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.
14 #include "qemu-common.h"
15 #include "block_int.h"
17 #include "qemu-queue.h"
18 #include "block-migration.h"
21 #define BLOCK_SIZE (BDRV_SECTORS_PER_DIRTY_CHUNK << BDRV_SECTOR_BITS)
23 #define BLK_MIG_FLAG_DEVICE_BLOCK 0x01
24 #define BLK_MIG_FLAG_EOS 0x02
26 #define MAX_IS_ALLOCATED_SEARCH 65536
27 #define MAX_BLOCKS_READ 10000
28 #define BLOCKS_READ_CHANGE 100
29 #define INITIAL_BLOCKS_READ 100
31 //#define DEBUG_BLK_MIGRATION
33 #ifdef DEBUG_BLK_MIGRATION
34 #define dprintf(fmt, ...) \
35 do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
37 #define dprintf(fmt, ...) \
41 typedef struct BlkMigDevState
{
46 int64_t total_sectors
;
48 QSIMPLEQ_ENTRY(BlkMigDevState
) entry
;
51 typedef struct BlkMigBlock
{
57 BlockDriverAIOCB
*aiocb
;
59 QSIMPLEQ_ENTRY(BlkMigBlock
) entry
;
62 typedef struct BlkMigState
{
65 QSIMPLEQ_HEAD(bmds_list
, BlkMigDevState
) bmds_list
;
66 QSIMPLEQ_HEAD(blk_list
, BlkMigBlock
) blk_list
;
70 int64_t print_completion
;
73 static BlkMigState block_mig_state
;
75 static void blk_send(QEMUFile
*f
, BlkMigBlock
* blk
)
79 /* sector number and flags */
80 qemu_put_be64(f
, (blk
->sector
<< BDRV_SECTOR_BITS
)
81 | BLK_MIG_FLAG_DEVICE_BLOCK
);
84 len
= strlen(blk
->bmds
->bs
->device_name
);
85 qemu_put_byte(f
, len
);
86 qemu_put_buffer(f
, (uint8_t *)blk
->bmds
->bs
->device_name
, len
);
88 qemu_put_buffer(f
, blk
->buf
, BLOCK_SIZE
);
91 static void blk_mig_read_cb(void *opaque
, int ret
)
93 BlkMigBlock
*blk
= opaque
;
97 QSIMPLEQ_INSERT_TAIL(&block_mig_state
.blk_list
, blk
, entry
);
99 block_mig_state
.submitted
--;
100 block_mig_state
.read_done
++;
101 assert(block_mig_state
.submitted
>= 0);
104 static int mig_save_device_bulk(QEMUFile
*f
, BlkMigDevState
*bmds
, int is_async
)
106 int64_t total_sectors
= bmds
->total_sectors
;
107 int64_t cur_sector
= bmds
->cur_sector
;
108 BlockDriverState
*bs
= bmds
->bs
;
112 if (bmds
->shared_base
) {
113 while (cur_sector
< total_sectors
&&
114 !bdrv_is_allocated(bs
, cur_sector
, MAX_IS_ALLOCATED_SEARCH
,
116 cur_sector
+= nr_sectors
;
120 if (cur_sector
>= total_sectors
) {
121 bmds
->cur_sector
= total_sectors
;
125 if (cur_sector
>= block_mig_state
.print_completion
) {
126 printf("Completed %" PRId64
" %%\r", cur_sector
* 100 / total_sectors
);
128 block_mig_state
.print_completion
+=
129 (BDRV_SECTORS_PER_DIRTY_CHUNK
* 10000);
132 cur_sector
&= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK
- 1);
134 /* we are going to transfer a full block even if it is not allocated */
135 nr_sectors
= BDRV_SECTORS_PER_DIRTY_CHUNK
;
137 if (total_sectors
- cur_sector
< BDRV_SECTORS_PER_DIRTY_CHUNK
) {
138 nr_sectors
= total_sectors
- cur_sector
;
141 blk
= qemu_malloc(sizeof(BlkMigBlock
));
142 blk
->buf
= qemu_malloc(BLOCK_SIZE
);
144 blk
->sector
= cur_sector
;
147 blk
->iov
.iov_base
= blk
->buf
;
148 blk
->iov
.iov_len
= nr_sectors
* BDRV_SECTOR_SIZE
