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/types.h>
31 #ifdef RTE_LIBRTE_VHOST_NUMA
35 #include <rte_common.h>
36 #include <rte_malloc.h>
41 #include "vhost_user.h"
43 #define VIRTIO_MIN_MTU 68
44 #define VIRTIO_MAX_MTU 65535
46 static const char *vhost_message_str[VHOST_USER_MAX] = {
47 [VHOST_USER_NONE] = "VHOST_USER_NONE",
48 [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
49 [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
50 [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
51 [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
52 [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
53 [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
54 [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
55 [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
56 [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
57 [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
58 [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
59 [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
60 [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
61 [VHOST_USER_SET_VRING_ERR] = "VHOST_USER_SET_VRING_ERR",
62 [VHOST_USER_GET_PROTOCOL_FEATURES] = "VHOST_USER_GET_PROTOCOL_FEATURES",
63 [VHOST_USER_SET_PROTOCOL_FEATURES] = "VHOST_USER_SET_PROTOCOL_FEATURES",
64 [VHOST_USER_GET_QUEUE_NUM] = "VHOST_USER_GET_QUEUE_NUM",
65 [VHOST_USER_SET_VRING_ENABLE] = "VHOST_USER_SET_VRING_ENABLE",
66 [VHOST_USER_SEND_RARP] = "VHOST_USER_SEND_RARP",
67 [VHOST_USER_NET_SET_MTU] = "VHOST_USER_NET_SET_MTU",
68 [VHOST_USER_SET_SLAVE_REQ_FD] = "VHOST_USER_SET_SLAVE_REQ_FD",
69 [VHOST_USER_IOTLB_MSG] = "VHOST_USER_IOTLB_MSG",
70 [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
71 [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
80 ret = fstat(fd, &stat);
81 return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
85 free_mem_region(struct virtio_net *dev)
88 struct rte_vhost_mem_region *reg;
90 if (!dev || !dev->mem)
93 for (i = 0; i < dev->mem->nregions; i++) {
94 reg = &dev->mem->regions[i];
95 if (reg->host_user_addr) {
96 munmap(reg->mmap_addr, reg->mmap_size);
103 vhost_backend_cleanup(struct virtio_net *dev)
106 free_mem_region(dev);
111 free(dev->guest_pages);
112 dev->guest_pages = NULL;
115 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
119 if (dev->slave_req_fd >= 0) {
120 close(dev->slave_req_fd);
121 dev->slave_req_fd = -1;
126 * This function just returns success at the moment unless
127 * the device hasn't been initialised.
130 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
131 struct VhostUserMsg *msg __rte_unused,
132 int main_fd __rte_unused)
138 vhost_user_reset_owner(struct virtio_net **pdev,
139 struct VhostUserMsg *msg __rte_unused,
140 int main_fd __rte_unused)
142 struct virtio_net *dev = *pdev;
143 vhost_destroy_device_notify(dev);
145 cleanup_device(dev, 0);
151 * The features that we support are requested.
154 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
155 int main_fd __rte_unused)
157 struct virtio_net *dev = *pdev;
158 uint64_t features = 0;
160 rte_vhost_driver_get_features(dev->ifname, &features);
162 msg->payload.u64 = features;
163 msg->size = sizeof(msg->payload.u64);
166 return VH_RESULT_REPLY;
170 * The queue number that we support are requested.
173 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
174 int main_fd __rte_unused)
176 struct virtio_net *dev = *pdev;
177 uint32_t queue_num = 0;
179 rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
181 msg->payload.u64 = (uint64_t)queue_num;
182 msg->size = sizeof(msg->payload.u64);
185 return VH_RESULT_REPLY;
189 * We receive the negotiated features supported by us and the virtio device.
192 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
193 int main_fd __rte_unused)
195 struct virtio_net *dev = *pdev;
196 uint64_t features = msg->payload.u64;
197 uint64_t vhost_features = 0;
198 struct rte_vdpa_device *vdpa_dev;
201 rte_vhost_driver_get_features(dev->ifname, &vhost_features);
202 if (features & ~vhost_features) {
203 RTE_LOG(ERR, VHOST_CONFIG,
204 "(%d) received invalid negotiated features.\n",
206 return VH_RESULT_ERR;
209 if (dev->flags & VIRTIO_DEV_RUNNING) {
210 if (dev->features == features)
214 * Error out if master tries to change features while device is
215 * in running state. The exception being VHOST_F_LOG_ALL, which
216 * is enabled when the live-migration starts.
218 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
219 RTE_LOG(ERR, VHOST_CONFIG,
220 "(%d) features changed while device is running.\n",
222 return VH_RESULT_ERR;
225 if (dev->notify_ops->features_changed)
226 dev->notify_ops->features_changed(dev->vid, features);
229 dev->features = features;
231 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
232 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
234 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
236 VHOST_LOG_DEBUG(VHOST_CONFIG,
237 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
239 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
240 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
242 if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
243 !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
245 * Remove all but first queue pair if MQ hasn't been
246 * negotiated. This is safe because the device is not
247 * running at this stage.
249 while (dev->nr_vring > 2) {
250 struct vhost_virtqueue *vq;
252 vq = dev->virtqueue[--dev->nr_vring];
256 dev->virtqueue[dev->nr_vring] = NULL;
262 did = dev->vdpa_dev_id;
263 vdpa_dev = rte_vdpa_get_device(did);
264 if (vdpa_dev && vdpa_dev->ops->set_features)
265 vdpa_dev->ops->set_features(dev->vid);
271 * The virtio device sends us the size of the descriptor ring.
