1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2016 Intel Corporation
11 #include <sys/types.h>
14 #ifdef RTE_LIBRTE_VHOST_NUMA
18 #include <rte_common.h>
19 #include <rte_malloc.h>
24 #include "vhost_user.h"
26 #define VIRTIO_MIN_MTU 68
27 #define VIRTIO_MAX_MTU 65535
29 static const char *vhost_message_str[VHOST_USER_MAX] = {
30 [VHOST_USER_NONE] = "VHOST_USER_NONE",
31 [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
32 [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
33 [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
34 [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
35 [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
36 [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
37 [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
38 [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
39 [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
40 [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
41 [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
42 [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
43 [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
44 [VHOST_USER_SET_VRING_ERR] = "VHOST_USER_SET_VRING_ERR",
45 [VHOST_USER_GET_PROTOCOL_FEATURES] = "VHOST_USER_GET_PROTOCOL_FEATURES",
46 [VHOST_USER_SET_PROTOCOL_FEATURES] = "VHOST_USER_SET_PROTOCOL_FEATURES",
47 [VHOST_USER_GET_QUEUE_NUM] = "VHOST_USER_GET_QUEUE_NUM",
48 [VHOST_USER_SET_VRING_ENABLE] = "VHOST_USER_SET_VRING_ENABLE",
49 [VHOST_USER_SEND_RARP] = "VHOST_USER_SEND_RARP",
50 [VHOST_USER_NET_SET_MTU] = "VHOST_USER_NET_SET_MTU",
51 [VHOST_USER_SET_SLAVE_REQ_FD] = "VHOST_USER_SET_SLAVE_REQ_FD",
52 [VHOST_USER_IOTLB_MSG] = "VHOST_USER_IOTLB_MSG",
61 ret = fstat(fd, &stat);
62 return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
66 free_mem_region(struct virtio_net *dev)
69 struct rte_vhost_mem_region *reg;
71 if (!dev || !dev->mem)
74 for (i = 0; i < dev->mem->nregions; i++) {
75 reg = &dev->mem->regions[i];
76 if (reg->host_user_addr) {
77 munmap(reg->mmap_addr, reg->mmap_size);
84 vhost_backend_cleanup(struct virtio_net *dev)
92 free(dev->guest_pages);
93 dev->guest_pages = NULL;
96 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
100 if (dev->slave_req_fd >= 0) {
101 close(dev->slave_req_fd);
102 dev->slave_req_fd = -1;
107 * This function just returns success at the moment unless
108 * the device hasn't been initialised.
111 vhost_user_set_owner(void)
117 vhost_user_reset_owner(struct virtio_net *dev)
119 if (dev->flags & VIRTIO_DEV_RUNNING) {
120 dev->flags &= ~VIRTIO_DEV_RUNNING;
121 dev->notify_ops->destroy_device(dev->vid);
124 cleanup_device(dev, 0);
130 * The features that we support are requested.
133 vhost_user_get_features(struct virtio_net *dev)
135 uint64_t features = 0;
137 rte_vhost_driver_get_features(dev->ifname, &features);
142 * We receive the negotiated features supported by us and the virtio device.
145 vhost_user_set_features(struct virtio_net *dev, uint64_t features)
147 uint64_t vhost_features = 0;
149 rte_vhost_driver_get_features(dev->ifname, &vhost_features);
150 if (features & ~vhost_features) {
151 RTE_LOG(ERR, VHOST_CONFIG,
152 "(%d) received invalid negotiated features.\n",
157 if (dev->flags & VIRTIO_DEV_RUNNING) {
158 if (dev->features == features)
162 * Error out if master tries to change features while device is
163 * in running state. The exception being VHOST_F_LOG_ALL, which
164 * is enabled when the live-migration starts.
166 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
167 RTE_LOG(ERR, VHOST_CONFIG,
168 "(%d) features changed while device is running.\n",
173 if (dev->notify_ops->features_changed)
174 dev->notify_ops->features_changed(dev->vid, features);
177 dev->features = features;
179 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
180 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
182 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
184 LOG_DEBUG(VHOST_CONFIG,
185 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
187 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
188 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
190 if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
191 !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
193 * Remove all but first queue pair if MQ hasn't been
194 * negotiated. This is safe because the device is not
195 * running at this stage.
197 while (dev->nr_vring > 2) {
198 struct vhost_virtqueue *vq;
200 vq = dev->virtqueue[--dev->nr_vring];
204 dev->virtqueue[dev->nr_vring] = NULL;
214 * The virtio device sends us the size of the descriptor ring.
