1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2018 Intel Corporation
7 * The vhost-user protocol connection is an external interface, so it must be
8 * robust against invalid inputs.
10 * This is important because the vhost-user master is only one step removed
11 * from the guest. Malicious guests that have escaped will then launch further
12 * attacks from the vhost-user master.
14 * Even in deployments where guests are trusted, a bug in the vhost-user master
15 * can still cause invalid messages to be sent. Such messages must not
16 * compromise the stability of the DPDK application by causing crashes, memory
17 * corruption, or other problematic behavior.
19 * Do not assume received VhostUserMsg fields contain sensible values!
28 #include <sys/ioctl.h>
30 #include <sys/types.h>
32 #include <sys/syscall.h>
34 #ifdef RTE_LIBRTE_VHOST_NUMA
37 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
38 #include <linux/userfaultfd.h>
41 #include <rte_common.h>
42 #include <rte_malloc.h>
47 #include "vhost_user.h"
49 #define VIRTIO_MIN_MTU 68
50 #define VIRTIO_MAX_MTU 65535
52 static const char *vhost_message_str[VHOST_USER_MAX] = {
53 [VHOST_USER_NONE] = "VHOST_USER_NONE",
54 [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
55 [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
56 [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
57 [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
58 [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
59 [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
60 [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
61 [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
62 [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
63 [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
64 [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
65 [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
66 [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
67 [VHOST_USER_SET_VRING_ERR] = "VHOST_USER_SET_VRING_ERR",
68 [VHOST_USER_GET_PROTOCOL_FEATURES] = "VHOST_USER_GET_PROTOCOL_FEATURES",
69 [VHOST_USER_SET_PROTOCOL_FEATURES] = "VHOST_USER_SET_PROTOCOL_FEATURES",
70 [VHOST_USER_GET_QUEUE_NUM] = "VHOST_USER_GET_QUEUE_NUM",
71 [VHOST_USER_SET_VRING_ENABLE] = "VHOST_USER_SET_VRING_ENABLE",
72 [VHOST_USER_SEND_RARP] = "VHOST_USER_SEND_RARP",
73 [VHOST_USER_NET_SET_MTU] = "VHOST_USER_NET_SET_MTU",
74 [VHOST_USER_SET_SLAVE_REQ_FD] = "VHOST_USER_SET_SLAVE_REQ_FD",
75 [VHOST_USER_IOTLB_MSG] = "VHOST_USER_IOTLB_MSG",
76 [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
77 [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
78 [VHOST_USER_POSTCOPY_ADVISE] = "VHOST_USER_POSTCOPY_ADVISE",
79 [VHOST_USER_POSTCOPY_LISTEN] = "VHOST_USER_POSTCOPY_LISTEN",
80 [VHOST_USER_POSTCOPY_END] = "VHOST_USER_POSTCOPY_END",
83 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
84 static int read_vhost_message(int sockfd, struct VhostUserMsg *msg);
92 ret = fstat(fd, &stat);
93 return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
97 free_mem_region(struct virtio_net *dev)
100 struct rte_vhost_mem_region *reg;
102 if (!dev || !dev->mem)
105 for (i = 0; i < dev->mem->nregions; i++) {
106 reg = &dev->mem->regions[i];
107 if (reg->host_user_addr) {
108 munmap(reg->mmap_addr, reg->mmap_size);
115 vhost_backend_cleanup(struct virtio_net *dev)
118 free_mem_region(dev);
123 free(dev->guest_pages);
124 dev->guest_pages = NULL;
127 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
131 if (dev->slave_req_fd >= 0) {
132 close(dev->slave_req_fd);
133 dev->slave_req_fd = -1;
136 if (dev->postcopy_ufd >= 0) {
137 close(dev->postcopy_ufd);
138 dev->postcopy_ufd = -1;
141 dev->postcopy_listening = 0;
145 * This function just returns success at the moment unless
146 * the device hasn't been initialised.
149 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
150 struct VhostUserMsg *msg __rte_unused,
151 int main_fd __rte_unused)
157 vhost_user_reset_owner(struct virtio_net **pdev,
158 struct VhostUserMsg *msg __rte_unused,
159 int main_fd __rte_unused)
161 struct virtio_net *dev = *pdev;
162 vhost_destroy_device_notify(dev);
164 cleanup_device(dev, 0);
170 * The features that we support are requested.
173 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
174 int main_fd __rte_unused)
176 struct virtio_net *dev = *pdev;
177 uint64_t features = 0;
179 rte_vhost_driver_get_features(dev->ifname, &features);
181 msg->payload.u64 = features;
182 msg->size = sizeof(msg->payload.u64);
185 return VH_RESULT_REPLY;
189 * The queue number that we support are requested.
192 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
193 int main_fd __rte_unused)
195 struct virtio_net *dev = *pdev;
196 uint32_t queue_num = 0;
198 rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
200 msg->payload.u64 = (uint64_t)queue_num;
201 msg->size = sizeof(msg->payload.u64);
204 return VH_RESULT_REPLY;
208 * We receive the negotiated features supported by us and the virtio device.
211 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
212 int main_fd __rte_unused)
214 struct virtio_net *dev = *pdev;
215 uint64_t features = msg->payload.u64;
216 uint64_t vhost_features = 0;
217 struct rte_vdpa_device *vdpa_dev;
220 rte_vhost_driver_get_features(dev->ifname, &vhost_features);
221 if (features & ~vhost_features) {
222 RTE_LOG(ERR, VHOST_CONFIG,
223 "(%d) received invalid negotiated features.\n",
225 return VH_RESULT_ERR;
228 if (dev->flags & VIRTIO_DEV_RUNNING) {
229 if (dev->features == features)
233 * Error out if master tries to change features while device is
234 * in running state. The exception being VHOST_F_LOG_ALL, which
235 * is enabled when the live-migration starts.
237 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
238 RTE_LOG(ERR, VHOST_CONFIG,
239 "(%d) features changed while device is running.\n",
241 return VH_RESULT_ERR;
244 if (dev->notify_ops->features_changed)
245 dev->notify_ops->features_changed(dev->vid, features);
248 dev->features = features;
250 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
251 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
253 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
255 VHOST_LOG_DEBUG(VHOST_CONFIG,
256 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
258 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
259 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
261 if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
262 !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
264 * Remove all but first queue pair if MQ hasn't been
265 * negotiated. This is safe because the device is not
266 * running at this stage.
