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-common.h"
17 #include "block/block.h"
18 #include "qemu/error-report.h"
19 #include "qemu/main-loop.h"
21 #include "qemu/queue.h"
22 #include "qemu/timer.h"
23 #include "migration/block.h"
24 #include "migration/migration.h"
25 #include "sysemu/blockdev.h"
28 #define BLOCK_SIZE (1 << 20)
29 #define BDRV_SECTORS_PER_DIRTY_CHUNK (BLOCK_SIZE >> BDRV_SECTOR_BITS)
31 #define BLK_MIG_FLAG_DEVICE_BLOCK 0x01
32 #define BLK_MIG_FLAG_EOS 0x02
33 #define BLK_MIG_FLAG_PROGRESS 0x04
34 #define BLK_MIG_FLAG_ZERO_BLOCK 0x08
36 #define MAX_IS_ALLOCATED_SEARCH 65536
38 //#define DEBUG_BLK_MIGRATION
40 #ifdef DEBUG_BLK_MIGRATION
41 #define DPRINTF(fmt, ...) \
42 do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
44 #define DPRINTF(fmt, ...) \
48 typedef struct BlkMigDevState {
49 /* Written during setup phase. Can be read without a lock. */
52 int64_t total_sectors;
53 QSIMPLEQ_ENTRY(BlkMigDevState) entry;
55 /* Only used by migration thread. Does not need a lock. */
60 /* Protected by block migration lock. */
61 unsigned long *aio_bitmap;
62 int64_t completed_sectors;
63 BdrvDirtyBitmap *dirty_bitmap;
67 typedef struct BlkMigBlock {
68 /* Only used by migration thread. */
77 /* Protected by block migration lock. */
79 QSIMPLEQ_ENTRY(BlkMigBlock) entry;
82 typedef struct BlkMigState {
83 /* Written during setup phase. Can be read without a lock. */
86 QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
87 int64_t total_sector_sum;
90 /* Protected by lock. */
91 QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
95 /* Only used by migration thread. Does not need a lock. */
100 /* Lock must be taken _inside_ the iothread lock. */
104 static BlkMigState block_mig_state;
106 static void blk_mig_lock(void)
108 qemu_mutex_lock(&block_mig_state.lock);
111 static void blk_mig_unlock(void)
113 qemu_mutex_unlock(&block_mig_state.lock);
116 /* Must run outside of the iothread lock during the bulk phase,
117 * or the VM will stall.
120 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
123 uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
125 if (block_mig_state.zero_blocks &&
126 buffer_is_zero(blk->buf, BLOCK_SIZE)) {
127 flags |= BLK_MIG_FLAG_ZERO_BLOCK;
130 /* sector number and flags */
131 qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
135 len = strlen(bdrv_get_device_name(blk->bmds->bs));
136 qemu_put_byte(f, len);
137 qemu_put_buffer(f, (uint8_t *)bdrv_get_device_name(blk->bmds->bs), len);
139 /* if a block is zero we need to flush here since the network
140 * bandwidth is now a lot higher than the storage device bandwidth.
141 * thus if we queue zero blocks we slow down the migration */
142 if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
147 qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
150 int blk_mig_active(void)
152 return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
155 uint64_t blk_mig_bytes_transferred(void)
157 BlkMigDevState *bmds;
161 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
162 sum += bmds->completed_sectors;
165 return sum << BDRV_SECTOR_BITS;
168 uint64_t blk_mig_bytes_remaining(void)
170 return blk_mig_bytes_total() - blk_mig_bytes_transferred();
173 uint64_t blk_mig_bytes_total(void)
175 BlkMigDevState *bmds;
178 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
179 sum += bmds->total_sectors;
181 return sum << BDRV_SECTOR_BITS;
185 /* Called with migration lock held. */
187 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
189 int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
191 if (sector < bdrv_nb_sectors(bmds->bs)) {
192 return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
193 (1UL << (chunk % (sizeof(unsigned long) * 8))));
199 /* Called with migration lock held. */
201 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
202 int nb_sectors, int set)
205 unsigned long val, idx, bit;
207 start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
208 end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
210 for (; start <= end; start++) {
211 idx = start / (sizeof(unsigned long) * 8);
212 bit = start % (sizeof(unsigned long) * 8);
213 val = bmds->aio_bitmap[idx];
217 val &= ~(1UL << bit);
219 bmds->aio_bitmap[idx] = val;
223 static void alloc_aio_bitmap(BlkMigDevState *bmds)
225 BlockDriverState *bs = bmds->bs;
228 bitmap_size = bdrv_nb_sectors(bs) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
229 bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
