
965 struct ibv_port_attr port_attr;
966 struct ibv_pd *pd = NULL;
967 struct mlx4_priv *priv = NULL;
968 struct rte_eth_dev *eth_dev = NULL;
969 struct ether_addr mac;
970 char name[RTE_ETH_NAME_MAX_LEN];
972 /* If port is not enabled, skip. */
973 if (!(conf.ports.enabled & (1 << i)))
975 DEBUG("using port %u", port);
976 ctx = mlx4_glue->open_device(ibv_dev);
981 snprintf(name, sizeof(name), "%s port %u",
982 mlx4_glue->get_device_name(ibv_dev), port);
983 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
984 eth_dev = rte_eth_dev_attach_secondary(name);
985 if (eth_dev == NULL) {
986 ERROR("can not attach rte ethdev");
991 priv = eth_dev->data->dev_private;
992 if (!priv->verbs_alloc_ctx.enabled) {
993 ERROR("secondary process is not supported"
994 " due to lack of external allocator"
1000 eth_dev->device = &pci_dev->device;
1001 eth_dev->dev_ops = &mlx4_dev_sec_ops;
1002 /* Receive command fd from primary process. */
1003 err = mlx4_mp_req_verbs_cmd_fd(eth_dev);
1008 /* Remap UAR for Tx queues. */
1009 err = mlx4_tx_uar_remap(eth_dev, err);
1015 * Ethdev pointer is still required as input since
1016 * the primary device is not accessible from the
1017 * secondary process.
1019 eth_dev->tx_pkt_burst = mlx4_tx_burst;
1020 eth_dev->rx_pkt_burst = mlx4_rx_burst;
1021 claim_zero(mlx4_glue->close_device(ctx));
1022 rte_eth_copy_pci_info(eth_dev, pci_dev);
1023 rte_eth_dev_probing_finish(eth_dev);
1026 /* Check port status. */
1027 err = mlx4_glue->query_port(ctx, port, &port_attr);
1030 ERROR("port query failed: %s", strerror(err));
1033 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1035 ERROR("port %d is not configured in Ethernet mode",
1039 if (port_attr.state != IBV_PORT_ACTIVE)
1040 DEBUG("port %d is not active: \"%s\" (%d)",
1041 port, mlx4_glue->port_state_str(port_attr.state),
1043 /* Make asynchronous FD non-blocking to handle interrupts. */
1044 err = mlx4_fd_set_non_blocking(ctx->async_fd);
1046 ERROR("cannot make asynchronous FD non-blocking: %s",
1050 /* Allocate protection domain. */
1051 pd = mlx4_glue->alloc_pd(ctx);
1054 ERROR("PD allocation failure");
1057 /* from rte_ethdev.c */
1058 priv = rte_zmalloc("ethdev private structure",
1060 RTE_CACHE_LINE_SIZE);
1063 ERROR("priv allocation failure");
1067 priv->device_attr = device_attr;
1070 priv->mtu = ETHER_MTU;
1072 priv->hw_csum = !!(device_attr.device_cap_flags &
1073 IBV_DEVICE_RAW_IP_CSUM);
1074 DEBUG("checksum offloading is %ssupported",
1075 (priv->hw_csum ? "" : "not "));
1076 /* Only ConnectX-3 Pro supports tunneling. */
1077 priv->hw_csum_l2tun =
1079 (device_attr.vendor_part_id ==
1080 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
1081 DEBUG("L2 tunnel checksum offloads are %ssupported",
1082 priv->hw_csum_l2tun ? "" : "not ");
1083 priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd,
1085 DEBUG("supported RSS hash fields mask: %016" PRIx64,
1087 priv->hw_rss_max_qps =
1088 device_attr_ex.rss_caps.max_rwq_indirection_table_size;
1089 DEBUG("MAX RSS queues %d", priv->hw_rss_max_qps);
1090 priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
1091 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1092 DEBUG("FCS stripping toggling is %ssupported",
1093 priv->hw_fcs_strip ? "" : "not ");
1095 ((device_attr_ex.tso_caps.max_tso > 0) &&
1096 (device_attr_ex.tso_caps.supported_qpts &
1097 (1 << IBV_QPT_RAW_PACKET)));
1099 priv->tso_max_payload_sz =
1100 device_attr_ex.tso_caps.max_tso;
1101 DEBUG("TSO is %ssupported",
1102 priv->tso ? "" : "not ");
1103 /* Configure the first MAC address by default. */
1104 err = mlx4_get_mac(priv, &mac.addr_bytes);
