1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2018 Intel Corporation
7 * The vhost-user protocol connection is an external interface, so it must be
8 * robust against invalid inputs.
10 * This is important because the vhost-user master is only one step removed
11 * from the guest. Malicious guests that have escaped will then launch further
12 * attacks from the vhost-user master.
14 * Even in deployments where guests are trusted, a bug in the vhost-user master
15 * can still cause invalid messages to be sent. Such messages must not
16 * compromise the stability of the DPDK application by causing crashes, memory
17 * corruption, or other problematic behavior.
19 * Do not assume received VhostUserMsg fields contain sensible values!
28 #include <sys/ioctl.h>
30 #include <sys/types.h>
32 #include <sys/syscall.h>
34 #ifdef RTE_LIBRTE_VHOST_NUMA
37 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
38 #include <linux/userfaultfd.h>
41 #include <rte_common.h>
42 #include <rte_malloc.h>
47 #include "vhost_user.h"
49 #define VIRTIO_MIN_MTU 68
50 #define VIRTIO_MAX_MTU 65535
52 static const char *vhost_message_str[VHOST_USER_MAX] = {
53 [VHOST_USER_NONE] = "VHOST_USER_NONE",
54 [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
55 [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
56 [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
57 [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
58 [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
59 [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
60 [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
61 [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
62 [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
63 [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
64 [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
65 [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
66 [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
67 [VHOST_USER_SET_VRING_ERR] = "VHOST_USER_SET_VRING_ERR",
68 [VHOST_USER_GET_PROTOCOL_FEATURES] = "VHOST_USER_GET_PROTOCOL_FEATURES",
69 [VHOST_USER_SET_PROTOCOL_FEATURES] = "VHOST_USER_SET_PROTOCOL_FEATURES",
70 [VHOST_USER_GET_QUEUE_NUM] = "VHOST_USER_GET_QUEUE_NUM",
71 [VHOST_USER_SET_VRING_ENABLE] = "VHOST_USER_SET_VRING_ENABLE",
72 [VHOST_USER_SEND_RARP] = "VHOST_USER_SEND_RARP",
73 [VHOST_USER_NET_SET_MTU] = "VHOST_USER_NET_SET_MTU",
74 [VHOST_USER_SET_SLAVE_REQ_FD] = "VHOST_USER_SET_SLAVE_REQ_FD",
75 [VHOST_USER_IOTLB_MSG] = "VHOST_USER_IOTLB_MSG",
76 [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
77 [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
78 [VHOST_USER_POSTCOPY_ADVISE] = "VHOST_USER_POSTCOPY_ADVISE",
79 [VHOST_USER_POSTCOPY_LISTEN] = "VHOST_USER_POSTCOPY_LISTEN",
88 ret = fstat(fd, &stat);
89 return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
93 free_mem_region(struct virtio_net *dev)
96 struct rte_vhost_mem_region *reg;
98 if (!dev || !dev->mem)
101 for (i = 0; i < dev->mem->nregions; i++) {
102 reg = &dev->mem->regions[i];
103 if (reg->host_user_addr) {
104 munmap(reg->mmap_addr, reg->mmap_size);
111 vhost_backend_cleanup(struct virtio_net *dev)
114 free_mem_region(dev);
119 free(dev->guest_pages);
120 dev->guest_pages = NULL;
123 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
127 if (dev->slave_req_fd >= 0) {
128 close(dev->slave_req_fd);
129 dev->slave_req_fd = -1;
132 if (dev->postcopy_ufd >= 0) {
133 close(dev->postcopy_ufd);
134 dev->postcopy_ufd = -1;
137 dev->postcopy_listening = 0;
141 * This function just returns success at the moment unless
142 * the device hasn't been initialised.
145 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
146 struct VhostUserMsg *msg __rte_unused,
147 int main_fd __rte_unused)
153 vhost_user_reset_owner(struct virtio_net **pdev,
154 struct VhostUserMsg *msg __rte_unused,
155 int main_fd __rte_unused)
157 struct virtio_net *dev = *pdev;
158 vhost_destroy_device_notify(dev);
160 cleanup_device(dev, 0);
166 * The features that we support are requested.
169 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
170 int main_fd __rte_unused)
172 struct virtio_net *dev = *pdev;
173 uint64_t features = 0;
175 rte_vhost_driver_get_features(dev->ifname, &features);
177 msg->payload.u64 = features;
178 msg->size = sizeof(msg->payload.u64);
181 return VH_RESULT_REPLY;
185 * The queue number that we support are requested.
188 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
189 int main_fd __rte_unused)
191 struct virtio_net *dev = *pdev;
192 uint32_t queue_num = 0;
194 rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
196 msg->payload.u64 = (uint64_t)queue_num;
197 msg->size = sizeof(msg->payload.u64);
200 return VH_RESULT_REPLY;
204 * We receive the negotiated features supported by us and the virtio device.
207 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
208 int main_fd __rte_unused)
210 struct virtio_net *dev = *pdev;
211 uint64_t features = msg->payload.u64;
212 uint64_t vhost_features = 0;
213 struct rte_vdpa_device *vdpa_dev;
216 rte_vhost_driver_get_features(dev->ifname, &vhost_features);
217 if (features & ~vhost_features) {
218 RTE_LOG(ERR, VHOST_CONFIG,
219 "(%d) received invalid negotiated features.\n",
221 return VH_RESULT_ERR;
224 if (dev->flags & VIRTIO_DEV_RUNNING) {
225 if (dev->features == features)
229 * Error out if master tries to change features while device is
230 * in running state. The exception being VHOST_F_LOG_ALL, which
231 * is enabled when the live-migration starts.
233 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
234 RTE_LOG(ERR, VHOST_CONFIG,
235 "(%d) features changed while device is running.\n",
237 return VH_RESULT_ERR;
240 if (dev->notify_ops->features_changed)
241 dev->notify_ops->features_changed(dev->vid, features);
244 dev->features = features;
246 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
247 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
249 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
251 VHOST_LOG_DEBUG(VHOST_CONFIG,
252 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
254 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
255 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
257 if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
258 !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
260 * Remove all but first queue pair if MQ hasn't been
261 * negotiated. This is safe because the device is not
262 * running at this stage.