231 bmds->aio_bitmap = g_malloc0(bitmap_size);
234 /* Never hold migration lock when yielding to the main loop! */
236 static void blk_mig_read_cb(void *opaque, int ret)
238 BlkMigBlock *blk = opaque;
243 QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
244 bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
246 block_mig_state.submitted--;
247 block_mig_state.read_done++;
248 assert(block_mig_state.submitted >= 0);
252 /* Called with no lock taken. */
254 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
256 int64_t total_sectors = bmds->total_sectors;
257 int64_t cur_sector = bmds->cur_sector;
258 BlockDriverState *bs = bmds->bs;
262 if (bmds->shared_base) {
263 qemu_mutex_lock_iothread();
264 while (cur_sector < total_sectors &&
265 !bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH,
267 cur_sector += nr_sectors;
269 qemu_mutex_unlock_iothread();
272 if (cur_sector >= total_sectors) {
273 bmds->cur_sector = bmds->completed_sectors = total_sectors;
277 bmds->completed_sectors = cur_sector;
279 cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
281 /* we are going to transfer a full block even if it is not allocated */
282 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
284 if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
285 nr_sectors = total_sectors - cur_sector;
288 blk = g_new(BlkMigBlock, 1);
289 blk->buf = g_malloc(BLOCK_SIZE);
291 blk->sector = cur_sector;
292 blk->nr_sectors = nr_sectors;
294 blk->iov.iov_base = blk->buf;
295 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
296 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
299 block_mig_state.submitted++;
302 qemu_mutex_lock_iothread();
303 blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov,
304 nr_sectors, blk_mig_read_cb, blk);
306 bdrv_reset_dirty(bs, cur_sector, nr_sectors);
307 qemu_mutex_unlock_iothread();
309 bmds->cur_sector = cur_sector + nr_sectors;
310 return (bmds->cur_sector >= total_sectors);
313 /* Called with iothread lock taken. */
315 static int set_dirty_tracking(void)
317 BlkMigDevState *bmds;
320 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
321 bmds->dirty_bitmap = bdrv_create_dirty_bitmap(bmds->bs, BLOCK_SIZE,
323 if (!bmds->dirty_bitmap) {
331 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
332 if (bmds->dirty_bitmap) {
333 bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap);
339 static void unset_dirty_tracking(void)
341 BlkMigDevState *bmds;
343 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
344 bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap);
348 static void init_blk_migration(QEMUFile *f)
350 BlockDriverState *bs;
351 BlkMigDevState *bmds;
354 block_mig_state.submitted = 0;
355 block_mig_state.read_done = 0;
356 block_mig_state.transferred = 0;
357 block_mig_state.total_sector_sum = 0;
358 block_mig_state.prev_progress = -1;
359 block_mig_state.bulk_completed = 0;
360 block_mig_state.zero_blocks = migrate_zero_blocks();
362 for (bs = bdrv_next(NULL); bs; bs = bdrv_next(bs)) {
363 if (bdrv_is_read_only(bs)) {
367 sectors = bdrv_nb_sectors(bs);
372 bmds = g_new0(BlkMigDevState, 1);
374 bmds->bulk_completed = 0;
375 bmds->total_sectors = sectors;
376 bmds->completed_sectors = 0;
377 bmds->shared_base = block_mig_state.shared_base;
378 alloc_aio_bitmap(bmds);
379 error_setg(&bmds->blocker, "block device is in use by migration");
380 bdrv_op_block_all(bs, bmds->blocker);
383 block_mig_state.total_sector_sum += sectors;
385 if (bmds->shared_base) {
386 DPRINTF("Start migration for %s with shared base image\n",
387 bdrv_get_device_name(bs));
389 DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
392 QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
396 /* Called with no lock taken. */
398 static int blk_mig_save_bulked_block(QEMUFile *f)
400 int64_t completed_sector_sum = 0;
401 BlkMigDevState *bmds;
405 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
406 if (bmds->bulk_completed == 0) {
407 if (mig_save_device_bulk(f, bmds) == 1) {
408 /* completed bulk section for this device */
409 bmds->bulk_completed = 1;
411 completed_sector_sum += bmds->completed_sectors;
415 completed_sector_sum += bmds->completed_sectors;
419 if (block_mig_state.total_sector_sum != 0) {
420 progress = completed_sector_sum * 100 /
421 block_mig_state.total_sector_sum;
425 if (progress != block_mig_state.prev_progress) {
426 block_mig_state.prev_progress = progress;
427 qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
