Merge remote-tracking branch 'remotes/dgibson/tags/ppc-for-2.8-20161017' into staging
[qemu/ar7.git] / migration / migration.c
blob4d417b76cf8792f86d1998bea0034fea0037b856
1 /*
2 * QEMU live migration
4 * Copyright IBM, Corp. 2008
6 * Authors:
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"
27 #include "qemu/rcu.h"
28 #include "migration/block.h"
29 #include "migration/postcopy-ram.h"
30 #include "qemu/thread.h"
31 #include "qmp-commands.h"
32 #include "trace.h"
33 #include "qapi-event.h"
34 #include "qom/cpu.h"
35 #include "exec/memory.h"
36 #include "exec/address-spaces.h"
37 #include "io/channel-buffer.h"
38 #include "io/channel-tls.h"
40 #define MAX_THROTTLE (32 << 20) /* Migration transfer speed throttling */
42 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
43 * data. */
44 #define BUFFER_DELAY 100
45 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
47 /* Time in milliseconds we are allowed to stop the source,
48 * for sending the last part */
49 #define DEFAULT_MIGRATE_SET_DOWNTIME 300
51 /* Default compression thread count */
52 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
53 /* Default decompression thread count, usually decompression is at
54 * least 4 times as fast as compression.*/
55 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2
56 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */
57 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1
58 /* Define default autoconverge cpu throttle migration parameters */
59 #define DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL 20
60 #define DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT 10
62 /* Migration XBZRLE default cache size */
63 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
65 static NotifierList migration_state_notifiers =
66 NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
68 static bool deferred_incoming;
71 * Current state of incoming postcopy; note this is not part of
72 * MigrationIncomingState since it's state is used during cleanup
73 * at the end as MIS is being freed.
75 static PostcopyState incoming_postcopy_state;
77 /* When we add fault tolerance, we could have several
78 migrations at once. For now we don't need to add
79 dynamic creation of migration */
81 /* For outgoing */
82 MigrationState *migrate_get_current(void)
84 static bool once;
85 static MigrationState current_migration = {
86 .state = MIGRATION_STATUS_NONE,
87 .xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE,
88 .mbps = -1,
89 .parameters = {
90 .compress_level = DEFAULT_MIGRATE_COMPRESS_LEVEL,
91 .compress_threads = DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT,
92 .decompress_threads = DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT,
93 .cpu_throttle_initial = DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL,
94 .cpu_throttle_increment = DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT,
95 .max_bandwidth = MAX_THROTTLE,
96 .downtime_limit = DEFAULT_MIGRATE_SET_DOWNTIME,
100 if (!once) {
101 qemu_mutex_init(&current_migration.src_page_req_mutex);
102 once = true;
104 return &current_migration;
107 /* For incoming */
108 static MigrationIncomingState *mis_current;
110 MigrationIncomingState *migration_incoming_get_current(void)
112 return mis_current;
115 MigrationIncomingState *migration_incoming_state_new(QEMUFile* f)
117 mis_current = g_new0(MigrationIncomingState, 1);
118 mis_current->from_src_file = f;
119 mis_current->state = MIGRATION_STATUS_NONE;
120 QLIST_INIT(&mis_current->loadvm_handlers);
121 qemu_mutex_init(&mis_current->rp_mutex);
122 qemu_event_init(&mis_current->main_thread_load_event, false);
124 return mis_current;
127 void migration_incoming_state_destroy(void)
129 qemu_event_destroy(&mis_current->main_thread_load_event);
130 loadvm_free_handlers(mis_current);
131 g_free(mis_current);
132 mis_current = NULL;
136 typedef struct {
137 bool optional;
138 uint32_t size;
139 uint8_t runstate[100];
140 RunState state;
141 bool received;
142 } GlobalState;
144 static GlobalState global_state;
146 int global_state_store(void)
148 if (!runstate_store((char *)global_state.runstate,
149 sizeof(global_state.runstate))) {
150 error_report("runstate name too big: %s", global_state.runstate);
151 trace_migrate_state_too_big();
152 return -EINVAL;
154 return 0;
157 void global_state_store_running(void)
159 const char *state = RunState_lookup[RUN_STATE_RUNNING];
160 strncpy((char *)global_state.runstate,
161 state, sizeof(global_state.runstate));
164 static bool global_state_received(void)
166 return global_state.received;
169 static RunState global_state_get_runstate(void)
171 return global_state.state;
174 void global_state_set_optional(void)
176 global_state.optional = true;
179 static bool global_state_needed(void *opaque)
181 GlobalState *s = opaque;
182 char *runstate = (char *)s->runstate;
184 /* If it is not optional, it is mandatory */
186 if (s->optional == false) {
187 return true;
190 /* If state is running or paused, it is not needed */
192 if (strcmp(runstate, "running") == 0 ||
193 strcmp(runstate, "paused") == 0) {
194 return false;
197 /* for any other state it is needed */
198 return true;
201 static int global_state_post_load(void *opaque, int version_id)
203 GlobalState *s = opaque;
204 Error *local_err = NULL;
205 int r;
206 char *runstate = (char *)s->runstate;
208 s->received = true;
209 trace_migrate_global_state_post_load(runstate);
211 r = qapi_enum_parse(RunState_lookup, runstate, RUN_STATE__MAX,
212 -1, &local_err);
214 if (r == -1) {
215 if (local_err) {
216 error_report_err(local_err);
218 return -EINVAL;
220 s->state = r;
222 return 0;
225 static void global_state_pre_save(void *opaque)
227 GlobalState *s = opaque;
229 trace_migrate_global_state_pre_save((char *)s->runstate);
230 s->size = strlen((char *)s->runstate) + 1;
233 static const VMStateDescription vmstate_globalstate = {
234 .name = "globalstate",
235 .version_id = 1,
236 .minimum_version_id = 1,
237 .post_load = global_state_post_load,
238 .pre_save = global_state_pre_save,
239 .needed = global_state_needed,
240 .fields = (VMStateField[]) {
241 VMSTATE_UINT32(size, GlobalState),
242 VMSTATE_BUFFER(runstate, GlobalState),
243 VMSTATE_END_OF_LIST()
247 void register_global_state(void)
249 /* We would use it independently that we receive it */
250 strcpy((char *)&global_state.runstate, "");
251 global_state.received = false;
252 vmstate_register(NULL, 0, &vmstate_globalstate, &global_state);
255 static void migrate_generate_event(int new_state)
257 if (migrate_use_events()) {
258 qapi_event_send_migration(new_state, &error_abort);
263 * Called on -incoming with a defer: uri.
264 * The migration can be started later after any parameters have been
265 * changed.
267 static void deferred_incoming_migration(Error **errp)
269 if (deferred_incoming) {
270 error_setg(errp, "Incoming migration already deferred");
272 deferred_incoming = true;
275 /* Request a range of pages from the source VM at the given
276 * start address.