264 while (dev->nr_vring > 2) {
265 struct vhost_virtqueue *vq;
267 vq = dev->virtqueue[--dev->nr_vring];
271 dev->virtqueue[dev->nr_vring] = NULL;
277 did = dev->vdpa_dev_id;
278 vdpa_dev = rte_vdpa_get_device(did);
279 if (vdpa_dev && vdpa_dev->ops->set_features)
280 vdpa_dev->ops->set_features(dev->vid);
286 * The virtio device sends us the size of the descriptor ring.
289 vhost_user_set_vring_num(struct virtio_net **pdev,
290 struct VhostUserMsg *msg,
291 int main_fd __rte_unused)
293 struct virtio_net *dev = *pdev;
294 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
296 vq->size = msg->payload.state.num;
298 /* VIRTIO 1.0, 2.4 Virtqueues says:
300 * Queue Size value is always a power of 2. The maximum Queue Size
303 if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
304 RTE_LOG(ERR, VHOST_CONFIG,
305 "invalid virtqueue size %u\n", vq->size);
306 return VH_RESULT_ERR;
309 if (dev->dequeue_zero_copy) {
311 vq->last_zmbuf_idx = 0;
312 vq->zmbuf_size = vq->size;
313 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
314 sizeof(struct zcopy_mbuf), 0);
315 if (vq->zmbufs == NULL) {
316 RTE_LOG(WARNING, VHOST_CONFIG,
317 "failed to allocate mem for zero copy; "
318 "zero copy is force disabled\n");
319 dev->dequeue_zero_copy = 0;
321 TAILQ_INIT(&vq->zmbuf_list);
324 if (vq_is_packed(dev)) {
325 vq->shadow_used_packed = rte_malloc(NULL,
327 sizeof(struct vring_used_elem_packed),
328 RTE_CACHE_LINE_SIZE);
329 if (!vq->shadow_used_packed) {
330 RTE_LOG(ERR, VHOST_CONFIG,
331 "failed to allocate memory for shadow used ring.\n");
332 return VH_RESULT_ERR;
336 vq->shadow_used_split = rte_malloc(NULL,
337 vq->size * sizeof(struct vring_used_elem),
338 RTE_CACHE_LINE_SIZE);
339 if (!vq->shadow_used_split) {
340 RTE_LOG(ERR, VHOST_CONFIG,
341 "failed to allocate memory for shadow used ring.\n");
342 return VH_RESULT_ERR;
346 vq->batch_copy_elems = rte_malloc(NULL,
347 vq->size * sizeof(struct batch_copy_elem),
348 RTE_CACHE_LINE_SIZE);
349 if (!vq->batch_copy_elems) {
350 RTE_LOG(ERR, VHOST_CONFIG,
351 "failed to allocate memory for batching copy.\n");
352 return VH_RESULT_ERR;
359 * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
360 * same numa node as the memory of vring descriptor.
362 #ifdef RTE_LIBRTE_VHOST_NUMA
363 static struct virtio_net*
364 numa_realloc(struct virtio_net *dev, int index)
366 int oldnode, newnode;
367 struct virtio_net *old_dev;
368 struct vhost_virtqueue *old_vq, *vq;
369 struct zcopy_mbuf *new_zmbuf;
370 struct vring_used_elem *new_shadow_used_split;
371 struct vring_used_elem_packed *new_shadow_used_packed;
372 struct batch_copy_elem *new_batch_copy_elems;
376 vq = old_vq = dev->virtqueue[index];
378 ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
379 MPOL_F_NODE | MPOL_F_ADDR);
381 /* check if we need to reallocate vq */
382 ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
383 MPOL_F_NODE | MPOL_F_ADDR);
385 RTE_LOG(ERR, VHOST_CONFIG,
386 "Unable to get vq numa information.\n");
389 if (oldnode != newnode) {
390 RTE_LOG(INFO, VHOST_CONFIG,
391 "reallocate vq from %d to %d node\n", oldnode, newnode);
392 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
396 memcpy(vq, old_vq, sizeof(*vq));
397 TAILQ_INIT(&vq->zmbuf_list);
399 if (dev->dequeue_zero_copy) {
400 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
401 sizeof(struct zcopy_mbuf), 0, newnode);
403 rte_free(vq->zmbufs);
404 vq->zmbufs = new_zmbuf;
408 if (vq_is_packed(dev)) {
409 new_shadow_used_packed = rte_malloc_socket(NULL,
411 sizeof(struct vring_used_elem_packed),
414 if (new_shadow_used_packed) {
415 rte_free(vq->shadow_used_packed);
416 vq->shadow_used_packed = new_shadow_used_packed;
419 new_shadow_used_split = rte_malloc_socket(NULL,
421 sizeof(struct vring_used_elem),
424 if (new_shadow_used_split) {
425 rte_free(vq->shadow_used_split);
426 vq->shadow_used_split = new_shadow_used_split;
430 new_batch_copy_elems = rte_malloc_socket(NULL,
431 vq->size * sizeof(struct batch_copy_elem),
434 if (new_batch_copy_elems) {
435 rte_free(vq->batch_copy_elems);
436 vq->batch_copy_elems = new_batch_copy_elems;
442 /* check if we need to reallocate dev */
443 ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
444 MPOL_F_NODE | MPOL_F_ADDR);
446 RTE_LOG(ERR, VHOST_CONFIG,
447 "Unable to get dev numa information.\n");
450 if (oldnode != newnode) {
451 RTE_LOG(INFO, VHOST_CONFIG,
452 "reallocate dev from %d to %d node\n",
454 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
460 memcpy(dev, old_dev, sizeof(*dev));
465 dev->virtqueue[index] = vq;
466 vhost_devices[dev->vid] = dev;
469 vhost_user_iotlb_init(dev, index);
474 static struct virtio_net*
475 numa_realloc(struct virtio_net *dev, int index __rte_unused)
481 /* Converts QEMU virtual address to Vhost virtual address. */
483 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
485 struct rte_vhost_mem_region *r;
488 /* Find the region where the address lives. */
489 for (i = 0; i < dev->mem->nregions; i++) {
490 r = &dev->mem->regions[i];
492 if (qva >= r->guest_user_addr &&
493 qva < r->guest_user_addr + r->size) {
495 if (unlikely(*len > r->guest_user_addr + r->size - qva))
496 *len = r->guest_user_addr + r->size - qva;
498 return qva - r->guest_user_addr +
509 * Converts ring address to Vhost virtual address.
