4 * Copyright Red Hat, Inc. 2010
7 * Michael S. Tsirkin <mst@redhat.com>
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
12 * Contributions after 2012-01-13 are licensed under the terms of the
13 * GNU GPL, version 2 or (at your option) any later version.
16 #include "qemu/osdep.h"
17 #include "qapi/error.h"
18 #include "hw/virtio/vhost.h"
20 #include "qemu/atomic.h"
21 #include "qemu/range.h"
22 #include "qemu/error-report.h"
23 #include "qemu/memfd.h"
24 #include <linux/vhost.h>
25 #include "exec/address-spaces.h"
26 #include "hw/virtio/virtio-bus.h"
27 #include "hw/virtio/virtio-access.h"
28 #include "migration/migration.h"
29 #include "sysemu/dma.h"
31 /* enabled until disconnected backend stabilizes */
32 #define _VHOST_DEBUG 1
35 #define VHOST_OPS_DEBUG(fmt, ...) \
36 do { error_report(fmt ": %s (%d)", ## __VA_ARGS__, \
37 strerror(errno), errno); } while (0)
39 #define VHOST_OPS_DEBUG(fmt, ...) \
43 static struct vhost_log
*vhost_log
;
44 static struct vhost_log
*vhost_log_shm
;
46 static unsigned int used_memslots
;
47 static QLIST_HEAD(, vhost_dev
) vhost_devices
=
48 QLIST_HEAD_INITIALIZER(vhost_devices
);
50 bool vhost_has_free_slot(void)
52 unsigned int slots_limit
= ~0U;
53 struct vhost_dev
*hdev
;
55 QLIST_FOREACH(hdev
, &vhost_devices
, entry
) {
56 unsigned int r
= hdev
->vhost_ops
->vhost_backend_memslots_limit(hdev
);
57 slots_limit
= MIN(slots_limit
, r
);
59 return slots_limit
> used_memslots
;
62 static void vhost_dev_sync_region(struct vhost_dev
*dev
,
63 MemoryRegionSection
*section
,
64 uint64_t mfirst
, uint64_t mlast
,
65 uint64_t rfirst
, uint64_t rlast
)
67 vhost_log_chunk_t
*log
= dev
->log
->log
;
69 uint64_t start
= MAX(mfirst
, rfirst
);
70 uint64_t end
= MIN(mlast
, rlast
);
71 vhost_log_chunk_t
*from
= log
+ start
/ VHOST_LOG_CHUNK
;
72 vhost_log_chunk_t
*to
= log
+ end
/ VHOST_LOG_CHUNK
+ 1;
73 uint64_t addr
= (start
/ VHOST_LOG_CHUNK
) * VHOST_LOG_CHUNK
;
78 assert(end
/ VHOST_LOG_CHUNK
< dev
->log_size
);
79 assert(start
/ VHOST_LOG_CHUNK
< dev
->log_size
);
81 for (;from
< to
; ++from
) {
82 vhost_log_chunk_t log
;
83 /* We first check with non-atomic: much cheaper,
84 * and we expect non-dirty to be the common case. */
86 addr
+= VHOST_LOG_CHUNK
;
89 /* Data must be read atomically. We don't really need barrier semantics
90 * but it's easier to use atomic_* than roll our own. */
91 log
= atomic_xchg(from
, 0);
95 hwaddr section_offset
;
97 page_addr
= addr
+ bit
* VHOST_LOG_PAGE
;
98 section_offset
= page_addr
- section
->offset_within_address_space
;
99 mr_offset
= section_offset
+ section
->offset_within_region
;
100 memory_region_set_dirty(section
->mr
, mr_offset
, VHOST_LOG_PAGE
);
101 log
&= ~(0x1ull
<< bit
);
103 addr
+= VHOST_LOG_CHUNK
;
107 static int vhost_sync_dirty_bitmap(struct vhost_dev
*dev
,
108 MemoryRegionSection
*section
,
116 if (!dev
->log_enabled
|| !dev
->started
) {
119 start_addr
= section
->offset_within_address_space
;
120 end_addr
= range_get_last(start_addr
, int128_get64(section
->size
));
121 start_addr
= MAX(first
, start_addr
);
122 end_addr
= MIN(last
, end_addr
);
124 for (i
= 0; i
< dev
->mem
->nregions
; ++i
) {
125 struct vhost_memory_region
*reg
= dev
->mem
->regions
+ i
;
126 vhost_dev_sync_region(dev
, section
, start_addr
, end_addr
,
127 reg
->guest_phys_addr
,
128 range_get_last(reg
->guest_phys_addr
,
131 for (i
= 0; i
< dev
->nvqs
; ++i
) {
132 struct vhost_virtqueue
*vq
= dev
->vqs
+ i
;
133 vhost_dev_sync_region(dev
, section
, start_addr
, end_addr
, vq
->used_phys
,
134 range_get_last(vq
->used_phys
, vq
->used_size
));
139 static void vhost_log_sync(MemoryListener
*listener
,
140 MemoryRegionSection
*section
)
142 struct vhost_dev
*dev
= container_of(listener
, struct vhost_dev
,
144 vhost_sync_dirty_bitmap(dev
, section
, 0x0, ~0x0ULL
);
147 static void vhost_log_sync_range(struct vhost_dev
*dev
,
148 hwaddr first
, hwaddr last
)
151 /* FIXME: this is N^2 in number of sections */
152 for (i
= 0; i
< dev
->n_mem_sections
; ++i
) {
153 MemoryRegionSection
*section
= &dev
->mem_sections
[i
];
154 vhost_sync_dirty_bitmap(dev
, section
, first
, last
);
158 /* Assign/unassign. Keep an unsorted array of non-overlapping
159 * memory regions in dev->mem. */
160 static void vhost_dev_unassign_memory(struct vhost_dev
*dev
,
164 int from
, to
, n
= dev
->mem
->nregions
;
165 /* Track overlapping/split regions for sanity checking. */
166 int overlap_start
= 0, overlap_end
= 0, overlap_middle
= 0, split
= 0;
168 for (from
= 0, to
= 0; from
