1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2016 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/types.h>
31 #ifdef RTE_LIBRTE_VHOST_NUMA
35 #include <rte_common.h>
36 #include <rte_malloc.h>
41 #include "vhost_user.h"
43 #define VIRTIO_MIN_MTU 68
44 #define VIRTIO_MAX_MTU 65535
46 static const char *vhost_message_str[VHOST_USER_MAX] = {
47 [VHOST_USER_NONE] = "VHOST_USER_NONE",
48 [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
49 [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
50 [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
51 [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
52 [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
53 [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
54 [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
55 [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
56 [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
57 [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
58 [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
59 [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
60 [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
61 [VHOST_USER_SET_VRING_ERR] = "VHOST_USER_SET_VRING_ERR",
62 [VHOST_USER_GET_PROTOCOL_FEATURES] = "VHOST_USER_GET_PROTOCOL_FEATURES",
63 [VHOST_USER_SET_PROTOCOL_FEATURES] = "VHOST_USER_SET_PROTOCOL_FEATURES",
64 [VHOST_USER_GET_QUEUE_NUM] = "VHOST_USER_GET_QUEUE_NUM",
65 [VHOST_USER_SET_VRING_ENABLE] = "VHOST_USER_SET_VRING_ENABLE",
66 [VHOST_USER_SEND_RARP] = "VHOST_USER_SEND_RARP",
67 [VHOST_USER_NET_SET_MTU] = "VHOST_USER_NET_SET_MTU",
68 [VHOST_USER_SET_SLAVE_REQ_FD] = "VHOST_USER_SET_SLAVE_REQ_FD",
69 [VHOST_USER_IOTLB_MSG] = "VHOST_USER_IOTLB_MSG",
78 ret = fstat(fd, &stat);
79 return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
83 free_mem_region(struct virtio_net *dev)
86 struct rte_vhost_mem_region *reg;
88 if (!dev || !dev->mem)
91 for (i = 0; i < dev->mem->nregions; i++) {
92 reg = &dev->mem->regions[i];
93 if (reg->host_user_addr) {
94 munmap(reg->mmap_addr, reg->mmap_size);
101 vhost_backend_cleanup(struct virtio_net *dev)
104 free_mem_region(dev);
109 free(dev->guest_pages);
110 dev->guest_pages = NULL;
113 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
117 if (dev->slave_req_fd >= 0) {
118 close(dev->slave_req_fd);
119 dev->slave_req_fd = -1;
124 * This function just returns success at the moment unless
125 * the device hasn't been initialised.
128 vhost_user_set_owner(void)
134 vhost_user_reset_owner(struct virtio_net *dev)
136 if (dev->flags & VIRTIO_DEV_RUNNING) {
137 dev->flags &= ~VIRTIO_DEV_RUNNING;
138 dev->notify_ops->destroy_device(dev->vid);
141 cleanup_device(dev, 0);
147 * The features that we support are requested.
150 vhost_user_get_features(struct virtio_net *dev)
152 uint64_t features = 0;
154 rte_vhost_driver_get_features(dev->ifname, &features);
159 * We receive the negotiated features supported by us and the virtio device.
162 vhost_user_set_features(struct virtio_net *dev, uint64_t features)
164 uint64_t vhost_features = 0;
166 rte_vhost_driver_get_features(dev->ifname, &vhost_features);
167 if (features & ~vhost_features) {
168 RTE_LOG(ERR, VHOST_CONFIG,
169 "(%d) received invalid negotiated features.\n",
174 if (dev->flags & VIRTIO_DEV_RUNNING) {
175 if (dev->features == features)
179 * Error out if master tries to change features while device is
180 * in running state. The exception being VHOST_F_LOG_ALL, which
181 * is enabled when the live-migration starts.
183 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
184 RTE_LOG(ERR, VHOST_CONFIG,
185 "(%d) features changed while device is running.\n",
190 if (dev->notify_ops->features_changed)
191 dev->notify_ops->features_changed(dev->vid, features);
194 dev->features = features;
196 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
197 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
199 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
201 VHOST_LOG_DEBUG(VHOST_CONFIG,
202 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
204 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
205 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
207 if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
208 !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
210 * Remove all but first queue pair if MQ hasn't been
211 * negotiated. This is safe because the device is not
212 * running at this stage.
214 while (dev->nr_vring > 2) {
215 struct vhost_virtqueue *vq;
217 vq = dev->virtqueue[--dev->nr_vring];
221 dev->virtqueue[dev->nr_vring] = NULL;
231 * The virtio device sends us the size of the descriptor ring.
