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/osdep.h"
17 #include "qemu/cutils.h"
18 #include "qemu/error-report.h"
19 #include "qemu/main-loop.h"
20 #include "migration/migration.h"
21 #include "migration/qemu-file.h"
22 #include "sysemu/sysemu.h"
23 #include "block/block.h"
24 #include "qapi/qmp/qerror.h"
25 #include "qapi/util.h"
26 #include "qemu/sockets.h"
28 #include "migration/block.h"
29 #include "migration/postcopy-ram.h"
30 #include "qemu/thread.h"
31 #include "qmp-commands.h"
33 #include "qapi-event.h"
35 #include "exec/memory.h"
36 #include "exec/address-spaces.h"
37 #include "io/channel-buffer.h"
38 #include "io/channel-tls.h"
39 #include "migration/colo.h"
41 #define MAX_THROTTLE (32 << 20) /* Migration transfer speed throttling */
43 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
45 #define BUFFER_DELAY 100
46 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
48 /* Time in milliseconds we are allowed to stop the source,
49 * for sending the last part */
50 #define DEFAULT_MIGRATE_SET_DOWNTIME 300
52 /* Default compression thread count */
53 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
54 /* Default decompression thread count, usually decompression is at
55 * least 4 times as fast as compression.*/
56 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2
57 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */
58 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1
59 /* Define default autoconverge cpu throttle migration parameters */
60 #define DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL 20
61 #define DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT 10
63 /* Migration XBZRLE default cache size */
64 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
66 /* The delay time (in ms) between two COLO checkpoints
67 * Note: Please change this default value to 10000 when we support hybrid mode.
69 #define DEFAULT_MIGRATE_X_CHECKPOINT_DELAY 200
71 static NotifierList migration_state_notifiers
=
72 NOTIFIER_LIST_INITIALIZER(migration_state_notifiers
);
74 static bool deferred_incoming
;
77 * Current state of incoming postcopy; note this is not part of
78 * MigrationIncomingState since it's state is used during cleanup
79 * at the end as MIS is being freed.
81 static PostcopyState incoming_postcopy_state
;
83 /* When we add fault tolerance, we could have several
84 migrations at once. For now we don't need to add
85 dynamic creation of migration */
88 MigrationState
*migrate_get_current(void)
91 static MigrationState current_migration
= {
92 .state
= MIGRATION_STATUS_NONE
,
93 .xbzrle_cache_size
= DEFAULT_MIGRATE_CACHE_SIZE
,
96 .compress_level
= DEFAULT_MIGRATE_COMPRESS_LEVEL
,
97 .compress_threads
= DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT
,
98 .decompress_threads
= DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT
,
99 .cpu_throttle_initial
= DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL
,
100 .cpu_throttle_increment
= DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT
,
101 .max_bandwidth
= MAX_THROTTLE
,
102 .downtime_limit
= DEFAULT_MIGRATE_SET_DOWNTIME
,
103 .x_checkpoint_delay
= DEFAULT_MIGRATE_X_CHECKPOINT_DELAY
,
108 qemu_mutex_init(¤t_migration
.src_page_req_mutex
);
111 return ¤t_migration
;
114 MigrationIncomingState
*migration_incoming_get_current(void)
117 static MigrationIncomingState mis_current
;
120 mis_current
.state
= MIGRATION_STATUS_NONE
;
121 memset(&mis_current
, 0, sizeof(MigrationIncomingState
));
122 QLIST_INIT(&mis_current
.loadvm_handlers
);
123 qemu_mutex_init(&mis_current
.rp_mutex
);
124 qemu_event_init(&mis_current
.main_thread_load_event
, false);
130 void migration_incoming_state_destroy(void)
132 struct MigrationIncomingState
*mis
= migration_incoming_get_current();
134 qemu_event_destroy(&mis
->main_thread_load_event
);
135 loadvm_free_handlers(mis
);
142 uint8_t runstate
[100];
147 static GlobalState global_state
;
149 int global_state_store(void)
151 if (!runstate_store((char *)global_state
.runstate
,
152 sizeof(global_state
.runstate
))) {
153 error_report("runstate name too big: %s", global_state
.runstate
);
154 trace_migrate_state_too_big();
160 void global_state_store_running(void)
162 const char *state
= RunState_lookup
[RUN_STATE_RUNNING
];
163 strncpy((char *)global_state
.runstate
,
164 state
, sizeof(global_state
.runstate
));
167 static bool global_state_received(void)
169 return global_state
.received
;
172 static RunState
global_state_get_runstate(void)
174 return global_state
.state
;
177 void global_state_set_optional(void)
179 global_state
.optional
= true;
182 static bool global_state_needed(void *opaque
)
184 GlobalState
*s
= opaque
;
185 char *runstate
= (char *)s
->runstate
;
187 /* If it is not optional, it is mandatory */
189 if (s
->optional
== false) {
193 /* If state is running or paused, it is not needed */
195 if (strcmp(runstate
, "running") == 0 ||
196 strcmp(runstate
, "paused") == 0) {
200 /* for any other state it is needed */
204 static int global_state_post_load(void *opaque
, int version_id
)
206 GlobalState
*s
= opaque
;
207 Error
*local_err
= NULL
;
209 char *runstate
= (char *)s
->runstate
;
212 trace_migrate_global_state_post_load(runstate
);
214 r
= qapi_enum_parse(RunState_lookup
, runstate
, RUN_STATE__MAX
,
219 error_report_err(local_err
);
228 static void global_state_pre_save(void *opaque
)
230 GlobalState
*s
= opaque
;
232 trace_migrate_global_state_pre_save((char *)s
->runstate
);
233 s
->size
= strlen((char *)s
->runstate
) + 1;
236 static const VMStateDescription vmstate_globalstate
= {
237 .name
= "globalstate",
239 .minimum_version_id
= 1,
240 .post_load
= global_state_post_load
,
241 .pre_save
= global_state_pre_save
,
242 .needed
= global_state_needed
,
243 .fields
= (VMStateField
[]) {
244 VMSTATE_UINT32(size
, GlobalState
),
245 VMSTATE_BUFFER(runstate
, GlobalState
),
246 VMSTATE_END_OF_LIST()
250 void register_global_state(void)
252 /* We would use it independently that we receive it */
253 strcpy((char *)&global_state
.runstate
, "");
254 global_state
.received
= false;
255 vmstate_register(NULL
, 0, &vmstate_globalstate
, &global_state
);
258 static void migrate_generate_event(int new_state
)
260 if (migrate_use_events()) {
261 qapi_event_send_migration(new_state
, &error_abort
);
266 * Called on -incoming with a defer: uri.
