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>
40 #ifdef F_ADD_SEALS /* if file sealing is supported, so is memfd */
41 #include <linux/memfd.h>
42 #define MEMFD_SUPPORTED
45 #include <rte_common.h>
46 #include <rte_malloc.h>
51 #include "vhost_user.h"
53 #define VIRTIO_MIN_MTU 68
54 #define VIRTIO_MAX_MTU 65535
56 #define INFLIGHT_ALIGNMENT 64
57 #define INFLIGHT_VERSION 0x1
59 static const char *vhost_message_str[VHOST_USER_MAX] = {
60 [VHOST_USER_NONE] = "VHOST_USER_NONE",
61 [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
62 [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
63 [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
64 [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
65 [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
66 [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
67 [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
68 [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
69 [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
70 [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
71 [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
72 [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
73 [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
74 [VHOST_USER_SET_VRING_ERR] = "VHOST_USER_SET_VRING_ERR",
75 [VHOST_USER_GET_PROTOCOL_FEATURES] = "VHOST_USER_GET_PROTOCOL_FEATURES",
76 [VHOST_USER_SET_PROTOCOL_FEATURES] = "VHOST_USER_SET_PROTOCOL_FEATURES",
77 [VHOST_USER_GET_QUEUE_NUM] = "VHOST_USER_GET_QUEUE_NUM",
78 [VHOST_USER_SET_VRING_ENABLE] = "VHOST_USER_SET_VRING_ENABLE",
79 [VHOST_USER_SEND_RARP] = "VHOST_USER_SEND_RARP",
80 [VHOST_USER_NET_SET_MTU] = "VHOST_USER_NET_SET_MTU",
81 [VHOST_USER_SET_SLAVE_REQ_FD] = "VHOST_USER_SET_SLAVE_REQ_FD",
82 [VHOST_USER_IOTLB_MSG] = "VHOST_USER_IOTLB_MSG",
83 [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
84 [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
85 [VHOST_USER_POSTCOPY_ADVISE] = "VHOST_USER_POSTCOPY_ADVISE",
86 [VHOST_USER_POSTCOPY_LISTEN] = "VHOST_USER_POSTCOPY_LISTEN",
87 [VHOST_USER_POSTCOPY_END] = "VHOST_USER_POSTCOPY_END",
88 [VHOST_USER_GET_INFLIGHT_FD] = "VHOST_USER_GET_INFLIGHT_FD",
89 [VHOST_USER_SET_INFLIGHT_FD] = "VHOST_USER_SET_INFLIGHT_FD",
92 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
93 static int read_vhost_message(int sockfd, struct VhostUserMsg *msg);
101 ret = fstat(fd, &stat);
102 return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
106 * Reclaim all the outstanding zmbufs for a virtqueue.
109 drain_zmbuf_list(struct vhost_virtqueue *vq)
111 struct zcopy_mbuf *zmbuf, *next;
113 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
114 zmbuf != NULL; zmbuf = next) {
115 next = TAILQ_NEXT(zmbuf, next);
117 while (!mbuf_is_consumed(zmbuf->mbuf))
120 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
121 restore_mbuf(zmbuf->mbuf);
122 rte_pktmbuf_free(zmbuf->mbuf);
129 free_mem_region(struct virtio_net *dev)
132 struct rte_vhost_mem_region *reg;
133 struct vhost_virtqueue *vq;
135 if (!dev || !dev->mem)
138 if (dev->dequeue_zero_copy) {
139 for (i = 0; i < dev->nr_vring; i++) {
140 vq = dev->virtqueue[i];
142 drain_zmbuf_list(vq);
146 for (i = 0; i < dev->mem->nregions; i++) {
147 reg = &dev->mem->regions[i];
148 if (reg->host_user_addr) {
149 munmap(reg->mmap_addr, reg->mmap_size);
156 vhost_backend_cleanup(struct virtio_net *dev)
159 free_mem_region(dev);
164 free(dev->guest_pages);
165 dev->guest_pages = NULL;
168 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
172 if (dev->inflight_info) {
173 if (dev->inflight_info->addr) {
174 munmap(dev->inflight_info->addr,
175 dev->inflight_info->size);
176 dev->inflight_info->addr = NULL;
179 if (dev->inflight_info->fd > 0) {
180 close(dev->inflight_info->fd);
181 dev->inflight_info->fd = -1;
184 free(dev->inflight_info);
185 dev->inflight_info = NULL;
188 if (dev->slave_req_fd >= 0) {
189 close(dev->slave_req_fd);
190 dev->slave_req_fd = -1;
193 if (dev->postcopy_ufd >= 0) {
194 close(dev->postcopy_ufd);
195 dev->postcopy_ufd = -1;
198 dev->postcopy_listening = 0;
202 * This function just returns success at the moment unless
203 * the device hasn't been initialised.
206 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
207 struct VhostUserMsg *msg __rte_unused,
208 int main_fd __rte_unused)
210 return RTE_VHOST_MSG_RESULT_OK;
214 vhost_user_reset_owner(struct virtio_net **pdev,
215 struct VhostUserMsg *msg __rte_unused,
216 int main_fd __rte_unused)
218 struct virtio_net *dev = *pdev;
219 vhost_destroy_device_notify(dev);
221 cleanup_device(dev, 0);
223 return RTE_VHOST_MSG_RESULT_OK;
227 * The features that we support are requested.
230 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
231 int main_fd __rte_unused)
233 struct virtio_net *dev = *pdev;
234 uint64_t features = 0;
236 rte_vhost_driver_get_features(dev->ifname, &features);
238 msg->payload.u64 = features;
239 msg->size = sizeof(msg->payload.u64);
242 return RTE_VHOST_MSG_RESULT_REPLY;
246 * The queue number that we support are requested.
249 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
250 int main_fd __rte_unused)
252 struct virtio_net *dev = *pdev;
253 uint32_t queue_num = 0;
255 rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
257 msg->payload.u64 = (uint64_t)queue_num;
258 msg->size = sizeof(msg->payload.u64);
261 return RTE_VHOST_MSG_RESULT_REPLY;
265 * We receive the negotiated features supported by us and the virtio device.
268 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
269 int main_fd __rte_unused)
271 struct virtio_net *dev = *pdev;
272 uint64_t features = msg->payload.u64;
273 uint64_t vhost_features = 0;
274 struct rte_vdpa_device *vdpa_dev;
277 rte_vhost_driver_get_features(dev->ifname, &vhost_features);
278 if (features & ~vhost_features) {
279 RTE_LOG(ERR, VHOST_CONFIG,
280 "(%d) received invalid negotiated features.\n",
282 return RTE_VHOST_MSG_RESULT_ERR;
285 if (dev->flags & VIRTIO_DEV_RUNNING) {
286 if (dev->features == features)
287 return RTE_VHOST_MSG_RESULT_OK;
290 * Error out if master tries to change features while device is
291 * in running state. The exception being VHOST_F_LOG_ALL, which
292 * is enabled when the live-migration starts.
294 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
295 RTE_LOG(ERR, VHOST_CONFIG,
296 "(%d) features changed while device is running.\n",
298 return RTE_VHOST_MSG_RESULT_ERR;
301 if (dev->notify_ops->features_changed)
302 dev->notify_ops->features_changed(dev->vid, features);
305 dev->features = features;
307 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
308 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
310 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
312 RTE_LOG(INFO, VHOST_CONFIG,
313 "negotiated Virtio features: 0x%" PRIx64 "\n", dev->features);
314 VHOST_LOG_DEBUG(VHOST_CONFIG,
315 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
317 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
318 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
320 if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
321 !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
323 * Remove all but first queue pair if MQ hasn't been
324 * negotiated. This is safe because the device is not
325 * running at this stage.
327 while (dev->nr_vring > 2) {
328 struct vhost_virtqueue *vq;
330 vq = dev->virtqueue[--dev->nr_vring];
334 dev->virtqueue[dev->nr_vring] = NULL;
336 cleanup_vq_inflight(dev, vq);
341 did = dev->vdpa_dev_id;
342 vdpa_dev = rte_vdpa_get_device(did);
343 if (vdpa_dev && vdpa_dev->ops->set_features)
344 vdpa_dev->ops->set_features(dev->vid);
346 return RTE_VHOST_MSG_RESULT_OK;
350 * The virtio device sends us the size of the descriptor ring.
353 vhost_user_set_vring_num(struct virtio_net **pdev,
354 struct VhostUserMsg *msg,
355 int main_fd __rte_unused)
357 struct virtio_net *dev = *pdev;
358 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
360 vq->size = msg->payload.state.num;
362 /* VIRTIO 1.0, 2.4 Virtqueues says:
364 * Queue Size value is always a power of 2. The maximum Queue Size
367 * VIRTIO 1.1 2.7 Virtqueues says:
369 * Packed virtqueues support up to 2^15 entries each.
