1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2019 Mellanox Technologies, Ltd
6 #include <rte_malloc.h>
9 #include <rte_bus_pci.h>
12 #include <mlx5_glue.h>
13 #include <mlx5_common.h>
14 #include <mlx5_devx_cmds.h>
18 #include "mlx5_vdpa_utils.h"
19 #include "mlx5_vdpa.h"
22 #define MLX5_VDPA_DEFAULT_FEATURES ((1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
23 (1ULL << VIRTIO_F_ANY_LAYOUT) | \
24 (1ULL << VIRTIO_NET_F_MQ) | \
25 (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
26 (1ULL << VIRTIO_F_ORDER_PLATFORM) | \
27 (1ULL << VHOST_F_LOG_ALL))
29 #define MLX5_VDPA_PROTOCOL_FEATURES \
30 ((1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ) | \
31 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD) | \
32 (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER) | \
33 (1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD) | \
34 (1ULL << VHOST_USER_PROTOCOL_F_MQ))
36 TAILQ_HEAD(mlx5_vdpa_privs, mlx5_vdpa_priv) priv_list =
37 TAILQ_HEAD_INITIALIZER(priv_list);
38 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
39 int mlx5_vdpa_logtype;
41 static struct mlx5_vdpa_priv *
42 mlx5_vdpa_find_priv_resource_by_did(int did)
44 struct mlx5_vdpa_priv *priv;
47 pthread_mutex_lock(&priv_list_lock);
48 TAILQ_FOREACH(priv, &priv_list, next) {
49 if (did == priv->id) {
54 pthread_mutex_unlock(&priv_list_lock);
56 DRV_LOG(ERR, "Invalid device id: %d.", did);
64 mlx5_vdpa_get_queue_num(int did, uint32_t *queue_num)
66 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
69 DRV_LOG(ERR, "Invalid device id: %d.", did);
72 *queue_num = priv->caps.max_num_virtio_queues;
77 mlx5_vdpa_get_vdpa_features(int did, uint64_t *features)
79 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
82 DRV_LOG(ERR, "Invalid device id: %d.", did);
85 *features = MLX5_VDPA_DEFAULT_FEATURES;
86 if (priv->caps.virtio_queue_type & (1 << MLX5_VIRTQ_TYPE_PACKED))
87 *features |= (1ULL << VIRTIO_F_RING_PACKED);
88 if (priv->caps.tso_ipv4)
89 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO4);
90 if (priv->caps.tso_ipv6)
91 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO6);
92 if (priv->caps.tx_csum)
93 *features |= (1ULL << VIRTIO_NET_F_CSUM);
94 if (priv->caps.rx_csum)
95 *features |= (1ULL << VIRTIO_NET_F_GUEST_CSUM);
96 if (priv->caps.virtio_version_1_0)
97 *features |= (1ULL << VIRTIO_F_VERSION_1);
102 mlx5_vdpa_get_protocol_features(int did, uint64_t *features)
104 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
107 DRV_LOG(ERR, "Invalid device id: %d.", did);
110 *features = MLX5_VDPA_PROTOCOL_FEATURES;
115 mlx5_vdpa_set_vring_state(int vid, int vring, int state)
117 int did = rte_vhost_get_vdpa_device_id(vid);
118 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
121 DRV_LOG(ERR, "Invalid device id: %d.", did);
124 if (vring >= (int)priv->caps.max_num_virtio_queues * 2) {
125 DRV_LOG(ERR, "Too big vring id: %d.", vring);
128 return mlx5_vdpa_virtq_enable(priv, vring, state);
132 mlx5_vdpa_direct_db_prepare(struct mlx5_vdpa_priv *priv)
136 if (priv->direct_notifier) {
137 ret = rte_vhost_host_notifier_ctrl(priv->vid, false);
