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->features & (1ULL << VIRTIO_NET_F_MQ))) {
192 * Remove all but first queue pair if MQ hasn't been
193 * negotiated. This is safe because the device is not
194 * running at this stage.
196 while (dev->nr_vring > 2) {
197 struct vhost_virtqueue *vq;
199 vq = dev->virtqueue[--dev->nr_vring];
203 dev->virtqueue[dev->nr_vring] = NULL;
213 * The virtio device sends us the size of the descriptor ring.
216 vhost_user_set_vring_num(struct virtio_net *dev,
219 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
221 vq->size = msg->payload.state.num;
223 if (dev->dequeue_zero_copy) {
225 vq->last_zmbuf_idx = 0;
226 vq->zmbuf_size = vq->size;
227 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
228 sizeof(struct zcopy_mbuf), 0);
229 if (vq->zmbufs == NULL) {
230 RTE_LOG(WARNING, VHOST_CONFIG,
231 "failed to allocate mem for zero copy; "
232 "zero copy is force disabled\n");
233 dev->dequeue_zero_copy = 0;
235 TAILQ_INIT(&vq->zmbuf_list);
238 vq->shadow_used_ring = rte_malloc(NULL,
239 vq->size * sizeof(struct vring_used_elem),
240 RTE_CACHE_LINE_SIZE);
241 if (!vq->shadow_used_ring) {
242 RTE_LOG(ERR, VHOST_CONFIG,
243 "failed to allocate memory for shadow used ring.\n");
247 vq->batch_copy_elems = rte_malloc(NULL,
248 vq->size * sizeof(struct batch_copy_elem),
249 RTE_CACHE_LINE_SIZE);
250 if (!vq->batch_copy_elems) {
251 RTE_LOG(ERR, VHOST_CONFIG,
252 "failed to allocate memory for batching copy.\n");
260 * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
261 * same numa node as the memory of vring descriptor.
263 #ifdef RTE_LIBRTE_VHOST_NUMA
264 static struct virtio_net*
265 numa_realloc(struct virtio_net *dev, int index)
267 int oldnode, newnode;
268 struct virtio_net *old_dev;
269 struct vhost_virtqueue *old_vq, *vq;
270 struct zcopy_mbuf *new_zmbuf;
271 struct vring_used_elem *new_shadow_used_ring;
272 struct batch_copy_elem *new_batch_copy_elems;
276 vq = old_vq = dev->virtqueue[index];
278 ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
279 MPOL_F_NODE | MPOL_F_ADDR);
281 /* check if we need to reallocate vq */
282 ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
283 MPOL_F_NODE | MPOL_F_ADDR);
285 RTE_LOG(ERR, VHOST_CONFIG,
286 "Unable to get vq numa information.\n");
289 if (oldnode != newnode) {
290 RTE_LOG(INFO, VHOST_CONFIG,
291 "reallocate vq from %d to %d node\n", oldnode, newnode);
292 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
296 memcpy(vq, old_vq, sizeof(*vq));
297 TAILQ_INIT(&vq->zmbuf_list);
299 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
300 sizeof(struct zcopy_mbuf), 0, newnode);
302 rte_free(vq->zmbufs);
303 vq->zmbufs = new_zmbuf;
306 new_shadow_used_ring = rte_malloc_socket(NULL,
307 vq->size * sizeof(struct vring_used_elem),
310 if (new_shadow_used_ring) {
311 rte_free(vq->shadow_used_ring);
312 vq->shadow_used_ring = new_shadow_used_ring;
315 new_batch_copy_elems = rte_malloc_socket(NULL,
316 vq->size * sizeof(struct batch_copy_elem),
319 if (new_batch_copy_elems) {
320 rte_free(vq->batch_copy_elems);
321 vq->batch_copy_elems = new_batch_copy_elems;
327 /* check if we need to reallocate dev */
328 ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
329 MPOL_F_NODE | MPOL_F_ADDR);
331 RTE_LOG(ERR, VHOST_CONFIG,
332 "Unable to get dev numa information.\n");
335 if (oldnode != newnode) {
336 RTE_LOG(INFO, VHOST_CONFIG,
337 "reallocate dev from %d to %d node\n",
339 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
345 memcpy(dev, old_dev, sizeof(*dev));
350 dev->virtqueue[index] = vq;
351 vhost_devices[dev->vid] = dev;
354 vhost_user_iotlb_init(dev, index);
359 static struct virtio_net*
360 numa_realloc(struct virtio_net *dev, int index __rte_unused)
366 /* Converts QEMU virtual address to Vhost virtual address. */
368 qva_to_vva(struct virtio_net *dev, uint64_t qva)
370 struct rte_vhost_mem_region *reg;
373 /* Find the region where the address lives. */
374 for (i = 0; i < dev->mem->nregions; i++) {
375 reg = &dev->mem->regions[i];
377 if (qva >= reg->guest_user_addr &&
378 qva < reg->guest_user_addr + reg->size) {
379 return qva - reg->guest_user_addr +
389 * Converts ring address to Vhost virtual address.