1106 ERROR("cannot get MAC address, is mlx4_en loaded?"
1107 " (error: %s)", strerror(err));
1110 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1112 mac.addr_bytes[0], mac.addr_bytes[1],
1113 mac.addr_bytes[2], mac.addr_bytes[3],
1114 mac.addr_bytes[4], mac.addr_bytes[5]);
1115 /* Register MAC address. */
1119 char ifname[IF_NAMESIZE];
1121 if (mlx4_get_ifname(priv, &ifname) == 0)
1122 DEBUG("port %u ifname is \"%s\"",
1123 priv->port, ifname);
1125 DEBUG("port %u ifname is unknown", priv->port);
1128 /* Get actual MTU if possible. */
1129 mlx4_mtu_get(priv, &priv->mtu);
1130 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
1131 eth_dev = rte_eth_dev_allocate(name);
1132 if (eth_dev == NULL) {
1134 ERROR("can not allocate rte ethdev");
1137 eth_dev->data->dev_private = priv;
1138 eth_dev->data->mac_addrs = priv->mac;
1139 eth_dev->device = &pci_dev->device;
1140 rte_eth_copy_pci_info(eth_dev, pci_dev);
1141 /* Initialize local interrupt handle for current port. */
1142 priv->intr_handle = (struct rte_intr_handle){
1144 .type = RTE_INTR_HANDLE_EXT,
1147 * Override ethdev interrupt handle pointer with private
1148 * handle instead of that of the parent PCI device used by
1149 * default. This prevents it from being shared between all
1150 * ports of the same PCI device since each of them is
1151 * associated its own Verbs context.
1153 * Rx interrupts in particular require this as the PMD has
1154 * no control over the registration of queue interrupts
1155 * besides setting up eth_dev->intr_handle, the rest is
1156 * handled by rte_intr_rx_ctl().
1158 eth_dev->intr_handle = &priv->intr_handle;
1159 priv->dev_data = eth_dev->data;
1160 eth_dev->dev_ops = &mlx4_dev_ops;
1161 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
1162 /* Hint libmlx4 to use PMD allocator for data plane resources */
1163 struct mlx4dv_ctx_allocators alctr = {
1164 .alloc = &mlx4_alloc_verbs_buf,
1165 .free = &mlx4_free_verbs_buf,
1168 err = mlx4_glue->dv_set_context_attr
1169 (ctx, MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS,
1170 (void *)((uintptr_t)&alctr));
1172 WARN("Verbs external allocator is not supported");
1174 priv->verbs_alloc_ctx.enabled = 1;
1176 /* Bring Ethernet device up. */
1177 DEBUG("forcing Ethernet interface up");
1178 mlx4_dev_set_link_up(eth_dev);
1179 /* Update link status once if waiting for LSC. */
1180 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
1181 mlx4_link_update(eth_dev, 0);
1183 * Once the device is added to the list of memory event
1184 * callback, its global MR cache table cannot be expanded
1185 * on the fly because of deadlock. If it overflows, lookup
1186 * should be done by searching MR list linearly, which is slow.
1188 err = mlx4_mr_btree_init(&priv->mr.cache,
1189 MLX4_MR_BTREE_CACHE_N * 2,
1190 eth_dev->device->numa_node);
1192 /* rte_errno is already set. */
1195 /* Add device to memory callback list. */
1196 rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock);
1197 LIST_INSERT_HEAD(&mlx4_shared_data->mem_event_cb_list,
1198 priv, mem_event_cb);
1199 rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock);
1200 rte_eth_dev_probing_finish(eth_dev);
1204 if (eth_dev != NULL)
1205 eth_dev->data->dev_private = NULL;
1207 claim_zero(mlx4_glue->dealloc_pd(pd));
1209 claim_zero(mlx4_glue->close_device(ctx));
1210 if (eth_dev != NULL) {
1211 /* mac_addrs must not be freed because part of dev_private */
1212 eth_dev->data->mac_addrs = NULL;
1213 rte_eth_dev_release_port(eth_dev);
1218 * XXX if something went wrong in the loop above, there is a resource
1219 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1220 * long as the dpdk does not provide a way to deallocate a ethdev and a
1221 * way to enumerate the registered ethdevs to free the previous ones.