217 vhost_user_set_vring_num(struct virtio_net *dev,
220 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
222 vq->size = msg->payload.state.num;
224 if (dev->dequeue_zero_copy) {
226 vq->last_zmbuf_idx = 0;
227 vq->zmbuf_size = vq->size;
228 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
229 sizeof(struct zcopy_mbuf), 0);
230 if (vq->zmbufs == NULL) {
231 RTE_LOG(WARNING, VHOST_CONFIG,
232 "failed to allocate mem for zero copy; "
233 "zero copy is force disabled\n");
234 dev->dequeue_zero_copy = 0;
236 TAILQ_INIT(&vq->zmbuf_list);
239 vq->shadow_used_ring = rte_malloc(NULL,
240 vq->size * sizeof(struct vring_used_elem),
241 RTE_CACHE_LINE_SIZE);
242 if (!vq->shadow_used_ring) {
243 RTE_LOG(ERR, VHOST_CONFIG,
244 "failed to allocate memory for shadow used ring.\n");
248 vq->batch_copy_elems = rte_malloc(NULL,
249 vq->size * sizeof(struct batch_copy_elem),
250 RTE_CACHE_LINE_SIZE);
251 if (!vq->batch_copy_elems) {
252 RTE_LOG(ERR, VHOST_CONFIG,
253 "failed to allocate memory for batching copy.\n");
261 * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
262 * same numa node as the memory of vring descriptor.
264 #ifdef RTE_LIBRTE_VHOST_NUMA
265 static struct virtio_net*
266 numa_realloc(struct virtio_net *dev, int index)
268 int oldnode, newnode;
269 struct virtio_net *old_dev;
270 struct vhost_virtqueue *old_vq, *vq;
271 struct zcopy_mbuf *new_zmbuf;
272 struct vring_used_elem *new_shadow_used_ring;
273 struct batch_copy_elem *new_batch_copy_elems;
277 vq = old_vq = dev->virtqueue[index];
279 ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
280 MPOL_F_NODE | MPOL_F_ADDR);
282 /* check if we need to reallocate vq */
283 ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
284 MPOL_F_NODE | MPOL_F_ADDR);
286 RTE_LOG(ERR, VHOST_CONFIG,
287 "Unable to get vq numa information.\n");
290 if (oldnode != newnode) {
291 RTE_LOG(INFO, VHOST_CONFIG,
292 "reallocate vq from %d to %d node\n", oldnode, newnode);
293 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
297 memcpy(vq, old_vq, sizeof(*vq));
298 TAILQ_INIT(&vq->zmbuf_list);
300 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
301 sizeof(struct zcopy_mbuf), 0, newnode);
303 rte_free(vq->zmbufs);
304 vq->zmbufs = new_zmbuf;
307 new_shadow_used_ring = rte_malloc_socket(NULL,
308 vq->size * sizeof(struct vring_used_elem),
311 if (new_shadow_used_ring) {
312 rte_free(vq->shadow_used_ring);
313 vq->shadow_used_ring = new_shadow_used_ring;
316 new_batch_copy_elems = rte_malloc_socket(NULL,
317 vq->size * sizeof(struct batch_copy_elem),
320 if (new_batch_copy_elems) {
321 rte_free(vq->batch_copy_elems);
322 vq->batch_copy_elems = new_batch_copy_elems;
328 /* check if we need to reallocate dev */
329 ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
330 MPOL_F_NODE | MPOL_F_ADDR);
332 RTE_LOG(ERR, VHOST_CONFIG,
333 "Unable to get dev numa information.\n");
336 if (oldnode != newnode) {
337 RTE_LOG(INFO, VHOST_CONFIG,
338 "reallocate dev from %d to %d node\n",
340 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
346 memcpy(dev, old_dev, sizeof(*dev));
351 dev->virtqueue[index] = vq;
352 vhost_devices[dev->vid] = dev;
355 vhost_user_iotlb_init(dev, index);
360 static struct virtio_net*
361 numa_realloc(struct virtio_net *dev, int index __rte_unused)
367 /* Converts QEMU virtual address to Vhost virtual address. */
369 qva_to_vva(struct virtio_net *dev, uint64_t qva)
371 struct rte_vhost_mem_region *reg;
374 /* Find the region where the address lives. */
375 for (i = 0; i < dev->mem->nregions; i++) {
376 reg = &dev->mem->regions[i];
378 if (qva >= reg->guest_user_addr &&
379 qva < reg->guest_user_addr + reg->size) {
380 return qva - reg->guest_user_addr +
390 * Converts ring address to Vhost virtual address.
391 * If IOMMU is enabled, the ring address is a guest IO virtual address,
392 * else it is a QEMU virtual address.