274 vhost_user_set_vring_num(struct virtio_net **pdev,
275 struct VhostUserMsg *msg,
276 int main_fd __rte_unused)
278 struct virtio_net *dev = *pdev;
279 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
281 vq->size = msg->payload.state.num;
283 /* VIRTIO 1.0, 2.4 Virtqueues says:
285 * Queue Size value is always a power of 2. The maximum Queue Size
288 if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
289 RTE_LOG(ERR, VHOST_CONFIG,
290 "invalid virtqueue size %u\n", vq->size);
291 return VH_RESULT_ERR;
294 if (dev->dequeue_zero_copy) {
296 vq->last_zmbuf_idx = 0;
297 vq->zmbuf_size = vq->size;
298 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
299 sizeof(struct zcopy_mbuf), 0);
300 if (vq->zmbufs == NULL) {
301 RTE_LOG(WARNING, VHOST_CONFIG,
302 "failed to allocate mem for zero copy; "
303 "zero copy is force disabled\n");
304 dev->dequeue_zero_copy = 0;
306 TAILQ_INIT(&vq->zmbuf_list);
309 if (vq_is_packed(dev)) {
310 vq->shadow_used_packed = rte_malloc(NULL,
312 sizeof(struct vring_used_elem_packed),
313 RTE_CACHE_LINE_SIZE);
314 if (!vq->shadow_used_packed) {
315 RTE_LOG(ERR, VHOST_CONFIG,
316 "failed to allocate memory for shadow used ring.\n");
317 return VH_RESULT_ERR;
321 vq->shadow_used_split = rte_malloc(NULL,
322 vq->size * sizeof(struct vring_used_elem),
323 RTE_CACHE_LINE_SIZE);
324 if (!vq->shadow_used_split) {
325 RTE_LOG(ERR, VHOST_CONFIG,
326 "failed to allocate memory for shadow used ring.\n");
327 return VH_RESULT_ERR;
331 vq->batch_copy_elems = rte_malloc(NULL,
332 vq->size * sizeof(struct batch_copy_elem),
333 RTE_CACHE_LINE_SIZE);
334 if (!vq->batch_copy_elems) {
335 RTE_LOG(ERR, VHOST_CONFIG,
336 "failed to allocate memory for batching copy.\n");
337 return VH_RESULT_ERR;
344 * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
345 * same numa node as the memory of vring descriptor.
347 #ifdef RTE_LIBRTE_VHOST_NUMA
348 static struct virtio_net*
349 numa_realloc(struct virtio_net *dev, int index)
351 int oldnode, newnode;
352 struct virtio_net *old_dev;
353 struct vhost_virtqueue *old_vq, *vq;
354 struct zcopy_mbuf *new_zmbuf;
355 struct vring_used_elem *new_shadow_used_split;
356 struct vring_used_elem_packed *new_shadow_used_packed;
357 struct batch_copy_elem *new_batch_copy_elems;
361 vq = old_vq = dev->virtqueue[index];
363 ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
364 MPOL_F_NODE | MPOL_F_ADDR);
366 /* check if we need to reallocate vq */
367 ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
368 MPOL_F_NODE | MPOL_F_ADDR);
370 RTE_LOG(ERR, VHOST_CONFIG,
371 "Unable to get vq numa information.\n");
374 if (oldnode != newnode) {
375 RTE_LOG(INFO, VHOST_CONFIG,
376 "reallocate vq from %d to %d node\n", oldnode, newnode);
377 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
381 memcpy(vq, old_vq, sizeof(*vq));
382 TAILQ_INIT(&vq->zmbuf_list);
384 if (dev->dequeue_zero_copy) {
385 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
386 sizeof(struct zcopy_mbuf), 0, newnode);
388 rte_free(vq->zmbufs);
389 vq->zmbufs = new_zmbuf;
393 if (vq_is_packed(dev)) {
394 new_shadow_used_packed = rte_malloc_socket(NULL,
396 sizeof(struct vring_used_elem_packed),
399 if (new_shadow_used_packed) {
400 rte_free(vq->shadow_used_packed);
401 vq->shadow_used_packed = new_shadow_used_packed;
404 new_shadow_used_split = rte_malloc_socket(NULL,
406 sizeof(struct vring_used_elem),
409 if (new_shadow_used_split) {
410 rte_free(vq->shadow_used_split);
411 vq->shadow_used_split = new_shadow_used_split;
415 new_batch_copy_elems = rte_malloc_socket(NULL,
416 vq->size * sizeof(struct batch_copy_elem),
419 if (new_batch_copy_elems) {
420 rte_free(vq->batch_copy_elems);
421 vq->batch_copy_elems = new_batch_copy_elems;
427 /* check if we need to reallocate dev */
428 ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
429 MPOL_F_NODE | MPOL_F_ADDR);
431 RTE_LOG(ERR, VHOST_CONFIG,
432 "Unable to get dev numa information.\n");
435 if (oldnode != newnode) {
436 RTE_LOG(INFO, VHOST_CONFIG,
437 "reallocate dev from %d to %d node\n",
439 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
445 memcpy(dev, old_dev, sizeof(*dev));
450 dev->virtqueue[index] = vq;
451 vhost_devices[dev->vid] = dev;
454 vhost_user_iotlb_init(dev, index);
459 static struct virtio_net*
460 numa_realloc(struct virtio_net *dev, int index __rte_unused)
466 /* Converts QEMU virtual address to Vhost virtual address. */
468 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
470 struct rte_vhost_mem_region *r;
473 /* Find the region where the address lives. */
474 for (i = 0; i < dev->mem->nregions; i++) {
475 r = &dev->mem->regions[i];
477 if (qva >= r->guest_user_addr &&
478 qva < r->guest_user_addr + r->size) {
480 if (unlikely(*len > r->guest_user_addr + r->size - qva))
481 *len = r->guest_user_addr + r->size - qva;
483 return qva - r->guest_user_addr +
494 * Converts ring address to Vhost virtual address.
495 * If IOMMU is enabled, the ring address is a guest IO virtual address,
496 * else it is a QEMU virtual address.