390 * If IOMMU is enabled, the ring address is a guest IO virtual address,
391 * else it is a QEMU virtual address.
394 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
395 uint64_t ra, uint64_t size)
397 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
400 vva = vhost_user_iotlb_cache_find(vq, ra,
401 &size, VHOST_ACCESS_RW);
403 vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
408 return qva_to_vva(dev, ra);
411 static struct virtio_net *
412 translate_ring_addresses(struct virtio_net *dev, int vq_index)
414 struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
415 struct vhost_vring_addr *addr = &vq->ring_addrs;
417 /* The addresses are converted from QEMU virtual to Vhost virtual. */
418 if (vq->desc && vq->avail && vq->used)
421 vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
422 vq, addr->desc_user_addr, sizeof(struct vring_desc));
424 RTE_LOG(DEBUG, VHOST_CONFIG,
425 "(%d) failed to find desc ring address.\n",
430 dev = numa_realloc(dev, vq_index);
431 vq = dev->virtqueue[vq_index];
432 addr = &vq->ring_addrs;
434 vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
435 vq, addr->avail_user_addr, sizeof(struct vring_avail));
436 if (vq->avail == 0) {
437 RTE_LOG(DEBUG, VHOST_CONFIG,
438 "(%d) failed to find avail ring address.\n",
443 vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
444 vq, addr->used_user_addr, sizeof(struct vring_used));
446 RTE_LOG(DEBUG, VHOST_CONFIG,
447 "(%d) failed to find used ring address.\n",
452 if (vq->last_used_idx != vq->used->idx) {
453 RTE_LOG(WARNING, VHOST_CONFIG,
454 "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
455 "some packets maybe resent for Tx and dropped for Rx\n",
456 vq->last_used_idx, vq->used->idx);
457 vq->last_used_idx = vq->used->idx;
458 vq->last_avail_idx = vq->used->idx;
461 vq->log_guest_addr = addr->log_guest_addr;
463 LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
465 LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
466 dev->vid, vq->avail);
467 LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
469 LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
470 dev->vid, vq->log_guest_addr);
476 * The virtio device sends us the desc, used and avail ring addresses.
477 * This function then converts these to our address space.
480 vhost_user_set_vring_addr(struct virtio_net **pdev, VhostUserMsg *msg)
482 struct vhost_virtqueue *vq;
483 struct vhost_vring_addr *addr = &msg->payload.addr;
484 struct virtio_net *dev = *pdev;
486 if (dev->mem == NULL)
489 /* addr->index refers to the queue index. The txq 1, rxq is 0. */
490 vq = dev->virtqueue[msg->payload.addr.index];
493 * Rings addresses should not be interpreted as long as the ring is not
494 * started and enabled
496 memcpy(&vq->ring_addrs, addr, sizeof(*addr));
498 vring_invalidate(dev, vq);
500 if (vq->enabled && (dev->features &
501 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
502 dev = translate_ring_addresses(dev, msg->payload.state.index);
513 * The virtio device sends us the available ring last used index.
