root / block-migration.c @ b4bf0a9a
History | View | Annotate | Download (14.1 kB)
1 |
/*
|
---|---|
2 |
* QEMU live block migration
|
3 |
*
|
4 |
* Copyright IBM, Corp. 2009
|
5 |
*
|
6 |
* Authors:
|
7 |
* Liran Schour <lirans@il.ibm.com>
|
8 |
*
|
9 |
* This work is licensed under the terms of the GNU GPL, version 2. See
|
10 |
* the COPYING file in the top-level directory.
|
11 |
*
|
12 |
*/
|
13 |
|
14 |
#include "qemu-common.h" |
15 |
#include "block_int.h" |
16 |
#include "hw/hw.h" |
17 |
#include "qemu-queue.h" |
18 |
#include "monitor.h" |
19 |
#include "block-migration.h" |
20 |
#include <assert.h> |
21 |
|
22 |
#define BLOCK_SIZE (BDRV_SECTORS_PER_DIRTY_CHUNK << BDRV_SECTOR_BITS)
|
23 |
|
24 |
#define BLK_MIG_FLAG_DEVICE_BLOCK 0x01 |
25 |
#define BLK_MIG_FLAG_EOS 0x02 |
26 |
#define BLK_MIG_FLAG_PROGRESS 0x04 |
27 |
|
28 |
#define MAX_IS_ALLOCATED_SEARCH 65536 |
29 |
#define MAX_BLOCKS_READ 10000 |
30 |
#define BLOCKS_READ_CHANGE 100 |
31 |
#define INITIAL_BLOCKS_READ 100 |
32 |
|
33 |
//#define DEBUG_BLK_MIGRATION
|
34 |
|
35 |
#ifdef DEBUG_BLK_MIGRATION
|
36 |
#define dprintf(fmt, ...) \
|
37 |
do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0) |
38 |
#else
|
39 |
#define dprintf(fmt, ...) \
|
40 |
do { } while (0) |
41 |
#endif
|
42 |
|
43 |
typedef struct BlkMigDevState { |
44 |
BlockDriverState *bs; |
45 |
int bulk_completed;
|
46 |
int shared_base;
|
47 |
int64_t cur_sector; |
48 |
int64_t completed_sectors; |
49 |
int64_t total_sectors; |
50 |
int64_t dirty; |
51 |
QSIMPLEQ_ENTRY(BlkMigDevState) entry; |
52 |
} BlkMigDevState; |
53 |
|
54 |
typedef struct BlkMigBlock { |
55 |
uint8_t *buf; |
56 |
BlkMigDevState *bmds; |
57 |
int64_t sector; |
58 |
struct iovec iov;
|
59 |
QEMUIOVector qiov; |
60 |
BlockDriverAIOCB *aiocb; |
61 |
int ret;
|
62 |
QSIMPLEQ_ENTRY(BlkMigBlock) entry; |
63 |
} BlkMigBlock; |
64 |
|
65 |
typedef struct BlkMigState { |
66 |
int blk_enable;
|
67 |
int shared_base;
|
68 |
QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list; |
69 |
QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list; |
70 |
int submitted;
|
71 |
int read_done;
|
72 |
int transferred;
|
73 |
int64_t total_sector_sum; |
74 |
int prev_progress;
|
75 |
} BlkMigState; |
76 |
|
77 |
static BlkMigState block_mig_state;
|
78 |
|
79 |
static void blk_send(QEMUFile *f, BlkMigBlock * blk) |
80 |
{ |
81 |
int len;
|
82 |
|
83 |
/* sector number and flags */
|
84 |
qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS) |
85 |
| BLK_MIG_FLAG_DEVICE_BLOCK); |
86 |
|
87 |
/* device name */
|
88 |
len = strlen(blk->bmds->bs->device_name); |
89 |
qemu_put_byte(f, len); |
90 |
qemu_put_buffer(f, (uint8_t *)blk->bmds->bs->device_name, len); |
91 |
|
92 |
qemu_put_buffer(f, blk->buf, BLOCK_SIZE); |
93 |
} |
94 |
|
95 |
int blk_mig_active(void) |
96 |
{ |
97 |
return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
|
98 |
} |
99 |
|
100 |
uint64_t blk_mig_bytes_transferred(void)
|
101 |
{ |
102 |
BlkMigDevState *bmds; |
103 |
uint64_t sum = 0;
|
104 |
|
105 |
QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { |
106 |
sum += bmds->completed_sectors; |
107 |
} |
108 |
return sum << BDRV_SECTOR_BITS;