234 vhost_user_set_vring_num(struct virtio_net *dev,
237 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
239 vq->size = msg->payload.state.num;
241 /* VIRTIO 1.0, 2.4 Virtqueues says:
243 * Queue Size value is always a power of 2. The maximum Queue Size
246 if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
247 RTE_LOG(ERR, VHOST_CONFIG,
248 "invalid virtqueue size %u\n", vq->size);
252 if (dev->dequeue_zero_copy) {
254 vq->last_zmbuf_idx = 0;
255 vq->zmbuf_size = vq->size;
256 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
257 sizeof(struct zcopy_mbuf), 0);
258 if (vq->zmbufs == NULL) {
259 RTE_LOG(WARNING, VHOST_CONFIG,
260 "failed to allocate mem for zero copy; "
261 "zero copy is force disabled\n");
262 dev->dequeue_zero_copy = 0;
264 TAILQ_INIT(&vq->zmbuf_list);
267 vq->shadow_used_ring = rte_malloc(NULL,
268 vq->size * sizeof(struct vring_used_elem),
269 RTE_CACHE_LINE_SIZE);
270 if (!vq->shadow_used_ring) {
271 RTE_LOG(ERR, VHOST_CONFIG,
272 "failed to allocate memory for shadow used ring.\n");
276 vq->batch_copy_elems = rte_malloc(NULL,
277 vq->size * sizeof(struct batch_copy_elem),
278 RTE_CACHE_LINE_SIZE);
279 if (!vq->batch_copy_elems) {
280 RTE_LOG(ERR, VHOST_CONFIG,
281 "failed to allocate memory for batching copy.\n");
289 * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
290 * same numa node as the memory of vring descriptor.
292 #ifdef RTE_LIBRTE_VHOST_NUMA
293 static struct virtio_net*
294 numa_realloc(struct virtio_net *dev, int index)
296 int oldnode, newnode;
297 struct virtio_net *old_dev;
298 struct vhost_virtqueue *old_vq, *vq;
299 struct zcopy_mbuf *new_zmbuf;
300 struct vring_used_elem *new_shadow_used_ring;
301 struct batch_copy_elem *new_batch_copy_elems;
305 vq = old_vq = dev->virtqueue[index];
307 ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
308 MPOL_F_NODE | MPOL_F_ADDR);
310 /* check if we need to reallocate vq */
311 ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
312 MPOL_F_NODE | MPOL_F_ADDR);
314 RTE_LOG(ERR, VHOST_CONFIG,
315 "Unable to get vq numa information.\n");
318 if (oldnode != newnode) {
319 RTE_LOG(INFO, VHOST_CONFIG,
320 "reallocate vq from %d to %d node\n", oldnode, newnode);
321 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
325 memcpy(vq, old_vq, sizeof(*vq));
326 TAILQ_INIT(&vq->zmbuf_list);
328 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
329 sizeof(struct zcopy_mbuf), 0, newnode);
331 rte_free(vq->zmbufs);
332 vq->zmbufs = new_zmbuf;
335 new_shadow_used_ring = rte_malloc_socket(NULL,
336 vq->size * sizeof(struct vring_used_elem),
339 if (new_shadow_used_ring) {
340 rte_free(vq->shadow_used_ring);
341 vq->shadow_used_ring = new_shadow_used_ring;
344 new_batch_copy_elems = rte_malloc_socket(NULL,
345 vq->size * sizeof(struct batch_copy_elem),
348 if (new_batch_copy_elems) {
349 rte_free(vq->batch_copy_elems);
350 vq->batch_copy_elems = new_batch_copy_elems;
356 /* check if we need to reallocate dev */
357 ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
358 MPOL_F_NODE | MPOL_F_ADDR);
360 RTE_LOG(ERR, VHOST_CONFIG,
361 "Unable to get dev numa information.\n");
364 if (oldnode != newnode) {
365 RTE_LOG(INFO, VHOST_CONFIG,
366 "reallocate dev from %d to %d node\n",
368 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
374 memcpy(dev, old_dev, sizeof(*dev));
379 dev->virtqueue[index] = vq;
380 vhost_devices[dev->vid] = dev;
383 vhost_user_iotlb_init(dev, index);
388 static struct virtio_net*
389 numa_realloc(struct virtio_net *dev, int index __rte_unused)
395 /* Converts QEMU virtual address to Vhost virtual address. */
397 qva_to_vva(struct virtio_net *dev, uint64_t qva)
399 struct rte_vhost_mem_region *reg;
402 /* Find the region where the address lives. */
403 for (i = 0; i < dev->mem->nregions; i++) {
404 reg = &dev->mem->regions[i];
406 if (qva >= reg->guest_user_addr &&
407 qva < reg->guest_user_addr + reg->size) {
408 return qva - reg->guest_user_addr +
418 * Converts ring address to Vhost virtual address.
419 * If IOMMU is enabled, the ring address is a guest IO virtual address,
420 * else it is a QEMU virtual address.