516 vhost_user_set_vring_base(struct virtio_net *dev,
519 dev->virtqueue[msg->payload.state.index]->last_used_idx =
520 msg->payload.state.num;
521 dev->virtqueue[msg->payload.state.index]->last_avail_idx =
522 msg->payload.state.num;
528 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
529 uint64_t host_phys_addr, uint64_t size)
531 struct guest_page *page, *last_page;
533 if (dev->nr_guest_pages == dev->max_guest_pages) {
534 dev->max_guest_pages *= 2;
535 dev->guest_pages = realloc(dev->guest_pages,
536 dev->max_guest_pages * sizeof(*page));
539 if (dev->nr_guest_pages > 0) {
540 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
541 /* merge if the two pages are continuous */
542 if (host_phys_addr == last_page->host_phys_addr +
544 last_page->size += size;
549 page = &dev->guest_pages[dev->nr_guest_pages++];
550 page->guest_phys_addr = guest_phys_addr;
551 page->host_phys_addr = host_phys_addr;
556 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
559 uint64_t reg_size = reg->size;
560 uint64_t host_user_addr = reg->host_user_addr;
561 uint64_t guest_phys_addr = reg->guest_phys_addr;
562 uint64_t host_phys_addr;
565 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
566 size = page_size - (guest_phys_addr & (page_size - 1));
567 size = RTE_MIN(size, reg_size);
569 add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size);
570 host_user_addr += size;
571 guest_phys_addr += size;
574 while (reg_size > 0) {
575 size = RTE_MIN(reg_size, page_size);
576 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
578 add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size);
580 host_user_addr += size;
581 guest_phys_addr += size;
586 #ifdef RTE_LIBRTE_VHOST_DEBUG
587 /* TODO: enable it only in debug mode? */
589 dump_guest_pages(struct virtio_net *dev)
592 struct guest_page *page;
594 for (i = 0; i < dev->nr_guest_pages; i++) {
595 page = &dev->guest_pages[i];
597 RTE_LOG(INFO, VHOST_CONFIG,
598 "guest physical page region %u\n"
599 "\t guest_phys_addr: %" PRIx64 "\n"
600 "\t host_phys_addr : %" PRIx64 "\n"
601 "\t size : %" PRIx64 "\n",
603 page->guest_phys_addr,
604 page->host_phys_addr,
609 #define dump_guest_pages(dev)
613 vhost_memory_changed(struct VhostUserMemory *new,
614 struct rte_vhost_memory *old)
618 if (new->nregions != old->nregions)
621 for (i = 0; i < new->nregions; ++i) {
622 VhostUserMemoryRegion *new_r = &new->regions[i];
623 struct rte_vhost_mem_region *old_r = &old->regions[i];
625 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
627 if (new_r->memory_size != old_r->size)
629 if (new_r->userspace_addr != old_r->guest_user_addr)
637 vhost_user_set_mem_table(struct virtio_net *dev, struct VhostUserMsg *pmsg)
639 struct VhostUserMemory memory = pmsg->payload.memory;
640 struct rte_vhost_mem_region *reg;
643 uint64_t mmap_offset;
648 if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
649 RTE_LOG(INFO, VHOST_CONFIG,
650 "(%d) memory regions not changed\n", dev->vid);
652 for (i = 0; i < memory.nregions; i++)
659 free_mem_region(dev);
664 dev->nr_guest_pages = 0;
665 if (!dev->guest_pages) {
666 dev->max_guest_pages = 8;
667 dev->guest_pages = malloc(dev->max_guest_pages *
668 sizeof(struct guest_page));
669 if (dev->guest_pages == NULL) {
670 RTE_LOG(ERR, VHOST_CONFIG,
671 "(%d) failed to allocate memory "
672 "for dev->guest_pages\n",
678 dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
679 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
680 if (dev->mem == NULL) {
681 RTE_LOG(ERR, VHOST_CONFIG,
682 "(%d) failed to allocate memory for dev->mem\n",
686 dev->mem->nregions = memory.nregions;
688 for (i = 0; i < memory.nregions; i++) {
690 reg = &dev->mem->regions[i];
692 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
693 reg->guest_user_addr = memory.regions[i].userspace_addr;
694 reg->size = memory.regions[i].memory_size;
697 mmap_offset = memory.regions[i].mmap_offset;
698 mmap_size = reg->size + mmap_offset;
700 /* mmap() without flag of MAP_ANONYMOUS, should be called
701 * with length argument aligned with hugepagesz at older
702 * longterm version Linux, like 2.6.32 and 3.2.72, or
703 * mmap() will fail with EINVAL.