268 while (dev->nr_vring > 2) {
269 struct vhost_virtqueue *vq;
271 vq = dev->virtqueue[--dev->nr_vring];
275 dev->virtqueue[dev->nr_vring] = NULL;
281 did = dev->vdpa_dev_id;
282 vdpa_dev = rte_vdpa_get_device(did);
283 if (vdpa_dev && vdpa_dev->ops->set_features)
284 vdpa_dev->ops->set_features(dev->vid);
290 * The virtio device sends us the size of the descriptor ring.
293 vhost_user_set_vring_num(struct virtio_net **pdev,
294 struct VhostUserMsg *msg,
295 int main_fd __rte_unused)
297 struct virtio_net *dev = *pdev;
298 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
300 vq->size = msg->payload.state.num;
302 /* VIRTIO 1.0, 2.4 Virtqueues says:
304 * Queue Size value is always a power of 2. The maximum Queue Size
307 if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
308 RTE_LOG(ERR, VHOST_CONFIG,
309 "invalid virtqueue size %u\n", vq->size);
310 return VH_RESULT_ERR;
313 if (dev->dequeue_zero_copy) {
315 vq->last_zmbuf_idx = 0;
316 vq->zmbuf_size = vq->size;
317 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
318 sizeof(struct zcopy_mbuf), 0);
319 if (vq->zmbufs == NULL) {
320 RTE_LOG(WARNING, VHOST_CONFIG,
321 "failed to allocate mem for zero copy; "
322 "zero copy is force disabled\n");
323 dev->dequeue_zero_copy = 0;
325 TAILQ_INIT(&vq->zmbuf_list);
328 if (vq_is_packed(dev)) {
329 vq->shadow_used_packed = rte_malloc(NULL,
331 sizeof(struct vring_used_elem_packed),
332 RTE_CACHE_LINE_SIZE);
333 if (!vq->shadow_used_packed) {
334 RTE_LOG(ERR, VHOST_CONFIG,
335 "failed to allocate memory for shadow used ring.\n");
336 return VH_RESULT_ERR;
340 vq->shadow_used_split = rte_malloc(NULL,
341 vq->size * sizeof(struct vring_used_elem),
342 RTE_CACHE_LINE_SIZE);
343 if (!vq->shadow_used_split) {
344 RTE_LOG(ERR, VHOST_CONFIG,
345 "failed to allocate memory for shadow used ring.\n");
346 return VH_RESULT_ERR;
350 vq->batch_copy_elems = rte_malloc(NULL,
351 vq->size * sizeof(struct batch_copy_elem),
352 RTE_CACHE_LINE_SIZE);
353 if (!vq->batch_copy_elems) {
354 RTE_LOG(ERR, VHOST_CONFIG,
355 "failed to allocate memory for batching copy.\n");
356 return VH_RESULT_ERR;
363 * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
364 * same numa node as the memory of vring descriptor.
366 #ifdef RTE_LIBRTE_VHOST_NUMA
367 static struct virtio_net*
368 numa_realloc(struct virtio_net *dev, int index)
370 int oldnode, newnode;
371 struct virtio_net *old_dev;
372 struct vhost_virtqueue *old_vq, *vq;
373 struct zcopy_mbuf *new_zmbuf;
374 struct vring_used_elem *new_shadow_used_split;
375 struct vring_used_elem_packed *new_shadow_used_packed;
376 struct batch_copy_elem *new_batch_copy_elems;
380 vq = old_vq = dev->virtqueue[index];
382 ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
383 MPOL_F_NODE | MPOL_F_ADDR);
385 /* check if we need to reallocate vq */
386 ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
387 MPOL_F_NODE | MPOL_F_ADDR);
389 RTE_LOG(ERR, VHOST_CONFIG,
390 "Unable to get vq numa information.\n");
393 if (oldnode != newnode) {
394 RTE_LOG(INFO, VHOST_CONFIG,
395 "reallocate vq from %d to %d node\n", oldnode, newnode);
396 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
400 memcpy(vq, old_vq, sizeof(*vq));
401 TAILQ_INIT(&vq->zmbuf_list);
403 if (dev->dequeue_zero_copy) {
404 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
405 sizeof(struct zcopy_mbuf), 0, newnode);
407 rte_free(vq->zmbufs);
408 vq->zmbufs = new_zmbuf;
412 if (vq_is_packed(dev)) {
413 new_shadow_used_packed = rte_malloc_socket(NULL,
415 sizeof(struct vring_used_elem_packed),
418 if (new_shadow_used_packed) {
419 rte_free(vq->shadow_used_packed);
420 vq->shadow_used_packed = new_shadow_used_packed;
423 new_shadow_used_split = rte_malloc_socket(NULL,
425 sizeof(struct vring_used_elem),
428 if (new_shadow_used_split) {
429 rte_free(vq->shadow_used_split);
430 vq->shadow_used_split = new_shadow_used_split;
434 new_batch_copy_elems = rte_malloc_socket(NULL,
435 vq->size * sizeof(struct batch_copy_elem),
438 if (new_batch_copy_elems) {
439 rte_free(vq->batch_copy_elems);
440 vq->batch_copy_elems = new_batch_copy_elems;
446 /* check if we need to reallocate dev */
447 ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
448 MPOL_F_NODE | MPOL_F_ADDR);
450 RTE_LOG(ERR, VHOST_CONFIG,
451 "Unable to get dev numa information.\n");
454 if (oldnode != newnode) {
455 RTE_LOG(INFO, VHOST_CONFIG,
456 "reallocate dev from %d to %d node\n",
458 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
464 memcpy(dev, old_dev, sizeof(*dev));
469 dev->virtqueue[index] = vq;
470 vhost_devices[dev->vid] = dev;
473 vhost_user_iotlb_init(dev, index);
478 static struct virtio_net*
479 numa_realloc(struct virtio_net *dev, int index __rte_unused)
485 /* Converts QEMU virtual address to Vhost virtual address. */
487 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
489 struct rte_vhost_mem_region *r;
492 /* Find the region where the address lives. */
493 for (i = 0; i < dev->mem->nregions; i++) {
494 r = &dev->mem->regions[i];
496 if (qva >= r->guest_user_addr &&
497 qva < r->guest_user_addr + r->size) {
499 if (unlikely(*len > r->guest_user_addr + r->size - qva))
500 *len = r->guest_user_addr + r->size - qva;
502 return qva - r->guest_user_addr +
513 * Converts ring address to Vhost virtual address.
514 * If IOMMU is enabled, the ring address is a guest IO virtual address,
515 * else it is a QEMU virtual address.