267 * The migration can be started later after any parameters have been
270 static void deferred_incoming_migration(Error
**errp
)
272 if (deferred_incoming
) {
273 error_setg(errp
, "Incoming migration already deferred");
275 deferred_incoming
= true;
278 /* Request a range of pages from the source VM at the given
280 * rbname: Name of the RAMBlock to request the page in, if NULL it's the same
281 * as the last request (a name must have been given previously)
282 * Start: Address offset within the RB
283 * Len: Length in bytes required - must be a multiple of pagesize
285 void migrate_send_rp_req_pages(MigrationIncomingState
*mis
, const char *rbname
,
286 ram_addr_t start
, size_t len
)
288 uint8_t bufc
[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
289 size_t msglen
= 12; /* start + len */
291 *(uint64_t *)bufc
= cpu_to_be64((uint64_t)start
);
292 *(uint32_t *)(bufc
+ 8) = cpu_to_be32((uint32_t)len
);
295 int rbname_len
= strlen(rbname
);
296 assert(rbname_len
< 256);
298 bufc
[msglen
++] = rbname_len
;
299 memcpy(bufc
+ msglen
, rbname
, rbname_len
);
300 msglen
+= rbname_len
;
301 migrate_send_rp_message(mis
, MIG_RP_MSG_REQ_PAGES_ID
, msglen
, bufc
);
303 migrate_send_rp_message(mis
, MIG_RP_MSG_REQ_PAGES
, msglen
, bufc
);
307 void qemu_start_incoming_migration(const char *uri
, Error
**errp
)
311 qapi_event_send_migration(MIGRATION_STATUS_SETUP
, &error_abort
);
312 if (!strcmp(uri
, "defer")) {
313 deferred_incoming_migration(errp
);
314 } else if (strstart(uri
, "tcp:", &p
)) {
315 tcp_start_incoming_migration(p
, errp
);
317 } else if (strstart(uri
, "rdma:", &p
)) {
318 rdma_start_incoming_migration(p
, errp
);
320 } else if (strstart(uri
, "exec:", &p
)) {
321 exec_start_incoming_migration(p
, errp
);
322 } else if (strstart(uri
, "unix:", &p
)) {
323 unix_start_incoming_migration(p
, errp
);
324 } else if (strstart(uri
, "fd:", &p
)) {
325 fd_start_incoming_migration(p
, errp
);
327 error_setg(errp
, "unknown migration protocol: %s", uri
);
331 static void process_incoming_migration_bh(void *opaque
)
333 Error
*local_err
= NULL
;
334 MigrationIncomingState
*mis
= opaque
;
336 /* Make sure all file formats flush their mutable metadata */
337 bdrv_invalidate_cache_all(&local_err
);
339 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
340 MIGRATION_STATUS_FAILED
);
341 error_report_err(local_err
);
342 migrate_decompress_threads_join();
347 * This must happen after all error conditions are dealt with and
348 * we're sure the VM is going to be running on this host.
350 qemu_announce_self();
352 /* If global state section was not received or we are in running
353 state, we need to obey autostart. Any other state is set with
356 if (!global_state_received() ||
357 global_state_get_runstate() == RUN_STATE_RUNNING
) {
361 runstate_set(RUN_STATE_PAUSED
);
364 runstate_set(global_state_get_runstate());
366 migrate_decompress_threads_join();
368 * This must happen after any state changes since as soon as an external
369 * observer sees this event they might start to prod at the VM assuming
372 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
373 MIGRATION_STATUS_COMPLETED
);
374 qemu_bh_delete(mis
->bh
);
375 migration_incoming_state_destroy();
378 static void process_incoming_migration_co(void *opaque
)
380 QEMUFile
*f
= opaque
;
381 MigrationIncomingState
*mis
= migration_incoming_get_current();
385 mis
->from_src_file
= f
;
386 postcopy_state_set(POSTCOPY_INCOMING_NONE
);
387 migrate_set_state(&mis
->state
, MIGRATION_STATUS_NONE
,
388 MIGRATION_STATUS_ACTIVE
);
389 ret
= qemu_loadvm_state(f
);
391 ps
= postcopy_state_get();
392 trace_process_incoming_migration_co_end(ret
, ps
);
393 if (ps
!= POSTCOPY_INCOMING_NONE
) {
394 if (ps
== POSTCOPY_INCOMING_ADVISE
) {
396 * Where a migration had postcopy enabled (and thus went to advise)
397 * but managed to complete within the precopy period, we can use
400 postcopy_ram_incoming_cleanup(mis
);
401 } else if (ret
>= 0) {
403 * Postcopy was started, cleanup should happen at the end of the
406 trace_process_incoming_migration_co_postcopy_end_main();
409 /* Else if something went wrong then just fall out of the normal exit */
412 /* we get COLO info, and know if we are in COLO mode */
413 if (!ret
&& migration_incoming_enable_colo()) {
414 mis
->migration_incoming_co
= qemu_coroutine_self();
415 qemu_thread_create(&mis
->colo_incoming_thread
, "COLO incoming",
416 colo_process_incoming_thread
, mis
, QEMU_THREAD_JOINABLE
);
417 mis
->have_colo_incoming_thread
= true;
418 qemu_coroutine_yield();
420 /* Wait checkpoint incoming thread exit before free resource */
421 qemu_thread_join(&mis
->colo_incoming_thread
);
425 free_xbzrle_decoded_buf();
428 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
429 MIGRATION_STATUS_FAILED
);
430 error_report("load of migration failed: %s", strerror(-ret
));
431 migrate_decompress_threads_join();
435 mis
->bh
= qemu_bh_new(process_incoming_migration_bh
, mis
);
436 qemu_bh_schedule(mis
->bh
);
439 void migration_fd_process_incoming(QEMUFile
*f
)
441 Coroutine
*co
= qemu_coroutine_create(process_incoming_migration_co
, f
);
443 migrate_decompress_threads_create();
444 qemu_file_set_blocking(f
, false);
445 qemu_coroutine_enter(co
);
449 void migration_channel_process_incoming(MigrationState
*s
,
452 trace_migration_set_incoming_channel(
453 ioc
, object_get_typename(OBJECT(ioc
)));
455 if (s
->parameters
.tls_creds
&&
456 !object_dynamic_cast(OBJECT(ioc
),
457 TYPE_QIO_CHANNEL_TLS
)) {
458 Error
*local_err
= NULL
;
459 migration_tls_channel_process_incoming(s
, ioc
, &local_err
);
461 error_report_err(local_err
);
464 QEMUFile
*f
= qemu_fopen_channel_input(ioc
);
465 migration_fd_process_incoming(f
);
470 void migration_channel_connect(MigrationState
*s
,
472 const char *hostname
)
474 trace_migration_set_outgoing_channel(
475 ioc
, object_get_typename(OBJECT(ioc
)), hostname
);
477 if (s
->parameters
.tls_creds
&&
478 !object_dynamic_cast(OBJECT(ioc
),
479 TYPE_QIO_CHANNEL_TLS
)) {
480 Error
*local_err
= NULL
;
481 migration_tls_channel_connect(s
, ioc
, hostname
, &local_err
);
483 migrate_fd_error(s
, local_err
);
484 error_free(local_err
);
487 QEMUFile
*f
= qemu_fopen_channel_output(ioc
);
491 migrate_fd_connect(s
);
497 * Send a message on the return channel back to the source
500 void migrate_send_rp_message(MigrationIncomingState
*mis
,
501 enum mig_rp_message_type message_type
,
502 uint16_t len
, void *data
)
504 trace_migrate_send_rp_message((int)message_type
, len
);
505 qemu_mutex_lock(&mis
->rp_mutex
);
506 qemu_put_be16(mis
->to_src_file
, (unsigned int)message_type
);
507 qemu_put_be16(mis
->to_src_file
, len
);
508 qemu_put_buffer(mis
->to_src_file
, data
, len
);
509 qemu_fflush(mis
->to_src_file
);
510 qemu_mutex_unlock(&mis
->rp_mutex
);
514 * Send a 'SHUT' message on the return channel with the given value
515 * to indicate that we've finished with the RP. Non-0 value indicates
518 void migrate_send_rp_shut(MigrationIncomingState
*mis
,
523 buf
= cpu_to_be32(value
);
524 migrate_send_rp_message(mis
, MIG_RP_MSG_SHUT
, sizeof(buf
), &buf
);
528 * Send a 'PONG' message on the return channel with the given value
529 * (normally in response to a 'PING')
531 void migrate_send_rp_pong(MigrationIncomingState
*mis
,
536 buf
= cpu_to_be32(value
);
537 migrate_send_rp_message(mis
, MIG_RP_MSG_PONG
, sizeof(buf
), &buf
);