|
109 |
} |
110 |
|
111 |
uint64_t blk_mig_bytes_remaining(void)
|
112 |
{ |
113 |
return blk_mig_bytes_total() - blk_mig_bytes_transferred();
|
114 |
} |
115 |
|
116 |
uint64_t blk_mig_bytes_total(void)
|
117 |
{ |
118 |
BlkMigDevState *bmds; |
119 |
uint64_t sum = 0;
|
120 |
|
121 |
QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { |
122 |
sum += bmds->total_sectors; |
123 |
} |
124 |
return sum << BDRV_SECTOR_BITS;
|
125 |
} |
126 |
|
127 |
static void blk_mig_read_cb(void *opaque, int ret) |
128 |
{ |
129 |
BlkMigBlock *blk = opaque; |
130 |
|
131 |
blk->ret = ret; |
132 |
|
133 |
QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry); |
134 |
|
135 |
block_mig_state.submitted--; |
136 |
block_mig_state.read_done++; |
137 |
assert(block_mig_state.submitted >= 0);
|
138 |
} |
139 |
|
140 |
static int mig_save_device_bulk(Monitor *mon, QEMUFile *f, |
141 |
BlkMigDevState *bmds, int is_async)
|
142 |
{ |
143 |
int64_t total_sectors = bmds->total_sectors; |
144 |
int64_t cur_sector = bmds->cur_sector; |
145 |
BlockDriverState *bs = bmds->bs; |
146 |
BlkMigBlock *blk; |
147 |
int nr_sectors;
|
148 |
|
149 |
if (bmds->shared_base) {
|
150 |
while (cur_sector < total_sectors &&
|
151 |
!bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH, |
152 |
&nr_sectors)) { |
153 |
cur_sector += nr_sectors; |
154 |
} |
155 |
} |
156 |
|
157 |
if (cur_sector >= total_sectors) {
|
158 |
bmds->cur_sector = bmds->completed_sectors = total_sectors; |
159 |
return 1; |
160 |
} |
161 |
|
162 |
bmds->completed_sectors = cur_sector; |
163 |
|
164 |
cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
|
165 |
|
166 |
/* we are going to transfer a full block even if it is not allocated */
|
167 |
nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK; |
168 |
|
169 |
if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
|
170 |
nr_sectors = total_sectors - cur_sector; |
171 |
} |
172 |
|
173 |
blk = qemu_malloc(sizeof(BlkMigBlock));
|
174 |
blk->buf = qemu_malloc(BLOCK_SIZE); |
175 |
blk->bmds = bmds; |
176 |
blk->sector = cur_sector; |
177 |
|
178 |
if (is_async) {
|
179 |
blk->iov.iov_base = blk->buf; |
180 |
blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE; |
181 |
qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
|
182 |
|
183 |
blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov, |
184 |
nr_sectors, blk_mig_read_cb, blk); |
185 |
if (!blk->aiocb) {
|
186 |
goto error;
|
187 |
} |
188 |
block_mig_state.submitted++; |
189 |
} else {
|
190 |
if (bdrv_read(bs, cur_sector, blk->buf, nr_sectors) < 0) { |
191 |
goto error;
|
192 |
} |
193 |
blk_send(f, blk); |
194 |
|
195 |
qemu_free(blk->buf); |
196 |
qemu_free(blk); |
197 |
} |
198 |
|
199 |
bdrv_reset_dirty(bs, cur_sector, nr_sectors); |
200 |
bmds->cur_sector = cur_sector + nr_sectors; |
201 |
|
202 |
return (bmds->cur_sector >= total_sectors);
|
203 |
|
204 |
error:
|
205 |
monitor_printf(mon, "Error reading sector %" PRId64 "\n", cur_sector); |
206 |
qemu_file_set_error(f); |
207 |
qemu_free(blk->buf); |
208 |
qemu_free(blk); |
209 |
return 0; |
210 |
} |
211 |
|
212 |
static void set_dirty_tracking(int enable) |
213 |
{ |
214 |
BlkMigDevState *bmds; |
