2 * QEMU live block migration
4 * Copyright IBM, Corp. 2009
7 * Liran Schour <lirans@il.ibm.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 "qemu-common.h"
19 #include "block/block.h"
20 #include "qemu/error-report.h"
21 #include "qemu/main-loop.h"
23 #include "qemu/cutils.h"
24 #include "qemu/queue.h"
25 #include "qemu/timer.h"
27 #include "migration/misc.h"
28 #include "migration.h"
29 #include "migration/register.h"
30 #include "sysemu/blockdev.h"
31 #include "qemu-file.h"
32 #include "migration/vmstate.h"
33 #include "sysemu/block-backend.h"
35 #define BLOCK_SIZE (1 << 20)
36 #define BDRV_SECTORS_PER_DIRTY_CHUNK (BLOCK_SIZE >> BDRV_SECTOR_BITS)
38 #define BLK_MIG_FLAG_DEVICE_BLOCK 0x01
39 #define BLK_MIG_FLAG_EOS 0x02
40 #define BLK_MIG_FLAG_PROGRESS 0x04
41 #define BLK_MIG_FLAG_ZERO_BLOCK 0x08
43 #define MAX_IS_ALLOCATED_SEARCH 65536
45 #define MAX_INFLIGHT_IO 512
47 //#define DEBUG_BLK_MIGRATION
49 #ifdef DEBUG_BLK_MIGRATION
50 #define DPRINTF(fmt, ...) \
51 do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
53 #define DPRINTF(fmt, ...) \
57 typedef struct BlkMigDevState
{
58 /* Written during setup phase. Can be read without a lock. */
62 int64_t total_sectors
;
63 QSIMPLEQ_ENTRY(BlkMigDevState
) entry
;
66 /* Only used by migration thread. Does not need a lock. */
71 /* Data in the aio_bitmap is protected by block migration lock.
72 * Allocation and free happen during setup and cleanup respectively.
74 unsigned long *aio_bitmap
;
76 /* Protected by block migration lock. */
77 int64_t completed_sectors
;
79 /* During migration this is protected by iothread lock / AioContext.
80 * Allocation and free happen during setup and cleanup respectively.
82 BdrvDirtyBitmap
*dirty_bitmap
;
85 typedef struct BlkMigBlock
{
86 /* Only used by migration thread. */
95 /* Protected by block migration lock. */
97 QSIMPLEQ_ENTRY(BlkMigBlock
) entry
;
100 typedef struct BlkMigState
{
101 QSIMPLEQ_HEAD(bmds_list
, BlkMigDevState
) bmds_list
;
102 int64_t total_sector_sum
;
105 /* Protected by lock. */
106 QSIMPLEQ_HEAD(blk_list
, BlkMigBlock
) blk_list
;
110 /* Only used by migration thread. Does not need a lock. */
115 /* Lock must be taken _inside_ the iothread lock and any AioContexts. */
119 static BlkMigState block_mig_state
;
121 static void blk_mig_lock(void)
123 qemu_mutex_lock(&block_mig_state
.lock
);
126 static void blk_mig_unlock(void)
128 qemu_mutex_unlock(&block_mig_state
.lock
);
131 /* Must run outside of the iothread lock during the bulk phase,
132 * or the VM will stall.
135 static void blk_send(QEMUFile
*f
, BlkMigBlock
* blk
)
138 uint64_t flags
= BLK_MIG_FLAG_DEVICE_BLOCK
;
140 if (block_mig_state
.zero_blocks
&&
141 buffer_is_zero(blk
->buf
, BLOCK_SIZE
)) {
142 flags
|= BLK_MIG_FLAG_ZERO_BLOCK
;
145 /* sector number and flags */
146 qemu_put_be64(f
, (blk
->sector
<< BDRV_SECTOR_BITS
)
150 len
= strlen(blk
->bmds
->blk_name
);
151 qemu_put_byte(f
, len
);
152 qemu_put_buffer(f
, (uint8_t *) blk
->bmds
->blk_name
, len
);
154 /* if a block is zero we need to flush here since the network
155 * bandwidth is now a lot higher than the storage device bandwidth.