371 if (!vq_is_packed(dev)) {
372 if (vq->size & (vq->size - 1)) {
373 RTE_LOG(ERR, VHOST_CONFIG,
374 "invalid virtqueue size %u\n", vq->size);
375 return RTE_VHOST_MSG_RESULT_ERR;
379 if (vq->size > 32768) {
380 RTE_LOG(ERR, VHOST_CONFIG,
381 "invalid virtqueue size %u\n", vq->size);
382 return RTE_VHOST_MSG_RESULT_ERR;
385 if (dev->dequeue_zero_copy) {
387 vq->last_zmbuf_idx = 0;
388 vq->zmbuf_size = vq->size;
389 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
390 sizeof(struct zcopy_mbuf), 0);
391 if (vq->zmbufs == NULL) {
392 RTE_LOG(WARNING, VHOST_CONFIG,
393 "failed to allocate mem for zero copy; "
394 "zero copy is force disabled\n");
395 dev->dequeue_zero_copy = 0;
397 TAILQ_INIT(&vq->zmbuf_list);
400 if (vq_is_packed(dev)) {
401 vq->shadow_used_packed = rte_malloc(NULL,
403 sizeof(struct vring_used_elem_packed),
404 RTE_CACHE_LINE_SIZE);
405 if (!vq->shadow_used_packed) {
406 RTE_LOG(ERR, VHOST_CONFIG,
407 "failed to allocate memory for shadow used ring.\n");
408 return RTE_VHOST_MSG_RESULT_ERR;
412 vq->shadow_used_split = rte_malloc(NULL,
413 vq->size * sizeof(struct vring_used_elem),
414 RTE_CACHE_LINE_SIZE);
415 if (!vq->shadow_used_split) {
416 RTE_LOG(ERR, VHOST_CONFIG,
417 "failed to allocate memory for shadow used ring.\n");
418 return RTE_VHOST_MSG_RESULT_ERR;
422 vq->batch_copy_elems = rte_malloc(NULL,
423 vq->size * sizeof(struct batch_copy_elem),
424 RTE_CACHE_LINE_SIZE);
425 if (!vq->batch_copy_elems) {
426 RTE_LOG(ERR, VHOST_CONFIG,
427 "failed to allocate memory for batching copy.\n");
428 return RTE_VHOST_MSG_RESULT_ERR;
431 return RTE_VHOST_MSG_RESULT_OK;
435 * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
436 * same numa node as the memory of vring descriptor.
438 #ifdef RTE_LIBRTE_VHOST_NUMA
439 static struct virtio_net*
440 numa_realloc(struct virtio_net *dev, int index)
442 int oldnode, newnode;
443 struct virtio_net *old_dev;
444 struct vhost_virtqueue *old_vq, *vq;
445 struct zcopy_mbuf *new_zmbuf;
446 struct vring_used_elem *new_shadow_used_split;
447 struct vring_used_elem_packed *new_shadow_used_packed;
448 struct batch_copy_elem *new_batch_copy_elems;
451 if (dev->flags & VIRTIO_DEV_RUNNING)
455 vq = old_vq = dev->virtqueue[index];
457 ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
458 MPOL_F_NODE | MPOL_F_ADDR);
460 /* check if we need to reallocate vq */
461 ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
462 MPOL_F_NODE | MPOL_F_ADDR);
464 RTE_LOG(ERR, VHOST_CONFIG,
465 "Unable to get vq numa information.\n");
468 if (oldnode != newnode) {
469 RTE_LOG(INFO, VHOST_CONFIG,
470 "reallocate vq from %d to %d node\n", oldnode, newnode);
471 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
475 memcpy(vq, old_vq, sizeof(*vq));
476 TAILQ_INIT(&vq->zmbuf_list);
478 if (dev->dequeue_zero_copy) {
479 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
480 sizeof(struct zcopy_mbuf), 0, newnode);
482 rte_free(vq->zmbufs);
483 vq->zmbufs = new_zmbuf;
487 if (vq_is_packed(dev)) {
488 new_shadow_used_packed = rte_malloc_socket(NULL,
490 sizeof(struct vring_used_elem_packed),
493 if (new_shadow_used_packed) {
494 rte_free(vq->shadow_used_packed);
495 vq->shadow_used_packed = new_shadow_used_packed;
498 new_shadow_used_split = rte_malloc_socket(NULL,
500 sizeof(struct vring_used_elem),
503 if (new_shadow_used_split) {
504 rte_free(vq->shadow_used_split);
505 vq->shadow_used_split = new_shadow_used_split;
509 new_batch_copy_elems = rte_malloc_socket(NULL,
510 vq->size * sizeof(struct batch_copy_elem),
513 if (new_batch_copy_elems) {
514 rte_free(vq->batch_copy_elems);
515 vq->batch_copy_elems = new_batch_copy_elems;
521 /* check if we need to reallocate dev */
522 ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
523 MPOL_F_NODE | MPOL_F_ADDR);
525 RTE_LOG(ERR, VHOST_CONFIG,
526 "Unable to get dev numa information.\n");
529 if (oldnode != newnode) {
530 RTE_LOG(INFO, VHOST_CONFIG,
531 "reallocate dev from %d to %d node\n",
533 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
539 memcpy(dev, old_dev, sizeof(*dev));
544 dev->virtqueue[index] = vq;
545 vhost_devices[dev->vid] = dev;
548 vhost_user_iotlb_init(dev, index);
553 static struct virtio_net*
554 numa_realloc(struct virtio_net *dev, int index __rte_unused)
560 /* Converts QEMU virtual address to Vhost virtual address. */
562 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
564 struct rte_vhost_mem_region *r;
567 if (unlikely(!dev || !dev->mem))
570 /* Find the region where the address lives. */
571 for (i = 0; i < dev->mem->nregions; i++) {
572 r = &dev->mem->regions[i];
574 if (qva >= r->guest_user_addr &&
575 qva < r->guest_user_addr + r->size) {
577 if (unlikely(*len > r->guest_user_addr + r->size - qva))
578 *len = r->guest_user_addr + r->size - qva;
580 return qva - r->guest_user_addr +
592 * Converts ring address to Vhost virtual address.
593 * If IOMMU is enabled, the ring address is a guest IO virtual address,
594 * else it is a QEMU virtual address.
597 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
598 uint64_t ra, uint64_t *size)
600 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
602 uint64_t req_size = *size;
604 vva = vhost_user_iotlb_cache_find(vq, ra,
605 size, VHOST_ACCESS_RW);
606 if (req_size != *size)
607 vhost_user_iotlb_miss(dev, (ra + *size),
613 return qva_to_vva(dev, ra, size);
617 * Converts vring log address to GPA
618 * If IOMMU is enabled, the log address is IOVA
619 * If IOMMU not enabled, the log address is already GPA
622 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
625 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
626 const uint64_t exp_size = sizeof(struct vring_used) +
627 sizeof(struct vring_used_elem) * vq->size;
629 uint64_t size = exp_size;
631 hva = vhost_iova_to_vva(dev, vq, log_addr,
632 &size, VHOST_ACCESS_RW);
633 if (size != exp_size)
636 gpa = hva_to_gpa(dev, hva, exp_size);
638 RTE_LOG(ERR, VHOST_CONFIG,
639 "VQ: Failed to find GPA for log_addr: 0x%" PRIx64 " hva: 0x%" PRIx64 "\n",
649 static struct virtio_net *
650 translate_ring_addresses(struct virtio_net *dev, int vq_index)
652 struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
653 struct vhost_vring_addr *addr = &vq->ring_addrs;
654 uint64_t len, expected_len;
656 if (vq_is_packed(dev)) {
657 len = sizeof(struct vring_packed_desc) * vq->size;
658 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
659 ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
660 vq->log_guest_addr = 0;
661 if (vq->desc_packed == NULL ||
662 len != sizeof(struct vring_packed_desc) *
664 RTE_LOG(DEBUG, VHOST_CONFIG,
665 "(%d) failed to map desc_packed ring.\n",
670 dev = numa_realloc(dev, vq_index);
671 vq = dev->virtqueue[vq_index];
672 addr = &vq->ring_addrs;
674 len = sizeof(struct vring_packed_desc_event);
675 vq->driver_event = (struct vring_packed_desc_event *)
676 (uintptr_t)ring_addr_to_vva(dev,
677 vq, addr->avail_user_addr, &len);
678 if (vq->driver_event == NULL ||
679 len != sizeof(struct vring_packed_desc_event)) {
680 RTE_LOG(DEBUG, VHOST_CONFIG,
681 "(%d) failed to find driver area address.\n",
686 len = sizeof(struct vring_packed_desc_event);
687 vq->device_event = (struct vring_packed_desc_event *)
688 (uintptr_t)ring_addr_to_vva(dev,
689 vq, addr->used_user_addr, &len);
690 if (vq->device_event == NULL ||
691 len != sizeof(struct vring_packed_desc_event)) {
692 RTE_LOG(DEBUG, VHOST_CONFIG,
693 "(%d) failed to find device area address.\n",
702 /* The addresses are converted from QEMU virtual to Vhost virtual. */
703 if (vq->desc && vq->avail && vq->used)
706 len = sizeof(struct vring_desc) * vq->size;
707 vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
708 vq, addr->desc_user_addr, &len);
709 if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
710 RTE_LOG(DEBUG, VHOST_CONFIG,
711 "(%d) failed to map desc ring.\n",
716 dev = numa_realloc(dev, vq_index);
717 vq = dev->virtqueue[vq_index];
718 addr = &vq->ring_addrs;
720 len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
721 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
722 len += sizeof(uint16_t);
724 vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
725 vq, addr->avail_user_addr, &len);
726 if (vq->avail == 0 || len != expected_len) {
727 RTE_LOG(DEBUG, VHOST_CONFIG,
728 "(%d) failed to map avail ring.\n",
733 len = sizeof(struct vring_used) +
734 sizeof(struct vring_used_elem) * vq->size;
735 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
736 len += sizeof(uint16_t);
738 vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
739 vq, addr->used_user_addr, &len);
740 if (vq->used == 0 || len != expected_len) {
741 RTE_LOG(DEBUG, VHOST_CONFIG,
742 "(%d) failed to map used ring.\n",
747 if (vq->last_used_idx != vq->used->idx) {
748 RTE_LOG(WARNING, VHOST_CONFIG,
749 "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
750 "some packets maybe resent for Tx and dropped for Rx\n",
751 vq->last_used_idx, vq->used->idx);
752 vq->last_used_idx = vq->used->idx;
753 vq->last_avail_idx = vq->used->idx;
757 translate_log_addr(dev, vq, addr->log_guest_addr);
758 if (vq->log_guest_addr == 0) {
759 RTE_LOG(DEBUG, VHOST_CONFIG,
760 "(%d) failed to map log_guest_addr .\n",
766 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
768 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
769 dev->vid, vq->avail);
770 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
772 VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
773 dev->vid, vq->log_guest_addr);
779 * The virtio device sends us the desc, used and avail ring addresses.
