4 * Copyright IBM, Corp. 2008
7 * Anthony Liguori <aliguori@us.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 "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"
39 #define MAX_THROTTLE (32 << 20) /* Migration speed throttling */
41 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
43 #define BUFFER_DELAY 100
44 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
46 /* Migration XBZRLE default cache size */
47 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
49 static NotifierList migration_state_notifiers
=
50 NOTIFIER_LIST_INITIALIZER(migration_state_notifiers
);
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 */
56 MigrationState
*migrate_get_current(void)
58 static MigrationState current_migration
= {
59 .state
= MIG_STATE_NONE
,
60 .bandwidth_limit
= MAX_THROTTLE
,
61 .xbzrle_cache_size
= DEFAULT_MIGRATE_CACHE_SIZE
,
65 return ¤t_migration
;
68 void qemu_start_incoming_migration(const char *uri
, Error
**errp
)
72 if (strstart(uri
, "tcp:", &p
))
73 tcp_start_incoming_migration(p
, errp
);
75 else if (strstart(uri
, "rdma:", &p
))
76 rdma_start_incoming_migration(p
, errp
);
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
);
87 error_setg(errp
, "unknown migration protocol: %s", uri
);
91 static void process_incoming_migration_co(void *opaque
)
94 Error
*local_err
= NULL
;
97 ret
= qemu_loadvm_state(f
);
99 free_xbzrle_decoded_buf();
101 error_report("load of migration failed: %s", strerror(-ret
));
104 qemu_announce_self();
106 /* Make sure all file formats flush their mutable metadata */
107 bdrv_invalidate_cache_all(&local_err
);
109 qerror_report_err(local_err
);
110 error_free(local_err
);
117 runstate_set(RUN_STATE_PAUSED
);
121 void process_incoming_migration(QEMUFile
*f
)
123 Coroutine
*co
= qemu_coroutine_create(process_incoming_migration_co
);
124 int fd
= qemu_get_fd(f
);
127 qemu_set_nonblock(fd
);
128 qemu_coroutine_enter(co
, f
);
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;
137 uint64_t migrate_max_downtime(void)
142 MigrationCapabilityStatusList
*qmp_query_migrate_capabilities(Error
**errp
)
144 MigrationCapabilityStatusList
*head
= NULL
;
145 MigrationCapabilityStatusList
*caps
;
146 MigrationState
*s
= migrate_get_current();
149 caps
= NULL
; /* silence compiler warning */
150 for (i
= 0; i
< MIGRATION_CAPABILITY_MAX
; i
++) {
152 head
= g_malloc0(sizeof(*caps
));
155 caps
->next
= g_malloc0(sizeof(*caps
));
159 g_malloc(sizeof(*caps
->value
));
160 caps
->value
->capability
= i
;
161 caps
->value
->state
= s
->enabled_capabilities
[i
];
167 static void get_xbzrle_cache_stats(MigrationInfo
*info
)
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();
181 MigrationInfo
*qmp_query_migrate(Error
**errp
)
183 MigrationInfo
*info
= g_malloc0(sizeof(*info
));
184 MigrationState
*s
= migrate_get_current();
188 /* no migration has happened ever */
190 case MIG_STATE_SETUP
:
191 info
->has_status
= true;
192 info
->status
= g_strdup("setup");
193 info
->has_total_time
= false;
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
)
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
;
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
;
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();
228 get_xbzrle_cache_stats(info
);
230 case MIG_STATE_COMPLETED
:
231 get_xbzrle_cache_stats(info
);
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
;
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
;
254 case MIG_STATE_ERROR
:
255 info
->has_status
= true;
256 info
->status
= g_strdup("failed");
258 case MIG_STATE_CANCELLED
:
259 info