510 * If IOMMU is enabled, the ring address is a guest IO virtual address,
511 * else it is a QEMU virtual address.
514 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
515 uint64_t ra, uint64_t *size)
517 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
520 vva = vhost_user_iotlb_cache_find(vq, ra,
521 size, VHOST_ACCESS_RW);
523 vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
528 return qva_to_vva(dev, ra, size);
531 static struct virtio_net *
532 translate_ring_addresses(struct virtio_net *dev, int vq_index)
534 struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
535 struct vhost_vring_addr *addr = &vq->ring_addrs;
538 if (vq_is_packed(dev)) {
539 len = sizeof(struct vring_packed_desc) * vq->size;
540 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
541 ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
542 vq->log_guest_addr = 0;
543 if (vq->desc_packed == NULL ||
544 len != sizeof(struct vring_packed_desc) *
546 RTE_LOG(DEBUG, VHOST_CONFIG,
547 "(%d) failed to map desc_packed ring.\n",
552 dev = numa_realloc(dev, vq_index);
553 vq = dev->virtqueue[vq_index];
554 addr = &vq->ring_addrs;
556 len = sizeof(struct vring_packed_desc_event);
557 vq->driver_event = (struct vring_packed_desc_event *)
558 (uintptr_t)ring_addr_to_vva(dev,
559 vq, addr->avail_user_addr, &len);
560 if (vq->driver_event == NULL ||
561 len != sizeof(struct vring_packed_desc_event)) {
562 RTE_LOG(DEBUG, VHOST_CONFIG,
563 "(%d) failed to find driver area address.\n",
568 len = sizeof(struct vring_packed_desc_event);
569 vq->device_event = (struct vring_packed_desc_event *)
570 (uintptr_t)ring_addr_to_vva(dev,
571 vq, addr->used_user_addr, &len);
572 if (vq->device_event == NULL ||
573 len != sizeof(struct vring_packed_desc_event)) {
574 RTE_LOG(DEBUG, VHOST_CONFIG,
575 "(%d) failed to find device area address.\n",
583 /* The addresses are converted from QEMU virtual to Vhost virtual. */
584 if (vq->desc && vq->avail && vq->used)
587 len = sizeof(struct vring_desc) * vq->size;
588 vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
589 vq, addr->desc_user_addr, &len);
590 if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
591 RTE_LOG(DEBUG, VHOST_CONFIG,
592 "(%d) failed to map desc ring.\n",
597 dev = numa_realloc(dev, vq_index);
598 vq = dev->virtqueue[vq_index];
599 addr = &vq->ring_addrs;
601 len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
602 vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
603 vq, addr->avail_user_addr, &len);
604 if (vq->avail == 0 ||
605 len != sizeof(struct vring_avail) +
606 sizeof(uint16_t) * vq->size) {
607 RTE_LOG(DEBUG, VHOST_CONFIG,
608 "(%d) failed to map avail ring.\n",
613 len = sizeof(struct vring_used) +
614 sizeof(struct vring_used_elem) * vq->size;
615 vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
616 vq, addr->used_user_addr, &len);
617 if (vq->used == 0 || len != sizeof(struct vring_used) +
618 sizeof(struct vring_used_elem) * vq->size) {
619 RTE_LOG(DEBUG, VHOST_CONFIG,
620 "(%d) failed to map used ring.\n",
625 if (vq->last_used_idx != vq->used->idx) {
626 RTE_LOG(WARNING, VHOST_CONFIG,
627 "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
628 "some packets maybe resent for Tx and dropped for Rx\n",
629 vq->last_used_idx, vq->used->idx);
630 vq->last_used_idx = vq->used->idx;
631 vq->last_avail_idx = vq->used->idx;
634 vq->log_guest_addr = addr->log_guest_addr;
636 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
638 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
639 dev->vid, vq->avail);
640 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
642 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
643 dev->vid, vq->log_guest_addr);
649 * The virtio device sends us the desc, used and avail ring addresses.
650 * This function then converts these to our address space.
653 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
654 int main_fd __rte_unused)
656 struct virtio_net *dev = *pdev;
657 struct vhost_virtqueue *vq;
658 struct vhost_vring_addr *addr = &msg->payload.addr;
660 if (dev->mem == NULL)
661 return VH_RESULT_ERR;
663 /* addr->index refers to the queue index. The txq 1, rxq is 0. */
664 vq = dev->virtqueue[msg->payload.addr.index];
667 * Rings addresses should not be interpreted as long as the ring is not
668 * started and enabled
670 memcpy(&vq->ring_addrs, addr, sizeof(*addr));
672 vring_invalidate(dev, vq);
674 if (vq->enabled && (dev->features &
675 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
676 dev = translate_ring_addresses(dev, msg->payload.addr.index);
678 return VH_RESULT_ERR;
687 * The virtio device sends us the available ring last used index.