< n
; ++from
, ++to
) {
169 struct vhost_memory_region
*reg
= dev
->mem
->regions
+ to
;
174 /* clone old region */
176 memcpy(reg
, dev
->mem
->regions
+ from
, sizeof *reg
);
179 /* No overlap is simple */
180 if (!ranges_overlap(reg
->guest_phys_addr
, reg
->memory_size
,
185 /* Split only happens if supplied region
186 * is in the middle of an existing one. Thus it can not
187 * overlap with any other existing region. */
190 reglast
= range_get_last(reg
->guest_phys_addr
, reg
->memory_size
);
191 memlast
= range_get_last(start_addr
, size
);
193 /* Remove whole region */
194 if (start_addr
<= reg
->guest_phys_addr
&& memlast
>= reglast
) {
195 --dev
->mem
->nregions
;
202 if (memlast
>= reglast
) {
203 reg
->memory_size
= start_addr
- reg
->guest_phys_addr
;
204 assert(reg
->memory_size
);
205 assert(!overlap_end
);
211 if (start_addr
<= reg
->guest_phys_addr
) {
212 change
= memlast
+ 1 - reg
->guest_phys_addr
;
213 reg
->memory_size
-= change
;
214 reg
->guest_phys_addr
+= change
;
215 reg
->userspace_addr
+= change
;
216 assert(reg
->memory_size
);
217 assert(!overlap_start
);
222 /* This only happens if supplied region
223 * is in the middle of an existing one. Thus it can not
224 * overlap with any other existing region. */
225 assert(!overlap_start
);
226 assert(!overlap_end
);
227 assert(!overlap_middle
);
228 /* Split region: shrink first part, shift second part. */
229 memcpy(dev
->mem
->regions
+ n
, reg
, sizeof *reg
);
230 reg
->memory_size
= start_addr
- reg
->guest_phys_addr
;
231 assert(reg
->memory_size
);
232 change
= memlast
+ 1 - reg
->guest_phys_addr
;
233 reg
= dev
->mem
->regions
+ n
;
234 reg
->memory_size
-= change
;
235 assert(reg
->memory_size
);
236 reg
->guest_phys_addr
+= change
;
237 reg
->userspace_addr
+= change
;
238 /* Never add more than 1 region */
239 assert(dev
->mem
->nregions
== n
);
240 ++dev
->mem
->nregions
;
245 /* Called after unassign, so no regions overlap the given range. */
246 static void vhost_dev_assign_memory(struct vhost_dev
*dev
,
252 struct vhost_memory_region
*merged
= NULL
;
253 for (from
= 0, to
= 0; from
< dev
->mem
->nregions
; ++from
, ++to
) {
254 struct vhost_memory_region
*reg
= dev
->mem
->regions
+ to
;
255 uint64_t prlast
, urlast
;
256 uint64_t pmlast
, umlast
;
259 /* clone old region */
261 memcpy(reg
, dev
->mem
->regions
+ from
, sizeof *reg
);
263 prlast
= range_get_last(reg
->guest_phys_addr
, reg
->memory_size
);
264 pmlast
= range_get_last(start_addr
, size
);
265 urlast
= range_get_last(reg
->userspace_addr
, reg
->memory_size
);
266 umlast
= range_get_last(uaddr
, size
);
268 /* check for overlapping regions: should never happen. */
269 assert(prlast
< start_addr
|| pmlast
< reg
->guest_phys_addr
);
270 /* Not an adjacent or overlapping region - do not merge. */
271 if ((prlast
+ 1 != start_addr
|| urlast
+ 1 != uaddr
) &&
272 (pmlast
+ 1 != reg
->guest_phys_addr
||
273 umlast
+ 1 != reg
->userspace_addr
)) {
277 if (dev
->vhost_ops
->vhost_backend_can_merge
&&
278 !dev
->vhost_ops
->vhost_backend_can_merge(dev
, uaddr
, size
,
290 u
= MIN(uaddr
, reg
->userspace_addr
);
291 s
= MIN(start_addr
, reg
->guest_phys_addr
);
292 e
= MAX(pmlast
, prlast
);
293 uaddr
= merged
->userspace_addr
= u
;
294 start_addr
= merged
->guest_phys_addr
= s
;
295 size
= merged
->memory_size
= e
- s
+ 1;
296 assert(merged
->memory_size
);
300 struct vhost_memory_region
*reg
= dev
->mem
->regions
+ to
;
301 memset(reg
, 0, sizeof *reg
);
302 reg
->memory_size
= size
;
303 assert(reg
->memory_size
);
304 reg
->guest_phys_addr
= start_addr
;
305 reg
->userspace_addr
= uaddr
;
308 assert(to
<= dev
->mem
->nregions
+ 1);
309 dev
->mem
->nregions
= to
;
312 static uint64_t vhost_get_log_size(struct vhost_dev
*dev
)
314 uint64_t log_size
= 0;
316 for (i
= 0; i
< dev
->mem
->nregions
; ++i
) {
317 struct vhost_memory_region
*reg
= dev
->mem
->regions
+ i
;
318 uint64_t last
= range_get_last(reg
->guest_phys_addr
,
320 log_size
= MAX(log_size
, last
/ VHOST_LOG_CHUNK
+ 1);
322 for (i
= 0; i
< dev
->nvqs
; ++i
) {
323 struct vhost_virtqueue
*vq
= dev
->vqs
+ i
;
324 uint64_t last
= vq
->used_phys
+ vq
->used_size
- 1;
325 log_size
= MAX(log_size
, last
/ VHOST_LOG_CHUNK
+ 1);
330 static struct vhost_log
*vhost_log_alloc(uint64_t size
, bool share
)
332 struct vhost_log
*log
;
333 uint64_t logsize
= size
* sizeof(*(log
->log
));
336 log
= g_new0(struct vhost_log
, 1);
338 log
->log
= qemu_memfd_alloc("vhost-log", logsize
,
339 F_SEAL_GROW
| F_SEAL_SHRINK
| F_SEAL_SEAL