780 * This function then converts these to our address space.
783 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
784 int main_fd __rte_unused)
786 struct virtio_net *dev = *pdev;
787 struct vhost_virtqueue *vq;
788 struct vhost_vring_addr *addr = &msg->payload.addr;
791 if (dev->mem == NULL)
792 return RTE_VHOST_MSG_RESULT_ERR;
794 /* addr->index refers to the queue index. The txq 1, rxq is 0. */
795 vq = dev->virtqueue[msg->payload.addr.index];
797 access_ok = vq->access_ok;
800 * Rings addresses should not be interpreted as long as the ring is not
801 * started and enabled
803 memcpy(&vq->ring_addrs, addr, sizeof(*addr));
805 vring_invalidate(dev, vq);
807 if ((vq->enabled && (dev->features &
808 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
810 dev = translate_ring_addresses(dev, msg->payload.addr.index);
812 return RTE_VHOST_MSG_RESULT_ERR;
817 return RTE_VHOST_MSG_RESULT_OK;
821 * The virtio device sends us the available ring last used index.
824 vhost_user_set_vring_base(struct virtio_net **pdev,
825 struct VhostUserMsg *msg,
826 int main_fd __rte_unused)
828 struct virtio_net *dev = *pdev;
829 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
830 uint64_t val = msg->payload.state.num;
832 if (vq_is_packed(dev)) {
834 * Bit[0:14]: avail index
835 * Bit[15]: avail wrap counter
837 vq->last_avail_idx = val & 0x7fff;
838 vq->avail_wrap_counter = !!(val & (0x1 << 15));
840 * Set used index to same value as available one, as
841 * their values should be the same since ring processing
842 * was stopped at get time.
844 vq->last_used_idx = vq->last_avail_idx;
845 vq->used_wrap_counter = vq->avail_wrap_counter;
847 vq->last_used_idx = msg->payload.state.num;
848 vq->last_avail_idx = msg->payload.state.num;
851 return RTE_VHOST_MSG_RESULT_OK;
855 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
856 uint64_t host_phys_addr, uint64_t size)
858 struct guest_page *page, *last_page;
859 struct guest_page *old_pages;
861 if (dev->nr_guest_pages == dev->max_guest_pages) {
862 dev->max_guest_pages *= 2;
863 old_pages = dev->guest_pages;
864 dev->guest_pages = realloc(dev->guest_pages,
865 dev->max_guest_pages * sizeof(*page));
866 if (!dev->guest_pages) {
867 RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
873 if (dev->nr_guest_pages > 0) {
874 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
875 /* merge if the two pages are continuous */
876 if (host_phys_addr == last_page->host_phys_addr +
878 last_page->size += size;
883 page = &dev->guest_pages[dev->nr_guest_pages++];
884 page->guest_phys_addr = guest_phys_addr;
885 page->host_phys_addr = host_phys_addr;
892 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
895 uint64_t reg_size = reg->size;
896 uint64_t host_user_addr = reg->host_user_addr;
897 uint64_t guest_phys_addr = reg->guest_phys_addr;
898 uint64_t host_phys_addr;
901 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
902 size = page_size - (guest_phys_addr & (page_size - 1));
903 size = RTE_MIN(size, reg_size);
905 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
908 host_user_addr += size;
909 guest_phys_addr += size;
912 while (reg_size > 0) {
913 size = RTE_MIN(reg_size, page_size);
914 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
916 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
920 host_user_addr += size;
921 guest_phys_addr += size;
928 #ifdef RTE_LIBRTE_VHOST_DEBUG
929 /* TODO: enable it only in debug mode? */
931 dump_guest_pages(struct virtio_net *dev)
934 struct guest_page *page;
936 for (i = 0; i < dev->nr_guest_pages; i++) {
937 page = &dev->guest_pages[i];
939 RTE_LOG(INFO, VHOST_CONFIG,
940 "guest physical page region %u\n"
941 "\t guest_phys_addr: %" PRIx64 "\n"
942 "\t host_phys_addr : %" PRIx64 "\n"
943 "\t size : %" PRIx64 "\n",
945 page->guest_phys_addr,
946 page->host_phys_addr,
951 #define dump_guest_pages(dev)
955 vhost_memory_changed(struct VhostUserMemory *new,
956 struct rte_vhost_memory *old)
960 if (new->nregions != old->nregions)
963 for (i = 0; i < new->nregions; ++i) {
964 VhostUserMemoryRegion *new_r = &new->regions[i];
965 struct rte_vhost_mem_region *old_r = &old->regions[i];
967 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
969 if (new_r->memory_size != old_r->size)
971 if (new_r->userspace_addr != old_r->guest_user_addr)
979 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
982 struct virtio_net *dev = *pdev;
983 struct VhostUserMemory *memory = &msg->payload.memory;
984 struct rte_vhost_mem_region *reg;
987 uint64_t mmap_offset;
993 if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
994 RTE_LOG(ERR, VHOST_CONFIG,
995 "too many memory regions (%u)\n", memory->nregions);
996 return RTE_VHOST_MSG_RESULT_ERR;
999 if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
1000 RTE_LOG(INFO, VHOST_CONFIG,
1001 "(%d) memory regions not changed\n", dev->vid);
1003 for (i = 0; i < memory->nregions; i++)
1006 return RTE_VHOST_MSG_RESULT_OK;
1010 free_mem_region(dev);
1015 /* Flush IOTLB cache as previous HVAs are now invalid */
1016 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1017 for (i = 0; i < dev->nr_vring; i++)
1018 vhost_user_iotlb_flush_all(dev->virtqueue[i]);
1020 dev->nr_guest_pages = 0;
1021 if (!dev->guest_pages) {
1022 dev->max_guest_pages = 8;
1023 dev->guest_pages = malloc(dev->max_guest_pages *
1024 sizeof(struct guest_page));
1025 if (dev->guest_pages == NULL) {
1026 RTE_LOG(ERR, VHOST_CONFIG,
1027 "(%d) failed to allocate memory "
1028 "for dev->guest_pages\n",
1030 return RTE_VHOST_MSG_RESULT_ERR;
1034 dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
1035 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
1036 if (dev->mem == NULL) {
1037 RTE_LOG(ERR, VHOST_CONFIG,
1038 "(%d) failed to allocate memory for dev->mem\n",
1040 return RTE_VHOST_MSG_RESULT_ERR;
1042 dev->mem->nregions = memory->nregions;
1044 for (i = 0; i < memory->nregions; i++) {
1046 reg = &dev->mem->regions[i];
1048 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1049 reg->guest_user_addr = memory->regions[i].userspace_addr;
1050 reg->size = memory->regions[i].memory_size;
1053 mmap_offset = memory->regions[i].mmap_offset;
1055 /* Check for memory_size + mmap_offset overflow */
1056 if (mmap_offset >= -reg->size) {
1057 RTE_LOG(ERR, VHOST_CONFIG,
1058 "mmap_offset (%#"PRIx64") and memory_size "
1059 "(%#"PRIx64") overflow\n",
1060 mmap_offset, reg->size);
1064 mmap_size = reg->size + mmap_offset;
1066 /* mmap() without flag of MAP_ANONYMOUS, should be called
1067 * with length argument aligned with hugepagesz at older
1068 * longterm version Linux, like 2.6.32 and 3.2.72, or
1069 * mmap() will fail with EINVAL.