705 * to avoid failure, make sure in caller to keep length
708 alignment = get_blk_size(fd);
709 if (alignment == (uint64_t)-1) {
710 RTE_LOG(ERR, VHOST_CONFIG,
711 "couldn't get hugepage size through fstat\n");
714 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
716 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
717 MAP_SHARED | MAP_POPULATE, fd, 0);
719 if (mmap_addr == MAP_FAILED) {
720 RTE_LOG(ERR, VHOST_CONFIG,
721 "mmap region %u failed.\n", i);
725 reg->mmap_addr = mmap_addr;
726 reg->mmap_size = mmap_size;
727 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
730 if (dev->dequeue_zero_copy)
731 add_guest_pages(dev, reg, alignment);
733 RTE_LOG(INFO, VHOST_CONFIG,
734 "guest memory region %u, size: 0x%" PRIx64 "\n"
735 "\t guest physical addr: 0x%" PRIx64 "\n"
736 "\t guest virtual addr: 0x%" PRIx64 "\n"
737 "\t host virtual addr: 0x%" PRIx64 "\n"
738 "\t mmap addr : 0x%" PRIx64 "\n"
739 "\t mmap size : 0x%" PRIx64 "\n"
740 "\t mmap align: 0x%" PRIx64 "\n"
741 "\t mmap off : 0x%" PRIx64 "\n",
743 reg->guest_phys_addr,
744 reg->guest_user_addr,
746 (uint64_t)(uintptr_t)mmap_addr,
752 dump_guest_pages(dev);
757 free_mem_region(dev);
764 vq_is_ready(struct vhost_virtqueue *vq)
766 return vq && vq->desc && vq->avail && vq->used &&
767 vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
768 vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
772 virtio_is_ready(struct virtio_net *dev)
774 struct vhost_virtqueue *vq;
777 if (dev->nr_vring == 0)
780 for (i = 0; i < dev->nr_vring; i++) {
781 vq = dev->virtqueue[i];
783 if (!vq_is_ready(vq))
787 RTE_LOG(INFO, VHOST_CONFIG,
788 "virtio is now ready for processing.\n");
793 vhost_user_set_vring_call(struct virtio_net *dev, struct VhostUserMsg *pmsg)
795 struct vhost_vring_file file;
796 struct vhost_virtqueue *vq;
798 file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
799 if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
800 file.fd = VIRTIO_INVALID_EVENTFD;
802 file.fd = pmsg->fds[0];
803 RTE_LOG(INFO, VHOST_CONFIG,
804 "vring call idx:%d file:%d\n", file.index, file.fd);
806 vq = dev->virtqueue[file.index];
810 vq->callfd = file.fd;
814 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *pmsg)
816 struct vhost_vring_file file;
817 struct vhost_virtqueue *vq;
818 struct virtio_net *dev = *pdev;
820 file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
821 if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
822 file.fd = VIRTIO_INVALID_EVENTFD;
824 file.fd = pmsg->fds[0];
825 RTE_LOG(INFO, VHOST_CONFIG,
826 "vring kick idx:%d file:%d\n", file.index, file.fd);
828 /* Interpret ring addresses only when ring is started. */
829 dev = translate_ring_addresses(dev, file.index);
835 vq = dev->virtqueue[file.index];
838 * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
839 * the ring starts already enabled. Otherwise, it is enabled via
840 * the SET_VRING_ENABLE message.