518 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
519 uint64_t ra, uint64_t *size)
521 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
524 vva = vhost_user_iotlb_cache_find(vq, ra,
525 size, VHOST_ACCESS_RW);
527 vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
532 return qva_to_vva(dev, ra, size);
535 static struct virtio_net *
536 translate_ring_addresses(struct virtio_net *dev, int vq_index)
538 struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
539 struct vhost_vring_addr *addr = &vq->ring_addrs;
542 if (vq_is_packed(dev)) {
543 len = sizeof(struct vring_packed_desc) * vq->size;
544 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
545 ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
546 vq->log_guest_addr = 0;
547 if (vq->desc_packed == NULL ||
548 len != sizeof(struct vring_packed_desc) *
550 RTE_LOG(DEBUG, VHOST_CONFIG,
551 "(%d) failed to map desc_packed ring.\n",
556 dev = numa_realloc(dev, vq_index);
557 vq = dev->virtqueue[vq_index];
558 addr = &vq->ring_addrs;
560 len = sizeof(struct vring_packed_desc_event);
561 vq->driver_event = (struct vring_packed_desc_event *)
562 (uintptr_t)ring_addr_to_vva(dev,
563 vq, addr->avail_user_addr, &len);
564 if (vq->driver_event == NULL ||
565 len != sizeof(struct vring_packed_desc_event)) {
566 RTE_LOG(DEBUG, VHOST_CONFIG,
567 "(%d) failed to find driver area address.\n",
572 len = sizeof(struct vring_packed_desc_event);
573 vq->device_event = (struct vring_packed_desc_event *)
574 (uintptr_t)ring_addr_to_vva(dev,
575 vq, addr->used_user_addr, &len);
576 if (vq->device_event == NULL ||
577 len != sizeof(struct vring_packed_desc_event)) {
578 RTE_LOG(DEBUG, VHOST_CONFIG,
579 "(%d) failed to find device area address.\n",
587 /* The addresses are converted from QEMU virtual to Vhost virtual. */
588 if (vq->desc && vq->avail && vq->used)
591 len = sizeof(struct vring_desc) * vq->size;
592 vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
593 vq, addr->desc_user_addr, &len);
594 if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
595 RTE_LOG(DEBUG, VHOST_CONFIG,
596 "(%d) failed to map desc ring.\n",
601 dev = numa_realloc(dev, vq_index);
602 vq = dev->virtqueue[vq_index];
603 addr = &vq->ring_addrs;
605 len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
606 vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
607 vq, addr->avail_user_addr, &len);
608 if (vq->avail == 0 ||
609 len != sizeof(struct vring_avail) +
610 sizeof(uint16_t) * vq->size) {
611 RTE_LOG(DEBUG, VHOST_CONFIG,
612 "(%d) failed to map avail ring.\n",
617 len = sizeof(struct vring_used) +
618 sizeof(struct vring_used_elem) * vq->size;
619 vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
620 vq, addr->used_user_addr, &len);
621 if (vq->used == 0 || len != sizeof(struct vring_used) +
622 sizeof(struct vring_used_elem) * vq->size) {
623 RTE_LOG(DEBUG, VHOST_CONFIG,
624 "(%d) failed to map used ring.\n",
629 if (vq->last_used_idx != vq->used->idx) {
630 RTE_LOG(WARNING, VHOST_CONFIG,
631 "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
632 "some packets maybe resent for Tx and dropped for Rx\n",
633 vq->last_used_idx, vq->used->idx);
634 vq->last_used_idx = vq->used->idx;
635 vq->last_avail_idx = vq->used->idx;
638 vq->log_guest_addr = addr->log_guest_addr;
640 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
642 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
643 dev->vid, vq->avail);
644 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
646 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
647 dev->vid, vq->log_guest_addr);
653 * The virtio device sends us the desc, used and avail ring addresses.
654 * This function then converts these to our address space.
657 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
658 int main_fd __rte_unused)
660 struct virtio_net *dev = *pdev;
661 struct vhost_virtqueue *vq;
662 struct vhost_vring_addr *addr = &msg->payload.addr;
664 if (dev->mem == NULL)
665 return VH_RESULT_ERR;
667 /* addr->index refers to the queue index. The txq 1, rxq is 0. */
668 vq = dev->virtqueue[msg->payload.addr.index];
671 * Rings addresses should not be interpreted as long as the ring is not
672 * started and enabled
674 memcpy(&vq->ring_addrs, addr, sizeof(*addr));
676 vring_invalidate(dev, vq);
678 if (vq->enabled && (dev->features &
679 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
680 dev = translate_ring_addresses(dev, msg->payload.addr.index);
682 return VH_RESULT_ERR;
691 * The virtio device sends us the available ring last used index.
694 vhost_user_set_vring_base(struct virtio_net **pdev,
695 struct VhostUserMsg *msg,
696 int main_fd __rte_unused)
698 struct virtio_net *dev = *pdev;
699 dev->virtqueue[msg->payload.state.index]->last_used_idx =
700 msg->payload.state.num;
701 dev->virtqueue[msg->payload.state.index]->last_avail_idx =
702 msg->payload.state.num;
708 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
709 uint64_t host_phys_addr, uint64_t size)
711 struct guest_page *page, *last_page;
713 if (dev->nr_guest_pages == dev->max_guest_pages) {
714 dev->max_guest_pages *= 2;
715 dev->guest_pages = realloc(dev->guest_pages,
716 dev->max_guest_pages * sizeof(*page));
717 if (!dev->guest_pages) {
718 RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
723 if (dev->nr_guest_pages > 0) {
724 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
725 /* merge if the two pages are continuous */
726 if (host_phys_addr == last_page->host_phys_addr +
728 last_page->size += size;
733 page = &dev->guest_pages[dev->nr_guest_pages++];
734 page->guest_phys_addr = guest_phys_addr;
735 page->host_phys_addr = host_phys_addr;
742 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
745 uint64_t reg_size = reg->size;
746 uint64_t host_user_addr = reg->host_user_addr;
747 uint64_t guest_phys_addr = reg->guest_phys_addr;
748 uint64_t host_phys_addr;
751 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
752 size = page_size - (guest_phys_addr & (page_size - 1));