1071 * to avoid failure, make sure in caller to keep length
1074 alignment = get_blk_size(fd);
1075 if (alignment == (uint64_t)-1) {
1076 RTE_LOG(ERR, VHOST_CONFIG,
1077 "couldn't get hugepage size through fstat\n");
1080 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1082 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1083 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1084 MAP_SHARED | populate, fd, 0);
1086 if (mmap_addr == MAP_FAILED) {
1087 RTE_LOG(ERR, VHOST_CONFIG,
1088 "mmap region %u failed.\n", i);
1092 reg->mmap_addr = mmap_addr;
1093 reg->mmap_size = mmap_size;
1094 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1097 if (dev->dequeue_zero_copy)
1098 if (add_guest_pages(dev, reg, alignment) < 0) {
1099 RTE_LOG(ERR, VHOST_CONFIG,
1100 "adding guest pages to region %u failed.\n",
1105 RTE_LOG(INFO, VHOST_CONFIG,
1106 "guest memory region %u, size: 0x%" PRIx64 "\n"
1107 "\t guest physical addr: 0x%" PRIx64 "\n"
1108 "\t guest virtual addr: 0x%" PRIx64 "\n"
1109 "\t host virtual addr: 0x%" PRIx64 "\n"
1110 "\t mmap addr : 0x%" PRIx64 "\n"
1111 "\t mmap size : 0x%" PRIx64 "\n"
1112 "\t mmap align: 0x%" PRIx64 "\n"
1113 "\t mmap off : 0x%" PRIx64 "\n",
1115 reg->guest_phys_addr,
1116 reg->guest_user_addr,
1117 reg->host_user_addr,
1118 (uint64_t)(uintptr_t)mmap_addr,
1123 if (dev->postcopy_listening) {
1125 * We haven't a better way right now than sharing
1126 * DPDK's virtual address with Qemu, so that Qemu can
1127 * retrieve the region offset when handling userfaults.
1129 memory->regions[i].userspace_addr =
1130 reg->host_user_addr;
1133 if (dev->postcopy_listening) {
1134 /* Send the addresses back to qemu */
1136 send_vhost_reply(main_fd, msg);
1138 /* Wait for qemu to acknolwedge it's got the addresses
1139 * we've got to wait before we're allowed to generate faults.
1141 VhostUserMsg ack_msg;
1142 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1143 RTE_LOG(ERR, VHOST_CONFIG,
1144 "Failed to read qemu ack on postcopy set-mem-table\n");
1147 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1148 RTE_LOG(ERR, VHOST_CONFIG,
1149 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1150 ack_msg.request.master);
1154 /* Now userfault register and we can use the memory */
1155 for (i = 0; i < memory->nregions; i++) {
1156 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1157 reg = &dev->mem->regions[i];
1158 struct uffdio_register reg_struct;
1161 * Let's register all the mmap'ed area to ensure
1162 * alignment on page boundary.
1164 reg_struct.range.start =
1165 (uint64_t)(uintptr_t)reg->mmap_addr;
1166 reg_struct.range.len = reg->mmap_size;
1167 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1169 if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1171 RTE_LOG(ERR, VHOST_CONFIG,
1172 "Failed to register ufd for region %d: (ufd = %d) %s\n",
1173 i, dev->postcopy_ufd,
1177 RTE_LOG(INFO, VHOST_CONFIG,
1178 "\t userfaultfd registered for range : "
1179 "%" PRIx64 " - %" PRIx64 "\n",
1180 (uint64_t)reg_struct.range.start,
1181 (uint64_t)reg_struct.range.start +
1182 (uint64_t)reg_struct.range.len - 1);
1189 for (i = 0; i < dev->nr_vring; i++) {
1190 struct vhost_virtqueue *vq = dev->virtqueue[i];
1192 if (vq->desc || vq->avail || vq->used) {
1194 * If the memory table got updated, the ring addresses
1195 * need to be translated again as virtual addresses have
1198 vring_invalidate(dev, vq);
1200 dev = translate_ring_addresses(dev, i);
1210 dump_guest_pages(dev);
1212 return RTE_VHOST_MSG_RESULT_OK;
1215 free_mem_region(dev);
1218 return RTE_VHOST_MSG_RESULT_ERR;
1222 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1229 if (vq_is_packed(dev))
1230 rings_ok = !!vq->desc_packed;
1232 rings_ok = vq->desc && vq->avail && vq->used;
1235 vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1236 vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1240 virtio_is_ready(struct virtio_net *dev)
1242 struct vhost_virtqueue *vq;
1245 if (dev->nr_vring == 0)
1248 for (i = 0; i < dev->nr_vring; i++) {
1249 vq = dev->virtqueue[i];
1251 if (!vq_is_ready(dev, vq))
1255 RTE_LOG(INFO, VHOST_CONFIG,
1256 "virtio is now ready for processing.\n");
1261 inflight_mem_alloc(const char *name, size_t size, int *fd)
1265 char fname[20] = "/tmp/memfd-XXXXXX";
1268 #ifdef MEMFD_SUPPORTED
1269 mfd = memfd_create(name, MFD_CLOEXEC);
1274 mfd = mkstemp(fname);
1276 RTE_LOG(ERR, VHOST_CONFIG,
1277 "failed to get inflight buffer fd\n");
1284 if (ftruncate(mfd, size) == -1) {
1285 RTE_LOG(ERR, VHOST_CONFIG,
1286 "failed to alloc inflight buffer\n");
1291 ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
1292 if (ptr == MAP_FAILED) {
1293 RTE_LOG(ERR, VHOST_CONFIG,
1294 "failed to mmap inflight buffer\n");
1304 get_pervq_shm_size_split(uint16_t queue_size)
1306 return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_split) *
1307 queue_size + sizeof(uint64_t) +
1308 sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
1312 get_pervq_shm_size_packed(uint16_t queue_size)
1314 return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_packed)
1315 * queue_size + sizeof(uint64_t) +
1316 sizeof(uint16_t) * 6 + sizeof(uint8_t) * 9,
1317 INFLIGHT_ALIGNMENT);
1321 vhost_user_get_inflight_fd(struct virtio_net **pdev,
1323 int main_fd __rte_unused)
1325 struct rte_vhost_inflight_info_packed *inflight_packed;
1326 uint64_t pervq_inflight_size, mmap_size;
1327 uint16_t num_queues, queue_size;
1328 struct virtio_net *dev = *pdev;
1332 if (msg->size != sizeof(msg->payload.inflight)) {
1333 RTE_LOG(ERR, VHOST_CONFIG,
1334 "invalid get_inflight_fd message size is %d\n",
1336 return RTE_VHOST_MSG_RESULT_ERR;
1339 if (dev->inflight_info == NULL) {
1340 dev->inflight_info = calloc(1,
1341 sizeof(struct inflight_mem_info));
1342 if (!dev->inflight_info) {
1343 RTE_LOG(ERR, VHOST_CONFIG,
1344 "failed to alloc dev inflight area\n");
1345 return RTE_VHOST_MSG_RESULT_ERR;
1349 num_queues = msg->payload.inflight.num_queues;
1350 queue_size = msg->payload.inflight.queue_size;
1352 RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd num_queues: %u\n",
1353 msg->payload.inflight.num_queues);
1354 RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd queue_size: %u\n",
1355 msg->payload.inflight.queue_size);
1357 if (vq_is_packed(dev))
1358 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1360 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1362 mmap_size = num_queues * pervq_inflight_size;
1363 addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
1365 RTE_LOG(ERR, VHOST_CONFIG,
1366 "failed to alloc vhost inflight area\n");
1367 msg->payload.inflight.mmap_size = 0;
1368 return RTE_VHOST_MSG_RESULT_ERR;
1370 memset(addr, 0, mmap_size);
1372 dev->inflight_info->addr = addr;
1373 dev->inflight_info->size = msg->payload.inflight.mmap_size = mmap_size;
1374 dev->inflight_info->fd = msg->fds[0] = fd;
1375 msg->payload.inflight.mmap_offset = 0;
1378 if (vq_is_packed(dev)) {
1379 for (i = 0; i < num_queues; i++) {
1381 (struct rte_vhost_inflight_info_packed *)addr;
1382 inflight_packed->used_wrap_counter = 1;
1383 inflight_packed->old_used_wrap_counter = 1;
1384 for (j = 0; j < queue_size; j++)
1385 inflight_packed->desc[j].next = j + 1;
1386 addr = (void *)((char *)addr + pervq_inflight_size);
1390 RTE_LOG(INFO, VHOST_CONFIG,
1391 "send inflight mmap_size: %"PRIu64"\n",
1392 msg->payload.inflight.mmap_size);
1393 RTE_LOG(INFO, VHOST_CONFIG,
1394 "send inflight mmap_offset: %"PRIu64"\n",
1395 msg->payload.inflight.mmap_offset);
1396 RTE_LOG(INFO, VHOST_CONFIG,
1397 "send inflight fd: %d\n", msg->fds[0]);
1399 return RTE_VHOST_MSG_RESULT_REPLY;
1403 vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
1404 int main_fd __rte_unused)
1406 uint64_t mmap_size, mmap_offset;
1407 uint16_t num_queues, queue_size;
1408 struct virtio_net *dev = *pdev;
1409 uint32_t pervq_inflight_size;
1410 struct vhost_virtqueue *vq;
1415 if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
1416 RTE_LOG(ERR, VHOST_CONFIG,
1417 "invalid set_inflight_fd message size is %d,fd is %d\n",
1419 return RTE_VHOST_MSG_RESULT_ERR;
1422 mmap_size = msg->payload.inflight.mmap_size;
1423 mmap_offset = msg->payload.inflight.mmap_offset;
1424 num_queues = msg->payload.inflight.num_queues;
1425 queue_size = msg->payload.inflight.queue_size;
1427 if (vq_is_packed(dev))