540 MigrationCapabilityStatusList
*qmp_query_migrate_capabilities(Error
**errp
)
542 MigrationCapabilityStatusList
*head
= NULL
;
543 MigrationCapabilityStatusList
*caps
;
544 MigrationState
*s
= migrate_get_current();
547 caps
= NULL
; /* silence compiler warning */
548 for (i
= 0; i
< MIGRATION_CAPABILITY__MAX
; i
++) {
549 if (i
== MIGRATION_CAPABILITY_X_COLO
&& !colo_supported()) {
553 head
= g_malloc0(sizeof(*caps
));
556 caps
->next
= g_malloc0(sizeof(*caps
));
560 g_malloc(sizeof(*caps
->value
));
561 caps
->value
->capability
= i
;
562 caps
->value
->state
= s
->enabled_capabilities
[i
];
568 MigrationParameters
*qmp_query_migrate_parameters(Error
**errp
)
570 MigrationParameters
*params
;
571 MigrationState
*s
= migrate_get_current();
573 params
= g_malloc0(sizeof(*params
));
574 params
->has_compress_level
= true;
575 params
->compress_level
= s
->parameters
.compress_level
;
576 params
->has_compress_threads
= true;
577 params
->compress_threads
= s
->parameters
.compress_threads
;
578 params
->has_decompress_threads
= true;
579 params
->decompress_threads
= s
->parameters
.decompress_threads
;
580 params
->has_cpu_throttle_initial
= true;
581 params
->cpu_throttle_initial
= s
->parameters
.cpu_throttle_initial
;
582 params
->has_cpu_throttle_increment
= true;
583 params
->cpu_throttle_increment
= s
->parameters
.cpu_throttle_increment
;
584 params
->has_tls_creds
= !!s
->parameters
.tls_creds
;
585 params
->tls_creds
= g_strdup(s
->parameters
.tls_creds
);
586 params
->has_tls_hostname
= !!s
->parameters
.tls_hostname
;
587 params
->tls_hostname
= g_strdup(s
->parameters
.tls_hostname
);
588 params
->has_max_bandwidth
= true;
589 params
->max_bandwidth
= s
->parameters
.max_bandwidth
;
590 params
->has_downtime_limit
= true;
591 params
->downtime_limit
= s
->parameters
.downtime_limit
;
592 params
->has_x_checkpoint_delay
= true;
593 params
->x_checkpoint_delay
= s
->parameters
.x_checkpoint_delay
;
599 * Return true if we're already in the middle of a migration
600 * (i.e. any of the active or setup states)
602 static bool migration_is_setup_or_active(int state
)
605 case MIGRATION_STATUS_ACTIVE
:
606 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
607 case MIGRATION_STATUS_SETUP
:
616 static void get_xbzrle_cache_stats(MigrationInfo
*info
)
618 if (migrate_use_xbzrle()) {
619 info
->has_xbzrle_cache
= true;
620 info
->xbzrle_cache
= g_malloc0(sizeof(*info
->xbzrle_cache
));
621 info
->xbzrle_cache
->cache_size
= migrate_xbzrle_cache_size();
622 info
->xbzrle_cache
->bytes
= xbzrle_mig_bytes_transferred();
623 info
->xbzrle_cache
->pages
= xbzrle_mig_pages_transferred();
624 info
->xbzrle_cache
->cache_miss
= xbzrle_mig_pages_cache_miss();
625 info
->xbzrle_cache
->cache_miss_rate
= xbzrle_mig_cache_miss_rate();
626 info
->xbzrle_cache
->overflow
= xbzrle_mig_pages_overflow();
630 static void populate_ram_info(MigrationInfo
*info
, MigrationState
*s
)
632 info
->has_ram
= true;
633 info
->ram
= g_malloc0(sizeof(*info
->ram
));
634 info
->ram
->transferred
= ram_bytes_transferred();
635 info
->ram
->total
= ram_bytes_total();
636 info
->ram
->duplicate
= dup_mig_pages_transferred();
637 info
->ram
->skipped
= skipped_mig_pages_transferred();
638 info
->ram
->normal
= norm_mig_pages_transferred();
639 info
->ram
->normal_bytes
= norm_mig_bytes_transferred();
640 info
->ram
->mbps
= s
->mbps
;
641 info
->ram
->dirty_sync_count
= s
->dirty_sync_count
;
642 info
->ram
->postcopy_requests
= s
->postcopy_requests
;
644 if (s
->state
!= MIGRATION_STATUS_COMPLETED
) {
645 info
->ram
->remaining
= ram_bytes_remaining();
646 info
->ram
->dirty_pages_rate
= s
->dirty_pages_rate
;
650 MigrationInfo
*qmp_query_migrate(Error
**errp
)
652 MigrationInfo
*info
= g_malloc0(sizeof(*info
));
653 MigrationState
*s
= migrate_get_current();
656 case MIGRATION_STATUS_NONE
:
657 /* no migration has happened ever */
659 case MIGRATION_STATUS_SETUP
:
660 info
->has_status
= true;
661 info
->has_total_time
= false;
663 case MIGRATION_STATUS_ACTIVE
:
664 case MIGRATION_STATUS_CANCELLING
:
665 info
->has_status
= true;
666 info
->has_total_time
= true;
667 info
->total_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
)
669 info
->has_expected_downtime
= true;
670 info
->expected_downtime
= s
->expected_downtime
;
671 info
->has_setup_time
= true;
672 info
->setup_time
= s
->setup_time
;
674 populate_ram_info(info
, s
);
676 if (blk_mig_active()) {
677 info
->has_disk
= true;
678 info
->disk
= g_malloc0(sizeof(*info
->disk
));
679 info
->disk
->transferred
= blk_mig_bytes_transferred();
680 info
->disk
->remaining
= blk_mig_bytes_remaining();
681 info
->disk
->total
= blk_mig_bytes_total();
684 if (cpu_throttle_active()) {
685 info
->has_cpu_throttle_percentage
= true;
686 info
->cpu_throttle_percentage
= cpu_throttle_get_percentage();
689 get_xbzrle_cache_stats(info
);
691 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
692 /* Mostly the same as active; TODO add some postcopy stats */
693 info
->has_status
= true;
694 info
->has_total_time
= true;
695 info
->total_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
)
697 info
->has_expected_downtime
= true;
698 info
->expected_downtime
= s
->expected_downtime
;
699 info
->has_setup_time
= true;
700 info
->setup_time
= s
->setup_time
;
702 populate_ram_info(info
, s
);
704 if (blk_mig_active()) {
705 info
->has_disk
= true;
706 info
->disk
= g_malloc0(sizeof(*info
->disk
));
707 info
->disk
->transferred
= blk_mig_bytes_transferred();
708 info
->disk
->remaining
= blk_mig_bytes_remaining();
709 info
->disk
->total
= blk_mig_bytes_total();
712 get_xbzrle_cache_stats(info
);
714 case MIGRATION_STATUS_COLO
:
715 info
->has_status
= true;
716 /* TODO: display COLO specific information (checkpoint info etc.) */
718 case MIGRATION_STATUS_COMPLETED
:
719 get_xbzrle_cache_stats(info
);
721 info
->has_status
= true;
722 info
->has_total_time
= true;
723 info
->total_time
= s
->total_time
;
724 info
->has_downtime
= true;
725 info
->downtime
= s
->downtime
;
726 info
->has_setup_time
= true;
727 info
->setup_time
= s
->setup_time
;
729 populate_ram_info(info
, s
);
731 case MIGRATION_STATUS_FAILED
:
732 info
->has_status
= true;
734 info
->has_error_desc
= true;
735 info
->error_desc
= g_strdup(error_get_pretty(s
->error
));
738 case MIGRATION_STATUS_CANCELLED
:
739 info
->has_status
= true;
742 info
->status
= s
->state
;
747 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList
*params
,
750 MigrationState
*s
= migrate_get_current();
751 MigrationCapabilityStatusList
*cap
;
752 bool old_postcopy_cap
= migrate_postcopy_ram();
754 if (migration_is_setup_or_active(s
->state
)) {
755 error_setg(errp
, QERR_MIGRATION_ACTIVE
);
759 for (cap
= params
; cap
; cap
= cap
->next
) {
760 if (cap
->value
->capability
== MIGRATION_CAPABILITY_X_COLO
) {
761 if (!colo_supported()) {
762 error_setg(errp
, "COLO is not currently supported, please"
763 " configure with --enable-colo option in order to"
764 " support COLO feature");
768 s
->enabled_capabilities
[cap
->value
->capability
] = cap
->value
->state
;
771 if (migrate_postcopy_ram()) {
772 if (migrate_use_compression()) {
773 /* The decompression threads asynchronously write into RAM
774 * rather than use the atomic copies needed to avoid
775 * userfaulting. It should be possible to fix the decompression
776 * threads for compatibility in future.