->has_status
= true;
260 info
->status
= g_strdup("cancelled");
267 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList
*params
,
270 MigrationState
*s
= migrate_get_current();
271 MigrationCapabilityStatusList
*cap
;
273 if (s
->state
== MIG_STATE_ACTIVE
|| s
->state
== MIG_STATE_SETUP
) {
274 error_set(errp
, QERR_MIGRATION_ACTIVE
);
278 for (cap
= params
; cap
; cap
= cap
->next
) {
279 s
->enabled_capabilities
[cap
->value
->capability
] = cap
->value
->state
;
283 /* shared migration helpers */
285 static void migrate_set_state(MigrationState
*s
, int old_state
, int new_state
)
287 if (atomic_cmpxchg(&s
->state
, old_state
, new_state
) == new_state
) {
288 trace_migrate_set_state(new_state
);
292 static void migrate_fd_cleanup(void *opaque
)
294 MigrationState
*s
= opaque
;
296 qemu_bh_delete(s
->cleanup_bh
);
297 s
->cleanup_bh
= NULL
;
300 trace_migrate_fd_cleanup();
301 qemu_mutex_unlock_iothread();
302 qemu_thread_join(&s
->thread
);
303 qemu_mutex_lock_iothread();
305 qemu_fclose(s
->file
);
309 assert(s
->state
!= MIG_STATE_ACTIVE
);
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
);
318 notifier_list_notify(&migration_state_notifiers
, s
);
321 void migrate_fd_error(MigrationState
*s
)
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
);
330 static void migrate_fd_cancel(MigrationState
*s
)
333 QEMUFile
*f
= migrate_get_current()->file
;
334 trace_migrate_fd_cancel();
337 old_state
= s
->state
;
338 if (old_state
!= MIG_STATE_SETUP
&& old_state
!= MIG_STATE_ACTIVE
) {
341 migrate_set_state(s
, old_state
, MIG_STATE_CANCELLING
);
342 } while (s
->state
!= MIG_STATE_CANCELLING
);
345 * If we're unlucky the migration code might be stuck somewhere in a
346 * send/write while the network has failed and is waiting to timeout;
347 * if we've got shutdown(2) available then we can force it to quit.
348 * The outgoing qemu file gets closed in migrate_fd_cleanup that is
349 * called in a bh, so there is no race against this cancel.
351 if (s
->state
== MIG_STATE_CANCELLING
&& f
) {
352 qemu_file_shutdown(f
);
356 void add_migration_state_change_notifier(Notifier
*notify
)
358 notifier_list_add(&migration_state_notifiers
, notify
);
361 void remove_migration_state_change_notifier(Notifier
*notify
)
363 notifier_remove(notify
);
366 bool migration_in_setup(MigrationState
*s
)
368 return s
->state
== MIG_STATE_SETUP
;
371 bool migration_has_finished(MigrationState
*s
)
373 return s
->state
== MIG_STATE_COMPLETED
;
376 bool migration_has_failed(MigrationState
*s
)
378 return (s
->state
== MIG_STATE_CANCELLED
||
379 s
->state
== MIG_STATE_ERROR
);
382 static MigrationState
*migrate_init(const MigrationParams
*params
)
384 MigrationState
*s
= migrate_get_current();
385 int64_t bandwidth_limit
= s
->bandwidth_limit
;
386 bool enabled_capabilities
[MIGRATION_CAPABILITY_MAX
];
387 int64_t xbzrle_cache_size
= s
->xbzrle_cache_size
;
389 memcpy(enabled_capabilities
, s
->enabled_capabilities
,
390 sizeof(enabled_capabilities
));
392 memset(s
, 0, sizeof(*s
));
394 memcpy(s
->enabled_capabilities
, enabled_capabilities
,
395 sizeof(enabled_capabilities
));
396 s
->xbzrle_cache_size
= xbzrle_cache_size
;
398 s
->bandwidth_limit
= bandwidth_limit
;
399 s
->state
= MIG_STATE_SETUP
;
400 trace_migrate_set_state(MIG_STATE_SETUP
);
402 s
->total_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
406 static GSList
*migration_blockers
;
408 void migrate_add_blocker(Error
*reason
)
410 migration_blockers
= g_slist_prepend(migration_blockers
, reason
);
413 void migrate_del_blocker(Error
*reason
)
415 migration_blockers
