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",
80 [VHOST_USER_POSTCOPY_END] = "VHOST_USER_POSTCOPY_END",
83 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
84 static int read_vhost_message(int sockfd, struct VhostUserMsg *msg);
92 ret = fstat(fd, &stat);
93 return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
97 * Reclaim all the outstanding zmbufs for a virtqueue.
100 drain_zmbuf_list(struct vhost_virtqueue *vq)
102 struct zcopy_mbuf *zmbuf, *next;
104 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
105 zmbuf != NULL; zmbuf = next) {
106 next = TAILQ_NEXT(zmbuf, next);
108 while (!mbuf_is_consumed(zmbuf->mbuf))
111 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
112 restore_mbuf(zmbuf->mbuf);
113 rte_pktmbuf_free(zmbuf->mbuf);
120 free_mem_region(struct virtio_net *dev)
123 struct rte_vhost_mem_region *reg;
124 struct vhost_virtqueue *vq;
126 if (!dev || !dev->mem)
129 if (dev->dequeue_zero_copy) {
130 for (i = 0; i < dev->nr_vring; i++) {
131 vq = dev->virtqueue[i];
133 drain_zmbuf_list(vq);
137 for (i = 0; i < dev->mem->nregions; i++) {
138 reg = &dev->mem->regions[i];
139 if (reg->host_user_addr) {
140 munmap(reg->mmap_addr, reg->mmap_size);
147 vhost_backend_cleanup(struct virtio_net *dev)
150 free_mem_region(dev);
155 free(dev->guest_pages);
156 dev->guest_pages = NULL;
159 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
163 if (dev->slave_req_fd >= 0) {
164 close(dev->slave_req_fd);
165 dev->slave_req_fd = -1;
168 if (dev->postcopy_ufd >= 0) {
169 close(dev->postcopy_ufd);
170 dev->postcopy_ufd = -1;
173 dev->postcopy_listening = 0;
177 * This function just returns success at the moment unless
178 * the device hasn't been initialised.
181 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
182 struct VhostUserMsg *msg __rte_unused,
183 int main_fd __rte_unused)
185 return RTE_VHOST_MSG_RESULT_OK;
189 vhost_user_reset_owner(struct virtio_net **pdev,
190 struct VhostUserMsg *msg __rte_unused,
191 int main_fd __rte_unused)
193 struct virtio_net *dev = *pdev;
194 vhost_destroy_device_notify(dev);
196 cleanup_device(dev, 0);
198 return RTE_VHOST_MSG_RESULT_OK;
202 * The features that we support are requested.
205 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
206 int main_fd __rte_unused)
208 struct virtio_net *dev = *pdev;
209 uint64_t features = 0;
211 rte_vhost_driver_get_features(dev->ifname, &features);
213 msg->payload.u64 = features;
214 msg->size = sizeof(msg->payload.u64);
217 return RTE_VHOST_MSG_RESULT_REPLY;
221 * The queue number that we support are requested.
224 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
225 int main_fd __rte_unused)
227 struct virtio_net *dev = *pdev;
228 uint32_t queue_num = 0;
230 rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
232 msg->payload.u64 = (uint64_t)queue_num;
233 msg->size = sizeof(msg->payload.u64);
236 return RTE_VHOST_MSG_RESULT_REPLY;
240 * We receive the negotiated features supported by us and the virtio device.
243 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
244 int main_fd __rte_unused)
246 struct virtio_net *dev = *pdev;
247 uint64_t features = msg->payload.u64;
248 uint64_t vhost_features = 0;
249 struct rte_vdpa_device *vdpa_dev;
252 rte_vhost_driver_get_features(dev->ifname, &vhost_features);
253 if (features & ~vhost_features) {
254 RTE_LOG(ERR, VHOST_CONFIG,
255 "(%d) received invalid negotiated features.\n",
257 return RTE_VHOST_MSG_RESULT_ERR;
260 if (dev->flags & VIRTIO_DEV_RUNNING) {
261 if (dev->features == features)
262 return RTE_VHOST_MSG_RESULT_OK;
265 * Error out if master tries to change features while device is
266 * in running state. The exception being VHOST_F_LOG_ALL, which
267 * is enabled when the live-migration starts.
269 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
270 RTE_LOG(ERR, VHOST_CONFIG,
271 "(%d) features changed while device is running.\n",
273 return RTE_VHOST_MSG_RESULT_ERR;
276 if (dev->notify_ops->features_changed)
277 dev->notify_ops->features_changed(dev->vid, features);
280 dev->features = features;
282 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
283 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
285 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
287 VHOST_LOG_DEBUG(VHOST_CONFIG,
288 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
290 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
291 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
293 if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
294 !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
296 * Remove all but first queue pair if MQ hasn't been
297 * negotiated. This is safe because the device is not
298 * running at this stage.
300 while (dev->nr_vring > 2) {
301 struct vhost_virtqueue *vq;
303 vq = dev->virtqueue[--dev->nr_vring];
307 dev->virtqueue[dev->nr_vring] = NULL;
313 did = dev->vdpa_dev_id;
314 vdpa_dev = rte_vdpa_get_device(did);
315 if (vdpa_dev && vdpa_dev->ops->set_features)
316 vdpa_dev->ops->set_features(dev->vid);
318 return RTE_VHOST_MSG_RESULT_OK;
322 * The virtio device sends us the size of the descriptor ring.
325 vhost_user_set_vring_num(struct virtio_net **pdev,
326 struct VhostUserMsg *msg,
327 int main_fd __rte_unused)
329 struct virtio_net *dev = *pdev;
330 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
332 vq->size = msg->payload.state.num;
334 /* VIRTIO 1.0, 2.4 Virtqueues says:
336 * Queue Size value is always a power of 2. The maximum Queue Size
339 if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
340 RTE_LOG(ERR, VHOST_CONFIG,
341 "invalid virtqueue size %u\n", vq->size);
342 return RTE_VHOST_MSG_RESULT_ERR;
345 if (dev->dequeue_zero_copy) {
347 vq->last_zmbuf_idx = 0;
348 vq->zmbuf_size = vq->size;
349 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
350 sizeof(struct zcopy_mbuf), 0);
351 if (vq->zmbufs == NULL) {
352 RTE_LOG(WARNING, VHOST_CONFIG,
353 "failed to allocate mem for zero copy; "
354 "zero copy is force disabled\n");
355 dev->dequeue_zero_copy = 0;
357 TAILQ_INIT(&vq->zmbuf_list);
360 if (vq_is_packed(dev)) {
361 vq->shadow_used_packed = rte_malloc(NULL,
363 sizeof(struct vring_used_elem_packed),
364 RTE_CACHE_LINE_SIZE);
365 if (!vq->shadow_used_packed) {
366 RTE_LOG(ERR, VHOST_CONFIG,
367 "failed to allocate memory for shadow used ring.\n");
368 return RTE_VHOST_MSG_RESULT_ERR;
372 vq->shadow_used_split = rte_malloc(NULL,
373 vq->size * sizeof(struct vring_used_elem),
374 RTE_CACHE_LINE_SIZE);
375 if (!vq->shadow_used_split) {
376 RTE_LOG(ERR, VHOST_CONFIG,
377 "failed to allocate memory for shadow used ring.\n");
378 return RTE_VHOST_MSG_RESULT_ERR;
382 vq->batch_copy_elems = rte_malloc(NULL,
383 vq->size * sizeof(struct batch_copy_elem),
384 RTE_CACHE_LINE_SIZE);
385 if (!vq->batch_copy_elems) {
386 RTE_LOG(ERR, VHOST_CONFIG,
387 "failed to allocate memory for batching copy.\n");
388 return RTE_VHOST_MSG_RESULT_ERR;
391 return RTE_VHOST_MSG_RESULT_OK;
395 * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
396 * same numa node as the memory of vring descriptor.