423 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
424 uint64_t ra, uint64_t size)
426 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
429 vva = vhost_user_iotlb_cache_find(vq, ra,
430 &size, VHOST_ACCESS_RW);
432 vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
437 return qva_to_vva(dev, ra);
440 static struct virtio_net *
441 translate_ring_addresses(struct virtio_net *dev, int vq_index)
443 struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
444 struct vhost_vring_addr *addr = &vq->ring_addrs;
446 /* The addresses are converted from QEMU virtual to Vhost virtual. */
447 if (vq->desc && vq->avail && vq->used)
450 vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
451 vq, addr->desc_user_addr, sizeof(struct vring_desc));
453 RTE_LOG(DEBUG, VHOST_CONFIG,
454 "(%d) failed to find desc ring address.\n",
459 dev = numa_realloc(dev, vq_index);
460 vq = dev->virtqueue[vq_index];
461 addr = &vq->ring_addrs;
463 vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
464 vq, addr->avail_user_addr, sizeof(struct vring_avail));
465 if (vq->avail == 0) {
466 RTE_LOG(DEBUG, VHOST_CONFIG,
467 "(%d) failed to find avail ring address.\n",
472 vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
473 vq, addr->used_user_addr, sizeof(struct vring_used));
475 RTE_LOG(DEBUG, VHOST_CONFIG,
476 "(%d) failed to find used ring address.\n",
481 if (vq->last_used_idx != vq->used->idx) {
482 RTE_LOG(WARNING, VHOST_CONFIG,
483 "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
484 "some packets maybe resent for Tx and dropped for Rx\n",
485 vq->last_used_idx, vq->used->idx);
486 vq->last_used_idx = vq->used->idx;
487 vq->last_avail_idx = vq->used->idx;
490 vq->log_guest_addr = addr->log_guest_addr;
492 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
494 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
495 dev->vid, vq->avail);
496 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
498 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
499 dev->vid, vq->log_guest_addr);
505 * The virtio device sends us the desc, used and avail ring addresses.
506 * This function then converts these to our address space.
509 vhost_user_set_vring_addr(struct virtio_net **pdev, VhostUserMsg *msg)
511 struct vhost_virtqueue *vq;
512 struct vhost_vring_addr *addr = &msg->payload.addr;
513 struct virtio_net *dev = *pdev;
515 if (dev->mem == NULL)
518 /* addr->index refers to the queue index. The txq 1, rxq is 0. */
519 vq = dev->virtqueue[msg->payload.addr.index];
522 * Rings addresses should not be interpreted as long as the ring is not
523 * started and enabled
525 memcpy(&vq->ring_addrs, addr, sizeof(*addr));
527 vring_invalidate(dev, vq);
529 if (vq->enabled && (dev->features &
530 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
531 dev = translate_ring_addresses(dev, msg->payload.addr.index);
542 * The virtio device sends us the available ring last used index.
545 vhost_user_set_vring_base(struct virtio_net *dev,
548 dev->virtqueue[msg->payload.state.index]->last_used_idx =
549 msg->payload.state.num;
550 dev->virtqueue[msg->payload.state.index]->last_avail_idx =
551 msg->payload.state.num;
557 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
558 uint64_t host_phys_addr, uint64_t size)
560 struct guest_page *page, *last_page;
562 if (dev->nr_guest_pages == dev->max_guest_pages) {
563 dev->max_guest_pages *= 2;
564 dev->guest_pages = realloc(dev->guest_pages,
565 dev->max_guest_pages * sizeof(*page));
566 if (!dev->guest_pages) {
567 RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
572 if (dev->nr_guest_pages > 0) {
573 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
574 /* merge if the two pages are continuous */
575 if (host_phys_addr == last_page->host_phys_addr +
577 last_page->size += size;
582 page = &dev->guest_pages[dev->nr_guest_pages++];
583 page->guest_phys_addr = guest_phys_addr;
584 page->host_phys_addr = host_phys_addr;
591 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
594 uint64_t reg_size = reg->size;
595 uint64_t host_user_addr = reg->host_user_addr;
596 uint64_t guest_phys_addr = reg->guest_phys_addr;
597 uint64_t host_phys_addr;
600 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
601 size = page_size - (guest_phys_addr & (page_size - 1));
602 size = RTE_MIN(size, reg_size);