;
149 qemu_iovec_init_external(&blk
->qiov
, &blk
->iov
, 1);
151 blk
->aiocb
= bdrv_aio_readv(bs
, cur_sector
, &blk
->qiov
,
152 nr_sectors
, blk_mig_read_cb
, blk
);
156 block_mig_state
.submitted
++;
158 if (bdrv_read(bs
, cur_sector
, blk
->buf
, nr_sectors
) < 0) {
167 bdrv_reset_dirty(bs
, cur_sector
, nr_sectors
);
168 bmds
->cur_sector
= cur_sector
+ nr_sectors
;
170 return (bmds
->cur_sector
>= total_sectors
);
173 printf("Error reading sector %" PRId64
"\n", cur_sector
);
174 qemu_file_set_error(f
);
180 static void set_dirty_tracking(int enable
)
182 BlkMigDevState
*bmds
;
184 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
185 bdrv_set_dirty_tracking(bmds
->bs
, enable
);
189 static void init_blk_migration(QEMUFile
*f
)
191 BlkMigDevState
*bmds
;
192 BlockDriverState
*bs
;
194 block_mig_state
.submitted
= 0;
195 block_mig_state
.read_done
= 0;
196 block_mig_state
.transferred
= 0;
197 block_mig_state
.print_completion
= 0;
199 for (bs
= bdrv_first
; bs
!= NULL
; bs
= bs
->next
) {
200 if (bs
->type
== BDRV_TYPE_HD
) {
201 bmds
= qemu_mallocz(sizeof(BlkMigDevState
));
203 bmds
->bulk_completed
= 0;
204 bmds
->total_sectors
= bdrv_getlength(bs
) >> BDRV_SECTOR_BITS
;
205 bmds
->shared_base
= block_mig_state
.shared_base
;
207 if (bmds
->shared_base
) {
208 printf("Start migration for %s with shared base image\n",
211 printf("Start full migration for %s\n", bs
->device_name
);
214 QSIMPLEQ_INSERT_TAIL(&block_mig_state
.bmds_list
, bmds
, entry
);
219 static int blk_mig_save_bulked_block(QEMUFile
*f
, int is_async
)
221 BlkMigDevState
*bmds
;
223 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
224 if (bmds
->bulk_completed
== 0) {
225 if (mig_save_device_bulk(f
, bmds
, is_async
) == 1) {
226 /* completed bulk section for this device */
227 bmds
->bulk_completed
= 1;
233 /* we reached here means bulk is completed */
237 #define MAX_NUM_BLOCKS 4
239 static void blk_mig_save_dirty_blocks(QEMUFile
*f
)
241 BlkMigDevState
*bmds
;
245 blk
.buf
= qemu_malloc(BLOCK_SIZE
);
247 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
248 for (sector
= 0; sector
< bmds
->cur_sector
;) {
249 if (bdrv_get_dirty(bmds
->bs
, sector
)) {
250 if (bdrv_read(bmds
->bs
, sector
, blk
.buf
,
251 BDRV_SECTORS_PER_DIRTY_CHUNK
) < 0) {
252 printf("Error reading sector %" PRId64
"\n", sector
);
253 qemu_file_set_error(f
);
261 bdrv_reset_dirty(bmds
->bs
, sector
,
262 BDRV_SECTORS_PER_DIRTY_CHUNK
);
264 sector
+= BDRV_SECTORS_PER_DIRTY_CHUNK
;
271 static void flush_blks(QEMUFile
* f
)
275 dprintf("%s Enter submitted %d read_done %d transferred %d\n",
276 __FUNCTION__
, block_mig_state
.submitted
, block_mig_state
.read_done
,
277 block_mig_state
.transferred
);
279 while ((blk
= QSIMPLEQ_FIRST(&block_mig_state
.blk_list
)) != NULL
) {
280 if (qemu_file_rate_limit(f
)) {
284 qemu_file_set_error(f
);
289 QSIMPLEQ_REMOVE_HEAD(&block_mig_state
.blk_list
, entry
);
293 block_mig_state
.read_done
