Add Migration maintainer
[qemu.git] / migration.c
1 /*
2 * QEMU live migration
3 *
4 * Copyright IBM, Corp. 2008
5 *
6 * Authors:
7 * Anthony Liguori <aliguori@us.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 * Contributions after 2012-01-13 are licensed under the terms of the
13 * GNU GPL, version 2 or (at your option) any later version.
14 */
15
16 #include "qemu-common.h"
17 #include "qemu/main-loop.h"
18 #include "migration/migration.h"
19 #include "monitor/monitor.h"
20 #include "migration/qemu-file.h"
21 #include "sysemu/sysemu.h"
22 #include "block/block.h"
23 #include "qemu/sockets.h"
24 #include "migration/block.h"
25 #include "qemu/thread.h"
26 #include "qmp-commands.h"
27 #include "trace.h"
28
29 enum {
30 MIG_STATE_ERROR = -1,
31 MIG_STATE_NONE,
32 MIG_STATE_SETUP,
33 MIG_STATE_CANCELLING,
34 MIG_STATE_CANCELLED,
35 MIG_STATE_ACTIVE,
36 MIG_STATE_COMPLETED,
37 };
38
39 #define MAX_THROTTLE (32 << 20) /* Migration speed throttling */
40
41 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
42 * data. */
43 #define BUFFER_DELAY 100
44 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
45
46 /* Migration XBZRLE default cache size */
47 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
48
49 static NotifierList migration_state_notifiers =
50 NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
51
52 /* When we add fault tolerance, we could have several
53 migrations at once. For now we don't need to add
54 dynamic creation of migration */
55
56 MigrationState *migrate_get_current(void)
57 {
58 static MigrationState current_migration = {
59 .state = MIG_STATE_NONE,
60 .bandwidth_limit = MAX_THROTTLE,
61 .xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE,
62 .mbps = -1,
63 };
64
65 return &current_migration;
66 }
67
68 void qemu_start_incoming_migration(const char *uri, Error **errp)
69 {
70 const char *p;
71
72 if (strstart(uri, "tcp:", &p))
73 tcp_start_incoming_migration(p, errp);
74 #ifdef CONFIG_RDMA
75 else if (strstart(uri, "rdma:", &p))
76 rdma_start_incoming_migration(p, errp);
77 #endif
78 #if !defined(WIN32)
79 else if (strstart(uri, "exec:", &p))
80 exec_start_incoming_migration(p, errp);
81 else if (strstart(uri, "unix:", &p))
82 unix_start_incoming_migration(p, errp);
83 else if (strstart(uri, "fd:", &p))
84 fd_start_incoming_migration(p, errp);
85 #endif
86 else {
87 error_setg(errp, "unknown migration protocol: %s", uri);
88 }
89 }
90
91 static void process_incoming_migration_co(void *opaque)
92 {
93 QEMUFile *f = opaque;
94 Error *local_err = NULL;
95 int ret;
96
97 ret = qemu_loadvm_state(f);
98 qemu_fclose(f);
99 free_xbzrle_decoded_buf();
100 if (ret < 0) {
101 error_report("load of migration failed: %s", strerror(-ret));
102 exit(EXIT_FAILURE);
103 }
104 qemu_announce_self();
105
106 /* Make sure all file formats flush their mutable metadata */
107 bdrv_invalidate_cache_all(&local_err);
108 if (local_err) {
109 qerror_report_err(local_err);
110 error_free(local_err);
111 exit(EXIT_FAILURE);
112 }
113
114 if (autostart) {
115 vm_start();
116 } else {
117 runstate_set(RUN_STATE_PAUSED);
118 }
119 }
120
121 void process_incoming_migration(QEMUFile *f)
122 {
123 Coroutine *co = qemu_coroutine_create(process_incoming_migration_co);
124 int fd = qemu_get_fd(f);
125
126 assert(fd != -1);
127 qemu_set_nonblock(fd);
128 qemu_coroutine_enter(co, f);
129 }
130
131 /* amount of nanoseconds we are willing to wait for migration to be down.