604 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
607 host_user_addr += size;
608 guest_phys_addr += size;
611 while (reg_size > 0) {
612 size = RTE_MIN(reg_size, page_size);
613 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
615 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
619 host_user_addr += size;
620 guest_phys_addr += size;
627 #ifdef RTE_LIBRTE_VHOST_DEBUG
628 /* TODO: enable it only in debug mode? */
630 dump_guest_pages(struct virtio_net *dev)
633 struct guest_page *page;
635 for (i = 0; i < dev->nr_guest_pages; i++) {
636 page = &dev->guest_pages[i];
638 RTE_LOG(INFO, VHOST_CONFIG,
639 "guest physical page region %u\n"
640 "\t guest_phys_addr: %" PRIx64 "\n"
641 "\t host_phys_addr : %" PRIx64 "\n"
642 "\t size : %" PRIx64 "\n",
644 page->guest_phys_addr,
645 page->host_phys_addr,
650 #define dump_guest_pages(dev)
654 vhost_memory_changed(struct VhostUserMemory *new,
655 struct rte_vhost_memory *old)
659 if (new->nregions != old->nregions)
662 for (i = 0; i < new->nregions; ++i) {
663 VhostUserMemoryRegion *new_r = &new->regions[i];
664 struct rte_vhost_mem_region *old_r = &old->regions[i];
666 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
668 if (new_r->memory_size != old_r->size)
670 if (new_r->userspace_addr != old_r->guest_user_addr)
678 vhost_user_set_mem_table(struct virtio_net *dev, struct VhostUserMsg *pmsg)
680 struct VhostUserMemory memory = pmsg->payload.memory;
681 struct rte_vhost_mem_region *reg;
684 uint64_t mmap_offset;
689 if (memory.nregions > VHOST_MEMORY_MAX_NREGIONS) {
690 RTE_LOG(ERR, VHOST_CONFIG,
691 "too many memory regions (%u)\n", memory.nregions);
695 if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
696 RTE_LOG(INFO, VHOST_CONFIG,
697 "(%d) memory regions not changed\n", dev->vid);
699 for (i = 0; i < memory.nregions; i++)
706 free_mem_region(dev);
711 dev->nr_guest_pages = 0;
712 if (!dev->guest_pages) {
713 dev->max_guest_pages = 8;
714 dev->guest_pages = malloc(dev->max_guest_pages *
715 sizeof(struct guest_page));
716 if (dev->guest_pages == NULL) {
717 RTE_LOG(ERR, VHOST_CONFIG,
718 "(%d) failed to allocate memory "
719 "for dev->guest_pages\n",
725 dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
726 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
727 if (dev->mem == NULL) {
728 RTE_LOG(ERR, VHOST_CONFIG,
729 "(%d) failed to allocate memory for dev->mem\n",
733 dev->mem->nregions = memory.nregions;
735 for (i = 0; i < memory.nregions; i++) {
737 reg = &dev->mem->regions[i];
739 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
740 reg->guest_user_addr = memory.regions[i].userspace_addr;
741 reg->size = memory.regions[i].memory_size;
744 mmap_offset = memory.regions[i].mmap_offset;
746 /* Check for memory_size + mmap_offset overflow */
747 if (mmap_offset >= -reg->size) {
748 RTE_LOG(ERR, VHOST_CONFIG,
749 "mmap_offset (%#"PRIx64") and memory_size "
750 "(%#"PRIx64") overflow\n",
751 mmap_offset, reg->size);
755 mmap_size = reg->size + mmap_offset;
757 /* mmap() without flag of MAP_ANONYMOUS, should be called
758 * with length argument aligned with hugepagesz at older
759 * longterm version Linux, like 2.6.32 and 3.2.72, or
760 * mmap() will fail with EINVAL.
762 * to avoid failure, make sure in caller to keep length
765 alignment = get_blk_size(fd);
766 if (alignment == (uint64_t)-1) {
767 RTE_LOG(ERR, VHOST_CONFIG,
768 "couldn't get hugepage size through fstat\n");
771 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
773 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
774 MAP_SHARED | MAP_POPULATE, fd, 0);
776 if (mmap_addr == MAP_FAILED) {
777 RTE_LOG(ERR, VHOST_CONFIG,
778 "mmap region %u failed.\n", i);
782 reg->mmap_addr = mmap_addr;
783 reg->mmap_size = mmap_size;
784 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
787 if (dev->dequeue_zero_copy)
788 if (add_guest_pages(dev, reg, alignment) < 0) {
789 RTE_LOG(ERR, VHOST_CONFIG,
790 "adding guest pages to region %u failed.\n",
795 RTE_LOG(INFO, VHOST_CONFIG,
796 "guest memory region %u, size: 0x%" PRIx64 "\n"
797 "\t guest physical addr: 0x%" PRIx64 "\n"
798 "\t guest virtual addr: 0x%" PRIx64 "\n"
799 "\t host virtual addr: 0x%" PRIx64 "\n"
800 "\t mmap addr : 0x%" PRIx64 "\n"
801 "\t mmap size : 0x%" PRIx64 "\n"
802 "\t mmap align: 0x%" PRIx64 "\n"
803 "\t mmap off : 0x%" PRIx64 "\n",
805 reg->guest_phys_addr,
