4 * Copyright Red Hat, Inc. 2010
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
13 #include <sys/ioctl.h>
17 #include <linux/vhost.h>
19 static void vhost_dev_sync_region(struct vhost_dev *dev,
20 uint64_t mfirst, uint64_t mlast,
21 uint64_t rfirst, uint64_t rlast)
23 uint64_t start = MAX(mfirst, rfirst);
24 uint64_t end = MIN(mlast, rlast);
25 vhost_log_chunk_t *from = dev->log + start / VHOST_LOG_CHUNK;
26 vhost_log_chunk_t *to = dev->log + end / VHOST_LOG_CHUNK + 1;
27 uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK;
29 assert(end / VHOST_LOG_CHUNK < dev->log_size);
30 assert(start / VHOST_LOG_CHUNK < dev->log_size);
34 for (;from < to; ++from) {
35 vhost_log_chunk_t log;
37 /* We first check with non-atomic: much cheaper,
38 * and we expect non-dirty to be the common case. */
40 addr += VHOST_LOG_CHUNK;
43 /* Data must be read atomically. We don't really
44 * need the barrier semantics of __sync
45 * builtins, but it's easier to use them than
47 log = __sync_fetch_and_and(from, 0);
48 while ((bit = sizeof(log) > sizeof(int) ?
49 ffsll(log) : ffs(log))) {
52 ram_addr = cpu_get_physical_page_desc(addr + bit * VHOST_LOG_PAGE);
53 cpu_physical_memory_set_dirty(ram_addr);
54 log &= ~(0x1ull << bit);
56 addr += VHOST_LOG_CHUNK;
60 static int vhost_client_sync_dirty_bitmap(CPUPhysMemoryClient *client,
61 target_phys_addr_t start_addr,
62 target_phys_addr_t end_addr)
64 struct vhost_dev *dev = container_of(client, struct vhost_dev, client);
66 if (!dev->log_enabled || !dev->started) {
69 for (i = 0; i < dev->mem->nregions; ++i) {
70 struct vhost_memory_region *reg = dev->mem->regions + i;
71 vhost_dev_sync_region(dev, start_addr, end_addr,
73 range_get_last(reg->guest_phys_addr,
76 for (i = 0; i < dev->nvqs; ++i) {
77 struct vhost_virtqueue *vq = dev->vqs + i;
78 vhost_dev_sync_region(dev, start_addr, end_addr, vq->used_phys,
79 range_get_last(vq->used_phys, vq->used_size));
84 /* Assign/unassign. Keep an unsorted array of non-overlapping
85 * memory regions in dev->mem. */
86 static void vhost_dev_unassign_memory(struct vhost_dev *dev,
90 int from, to, n = dev->mem->nregions;
91 /* Track overlapping/split regions for sanity checking. */
92 int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0;
94 for (from = 0, to = 0; from < n; ++from, ++to) {
95 struct vhost_memory_region *reg = dev->mem->regions + to;
100 /* clone old region */
102 memcpy(reg, dev->mem->regions + from, sizeof *reg);
105 /* No overlap is simple */
106 if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
111 /* Split only happens if supplied region
112 * is in the middle of an existing one. Thus it can not
113 * overlap with any other existing region. */
116 reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
117 memlast = range_get_last(start_addr, size);
119 /* Remove whole region */
120 if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
121 --dev->mem->nregions;
129 if (memlast >= reglast) {
130 reg->memory_size = start_addr - reg->guest_phys_addr;
131 assert(reg->memory_size);
132 assert(!overlap_end);
138 if (start_addr <= reg->guest_phys_addr) {
139 change = memlast + 1 - reg->guest_phys_addr;
140 reg->memory_size -= change;
141 reg->guest_phys_addr += change;
142 reg->userspace_addr += change;
143 assert(reg->memory_size);
144 assert(!overlap_start);
149 /* This only happens if supplied region
150 * is in the middle of an existing one. Thus it can not
151 * overlap with any other existing region. */
152 assert(!overlap_start);
153 assert(!overlap_end);
154 assert(!overlap_middle);
155 /* Split region: shrink first part, shift second part. */
156 memcpy(dev->mem->regions + n, reg, sizeof *reg);
157 reg->memory_size = start_addr - reg->guest_phys_addr;
158 assert(reg->memory_size);
159 change = memlast + 1 - reg->guest_phys_addr;
160 reg = dev->mem->regions + n;
161 reg->memory_size -= change;
162 assert(reg->memory_size);
163 reg->guest_phys_addr += change;
164 reg->userspace_addr += change;
165 /* Never add more than 1 region */
166 assert(dev->mem->nregions == n);
167 ++dev->mem->nregions;
172 /* Called after unassign, so no regions overlap the given range. */