132 * the choice of nanoseconds is because it is the maximum resolution that
133 * get_clock() can achieve. It is an internal measure. All user-visible
134 * units must be in seconds */
135 static uint64_t max_downtime = 300000000;
136
137 uint64_t migrate_max_downtime(void)
138 {
139 return max_downtime;
140 }
141
142 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
143 {
144 MigrationCapabilityStatusList *head = NULL;
145 MigrationCapabilityStatusList *caps;
146 MigrationState *s = migrate_get_current();
147 int i;
148
149 caps = NULL; /* silence compiler warning */
150 for (i = 0; i < MIGRATION_CAPABILITY_MAX; i++) {
151 if (head == NULL) {
152 head = g_malloc0(sizeof(*caps));
153 caps = head;
154 } else {
155 caps->next = g_malloc0(sizeof(*caps));
156 caps = caps->next;
157 }
158 caps->value =
159 g_malloc(sizeof(*caps->value));
160 caps->value->capability = i;
161 caps->value->state = s->enabled_capabilities[i];
162 }
163
164 return head;
165 }
166
167 static void get_xbzrle_cache_stats(MigrationInfo *info)
168 {
169 if (migrate_use_xbzrle()) {
170 info->has_xbzrle_cache = true;
171 info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
172 info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
173 info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred();
174 info->xbzrle_cache->pages = xbzrle_mig_pages_transferred();
175 info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss();
176 info->xbzrle_cache->cache_miss_rate = xbzrle_mig_cache_miss_rate();
177 info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow();
178 }
179 }
180
181 MigrationInfo *qmp_query_migrate(Error **errp)
182 {
183 MigrationInfo *info = g_malloc0(sizeof(*info));
184 MigrationState *s = migrate_get_current();
185
186 switch (s->state) {
187 case MIG_STATE_NONE:
188 /* no migration has happened ever */
189 break;
190 case MIG_STATE_SETUP:
191 info->has_status = true;
192 info->status = g_strdup("setup");
193 info->has_total_time = false;
194 break;
195 case MIG_STATE_ACTIVE:
196 case MIG_STATE_CANCELLING:
197 info->has_status = true;
198 info->status = g_strdup("active");
199 info->has_total_time = true;
200 info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
201 - s->total_time;
202 info->has_expected_downtime = true;
203 info->expected_downtime = s->expected_downtime;
204 info->has_setup_time = true;
205 info->setup_time = s->setup_time;
206
207 info->has_ram = true;
208 info->ram = g_malloc0(sizeof(*info->ram));
209 info->ram->transferred = ram_bytes_transferred();
210 info->ram->remaining = ram_bytes_remaining();
211 info->ram->total = ram_bytes_total();
212 info->ram->duplicate = dup_mig_pages_transferred();
213 info->ram->skipped = skipped_mig_pages_transferred();
214 info->ram->normal = norm_mig_pages_transferred();
215 info->ram->normal_bytes = norm_mig_bytes_transferred();
216 info->ram->dirty_pages_rate = s->dirty_pages_rate;
217 info->ram->mbps = s->mbps;
218 info->ram->dirty_sync_count = s->dirty_sync_count;
219
220 if (blk_mig_active()) {
221 info->has_disk = true;
222 info->disk = g_malloc0(sizeof(*info->disk));
223 info->disk->transferred = blk_mig_bytes_transferred();
224 info->disk->remaining = blk_mig_bytes_remaining();
225 info->disk->total = blk_mig_bytes_total();
226 }
227
228 get_xbzrle_cache_stats(info);
229 break;
230 case MIG_STATE_COMPLETED:
231 get_xbzrle_cache_stats(info);
232
233 info->has_status = true;
234 info->status = g_strdup("completed");
235 info->has_total_time = true;
236 info->total_time = s->total_time;
237 info->has_downtime = true;
238 info->downtime = s->downtime;
239 info->has_setup_time = true;
240 info->setup_time = s->setup_time;
241
242 info->has_ram = true;
243 info->ram = g_malloc0(sizeof(*info->ram));
244 info->ram->transferred = ram_bytes_transferred();
245 info->ram->remaining = 0;
246 info->ram->total = ram_bytes_total();
247 info->ram->duplicate = dup_mig_pages_transferred();
248 info->ram->skipped = skipped_mig_pages_transferred();
249 info->ram->normal = norm_mig_pages_transferred();
250 info->ram->normal_bytes = norm_mig_bytes_transferred();
