]> git.proxmox.com Git - qemu.git/blob - block-migration.c
block migration: do not submit multiple AIOs for same sector
[qemu.git] / block-migration.c
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 "qemu-timer.h"
19 #include "monitor.h"
20 #include "block-migration.h"
21 #include "migration.h"
22 #include <assert.h>
23
24 #define BLOCK_SIZE (BDRV_SECTORS_PER_DIRTY_CHUNK << BDRV_SECTOR_BITS)
25
26 #define BLK_MIG_FLAG_DEVICE_BLOCK 0x01
27 #define BLK_MIG_FLAG_EOS 0x02
28 #define BLK_MIG_FLAG_PROGRESS 0x04
29
30 #define MAX_IS_ALLOCATED_SEARCH 65536
31
32 //#define DEBUG_BLK_MIGRATION
33
34 #ifdef DEBUG_BLK_MIGRATION
35 #define DPRINTF(fmt, ...) \
36 do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
37 #else
38 #define DPRINTF(fmt, ...) \
39 do { } while (0)
40 #endif
41
42 typedef struct BlkMigDevState {
43 BlockDriverState *bs;
44 int bulk_completed;
45 int shared_base;
46 int64_t cur_sector;
47 int64_t cur_dirty;
48 int64_t completed_sectors;
49 int64_t total_sectors;
50 int64_t dirty;
51 QSIMPLEQ_ENTRY(BlkMigDevState) entry;
52 unsigned long *aio_bitmap;
53 } BlkMigDevState;
54
55 typedef struct BlkMigBlock {
56 uint8_t *buf;
57 BlkMigDevState *bmds;
58 int64_t sector;
59 int nr_sectors;
60 struct iovec iov;
61 QEMUIOVector qiov;
62 BlockDriverAIOCB *aiocb;
63 int ret;
64 int64_t time;
65 QSIMPLEQ_ENTRY(BlkMigBlock) entry;
66 } BlkMigBlock;
67
68 typedef struct BlkMigState {
69 int blk_enable;
70 int shared_base;
71 QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
72 QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
73 int submitted;
74 int read_done;
75 int transferred;
76 int64_t total_sector_sum;
77 int prev_progress;
78 int bulk_completed;
79 long double total_time;
80 int reads;
81 } BlkMigState;
82
83 static BlkMigState block_mig_state;
84
85 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
86 {
87 int len;
88
89 /* sector number and flags */
90 qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
91 | BLK_MIG_FLAG_DEVICE_BLOCK);
92
93 /* device name */
94 len = strlen(blk->bmds->bs->device_name);
95 qemu_put_byte(f, len);
96 qemu_put_buffer(f, (uint8_t *)blk->bmds->bs->device_name, len);
97
98 qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
99 }
100
101 int blk_mig_active(void)
102 {
103 return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
104 }
105
106 uint64_t blk_mig_bytes_transferred(void)
107 {
108 BlkMigDevState *bmds;
109 uint64_t sum = 0;
110
111 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
112 sum += bmds->completed_sectors;
113 }
114 return sum << BDRV_SECTOR_BITS;
115 }
116
117 uint64_t blk_mig_bytes_remaining(void)
118 {
119 return blk_mig_bytes_total() - blk_mig_bytes_transferred();
120 }
121
122 uint64_t blk_mig_bytes_total(void)
123 {
124 BlkMigDevState *bmds;
125 uint64_t sum = 0;
126
127 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
128 sum += bmds->total_sectors;
129 }
130 return sum << BDRV_SECTOR_BITS;
131 }
132
133 static inline void add_avg_read_time(int64_t time)
134 {
135 block_mig_state.reads++;
136 block_mig_state.total_time += time;
137 }
138
139 static inline long double compute_read_bwidth(void)
140 {
141 assert(block_mig_state.total_time != 0);
142 return (block_mig_state.reads * BLOCK_SIZE)/ block_mig_state.total_time;
143 }
144
145 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
146 {
147 int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
148
149 if (bmds->aio_bitmap &&
150 (sector << BDRV_SECTOR_BITS) < bdrv_getlength(bmds->bs)) {
151 return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