398 #ifdef RTE_LIBRTE_VHOST_NUMA
399 static struct virtio_net*
400 numa_realloc(struct virtio_net *dev, int index)
402 int oldnode, newnode;
403 struct virtio_net *old_dev;
404 struct vhost_virtqueue *old_vq, *vq;
405 struct zcopy_mbuf *new_zmbuf;
406 struct vring_used_elem *new_shadow_used_split;
407 struct vring_used_elem_packed *new_shadow_used_packed;
408 struct batch_copy_elem *new_batch_copy_elems;
412 vq = old_vq = dev->virtqueue[index];
414 ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
415 MPOL_F_NODE | MPOL_F_ADDR);
417 /* check if we need to reallocate vq */
418 ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
419 MPOL_F_NODE | MPOL_F_ADDR);
421 RTE_LOG(ERR, VHOST_CONFIG,
422 "Unable to get vq numa information.\n");
425 if (oldnode != newnode) {
426 RTE_LOG(INFO, VHOST_CONFIG,
427 "reallocate vq from %d to %d node\n", oldnode, newnode);
428 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
432 memcpy(vq, old_vq, sizeof(*vq));
433 TAILQ_INIT(&vq->zmbuf_list);
435 if (dev->dequeue_zero_copy) {
436 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
437 sizeof(struct zcopy_mbuf), 0, newnode);
439 rte_free(vq->zmbufs);
440 vq->zmbufs = new_zmbuf;
444 if (vq_is_packed(dev)) {
445 new_shadow_used_packed = rte_malloc_socket(NULL,
447 sizeof(struct vring_used_elem_packed),
450 if (new_shadow_used_packed) {
451 rte_free(vq->shadow_used_packed);
452 vq->shadow_used_packed = new_shadow_used_packed;
455 new_shadow_used_split = rte_malloc_socket(NULL,
457 sizeof(struct vring_used_elem),
460 if (new_shadow_used_split) {
461 rte_free(vq->shadow_used_split);
462 vq->shadow_used_split = new_shadow_used_split;
466 new_batch_copy_elems = rte_malloc_socket(NULL,
467 vq->size * sizeof(struct batch_copy_elem),
470 if (new_batch_copy_elems) {
471 rte_free(vq->batch_copy_elems);
472 vq->batch_copy_elems = new_batch_copy_elems;
478 /* check if we need to reallocate dev */
479 ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
480 MPOL_F_NODE | MPOL_F_ADDR);
482 RTE_LOG(ERR, VHOST_CONFIG,
483 "Unable to get dev numa information.\n");
486 if (oldnode != newnode) {
487 RTE_LOG(INFO, VHOST_CONFIG,
488 "reallocate dev from %d to %d node\n",
490 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
496 memcpy(dev, old_dev, sizeof(*dev));
501 dev->virtqueue[index] = vq;
502 vhost_devices[dev->vid] = dev;
505 vhost_user_iotlb_init(dev, index);
510 static struct virtio_net*
511 numa_realloc(struct virtio_net *dev, int index __rte_unused)
517 /* Converts QEMU virtual address to Vhost virtual address. */
519 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
521 struct rte_vhost_mem_region *r;
524 if (unlikely(!dev || !dev->mem))
527 /* Find the region where the address lives. */
528 for (i = 0; i < dev->mem->nregions; i++) {
529 r = &dev->mem->regions[i];
531 if (qva >= r->guest_user_addr &&
532 qva < r->guest_user_addr + r->size) {
534 if (unlikely(*len > r->guest_user_addr + r->size - qva))
535 *len = r->guest_user_addr + r->size - qva;
537 return qva - r->guest_user_addr +
549 * Converts ring address to Vhost virtual address.
550 * If IOMMU is enabled, the ring address is a guest IO virtual address,
551 * else it is a QEMU virtual address.