1428 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1430 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1432 RTE_LOG(INFO, VHOST_CONFIG,
1433 "set_inflight_fd mmap_size: %"PRIu64"\n", mmap_size);
1434 RTE_LOG(INFO, VHOST_CONFIG,
1435 "set_inflight_fd mmap_offset: %"PRIu64"\n", mmap_offset);
1436 RTE_LOG(INFO, VHOST_CONFIG,
1437 "set_inflight_fd num_queues: %u\n", num_queues);
1438 RTE_LOG(INFO, VHOST_CONFIG,
1439 "set_inflight_fd queue_size: %u\n", queue_size);
1440 RTE_LOG(INFO, VHOST_CONFIG,
1441 "set_inflight_fd fd: %d\n", fd);
1442 RTE_LOG(INFO, VHOST_CONFIG,
1443 "set_inflight_fd pervq_inflight_size: %d\n",
1444 pervq_inflight_size);
1446 if (!dev->inflight_info) {
1447 dev->inflight_info = calloc(1,
1448 sizeof(struct inflight_mem_info));
1449 if (dev->inflight_info == NULL) {
1450 RTE_LOG(ERR, VHOST_CONFIG,
1451 "failed to alloc dev inflight area\n");
1452 return RTE_VHOST_MSG_RESULT_ERR;
1456 if (dev->inflight_info->addr)
1457 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1459 addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1461 if (addr == MAP_FAILED) {
1462 RTE_LOG(ERR, VHOST_CONFIG, "failed to mmap share memory.\n");
1463 return RTE_VHOST_MSG_RESULT_ERR;
1466 if (dev->inflight_info->fd)
1467 close(dev->inflight_info->fd);
1469 dev->inflight_info->fd = fd;
1470 dev->inflight_info->addr = addr;
1471 dev->inflight_info->size = mmap_size;
1473 for (i = 0; i < num_queues; i++) {
1474 vq = dev->virtqueue[i];
1475 if (vq_is_packed(dev)) {
1476 vq->inflight_packed = addr;
1477 vq->inflight_packed->desc_num = queue_size;
1479 vq->inflight_split = addr;
1480 vq->inflight_split->desc_num = queue_size;
1482 addr = (void *)((char *)addr + pervq_inflight_size);
1485 return RTE_VHOST_MSG_RESULT_OK;
1489 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1490 int main_fd __rte_unused)
1492 struct virtio_net *dev = *pdev;
1493 struct vhost_vring_file file;
1494 struct vhost_virtqueue *vq;
1496 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1497 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1498 file.fd = VIRTIO_INVALID_EVENTFD;
1500 file.fd = msg->fds[0];
1501 RTE_LOG(INFO, VHOST_CONFIG,
1502 "vring call idx:%d file:%d\n", file.index, file.fd);
1504 vq = dev->virtqueue[file.index];
1505 if (vq->callfd >= 0)
1508 vq->callfd = file.fd;
1510 return RTE_VHOST_MSG_RESULT_OK;
1513 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1514 struct VhostUserMsg *msg,
1515 int main_fd __rte_unused)
1517 if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1519 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1521 return RTE_VHOST_MSG_RESULT_OK;
1525 resubmit_desc_compare(const void *a, const void *b)
1527 const struct rte_vhost_resubmit_desc *desc0 = a;
1528 const struct rte_vhost_resubmit_desc *desc1 = b;
1530 if (desc1->counter > desc0->counter)
1537 vhost_check_queue_inflights_split(struct virtio_net *dev,
1538 struct vhost_virtqueue *vq)
1541 uint16_t resubmit_num = 0, last_io, num;
1542 struct vring_used *used = vq->used;
1543 struct rte_vhost_resubmit_info *resubmit;
1544 struct rte_vhost_inflight_info_split *inflight_split;
1546 if (!(dev->protocol_features &
1547 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1548 return RTE_VHOST_MSG_RESULT_OK;
1550 if ((!vq->inflight_split))
1551 return RTE_VHOST_MSG_RESULT_ERR;
1553 if (!vq->inflight_split->version) {
1554 vq->inflight_split->version = INFLIGHT_VERSION;
1555 return RTE_VHOST_MSG_RESULT_OK;
1558 if (vq->resubmit_inflight)
1559 return RTE_VHOST_MSG_RESULT_OK;
1561 inflight_split = vq->inflight_split;
1562 vq->global_counter = 0;
1563 last_io = inflight_split->last_inflight_io;
1565 if (inflight_split->used_idx != used->idx) {
1566 inflight_split->desc[last_io].inflight = 0;
1568 inflight_split->used_idx = used->idx;
1571 for (i = 0; i < inflight_split->desc_num; i++) {
1572 if (inflight_split->desc[i].inflight == 1)
1576 vq->last_avail_idx += resubmit_num;
1579 resubmit = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1581 RTE_LOG(ERR, VHOST_CONFIG,
1582 "failed to allocate memory for resubmit info.\n");
1583 return RTE_VHOST_MSG_RESULT_ERR;
1586 resubmit->resubmit_list = calloc(resubmit_num,
1587 sizeof(struct rte_vhost_resubmit_desc));
1588 if (!resubmit->resubmit_list) {
1589 RTE_LOG(ERR, VHOST_CONFIG,
1590 "failed to allocate memory for inflight desc.\n");
1592 return RTE_VHOST_MSG_RESULT_ERR;
1596 for (i = 0; i < vq->inflight_split->desc_num; i++) {
1597 if (vq->inflight_split->desc[i].inflight == 1) {
1598 resubmit->resubmit_list[num].index = i;
1599 resubmit->resubmit_list[num].counter =
1600 inflight_split->desc[i].counter;
1604 resubmit->resubmit_num = num;
1606 if (resubmit->resubmit_num > 1)
1607 qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1608 sizeof(struct rte_vhost_resubmit_desc),
1609 resubmit_desc_compare);
1611 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1612 vq->resubmit_inflight = resubmit;
1615 return RTE_VHOST_MSG_RESULT_OK;
1619 vhost_check_queue_inflights_packed(struct virtio_net *dev,
1620 struct vhost_virtqueue *vq)
1623 uint16_t resubmit_num = 0, old_used_idx, num;
1624 struct rte_vhost_resubmit_info *resubmit;
1625 struct rte_vhost_inflight_info_packed *inflight_packed;
1627 if (!(dev->protocol_features &
1628 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1629 return RTE_VHOST_MSG_RESULT_OK;
1631 if ((!vq->inflight_packed))
1632 return RTE_VHOST_MSG_RESULT_ERR;
1634 if (!vq->inflight_packed->version) {
1635 vq->inflight_packed->version = INFLIGHT_VERSION;
1636 return RTE_VHOST_MSG_RESULT_OK;
1639 if (vq->resubmit_inflight)
1640 return RTE_VHOST_MSG_RESULT_OK;
1642 inflight_packed = vq->inflight_packed;
1643 vq->global_counter = 0;
1644 old_used_idx = inflight_packed->old_used_idx;
1646 if (inflight_packed->used_idx != old_used_idx) {
1647 if (inflight_packed->desc[old_used_idx].inflight == 0) {
1648 inflight_packed->old_used_idx =
1649 inflight_packed->used_idx;
1650 inflight_packed->old_used_wrap_counter =
1651 inflight_packed->used_wrap_counter;
1652 inflight_packed->old_free_head =
1653 inflight_packed->free_head;
1655 inflight_packed->used_idx =
1656 inflight_packed->old_used_idx;
1657 inflight_packed->used_wrap_counter =
1658 inflight_packed->old_used_wrap_counter;
1659 inflight_packed->free_head =
1660 inflight_packed->old_free_head;
1664 for (i = 0; i < inflight_packed->desc_num; i++) {
1665 if (inflight_packed->desc[i].inflight == 1)
1670 resubmit = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1671 if (resubmit == NULL) {
1672 RTE_LOG(ERR, VHOST_CONFIG,
1673 "failed to allocate memory for resubmit info.\n");
1674 return RTE_VHOST_MSG_RESULT_ERR;
1677 resubmit->resubmit_list = calloc(resubmit_num,
1678 sizeof(struct rte_vhost_resubmit_desc));
1679 if (resubmit->resubmit_list == NULL) {
1680 RTE_LOG(ERR, VHOST_CONFIG,
1681 "failed to allocate memory for resubmit desc.\n");
1683 return RTE_VHOST_MSG_RESULT_ERR;
1687 for (i = 0; i < inflight_packed->desc_num; i++) {
1688 if (vq->inflight_packed->desc[i].inflight == 1) {
1689 resubmit->resubmit_list[num].index = i;
1690 resubmit->resubmit_list[num].counter =
1691 inflight_packed->desc[i].counter;
1695 resubmit->resubmit_num = num;
1697 if (resubmit->resubmit_num > 1)
1698 qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1699 sizeof(struct rte_vhost_resubmit_desc),
1700 resubmit_desc_compare);
1702 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1703 vq->resubmit_inflight = resubmit;
1706 return RTE_VHOST_MSG_RESULT_OK;
1710 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1711 int main_fd __rte_unused)
1713 struct virtio_net *dev = *pdev;
1714 struct vhost_vring_file file;
1715 struct vhost_virtqueue *vq;
1717 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1718 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1719 file.fd = VIRTIO_INVALID_EVENTFD;
1721 file.fd = msg->fds[0];
1722 RTE_LOG(INFO, VHOST_CONFIG,
1723 "vring kick idx:%d file:%d\n", file.index, file.fd);
1725 /* Interpret ring addresses only when ring is started. */
1726 dev = translate_ring_addresses(dev, file.index);
1728 return RTE_VHOST_MSG_RESULT_ERR;
1732 vq = dev->virtqueue[file.index];
1735 * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1736 * the ring starts already enabled. Otherwise, it is enabled via
1737 * the SET_VRING_ENABLE message.