173 static void vhost_dev_assign_memory(struct vhost_dev *dev,
179 struct vhost_memory_region *merged = NULL;
180 for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) {
181 struct vhost_memory_region *reg = dev->mem->regions + to;
182 uint64_t prlast, urlast;
183 uint64_t pmlast, umlast;
186 /* clone old region */
188 memcpy(reg, dev->mem->regions + from, sizeof *reg);
190 prlast = range_get_last(reg->guest_phys_addr, reg->memory_size);
191 pmlast = range_get_last(start_addr, size);
192 urlast = range_get_last(reg->userspace_addr, reg->memory_size);
193 umlast = range_get_last(uaddr, size);
195 /* check for overlapping regions: should never happen. */
196 assert(prlast < start_addr || pmlast < reg->guest_phys_addr);
197 /* Not an adjacent or overlapping region - do not merge. */
198 if ((prlast + 1 != start_addr || urlast + 1 != uaddr) &&
199 (pmlast + 1 != reg->guest_phys_addr ||
200 umlast + 1 != reg->userspace_addr)) {
210 u = MIN(uaddr, reg->userspace_addr);
211 s = MIN(start_addr, reg->guest_phys_addr);
212 e = MAX(pmlast, prlast);
213 uaddr = merged->userspace_addr = u;
214 start_addr = merged->guest_phys_addr = s;
215 size = merged->memory_size = e - s + 1;
216 assert(merged->memory_size);
220 struct vhost_memory_region *reg = dev->mem->regions + to;
221 memset(reg, 0, sizeof *reg);
222 reg->memory_size = size;
223 assert(reg->memory_size);
224 reg->guest_phys_addr = start_addr;
225 reg->userspace_addr = uaddr;
228 assert(to <= dev->mem->nregions + 1);
229 dev->mem->nregions = to;
232 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
234 uint64_t log_size = 0;
236 for (i = 0; i < dev->mem->nregions; ++i) {
237 struct vhost_memory_region *reg = dev->mem->regions + i;
238 uint64_t last = range_get_last(reg->guest_phys_addr,
240 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
242 for (i = 0; i < dev->nvqs; ++i) {
243 struct vhost_virtqueue *vq = dev->vqs + i;
244 uint64_t last = vq->used_phys + vq->used_size - 1;
245 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
250 static inline void vhost_dev_log_resize(struct vhost_dev* dev, uint64_t size)
252 vhost_log_chunk_t *log;
256 log = qemu_mallocz(size * sizeof *log);
260 log_base = (uint64_t)(unsigned long)log;
261 r = ioctl(dev->control, VHOST_SET_LOG_BASE, &log_base);
263 vhost_client_sync_dirty_bitmap(&dev->client, 0,
264 (target_phys_addr_t)~0x0ull);
269 dev->log_size = size;
272 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
277 for (i = 0; i < dev->nvqs; ++i) {
278 struct vhost_virtqueue *vq = dev->vqs + i;
279 target_phys_addr_t l;
282 if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
286 p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
287 if (!p || l != vq->ring_size) {
288 fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
292 fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
295 cpu_physical_memory_unmap(p, l, 0, 0);
300 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
304 int i, n = dev->mem->nregions;
305 for (i = 0; i < n; ++i) {
306 struct vhost_memory_region *reg = dev->mem->regions + i;
307 if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
315 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
320 struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
328 reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
329 memlast = range_get_last(start_addr, size);
331 /* Need to extend region? */
332 if (start_addr < reg->guest_phys_addr || memlast > reglast) {
335 /* userspace_addr changed? */
336 return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
339 static void vhost_client_set_memory(CPUPhysMemoryClient *client,
340 target_phys_addr_t start_addr,
342 ram_addr_t phys_offset,
345 struct vhost_dev *dev = container_of(client, struct vhost_dev, client);
346 ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
347 int s = offsetof(struct vhost_memory, regions) +
348 (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
352 dev->mem = qemu_realloc(dev->mem, s);
355 flags = IO_MEM_UNASSIGNED;
360 /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
361 if (flags == IO_MEM_RAM) {
362 if (!vhost_dev_cmp_memory(dev, start_addr, size,
363 (uintptr_t)qemu_get_ram_ptr(phys_offset))) {