554 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
555 uint64_t ra, uint64_t *size)
557 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
560 vva = vhost_user_iotlb_cache_find(vq, ra,
561 size, VHOST_ACCESS_RW);
563 vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
568 return qva_to_vva(dev, ra, size);
571 static struct virtio_net *
572 translate_ring_addresses(struct virtio_net *dev, int vq_index)
574 struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
575 struct vhost_vring_addr *addr = &vq->ring_addrs;
576 uint64_t len, expected_len;
578 if (vq_is_packed(dev)) {
579 len = sizeof(struct vring_packed_desc) * vq->size;
580 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
581 ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
582 vq->log_guest_addr = 0;
583 if (vq->desc_packed == NULL ||
584 len != sizeof(struct vring_packed_desc) *
586 RTE_LOG(DEBUG, VHOST_CONFIG,
587 "(%d) failed to map desc_packed ring.\n",
592 dev = numa_realloc(dev, vq_index);
593 vq = dev->virtqueue[vq_index];
594 addr = &vq->ring_addrs;
596 len = sizeof(struct vring_packed_desc_event);
597 vq->driver_event = (struct vring_packed_desc_event *)
598 (uintptr_t)ring_addr_to_vva(dev,
599 vq, addr->avail_user_addr, &len);
600 if (vq->driver_event == NULL ||
601 len != sizeof(struct vring_packed_desc_event)) {
602 RTE_LOG(DEBUG, VHOST_CONFIG,
603 "(%d) failed to find driver area address.\n",
608 len = sizeof(struct vring_packed_desc_event);
609 vq->device_event = (struct vring_packed_desc_event *)
610 (uintptr_t)ring_addr_to_vva(dev,
611 vq, addr->used_user_addr, &len);
612 if (vq->device_event == NULL ||
613 len != sizeof(struct vring_packed_desc_event)) {
614 RTE_LOG(DEBUG, VHOST_CONFIG,
615 "(%d) failed to find device area address.\n",
623 /* The addresses are converted from QEMU virtual to Vhost virtual. */
624 if (vq->desc && vq->avail && vq->used)
627 len = sizeof(struct vring_desc) * vq->size;
628 vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
629 vq, addr->desc_user_addr, &len);
630 if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
631 RTE_LOG(DEBUG, VHOST_CONFIG,
632 "(%d) failed to map desc ring.\n",
637 dev = numa_realloc(dev, vq_index);
638 vq = dev->virtqueue[vq_index];
639 addr = &vq->ring_addrs;
641 len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
642 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
643 len += sizeof(uint16_t);
645 vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
646 vq, addr->avail_user_addr, &len);
647 if (vq->avail == 0 || len != expected_len) {
648 RTE_LOG(DEBUG, VHOST_CONFIG,
649 "(%d) failed to map avail ring.\n",
654 len = sizeof(struct vring_used) +
655 sizeof(struct vring_used_elem) * vq->size;
656 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
657 len += sizeof(uint16_t);
659 vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
660 vq, addr->used_user_addr, &len);
661 if (vq->used == 0 || len != expected_len) {
662 RTE_LOG(DEBUG, VHOST_CONFIG,
663 "(%d) failed to map used ring.\n",
668 if (vq->last_used_idx != vq->used->idx) {
669 RTE_LOG(WARNING, VHOST_CONFIG,
670 "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
671 "some packets maybe resent for Tx and dropped for Rx\n",
672 vq->last_used_idx, vq->used->idx);
673 vq->last_used_idx = vq->used->idx;
674 vq->last_avail_idx = vq->used->idx;
677 vq->log_guest_addr = addr->log_guest_addr;
679 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
681 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
682 dev->vid, vq->avail);
683 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
685 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
686 dev->vid, vq->log_guest_addr);
692 * The virtio device sends us the desc, used and avail ring addresses.
693 * This function then converts these to our address space.
696 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
697 int main_fd __rte_unused)
699 struct virtio_net *dev = *pdev;
700 struct vhost_virtqueue *vq;
701 struct vhost_vring_addr *addr = &msg->payload.addr;
703 if (dev->mem == NULL)
704 return RTE_VHOST_MSG_RESULT_ERR;
706 /* addr->index refers to the queue index. The txq 1, rxq is 0. */
707 vq = dev->virtqueue[msg->payload.addr.index];
710 * Rings addresses should not be interpreted as long as the ring is not
711 * started and enabled
713 memcpy(&vq->ring_addrs, addr, sizeof(*addr));
715 vring_invalidate(dev, vq);
717 if (vq->enabled && (dev->features &
718 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
719 dev = translate_ring_addresses(dev, msg->payload.addr.index);
721 return RTE_VHOST_MSG_RESULT_ERR;
726 return RTE_VHOST_MSG_RESULT_OK;
730 * The virtio device sends us the available ring last used index.
733 vhost_user_set_vring_base(struct virtio_net **pdev,
734 struct VhostUserMsg *msg,
735 int main_fd __rte_unused)
737 struct virtio_net *dev = *pdev;
738 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
739 uint64_t val = msg->payload.state.num;
741 if (vq_is_packed(dev)) {
743 * Bit[0:14]: avail index
744 * Bit[15]: avail wrap counter
746 vq->last_avail_idx = val & 0x7fff;
747 vq->avail_wrap_counter = !!(val & (0x1 << 15));
749 * Set used index to same value as available one, as
750 * their values should be the same since ring processing
751 * was stopped at get time.
753 vq->last_used_idx = vq->last_avail_idx;
754 vq->used_wrap_counter = vq->avail_wrap_counter;
756 vq->last_used_idx = msg->payload.state.num;
757 vq->last_avail_idx = msg->payload.state.num;
760 return RTE_VHOST_MSG_RESULT_OK;
764 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
765 uint64_t host_phys_addr, uint64_t size)
767 struct guest_page *page, *last_page;
768 struct guest_page *old_pages;
770 if (dev->nr_guest_pages == dev->max_guest_pages) {
771 dev->max_guest_pages *= 2;
772 old_pages = dev->guest_pages;
773 dev->guest_pages = realloc(dev->guest_pages,
774 dev->max_guest_pages * sizeof(*page));
775 if (!dev->guest_pages) {
776 RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
782 if (dev->nr_guest_pages > 0) {
783 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
784 /* merge if the two pages are continuous */
785 if (host_phys_addr == last_page->host_phys_addr +
787 last_page->size += size;
792 page = &dev->guest_pages[dev->nr_guest_pages++];
793 page->guest_phys_addr = guest_phys_addr;
794 page->host_phys_addr = host_phys_addr;
801 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
