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);
96 close_msg_fds(struct VhostUserMsg *msg)
100 for (i = 0; i < msg->fd_num; i++)
105 * Ensure the expected number of FDs is received,
106 * close all FDs and return an error if this is not the case.
109 validate_msg_fds(struct VhostUserMsg *msg, int expected_fds)
111 if (msg->fd_num == expected_fds)
114 VHOST_LOG_CONFIG(ERR,
115 " Expect %d FDs for request %s, received %d\n",
117 vhost_message_str[msg->request.master],
131 ret = fstat(fd, &stat);
132 return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
136 * Reclaim all the outstanding zmbufs for a virtqueue.
139 drain_zmbuf_list(struct vhost_virtqueue *vq)
141 struct zcopy_mbuf *zmbuf, *next;
143 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
144 zmbuf != NULL; zmbuf = next) {
145 next = TAILQ_NEXT(zmbuf, next);
147 while (!mbuf_is_consumed(zmbuf->mbuf))
150 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
151 restore_mbuf(zmbuf->mbuf);
152 rte_pktmbuf_free(zmbuf->mbuf);
159 free_mem_region(struct virtio_net *dev)
162 struct rte_vhost_mem_region *reg;
163 struct vhost_virtqueue *vq;
165 if (!dev || !dev->mem)
168 if (dev->dequeue_zero_copy) {
169 for (i = 0; i < dev->nr_vring; i++) {
170 vq = dev->virtqueue[i];
172 drain_zmbuf_list(vq);
176 for (i = 0; i < dev->mem->nregions; i++) {
177 reg = &dev->mem->regions[i];
178 if (reg->host_user_addr) {
179 munmap(reg->mmap_addr, reg->mmap_size);
186 vhost_backend_cleanup(struct virtio_net *dev)
189 free_mem_region(dev);
194 rte_free(dev->guest_pages);
195 dev->guest_pages = NULL;
198 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
202 if (dev->inflight_info) {
203 if (dev->inflight_info->addr) {
204 munmap(dev->inflight_info->addr,
205 dev->inflight_info->size);
206 dev->inflight_info->addr = NULL;
209 if (dev->inflight_info->fd > 0) {
210 close(dev->inflight_info->fd);
211 dev->inflight_info->fd = -1;
214 free(dev->inflight_info);
215 dev->inflight_info = NULL;
218 if (dev->slave_req_fd >= 0) {
219 close(dev->slave_req_fd);
220 dev->slave_req_fd = -1;
223 if (dev->postcopy_ufd >= 0) {
224 close(dev->postcopy_ufd);
225 dev->postcopy_ufd = -1;
228 dev->postcopy_listening = 0;
232 * This function just returns success at the moment unless
233 * the device hasn't been initialised.
236 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
237 struct VhostUserMsg *msg,
238 int main_fd __rte_unused)
240 if (validate_msg_fds(msg, 0) != 0)
241 return RTE_VHOST_MSG_RESULT_ERR;
243 return RTE_VHOST_MSG_RESULT_OK;
247 vhost_user_reset_owner(struct virtio_net **pdev,
248 struct VhostUserMsg *msg,
249 int main_fd __rte_unused)
251 struct virtio_net *dev = *pdev;
253 if (validate_msg_fds(msg, 0) != 0)
254 return RTE_VHOST_MSG_RESULT_ERR;
256 vhost_destroy_device_notify(dev);
258 cleanup_device(dev, 0);
260 return RTE_VHOST_MSG_RESULT_OK;
264 * The features that we support are requested.
267 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
268 int main_fd __rte_unused)
270 struct virtio_net *dev = *pdev;
271 uint64_t features = 0;
273 if (validate_msg_fds(msg, 0) != 0)
274 return RTE_VHOST_MSG_RESULT_ERR;
276 rte_vhost_driver_get_features(dev->ifname, &features);
278 msg->payload.u64 = features;
279 msg->size = sizeof(msg->payload.u64);
282 return RTE_VHOST_MSG_RESULT_REPLY;
286 * The queue number that we support are requested.
289 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
290 int main_fd __rte_unused)
292 struct virtio_net *dev = *pdev;
293 uint32_t queue_num = 0;
295 if (validate_msg_fds(msg, 0) != 0)
296 return RTE_VHOST_MSG_RESULT_ERR;
298 rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
300 msg->payload.u64 = (uint64_t)queue_num;
301 msg->size = sizeof(msg->payload.u64);
304 return RTE_VHOST_MSG_RESULT_REPLY;
308 * We receive the negotiated features supported by us and the virtio device.
311 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
312 int main_fd __rte_unused)
314 struct virtio_net *dev = *pdev;
315 uint64_t features = msg->payload.u64;
316 uint64_t vhost_features = 0;
317 struct rte_vdpa_device *vdpa_dev;
320 if (validate_msg_fds(msg, 0) != 0)
321 return RTE_VHOST_MSG_RESULT_ERR;
323 rte_vhost_driver_get_features(dev->ifname, &vhost_features);
324 if (features & ~vhost_features) {
325 VHOST_LOG_CONFIG(ERR,
326 "(%d) received invalid negotiated features.\n",
328 return RTE_VHOST_MSG_RESULT_ERR;
331 if (dev->flags & VIRTIO_DEV_RUNNING) {
332 if (dev->features == features)
333 return RTE_VHOST_MSG_RESULT_OK;
336 * Error out if master tries to change features while device is
337 * in running state. The exception being VHOST_F_LOG_ALL, which
338 * is enabled when the live-migration starts.
340 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
341 VHOST_LOG_CONFIG(ERR,
342 "(%d) features changed while device is running.\n",
344 return RTE_VHOST_MSG_RESULT_ERR;
347 if (dev->notify_ops->features_changed)
348 dev->notify_ops->features_changed(dev->vid, features);
351 dev->features = features;
353 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
354 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
356 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
358 VHOST_LOG_CONFIG(INFO,
359 "negotiated Virtio features: 0x%" PRIx64 "\n", dev->features);
360 VHOST_LOG_CONFIG(DEBUG,
361 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
363 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
364 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
366 if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
367 !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
369 * Remove all but first queue pair if MQ hasn't been
370 * negotiated. This is safe because the device is not
371 * running at this stage.
373 while (dev->nr_vring > 2) {
374 struct vhost_virtqueue *vq;
376 vq = dev->virtqueue[--dev->nr_vring];
380 dev->virtqueue[dev->nr_vring] = NULL;
382 cleanup_vq_inflight(dev, vq);
387 did = dev->vdpa_dev_id;
388 vdpa_dev = rte_vdpa_get_device(did);
389 if (vdpa_dev && vdpa_dev->ops->set_features)
390 vdpa_dev->ops->set_features(dev->vid);
392 return RTE_VHOST_MSG_RESULT_OK;
396 * The virtio device sends us the size of the descriptor ring.
399 vhost_user_set_vring_num(struct virtio_net **pdev,
400 struct VhostUserMsg *msg,
401 int main_fd __rte_unused)
403 struct virtio_net *dev = *pdev;
404 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
406 if (validate_msg_fds(msg, 0) != 0)
407 return RTE_VHOST_MSG_RESULT_ERR;
409 vq->size = msg->payload.state.num;
411 /* VIRTIO 1.0, 2.4 Virtqueues says:
413 * Queue Size value is always a power of 2. The maximum Queue Size
416 * VIRTIO 1.1 2.7 Virtqueues says:
418 * Packed virtqueues support up to 2^15 entries each.
420 if (!vq_is_packed(dev)) {
421 if (vq->size & (vq->size - 1)) {
422 VHOST_LOG_CONFIG(ERR,
423 "invalid virtqueue size %u\n", vq->size);
424 return RTE_VHOST_MSG_RESULT_ERR;
428 if (vq->size > 32768) {
429 VHOST_LOG_CONFIG(ERR,
430 "invalid virtqueue size %u\n", vq->size);
431 return RTE_VHOST_MSG_RESULT_ERR;
434 if (dev->dequeue_zero_copy) {
436 vq->last_zmbuf_idx = 0;
437 vq->zmbuf_size = vq->size;
439 rte_free(vq->zmbufs);
440 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
441 sizeof(struct zcopy_mbuf), 0);
442 if (vq->zmbufs == NULL) {
443 VHOST_LOG_CONFIG(WARNING,
444 "failed to allocate mem for zero copy; "
445 "zero copy is force disabled\n");
446 dev->dequeue_zero_copy = 0;
448 TAILQ_INIT(&vq->zmbuf_list);
451 if (vq_is_packed(dev)) {
452 if (vq->shadow_used_packed)
453 rte_free(vq->shadow_used_packed);
454 vq->shadow_used_packed = rte_malloc(NULL,
456 sizeof(struct vring_used_elem_packed),
457 RTE_CACHE_LINE_SIZE);
458 if (!vq->shadow_used_packed) {
459 VHOST_LOG_CONFIG(ERR,
460 "failed to allocate memory for shadow used ring.\n");
461 return RTE_VHOST_MSG_RESULT_ERR;
465 if (vq->shadow_used_split)
466 rte_free(vq->shadow_used_split);
467 vq->shadow_used_split = rte_malloc(NULL,
468 vq->size * sizeof(struct vring_used_elem),
469 RTE_CACHE_LINE_SIZE);
470 if (!vq->shadow_used_split) {
471 VHOST_LOG_CONFIG(ERR,
472 "failed to allocate memory for shadow used ring.\n");
473 return RTE_VHOST_MSG_RESULT_ERR;
477 if (vq->batch_copy_elems)
478 rte_free(vq->batch_copy_elems);
479 vq->batch_copy_elems = rte_malloc(NULL,
480 vq->size * sizeof(struct batch_copy_elem),
481 RTE_CACHE_LINE_SIZE);
482 if (!vq->batch_copy_elems) {
483 VHOST_LOG_CONFIG(ERR,
484 "failed to allocate memory for batching copy.\n");
485 return RTE_VHOST_MSG_RESULT_ERR;
488 return RTE_VHOST_MSG_RESULT_OK;
492 * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
493 * same numa node as the memory of vring descriptor.