277 * rbname: Name of the RAMBlock to request the page in, if NULL it's the same
278 * as the last request (a name must have been given previously)
279 * Start: Address offset within the RB
280 * Len: Length in bytes required - must be a multiple of pagesize
282 void migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname,
283 ram_addr_t start, size_t len)
285 uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
286 size_t msglen = 12; /* start + len */
288 *(uint64_t *)bufc = cpu_to_be64((uint64_t)start);
289 *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len);
291 if (rbname) {
292 int rbname_len = strlen(rbname);
293 assert(rbname_len < 256);
295 bufc[msglen++] = rbname_len;
296 memcpy(bufc + msglen, rbname, rbname_len);
297 msglen += rbname_len;
298 migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES_ID, msglen, bufc);
299 } else {
300 migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES, msglen, bufc);
304 void qemu_start_incoming_migration(const char *uri, Error **errp)
306 const char *p;
308 qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort);
309 if (!strcmp(uri, "defer")) {
310 deferred_incoming_migration(errp);
311 } else if (strstart(uri, "tcp:", &p)) {
312 tcp_start_incoming_migration(p, errp);
313 #ifdef CONFIG_RDMA
314 } else if (strstart(uri, "rdma:", &p)) {
315 rdma_start_incoming_migration(p, errp);
316 #endif
317 } else if (strstart(uri, "exec:", &p)) {
318 exec_start_incoming_migration(p, errp);
319 } else if (strstart(uri, "unix:", &p)) {
320 unix_start_incoming_migration(p, errp);
321 } else if (strstart(uri, "fd:", &p)) {
322 fd_start_incoming_migration(p, errp);
323 } else {
324 error_setg(errp, "unknown migration protocol: %s", uri);
328 static void process_incoming_migration_bh(void *opaque)
330 Error *local_err = NULL;
331 MigrationIncomingState *mis = opaque;
333 /* Make sure all file formats flush their mutable metadata */
334 bdrv_invalidate_cache_all(&local_err);
335 if (local_err) {
336 migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
337 MIGRATION_STATUS_FAILED);
338 error_report_err(local_err);
339 migrate_decompress_threads_join();
340 exit(EXIT_FAILURE);
344 * This must happen after all error conditions are dealt with and
345 * we're sure the VM is going to be running on this host.
347 qemu_announce_self();
349 /* If global state section was not received or we are in running
350 state, we need to obey autostart. Any other state is set with
351 runstate_set. */
353 if (!global_state_received() ||
354 global_state_get_runstate() == RUN_STATE_RUNNING) {
355 if (autostart) {
356 vm_start();
357 } else {
358 runstate_set(RUN_STATE_PAUSED);
360 } else {
361 runstate_set(global_state_get_runstate());
363 migrate_decompress_threads_join();
365 * This must happen after any state changes since as soon as an external
366 * observer sees this event they might start to prod at the VM assuming
367 * it's ready to use.
369 migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
370 MIGRATION_STATUS_COMPLETED);
371 qemu_bh_delete(mis->bh);
372 migration_incoming_state_destroy();
375 static void process_incoming_migration_co(void *opaque)
377 QEMUFile *f = opaque;
378 MigrationIncomingState *mis;
379 PostcopyState ps;
380 int ret;
382 mis = migration_incoming_state_new(f);
383 postcopy_state_set(POSTCOPY_INCOMING_NONE);
384 migrate_set_state(&mis->state, MIGRATION_STATUS_NONE,
385 MIGRATION_STATUS_ACTIVE);
386 ret = qemu_loadvm_state(f);
388 ps = postcopy_state_get();
389 trace_process_incoming_migration_co_end(ret, ps);
390 if (ps != POSTCOPY_INCOMING_NONE) {
391 if (ps == POSTCOPY_INCOMING_ADVISE) {
393 * Where a migration had postcopy enabled (and thus went to advise)
394 * but managed to complete within the precopy period, we can use
395 * the normal exit.
397 postcopy_ram_incoming_cleanup(mis);
398 } else if (ret >= 0) {
400 * Postcopy was started, cleanup should happen at the end of the
401 * postcopy thread.
403 trace_process_incoming_migration_co_postcopy_end_main();
404 return;
406 /* Else if something went wrong then just fall out of the normal exit */
409 qemu_fclose(f);
410 free_xbzrle_decoded_buf();
412 if (ret < 0) {
413 migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
414 MIGRATION_STATUS_FAILED);
415 error_report("load of migration failed: %s", strerror(-ret));
416 migrate_decompress_threads_join();
417 exit(EXIT_FAILURE);
420 mis->bh = qemu_bh_new(process_incoming_migration_bh, mis);
421 qemu_bh_schedule(mis->bh);
424 void migration_fd_process_incoming(QEMUFile *f)
426 Coroutine *co = qemu_coroutine_create(process_incoming_migration_co, f);
428 migrate_decompress_threads_create();
429 qemu_file_set_blocking(f, false);
430 qemu_coroutine_enter(co);
434 void migration_channel_process_incoming(MigrationState *s,
435 QIOChannel *ioc)
437 trace_migration_set_incoming_channel(
438 ioc, object_get_typename(OBJECT(ioc)));
440 if (s->parameters.tls_creds &&
441 !object_dynamic_cast(OBJECT(ioc),
442 TYPE_QIO_CHANNEL_TLS)) {
443 Error *local_err = NULL;
444 migration_tls_channel_process_incoming(s, ioc, &local_err);
445 if (local_err) {
446 error_report_err(local_err);
448 } else {
449 QEMUFile *f = qemu_fopen_channel_input(ioc);
450 migration_fd_process_incoming(f);
455 void migration_channel_connect(MigrationState *s,
456 QIOChannel *ioc,
457 const char *hostname)
459 trace_migration_set_outgoing_channel(
460 ioc, object_get_typename(OBJECT(ioc)), hostname);
462 if (s->parameters.tls_creds &&
463 !object_dynamic_cast(OBJECT(ioc),
464 TYPE_QIO_CHANNEL_TLS)) {
465 Error *local_err = NULL;
466 migration_tls_channel_connect(s, ioc, hostname, &local_err);
467 if (local_err) {
468 migrate_fd_error(s, local_err);
469 error_free(local_err);
471 } else {
472 QEMUFile *f = qemu_fopen_channel_output(ioc);
474 s->to_dst_file = f;
476 migrate_fd_connect(s);
482 * Send a message on the return channel back to the source
483 * of the migration.
485 void migrate_send_rp_message(MigrationIncomingState *mis,
486 enum mig_rp_message_type message_type,
487 uint16_t len, void *data)
489 trace_migrate_send_rp_message((int)message_type, len);
490 qemu_mutex_lock(&mis->rp_mutex);
491 qemu_put_be16(mis->to_src_file, (unsigned int)message_type);
492 qemu_put_be16(mis->to_src_file, len);
493 qemu_put_buffer(mis->to_src_file, data, len);
494 qemu_fflush(mis->to_src_file);
495 qemu_mutex_unlock(&mis->rp_mutex);
499 * Send a 'SHUT' message on the return channel with the given value
500 * to indicate that we've finished with the RP. Non-0 value indicates
501 * error.