778 error_report("Postcopy is not currently compatible with "
780 s
->enabled_capabilities
[MIGRATION_CAPABILITY_POSTCOPY_RAM
] =
783 /* This check is reasonably expensive, so only when it's being
784 * set the first time, also it's only the destination that needs
787 if (!old_postcopy_cap
&& runstate_check(RUN_STATE_INMIGRATE
) &&
788 !postcopy_ram_supported_by_host()) {
789 /* postcopy_ram_supported_by_host will have emitted a more
792 error_report("Postcopy is not supported");
793 s
->enabled_capabilities
[MIGRATION_CAPABILITY_POSTCOPY_RAM
] =
799 void qmp_migrate_set_parameters(MigrationParameters
*params
, Error
**errp
)
801 MigrationState
*s
= migrate_get_current();
803 if (params
->has_compress_level
&&
804 (params
->compress_level
< 0 || params
->compress_level
> 9)) {
805 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
, "compress_level",
806 "is invalid, it should be in the range of 0 to 9");
809 if (params
->has_compress_threads
&&
810 (params
->compress_threads
< 1 || params
->compress_threads
> 255)) {
811 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
,
813 "is invalid, it should be in the range of 1 to 255");
816 if (params
->has_decompress_threads
&&
817 (params
->decompress_threads
< 1 || params
->decompress_threads
> 255)) {
818 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
,
819 "decompress_threads",
820 "is invalid, it should be in the range of 1 to 255");
823 if (params
->has_cpu_throttle_initial
&&
824 (params
->cpu_throttle_initial
< 1 ||
825 params
->cpu_throttle_initial
> 99)) {
826 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
,
827 "cpu_throttle_initial",
828 "an integer in the range of 1 to 99");
831 if (params
->has_cpu_throttle_increment
&&
832 (params
->cpu_throttle_increment
< 1 ||
833 params
->cpu_throttle_increment
> 99)) {
834 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
,
835 "cpu_throttle_increment",
836 "an integer in the range of 1 to 99");
839 if (params
->has_max_bandwidth
&&
840 (params
->max_bandwidth
< 0 || params
->max_bandwidth
> SIZE_MAX
)) {
841 error_setg(errp
, "Parameter 'max_bandwidth' expects an integer in the"
842 " range of 0 to %zu bytes/second", SIZE_MAX
);
845 if (params
->has_downtime_limit
&&
846 (params
->downtime_limit
< 0 || params
->downtime_limit
> 2000000)) {
847 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
,
849 "an integer in the range of 0 to 2000000 milliseconds");
852 if (params
->has_x_checkpoint_delay
&& (params
->x_checkpoint_delay
< 0)) {
853 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
,
854 "x_checkpoint_delay",
855 "is invalid, it should be positive");
858 if (params
->has_compress_level
) {
859 s
->parameters
.compress_level
= params
->compress_level
;
861 if (params
->has_compress_threads
) {
862 s
->parameters
.compress_threads
= params
->compress_threads
;
864 if (params
->has_decompress_threads
) {
865 s
->parameters
.decompress_threads
= params
->decompress_threads
;
867 if (params
->has_cpu_throttle_initial
) {
868 s
->parameters
.cpu_throttle_initial
= params
->cpu_throttle_initial
;
870 if (params
->has_cpu_throttle_increment
) {
871 s
->parameters
.cpu_throttle_increment
= params
->cpu_throttle_increment
;
873 if (params
->has_tls_creds
) {
874 g_free(s
->parameters
.tls_creds
);
875 s
->parameters
.tls_creds
= g_strdup(params
->tls_creds
);
877 if (params
->has_tls_hostname
) {
878 g_free(s
->parameters
.tls_hostname
);
879 s
->parameters
.tls_hostname
= g_strdup(params
->tls_hostname
);
881 if (params
->has_max_bandwidth
) {
882 s
->parameters
.max_bandwidth
= params
->max_bandwidth
;
883 if (s
->to_dst_file
) {
884 qemu_file_set_rate_limit(s
->to_dst_file
,
885 s
->parameters
.max_bandwidth
/ XFER_LIMIT_RATIO
);
888 if (params
->has_downtime_limit
) {
889 s
->parameters
.downtime_limit
= params
->downtime_limit
;
892 if (params
->has_x_checkpoint_delay
) {
893 s
->parameters
.x_checkpoint_delay
= params
->x_checkpoint_delay
;
894 if (migration_in_colo_state()) {
895 colo_checkpoint_notify(s
);
901 void qmp_migrate_start_postcopy(Error
**errp
)
903 MigrationState
*s
= migrate_get_current();
905 if (!migrate_postcopy_ram()) {
906 error_setg(errp
, "Enable postcopy with migrate_set_capability before"
907 " the start of migration");
911 if (s
->state
== MIGRATION_STATUS_NONE
) {
912 error_setg(errp
, "Postcopy must be started after migration has been"
917 * we don't error if migration has finished since that would be racy
918 * with issuing this command.
920 atomic_set(&s
->start_postcopy
, true);
923 /* shared migration helpers */
925 void migrate_set_state(int *state
, int old_state
, int new_state
)
927 if (atomic_cmpxchg(state
, old_state
, new_state
) == old_state
) {
928 trace_migrate_set_state(new_state
);
929 migrate_generate_event(new_state
);
933 static void migrate_fd_cleanup(void *opaque
)
935 MigrationState
*s
= opaque
;
937 qemu_bh_delete(s
->cleanup_bh
);
938 s
->cleanup_bh
= NULL
;
942 if (s
->to_dst_file
) {
943 trace_migrate_fd_cleanup();
944 qemu_mutex_unlock_iothread();
945 if (s
->migration_thread_running
) {
946 qemu_thread_join(&s
->thread
);
947 s
->migration_thread_running
= false;
949 qemu_mutex_lock_iothread();
951 migrate_compress_threads_join();
952 qemu_fclose(s
->to_dst_file
);
953 s
->to_dst_file
= NULL
;
956 assert((s
->state
!= MIGRATION_STATUS_ACTIVE
) &&
957 (s
->state
!= MIGRATION_STATUS_POSTCOPY_ACTIVE
));
959 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
960 migrate_set_state(&s
->state
, MIGRATION_STATUS_CANCELLING
,
961 MIGRATION_STATUS_CANCELLED
);
964 notifier_list_notify(&migration_state_notifiers
, s
);
967 void migrate_fd_error(MigrationState
*s
, const Error
*error
)
969 trace_migrate_fd_error(error_get_pretty(error
));
970 assert(s
->to_dst_file
== NULL
);
971 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
972 MIGRATION_STATUS_FAILED
);
974 s
->error
= error_copy(error
);
976 notifier_list_notify(&migration_state_notifiers
, s
);
979 static void migrate_fd_cancel(MigrationState
*s
)
982 QEMUFile
*f
= migrate_get_current()->to_dst_file
;
983 trace_migrate_fd_cancel();
985 if (s
->rp_state
.from_dst_file
) {
986 /* shutdown the rp socket, so causing the rp thread to shutdown */
987 qemu_file_shutdown(s
->rp_state
.from_dst_file
);
991 old_state
= s
->state
;
992 if (!migration_is_setup_or_active(old_state
)) {
995 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_CANCELLING
);
996 } while (s
->state
!= MIGRATION_STATUS_CANCELLING
);
999 * If we're unlucky the migration code might be stuck somewhere in a
1000 * send/write while the network has failed and is waiting to timeout;
1001 * if we've got shutdown(2) available then we can force it to quit.
1002 * The outgoing qemu file gets closed in migrate_fd_cleanup that is
1003 * called in a bh, so there is no race against this cancel.