495 #ifdef RTE_LIBRTE_VHOST_NUMA
496 static struct virtio_net*
497 numa_realloc(struct virtio_net *dev, int index)
499 int oldnode, newnode;
500 struct virtio_net *old_dev;
501 struct vhost_virtqueue *old_vq, *vq;
502 struct zcopy_mbuf *new_zmbuf;
503 struct vring_used_elem *new_shadow_used_split;
504 struct vring_used_elem_packed *new_shadow_used_packed;
505 struct batch_copy_elem *new_batch_copy_elems;
508 if (dev->flags & VIRTIO_DEV_RUNNING)
512 vq = old_vq = dev->virtqueue[index];
514 ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
515 MPOL_F_NODE | MPOL_F_ADDR);
517 /* check if we need to reallocate vq */
518 ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
519 MPOL_F_NODE | MPOL_F_ADDR);
521 VHOST_LOG_CONFIG(ERR,
522 "Unable to get vq numa information.\n");
525 if (oldnode != newnode) {
526 VHOST_LOG_CONFIG(INFO,
527 "reallocate vq from %d to %d node\n", oldnode, newnode);
528 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
532 memcpy(vq, old_vq, sizeof(*vq));
533 TAILQ_INIT(&vq->zmbuf_list);
535 if (dev->dequeue_zero_copy) {
536 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
537 sizeof(struct zcopy_mbuf), 0, newnode);
539 rte_free(vq->zmbufs);
540 vq->zmbufs = new_zmbuf;
544 if (vq_is_packed(dev)) {
545 new_shadow_used_packed = rte_malloc_socket(NULL,
547 sizeof(struct vring_used_elem_packed),
550 if (new_shadow_used_packed) {
551 rte_free(vq->shadow_used_packed);
552 vq->shadow_used_packed = new_shadow_used_packed;
555 new_shadow_used_split = rte_malloc_socket(NULL,
557 sizeof(struct vring_used_elem),
560 if (new_shadow_used_split) {
561 rte_free(vq->shadow_used_split);
562 vq->shadow_used_split = new_shadow_used_split;
566 new_batch_copy_elems = rte_malloc_socket(NULL,
567 vq->size * sizeof(struct batch_copy_elem),
570 if (new_batch_copy_elems) {
571 rte_free(vq->batch_copy_elems);
572 vq->batch_copy_elems = new_batch_copy_elems;
578 /* check if we need to reallocate dev */
579 ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
580 MPOL_F_NODE | MPOL_F_ADDR);
582 VHOST_LOG_CONFIG(ERR,
583 "Unable to get dev numa information.\n");
586 if (oldnode != newnode) {
587 VHOST_LOG_CONFIG(INFO,
588 "reallocate dev from %d to %d node\n",
590 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
596 memcpy(dev, old_dev, sizeof(*dev));
601 dev->virtqueue[index] = vq;
602 vhost_devices[dev->vid] = dev;
605 vhost_user_iotlb_init(dev, index);
610 static struct virtio_net*
611 numa_realloc(struct virtio_net *dev, int index __rte_unused)
617 /* Converts QEMU virtual address to Vhost virtual address. */
619 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
621 struct rte_vhost_mem_region *r;
624 if (unlikely(!dev || !dev->mem))
627 /* Find the region where the address lives. */
628 for (i = 0; i < dev->mem->nregions; i++) {
629 r = &dev->mem->regions[i];
631 if (qva >= r->guest_user_addr &&
632 qva < r->guest_user_addr + r->size) {
634 if (unlikely(*len > r->guest_user_addr + r->size - qva))
635 *len = r->guest_user_addr + r->size - qva;
637 return qva - r->guest_user_addr +
649 * Converts ring address to Vhost virtual address.
650 * If IOMMU is enabled, the ring address is a guest IO virtual address,
651 * else it is a QEMU virtual address.
654 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
655 uint64_t ra, uint64_t *size)
657 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
660 vhost_user_iotlb_rd_lock(vq);
661 vva = vhost_iova_to_vva(dev, vq, ra,
662 size, VHOST_ACCESS_RW);
663 vhost_user_iotlb_rd_unlock(vq);
668 return qva_to_vva(dev, ra, size);
672 log_addr_to_gpa(struct virtio_net *dev, struct vhost_virtqueue *vq)
676 vhost_user_iotlb_rd_lock(vq);
677 log_gpa = translate_log_addr(dev, vq, vq->ring_addrs.log_guest_addr);
678 vhost_user_iotlb_rd_unlock(vq);
683 static struct virtio_net *
684 translate_ring_addresses(struct virtio_net *dev, int vq_index)
686 struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
687 struct vhost_vring_addr *addr = &vq->ring_addrs;
688 uint64_t len, expected_len;
690 if (addr->flags & (1 << VHOST_VRING_F_LOG)) {
692 log_addr_to_gpa(dev, vq);
693 if (vq->log_guest_addr == 0) {
694 VHOST_LOG_CONFIG(DEBUG,
695 "(%d) failed to map log_guest_addr.\n",
701 if (vq_is_packed(dev)) {
702 len = sizeof(struct vring_packed_desc) * vq->size;
703 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
704 ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
705 if (vq->desc_packed == NULL ||
706 len != sizeof(struct vring_packed_desc) *
708 VHOST_LOG_CONFIG(DEBUG,
709 "(%d) failed to map desc_packed ring.\n",
714 dev = numa_realloc(dev, vq_index);
715 vq = dev->virtqueue[vq_index];
716 addr = &vq->ring_addrs;
718 len = sizeof(struct vring_packed_desc_event);
719 vq->driver_event = (struct vring_packed_desc_event *)
720 (uintptr_t)ring_addr_to_vva(dev,
721 vq, addr->avail_user_addr, &len);
722 if (vq->driver_event == NULL ||
723 len != sizeof(struct vring_packed_desc_event)) {
724 VHOST_LOG_CONFIG(DEBUG,
725 "(%d) failed to find driver area address.\n",
730 len = sizeof(struct vring_packed_desc_event);
731 vq->device_event = (struct vring_packed_desc_event *)
732 (uintptr_t)ring_addr_to_vva(dev,
733 vq, addr->used_user_addr, &len);
734 if (vq->device_event == NULL ||
735 len != sizeof(struct vring_packed_desc_event)) {
736 VHOST_LOG_CONFIG(DEBUG,
737 "(%d) failed to find device area address.\n",
746 /* The addresses are converted from QEMU virtual to Vhost virtual. */
747 if (vq->desc && vq->avail && vq->used)
750 len = sizeof(struct vring_desc) * vq->size;
751 vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
752 vq, addr->desc_user_addr, &len);
753 if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
754 VHOST_LOG_CONFIG(DEBUG,
755 "(%d) failed to map desc ring.\n",
760 dev = numa_realloc(dev, vq_index);
761 vq = dev->virtqueue[vq_index];
762 addr = &vq->ring_addrs;
764 len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
765 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
766 len += sizeof(uint16_t);
768 vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
769 vq, addr->avail_user_addr, &len);
770 if (vq->avail == 0 || len != expected_len) {
771 VHOST_LOG_CONFIG(DEBUG,
772 "(%d) failed to map avail ring.\n",
777 len = sizeof(struct vring_used) +
778 sizeof(struct vring_used_elem) * vq->size;
779 if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
780 len += sizeof(uint16_t);
782 vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
783 vq, addr->used_user_addr, &len);
784 if (vq->used == 0 || len != expected_len) {
785 VHOST_LOG_CONFIG(DEBUG,
786 "(%d) failed to map used ring.\n",
791 if (vq->last_used_idx != vq->used->idx) {
792 VHOST_LOG_CONFIG(WARNING,
793 "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
794 "some packets maybe resent for Tx and dropped for Rx\n",
795 vq->last_used_idx, vq->used->idx);
796 vq->last_used_idx = vq->used->idx;
797 vq->last_avail_idx = vq->used->idx;
802 VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address desc: %p\n",
804 VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address avail: %p\n",
805 dev->vid, vq->avail);
806 VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address used: %p\n",
808 VHOST_LOG_CONFIG(DEBUG, "(%d) log_guest_addr: %" PRIx64 "\n",
809 dev->vid, vq->log_guest_addr);
815 * The virtio device sends us the desc, used and avail ring addresses.
816 * This function then converts these to our address space.
819 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
820 int main_fd __rte_unused)
822 struct virtio_net *dev = *pdev;
823 struct vhost_virtqueue *vq;
824 struct vhost_vring_addr *addr = &msg->payload.addr;
827 if (validate_msg_fds(msg, 0) != 0)
828 return RTE_VHOST_MSG_RESULT_ERR;
830 if (dev->mem == NULL)
831 return RTE_VHOST_MSG_RESULT_ERR;
833 /* addr->index refers to the queue index. The txq 1, rxq is 0. */
834 vq = dev->virtqueue[msg->payload.addr.index];
836 access_ok = vq->access_ok;
839 * Rings addresses should not be interpreted as long as the ring is not
840 * started and enabled
842 memcpy(&vq->ring_addrs, addr, sizeof(*addr));
844 vring_invalidate(dev, vq);
846 if ((vq->enabled && (dev->features &
847 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
849 dev = translate_ring_addresses(dev, msg->payload.addr.index);
851 return RTE_VHOST_MSG_RESULT_ERR;
856 return RTE_VHOST_MSG_RESULT_OK;
860 * The virtio device sends us the available ring last used index.