156 * thus if we queue zero blocks we slow down the migration */
157 if (flags
& BLK_MIG_FLAG_ZERO_BLOCK
) {
162 qemu_put_buffer(f
, blk
->buf
, BLOCK_SIZE
);
165 int blk_mig_active(void)
167 return !QSIMPLEQ_EMPTY(&block_mig_state
.bmds_list
);
170 uint64_t blk_mig_bytes_transferred(void)
172 BlkMigDevState
*bmds
;
176 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
177 sum
+= bmds
->completed_sectors
;
180 return sum
<< BDRV_SECTOR_BITS
;
183 uint64_t blk_mig_bytes_remaining(void)
185 return blk_mig_bytes_total() - blk_mig_bytes_transferred();
188 uint64_t blk_mig_bytes_total(void)
190 BlkMigDevState
*bmds
;
193 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
194 sum
+= bmds
->total_sectors
;
196 return sum
<< BDRV_SECTOR_BITS
;
200 /* Called with migration lock held. */
202 static int bmds_aio_inflight(BlkMigDevState
*bmds
, int64_t sector
)
204 int64_t chunk
= sector
/ (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK
;
206 if (sector
< blk_nb_sectors(bmds
->blk
)) {
207 return !!(bmds
->aio_bitmap
[chunk
/ (sizeof(unsigned long) * 8)] &
208 (1UL << (chunk
% (sizeof(unsigned long) * 8))));
214 /* Called with migration lock held. */
216 static void bmds_set_aio_inflight(BlkMigDevState
*bmds
, int64_t sector_num
,
217 int nb_sectors
, int set
)
220 unsigned long val
, idx
, bit
;
222 start
= sector_num
/ BDRV_SECTORS_PER_DIRTY_CHUNK
;
223 end
= (sector_num
+ nb_sectors
- 1) / BDRV_SECTORS_PER_DIRTY_CHUNK
;
225 for (; start
<= end
; start
++) {
226 idx
= start
/ (sizeof(unsigned long) * 8);
227 bit
= start
% (sizeof(unsigned long) * 8);
228 val
= bmds
->aio_bitmap
[idx
];
232 val
&= ~(1UL << bit
);
234 bmds
->aio_bitmap
[idx
] = val
;
238 static void alloc_aio_bitmap(BlkMigDevState
*bmds
)
240 BlockBackend
*bb
= bmds
->blk
;
243 bitmap_size
= blk_nb_sectors(bb
) + BDRV_SECTORS_PER_DIRTY_CHUNK
* 8 - 1;
244 bitmap_size
/= BDRV_SECTORS_PER_DIRTY_CHUNK
* 8;
246 bmds
->aio_bitmap
= g_malloc0(bitmap_size
);
249 /* Never hold migration lock when yielding to the main loop! */
251 static void blk_mig_read_cb(void *opaque
, int ret
)
253 BlkMigBlock
*blk
= opaque
;
258 QSIMPLEQ_INSERT_TAIL(&block_mig_state
.blk_list
, blk
, entry
);
259 bmds_set_aio_inflight(blk
->bmds
, blk
->sector
, blk
->nr_sectors
, 0);
261 block_mig_state
.submitted
--;
262 block_mig_state
.read_done
++;
263 assert(block_mig_state
.submitted
>= 0);
267 /* Called with no lock taken. */
269 static int mig_save_device_bulk(QEMUFile
*f
, BlkMigDevState
*bmds
)
271 int64_t total_sectors
= bmds
->total_sectors
;
272 int64_t cur_sector
= bmds
->cur_sector
;
273 BlockBackend
*bb
= bmds
->blk
;
277 if (bmds
->shared_base
) {
278 qemu_mutex_lock_iothread();
279 aio_context_acquire(blk_get_aio_context(bb
));
280 /* Skip unallocated sectors; intentionally treats failure as
281 * an allocated sector */
282 while (cur_sector
< total_sectors
&&
283 !bdrv_is_allocated(blk_bs(bb
), cur_sector
,
284 MAX_IS_ALLOCATED_SEARCH
, &nr_sectors
)) {
285 cur_sector
+= nr_sectors
;
287 aio_context_release(blk_get_aio_context(bb
));
288 qemu_mutex_unlock_iothread();
291 if (cur_sector
>= total_sectors
) {