1005 if (s
->state
== MIGRATION_STATUS_CANCELLING
&& f
) {
1006 qemu_file_shutdown(f
);
1008 if (s
->state
== MIGRATION_STATUS_CANCELLING
&& s
->block_inactive
) {
1009 Error
*local_err
= NULL
;
1011 bdrv_invalidate_cache_all(&local_err
);
1013 error_report_err(local_err
);
1015 s
->block_inactive
= false;
1020 void add_migration_state_change_notifier(Notifier
*notify
)
1022 notifier_list_add(&migration_state_notifiers
, notify
);
1025 void remove_migration_state_change_notifier(Notifier
*notify
)
1027 notifier_remove(notify
);
1030 bool migration_in_setup(MigrationState
*s
)
1032 return s
->state
== MIGRATION_STATUS_SETUP
;
1035 bool migration_has_finished(MigrationState
*s
)
1037 return s
->state
== MIGRATION_STATUS_COMPLETED
;
1040 bool migration_has_failed(MigrationState
*s
)
1042 return (s
->state
== MIGRATION_STATUS_CANCELLED
||
1043 s
->state
== MIGRATION_STATUS_FAILED
);
1046 bool migration_in_postcopy(MigrationState
*s
)
1048 return (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
1051 bool migration_in_postcopy_after_devices(MigrationState
*s
)
1053 return migration_in_postcopy(s
) && s
->postcopy_after_devices
;
1056 bool migration_is_idle(MigrationState
*s
)
1059 s
= migrate_get_current();
1063 case MIGRATION_STATUS_NONE
:
1064 case MIGRATION_STATUS_CANCELLED
:
1065 case MIGRATION_STATUS_COMPLETED
:
1066 case MIGRATION_STATUS_FAILED
:
1068 case MIGRATION_STATUS_SETUP
:
1069 case MIGRATION_STATUS_CANCELLING
:
1070 case MIGRATION_STATUS_ACTIVE
:
1071 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1072 case MIGRATION_STATUS_COLO
:
1074 case MIGRATION_STATUS__MAX
:
1075 g_assert_not_reached();
1081 MigrationState
*migrate_init(const MigrationParams
*params
)
1083 MigrationState
*s
= migrate_get_current();
1086 * Reinitialise all migration state, except
1087 * parameters/capabilities that the user set, and
1093 s
->to_dst_file
= NULL
;
1094 s
->state
= MIGRATION_STATUS_NONE
;
1095 s
->params
= *params
;
1096 s
->rp_state
.from_dst_file
= NULL
;
1097 s
->rp_state
.error
= false;
1100 s
->expected_downtime
= 0;
1101 s
->dirty_pages_rate
= 0;
1102 s
->dirty_bytes_rate
= 0;
1104 s
->dirty_sync_count
= 0;
1105 s
->start_postcopy
= false;
1106 s
->postcopy_after_devices
= false;
1107 s
->postcopy_requests
= 0;
1108 s
->migration_thread_running
= false;
1109 s
->last_req_rb
= NULL
;
1110 error_free(s
->error
);
1113 migrate_set_state(&s
->state
, MIGRATION_STATUS_NONE
, MIGRATION_STATUS_SETUP
);
1115 QSIMPLEQ_INIT(&s
->src_page_requests
);
1117 s
->total_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1121 static GSList
*migration_blockers
;
1123 int migrate_add_blocker(Error
*reason
, Error
**errp
)
1125 if (only_migratable
) {
1126 error_propagate(errp
, error_copy(reason
));
1127 error_prepend(errp
, "disallowing migration blocker "
1128 "(--only_migratable) for: ");
1132 if (migration_is_idle(NULL
)) {
1133 migration_blockers
= g_slist_prepend(migration_blockers
, reason
);
1137 error_propagate(errp
, error_copy(reason
));
1138 error_prepend(errp
, "disallowing migration blocker (migration in "
1143 void migrate_del_blocker(Error
*reason
)
1145 migration_blockers
= g_slist_remove(migration_blockers
, reason
);
1148 void qmp_migrate_incoming(const char *uri
, Error
**errp
)
1150 Error
*local_err
= NULL
;
1151 static bool once
= true;
1153 if (!deferred_incoming
) {
1154 error_setg(errp
, "For use with '-incoming defer'");
1158 error_setg(errp
, "The incoming migration has already been started");
1161 qemu_start_incoming_migration(uri
, &local_err
);
1164 error_propagate(errp
, local_err
);
1171 bool migration_is_blocked(Error
**errp
)
1173 if (qemu_savevm_state_blocked(errp
)) {
1177 if (migration_blockers
) {
1178 *errp
= error_copy(migration_blockers
->data
);
1185 void qmp_migrate(const char *uri
, bool has_blk
, bool blk
,
1186 bool has_inc
, bool inc
, bool has_detach
, bool detach
,
1189 Error
*local_err
= NULL
;
1190 MigrationState
*s
= migrate_get_current();
1191 MigrationParams params
;
1194 params
.blk
= has_blk
&& blk
;
1195 params
.shared
= has_inc
&& inc
;
1197 if (migration_is_setup_or_active(s
->state
) ||
1198 s
->state
== MIGRATION_STATUS_CANCELLING
||
1199 s
->state
== MIGRATION_STATUS_COLO
) {
1200 error_setg(errp
, QERR_MIGRATION_ACTIVE
);
1203 if (runstate_check(RUN_STATE_INMIGRATE
)) {
1204 error_setg(errp
, "Guest is waiting for an incoming migration");
1208 if (migration_is_blocked(errp
)) {
1212 s
= migrate_init(¶ms
);
1214 if (strstart(uri
, "tcp:", &p
)) {
1215 tcp_start_outgoing_migration(s
, p
, &local_err
);
1217 } else if (strstart(uri
, "rdma:", &p
)) {
1218 rdma_start_outgoing_migration(s
, p
, &local_err
);
1220 } else if (strstart(uri
, "exec:", &p
)) {
1221 exec_start_outgoing_migration(s
, p
, &local_err
);
1222 } else if (strstart(uri
, "unix:", &p
)) {
1223 unix_start_outgoing_migration(s
, p
, &local_err
);
1224 } else if (strstart(uri
, "fd:", &p
)) {
1225 fd_start_outgoing_migration(s
, p
, &local_err
);
1227 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
, "uri",
1228 "a valid migration protocol");
1229 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
1230 MIGRATION_STATUS_FAILED
);
1235 migrate_fd_error(s
, local_err
);
1236 error_propagate(errp
, local_err
);
1241 void qmp_migrate_cancel(Error
**errp
)
1243 migrate_fd_cancel(migrate_get_current());
1246 void qmp_migrate_set_cache_size(int64_t value
, Error
**errp
)
1248 MigrationState
*s
= migrate_get_current();
1251 /* Check for truncation */
1252 if (value
!= (size_t)value
) {
1253 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
, "cache size",
1254 "exceeding address space");
1258 /* Cache should not be larger than guest ram size */
1259 if (value
> ram_bytes_total()) {
1260 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
, "cache size",
1261 "exceeds guest ram size ");
1265 new_size
= xbzrle_cache_resize(value
);
1267 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
, "cache size",
1268 "is smaller than page size");
1272 s
->xbzrle_cache_size
= new_size
;
1275 int64_t qmp_query_migrate_cache_size(Error
**errp
)
1277 return migrate_xbzrle_cache_size();
1280 void qmp_migrate_set_speed(int64_t value
, Error
**errp
)
1282 MigrationParameters p
= {
1283 .has_max_bandwidth
= true,
1284 .max_bandwidth
= value
,
1287 qmp_migrate_set_parameters(&p
, errp
);
1290 void qmp_migrate_set_downtime(double value
, Error
**errp
)
1292 value
*= 1000; /* Convert to milliseconds */
1293 value
= MAX(0, MIN(INT64_MAX
, value
));
1295 MigrationParameters p
= {
1296 .has_downtime_limit
= true,
1297 .downtime_limit
= value
,
1300 qmp_migrate_set_parameters(&p
, errp
);
1303 bool migrate_release_ram(void)
1307 s
= migrate_get_current();
1309 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_RELEASE_RAM
];
1312 bool migrate_postcopy_ram(void)
1316 s
= migrate_get_current();
1318 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_POSTCOPY_RAM
];
1321 bool migrate_auto_converge(void)
1325 s
= migrate_get_current();
1327 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_AUTO_CONVERGE
];
1330 bool migrate_zero_blocks(void)
1334 s
= migrate_get_current();
1336 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_ZERO_BLOCKS
];
1339 bool migrate_use_compression(void)
1343 s
= migrate_get_current();
1345 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_COMPRESS
];
1348 int migrate_compress_level(void)
1352 s
= migrate_get_current();
1354 return s
->parameters
.compress_level
;
1357 int migrate_compress_threads(void)
1361 s
= migrate_get_current();
1363 return s
->parameters
.compress_threads
;
1366 int migrate_decompress_threads(void)
1370 s
= migrate_get_current();
1372 return s
->parameters
.decompress_threads
;
1375 bool migrate_use_events(void)
1379 s
= migrate_get_current();
1381 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_EVENTS
];
1384 int migrate_use_xbzrle(void)
1388 s
= migrate_get_current();
1390 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_XBZRLE
];
1393 int64_t migrate_xbzrle_cache_size(void)
1397 s
= migrate_get_current();
1399 return s
->xbzrle_cache_size
;
1402 /* migration thread support */
1404 * Something bad happened to the RP stream, mark an error
1405 * The caller shall print or trace something to indicate why
1407 static void mark_source_rp_bad(MigrationState
*s
)
1409 s
->rp_state
.error
= true;
1412 static struct rp_cmd_args
{
1413 ssize_t len
; /* -1 = variable */
1416 [MIG_RP_MSG_INVALID
] = { .len
= -1, .name
= "INVALID" },
1417 [MIG_RP_MSG_SHUT
] = { .len
= 4, .name
= "SHUT" },
1418 [MIG_RP_MSG_PONG
] = { .len
= 4, .name
= "PONG" },
1419 [MIG_RP_MSG_REQ_PAGES
] = { .len
= 12, .name
= "REQ_PAGES" },
1420 [MIG_RP_MSG_REQ_PAGES_ID
] = { .len
= -1, .name
= "REQ_PAGES_ID" },
1421 [MIG_RP_MSG_MAX
] = { .len
= -1, .name
= "MAX" },
1425 * Process a request for pages received on the return path,
1426 * We're allowed to send more than requested (e.g. to round to our page size)
1427 * and we don't need to send pages that have already been sent.