= g_slist_remove(migration_blockers
, reason
);
418 void qmp_migrate(const char *uri
, bool has_blk
, bool blk
,
419 bool has_inc
, bool inc
, bool has_detach
, bool detach
,
422 Error
*local_err
= NULL
;
423 MigrationState
*s
= migrate_get_current();
424 MigrationParams params
;
427 params
.blk
= has_blk
&& blk
;
428 params
.shared
= has_inc
&& inc
;
430 if (s
->state
== MIG_STATE_ACTIVE
|| s
->state
== MIG_STATE_SETUP
||
431 s
->state
== MIG_STATE_CANCELLING
) {
432 error_set(errp
, QERR_MIGRATION_ACTIVE
);
436 if (runstate_check(RUN_STATE_INMIGRATE
)) {
437 error_setg(errp
, "Guest is waiting for an incoming migration");
441 if (qemu_savevm_state_blocked(errp
)) {
445 if (migration_blockers
) {
446 *errp
= error_copy(migration_blockers
->data
);
450 s
= migrate_init(¶ms
);
452 if (strstart(uri
, "tcp:", &p
)) {
453 tcp_start_outgoing_migration(s
, p
, &local_err
);
455 } else if (strstart(uri
, "rdma:", &p
)) {
456 rdma_start_outgoing_migration(s
, p
, &local_err
);
459 } else if (strstart(uri
, "exec:", &p
)) {
460 exec_start_outgoing_migration(s
, p
, &local_err
);
461 } else if (strstart(uri
, "unix:", &p
)) {
462 unix_start_outgoing_migration(s
, p
, &local_err
);
463 } else if (strstart(uri
, "fd:", &p
)) {
464 fd_start_outgoing_migration(s
, p
, &local_err
);
467 error_set(errp
, QERR_INVALID_PARAMETER_VALUE
, "uri", "a valid migration protocol");
468 s
->state
= MIG_STATE_ERROR
;
474 error_propagate(errp
, local_err
);
479 void qmp_migrate_cancel(Error
**errp
)
481 migrate_fd_cancel(migrate_get_current());
484 void qmp_migrate_set_cache_size(int64_t value
, Error
**errp
)
486 MigrationState
*s
= migrate_get_current();
489 /* Check for truncation */
490 if (value
!= (size_t)value
) {
491 error_set(errp
, QERR_INVALID_PARAMETER_VALUE
, "cache size",
492 "exceeding address space");
496 /* Cache should not be larger than guest ram size */
497 if (value
> ram_bytes_total()) {
498 error_set(errp
, QERR_INVALID_PARAMETER_VALUE
, "cache size",
499 "exceeds guest ram size ");
503 new_size
= xbzrle_cache_resize(value
);
505 error_set(errp
, QERR_INVALID_PARAMETER_VALUE
, "cache size",
506 "is smaller than page size");
510 s
->xbzrle_cache_size
= new_size
;
513 int64_t qmp_query_migrate_cache_size(Error
**errp
)
515 return migrate_xbzrle_cache_size();
518 void qmp_migrate_set_speed(int64_t value
, Error
**errp
)
525 if (value
> SIZE_MAX
) {
529 s
= migrate_get_current();
530 s
->bandwidth_limit
= value
;
532 qemu_file_set_rate_limit(s
->file
, s
->bandwidth_limit
/ XFER_LIMIT_RATIO
);
536 void qmp_migrate_set_downtime(double value
, Error
**errp
)
539 value
= MAX(0, MIN(UINT64_MAX
, value
));
540 max_downtime
= (uint64_t)value
;
543 bool migrate_rdma_pin_all(void)
547 s
= migrate_get_current();
549 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
];
552 bool migrate_auto_converge(void)
556 s
= migrate_get_current();
558 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_AUTO_CONVERGE
];
561 bool migrate_zero_blocks(void)
565 s
= migrate_get_current();
567 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_ZERO_BLOCKS
];
570 int migrate_use_xbzrle(void)
574 s
= migrate_get_current();
576 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_XBZRLE
];
579 int64_t migrate_xbzrle_cache_size(void)
583 s
= migrate_get_current();
585 return s
->xbzrle_cache_size
;
588 /* migration thread support */
590 static void *migration_thread(void *opaque
)
592 MigrationState
*s
= opaque
;
593 int64_t initial_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
594 int64_t setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