806 reg->guest_user_addr,
808 (uint64_t)(uintptr_t)mmap_addr,
814 dump_guest_pages(dev);
819 free_mem_region(dev);
826 vq_is_ready(struct vhost_virtqueue *vq)
828 return vq && vq->desc && vq->avail && vq->used &&
829 vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
830 vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
834 virtio_is_ready(struct virtio_net *dev)
836 struct vhost_virtqueue *vq;
839 if (dev->nr_vring == 0)
842 for (i = 0; i < dev->nr_vring; i++) {
843 vq = dev->virtqueue[i];
845 if (!vq_is_ready(vq))
849 RTE_LOG(INFO, VHOST_CONFIG,
850 "virtio is now ready for processing.\n");
855 vhost_user_set_vring_call(struct virtio_net *dev, struct VhostUserMsg *pmsg)
857 struct vhost_vring_file file;
858 struct vhost_virtqueue *vq;
860 file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
861 if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
862 file.fd = VIRTIO_INVALID_EVENTFD;
864 file.fd = pmsg->fds[0];
865 RTE_LOG(INFO, VHOST_CONFIG,
866 "vring call idx:%d file:%d\n", file.index, file.fd);
868 vq = dev->virtqueue[file.index];
872 vq->callfd = file.fd;
876 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *pmsg)
878 struct vhost_vring_file file;
879 struct vhost_virtqueue *vq;
880 struct virtio_net *dev = *pdev;
882 file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
883 if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
884 file.fd = VIRTIO_INVALID_EVENTFD;
886 file.fd = pmsg->fds[0];
887 RTE_LOG(INFO, VHOST_CONFIG,
888 "vring kick idx:%d file:%d\n", file.index, file.fd);
890 /* Interpret ring addresses only when ring is started. */
891 dev = translate_ring_addresses(dev, file.index);
897 vq = dev->virtqueue[file.index];
900 * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
901 * the ring starts already enabled. Otherwise, it is enabled via
902 * the SET_VRING_ENABLE message.
904 if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
909 vq->kickfd = file.fd;
913 free_zmbufs(struct vhost_virtqueue *vq)
915 struct zcopy_mbuf *zmbuf, *next;
917 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
918 zmbuf != NULL; zmbuf = next) {
919 next = TAILQ_NEXT(zmbuf, next);
921 rte_pktmbuf_free(zmbuf->mbuf);
922 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
925 rte_free(vq->zmbufs);
929 * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
932 vhost_user_get_vring_base(struct virtio_net *dev,
935 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
937 /* We have to stop the queue (virtio) if it is running. */
938 if (dev->flags & VIRTIO_DEV_RUNNING) {
939 dev->flags &= ~VIRTIO_DEV_RUNNING;
940 dev->notify_ops->destroy_device(dev->vid);
943 dev->flags &= ~VIRTIO_DEV_READY;
945 /* Here we are safe to get the last used index */
946 msg->payload.state.num = vq->last_used_idx;
948 RTE_LOG(INFO, VHOST_CONFIG,
949 "vring base idx:%d file:%d\n", msg->payload.state.index,
950 msg->payload.state.num);
952 * Based on current qemu vhost-user implementation, this message is
953 * sent and only sent in vhost_vring_stop.
954 * TODO: cleanup the vring, it isn't usable since here.
959 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
964 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
966 if (dev->dequeue_zero_copy)
968 rte_free(vq->shadow_used_ring);
969 vq->shadow_used_ring = NULL;
971 rte_free(vq->batch_copy_elems);
972 vq->batch_copy_elems = NULL;
978 * when virtio queues are ready to work, qemu will send us to
979 * enable the virtio queue pair.
982 vhost_user_set_vring_enable(struct virtio_net *dev,
985 int enable = (int)msg->payload.state.num;
987 RTE_LOG(INFO, VHOST_CONFIG,
988 "set queue enable: %d to qp idx: %d\n",
989 enable, msg->payload.state.index);
991 if (dev->notify_ops->vring_state_changed)
992 dev->notify_ops->vring_state_changed(dev->vid,
993 msg->payload.state.index, enable);
995 dev->virtqueue[msg->payload.state.index]->enabled = enable;
1001 vhost_user_get_protocol_features(struct virtio_net *dev,
1002 struct VhostUserMsg *msg)
1004 uint64_t features, protocol_features = VHOST_USER_PROTOCOL_FEATURES;
1006 rte_vhost_driver_get_features(dev->ifname, &features);
1009 * REPLY_ACK protocol feature is only mandatory for now
1010 * for IOMMU feature. If IOMMU is explicitly disabled by the
1011 * application, disable also REPLY_ACK feature for older buggy
1012 * Qemu versions (from v2.7.0 to v2.9.0).