395 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
396 uint64_t ra, uint64_t size)
398 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
401 vva = vhost_user_iotlb_cache_find(vq, ra,
402 &size, VHOST_ACCESS_RW);
404 vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
409 return qva_to_vva(dev, ra);
412 static struct virtio_net *
413 translate_ring_addresses(struct virtio_net *dev, int vq_index)
415 struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
416 struct vhost_vring_addr *addr = &vq->ring_addrs;
418 /* The addresses are converted from QEMU virtual to Vhost virtual. */
419 if (vq->desc && vq->avail && vq->used)
422 vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
423 vq, addr->desc_user_addr, sizeof(struct vring_desc));
425 RTE_LOG(DEBUG, VHOST_CONFIG,
426 "(%d) failed to find desc ring address.\n",
431 dev = numa_realloc(dev, vq_index);
432 vq = dev->virtqueue[vq_index];
433 addr = &vq->ring_addrs;
435 vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
436 vq, addr->avail_user_addr, sizeof(struct vring_avail));
437 if (vq->avail == 0) {
438 RTE_LOG(DEBUG, VHOST_CONFIG,
439 "(%d) failed to find avail ring address.\n",
444 vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
445 vq, addr->used_user_addr, sizeof(struct vring_used));
447 RTE_LOG(DEBUG, VHOST_CONFIG,
448 "(%d) failed to find used ring address.\n",
453 if (vq->last_used_idx != vq->used->idx) {
454 RTE_LOG(WARNING, VHOST_CONFIG,
455 "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
456 "some packets maybe resent for Tx and dropped for Rx\n",
457 vq->last_used_idx, vq->used->idx);
458 vq->last_used_idx = vq->used->idx;
459 vq->last_avail_idx = vq->used->idx;
462 vq->log_guest_addr = addr->log_guest_addr;
464 LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
466 LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
467 dev->vid, vq->avail);
468 LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
470 LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
471 dev->vid, vq->log_guest_addr);
477 * The virtio device sends us the desc, used and avail ring addresses.
478 * This function then converts these to our address space.
481 vhost_user_set_vring_addr(struct virtio_net **pdev, VhostUserMsg *msg)
483 struct vhost_virtqueue *vq;
484 struct vhost_vring_addr *addr = &msg->payload.addr;
485 struct virtio_net *dev = *pdev;
487 if (dev->mem == NULL)
490 /* addr->index refers to the queue index. The txq 1, rxq is 0. */
491 vq = dev->virtqueue[msg->payload.addr.index];
494 * Rings addresses should not be interpreted as long as the ring is not
495 * started and enabled
497 memcpy(&vq->ring_addrs, addr, sizeof(*addr));
499 vring_invalidate(dev, vq);
501 if (vq->enabled && (dev->features &
502 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
503 dev = translate_ring_addresses(dev, msg->payload.state.index);
514 * The virtio device sends us the available ring last used index.
517 vhost_user_set_vring_base(struct virtio_net *dev,
520 dev->virtqueue[msg->payload.state.index]->last_used_idx =
521 msg->payload.state.num;
522 dev->virtqueue[msg->payload.state.index]->last_avail_idx =
523 msg->payload.state.num;
529 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
530 uint64_t host_phys_addr, uint64_t size)
532 struct guest_page *page, *last_page;
534 if (dev->nr_guest_pages == dev->max_guest_pages) {
535 dev->max_guest_pages *= 2;
536 dev->guest_pages = realloc(dev->guest_pages,
537 dev->max_guest_pages * sizeof(*page));
540 if (dev->nr_guest_pages > 0) {
541 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
542 /* merge if the two pages are continuous */
543 if (host_phys_addr == last_page->host_phys_addr +
545 last_page->size += size;
550 page = &dev->guest_pages[dev->nr_guest_pages++];
551 page->guest_phys_addr = guest_phys_addr;
552 page->host_phys_addr = host_phys_addr;
557 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
560 uint64_t reg_size = reg->size;
561 uint64_t host_user_addr = reg->host_user_addr;
562 uint64_t guest_phys_addr = reg->guest_phys_addr;
563 uint64_t host_phys_addr;
566 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