215 |
|
216 |
QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { |
217 |
bdrv_set_dirty_tracking(bmds->bs, enable); |
218 |
} |
219 |
} |
220 |
|
221 |
static void init_blk_migration(Monitor *mon, QEMUFile *f) |
222 |
{ |
223 |
BlkMigDevState *bmds; |
224 |
BlockDriverState *bs; |
225 |
int64_t sectors; |
226 |
|
227 |
block_mig_state.submitted = 0;
|
228 |
block_mig_state.read_done = 0;
|
229 |
block_mig_state.transferred = 0;
|
230 |
block_mig_state.total_sector_sum = 0;
|
231 |
block_mig_state.prev_progress = -1;
|
232 |
|
233 |
for (bs = bdrv_first; bs != NULL; bs = bs->next) { |
234 |
if (bs->type == BDRV_TYPE_HD) {
|
235 |
sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS; |
236 |
if (sectors == 0) { |
237 |
continue;
|
238 |
} |
239 |
|
240 |
bmds = qemu_mallocz(sizeof(BlkMigDevState));
|
241 |
bmds->bs = bs; |
242 |
bmds->bulk_completed = 0;
|
243 |
bmds->total_sectors = sectors; |
244 |
bmds->completed_sectors = 0;
|
245 |
bmds->shared_base = block_mig_state.shared_base; |
246 |
|
247 |
block_mig_state.total_sector_sum += sectors; |
248 |
|
249 |
if (bmds->shared_base) {
|
250 |
monitor_printf(mon, "Start migration for %s with shared base "
|
251 |
"image\n",
|
252 |
bs->device_name); |
253 |
} else {
|
254 |
monitor_printf(mon, "Start full migration for %s\n",
|
255 |
bs->device_name); |
256 |
} |
257 |
|
258 |
QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry); |
259 |
} |
260 |
} |
261 |
} |
262 |
|
263 |
static int blk_mig_save_bulked_block(Monitor *mon, QEMUFile *f, int is_async) |
264 |
{ |
265 |
int64_t completed_sector_sum = 0;
|
266 |
BlkMigDevState *bmds; |
267 |
int progress;
|
268 |
int ret = 0; |
269 |
|
270 |
QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { |
271 |
if (bmds->bulk_completed == 0) { |
272 |
if (mig_save_device_bulk(mon, f, bmds, is_async) == 1) { |
273 |
/* completed bulk section for this device */
|
274 |
bmds->bulk_completed = 1;
|
275 |
} |
276 |
completed_sector_sum += bmds->completed_sectors; |
277 |
ret = 1;
|
278 |
break;
|
279 |
} else {
|
280 |
completed_sector_sum += bmds->completed_sectors; |
281 |
} |
282 |
} |
283 |
|
284 |
progress = completed_sector_sum * 100 / block_mig_state.total_sector_sum;
|
285 |
if (progress != block_mig_state.prev_progress) {
|
286 |
block_mig_state.prev_progress = progress; |
287 |
qemu_put_be64(f, (progress << BDRV_SECTOR_BITS) |
288 |
| BLK_MIG_FLAG_PROGRESS); |
289 |
monitor_printf(mon, "Completed %d %%\r", progress);
|
290 |
monitor_flush(mon); |
291 |
} |
292 |
|
293 |
return ret;
|
294 |
} |
295 |
|
296 |
#define MAX_NUM_BLOCKS 4 |
297 |
|
298 |
static void blk_mig_save_dirty_blocks(Monitor *mon, QEMUFile *f) |
299 |
{ |
300 |
BlkMigDevState *bmds; |
301 |
BlkMigBlock blk; |
302 |
int64_t sector; |
303 |
|
304 |
blk.buf = qemu_malloc(BLOCK_SIZE); |
305 |
|
306 |
QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { |
307 |
for (sector = 0; sector < bmds->cur_sector;) { |
308 |
if (bdrv_get_dirty(bmds->bs, sector)) {
|
309 |
if (bdrv_read(bmds->bs, sector, blk.buf,
|
310 |
BDRV_SECTORS_PER_DIRTY_CHUNK) < 0) {
|
311 |
monitor_printf(mon, "Error reading sector %" PRId64 "\n", |
312 |
sector); |
313 |