1014 if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1015 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1017 msg->payload.u64 = protocol_features;
1018 msg->size = sizeof(msg->payload.u64);
1022 vhost_user_set_protocol_features(struct virtio_net *dev,
1023 uint64_t protocol_features)
1025 if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES)
1028 dev->protocol_features = protocol_features;
1032 vhost_user_set_log_base(struct virtio_net *dev, struct VhostUserMsg *msg)
1034 int fd = msg->fds[0];
1039 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1043 if (msg->size != sizeof(VhostUserLog)) {
1044 RTE_LOG(ERR, VHOST_CONFIG,
1045 "invalid log base msg size: %"PRId32" != %d\n",
1046 msg->size, (int)sizeof(VhostUserLog));
1050 size = msg->payload.log.mmap_size;
1051 off = msg->payload.log.mmap_offset;
1053 /* Don't allow mmap_offset to point outside the mmap region */
1055 RTE_LOG(ERR, VHOST_CONFIG,
1056 "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1061 RTE_LOG(INFO, VHOST_CONFIG,
1062 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1066 * mmap from 0 to workaround a hugepage mmap bug: mmap will
1067 * fail when offset is not page size aligned.
1069 addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1071 if (addr == MAP_FAILED) {
1072 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1077 * Free previously mapped log memory on occasionally
1078 * multiple VHOST_USER_SET_LOG_BASE.
1080 if (dev->log_addr) {
1081 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1083 dev->log_addr = (uint64_t)(uintptr_t)addr;
1084 dev->log_base = dev->log_addr + off;
1085 dev->log_size = size;
1091 * An rarp packet is constructed and broadcasted to notify switches about
1092 * the new location of the migrated VM, so that packets from outside will
1093 * not be lost after migration.
1095 * However, we don't actually "send" a rarp packet here, instead, we set
1096 * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1099 vhost_user_send_rarp(struct virtio_net *dev, struct VhostUserMsg *msg)
1101 uint8_t *mac = (uint8_t *)&msg->payload.u64;
1103 RTE_LOG(DEBUG, VHOST_CONFIG,
1104 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1105 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1106 memcpy(dev->mac.addr_bytes, mac, 6);
1109 * Set the flag to inject a RARP broadcast packet at
1110 * rte_vhost_dequeue_burst().
1112 * rte_smp_wmb() is for making sure the mac is copied
1113 * before the flag is set.
1116 rte_atomic16_set(&dev->broadcast_rarp, 1);
1122 vhost_user_net_set_mtu(struct virtio_net *dev, struct VhostUserMsg *msg)
1124 if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1125 msg->payload.u64 > VIRTIO_MAX_MTU) {
1126 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1132 dev->mtu = msg->payload.u64;
1138 vhost_user_set_req_fd(struct virtio_net *dev, struct VhostUserMsg *msg)
1140 int fd = msg->fds[0];
1143 RTE_LOG(ERR, VHOST_CONFIG,
1144 "Invalid file descriptor for slave channel (%d)\n",
1149 dev->slave_req_fd = fd;
1155 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1157 struct vhost_vring_addr *ra;
1158 uint64_t start, end;
1161 end = start + imsg->size;
1163 ra = &vq->ring_addrs;
1164 if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1166 if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1168 if (ra->used_user_addr >= start && ra->used_user_addr < end)
1175 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1176 struct vhost_iotlb_msg *imsg)
1178 uint64_t istart, iend, vstart, vend;
1180 istart = imsg->iova;
1181 iend = istart + imsg->size - 1;
1183 vstart = (uintptr_t)vq->desc;
1184 vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1185 if (vstart <= iend && istart <= vend)
1188 vstart = (uintptr_t)vq->avail;
1189 vend = vstart + sizeof(struct vring_avail);
1190 vend += sizeof(uint16_t) * vq->size - 1;
1191 if (vstart <= iend && istart <= vend)
1194 vstart = (uintptr_t)vq->used;
1195 vend = vstart + sizeof(struct vring_used);
1196 vend += sizeof(struct vring_used_elem) * vq->size - 1;
1197 if (vstart <= iend && istart <= vend)
1204 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg)
1206 struct virtio_net *dev = *pdev;
1207 struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1211 switch (imsg->type) {