292 bmds
->cur_sector
= bmds
->completed_sectors
= total_sectors
;
296 bmds
->completed_sectors
= cur_sector
;
298 cur_sector
&= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK
- 1);
300 /* we are going to transfer a full block even if it is not allocated */
301 nr_sectors
= BDRV_SECTORS_PER_DIRTY_CHUNK
;
303 if (total_sectors
- cur_sector
< BDRV_SECTORS_PER_DIRTY_CHUNK
) {
304 nr_sectors
= total_sectors
- cur_sector
;
307 blk
= g_new(BlkMigBlock
, 1);
308 blk
->buf
= g_malloc(BLOCK_SIZE
);
310 blk
->sector
= cur_sector
;
311 blk
->nr_sectors
= nr_sectors
;
313 blk
->iov
.iov_base
= blk
->buf
;
314 blk
->iov
.iov_len
= nr_sectors
* BDRV_SECTOR_SIZE
;
315 qemu_iovec_init_external(&blk
->qiov
, &blk
->iov
, 1);
318 block_mig_state
.submitted
++;
321 /* We do not know if bs is under the main thread (and thus does
322 * not acquire the AioContext when doing AIO) or rather under
323 * dataplane. Thus acquire both the iothread mutex and the
326 * This is ugly and will disappear when we make bdrv_* thread-safe,
327 * without the need to acquire the AioContext.
329 qemu_mutex_lock_iothread();
330 aio_context_acquire(blk_get_aio_context(bmds
->blk
));
331 blk
->aiocb
= blk_aio_preadv(bb
, cur_sector
* BDRV_SECTOR_SIZE
, &blk
->qiov
,
332 0, blk_mig_read_cb
, blk
);
334 bdrv_reset_dirty_bitmap(bmds
->dirty_bitmap
, cur_sector
, nr_sectors
);
335 aio_context_release(blk_get_aio_context(bmds
->blk
));
336 qemu_mutex_unlock_iothread();
338 bmds
->cur_sector
= cur_sector
+ nr_sectors
;
339 return (bmds
->cur_sector
>= total_sectors
);
342 /* Called with iothread lock taken. */
344 static int set_dirty_tracking(void)
346 BlkMigDevState
*bmds
;
349 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
350 aio_context_acquire(blk_get_aio_context(bmds
->blk
));
351 bmds
->dirty_bitmap
= bdrv_create_dirty_bitmap(blk_bs(bmds
->blk
),
352 BLOCK_SIZE
, NULL
, NULL
);
353 aio_context_release(blk_get_aio_context(bmds
->blk
));
354 if (!bmds
->dirty_bitmap
) {
362 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
363 if (bmds
->dirty_bitmap
) {
364 aio_context_acquire(blk_get_aio_context(bmds
->blk
));
365 bdrv_release_dirty_bitmap(blk_bs(bmds
->blk
), bmds
->dirty_bitmap
);
366 aio_context_release(blk_get_aio_context(bmds
->blk
));
372 /* Called with iothread lock taken. */
374 static void unset_dirty_tracking(void)
376 BlkMigDevState
*bmds
;
378 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
379 aio_context_acquire(blk_get_aio_context(bmds
->blk
));
380 bdrv_release_dirty_bitmap(blk_bs(bmds
->blk
), bmds
->dirty_bitmap
);
381 aio_context_release(blk_get_aio_context(bmds
->blk
));
385 static int init_blk_migration(QEMUFile
*f
)
387 BlockDriverState
*bs
;
388 BlkMigDevState
*bmds
;
393 BlkMigDevState
*bmds
;
394 BlockDriverState
*bs
;
396 Error
*local_err
= NULL
;
399 block_mig_state
.submitted
= 0;
400 block_mig_state
.read_done
= 0;
401 block_mig_state
.transferred
= 0;
402 block_mig_state
.total_sector_sum
= 0;
403 block_mig_state
.prev_progress
= -1;
404 block_mig_state
.bulk_completed
= 0;
405 block_mig_state
.zero_blocks
= migrate_zero_blocks();
407 for (bs
= bdrv_first(&it