753 size = RTE_MIN(size, reg_size);
755 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
758 host_user_addr += size;
759 guest_phys_addr += size;
762 while (reg_size > 0) {
763 size = RTE_MIN(reg_size, page_size);
764 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
766 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
770 host_user_addr += size;
771 guest_phys_addr += size;
778 #ifdef RTE_LIBRTE_VHOST_DEBUG
779 /* TODO: enable it only in debug mode? */
781 dump_guest_pages(struct virtio_net *dev)
784 struct guest_page *page;
786 for (i = 0; i < dev->nr_guest_pages; i++) {
787 page = &dev->guest_pages[i];
789 RTE_LOG(INFO, VHOST_CONFIG,
790 "guest physical page region %u\n"
791 "\t guest_phys_addr: %" PRIx64 "\n"
792 "\t host_phys_addr : %" PRIx64 "\n"
793 "\t size : %" PRIx64 "\n",
795 page->guest_phys_addr,
796 page->host_phys_addr,
801 #define dump_guest_pages(dev)
805 vhost_memory_changed(struct VhostUserMemory *new,
806 struct rte_vhost_memory *old)
810 if (new->nregions != old->nregions)
813 for (i = 0; i < new->nregions; ++i) {
814 VhostUserMemoryRegion *new_r = &new->regions[i];
815 struct rte_vhost_mem_region *old_r = &old->regions[i];
817 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
819 if (new_r->memory_size != old_r->size)
821 if (new_r->userspace_addr != old_r->guest_user_addr)
829 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
832 struct virtio_net *dev = *pdev;
833 struct VhostUserMemory *memory = &msg->payload.memory;
834 struct rte_vhost_mem_region *reg;
837 uint64_t mmap_offset;
843 if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
844 RTE_LOG(ERR, VHOST_CONFIG,
845 "too many memory regions (%u)\n", memory->nregions);
846 return VH_RESULT_ERR;
849 if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
850 RTE_LOG(INFO, VHOST_CONFIG,
851 "(%d) memory regions not changed\n", dev->vid);
853 for (i = 0; i < memory->nregions; i++)
860 free_mem_region(dev);
865 /* Flush IOTLB cache as previous HVAs are now invalid */
866 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
867 for (i = 0; i < dev->nr_vring; i++)
868 vhost_user_iotlb_flush_all(dev->virtqueue[i]);
870 dev->nr_guest_pages = 0;
871 if (!dev->guest_pages) {
872 dev->max_guest_pages = 8;
873 dev->guest_pages = malloc(dev->max_guest_pages *
874 sizeof(struct guest_page));
875 if (dev->guest_pages == NULL) {
876 RTE_LOG(ERR, VHOST_CONFIG,
877 "(%d) failed to allocate memory "
878 "for dev->guest_pages\n",
880 return VH_RESULT_ERR;
884 dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
885 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
886 if (dev->mem == NULL) {
887 RTE_LOG(ERR, VHOST_CONFIG,
888 "(%d) failed to allocate memory for dev->mem\n",
890 return VH_RESULT_ERR;
892 dev->mem->nregions = memory->nregions;
894 for (i = 0; i < memory->nregions; i++) {
896 reg = &dev->mem->regions[i];
898 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
899 reg->guest_user_addr = memory->regions[i].userspace_addr;
900 reg->size = memory->regions[i].memory_size;
903 mmap_offset = memory->regions[i].mmap_offset;
905 /* Check for memory_size + mmap_offset overflow */
906 if (mmap_offset >= -reg->size) {
907 RTE_LOG(ERR, VHOST_CONFIG,
908 "mmap_offset (%#"PRIx64") and memory_size "
909 "(%#"PRIx64") overflow\n",
910 mmap_offset, reg->size);
914 mmap_size = reg->size + mmap_offset;
916 /* mmap() without flag of MAP_ANONYMOUS, should be called
917 * with length argument aligned with hugepagesz at older
918 * longterm version Linux, like 2.6.32 and 3.2.72, or
919 * mmap() will fail with EINVAL.
921 * to avoid failure, make sure in caller to keep length
924 alignment = get_blk_size(fd);
925 if (alignment == (uint64_t)-1) {
926 RTE_LOG(ERR, VHOST_CONFIG,
927 "couldn't get hugepage size through fstat\n");
930 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
932 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
933 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
934 MAP_SHARED | populate, fd, 0);
936 if (mmap_addr == MAP_FAILED) {
937 RTE_LOG(ERR, VHOST_CONFIG,
938 "mmap region %u failed.\n", i);
942 reg->mmap_addr = mmap_addr;
943 reg->mmap_size = mmap_size;
944 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
947 if (dev->dequeue_zero_copy)
948 if (add_guest_pages(dev, reg, alignment) < 0) {
949 RTE_LOG(ERR, VHOST_CONFIG,
950 "adding guest pages to region %u failed.\n",
955 RTE_LOG(INFO, VHOST_CONFIG,
956 "guest memory region %u, size: 0x%" PRIx64 "\n"
957 "\t guest physical addr: 0x%" PRIx64 "\n"
958 "\t guest virtual addr: 0x%" PRIx64 "\n"
959 "\t host virtual addr: 0x%" PRIx64 "\n"
960 "\t mmap addr : 0x%" PRIx64 "\n"
961 "\t mmap size : 0x%" PRIx64 "\n"
962 "\t mmap align: 0x%" PRIx64 "\n"
963 "\t mmap off : 0x%" PRIx64 "\n",
965 reg->guest_phys_addr,
966 reg->guest_user_addr,
968 (uint64_t)(uintptr_t)mmap_addr,
973 if (dev->postcopy_listening) {
975 * We haven't a better way right now than sharing
976 * DPDK's virtual address with Qemu, so that Qemu can
977 * retrieve the region offset when handling userfaults.
979 memory->regions[i].userspace_addr =
983 if (dev->postcopy_listening) {
984 /* Send the addresses back to qemu */
986 send_vhost_reply(main_fd, msg);
988 /* Wait for qemu to acknolwedge it's got the addresses
989 * we've got to wait before we're allowed to generate faults.
991 VhostUserMsg ack_msg;
992 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
993 RTE_LOG(ERR, VHOST_CONFIG,
994 "Failed to read qemu ack on postcopy set-mem-table\n");
997 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
998 RTE_LOG(ERR, VHOST_CONFIG,
999 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1000 ack_msg.request.master);
1004 /* Now userfault register and we can use the memory */
1005 for (i = 0; i < memory->nregions; i++) {
1006 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1007 reg = &dev->mem->regions[i];
1008 struct uffdio_register reg_struct;
1011 * Let's register all the mmap'ed area to ensure
1012 * alignment on page boundary.