499 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
500 uint64_t ra, uint64_t *size)
502 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
505 vva = vhost_user_iotlb_cache_find(vq, ra,
506 size, VHOST_ACCESS_RW);
508 vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
513 return qva_to_vva(dev, ra, size);
516 static struct virtio_net *
517 translate_ring_addresses(struct virtio_net *dev, int vq_index)
519 struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
520 struct vhost_vring_addr *addr = &vq->ring_addrs;
523 if (vq_is_packed(dev)) {
524 len = sizeof(struct vring_packed_desc) * vq->size;
525 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
526 ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
527 vq->log_guest_addr = 0;
528 if (vq->desc_packed == NULL ||
529 len != sizeof(struct vring_packed_desc) *
531 RTE_LOG(DEBUG, VHOST_CONFIG,
532 "(%d) failed to map desc_packed ring.\n",
537 dev = numa_realloc(dev, vq_index);
538 vq = dev->virtqueue[vq_index];
539 addr = &vq->ring_addrs;
541 len = sizeof(struct vring_packed_desc_event);
542 vq->driver_event = (struct vring_packed_desc_event *)
543 (uintptr_t)ring_addr_to_vva(dev,
544 vq, addr->avail_user_addr, &len);
545 if (vq->driver_event == NULL ||
546 len != sizeof(struct vring_packed_desc_event)) {
547 RTE_LOG(DEBUG, VHOST_CONFIG,
548 "(%d) failed to find driver area address.\n",
553 len = sizeof(struct vring_packed_desc_event);
554 vq->device_event = (struct vring_packed_desc_event *)
555 (uintptr_t)ring_addr_to_vva(dev,
556 vq, addr->used_user_addr, &len);
557 if (vq->device_event == NULL ||
558 len != sizeof(struct vring_packed_desc_event)) {
559 RTE_LOG(DEBUG, VHOST_CONFIG,
560 "(%d) failed to find device area address.\n",
568 /* The addresses are converted from QEMU virtual to Vhost virtual. */
569 if (vq->desc && vq->avail && vq->used)
572 len = sizeof(struct vring_desc) * vq->size;
573 vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
574 vq, addr->desc_user_addr, &len);
575 if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
576 RTE_LOG(DEBUG, VHOST_CONFIG,
577 "(%d) failed to map desc ring.\n",
582 dev = numa_realloc(dev, vq_index);
583 vq = dev->virtqueue[vq_index];
584 addr = &vq->ring_addrs;
586 len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
587 vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
588 vq, addr->avail_user_addr, &len);
589 if (vq->avail == 0 ||
590 len != sizeof(struct vring_avail) +
591 sizeof(uint16_t) * vq->size) {
592 RTE_LOG(DEBUG, VHOST_CONFIG,
593 "(%d) failed to map avail ring.\n",
598 len = sizeof(struct vring_used) +
599 sizeof(struct vring_used_elem) * vq->size;
600 vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
601 vq, addr->used_user_addr, &len);
602 if (vq->used == 0 || len != sizeof(struct vring_used) +
603 sizeof(struct vring_used_elem) * vq->size) {
604 RTE_LOG(DEBUG, VHOST_CONFIG,
605 "(%d) failed to map used ring.\n",
610 if (vq->last_used_idx != vq->used->idx) {
611 RTE_LOG(WARNING, VHOST_CONFIG,
612 "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
613 "some packets maybe resent for Tx and dropped for Rx\n",
614 vq->last_used_idx, vq->used->idx);
615 vq->last_used_idx = vq->used->idx;
616 vq->last_avail_idx = vq->used->idx;
619 vq->log_guest_addr = addr->log_guest_addr;
621 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
623 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
624 dev->vid, vq->avail);
625 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
627 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
628 dev->vid, vq->log_guest_addr);
634 * The virtio device sends us the desc, used and avail ring addresses.
635 * This function then converts these to our address space.
638 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
639 int main_fd __rte_unused)
641 struct virtio_net *dev = *pdev;
642 struct vhost_virtqueue *vq;
643 struct vhost_vring_addr *addr = &msg->payload.addr;
645 if (dev->mem == NULL)
646 return VH_RESULT_ERR;
648 /* addr->index refers to the queue index. The txq 1, rxq is 0. */
649 vq = dev->virtqueue[msg->payload.addr.index];
652 * Rings addresses should not be interpreted as long as the ring is not
653 * started and enabled
655 memcpy(&vq->ring_addrs, addr, sizeof(*addr));
657 vring_invalidate(dev, vq);
659 if (vq->enabled && (dev->features &
660 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
661 dev = translate_ring_addresses(dev, msg->payload.addr.index);
663 return VH_RESULT_ERR;
672 * The virtio device sends us the available ring last used index.