842 if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
847 vq->kickfd = file.fd;
851 free_zmbufs(struct vhost_virtqueue *vq)
853 struct zcopy_mbuf *zmbuf, *next;
855 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
856 zmbuf != NULL; zmbuf = next) {
857 next = TAILQ_NEXT(zmbuf, next);
859 rte_pktmbuf_free(zmbuf->mbuf);
860 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
863 rte_free(vq->zmbufs);
867 * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
870 vhost_user_get_vring_base(struct virtio_net *dev,
873 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
875 /* We have to stop the queue (virtio) if it is running. */
876 if (dev->flags & VIRTIO_DEV_RUNNING) {
877 dev->flags &= ~VIRTIO_DEV_RUNNING;
878 dev->notify_ops->destroy_device(dev->vid);
881 dev->flags &= ~VIRTIO_DEV_READY;
883 /* Here we are safe to get the last used index */
884 msg->payload.state.num = vq->last_used_idx;
886 RTE_LOG(INFO, VHOST_CONFIG,
887 "vring base idx:%d file:%d\n", msg->payload.state.index,
888 msg->payload.state.num);
890 * Based on current qemu vhost-user implementation, this message is
891 * sent and only sent in vhost_vring_stop.
892 * TODO: cleanup the vring, it isn't usable since here.
897 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
899 if (dev->dequeue_zero_copy)
901 rte_free(vq->shadow_used_ring);
902 vq->shadow_used_ring = NULL;
904 rte_free(vq->batch_copy_elems);
905 vq->batch_copy_elems = NULL;
911 * when virtio queues are ready to work, qemu will send us to
912 * enable the virtio queue pair.
915 vhost_user_set_vring_enable(struct virtio_net *dev,
918 int enable = (int)msg->payload.state.num;
920 RTE_LOG(INFO, VHOST_CONFIG,
921 "set queue enable: %d to qp idx: %d\n",
922 enable, msg->payload.state.index);
924 if (dev->notify_ops->vring_state_changed)
925 dev->notify_ops->vring_state_changed(dev->vid,
926 msg->payload.state.index, enable);
928 dev->virtqueue[msg->payload.state.index]->enabled = enable;
934 vhost_user_get_protocol_features(struct virtio_net *dev,
935 struct VhostUserMsg *msg)
937 uint64_t features, protocol_features = VHOST_USER_PROTOCOL_FEATURES;
939 rte_vhost_driver_get_features(dev->ifname, &features);
942 * REPLY_ACK protocol feature is only mandatory for now
943 * for IOMMU feature. If IOMMU is explicitly disabled by the
944 * application, disable also REPLY_ACK feature for older buggy
945 * Qemu versions (from v2.7.0 to v2.9.0).
947 if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
948 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
950 msg->payload.u64 = protocol_features;
951 msg->size = sizeof(msg->payload.u64);
955 vhost_user_set_protocol_features(struct virtio_net *dev,
956 uint64_t protocol_features)
958 if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES)
961 dev->protocol_features = protocol_features;
965 vhost_user_set_log_base(struct virtio_net *dev, struct VhostUserMsg *msg)
967 int fd = msg->fds[0];
972 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
976 if (msg->size != sizeof(VhostUserLog)) {
977 RTE_LOG(ERR, VHOST_CONFIG,
978 "invalid log base msg size: %"PRId32" != %d\n",
979 msg->size, (int)sizeof(VhostUserLog));
983 size = msg->payload.log.mmap_size;
984 off = msg->payload.log.mmap_offset;
985 RTE_LOG(INFO, VHOST_CONFIG,
986 "log mmap size: %"PRId64", offset: %"PRId64"\n",
990 * mmap from 0 to workaround a hugepage mmap bug: mmap will
991 * fail when offset is not page size aligned.
993 addr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
995 if (addr == MAP_FAILED) {
996 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1001 * Free previously mapped log memory on occasionally
1002 * multiple VHOST_USER_SET_LOG_BASE.
1004 if (dev->log_addr) {
1005 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1007 dev->log_addr = (uint64_t)(uintptr_t)addr;
1008 dev->log_base = dev->log_addr + off;
1009 dev->log_size = size;
1015 * An rarp packet is constructed and broadcasted to notify switches about
1016 * the new location of the migrated VM, so that packets from outside will
1017 * not be lost after migration.
1019 * However, we don't actually "send" a rarp packet here, instead, we set
1020 * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1023 vhost_user_send_rarp(struct virtio_net *dev, struct VhostUserMsg *msg)
1025 uint8_t *mac = (uint8_t *)&msg->payload.u64;
1027 RTE_LOG(DEBUG, VHOST_CONFIG,
1028 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1029 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1030 memcpy(dev->mac.addr_bytes, mac, 6);
1033 * Set the flag to inject a RARP broadcast packet at
1034 * rte_vhost_dequeue_burst().