1014 reg_struct.range.start =
1015 (uint64_t)(uintptr_t)reg->mmap_addr;
1016 reg_struct.range.len = reg->mmap_size;
1017 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1019 if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1021 RTE_LOG(ERR, VHOST_CONFIG,
1022 "Failed to register ufd for region %d: (ufd = %d) %s\n",
1023 i, dev->postcopy_ufd,
1027 RTE_LOG(INFO, VHOST_CONFIG,
1028 "\t userfaultfd registered for range : %llx - %llx\n",
1029 reg_struct.range.start,
1030 reg_struct.range.start +
1031 reg_struct.range.len - 1);
1038 for (i = 0; i < dev->nr_vring; i++) {
1039 struct vhost_virtqueue *vq = dev->virtqueue[i];
1041 if (vq->desc || vq->avail || vq->used) {
1043 * If the memory table got updated, the ring addresses
1044 * need to be translated again as virtual addresses have
1047 vring_invalidate(dev, vq);
1049 dev = translate_ring_addresses(dev, i);
1059 dump_guest_pages(dev);
1061 return VH_RESULT_OK;
1064 free_mem_region(dev);
1067 return VH_RESULT_ERR;
1071 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1078 if (vq_is_packed(dev))
1079 rings_ok = !!vq->desc_packed;
1081 rings_ok = vq->desc && vq->avail && vq->used;
1084 vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1085 vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1089 virtio_is_ready(struct virtio_net *dev)
1091 struct vhost_virtqueue *vq;
1094 if (dev->nr_vring == 0)
1097 for (i = 0; i < dev->nr_vring; i++) {
1098 vq = dev->virtqueue[i];
1100 if (!vq_is_ready(dev, vq))
1104 RTE_LOG(INFO, VHOST_CONFIG,
1105 "virtio is now ready for processing.\n");
1110 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1111 int main_fd __rte_unused)
1113 struct virtio_net *dev = *pdev;
1114 struct vhost_vring_file file;
1115 struct vhost_virtqueue *vq;
1117 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1118 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1119 file.fd = VIRTIO_INVALID_EVENTFD;
1121 file.fd = msg->fds[0];
1122 RTE_LOG(INFO, VHOST_CONFIG,
1123 "vring call idx:%d file:%d\n", file.index, file.fd);
1125 vq = dev->virtqueue[file.index];
1126 if (vq->callfd >= 0)
1129 vq->callfd = file.fd;
1131 return VH_RESULT_OK;
1134 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1135 struct VhostUserMsg *msg,
1136 int main_fd __rte_unused)
1138 if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1140 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1142 return VH_RESULT_OK;
1146 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1147 int main_fd __rte_unused)
1149 struct virtio_net *dev = *pdev;
1150 struct vhost_vring_file file;
1151 struct vhost_virtqueue *vq;
1153 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1154 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1155 file.fd = VIRTIO_INVALID_EVENTFD;
1157 file.fd = msg->fds[0];
1158 RTE_LOG(INFO, VHOST_CONFIG,
1159 "vring kick idx:%d file:%d\n", file.index, file.fd);
1161 /* Interpret ring addresses only when ring is started. */
1162 dev = translate_ring_addresses(dev, file.index);
1164 return VH_RESULT_ERR;
1168 vq = dev->virtqueue[file.index];
1171 * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1172 * the ring starts already enabled. Otherwise, it is enabled via
1173 * the SET_VRING_ENABLE message.
1175 if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
1178 if (vq->kickfd >= 0)
1180 vq->kickfd = file.fd;
1182 return VH_RESULT_OK;
1186 free_zmbufs(struct vhost_virtqueue *vq)
1188 struct zcopy_mbuf *zmbuf, *next;
1190 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1191 zmbuf != NULL; zmbuf = next) {
1192 next = TAILQ_NEXT(zmbuf, next);
1194 rte_pktmbuf_free(zmbuf->mbuf);
1195 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1198 rte_free(vq->zmbufs);
1202 * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1205 vhost_user_get_vring_base(struct virtio_net **pdev,
1206 struct VhostUserMsg *msg,
1207 int main_fd __rte_unused)
1209 struct virtio_net *dev = *pdev;
1210 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1212 /* We have to stop the queue (virtio) if it is running. */
1213 vhost_destroy_device_notify(dev);
1215 dev->flags &= ~VIRTIO_DEV_READY;
1216 dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1218 /* Here we are safe to get the last avail index */
1219 msg->payload.state.num = vq->last_avail_idx;
1221 RTE_LOG(INFO, VHOST_CONFIG,
1222 "vring base idx:%d file:%d\n", msg->payload.state.index,
1223 msg->payload.state.num);
1225 * Based on current qemu vhost-user implementation, this message is
1226 * sent and only sent in vhost_vring_stop.
1227 * TODO: cleanup the vring, it isn't usable since here.
1229 if (vq->kickfd >= 0)
1232 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1234 if (vq->callfd >= 0)
1237 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1239 if (dev->dequeue_zero_copy)
1241 if (vq_is_packed(dev)) {
1242 rte_free(vq->shadow_used_packed);
1243 vq->shadow_used_packed = NULL;
1245 rte_free(vq->shadow_used_split);
1246 vq->shadow_used_split = NULL;
1249 rte_free(vq->batch_copy_elems);
1250 vq->batch_copy_elems = NULL;
1252 msg->size = sizeof(msg->payload.state);
1255 return VH_RESULT_REPLY;
1259 * when virtio queues are ready to work, qemu will send us to
1260 * enable the virtio queue pair.
1263 vhost_user_set_vring_enable(struct virtio_net **pdev,
1264 struct VhostUserMsg *msg,
1265 int main_fd __rte_unused)
1267 struct virtio_net *dev = *pdev;
1268 int enable = (int)msg->payload.state.num;
1269 int index = (int)msg->payload.state.index;
1270 struct rte_vdpa_device *vdpa_dev;
1273 RTE_LOG(INFO, VHOST_CONFIG,
1274 "set queue enable: %d to qp idx: %d\n",
1277 did = dev->vdpa_dev_id;
1278 vdpa_dev = rte_vdpa_get_device(did);
1279 if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1280 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1282 if (dev->notify_ops->vring_state_changed)
1283 dev->notify_ops->vring_state_changed(dev->vid,
1286 dev->virtqueue[index]->enabled = enable;
1288 return VH_RESULT_OK;
1292 vhost_user_get_protocol_features(struct virtio_net **pdev,
1293 struct VhostUserMsg *msg,
1294 int main_fd __rte_unused)
1296 struct virtio_net *dev = *pdev;
1297 uint64_t features, protocol_features;
1299 rte_vhost_driver_get_features(dev->ifname, &features);
1300 rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1303 * REPLY_ACK protocol feature is only mandatory for now
1304 * for IOMMU feature. If IOMMU is explicitly disabled by the
1305 * application, disable also REPLY_ACK feature for older buggy
1306 * Qemu versions (from v2.7.0 to v2.9.0).