863 vhost_user_set_vring_base(struct virtio_net **pdev,
864 struct VhostUserMsg *msg,
865 int main_fd __rte_unused)
867 struct virtio_net *dev = *pdev;
868 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
869 uint64_t val = msg->payload.state.num;
871 if (validate_msg_fds(msg, 0) != 0)
872 return RTE_VHOST_MSG_RESULT_ERR;
874 if (vq_is_packed(dev)) {
876 * Bit[0:14]: avail index
877 * Bit[15]: avail wrap counter
879 vq->last_avail_idx = val & 0x7fff;
880 vq->avail_wrap_counter = !!(val & (0x1 << 15));
882 * Set used index to same value as available one, as
883 * their values should be the same since ring processing
884 * was stopped at get time.
886 vq->last_used_idx = vq->last_avail_idx;
887 vq->used_wrap_counter = vq->avail_wrap_counter;
889 vq->last_used_idx = msg->payload.state.num;
890 vq->last_avail_idx = msg->payload.state.num;
893 return RTE_VHOST_MSG_RESULT_OK;
897 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
898 uint64_t host_phys_addr, uint64_t size)
900 struct guest_page *page, *last_page;
901 struct guest_page *old_pages;
903 if (dev->nr_guest_pages == dev->max_guest_pages) {
904 dev->max_guest_pages *= 2;
905 old_pages = dev->guest_pages;
906 dev->guest_pages = rte_realloc(dev->guest_pages,
907 dev->max_guest_pages * sizeof(*page),
908 RTE_CACHE_LINE_SIZE);
909 if (dev->guest_pages == NULL) {
910 VHOST_LOG_CONFIG(ERR, "cannot realloc guest_pages\n");
916 if (dev->nr_guest_pages > 0) {
917 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
918 /* merge if the two pages are continuous */
919 if (host_phys_addr == last_page->host_phys_addr +
921 last_page->size += size;
926 page = &dev->guest_pages[dev->nr_guest_pages++];
927 page->guest_phys_addr = guest_phys_addr;
928 page->host_phys_addr = host_phys_addr;
935 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
938 uint64_t reg_size = reg->size;
939 uint64_t host_user_addr = reg->host_user_addr;
940 uint64_t guest_phys_addr = reg->guest_phys_addr;
941 uint64_t host_phys_addr;
944 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
945 size = page_size - (guest_phys_addr & (page_size - 1));
946 size = RTE_MIN(size, reg_size);
948 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
951 host_user_addr += size;
952 guest_phys_addr += size;
955 while (reg_size > 0) {
956 size = RTE_MIN(reg_size, page_size);
957 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
959 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
963 host_user_addr += size;
964 guest_phys_addr += size;
971 #ifdef RTE_LIBRTE_VHOST_DEBUG
972 /* TODO: enable it only in debug mode? */
974 dump_guest_pages(struct virtio_net *dev)
977 struct guest_page *page;
979 for (i = 0; i < dev->nr_guest_pages; i++) {
980 page = &dev->guest_pages[i];
982 VHOST_LOG_CONFIG(INFO,
983 "guest physical page region %u\n"
984 "\t guest_phys_addr: %" PRIx64 "\n"
985 "\t host_phys_addr : %" PRIx64 "\n"
986 "\t size : %" PRIx64 "\n",
988 page->guest_phys_addr,
989 page->host_phys_addr,
994 #define dump_guest_pages(dev)
998 vhost_memory_changed(struct VhostUserMemory *new,
999 struct rte_vhost_memory *old)
1003 if (new->nregions != old->nregions)
1006 for (i = 0; i < new->nregions; ++i) {
1007 VhostUserMemoryRegion *new_r = &new->regions[i];
1008 struct rte_vhost_mem_region *old_r = &old->regions[i];
1010 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
1012 if (new_r->memory_size != old_r->size)
1014 if (new_r->userspace_addr != old_r->guest_user_addr)
1022 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
1025 struct virtio_net *dev = *pdev;
1026 struct VhostUserMemory *memory = &msg->payload.memory;
1027 struct rte_vhost_mem_region *reg;
1030 uint64_t mmap_offset;
1036 if (validate_msg_fds(msg, memory->nregions) != 0)
1037 return RTE_VHOST_MSG_RESULT_ERR;
1039 if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
1040 VHOST_LOG_CONFIG(ERR,
1041 "too many memory regions (%u)\n", memory->nregions);
1042 return RTE_VHOST_MSG_RESULT_ERR;
1045 if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
1046 VHOST_LOG_CONFIG(INFO,
1047 "(%d) memory regions not changed\n", dev->vid);
1051 return RTE_VHOST_MSG_RESULT_OK;
1055 free_mem_region(dev);
1060 /* Flush IOTLB cache as previous HVAs are now invalid */
1061 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1062 for (i = 0; i < dev->nr_vring; i++)
1063 vhost_user_iotlb_flush_all(dev->virtqueue[i]);
1065 dev->nr_guest_pages = 0;
1066 if (dev->guest_pages == NULL) {
1067 dev->max_guest_pages = 8;
1068 dev->guest_pages = rte_zmalloc(NULL,
1069 dev->max_guest_pages *
1070 sizeof(struct guest_page),
1071 RTE_CACHE_LINE_SIZE);
1072 if (dev->guest_pages == NULL) {
1073 VHOST_LOG_CONFIG(ERR,
1074 "(%d) failed to allocate memory "
1075 "for dev->guest_pages\n",
1077 return RTE_VHOST_MSG_RESULT_ERR;
1081 dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
1082 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
1083 if (dev->mem == NULL) {
1084 VHOST_LOG_CONFIG(ERR,
1085 "(%d) failed to allocate memory for dev->mem\n",
1087 return RTE_VHOST_MSG_RESULT_ERR;
1089 dev->mem->nregions = memory->nregions;
1091 for (i = 0; i < memory->nregions; i++) {
1093 reg = &dev->mem->regions[i];
1095 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1096 reg->guest_user_addr = memory->regions[i].userspace_addr;
1097 reg->size = memory->regions[i].memory_size;
1100 mmap_offset = memory->regions[i].mmap_offset;
1102 /* Check for memory_size + mmap_offset overflow */
1103 if (mmap_offset >= -reg->size) {
1104 VHOST_LOG_CONFIG(ERR,
1105 "mmap_offset (%#"PRIx64") and memory_size "
1106 "(%#"PRIx64") overflow\n",
1107 mmap_offset, reg->size);
1111 mmap_size = reg->size + mmap_offset;
1113 /* mmap() without flag of MAP_ANONYMOUS, should be called
1114 * with length argument aligned with hugepagesz at older
1115 * longterm version Linux, like 2.6.32 and 3.2.72, or
1116 * mmap() will fail with EINVAL.
1118 * to avoid failure, make sure in caller to keep length
1121 alignment = get_blk_size(fd);
1122 if (alignment == (uint64_t)-1) {
1123 VHOST_LOG_CONFIG(ERR,
1124 "couldn't get hugepage size through fstat\n");
1127 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1128 if (mmap_size == 0) {
1130 * It could happen if initial mmap_size + alignment
1131 * overflows the sizeof uint64, which could happen if
1132 * either mmap_size or alignment value is wrong.
1134 * mmap() kernel implementation would return an error,
1135 * but better catch it before and provide useful info
1138 VHOST_LOG_CONFIG(ERR, "mmap size (0x%" PRIx64 ") "
1139 "or alignment (0x%" PRIx64 ") is invalid\n",
1140 reg->size + mmap_offset, alignment);
1144 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1145 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1146 MAP_SHARED | populate, fd, 0);
1148 if (mmap_addr == MAP_FAILED) {
1149 VHOST_LOG_CONFIG(ERR,
1150 "mmap region %u failed.\n", i);
1154 reg->mmap_addr = mmap_addr;
1155 reg->mmap_size = mmap_size;
1156 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1159 if (dev->dequeue_zero_copy)
1160 if (add_guest_pages(dev, reg, alignment) < 0) {
1161 VHOST_LOG_CONFIG(ERR,
1162 "adding guest pages to region %u failed.\n",
1167 VHOST_LOG_CONFIG(INFO,
1168 "guest memory region %u, size: 0x%" PRIx64 "\n"
1169 "\t guest physical addr: 0x%" PRIx64 "\n"
1170 "\t guest virtual addr: 0x%" PRIx64 "\n"
1171 "\t host virtual addr: 0x%" PRIx64 "\n"
1172 "\t mmap addr : 0x%" PRIx64 "\n"
1173 "\t mmap size : 0x%" PRIx64 "\n"
1174 "\t mmap align: 0x%" PRIx64 "\n"
1175 "\t mmap off : 0x%" PRIx64 "\n",
1177 reg->guest_phys_addr,
1178 reg->guest_user_addr,
1179 reg->host_user_addr,
1180 (uint64_t)(uintptr_t)mmap_addr,
1185 if (dev->postcopy_listening) {
1187 * We haven't a better way right now than sharing
1188 * DPDK's virtual address with Qemu, so that Qemu can
1189 * retrieve the region offset when handling userfaults.