364 /* Region exists with same address. Nothing to do. */
368 if (!vhost_dev_find_reg(dev, start_addr, size)) {
369 /* Removing region that we don't access. Nothing to do. */
374 vhost_dev_unassign_memory(dev, start_addr, size);
375 if (flags == IO_MEM_RAM) {
376 /* Add given mapping, merging adjacent regions if any */
377 vhost_dev_assign_memory(dev, start_addr, size,
378 (uintptr_t)qemu_get_ram_ptr(phys_offset));
380 /* Remove old mapping for this memory, if any. */
381 vhost_dev_unassign_memory(dev, start_addr, size);
389 r = vhost_verify_ring_mappings(dev, start_addr, size);
393 if (!dev->log_enabled) {
394 r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
398 log_size = vhost_get_log_size(dev);
399 /* We allocate an extra 4K bytes to log,
400 * to reduce the * number of reallocations. */
401 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
402 /* To log more, must increase log size before table update. */
403 if (dev->log_size < log_size) {
404 vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
406 r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
408 /* To log less, can only decrease log size after table update. */
409 if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
410 vhost_dev_log_resize(dev, log_size);
414 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
415 struct vhost_virtqueue *vq,
416 unsigned idx, bool enable_log)
418 struct vhost_vring_addr addr = {
420 .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
421 .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
422 .used_user_addr = (uint64_t)(unsigned long)vq->used,
423 .log_guest_addr = vq->used_phys,
424 .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
426 int r = ioctl(dev->control, VHOST_SET_VRING_ADDR, &addr);
433 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
435 uint64_t features = dev->acked_features;
438 features |= 0x1 << VHOST_F_LOG_ALL;
440 r = ioctl(dev->control, VHOST_SET_FEATURES, &features);
441 return r < 0 ? -errno : 0;
444 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
447 r = vhost_dev_set_features(dev, enable_log);
451 for (i = 0; i < dev->nvqs; ++i) {
452 r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
460 for (; i >= 0; --i) {
461 t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
465 t = vhost_dev_set_features(dev, dev->log_enabled);
471 static int vhost_client_migration_log(CPUPhysMemoryClient *client,
474 struct vhost_dev *dev = container_of(client, struct vhost_dev, client);
476 if (!!enable == dev->log_enabled) {
480 dev->log_enabled = enable;
484 r = vhost_dev_set_log(dev, false);
494 vhost_dev_log_resize(dev, vhost_get_log_size(dev));
495 r = vhost_dev_set_log(dev, true);
500 dev->log_enabled = enable;
504 static int vhost_virtqueue_init(struct vhost_dev *dev,
505 struct VirtIODevice *vdev,
506 struct vhost_virtqueue *vq,
509 target_phys_addr_t s, l, a;
511 struct vhost_vring_file file = {
514 struct vhost_vring_state state = {
517 struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
519 if (!vdev->binding->set_host_notifier) {
520 fprintf(stderr, "binding does not support host notifiers\n");
524 vq->num = state.num = virtio_queue_get_num(vdev, idx);
525 r = ioctl(dev->control, VHOST_SET_VRING_NUM, &state);
530 state.num = virtio_queue_get_last_avail_idx(vdev, idx);
531 r = ioctl(dev->control, VHOST_SET_VRING_BASE, &state);
536 s = l = virtio_queue_get_desc_size(vdev, idx);
537 a = virtio_queue_get_desc_addr(vdev, idx);
538 vq->desc = cpu_physical_memory_map(a, &l, 0);
539 if (!vq->desc || l != s) {
541 goto fail_alloc_desc;
543 s = l = virtio_queue_get_avail_size(vdev, idx);
544 a = virtio_queue_get_avail_addr(vdev, idx);
545 vq->avail = cpu_physical_memory_map(a, &l, 0);
546 if (!vq->avail || l != s) {
548 goto fail_alloc_avail;
550 vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
551 vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
552 vq->used = cpu_physical_memory_map(a, &l, 1);
553 if (!vq->used || l != s) {
555 goto fail_alloc_used;
558 vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
559 vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
560 vq->ring = cpu_physical_memory_map(a, &l, 1);