804 uint64_t reg_size = reg->size;
805 uint64_t host_user_addr = reg->host_user_addr;
806 uint64_t guest_phys_addr = reg->guest_phys_addr;
807 uint64_t host_phys_addr;
810 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
811 size = page_size - (guest_phys_addr & (page_size - 1));
812 size = RTE_MIN(size, reg_size);
814 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
817 host_user_addr += size;
818 guest_phys_addr += size;
821 while (reg_size > 0) {
822 size = RTE_MIN(reg_size, page_size);
823 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
825 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
829 host_user_addr += size;
830 guest_phys_addr += size;
837 #ifdef RTE_LIBRTE_VHOST_DEBUG
838 /* TODO: enable it only in debug mode? */
840 dump_guest_pages(struct virtio_net *dev)
843 struct guest_page *page;
845 for (i = 0; i < dev->nr_guest_pages; i++) {
846 page = &dev->guest_pages[i];
848 RTE_LOG(INFO, VHOST_CONFIG,
849 "guest physical page region %u\n"
850 "\t guest_phys_addr: %" PRIx64 "\n"
851 "\t host_phys_addr : %" PRIx64 "\n"
852 "\t size : %" PRIx64 "\n",
854 page->guest_phys_addr,
855 page->host_phys_addr,
860 #define dump_guest_pages(dev)
864 vhost_memory_changed(struct VhostUserMemory *new,
865 struct rte_vhost_memory *old)
869 if (new->nregions != old->nregions)
872 for (i = 0; i < new->nregions; ++i) {
873 VhostUserMemoryRegion *new_r = &new->regions[i];
874 struct rte_vhost_mem_region *old_r = &old->regions[i];
876 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
878 if (new_r->memory_size != old_r->size)
880 if (new_r->userspace_addr != old_r->guest_user_addr)
888 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
891 struct virtio_net *dev = *pdev;
892 struct VhostUserMemory *memory = &msg->payload.memory;
893 struct rte_vhost_mem_region *reg;
896 uint64_t mmap_offset;
902 if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
903 RTE_LOG(ERR, VHOST_CONFIG,
904 "too many memory regions (%u)\n", memory->nregions);
905 return RTE_VHOST_MSG_RESULT_ERR;
908 if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
909 RTE_LOG(INFO, VHOST_CONFIG,
910 "(%d) memory regions not changed\n", dev->vid);
912 for (i = 0; i < memory->nregions; i++)
915 return RTE_VHOST_MSG_RESULT_OK;
919 free_mem_region(dev);
924 /* Flush IOTLB cache as previous HVAs are now invalid */
925 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
926 for (i = 0; i < dev->nr_vring; i++)
927 vhost_user_iotlb_flush_all(dev->virtqueue[i]);
929 dev->nr_guest_pages = 0;
930 if (!dev->guest_pages) {
931 dev->max_guest_pages = 8;
932 dev->guest_pages = malloc(dev->max_guest_pages *
933 sizeof(struct guest_page));
934 if (dev->guest_pages == NULL) {
935 RTE_LOG(ERR, VHOST_CONFIG,
936 "(%d) failed to allocate memory "
937 "for dev->guest_pages\n",
939 return RTE_VHOST_MSG_RESULT_ERR;
943 dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
944 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
945 if (dev->mem == NULL) {
946 RTE_LOG(ERR, VHOST_CONFIG,
947 "(%d) failed to allocate memory for dev->mem\n",
949 return RTE_VHOST_MSG_RESULT_ERR;
951 dev->mem->nregions = memory->nregions;
953 for (i = 0; i < memory->nregions; i++) {
955 reg = &dev->mem->regions[i];
957 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
958 reg->guest_user_addr = memory->regions[i].userspace_addr;
959 reg->size = memory->regions[i].memory_size;
962 mmap_offset = memory->regions[i].mmap_offset;
964 /* Check for memory_size + mmap_offset overflow */
965 if (mmap_offset >= -reg->size) {
966 RTE_LOG(ERR, VHOST_CONFIG,
967 "mmap_offset (%#"PRIx64") and memory_size "
968 "(%#"PRIx64") overflow\n",
969 mmap_offset, reg->size);
973 mmap_size = reg->size + mmap_offset;
975 /* mmap() without flag of MAP_ANONYMOUS, should be called
976 * with length argument aligned with hugepagesz at older
977 * longterm version Linux, like 2.6.32 and 3.2.72, or
978 * mmap() will fail with EINVAL.
980 * to avoid failure, make sure in caller to keep length
983 alignment = get_blk_size(fd);
984 if (alignment == (uint64_t)-1) {
985 RTE_LOG(ERR, VHOST_CONFIG,
986 "couldn't get hugepage size through fstat\n");
989 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
991 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
992 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
993 MAP_SHARED | populate, fd, 0);
995 if (mmap_addr == MAP_FAILED) {
996 RTE_LOG(ERR, VHOST_CONFIG,
997 "mmap region %u failed.\n", i);
1001 reg->mmap_addr = mmap_addr;
1002 reg->mmap_size = mmap_size;
1003 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1006 if (dev->dequeue_zero_copy)
1007 if (add_guest_pages(dev, reg, alignment) < 0) {
1008 RTE_LOG(ERR, VHOST_CONFIG,
1009 "adding guest pages to region %u failed.\n",
1014 RTE_LOG(INFO, VHOST_CONFIG,
1015 "guest memory region %u, size: 0x%" PRIx64 "\n"
1016 "\t guest physical addr: 0x%" PRIx64 "\n"
1017 "\t guest virtual addr: 0x%" PRIx64 "\n"
1018 "\t host virtual addr: 0x%" PRIx64 "\n"
1019 "\t mmap addr : 0x%" PRIx64 "\n"
1020 "\t mmap size : 0x%" PRIx64 "\n"
1021 "\t mmap align: 0x%" PRIx64 "\n"
1022 "\t mmap off : 0x%" PRIx64 "\n",
1024 reg->guest_phys_addr,
1025 reg->guest_user_addr,
1026 reg->host_user_addr,
1027 (uint64_t)(uintptr_t)mmap_addr,
1032 if (dev->postcopy_listening) {
1034 * We haven't a better way right now than sharing
1035 * DPDK's virtual address with Qemu, so that Qemu can
1036 * retrieve the region offset when handling userfaults.
1038 memory->regions[i].userspace_addr =
1039 reg->host_user_addr;
1042 if (dev->postcopy_listening) {
1043 /* Send the addresses back to qemu */
1045 send_vhost_reply(main_fd, msg);
1047 /* Wait for qemu to acknolwedge it's got the addresses
1048 * we've got to wait before we're allowed to generate faults.
1050 VhostUserMsg ack_msg;
1051 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1052 RTE_LOG(ERR, VHOST_CONFIG,
1053 "Failed to read qemu ack on postcopy set-mem-table\n");
1056 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1057 RTE_LOG(ERR, VHOST_CONFIG,
1058 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1059 ack_msg.request.master);
1063 /* Now userfault register and we can use the memory */
1064 for (i = 0; i < memory->nregions; i++) {
1065 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1066 reg = &dev->mem->regions[i];
1067 struct uffdio_register reg_struct;
1070 * Let's register all the mmap'ed area to ensure
1071 * alignment on page boundary.