1191 memory->regions[i].userspace_addr =
1192 reg->host_user_addr;
1195 if (dev->postcopy_listening) {
1196 /* Send the addresses back to qemu */
1198 send_vhost_reply(main_fd, msg);
1200 /* Wait for qemu to acknolwedge it's got the addresses
1201 * we've got to wait before we're allowed to generate faults.
1203 VhostUserMsg ack_msg;
1204 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1205 VHOST_LOG_CONFIG(ERR,
1206 "Failed to read qemu ack on postcopy set-mem-table\n");
1210 if (validate_msg_fds(&ack_msg, 0) != 0)
1213 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1214 VHOST_LOG_CONFIG(ERR,
1215 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1216 ack_msg.request.master);
1220 /* Now userfault register and we can use the memory */
1221 for (i = 0; i < memory->nregions; i++) {
1222 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1223 reg = &dev->mem->regions[i];
1224 struct uffdio_register reg_struct;
1227 * Let's register all the mmap'ed area to ensure
1228 * alignment on page boundary.
1230 reg_struct.range.start =
1231 (uint64_t)(uintptr_t)reg->mmap_addr;
1232 reg_struct.range.len = reg->mmap_size;
1233 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1235 if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1237 VHOST_LOG_CONFIG(ERR,
1238 "Failed to register ufd for region %d: (ufd = %d) %s\n",
1239 i, dev->postcopy_ufd,
1243 VHOST_LOG_CONFIG(INFO,
1244 "\t userfaultfd registered for range : "
1245 "%" PRIx64 " - %" PRIx64 "\n",
1246 (uint64_t)reg_struct.range.start,
1247 (uint64_t)reg_struct.range.start +
1248 (uint64_t)reg_struct.range.len - 1);
1255 for (i = 0; i < dev->nr_vring; i++) {
1256 struct vhost_virtqueue *vq = dev->virtqueue[i];
1258 if (vq->desc || vq->avail || vq->used) {
1260 * If the memory table got updated, the ring addresses
1261 * need to be translated again as virtual addresses have
1264 vring_invalidate(dev, vq);
1266 dev = translate_ring_addresses(dev, i);
1276 dump_guest_pages(dev);
1278 return RTE_VHOST_MSG_RESULT_OK;
1281 free_mem_region(dev);
1284 return RTE_VHOST_MSG_RESULT_ERR;
1288 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1295 if (vq_is_packed(dev))
1296 rings_ok = vq->desc_packed && vq->driver_event &&
1299 rings_ok = vq->desc && vq->avail && vq->used;
1302 vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1303 vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1307 virtio_is_ready(struct virtio_net *dev)
1309 struct vhost_virtqueue *vq;
1312 if (dev->nr_vring == 0)
1315 for (i = 0; i < dev->nr_vring; i++) {
1316 vq = dev->virtqueue[i];
1318 if (!vq_is_ready(dev, vq))
1322 VHOST_LOG_CONFIG(INFO,
1323 "virtio is now ready for processing.\n");
1328 inflight_mem_alloc(const char *name, size_t size, int *fd)
1332 char fname[20] = "/tmp/memfd-XXXXXX";
1335 #ifdef MEMFD_SUPPORTED
1336 mfd = memfd_create(name, MFD_CLOEXEC);
1341 mfd = mkstemp(fname);
1343 VHOST_LOG_CONFIG(ERR,
1344 "failed to get inflight buffer fd\n");
1351 if (ftruncate(mfd, size) == -1) {
1352 VHOST_LOG_CONFIG(ERR,
1353 "failed to alloc inflight buffer\n");
1358 ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
1359 if (ptr == MAP_FAILED) {
1360 VHOST_LOG_CONFIG(ERR,
1361 "failed to mmap inflight buffer\n");
1371 get_pervq_shm_size_split(uint16_t queue_size)
1373 return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_split) *
1374 queue_size + sizeof(uint64_t) +
1375 sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
1379 get_pervq_shm_size_packed(uint16_t queue_size)
1381 return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_packed)
1382 * queue_size + sizeof(uint64_t) +
1383 sizeof(uint16_t) * 6 + sizeof(uint8_t) * 9,
1384 INFLIGHT_ALIGNMENT);
1388 vhost_user_get_inflight_fd(struct virtio_net **pdev,
1390 int main_fd __rte_unused)
1392 struct rte_vhost_inflight_info_packed *inflight_packed;
1393 uint64_t pervq_inflight_size, mmap_size;
1394 uint16_t num_queues, queue_size;
1395 struct virtio_net *dev = *pdev;
1399 if (msg->size != sizeof(msg->payload.inflight)) {
1400 VHOST_LOG_CONFIG(ERR,
1401 "invalid get_inflight_fd message size is %d\n",
1403 return RTE_VHOST_MSG_RESULT_ERR;
1406 if (dev->inflight_info == NULL) {
1407 dev->inflight_info = calloc(1,
1408 sizeof(struct inflight_mem_info));
1409 if (!dev->inflight_info) {
1410 VHOST_LOG_CONFIG(ERR,
1411 "failed to alloc dev inflight area\n");
1412 return RTE_VHOST_MSG_RESULT_ERR;
1416 num_queues = msg->payload.inflight.num_queues;
1417 queue_size = msg->payload.inflight.queue_size;
1419 VHOST_LOG_CONFIG(INFO, "get_inflight_fd num_queues: %u\n",
1420 msg->payload.inflight.num_queues);
1421 VHOST_LOG_CONFIG(INFO, "get_inflight_fd queue_size: %u\n",
1422 msg->payload.inflight.queue_size);
1424 if (vq_is_packed(dev))
1425 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1427 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1429 mmap_size = num_queues * pervq_inflight_size;
1430 addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
1432 VHOST_LOG_CONFIG(ERR,
1433 "failed to alloc vhost inflight area\n");
1434 msg->payload.inflight.mmap_size = 0;
1435 return RTE_VHOST_MSG_RESULT_ERR;
1437 memset(addr, 0, mmap_size);
1439 dev->inflight_info->addr = addr;
1440 dev->inflight_info->size = msg->payload.inflight.mmap_size = mmap_size;
1441 dev->inflight_info->fd = msg->fds[0] = fd;
1442 msg->payload.inflight.mmap_offset = 0;
1445 if (vq_is_packed(dev)) {
1446 for (i = 0; i < num_queues; i++) {
1448 (struct rte_vhost_inflight_info_packed *)addr;
1449 inflight_packed->used_wrap_counter = 1;
1450 inflight_packed->old_used_wrap_counter = 1;
1451 for (j = 0; j < queue_size; j++)
1452 inflight_packed->desc[j].next = j + 1;
1453 addr = (void *)((char *)addr + pervq_inflight_size);
1457 VHOST_LOG_CONFIG(INFO,
1458 "send inflight mmap_size: %"PRIu64"\n",
1459 msg->payload.inflight.mmap_size);
1460 VHOST_LOG_CONFIG(INFO,
1461 "send inflight mmap_offset: %"PRIu64"\n",
1462 msg->payload.inflight.mmap_offset);
1463 VHOST_LOG_CONFIG(INFO,
1464 "send inflight fd: %d\n", msg->fds[0]);
1466 return RTE_VHOST_MSG_RESULT_REPLY;
1470 vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
1471 int main_fd __rte_unused)
1473 uint64_t mmap_size, mmap_offset;
1474 uint16_t num_queues, queue_size;
1475 struct virtio_net *dev = *pdev;
1476 uint32_t pervq_inflight_size;
1477 struct vhost_virtqueue *vq;
1482 if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
1483 VHOST_LOG_CONFIG(ERR,
1484 "invalid set_inflight_fd message size is %d,fd is %d\n",
1486 return RTE_VHOST_MSG_RESULT_ERR;
1489 mmap_size = msg->payload.inflight.mmap_size;
1490 mmap_offset = msg->payload.inflight.mmap_offset;
1491 num_queues = msg->payload.inflight.num_queues;
1492 queue_size = msg->payload.inflight.queue_size;
1494 if (vq_is_packed(dev))
1495 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1497 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1499 VHOST_LOG_CONFIG(INFO,
1500 "set_inflight_fd mmap_size: %"PRIu64"\n", mmap_size);
1501 VHOST_LOG_CONFIG(INFO,
1502 "set_inflight_fd mmap_offset: %"PRIu64"\n", mmap_offset);
1503 VHOST_LOG_CONFIG(INFO,
1504 "set_inflight_fd num_queues: %u\n", num_queues);
1505 VHOST_LOG_CONFIG(INFO,
1506 "set_inflight_fd queue_size: %u\n", queue_size);
1507 VHOST_LOG_CONFIG(INFO,
1508 "set_inflight_fd fd: %d\n", fd);
1509 VHOST_LOG_CONFIG(INFO,
1510 "set_inflight_fd pervq_inflight_size: %d\n",
1511 pervq_inflight_size);
1513 if (!dev->inflight_info) {
1514 dev->inflight_info = calloc(1,
1515 sizeof(struct inflight_mem_info));
1516 if (dev->inflight_info == NULL) {
1517 VHOST_LOG_CONFIG(ERR,
1518 "failed to alloc dev inflight area\n");
1519 return RTE_VHOST_MSG_RESULT_ERR;
1523 if (dev->inflight_info->addr)
1524 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1526 addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1528 if (addr == MAP_FAILED) {
1529 VHOST_LOG_CONFIG(ERR, "failed to mmap share memory.\n");
1530 return RTE_VHOST_MSG_RESULT_ERR;
1533 if (dev->inflight_info->fd)
1534 close(dev->inflight_info->fd);
1536 dev->inflight_info->fd = fd;
1537 dev->inflight_info->addr = addr;
1538 dev->inflight_info->size = mmap_size;
1540 for (i = 0; i < num_queues; i++) {
1541 vq = dev->virtqueue[i];
1542 if (vq_is_packed(dev)) {
1543 vq->inflight_packed = addr;
1544 vq->inflight_packed->desc_num = queue_size;
1546 vq->inflight_split = addr;
1547 vq->inflight_split->desc_num = queue_size;
1549 addr = (void *)((char *)addr + pervq_inflight_size);
1552 return RTE_VHOST_MSG_RESULT_OK;
1556 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1557 int main_fd __rte_unused)
1559 struct virtio_net *dev = *pdev;
1560 struct vhost_vring_file file;
1561 struct vhost_virtqueue *vq;
1564 expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1565 if (validate_msg_fds(msg, expected_fds) != 0)
1566 return RTE_VHOST_MSG_RESULT_ERR;
1568 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1569 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1570 file.fd = VIRTIO_INVALID_EVENTFD;
1572 file.fd = msg->fds[0];
1573 VHOST_LOG_CONFIG(INFO,
1574 "vring call idx:%d file:%d\n", file.index, file.fd);
1576 vq = dev->virtqueue[file.index];
1577 if (vq->callfd >= 0)
1580 vq->callfd = file.fd;
1582 return RTE_VHOST_MSG_RESULT_OK;
1585 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1586 struct VhostUserMsg *msg,
1587 int main_fd __rte_unused)
1591 expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1592 if (validate_msg_fds(msg, expected_fds) != 0)
1593 return RTE_VHOST_MSG_RESULT_ERR;
1595 if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1597 VHOST_LOG_CONFIG(INFO, "not implemented\n");
1599 return RTE_VHOST_MSG_RESULT_OK;
1603 resubmit_desc_compare(const void *a, const void *b)
1605 const struct rte_vhost_resubmit_desc *desc0 = a;
1606 const struct rte_vhost_resubmit_desc *desc1 = b;
1608 if (desc1->counter > desc0->counter)
1615 vhost_check_queue_inflights_split(struct virtio_net *dev,
1616 struct vhost_virtqueue *vq)
1619 uint16_t resubmit_num = 0, last_io, num;
1620 struct vring_used *used = vq->used;
1621 struct rte_vhost_resubmit_info *resubmit;
1622 struct rte_vhost_inflight_info_split *inflight_split;
1624 if (!(dev->protocol_features &
1625 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1626 return RTE_VHOST_MSG_RESULT_OK;
1628 /* The frontend may still not support the inflight feature
1629 * although we negotiate the protocol feature.