251 info->ram->mbps = s->mbps;
252 info->ram->dirty_sync_count = s->dirty_sync_count;
253 break;
254 case MIG_STATE_ERROR:
255 info->has_status = true;
256 info->status = g_strdup("failed");
257 break;
258 case MIG_STATE_CANCELLED:
259 info->has_status = true;
260 info->status = g_strdup("cancelled");
261 break;
262 }
263
264 return info;
265 }
266
267 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
268 Error **errp)
269 {
270 MigrationState *s = migrate_get_current();
271 MigrationCapabilityStatusList *cap;
272
273 if (s->state == MIG_STATE_ACTIVE || s->state == MIG_STATE_SETUP) {
274 error_set(errp, QERR_MIGRATION_ACTIVE);
275 return;
276 }
277
278 for (cap = params; cap; cap = cap->next) {
279 s->enabled_capabilities[cap->value->capability] = cap->value->state;
280 }
281 }
282
283 /* shared migration helpers */
284
285 static void migrate_set_state(MigrationState *s, int old_state, int new_state)
286 {
287 if (atomic_cmpxchg(&s->state, old_state, new_state) == new_state) {
288 trace_migrate_set_state(new_state);
289 }
290 }
291
292 static void migrate_fd_cleanup(void *opaque)
293 {
294 MigrationState *s = opaque;
295
296 qemu_bh_delete(s->cleanup_bh);
297 s->cleanup_bh = NULL;
298
299 if (s->file) {
300 trace_migrate_fd_cleanup();
301 qemu_mutex_unlock_iothread();
302 qemu_thread_join(&s->thread);
303 qemu_mutex_lock_iothread();
304
305 qemu_fclose(s->file);
306 s->file = NULL;
307 }
308
309 assert(s->state != MIG_STATE_ACTIVE);
310
311 if (s->state != MIG_STATE_COMPLETED) {
312 qemu_savevm_state_cancel();
313 if (s->state == MIG_STATE_CANCELLING) {
314 migrate_set_state(s, MIG_STATE_CANCELLING, MIG_STATE_CANCELLED);
315 }
316 }
317
318 notifier_list_notify(&migration_state_notifiers, s);
319 }
320
321 void migrate_fd_error(MigrationState *s)
322 {
323 trace_migrate_fd_error();
324 assert(s->file == NULL);
325 s->state = MIG_STATE_ERROR;
326 trace_migrate_set_state(MIG_STATE_ERROR);
327 notifier_list_notify(&migration_state_notifiers, s);
328 }
329
330 static void migrate_fd_cancel(MigrationState *s)
331 {
332 int old_state ;
333 trace_migrate_fd_cancel();
334
335 do {
336 old_state = s->state;
337 if (old_state != MIG_STATE_SETUP && old_state != MIG_STATE_ACTIVE) {
338 break;
339 }
340 migrate_set_state(s, old_state, MIG_STATE_CANCELLING);
341 } while (s->state != MIG_STATE_CANCELLING);
342 }
343
344 void add_migration_state_change_notifier(Notifier *notify)
345 {
346 notifier_list_add(&migration_state_notifiers, notify);
347 }
348
349 void remove_migration_state_change_notifier(Notifier *notify)
350 {
351 notifier_remove(notify);
352 }
353
354 bool migration_in_setup(MigrationState *s)
355 {
356 return s->state == MIG_STATE_SETUP;
357 }
358
359 bool migration_has_finished(MigrationState *s)
360 {
361 return s->state == MIG_STATE_COMPLETED;
362 }
363
364 bool migration_has_failed(MigrationState *s)
365 {
366 return (s->state == MIG_STATE_CANCELLED ||
367 s->state == MIG_STATE_ERROR);
368 }
369
370 static MigrationState *migrate_init(const MigrationParams *params)
371 {
372 MigrationState *s = migrate_get_current();
373 int64_t bandwidth_limit = s->bandwidth_limit;
374 bool enabled_capabilities[MIGRATION_CAPABILITY_MAX];
375 int64_t xbzrle_cache_size = s->xbzrle_cache_size;
376
377 memcpy(enabled_capabilities, s->enabled_capabilities,
378 sizeof(enabled_capabilities));
379
380 memset(s, 0, sizeof(*s));
381 s->params = *params;
382 memcpy(s->enabled_capabilities, enabled_capabilities,
383 sizeof(enabled_capabilities));
384 s->xbzrle_cache_size = xbzrle_cache_size;
385
386 s->bandwidth_limit = bandwidth_limit;
387 s->state = MIG_STATE_SETUP;
388 trace_migrate_set_state(MIG_STATE_SETUP);
389
390 s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
391 return s;
392 }
393
394 static GSList *migration_blockers;
395
396 void migrate_add_blocker(Error *reason)
397 {
398 migration_blockers = g_slist_prepend(migration_blockers, reason);
399 }
400
401 void migrate_del_blocker(Error *reason)
402 {
403 migration_blockers = g_slist_remove(migration_blockers, reason);
404 }
405