675 vhost_user_set_vring_base(struct virtio_net **pdev,
676 struct VhostUserMsg *msg,
677 int main_fd __rte_unused)
679 struct virtio_net *dev = *pdev;
680 dev->virtqueue[msg->payload.state.index]->last_used_idx =
681 msg->payload.state.num;
682 dev->virtqueue[msg->payload.state.index]->last_avail_idx =
683 msg->payload.state.num;
689 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
690 uint64_t host_phys_addr, uint64_t size)
692 struct guest_page *page, *last_page;
694 if (dev->nr_guest_pages == dev->max_guest_pages) {
695 dev->max_guest_pages *= 2;
696 dev->guest_pages = realloc(dev->guest_pages,
697 dev->max_guest_pages * sizeof(*page));
698 if (!dev->guest_pages) {
699 RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
704 if (dev->nr_guest_pages > 0) {
705 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
706 /* merge if the two pages are continuous */
707 if (host_phys_addr == last_page->host_phys_addr +
709 last_page->size += size;
714 page = &dev->guest_pages[dev->nr_guest_pages++];
715 page->guest_phys_addr = guest_phys_addr;
716 page->host_phys_addr = host_phys_addr;
723 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
726 uint64_t reg_size = reg->size;
727 uint64_t host_user_addr = reg->host_user_addr;
728 uint64_t guest_phys_addr = reg->guest_phys_addr;
729 uint64_t host_phys_addr;
732 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
733 size = page_size - (guest_phys_addr & (page_size - 1));
734 size = RTE_MIN(size, reg_size);
736 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
739 host_user_addr += size;
740 guest_phys_addr += size;
743 while (reg_size > 0) {
744 size = RTE_MIN(reg_size, page_size);
745 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
747 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
751 host_user_addr += size;
752 guest_phys_addr += size;
759 #ifdef RTE_LIBRTE_VHOST_DEBUG
760 /* TODO: enable it only in debug mode? */
762 dump_guest_pages(struct virtio_net *dev)
765 struct guest_page *page;
767 for (i = 0; i < dev->nr_guest_pages; i++) {
768 page = &dev->guest_pages[i];
770 RTE_LOG(INFO, VHOST_CONFIG,
771 "guest physical page region %u\n"
772 "\t guest_phys_addr: %" PRIx64 "\n"
773 "\t host_phys_addr : %" PRIx64 "\n"
774 "\t size : %" PRIx64 "\n",
776 page->guest_phys_addr,
777 page->host_phys_addr,
782 #define dump_guest_pages(dev)
786 vhost_memory_changed(struct VhostUserMemory *new,
787 struct rte_vhost_memory *old)
791 if (new->nregions != old->nregions)
794 for (i = 0; i < new->nregions; ++i) {
795 VhostUserMemoryRegion *new_r = &new->regions[i];
796 struct rte_vhost_mem_region *old_r = &old->regions[i];
798 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
800 if (new_r->memory_size != old_r->size)
802 if (new_r->userspace_addr != old_r->guest_user_addr)
810 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
811 int main_fd __rte_unused)
813 struct virtio_net *dev = *pdev;
814 struct VhostUserMemory memory = msg->payload.memory;
815 struct rte_vhost_mem_region *reg;
818 uint64_t mmap_offset;
824 if (memory.nregions > VHOST_MEMORY_MAX_NREGIONS) {
825 RTE_LOG(ERR, VHOST_CONFIG,
826 "too many memory regions (%u)\n", memory.nregions);
827 return VH_RESULT_ERR;
830 if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
831 RTE_LOG(INFO, VHOST_CONFIG,
832 "(%d) memory regions not changed\n", dev->vid);
834 for (i = 0; i < memory.nregions; i++)
841 free_mem_region(dev);
846 /* Flush IOTLB cache as previous HVAs are now invalid */
847 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
848 for (i = 0; i < dev->nr_vring; i++)
849 vhost_user_iotlb_flush_all(dev->virtqueue[i]);
851 dev->nr_guest_pages = 0;
852 if (!dev->guest_pages) {
853 dev->max_guest_pages = 8;
854 dev->guest_pages = malloc(dev->max_guest_pages *
855 sizeof(struct guest_page));
856 if (dev->guest_pages == NULL) {
857 RTE_LOG(ERR, VHOST_CONFIG,
858 "(%d) failed to allocate memory "
859 "for dev->guest_pages\n",
861 return VH_RESULT_ERR;
865 dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
866 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
867 if (dev->mem == NULL) {
868 RTE_LOG(ERR, VHOST_CONFIG,
869 "(%d) failed to allocate memory for dev->mem\n",
871 return VH_RESULT_ERR;
873 dev->mem->nregions = memory.nregions;
875 for (i = 0; i < memory.nregions; i++) {
877 reg = &dev->mem->regions[i];
879 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
880 reg->guest_user_addr = memory.regions[i].userspace_addr;
881 reg->size = memory.regions[i].memory_size;
884 mmap_offset = memory.regions[i].mmap_offset;
886 /* Check for memory_size + mmap_offset overflow */
887 if (mmap_offset >= -reg->size) {
888 RTE_LOG(ERR, VHOST_CONFIG,
889 "mmap_offset (%#"PRIx64") and memory_size "
890 "(%#"PRIx64") overflow\n",
891 mmap_offset, reg->size);
895 mmap_size = reg->size + mmap_offset;
897 /* mmap() without flag of MAP_ANONYMOUS, should be called
898 * with length argument aligned with hugepagesz at older
899 * longterm version Linux, like 2.6.32 and 3.2.72, or
900 * mmap() will fail with EINVAL.
902 * to avoid failure, make sure in caller to keep length
905 alignment = get_blk_size(fd);
906 if (alignment == (uint64_t)-1) {
907 RTE_LOG(ERR, VHOST_CONFIG,
908 "couldn't get hugepage size through fstat\n");
911 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
913 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
914 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
915 MAP_SHARED | populate, fd, 0);
917 if (mmap_addr == MAP_FAILED) {
918 RTE_LOG(ERR, VHOST_CONFIG,
919 "mmap region %u failed.\n", i);
923 reg->mmap_addr = mmap_addr;
924 reg->mmap_size = mmap_size;
925 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
928 if (dev->dequeue_zero_copy)
929 if (add_guest_pages(dev, reg, alignment) < 0) {
930 RTE_LOG(ERR, VHOST_CONFIG,
931 "adding guest pages to region %u failed.\n",
936 RTE_LOG(INFO, VHOST_CONFIG,
937 "guest memory region %u, size: 0x%" PRIx64 "\n"
938 "\t guest physical addr: 0x%" PRIx64 "\n"
939 "\t guest virtual addr: 0x%" PRIx64 "\n"
940 "\t host virtual addr: 0x%" PRIx64 "\n"
941 "\t mmap addr : 0x%" PRIx64 "\n"