690 vhost_user_set_vring_base(struct virtio_net **pdev,
691 struct VhostUserMsg *msg,
692 int main_fd __rte_unused)
694 struct virtio_net *dev = *pdev;
695 dev->virtqueue[msg->payload.state.index]->last_used_idx =
696 msg->payload.state.num;
697 dev->virtqueue[msg->payload.state.index]->last_avail_idx =
698 msg->payload.state.num;
704 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
705 uint64_t host_phys_addr, uint64_t size)
707 struct guest_page *page, *last_page;
709 if (dev->nr_guest_pages == dev->max_guest_pages) {
710 dev->max_guest_pages *= 2;
711 dev->guest_pages = realloc(dev->guest_pages,
712 dev->max_guest_pages * sizeof(*page));
713 if (!dev->guest_pages) {
714 RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
719 if (dev->nr_guest_pages > 0) {
720 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
721 /* merge if the two pages are continuous */
722 if (host_phys_addr == last_page->host_phys_addr +
724 last_page->size += size;
729 page = &dev->guest_pages[dev->nr_guest_pages++];
730 page->guest_phys_addr = guest_phys_addr;
731 page->host_phys_addr = host_phys_addr;
738 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
741 uint64_t reg_size = reg->size;
742 uint64_t host_user_addr = reg->host_user_addr;
743 uint64_t guest_phys_addr = reg->guest_phys_addr;
744 uint64_t host_phys_addr;
747 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
748 size = page_size - (guest_phys_addr & (page_size - 1));
749 size = RTE_MIN(size, reg_size);
751 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
754 host_user_addr += size;
755 guest_phys_addr += size;
758 while (reg_size > 0) {
759 size = RTE_MIN(reg_size, page_size);
760 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
762 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
766 host_user_addr += size;
767 guest_phys_addr += size;
774 #ifdef RTE_LIBRTE_VHOST_DEBUG
775 /* TODO: enable it only in debug mode? */
777 dump_guest_pages(struct virtio_net *dev)
780 struct guest_page *page;
782 for (i = 0; i < dev->nr_guest_pages; i++) {
783 page = &dev->guest_pages[i];
785 RTE_LOG(INFO, VHOST_CONFIG,
786 "guest physical page region %u\n"
787 "\t guest_phys_addr: %" PRIx64 "\n"
788 "\t host_phys_addr : %" PRIx64 "\n"
789 "\t size : %" PRIx64 "\n",
791 page->guest_phys_addr,
792 page->host_phys_addr,
797 #define dump_guest_pages(dev)
801 vhost_memory_changed(struct VhostUserMemory *new,
802 struct rte_vhost_memory *old)
806 if (new->nregions != old->nregions)
809 for (i = 0; i < new->nregions; ++i) {
810 VhostUserMemoryRegion *new_r = &new->regions[i];
811 struct rte_vhost_mem_region *old_r = &old->regions[i];
813 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
815 if (new_r->memory_size != old_r->size)
817 if (new_r->userspace_addr != old_r->guest_user_addr)
825 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
826 int main_fd __rte_unused)
828 struct virtio_net *dev = *pdev;
829 struct VhostUserMemory memory = msg->payload.memory;
830 struct rte_vhost_mem_region *reg;
833 uint64_t mmap_offset;
839 if (memory.nregions > VHOST_MEMORY_MAX_NREGIONS) {
840 RTE_LOG(ERR, VHOST_CONFIG,
841 "too many memory regions (%u)\n", memory.nregions);
842 return VH_RESULT_ERR;
845 if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
846 RTE_LOG(INFO, VHOST_CONFIG,
847 "(%d) memory regions not changed\n", dev->vid);
849 for (i = 0; i < memory.nregions; i++)
856 free_mem_region(dev);
861 /* Flush IOTLB cache as previous HVAs are now invalid */
862 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
863 for (i = 0; i < dev->nr_vring; i++)
864 vhost_user_iotlb_flush_all(dev->virtqueue[i]);
866 dev->nr_guest_pages = 0;
867 if (!dev->guest_pages) {
868 dev->max_guest_pages = 8;
869 dev->guest_pages = malloc(dev->max_guest_pages *
870 sizeof(struct guest_page));
871 if (dev->guest_pages == NULL) {
872 RTE_LOG(ERR, VHOST_CONFIG,
873 "(%d) failed to allocate memory "
874 "for dev->guest_pages\n",
876 return VH_RESULT_ERR;
880 dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
881 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
882 if (dev->mem == NULL) {
883 RTE_LOG(ERR, VHOST_CONFIG,
884 "(%d) failed to allocate memory for dev->mem\n",
886 return VH_RESULT_ERR;
888 dev->mem->nregions = memory.nregions;
890 for (i = 0; i < memory.nregions; i++) {
892 reg = &dev->mem->regions[i];
894 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
895 reg->guest_user_addr = memory.regions[i].userspace_addr;
896 reg->size = memory.regions[i].memory_size;
899 mmap_offset = memory.regions[i].mmap_offset;
901 /* Check for memory_size + mmap_offset overflow */
902 if (mmap_offset >= -reg->size) {
903 RTE_LOG(ERR, VHOST_CONFIG,
904 "mmap_offset (%#"PRIx64") and memory_size "
905 "(%#"PRIx64") overflow\n",
906 mmap_offset, reg->size);
910 mmap_size = reg->size + mmap_offset;
912 /* mmap() without flag of MAP_ANONYMOUS, should be called
913 * with length argument aligned with hugepagesz at older
914 * longterm version Linux, like 2.6.32 and 3.2.72, or
915 * mmap() will fail with EINVAL.
917 * to avoid failure, make sure in caller to keep length
920 alignment = get_blk_size(fd);
921 if (alignment == (uint64_t)-1) {
922 RTE_LOG(ERR, VHOST_CONFIG,
923 "couldn't get hugepage size through fstat\n");
926 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
928 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
929 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
930 MAP_SHARED | populate, fd, 0);
932 if (mmap_addr == MAP_FAILED) {
933 RTE_LOG(ERR, VHOST_CONFIG,
934 "mmap region %u failed.\n", i);
938 reg->mmap_addr = mmap_addr;
939 reg->mmap_size = mmap_size;
940 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
943 if (dev->dequeue_zero_copy)
944 if (add_guest_pages(dev, reg, alignment) < 0) {
945 RTE_LOG(ERR, VHOST_CONFIG,
946 "adding guest pages to region %u failed.\n",
951 RTE_LOG(INFO, VHOST_CONFIG,
952 "guest memory region %u, size: 0x%" PRIx64 "\n"
953 "\t guest physical addr: 0x%" PRIx64 "\n"
954 "\t guest virtual addr: 0x%" PRIx64 "\n"
955 "\t host virtual addr: 0x%" PRIx64 "\n"
956 "\t mmap addr : 0x%" PRIx64 "\n"