1212 case VHOST_IOTLB_UPDATE:
1213 vva = qva_to_vva(dev, imsg->uaddr);
1217 for (i = 0; i < dev->nr_vring; i++) {
1218 struct vhost_virtqueue *vq = dev->virtqueue[i];
1220 vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1221 imsg->size, imsg->perm);
1223 if (is_vring_iotlb_update(vq, imsg))
1224 *pdev = dev = translate_ring_addresses(dev, i);
1227 case VHOST_IOTLB_INVALIDATE:
1228 for (i = 0; i < dev->nr_vring; i++) {
1229 struct vhost_virtqueue *vq = dev->virtqueue[i];
1231 vhost_user_iotlb_cache_remove(vq, imsg->iova,
1234 if (is_vring_iotlb_invalidate(vq, imsg))
1235 vring_invalidate(dev, vq);
1239 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1247 /* return bytes# of read on success or negative val on failure. */
1249 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1253 ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1254 msg->fds, VHOST_MEMORY_MAX_NREGIONS);
1258 if (msg && msg->size) {
1259 if (msg->size > sizeof(msg->payload)) {
1260 RTE_LOG(ERR, VHOST_CONFIG,
1261 "invalid msg size: %d\n", msg->size);
1264 ret = read(sockfd, &msg->payload, msg->size);
1267 if (ret != (int)msg->size) {
1268 RTE_LOG(ERR, VHOST_CONFIG,
1269 "read control message failed\n");
1278 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1283 return send_fd_message(sockfd, (char *)msg,
1284 VHOST_USER_HDR_SIZE + msg->size, NULL, 0);
1288 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1293 msg->flags &= ~VHOST_USER_VERSION_MASK;
1294 msg->flags &= ~VHOST_USER_NEED_REPLY;
1295 msg->flags |= VHOST_USER_VERSION;
1296 msg->flags |= VHOST_USER_REPLY_MASK;
1298 return send_vhost_message(sockfd, msg);
1302 * Allocate a queue pair if it hasn't been allocated yet
1305 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev, VhostUserMsg *msg)
1309 switch (msg->request.master) {
1310 case VHOST_USER_SET_VRING_KICK:
1311 case VHOST_USER_SET_VRING_CALL:
1312 case VHOST_USER_SET_VRING_ERR:
1313 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1315 case VHOST_USER_SET_VRING_NUM:
1316 case VHOST_USER_SET_VRING_BASE:
1317 case VHOST_USER_SET_VRING_ENABLE:
1318 vring_idx = msg->payload.state.index;
1320 case VHOST_USER_SET_VRING_ADDR:
1321 vring_idx = msg->payload.addr.index;
1327 if (vring_idx >= VHOST_MAX_VRING) {
1328 RTE_LOG(ERR, VHOST_CONFIG,
1329 "invalid vring index: %u\n", vring_idx);
1333 if (dev->virtqueue[vring_idx])
1336 return alloc_vring_queue(dev, vring_idx);
1340 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1343 unsigned int vq_num = 0;
1345 while (vq_num < dev->nr_vring) {
1346 struct vhost_virtqueue *vq = dev->virtqueue[i];
1349 rte_spinlock_lock(&vq->access_lock);
1357 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1360 unsigned int vq_num = 0;
1362 while (vq_num < dev->nr_vring) {
1363 struct vhost_virtqueue *vq = dev->virtqueue[i];
1366 rte_spinlock_unlock(&vq->access_lock);
1374 vhost_user_msg_handler(int vid, int fd)
1376 struct virtio_net *dev;
1377 struct VhostUserMsg msg;
1379 int unlock_required = 0;
1381 dev = get_device(vid);
1385 if (!dev->notify_ops) {
1386 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1387 if (!dev->notify_ops) {
1388 RTE_LOG(ERR, VHOST_CONFIG,
1389 "failed to get callback ops for driver %s\n",
1395 ret = read_vhost_message(fd, &msg);
1396 if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1398 RTE_LOG(ERR, VHOST_CONFIG,
1399 "vhost read message failed\n");
1401 RTE_LOG(INFO, VHOST_CONFIG,
1402 "vhost peer closed\n");
1404 RTE_LOG(ERR, VHOST_CONFIG,
1405 "vhost read incorrect message\n");
1411 if (msg.request.master != VHOST_USER_IOTLB_MSG)
1412 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1413 vhost_message_str[msg.request.master]);
1415 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1416 vhost_message_str[msg.request.master]);
1418 ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1420 RTE_LOG(ERR, VHOST_CONFIG,
1421 "failed to alloc queue\n");
1426 * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1427 * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1428 * and device is destroyed. destroy_device waits for queues to be
1429 * inactive, so it is safe. Otherwise taking the access_lock
1430 * would cause a dead lock.