567 size = page_size - (guest_phys_addr & (page_size - 1));
568 size = RTE_MIN(size, reg_size);
570 add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size);
571 host_user_addr += size;
572 guest_phys_addr += size;
575 while (reg_size > 0) {
576 size = RTE_MIN(reg_size, page_size);
577 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
579 add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size);
581 host_user_addr += size;
582 guest_phys_addr += size;
587 #ifdef RTE_LIBRTE_VHOST_DEBUG
588 /* TODO: enable it only in debug mode? */
590 dump_guest_pages(struct virtio_net *dev)
593 struct guest_page *page;
595 for (i = 0; i < dev->nr_guest_pages; i++) {
596 page = &dev->guest_pages[i];
598 RTE_LOG(INFO, VHOST_CONFIG,
599 "guest physical page region %u\n"
600 "\t guest_phys_addr: %" PRIx64 "\n"
601 "\t host_phys_addr : %" PRIx64 "\n"
602 "\t size : %" PRIx64 "\n",
604 page->guest_phys_addr,
605 page->host_phys_addr,
610 #define dump_guest_pages(dev)
614 vhost_memory_changed(struct VhostUserMemory *new,
615 struct rte_vhost_memory *old)
619 if (new->nregions != old->nregions)
622 for (i = 0; i < new->nregions; ++i) {
623 VhostUserMemoryRegion *new_r = &new->regions[i];
624 struct rte_vhost_mem_region *old_r = &old->regions[i];
626 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
628 if (new_r->memory_size != old_r->size)
630 if (new_r->userspace_addr != old_r->guest_user_addr)
638 vhost_user_set_mem_table(struct virtio_net *dev, struct VhostUserMsg *pmsg)
640 struct VhostUserMemory memory = pmsg->payload.memory;
641 struct rte_vhost_mem_region *reg;
644 uint64_t mmap_offset;
649 if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
650 RTE_LOG(INFO, VHOST_CONFIG,
651 "(%d) memory regions not changed\n", dev->vid);
653 for (i = 0; i < memory.nregions; i++)
660 free_mem_region(dev);
665 dev->nr_guest_pages = 0;
666 if (!dev->guest_pages) {
667 dev->max_guest_pages = 8;
668 dev->guest_pages = malloc(dev->max_guest_pages *
669 sizeof(struct guest_page));
670 if (dev->guest_pages == NULL) {
671 RTE_LOG(ERR, VHOST_CONFIG,
672 "(%d) failed to allocate memory "
673 "for dev->guest_pages\n",
679 dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
680 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
681 if (dev->mem == NULL) {
682 RTE_LOG(ERR, VHOST_CONFIG,
683 "(%d) failed to allocate memory for dev->mem\n",
687 dev->mem->nregions = memory.nregions;
689 for (i = 0; i < memory.nregions; i++) {
691 reg = &dev->mem->regions[i];
693 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
694 reg->guest_user_addr = memory.regions[i].userspace_addr;
695 reg->size = memory.regions[i].memory_size;
698 mmap_offset = memory.regions[i].mmap_offset;
699 mmap_size = reg->size + mmap_offset;
701 /* mmap() without flag of MAP_ANONYMOUS, should be called
702 * with length argument aligned with hugepagesz at older
703 * longterm version Linux, like 2.6.32 and 3.2.72, or
704 * mmap() will fail with EINVAL.
706 * to avoid failure, make sure in caller to keep length
709 alignment = get_blk_size(fd);
710 if (alignment == (uint64_t)-1) {
711 RTE_LOG(ERR, VHOST_CONFIG,
712 "couldn't get hugepage size through fstat\n");
715 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
717 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
718 MAP_SHARED | MAP_POPULATE, fd, 0);
720 if (mmap_addr == MAP_FAILED) {
721 RTE_LOG(ERR, VHOST_CONFIG,
722 "mmap region %u failed.\n", i);
726 reg->mmap_addr = mmap_addr;
727 reg->mmap_size = mmap_size;
728 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
731 if (dev->dequeue_zero_copy)
732 add_guest_pages(dev, reg, alignment);
734 RTE_LOG(INFO, VHOST_CONFIG,
735 "guest memory region %u, size: 0x%" PRIx64 "\n"
736 "\t guest physical addr: 0x%" PRIx64 "\n"
737 "\t guest virtual addr: 0x%" PRIx64 "\n"
738 "\t host virtual addr: 0x%" PRIx64 "\n"
739 "\t mmap addr : 0x%" PRIx64 "\n"
740 "\t mmap size : 0x%" PRIx64 "\n"
741 "\t mmap align: 0x%" PRIx64 "\n"
742 "\t mmap off : 0x%" PRIx64 "\n",
744 reg->guest_phys_addr,
745 reg->guest_user_addr,
747 (uint64_t)(uintptr_t)mmap_addr,
753 dump_guest_pages(dev);
758 free_mem_region(dev);