561 if (!vq->ring || l != s) {
563 goto fail_alloc_ring;
566 r = vhost_virtqueue_set_addr(dev, vq, idx, dev->log_enabled);
571 r = vdev->binding->set_host_notifier(vdev->binding_opaque, idx, true);
573 fprintf(stderr, "Error binding host notifier: %d\n", -r);
574 goto fail_host_notifier;
577 file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
578 r = ioctl(dev->control, VHOST_SET_VRING_KICK, &file);
584 file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
585 r = ioctl(dev->control, VHOST_SET_VRING_CALL, &file);
595 vdev->binding->set_host_notifier(vdev->binding_opaque, idx, false);
598 cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
601 cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
604 cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
607 cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
613 static void vhost_virtqueue_cleanup(struct vhost_dev *dev,
614 struct VirtIODevice *vdev,
615 struct vhost_virtqueue *vq,
618 struct vhost_vring_state state = {
622 r = vdev->binding->set_host_notifier(vdev->binding_opaque, idx, false);
624 fprintf(stderr, "vhost VQ %d host cleanup failed: %d\n", idx, r);
628 r = ioctl(dev->control, VHOST_GET_VRING_BASE, &state);
630 fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
633 virtio_queue_set_last_avail_idx(vdev, idx, state.num);
635 cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
636 0, virtio_queue_get_ring_size(vdev, idx));
637 cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
638 1, virtio_queue_get_used_size(vdev, idx));
639 cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
640 0, virtio_queue_get_avail_size(vdev, idx));
641 cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
642 0, virtio_queue_get_desc_size(vdev, idx));
645 int vhost_dev_init(struct vhost_dev *hdev, int devfd, bool force)
650 hdev->control = devfd;
652 hdev->control = open("/dev/vhost-net", O_RDWR);
653 if (hdev->control < 0) {
657 r = ioctl(hdev->control, VHOST_SET_OWNER, NULL);
662 r = ioctl(hdev->control, VHOST_GET_FEATURES, &features);
666 hdev->features = features;
668 hdev->client.set_memory = vhost_client_set_memory;
669 hdev->client.sync_dirty_bitmap = vhost_client_sync_dirty_bitmap;
670 hdev->client.migration_log = vhost_client_migration_log;
671 hdev->client.log_start = NULL;
672 hdev->client.log_stop = NULL;
673 hdev->mem = qemu_mallocz(offsetof(struct vhost_memory, regions));
676 hdev->log_enabled = false;
677 hdev->started = false;
678 cpu_register_phys_memory_client(&hdev->client);
683 close(hdev->control);
687 void vhost_dev_cleanup(struct vhost_dev *hdev)
689 cpu_unregister_phys_memory_client(&hdev->client);
690 qemu_free(hdev->mem);
691 close(hdev->control);
694 bool vhost_dev_query(struct vhost_dev *hdev, VirtIODevice *vdev)
696 return !vdev->binding->query_guest_notifiers ||
697 vdev->binding->query_guest_notifiers(vdev->binding_opaque) ||
701 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
704 if (!vdev->binding->set_guest_notifiers) {
705 fprintf(stderr, "binding does not support guest notifiers\n");
710 r = vdev->binding->set_guest_notifiers(vdev->binding_opaque, true);
712 fprintf(stderr, "Error binding guest notifier: %d\n", -r);
716 r = vhost_dev_set_features(hdev, hdev->log_enabled);
720 r = ioctl(hdev->control, VHOST_SET_MEM_TABLE, hdev->mem);
725 for (i = 0; i < hdev->nvqs; ++i) {
726 r = vhost_virtqueue_init(hdev,
735 if (hdev->log_enabled) {
736 hdev->log_size = vhost_get_log_size(hdev);
737 hdev->log = hdev->log_size ?
738 qemu_mallocz(hdev->log_size * sizeof *hdev->log) : NULL;
739 r = ioctl(hdev->control, VHOST_SET_LOG_BASE,
740 (uint64_t)(unsigned long)hdev->log);
747 hdev->started = true;
753 vhost_virtqueue_cleanup(hdev,
760 vdev->binding->set_guest_notifiers(vdev->binding_opaque, false);
766 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
770 for (i = 0; i < hdev->nvqs; ++i) {
771 vhost_virtqueue_cleanup(hdev,
776 vhost_client_sync_dirty_bitmap(&hdev->client, 0,
777 (target_phys_addr_t)~0x0ull);
778 r = vdev->binding->set_guest_notifiers(vdev->binding_opaque, false);
780 fprintf(stderr, "vhost guest notifier cleanup failed: %d\n", r);
785 hdev->started = false;
786 qemu_free(hdev->log);