152 (1UL << (chunk % (sizeof(unsigned long) * 8))));
153 } else {
154 return 0;
155 }
156 }
157
158 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
159 int nb_sectors, int set)
160 {
161 int64_t start, end;
162 unsigned long val, idx, bit;
163
164 start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
165 end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
166
167 for (; start <= end; start++) {
168 idx = start / (sizeof(unsigned long) * 8);
169 bit = start % (sizeof(unsigned long) * 8);
170 val = bmds->aio_bitmap[idx];
171 if (set) {
172 if (!(val & (1UL << bit))) {
173 val |= 1UL << bit;
174 }
175 } else {
176 if (val & (1UL << bit)) {
177 val &= ~(1UL << bit);
178 }
179 }
180 bmds->aio_bitmap[idx] = val;
181 }
182 }
183
184 static void alloc_aio_bitmap(BlkMigDevState *bmds)
185 {
186 BlockDriverState *bs = bmds->bs;
187 int64_t bitmap_size;
188
189 bitmap_size = (bdrv_getlength(bs) >> BDRV_SECTOR_BITS) +
190 BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
191 bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
192
193 bmds->aio_bitmap = qemu_mallocz(bitmap_size);
194 }
195
196 static void blk_mig_read_cb(void *opaque, int ret)
197 {
198 BlkMigBlock *blk = opaque;
199
200 blk->ret = ret;
201
202 blk->time = qemu_get_clock_ns(rt_clock) - blk->time;
203
204 add_avg_read_time(blk->time);
205
206 QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
207 bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
208
209 block_mig_state.submitted--;
210 block_mig_state.read_done++;
211 assert(block_mig_state.submitted >= 0);
212 }
213
214 static int mig_save_device_bulk(Monitor *mon, QEMUFile *f,
215 BlkMigDevState *bmds)
216 {
217 int64_t total_sectors = bmds->total_sectors;
218 int64_t cur_sector = bmds->cur_sector;
219 BlockDriverState *bs = bmds->bs;
220 BlkMigBlock *blk;
221 int nr_sectors;
222
223 if (bmds->shared_base) {
224 while (cur_sector < total_sectors &&
225 !bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH,
226 &nr_sectors)) {
227 cur_sector += nr_sectors;
228 }
229 }
230
231 if (cur_sector >= total_sectors) {
232 bmds->cur_sector = bmds->completed_sectors = total_sectors;
233 return 1;
234 }
235
236 bmds->completed_sectors = cur_sector;
237
238 cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
239
240 /* we are going to transfer a full block even if it is not allocated */
241 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
242
243 if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
244 nr_sectors = total_sectors - cur_sector;
245 }
246
247 blk = qemu_malloc(sizeof(BlkMigBlock));
248 blk->buf = qemu_malloc(BLOCK_SIZE);
249 blk->bmds = bmds;
250 blk->sector = cur_sector;
251 blk->nr_sectors = nr_sectors;
252
253 blk->iov.iov_base = blk->buf;
254 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
255 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
256
257 blk->time = qemu_get_clock_ns(rt_clock);
258
259 blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov,
260 nr_sectors, blk_mig_read_cb, blk);
261 if (!blk->aiocb) {
262 goto error;
263 }
264 block_mig_state.submitted++;
265
266 bdrv_reset_dirty(bs, cur_sector, nr_sectors);
267 bmds->cur_sector = cur_sector + nr_sectors;
268
269 return (bmds->cur_sector >= total_sectors);
270
271 error:
272 monitor_printf(mon, "Error reading sector %" PRId64 "\n", cur_sector);
273 qemu_file_set_error(f);
274 qemu_free(blk->buf);
275 qemu_free(blk);
276 return 0;
277 }
278
279 static void set_dirty_tracking(int enable)
280 {
281 BlkMigDevState *bmds;
282
283 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