765 vq_is_ready(struct vhost_virtqueue *vq)
767 return vq && vq->desc && vq->avail && vq->used &&
768 vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
769 vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
773 virtio_is_ready(struct virtio_net *dev)
775 struct vhost_virtqueue *vq;
778 if (dev->nr_vring == 0)
781 for (i = 0; i < dev->nr_vring; i++) {
782 vq = dev->virtqueue[i];
784 if (!vq_is_ready(vq))
788 RTE_LOG(INFO, VHOST_CONFIG,
789 "virtio is now ready for processing.\n");
794 vhost_user_set_vring_call(struct virtio_net *dev, struct VhostUserMsg *pmsg)
796 struct vhost_vring_file file;
797 struct vhost_virtqueue *vq;
799 file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
800 if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
801 file.fd = VIRTIO_INVALID_EVENTFD;
803 file.fd = pmsg->fds[0];
804 RTE_LOG(INFO, VHOST_CONFIG,
805 "vring call idx:%d file:%d\n", file.index, file.fd);
807 vq = dev->virtqueue[file.index];
811 vq->callfd = file.fd;
815 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *pmsg)
817 struct vhost_vring_file file;
818 struct vhost_virtqueue *vq;
819 struct virtio_net *dev = *pdev;
821 file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
822 if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
823 file.fd = VIRTIO_INVALID_EVENTFD;
825 file.fd = pmsg->fds[0];
826 RTE_LOG(INFO, VHOST_CONFIG,
827 "vring kick idx:%d file:%d\n", file.index, file.fd);
829 /* Interpret ring addresses only when ring is started. */
830 dev = translate_ring_addresses(dev, file.index);
836 vq = dev->virtqueue[file.index];
839 * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
840 * the ring starts already enabled. Otherwise, it is enabled via
841 * the SET_VRING_ENABLE message.
843 if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
848 vq->kickfd = file.fd;
852 free_zmbufs(struct vhost_virtqueue *vq)
854 struct zcopy_mbuf *zmbuf, *next;
856 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
857 zmbuf != NULL; zmbuf = next) {
858 next = TAILQ_NEXT(zmbuf, next);
860 rte_pktmbuf_free(zmbuf->mbuf);
861 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
864 rte_free(vq->zmbufs);
868 * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
871 vhost_user_get_vring_base(struct virtio_net *dev,
874 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
876 /* We have to stop the queue (virtio) if it is running. */
877 if (dev->flags & VIRTIO_DEV_RUNNING) {
878 dev->flags &= ~VIRTIO_DEV_RUNNING;
879 dev->notify_ops->destroy_device(dev->vid);
882 dev->flags &= ~VIRTIO_DEV_READY;
884 /* Here we are safe to get the last used index */
885 msg->payload.state.num = vq->last_used_idx;
887 RTE_LOG(INFO, VHOST_CONFIG,
888 "vring base idx:%d file:%d\n", msg->payload.state.index,
889 msg->payload.state.num);
891 * Based on current qemu vhost-user implementation, this message is
892 * sent and only sent in vhost_vring_stop.
893 * TODO: cleanup the vring, it isn't usable since here.
898 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
900 if (dev->dequeue_zero_copy)
902 rte_free(vq->shadow_used_ring);
903 vq->shadow_used_ring = NULL;
905 rte_free(vq->batch_copy_elems);
906 vq->batch_copy_elems = NULL;
912 * when virtio queues are ready to work, qemu will send us to
913 * enable the virtio queue pair.
916 vhost_user_set_vring_enable(struct virtio_net *dev,
919 int enable = (int)msg->payload.state.num;
921 RTE_LOG(INFO, VHOST_CONFIG,
922 "set queue enable: %d to qp idx: %d\n",
923 enable, msg->payload.state.index);
925 if (dev->notify_ops->vring_state_changed)
926 dev->notify_ops->vring_state_changed(dev->vid,
927 msg->payload.state.index, enable);
929 dev->virtqueue[msg->payload.state.index]->enabled = enable;
935 vhost_user_get_protocol_features(struct virtio_net *dev,
936 struct VhostUserMsg *msg)
938 uint64_t features, protocol_features = VHOST_USER_PROTOCOL_FEATURES;
940 rte_vhost_driver_get_features(dev->ifname, &features);
943 * REPLY_ACK protocol feature is only mandatory for now
944 * for IOMMU feature. If IOMMU is explicitly disabled by the
945 * application, disable also REPLY_ACK feature for older buggy
946 * Qemu versions (from v2.7.0 to v2.9.0).