957 "\t mmap size : 0x%" PRIx64 "\n"
958 "\t mmap align: 0x%" PRIx64 "\n"
959 "\t mmap off : 0x%" PRIx64 "\n",
961 reg->guest_phys_addr,
962 reg->guest_user_addr,
964 (uint64_t)(uintptr_t)mmap_addr,
970 for (i = 0; i < dev->nr_vring; i++) {
971 struct vhost_virtqueue *vq = dev->virtqueue[i];
973 if (vq->desc || vq->avail || vq->used) {
975 * If the memory table got updated, the ring addresses
976 * need to be translated again as virtual addresses have
979 vring_invalidate(dev, vq);
981 dev = translate_ring_addresses(dev, i);
991 dump_guest_pages(dev);
996 free_mem_region(dev);
999 return VH_RESULT_ERR;
1003 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1010 if (vq_is_packed(dev))
1011 rings_ok = !!vq->desc_packed;
1013 rings_ok = vq->desc && vq->avail && vq->used;
1016 vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1017 vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1021 virtio_is_ready(struct virtio_net *dev)
1023 struct vhost_virtqueue *vq;
1026 if (dev->nr_vring == 0)
1029 for (i = 0; i < dev->nr_vring; i++) {
1030 vq = dev->virtqueue[i];
1032 if (!vq_is_ready(dev, vq))
1036 RTE_LOG(INFO, VHOST_CONFIG,
1037 "virtio is now ready for processing.\n");
1042 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1043 int main_fd __rte_unused)
1045 struct virtio_net *dev = *pdev;
1046 struct vhost_vring_file file;
1047 struct vhost_virtqueue *vq;
1049 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1050 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1051 file.fd = VIRTIO_INVALID_EVENTFD;
1053 file.fd = msg->fds[0];
1054 RTE_LOG(INFO, VHOST_CONFIG,
1055 "vring call idx:%d file:%d\n", file.index, file.fd);
1057 vq = dev->virtqueue[file.index];
1058 if (vq->callfd >= 0)
1061 vq->callfd = file.fd;
1063 return VH_RESULT_OK;
1066 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1067 struct VhostUserMsg *msg,
1068 int main_fd __rte_unused)
1070 if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1072 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1074 return VH_RESULT_OK;
1078 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1079 int main_fd __rte_unused)
1081 struct virtio_net *dev = *pdev;
1082 struct vhost_vring_file file;
1083 struct vhost_virtqueue *vq;
1085 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1086 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1087 file.fd = VIRTIO_INVALID_EVENTFD;
1089 file.fd = msg->fds[0];
1090 RTE_LOG(INFO, VHOST_CONFIG,
1091 "vring kick idx:%d file:%d\n", file.index, file.fd);
1093 /* Interpret ring addresses only when ring is started. */
1094 dev = translate_ring_addresses(dev, file.index);
1096 return VH_RESULT_ERR;
1100 vq = dev->virtqueue[file.index];
1103 * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1104 * the ring starts already enabled. Otherwise, it is enabled via
1105 * the SET_VRING_ENABLE message.
1107 if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
1110 if (vq->kickfd >= 0)
1112 vq->kickfd = file.fd;
1114 return VH_RESULT_OK;
1118 free_zmbufs(struct vhost_virtqueue *vq)
1120 struct zcopy_mbuf *zmbuf, *next;
1122 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1123 zmbuf != NULL; zmbuf = next) {
1124 next = TAILQ_NEXT(zmbuf, next);
1126 rte_pktmbuf_free(zmbuf->mbuf);
1127 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1130 rte_free(vq->zmbufs);
1134 * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1137 vhost_user_get_vring_base(struct virtio_net **pdev,
1138 struct VhostUserMsg *msg,
1139 int main_fd __rte_unused)
1141 struct virtio_net *dev = *pdev;
1142 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1144 /* We have to stop the queue (virtio) if it is running. */
1145 vhost_destroy_device_notify(dev);
1147 dev->flags &= ~VIRTIO_DEV_READY;
1148 dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1150 /* Here we are safe to get the last avail index */
1151 msg->payload.state.num = vq->last_avail_idx;
1153 RTE_LOG(INFO, VHOST_CONFIG,
1154 "vring base idx:%d file:%d\n", msg->payload.state.index,
1155 msg->payload.state.num);
1157 * Based on current qemu vhost-user implementation, this message is
1158 * sent and only sent in vhost_vring_stop.
1159 * TODO: cleanup the vring, it isn't usable since here.
1161 if (vq->kickfd >= 0)
1164 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1166 if (vq->callfd >= 0)
1169 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1171 if (dev->dequeue_zero_copy)
1173 if (vq_is_packed(dev)) {
1174 rte_free(vq->shadow_used_packed);
1175 vq->shadow_used_packed = NULL;
1177 rte_free(vq->shadow_used_split);
1178 vq->shadow_used_split = NULL;
1181 rte_free(vq->batch_copy_elems);
1182 vq->batch_copy_elems = NULL;
1184 msg->size = sizeof(msg->payload.state);
1187 return VH_RESULT_REPLY;
1191 * when virtio queues are ready to work, qemu will send us to
1192 * enable the virtio queue pair.
1195 vhost_user_set_vring_enable(struct virtio_net **pdev,
1196 struct VhostUserMsg *msg,
1197 int main_fd __rte_unused)
1199 struct virtio_net *dev = *pdev;
1200 int enable = (int)msg->payload.state.num;
1201 int index = (int)msg->payload.state.index;
1202 struct rte_vdpa_device *vdpa_dev;
1205 RTE_LOG(INFO, VHOST_CONFIG,
1206 "set queue enable: %d to qp idx: %d\n",
1209 did = dev->vdpa_dev_id;
1210 vdpa_dev = rte_vdpa_get_device(did);
1211 if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1212 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1214 if (dev->notify_ops->vring_state_changed)
1215 dev->notify_ops->vring_state_changed(dev->vid,
1218 dev->virtqueue[index]->enabled = enable;
1220 return VH_RESULT_OK;
1224 vhost_user_get_protocol_features(struct virtio_net **pdev,
1225 struct VhostUserMsg *msg,
1226 int main_fd __rte_unused)
1228 struct virtio_net *dev = *pdev;
1229 uint64_t features, protocol_features;
1231 rte_vhost_driver_get_features(dev->ifname, &features);
1232 rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1235 * REPLY_ACK protocol feature is only mandatory for now
1236 * for IOMMU feature. If IOMMU is explicitly disabled by the
1237 * application, disable also REPLY_ACK feature for older buggy
1238 * Qemu versions (from v2.7.0 to v2.9.0).