1739 if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1741 if (dev->notify_ops->vring_state_changed)
1742 dev->notify_ops->vring_state_changed(
1743 dev->vid, file.index, 1);
1746 if (vq->kickfd >= 0)
1748 vq->kickfd = file.fd;
1750 if (vq_is_packed(dev)) {
1751 if (vhost_check_queue_inflights_packed(dev, vq)) {
1752 RTE_LOG(ERR, VHOST_CONFIG,
1753 "failed to inflights for vq: %d\n", file.index);
1754 return RTE_VHOST_MSG_RESULT_ERR;
1757 if (vhost_check_queue_inflights_split(dev, vq)) {
1758 RTE_LOG(ERR, VHOST_CONFIG,
1759 "failed to inflights for vq: %d\n", file.index);
1760 return RTE_VHOST_MSG_RESULT_ERR;
1764 return RTE_VHOST_MSG_RESULT_OK;
1768 free_zmbufs(struct vhost_virtqueue *vq)
1770 drain_zmbuf_list(vq);
1772 rte_free(vq->zmbufs);
1776 * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1779 vhost_user_get_vring_base(struct virtio_net **pdev,
1780 struct VhostUserMsg *msg,
1781 int main_fd __rte_unused)
1783 struct virtio_net *dev = *pdev;
1784 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1787 /* We have to stop the queue (virtio) if it is running. */
1788 vhost_destroy_device_notify(dev);
1790 dev->flags &= ~VIRTIO_DEV_READY;
1791 dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1793 /* Here we are safe to get the indexes */
1794 if (vq_is_packed(dev)) {
1796 * Bit[0:14]: avail index
1797 * Bit[15]: avail wrap counter
1799 val = vq->last_avail_idx & 0x7fff;
1800 val |= vq->avail_wrap_counter << 15;
1801 msg->payload.state.num = val;
1803 msg->payload.state.num = vq->last_avail_idx;
1806 RTE_LOG(INFO, VHOST_CONFIG,
1807 "vring base idx:%d file:%d\n", msg->payload.state.index,
1808 msg->payload.state.num);
1810 * Based on current qemu vhost-user implementation, this message is
1811 * sent and only sent in vhost_vring_stop.
1812 * TODO: cleanup the vring, it isn't usable since here.
1814 if (vq->kickfd >= 0)
1817 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1819 if (vq->callfd >= 0)
1822 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1824 vq->signalled_used_valid = false;
1826 if (dev->dequeue_zero_copy)
1828 if (vq_is_packed(dev)) {
1829 rte_free(vq->shadow_used_packed);
1830 vq->shadow_used_packed = NULL;
1832 rte_free(vq->shadow_used_split);
1833 vq->shadow_used_split = NULL;
1836 rte_free(vq->batch_copy_elems);
1837 vq->batch_copy_elems = NULL;
1839 msg->size = sizeof(msg->payload.state);
1842 vring_invalidate(dev, vq);
1844 return RTE_VHOST_MSG_RESULT_REPLY;
1848 * when virtio queues are ready to work, qemu will send us to
1849 * enable the virtio queue pair.
1852 vhost_user_set_vring_enable(struct virtio_net **pdev,
1853 struct VhostUserMsg *msg,
1854 int main_fd __rte_unused)
1856 struct virtio_net *dev = *pdev;
1857 int enable = (int)msg->payload.state.num;
1858 int index = (int)msg->payload.state.index;
1859 struct rte_vdpa_device *vdpa_dev;
1862 RTE_LOG(INFO, VHOST_CONFIG,
1863 "set queue enable: %d to qp idx: %d\n",
1866 did = dev->vdpa_dev_id;
1867 vdpa_dev = rte_vdpa_get_device(did);
1868 if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1869 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1871 if (dev->notify_ops->vring_state_changed)
1872 dev->notify_ops->vring_state_changed(dev->vid,
1875 /* On disable, rings have to be stopped being processed. */
1876 if (!enable && dev->dequeue_zero_copy)
1877 drain_zmbuf_list(dev->virtqueue[index]);
1879 dev->virtqueue[index]->enabled = enable;
1881 return RTE_VHOST_MSG_RESULT_OK;
1885 vhost_user_get_protocol_features(struct virtio_net **pdev,
1886 struct VhostUserMsg *msg,
1887 int main_fd __rte_unused)
1889 struct virtio_net *dev = *pdev;
1890 uint64_t features, protocol_features;
1892 rte_vhost_driver_get_features(dev->ifname, &features);
1893 rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1896 * REPLY_ACK protocol feature is only mandatory for now
1897 * for IOMMU feature. If IOMMU is explicitly disabled by the
1898 * application, disable also REPLY_ACK feature for older buggy
1899 * Qemu versions (from v2.7.0 to v2.9.0).
1901 if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1902 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1904 msg->payload.u64 = protocol_features;
1905 msg->size = sizeof(msg->payload.u64);
1908 return RTE_VHOST_MSG_RESULT_REPLY;
1912 vhost_user_set_protocol_features(struct virtio_net **pdev,
1913 struct VhostUserMsg *msg,
1914 int main_fd __rte_unused)
1916 struct virtio_net *dev = *pdev;
1917 uint64_t protocol_features = msg->payload.u64;
1918 uint64_t slave_protocol_features = 0;
1920 rte_vhost_driver_get_protocol_features(dev->ifname,
1921 &slave_protocol_features);
1922 if (protocol_features & ~slave_protocol_features) {
1923 RTE_LOG(ERR, VHOST_CONFIG,
1924 "(%d) received invalid protocol features.\n",
1926 return RTE_VHOST_MSG_RESULT_ERR;
1929 dev->protocol_features = protocol_features;
1930 RTE_LOG(INFO, VHOST_CONFIG,
1931 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
1932 dev->protocol_features);
1934 return RTE_VHOST_MSG_RESULT_OK;
1938 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1939 int main_fd __rte_unused)
1941 struct virtio_net *dev = *pdev;
1942 int fd = msg->fds[0];
1947 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1948 return RTE_VHOST_MSG_RESULT_ERR;
1951 if (msg->size != sizeof(VhostUserLog)) {
1952 RTE_LOG(ERR, VHOST_CONFIG,
1953 "invalid log base msg size: %"PRId32" != %d\n",
1954 msg->size, (int)sizeof(VhostUserLog));
1955 return RTE_VHOST_MSG_RESULT_ERR;
1958 size = msg->payload.log.mmap_size;
1959 off = msg->payload.log.mmap_offset;
1961 /* Don't allow mmap_offset to point outside the mmap region */
1963 RTE_LOG(ERR, VHOST_CONFIG,
1964 "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1966 return RTE_VHOST_MSG_RESULT_ERR;
1969 RTE_LOG(INFO, VHOST_CONFIG,
1970 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1974 * mmap from 0 to workaround a hugepage mmap bug: mmap will
1975 * fail when offset is not page size aligned.
1977 addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1979 if (addr == MAP_FAILED) {
1980 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1981 return RTE_VHOST_MSG_RESULT_ERR;
1985 * Free previously mapped log memory on occasionally
1986 * multiple VHOST_USER_SET_LOG_BASE.
1988 if (dev->log_addr) {
1989 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1991 dev->log_addr = (uint64_t)(uintptr_t)addr;
1992 dev->log_base = dev->log_addr + off;
1993 dev->log_size = size;
1996 * The spec is not clear about it (yet), but QEMU doesn't expect
1997 * any payload in the reply.