qemu_file_set_error(f); |
314 |
qemu_free(blk.buf); |
315 |
return;
|
316 |
} |
317 |
blk.bmds = bmds; |
318 |
blk.sector = sector; |
319 |
blk_send(f, &blk); |
320 |
|
321 |
bdrv_reset_dirty(bmds->bs, sector, |
322 |
BDRV_SECTORS_PER_DIRTY_CHUNK); |
323 |
} |
324 |
sector += BDRV_SECTORS_PER_DIRTY_CHUNK; |
325 |
} |
326 |
} |
327 |
|
328 |
qemu_free(blk.buf); |
329 |
} |
330 |
|
331 |
static void flush_blks(QEMUFile* f) |
332 |
{ |
333 |
BlkMigBlock *blk; |
334 |
|
335 |
dprintf("%s Enter submitted %d read_done %d transferred %d\n",
|
336 |
__FUNCTION__, block_mig_state.submitted, block_mig_state.read_done, |
337 |
block_mig_state.transferred); |
338 |
|
339 |
while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) { |
340 |
if (qemu_file_rate_limit(f)) {
|
341 |
break;
|
342 |
} |
343 |
if (blk->ret < 0) { |
344 |
qemu_file_set_error(f); |
345 |
break;
|
346 |
} |
347 |
blk_send(f, blk); |
348 |
|
349 |
QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry); |
350 |
qemu_free(blk->buf); |
351 |
qemu_free(blk); |
352 |
|
353 |
block_mig_state.read_done--; |
354 |
block_mig_state.transferred++; |
355 |
assert(block_mig_state.read_done >= 0);
|
356 |
} |
357 |
|
358 |
dprintf("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
|
359 |
block_mig_state.submitted, block_mig_state.read_done, |
360 |
block_mig_state.transferred); |
361 |
} |
362 |
|
363 |
static int is_stage2_completed(void) |
364 |
{ |
365 |
BlkMigDevState *bmds; |
366 |
|
367 |
if (block_mig_state.submitted > 0) { |
368 |
return 0; |
369 |
} |
370 |
|
371 |
QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { |
372 |
if (bmds->bulk_completed == 0) { |
373 |
return 0; |
374 |
} |
375 |
} |
376 |
|
377 |
return 1; |
378 |
} |
379 |
|
380 |
static void blk_mig_cleanup(Monitor *mon) |
381 |
{ |
382 |
BlkMigDevState *bmds; |
383 |
BlkMigBlock *blk; |
384 |
|
385 |
while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) { |
386 |
QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry); |
387 |
qemu_free(bmds); |
388 |
} |
389 |
|
390 |
while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) { |
391 |
QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry); |
392 |
qemu_free(blk->buf); |
393 |
qemu_free(blk); |
394 |
} |
395 |
|
396 |
set_dirty_tracking(0);
|
397 |
|
398 |
monitor_printf(mon, "\n");
|
399 |
} |
400 |
|
401 |
static int block_save_live(Monitor *mon, QEMUFile *f, int stage, void *opaque) |
402 |
{ |
403 |
dprintf("Enter save live stage %d submitted %d transferred %d\n",
|
404 |
stage, block_mig_state.submitted, block_mig_state.transferred); |
405 |
|
406 |
if (stage < 0) { |
407 |
blk_mig_cleanup(mon); |
408 |
return 0; |
409 |
} |
410 |
|
411 |
if (block_mig_state.blk_enable != 1) { |
412 |
/* no need to migrate storage */
|
413 |
qemu_put_be64(f, BLK_MIG_FLAG_EOS); |
414 |
return 1; |
415 |
} |
416 |
|
417 |
if (stage == 1) { |
418 |
init_blk_migration(mon, f); |
419 |
|
420 |
/* start track dirty blocks */
|
421 |
set_dirty_tracking(1);
|
422 |
} |
423 |
|
424 |
flush_blks(f); |
425 |
|
426 |
if (qemu_file_has_error(f)) {
|
427 |
blk_mig_cleanup(mon); |
428 |
return 0; |
429 |
} |
430 |
|
431 |
/* control the rate of transfer */