1631 if ((!vq->inflight_split))
1632 return RTE_VHOST_MSG_RESULT_OK;
1634 if (!vq->inflight_split->version) {
1635 vq->inflight_split->version = INFLIGHT_VERSION;
1636 return RTE_VHOST_MSG_RESULT_OK;
1639 if (vq->resubmit_inflight)
1640 return RTE_VHOST_MSG_RESULT_OK;
1642 inflight_split = vq->inflight_split;
1643 vq->global_counter = 0;
1644 last_io = inflight_split->last_inflight_io;
1646 if (inflight_split->used_idx != used->idx) {
1647 inflight_split->desc[last_io].inflight = 0;
1649 inflight_split->used_idx = used->idx;
1652 for (i = 0; i < inflight_split->desc_num; i++) {
1653 if (inflight_split->desc[i].inflight == 1)
1657 vq->last_avail_idx += resubmit_num;
1660 resubmit = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1662 VHOST_LOG_CONFIG(ERR,
1663 "failed to allocate memory for resubmit info.\n");
1664 return RTE_VHOST_MSG_RESULT_ERR;
1667 resubmit->resubmit_list = calloc(resubmit_num,
1668 sizeof(struct rte_vhost_resubmit_desc));
1669 if (!resubmit->resubmit_list) {
1670 VHOST_LOG_CONFIG(ERR,
1671 "failed to allocate memory for inflight desc.\n");
1673 return RTE_VHOST_MSG_RESULT_ERR;
1677 for (i = 0; i < vq->inflight_split->desc_num; i++) {
1678 if (vq->inflight_split->desc[i].inflight == 1) {
1679 resubmit->resubmit_list[num].index = i;
1680 resubmit->resubmit_list[num].counter =
1681 inflight_split->desc[i].counter;
1685 resubmit->resubmit_num = num;
1687 if (resubmit->resubmit_num > 1)
1688 qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1689 sizeof(struct rte_vhost_resubmit_desc),
1690 resubmit_desc_compare);
1692 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1693 vq->resubmit_inflight = resubmit;
1696 return RTE_VHOST_MSG_RESULT_OK;
1700 vhost_check_queue_inflights_packed(struct virtio_net *dev,
1701 struct vhost_virtqueue *vq)
1704 uint16_t resubmit_num = 0, old_used_idx, num;
1705 struct rte_vhost_resubmit_info *resubmit;
1706 struct rte_vhost_inflight_info_packed *inflight_packed;
1708 if (!(dev->protocol_features &
1709 (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1710 return RTE_VHOST_MSG_RESULT_OK;
1712 /* The frontend may still not support the inflight feature
1713 * although we negotiate the protocol feature.
1715 if ((!vq->inflight_packed))
1716 return RTE_VHOST_MSG_RESULT_OK;
1718 if (!vq->inflight_packed->version) {
1719 vq->inflight_packed->version = INFLIGHT_VERSION;
1720 return RTE_VHOST_MSG_RESULT_OK;
1723 if (vq->resubmit_inflight)
1724 return RTE_VHOST_MSG_RESULT_OK;
1726 inflight_packed = vq->inflight_packed;
1727 vq->global_counter = 0;
1728 old_used_idx = inflight_packed->old_used_idx;
1730 if (inflight_packed->used_idx != old_used_idx) {
1731 if (inflight_packed->desc[old_used_idx].inflight == 0) {
1732 inflight_packed->old_used_idx =
1733 inflight_packed->used_idx;
1734 inflight_packed->old_used_wrap_counter =
1735 inflight_packed->used_wrap_counter;
1736 inflight_packed->old_free_head =
1737 inflight_packed->free_head;
1739 inflight_packed->used_idx =
1740 inflight_packed->old_used_idx;
1741 inflight_packed->used_wrap_counter =
1742 inflight_packed->old_used_wrap_counter;
1743 inflight_packed->free_head =
1744 inflight_packed->old_free_head;
1748 for (i = 0; i < inflight_packed->desc_num; i++) {
1749 if (inflight_packed->desc[i].inflight == 1)
1754 resubmit = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1755 if (resubmit == NULL) {
1756 VHOST_LOG_CONFIG(ERR,
1757 "failed to allocate memory for resubmit info.\n");
1758 return RTE_VHOST_MSG_RESULT_ERR;
1761 resubmit->resubmit_list = calloc(resubmit_num,
1762 sizeof(struct rte_vhost_resubmit_desc));
1763 if (resubmit->resubmit_list == NULL) {
1764 VHOST_LOG_CONFIG(ERR,
1765 "failed to allocate memory for resubmit desc.\n");
1767 return RTE_VHOST_MSG_RESULT_ERR;
1771 for (i = 0; i < inflight_packed->desc_num; i++) {
1772 if (vq->inflight_packed->desc[i].inflight == 1) {
1773 resubmit->resubmit_list[num].index = i;
1774 resubmit->resubmit_list[num].counter =
1775 inflight_packed->desc[i].counter;
1779 resubmit->resubmit_num = num;
1781 if (resubmit->resubmit_num > 1)
1782 qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1783 sizeof(struct rte_vhost_resubmit_desc),
1784 resubmit_desc_compare);
1786 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1787 vq->resubmit_inflight = resubmit;
1790 return RTE_VHOST_MSG_RESULT_OK;
1794 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1795 int main_fd __rte_unused)
1797 struct virtio_net *dev = *pdev;
1798 struct vhost_vring_file file;
1799 struct vhost_virtqueue *vq;
1802 expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1803 if (validate_msg_fds(msg, expected_fds) != 0)
1804 return RTE_VHOST_MSG_RESULT_ERR;
1806 file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1807 if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1808 file.fd = VIRTIO_INVALID_EVENTFD;
1810 file.fd = msg->fds[0];
1811 VHOST_LOG_CONFIG(INFO,
1812 "vring kick idx:%d file:%d\n", file.index, file.fd);
1814 /* Interpret ring addresses only when ring is started. */
1815 dev = translate_ring_addresses(dev, file.index);
1817 return RTE_VHOST_MSG_RESULT_ERR;
1821 vq = dev->virtqueue[file.index];
1824 * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1825 * the ring starts already enabled. Otherwise, it is enabled via
1826 * the SET_VRING_ENABLE message.
