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);
119 struct mlx5_vdpa_virtq *virtq = NULL;
122 DRV_LOG(ERR, "Invalid device id: %d.", did);
125 SLIST_FOREACH(virtq, &priv->virtq_list, next)
126 if (virtq->index == vring)
129 DRV_LOG(ERR, "Invalid or unconfigured vring id: %d.", vring);
132 return mlx5_vdpa_virtq_enable(virtq, state);
136 mlx5_vdpa_direct_db_prepare(struct mlx5_vdpa_priv *priv)
140 if (priv->direct_notifier) {
141 ret = rte_vhost_host_notifier_ctrl(priv->vid, false);
143 DRV_LOG(INFO, "Direct HW notifier FD cannot be "
144 "destroyed for device %d: %d.", priv->vid, ret);
147 priv->direct_notifier = 0;
149 ret = rte_vhost_host_notifier_ctrl(priv->vid, true);
151 DRV_LOG(INFO, "Direct HW notifier FD cannot be configured for"
152 " device %d: %d.", priv->vid, ret);
154 priv->direct_notifier = 1;
159 mlx5_vdpa_features_set(int vid)
161 int did = rte_vhost_get_vdpa_device_id(vid);
162 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
163 uint64_t log_base, log_size;
168 DRV_LOG(ERR, "Invalid device id: %d.", did);
171 ret = rte_vhost_get_negotiated_features(vid, &features);
173 DRV_LOG(ERR, "Failed to get negotiated features.");
176 if (RTE_VHOST_NEED_LOG(features)) {
177 ret = rte_vhost_get_log_base(vid, &log_base, &log_size);
179 DRV_LOG(ERR, "Failed to get log base.");
182 ret = mlx5_vdpa_dirty_bitmap_set(priv, log_base, log_size);
184 DRV_LOG(ERR, "Failed to set dirty bitmap.");
187 DRV_LOG(INFO, "mlx5 vdpa: enabling dirty logging...");
188 ret = mlx5_vdpa_logging_enable(priv, 1);
190 DRV_LOG(ERR, "Failed t enable dirty logging.");
198 mlx5_vdpa_dev_close(int vid)
200 int did = rte_vhost_get_vdpa_device_id(vid);
201 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
205 DRV_LOG(ERR, "Invalid device id: %d.", did);
208 if (priv->configured)
209 ret |= mlx5_vdpa_lm_log(priv);
210 mlx5_vdpa_cqe_event_unset(priv);
211 ret |= mlx5_vdpa_steer_unset(priv);
212 mlx5_vdpa_virtqs_release(priv);
213 mlx5_vdpa_event_qp_global_release(priv);
214 mlx5_vdpa_mem_dereg(priv);
215 priv->configured = 0;
217 DRV_LOG(INFO, "vDPA device %d was closed.", vid);
222 mlx5_vdpa_dev_config(int vid)
224 int did = rte_vhost_get_vdpa_device_id(vid);
225 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
228 DRV_LOG(ERR, "Invalid device id: %d.", did);
231 if (priv->configured && mlx5_vdpa_dev_close(vid)) {
232 DRV_LOG(ERR, "Failed to reconfigure vid %d.", vid);
236 if (mlx5_vdpa_mem_register(priv) || mlx5_vdpa_direct_db_prepare(priv) ||
237 mlx5_vdpa_virtqs_prepare(priv) || mlx5_vdpa_steer_setup(priv) ||
238 mlx5_vdpa_cqe_event_setup(priv)) {
239 mlx5_vdpa_dev_close(vid);
242 priv->configured = 1;
243 DRV_LOG(INFO, "vDPA device %d was configured.", vid);
248 mlx5_vdpa_get_device_fd(int vid)
250 int did = rte_vhost_get_vdpa_device_id(vid);
251 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
254 DRV_LOG(ERR, "Invalid device id: %d.", did);
257 return priv->ctx->cmd_fd;
261 mlx5_vdpa_get_notify_area(int vid, int qid, uint64_t *offset, uint64_t *size)
263 int did = rte_vhost_get_vdpa_device_id(vid);
264 struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
268 DRV_LOG(ERR, "Invalid device id: %d.", did);
272 DRV_LOG(ERR, "VAR was not created for device %d, is the device"