,
341 memset(log
->log
, 0, logsize
);
343 log
->log
= g_malloc0(logsize
);
353 static struct vhost_log
*vhost_log_get(uint64_t size
, bool share
)
355 struct vhost_log
*log
= share
? vhost_log_shm
: vhost_log
;
357 if (!log
|| log
->size
!= size
) {
358 log
= vhost_log_alloc(size
, share
);
371 static void vhost_log_put(struct vhost_dev
*dev
, bool sync
)
373 struct vhost_log
*log
= dev
->log
;
382 if (log
->refcnt
== 0) {
383 /* Sync only the range covered by the old log */
384 if (dev
->log_size
&& sync
) {
385 vhost_log_sync_range(dev
, 0, dev
->log_size
* VHOST_LOG_CHUNK
- 1);
388 if (vhost_log
== log
) {
391 } else if (vhost_log_shm
== log
) {
392 qemu_memfd_free(log
->log
, log
->size
* sizeof(*(log
->log
)),
394 vhost_log_shm
= NULL
;
401 static bool vhost_dev_log_is_shared(struct vhost_dev
*dev
)
403 return dev
->vhost_ops
->vhost_requires_shm_log
&&
404 dev
->vhost_ops
->vhost_requires_shm_log(dev
);
407 static inline void vhost_dev_log_resize(struct vhost_dev
*dev
, uint64_t size
)
409 struct vhost_log
*log
= vhost_log_get(size
, vhost_dev_log_is_shared(dev
));
410 uint64_t log_base
= (uintptr_t)log
->log
;
413 /* inform backend of log switching, this must be done before
414 releasing the current log, to ensure no logging is lost */
415 r
= dev
->vhost_ops
->vhost_set_log_base(dev
, log_base
, log
);
417 VHOST_OPS_DEBUG("vhost_set_log_base failed");
420 vhost_log_put(dev
, true);
422 dev
->log_size
= size
;
425 static int vhost_dev_has_iommu(struct vhost_dev
*dev
)
427 VirtIODevice
*vdev
= dev
->vdev
;
428 AddressSpace
*dma_as
= vdev
->dma_as
;
430 return memory_region_is_iommu(dma_as
->root
) &&
431 virtio_host_has_feature(vdev
, VIRTIO_F_IOMMU_PLATFORM
);
434 static void *vhost_memory_map(struct vhost_dev
*dev
, hwaddr addr
,
435 hwaddr
*plen
, int is_write
)
437 if (!vhost_dev_has_iommu(dev
)) {
438 return cpu_physical_memory_map(addr
, plen
, is_write
);
440 return (void *)(uintptr_t)addr
;
444 static void vhost_memory_unmap(struct vhost_dev
*dev
, void *buffer
,
445 hwaddr len
, int is_write
,
448 if (!vhost_dev_has_iommu(dev
)) {
449 cpu_physical_memory_unmap(buffer
, len
, is_write
, access_len
);
453 static int vhost_verify_ring_part_mapping(struct vhost_dev
*dev
,
464 if (!ranges_overlap(start_addr
, size
, part_addr
, part_size
)) {
468 p
= vhost_memory_map(dev
, part_addr
, &l
, 1);
469 if (!p
|| l
!= part_size
) {
475 vhost_memory_unmap(dev
, p
, l
, 0, 0);
479 static int vhost_verify_ring_mappings(struct vhost_dev
*dev
,
485 const char *part_name
[] = {
491 for (i
= 0; i
< dev
->nvqs
; ++i
) {
492 struct vhost_virtqueue
*vq
= dev
->vqs
+ i
;
495 r
= vhost_verify_ring_part_mapping(dev
, vq
->desc
, vq
->desc_phys
,
496 vq
->desc_size
, start_addr
, size
);
502 r
= vhost_verify_ring_part_mapping(dev
, vq
->avail
, vq
->avail_phys
,
503 vq
->avail_size
, start_addr
, size
);
509 r
= vhost_verify_ring_part_mapping(dev
, vq
->used
, vq
->used_phys
,
510 vq
->used_size
, start_addr
, size
);
517 error_report("Unable to map %s for ring %d", part_name
[j
], i
);
518 } else if (r
== -EBUSY
) {
519 error_report("%s relocated for ring %d", part_name
[j
], i
);
524 static struct vhost_memory_region
*vhost_dev_find_reg(struct vhost_dev
*dev
,
528 int i
, n
= dev
->mem
->nregions
;
529 for (i
= 0; i
< n
; ++i
) {
530 struct vhost_memory_region
*reg
= dev
->mem
->regions
+ i
;
531 if (ranges_overlap(reg
->guest_phys_addr
, reg
->memory_size
,
539 static bool vhost_dev_cmp_memory(struct vhost_dev
*dev
,
544 struct vhost_memory_region
*reg
= vhost_dev_find_reg(dev
, start_addr
, size
);
552 reglast
= range_get_last(reg
->guest_phys_addr
, reg
->memory_size
);
553 memlast
= range_get_last(start_addr
, size
);
555 /* Need to extend region? */
556 if (start_addr
< reg
->guest_phys_addr
|| memlast
> reglast
) {
559 /* userspace_addr changed? */
560 return uaddr
!= reg
->userspace_addr
+ start_addr
- reg
->guest_phys_addr
;
563 static void vhost_set_memory(MemoryListener
*listener
,
564 MemoryRegionSection
*section
,
567 struct vhost_dev
*dev
= container_of(listener
, struct vhost_dev
,
569 hwaddr start_addr
= section
->offset_within_address_space
;
570 ram_addr_t size
= int128_get64(section
->size
);
572 memory_region_get_dirty_log_mask(section
->mr
) & ~(1 << DIRTY_MEMORY_MIGRATION
);
573 int s
= offsetof(struct vhost_memory
, regions
) +
574 (dev
->mem
->nregions
+ 1) * sizeof dev
->mem
->regions
[0];
577 dev
->mem
= g_realloc(dev
->mem
, s
);
585 /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
586 ram