1073 reg_struct.range.start =
1074 (uint64_t)(uintptr_t)reg->mmap_addr;
1075 reg_struct.range.len = reg->mmap_size;
1076 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1078 if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1080 RTE_LOG(ERR, VHOST_CONFIG,
1081 "Failed to register ufd for region %d: (ufd = %d) %s\n",
1082 i, dev->postcopy_ufd,
1086 RTE_LOG(INFO, VHOST_CONFIG,
1087 "\t userfaultfd registered for range : %llx - %llx\n",
1088 reg_struct.range.start,
1089 reg_struct.range.start +
1090 reg_struct.range.len - 1);
1097 for (i = 0; i < dev->nr_vring; i++) {
1098 struct vhost_virtqueue *vq = dev->virtqueue[i];
1100 if (vq->desc || vq->avail || vq->used) {
1102 * If the memory table got updated, the ring addresses
1103 * need to be translated again as virtual addresses have
1106 vring_invalidate(dev, vq);
1108 dev = translate_ring_addresses(dev, i);
1118 dump_guest_pages(dev);
1120 return RTE_VHOST_MSG_RESULT_OK;
1123 free_mem_region(dev);
1126 return RTE_VHOST_MSG_RESULT_ERR;
1130 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1137 if (vq_is_packed(dev))
1138 rings_ok = !!vq->desc_packed;
1140 rings_ok = vq->desc && vq->avail && vq->used;
1143 vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1144 vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1148 virtio_is_ready(struct virtio_net *dev)
1150 struct vhost_virtqueue *vq;
1153 if (dev->nr_vring == 0)
1156 for (i = 0; i < dev->nr_vring; i++) {
1157 vq = dev->virtqueue[i];
1159 if (!vq_is_ready(dev, vq))
1163 RTE_LOG(INFO, VHOST_CONFIG,
1164 "virtio is now ready for processing.\n");
1169 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1170 int main_fd __rte_unused)
1172 struct virtio_net *dev = *pdev;
1173 struct vhost_vring_file file;
1174 struct vhost_virtqueue *vq;
1176 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1177 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1178 file.fd = VIRTIO_INVALID_EVENTFD;
1180 file.fd = msg->fds[0];
1181 RTE_LOG(INFO, VHOST_CONFIG,
1182 "vring call idx:%d file:%d\n", file.index, file.fd);
1184 vq = dev->virtqueue[file.index];
1185 if (vq->callfd >= 0)
1188 vq->callfd = file.fd;
1190 return RTE_VHOST_MSG_RESULT_OK;
1193 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1194 struct VhostUserMsg *msg,
1195 int main_fd __rte_unused)
1197 if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1199 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1201 return RTE_VHOST_MSG_RESULT_OK;
1205 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1206 int main_fd __rte_unused)
1208 struct virtio_net *dev = *pdev;
1209 struct vhost_vring_file file;
1210 struct vhost_virtqueue *vq;
1212 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1213 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1214 file.fd = VIRTIO_INVALID_EVENTFD;
1216 file.fd = msg->fds[0];
1217 RTE_LOG(INFO, VHOST_CONFIG,
1218 "vring kick idx:%d file:%d\n", file.index, file.fd);
1220 /* Interpret ring addresses only when ring is started. */
1221 dev = translate_ring_addresses(dev, file.index);
1223 return RTE_VHOST_MSG_RESULT_ERR;
1227 vq = dev->virtqueue[file.index];
1230 * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1231 * the ring starts already enabled. Otherwise, it is enabled via
1232 * the SET_VRING_ENABLE message.
1234 if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
1237 if (vq->kickfd >= 0)
1239 vq->kickfd = file.fd;
1241 return RTE_VHOST_MSG_RESULT_OK;
1245 free_zmbufs(struct vhost_virtqueue *vq)
1247 drain_zmbuf_list(vq);
1249 rte_free(vq->zmbufs);
1253 * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1256 vhost_user_get_vring_base(struct virtio_net **pdev,
1257 struct VhostUserMsg *msg,
1258 int main_fd __rte_unused)
1260 struct virtio_net *dev = *pdev;
1261 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1264 /* We have to stop the queue (virtio) if it is running. */
1265 vhost_destroy_device_notify(dev);
1267 dev->flags &= ~VIRTIO_DEV_READY;
1268 dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1270 /* Here we are safe to get the indexes */
1271 if (vq_is_packed(dev)) {
1273 * Bit[0:14]: avail index
1274 * Bit[15]: avail wrap counter
1276 val = vq->last_avail_idx & 0x7fff;
1277 val |= vq->avail_wrap_counter << 15;
1278 msg->payload.state.num = val;
1280 msg->payload.state.num = vq->last_avail_idx;
1283 RTE_LOG(INFO, VHOST_CONFIG,
1284 "vring base idx:%d file:%d\n", msg->payload.state.index,
1285 msg->payload.state.num);
1287 * Based on current qemu vhost-user implementation, this message is
1288 * sent and only sent in vhost_vring_stop.
1289 * TODO: cleanup the vring, it isn't usable since here.
1291 if (vq->kickfd >= 0)
1294 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1296 if (vq->callfd >= 0)
1299 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1301 if (dev->dequeue_zero_copy)
1303 if (vq_is_packed(dev)) {
1304 rte_free(vq->shadow_used_packed);
1305 vq->shadow_used_packed = NULL;
1307 rte_free(vq->shadow_used_split);
1308 vq->shadow_used_split = NULL;
1311 rte_free(vq->batch_copy_elems);
1312 vq->batch_copy_elems = NULL;
1314 msg->size = sizeof(msg->payload.state);
1317 return RTE_VHOST_MSG_RESULT_REPLY;
1321 * when virtio queues are ready to work, qemu will send us to
1322 * enable the virtio queue pair.
1325 vhost_user_set_vring_enable(struct virtio_net **pdev,
1326 struct VhostUserMsg *msg,
1327 int main_fd __rte_unused)
1329 struct virtio_net *dev = *pdev;
1330 int enable = (int)msg->payload.state.num;
1331 int index = (int)msg->payload.state.index;
1332 struct rte_vdpa_device *vdpa_dev;
1335 RTE_LOG(INFO, VHOST_CONFIG,
1336 "set queue enable: %d to qp idx: %d\n",
1339 did = dev->vdpa_dev_id;
1340 vdpa_dev = rte_vdpa_get_device(did);
1341 if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1342 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1344 if (dev->notify_ops->vring_state_changed)
1345 dev->notify_ops->vring_state_changed(dev->vid,
1348 /* On disable, rings have to be stopped being processed. */
1349 if (!enable && dev->dequeue_zero_copy)
1350 drain_zmbuf_list(dev->virtqueue[index]);
1352 dev->virtqueue[index]->enabled = enable;
1354 return RTE_VHOST_MSG_RESULT_OK;
1358 vhost_user_get_protocol_features(struct virtio_net **pdev,
1359 struct VhostUserMsg *msg,
1360 int main_fd __rte_unused)
1362 struct virtio_net *dev = *pdev;
1363 uint64_t features, protocol_features;
1365 rte_vhost_driver_get_features(dev->ifname, &features);
1366 rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1369 * REPLY_ACK protocol feature is only mandatory for now
1370 * for IOMMU feature. If IOMMU is explicitly disabled by the
1371 * application, disable also REPLY_ACK feature for older buggy
1372 * Qemu versions (from v2.7.0 to v2.9.0).