1308 if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1309 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1311 msg->payload.u64 = protocol_features;
1312 msg->size = sizeof(msg->payload.u64);
1315 return VH_RESULT_REPLY;
1319 vhost_user_set_protocol_features(struct virtio_net **pdev,
1320 struct VhostUserMsg *msg,
1321 int main_fd __rte_unused)
1323 struct virtio_net *dev = *pdev;
1324 uint64_t protocol_features = msg->payload.u64;
1325 uint64_t slave_protocol_features = 0;
1327 rte_vhost_driver_get_protocol_features(dev->ifname,
1328 &slave_protocol_features);
1329 if (protocol_features & ~slave_protocol_features) {
1330 RTE_LOG(ERR, VHOST_CONFIG,
1331 "(%d) received invalid protocol features.\n",
1333 return VH_RESULT_ERR;
1336 dev->protocol_features = protocol_features;
1338 return VH_RESULT_OK;
1342 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1343 int main_fd __rte_unused)
1345 struct virtio_net *dev = *pdev;
1346 int fd = msg->fds[0];
1351 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1352 return VH_RESULT_ERR;
1355 if (msg->size != sizeof(VhostUserLog)) {
1356 RTE_LOG(ERR, VHOST_CONFIG,
1357 "invalid log base msg size: %"PRId32" != %d\n",
1358 msg->size, (int)sizeof(VhostUserLog));
1359 return VH_RESULT_ERR;
1362 size = msg->payload.log.mmap_size;
1363 off = msg->payload.log.mmap_offset;
1365 /* Don't allow mmap_offset to point outside the mmap region */
1367 RTE_LOG(ERR, VHOST_CONFIG,
1368 "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1370 return VH_RESULT_ERR;
1373 RTE_LOG(INFO, VHOST_CONFIG,
1374 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1378 * mmap from 0 to workaround a hugepage mmap bug: mmap will
1379 * fail when offset is not page size aligned.
1381 addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1383 if (addr == MAP_FAILED) {
1384 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1385 return VH_RESULT_ERR;
1389 * Free previously mapped log memory on occasionally
1390 * multiple VHOST_USER_SET_LOG_BASE.
1392 if (dev->log_addr) {
1393 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1395 dev->log_addr = (uint64_t)(uintptr_t)addr;
1396 dev->log_base = dev->log_addr + off;
1397 dev->log_size = size;
1400 * The spec is not clear about it (yet), but QEMU doesn't expect
1401 * any payload in the reply.
1406 return VH_RESULT_REPLY;
1409 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1410 struct VhostUserMsg *msg,
1411 int main_fd __rte_unused)
1414 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1416 return VH_RESULT_OK;
1420 * An rarp packet is constructed and broadcasted to notify switches about
1421 * the new location of the migrated VM, so that packets from outside will
1422 * not be lost after migration.
1424 * However, we don't actually "send" a rarp packet here, instead, we set
1425 * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1428 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1429 int main_fd __rte_unused)
1431 struct virtio_net *dev = *pdev;
1432 uint8_t *mac = (uint8_t *)&msg->payload.u64;
1433 struct rte_vdpa_device *vdpa_dev;
1436 RTE_LOG(DEBUG, VHOST_CONFIG,
1437 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1438 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1439 memcpy(dev->mac.addr_bytes, mac, 6);
1442 * Set the flag to inject a RARP broadcast packet at
1443 * rte_vhost_dequeue_burst().
1445 * rte_smp_wmb() is for making sure the mac is copied
1446 * before the flag is set.
1449 rte_atomic16_set(&dev->broadcast_rarp, 1);
1450 did = dev->vdpa_dev_id;
1451 vdpa_dev = rte_vdpa_get_device(did);
1452 if (vdpa_dev && vdpa_dev->ops->migration_done)
1453 vdpa_dev->ops->migration_done(dev->vid);
1455 return VH_RESULT_OK;
1459 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1460 int main_fd __rte_unused)
1462 struct virtio_net *dev = *pdev;
1463 if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1464 msg->payload.u64 > VIRTIO_MAX_MTU) {
1465 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1468 return VH_RESULT_ERR;
1471 dev->mtu = msg->payload.u64;
1473 return VH_RESULT_OK;
1477 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1478 int main_fd __rte_unused)
1480 struct virtio_net *dev = *pdev;
1481 int fd = msg->fds[0];
1484 RTE_LOG(ERR, VHOST_CONFIG,
1485 "Invalid file descriptor for slave channel (%d)\n",
1487 return VH_RESULT_ERR;
1490 dev->slave_req_fd = fd;
1492 return VH_RESULT_OK;
1496 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1498 struct vhost_vring_addr *ra;
1499 uint64_t start, end;
1502 end = start + imsg->size;
1504 ra = &vq->ring_addrs;
1505 if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1507 if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1509 if (ra->used_user_addr >= start && ra->used_user_addr < end)
1516 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1517 struct vhost_iotlb_msg *imsg)
1519 uint64_t istart, iend, vstart, vend;
1521 istart = imsg->iova;
1522 iend = istart + imsg->size - 1;
1524 vstart = (uintptr_t)vq->desc;
1525 vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1526 if (vstart <= iend && istart <= vend)
1529 vstart = (uintptr_t)vq->avail;
1530 vend = vstart + sizeof(struct vring_avail);
1531 vend += sizeof(uint16_t) * vq->size - 1;
1532 if (vstart <= iend && istart <= vend)
1535 vstart = (uintptr_t)vq->used;
1536 vend = vstart + sizeof(struct vring_used);
1537 vend += sizeof(struct vring_used_elem) * vq->size - 1;
1538 if (vstart <= iend && istart <= vend)
1545 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1546 int main_fd __rte_unused)
1548 struct virtio_net *dev = *pdev;
1549 struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1553 switch (imsg->type) {
1554 case VHOST_IOTLB_UPDATE:
1556 vva = qva_to_vva(dev, imsg->uaddr, &len);
1558 return VH_RESULT_ERR;
1560 for (i = 0; i < dev->nr_vring; i++) {
1561 struct vhost_virtqueue *vq = dev->virtqueue[i];
1563 vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1566 if (is_vring_iotlb_update(vq, imsg))
1567 *pdev = dev = translate_ring_addresses(dev, i);
1570 case VHOST_IOTLB_INVALIDATE:
1571 for (i = 0; i < dev->nr_vring; i++) {
1572 struct vhost_virtqueue *vq = dev->virtqueue[i];
1574 vhost_user_iotlb_cache_remove(vq, imsg->iova,
1577 if (is_vring_iotlb_invalidate(vq, imsg))
1578 vring_invalidate(dev, vq);
1582 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1584 return VH_RESULT_ERR;
1587 return VH_RESULT_OK;
1591 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1592 struct VhostUserMsg *msg,
1593 int main_fd __rte_unused)
1595 struct virtio_net *dev = *pdev;
1596 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1597 struct uffdio_api api_struct;
1599 dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1601 if (dev->postcopy_ufd == -1) {