428 | BLK_MIG_FLAG_PROGRESS);
429 DPRINTF("Completed %d %%\r", progress);
435 static void blk_mig_reset_dirty_cursor(void)
437 BlkMigDevState *bmds;
439 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
444 /* Called with iothread lock taken. */
446 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
450 int64_t total_sectors = bmds->total_sectors;
455 for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
457 if (bmds_aio_inflight(bmds, sector)) {
463 if (bdrv_get_dirty(bmds->bs, bmds->dirty_bitmap, sector)) {
465 if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
466 nr_sectors = total_sectors - sector;
468 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
470 blk = g_new(BlkMigBlock, 1);
471 blk->buf = g_malloc(BLOCK_SIZE);
473 blk->sector = sector;
474 blk->nr_sectors = nr_sectors;
477 blk->iov.iov_base = blk->buf;
478 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
479 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
481 blk->aiocb = bdrv_aio_readv(bmds->bs, sector, &blk->qiov,
482 nr_sectors, blk_mig_read_cb, blk);
485 block_mig_state.submitted++;
486 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
489 ret = bdrv_read(bmds->bs, sector, blk->buf, nr_sectors);
499 bdrv_reset_dirty(bmds->bs, sector, nr_sectors);
502 sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
503 bmds->cur_dirty = sector;
506 return (bmds->cur_dirty >= bmds->total_sectors);
509 DPRINTF("Error reading sector %" PRId64 "\n", sector);
515 /* Called with iothread lock taken.
518 * 0: too much data for max_downtime
519 * 1: few enough data for max_downtime
521 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
523 BlkMigDevState *bmds;
526 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
527 ret = mig_save_device_dirty(f, bmds, is_async);
536 /* Called with no locks taken. */
538 static int flush_blks(QEMUFile *f)
543 DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
544 __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
545 block_mig_state.transferred);
548 while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
549 if (qemu_file_rate_limit(f)) {
557 QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
565 block_mig_state.read_done--;
566 block_mig_state.transferred++;
567 assert(block_mig_state.read_done >= 0);
571 DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
572 block_mig_state.submitted, block_mig_state.read_done,
573 block_mig_state.transferred);
577 /* Called with iothread lock taken. */
579 static int64_t get_remaining_dirty(void)
581 BlkMigDevState *bmds;
584 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
585 dirty += bdrv_get_dirty_count(bmds->bs, bmds->dirty_bitmap);
588 return dirty << BDRV_SECTOR_BITS;
591 /* Called with iothread lock taken. */
593 static void blk_mig_cleanup(void)
595 BlkMigDevState *bmds;
600 unset_dirty_tracking();
603 while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
604 QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
605 bdrv_op_unblock_all(bmds->bs, bmds->blocker);
606 error_free(bmds->blocker);
607 bdrv_unref(bmds->bs);
608 g_free(bmds->aio_bitmap);
612 while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
613 QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
620 static void block_migration_cancel(void *opaque)
625 static int block_save_setup(QEMUFile *f, void *opaque)
629 DPRINTF("Enter save live setup submitted %d transferred %d\n",
630 block_mig_state.submitted, block_mig_state.transferred);
632 qemu_mutex_lock_iothread();
633 init_blk_migration(f);
635 /* start track dirty blocks */
636 ret = set_dirty_tracking();
639 qemu_mutex_unlock_iothread();
643 qemu_mutex_unlock_iothread();
646 blk_mig_reset_dirty_cursor();
647 qemu_put_be64(f, BLK_MIG_FLAG_EOS);
652 static int block_save_iterate(QEMUFile *f, void *opaque)
655 int64_t last_ftell = qemu_ftell(f);
657 DPRINTF("Enter save live iterate submitted %d transferred %d\n",
658 block_mig_state.submitted, block_mig_state.transferred);
665 blk_mig_reset_dirty_cursor();
667 /* control the rate of transfer */
669 while ((block_mig_state.submitted +
670 block_mig_state.read_done) * BLOCK_SIZE <
671 qemu_file_get_rate_limit(f)) {
673 if (block_mig_state.bulk_completed == 0) {
674 /* first finish the bulk phase */
675 if (blk_mig_save_bulked_block(f) == 0) {
676 /* finished saving bulk on all devices */
677 block_mig_state.bulk_completed = 1;
681 /* Always called with iothread lock taken for
682 * simplicity, block_save_complete also calls it.