1374 if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1375 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1377 msg->payload.u64 = protocol_features;
1378 msg->size = sizeof(msg->payload.u64);
1381 return RTE_VHOST_MSG_RESULT_REPLY;
1385 vhost_user_set_protocol_features(struct virtio_net **pdev,
1386 struct VhostUserMsg *msg,
1387 int main_fd __rte_unused)
1389 struct virtio_net *dev = *pdev;
1390 uint64_t protocol_features = msg->payload.u64;
1391 uint64_t slave_protocol_features = 0;
1393 rte_vhost_driver_get_protocol_features(dev->ifname,
1394 &slave_protocol_features);
1395 if (protocol_features & ~slave_protocol_features) {
1396 RTE_LOG(ERR, VHOST_CONFIG,
1397 "(%d) received invalid protocol features.\n",
1399 return RTE_VHOST_MSG_RESULT_ERR;
1402 dev->protocol_features = protocol_features;
1404 return RTE_VHOST_MSG_RESULT_OK;
1408 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1409 int main_fd __rte_unused)
1411 struct virtio_net *dev = *pdev;
1412 int fd = msg->fds[0];
1417 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1418 return RTE_VHOST_MSG_RESULT_ERR;
1421 if (msg->size != sizeof(VhostUserLog)) {
1422 RTE_LOG(ERR, VHOST_CONFIG,
1423 "invalid log base msg size: %"PRId32" != %d\n",
1424 msg->size, (int)sizeof(VhostUserLog));
1425 return RTE_VHOST_MSG_RESULT_ERR;
1428 size = msg->payload.log.mmap_size;
1429 off = msg->payload.log.mmap_offset;
1431 /* Don't allow mmap_offset to point outside the mmap region */
1433 RTE_LOG(ERR, VHOST_CONFIG,
1434 "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1436 return RTE_VHOST_MSG_RESULT_ERR;
1439 RTE_LOG(INFO, VHOST_CONFIG,
1440 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1444 * mmap from 0 to workaround a hugepage mmap bug: mmap will
1445 * fail when offset is not page size aligned.
1447 addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1449 if (addr == MAP_FAILED) {
1450 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1451 return RTE_VHOST_MSG_RESULT_ERR;
1455 * Free previously mapped log memory on occasionally
1456 * multiple VHOST_USER_SET_LOG_BASE.
1458 if (dev->log_addr) {
1459 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1461 dev->log_addr = (uint64_t)(uintptr_t)addr;
1462 dev->log_base = dev->log_addr + off;
1463 dev->log_size = size;
1466 * The spec is not clear about it (yet), but QEMU doesn't expect
1467 * any payload in the reply.
1472 return RTE_VHOST_MSG_RESULT_REPLY;
1475 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1476 struct VhostUserMsg *msg,
1477 int main_fd __rte_unused)
1480 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1482 return RTE_VHOST_MSG_RESULT_OK;
1486 * An rarp packet is constructed and broadcasted to notify switches about
1487 * the new location of the migrated VM, so that packets from outside will
1488 * not be lost after migration.
1490 * However, we don't actually "send" a rarp packet here, instead, we set
1491 * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1494 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1495 int main_fd __rte_unused)
1497 struct virtio_net *dev = *pdev;
1498 uint8_t *mac = (uint8_t *)&msg->payload.u64;
1499 struct rte_vdpa_device *vdpa_dev;
1502 RTE_LOG(DEBUG, VHOST_CONFIG,
1503 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1504 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1505 memcpy(dev->mac.addr_bytes, mac, 6);
1508 * Set the flag to inject a RARP broadcast packet at
1509 * rte_vhost_dequeue_burst().
1511 * rte_smp_wmb() is for making sure the mac is copied
1512 * before the flag is set.
1515 rte_atomic16_set(&dev->broadcast_rarp, 1);
1516 did = dev->vdpa_dev_id;
1517 vdpa_dev = rte_vdpa_get_device(did);
1518 if (vdpa_dev && vdpa_dev->ops->migration_done)
1519 vdpa_dev->ops->migration_done(dev->vid);
1521 return RTE_VHOST_MSG_RESULT_OK;
1525 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1526 int main_fd __rte_unused)
1528 struct virtio_net *dev = *pdev;
1529 if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1530 msg->payload.u64 > VIRTIO_MAX_MTU) {
1531 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1534 return RTE_VHOST_MSG_RESULT_ERR;
1537 dev->mtu = msg->payload.u64;
1539 return RTE_VHOST_MSG_RESULT_OK;
1543 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1544 int main_fd __rte_unused)
1546 struct virtio_net *dev = *pdev;
1547 int fd = msg->fds[0];
1550 RTE_LOG(ERR, VHOST_CONFIG,
1551 "Invalid file descriptor for slave channel (%d)\n",
1553 return RTE_VHOST_MSG_RESULT_ERR;
1556 dev->slave_req_fd = fd;
1558 return RTE_VHOST_MSG_RESULT_OK;
1562 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1564 struct vhost_vring_addr *ra;
1565 uint64_t start, end;
1568 end = start + imsg->size;
1570 ra = &vq->ring_addrs;
1571 if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1573 if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1575 if (ra->used_user_addr >= start && ra->used_user_addr < end)
1582 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1583 struct vhost_iotlb_msg *imsg)
1585 uint64_t istart, iend, vstart, vend;
1587 istart = imsg->iova;
1588 iend = istart + imsg->size - 1;
1590 vstart = (uintptr_t)vq->desc;
1591 vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1592 if (vstart <= iend && istart <= vend)
1595 vstart = (uintptr_t)vq->avail;
1596 vend = vstart + sizeof(struct vring_avail);
1597 vend += sizeof(uint16_t) * vq->size - 1;
1598 if (vstart <= iend && istart <= vend)
1601 vstart = (uintptr_t)vq->used;
1602 vend = vstart + sizeof(struct vring_used);
1603 vend += sizeof(struct vring_used_elem) * vq->size - 1;
1604 if (vstart <= iend && istart <= vend)
1611 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1612 int main_fd __rte_unused)
1614 struct virtio_net *dev = *pdev;
1615 struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1619 switch (imsg->type) {
1620 case VHOST_IOTLB_UPDATE:
1622 vva = qva_to_vva(dev, imsg->uaddr, &len);
1624 return RTE_VHOST_MSG_RESULT_ERR;
1626 for (i = 0; i < dev->nr_vring; i++) {
1627 struct vhost_virtqueue *vq = dev->virtqueue[i];
1629 vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1632 if (is_vring_iotlb_update(vq, imsg))
1633 *pdev = dev = translate_ring_addresses(dev, i);
1636 case VHOST_IOTLB_INVALIDATE:
1637 for (i = 0; i < dev->nr_vring; i++) {
1638 struct vhost_virtqueue *vq = dev->virtqueue[i];
1640 vhost_user_iotlb_cache_remove(vq, imsg->iova,
1643 if (is_vring_iotlb_invalidate(vq, imsg))
1644 vring_invalidate(dev, vq);
1648 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1650 return RTE_VHOST_MSG_RESULT_ERR;
1653 return RTE_VHOST_MSG_RESULT_OK;
1657 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1658 struct VhostUserMsg *msg,
1659 int main_fd __rte_unused)
1661 struct virtio_net *dev = *pdev;
1662 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1663 struct uffdio_api api_struct;
1665 dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1667 if (dev->postcopy_ufd == -1) {