948 if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
949 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
951 msg->payload.u64 = protocol_features;
952 msg->size = sizeof(msg->payload.u64);
956 vhost_user_set_protocol_features(struct virtio_net *dev,
957 uint64_t protocol_features)
959 if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES)
962 dev->protocol_features = protocol_features;
966 vhost_user_set_log_base(struct virtio_net *dev, struct VhostUserMsg *msg)
968 int fd = msg->fds[0];
973 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
977 if (msg->size != sizeof(VhostUserLog)) {
978 RTE_LOG(ERR, VHOST_CONFIG,
979 "invalid log base msg size: %"PRId32" != %d\n",
980 msg->size, (int)sizeof(VhostUserLog));
984 size = msg->payload.log.mmap_size;
985 off = msg->payload.log.mmap_offset;
986 RTE_LOG(INFO, VHOST_CONFIG,
987 "log mmap size: %"PRId64", offset: %"PRId64"\n",
991 * mmap from 0 to workaround a hugepage mmap bug: mmap will
992 * fail when offset is not page size aligned.
994 addr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
996 if (addr == MAP_FAILED) {
997 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1002 * Free previously mapped log memory on occasionally
1003 * multiple VHOST_USER_SET_LOG_BASE.
1005 if (dev->log_addr) {
1006 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1008 dev->log_addr = (uint64_t)(uintptr_t)addr;
1009 dev->log_base = dev->log_addr + off;
1010 dev->log_size = size;
1016 * An rarp packet is constructed and broadcasted to notify switches about
1017 * the new location of the migrated VM, so that packets from outside will
1018 * not be lost after migration.
1020 * However, we don't actually "send" a rarp packet here, instead, we set
1021 * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1024 vhost_user_send_rarp(struct virtio_net *dev, struct VhostUserMsg *msg)
1026 uint8_t *mac = (uint8_t *)&msg->payload.u64;
1028 RTE_LOG(DEBUG, VHOST_CONFIG,
1029 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1030 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1031 memcpy(dev->mac.addr_bytes, mac, 6);
1034 * Set the flag to inject a RARP broadcast packet at
1035 * rte_vhost_dequeue_burst().
1037 * rte_smp_wmb() is for making sure the mac is copied
1038 * before the flag is set.
1041 rte_atomic16_set(&dev->broadcast_rarp, 1);
1047 vhost_user_net_set_mtu(struct virtio_net *dev, struct VhostUserMsg *msg)
1049 if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1050 msg->payload.u64 > VIRTIO_MAX_MTU) {
1051 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1057 dev->mtu = msg->payload.u64;
1063 vhost_user_set_req_fd(struct virtio_net *dev, struct VhostUserMsg *msg)
1065 int fd = msg->fds[0];
1068 RTE_LOG(ERR, VHOST_CONFIG,
1069 "Invalid file descriptor for slave channel (%d)\n",
1074 dev->slave_req_fd = fd;
1080 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1082 struct vhost_vring_addr *ra;
1083 uint64_t start, end;
1086 end = start + imsg->size;
1088 ra = &vq->ring_addrs;
1089 if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1091 if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1093 if (ra->used_user_addr >= start && ra->used_user_addr < end)
1100 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1101 struct vhost_iotlb_msg *imsg)
1103 uint64_t istart, iend, vstart, vend;
1105 istart = imsg->iova;
1106 iend = istart + imsg->size - 1;
1108 vstart = (uintptr_t)vq->desc;
1109 vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1110 if (vstart <= iend && istart <= vend)
1113 vstart = (uintptr_t)vq->avail;
1114 vend = vstart + sizeof(struct vring_avail);
1115 vend += sizeof(uint16_t) * vq->size - 1;
1116 if (vstart <= iend && istart <= vend)
1119 vstart = (uintptr_t)vq->used;
1120 vend = vstart + sizeof(struct vring_used);
1121 vend += sizeof(struct vring_used_elem) * vq->size - 1;
1122 if (vstart <= iend && istart <= vend)
1129 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg)
1131 struct virtio_net *dev = *pdev;
1132 struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1136 switch (imsg->type) {
1137 case VHOST_IOTLB_UPDATE:
1138 vva = qva_to_vva(dev, imsg->uaddr);