--;
294 block_mig_state
.transferred
++;
295 assert(block_mig_state
.read_done
>= 0);
298 dprintf("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__
,
299 block_mig_state
.submitted
, block_mig_state
.read_done
,
300 block_mig_state
.transferred
);
303 static int is_stage2_completed(void)
305 BlkMigDevState
*bmds
;
307 if (block_mig_state
.submitted
> 0) {
311 QSIMPLEQ_FOREACH(bmds
, &block_mig_state
.bmds_list
, entry
) {
312 if (bmds
->bulk_completed
== 0) {
320 static int block_save_live(QEMUFile
*f
, int stage
, void *opaque
)
322 dprintf("Enter save live stage %d submitted %d transferred %d\n",
323 stage
, block_mig_state
.submitted
, block_mig_state
.transferred
);
325 if (block_mig_state
.blk_enable
!= 1) {
326 /* no need to migrate storage */
327 qemu_put_be64(f
, BLK_MIG_FLAG_EOS
);
332 init_blk_migration(f
);
334 /* start track dirty blocks */
335 set_dirty_tracking(1);
340 if (qemu_file_has_error(f
)) {
341 set_dirty_tracking(0);
345 /* control the rate of transfer */
346 while ((block_mig_state
.submitted
+
347 block_mig_state
.read_done
) * BLOCK_SIZE
<
348 qemu_file_get_rate_limit(f
)) {
349 if (blk_mig_save_bulked_block(f
, 1) == 0) {
350 /* no more bulk blocks for now */
357 if (qemu_file_has_error(f
)) {
358 set_dirty_tracking(0);
363 while (blk_mig_save_bulked_block(f
, 0) != 0) {
367 blk_mig_save_dirty_blocks(f
);
369 /* stop track dirty blocks */
370 set_dirty_tracking(0);
372 if (qemu_file_has_error(f
)) {
376 printf("\nBlock migration completed\n");
379 qemu_put_be64(f
, BLK_MIG_FLAG_EOS
);
381 return ((stage
== 2) && is_stage2_completed());
384 static int block_load(QEMUFile
*f
, void *opaque
, int version_id
)
387 char device_name
[256];
389 BlockDriverState
*bs
;
393 addr
= qemu_get_be64(f
);
395 flags
= addr
& ~BDRV_SECTOR_MASK
;
396 addr
>>= BDRV_SECTOR_BITS
;
398 if (flags
& BLK_MIG_FLAG_DEVICE_BLOCK
) {
399 /* get device name */
400 len
= qemu_get_byte(f
);
401 qemu_get_buffer(f
, (uint8_t *)device_name
, len
);
402 device_name
[len
] = '\0';
404 bs
= bdrv_find(device_name
);
406 fprintf(stderr
, "Error unknown block device %s\n",
411 buf
= qemu_malloc(BLOCK_SIZE
);
413 qemu_get_buffer(f
, buf
, BLOCK_SIZE
);
414 bdrv_write(bs
, addr
, buf
, BDRV_SECTORS_PER_DIRTY_CHUNK
);
417 } else if (!(flags
& BLK_MIG_FLAG_EOS
)) {
418 fprintf(stderr
, "Unknown flags\n");
421 if (qemu_file_has_error(f
)) {
424 } while (!(flags
& BLK_MIG_FLAG_EOS
));
429 static void block_set_params(int blk_enable
, int shared_base
, void *opaque
)
431 block_mig_state
.blk_enable
= blk_enable
;
432 block_mig_state
.shared_base
= shared_base
;
434 /* shared base means that blk_enable = 1 */
435 block_mig_state
.blk_enable
|= shared_base
;
438 void blk_mig_init(void)
440 QSIMPLEQ_INIT(&block_mig_state
.bmds_list
);
441 QSIMPLEQ_INIT(&block_mig_state
.blk_list
);
443 register_savevm_live("block", 0, 1, block_set_params
, block_save_live
,
444 NULL
, block_load
, &block_mig_state
);