406 void qmp_migrate(const char *uri, bool has_blk, bool blk,
407 bool has_inc, bool inc, bool has_detach, bool detach,
408 Error **errp)
409 {
410 Error *local_err = NULL;
411 MigrationState *s = migrate_get_current();
412 MigrationParams params;
413 const char *p;
414
415 params.blk = has_blk && blk;
416 params.shared = has_inc && inc;
417
418 if (s->state == MIG_STATE_ACTIVE || s->state == MIG_STATE_SETUP ||
419 s->state == MIG_STATE_CANCELLING) {
420 error_set(errp, QERR_MIGRATION_ACTIVE);
421 return;
422 }
423
424 if (runstate_check(RUN_STATE_INMIGRATE)) {
425 error_setg(errp, "Guest is waiting for an incoming migration");
426 return;
427 }
428
429 if (qemu_savevm_state_blocked(errp)) {
430 return;
431 }
432
433 if (migration_blockers) {
434 *errp = error_copy(migration_blockers->data);
435 return;
436 }
437
438 s = migrate_init(&params);
439
440 if (strstart(uri, "tcp:", &p)) {
441 tcp_start_outgoing_migration(s, p, &local_err);
442 #ifdef CONFIG_RDMA
443 } else if (strstart(uri, "rdma:", &p)) {
444 rdma_start_outgoing_migration(s, p, &local_err);
445 #endif
446 #if !defined(WIN32)
447 } else if (strstart(uri, "exec:", &p)) {
448 exec_start_outgoing_migration(s, p, &local_err);
449 } else if (strstart(uri, "unix:", &p)) {
450 unix_start_outgoing_migration(s, p, &local_err);
451 } else if (strstart(uri, "fd:", &p)) {
452 fd_start_outgoing_migration(s, p, &local_err);
453 #endif
454 } else {
455 error_set(errp, QERR_INVALID_PARAMETER_VALUE, "uri", "a valid migration protocol");
456 s->state = MIG_STATE_ERROR;
457 return;
458 }
459
460 if (local_err) {
461 migrate_fd_error(s);
462 error_propagate(errp, local_err);
463 return;
464 }
465 }
466
467 void qmp_migrate_cancel(Error **errp)
468 {
469 migrate_fd_cancel(migrate_get_current());
470 }
471
472 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
473 {
474 MigrationState *s = migrate_get_current();
475 int64_t new_size;
476
477 /* Check for truncation */
478 if (value != (size_t)value) {
479 error_set(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
480 "exceeding address space");
481 return;
482 }
483
484 /* Cache should not be larger than guest ram size */
485 if (value > ram_bytes_total()) {
486 error_set(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
487 "exceeds guest ram size ");
488 return;
489 }
490
491 new_size = xbzrle_cache_resize(value);
492 if (new_size < 0) {
493 error_set(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
494 "is smaller than page size");
495 return;
496 }
497
498 s->xbzrle_cache_size = new_size;
499 }
500
501 int64_t qmp_query_migrate_cache_size(Error **errp)
502 {
503 return migrate_xbzrle_cache_size();
504 }
505
506 void qmp_migrate_set_speed(int64_t value, Error **errp)
507 {
508 MigrationState *s;
509
510 if (value < 0) {
511 value = 0;
512 }
513 if (value > SIZE_MAX) {
514 value = SIZE_MAX;
515 }
516
517 s = migrate_get_current();
518 s->bandwidth_limit = value;
519 if (s->file) {
520 qemu_file_set_rate_limit(s->file, s->bandwidth_limit / XFER_LIMIT_RATIO);
521 }
522 }
523
524 void qmp_migrate_set_downtime(double value, Error **errp)
525 {
526 value *= 1e9;
527 value = MAX(0, MIN(UINT64_MAX, value));
528 max_downtime = (uint64_t)value;
529 }
530
531 bool migrate_rdma_pin_all(void)
532 {
533 MigrationState *s;
534
535 s = migrate_get_current();
536
537 return s->enabled_capabilities[MIGRATION_CAPABILITY_RDMA_PIN_ALL];
538 }
539
540 bool migrate_auto_converge(void)
541 {
542 MigrationState *s;
543
544 s = migrate_get_current();
545
546 return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
547 }
548
549 bool migrate_zero_blocks(void)
550 {
551 MigrationState *s;
552
553 s = migrate_get_current();
554
555 return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
556 }
557
558 int migrate_use_xbzrle(void)
559 {
560 MigrationState *s;
561
562 s = migrate_get_current();
563
564 return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
565 }
566
567 int64_t migrate_xbzrle_cache_size(void)
568 {
569 MigrationState *s;
570
571 s = migrate_get_current();
572
573 return s->xbzrle_cache_size;