273 " configured?.", did);
276 *offset = priv->var->mmap_off;
277 *size = priv->var->length;
281 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
282 .get_queue_num = mlx5_vdpa_get_queue_num,
283 .get_features = mlx5_vdpa_get_vdpa_features,
284 .get_protocol_features = mlx5_vdpa_get_protocol_features,
285 .dev_conf = mlx5_vdpa_dev_config,
286 .dev_close = mlx5_vdpa_dev_close,
287 .set_vring_state = mlx5_vdpa_set_vring_state,
288 .set_features = mlx5_vdpa_features_set,
289 .migration_done = NULL,
290 .get_vfio_group_fd = NULL,
291 .get_vfio_device_fd = mlx5_vdpa_get_device_fd,
292 .get_notify_area = mlx5_vdpa_get_notify_area,
295 static struct ibv_device *
296 mlx5_vdpa_get_ib_device_match(struct rte_pci_addr *addr)
299 struct ibv_device **ibv_list = mlx5_glue->get_device_list(&n);
300 struct ibv_device *ibv_match = NULL;
307 struct rte_pci_addr pci_addr;
309 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[n]->name);
310 if (mlx5_dev_to_pci_addr(ibv_list[n]->ibdev_path, &pci_addr))
312 if (memcmp(addr, &pci_addr, sizeof(pci_addr)))
314 ibv_match = ibv_list[n];
319 mlx5_glue->free_device_list(ibv_list);
323 /* Try to disable ROCE by Netlink\Devlink. */
325 mlx5_vdpa_nl_roce_disable(const char *addr)
327 int nlsk_fd = mlx5_nl_init(NETLINK_GENERIC);
334 devlink_id = mlx5_nl_devlink_family_id_get(nlsk_fd);
335 if (devlink_id < 0) {
337 DRV_LOG(DEBUG, "Failed to get devlink id for ROCE operations by"
341 ret = mlx5_nl_enable_roce_get(nlsk_fd, devlink_id, addr, &enable);
343 DRV_LOG(DEBUG, "Failed to get ROCE enable by Netlink: %d.",
346 } else if (!enable) {
347 DRV_LOG(INFO, "ROCE has already disabled(Netlink).");
350 ret = mlx5_nl_enable_roce_set(nlsk_fd, devlink_id, addr, 0);
352 DRV_LOG(DEBUG, "Failed to disable ROCE by Netlink: %d.", ret);
354 DRV_LOG(INFO, "ROCE is disabled by Netlink successfully.");
360 /* Try to disable ROCE by sysfs. */
362 mlx5_vdpa_sys_roce_disable(const char *addr)
368 MKSTR(file_p, "/sys/bus/pci/devices/%s/roce_enable", addr);
369 file_o = fopen(file_p, "rb");
374 ret = fscanf(file_o, "%d", &enable);
379 } else if (!enable) {
381 DRV_LOG(INFO, "ROCE has already disabled(sysfs).");
385 file_o = fopen(file_p, "wb");
390 fprintf(file_o, "0\n");
394 DRV_LOG(DEBUG, "Failed to disable ROCE by sysfs: %d.", ret);
396 DRV_LOG(INFO, "ROCE is disabled by sysfs successfully.");
401 #define MLX5_VDPA_MAX_RETRIES 20
402 #define MLX5_VDPA_USEC 1000
404 mlx5_vdpa_roce_disable(struct rte_pci_addr *addr, struct ibv_device **ibv)
406 char addr_name[64] = {0};
408 rte_pci_device_name(addr, addr_name, sizeof(addr_name));
409 /* Firstly try to disable ROCE by Netlink and fallback to sysfs. */
410 if (mlx5_vdpa_nl_roce_disable(addr_name) == 0 ||
411 mlx5_vdpa_sys_roce_disable(addr_name) == 0) {
413 * Succeed to disable ROCE, wait for the IB device to appear
414 * again after reload.
417 struct ibv_device *ibv_new;
419 for (r = MLX5_VDPA_MAX_RETRIES; r; r--) {
420 ibv_new = mlx5_vdpa_get_ib_device_match(addr);
425 usleep(MLX5_VDPA_USEC);
427 DRV_LOG(ERR, "Cannot much device %s after ROCE disable, "
428 "retries exceed %d", addr_name, MLX5_VDPA_MAX_RETRIES);
435 * DPDK callback to register a PCI device.