= memory_region_get_ram_ptr(section
->mr
) + section
->offset_within_region
;
588 if (!vhost_dev_cmp_memory(dev
, start_addr
, size
, (uintptr_t)ram
)) {
589 /* Region exists with same address. Nothing to do. */
593 if (!vhost_dev_find_reg(dev
, start_addr
, size
)) {
594 /* Removing region that we don't access. Nothing to do. */
599 vhost_dev_unassign_memory(dev
, start_addr
, size
);
601 /* Add given mapping, merging adjacent regions if any */
602 vhost_dev_assign_memory(dev
, start_addr
, size
, (uintptr_t)ram
);
604 /* Remove old mapping for this memory, if any. */
605 vhost_dev_unassign_memory(dev
, start_addr
, size
);
607 dev
->mem_changed_start_addr
= MIN(dev
->mem_changed_start_addr
, start_addr
);
608 dev
->mem_changed_end_addr
= MAX(dev
->mem_changed_end_addr
, start_addr
+ size
- 1);
609 dev
->memory_changed
= true;
610 used_memslots
= dev
->mem
->nregions
;
613 static bool vhost_section(MemoryRegionSection
*section
)
615 return memory_region_is_ram(section
->mr
);
618 static void vhost_begin(MemoryListener
*listener
)
620 struct vhost_dev
*dev
= container_of(listener
, struct vhost_dev
,
622 dev
->mem_changed_end_addr
= 0;
623 dev
->mem_changed_start_addr
= -1;
626 static void vhost_commit(MemoryListener
*listener
)
628 struct vhost_dev
*dev
= container_of(listener
, struct vhost_dev
,
630 hwaddr start_addr
= 0;
635 if (!dev
->memory_changed
) {
641 if (dev
->mem_changed_start_addr
> dev
->mem_changed_end_addr
) {
646 start_addr
= dev
->mem_changed_start_addr
;
647 size
= dev
->mem_changed_end_addr
- dev
->mem_changed_start_addr
+ 1;
649 r
= vhost_verify_ring_mappings(dev
, start_addr
, size
);
653 if (!dev
->log_enabled
) {
654 r
= dev
->vhost_ops
->vhost_set_mem_table(dev
, dev
->mem
);
656 VHOST_OPS_DEBUG("vhost_set_mem_table failed");
658 dev
->memory_changed
= false;
661 log_size
= vhost_get_log_size(dev
);
662 /* We allocate an extra 4K bytes to log,
663 * to reduce the * number of reallocations. */
664 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
665 /* To log more, must increase log size before table update. */
666 if (dev
->log_size
< log_size
) {
667 vhost_dev_log_resize(dev
, log_size
+ VHOST_LOG_BUFFER
);
669 r
= dev
->vhost_ops
->vhost_set_mem_table(dev
, dev
->mem
);
671 VHOST_OPS_DEBUG("vhost_set_mem_table failed");
673 /* To log less, can only decrease log size after table update. */
674 if (dev
->log_size
> log_size
+ VHOST_LOG_BUFFER
) {
675 vhost_dev_log_resize(dev
, log_size
);
677 dev
->memory_changed
= false;
680 static void vhost_region_add(MemoryListener
*listener
,
681 MemoryRegionSection
*section
)
683 struct vhost_dev
*dev
= container_of(listener
, struct vhost_dev
,
686 if (!vhost_section(section
)) {
690 ++dev
->n_mem_sections
;
691 dev
->mem_sections
= g_renew(MemoryRegionSection
, dev
->mem_sections
,
692 dev
->n_mem_sections
);
693 dev
->mem_sections
[dev
->n_mem_sections
- 1] = *section
;
694 memory_region_ref(section
->mr
);
695 vhost_set_memory(listener
, section
, true);
698 static void vhost_region_del(MemoryListener
*listener
,
699 MemoryRegionSection
*section
)
701 struct vhost_dev
*dev
= container_of(listener
, struct vhost_dev
,
705 if (!vhost_section(section
)) {
709 vhost_set_memory(listener
, section
, false);
710 memory_region_unref(section
->mr
);
711 for (i
= 0; i
< dev
->n_mem_sections
; ++i
) {
712 if (dev
->mem_sections
[i
].offset_within_address_space
713 == section
->offset_within_address_space
) {
714 --dev
->n_mem_sections
;
715 memmove(&dev
->mem_sections
[i
], &dev
->mem_sections
[i
+1],
716 (dev
->n_mem_sections
- i
) * sizeof(*dev
->mem_sections
));
722 static void vhost_region_nop(MemoryListener
*listener
,
723 MemoryRegionSection
*section
)
727 static int vhost_virtqueue_set_addr(struct vhost_dev
*dev
,
728 struct vhost_virtqueue
*vq
,
729 unsigned idx
, bool enable_log
)
731 struct vhost_vring_addr addr
= {
733 .desc_user_addr
= (uint64_t)(unsigned long)vq
->desc
,
734 .avail_user_addr
= (uint64_t)(unsigned long)vq
->avail
,
735 .used_user_addr
= (uint64_t)(unsigned long)vq
->used
,
736 .log_guest_addr
= vq
->used_phys
,
737 .flags
= enable_log
? (1 << VHOST_VRING_F_LOG
) : 0,
739 int r
= dev
->vhost_ops
->vhost_set_vring_addr(dev
, &addr
);
741 VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
747 static int vhost_dev_set_features(struct vhost_dev
*dev
,
750 uint64_t features
= dev
->acked_features
;
753 features
|= 0x1ULL
<< VHOST_F_LOG_ALL
;
755 r
= dev
->vhost_ops
->vhost_set_features(dev
, features
);
757 VHOST_OPS_DEBUG("vhost_set_features failed");