1142 for (i = 0; i < dev->nr_vring; i++) {
1143 struct vhost_virtqueue *vq = dev->virtqueue[i];
1145 vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1146 imsg->size, imsg->perm);
1148 if (is_vring_iotlb_update(vq, imsg))
1149 *pdev = dev = translate_ring_addresses(dev, i);
1152 case VHOST_IOTLB_INVALIDATE:
1153 for (i = 0; i < dev->nr_vring; i++) {
1154 struct vhost_virtqueue *vq = dev->virtqueue[i];
1156 vhost_user_iotlb_cache_remove(vq, imsg->iova,
1159 if (is_vring_iotlb_invalidate(vq, imsg))
1160 vring_invalidate(dev, vq);
1164 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1172 /* return bytes# of read on success or negative val on failure. */
1174 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1178 ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1179 msg->fds, VHOST_MEMORY_MAX_NREGIONS);
1183 if (msg && msg->size) {
1184 if (msg->size > sizeof(msg->payload)) {
1185 RTE_LOG(ERR, VHOST_CONFIG,
1186 "invalid msg size: %d\n", msg->size);
1189 ret = read(sockfd, &msg->payload, msg->size);
1192 if (ret != (int)msg->size) {
1193 RTE_LOG(ERR, VHOST_CONFIG,
1194 "read control message failed\n");
1203 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1208 return send_fd_message(sockfd, (char *)msg,
1209 VHOST_USER_HDR_SIZE + msg->size, NULL, 0);
1213 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1218 msg->flags &= ~VHOST_USER_VERSION_MASK;
1219 msg->flags &= ~VHOST_USER_NEED_REPLY;
1220 msg->flags |= VHOST_USER_VERSION;
1221 msg->flags |= VHOST_USER_REPLY_MASK;
1223 return send_vhost_message(sockfd, msg);
1227 * Allocate a queue pair if it hasn't been allocated yet
1230 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev, VhostUserMsg *msg)
1234 switch (msg->request.master) {
1235 case VHOST_USER_SET_VRING_KICK:
1236 case VHOST_USER_SET_VRING_CALL:
1237 case VHOST_USER_SET_VRING_ERR:
1238 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1240 case VHOST_USER_SET_VRING_NUM:
1241 case VHOST_USER_SET_VRING_BASE:
1242 case VHOST_USER_SET_VRING_ENABLE:
1243 vring_idx = msg->payload.state.index;
1245 case VHOST_USER_SET_VRING_ADDR:
1246 vring_idx = msg->payload.addr.index;
1252 if (vring_idx >= VHOST_MAX_VRING) {
1253 RTE_LOG(ERR, VHOST_CONFIG,
1254 "invalid vring index: %u\n", vring_idx);
1258 if (dev->virtqueue[vring_idx])
1261 return alloc_vring_queue(dev, vring_idx);
1265 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1268 unsigned int vq_num = 0;
1270 while (vq_num < dev->nr_vring) {
1271 struct vhost_virtqueue *vq = dev->virtqueue[i];
1274 rte_spinlock_lock(&vq->access_lock);
1282 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1285 unsigned int vq_num = 0;
1287 while (vq_num < dev->nr_vring) {
1288 struct vhost_virtqueue *vq = dev->virtqueue[i];
1291 rte_spinlock_unlock(&vq->access_lock);
1299 vhost_user_msg_handler(int vid, int fd)
1301 struct virtio_net *dev;
1302 struct VhostUserMsg msg;
1304 int unlock_required = 0;
1306 dev = get_device(vid);
1310 if (!dev->notify_ops) {
1311 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1312 if (!dev->notify_ops) {
1313 RTE_LOG(ERR, VHOST_CONFIG,
1314 "failed to get callback ops for driver %s\n",
1320 ret = read_vhost_message(fd, &msg);
1321 if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1323 RTE_LOG(ERR, VHOST_CONFIG,
1324 "vhost read message failed\n");
1326 RTE_LOG(INFO, VHOST_CONFIG,
1327 "vhost peer closed\n");
1329 RTE_LOG(ERR, VHOST_CONFIG,
1330 "vhost read incorrect message\n");
1336 if (msg.request.master != VHOST_USER_IOTLB_MSG)
1337 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1338 vhost_message_str[msg.request.master]);
1340 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1341 vhost_message_str[msg.request.master]);
1343 ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1345 RTE_LOG(ERR, VHOST_CONFIG,
1346 "failed to alloc queue\n");
1351 * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1352 * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1353 * and device is destroyed. destroy_device waits for queues to be
1354 * inactive, so it is safe. Otherwise taking the access_lock
1355 * would cause a dead lock.