503 void migrate_send_rp_shut(MigrationIncomingState *mis,
504 uint32_t value)
506 uint32_t buf;
508 buf = cpu_to_be32(value);
509 migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf);
513 * Send a 'PONG' message on the return channel with the given value
514 * (normally in response to a 'PING')
516 void migrate_send_rp_pong(MigrationIncomingState *mis,
517 uint32_t value)
519 uint32_t buf;
521 buf = cpu_to_be32(value);
522 migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf);
525 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
527 MigrationCapabilityStatusList *head = NULL;
528 MigrationCapabilityStatusList *caps;
529 MigrationState *s = migrate_get_current();
530 int i;
532 caps = NULL; /* silence compiler warning */
533 for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
534 if (head == NULL) {
535 head = g_malloc0(sizeof(*caps));
536 caps = head;
537 } else {
538 caps->next = g_malloc0(sizeof(*caps));
539 caps = caps->next;
541 caps->value =
542 g_malloc(sizeof(*caps->value));
543 caps->value->capability = i;
544 caps->value->state = s->enabled_capabilities[i];
547 return head;
550 MigrationParameters *qmp_query_migrate_parameters(Error **errp)
552 MigrationParameters *params;
553 MigrationState *s = migrate_get_current();
555 params = g_malloc0(sizeof(*params));
556 params->has_compress_level = true;
557 params->compress_level = s->parameters.compress_level;
558 params->has_compress_threads = true;
559 params->compress_threads = s->parameters.compress_threads;
560 params->has_decompress_threads = true;
561 params->decompress_threads = s->parameters.decompress_threads;
562 params->has_cpu_throttle_initial = true;
563 params->cpu_throttle_initial = s->parameters.cpu_throttle_initial;
564 params->has_cpu_throttle_increment = true;
565 params->cpu_throttle_increment = s->parameters.cpu_throttle_increment;
566 params->has_tls_creds = !!s->parameters.tls_creds;
567 params->tls_creds = g_strdup(s->parameters.tls_creds);
568 params->has_tls_hostname = !!s->parameters.tls_hostname;
569 params->tls_hostname = g_strdup(s->parameters.tls_hostname);
570 params->has_max_bandwidth = true;
571 params->max_bandwidth = s->parameters.max_bandwidth;
572 params->has_downtime_limit = true;
573 params->downtime_limit = s->parameters.downtime_limit;
575 return params;
579 * Return true if we're already in the middle of a migration
580 * (i.e. any of the active or setup states)
582 static bool migration_is_setup_or_active(int state)
584 switch (state) {
585 case MIGRATION_STATUS_ACTIVE:
586 case MIGRATION_STATUS_POSTCOPY_ACTIVE:
587 case MIGRATION_STATUS_SETUP:
588 return true;
590 default:
591 return false;
596 static void get_xbzrle_cache_stats(MigrationInfo *info)
598 if (migrate_use_xbzrle()) {
599 info->has_xbzrle_cache = true;
600 info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
601 info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
602 info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred();
603 info->xbzrle_cache->pages = xbzrle_mig_pages_transferred();
604 info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss();
605 info->xbzrle_cache->cache_miss_rate = xbzrle_mig_cache_miss_rate();
606 info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow();
610 static void populate_ram_info(MigrationInfo *info, MigrationState *s)
612 info->has_ram = true;
613 info->ram = g_malloc0(sizeof(*info->ram));
614 info->ram->transferred = ram_bytes_transferred();
615 info->ram->total = ram_bytes_total();
616 info->ram->duplicate = dup_mig_pages_transferred();
617 info->ram->skipped = skipped_mig_pages_transferred();
618 info->ram->normal = norm_mig_pages_transferred();
619 info->ram->normal_bytes = norm_mig_bytes_transferred();
620 info->ram->mbps = s->mbps;
621 info->ram->dirty_sync_count = s->dirty_sync_count;
622 info->ram->postcopy_requests = s->postcopy_requests;
624 if (s->state != MIGRATION_STATUS_COMPLETED) {
625 info->ram->remaining = ram_bytes_remaining();
626 info->ram->dirty_pages_rate = s->dirty_pages_rate;
630 MigrationInfo *qmp_query_migrate(Error **errp)
632 MigrationInfo *info = g_malloc0(sizeof(*info));
633 MigrationState *s = migrate_get_current();
635 switch (s->state) {
636 case MIGRATION_STATUS_NONE:
637 /* no migration has happened ever */
638 break;
639 case MIGRATION_STATUS_SETUP:
640 info->has_status = true;
641 info->has_total_time = false;
642 break;
643 case MIGRATION_STATUS_ACTIVE:
644 case MIGRATION_STATUS_CANCELLING:
645 info->has_status = true;
646 info->has_total_time = true;
647 info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
648 - s->total_time;
649 info->has_expected_downtime = true;
650 info->expected_downtime = s->expected_downtime;
651 info->has_setup_time = true;
652 info->setup_time = s->setup_time;
654 populate_ram_info(info, s);
656 if (blk_mig_active()) {
657 info->has_disk = true;
658 info->disk = g_malloc0(sizeof(*info->disk));
659 info->disk->transferred = blk_mig_bytes_transferred();
660 info->disk->remaining = blk_mig_bytes_remaining();
661 info->disk->total = blk_mig_bytes_total();
664 if (cpu_throttle_active()) {
665 info->has_cpu_throttle_percentage = true;
666 info->cpu_throttle_percentage = cpu_throttle_get_percentage();
669 get_xbzrle_cache_stats(info);
670 break;
671 case MIGRATION_STATUS_POSTCOPY_ACTIVE:
672 /* Mostly the same as active; TODO add some postcopy stats */
673 info->has_status = true;
674 info->has_total_time = true;
675 info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
676 - s->total_time;
677 info->has_expected_downtime = true;
678 info->expected_downtime = s->expected_downtime;
679 info->has_setup_time = true;
680 info->setup_time = s->setup_time;
682 populate_ram_info(info, s);
684 if (blk_mig_active()) {
685 info->has_disk = true;
686 info->disk = g_malloc0(sizeof(*info->disk));
687 info->disk->transferred = blk_mig_bytes_transferred();
688 info->disk->remaining = blk_mig_bytes_remaining();
689 info->disk->total = blk_mig_bytes_total();
692 get_xbzrle_cache_stats(info);
693 break;
694 case MIGRATION_STATUS_COMPLETED:
695 get_xbzrle_cache_stats(info);
697 info->has_status = true;
698 info->has_total_time = true;
699 info->total_time = s->total_time;
700 info->has_downtime = true;
701 info->downtime = s->downtime;
702 info->has_setup_time = true;
703 info->setup_time = s->setup_time;
705 populate_ram_info(info, s);
706 break;
707 case MIGRATION_STATUS_FAILED:
708 info->has_status = true;
709 if (s->error) {
710 info->has_error_desc = true;
711 info->error_desc = g_strdup(error_get_pretty(s->error));
713 break;
714 case MIGRATION_STATUS_CANCELLED:
715 info->has_status = true;
716 break;
718 info->status = s->state;
720 return info;
723 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
724 Error **errp)
726 MigrationState *s = migrate_get_current();
727 MigrationCapabilityStatusList *cap;
728 bool old_postcopy_cap = migrate_postcopy_ram();
730 if (migration_is_setup_or_active(s->state)) {
731 error_setg(errp, QERR_MIGRATION_ACTIVE);
732 return;
735 for (cap = params; cap; cap = cap->next) {
736 s->enabled_capabilities[cap->value->capability] = cap->value->state;
739 if (migrate_postcopy_ram()) {
740 if (migrate_use_compression()) {
741 /* The decompression threads asynchronously write into RAM
742 * rather than use the atomic copies needed to avoid
743 * userfaulting. It should be possible to fix the decompression
744 * threads for compatibility in future.