); bs
; bs
= bdrv_next(&it
)) {
410 bmds_bs
= g_malloc0(num_bs
* sizeof(*bmds_bs
));
412 for (i
= 0, bs
= bdrv_first(&it
); bs
; bs
= bdrv_next(&it
), i
++) {
413 if (bdrv_is_read_only(bs
)) {
417 sectors
= bdrv_nb_sectors(bs
);
423 bmds
= g_new0(BlkMigDevState
, 1);
424 bmds
->blk
= blk_new(BLK_PERM_CONSISTENT_READ
, BLK_PERM_ALL
);
425 bmds
->blk_name
= g_strdup(bdrv_get_device_name(bs
));
426 bmds
->bulk_completed
= 0;
427 bmds
->total_sectors
= sectors
;
428 bmds
->completed_sectors
= 0;
429 bmds
->shared_base
= migrate_use_block_incremental();
432 bmds_bs
[i
].bmds
= bmds
;
435 block_mig_state
.total_sector_sum
+= sectors
;
437 if (bmds
->shared_base
) {
438 DPRINTF("Start migration for %s with shared base image\n",
439 bdrv_get_device_name(bs
));
441 DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs
));
444 QSIMPLEQ_INSERT_TAIL(&block_mig_state
.bmds_list
, bmds
, entry
);
447 /* Can only insert new BDSes now because doing so while iterating block
448 * devices may end up in a deadlock (iterating the new BDSes, too). */
449 for (i
= 0; i
< num_bs
; i
++) {
450 BlkMigDevState
*bmds
= bmds_bs
[i
].bmds
;
451 BlockDriverState
*bs
= bmds_bs
[i
].bs
;
454 ret
= blk_insert_bs(bmds
->blk
, bs
, &local_err
);
456 error_report_err(local_err
);
460 alloc_aio_bitmap(bmds
);
461 error_setg(&bmds
->blocker
, "block device is in use by migration");
462 bdrv_op_block_all(bs
, bmds
->blocker
);
472 /* Called with no lock taken. */
474 static int blk_mig_save_bulked_block(QEMUFile
*f
)
476 int64_t completed_sector_sum
= 0;
477 BlkMigDevState
*bmds
;
481 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
482 if (bmds
->bulk_completed
== 0) {
483 if (mig_save_device_bulk(f
, bmds
) == 1) {
484 /* completed bulk section for this device */
485 bmds
->bulk_completed
= 1;
487 completed_sector_sum
+= bmds
->completed_sectors
;
491 completed_sector_sum
+= bmds
->completed_sectors
;
495 if (block_mig_state
.total_sector_sum
!= 0) {
496 progress
= completed_sector_sum
* 100 /
497 block_mig_state
.total_sector_sum
;
501 if (progress
!= block_mig_state
.prev_progress
) {
502 block_mig_state
.prev_progress
= progress
;
503 qemu_put_be64(f
, (progress
<< BDRV_SECTOR_BITS
)
504 | BLK_MIG_FLAG_PROGRESS
);
505 DPRINTF("Completed %d %%\r", progress
);
511 static void blk_mig_reset_dirty_cursor(void)
513 BlkMigDevState
*bmds
;
515 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
520 /* Called with iothread lock and AioContext taken. */
522 static int mig_save_device_dirty(QEMUFile
*f
, BlkMigDevState
*bmds
,
526 BlockDriverState
*bs
= blk_bs(bmds
->blk
);
527 int64_t total_sectors
= bmds
->total_sectors
;
532 for (sector
= bmds
->cur_dirty
; sector
< bmds
->total_sectors
;) {
534 if (bmds_aio_inflight(bmds
, sector
)) {
536 blk_drain(bmds
->blk
);
540 if (bdrv_get_dirty(bs
, bmds
->dirty_bitmap
, sector
)) {
542 if (total_sectors
- sector
< BDRV_SECTORS_PER_DIRTY_CHUNK
) {
543 nr_sectors
= total_sectors
- sector
;
545 nr_sectors
= BDRV_SECTORS_PER_DIRTY_CHUNK
;
547 blk
= g_new(BlkMigBlock
, 1);
548 blk
->buf
= g_malloc(BLOCK_SIZE
);
550 blk
->sector
= sector
;
551 blk