1432 switch (msg.request.master) {
1433 case VHOST_USER_SET_FEATURES:
1434 case VHOST_USER_SET_PROTOCOL_FEATURES:
1435 case VHOST_USER_SET_OWNER:
1436 case VHOST_USER_SET_MEM_TABLE:
1437 case VHOST_USER_SET_LOG_BASE:
1438 case VHOST_USER_SET_LOG_FD:
1439 case VHOST_USER_SET_VRING_NUM:
1440 case VHOST_USER_SET_VRING_ADDR:
1441 case VHOST_USER_SET_VRING_BASE:
1442 case VHOST_USER_SET_VRING_KICK:
1443 case VHOST_USER_SET_VRING_CALL:
1444 case VHOST_USER_SET_VRING_ERR:
1445 case VHOST_USER_SET_VRING_ENABLE:
1446 case VHOST_USER_SEND_RARP:
1447 case VHOST_USER_NET_SET_MTU:
1448 case VHOST_USER_SET_SLAVE_REQ_FD:
1449 vhost_user_lock_all_queue_pairs(dev);
1450 unlock_required = 1;
1457 switch (msg.request.master) {
1458 case VHOST_USER_GET_FEATURES:
1459 msg.payload.u64 = vhost_user_get_features(dev);
1460 msg.size = sizeof(msg.payload.u64);
1461 send_vhost_reply(fd, &msg);
1463 case VHOST_USER_SET_FEATURES:
1464 ret = vhost_user_set_features(dev, msg.payload.u64);
1469 case VHOST_USER_GET_PROTOCOL_FEATURES:
1470 vhost_user_get_protocol_features(dev, &msg);
1471 send_vhost_reply(fd, &msg);
1473 case VHOST_USER_SET_PROTOCOL_FEATURES:
1474 vhost_user_set_protocol_features(dev, msg.payload.u64);
1477 case VHOST_USER_SET_OWNER:
1478 vhost_user_set_owner();
1480 case VHOST_USER_RESET_OWNER:
1481 vhost_user_reset_owner(dev);
1484 case VHOST_USER_SET_MEM_TABLE:
1485 ret = vhost_user_set_mem_table(dev, &msg);
1488 case VHOST_USER_SET_LOG_BASE:
1489 vhost_user_set_log_base(dev, &msg);
1491 /* it needs a reply */
1492 msg.size = sizeof(msg.payload.u64);
1493 send_vhost_reply(fd, &msg);
1495 case VHOST_USER_SET_LOG_FD:
1497 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1500 case VHOST_USER_SET_VRING_NUM:
1501 vhost_user_set_vring_num(dev, &msg);
1503 case VHOST_USER_SET_VRING_ADDR:
1504 vhost_user_set_vring_addr(&dev, &msg);
1506 case VHOST_USER_SET_VRING_BASE:
1507 vhost_user_set_vring_base(dev, &msg);
1510 case VHOST_USER_GET_VRING_BASE:
1511 vhost_user_get_vring_base(dev, &msg);
1512 msg.size = sizeof(msg.payload.state);
1513 send_vhost_reply(fd, &msg);
1516 case VHOST_USER_SET_VRING_KICK:
1517 vhost_user_set_vring_kick(&dev, &msg);
1519 case VHOST_USER_SET_VRING_CALL:
1520 vhost_user_set_vring_call(dev, &msg);
1523 case VHOST_USER_SET_VRING_ERR:
1524 if (!(msg.payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1526 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1529 case VHOST_USER_GET_QUEUE_NUM:
1530 msg.payload.u64 = VHOST_MAX_QUEUE_PAIRS;
1531 msg.size = sizeof(msg.payload.u64);
1532 send_vhost_reply(fd, &msg);
1535 case VHOST_USER_SET_VRING_ENABLE:
1536 vhost_user_set_vring_enable(dev, &msg);
1538 case VHOST_USER_SEND_RARP:
1539 vhost_user_send_rarp(dev, &msg);
1542 case VHOST_USER_NET_SET_MTU:
1543 ret = vhost_user_net_set_mtu(dev, &msg);
1546 case VHOST_USER_SET_SLAVE_REQ_FD:
1547 ret = vhost_user_set_req_fd(dev, &msg);
1550 case VHOST_USER_IOTLB_MSG:
1551 ret = vhost_user_iotlb_msg(&dev, &msg);
1560 if (unlock_required)
1561 vhost_user_unlock_all_queue_pairs(dev);
1563 if (msg.flags & VHOST_USER_NEED_REPLY) {
1564 msg.payload.u64 = !!ret;
1565 msg.size = sizeof(msg.payload.u64);
1566 send_vhost_reply(fd, &msg);
1569 if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1570 dev->flags |= VIRTIO_DEV_READY;
1572 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1573 if (dev->dequeue_zero_copy) {
1574 RTE_LOG(INFO, VHOST_CONFIG,
1575 "dequeue zero copy is enabled\n");
1578 if (dev->notify_ops->new_device(dev->vid) == 0)
1579 dev->flags |= VIRTIO_DEV_RUNNING;
1587 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
1590 struct VhostUserMsg msg = {
1591 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
1592 .flags = VHOST_USER_VERSION,
1593 .size = sizeof(msg.payload.iotlb),
1597 .type = VHOST_IOTLB_MISS,
1601 ret = send_vhost_message(dev->slave_req_fd, &msg);
1603 RTE_LOG(ERR, VHOST_CONFIG,
1604 "Failed to send IOTLB miss message (%d)\n",