746 error_report("Postcopy is not currently compatible with "
747 "compression");
748 s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] =
749 false;
751 /* This check is reasonably expensive, so only when it's being
752 * set the first time, also it's only the destination that needs
753 * special support.
755 if (!old_postcopy_cap && runstate_check(RUN_STATE_INMIGRATE) &&
756 !postcopy_ram_supported_by_host()) {
757 /* postcopy_ram_supported_by_host will have emitted a more
758 * detailed message
760 error_report("Postcopy is not supported");
761 s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] =
762 false;
767 void qmp_migrate_set_parameters(MigrationParameters *params, Error **errp)
769 MigrationState *s = migrate_get_current();
771 if (params->has_compress_level &&
772 (params->compress_level < 0 || params->compress_level > 9)) {
773 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level",
774 "is invalid, it should be in the range of 0 to 9");
775 return;
777 if (params->has_compress_threads &&
778 (params->compress_threads < 1 || params->compress_threads > 255)) {
779 error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
780 "compress_threads",
781 "is invalid, it should be in the range of 1 to 255");
782 return;
784 if (params->has_decompress_threads &&
785 (params->decompress_threads < 1 || params->decompress_threads > 255)) {
786 error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
787 "decompress_threads",
788 "is invalid, it should be in the range of 1 to 255");
789 return;
791 if (params->has_cpu_throttle_initial &&
792 (params->cpu_throttle_initial < 1 ||
793 params->cpu_throttle_initial > 99)) {
794 error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
795 "cpu_throttle_initial",
796 "an integer in the range of 1 to 99");
797 return;
799 if (params->has_cpu_throttle_increment &&
800 (params->cpu_throttle_increment < 1 ||
801 params->cpu_throttle_increment > 99)) {
802 error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
803 "cpu_throttle_increment",
804 "an integer in the range of 1 to 99");
805 return;
807 if (params->has_max_bandwidth &&
808 (params->max_bandwidth < 0 || params->max_bandwidth > SIZE_MAX)) {
809 error_setg(errp, "Parameter 'max_bandwidth' expects an integer in the"
810 " range of 0 to %zu bytes/second", SIZE_MAX);
811 return;
813 if (params->has_downtime_limit &&
814 (params->downtime_limit < 0 || params->downtime_limit > 2000000)) {
815 error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
816 "downtime_limit",
817 "an integer in the range of 0 to 2000000 milliseconds");
818 return;
821 if (params->has_compress_level) {
822 s->parameters.compress_level = params->compress_level;
824 if (params->has_compress_threads) {
825 s->parameters.compress_threads = params->compress_threads;
827 if (params->has_decompress_threads) {
828 s->parameters.decompress_threads = params->decompress_threads;
830 if (params->has_cpu_throttle_initial) {
831 s->parameters.cpu_throttle_initial = params->cpu_throttle_initial;
833 if (params->has_cpu_throttle_increment) {
834 s->parameters.cpu_throttle_increment = params->cpu_throttle_increment;
836 if (params->has_tls_creds) {
837 g_free(s->parameters.tls_creds);
838 s->parameters.tls_creds = g_strdup(params->tls_creds);
840 if (params->has_tls_hostname) {
841 g_free(s->parameters.tls_hostname);
842 s->parameters.tls_hostname = g_strdup(params->tls_hostname);
844 if (params->has_max_bandwidth) {
845 s->parameters.max_bandwidth = params->max_bandwidth;
846 if (s->to_dst_file) {
847 qemu_file_set_rate_limit(s->to_dst_file,
848 s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
851 if (params->has_downtime_limit) {
852 s->parameters.downtime_limit = params->downtime_limit;
857 void qmp_migrate_start_postcopy(Error **errp)
859 MigrationState *s = migrate_get_current();
861 if (!migrate_postcopy_ram()) {
862 error_setg(errp, "Enable postcopy with migrate_set_capability before"
863 " the start of migration");
864 return;
867 if (s->state == MIGRATION_STATUS_NONE) {
868 error_setg(errp, "Postcopy must be started after migration has been"
869 " started");
870 return;
873 * we don't error if migration has finished since that would be racy
874 * with issuing this command.
876 atomic_set(&s->start_postcopy, true);
879 /* shared migration helpers */
881 void migrate_set_state(int *state, int old_state, int new_state)
883 if (atomic_cmpxchg(state, old_state, new_state) == old_state) {
884 trace_migrate_set_state(new_state);
885 migrate_generate_event(new_state);
889 static void migrate_fd_cleanup(void *opaque)
891 MigrationState *s = opaque;
893 qemu_bh_delete(s->cleanup_bh);
894 s->cleanup_bh = NULL;
896 flush_page_queue(s);
898 if (s->to_dst_file) {
899 trace_migrate_fd_cleanup();
900 qemu_mutex_unlock_iothread();
901 if (s->migration_thread_running) {
902 qemu_thread_join(&s->thread);
903 s->migration_thread_running = false;
905 qemu_mutex_lock_iothread();
907 migrate_compress_threads_join();
908 qemu_fclose(s->to_dst_file);
909 s->to_dst_file = NULL;
912 assert((s->state != MIGRATION_STATUS_ACTIVE) &&
913 (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE));
915 if (s->state == MIGRATION_STATUS_CANCELLING) {
916 migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING,
917 MIGRATION_STATUS_CANCELLED);
920 notifier_list_notify(&migration_state_notifiers, s);
923 void migrate_fd_error(MigrationState *s, const Error *error)
925 trace_migrate_fd_error(error ? error_get_pretty(error) : "");
926 assert(s->to_dst_file == NULL);
927 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
928 MIGRATION_STATUS_FAILED);
929 if (!s->error) {
930 s->error = error_copy(error);
932 notifier_list_notify(&migration_state_notifiers, s);
935 static void migrate_fd_cancel(MigrationState *s)
937 int old_state ;
938 QEMUFile *f = migrate_get_current()->to_dst_file;
939 trace_migrate_fd_cancel();
941 if (s->rp_state.from_dst_file) {
942 /* shutdown the rp socket, so causing the rp thread to shutdown */
943 qemu_file_shutdown(s->rp_state.from_dst_file);
946 do {
947 old_state = s->state;
948 if (!migration_is_setup_or_active(old_state)) {
949 break;
951 migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING);
952 } while (s->state != MIGRATION_STATUS_CANCELLING);
955 * If we're unlucky the migration code might be stuck somewhere in a
956 * send/write while the network has failed and is waiting to timeout;
957 * if we've got shutdown(2) available then we can force it to quit.