595 int64_t initial_bytes
= 0;
596 int64_t max_size
= 0;
597 int64_t start_time
= initial_time
;
598 bool old_vm_running
= false;
600 qemu_savevm_state_begin(s
->file
, &s
->params
);
602 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
603 migrate_set_state(s
, MIG_STATE_SETUP
, MIG_STATE_ACTIVE
);
605 while (s
->state
== MIG_STATE_ACTIVE
) {
606 int64_t current_time
;
607 uint64_t pending_size
;
609 if (!qemu_file_rate_limit(s
->file
)) {
610 pending_size
= qemu_savevm_state_pending(s
->file
, max_size
);
611 trace_migrate_pending(pending_size
, max_size
);
612 if (pending_size
&& pending_size
>= max_size
) {
613 qemu_savevm_state_iterate(s
->file
);
617 qemu_mutex_lock_iothread();
618 start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
619 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER
);
620 old_vm_running
= runstate_is_running();
622 ret
= vm_stop_force_state(RUN_STATE_FINISH_MIGRATE
);
624 qemu_file_set_rate_limit(s
->file
, INT64_MAX
);
625 qemu_savevm_state_complete(s
->file
);
627 qemu_mutex_unlock_iothread();
630 migrate_set_state(s
, MIG_STATE_ACTIVE
, MIG_STATE_ERROR
);
634 if (!qemu_file_get_error(s
->file
)) {
635 migrate_set_state(s
, MIG_STATE_ACTIVE
, MIG_STATE_COMPLETED
);
641 if (qemu_file_get_error(s
->file
)) {
642 migrate_set_state(s
, MIG_STATE_ACTIVE
, MIG_STATE_ERROR
);
645 current_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
646 if (current_time
>= initial_time
+ BUFFER_DELAY
) {
647 uint64_t transferred_bytes
= qemu_ftell(s
->file
) - initial_bytes
;
648 uint64_t time_spent
= current_time
- initial_time
;
649 double bandwidth
= transferred_bytes
/ time_spent
;
650 max_size
= bandwidth
* migrate_max_downtime() / 1000000;
652 s
->mbps
= time_spent
? (((double) transferred_bytes
* 8.0) /
653 ((double) time_spent
/ 1000.0)) / 1000.0 / 1000.0 : -1;
655 trace_migrate_transferred(transferred_bytes
, time_spent
,
656 bandwidth
, max_size
);
657 /* if we haven't sent anything, we don't want to recalculate
658 10000 is a small enough number for our purposes */
659 if (s
->dirty_bytes_rate
&& transferred_bytes
> 10000) {
660 s
->expected_downtime
= s
->dirty_bytes_rate
/ bandwidth
;
663 qemu_file_reset_rate_limit(s
->file
);
664 initial_time
= current_time
;
665 initial_bytes
= qemu_ftell(s
->file
);
667 if (qemu_file_rate_limit(s
->file
)) {
668 /* usleep expects microseconds */
669 g_usleep((initial_time
+ BUFFER_DELAY
- current_time
)*1000);
673 qemu_mutex_lock_iothread();
674 if (s
->state
== MIG_STATE_COMPLETED
) {
675 int64_t end_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
676 uint64_t transferred_bytes
= qemu_ftell(s
->file
);
677 s
->total_time
= end_time
- s
->total_time
;
678 s
->downtime
= end_time
- start_time
;
680 s
->mbps
= (((double) transferred_bytes
* 8.0) /
681 ((double) s
->total_time
)) / 1000;
683 runstate_set(RUN_STATE_POSTMIGRATE
);
685 if (old_vm_running
) {
689 qemu_bh_schedule(s
->cleanup_bh
);
690 qemu_mutex_unlock_iothread();
695 void migrate_fd_connect(MigrationState
*s
)
697 s
->state
= MIG_STATE_SETUP
;
698 trace_migrate_set_state(MIG_STATE_SETUP
);
700 /* This is a best 1st approximation. ns to ms */
701 s
->expected_downtime
= max_downtime
/1000000;
702 s
->cleanup_bh
= qemu_bh_new(migrate_fd_cleanup
, s
);
704 qemu_file_set_rate_limit(s
->file
,
705 s
->bandwidth_limit
/ XFER_LIMIT_RATIO
);
707 /* Notify before starting migration thread */
708 notifier_list_notify(&migration_state_notifiers
, s
);
710 qemu_thread_create(&s
->thread
, "migration", migration_thread
, s
,
711 QEMU_THREAD_JOINABLE
);