942 "\t mmap size : 0x%" PRIx64 "\n"
943 "\t mmap align: 0x%" PRIx64 "\n"
944 "\t mmap off : 0x%" PRIx64 "\n",
946 reg->guest_phys_addr,
947 reg->guest_user_addr,
949 (uint64_t)(uintptr_t)mmap_addr,
955 for (i = 0; i < dev->nr_vring; i++) {
956 struct vhost_virtqueue *vq = dev->virtqueue[i];
958 if (vq->desc || vq->avail || vq->used) {
960 * If the memory table got updated, the ring addresses
961 * need to be translated again as virtual addresses have
964 vring_invalidate(dev, vq);
966 dev = translate_ring_addresses(dev, i);
976 dump_guest_pages(dev);
981 free_mem_region(dev);
984 return VH_RESULT_ERR;
988 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
995 if (vq_is_packed(dev))
996 rings_ok = !!vq->desc_packed;
998 rings_ok = vq->desc && vq->avail && vq->used;
1001 vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1002 vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1006 virtio_is_ready(struct virtio_net *dev)
1008 struct vhost_virtqueue *vq;
1011 if (dev->nr_vring == 0)
1014 for (i = 0; i < dev->nr_vring; i++) {
1015 vq = dev->virtqueue[i];
1017 if (!vq_is_ready(dev, vq))
1021 RTE_LOG(INFO, VHOST_CONFIG,
1022 "virtio is now ready for processing.\n");
1027 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1028 int main_fd __rte_unused)
1030 struct virtio_net *dev = *pdev;
1031 struct vhost_vring_file file;
1032 struct vhost_virtqueue *vq;
1034 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1035 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1036 file.fd = VIRTIO_INVALID_EVENTFD;
1038 file.fd = msg->fds[0];
1039 RTE_LOG(INFO, VHOST_CONFIG,
1040 "vring call idx:%d file:%d\n", file.index, file.fd);
1042 vq = dev->virtqueue[file.index];
1043 if (vq->callfd >= 0)
1046 vq->callfd = file.fd;
1048 return VH_RESULT_OK;
1051 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1052 struct VhostUserMsg *msg,
1053 int main_fd __rte_unused)
1055 if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1057 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1059 return VH_RESULT_OK;
1063 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1064 int main_fd __rte_unused)
1066 struct virtio_net *dev = *pdev;
1067 struct vhost_vring_file file;
1068 struct vhost_virtqueue *vq;
1070 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1071 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1072 file.fd = VIRTIO_INVALID_EVENTFD;
1074 file.fd = msg->fds[0];
1075 RTE_LOG(INFO, VHOST_CONFIG,
1076 "vring kick idx:%d file:%d\n", file.index, file.fd);
1078 /* Interpret ring addresses only when ring is started. */
1079 dev = translate_ring_addresses(dev, file.index);
1081 return VH_RESULT_ERR;
1085 vq = dev->virtqueue[file.index];
1088 * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1089 * the ring starts already enabled. Otherwise, it is enabled via
1090 * the SET_VRING_ENABLE message.
1092 if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
1095 if (vq->kickfd >= 0)
1097 vq->kickfd = file.fd;
1099 return VH_RESULT_OK;
1103 free_zmbufs(struct vhost_virtqueue *vq)
1105 struct zcopy_mbuf *zmbuf, *next;
1107 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1108 zmbuf != NULL; zmbuf = next) {
1109 next = TAILQ_NEXT(zmbuf, next);
1111 rte_pktmbuf_free(zmbuf->mbuf);
1112 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1115 rte_free(vq->zmbufs);
1119 * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1122 vhost_user_get_vring_base(struct virtio_net **pdev,
1123 struct VhostUserMsg *msg,
1124 int main_fd __rte_unused)
1126 struct virtio_net *dev = *pdev;
1127 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1129 /* We have to stop the queue (virtio) if it is running. */
1130 vhost_destroy_device_notify(dev);
1132 dev->flags &= ~VIRTIO_DEV_READY;
1133 dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1135 /* Here we are safe to get the last avail index */
1136 msg->payload.state.num = vq->last_avail_idx;
1138 RTE_LOG(INFO, VHOST_CONFIG,
1139 "vring base idx:%d file:%d\n", msg->payload.state.index,
1140 msg->payload.state.num);
1142 * Based on current qemu vhost-user implementation, this message is
1143 * sent and only sent in vhost_vring_stop.
1144 * TODO: cleanup the vring, it isn't usable since here.
1146 if (vq->kickfd >= 0)
1149 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1151 if (vq->callfd >= 0)
1154 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1156 if (dev->dequeue_zero_copy)
1158 if (vq_is_packed(dev)) {
1159 rte_free(vq->shadow_used_packed);
1160 vq->shadow_used_packed = NULL;
1162 rte_free(vq->shadow_used_split);
1163 vq->shadow_used_split = NULL;
1166 rte_free(vq->batch_copy_elems);
1167 vq->batch_copy_elems = NULL;
1169 msg->size = sizeof(msg->payload.state);
1172 return VH_RESULT_REPLY;
1176 * when virtio queues are ready to work, qemu will send us to
1177 * enable the virtio queue pair.
1180 vhost_user_set_vring_enable(struct virtio_net **pdev,
1181 struct VhostUserMsg *msg,
1182 int main_fd __rte_unused)
1184 struct virtio_net *dev = *pdev;
1185 int enable = (int)msg->payload.state.num;
1186 int index = (int)msg->payload.state.index;
1187 struct rte_vdpa_device *vdpa_dev;
1190 RTE_LOG(INFO, VHOST_CONFIG,
1191 "set queue enable: %d to qp idx: %d\n",
1194 did = dev->vdpa_dev_id;
1195 vdpa_dev = rte_vdpa_get_device(did);
1196 if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1197 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1199 if (dev->notify_ops->vring_state_changed)
1200 dev->notify_ops->vring_state_changed(dev->vid,
1203 dev->virtqueue[index]->enabled = enable;
1205 return VH_RESULT_OK;
1209 vhost_user_get_protocol_features(struct virtio_net **pdev,
1210 struct VhostUserMsg *msg,
1211 int main_fd __rte_unused)
1213 struct virtio_net *dev = *pdev;
1214 uint64_t features, protocol_features;
1216 rte_vhost_driver_get_features(dev->ifname, &features);
1217 rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1220 * REPLY_ACK protocol feature is only mandatory for now
1221 * for IOMMU feature. If IOMMU is explicitly disabled by the
1222 * application, disable also REPLY_ACK feature for older buggy
1223 * Qemu versions (from v2.7.0 to v2.9.0).