284 bdrv_set_dirty_tracking(bmds->bs, enable);
285 }
286 }
287
288 static void init_blk_migration_it(void *opaque, BlockDriverState *bs)
289 {
290 Monitor *mon = opaque;
291 BlkMigDevState *bmds;
292 int64_t sectors;
293
294 if (!bdrv_is_read_only(bs)) {
295 sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
296 if (sectors <= 0) {
297 return;
298 }
299
300 bmds = qemu_mallocz(sizeof(BlkMigDevState));
301 bmds->bs = bs;
302 bmds->bulk_completed = 0;
303 bmds->total_sectors = sectors;
304 bmds->completed_sectors = 0;
305 bmds->shared_base = block_mig_state.shared_base;
306 alloc_aio_bitmap(bmds);
307
308 block_mig_state.total_sector_sum += sectors;
309
310 if (bmds->shared_base) {
311 monitor_printf(mon, "Start migration for %s with shared base "
312 "image\n",
313 bs->device_name);
314 } else {
315 monitor_printf(mon, "Start full migration for %s\n",
316 bs->device_name);
317 }
318
319 QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
320 }
321 }
322
323 static void init_blk_migration(Monitor *mon, QEMUFile *f)
324 {
325 block_mig_state.submitted = 0;
326 block_mig_state.read_done = 0;
327 block_mig_state.transferred = 0;
328 block_mig_state.total_sector_sum = 0;
329 block_mig_state.prev_progress = -1;
330 block_mig_state.bulk_completed = 0;
331 block_mig_state.total_time = 0;
332 block_mig_state.reads = 0;
333
334 bdrv_iterate(init_blk_migration_it, mon);
335 }
336
337 static int blk_mig_save_bulked_block(Monitor *mon, QEMUFile *f)
338 {
339 int64_t completed_sector_sum = 0;
340 BlkMigDevState *bmds;
341 int progress;
342 int ret = 0;
343
344 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
345 if (bmds->bulk_completed == 0) {
346 if (mig_save_device_bulk(mon, f, bmds) == 1) {
347 /* completed bulk section for this device */
348 bmds->bulk_completed = 1;
349 }
350 completed_sector_sum += bmds->completed_sectors;
351 ret = 1;
352 break;
353 } else {
354 completed_sector_sum += bmds->completed_sectors;
355 }
356 }
357
358 progress = completed_sector_sum * 100 / block_mig_state.total_sector_sum;
359 if (progress != block_mig_state.prev_progress) {
360 block_mig_state.prev_progress = progress;
361 qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
362 | BLK_MIG_FLAG_PROGRESS);
363 monitor_printf(mon, "Completed %d %%\r", progress);
364 monitor_flush(mon);
365 }
366
367 return ret;
368 }
369
370 static void blk_mig_reset_dirty_cursor(void)
371 {
372 BlkMigDevState *bmds;
373
374 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
375 bmds->cur_dirty = 0;
376 }
377 }
378
379 static int mig_save_device_dirty(Monitor *mon, QEMUFile *f,
380 BlkMigDevState *bmds, int is_async)
381 {
382 BlkMigBlock *blk;
383 int64_t total_sectors = bmds->total_sectors;
384 int64_t sector;
385 int nr_sectors;
386
387 for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
388 if (bmds_aio_inflight(bmds, sector))
389 qemu_aio_flush();
390 if (bdrv_get_dirty(bmds->bs, sector)) {
391
392 if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
393 nr_sectors = total_sectors - sector;
394 } else {
395 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
396 }
397 blk = qemu_malloc(sizeof(BlkMigBlock));
398 blk->buf = qemu_malloc(BLOCK_SIZE);
399 blk->bmds = bmds;
400 blk->sector = sector;
401 blk->nr_sectors = nr_sectors;
402
403 if (is_async) {
404 blk->iov.iov_base = blk->buf;
405 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
406 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
407
408 blk->time = qemu_get_clock_ns(rt_clock);
409
410 blk->aiocb = bdrv_aio_readv(bmds->bs, sector, &blk->qiov,