958 * The outgoing qemu file gets closed in migrate_fd_cleanup that is
959 * called in a bh, so there is no race against this cancel.
961 if (s->state == MIGRATION_STATUS_CANCELLING && f) {
962 qemu_file_shutdown(f);
966 void add_migration_state_change_notifier(Notifier *notify)
968 notifier_list_add(&migration_state_notifiers, notify);
971 void remove_migration_state_change_notifier(Notifier *notify)
973 notifier_remove(notify);
976 bool migration_in_setup(MigrationState *s)
978 return s->state == MIGRATION_STATUS_SETUP;
981 bool migration_has_finished(MigrationState *s)
983 return s->state == MIGRATION_STATUS_COMPLETED;
986 bool migration_has_failed(MigrationState *s)
988 return (s->state == MIGRATION_STATUS_CANCELLED ||
989 s->state == MIGRATION_STATUS_FAILED);
992 bool migration_in_postcopy(MigrationState *s)
994 return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
997 bool migration_in_postcopy_after_devices(MigrationState *s)
999 return migration_in_postcopy(s) && s->postcopy_after_devices;
1002 MigrationState *migrate_init(const MigrationParams *params)
1004 MigrationState *s = migrate_get_current();
1007 * Reinitialise all migration state, except
1008 * parameters/capabilities that the user set, and
1009 * locks.
1011 s->bytes_xfer = 0;
1012 s->xfer_limit = 0;
1013 s->cleanup_bh = 0;
1014 s->to_dst_file = NULL;
1015 s->state = MIGRATION_STATUS_NONE;
1016 s->params = *params;
1017 s->rp_state.from_dst_file = NULL;
1018 s->rp_state.error = false;
1019 s->mbps = 0.0;
1020 s->downtime = 0;
1021 s->expected_downtime = 0;
1022 s->dirty_pages_rate = 0;
1023 s->dirty_bytes_rate = 0;
1024 s->setup_time = 0;
1025 s->dirty_sync_count = 0;
1026 s->start_postcopy = false;
1027 s->postcopy_after_devices = false;
1028 s->postcopy_requests = 0;
1029 s->migration_thread_running = false;
1030 s->last_req_rb = NULL;
1031 error_free(s->error);
1032 s->error = NULL;
1034 migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP);
1036 QSIMPLEQ_INIT(&s->src_page_requests);
1038 s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1039 return s;
1042 static GSList *migration_blockers;
1044 void migrate_add_blocker(Error *reason)
1046 migration_blockers = g_slist_prepend(migration_blockers, reason);
1049 void migrate_del_blocker(Error *reason)
1051 migration_blockers = g_slist_remove(migration_blockers, reason);
1054 void qmp_migrate_incoming(const char *uri, Error **errp)
1056 Error *local_err = NULL;
1057 static bool once = true;
1059 if (!deferred_incoming) {
1060 error_setg(errp, "For use with '-incoming defer'");
1061 return;
1063 if (!once) {
1064 error_setg(errp, "The incoming migration has already been started");
1067 qemu_start_incoming_migration(uri, &local_err);
1069 if (local_err) {
1070 error_propagate(errp, local_err);
1071 return;
1074 once = false;
1077 bool migration_is_blocked(Error **errp)
1079 if (qemu_savevm_state_blocked(errp)) {
1080 return true;
1083 if (migration_blockers) {
1084 *errp = error_copy(migration_blockers->data);
1085 return true;
1088 return false;
1091 void qmp_migrate(const char *uri, bool has_blk, bool blk,
1092 bool has_inc, bool inc, bool has_detach, bool detach,
1093 Error **errp)
1095 Error *local_err = NULL;
1096 MigrationState *s = migrate_get_current();
1097 MigrationParams params;
1098 const char *p;
1100 params.blk = has_blk && blk;
1101 params.shared = has_inc && inc;
1103 if (migration_is_setup_or_active(s->state) ||
1104 s->state == MIGRATION_STATUS_CANCELLING) {
1105 error_setg(errp, QERR_MIGRATION_ACTIVE);
1106 return;
1108 if (runstate_check(RUN_STATE_INMIGRATE)) {
1109 error_setg(errp, "Guest is waiting for an incoming migration");
1110 return;
1113 if (migration_is_blocked(errp)) {
1114 return;
1117 s = migrate_init(&params);
1119 if (strstart(uri, "tcp:", &p)) {
1120 tcp_start_outgoing_migration(s, p, &local_err);
1121 #ifdef CONFIG_RDMA
1122 } else if (strstart(uri, "rdma:", &p)) {
1123 rdma_start_outgoing_migration(s, p, &local_err);
1124 #endif
1125 } else if (strstart(uri, "exec:", &p)) {
1126 exec_start_outgoing_migration(s, p, &local_err);
1127 } else if (strstart(uri, "unix:", &p)) {
1128 unix_start_outgoing_migration(s, p, &local_err);
1129 } else if (strstart(uri, "fd:", &p)) {
1130 fd_start_outgoing_migration(s, p, &local_err);
1131 } else {
1132 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri",
1133 "a valid migration protocol");
1134 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1135 MIGRATION_STATUS_FAILED);
1136 return;
1139 if (local_err) {
1140 migrate_fd_error(s, local_err);
1141 error_propagate(errp, local_err);
1142 return;
1146 void qmp_migrate_cancel(Error **errp)
1148 migrate_fd_cancel(migrate_get_current());
1151 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
1153 MigrationState *s = migrate_get_current();
1154 int64_t new_size;
1156 /* Check for truncation */
1157 if (value != (size_t)value) {
1158 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1159 "exceeding address space");
1160 return;
1163 /* Cache should not be larger than guest ram size */
1164 if (value > ram_bytes_total()) {
1165 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1166 "exceeds guest ram size ");
1167 return;
1170 new_size = xbzrle_cache_resize(value);
1171 if (new_size < 0) {
1172 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1173 "is smaller than page size");
1174 return;
1177 s->xbzrle_cache_size = new_size;
1180 int64_t qmp_query_migrate_cache_size(Error **errp)
1182 return migrate_xbzrle_cache_size();
1185 void qmp_migrate_set_speed(int64_t value, Error **errp)
1187 MigrationParameters p = {
1188 .has_max_bandwidth = true,
1189 .max_bandwidth = value,
1192 qmp_migrate_set_parameters(&p, errp);
1195 void qmp_migrate_set_downtime(double value, Error **errp)
1197 value *= 1000; /* Convert to milliseconds */
1198 value = MAX(0, MIN(INT64_MAX, value));
1200 MigrationParameters p = {
1201 .has_downtime_limit = true,
1202 .downtime_limit = value,
1205 qmp_migrate_set_parameters(&p, errp);
1208 bool migrate_postcopy_ram(void)
1210 MigrationState *s;
1212 s = migrate_get_current();
1214 return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM];
1217 bool migrate_auto_converge(void)
1219 MigrationState *s;
1221 s = migrate_get_current();
1223 return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
1226 bool migrate_zero_blocks(void)
1228 MigrationState *s;
1230 s = migrate_get_current();
1232 return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
1235 bool migrate_use_compression(void)
1237 MigrationState *s;