437 * This function spawns vdpa device out of a given PCI device.
440 * PCI driver structure (mlx5_vpda_driver).
442 * PCI device information.
445 * 0 on success, 1 to skip this driver, a negative errno value otherwise
446 * and rte_errno is set.
449 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
450 struct rte_pci_device *pci_dev __rte_unused)
452 struct ibv_device *ibv;
453 struct mlx5_vdpa_priv *priv = NULL;
454 struct ibv_context *ctx = NULL;
455 struct mlx5_hca_attr attr;
458 if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
459 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
463 ibv = mlx5_vdpa_get_ib_device_match(&pci_dev->addr);
465 DRV_LOG(ERR, "No matching IB device for PCI slot "
466 PCI_PRI_FMT ".", pci_dev->addr.domain,
467 pci_dev->addr.bus, pci_dev->addr.devid,
468 pci_dev->addr.function);
471 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
474 if (mlx5_vdpa_roce_disable(&pci_dev->addr, &ibv) != 0) {
475 DRV_LOG(WARNING, "Failed to disable ROCE for \"%s\".",
479 ctx = mlx5_glue->dv_open_device(ibv);
481 DRV_LOG(ERR, "Failed to open IB device \"%s\".", ibv->name);
485 priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv),
486 RTE_CACHE_LINE_SIZE);
488 DRV_LOG(ERR, "Failed to allocate private memory.");
492 ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
494 DRV_LOG(ERR, "Unable to read HCA capabilities.");
498 if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
499 DRV_LOG(ERR, "Not enough capabilities to support vdpa,"
500 " maybe old FW/OFED version?");
504 priv->caps = attr.vdpa;
505 priv->log_max_rqt_size = attr.log_max_rqt_size;
508 priv->dev_addr.pci_addr = pci_dev->addr;
509 priv->dev_addr.type = VDPA_ADDR_PCI;
510 priv->var = mlx5_glue->dv_alloc_var(ctx, 0);
512 DRV_LOG(ERR, "Failed to allocate VAR %u.\n", errno);
515 priv->id = rte_vdpa_register_device(&priv->dev_addr, &mlx5_vdpa_ops);
517 DRV_LOG(ERR, "Failed to register vDPA device.");
518 rte_errno = rte_errno ? rte_errno : EINVAL;
521 SLIST_INIT(&priv->mr_list);
522 SLIST_INIT(&priv->virtq_list);
523 pthread_mutex_lock(&priv_list_lock);
524 TAILQ_INSERT_TAIL(&priv_list, priv, next);
525 pthread_mutex_unlock(&priv_list_lock);
531 mlx5_glue->dv_free_var(priv->var);
535 mlx5_glue->close_device(ctx);
540 * DPDK callback to remove a PCI device.
542 * This function removes all vDPA devices belong to a given PCI device.
545 * Pointer to the PCI device.
548 * 0 on success, the function cannot fail.
551 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
553 struct mlx5_vdpa_priv *priv = NULL;
556 pthread_mutex_lock(&priv_list_lock);
557 TAILQ_FOREACH(priv, &priv_list, next) {
558 if (memcmp(&priv->dev_addr.pci_addr, &pci_dev->addr,
559 sizeof(pci_dev->addr)) == 0) {
565 TAILQ_REMOVE(&priv_list, priv, next);
566 pthread_mutex_unlock(&priv_list_lock);
568 if (priv->configured)
569 mlx5_vdpa_dev_close(priv->vid);
571 mlx5_glue->dv_free_var(priv->var);
574 mlx5_glue->close_device(priv->ctx);
580 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
582 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
583 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
586 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
587 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
590 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
591 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
594 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
595 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
598 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
599 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
606 static struct rte_pci_driver mlx5_vdpa_driver = {
610 .id_table = mlx5_vdpa_pci_id_map,
611 .probe = mlx5_vdpa_pci_probe,
612 .remove = mlx5_vdpa_pci_remove,
617 * Driver initialization routine.
619 RTE_INIT(rte_mlx5_vdpa_init)
621 /* Initialize common log type. */
622 mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
623 if (mlx5_vdpa_logtype >= 0)
624 rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
626 rte_pci_register(&mlx5_vdpa_driver);
629 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
630 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
631 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");