1429 static void migrate_handle_rp_req_pages(MigrationState
*ms
, const char* rbname
,
1430 ram_addr_t start
, size_t len
)
1432 long our_host_ps
= getpagesize();
1434 trace_migrate_handle_rp_req_pages(rbname
, start
, len
);
1437 * Since we currently insist on matching page sizes, just sanity check
1438 * we're being asked for whole host pages.
1440 if (start
& (our_host_ps
-1) ||
1441 (len
& (our_host_ps
-1))) {
1442 error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1443 " len: %zd", __func__
, start
, len
);
1444 mark_source_rp_bad(ms
);
1448 if (ram_save_queue_pages(ms
, rbname
, start
, len
)) {
1449 mark_source_rp_bad(ms
);
1454 * Handles messages sent on the return path towards the source VM
1457 static void *source_return_path_thread(void *opaque
)
1459 MigrationState
*ms
= opaque
;
1460 QEMUFile
*rp
= ms
->rp_state
.from_dst_file
;
1461 uint16_t header_len
, header_type
;
1463 uint32_t tmp32
, sibling_error
;
1464 ram_addr_t start
= 0; /* =0 to silence warning */
1465 size_t len
= 0, expected_len
;
1468 trace_source_return_path_thread_entry();
1469 while (!ms
->rp_state
.error
&& !qemu_file_get_error(rp
) &&
1470 migration_is_setup_or_active(ms
->state
)) {
1471 trace_source_return_path_thread_loop_top();
1472 header_type
= qemu_get_be16(rp
);
1473 header_len
= qemu_get_be16(rp
);
1475 if (header_type
>= MIG_RP_MSG_MAX
||
1476 header_type
== MIG_RP_MSG_INVALID
) {
1477 error_report("RP: Received invalid message 0x%04x length 0x%04x",
1478 header_type
, header_len
);
1479 mark_source_rp_bad(ms
);
1483 if ((rp_cmd_args
[header_type
].len
!= -1 &&
1484 header_len
!= rp_cmd_args
[header_type
].len
) ||
1485 header_len
> sizeof(buf
)) {
1486 error_report("RP: Received '%s' message (0x%04x) with"
1487 "incorrect length %d expecting %zu",
1488 rp_cmd_args
[header_type
].name
, header_type
, header_len
,
1489 (size_t)rp_cmd_args
[header_type
].len
);
1490 mark_source_rp_bad(ms
);
1494 /* We know we've got a valid header by this point */
1495 res
= qemu_get_buffer(rp
, buf
, header_len
);
1496 if (res
!= header_len
) {
1497 error_report("RP: Failed reading data for message 0x%04x"
1498 " read %d expected %d",
1499 header_type
, res
, header_len
);
1500 mark_source_rp_bad(ms
);
1504 /* OK, we have the message and the data */
1505 switch (header_type
) {
1506 case MIG_RP_MSG_SHUT
:
1507 sibling_error
= ldl_be_p(buf
);
1508 trace_source_return_path_thread_shut(sibling_error
);
1509 if (sibling_error
) {
1510 error_report("RP: Sibling indicated error %d", sibling_error
);
1511 mark_source_rp_bad(ms
);
1514 * We'll let the main thread deal with closing the RP
1515 * we could do a shutdown(2) on it, but we're the only user
1516 * anyway, so there's nothing gained.
1520 case MIG_RP_MSG_PONG
:
1521 tmp32
= ldl_be_p(buf
);
1522 trace_source_return_path_thread_pong(tmp32
);
1525 case MIG_RP_MSG_REQ_PAGES
:
1526 start
= ldq_be_p(buf
);
1527 len
= ldl_be_p(buf
+ 8);
1528 migrate_handle_rp_req_pages(ms
, NULL
, start
, len
);
1531 case MIG_RP_MSG_REQ_PAGES_ID
:
1532 expected_len
= 12 + 1; /* header + termination */
1534 if (header_len
>= expected_len
) {
1535 start
= ldq_be_p(buf
);
1536 len
= ldl_be_p(buf
+ 8);
1537 /* Now we expect an idstr */
1538 tmp32
= buf
[12]; /* Length of the following idstr */
1539 buf
[13 + tmp32
] = '\0';
1540 expected_len
+= tmp32
;
1542 if (header_len
!= expected_len
) {
1543 error_report("RP: Req_Page_id with length %d expecting %zd",
1544 header_len
, expected_len
);
1545 mark_source_rp_bad(ms
);
1548 migrate_handle_rp_req_pages(ms
, (char *)&buf
[13], start
, len
);
1555 if (qemu_file_get_error(rp
)) {
1556 trace_source_return_path_thread_bad_end();
1557 mark_source_rp_bad(ms
);
1560 trace_source_return_path_thread_end();
1562 ms
->rp_state
.from_dst_file
= NULL
;
1567 static int open_return_path_on_source(MigrationState
*ms
)
1570 ms
->rp_state
.from_dst_file
= qemu_file_get_return_path(ms
->to_dst_file
);
1571 if (!ms
->rp_state
.from_dst_file
) {
1575 trace_open_return_path_on_source();
1576 qemu_thread_create(&ms
->rp_state
.rp_thread
, "return path",
1577 source_return_path_thread
, ms
, QEMU_THREAD_JOINABLE
);
1579 trace_open_return_path_on_source_continue();
1584 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
1585 static int await_return_path_close_on_source(MigrationState
*ms
)
1588 * If this is a normal exit then the destination will send a SHUT and the
1589 * rp_thread will exit, however if there's an error we need to cause
1592 if (qemu_file_get_error(ms
->to_dst_file
) && ms
->rp_state
.from_dst_file
) {
1594 * shutdown(2), if we have it, will cause it to unblock if it's stuck
1595 * waiting for the destination.