1240 if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1241 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1243 msg->payload.u64 = protocol_features;
1244 msg->size = sizeof(msg->payload.u64);
1247 return VH_RESULT_REPLY;
1251 vhost_user_set_protocol_features(struct virtio_net **pdev,
1252 struct VhostUserMsg *msg,
1253 int main_fd __rte_unused)
1255 struct virtio_net *dev = *pdev;
1256 uint64_t protocol_features = msg->payload.u64;
1257 if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES) {
1258 RTE_LOG(ERR, VHOST_CONFIG,
1259 "(%d) received invalid protocol features.\n",
1261 return VH_RESULT_ERR;
1264 dev->protocol_features = protocol_features;
1266 return VH_RESULT_OK;
1270 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1271 int main_fd __rte_unused)
1273 struct virtio_net *dev = *pdev;
1274 int fd = msg->fds[0];
1279 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1280 return VH_RESULT_ERR;
1283 if (msg->size != sizeof(VhostUserLog)) {
1284 RTE_LOG(ERR, VHOST_CONFIG,
1285 "invalid log base msg size: %"PRId32" != %d\n",
1286 msg->size, (int)sizeof(VhostUserLog));
1287 return VH_RESULT_ERR;
1290 size = msg->payload.log.mmap_size;
1291 off = msg->payload.log.mmap_offset;
1293 /* Don't allow mmap_offset to point outside the mmap region */
1295 RTE_LOG(ERR, VHOST_CONFIG,
1296 "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1298 return VH_RESULT_ERR;
1301 RTE_LOG(INFO, VHOST_CONFIG,
1302 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1306 * mmap from 0 to workaround a hugepage mmap bug: mmap will
1307 * fail when offset is not page size aligned.
1309 addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1311 if (addr == MAP_FAILED) {
1312 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1313 return VH_RESULT_ERR;
1317 * Free previously mapped log memory on occasionally
1318 * multiple VHOST_USER_SET_LOG_BASE.
1320 if (dev->log_addr) {
1321 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1323 dev->log_addr = (uint64_t)(uintptr_t)addr;
1324 dev->log_base = dev->log_addr + off;
1325 dev->log_size = size;
1328 * The spec is not clear about it (yet), but QEMU doesn't expect
1329 * any payload in the reply.
1334 return VH_RESULT_REPLY;
1337 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1338 struct VhostUserMsg *msg,
1339 int main_fd __rte_unused)
1342 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1344 return VH_RESULT_OK;
1348 * An rarp packet is constructed and broadcasted to notify switches about
1349 * the new location of the migrated VM, so that packets from outside will
1350 * not be lost after migration.
1352 * However, we don't actually "send" a rarp packet here, instead, we set
1353 * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1356 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1357 int main_fd __rte_unused)
1359 struct virtio_net *dev = *pdev;
1360 uint8_t *mac = (uint8_t *)&msg->payload.u64;
1361 struct rte_vdpa_device *vdpa_dev;
1364 RTE_LOG(DEBUG, VHOST_CONFIG,
1365 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1366 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1367 memcpy(dev->mac.addr_bytes, mac, 6);
1370 * Set the flag to inject a RARP broadcast packet at
1371 * rte_vhost_dequeue_burst().
1373 * rte_smp_wmb() is for making sure the mac is copied
1374 * before the flag is set.
1377 rte_atomic16_set(&dev->broadcast_rarp, 1);
1378 did = dev->vdpa_dev_id;
1379 vdpa_dev = rte_vdpa_get_device(did);
1380 if (vdpa_dev && vdpa_dev->ops->migration_done)
1381 vdpa_dev->ops->migration_done(dev->vid);
1383 return VH_RESULT_OK;
1387 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1388 int main_fd __rte_unused)
1390 struct virtio_net *dev = *pdev;
1391 if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1392 msg->payload.u64 > VIRTIO_MAX_MTU) {
1393 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1396 return VH_RESULT_ERR;
1399 dev->mtu = msg->payload.u64;
1401 return VH_RESULT_OK;
1405 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1406 int main_fd __rte_unused)
1408 struct virtio_net *dev = *pdev;
1409 int fd = msg->fds[0];
1412 RTE_LOG(ERR, VHOST_CONFIG,
1413 "Invalid file descriptor for slave channel (%d)\n",
1415 return VH_RESULT_ERR;
1418 dev->slave_req_fd = fd;
1420 return VH_RESULT_OK;
1424 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1426 struct vhost_vring_addr *ra;
1427 uint64_t start, end;
1430 end = start + imsg->size;
1432 ra = &vq->ring_addrs;
1433 if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1435 if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1437 if (ra->used_user_addr >= start && ra->used_user_addr < end)
1444 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1445 struct vhost_iotlb_msg *imsg)
1447 uint64_t istart, iend, vstart, vend;
1449 istart = imsg->iova;
1450 iend = istart + imsg->size - 1;
1452 vstart = (uintptr_t)vq->desc;
1453 vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1454 if (vstart <= iend && istart <= vend)
1457 vstart = (uintptr_t)vq->avail;
1458 vend = vstart + sizeof(struct vring_avail);
1459 vend += sizeof(uint16_t) * vq->size - 1;
1460 if (vstart <= iend && istart <= vend)
1463 vstart = (uintptr_t)vq->used;
1464 vend = vstart + sizeof(struct vring_used);
1465 vend += sizeof(struct vring_used_elem) * vq->size - 1;
1466 if (vstart <= iend && istart <= vend)
1473 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1474 int main_fd __rte_unused)
1476 struct virtio_net *dev = *pdev;
1477 struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1481 switch (imsg->type) {
1482 case VHOST_IOTLB_UPDATE:
1484 vva = qva_to_vva(dev, imsg->uaddr, &len);
1486 return VH_RESULT_ERR;
1488 for (i = 0; i < dev->nr_vring; i++) {
1489 struct vhost_virtqueue *vq = dev->virtqueue[i];
1491 vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1494 if (is_vring_iotlb_update(vq, imsg))
1495 *pdev = dev = translate_ring_addresses(dev, i);
1498 case VHOST_IOTLB_INVALIDATE:
1499 for (i = 0; i < dev->nr_vring; i++) {
1500 struct vhost_virtqueue *vq = dev->virtqueue[i];
1502 vhost_user_iotlb_cache_remove(vq, imsg->iova,
1505 if (is_vring_iotlb_invalidate(vq, imsg))
1506 vring_invalidate(dev, vq);
1510 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1512 return VH_RESULT_ERR;
1515 return VH_RESULT_OK;
1519 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1520 struct VhostUserMsg *msg,
1521 int main_fd __rte_unused)
1523 struct virtio_net *dev = *pdev;
1524 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1525 struct uffdio_api api_struct;