1239 s = migrate_get_current();
1241 return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS];
1244 int migrate_compress_level(void)
1246 MigrationState *s;
1248 s = migrate_get_current();
1250 return s->parameters.compress_level;
1253 int migrate_compress_threads(void)
1255 MigrationState *s;
1257 s = migrate_get_current();
1259 return s->parameters.compress_threads;
1262 int migrate_decompress_threads(void)
1264 MigrationState *s;
1266 s = migrate_get_current();
1268 return s->parameters.decompress_threads;
1271 bool migrate_use_events(void)
1273 MigrationState *s;
1275 s = migrate_get_current();
1277 return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS];
1280 int migrate_use_xbzrle(void)
1282 MigrationState *s;
1284 s = migrate_get_current();
1286 return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
1289 int64_t migrate_xbzrle_cache_size(void)
1291 MigrationState *s;
1293 s = migrate_get_current();
1295 return s->xbzrle_cache_size;
1298 /* migration thread support */
1300 * Something bad happened to the RP stream, mark an error
1301 * The caller shall print or trace something to indicate why
1303 static void mark_source_rp_bad(MigrationState *s)
1305 s->rp_state.error = true;
1308 static struct rp_cmd_args {
1309 ssize_t len; /* -1 = variable */
1310 const char *name;
1311 } rp_cmd_args[] = {
1312 [MIG_RP_MSG_INVALID] = { .len = -1, .name = "INVALID" },
1313 [MIG_RP_MSG_SHUT] = { .len = 4, .name = "SHUT" },
1314 [MIG_RP_MSG_PONG] = { .len = 4, .name = "PONG" },
1315 [MIG_RP_MSG_REQ_PAGES] = { .len = 12, .name = "REQ_PAGES" },
1316 [MIG_RP_MSG_REQ_PAGES_ID] = { .len = -1, .name = "REQ_PAGES_ID" },
1317 [MIG_RP_MSG_MAX] = { .len = -1, .name = "MAX" },
1321 * Process a request for pages received on the return path,
1322 * We're allowed to send more than requested (e.g. to round to our page size)
1323 * and we don't need to send pages that have already been sent.
1325 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname,
1326 ram_addr_t start, size_t len)
1328 long our_host_ps = getpagesize();
1330 trace_migrate_handle_rp_req_pages(rbname, start, len);
1333 * Since we currently insist on matching page sizes, just sanity check
1334 * we're being asked for whole host pages.
1336 if (start & (our_host_ps-1) ||
1337 (len & (our_host_ps-1))) {
1338 error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1339 " len: %zd", __func__, start, len);
1340 mark_source_rp_bad(ms);
1341 return;
1344 if (ram_save_queue_pages(ms, rbname, start, len)) {
1345 mark_source_rp_bad(ms);
1350 * Handles messages sent on the return path towards the source VM
1353 static void *source_return_path_thread(void *opaque)
1355 MigrationState *ms = opaque;
1356 QEMUFile *rp = ms->rp_state.from_dst_file;
1357 uint16_t header_len, header_type;
1358 uint8_t buf[512];
1359 uint32_t tmp32, sibling_error;
1360 ram_addr_t start = 0; /* =0 to silence warning */
1361 size_t len = 0, expected_len;
1362 int res;
1364 trace_source_return_path_thread_entry();
1365 while (!ms->rp_state.error && !qemu_file_get_error(rp) &&
1366 migration_is_setup_or_active(ms->state)) {
1367 trace_source_return_path_thread_loop_top();
1368 header_type = qemu_get_be16(rp);
1369 header_len = qemu_get_be16(rp);
1371 if (header_type >= MIG_RP_MSG_MAX ||
1372 header_type == MIG_RP_MSG_INVALID) {
1373 error_report("RP: Received invalid message 0x%04x length 0x%04x",
1374 header_type, header_len);
1375 mark_source_rp_bad(ms);
1376 goto out;
1379 if ((rp_cmd_args[header_type].len != -1 &&
1380 header_len != rp_cmd_args[header_type].len) ||
1381 header_len > sizeof(buf)) {
1382 error_report("RP: Received '%s' message (0x%04x) with"
1383 "incorrect length %d expecting %zu",
1384 rp_cmd_args[header_type].name, header_type, header_len,
1385 (size_t)rp_cmd_args[header_type].len);
1386 mark_source_rp_bad(ms);
1387 goto out;
1390 /* We know we've got a valid header by this point */
1391 res = qemu_get_buffer(rp, buf, header_len);
1392 if (res != header_len) {
1393 error_report("RP: Failed reading data for message 0x%04x"
1394 " read %d expected %d",
1395 header_type, res, header_len);
1396 mark_source_rp_bad(ms);
1397 goto out;
1400 /* OK, we have the message and the data */
1401 switch (header_type) {
1402 case MIG_RP_MSG_SHUT:
1403 sibling_error = ldl_be_p(buf);
1404 trace_source_return_path_thread_shut(sibling_error);
1405 if (sibling_error) {
1406 error_report("RP: Sibling indicated error %d", sibling_error);
1407 mark_source_rp_bad(ms);
1410 * We'll let the main thread deal with closing the RP
1411 * we could do a shutdown(2) on it, but we're the only user
1412 * anyway, so there's nothing gained.
1414 goto out;
1416 case MIG_RP_MSG_PONG:
1417 tmp32 = ldl_be_p(buf);
1418 trace_source_return_path_thread_pong(tmp32);
1419 break;
1421 case MIG_RP_MSG_REQ_PAGES:
1422 start = ldq_be_p(buf);
1423 len = ldl_be_p(buf + 8);
1424 migrate_handle_rp_req_pages(ms, NULL, start, len);
1425 break;
1427 case MIG_RP_MSG_REQ_PAGES_ID:
1428 expected_len = 12 + 1; /* header + termination */
1430 if (header_len >= expected_len) {
1431 start = ldq_be_p(buf);
1432 len = ldl_be_p(buf + 8);
1433 /* Now we expect an idstr */
1434 tmp32 = buf[12]; /* Length of the following idstr */
1435 buf[13 + tmp32] = '\0';
1436 expected_len += tmp32;
1438 if (header_len != expected_len) {
1439 error_report("RP: Req_Page_id with length %d expecting %zd",
1440 header_len, expected_len);
1441 mark_source_rp_bad(ms);
1442 goto out;
1444 migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len);
1445 break;
1447 default:
1448 break;
1451 if (qemu_file_get_error(rp)) {
1452 trace_source_return_path_thread_bad_end();
1453 mark_source_rp_bad(ms);
1456 trace_source_return_path_thread_end();
1457 out:
1458 ms->rp_state.from_dst_file = NULL;
1459 qemu_fclose(rp);
1460 return NULL;
1463 static int open_return_path_on_source(MigrationState *ms)
1466 ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file);
1467 if (!ms->rp_state.from_dst_file) {
1468 return -1;
1471 trace_open_return_path_on_source();
1472 qemu_thread_create(&ms->rp_state.rp_thread, "return path",
1473 source_return_path_thread, ms, QEMU_THREAD_JOINABLE);
1475 trace_open_return_path_on_source_continue();
1477 return 0;
1480 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
1481 static int await_return_path_close_on_source(MigrationState *ms)
1484 * If this is a normal exit then the destination will send a SHUT and the
1485 * rp_thread will exit, however if there's an error we need to cause
1486 * it to exit.