1828 if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1830 if (dev->notify_ops->vring_state_changed)
1831 dev->notify_ops->vring_state_changed(
1832 dev->vid, file.index, 1);
1835 if (vq->kickfd >= 0)
1837 vq->kickfd = file.fd;
1839 if (vq_is_packed(dev)) {
1840 if (vhost_check_queue_inflights_packed(dev, vq)) {
1841 VHOST_LOG_CONFIG(ERR,
1842 "failed to inflights for vq: %d\n", file.index);
1843 return RTE_VHOST_MSG_RESULT_ERR;
1846 if (vhost_check_queue_inflights_split(dev, vq)) {
1847 VHOST_LOG_CONFIG(ERR,
1848 "failed to inflights for vq: %d\n", file.index);
1849 return RTE_VHOST_MSG_RESULT_ERR;
1853 return RTE_VHOST_MSG_RESULT_OK;
1857 free_zmbufs(struct vhost_virtqueue *vq)
1859 drain_zmbuf_list(vq);
1861 rte_free(vq->zmbufs);
1865 * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1868 vhost_user_get_vring_base(struct virtio_net **pdev,
1869 struct VhostUserMsg *msg,
1870 int main_fd __rte_unused)
1872 struct virtio_net *dev = *pdev;
1873 struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1876 if (validate_msg_fds(msg, 0) != 0)
1877 return RTE_VHOST_MSG_RESULT_ERR;
1879 /* We have to stop the queue (virtio) if it is running. */
1880 vhost_destroy_device_notify(dev);
1882 dev->flags &= ~VIRTIO_DEV_READY;
1883 dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1885 /* Here we are safe to get the indexes */
1886 if (vq_is_packed(dev)) {
1888 * Bit[0:14]: avail index
1889 * Bit[15]: avail wrap counter
1891 val = vq->last_avail_idx & 0x7fff;
1892 val |= vq->avail_wrap_counter << 15;
1893 msg->payload.state.num = val;
1895 msg->payload.state.num = vq->last_avail_idx;
1898 VHOST_LOG_CONFIG(INFO,
1899 "vring base idx:%d file:%d\n", msg->payload.state.index,
1900 msg->payload.state.num);
1902 * Based on current qemu vhost-user implementation, this message is
1903 * sent and only sent in vhost_vring_stop.
1904 * TODO: cleanup the vring, it isn't usable since here.
1906 if (vq->kickfd >= 0)
1909 vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1911 if (vq->callfd >= 0)
1914 vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1916 vq->signalled_used_valid = false;
1918 if (dev->dequeue_zero_copy)
1920 if (vq_is_packed(dev)) {
1921 rte_free(vq->shadow_used_packed);
1922 vq->shadow_used_packed = NULL;
1924 rte_free(vq->shadow_used_split);
1925 vq->shadow_used_split = NULL;
1928 rte_free(vq->batch_copy_elems);
1929 vq->batch_copy_elems = NULL;
1931 msg->size = sizeof(msg->payload.state);
1934 vring_invalidate(dev, vq);
1936 return RTE_VHOST_MSG_RESULT_REPLY;
1940 * when virtio queues are ready to work, qemu will send us to
1941 * enable the virtio queue pair.
1944 vhost_user_set_vring_enable(struct virtio_net **pdev,
1945 struct VhostUserMsg *msg,
1946 int main_fd __rte_unused)
1948 struct virtio_net *dev = *pdev;
1949 int enable = (int)msg->payload.state.num;
1950 int index = (int)msg->payload.state.index;
1951 struct rte_vdpa_device *vdpa_dev;
1954 if (validate_msg_fds(msg, 0) != 0)
1955 return RTE_VHOST_MSG_RESULT_ERR;
1957 VHOST_LOG_CONFIG(INFO,
1958 "set queue enable: %d to qp idx: %d\n",
1961 did = dev->vdpa_dev_id;
1962 vdpa_dev = rte_vdpa_get_device(did);
1963 if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1964 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1966 if (dev->notify_ops->vring_state_changed)
1967 dev->notify_ops->vring_state_changed(dev->vid,
1970 /* On disable, rings have to be stopped being processed. */
1971 if (!enable && dev->dequeue_zero_copy)
1972 drain_zmbuf_list(dev->virtqueue[index]);
1974 dev->virtqueue[index]->enabled = enable;
1976 return RTE_VHOST_MSG_RESULT_OK;
1980 vhost_user_get_protocol_features(struct virtio_net **pdev,
1981 struct VhostUserMsg *msg,
1982 int main_fd __rte_unused)
1984 struct virtio_net *dev = *pdev;
1985 uint64_t features, protocol_features;
1987 if (validate_msg_fds(msg, 0) != 0)
1988 return RTE_VHOST_MSG_RESULT_ERR;
1990 rte_vhost_driver_get_features(dev->ifname, &features);
1991 rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1994 * REPLY_ACK protocol feature is only mandatory for now
1995 * for IOMMU feature. If IOMMU is explicitly disabled by the
1996 * application, disable also REPLY_ACK feature for older buggy
1997 * Qemu versions (from v2.7.0 to v2.9.0).
1999 if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
2000 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
2002 msg->payload.u64 = protocol_features;
2003 msg->size = sizeof(msg->payload.u64);
2006 return RTE_VHOST_MSG_RESULT_REPLY;
2010 vhost_user_set_protocol_features(struct virtio_net **pdev,
2011 struct VhostUserMsg *msg,
2012 int main_fd __rte_unused)
2014 struct virtio_net *dev = *pdev;
2015 uint64_t protocol_features = msg->payload.u64;
2016 uint64_t slave_protocol_features = 0;
2018 if (validate_msg_fds(msg, 0) != 0)
2019 return RTE_VHOST_MSG_RESULT_ERR;
2021 rte_vhost_driver_get_protocol_features(dev->ifname,
2022 &slave_protocol_features);
2023 if (protocol_features & ~slave_protocol_features) {
2024 VHOST_LOG_CONFIG(ERR,
2025 "(%d) received invalid protocol features.\n",
2027 return RTE_VHOST_MSG_RESULT_ERR;
2030 dev->protocol_features = protocol_features;
2031 VHOST_LOG_CONFIG(INFO,
2032 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
2033 dev->protocol_features);
2035 return RTE_VHOST_MSG_RESULT_OK;
2039 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
2040 int main_fd __rte_unused)
2042 struct virtio_net *dev = *pdev;
2043 int fd = msg->fds[0];
2047 if (validate_msg_fds(msg, 1) != 0)
2048 return RTE_VHOST_MSG_RESULT_ERR;
2051 VHOST_LOG_CONFIG(ERR, "invalid log fd: %d\n", fd);
2052 return RTE_VHOST_MSG_RESULT_ERR;
2055 if (msg->size != sizeof(VhostUserLog)) {
2056 VHOST_LOG_CONFIG(ERR,
2057 "invalid log base msg size: %"PRId32" != %d\n",
2058 msg->size, (int)sizeof(VhostUserLog));
2059 return RTE_VHOST_MSG_RESULT_ERR;
2062 size = msg->payload.log.mmap_size;
2063 off = msg->payload.log.mmap_offset;
2065 /* Don't allow mmap_offset to point outside the mmap region */
2067 VHOST_LOG_CONFIG(ERR,
2068 "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
2070 return RTE_VHOST_MSG_RESULT_ERR;
2073 VHOST_LOG_CONFIG(INFO,
2074 "log mmap size: %"PRId64", offset: %"PRId64"\n",
2078 * mmap from 0 to workaround a hugepage mmap bug: mmap will
2079 * fail when offset is not page size aligned.
2081 addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
2083 if (addr == MAP_FAILED) {
2084 VHOST_LOG_CONFIG(ERR, "mmap log base failed!\n");
2085 return RTE_VHOST_MSG_RESULT_ERR;
2089 * Free previously mapped log memory on occasionally
2090 * multiple VHOST_USER_SET_LOG_BASE.
2092 if (dev->log_addr) {
2093 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
2095 dev->log_addr = (uint64_t)(uintptr_t)addr;
2096 dev->log_base = dev->log_addr + off;
2097 dev->log_size = size;
2100 * The spec is not clear about it (yet), but QEMU doesn't expect
2101 * any payload in the reply.
2106 return RTE_VHOST_MSG_RESULT_REPLY;
2109 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
2110 struct VhostUserMsg *msg,
2111 int main_fd __rte_unused)
2113 if (validate_msg_fds(msg, 1) != 0)
2114 return RTE_VHOST_MSG_RESULT_ERR;
2117 VHOST_LOG_CONFIG(INFO, "not implemented.\n");
2119 return RTE_VHOST_MSG_RESULT_OK;
2123 * An rarp packet is constructed and broadcasted to notify switches about
2124 * the new location of the migrated VM, so that packets from outside will
2125 * not be lost after migration.
2127 * However, we don't actually "send" a rarp packet here, instead, we set
2128 * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
2131 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
2132 int main_fd __rte_unused)
2134 struct virtio_net *dev = *pdev;
2135 uint8_t *mac = (uint8_t *)&msg->payload.u64;
2136 struct rte_vdpa_device *vdpa_dev;
2139 if (validate_msg_fds(msg, 0) != 0)
2140 return RTE_VHOST_MSG_RESULT_ERR;
2142 VHOST_LOG_CONFIG(DEBUG,
2143 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
2144 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
2145 memcpy(dev->mac.addr_bytes, mac, 6);
2148 * Set the flag to inject a RARP broadcast packet at
2149 * rte_vhost_dequeue_burst().
2151 * __ATOMIC_RELEASE ordering is for making sure the mac is
2152 * copied before the flag is set.