1527 dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1529 if (dev->postcopy_ufd == -1) {
1530 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1532 return VH_RESULT_ERR;
1534 api_struct.api = UFFD_API;
1535 api_struct.features = 0;
1536 if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1537 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1539 close(dev->postcopy_ufd);
1540 dev->postcopy_ufd = -1;
1541 return VH_RESULT_ERR;
1543 msg->fds[0] = dev->postcopy_ufd;
1546 return VH_RESULT_REPLY;
1548 dev->postcopy_ufd = -1;
1551 return VH_RESULT_ERR;
1556 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1557 struct VhostUserMsg *msg __rte_unused,
1558 int main_fd __rte_unused)
1560 struct virtio_net *dev = *pdev;
1562 if (dev->mem && dev->mem->nregions) {
1563 RTE_LOG(ERR, VHOST_CONFIG,
1564 "Regions already registered at postcopy-listen\n");
1565 return VH_RESULT_ERR;
1567 dev->postcopy_listening = 1;
1569 return VH_RESULT_OK;
1572 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1573 struct VhostUserMsg *msg,
1575 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1576 [VHOST_USER_NONE] = NULL,
1577 [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1578 [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1579 [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1580 [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1581 [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1582 [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1583 [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1584 [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1585 [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1586 [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1587 [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1588 [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1589 [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1590 [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1591 [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1592 [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1593 [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1594 [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1595 [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1596 [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1597 [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1598 [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1599 [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1600 [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1604 /* return bytes# of read on success or negative val on failure. */
1606 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1610 ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1611 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1615 if (msg && msg->size) {
1616 if (msg->size > sizeof(msg->payload)) {
1617 RTE_LOG(ERR, VHOST_CONFIG,
1618 "invalid msg size: %d\n", msg->size);
1621 ret = read(sockfd, &msg->payload, msg->size);
1624 if (ret != (int)msg->size) {
1625 RTE_LOG(ERR, VHOST_CONFIG,
1626 "read control message failed\n");
1635 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1640 return send_fd_message(sockfd, (char *)msg,
1641 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1645 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1650 msg->flags &= ~VHOST_USER_VERSION_MASK;
1651 msg->flags &= ~VHOST_USER_NEED_REPLY;
1652 msg->flags |= VHOST_USER_VERSION;
1653 msg->flags |= VHOST_USER_REPLY_MASK;
1655 return send_vhost_message(sockfd, msg);
1659 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1663 if (msg->flags & VHOST_USER_NEED_REPLY)
1664 rte_spinlock_lock(&dev->slave_req_lock);
1666 ret = send_vhost_message(dev->slave_req_fd, msg);
1667 if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1668 rte_spinlock_unlock(&dev->slave_req_lock);
1674 * Allocate a queue pair if it hasn't been allocated yet
1677 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1678 struct VhostUserMsg *msg)
1682 switch (msg->request.master) {
1683 case VHOST_USER_SET_VRING_KICK:
1684 case VHOST_USER_SET_VRING_CALL:
1685 case VHOST_USER_SET_VRING_ERR:
1686 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1688 case VHOST_USER_SET_VRING_NUM:
1689 case VHOST_USER_SET_VRING_BASE:
1690 case VHOST_USER_SET_VRING_ENABLE:
1691 vring_idx = msg->payload.state.index;
1693 case VHOST_USER_SET_VRING_ADDR:
1694 vring_idx = msg->payload.addr.index;
1700 if (vring_idx >= VHOST_MAX_VRING) {
1701 RTE_LOG(ERR, VHOST_CONFIG,
1702 "invalid vring index: %u\n", vring_idx);
1706 if (dev->virtqueue[vring_idx])
1709 return alloc_vring_queue(dev, vring_idx);
1713 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1716 unsigned int vq_num = 0;
1718 while (vq_num < dev->nr_vring) {
1719 struct vhost_virtqueue *vq = dev->virtqueue[i];
1722 rte_spinlock_lock(&vq->access_lock);
1730 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1733 unsigned int vq_num = 0;
1735 while (vq_num < dev->nr_vring) {
1736 struct vhost_virtqueue *vq = dev->virtqueue[i];
1739 rte_spinlock_unlock(&vq->access_lock);
1747 vhost_user_msg_handler(int vid, int fd)
1749 struct virtio_net *dev;
1750 struct VhostUserMsg msg;
1751 struct rte_vdpa_device *vdpa_dev;
1754 int unlock_required = 0;
1755 uint32_t skip_master = 0;
1758 dev = get_device(vid);
1762 if (!dev->notify_ops) {
1763 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1764 if (!dev->notify_ops) {
1765 RTE_LOG(ERR, VHOST_CONFIG,
1766 "failed to get callback ops for driver %s\n",
1772 ret = read_vhost_message(fd, &msg);
1773 if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1775 RTE_LOG(ERR, VHOST_CONFIG,
1776 "vhost read message failed\n");
1778 RTE_LOG(INFO, VHOST_CONFIG,
1779 "vhost peer closed\n");
1781 RTE_LOG(ERR, VHOST_CONFIG,
1782 "vhost read incorrect message\n");
1788 if (msg.request.master != VHOST_USER_IOTLB_MSG)
1789 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1790 vhost_message_str[msg.request.master]);
1792 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1793 vhost_message_str[msg.request.master]);
1795 ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1797 RTE_LOG(ERR, VHOST_CONFIG,
1798 "failed to alloc queue\n");
1803 * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1804 * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1805 * and device is destroyed. destroy_device waits for queues to be
1806 * inactive, so it is safe. Otherwise taking the access_lock
1807 * would cause a dead lock.