1597 qemu_file_shutdown(ms
->rp_state
.from_dst_file
);
1598 mark_source_rp_bad(ms
);
1600 trace_await_return_path_close_on_source_joining();
1601 qemu_thread_join(&ms
->rp_state
.rp_thread
);
1602 trace_await_return_path_close_on_source_close();
1603 return ms
->rp_state
.error
;
1607 * Switch from normal iteration to postcopy
1608 * Returns non-0 on error
1610 static int postcopy_start(MigrationState
*ms
, bool *old_vm_running
)
1613 QIOChannelBuffer
*bioc
;
1615 int64_t time_at_stop
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1616 bool restart_block
= false;
1617 migrate_set_state(&ms
->state
, MIGRATION_STATUS_ACTIVE
,
1618 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
1620 trace_postcopy_start();
1621 qemu_mutex_lock_iothread();
1622 trace_postcopy_start_set_run();
1624 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER
);
1625 *old_vm_running
= runstate_is_running();
1626 global_state_store();
1627 ret
= vm_stop_force_state(RUN_STATE_FINISH_MIGRATE
);
1632 ret
= bdrv_inactivate_all();
1636 restart_block
= true;
1639 * Cause any non-postcopiable, but iterative devices to
1640 * send out their final data.
1642 qemu_savevm_state_complete_precopy(ms
->to_dst_file
, true);
1645 * in Finish migrate and with the io-lock held everything should
1646 * be quiet, but we've potentially still got dirty pages and we
1647 * need to tell the destination to throw any pages it's already received
1650 if (ram_postcopy_send_discard_bitmap(ms
)) {
1651 error_report("postcopy send discard bitmap failed");
1656 * send rest of state - note things that are doing postcopy
1657 * will notice we're in POSTCOPY_ACTIVE and not actually
1658 * wrap their state up here
1660 qemu_file_set_rate_limit(ms
->to_dst_file
, INT64_MAX
);
1661 /* Ping just for debugging, helps line traces up */
1662 qemu_savevm_send_ping(ms
->to_dst_file
, 2);
1665 * While loading the device state we may trigger page transfer
1666 * requests and the fd must be free to process those, and thus
1667 * the destination must read the whole device state off the fd before
1668 * it starts processing it. Unfortunately the ad-hoc migration format
1669 * doesn't allow the destination to know the size to read without fully
1670 * parsing it through each devices load-state code (especially the open
1671 * coded devices that use get/put).
1672 * So we wrap the device state up in a package with a length at the start;
1673 * to do this we use a qemu_buf to hold the whole of the device state.
1675 bioc
= qio_channel_buffer_new(4096);
1676 qio_channel_set_name(QIO_CHANNEL(bioc
), "migration-postcopy-buffer");
1677 fb
= qemu_fopen_channel_output(QIO_CHANNEL(bioc
));
1678 object_unref(OBJECT(bioc
));
1681 * Make sure the receiver can get incoming pages before we send the rest
1684 qemu_savevm_send_postcopy_listen(fb
);
1686 qemu_savevm_state_complete_precopy(fb
, false);
1687 qemu_savevm_send_ping(fb
, 3);
1689 qemu_savevm_send_postcopy_run(fb
);
1691 /* <><> end of stuff going into the package */
1693 /* Last point of recovery; as soon as we send the package the destination
1694 * can open devices and potentially start running.
1695 * Lets just check again we've not got any errors.
1697 ret
= qemu_file_get_error(ms
->to_dst_file
);
1699 error_report("postcopy_start: Migration stream errored (pre package)");
1703 restart_block
= false;
1705 /* Now send that blob */
1706 if (qemu_savevm_send_packaged(ms
->to_dst_file
, bioc
->data
, bioc
->usage
)) {
1711 /* Send a notify to give a chance for anything that needs to happen
1712 * at the transition to postcopy and after the device state; in particular
1713 * spice needs to trigger a transition now
1715 ms
->postcopy_after_devices
= true;
1716 notifier_list_notify(&migration_state_notifiers
, ms
);
1718 ms
->downtime
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
) - time_at_stop
;
1720 qemu_mutex_unlock_iothread();
1723 * Although this ping is just for debug, it could potentially be
1724 * used for getting a better measurement of downtime at the source.
1726 qemu_savevm_send_ping(ms
->to_dst_file
, 4);
1728 if (migrate_release_ram()) {
1729 ram_postcopy_migrated_memory_release(ms
);
1732 ret
= qemu_file_get_error(ms
->to_dst_file
);
1734 error_report("postcopy_start: Migration stream errored");
1735 migrate_set_state(&ms
->state
, MIGRATION_STATUS_POSTCOPY_ACTIVE
,
1736 MIGRATION_STATUS_FAILED
);
1744 migrate_set_state(&ms
->state
, MIGRATION_STATUS_POSTCOPY_ACTIVE
,
1745 MIGRATION_STATUS_FAILED
);
1746 if (restart_block
) {
1747 /* A failure happened early enough that we know the destination hasn't
1748 * accessed block devices, so we're safe to recover.
1750 Error
*local_err
= NULL
;
1752 bdrv_invalidate_cache_all(&local_err
);
1754 error_report_err(local_err
);
1757 qemu_mutex_unlock_iothread();
1762 * migration_completion: Used by migration_thread when there's not much left.
1763 * The caller 'breaks' the loop when this returns.
1765 * @s: Current migration state
1766 * @current_active_state: The migration state we expect to be in
1767 * @*old_vm_running: Pointer to old_vm_running flag
1768 * @*start_time: Pointer to time to update
1770 static void migration_completion(MigrationState
*s
, int current_active_state
,
1771 bool *old_vm_running
,
1772 int64_t *start_time
)
1776 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
1777 qemu_mutex_lock_iothread();
1778 *start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1779 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER
);
1780 *old_vm_running
= runstate_is_running();
1781 ret
= global_state_store();
1784 ret
= vm_stop_force_state(RUN_STATE_FINISH_MIGRATE
);
1786 * Don't mark the image with BDRV_O_INACTIVE flag if
1787 * we will go into COLO stage later.
1789 if (ret
>= 0 && !migrate_colo_enabled()) {
1790 ret
= bdrv_inactivate_all();
1793 qemu_file_set_rate_limit(s
->to_dst_file
, INT64_MAX
);
1794 qemu_savevm_state_complete_precopy(s
->to_dst_file
, false);
1795 s
->block_inactive
= true;
1798 qemu_mutex_unlock_iothread();
1803 } else if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
1804 trace_migration_completion_postcopy_end();
1806 qemu_savevm_state_complete_postcopy(s
->to_dst_file
);
1807 trace_migration_completion_postcopy_end_after_complete();
1811 * If rp was opened we must clean up the thread before
1812 * cleaning everything else up (since if there are no failures
1813 * it will wait for the destination to send it's status in
1815 * Postcopy opens rp if enabled (even if it's not avtivated)
1817 if (migrate_postcopy_ram()) {
1819 trace_migration_completion_postcopy_end_before_rp();
1820 rp_error
= await_return_path_close_on_source(s
);
1821 trace_migration_completion_postcopy_end_after_rp(rp_error
);
1823 goto fail_invalidate
;
1827 if (qemu_file_get_error(s
->to_dst_file
)) {
1828 trace_migration_completion_file_err();
1829 goto fail_invalidate
;
1832 if (!migrate_colo_enabled()) {
1833 migrate_set_state(&s
->state
, current_active_state
,
1834 MIGRATION_STATUS_COMPLETED
);
1840 /* If not doing postcopy, vm_start() will be called: let's regain
1841 * control on images.
1843 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
1844 Error
*local_err
= NULL
;
1846 qemu_mutex_lock_iothread();
1847 bdrv_invalidate_cache_all(&local_err
);
1849 error_report_err(local_err
);
1851 s
->block_inactive
= false;
1853 qemu_mutex_unlock_iothread();
1857 migrate_set_state(&s
->state
, current_active_state
,
1858 MIGRATION_STATUS_FAILED
);
1861 bool migrate_colo_enabled(void)
1863 MigrationState
*s
= migrate_get_current();
1864 return s
->enabled_capabilities
[MIGRATION_CAPABILITY_X_COLO
];
1868 * Master migration thread on the source VM.