759 return r
< 0 ? -errno
: 0;
762 static int vhost_dev_set_log(struct vhost_dev
*dev
, bool enable_log
)
765 r
= vhost_dev_set_features(dev
, enable_log
);
769 for (i
= 0; i
< dev
->nvqs
; ++i
) {
770 idx
= dev
->vhost_ops
->vhost_get_vq_index(dev
, dev
->vq_index
+ i
);
771 r
= vhost_virtqueue_set_addr(dev
, dev
->vqs
+ i
, idx
,
779 for (; i
>= 0; --i
) {
780 idx
= dev
->vhost_ops
->vhost_get_vq_index(dev
, dev
->vq_index
+ i
);
781 vhost_virtqueue_set_addr(dev
, dev
->vqs
+ i
, idx
,
784 vhost_dev_set_features(dev
, dev
->log_enabled
);
789 static int vhost_migration_log(MemoryListener
*listener
, int enable
)
791 struct vhost_dev
*dev
= container_of(listener
, struct vhost_dev
,
794 if (!!enable
== dev
->log_enabled
) {
798 dev
->log_enabled
= enable
;
802 r
= vhost_dev_set_log(dev
, false);
806 vhost_log_put(dev
, false);
808 vhost_dev_log_resize(dev
, vhost_get_log_size(dev
));
809 r
= vhost_dev_set_log(dev
, true);
814 dev
->log_enabled
= enable
;
818 static void vhost_log_global_start(MemoryListener
*listener
)
822 r
= vhost_migration_log(listener
, true);
828 static void vhost_log_global_stop(MemoryListener
*listener
)
832 r
= vhost_migration_log(listener
, false);
838 static void vhost_log_start(MemoryListener
*listener
,
839 MemoryRegionSection
*section
,
842 /* FIXME: implement */
845 static void vhost_log_stop(MemoryListener
*listener
,
846 MemoryRegionSection
*section
,
849 /* FIXME: implement */
852 /* The vhost driver natively knows how to handle the vrings of non
853 * cross-endian legacy devices and modern devices. Only legacy devices
854 * exposed to a bi-endian guest may require the vhost driver to use a
855 * specific endianness.
857 static inline bool vhost_needs_vring_endian(VirtIODevice
*vdev
)
859 if (virtio_vdev_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
862 #ifdef HOST_WORDS_BIGENDIAN
863 return vdev
->device_endian
== VIRTIO_DEVICE_ENDIAN_LITTLE
;
865 return vdev
->device_endian
== VIRTIO_DEVICE_ENDIAN_BIG
;
869 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev
*dev
,
873 struct vhost_vring_state s
= {
874 .index
= vhost_vq_index
,
878 if (!dev
->vhost_ops
->vhost_set_vring_endian(dev
, &s
)) {
882 VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
883 if (errno
== ENOTTY
) {
884 error_report("vhost does not support cross-endian");
891 static int vhost_memory_region_lookup(struct vhost_dev
*hdev
,
892 uint64_t gpa
, uint64_t *uaddr
,
897 for (i
= 0; i
< hdev
->mem
->nregions
; i
++) {
898 struct vhost_memory_region
*reg
= hdev
->mem
->regions
+ i
;
900 if (gpa
>= reg
->guest_phys_addr
&&
901 reg
->guest_phys_addr
+ reg
->memory_size
> gpa
) {
902 *uaddr
= reg
->userspace_addr
+ gpa
- reg
->guest_phys_addr
;
903 *len
= reg
->guest_phys_addr
+ reg
->memory_size
- gpa
;
911 void vhost_device_iotlb_miss(struct vhost_dev
*dev
, uint64_t iova
, int write
)
918 iotlb
= address_space_get_iotlb_entry(dev
->vdev
->dma_as
,
920 if (iotlb
.target_as
!= NULL
) {
921 if (vhost_memory_region_lookup(dev
, iotlb
.translated_addr
,
923 error_report("Fail to lookup the translated address "
924 "%"PRIx64
, iotlb
.translated_addr
);
928 len
= MIN(iotlb
.addr_mask
+ 1, len
);
929 iova
= iova
& ~iotlb
.addr_mask
;
931 if (dev
->vhost_ops
->vhost_update_device_iotlb(dev
, iova
, uaddr
,
933 error_report("Fail to update device iotlb");
941 static int vhost_virtqueue_start(struct vhost_dev
*dev
,
942 struct VirtIODevice
*vdev
,
943 struct vhost_virtqueue
*vq
,
946 BusState
*qbus
= BUS(qdev_get_parent_bus(DEVICE(vdev
)));
947 VirtioBusState
*vbus
= VIRTIO_BUS(qbus
);
948 VirtioBusClass
*k
= VIRTIO_BUS_GET_CLASS(vbus
);
951 int vhost_vq_index
= dev
->vhost_ops
->vhost_get_vq_index(dev
, idx
);
952 struct vhost_vring_file file
= {
953 .index
= vhost_vq_index
955 struct vhost_vring_state state
= {
956 .index
= vhost_vq_index
958 struct VirtQueue
*vvq
= virtio_get_queue(vdev
, idx
);
961 vq
->num
= state
.num
= virtio_queue_get_num(vdev
, idx
);
962 r
= dev
->vhost_ops
->vhost_set_vring_num(dev
, &state
);
964 VHOST_OPS_DEBUG("vhost_set_vring_num failed");
968 state
.num
= virtio_queue_get_last_avail_idx(vdev
, idx
);
969 r
= dev
->vhost_ops
->vhost_set_vring_base(dev
, &state
);
971 VHOST_OPS_DEBUG("vhost_set_vring_base failed");
975 if (vhost_needs_vring_endian(vdev
)) {
976 r
= vhost_virtqueue_set_vring_endian_legacy(dev
,
977 virtio_is_big_endian(vdev
),
984 vq
->desc_size
= s
= l
= virtio_queue_get_desc_size(vdev
, idx
);
985 vq
->desc_phys
= a
= virtio_queue_get_desc_addr(vdev