2002 return RTE_VHOST_MSG_RESULT_REPLY;
2005 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
2006 struct VhostUserMsg *msg,
2007 int main_fd __rte_unused)
2010 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
2012 return RTE_VHOST_MSG_RESULT_OK;
2016 * An rarp packet is constructed and broadcasted to notify switches about
2017 * the new location of the migrated VM, so that packets from outside will
2018 * not be lost after migration.
2020 * However, we don't actually "send" a rarp packet here, instead, we set
2021 * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
2024 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
2025 int main_fd __rte_unused)
2027 struct virtio_net *dev = *pdev;
2028 uint8_t *mac = (uint8_t *)&msg->payload.u64;
2029 struct rte_vdpa_device *vdpa_dev;
2032 RTE_LOG(DEBUG, VHOST_CONFIG,
2033 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
2034 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
2035 memcpy(dev->mac.addr_bytes, mac, 6);
2038 * Set the flag to inject a RARP broadcast packet at
2039 * rte_vhost_dequeue_burst().
2041 * rte_smp_wmb() is for making sure the mac is copied
2042 * before the flag is set.
2045 rte_atomic16_set(&dev->broadcast_rarp, 1);
2046 did = dev->vdpa_dev_id;
2047 vdpa_dev = rte_vdpa_get_device(did);
2048 if (vdpa_dev && vdpa_dev->ops->migration_done)
2049 vdpa_dev->ops->migration_done(dev->vid);
2051 return RTE_VHOST_MSG_RESULT_OK;
2055 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
2056 int main_fd __rte_unused)
2058 struct virtio_net *dev = *pdev;
2059 if (msg->payload.u64 < VIRTIO_MIN_MTU ||
2060 msg->payload.u64 > VIRTIO_MAX_MTU) {
2061 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
2064 return RTE_VHOST_MSG_RESULT_ERR;
2067 dev->mtu = msg->payload.u64;
2069 return RTE_VHOST_MSG_RESULT_OK;
2073 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
2074 int main_fd __rte_unused)
2076 struct virtio_net *dev = *pdev;
2077 int fd = msg->fds[0];
2080 RTE_LOG(ERR, VHOST_CONFIG,
2081 "Invalid file descriptor for slave channel (%d)\n",
2083 return RTE_VHOST_MSG_RESULT_ERR;
2086 if (dev->slave_req_fd >= 0)
2087 close(dev->slave_req_fd);
2089 dev->slave_req_fd = fd;
2091 return RTE_VHOST_MSG_RESULT_OK;
2095 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2097 struct vhost_vring_addr *ra;
2098 uint64_t start, end, len;
2101 end = start + imsg->size;
2103 ra = &vq->ring_addrs;
2104 len = sizeof(struct vring_desc) * vq->size;
2105 if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2108 len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
2109 if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2112 len = sizeof(struct vring_used) +
2113 sizeof(struct vring_used_elem) * vq->size;
2114 if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2121 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2123 struct vhost_vring_addr *ra;
2124 uint64_t start, end, len;
2127 end = start + imsg->size;
2129 ra = &vq->ring_addrs;
2130 len = sizeof(struct vring_packed_desc) * vq->size;
2131 if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2134 len = sizeof(struct vring_packed_desc_event);
2135 if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2138 len = sizeof(struct vring_packed_desc_event);
2139 if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2145 static int is_vring_iotlb(struct virtio_net *dev,
2146 struct vhost_virtqueue *vq,
2147 struct vhost_iotlb_msg *imsg)
2149 if (vq_is_packed(dev))
2150 return is_vring_iotlb_packed(vq, imsg);
2152 return is_vring_iotlb_split(vq, imsg);
2156 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
2157 int main_fd __rte_unused)
2159 struct virtio_net *dev = *pdev;
2160 struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
2164 switch (imsg->type) {
2165 case VHOST_IOTLB_UPDATE:
2167 vva = qva_to_vva(dev, imsg->uaddr, &len);
2169 return RTE_VHOST_MSG_RESULT_ERR;
2171 for (i = 0; i < dev->nr_vring; i++) {
2172 struct vhost_virtqueue *vq = dev->virtqueue[i];
2174 vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
2177 if (is_vring_iotlb(dev, vq, imsg))
2178 *pdev = dev = translate_ring_addresses(dev, i);
2181 case VHOST_IOTLB_INVALIDATE:
2182 for (i = 0; i < dev->nr_vring; i++) {
2183 struct vhost_virtqueue *vq = dev->virtqueue[i];
2185 vhost_user_iotlb_cache_remove(vq, imsg->iova,
2188 if (is_vring_iotlb(dev, vq, imsg))
2189 vring_invalidate(dev, vq);
2193 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
2195 return RTE_VHOST_MSG_RESULT_ERR;
2198 return RTE_VHOST_MSG_RESULT_OK;
2202 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
2203 struct VhostUserMsg *msg,
2204 int main_fd __rte_unused)
2206 struct virtio_net *dev = *pdev;
2207 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
2208 struct uffdio_api api_struct;
2210 dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
2212 if (dev->postcopy_ufd == -1) {
2213 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
2215 return RTE_VHOST_MSG_RESULT_ERR;
2217 api_struct.api = UFFD_API;
2218 api_struct.features = 0;
2219 if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
2220 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
2222 close(dev->postcopy_ufd);
2223 dev->postcopy_ufd = -1;
2224 return RTE_VHOST_MSG_RESULT_ERR;
2226 msg->fds[0] = dev->postcopy_ufd;
2229 return RTE_VHOST_MSG_RESULT_REPLY;
2231 dev->postcopy_ufd = -1;
2234 return RTE_VHOST_MSG_RESULT_ERR;
2239 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
2240 struct VhostUserMsg *msg __rte_unused,
2241 int main_fd __rte_unused)
2243 struct virtio_net *dev = *pdev;
2245 if (dev->mem && dev->mem->nregions) {
2246 RTE_LOG(ERR, VHOST_CONFIG,
2247 "Regions already registered at postcopy-listen\n");
2248 return RTE_VHOST_MSG_RESULT_ERR;
2250 dev->postcopy_listening = 1;
2252 return RTE_VHOST_MSG_RESULT_OK;
2256 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
2257 int main_fd __rte_unused)
2259 struct virtio_net *dev = *pdev;
2261 dev->postcopy_listening = 0;
2262 if (dev->postcopy_ufd >= 0) {
2263 close(dev->postcopy_ufd);
2264 dev->postcopy_ufd = -1;
2267 msg->payload.u64 = 0;
2268 msg->size = sizeof(msg->payload.u64);
2271 return RTE_VHOST_MSG_RESULT_REPLY;
2274 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
2275 struct VhostUserMsg *msg,
2277 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
2278 [VHOST_USER_NONE] = NULL,
2279 [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
2280 [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
2281 [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
2282 [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
2283 [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
2284 [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
2285 [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
2286 [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
2287 [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
2288 [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
2289 [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
2290 [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
2291 [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
2292 [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
2293 [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
2294 [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
2295 [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
2296 [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
2297 [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
2298 [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
2299 [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
2300 [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
2301 [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
2302 [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
2303 [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
2304 [VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
2305 [VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
2308 /* return bytes# of read on success or negative val on failure. */
2310 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
2314 ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
2315 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
2320 if (msg->size > sizeof(msg->payload)) {
2321 RTE_LOG(ERR, VHOST_CONFIG,
2322 "invalid msg size: %d\n", msg->size);
2325 ret = read(sockfd, &msg->payload, msg->size);
2328 if (ret != (int)msg->size) {
2329 RTE_LOG(ERR, VHOST_CONFIG,
2330 "read control message failed\n");
2339 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
2344 return send_fd_message(sockfd, (char *)msg,
2345 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
2349 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
2354 msg->flags &= ~VHOST_USER_VERSION_MASK;
2355 msg->flags &= ~VHOST_USER_NEED_REPLY;
2356 msg->flags |= VHOST_USER_VERSION;
2357 msg->flags |= VHOST_USER_REPLY_MASK;
2359 return send_vhost_message(sockfd, msg);
2363 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
2367 if (msg->flags & VHOST_USER_NEED_REPLY)
2368 rte_spinlock_lock(&dev->slave_req_lock);
2370 ret = send_vhost_message(dev->slave_req_fd, msg);
2371 if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
2372 rte_spinlock_unlock(&dev->slave_req_lock);
2378 * Allocate a queue pair if it hasn't been allocated yet
2381 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
2382 struct VhostUserMsg *msg)
2386 switch (msg->request.master) {
2387 case VHOST_USER_SET_VRING_KICK:
2388 case VHOST_USER_SET_VRING_CALL:
2389 case VHOST_USER_SET_VRING_ERR:
2390 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2392 case VHOST_USER_SET_VRING_NUM:
2393 case VHOST_USER_SET_VRING_BASE:
2394 case VHOST_USER_SET_VRING_ENABLE:
2395 vring_idx = msg->payload.state.index;
2397 case VHOST_USER_SET_VRING_ADDR:
2398 vring_idx = msg->payload.addr.index;
2404 if (vring_idx >= VHOST_MAX_VRING) {
2405 RTE_LOG(ERR, VHOST_CONFIG,
2406 "invalid vring index: %u\n", vring_idx);
2410 if (dev->virtqueue[vring_idx])
2413 return alloc_vring_queue(dev, vring_idx);
2417 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
2420 unsigned int vq_num = 0;
2422 while (vq_num < dev->nr_vring) {
2423 struct vhost_virtqueue *vq = dev->virtqueue[i];
2426 rte_spinlock_lock(&vq->access_lock);
2434 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
2437 unsigned int vq_num = 0;
2439 while (vq_num < dev->nr_vring) {
2440 struct vhost_virtqueue *vq = dev->virtqueue[i];
2443 rte_spinlock_unlock(&vq->access_lock);
2451 vhost_user_msg_handler(int vid, int fd)
2453 struct virtio_net *dev;
2454 struct VhostUserMsg msg;
2455 struct rte_vdpa_device *vdpa_dev;
2458 int unlock_required = 0;
2462 dev = get_device(vid);
2466 if (!dev->notify_ops) {
2467 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2468 if (!dev->notify_ops) {
2469 RTE_LOG(ERR, VHOST_CONFIG,
2470 "failed to get callback ops for driver %s\n",
2476 ret = read_vhost_message(fd, &msg);
2479 RTE_LOG(ERR, VHOST_CONFIG,
2480 "vhost read message failed\n");
2482 RTE_LOG(INFO, VHOST_CONFIG,
2483 "vhost peer closed\n");
2489 request = msg.request.master;
2490 if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
2491 vhost_message_str[request]) {
2492 if (request != VHOST_USER_IOTLB_MSG)
2493 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
2494 vhost_message_str[request]);
2496 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
2497 vhost_message_str[request]);
2499 RTE_LOG(DEBUG, VHOST_CONFIG, "External request %d\n", request);
2502 ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
2504 RTE_LOG(ERR, VHOST_CONFIG,
2505 "failed to alloc queue\n");
2510 * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2511 * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2512 * and device is destroyed. destroy_device waits for queues to be
2513 * inactive, so it is safe. Otherwise taking the access_lock
2514 * would cause a dead lock.