1869 * It drives the migration and pumps the data down the outgoing channel.
1871 static void *migration_thread(void *opaque
)
1873 MigrationState
*s
= opaque
;
1874 /* Used by the bandwidth calcs, updated later */
1875 int64_t initial_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1876 int64_t setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
1877 int64_t initial_bytes
= 0;
1878 int64_t max_size
= 0;
1879 int64_t start_time
= initial_time
;
1881 bool old_vm_running
= false;
1882 bool entered_postcopy
= false;
1883 /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */
1884 enum MigrationStatus current_active_state
= MIGRATION_STATUS_ACTIVE
;
1885 bool enable_colo
= migrate_colo_enabled();
1887 rcu_register_thread();
1889 qemu_savevm_state_header(s
->to_dst_file
);
1891 if (migrate_postcopy_ram()) {
1892 /* Now tell the dest that it should open its end so it can reply */
1893 qemu_savevm_send_open_return_path(s
->to_dst_file
);
1895 /* And do a ping that will make stuff easier to debug */
1896 qemu_savevm_send_ping(s
->to_dst_file
, 1);
1899 * Tell the destination that we *might* want to do postcopy later;
1900 * if the other end can't do postcopy it should fail now, nice and
1903 qemu_savevm_send_postcopy_advise(s
->to_dst_file
);
1906 qemu_savevm_state_begin(s
->to_dst_file
, &s
->params
);
1908 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
1909 current_active_state
= MIGRATION_STATUS_ACTIVE
;
1910 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
1911 MIGRATION_STATUS_ACTIVE
);
1913 trace_migration_thread_setup_complete();
1915 while (s
->state
== MIGRATION_STATUS_ACTIVE
||
1916 s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
1917 int64_t current_time
;
1918 uint64_t pending_size
;
1920 if (!qemu_file_rate_limit(s
->to_dst_file
)) {
1921 uint64_t pend_post
, pend_nonpost
;
1923 qemu_savevm_state_pending(s
->to_dst_file
, max_size
, &pend_nonpost
,
1925 pending_size
= pend_nonpost
+ pend_post
;
1926 trace_migrate_pending(pending_size
, max_size
,
1927 pend_post
, pend_nonpost
);
1928 if (pending_size
&& pending_size
>= max_size
) {
1929 /* Still a significant amount to transfer */
1931 if (migrate_postcopy_ram() &&
1932 s
->state
!= MIGRATION_STATUS_POSTCOPY_ACTIVE
&&
1933 pend_nonpost
<= max_size
&&
1934 atomic_read(&s
->start_postcopy
)) {
1936 if (!postcopy_start(s
, &old_vm_running
)) {
1937 current_active_state
= MIGRATION_STATUS_POSTCOPY_ACTIVE
;
1938 entered_postcopy
= true;
1943 /* Just another iteration step */
1944 qemu_savevm_state_iterate(s
->to_dst_file
, entered_postcopy
);
1946 trace_migration_thread_low_pending(pending_size
);
1947 migration_completion(s
, current_active_state
,
1948 &old_vm_running
, &start_time
);
1953 if (qemu_file_get_error(s
->to_dst_file
)) {
1954 migrate_set_state(&s
->state
, current_active_state
,
1955 MIGRATION_STATUS_FAILED
);
1956 trace_migration_thread_file_err();
1959 current_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1960 if (current_time
>= initial_time
+ BUFFER_DELAY
) {
1961 uint64_t transferred_bytes
= qemu_ftell(s
->to_dst_file
) -
1963 uint64_t time_spent
= current_time
- initial_time
;
1964 double bandwidth
= (double)transferred_bytes
/ time_spent
;
1965 max_size
= bandwidth
* s
->parameters
.downtime_limit
;
1967 s
->mbps
= (((double) transferred_bytes
* 8.0) /
1968 ((double) time_spent
/ 1000.0)) / 1000.0 / 1000.0;
1970 trace_migrate_transferred(transferred_bytes
, time_spent
,
1971 bandwidth
, max_size
);
1972 /* if we haven't sent anything, we don't want to recalculate
1973 10000 is a small enough number for our purposes */
1974 if (s
->dirty_bytes_rate
&& transferred_bytes
> 10000) {
1975 s
->expected_downtime
= s
->dirty_bytes_rate
/ bandwidth
;
1978 qemu_file_reset_rate_limit(s
->to_dst_file
);
1979 initial_time
= current_time
;
1980 initial_bytes
= qemu_ftell(s
->to_dst_file
);
1982 if (qemu_file_rate_limit(s
->to_dst_file
)) {
1983 /* usleep expects microseconds */
1984 g_usleep((initial_time
+ BUFFER_DELAY
- current_time
)*1000);
1988 trace_migration_thread_after_loop();
1989 /* If we enabled cpu throttling for auto-converge, turn it off. */
1990 cpu_throttle_stop();
1991 end_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1993 qemu_mutex_lock_iothread();
1995 * The resource has been allocated by migration will be reused in COLO
1996 * process, so don't release them.
1999 qemu_savevm_state_cleanup();
2001 if (s
->state
== MIGRATION_STATUS_COMPLETED
) {
2002 uint64_t transferred_bytes
= qemu_ftell(s
->to_dst_file
);
2003 s
->total_time
= end_time
- s
->total_time
;
2004 if (!entered_postcopy
) {
2005 s
->downtime
= end_time
- start_time
;
2007 if (s
->total_time
) {
2008 s
->mbps
= (((double) transferred_bytes
* 8.0) /
2009 ((double) s
->total_time
)) / 1000;
2011 runstate_set(RUN_STATE_POSTMIGRATE
);
2013 if (s
->state
== MIGRATION_STATUS_ACTIVE
&& enable_colo
) {
2014 migrate_start_colo_process(s
);
2015 qemu_savevm_state_cleanup();
2017 * Fixme: we will run VM in COLO no matter its old running state.
2018 * After exited COLO, we will keep running.
2020 old_vm_running
= true;
2022 if (old_vm_running
&& !entered_postcopy
) {
2025 if (runstate_check(RUN_STATE_FINISH_MIGRATE
)) {
2026 runstate_set(RUN_STATE_POSTMIGRATE
);
2030 qemu_bh_schedule(s
->cleanup_bh
);
2031 qemu_mutex_unlock_iothread();
2033 rcu_unregister_thread();
2037 void migrate_fd_connect(MigrationState
*s
)
2039 s
->expected_downtime
= s
->parameters
.downtime_limit
;
2040 s
->cleanup_bh
= qemu_bh_new(migrate_fd_cleanup
, s
);
2042 qemu_file_set_blocking(s
->to_dst_file
, true);
2043 qemu_file_set_rate_limit(s
->to_dst_file
,
2044 s
->parameters
.max_bandwidth
/ XFER_LIMIT_RATIO
);
2046 /* Notify before starting migration thread */
2047 notifier_list_notify(&migration_state_notifiers
, s
);
2050 * Open the return path; currently for postcopy but other things might
2053 if (migrate_postcopy_ram()) {
2054 if (open_return_path_on_source(s
)) {
2055 error_report("Unable to open return-path for postcopy");
2056 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
2057 MIGRATION_STATUS_FAILED
);
2058 migrate_fd_cleanup(s
);
2063 migrate_compress_threads_create();
2064 qemu_thread_create(&s
->thread
, "live_migration", migration_thread
, s
,
2065 QEMU_THREAD_JOINABLE
);
2066 s
->migration_thread_running
= true;
2069 PostcopyState
postcopy_state_get(void)
2071 return atomic_mb_read(&incoming_postcopy_state
);
2074 /* Set the state and return the old state */
2075 PostcopyState
postcopy_state_set(PostcopyState new_state
)
2077 return atomic_xchg(&incoming_postcopy_state
, new_state
);