139 DRV_LOG(INFO, "Direct HW notifier FD cannot be "
140 "destroyed for device %d: %d.", priv->vid, ret);
143 priv->direct_notifier = 0;
145 ret = rte_vhost_host_notifier_ctrl(priv->vid, true);
147 DRV_LOG(INFO, "Direct HW notifier FD cannot be configured for"
148 " device %d: %d.", priv->vid, ret);
150 priv->direct_notifier = 1;
155 mlx5_vdpa_features_set(int vid)
157 int did = rte_vhost_get_vdpa_device_id(vid);
158 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
159 uint64_t log_base, log_size;
164 DRV_LOG(ERR, "Invalid device id: %d.", did);
167 ret = rte_vhost_get_negotiated_features(vid, &features);
169 DRV_LOG(ERR, "Failed to get negotiated features.");
172 if (RTE_VHOST_NEED_LOG(features)) {
173 ret = rte_vhost_get_log_base(vid, &log_base, &log_size);
175 DRV_LOG(ERR, "Failed to get log base.");
178 ret = mlx5_vdpa_dirty_bitmap_set(priv, log_base, log_size);
180 DRV_LOG(ERR, "Failed to set dirty bitmap.");
183 DRV_LOG(INFO, "mlx5 vdpa: enabling dirty logging...");
184 ret = mlx5_vdpa_logging_enable(priv, 1);
186 DRV_LOG(ERR, "Failed t enable dirty logging.");
194 mlx5_vdpa_dev_close(int vid)
196 int did = rte_vhost_get_vdpa_device_id(vid);
197 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
201 DRV_LOG(ERR, "Invalid device id: %d.", did);
204 if (priv->configured)
205 ret |= mlx5_vdpa_lm_log(priv);
206 mlx5_vdpa_cqe_event_unset(priv);
207 mlx5_vdpa_steer_unset(priv);
208 mlx5_vdpa_virtqs_release(priv);
209 mlx5_vdpa_event_qp_global_release(priv);
210 mlx5_vdpa_mem_dereg(priv);
211 priv->configured = 0;
213 DRV_LOG(INFO, "vDPA device %d was closed.", vid);
218 mlx5_vdpa_dev_config(int vid)
220 int did = rte_vhost_get_vdpa_device_id(vid);
221 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
224 DRV_LOG(ERR, "Invalid device id: %d.", did);
227 if (priv->configured && mlx5_vdpa_dev_close(vid)) {
228 DRV_LOG(ERR, "Failed to reconfigure vid %d.", vid);
232 if (mlx5_vdpa_mem_register(priv) || mlx5_vdpa_direct_db_prepare(priv) ||
233 mlx5_vdpa_virtqs_prepare(priv) || mlx5_vdpa_steer_setup(priv) ||
234 mlx5_vdpa_cqe_event_setup(priv)) {
235 mlx5_vdpa_dev_close(vid);
238 priv->configured = 1;
239 DRV_LOG(INFO, "vDPA device %d was configured.", vid);
244 mlx5_vdpa_get_device_fd(int vid)
246 int did = rte_vhost_get_vdpa_device_id(vid);
247 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
250 DRV_LOG(ERR, "Invalid device id: %d.", did);
253 return priv->ctx->cmd_fd;
257 mlx5_vdpa_get_notify_area(int vid, int qid, uint64_t *offset, uint64_t *size)
259 int did = rte_vhost_get_vdpa_device_id(vid);
260 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
264 DRV_LOG(ERR, "Invalid device id: %d.", did);
268 DRV_LOG(ERR, "VAR was not created for device %d, is the device"
269 " configured?.", did);
272 *offset = priv->var->mmap_off;
273 *size = priv->var->length;
277 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
278 .get_queue_num = mlx5_vdpa_get_queue_num,
279 .get_features = mlx5_vdpa_get_vdpa_features,
280 .get_protocol_features = mlx5_vdpa_get_protocol_features,
281 .dev_conf = mlx5_vdpa_dev_config,