574 }
575
576 /* migration thread support */
577
578 static void *migration_thread(void *opaque)
579 {
580 MigrationState *s = opaque;
581 int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
582 int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
583 int64_t initial_bytes = 0;
584 int64_t max_size = 0;
585 int64_t start_time = initial_time;
586 bool old_vm_running = false;
587
588 qemu_savevm_state_begin(s->file, &s->params);
589
590 s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
591 migrate_set_state(s, MIG_STATE_SETUP, MIG_STATE_ACTIVE);
592
593 while (s->state == MIG_STATE_ACTIVE) {
594 int64_t current_time;
595 uint64_t pending_size;
596
597 if (!qemu_file_rate_limit(s->file)) {
598 pending_size = qemu_savevm_state_pending(s->file, max_size);
599 trace_migrate_pending(pending_size, max_size);
600 if (pending_size && pending_size >= max_size) {
601 qemu_savevm_state_iterate(s->file);
602 } else {
603 int ret;
604
605 qemu_mutex_lock_iothread();
606 start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
607 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
608 old_vm_running = runstate_is_running();
609
610 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
611 if (ret >= 0) {
612 qemu_file_set_rate_limit(s->file, INT64_MAX);
613 qemu_savevm_state_complete(s->file);
614 }
615 qemu_mutex_unlock_iothread();
616
617 if (ret < 0) {
618 migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_ERROR);
619 break;
620 }
621
622 if (!qemu_file_get_error(s->file)) {
623 migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_COMPLETED);
624 break;
625 }
626 }
627 }
628
629 if (qemu_file_get_error(s->file)) {
630 migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_ERROR);
631 break;
632 }
633 current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
634 if (current_time >= initial_time + BUFFER_DELAY) {
635 uint64_t transferred_bytes = qemu_ftell(s->file) - initial_bytes;
636 uint64_t time_spent = current_time - initial_time;
637 double bandwidth = transferred_bytes / time_spent;
638 max_size = bandwidth * migrate_max_downtime() / 1000000;
639
640 s->mbps = time_spent ? (((double) transferred_bytes * 8.0) /
641 ((double) time_spent / 1000.0)) / 1000.0 / 1000.0 : -1;
642
643 trace_migrate_transferred(transferred_bytes, time_spent,
644 bandwidth, max_size);
645 /* if we haven't sent anything, we don't want to recalculate
646 10000 is a small enough number for our purposes */
647 if (s->dirty_bytes_rate && transferred_bytes > 10000) {
648 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
649 }
650
651 qemu_file_reset_rate_limit(s->file);
652 initial_time = current_time;
653 initial_bytes = qemu_ftell(s->file);
654 }
655 if (qemu_file_rate_limit(s->file)) {
656 /* usleep expects microseconds */
657 g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
658 }
659 }
660
661 qemu_mutex_lock_iothread();
662 if (s->state == MIG_STATE_COMPLETED) {
663 int64_t end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
664 uint64_t transferred_bytes = qemu_ftell(s->file);
665 s->total_time = end_time - s->total_time;
666 s->downtime = end_time - start_time;
667 if (s->total_time) {
668 s->mbps = (((double) transferred_bytes * 8.0) /
669 ((double) s->total_time)) / 1000;
670 }
671 runstate_set(RUN_STATE_POSTMIGRATE);
672 } else {
673 if (old_vm_running) {
674 vm_start();
675 }
676 }
677 qemu_bh_schedule(s->cleanup_bh);
678 qemu_mutex_unlock_iothread();
679
680 return NULL;
681 }
682
683 void migrate_fd_connect(MigrationState *s)
684 {
685 s->state = MIG_STATE_SETUP;
686 trace_migrate_set_state(MIG_STATE_SETUP);
687
688 /* This is a best 1st approximation. ns to ms */
689 s->expected_downtime = max_downtime/1000000;
690 s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
691
692 qemu_file_set_rate_limit(s->file,
693 s->bandwidth_limit / XFER_LIMIT_RATIO);
694
695 /* Notify before starting migration thread */
696 notifier_list_notify(&migration_state_notifiers, s);
697
698 qemu_thread_create(&s->thread, "migration", migration_thread, s,
699 QEMU_THREAD_JOINABLE);
700 }