1488 if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
1490 * shutdown(2), if we have it, will cause it to unblock if it's stuck
1491 * waiting for the destination.
1493 qemu_file_shutdown(ms->rp_state.from_dst_file);
1494 mark_source_rp_bad(ms);
1496 trace_await_return_path_close_on_source_joining();
1497 qemu_thread_join(&ms->rp_state.rp_thread);
1498 trace_await_return_path_close_on_source_close();
1499 return ms->rp_state.error;
1503 * Switch from normal iteration to postcopy
1504 * Returns non-0 on error
1506 static int postcopy_start(MigrationState *ms, bool *old_vm_running)
1508 int ret;
1509 QIOChannelBuffer *bioc;
1510 QEMUFile *fb;
1511 int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1512 migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE,
1513 MIGRATION_STATUS_POSTCOPY_ACTIVE);
1515 trace_postcopy_start();
1516 qemu_mutex_lock_iothread();
1517 trace_postcopy_start_set_run();
1519 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1520 *old_vm_running = runstate_is_running();
1521 global_state_store();
1522 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1523 if (ret < 0) {
1524 goto fail;
1527 ret = bdrv_inactivate_all();
1528 if (ret < 0) {
1529 goto fail;
1533 * Cause any non-postcopiable, but iterative devices to
1534 * send out their final data.
1536 qemu_savevm_state_complete_precopy(ms->to_dst_file, true);
1539 * in Finish migrate and with the io-lock held everything should
1540 * be quiet, but we've potentially still got dirty pages and we
1541 * need to tell the destination to throw any pages it's already received
1542 * that are dirty
1544 if (ram_postcopy_send_discard_bitmap(ms)) {
1545 error_report("postcopy send discard bitmap failed");
1546 goto fail;
1550 * send rest of state - note things that are doing postcopy
1551 * will notice we're in POSTCOPY_ACTIVE and not actually
1552 * wrap their state up here
1554 qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX);
1555 /* Ping just for debugging, helps line traces up */
1556 qemu_savevm_send_ping(ms->to_dst_file, 2);
1559 * While loading the device state we may trigger page transfer
1560 * requests and the fd must be free to process those, and thus
1561 * the destination must read the whole device state off the fd before
1562 * it starts processing it. Unfortunately the ad-hoc migration format
1563 * doesn't allow the destination to know the size to read without fully
1564 * parsing it through each devices load-state code (especially the open
1565 * coded devices that use get/put).
1566 * So we wrap the device state up in a package with a length at the start;
1567 * to do this we use a qemu_buf to hold the whole of the device state.
1569 bioc = qio_channel_buffer_new(4096);
1570 fb = qemu_fopen_channel_output(QIO_CHANNEL(bioc));
1571 object_unref(OBJECT(bioc));
1574 * Make sure the receiver can get incoming pages before we send the rest
1575 * of the state
1577 qemu_savevm_send_postcopy_listen(fb);
1579 qemu_savevm_state_complete_precopy(fb, false);
1580 qemu_savevm_send_ping(fb, 3);
1582 qemu_savevm_send_postcopy_run(fb);
1584 /* <><> end of stuff going into the package */
1586 /* Now send that blob */
1587 if (qemu_savevm_send_packaged(ms->to_dst_file, bioc->data, bioc->usage)) {
1588 goto fail_closefb;
1590 qemu_fclose(fb);
1592 /* Send a notify to give a chance for anything that needs to happen
1593 * at the transition to postcopy and after the device state; in particular
1594 * spice needs to trigger a transition now
1596 ms->postcopy_after_devices = true;
1597 notifier_list_notify(&migration_state_notifiers, ms);
1599 ms->downtime = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop;
1601 qemu_mutex_unlock_iothread();
1604 * Although this ping is just for debug, it could potentially be
1605 * used for getting a better measurement of downtime at the source.
1607 qemu_savevm_send_ping(ms->to_dst_file, 4);
1609 ret = qemu_file_get_error(ms->to_dst_file);
1610 if (ret) {
1611 error_report("postcopy_start: Migration stream errored");
1612 migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1613 MIGRATION_STATUS_FAILED);
1616 return ret;
1618 fail_closefb:
1619 qemu_fclose(fb);
1620 fail:
1621 migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1622 MIGRATION_STATUS_FAILED);
1623 qemu_mutex_unlock_iothread();
1624 return -1;
1628 * migration_completion: Used by migration_thread when there's not much left.
1629 * The caller 'breaks' the loop when this returns.
1631 * @s: Current migration state
1632 * @current_active_state: The migration state we expect to be in
1633 * @*old_vm_running: Pointer to old_vm_running flag
1634 * @*start_time: Pointer to time to update
1636 static void migration_completion(MigrationState *s, int current_active_state,
1637 bool *old_vm_running,
1638 int64_t *start_time)
1640 int ret;
1642 if (s->state == MIGRATION_STATUS_ACTIVE) {
1643 qemu_mutex_lock_iothread();
1644 *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1645 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1646 *old_vm_running = runstate_is_running();
1647 ret = global_state_store();
1649 if (!ret) {
1650 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1651 if (ret >= 0) {
1652 ret = bdrv_inactivate_all();
1654 if (ret >= 0) {
1655 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX);
1656 qemu_savevm_state_complete_precopy(s->to_dst_file, false);
1659 qemu_mutex_unlock_iothread();
1661 if (ret < 0) {
1662 goto fail;
1664 } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1665 trace_migration_completion_postcopy_end();
1667 qemu_savevm_state_complete_postcopy(s->to_dst_file);
1668 trace_migration_completion_postcopy_end_after_complete();
1672 * If rp was opened we must clean up the thread before
1673 * cleaning everything else up (since if there are no failures
1674 * it will wait for the destination to send it's status in
1675 * a SHUT command).
1676 * Postcopy opens rp if enabled (even if it's not avtivated)
1678 if (migrate_postcopy_ram()) {
1679 int rp_error;
1680 trace_migration_completion_postcopy_end_before_rp();
1681 rp_error = await_return_path_close_on_source(s);
1682 trace_migration_completion_postcopy_end_after_rp(rp_error);
1683 if (rp_error) {
1684 goto fail_invalidate;
1688 if (qemu_file_get_error(s->to_dst_file)) {
1689 trace_migration_completion_file_err();
1690 goto fail_invalidate;
1693 migrate_set_state(&s->state, current_active_state,
1694 MIGRATION_STATUS_COMPLETED);
1695 return;
1697 fail_invalidate:
1698 /* If not doing postcopy, vm_start() will be called: let's regain
1699 * control on images.