1602 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1604 return VH_RESULT_ERR;
1606 api_struct.api = UFFD_API;
1607 api_struct.features = 0;
1608 if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1609 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1611 close(dev->postcopy_ufd);
1612 dev->postcopy_ufd = -1;
1613 return VH_RESULT_ERR;
1615 msg->fds[0] = dev->postcopy_ufd;
1618 return VH_RESULT_REPLY;
1620 dev->postcopy_ufd = -1;
1623 return VH_RESULT_ERR;
1628 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1629 struct VhostUserMsg *msg __rte_unused,
1630 int main_fd __rte_unused)
1632 struct virtio_net *dev = *pdev;
1634 if (dev->mem && dev->mem->nregions) {
1635 RTE_LOG(ERR, VHOST_CONFIG,
1636 "Regions already registered at postcopy-listen\n");
1637 return VH_RESULT_ERR;
1639 dev->postcopy_listening = 1;
1641 return VH_RESULT_OK;
1645 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1646 int main_fd __rte_unused)
1648 struct virtio_net *dev = *pdev;
1650 dev->postcopy_listening = 0;
1651 if (dev->postcopy_ufd >= 0) {
1652 close(dev->postcopy_ufd);
1653 dev->postcopy_ufd = -1;
1656 msg->payload.u64 = 0;
1657 msg->size = sizeof(msg->payload.u64);
1660 return VH_RESULT_REPLY;
1663 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1664 struct VhostUserMsg *msg,
1666 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1667 [VHOST_USER_NONE] = NULL,
1668 [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1669 [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1670 [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1671 [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1672 [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1673 [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1674 [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1675 [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1676 [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1677 [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1678 [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1679 [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1680 [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1681 [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1682 [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1683 [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1684 [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1685 [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1686 [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1687 [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1688 [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1689 [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1690 [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1691 [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1692 [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1696 /* return bytes# of read on success or negative val on failure. */
1698 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1702 ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1703 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1707 if (msg && msg->size) {
1708 if (msg->size > sizeof(msg->payload)) {
1709 RTE_LOG(ERR, VHOST_CONFIG,
1710 "invalid msg size: %d\n", msg->size);
1713 ret = read(sockfd, &msg->payload, msg->size);
1716 if (ret != (int)msg->size) {
1717 RTE_LOG(ERR, VHOST_CONFIG,
1718 "read control message failed\n");
1727 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1732 return send_fd_message(sockfd, (char *)msg,
1733 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1737 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1742 msg->flags &= ~VHOST_USER_VERSION_MASK;
1743 msg->flags &= ~VHOST_USER_NEED_REPLY;
1744 msg->flags |= VHOST_USER_VERSION;
1745 msg->flags |= VHOST_USER_REPLY_MASK;
1747 return send_vhost_message(sockfd, msg);
1751 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1755 if (msg->flags & VHOST_USER_NEED_REPLY)
1756 rte_spinlock_lock(&dev->slave_req_lock);
1758 ret = send_vhost_message(dev->slave_req_fd, msg);
1759 if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1760 rte_spinlock_unlock(&dev->slave_req_lock);
1766 * Allocate a queue pair if it hasn't been allocated yet
1769 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1770 struct VhostUserMsg *msg)
1774 switch (msg->request.master) {
1775 case VHOST_USER_SET_VRING_KICK:
1776 case VHOST_USER_SET_VRING_CALL:
1777 case VHOST_USER_SET_VRING_ERR:
1778 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1780 case VHOST_USER_SET_VRING_NUM:
1781 case VHOST_USER_SET_VRING_BASE:
1782 case VHOST_USER_SET_VRING_ENABLE:
1783 vring_idx = msg->payload.state.index;
1785 case VHOST_USER_SET_VRING_ADDR:
1786 vring_idx = msg->payload.addr.index;
1792 if (vring_idx >= VHOST_MAX_VRING) {
1793 RTE_LOG(ERR, VHOST_CONFIG,
1794 "invalid vring index: %u\n", vring_idx);
1798 if (dev->virtqueue[vring_idx])
1801 return alloc_vring_queue(dev, vring_idx);
1805 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1808 unsigned int vq_num = 0;
1810 while (vq_num < dev->nr_vring) {
1811 struct vhost_virtqueue *vq = dev->virtqueue[i];
1814 rte_spinlock_lock(&vq->access_lock);
1822 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1825 unsigned int vq_num = 0;
1827 while (vq_num < dev->nr_vring) {
1828 struct vhost_virtqueue *vq = dev->virtqueue[i];
1831 rte_spinlock_unlock(&vq->access_lock);
1839 vhost_user_msg_handler(int vid, int fd)
1841 struct virtio_net *dev;
1842 struct VhostUserMsg msg;
1843 struct rte_vdpa_device *vdpa_dev;
1846 int unlock_required = 0;
1847 uint32_t skip_master = 0;
1850 dev = get_device(vid);
1854 if (!dev->notify_ops) {
1855 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1856 if (!dev->notify_ops) {
1857 RTE_LOG(ERR, VHOST_CONFIG,
1858 "failed to get callback ops for driver %s\n",
1864 ret = read_vhost_message(fd, &msg);
1865 if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1867 RTE_LOG(ERR, VHOST_CONFIG,
1868 "vhost read message failed\n");
1870 RTE_LOG(INFO, VHOST_CONFIG,
1871 "vhost peer closed\n");
1873 RTE_LOG(ERR, VHOST_CONFIG,
1874 "vhost read incorrect message\n");
1880 if (msg.request.master != VHOST_USER_IOTLB_MSG)
1881 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1882 vhost_message_str[msg.request.master]);
1884 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1885 vhost_message_str[msg.request.master]);
1887 ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1889 RTE_LOG(ERR, VHOST_CONFIG,
1890 "failed to alloc queue\n");
1895 * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1896 * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1897 * and device is destroyed. destroy_device waits for queues to be
1898 * inactive, so it is safe. Otherwise taking the access_lock
1899 * would cause a dead lock.