1225 claim_zero(mlx4_glue->close_device(attr_ctx));
1227 mlx4_glue->free_device_list(list);
1233 static const struct rte_pci_id mlx4_pci_id_map[] = {
1235 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1236 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
1239 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1240 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
1243 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1244 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
1251 static struct rte_pci_driver mlx4_driver = {
1253 .name = MLX4_DRIVER_NAME
1255 .id_table = mlx4_pci_id_map,
1256 .probe = mlx4_pci_probe,
1257 .drv_flags = RTE_PCI_DRV_INTR_LSC |
1258 RTE_PCI_DRV_INTR_RMV,
1261 #ifdef RTE_IBVERBS_LINK_DLOPEN
1264 * Suffix RTE_EAL_PMD_PATH with "-glue".
1266 * This function performs a sanity check on RTE_EAL_PMD_PATH before
1267 * suffixing its last component.
1270 * Output buffer, should be large enough otherwise NULL is returned.
1275 * Pointer to @p buf or @p NULL in case suffix cannot be appended.
1278 mlx4_glue_path(char *buf, size_t size)
1280 static const char *const bad[] = { "/", ".", "..", NULL };
1281 const char *path = RTE_EAL_PMD_PATH;
1282 size_t len = strlen(path);
1286 while (len && path[len - 1] == '/')
1288 for (off = len; off && path[off - 1] != '/'; --off)
1290 for (i = 0; bad[i]; ++i)
1291 if (!strncmp(path + off, bad[i], (int)(len - off)))
1293 i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1294 if (i == -1 || (size_t)i >= size)
1298 ERROR("unable to append \"-glue\" to last component of"
1299 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1300 " please re-configure DPDK");
1305 * Initialization routine for run-time dependency on rdma-core.
1308 mlx4_glue_init(void)
1310 char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1311 const char *path[] = {
1313 * A basic security check is necessary before trusting
1314 * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1316 (geteuid() == getuid() && getegid() == getgid() ?
1317 getenv("MLX4_GLUE_PATH") : NULL),
1319 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1320 * variant, otherwise let dlopen() look up libraries on its
1323 (*RTE_EAL_PMD_PATH ?
1324 mlx4_glue_path(glue_path, sizeof(glue_path)) : ""),
1327 void *handle = NULL;
1331 while (!handle && i != RTE_DIM(path)) {
1340 end = strpbrk(path[i], ":;");
1342 end = path[i] + strlen(path[i]);
1343 len = end - path[i];
1348 ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE,
1350 (!len || *(end - 1) == '/') ? "" : "/");
1353 if (sizeof(name) != (size_t)ret + 1)
1355 DEBUG("looking for rdma-core glue as \"%s\"", name);
1356 handle = dlopen(name, RTLD_LAZY);
1367 WARN("cannot load glue library: %s", dlmsg);
1370 sym = dlsym(handle, "mlx4_glue");
1371 if (!sym || !*sym) {
1375 ERROR("cannot resolve glue symbol: %s", dlmsg);
1383 WARN("cannot initialize PMD due to missing run-time"
1384 " dependency on rdma-core libraries (libibverbs,"
1392 * Driver initialization routine.
1394 RTE_INIT(rte_mlx4_pmd_init)
1397 * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
1398 * want to get success errno value in case of calling them
1399 * when the device was removed.
1401 setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
1403 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1404 * huge pages. Calling ibv_fork_init() during init allows
1405 * applications to use fork() safely for purposes other than
1406 * using this PMD, which is not supported in forked processes.
1408 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1409 #ifdef RTE_IBVERBS_LINK_DLOPEN
1410 if (mlx4_glue_init())
1415 /* Glue structure must not contain any NULL pointers. */
1419 for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i)
1420 assert(((const void *const *)mlx4_glue)[i]);
1423 if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) {
1424 ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required",
1425 mlx4_glue->version, MLX4_GLUE_VERSION);
1428 mlx4_glue->fork_init();
1429 rte_pci_register(&mlx4_driver);
1432 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
1433 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
1434 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
1435 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");