->nr_sectors
= nr_sectors
;
554 blk
->iov
.iov_base
= blk
->buf
;
555 blk
->iov
.iov_len
= nr_sectors
* BDRV_SECTOR_SIZE
;
556 qemu_iovec_init_external(&blk
->qiov
, &blk
->iov
, 1);
558 blk
->aiocb
= blk_aio_preadv(bmds
->blk
,
559 sector
* BDRV_SECTOR_SIZE
,
560 &blk
->qiov
, 0, blk_mig_read_cb
,
564 block_mig_state
.submitted
++;
565 bmds_set_aio_inflight(bmds
, sector
, nr_sectors
, 1);
568 ret
= blk_pread(bmds
->blk
, sector
* BDRV_SECTOR_SIZE
, blk
->buf
,
569 nr_sectors
* BDRV_SECTOR_SIZE
);
579 bdrv_reset_dirty_bitmap(bmds
->dirty_bitmap
, sector
, nr_sectors
);
580 sector
+= nr_sectors
;
581 bmds
->cur_dirty
= sector
;
585 sector
+= BDRV_SECTORS_PER_DIRTY_CHUNK
;
586 bmds
->cur_dirty
= sector
;
589 return (bmds
->cur_dirty
>= bmds
->total_sectors
);
592 DPRINTF("Error reading sector %" PRId64
"\n", sector
);
598 /* Called with iothread lock taken.
601 * 0: too much data for max_downtime
602 * 1: few enough data for max_downtime
604 static int blk_mig_save_dirty_block(QEMUFile
*f
, int is_async
)
606 BlkMigDevState
*bmds
;
609 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
610 aio_context_acquire(blk_get_aio_context(bmds
->blk
));
611 ret
= mig_save_device_dirty(f
, bmds
, is_async
);
612 aio_context_release(blk_get_aio_context(bmds
->blk
));
621 /* Called with no locks taken. */
623 static int flush_blks(QEMUFile
*f
)
628 DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
629 __FUNCTION__
, block_mig_state
.submitted
, block_mig_state
.read_done
,
630 block_mig_state
.transferred
);
633 while ((blk
= QSIMPLEQ_FIRST(&block_mig_state
.blk_list
)) != NULL
) {
634 if (qemu_file_rate_limit(f
)) {
642 QSIMPLEQ_REMOVE_HEAD(&block_mig_state
.blk_list
, entry
);
650 block_mig_state
.read_done
--;
651 block_mig_state
.transferred
++;
652 assert(block_mig_state
.read_done
>= 0);
656 DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__
,
657 block_mig_state
.submitted
, block_mig_state
.read_done
,
658 block_mig_state
.transferred
);
662 /* Called with iothread lock taken. */
664 static int64_t get_remaining_dirty(void)
666 BlkMigDevState
*bmds
;
669 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
670 aio_context_acquire(blk_get_aio_context(bmds
->blk
));
671 dirty
+= bdrv_get_dirty_count(bmds
->dirty_bitmap
);
672 aio_context_release(blk_get_aio_context(bmds
->blk
));
675 return dirty
<< BDRV_SECTOR_BITS
;
680 /* Called with iothread lock taken. */
681 static void block_migration_cleanup_bmds(void)
683 BlkMigDevState
*bmds
;
686 unset_dirty_tracking();
688 while ((bmds
= QSIMPLEQ_FIRST(&block_mig_state
.bmds_list
)) != NULL
) {
689 QSIMPLEQ_REMOVE_HEAD(&block_mig_state
.bmds_list
, entry
);
690 bdrv_op_unblock_all(blk_bs(bmds
->blk
), bmds
->blocker
);
691 error_free(bmds
->blocker
);
693 /* Save ctx, because bmds->blk can disappear during blk_unref. */
694 ctx
= blk_get_aio_context(bmds
->blk
);
695 aio_context_acquire(ctx
);
696 blk_unref(bmds
->blk
);
697 aio_context_release(ctx
);
699 g_free(bmds
->blk_name
);
700 g_free(bmds
->aio_bitmap
);
705 /* Called with iothread lock taken. */
706 static void block_migration_cleanup(void *opaque
)