, idx
);
986 vq
->desc
= vhost_memory_map(dev
, a
, &l
, 0);
987 if (!vq
->desc
|| l
!= s
) {
989 goto fail_alloc_desc
;
991 vq
->avail_size
= s
= l
= virtio_queue_get_avail_size(vdev
, idx
);
992 vq
->avail_phys
= a
= virtio_queue_get_avail_addr(vdev
, idx
);
993 vq
->avail
= vhost_memory_map(dev
, a
, &l
, 0);
994 if (!vq
->avail
|| l
!= s
) {
996 goto fail_alloc_avail
;
998 vq
->used_size
= s
= l
= virtio_queue_get_used_size(vdev
, idx
);
999 vq
->used_phys
= a
= virtio_queue_get_used_addr(vdev
, idx
);
1000 vq
->used
= vhost_memory_map(dev
, a
, &l
, 1);
1001 if (!vq
->used
|| l
!= s
) {
1003 goto fail_alloc_used
;
1006 r
= vhost_virtqueue_set_addr(dev
, vq
, vhost_vq_index
, dev
->log_enabled
);
1012 file
.fd
= event_notifier_get_fd(virtio_queue_get_host_notifier(vvq
));
1013 r
= dev
->vhost_ops
->vhost_set_vring_kick(dev
, &file
);
1015 VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1020 /* Clear and discard previous events if any. */
1021 event_notifier_test_and_clear(&vq
->masked_notifier
);
1023 /* Init vring in unmasked state, unless guest_notifier_mask
1026 if (!vdev
->use_guest_notifier_mask
) {
1027 /* TODO: check and handle errors. */
1028 vhost_virtqueue_mask(dev
, vdev
, idx
, false);
1031 if (k
->query_guest_notifiers
&&
1032 k
->query_guest_notifiers(qbus
->parent
) &&
1033 virtio_queue_vector(vdev
, idx
) == VIRTIO_NO_VECTOR
) {
1035 r
= dev
->vhost_ops
->vhost_set_vring_call(dev
, &file
);
1046 vhost_memory_unmap(dev
, vq
->used
, virtio_queue_get_used_size(vdev
, idx
),
1049 vhost_memory_unmap(dev
, vq
->avail
, virtio_queue_get_avail_size(vdev
, idx
),
1052 vhost_memory_unmap(dev
, vq
->desc
, virtio_queue_get_desc_size(vdev
, idx
),
1058 static void vhost_virtqueue_stop(struct vhost_dev
*dev
,
1059 struct VirtIODevice
*vdev
,
1060 struct vhost_virtqueue
*vq
,
1063 int vhost_vq_index
= dev
->vhost_ops
->vhost_get_vq_index(dev
, idx
);
1064 struct vhost_vring_state state
= {
1065 .index
= vhost_vq_index
,
1069 r
= dev
->vhost_ops
->vhost_get_vring_base(dev
, &state
);
1071 VHOST_OPS_DEBUG("vhost VQ %d ring restore failed: %d", idx
, r
);
1073 virtio_queue_set_last_avail_idx(vdev
, idx
, state
.num
);
1075 virtio_queue_invalidate_signalled_used(vdev
, idx
);
1076 virtio_queue_update_used_idx(vdev
, idx
);
1078 /* In the cross-endian case, we need to reset the vring endianness to
1079 * native as legacy devices expect so by default.
1081 if (vhost_needs_vring_endian(vdev
)) {
1082 vhost_virtqueue_set_vring_endian_legacy(dev
,
1083 !virtio_is_big_endian(vdev
),
1087 vhost_memory_unmap(dev
, vq
->used
, virtio_queue_get_used_size(vdev
, idx
),
1088 1, virtio_queue_get_used_size(vdev
, idx
));
1089 vhost_memory_unmap(dev
, vq
->avail
, virtio_queue_get_avail_size(vdev
, idx
),
1090 0, virtio_queue_get_avail_size(vdev
, idx
));
1091 vhost_memory_unmap(dev
, vq
->desc
, virtio_queue_get_desc_size(vdev
, idx
),
1092 0, virtio_queue_get_desc_size(vdev
, idx
));
1095 static void vhost_eventfd_add(MemoryListener
*listener
,
1096 MemoryRegionSection
*section
,
1097 bool match_data
, uint64_t data
, EventNotifier
*e
)
1101 static void vhost_eventfd_del(MemoryListener
*listener
,
1102 MemoryRegionSection
*section
,
1103 bool match_data
, uint64_t data
, EventNotifier
*e
)
1107 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev
*dev
,
1108 int n
, uint32_t timeout
)
1110 int vhost_vq_index
= dev
->vhost_ops
->vhost_get_vq_index(dev
, n
);
1111 struct vhost_vring_state state
= {
1112 .index
= vhost_vq_index
,
1117 if (!dev
->vhost_ops
->vhost_set_vring_busyloop_timeout
) {
1121 r
= dev
->vhost_ops
->vhost_set_vring_busyloop_timeout(dev
, &state
);
1123 VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1130 static int vhost_virtqueue_init(struct vhost_dev
*dev
,
1131 struct vhost_virtqueue
*vq
, int n
)
1133 int vhost_vq_index
= dev
->vhost_ops
->vhost_get_vq_index(dev
, n
);
1134 struct vhost_vring_file file
= {
1135 .index
= vhost_vq_index
,
1137 int r
= event_notifier_init(&vq
->masked_notifier
, 0);
1142 file
.fd
= event_notifier_get_fd(&vq
->masked_notifier
);
1143 r
= dev
->vhost_ops
->vhost_set_vring_call(dev
, &file
);
1145 VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1154 event_notifier_cleanup(&vq
->masked_notifier
);
1158 static void vhost_virtqueue_cleanup(struct vhost_virtqueue
*vq
)
1160 event_notifier_cleanup(&vq
->masked_notifier
);
1163 static void vhost_iommu_unmap_notify(IOMMUNotifier
*n
, IOMMUTLBEntry
*iotlb
)
1165 struct vhost_dev
*hdev
= container_of(n
, struct vhost_dev