1225 if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1226 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1228 msg->payload.u64 = protocol_features;
1229 msg->size = sizeof(msg->payload.u64);
1232 return VH_RESULT_REPLY;
1236 vhost_user_set_protocol_features(struct virtio_net **pdev,
1237 struct VhostUserMsg *msg,
1238 int main_fd __rte_unused)
1240 struct virtio_net *dev = *pdev;
1241 uint64_t protocol_features = msg->payload.u64;
1242 if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES) {
1243 RTE_LOG(ERR, VHOST_CONFIG,
1244 "(%d) received invalid protocol features.\n",
1246 return VH_RESULT_ERR;
1249 dev->protocol_features = protocol_features;
1251 return VH_RESULT_OK;
1255 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1256 int main_fd __rte_unused)
1258 struct virtio_net *dev = *pdev;
1259 int fd = msg->fds[0];
1264 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1265 return VH_RESULT_ERR;
1268 if (msg->size != sizeof(VhostUserLog)) {
1269 RTE_LOG(ERR, VHOST_CONFIG,
1270 "invalid log base msg size: %"PRId32" != %d\n",
1271 msg->size, (int)sizeof(VhostUserLog));
1272 return VH_RESULT_ERR;
1275 size = msg->payload.log.mmap_size;
1276 off = msg->payload.log.mmap_offset;
1278 /* Don't allow mmap_offset to point outside the mmap region */
1280 RTE_LOG(ERR, VHOST_CONFIG,
1281 "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1283 return VH_RESULT_ERR;
1286 RTE_LOG(INFO, VHOST_CONFIG,
1287 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1291 * mmap from 0 to workaround a hugepage mmap bug: mmap will
1292 * fail when offset is not page size aligned.
1294 addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1296 if (addr == MAP_FAILED) {
1297 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1298 return VH_RESULT_ERR;
1302 * Free previously mapped log memory on occasionally
1303 * multiple VHOST_USER_SET_LOG_BASE.
1305 if (dev->log_addr) {
1306 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1308 dev->log_addr = (uint64_t)(uintptr_t)addr;
1309 dev->log_base = dev->log_addr + off;
1310 dev->log_size = size;
1313 * The spec is not clear about it (yet), but QEMU doesn't expect
1314 * any payload in the reply.
1319 return VH_RESULT_REPLY;
1322 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1323 struct VhostUserMsg *msg,
1324 int main_fd __rte_unused)
1327 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1329 return VH_RESULT_OK;
1333 * An rarp packet is constructed and broadcasted to notify switches about
1334 * the new location of the migrated VM, so that packets from outside will
1335 * not be lost after migration.
1337 * However, we don't actually "send" a rarp packet here, instead, we set
1338 * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1341 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1342 int main_fd __rte_unused)
1344 struct virtio_net *dev = *pdev;
1345 uint8_t *mac = (uint8_t *)&msg->payload.u64;
1346 struct rte_vdpa_device *vdpa_dev;
1349 RTE_LOG(DEBUG, VHOST_CONFIG,
1350 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1351 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1352 memcpy(dev->mac.addr_bytes, mac, 6);
1355 * Set the flag to inject a RARP broadcast packet at
1356 * rte_vhost_dequeue_burst().
1358 * rte_smp_wmb() is for making sure the mac is copied
1359 * before the flag is set.
1362 rte_atomic16_set(&dev->broadcast_rarp, 1);
1363 did = dev->vdpa_dev_id;
1364 vdpa_dev = rte_vdpa_get_device(did);
1365 if (vdpa_dev && vdpa_dev->ops->migration_done)
1366 vdpa_dev->ops->migration_done(dev->vid);
1368 return VH_RESULT_OK;
1372 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1373 int main_fd __rte_unused)
1375 struct virtio_net *dev = *pdev;
1376 if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1377 msg->payload.u64 > VIRTIO_MAX_MTU) {
1378 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1381 return VH_RESULT_ERR;
1384 dev->mtu = msg->payload.u64;
1386 return VH_RESULT_OK;
1390 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1391 int main_fd __rte_unused)
1393 struct virtio_net *dev = *pdev;
1394 int fd = msg->fds[0];
1397 RTE_LOG(ERR, VHOST_CONFIG,
1398 "Invalid file descriptor for slave channel (%d)\n",
1400 return VH_RESULT_ERR;
1403 dev->slave_req_fd = fd;
1405 return VH_RESULT_OK;
1409 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1411 struct vhost_vring_addr *ra;
1412 uint64_t start, end;
1415 end = start + imsg->size;
1417 ra = &vq->ring_addrs;
1418 if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1420 if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1422 if (ra->used_user_addr >= start && ra->used_user_addr < end)
1429 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1430 struct vhost_iotlb_msg *imsg)
1432 uint64_t istart, iend, vstart, vend;
1434 istart = imsg->iova;
1435 iend = istart + imsg->size - 1;
1437 vstart = (uintptr_t)vq->desc;
1438 vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1439 if (vstart <= iend && istart <= vend)
1442 vstart = (uintptr_t)vq->avail;
1443 vend = vstart + sizeof(struct vring_avail);
1444 vend += sizeof(uint16_t) * vq->size - 1;
1445 if (vstart <= iend && istart <= vend)
1448 vstart = (uintptr_t)vq->used;
1449 vend = vstart + sizeof(struct vring_used);
1450 vend += sizeof(struct vring_used_elem) * vq->size - 1;
1451 if (vstart <= iend && istart <= vend)
1458 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1459 int main_fd __rte_unused)
1461 struct virtio_net *dev = *pdev;
1462 struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1466 switch (imsg->type) {
1467 case VHOST_IOTLB_UPDATE:
1469 vva = qva_to_vva(dev, imsg->uaddr, &len);
1471 return VH_RESULT_ERR;
1473 for (i = 0; i < dev->nr_vring; i++) {
1474 struct vhost_virtqueue *vq = dev->virtqueue[i];
1476 vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1479 if (is_vring_iotlb_update(vq, imsg))