1809 switch (msg.request.master) {
1810 case VHOST_USER_SET_FEATURES:
1811 case VHOST_USER_SET_PROTOCOL_FEATURES:
1812 case VHOST_USER_SET_OWNER:
1813 case VHOST_USER_SET_MEM_TABLE:
1814 case VHOST_USER_SET_LOG_BASE:
1815 case VHOST_USER_SET_LOG_FD:
1816 case VHOST_USER_SET_VRING_NUM:
1817 case VHOST_USER_SET_VRING_ADDR:
1818 case VHOST_USER_SET_VRING_BASE:
1819 case VHOST_USER_SET_VRING_KICK:
1820 case VHOST_USER_SET_VRING_CALL:
1821 case VHOST_USER_SET_VRING_ERR:
1822 case VHOST_USER_SET_VRING_ENABLE:
1823 case VHOST_USER_SEND_RARP:
1824 case VHOST_USER_NET_SET_MTU:
1825 case VHOST_USER_SET_SLAVE_REQ_FD:
1826 vhost_user_lock_all_queue_pairs(dev);
1827 unlock_required = 1;
1834 if (dev->extern_ops.pre_msg_handle) {
1835 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
1836 (void *)&msg, &skip_master);
1837 if (ret == VH_RESULT_ERR)
1839 else if (ret == VH_RESULT_REPLY)
1840 send_vhost_reply(fd, &msg);
1843 goto skip_to_post_handle;
1846 request = msg.request.master;
1847 if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
1848 if (!vhost_message_handlers[request])
1849 goto skip_to_post_handle;
1850 ret = vhost_message_handlers[request](&dev, &msg, fd);
1854 RTE_LOG(ERR, VHOST_CONFIG,
1855 "Processing %s failed.\n",
1856 vhost_message_str[request]);
1859 RTE_LOG(DEBUG, VHOST_CONFIG,
1860 "Processing %s succeeded.\n",
1861 vhost_message_str[request]);
1863 case VH_RESULT_REPLY:
1864 RTE_LOG(DEBUG, VHOST_CONFIG,
1865 "Processing %s succeeded and needs reply.\n",
1866 vhost_message_str[request]);
1867 send_vhost_reply(fd, &msg);
1871 RTE_LOG(ERR, VHOST_CONFIG,
1872 "Requested invalid message type %d.\n", request);
1873 ret = VH_RESULT_ERR;
1876 skip_to_post_handle:
1877 if (ret != VH_RESULT_ERR && dev->extern_ops.post_msg_handle) {
1878 ret = (*dev->extern_ops.post_msg_handle)(
1879 dev->vid, (void *)&msg);
1880 if (ret == VH_RESULT_ERR)
1882 else if (ret == VH_RESULT_REPLY)
1883 send_vhost_reply(fd, &msg);
1887 if (unlock_required)
1888 vhost_user_unlock_all_queue_pairs(dev);
1891 * If the request required a reply that was already sent,
1892 * this optional reply-ack won't be sent as the
1893 * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
1895 if (msg.flags & VHOST_USER_NEED_REPLY) {
1896 msg.payload.u64 = ret == VH_RESULT_ERR;
1897 msg.size = sizeof(msg.payload.u64);
1899 send_vhost_reply(fd, &msg);
1900 } else if (ret == VH_RESULT_ERR) {
1901 RTE_LOG(ERR, VHOST_CONFIG,
1902 "vhost message handling failed.\n");
1906 if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1907 dev->flags |= VIRTIO_DEV_READY;
1909 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1910 if (dev->dequeue_zero_copy) {
1911 RTE_LOG(INFO, VHOST_CONFIG,
1912 "dequeue zero copy is enabled\n");
1915 if (dev->notify_ops->new_device(dev->vid) == 0)
1916 dev->flags |= VIRTIO_DEV_RUNNING;
1920 did = dev->vdpa_dev_id;
1921 vdpa_dev = rte_vdpa_get_device(did);
1922 if (vdpa_dev && virtio_is_ready(dev) &&
1923 !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
1924 msg.request.master == VHOST_USER_SET_VRING_ENABLE) {
1925 if (vdpa_dev->ops->dev_conf)
1926 vdpa_dev->ops->dev_conf(dev->vid);
1927 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
1928 if (vhost_user_host_notifier_ctrl(dev->vid, true) != 0) {
1929 RTE_LOG(INFO, VHOST_CONFIG,
1930 "(%d) software relay is used for vDPA, performance may be low.\n",
1938 static int process_slave_message_reply(struct virtio_net *dev,
1939 const struct VhostUserMsg *msg)
1941 struct VhostUserMsg msg_reply;
1944 if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
1947 if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
1952 if (msg_reply.request.slave != msg->request.slave) {
1953 RTE_LOG(ERR, VHOST_CONFIG,
1954 "Received unexpected msg type (%u), expected %u\n",
1955 msg_reply.request.slave, msg->request.slave);
1960 ret = msg_reply.payload.u64 ? -1 : 0;
1963 rte_spinlock_unlock(&dev->slave_req_lock);
1968 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
1971 struct VhostUserMsg msg = {
1972 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
1973 .flags = VHOST_USER_VERSION,
1974 .size = sizeof(msg.payload.iotlb),
1978 .type = VHOST_IOTLB_MISS,
1982 ret = send_vhost_message(dev->slave_req_fd, &msg);
1984 RTE_LOG(ERR, VHOST_CONFIG,
1985 "Failed to send IOTLB miss message (%d)\n",
1993 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
1999 struct VhostUserMsg msg = {
2000 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2001 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2002 .size = sizeof(msg.payload.area),
2004 .u64 = index & VHOST_USER_VRING_IDX_MASK,
2011 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2017 ret = send_vhost_slave_message(dev, &msg);
2019 RTE_LOG(ERR, VHOST_CONFIG,
2020 "Failed to set host notifier (%d)\n", ret);
2024 return process_slave_message_reply(dev, &msg);
2027 int vhost_user_host_notifier_ctrl(int vid, bool enable)
2029 struct virtio_net *dev;
2030 struct rte_vdpa_device *vdpa_dev;
2031 int vfio_device_fd, did, ret = 0;
2032 uint64_t offset, size;
2035 dev = get_device(vid);
2039 did = dev->vdpa_dev_id;
2043 if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2044 !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2045 !(dev->protocol_features &
2046 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2047 !(dev->protocol_features &
2048 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2049 !(dev->protocol_features &
2050 (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2053 vdpa_dev = rte_vdpa_get_device(did);
2057 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2058 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2060 vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2061 if (vfio_device_fd < 0)
2065 for (i = 0; i < dev->nr_vring; i++) {
2066 if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2072 if (vhost_user_slave_set_vring_host_notifier(dev, i,
2073 vfio_device_fd, offset, size) < 0) {
2080 for (i = 0; i < dev->nr_vring; i++) {
2081 vhost_user_slave_set_vring_host_notifier(dev, i, -1,