1036 * rte_smp_wmb() is for making sure the mac is copied
1037 * before the flag is set.
1040 rte_atomic16_set(&dev->broadcast_rarp, 1);
1046 vhost_user_net_set_mtu(struct virtio_net *dev, struct VhostUserMsg *msg)
1048 if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1049 msg->payload.u64 > VIRTIO_MAX_MTU) {
1050 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1056 dev->mtu = msg->payload.u64;
1062 vhost_user_set_req_fd(struct virtio_net *dev, struct VhostUserMsg *msg)
1064 int fd = msg->fds[0];
1067 RTE_LOG(ERR, VHOST_CONFIG,
1068 "Invalid file descriptor for slave channel (%d)\n",
1073 dev->slave_req_fd = fd;
1079 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1081 struct vhost_vring_addr *ra;
1082 uint64_t start, end;
1085 end = start + imsg->size;
1087 ra = &vq->ring_addrs;
1088 if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1090 if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1092 if (ra->used_user_addr >= start && ra->used_user_addr < end)
1099 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1100 struct vhost_iotlb_msg *imsg)
1102 uint64_t istart, iend, vstart, vend;
1104 istart = imsg->iova;
1105 iend = istart + imsg->size - 1;
1107 vstart = (uintptr_t)vq->desc;
1108 vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1109 if (vstart <= iend && istart <= vend)
1112 vstart = (uintptr_t)vq->avail;
1113 vend = vstart + sizeof(struct vring_avail);
1114 vend += sizeof(uint16_t) * vq->size - 1;
1115 if (vstart <= iend && istart <= vend)
1118 vstart = (uintptr_t)vq->used;
1119 vend = vstart + sizeof(struct vring_used);
1120 vend += sizeof(struct vring_used_elem) * vq->size - 1;
1121 if (vstart <= iend && istart <= vend)
1128 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg)
1130 struct virtio_net *dev = *pdev;
1131 struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1135 switch (imsg->type) {
1136 case VHOST_IOTLB_UPDATE:
1137 vva = qva_to_vva(dev, imsg->uaddr);
1141 for (i = 0; i < dev->nr_vring; i++) {
1142 struct vhost_virtqueue *vq = dev->virtqueue[i];
1144 vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1145 imsg->size, imsg->perm);
1147 if (is_vring_iotlb_update(vq, imsg))
1148 *pdev = dev = translate_ring_addresses(dev, i);
1151 case VHOST_IOTLB_INVALIDATE:
1152 for (i = 0; i < dev->nr_vring; i++) {
1153 struct vhost_virtqueue *vq = dev->virtqueue[i];
1155 vhost_user_iotlb_cache_remove(vq, imsg->iova,
1158 if (is_vring_iotlb_invalidate(vq, imsg))
1159 vring_invalidate(dev, vq);
1163 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1171 /* return bytes# of read on success or negative val on failure. */
1173 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1177 ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1178 msg->fds, VHOST_MEMORY_MAX_NREGIONS);
1182 if (msg && msg->size) {
1183 if (msg->size > sizeof(msg->payload)) {
1184 RTE_LOG(ERR, VHOST_CONFIG,
1185 "invalid msg size: %d\n", msg->size);
1188 ret = read(sockfd, &msg->payload, msg->size);
1191 if (ret != (int)msg->size) {
1192 RTE_LOG(ERR, VHOST_CONFIG,
1193 "read control message failed\n");
1202 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1207 return send_fd_message(sockfd, (char *)msg,
1208 VHOST_USER_HDR_SIZE + msg->size, NULL, 0);
1212 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1217 msg->flags &= ~VHOST_USER_VERSION_MASK;