712 block_migration_cleanup_bmds();
715 while ((blk
= QSIMPLEQ_FIRST(&block_mig_state
.blk_list
)) != NULL
) {
716 QSIMPLEQ_REMOVE_HEAD(&block_mig_state
.blk_list
, entry
);
723 static int block_save_setup(QEMUFile
*f
, void *opaque
)
727 DPRINTF("Enter save live setup submitted %d transferred %d\n",
728 block_mig_state
.submitted
, block_mig_state
.transferred
);
730 qemu_mutex_lock_iothread();
731 ret
= init_blk_migration(f
);
733 qemu_mutex_unlock_iothread();
737 /* start track dirty blocks */
738 ret
= set_dirty_tracking();
740 qemu_mutex_unlock_iothread();
747 blk_mig_reset_dirty_cursor();
748 qemu_put_be64(f
, BLK_MIG_FLAG_EOS
);
753 static int block_save_iterate(QEMUFile
*f
, void *opaque
)
756 int64_t last_ftell
= qemu_ftell(f
);
759 DPRINTF("Enter save live iterate submitted %d transferred %d\n",
760 block_mig_state
.submitted
, block_mig_state
.transferred
);
767 blk_mig_reset_dirty_cursor();
769 /* control the rate of transfer */
771 while ((block_mig_state
.submitted
+
772 block_mig_state
.read_done
) * BLOCK_SIZE
<
773 qemu_file_get_rate_limit(f
) &&
774 (block_mig_state
.submitted
+
775 block_mig_state
.read_done
) <
778 if (block_mig_state
.bulk_completed
== 0) {
779 /* first finish the bulk phase */
780 if (blk_mig_save_bulked_block(f
) == 0) {
781 /* finished saving bulk on all devices */
782 block_mig_state
.bulk_completed
= 1;
786 /* Always called with iothread lock taken for
787 * simplicity, block_save_complete also calls it.
789 qemu_mutex_lock_iothread();
790 ret
= blk_mig_save_dirty_block(f
, 1);
791 qemu_mutex_unlock_iothread();
798 /* no more dirty blocks */
809 qemu_put_be64(f
, BLK_MIG_FLAG_EOS
);
810 delta_ftell
= qemu_ftell(f
) - last_ftell
;
811 if (delta_ftell
> 0) {
813 } else if (delta_ftell
< 0) {
820 /* Called with iothread lock taken. */
822 static int block_save_complete(QEMUFile
*f
, void *opaque
)
826 DPRINTF("Enter save live complete submitted %d transferred %d\n",
827 block_mig_state
.submitted
, block_mig_state
.transferred
);
834 blk_mig_reset_dirty_cursor();
836 /* we know for sure that save bulk is completed and
837 all async read completed */
839 assert(block_mig_state
.submitted
== 0);
843 ret
= blk_mig_save_dirty_block(f
, 0);
849 /* report completion */
850 qemu_put_be64(f
, (100 << BDRV_SECTOR_BITS
) | BLK_MIG_FLAG_PROGRESS
);
852 DPRINTF("Block migration completed\n");
854 qemu_put_be64(f
, BLK_MIG_FLAG_EOS
);
856 /* Make sure that our BlockBackends are gone, so that the block driver
857 * nodes can be inactivated. */
858 block_migration_cleanup_bmds();
863 static void block_save_pending(QEMUFile
*f
, void *opaque
, uint64_t max_size
,
864 uint64_t *non_postcopiable_pending
,
865 uint64_t *postcopiable_pending
)
867 /* Estimate pending number of bytes to send */
870 qemu_mutex_lock_iothread();
871 pending
= get_remaining_dirty();
872 qemu_mutex_unlock_iothread();
875 pending
+= block_mig_state
.submitted
* BLOCK_SIZE
+
876 block_mig_state
.read_done
* BLOCK_SIZE
;
879 /* Report at least one block pending during bulk phase */
880 if (pending
<= max_size
&& !block_mig_state
.bulk_completed
) {
881 pending
= max_size