1480 *pdev = dev = translate_ring_addresses(dev, i);
1483 case VHOST_IOTLB_INVALIDATE:
1484 for (i = 0; i < dev->nr_vring; i++) {
1485 struct vhost_virtqueue *vq = dev->virtqueue[i];
1487 vhost_user_iotlb_cache_remove(vq, imsg->iova,
1490 if (is_vring_iotlb_invalidate(vq, imsg))
1491 vring_invalidate(dev, vq);
1495 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1497 return VH_RESULT_ERR;
1500 return VH_RESULT_OK;
1503 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1504 struct VhostUserMsg *msg,
1506 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1507 [VHOST_USER_NONE] = NULL,
1508 [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1509 [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1510 [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1511 [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1512 [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1513 [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1514 [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1515 [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1516 [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1517 [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1518 [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1519 [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1520 [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1521 [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1522 [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1523 [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1524 [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1525 [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1526 [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1527 [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1528 [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1529 [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1533 /* return bytes# of read on success or negative val on failure. */
1535 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1539 ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1540 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1544 if (msg && msg->size) {
1545 if (msg->size > sizeof(msg->payload)) {
1546 RTE_LOG(ERR, VHOST_CONFIG,
1547 "invalid msg size: %d\n", msg->size);
1550 ret = read(sockfd, &msg->payload, msg->size);
1553 if (ret != (int)msg->size) {
1554 RTE_LOG(ERR, VHOST_CONFIG,
1555 "read control message failed\n");
1564 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1569 return send_fd_message(sockfd, (char *)msg,
1570 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1574 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1579 msg->flags &= ~VHOST_USER_VERSION_MASK;
1580 msg->flags &= ~VHOST_USER_NEED_REPLY;
1581 msg->flags |= VHOST_USER_VERSION;
1582 msg->flags |= VHOST_USER_REPLY_MASK;
1584 return send_vhost_message(sockfd, msg);
1588 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1592 if (msg->flags & VHOST_USER_NEED_REPLY)
1593 rte_spinlock_lock(&dev->slave_req_lock);
1595 ret = send_vhost_message(dev->slave_req_fd, msg);
1596 if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1597 rte_spinlock_unlock(&dev->slave_req_lock);
1603 * Allocate a queue pair if it hasn't been allocated yet
1606 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1607 struct VhostUserMsg *msg)
1611 switch (msg->request.master) {
1612 case VHOST_USER_SET_VRING_KICK:
1613 case VHOST_USER_SET_VRING_CALL:
1614 case VHOST_USER_SET_VRING_ERR:
1615 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1617 case VHOST_USER_SET_VRING_NUM:
1618 case VHOST_USER_SET_VRING_BASE:
1619 case VHOST_USER_SET_VRING_ENABLE:
1620 vring_idx = msg->payload.state.index;
1622 case VHOST_USER_SET_VRING_ADDR:
1623 vring_idx = msg->payload.addr.index;
1629 if (vring_idx >= VHOST_MAX_VRING) {
1630 RTE_LOG(ERR, VHOST_CONFIG,
1631 "invalid vring index: %u\n", vring_idx);
1635 if (dev->virtqueue[vring_idx])
1638 return alloc_vring_queue(dev, vring_idx);
1642 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1645 unsigned int vq_num = 0;
1647 while (vq_num < dev->nr_vring) {
1648 struct vhost_virtqueue *vq = dev->virtqueue[i];
1651 rte_spinlock_lock(&vq->access_lock);
1659 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1662 unsigned int vq_num = 0;
1664 while (vq_num < dev->nr_vring) {
1665 struct vhost_virtqueue *vq = dev->virtqueue[i];
1668 rte_spinlock_unlock(&vq->access_lock);
1676 vhost_user_msg_handler(int vid, int fd)
1678 struct virtio_net *dev;
1679 struct VhostUserMsg msg;
1680 struct rte_vdpa_device *vdpa_dev;
1683 int unlock_required = 0;
1684 uint32_t skip_master = 0;
1687 dev = get_device(vid);
1691 if (!dev->notify_ops) {
1692 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1693 if (!dev->notify_ops) {
1694 RTE_LOG(ERR, VHOST_CONFIG,
1695 "failed to get callback ops for driver %s\n",
1701 ret = read_vhost_message(fd, &msg);
1702 if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1704 RTE_LOG(ERR, VHOST_CONFIG,
1705 "vhost read message failed\n");
1707 RTE_LOG(INFO, VHOST_CONFIG,
1708 "vhost peer closed\n");
1710 RTE_LOG(ERR, VHOST_CONFIG,
1711 "vhost read incorrect message\n");
1717 if (msg.request.master != VHOST_USER_IOTLB_MSG)
1718 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1719 vhost_message_str[msg.request.master]);
1721 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1722 vhost_message_str[msg.request.master]);
1724 ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1726 RTE_LOG(ERR, VHOST_CONFIG,
1727 "failed to alloc queue\n");
1732 * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1733 * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1734 * and device is destroyed. destroy_device waits for queues to be
1735 * inactive, so it is safe. Otherwise taking the access_lock
1736 * would cause a dead lock.