|
432 |
while ((block_mig_state.submitted +
|
433 |
block_mig_state.read_done) * BLOCK_SIZE < |
434 |
qemu_file_get_rate_limit(f)) { |
435 |
if (blk_mig_save_bulked_block(mon, f, 1) == 0) { |
436 |
/* no more bulk blocks for now */
|
437 |
break;
|
438 |
} |
439 |
} |
440 |
|
441 |
flush_blks(f); |
442 |
|
443 |
if (qemu_file_has_error(f)) {
|
444 |
blk_mig_cleanup(mon); |
445 |
return 0; |
446 |
} |
447 |
|
448 |
if (stage == 3) { |
449 |
while (blk_mig_save_bulked_block(mon, f, 0) != 0) { |
450 |
/* empty */
|
451 |
} |
452 |
|
453 |
blk_mig_save_dirty_blocks(mon, f); |
454 |
blk_mig_cleanup(mon); |
455 |
|
456 |
/* report completion */
|
457 |
qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
|
458 |
|
459 |
if (qemu_file_has_error(f)) {
|
460 |
return 0; |
461 |
} |
462 |
|
463 |
monitor_printf(mon, "Block migration completed\n");
|
464 |
} |
465 |
|
466 |
qemu_put_be64(f, BLK_MIG_FLAG_EOS); |
467 |
|
468 |
return ((stage == 2) && is_stage2_completed()); |
469 |
} |
470 |
|
471 |
static int block_load(QEMUFile *f, void *opaque, int version_id) |
472 |
{ |
473 |
static int banner_printed; |
474 |
int len, flags;
|
475 |
char device_name[256]; |
476 |
int64_t addr; |
477 |
BlockDriverState *bs; |
478 |
uint8_t *buf; |
479 |
|
480 |
do {
|
481 |
addr = qemu_get_be64(f); |
482 |
|
483 |
flags = addr & ~BDRV_SECTOR_MASK; |
484 |
addr >>= BDRV_SECTOR_BITS; |
485 |
|
486 |
if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
|
487 |
/* get device name */
|
488 |
len = qemu_get_byte(f); |
489 |
qemu_get_buffer(f, (uint8_t *)device_name, len); |
490 |
device_name[len] = '\0';
|
491 |
|
492 |
bs = bdrv_find(device_name); |
493 |
if (!bs) {
|
494 |
fprintf(stderr, "Error unknown block device %s\n",
|
495 |
device_name); |
496 |
return -EINVAL;
|
497 |
} |
498 |
|
499 |
buf = qemu_malloc(BLOCK_SIZE); |
500 |
|
501 |
qemu_get_buffer(f, buf, BLOCK_SIZE); |
502 |
bdrv_write(bs, addr, buf, BDRV_SECTORS_PER_DIRTY_CHUNK); |
503 |
|
504 |
qemu_free(buf); |
505 |
} else if (flags & BLK_MIG_FLAG_PROGRESS) { |
506 |
if (!banner_printed) {
|
507 |
printf("Receiving block device images\n");
|
508 |
banner_printed = 1;
|
509 |
} |
510 |
printf("Completed %d %%%c", (int)addr, |
511 |
(addr == 100) ? '\n' : '\r'); |
512 |
fflush(stdout); |
513 |
} else if (!(flags & BLK_MIG_FLAG_EOS)) { |
514 |
fprintf(stderr, "Unknown flags\n");
|
515 |
return -EINVAL;
|
516 |
} |
517 |
if (qemu_file_has_error(f)) {
|
518 |
return -EIO;
|
519 |
} |
520 |
} while (!(flags & BLK_MIG_FLAG_EOS));
|
521 |
|
522 |
return 0; |
523 |
} |
524 |
|
525 |
static void block_set_params(int blk_enable, int shared_base, void *opaque) |
526 |
{ |
527 |
block_mig_state.blk_enable = blk_enable; |
528 |
block_mig_state.shared_base = shared_base; |
529 |
|
530 |
/* shared base means that blk_enable = 1 */
|
531 |
block_mig_state.blk_enable |= shared_base; |
532 |
} |
533 |
|
534 |
void blk_mig_init(void) |
535 |
{ |
536 |
QSIMPLEQ_INIT(&block_mig_state.bmds_list); |
537 |
QSIMPLEQ_INIT(&block_mig_state.blk_list); |
538 |
|
539 |
register_savevm_live("block", 0, 1, block_set_params, block_save_live, |
540 |
NULL, block_load, &block_mig_state);
|
541 |
} |