282 .dev_close = mlx5_vdpa_dev_close,
283 .set_vring_state = mlx5_vdpa_set_vring_state,
284 .set_features = mlx5_vdpa_features_set,
285 .migration_done = NULL,
286 .get_vfio_group_fd = NULL,
287 .get_vfio_device_fd = mlx5_vdpa_get_device_fd,
288 .get_notify_area = mlx5_vdpa_get_notify_area,
291 static struct ibv_device *
292 mlx5_vdpa_get_ib_device_match(struct rte_pci_addr *addr)
295 struct ibv_device **ibv_list = mlx5_glue->get_device_list(&n);
296 struct ibv_device *ibv_match = NULL;
303 struct rte_pci_addr pci_addr;
305 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[n]->name);
306 if (mlx5_dev_to_pci_addr(ibv_list[n]->ibdev_path, &pci_addr))
308 if (rte_pci_addr_cmp(addr, &pci_addr))
310 ibv_match = ibv_list[n];
315 mlx5_glue->free_device_list(ibv_list);
319 /* Try to disable ROCE by Netlink\Devlink. */
321 mlx5_vdpa_nl_roce_disable(const char *addr)
323 int nlsk_fd = mlx5_nl_init(NETLINK_GENERIC);
330 devlink_id = mlx5_nl_devlink_family_id_get(nlsk_fd);
331 if (devlink_id < 0) {
333 DRV_LOG(DEBUG, "Failed to get devlink id for ROCE operations by"
337 ret = mlx5_nl_enable_roce_get(nlsk_fd, devlink_id, addr, &enable);
339 DRV_LOG(DEBUG, "Failed to get ROCE enable by Netlink: %d.",
342 } else if (!enable) {
343 DRV_LOG(INFO, "ROCE has already disabled(Netlink).");
346 ret = mlx5_nl_enable_roce_set(nlsk_fd, devlink_id, addr, 0);
348 DRV_LOG(DEBUG, "Failed to disable ROCE by Netlink: %d.", ret);
350 DRV_LOG(INFO, "ROCE is disabled by Netlink successfully.");
356 /* Try to disable ROCE by sysfs. */
358 mlx5_vdpa_sys_roce_disable(const char *addr)
364 MKSTR(file_p, "/sys/bus/pci/devices/%s/roce_enable", addr);
365 file_o = fopen(file_p, "rb");
370 ret = fscanf(file_o, "%d", &enable);
375 } else if (!enable) {
377 DRV_LOG(INFO, "ROCE has already disabled(sysfs).");
381 file_o = fopen(file_p, "wb");
386 fprintf(file_o, "0\n");
390 DRV_LOG(DEBUG, "Failed to disable ROCE by sysfs: %d.", ret);
392 DRV_LOG(INFO, "ROCE is disabled by sysfs successfully.");
397 #define MLX5_VDPA_MAX_RETRIES 20
398 #define MLX5_VDPA_USEC 1000
400 mlx5_vdpa_roce_disable(struct rte_pci_addr *addr, struct ibv_device **ibv)
402 char addr_name[64] = {0};
404 rte_pci_device_name(addr, addr_name, sizeof(addr_name));
405 /* Firstly try to disable ROCE by Netlink and fallback to sysfs. */
406 if (mlx5_vdpa_nl_roce_disable(addr_name) == 0 ||
407 mlx5_vdpa_sys_roce_disable(addr_name) == 0) {
409 * Succeed to disable ROCE, wait for the IB device to appear
410 * again after reload.
413 struct ibv_device *ibv_new;
415 for (r = MLX5_VDPA_MAX_RETRIES; r; r--) {
416 ibv_new = mlx5_vdpa_get_ib_device_match(addr);
421 usleep(MLX5_VDPA_USEC);
423 DRV_LOG(ERR, "Cannot much device %s after ROCE disable, "
424 "retries exceed %d", addr_name, MLX5_VDPA_MAX_RETRIES);
431 * DPDK callback to register a PCI device.
433 * This function spawns vdpa device out of a given PCI device.
436 * PCI driver structure (mlx5_vpda_driver).
438 * PCI device information.
441 * 0 on success, 1 to skip this driver, a negative errno value otherwise
442 * and rte_errno is set.