1901 switch (msg.request.master) {
1902 case VHOST_USER_SET_FEATURES:
1903 case VHOST_USER_SET_PROTOCOL_FEATURES:
1904 case VHOST_USER_SET_OWNER:
1905 case VHOST_USER_SET_MEM_TABLE:
1906 case VHOST_USER_SET_LOG_BASE:
1907 case VHOST_USER_SET_LOG_FD:
1908 case VHOST_USER_SET_VRING_NUM:
1909 case VHOST_USER_SET_VRING_ADDR:
1910 case VHOST_USER_SET_VRING_BASE:
1911 case VHOST_USER_SET_VRING_KICK:
1912 case VHOST_USER_SET_VRING_CALL:
1913 case VHOST_USER_SET_VRING_ERR:
1914 case VHOST_USER_SET_VRING_ENABLE:
1915 case VHOST_USER_SEND_RARP:
1916 case VHOST_USER_NET_SET_MTU:
1917 case VHOST_USER_SET_SLAVE_REQ_FD:
1918 vhost_user_lock_all_queue_pairs(dev);
1919 unlock_required = 1;
1926 if (dev->extern_ops.pre_msg_handle) {
1927 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
1928 (void *)&msg, &skip_master);
1929 if (ret == VH_RESULT_ERR)
1931 else if (ret == VH_RESULT_REPLY)
1932 send_vhost_reply(fd, &msg);
1935 goto skip_to_post_handle;
1938 request = msg.request.master;
1939 if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
1940 if (!vhost_message_handlers[request])
1941 goto skip_to_post_handle;
1942 ret = vhost_message_handlers[request](&dev, &msg, fd);
1946 RTE_LOG(ERR, VHOST_CONFIG,
1947 "Processing %s failed.\n",
1948 vhost_message_str[request]);
1951 RTE_LOG(DEBUG, VHOST_CONFIG,
1952 "Processing %s succeeded.\n",
1953 vhost_message_str[request]);
1955 case VH_RESULT_REPLY:
1956 RTE_LOG(DEBUG, VHOST_CONFIG,
1957 "Processing %s succeeded and needs reply.\n",
1958 vhost_message_str[request]);
1959 send_vhost_reply(fd, &msg);
1963 RTE_LOG(ERR, VHOST_CONFIG,
1964 "Requested invalid message type %d.\n", request);
1965 ret = VH_RESULT_ERR;
1968 skip_to_post_handle:
1969 if (ret != VH_RESULT_ERR && dev->extern_ops.post_msg_handle) {
1970 ret = (*dev->extern_ops.post_msg_handle)(
1971 dev->vid, (void *)&msg);
1972 if (ret == VH_RESULT_ERR)
1974 else if (ret == VH_RESULT_REPLY)
1975 send_vhost_reply(fd, &msg);
1979 if (unlock_required)
1980 vhost_user_unlock_all_queue_pairs(dev);
1983 * If the request required a reply that was already sent,
1984 * this optional reply-ack won't be sent as the
1985 * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
1987 if (msg.flags & VHOST_USER_NEED_REPLY) {
1988 msg.payload.u64 = ret == VH_RESULT_ERR;
1989 msg.size = sizeof(msg.payload.u64);
1991 send_vhost_reply(fd, &msg);
1992 } else if (ret == VH_RESULT_ERR) {
1993 RTE_LOG(ERR, VHOST_CONFIG,
1994 "vhost message handling failed.\n");
1998 if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1999 dev->flags |= VIRTIO_DEV_READY;
2001 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2002 if (dev->dequeue_zero_copy) {
2003 RTE_LOG(INFO, VHOST_CONFIG,
2004 "dequeue zero copy is enabled\n");
2007 if (dev->notify_ops->new_device(dev->vid) == 0)
2008 dev->flags |= VIRTIO_DEV_RUNNING;
2012 did = dev->vdpa_dev_id;
2013 vdpa_dev = rte_vdpa_get_device(did);
2014 if (vdpa_dev && virtio_is_ready(dev) &&
2015 !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2016 msg.request.master == VHOST_USER_SET_VRING_ENABLE) {
2017 if (vdpa_dev->ops->dev_conf)
2018 vdpa_dev->ops->dev_conf(dev->vid);
2019 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2020 if (vhost_user_host_notifier_ctrl(dev->vid, true) != 0) {
2021 RTE_LOG(INFO, VHOST_CONFIG,
2022 "(%d) software relay is used for vDPA, performance may be low.\n",
2030 static int process_slave_message_reply(struct virtio_net *dev,
2031 const struct VhostUserMsg *msg)
2033 struct VhostUserMsg msg_reply;
2036 if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2039 if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2044 if (msg_reply.request.slave != msg->request.slave) {
2045 RTE_LOG(ERR, VHOST_CONFIG,
2046 "Received unexpected msg type (%u), expected %u\n",
2047 msg_reply.request.slave, msg->request.slave);
2052 ret = msg_reply.payload.u64 ? -1 : 0;
2055 rte_spinlock_unlock(&dev->slave_req_lock);
2060 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2063 struct VhostUserMsg msg = {
2064 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2065 .flags = VHOST_USER_VERSION,
2066 .size = sizeof(msg.payload.iotlb),
2070 .type = VHOST_IOTLB_MISS,
2074 ret = send_vhost_message(dev->slave_req_fd, &msg);
2076 RTE_LOG(ERR, VHOST_CONFIG,
2077 "Failed to send IOTLB miss message (%d)\n",
2085 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2091 struct VhostUserMsg msg = {
2092 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2093 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2094 .size = sizeof(msg.payload.area),
2096 .u64 = index & VHOST_USER_VRING_IDX_MASK,
2103 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2109 ret = send_vhost_slave_message(dev, &msg);
2111 RTE_LOG(ERR, VHOST_CONFIG,
2112 "Failed to set host notifier (%d)\n", ret);
2116 return process_slave_message_reply(dev, &msg);
2119 int vhost_user_host_notifier_ctrl(int vid, bool enable)
2121 struct virtio_net *dev;
2122 struct rte_vdpa_device *vdpa_dev;
2123 int vfio_device_fd, did, ret = 0;
2124 uint64_t offset, size;
2127 dev = get_device(vid);
2131 did = dev->vdpa_dev_id;
2135 if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2136 !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2137 !(dev->protocol_features &
2138 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2139 !(dev->protocol_features &
2140 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2141 !(dev->protocol_features &
2142 (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2145 vdpa_dev = rte_vdpa_get_device(did);
2149 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2150 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2152 vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2153 if (vfio_device_fd < 0)
2157 for (i = 0; i < dev->nr_vring; i++) {
2158 if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2164 if (vhost_user_slave_set_vring_host_notifier(dev, i,
2165 vfio_device_fd, offset, size) < 0) {
2172 for (i = 0; i < dev->nr_vring; i++) {
2173 vhost_user_slave_set_vring_host_notifier(dev, i, -1,