1738 switch (msg.request.master) {
1739 case VHOST_USER_SET_FEATURES:
1740 case VHOST_USER_SET_PROTOCOL_FEATURES:
1741 case VHOST_USER_SET_OWNER:
1742 case VHOST_USER_SET_MEM_TABLE:
1743 case VHOST_USER_SET_LOG_BASE:
1744 case VHOST_USER_SET_LOG_FD:
1745 case VHOST_USER_SET_VRING_NUM:
1746 case VHOST_USER_SET_VRING_ADDR:
1747 case VHOST_USER_SET_VRING_BASE:
1748 case VHOST_USER_SET_VRING_KICK:
1749 case VHOST_USER_SET_VRING_CALL:
1750 case VHOST_USER_SET_VRING_ERR:
1751 case VHOST_USER_SET_VRING_ENABLE:
1752 case VHOST_USER_SEND_RARP:
1753 case VHOST_USER_NET_SET_MTU:
1754 case VHOST_USER_SET_SLAVE_REQ_FD:
1755 vhost_user_lock_all_queue_pairs(dev);
1756 unlock_required = 1;
1763 if (dev->extern_ops.pre_msg_handle) {
1764 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
1765 (void *)&msg, &skip_master);
1766 if (ret == VH_RESULT_ERR)
1768 else if (ret == VH_RESULT_REPLY)
1769 send_vhost_reply(fd, &msg);
1772 goto skip_to_post_handle;
1775 request = msg.request.master;
1776 if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
1777 if (!vhost_message_handlers[request])
1778 goto skip_to_post_handle;
1779 ret = vhost_message_handlers[request](&dev, &msg, fd);
1783 RTE_LOG(ERR, VHOST_CONFIG,
1784 "Processing %s failed.\n",
1785 vhost_message_str[request]);
1788 RTE_LOG(DEBUG, VHOST_CONFIG,
1789 "Processing %s succeeded.\n",
1790 vhost_message_str[request]);
1792 case VH_RESULT_REPLY:
1793 RTE_LOG(DEBUG, VHOST_CONFIG,
1794 "Processing %s succeeded and needs reply.\n",
1795 vhost_message_str[request]);
1796 send_vhost_reply(fd, &msg);
1800 RTE_LOG(ERR, VHOST_CONFIG,
1801 "Requested invalid message type %d.\n", request);
1802 ret = VH_RESULT_ERR;
1805 skip_to_post_handle:
1806 if (ret != VH_RESULT_ERR && dev->extern_ops.post_msg_handle) {
1807 ret = (*dev->extern_ops.post_msg_handle)(
1808 dev->vid, (void *)&msg);
1809 if (ret == VH_RESULT_ERR)
1811 else if (ret == VH_RESULT_REPLY)
1812 send_vhost_reply(fd, &msg);
1816 if (unlock_required)
1817 vhost_user_unlock_all_queue_pairs(dev);
1820 * If the request required a reply that was already sent,
1821 * this optional reply-ack won't be sent as the
1822 * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
1824 if (msg.flags & VHOST_USER_NEED_REPLY) {
1825 msg.payload.u64 = ret == VH_RESULT_ERR;
1826 msg.size = sizeof(msg.payload.u64);
1828 send_vhost_reply(fd, &msg);
1829 } else if (ret == VH_RESULT_ERR) {
1830 RTE_LOG(ERR, VHOST_CONFIG,
1831 "vhost message handling failed.\n");
1835 if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1836 dev->flags |= VIRTIO_DEV_READY;
1838 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1839 if (dev->dequeue_zero_copy) {
1840 RTE_LOG(INFO, VHOST_CONFIG,
1841 "dequeue zero copy is enabled\n");
1844 if (dev->notify_ops->new_device(dev->vid) == 0)
1845 dev->flags |= VIRTIO_DEV_RUNNING;
1849 did = dev->vdpa_dev_id;
1850 vdpa_dev = rte_vdpa_get_device(did);
1851 if (vdpa_dev && virtio_is_ready(dev) &&
1852 !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
1853 msg.request.master == VHOST_USER_SET_VRING_ENABLE) {
1854 if (vdpa_dev->ops->dev_conf)
1855 vdpa_dev->ops->dev_conf(dev->vid);
1856 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
1857 if (vhost_user_host_notifier_ctrl(dev->vid, true) != 0) {
1858 RTE_LOG(INFO, VHOST_CONFIG,
1859 "(%d) software relay is used for vDPA, performance may be low.\n",
1867 static int process_slave_message_reply(struct virtio_net *dev,
1868 const struct VhostUserMsg *msg)
1870 struct VhostUserMsg msg_reply;
1873 if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
1876 if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
1881 if (msg_reply.request.slave != msg->request.slave) {
1882 RTE_LOG(ERR, VHOST_CONFIG,
1883 "Received unexpected msg type (%u), expected %u\n",
1884 msg_reply.request.slave, msg->request.slave);
1889 ret = msg_reply.payload.u64 ? -1 : 0;
1892 rte_spinlock_unlock(&dev->slave_req_lock);
1897 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
1900 struct VhostUserMsg msg = {
1901 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
1902 .flags = VHOST_USER_VERSION,
1903 .size = sizeof(msg.payload.iotlb),
1907 .type = VHOST_IOTLB_MISS,
1911 ret = send_vhost_message(dev->slave_req_fd, &msg);
1913 RTE_LOG(ERR, VHOST_CONFIG,
1914 "Failed to send IOTLB miss message (%d)\n",
1922 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
1928 struct VhostUserMsg msg = {
1929 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
1930 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
1931 .size = sizeof(msg.payload.area),
1933 .u64 = index & VHOST_USER_VRING_IDX_MASK,
1940 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
1946 ret = send_vhost_slave_message(dev, &msg);
1948 RTE_LOG(ERR, VHOST_CONFIG,
1949 "Failed to set host notifier (%d)\n", ret);
1953 return process_slave_message_reply(dev, &msg);
1956 int vhost_user_host_notifier_ctrl(int vid, bool enable)
1958 struct virtio_net *dev;
1959 struct rte_vdpa_device *vdpa_dev;
1960 int vfio_device_fd, did, ret = 0;
1961 uint64_t offset, size;
1964 dev = get_device(vid);
1968 did = dev->vdpa_dev_id;
1972 if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
1973 !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
1974 !(dev->protocol_features &
1975 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
1976 !(dev->protocol_features &
1977 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
1978 !(dev->protocol_features &
1979 (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
1982 vdpa_dev = rte_vdpa_get_device(did);
1986 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
1987 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
1989 vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
1990 if (vfio_device_fd < 0)
1994 for (i = 0; i < dev->nr_vring; i++) {
1995 if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2001 if (vhost_user_slave_set_vring_host_notifier(dev, i,
2002 vfio_device_fd, offset, size) < 0) {
2009 for (i = 0; i < dev->nr_vring; i++) {
2010 vhost_user_slave_set_vring_host_notifier(dev, i, -1,