445 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
446 struct rte_pci_device *pci_dev __rte_unused)
448 struct ibv_device *ibv;
449 struct mlx5_vdpa_priv *priv = NULL;
450 struct ibv_context *ctx = NULL;
451 struct mlx5_hca_attr attr;
454 if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
455 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
459 ibv = mlx5_vdpa_get_ib_device_match(&pci_dev->addr);
461 DRV_LOG(ERR, "No matching IB device for PCI slot "
462 PCI_PRI_FMT ".", pci_dev->addr.domain,
463 pci_dev->addr.bus, pci_dev->addr.devid,
464 pci_dev->addr.function);
467 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
470 if (mlx5_vdpa_roce_disable(&pci_dev->addr, &ibv) != 0) {
471 DRV_LOG(WARNING, "Failed to disable ROCE for \"%s\".",
475 ctx = mlx5_glue->dv_open_device(ibv);
477 DRV_LOG(ERR, "Failed to open IB device \"%s\".", ibv->name);
481 ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
483 DRV_LOG(ERR, "Unable to read HCA capabilities.");
486 } else if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
487 DRV_LOG(ERR, "Not enough capabilities to support vdpa, maybe "
488 "old FW/OFED version?");
492 priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv) +
493 sizeof(struct mlx5_vdpa_virtq) *
494 attr.vdpa.max_num_virtio_queues * 2,
495 RTE_CACHE_LINE_SIZE);
497 DRV_LOG(ERR, "Failed to allocate private memory.");
501 priv->caps = attr.vdpa;
502 priv->log_max_rqt_size = attr.log_max_rqt_size;
504 priv->dev_addr.pci_addr = pci_dev->addr;
505 priv->dev_addr.type = VDPA_ADDR_PCI;
506 priv->var = mlx5_glue->dv_alloc_var(ctx, 0);
508 DRV_LOG(ERR, "Failed to allocate VAR %u.\n", errno);
511 priv->id = rte_vdpa_register_device(&priv->dev_addr, &mlx5_vdpa_ops);
513 DRV_LOG(ERR, "Failed to register vDPA device.");
514 rte_errno = rte_errno ? rte_errno : EINVAL;
517 SLIST_INIT(&priv->mr_list);
518 pthread_mutex_lock(&priv_list_lock);
519 TAILQ_INSERT_TAIL(&priv_list, priv, next);
520 pthread_mutex_unlock(&priv_list_lock);
526 mlx5_glue->dv_free_var(priv->var);
530 mlx5_glue->close_device(ctx);
535 * DPDK callback to remove a PCI device.
537 * This function removes all vDPA devices belong to a given PCI device.
540 * Pointer to the PCI device.
543 * 0 on success, the function cannot fail.
546 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
548 struct mlx5_vdpa_priv *priv = NULL;
551 pthread_mutex_lock(&priv_list_lock);
552 TAILQ_FOREACH(priv, &priv_list, next) {
553 if (memcmp(&priv->dev_addr.pci_addr, &pci_dev->addr,
554 sizeof(pci_dev->addr)) == 0) {
560 TAILQ_REMOVE(&priv_list, priv, next);
561 pthread_mutex_unlock(&priv_list_lock);
563 if (priv->configured)
564 mlx5_vdpa_dev_close(priv->vid);
566 mlx5_glue->dv_free_var(priv->var);
569 mlx5_glue->close_device(priv->ctx);
575 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
577 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
578 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
581 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
582 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
585 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
586 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
589 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
590 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
593 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
594 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
601 static struct rte_pci_driver mlx5_vdpa_driver = {
605 .id_table = mlx5_vdpa_pci_id_map,
606 .probe = mlx5_vdpa_pci_probe,
607 .remove = mlx5_vdpa_pci_remove,
612 * Driver initialization routine.
614 RTE_INIT(rte_mlx5_vdpa_init)
616 /* Initialize common log type. */
617 mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
618 if (mlx5_vdpa_logtype >= 0)
619 rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
621 rte_pci_register(&mlx5_vdpa_driver);
624 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
625 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
626 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");