2154 __atomic_store_n(&dev->broadcast_rarp, 1, __ATOMIC_RELEASE);
2155 did = dev->vdpa_dev_id;
2156 vdpa_dev = rte_vdpa_get_device(did);
2157 if (vdpa_dev && vdpa_dev->ops->migration_done)
2158 vdpa_dev->ops->migration_done(dev->vid);
2160 return RTE_VHOST_MSG_RESULT_OK;
2164 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
2165 int main_fd __rte_unused)
2167 struct virtio_net *dev = *pdev;
2169 if (validate_msg_fds(msg, 0) != 0)
2170 return RTE_VHOST_MSG_RESULT_ERR;
2172 if (msg->payload.u64 < VIRTIO_MIN_MTU ||
2173 msg->payload.u64 > VIRTIO_MAX_MTU) {
2174 VHOST_LOG_CONFIG(ERR, "Invalid MTU size (%"PRIu64")\n",
2177 return RTE_VHOST_MSG_RESULT_ERR;
2180 dev->mtu = msg->payload.u64;
2182 return RTE_VHOST_MSG_RESULT_OK;
2186 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
2187 int main_fd __rte_unused)
2189 struct virtio_net *dev = *pdev;
2190 int fd = msg->fds[0];
2192 if (validate_msg_fds(msg, 1) != 0)
2193 return RTE_VHOST_MSG_RESULT_ERR;
2196 VHOST_LOG_CONFIG(ERR,
2197 "Invalid file descriptor for slave channel (%d)\n",
2199 return RTE_VHOST_MSG_RESULT_ERR;
2202 if (dev->slave_req_fd >= 0)
2203 close(dev->slave_req_fd);
2205 dev->slave_req_fd = fd;
2207 return RTE_VHOST_MSG_RESULT_OK;
2211 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2213 struct vhost_vring_addr *ra;
2214 uint64_t start, end, len;
2217 end = start + imsg->size;
2219 ra = &vq->ring_addrs;
2220 len = sizeof(struct vring_desc) * vq->size;
2221 if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2224 len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
2225 if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2228 len = sizeof(struct vring_used) +
2229 sizeof(struct vring_used_elem) * vq->size;
2230 if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2233 if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2234 len = sizeof(uint64_t);
2235 if (ra->log_guest_addr < end &&
2236 (ra->log_guest_addr + len) > start)
2244 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2246 struct vhost_vring_addr *ra;
2247 uint64_t start, end, len;
2250 end = start + imsg->size;
2252 ra = &vq->ring_addrs;
2253 len = sizeof(struct vring_packed_desc) * vq->size;
2254 if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2257 len = sizeof(struct vring_packed_desc_event);
2258 if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2261 len = sizeof(struct vring_packed_desc_event);
2262 if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2265 if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2266 len = sizeof(uint64_t);
2267 if (ra->log_guest_addr < end &&
2268 (ra->log_guest_addr + len) > start)
2275 static int is_vring_iotlb(struct virtio_net *dev,
2276 struct vhost_virtqueue *vq,
2277 struct vhost_iotlb_msg *imsg)
2279 if (vq_is_packed(dev))
2280 return is_vring_iotlb_packed(vq, imsg);
2282 return is_vring_iotlb_split(vq, imsg);
2286 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
2287 int main_fd __rte_unused)
2289 struct virtio_net *dev = *pdev;
2290 struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
2294 if (validate_msg_fds(msg, 0) != 0)
2295 return RTE_VHOST_MSG_RESULT_ERR;
2297 switch (imsg->type) {
2298 case VHOST_IOTLB_UPDATE:
2300 vva = qva_to_vva(dev, imsg->uaddr, &len);
2302 return RTE_VHOST_MSG_RESULT_ERR;
2304 for (i = 0; i < dev->nr_vring; i++) {
2305 struct vhost_virtqueue *vq = dev->virtqueue[i];
2307 vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
2310 if (is_vring_iotlb(dev, vq, imsg))
2311 *pdev = dev = translate_ring_addresses(dev, i);
2314 case VHOST_IOTLB_INVALIDATE:
2315 for (i = 0; i < dev->nr_vring; i++) {
2316 struct vhost_virtqueue *vq = dev->virtqueue[i];
2318 vhost_user_iotlb_cache_remove(vq, imsg->iova,
2321 if (is_vring_iotlb(dev, vq, imsg))
2322 vring_invalidate(dev, vq);
2326 VHOST_LOG_CONFIG(ERR, "Invalid IOTLB message type (%d)\n",
2328 return RTE_VHOST_MSG_RESULT_ERR;
2331 return RTE_VHOST_MSG_RESULT_OK;
2335 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
2336 struct VhostUserMsg *msg,
2337 int main_fd __rte_unused)
2339 struct virtio_net *dev = *pdev;
2340 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
2341 struct uffdio_api api_struct;
2343 if (validate_msg_fds(msg, 0) != 0)
2344 return RTE_VHOST_MSG_RESULT_ERR;
2346 dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
2348 if (dev->postcopy_ufd == -1) {
2349 VHOST_LOG_CONFIG(ERR, "Userfaultfd not available: %s\n",
2351 return RTE_VHOST_MSG_RESULT_ERR;
2353 api_struct.api = UFFD_API;
2354 api_struct.features = 0;
2355 if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
2356 VHOST_LOG_CONFIG(ERR, "UFFDIO_API ioctl failure: %s\n",
2358 close(dev->postcopy_ufd);
2359 dev->postcopy_ufd = -1;
2360 return RTE_VHOST_MSG_RESULT_ERR;
2362 msg->fds[0] = dev->postcopy_ufd;
2365 return RTE_VHOST_MSG_RESULT_REPLY;
2367 dev->postcopy_ufd = -1;
2370 return RTE_VHOST_MSG_RESULT_ERR;
2375 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
2376 struct VhostUserMsg *msg __rte_unused,
2377 int main_fd __rte_unused)
2379 struct virtio_net *dev = *pdev;
2381 if (validate_msg_fds(msg, 0) != 0)
2382 return RTE_VHOST_MSG_RESULT_ERR;
2384 if (dev->mem && dev->mem->nregions) {
2385 VHOST_LOG_CONFIG(ERR,
2386 "Regions already registered at postcopy-listen\n");
2387 return RTE_VHOST_MSG_RESULT_ERR;
2389 dev->postcopy_listening = 1;
2391 return RTE_VHOST_MSG_RESULT_OK;
2395 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
2396 int main_fd __rte_unused)
2398 struct virtio_net *dev = *pdev;
2400 if (validate_msg_fds(msg, 0) != 0)
2401 return RTE_VHOST_MSG_RESULT_ERR;
2403 dev->postcopy_listening = 0;
2404 if (dev->postcopy_ufd >= 0) {
2405 close(dev->postcopy_ufd);
2406 dev->postcopy_ufd = -1;
2409 msg->payload.u64 = 0;
2410 msg->size = sizeof(msg->payload.u64);
2413 return RTE_VHOST_MSG_RESULT_REPLY;
2416 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
2417 struct VhostUserMsg *msg,
2419 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
2420 [VHOST_USER_NONE] = NULL,
2421 [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
2422 [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
2423 [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
2424 [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
2425 [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
2426 [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
2427 [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
2428 [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
2429 [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
2430 [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
2431 [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
2432 [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
2433 [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
2434 [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
2435 [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
2436 [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
2437 [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
2438 [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
2439 [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
2440 [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
2441 [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
2442 [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
2443 [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
2444 [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
2445 [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
2446 [VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
2447 [VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
2450 /* return bytes# of read on success or negative val on failure. */
2452 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
2456 ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
2457 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
2460 } else if (ret != VHOST_USER_HDR_SIZE) {
2461 VHOST_LOG_CONFIG(ERR, "Unexpected header size read\n");
2467 if (msg->size > sizeof(msg->payload)) {
2468 VHOST_LOG_CONFIG(ERR,
2469 "invalid msg size: %d\n", msg->size);
2472 ret = read(sockfd, &msg->payload, msg->size);
2475 if (ret != (int)msg->size) {
2476 VHOST_LOG_CONFIG(ERR,
2477 "read control message failed\n");
2486 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
2491 return send_fd_message(sockfd, (char *)msg,
2492 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
2496 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
2501 msg->flags &= ~VHOST_USER_VERSION_MASK;
2502 msg->flags &= ~VHOST_USER_NEED_REPLY;
2503 msg->flags |= VHOST_USER_VERSION;
2504 msg->flags |= VHOST_USER_REPLY_MASK;
2506 return send_vhost_message(sockfd, msg);
2510 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
2514 if (msg->flags & VHOST_USER_NEED_REPLY)
2515 rte_spinlock_lock(&dev->slave_req_lock);
2517 ret = send_vhost_message(dev->slave_req_fd, msg);
2518 if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
2519 rte_spinlock_unlock(&dev->slave_req_lock);
2525 * Allocate a queue pair if it hasn't been allocated yet
2528 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
2529 struct VhostUserMsg *msg)
2533 switch (msg->request.master) {
2534 case VHOST_USER_SET_VRING_KICK:
2535 case VHOST_USER_SET_VRING_CALL:
2536 case VHOST_USER_SET_VRING_ERR:
2537 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2539 case VHOST_USER_SET_VRING_NUM:
2540 case VHOST_USER_SET_VRING_BASE:
2541 case VHOST_USER_SET_VRING_ENABLE:
2542 vring_idx = msg->payload.state.index;
2544 case VHOST_USER_SET_VRING_ADDR:
2545 vring_idx = msg->payload.addr.index;
2551 if (vring_idx >= VHOST_MAX_VRING) {
2552 VHOST_LOG_CONFIG(ERR,
2553 "invalid vring index: %u\n", vring_idx);
2557 if (dev->virtqueue[vring_idx])
2560 return alloc_vring_queue(dev, vring_idx);
2564 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
2567 unsigned int vq_num = 0;
2569 while (vq_num < dev->nr_vring) {
2570 struct vhost_virtqueue *vq = dev->virtqueue[i];
2573 rte_spinlock_lock(&vq->access_lock);
2581 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
2584 unsigned int vq_num = 0;
2586 while (vq_num < dev->nr_vring) {
2587 struct vhost_virtqueue *vq = dev->virtqueue[i];
2590 rte_spinlock_unlock(&vq->access_lock);
2598 vhost_user_msg_handler(int vid, int fd)
2600 struct virtio_net *dev;
2601 struct VhostUserMsg msg;
2602 struct rte_vdpa_device *vdpa_dev;
2605 int unlock_required = 0;
2609 dev = get_device(vid);
2613 if (!dev->notify_ops) {
2614 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2615 if (!dev->notify_ops) {
2616 VHOST_LOG_CONFIG(ERR,
2617 "failed to get callback ops for driver %s\n",
2623 ret = read_vhost_message(fd, &msg);
2626 VHOST_LOG_CONFIG(ERR,
2627 "vhost read message failed\n");
2629 VHOST_LOG_CONFIG(INFO,
2630 "vhost peer closed\n");
2636 request = msg.request.master;
2637 if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
2638 vhost_message_str[request]) {
2639 if (request != VHOST_USER_IOTLB_MSG)
2640 VHOST_LOG_CONFIG(INFO, "read message %s\n",
2641 vhost_message_str[request]);
2643 VHOST_LOG_CONFIG(DEBUG, "read message %s\n",
2644 vhost_message_str[request]);
2646 VHOST_LOG_CONFIG(DEBUG, "External request %d\n", request);
2649 ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
2651 VHOST_LOG_CONFIG(ERR,
2652 "failed to alloc queue\n");
2657 * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2658 * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2659 * and device is destroyed. destroy_device waits for queues to be
2660 * inactive, so it is safe. Otherwise taking the access_lock
2661 * would cause a dead lock.