684 qemu_mutex_lock_iothread();
685 ret = blk_mig_save_dirty_block(f, 1);
686 qemu_mutex_unlock_iothread();
693 /* no more dirty blocks */
704 qemu_put_be64(f, BLK_MIG_FLAG_EOS);
705 return qemu_ftell(f) - last_ftell;
708 /* Called with iothread lock taken. */
710 static int block_save_complete(QEMUFile *f, void *opaque)
714 DPRINTF("Enter save live complete submitted %d transferred %d\n",
715 block_mig_state.submitted, block_mig_state.transferred);
722 blk_mig_reset_dirty_cursor();
724 /* we know for sure that save bulk is completed and
725 all async read completed */
727 assert(block_mig_state.submitted == 0);
731 ret = blk_mig_save_dirty_block(f, 0);
737 /* report completion */
738 qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
740 DPRINTF("Block migration completed\n");
742 qemu_put_be64(f, BLK_MIG_FLAG_EOS);
748 static uint64_t block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size)
750 /* Estimate pending number of bytes to send */
753 qemu_mutex_lock_iothread();
755 pending = get_remaining_dirty() +
756 block_mig_state.submitted * BLOCK_SIZE +
757 block_mig_state.read_done * BLOCK_SIZE;
759 /* Report at least one block pending during bulk phase */
760 if (pending == 0 && !block_mig_state.bulk_completed) {
761 pending = BLOCK_SIZE;
764 qemu_mutex_unlock_iothread();
766 DPRINTF("Enter save live pending %" PRIu64 "\n", pending);
770 static int block_load(QEMUFile *f, void *opaque, int version_id)
772 static int banner_printed;
774 char device_name[256];
776 BlockDriverState *bs, *bs_prev = NULL;
778 int64_t total_sectors = 0;
783 addr = qemu_get_be64(f);
785 flags = addr & ~BDRV_SECTOR_MASK;
786 addr >>= BDRV_SECTOR_BITS;
788 if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
789 /* get device name */
790 len = qemu_get_byte(f);
791 qemu_get_buffer(f, (uint8_t *)device_name, len);
792 device_name[len] = '\0';
794 bs = bdrv_find(device_name);
796 fprintf(stderr, "Error unknown block device %s\n",
803 total_sectors = bdrv_nb_sectors(bs);
804 if (total_sectors <= 0) {
805 error_report("Error getting length of block device %s",
811 if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
812 nr_sectors = total_sectors - addr;
814 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
817 if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
818 ret = bdrv_write_zeroes(bs, addr, nr_sectors,
821 buf = g_malloc(BLOCK_SIZE);
822 qemu_get_buffer(f, buf, BLOCK_SIZE);
823 ret = bdrv_write(bs, addr, buf, nr_sectors);
830 } else if (flags & BLK_MIG_FLAG_PROGRESS) {
831 if (!banner_printed) {
832 printf("Receiving block device images\n");
835 printf("Completed %d %%%c", (int)addr,
836 (addr == 100) ? '\n' : '\r');
838 } else if (!(flags & BLK_MIG_FLAG_EOS)) {
839 fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
842 ret = qemu_file_get_error(f);
846 } while (!(flags & BLK_MIG_FLAG_EOS));
851 static void block_set_params(const MigrationParams *params, void *opaque)
853 block_mig_state.blk_enable = params->blk;
854 block_mig_state.shared_base = params->shared;
856 /* shared base means that blk_enable = 1 */
857 block_mig_state.blk_enable |= params->shared;
860 static bool block_is_active(void *opaque)
862 return block_mig_state.blk_enable == 1;
865 static SaveVMHandlers savevm_block_handlers = {
866 .set_params = block_set_params,
867 .save_live_setup = block_save_setup,
868 .save_live_iterate = block_save_iterate,
869 .save_live_complete = block_save_complete,
870 .save_live_pending = block_save_pending,
871 .load_state = block_load,
872 .cancel = block_migration_cancel,
873 .is_active = block_is_active,
876 void blk_mig_init(void)
878 QSIMPLEQ_INIT(&block_mig_state.bmds_list);
879 QSIMPLEQ_INIT(&block_mig_state.blk_list);
880 qemu_mutex_init(&block_mig_state.lock);
882 register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,