, n
);
1167 if (hdev
->vhost_ops
->vhost_invalidate_device_iotlb(hdev
,
1169 iotlb
->addr_mask
+ 1)) {
1170 error_report("Fail to invalidate device iotlb");
1174 int vhost_dev_init(struct vhost_dev
*hdev
, void *opaque
,
1175 VhostBackendType backend_type
, uint32_t busyloop_timeout
)
1178 int i
, r
, n_initialized_vqs
= 0;
1179 Error
*local_err
= NULL
;
1182 hdev
->migration_blocker
= NULL
;
1184 r
= vhost_set_backend_type(hdev
, backend_type
);
1187 r
= hdev
->vhost_ops
->vhost_backend_init(hdev
, opaque
);
1192 if (used_memslots
> hdev
->vhost_ops
->vhost_backend_memslots_limit(hdev
)) {
1193 error_report("vhost backend memory slots limit is less"
1194 " than current number of present memory slots");
1199 r
= hdev
->vhost_ops
->vhost_set_owner(hdev
);
1201 VHOST_OPS_DEBUG("vhost_set_owner failed");
1205 r
= hdev
->vhost_ops
->vhost_get_features(hdev
, &features
);
1207 VHOST_OPS_DEBUG("vhost_get_features failed");
1211 for (i
= 0; i
< hdev
->nvqs
; ++i
, ++n_initialized_vqs
) {
1212 r
= vhost_virtqueue_init(hdev
, hdev
->vqs
+ i
, hdev
->vq_index
+ i
);
1218 if (busyloop_timeout
) {
1219 for (i
= 0; i
< hdev
->nvqs
; ++i
) {
1220 r
= vhost_virtqueue_set_busyloop_timeout(hdev
, hdev
->vq_index
+ i
,
1228 hdev
->features
= features
;
1230 hdev
->memory_listener
= (MemoryListener
) {
1231 .begin
= vhost_begin
,
1232 .commit
= vhost_commit
,
1233 .region_add
= vhost_region_add
,
1234 .region_del
= vhost_region_del
,
1235 .region_nop
= vhost_region_nop
,
1236 .log_start
= vhost_log_start
,
1237 .log_stop
= vhost_log_stop
,
1238 .log_sync
= vhost_log_sync
,
1239 .log_global_start
= vhost_log_global_start
,
1240 .log_global_stop
= vhost_log_global_stop
,
1241 .eventfd_add
= vhost_eventfd_add
,
1242 .eventfd_del
= vhost_eventfd_del
,
1246 hdev
->n
.notify
= vhost_iommu_unmap_notify
;
1247 hdev
->n
.notifier_flags
= IOMMU_NOTIFIER_UNMAP
;
1249 if (hdev
->migration_blocker
== NULL
) {
1250 if (!(hdev
->features
& (0x1ULL
<< VHOST_F_LOG_ALL
))) {
1251 error_setg(&hdev
->migration_blocker
,
1252 "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1253 } else if (vhost_dev_log_is_shared(hdev
) && !qemu_memfd_check()) {
1254 error_setg(&hdev
->migration_blocker
,
1255 "Migration disabled: failed to allocate shared memory");
1259 if (hdev
->migration_blocker
!= NULL
) {
1260 r
= migrate_add_blocker(hdev
->migration_blocker
, &local_err
);
1262 error_report_err(local_err
);
1263 error_free(hdev
->migration_blocker
);
1268 hdev
->mem
= g_malloc0(offsetof(struct vhost_memory
, regions
));
1269 hdev
->n_mem_sections
= 0;
1270 hdev
->mem_sections
= NULL
;
1273 hdev
->log_enabled
= false;
1274 hdev
->started
= false;
1275 hdev
->memory_changed
= false;
1276 memory_listener_register(&hdev
->memory_listener
, &address_space_memory
);
1277 QLIST_INSERT_HEAD(&vhost_devices
, hdev
, entry
);
1282 vhost_virtqueue_set_busyloop_timeout(hdev
, hdev
->vq_index
+ i
, 0);
1285 hdev
->nvqs
= n_initialized_vqs
;
1286 vhost_dev_cleanup(hdev
);
1290 void vhost_dev_cleanup(struct vhost_dev
*hdev
)
1294 for (i
= 0; i
< hdev
->nvqs
; ++i
) {
1295 vhost_virtqueue_cleanup(hdev
->vqs
+ i
);
1298 /* those are only safe after successful init */
1299 memory_listener_unregister(&hdev
->memory_listener
);
1300 QLIST_REMOVE(hdev
, entry
);
1302 if (hdev
->migration_blocker
) {
1303 migrate_del_blocker(hdev
->migration_blocker
);
1304 error_free(hdev
->migration_blocker
);
1307 g_free(hdev
->mem_sections
);
1308 if (hdev
->vhost_ops
) {
1309 hdev
->vhost_ops
->vhost_backend_cleanup(hdev
);
1313 memset(hdev
, 0, sizeof(struct vhost_dev
));
1316 /* Stop processing guest IO notifications in qemu.
1317 * Start processing them in vhost in kernel.
1319 int vhost_dev_enable_notifiers(struct vhost_dev
*hdev
, VirtIODevice
*vdev
)
1321 BusState
*qbus
= BUS(qdev_get_parent_bus(DEVICE(vdev
)));
1324 /* We will pass the notifiers to the kernel, make sure that QEMU
1325 * doesn't interfere.
1327 r
= virtio_device_grab_ioeventfd(vdev
);
1329 error_report("binding does not support host notifiers");
1333 for (i
= 0; i
< hdev
->nvqs
; ++i
) {
1334 r
= virtio_bus_set_host_notifier(VIRTIO_BUS(qbus
), hdev
->vq_index
+ i
,
1337 error_report("vhost VQ %d notifier binding failed: %d", i
, -r
);
1345 e
= virtio_bus_set_host_notifier(VIRTIO_BUS(qbus
), hdev
->vq_index
+ i
,
1348 error_report("vhost VQ %d notifier cleanup error: %d", i
, -r
);
1352 virtio_device_release_ioeventfd(vdev
);
1357 /* Stop processing guest IO notifications in vhost.
1358 * Start processing them in qemu.