411 nr_sectors, blk_mig_read_cb, blk);
412 if (!blk->aiocb) {
413 goto error;
414 }
415 block_mig_state.submitted++;
416 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
417 } else {
418 if (bdrv_read(bmds->bs, sector, blk->buf,
419 nr_sectors) < 0) {
420 goto error;
421 }
422 blk_send(f, blk);
423
424 qemu_free(blk->buf);
425 qemu_free(blk);
426 }
427
428 bdrv_reset_dirty(bmds->bs, sector, nr_sectors);
429 break;
430 }
431 sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
432 bmds->cur_dirty = sector;
433 }
434
435 return (bmds->cur_dirty >= bmds->total_sectors);
436
437 error:
438 monitor_printf(mon, "Error reading sector %" PRId64 "\n", sector);
439 qemu_file_set_error(f);
440 qemu_free(blk->buf);
441 qemu_free(blk);
442 return 0;
443 }
444
445 static int blk_mig_save_dirty_block(Monitor *mon, QEMUFile *f, int is_async)
446 {
447 BlkMigDevState *bmds;
448 int ret = 0;
449
450 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
451 if (mig_save_device_dirty(mon, f, bmds, is_async) == 0) {
452 ret = 1;
453 break;
454 }
455 }
456
457 return ret;
458 }
459
460 static void flush_blks(QEMUFile* f)
461 {
462 BlkMigBlock *blk;
463
464 DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
465 __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
466 block_mig_state.transferred);
467
468 while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
469 if (qemu_file_rate_limit(f)) {
470 break;
471 }
472 if (blk->ret < 0) {
473 qemu_file_set_error(f);
474 break;
475 }
476 blk_send(f, blk);
477
478 QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
479 qemu_free(blk->buf);
480 qemu_free(blk);
481
482 block_mig_state.read_done--;
483 block_mig_state.transferred++;
484 assert(block_mig_state.read_done >= 0);
485 }
486
487 DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
488 block_mig_state.submitted, block_mig_state.read_done,
489 block_mig_state.transferred);
490 }
491
492 static int64_t get_remaining_dirty(void)
493 {
494 BlkMigDevState *bmds;
495 int64_t dirty = 0;
496
497 QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
498 dirty += bdrv_get_dirty_count(bmds->bs);
499 }
500
501 return dirty * BLOCK_SIZE;
502 }
503
504 static int is_stage2_completed(void)
505 {
506 int64_t remaining_dirty;
507 long double bwidth;
508
509 if (block_mig_state.bulk_completed == 1) {
510
511 remaining_dirty = get_remaining_dirty();
512 if (remaining_dirty == 0) {
513 return 1;
514 }
515
516 bwidth = compute_read_bwidth();
517
518 if ((remaining_dirty / bwidth) <=
519 migrate_max_downtime()) {
520 /* finish stage2 because we think that we can finish remaing work
521 below max_downtime */
522
523 return 1;
524 }
525 }
526
527 return 0;
528 }
529
530 static void blk_mig_cleanup(Monitor *mon)
531 {
532 BlkMigDevState *bmds;
533 BlkMigBlock *blk;
534
535 while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
536 QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
537 qemu_free(bmds->aio_bitmap);
538 qemu_free(bmds);
539 }
540
541 while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
542 QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
543 qemu_free(blk->buf);
544 qemu_free(blk);
545 }
546
547 set_dirty_tracking(0);
548
549 monitor_printf(mon, "\n");
550 }
551
552 static int block_save_live(Monitor *mon, QEMUFile *f, int stage, void *opaque)
553 {
554 DPRINTF("Enter save live stage %d submitted %d transferred %d\n",
555 stage, block_mig_state.submitted, block_mig_state.transferred);
556
557 if (stage < 0) {
558 blk_mig_cleanup(mon);
559 return 0;
560 }
561