+ BLOCK_SIZE
;
884 DPRINTF("Enter save live pending %" PRIu64
"\n", pending
);
885 /* We don't do postcopy */
886 *non_postcopiable_pending
+= pending
;
889 static int block_load(QEMUFile
*f
, void *opaque
, int version_id
)
891 static int banner_printed
;
893 char device_name
[256];
895 BlockBackend
*blk
, *blk_prev
= NULL
;;
896 Error
*local_err
= NULL
;
898 int64_t total_sectors
= 0;
902 int cluster_size
= BLOCK_SIZE
;
905 addr
= qemu_get_be64(f
);
907 flags
= addr
& ~BDRV_SECTOR_MASK
;
908 addr
>>= BDRV_SECTOR_BITS
;
910 if (flags
& BLK_MIG_FLAG_DEVICE_BLOCK
) {
911 /* get device name */
912 len
= qemu_get_byte(f
);
913 qemu_get_buffer(f
, (uint8_t *)device_name
, len
);
914 device_name
[len
] = '\0';
916 blk
= blk_by_name(device_name
);
918 fprintf(stderr
, "Error unknown block device %s\n",
923 if (blk
!= blk_prev
) {
925 total_sectors
= blk_nb_sectors(blk
);
926 if (total_sectors
<= 0) {
927 error_report("Error getting length of block device %s",
932 blk_invalidate_cache(blk
, &local_err
);
934 error_report_err(local_err
);
938 ret
= bdrv_get_info(blk_bs(blk
), &bdi
);
939 if (ret
== 0 && bdi
.cluster_size
> 0 &&
940 bdi
.cluster_size
<= BLOCK_SIZE
&&
941 BLOCK_SIZE
% bdi
.cluster_size
== 0) {
942 cluster_size
= bdi
.cluster_size
;
944 cluster_size
= BLOCK_SIZE
;
948 if (total_sectors
- addr
< BDRV_SECTORS_PER_DIRTY_CHUNK
) {
949 nr_sectors
= total_sectors
- addr
;
951 nr_sectors
= BDRV_SECTORS_PER_DIRTY_CHUNK
;
954 if (flags
& BLK_MIG_FLAG_ZERO_BLOCK
) {
955 ret
= blk_pwrite_zeroes(blk
, addr
* BDRV_SECTOR_SIZE
,
956 nr_sectors
* BDRV_SECTOR_SIZE
,
963 buf
= g_malloc(BLOCK_SIZE
);
964 qemu_get_buffer(f
, buf
, BLOCK_SIZE
);
965 for (i
= 0; i
< BLOCK_SIZE
/ cluster_size
; i
++) {
966 cur_addr
= addr
* BDRV_SECTOR_SIZE
+ i
* cluster_size
;
967 cur_buf
= buf
+ i
* cluster_size
;
969 if ((!block_mig_state
.zero_blocks
||
970 cluster_size
< BLOCK_SIZE
) &&
971 buffer_is_zero(cur_buf
, cluster_size
)) {
972 ret
= blk_pwrite_zeroes(blk
, cur_addr
,
976 ret
= blk_pwrite(blk
, cur_addr
, cur_buf
,
989 } else if (flags
& BLK_MIG_FLAG_PROGRESS
) {
990 if (!banner_printed
) {
991 printf("Receiving block device images\n");
994 printf("Completed %d %%%c", (int)addr
,
995 (addr
== 100) ? '\n' : '\r');
997 } else if (!(flags
& BLK_MIG_FLAG_EOS
)) {
998 fprintf(stderr
, "Unknown block migration flags: %#x\n", flags
);
1001 ret
= qemu_file_get_error(f
);
1005 } while (!(flags
& BLK_MIG_FLAG_EOS
));
1010 static bool block_is_active(void *opaque
)
1012 return migrate_use_block();
1015 static SaveVMHandlers savevm_block_handlers
= {
1016 .save_live_setup
= block_save_setup
,
1017 .save_live_iterate
= block_save_iterate
,
1018 .save_live_complete_precopy
= block_save_complete
,
1019 .save_live_pending
= block_save_pending
,
1020 .load_state
= block_load
,
1021 .cleanup
= block_migration_cleanup
,
1022 .is_active
= block_is_active
,
1025 void blk_mig_init(void)
1027 QSIMPLEQ_INIT(&block_mig_state
.bmds_list
);
1028 QSIMPLEQ_INIT(&block_mig_state
.blk_list
);
1029 qemu_mutex_init(&block_mig_state
.lock
);
1031 register_savevm_live(NULL
, "block", 0, 1, &savevm_block_handlers
,