2664 case VHOST_USER_SET_FEATURES:
2665 case VHOST_USER_SET_PROTOCOL_FEATURES:
2666 case VHOST_USER_SET_OWNER:
2667 case VHOST_USER_SET_MEM_TABLE:
2668 case VHOST_USER_SET_LOG_BASE:
2669 case VHOST_USER_SET_LOG_FD:
2670 case VHOST_USER_SET_VRING_NUM:
2671 case VHOST_USER_SET_VRING_ADDR:
2672 case VHOST_USER_SET_VRING_BASE:
2673 case VHOST_USER_SET_VRING_KICK:
2674 case VHOST_USER_SET_VRING_CALL:
2675 case VHOST_USER_SET_VRING_ERR:
2676 case VHOST_USER_SET_VRING_ENABLE:
2677 case VHOST_USER_SEND_RARP:
2678 case VHOST_USER_NET_SET_MTU:
2679 case VHOST_USER_SET_SLAVE_REQ_FD:
2680 vhost_user_lock_all_queue_pairs(dev);
2681 unlock_required = 1;
2689 if (dev->extern_ops.pre_msg_handle) {
2690 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2693 case RTE_VHOST_MSG_RESULT_REPLY:
2694 send_vhost_reply(fd, &msg);
2696 case RTE_VHOST_MSG_RESULT_ERR:
2697 case RTE_VHOST_MSG_RESULT_OK:
2699 goto skip_to_post_handle;
2700 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2706 if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2707 if (!vhost_message_handlers[request])
2708 goto skip_to_post_handle;
2709 ret = vhost_message_handlers[request](&dev, &msg, fd);
2712 case RTE_VHOST_MSG_RESULT_ERR:
2713 VHOST_LOG_CONFIG(ERR,
2714 "Processing %s failed.\n",
2715 vhost_message_str[request]);
2718 case RTE_VHOST_MSG_RESULT_OK:
2719 VHOST_LOG_CONFIG(DEBUG,
2720 "Processing %s succeeded.\n",
2721 vhost_message_str[request]);
2724 case RTE_VHOST_MSG_RESULT_REPLY:
2725 VHOST_LOG_CONFIG(DEBUG,
2726 "Processing %s succeeded and needs reply.\n",
2727 vhost_message_str[request]);
2728 send_vhost_reply(fd, &msg);
2736 skip_to_post_handle:
2737 if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2738 dev->extern_ops.post_msg_handle) {
2739 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2742 case RTE_VHOST_MSG_RESULT_REPLY:
2743 send_vhost_reply(fd, &msg);
2745 case RTE_VHOST_MSG_RESULT_ERR:
2746 case RTE_VHOST_MSG_RESULT_OK:
2748 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2754 if (unlock_required)
2755 vhost_user_unlock_all_queue_pairs(dev);
2757 /* If message was not handled at this stage, treat it as an error */
2759 VHOST_LOG_CONFIG(ERR,
2760 "vhost message (req: %d) was not handled.\n", request);
2761 close_msg_fds(&msg);
2762 ret = RTE_VHOST_MSG_RESULT_ERR;
2766 * If the request required a reply that was already sent,
2767 * this optional reply-ack won't be sent as the
2768 * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2770 if (msg.flags & VHOST_USER_NEED_REPLY) {
2771 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2772 msg.size = sizeof(msg.payload.u64);
2774 send_vhost_reply(fd, &msg);
2775 } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2776 VHOST_LOG_CONFIG(ERR,
2777 "vhost message handling failed.\n");
2781 if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2782 dev->flags |= VIRTIO_DEV_READY;
2784 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2785 if (dev->dequeue_zero_copy) {
2786 VHOST_LOG_CONFIG(INFO,
2787 "dequeue zero copy is enabled\n");
2790 if (dev->notify_ops->new_device(dev->vid) == 0)
2791 dev->flags |= VIRTIO_DEV_RUNNING;
2795 did = dev->vdpa_dev_id;
2796 vdpa_dev = rte_vdpa_get_device(did);
2797 if (vdpa_dev && virtio_is_ready(dev) &&
2798 !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2799 msg.request.master == VHOST_USER_SET_VRING_CALL) {
2800 if (vdpa_dev->ops->dev_conf)
2801 vdpa_dev->ops->dev_conf(dev->vid);
2802 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2808 static int process_slave_message_reply(struct virtio_net *dev,
2809 const struct VhostUserMsg *msg)
2811 struct VhostUserMsg msg_reply;
2814 if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2817 ret = read_vhost_message(dev->slave_req_fd, &msg_reply);
2820 VHOST_LOG_CONFIG(ERR,
2821 "vhost read slave message reply failed\n");
2823 VHOST_LOG_CONFIG(INFO,
2824 "vhost peer closed\n");
2830 if (msg_reply.request.slave != msg->request.slave) {
2831 VHOST_LOG_CONFIG(ERR,
2832 "Received unexpected msg type (%u), expected %u\n",
2833 msg_reply.request.slave, msg->request.slave);
2838 ret = msg_reply.payload.u64 ? -1 : 0;
2841 rte_spinlock_unlock(&dev->slave_req_lock);
2846 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2849 struct VhostUserMsg msg = {
2850 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2851 .flags = VHOST_USER_VERSION,
2852 .size = sizeof(msg.payload.iotlb),
2856 .type = VHOST_IOTLB_MISS,
2860 ret = send_vhost_message(dev->slave_req_fd, &msg);
2862 VHOST_LOG_CONFIG(ERR,
2863 "Failed to send IOTLB miss message (%d)\n",
2872 vhost_user_slave_config_change(struct virtio_net *dev, bool need_reply)
2875 struct VhostUserMsg msg = {
2876 .request.slave = VHOST_USER_SLAVE_CONFIG_CHANGE_MSG,
2877 .flags = VHOST_USER_VERSION,
2882 msg.flags |= VHOST_USER_NEED_REPLY;
2884 ret = send_vhost_slave_message(dev, &msg);
2886 VHOST_LOG_CONFIG(ERR,
2887 "Failed to send config change (%d)\n",
2892 return process_slave_message_reply(dev, &msg);
2896 rte_vhost_slave_config_change(int vid, bool need_reply)
2898 struct virtio_net *dev;
2900 dev = get_device(vid);
2904 return vhost_user_slave_config_change(dev, need_reply);
2907 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2913 struct VhostUserMsg msg = {
2914 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2915 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2916 .size = sizeof(msg.payload.area),
2918 .u64 = index & VHOST_USER_VRING_IDX_MASK,
2925 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2931 ret = send_vhost_slave_message(dev, &msg);
2933 VHOST_LOG_CONFIG(ERR,
2934 "Failed to set host notifier (%d)\n", ret);
2938 return process_slave_message_reply(dev, &msg);
2941 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2943 struct virtio_net *dev;
2944 struct rte_vdpa_device *vdpa_dev;
2945 int vfio_device_fd, did, ret = 0;
2946 uint64_t offset, size;
2949 dev = get_device(vid);
2953 did = dev->vdpa_dev_id;
2957 if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2958 !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2959 !(dev->protocol_features &
2960 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2961 !(dev->protocol_features &
2962 (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2963 !(dev->protocol_features &
2964 (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2967 vdpa_dev = rte_vdpa_get_device(did);
2971 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2972 RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2974 vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2975 if (vfio_device_fd < 0)
2979 for (i = 0; i < dev->nr_vring; i++) {
2980 if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2986 if (vhost_user_slave_set_vring_host_notifier(dev, i,
2987 vfio_device_fd, offset, size) < 0) {
2994 for (i = 0; i < dev->nr_vring; i++) {
2995 vhost_user_slave_set_vring_host_notifier(dev, i, -1,