1357 switch (msg.request.master) {
1358 case VHOST_USER_SET_FEATURES:
1359 case VHOST_USER_SET_PROTOCOL_FEATURES:
1360 case VHOST_USER_SET_OWNER:
1361 case VHOST_USER_SET_MEM_TABLE:
1362 case VHOST_USER_SET_LOG_BASE:
1363 case VHOST_USER_SET_LOG_FD:
1364 case VHOST_USER_SET_VRING_NUM:
1365 case VHOST_USER_SET_VRING_ADDR:
1366 case VHOST_USER_SET_VRING_BASE:
1367 case VHOST_USER_SET_VRING_KICK:
1368 case VHOST_USER_SET_VRING_CALL:
1369 case VHOST_USER_SET_VRING_ERR:
1370 case VHOST_USER_SET_VRING_ENABLE:
1371 case VHOST_USER_SEND_RARP:
1372 case VHOST_USER_NET_SET_MTU:
1373 case VHOST_USER_SET_SLAVE_REQ_FD:
1374 vhost_user_lock_all_queue_pairs(dev);
1375 unlock_required = 1;
1382 switch (msg.request.master) {
1383 case VHOST_USER_GET_FEATURES:
1384 msg.payload.u64 = vhost_user_get_features(dev);
1385 msg.size = sizeof(msg.payload.u64);
1386 send_vhost_reply(fd, &msg);
1388 case VHOST_USER_SET_FEATURES:
1389 ret = vhost_user_set_features(dev, msg.payload.u64);
1394 case VHOST_USER_GET_PROTOCOL_FEATURES:
1395 vhost_user_get_protocol_features(dev, &msg);
1396 send_vhost_reply(fd, &msg);
1398 case VHOST_USER_SET_PROTOCOL_FEATURES:
1399 vhost_user_set_protocol_features(dev, msg.payload.u64);
1402 case VHOST_USER_SET_OWNER:
1403 vhost_user_set_owner();
1405 case VHOST_USER_RESET_OWNER:
1406 vhost_user_reset_owner(dev);
1409 case VHOST_USER_SET_MEM_TABLE:
1410 ret = vhost_user_set_mem_table(dev, &msg);
1413 case VHOST_USER_SET_LOG_BASE:
1414 vhost_user_set_log_base(dev, &msg);
1416 /* it needs a reply */
1417 msg.size = sizeof(msg.payload.u64);
1418 send_vhost_reply(fd, &msg);
1420 case VHOST_USER_SET_LOG_FD:
1422 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1425 case VHOST_USER_SET_VRING_NUM:
1426 vhost_user_set_vring_num(dev, &msg);
1428 case VHOST_USER_SET_VRING_ADDR:
1429 vhost_user_set_vring_addr(&dev, &msg);
1431 case VHOST_USER_SET_VRING_BASE:
1432 vhost_user_set_vring_base(dev, &msg);
1435 case VHOST_USER_GET_VRING_BASE:
1436 vhost_user_get_vring_base(dev, &msg);
1437 msg.size = sizeof(msg.payload.state);
1438 send_vhost_reply(fd, &msg);
1441 case VHOST_USER_SET_VRING_KICK:
1442 vhost_user_set_vring_kick(&dev, &msg);
1444 case VHOST_USER_SET_VRING_CALL:
1445 vhost_user_set_vring_call(dev, &msg);
1448 case VHOST_USER_SET_VRING_ERR:
1449 if (!(msg.payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1451 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1454 case VHOST_USER_GET_QUEUE_NUM:
1455 msg.payload.u64 = VHOST_MAX_QUEUE_PAIRS;
1456 msg.size = sizeof(msg.payload.u64);
1457 send_vhost_reply(fd, &msg);
1460 case VHOST_USER_SET_VRING_ENABLE:
1461 vhost_user_set_vring_enable(dev, &msg);
1463 case VHOST_USER_SEND_RARP:
1464 vhost_user_send_rarp(dev, &msg);
1467 case VHOST_USER_NET_SET_MTU:
1468 ret = vhost_user_net_set_mtu(dev, &msg);
1471 case VHOST_USER_SET_SLAVE_REQ_FD:
1472 ret = vhost_user_set_req_fd(dev, &msg);
1475 case VHOST_USER_IOTLB_MSG:
1476 ret = vhost_user_iotlb_msg(&dev, &msg);
1485 if (unlock_required)
1486 vhost_user_unlock_all_queue_pairs(dev);
1488 if (msg.flags & VHOST_USER_NEED_REPLY) {
1489 msg.payload.u64 = !!ret;
1490 msg.size = sizeof(msg.payload.u64);
1491 send_vhost_reply(fd, &msg);
1494 if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1495 dev->flags |= VIRTIO_DEV_READY;
1497 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1498 if (dev->dequeue_zero_copy) {
1499 RTE_LOG(INFO, VHOST_CONFIG,
1500 "dequeue zero copy is enabled\n");
1503 if (dev->notify_ops->new_device(dev->vid) == 0)
1504 dev->flags |= VIRTIO_DEV_RUNNING;
1512 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
1515 struct VhostUserMsg msg = {
1516 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
1517 .flags = VHOST_USER_VERSION,
1518 .size = sizeof(msg.payload.iotlb),
1522 .type = VHOST_IOTLB_MISS,
1526 ret = send_vhost_message(dev->slave_req_fd, &msg);
1528 RTE_LOG(ERR, VHOST_CONFIG,
1529 "Failed to send IOTLB miss message (%d)\n",