1218 msg->flags &= ~VHOST_USER_NEED_REPLY;
1219 msg->flags |= VHOST_USER_VERSION;
1220 msg->flags |= VHOST_USER_REPLY_MASK;
1222 return send_vhost_message(sockfd, msg);
1226 * Allocate a queue pair if it hasn't been allocated yet
1229 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev, VhostUserMsg *msg)
1233 switch (msg->request.master) {
1234 case VHOST_USER_SET_VRING_KICK:
1235 case VHOST_USER_SET_VRING_CALL:
1236 case VHOST_USER_SET_VRING_ERR:
1237 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1239 case VHOST_USER_SET_VRING_NUM:
1240 case VHOST_USER_SET_VRING_BASE:
1241 case VHOST_USER_SET_VRING_ENABLE:
1242 vring_idx = msg->payload.state.index;
1244 case VHOST_USER_SET_VRING_ADDR:
1245 vring_idx = msg->payload.addr.index;
1251 if (vring_idx >= VHOST_MAX_VRING) {
1252 RTE_LOG(ERR, VHOST_CONFIG,
1253 "invalid vring index: %u\n", vring_idx);
1257 if (dev->virtqueue[vring_idx])
1260 return alloc_vring_queue(dev, vring_idx);
1264 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1267 unsigned int vq_num = 0;
1269 while (vq_num < dev->nr_vring) {
1270 struct vhost_virtqueue *vq = dev->virtqueue[i];
1273 rte_spinlock_lock(&vq->access_lock);
1281 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1284 unsigned int vq_num = 0;
1286 while (vq_num < dev->nr_vring) {
1287 struct vhost_virtqueue *vq = dev->virtqueue[i];
1290 rte_spinlock_unlock(&vq->access_lock);
1298 vhost_user_msg_handler(int vid, int fd)
1300 struct virtio_net *dev;
1301 struct VhostUserMsg msg;
1303 int unlock_required = 0;
1305 dev = get_device(vid);
1309 if (!dev->notify_ops) {
1310 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1311 if (!dev->notify_ops) {
1312 RTE_LOG(ERR, VHOST_CONFIG,
1313 "failed to get callback ops for driver %s\n",
1319 ret = read_vhost_message(fd, &msg);
1320 if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1322 RTE_LOG(ERR, VHOST_CONFIG,
1323 "vhost read message failed\n");
1325 RTE_LOG(INFO, VHOST_CONFIG,
1326 "vhost peer closed\n");
1328 RTE_LOG(ERR, VHOST_CONFIG,
1329 "vhost read incorrect message\n");
1335 if (msg.request.master != VHOST_USER_IOTLB_MSG)
1336 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1337 vhost_message_str[msg.request.master]);
1339 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1340 vhost_message_str[msg.request.master]);
1342 ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1344 RTE_LOG(ERR, VHOST_CONFIG,
1345 "failed to alloc queue\n");
1350 * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE,
1351 * since it is sent when virtio stops and device is destroyed.
1352 * destroy_device waits for queues to be inactive, so it is safe.
1353 * Otherwise taking the access_lock would cause a dead lock.
1355 switch (msg.request.master) {
1356 case VHOST_USER_SET_FEATURES:
1357 case VHOST_USER_SET_PROTOCOL_FEATURES:
1358 case VHOST_USER_SET_OWNER:
1359 case VHOST_USER_RESET_OWNER:
1360 case VHOST_USER_SET_MEM_TABLE:
1361 case VHOST_USER_SET_LOG_BASE:
1362 case VHOST_USER_SET_LOG_FD:
1363 case VHOST_USER_SET_VRING_NUM:
1364 case VHOST_USER_SET_VRING_ADDR:
1365 case VHOST_USER_SET_VRING_BASE:
1366 case VHOST_USER_SET_VRING_KICK:
1367 case VHOST_USER_SET_VRING_CALL:
1368 case VHOST_USER_SET_VRING_ERR:
1369 case VHOST_USER_SET_VRING_ENABLE:
1370 case VHOST_USER_SEND_RARP:
1371 case VHOST_USER_NET_SET_MTU:
1372 case VHOST_USER_SET_SLAVE_REQ_FD:
1373 vhost_user_lock_all_queue_pairs(dev);
1374 unlock_required = 1;
1381 switch (msg.request.master) {
1382 case VHOST_USER_GET_FEATURES:
1383 msg.payload.u64 = vhost_user_get_features(dev);
1384 msg.size = sizeof(msg.payload.u64);
1385 send_vhost_reply(fd, &msg);
1387 case VHOST_USER_SET_FEATURES:
1388 ret = vhost_user_set_features(dev, msg.payload.u64);
1393 case VHOST_USER_GET_PROTOCOL_FEATURES:
1394 vhost_user_get_protocol_features(dev, &msg);
1395 send_vhost_reply(fd, &msg);
1397 case VHOST_USER_SET_PROTOCOL_FEATURES:
1398 vhost_user_set_protocol_features(dev, msg.payload.u64);
1401 case VHOST_USER_SET_OWNER:
1402 vhost_user_set_owner();
1404 case VHOST_USER_RESET_OWNER:
1405 vhost_user_reset_owner(dev);
1408 case VHOST_USER_SET_MEM_TABLE:
1409 ret = vhost_user_set_mem_table(dev, &msg);
1412 case VHOST_USER_SET_LOG_BASE:
1413 vhost_user_set_log_base(dev, &msg);
1415 /* it needs a reply */
1416 msg.size = sizeof(msg.payload.u64);
1417 send_vhost_reply(fd, &msg);
1419 case VHOST_USER_SET_LOG_FD:
1421 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1424 case VHOST_USER_SET_VRING_NUM:
1425 vhost_user_set_vring_num(dev, &msg);
1427 case VHOST_USER_SET_VRING_ADDR:
1428 vhost_user_set_vring_addr(&dev, &msg);
1430 case VHOST_USER_SET_VRING_BASE:
1431 vhost_user_set_vring_base(dev, &msg);
1434 case VHOST_USER_GET_VRING_BASE:
1435 vhost_user_get_vring_base(dev, &msg);
1436 msg.size = sizeof(msg.payload.state);
1437 send_vhost_reply(fd, &msg);
1440 case VHOST_USER_SET_VRING_KICK:
1441 vhost_user_set_vring_kick(&dev, &msg);
1443 case VHOST_USER_SET_VRING_CALL:
1444 vhost_user_set_vring_call(dev, &msg);
1447 case VHOST_USER_SET_VRING_ERR:
1448 if (!(msg.payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1450 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1453 case VHOST_USER_GET_QUEUE_NUM:
1454 msg.payload.u64 = VHOST_MAX_QUEUE_PAIRS;
1455 msg.size = sizeof(msg.payload.u64);
1456 send_vhost_reply(fd, &msg);
1459 case VHOST_USER_SET_VRING_ENABLE:
1460 vhost_user_set_vring_enable(dev, &msg);
1462 case VHOST_USER_SEND_RARP:
1463 vhost_user_send_rarp(dev, &msg);
1466 case VHOST_USER_NET_SET_MTU:
1467 ret = vhost_user_net_set_mtu(dev, &msg);
1470 case VHOST_USER_SET_SLAVE_REQ_FD:
1471 ret = vhost_user_set_req_fd(dev, &msg);
1474 case VHOST_USER_IOTLB_MSG:
1475 ret = vhost_user_iotlb_msg(&dev, &msg);
1484 if (unlock_required)
1485 vhost_user_unlock_all_queue_pairs(dev);
1487 if (msg.flags & VHOST_USER_NEED_REPLY) {
1488 msg.payload.u64 = !!ret;
1489 msg.size = sizeof(msg.payload.u64);
1490 send_vhost_reply(fd, &msg);
1493 if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1494 dev->flags |= VIRTIO_DEV_READY;
1496 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1497 if (dev->dequeue_zero_copy) {
1498 RTE_LOG(INFO, VHOST_CONFIG,
1499 "dequeue zero copy is enabled\n");
1502 if (dev->notify_ops->new_device(dev->vid) == 0)
1503 dev->flags |= VIRTIO_DEV_RUNNING;
1511 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
1514 struct VhostUserMsg msg = {
1515 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
1516 .flags = VHOST_USER_VERSION,
1517 .size = sizeof(msg.payload.iotlb),
1521 .type = VHOST_IOTLB_MISS,
1525 ret = send_vhost_message(dev->slave_req_fd, &msg);
1527 RTE_LOG(ERR, VHOST_CONFIG,
1528 "Failed to send IOTLB miss message (%d)\n",