2517 case VHOST_USER_SET_FEATURES:
2518 case VHOST_USER_SET_PROTOCOL_FEATURES:
2519 case VHOST_USER_SET_OWNER:
2520 case VHOST_USER_SET_MEM_TABLE:
2521 case VHOST_USER_SET_LOG_BASE:
2522 case VHOST_USER_SET_LOG_FD:
2523 case VHOST_USER_SET_VRING_NUM:
2524 case VHOST_USER_SET_VRING_ADDR:
2525 case VHOST_USER_SET_VRING_BASE:
2526 case VHOST_USER_SET_VRING_KICK:
2527 case VHOST_USER_SET_VRING_CALL:
2528 case VHOST_USER_SET_VRING_ERR:
2529 case VHOST_USER_SET_VRING_ENABLE:
2530 case VHOST_USER_SEND_RARP:
2531 case VHOST_USER_NET_SET_MTU:
2532 case VHOST_USER_SET_SLAVE_REQ_FD:
2533 vhost_user_lock_all_queue_pairs(dev);
2534 unlock_required = 1;
2542 if (dev->extern_ops.pre_msg_handle) {
2543 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2546 case RTE_VHOST_MSG_RESULT_REPLY:
2547 send_vhost_reply(fd, &msg);
2549 case RTE_VHOST_MSG_RESULT_ERR:
2550 case RTE_VHOST_MSG_RESULT_OK:
2552 goto skip_to_post_handle;
2553 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2559 if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2560 if (!vhost_message_handlers[request])
2561 goto skip_to_post_handle;
2562 ret = vhost_message_handlers[request](&dev, &msg, fd);
2565 case RTE_VHOST_MSG_RESULT_ERR:
2566 RTE_LOG(ERR, VHOST_CONFIG,
2567 "Processing %s failed.\n",
2568 vhost_message_str[request]);
2571 case RTE_VHOST_MSG_RESULT_OK:
2572 RTE_LOG(DEBUG, VHOST_CONFIG,
2573 "Processing %s succeeded.\n",
2574 vhost_message_str[request]);
2577 case RTE_VHOST_MSG_RESULT_REPLY:
2578 RTE_LOG(DEBUG, VHOST_CONFIG,
2579 "Processing %s succeeded and needs reply.\n",
2580 vhost_message_str[request]);
2581 send_vhost_reply(fd, &msg);
2589 skip_to_post_handle:
2590 if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2591 dev->extern_ops.post_msg_handle) {
2592 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2595 case RTE_VHOST_MSG_RESULT_REPLY:
2596 send_vhost_reply(fd, &msg);
2598 case RTE_VHOST_MSG_RESULT_ERR:
2599 case RTE_VHOST_MSG_RESULT_OK:
2601 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2607 if (unlock_required)
2608 vhost_user_unlock_all_queue_pairs(dev);
2610 /* If message was not handled at this stage, treat it as an error */
2612 RTE_LOG(ERR, VHOST_CONFIG,
2613 "vhost message (req: %d) was not handled.\n", request);
2614 ret = RTE_VHOST_MSG_RESULT_ERR;
2618 * If the request required a reply that was already sent,
2619 * this optional reply-ack won't be sent as the
2620 * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2622 if (msg.flags & VHOST_USER_NEED_REPLY) {
2623 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2624 msg.size = sizeof(msg.payload.u64);
2626 send_vhost_reply(fd, &msg);
2627 } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2628 RTE_LOG(ERR, VHOST_CONFIG,
2629 "vhost message handling failed.\n");
2633 if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2634 dev->flags |= VIRTIO_DEV_READY;
2636 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2637 if (dev->dequeue_zero_copy) {
2638 RTE_LOG(INFO, VHOST_CONFIG,
2639 "dequeue zero copy is enabled\n");
2642 if (dev->notify_ops->new_device(dev->vid) == 0)
2643 dev->flags |= VIRTIO_DEV_RUNNING;
2647 did = dev->vdpa_dev_id;
2648 vdpa_dev = rte_vdpa_get_device(did);
2649 if (vdpa_dev && virtio_is_ready(dev) &&
2650 !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2651 msg.request.master == VHOST_USER_SET_VRING_CALL) {
2652 if (vdpa_dev->ops->dev_conf)
2653 vdpa_dev->ops->dev_conf(dev->vid);
2654 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2660 static int process_slave_message_reply(struct virtio_net *dev,
2661 const struct VhostUserMsg *msg)
2663 struct VhostUserMsg msg_reply;
2666 if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2669 if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2674 if (msg_reply.request.slave != msg->request.slave) {
2675 RTE_LOG(ERR, VHOST_CONFIG,
2676 "Received unexpected msg type (%u), expected %u\n",
2677 msg_reply.request.slave, msg->request.slave);
2682 ret = msg_reply.payload.u64 ? -1 : 0;
2685 rte_spinlock_unlock(&dev->slave_req_lock);
2690 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2693 struct VhostUserMsg msg = {
2694 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2695 .flags = VHOST_USER_VERSION,
2696 .size = sizeof(msg.payload.iotlb),
2700 .type = VHOST_IOTLB_MISS,
2704 ret = send_vhost_message(dev->slave_req_fd, &msg);
2706 RTE_LOG(ERR, VHOST_CONFIG,
2707 "Failed to send IOTLB miss message (%d)\n",
2715 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2721 struct VhostUserMsg msg = {
2722 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2723 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2724 .size = sizeof(msg.payload.area),
2726 .u64 = index & VHOST_USER_VRING_IDX_MASK,
2733 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2739 ret = send_vhost_slave_message(dev, &msg);
2741 RTE_LOG(ERR, VHOST_CONFIG,
2742 "Failed to set host notifier (%d)\n", ret);
2746 return process_slave_message_reply(dev, &msg);
2749 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2751 struct virtio_net *dev;
2752 struct rte_vdpa_device *vdpa_dev;
2753 int vfio_device_fd, did, ret = 0;
2754 uint64_t offset, size;
2757 dev = get_device(vid);
2761 did = dev->vdpa_dev_id;
2765 if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2766 !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2767 !(dev->protocol_features &
2768 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2769 !(dev->protocol_features &
2770 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2771 !(dev->protocol_features &
2772 (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2775 vdpa_dev = rte_vdpa_get_device(did);
2779 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2780 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2782 vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2783 if (vfio_device_fd < 0)
2787 for (i = 0; i < dev->nr_vring; i++) {
2788 if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2794 if (vhost_user_slave_set_vring_host_notifier(dev, i,
2795 vfio_device_fd, offset, size) < 0) {
2802 for (i = 0; i < dev->nr_vring; i++) {
2803 vhost_user_slave_set_vring_host_notifier(dev, i, -1,