1668 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1670 return RTE_VHOST_MSG_RESULT_ERR;
1672 api_struct.api = UFFD_API;
1673 api_struct.features = 0;
1674 if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1675 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1677 close(dev->postcopy_ufd);
1678 dev->postcopy_ufd = -1;
1679 return RTE_VHOST_MSG_RESULT_ERR;
1681 msg->fds[0] = dev->postcopy_ufd;
1684 return RTE_VHOST_MSG_RESULT_REPLY;
1686 dev->postcopy_ufd = -1;
1689 return RTE_VHOST_MSG_RESULT_ERR;
1694 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1695 struct VhostUserMsg *msg __rte_unused,
1696 int main_fd __rte_unused)
1698 struct virtio_net *dev = *pdev;
1700 if (dev->mem && dev->mem->nregions) {
1701 RTE_LOG(ERR, VHOST_CONFIG,
1702 "Regions already registered at postcopy-listen\n");
1703 return RTE_VHOST_MSG_RESULT_ERR;
1705 dev->postcopy_listening = 1;
1707 return RTE_VHOST_MSG_RESULT_OK;
1711 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1712 int main_fd __rte_unused)
1714 struct virtio_net *dev = *pdev;
1716 dev->postcopy_listening = 0;
1717 if (dev->postcopy_ufd >= 0) {
1718 close(dev->postcopy_ufd);
1719 dev->postcopy_ufd = -1;
1722 msg->payload.u64 = 0;
1723 msg->size = sizeof(msg->payload.u64);
1726 return RTE_VHOST_MSG_RESULT_REPLY;
1729 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1730 struct VhostUserMsg *msg,
1732 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1733 [VHOST_USER_NONE] = NULL,
1734 [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1735 [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1736 [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1737 [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1738 [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1739 [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1740 [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1741 [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1742 [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1743 [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1744 [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1745 [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1746 [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1747 [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1748 [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1749 [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1750 [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1751 [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1752 [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1753 [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1754 [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1755 [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1756 [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1757 [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1758 [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1762 /* return bytes# of read on success or negative val on failure. */
1764 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1768 ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1769 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1774 if (msg->size > sizeof(msg->payload)) {
1775 RTE_LOG(ERR, VHOST_CONFIG,
1776 "invalid msg size: %d\n", msg->size);
1779 ret = read(sockfd, &msg->payload, msg->size);
1782 if (ret != (int)msg->size) {
1783 RTE_LOG(ERR, VHOST_CONFIG,
1784 "read control message failed\n");
1793 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1798 return send_fd_message(sockfd, (char *)msg,
1799 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1803 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1808 msg->flags &= ~VHOST_USER_VERSION_MASK;
1809 msg->flags &= ~VHOST_USER_NEED_REPLY;
1810 msg->flags |= VHOST_USER_VERSION;
1811 msg->flags |= VHOST_USER_REPLY_MASK;
1813 return send_vhost_message(sockfd, msg);
1817 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1821 if (msg->flags & VHOST_USER_NEED_REPLY)
1822 rte_spinlock_lock(&dev->slave_req_lock);
1824 ret = send_vhost_message(dev->slave_req_fd, msg);
1825 if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1826 rte_spinlock_unlock(&dev->slave_req_lock);
1832 * Allocate a queue pair if it hasn't been allocated yet
1835 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1836 struct VhostUserMsg *msg)
1840 switch (msg->request.master) {
1841 case VHOST_USER_SET_VRING_KICK:
1842 case VHOST_USER_SET_VRING_CALL:
1843 case VHOST_USER_SET_VRING_ERR:
1844 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1846 case VHOST_USER_SET_VRING_NUM:
1847 case VHOST_USER_SET_VRING_BASE:
1848 case VHOST_USER_SET_VRING_ENABLE:
1849 vring_idx = msg->payload.state.index;
1851 case VHOST_USER_SET_VRING_ADDR:
1852 vring_idx = msg->payload.addr.index;
1858 if (vring_idx >= VHOST_MAX_VRING) {
1859 RTE_LOG(ERR, VHOST_CONFIG,
1860 "invalid vring index: %u\n", vring_idx);
1864 if (dev->virtqueue[vring_idx])
1867 return alloc_vring_queue(dev, vring_idx);
1871 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1874 unsigned int vq_num = 0;
1876 while (vq_num < dev->nr_vring) {
1877 struct vhost_virtqueue *vq = dev->virtqueue[i];
1880 rte_spinlock_lock(&vq->access_lock);
1888 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1891 unsigned int vq_num = 0;
1893 while (vq_num < dev->nr_vring) {
1894 struct vhost_virtqueue *vq = dev->virtqueue[i];
1897 rte_spinlock_unlock(&vq->access_lock);
1905 vhost_user_msg_handler(int vid, int fd)
1907 struct virtio_net *dev;
1908 struct VhostUserMsg msg;
1909 struct rte_vdpa_device *vdpa_dev;
1912 int unlock_required = 0;
1913 uint32_t skip_master = 0;
1916 dev = get_device(vid);
1920 if (!dev->notify_ops) {
1921 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1922 if (!dev->notify_ops) {
1923 RTE_LOG(ERR, VHOST_CONFIG,
1924 "failed to get callback ops for driver %s\n",
1930 ret = read_vhost_message(fd, &msg);
1931 if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1933 RTE_LOG(ERR, VHOST_CONFIG,
1934 "vhost read message failed\n");
1936 RTE_LOG(INFO, VHOST_CONFIG,
1937 "vhost peer closed\n");
1939 RTE_LOG(ERR, VHOST_CONFIG,
1940 "vhost read incorrect message\n");
1946 if (msg.request.master != VHOST_USER_IOTLB_MSG)
1947 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1948 vhost_message_str[msg.request.master]);
1950 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1951 vhost_message_str[msg.request.master]);
1953 ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1955 RTE_LOG(ERR, VHOST_CONFIG,
1956 "failed to alloc queue\n");
1961 * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1962 * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1963 * and device is destroyed. destroy_device waits for queues to be
1964 * inactive, so it is safe. Otherwise taking the access_lock
1965 * would cause a dead lock.