1701 if (s->state == MIGRATION_STATUS_ACTIVE) {
1702 Error *local_err = NULL;
1704 bdrv_invalidate_cache_all(&local_err);
1705 if (local_err) {
1706 error_report_err(local_err);
1710 fail:
1711 migrate_set_state(&s->state, current_active_state,
1712 MIGRATION_STATUS_FAILED);
1716 * Master migration thread on the source VM.
1717 * It drives the migration and pumps the data down the outgoing channel.
1719 static void *migration_thread(void *opaque)
1721 MigrationState *s = opaque;
1722 /* Used by the bandwidth calcs, updated later */
1723 int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1724 int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
1725 int64_t initial_bytes = 0;
1726 int64_t max_size = 0;
1727 int64_t start_time = initial_time;
1728 int64_t end_time;
1729 bool old_vm_running = false;
1730 bool entered_postcopy = false;
1731 /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */
1732 enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE;
1734 rcu_register_thread();
1736 qemu_savevm_state_header(s->to_dst_file);
1738 if (migrate_postcopy_ram()) {
1739 /* Now tell the dest that it should open its end so it can reply */
1740 qemu_savevm_send_open_return_path(s->to_dst_file);
1742 /* And do a ping that will make stuff easier to debug */
1743 qemu_savevm_send_ping(s->to_dst_file, 1);
1746 * Tell the destination that we *might* want to do postcopy later;
1747 * if the other end can't do postcopy it should fail now, nice and
1748 * early.
1750 qemu_savevm_send_postcopy_advise(s->to_dst_file);
1753 qemu_savevm_state_begin(s->to_dst_file, &s->params);
1755 s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
1756 current_active_state = MIGRATION_STATUS_ACTIVE;
1757 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1758 MIGRATION_STATUS_ACTIVE);
1760 trace_migration_thread_setup_complete();
1762 while (s->state == MIGRATION_STATUS_ACTIVE ||
1763 s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1764 int64_t current_time;
1765 uint64_t pending_size;
1767 if (!qemu_file_rate_limit(s->to_dst_file)) {
1768 uint64_t pend_post, pend_nonpost;
1770 qemu_savevm_state_pending(s->to_dst_file, max_size, &pend_nonpost,
1771 &pend_post);
1772 pending_size = pend_nonpost + pend_post;
1773 trace_migrate_pending(pending_size, max_size,
1774 pend_post, pend_nonpost);
1775 if (pending_size && pending_size >= max_size) {
1776 /* Still a significant amount to transfer */
1778 if (migrate_postcopy_ram() &&
1779 s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE &&
1780 pend_nonpost <= max_size &&
1781 atomic_read(&s->start_postcopy)) {
1783 if (!postcopy_start(s, &old_vm_running)) {
1784 current_active_state = MIGRATION_STATUS_POSTCOPY_ACTIVE;
1785 entered_postcopy = true;
1788 continue;
1790 /* Just another iteration step */
1791 qemu_savevm_state_iterate(s->to_dst_file, entered_postcopy);
1792 } else {
1793 trace_migration_thread_low_pending(pending_size);
1794 migration_completion(s, current_active_state,
1795 &old_vm_running, &start_time);
1796 break;
1800 if (qemu_file_get_error(s->to_dst_file)) {
1801 migrate_set_state(&s->state, current_active_state,
1802 MIGRATION_STATUS_FAILED);
1803 trace_migration_thread_file_err();
1804 break;
1806 current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1807 if (current_time >= initial_time + BUFFER_DELAY) {
1808 uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) -
1809 initial_bytes;
1810 uint64_t time_spent = current_time - initial_time;
1811 double bandwidth = (double)transferred_bytes / time_spent;
1812 max_size = bandwidth * s->parameters.downtime_limit;
1814 s->mbps = (((double) transferred_bytes * 8.0) /
1815 ((double) time_spent / 1000.0)) / 1000.0 / 1000.0;
1817 trace_migrate_transferred(transferred_bytes, time_spent,
1818 bandwidth, max_size);
1819 /* if we haven't sent anything, we don't want to recalculate
1820 10000 is a small enough number for our purposes */
1821 if (s->dirty_bytes_rate && transferred_bytes > 10000) {
1822 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
1825 qemu_file_reset_rate_limit(s->to_dst_file);
1826 initial_time = current_time;
1827 initial_bytes = qemu_ftell(s->to_dst_file);
1829 if (qemu_file_rate_limit(s->to_dst_file)) {
1830 /* usleep expects microseconds */
1831 g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
1835 trace_migration_thread_after_loop();
1836 /* If we enabled cpu throttling for auto-converge, turn it off. */
1837 cpu_throttle_stop();
1838 end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1840 qemu_mutex_lock_iothread();
1841 qemu_savevm_state_cleanup();
1842 if (s->state == MIGRATION_STATUS_COMPLETED) {
1843 uint64_t transferred_bytes = qemu_ftell(s->to_dst_file);
1844 s->total_time = end_time - s->total_time;
1845 if (!entered_postcopy) {
1846 s->downtime = end_time - start_time;
1848 if (s->total_time) {
1849 s->mbps = (((double) transferred_bytes * 8.0) /
1850 ((double) s->total_time)) / 1000;
1852 runstate_set(RUN_STATE_POSTMIGRATE);
1853 } else {
1854 if (old_vm_running && !entered_postcopy) {
1855 vm_start();
1856 } else {
1857 if (runstate_check(RUN_STATE_FINISH_MIGRATE)) {
1858 runstate_set(RUN_STATE_POSTMIGRATE);
1862 qemu_bh_schedule(s->cleanup_bh);
1863 qemu_mutex_unlock_iothread();
1865 rcu_unregister_thread();
1866 return NULL;
1869 void migrate_fd_connect(MigrationState *s)
1871 s->expected_downtime = s->parameters.downtime_limit;
1872 s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
1874 qemu_file_set_blocking(s->to_dst_file, true);
1875 qemu_file_set_rate_limit(s->to_dst_file,
1876 s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
1878 /* Notify before starting migration thread */
1879 notifier_list_notify(&migration_state_notifiers, s);
1882 * Open the return path; currently for postcopy but other things might
1883 * also want it.
1885 if (migrate_postcopy_ram()) {
1886 if (open_return_path_on_source(s)) {
1887 error_report("Unable to open return-path for postcopy");
1888 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1889 MIGRATION_STATUS_FAILED);
1890 migrate_fd_cleanup(s);
1891 return;
1895 migrate_compress_threads_create();
1896 qemu_thread_create(&s->thread, "migration", migration_thread, s,
1897 QEMU_THREAD_JOINABLE);
1898 s->migration_thread_running = true;
1901 PostcopyState postcopy_state_get(void)
1903 return atomic_mb_read(&incoming_postcopy_state);
1906 /* Set the state and return the old state */
1907 PostcopyState postcopy_state_set(PostcopyState new_state)
1909 return atomic_xchg(&incoming_postcopy_state, new_state);