562 if (block_mig_state.blk_enable != 1) {
563 /* no need to migrate storage */
564 qemu_put_be64(f, BLK_MIG_FLAG_EOS);
565 return 1;
566 }
567
568 if (stage == 1) {
569 init_blk_migration(mon, f);
570
571 /* start track dirty blocks */
572 set_dirty_tracking(1);
573 }
574
575 flush_blks(f);
576
577 if (qemu_file_has_error(f)) {
578 blk_mig_cleanup(mon);
579 return 0;
580 }
581
582 blk_mig_reset_dirty_cursor();
583
584 if (stage == 2) {
585 /* control the rate of transfer */
586 while ((block_mig_state.submitted +
587 block_mig_state.read_done) * BLOCK_SIZE <
588 qemu_file_get_rate_limit(f)) {
589 if (block_mig_state.bulk_completed == 0) {
590 /* first finish the bulk phase */
591 if (blk_mig_save_bulked_block(mon, f) == 0) {
592 /* finished saving bulk on all devices */
593 block_mig_state.bulk_completed = 1;
594 }
595 } else {
596 if (blk_mig_save_dirty_block(mon, f, 1) == 0) {
597 /* no more dirty blocks */
598 break;
599 }
600 }
601 }
602
603 flush_blks(f);
604
605 if (qemu_file_has_error(f)) {
606 blk_mig_cleanup(mon);
607 return 0;
608 }
609 }
610
611 if (stage == 3) {
612 /* we know for sure that save bulk is completed and
613 all async read completed */
614 assert(block_mig_state.submitted == 0);
615
616 while (blk_mig_save_dirty_block(mon, f, 0) != 0);
617 blk_mig_cleanup(mon);
618
619 /* report completion */
620 qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
621
622 if (qemu_file_has_error(f)) {
623 return 0;
624 }
625
626 monitor_printf(mon, "Block migration completed\n");
627 }
628
629 qemu_put_be64(f, BLK_MIG_FLAG_EOS);
630
631 return ((stage == 2) && is_stage2_completed());
632 }
633
634 static int block_load(QEMUFile *f, void *opaque, int version_id)
635 {
636 static int banner_printed;
637 int len, flags;
638 char device_name[256];
639 int64_t addr;
640 BlockDriverState *bs;
641 uint8_t *buf;
642
643 do {
644 addr = qemu_get_be64(f);
645
646 flags = addr & ~BDRV_SECTOR_MASK;
647 addr >>= BDRV_SECTOR_BITS;
648
649 if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
650 int ret;
651 /* get device name */
652 len = qemu_get_byte(f);
653 qemu_get_buffer(f, (uint8_t *)device_name, len);
654 device_name[len] = '\0';
655
656 bs = bdrv_find(device_name);
657 if (!bs) {
658 fprintf(stderr, "Error unknown block device %s\n",
659 device_name);
660 return -EINVAL;
661 }
662
663 buf = qemu_malloc(BLOCK_SIZE);
664
665 qemu_get_buffer(f, buf, BLOCK_SIZE);
666 ret = bdrv_write(bs, addr, buf, BDRV_SECTORS_PER_DIRTY_CHUNK);
667
668 qemu_free(buf);
669 if (ret < 0) {
670 return ret;
671 }
672 } else if (flags & BLK_MIG_FLAG_PROGRESS) {
673 if (!banner_printed) {
674 printf("Receiving block device images\n");
675 banner_printed = 1;
676 }
677 printf("Completed %d %%%c", (int)addr,
678 (addr == 100) ? '\n' : '\r');
679 fflush(stdout);
680 } else if (!(flags & BLK_MIG_FLAG_EOS)) {
681 fprintf(stderr, "Unknown flags\n");
682 return -EINVAL;
683 }
684 if (qemu_file_has_error(f)) {
685 return -EIO;
686 }
687 } while (!(flags & BLK_MIG_FLAG_EOS));
688
689 return 0;
690 }
691
692 static void block_set_params(int blk_enable, int shared_base, void *opaque)
693 {
694 block_mig_state.blk_enable = blk_enable;
695 block_mig_state.shared_base = shared_base;
696
697 /* shared base means that blk_enable = 1 */
698 block_mig_state.blk_enable |= shared_base;
699 }
700
701 void blk_mig_init(void)
702 {
703 QSIMPLEQ_INIT(&block_mig_state.bmds_list);
704 QSIMPLEQ_INIT(&block_mig_state.blk_list);
705
706 register_savevm_live(NULL, "block", 0, 1, block_set_params,
707 block_save_live, NULL, block_load, &block_mig_state);
708 }