1967 switch (msg.request.master) {
1968 case VHOST_USER_SET_FEATURES:
1969 case VHOST_USER_SET_PROTOCOL_FEATURES:
1970 case VHOST_USER_SET_OWNER:
1971 case VHOST_USER_SET_MEM_TABLE:
1972 case VHOST_USER_SET_LOG_BASE:
1973 case VHOST_USER_SET_LOG_FD:
1974 case VHOST_USER_SET_VRING_NUM:
1975 case VHOST_USER_SET_VRING_ADDR:
1976 case VHOST_USER_SET_VRING_BASE:
1977 case VHOST_USER_SET_VRING_KICK:
1978 case VHOST_USER_SET_VRING_CALL:
1979 case VHOST_USER_SET_VRING_ERR:
1980 case VHOST_USER_SET_VRING_ENABLE:
1981 case VHOST_USER_SEND_RARP:
1982 case VHOST_USER_NET_SET_MTU:
1983 case VHOST_USER_SET_SLAVE_REQ_FD:
1984 vhost_user_lock_all_queue_pairs(dev);
1985 unlock_required = 1;
1992 if (dev->extern_ops.pre_msg_handle) {
1993 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
1994 (void *)&msg, &skip_master);
1995 if (ret == RTE_VHOST_MSG_RESULT_ERR)
1997 else if (ret == RTE_VHOST_MSG_RESULT_REPLY)
1998 send_vhost_reply(fd, &msg);
2001 goto skip_to_post_handle;
2004 request = msg.request.master;
2005 if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2006 if (!vhost_message_handlers[request])
2007 goto skip_to_post_handle;
2008 ret = vhost_message_handlers[request](&dev, &msg, fd);
2011 case RTE_VHOST_MSG_RESULT_ERR:
2012 RTE_LOG(ERR, VHOST_CONFIG,
2013 "Processing %s failed.\n",
2014 vhost_message_str[request]);
2016 case RTE_VHOST_MSG_RESULT_OK:
2017 RTE_LOG(DEBUG, VHOST_CONFIG,
2018 "Processing %s succeeded.\n",
2019 vhost_message_str[request]);
2021 case RTE_VHOST_MSG_RESULT_REPLY:
2022 RTE_LOG(DEBUG, VHOST_CONFIG,
2023 "Processing %s succeeded and needs reply.\n",
2024 vhost_message_str[request]);
2025 send_vhost_reply(fd, &msg);
2029 RTE_LOG(ERR, VHOST_CONFIG,
2030 "Requested invalid message type %d.\n", request);
2031 ret = RTE_VHOST_MSG_RESULT_ERR;
2034 skip_to_post_handle:
2035 if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2036 dev->extern_ops.post_msg_handle) {
2037 ret = (*dev->extern_ops.post_msg_handle)(
2038 dev->vid, (void *)&msg);
2039 if (ret == RTE_VHOST_MSG_RESULT_ERR)
2041 else if (ret == RTE_VHOST_MSG_RESULT_REPLY)
2042 send_vhost_reply(fd, &msg);
2046 if (unlock_required)
2047 vhost_user_unlock_all_queue_pairs(dev);
2050 * If the request required a reply that was already sent,
2051 * this optional reply-ack won't be sent as the
2052 * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2054 if (msg.flags & VHOST_USER_NEED_REPLY) {
2055 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2056 msg.size = sizeof(msg.payload.u64);
2058 send_vhost_reply(fd, &msg);
2059 } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2060 RTE_LOG(ERR, VHOST_CONFIG,
2061 "vhost message handling failed.\n");
2065 if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2066 dev->flags |= VIRTIO_DEV_READY;
2068 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2069 if (dev->dequeue_zero_copy) {
2070 RTE_LOG(INFO, VHOST_CONFIG,
2071 "dequeue zero copy is enabled\n");
2074 if (dev->notify_ops->new_device(dev->vid) == 0)
2075 dev->flags |= VIRTIO_DEV_RUNNING;
2079 did = dev->vdpa_dev_id;
2080 vdpa_dev = rte_vdpa_get_device(did);
2081 if (vdpa_dev && virtio_is_ready(dev) &&
2082 !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2083 msg.request.master == VHOST_USER_SET_VRING_CALL) {
2084 if (vdpa_dev->ops->dev_conf)
2085 vdpa_dev->ops->dev_conf(dev->vid);
2086 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2092 static int process_slave_message_reply(struct virtio_net *dev,
2093 const struct VhostUserMsg *msg)
2095 struct VhostUserMsg msg_reply;
2098 if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2101 if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2106 if (msg_reply.request.slave != msg->request.slave) {
2107 RTE_LOG(ERR, VHOST_CONFIG,
2108 "Received unexpected msg type (%u), expected %u\n",
2109 msg_reply.request.slave, msg->request.slave);
2114 ret = msg_reply.payload.u64 ? -1 : 0;
2117 rte_spinlock_unlock(&dev->slave_req_lock);
2122 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2125 struct VhostUserMsg msg = {
2126 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2127 .flags = VHOST_USER_VERSION,
2128 .size = sizeof(msg.payload.iotlb),
2132 .type = VHOST_IOTLB_MISS,
2136 ret = send_vhost_message(dev->slave_req_fd, &msg);
2138 RTE_LOG(ERR, VHOST_CONFIG,
2139 "Failed to send IOTLB miss message (%d)\n",
2147 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2153 struct VhostUserMsg msg = {
2154 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2155 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2156 .size = sizeof(msg.payload.area),
2158 .u64 = index & VHOST_USER_VRING_IDX_MASK,
2165 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2171 ret = send_vhost_slave_message(dev, &msg);
2173 RTE_LOG(ERR, VHOST_CONFIG,
2174 "Failed to set host notifier (%d)\n", ret);
2178 return process_slave_message_reply(dev, &msg);
2181 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2183 struct virtio_net *dev;
2184 struct rte_vdpa_device *vdpa_dev;
2185 int vfio_device_fd, did, ret = 0;
2186 uint64_t offset, size;
2189 dev = get_device(vid);
2193 did = dev->vdpa_dev_id;
2197 if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2198 !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2199 !(dev->protocol_features &
2200 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2201 !(dev->protocol_features &
2202 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2203 !(dev->protocol_features &
2204 (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2207 vdpa_dev = rte_vdpa_get_device(did);
2211 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2212 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2214 vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2215 if (vfio_device_fd < 0)
2219 for (i = 0; i < dev->nr_vring; i++) {
2220 if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2226 if (vhost_user_slave_set_vring_host_notifier(dev, i,
2227 vfio_device_fd, offset, size) < 0) {
2234 for (i = 0; i < dev->nr_vring; i++) {
2235 vhost_user_slave_set_vring_host_notifier(dev, i, -1,