1359 * This might actually run the qemu handlers right away,
1360 * so virtio in qemu must be completely setup when this is called.
1362 void vhost_dev_disable_notifiers(struct vhost_dev
*hdev
, VirtIODevice
*vdev
)
1364 BusState
*qbus
= BUS(qdev_get_parent_bus(DEVICE(vdev
)));
1367 for (i
= 0; i
< hdev
->nvqs
; ++i
) {
1368 r
= virtio_bus_set_host_notifier(VIRTIO_BUS(qbus
), hdev
->vq_index
+ i
,
1371 error_report("vhost VQ %d notifier cleanup failed: %d", i
, -r
);
1375 virtio_device_release_ioeventfd(vdev
);
1378 /* Test and clear event pending status.
1379 * Should be called after unmask to avoid losing events.
1381 bool vhost_virtqueue_pending(struct vhost_dev
*hdev
, int n
)
1383 struct vhost_virtqueue
*vq
= hdev
->vqs
+ n
- hdev
->vq_index
;
1384 assert(n
>= hdev
->vq_index
&& n
< hdev
->vq_index
+ hdev
->nvqs
);
1385 return event_notifier_test_and_clear(&vq
->masked_notifier
);
1388 /* Mask/unmask events from this vq. */
1389 void vhost_virtqueue_mask(struct vhost_dev
*hdev
, VirtIODevice
*vdev
, int n
,
1392 struct VirtQueue
*vvq
= virtio_get_queue(vdev
, n
);
1393 int r
, index
= n
- hdev
->vq_index
;
1394 struct vhost_vring_file file
;
1396 /* should only be called after backend is connected */
1397 assert(hdev
->vhost_ops
);
1400 assert(vdev
->use_guest_notifier_mask
);
1401 file
.fd
= event_notifier_get_fd(&hdev
->vqs
[index
].masked_notifier
);
1403 file
.fd
= event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq
));
1406 file
.index
= hdev
->vhost_ops
->vhost_get_vq_index(hdev
, n
);
1407 r
= hdev
->vhost_ops
->vhost_set_vring_call(hdev
, &file
);
1409 VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1413 uint64_t vhost_get_features(struct vhost_dev
*hdev
, const int *feature_bits
,
1416 const int *bit
= feature_bits
;
1417 while (*bit
!= VHOST_INVALID_FEATURE_BIT
) {
1418 uint64_t bit_mask
= (1ULL << *bit
);
1419 if (!(hdev
->features
& bit_mask
)) {
1420 features
&= ~bit_mask
;
1427 void vhost_ack_features(struct vhost_dev
*hdev
, const int *feature_bits
,
1430 const int *bit
= feature_bits
;
1431 while (*bit
!= VHOST_INVALID_FEATURE_BIT
) {
1432 uint64_t bit_mask
= (1ULL << *bit
);
1433 if (features
& bit_mask
) {
1434 hdev
->acked_features
|= bit_mask
;
1440 /* Host notifiers must be enabled at this point. */
1441 int vhost_dev_start(struct vhost_dev
*hdev
, VirtIODevice
*vdev
)
1445 /* should only be called after backend is connected */
1446 assert(hdev
->vhost_ops
);
1448 hdev
->started
= true;
1451 r
= vhost_dev_set_features(hdev
, hdev
->log_enabled
);
1456 if (vhost_dev_has_iommu(hdev
)) {
1457 memory_region_register_iommu_notifier(vdev
->dma_as
->root
,
1461 r
= hdev
->vhost_ops
->vhost_set_mem_table(hdev
, hdev
->mem
);
1463 VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1467 for (i
= 0; i
< hdev
->nvqs
; ++i
) {
1468 r
= vhost_virtqueue_start(hdev
,
1471 hdev
->vq_index
+ i
);
1477 if (hdev
->log_enabled
) {
1480 hdev
->log_size
= vhost_get_log_size(hdev
);
1481 hdev
->log
= vhost_log_get(hdev
->log_size
,
1482 vhost_dev_log_is_shared(hdev
));
1483 log_base
= (uintptr_t)hdev
->log
->log
;
1484 r
= hdev
->vhost_ops
->vhost_set_log_base(hdev
,
1485 hdev
->log_size
? log_base
: 0,
1488 VHOST_OPS_DEBUG("vhost_set_log_base failed");
1494 if (vhost_dev_has_iommu(hdev
)) {
1495 hdev
->vhost_ops
->vhost_set_iotlb_callback(hdev
, true);
1497 /* Update used ring information for IOTLB to work correctly,
1498 * vhost-kernel code requires for this.*/
1499 for (i
= 0; i
< hdev
->nvqs
; ++i
) {
1500 struct vhost_virtqueue
*vq
= hdev
->vqs
+ i
;
1501 vhost_device_iotlb_miss(hdev
, vq
->used_phys
, true);
1506 vhost_log_put(hdev
, false);
1509 vhost_virtqueue_stop(hdev
,
1512 hdev
->vq_index
+ i
);
1519 hdev
->started
= false;
1523 /* Host notifiers must be enabled at this point. */
1524 void vhost_dev_stop(struct vhost_dev
*hdev
, VirtIODevice
*vdev
)
1528 /* should only be called after backend is connected */
1529 assert(hdev
->vhost_ops
);
1531 for (i
= 0; i
< hdev
->nvqs
; ++i
) {
1532 vhost_virtqueue_stop(hdev
,
1535 hdev
->vq_index
+ i
);
1538 if (vhost_dev_has_iommu(hdev
)) {
1539 hdev
->vhost_ops
->vhost_set_iotlb_callback(hdev
, false);
1540 memory_region_unregister_iommu_notifier(vdev
->dma_as
->root
,
1543 vhost_log_put(hdev
, true);
1544 hdev
->started
= false;
1548 int vhost_net_set_backend(struct vhost_dev
*hdev
,
1549 struct vhost_vring_file
*file
)
1551 if (hdev
->vhost_ops
->vhost_net_set_backend
) {
1552 return hdev
->vhost_ops
->vhost_net_set_backend(hdev
, file
);