xref: /openbmc/qemu/migration/migration.c (revision b917da4c)
1 /*
2  * QEMU live migration
3  *
4  * Copyright IBM, Corp. 2008
5  *
6  * Authors:
7  *  Anthony Liguori   <aliguori@us.ibm.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
12  * Contributions after 2012-01-13 are licensed under the terms of the
13  * GNU GPL, version 2 or (at your option) any later version.
14  */
15 
16 #include "qemu/osdep.h"
17 #include "qemu-common.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 
38 #define MAX_THROTTLE  (32 << 20)      /* Migration transfer speed throttling */
39 
40 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
41  * data. */
42 #define BUFFER_DELAY     100
43 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
44 
45 /* Default compression thread count */
46 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
47 /* Default decompression thread count, usually decompression is at
48  * least 4 times as fast as compression.*/
49 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2
50 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */
51 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1
52 /* Define default autoconverge cpu throttle migration parameters */
53 #define DEFAULT_MIGRATE_X_CPU_THROTTLE_INITIAL 20
54 #define DEFAULT_MIGRATE_X_CPU_THROTTLE_INCREMENT 10
55 
56 /* Migration XBZRLE default cache size */
57 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
58 
59 static NotifierList migration_state_notifiers =
60     NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
61 
62 static bool deferred_incoming;
63 
64 /*
65  * Current state of incoming postcopy; note this is not part of
66  * MigrationIncomingState since it's state is used during cleanup
67  * at the end as MIS is being freed.
68  */
69 static PostcopyState incoming_postcopy_state;
70 
71 /* When we add fault tolerance, we could have several
72    migrations at once.  For now we don't need to add
73    dynamic creation of migration */
74 
75 /* For outgoing */
76 MigrationState *migrate_get_current(void)
77 {
78     static bool once;
79     static MigrationState current_migration = {
80         .state = MIGRATION_STATUS_NONE,
81         .bandwidth_limit = MAX_THROTTLE,
82         .xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE,
83         .mbps = -1,
84         .parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL] =
85                 DEFAULT_MIGRATE_COMPRESS_LEVEL,
86         .parameters[MIGRATION_PARAMETER_COMPRESS_THREADS] =
87                 DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT,
88         .parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS] =
89                 DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT,
90         .parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INITIAL] =
91                 DEFAULT_MIGRATE_X_CPU_THROTTLE_INITIAL,
92         .parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INCREMENT] =
93                 DEFAULT_MIGRATE_X_CPU_THROTTLE_INCREMENT,
94     };
95 
96     if (!once) {
97         qemu_mutex_init(&current_migration.src_page_req_mutex);
98         once = true;
99     }
100     return &current_migration;
101 }
102 
103 /* For incoming */
104 static MigrationIncomingState *mis_current;
105 
106 MigrationIncomingState *migration_incoming_get_current(void)
107 {
108     return mis_current;
109 }
110 
111 MigrationIncomingState *migration_incoming_state_new(QEMUFile* f)
112 {
113     mis_current = g_new0(MigrationIncomingState, 1);
114     mis_current->from_src_file = f;
115     mis_current->state = MIGRATION_STATUS_NONE;
116     QLIST_INIT(&mis_current->loadvm_handlers);
117     qemu_mutex_init(&mis_current->rp_mutex);
118     qemu_event_init(&mis_current->main_thread_load_event, false);
119 
120     return mis_current;
121 }
122 
123 void migration_incoming_state_destroy(void)
124 {
125     qemu_event_destroy(&mis_current->main_thread_load_event);
126     loadvm_free_handlers(mis_current);
127     g_free(mis_current);
128     mis_current = NULL;
129 }
130 
131 
132 typedef struct {
133     bool optional;
134     uint32_t size;
135     uint8_t runstate[100];
136     RunState state;
137     bool received;
138 } GlobalState;
139 
140 static GlobalState global_state;
141 
142 int global_state_store(void)
143 {
144     if (!runstate_store((char *)global_state.runstate,
145                         sizeof(global_state.runstate))) {
146         error_report("runstate name too big: %s", global_state.runstate);
147         trace_migrate_state_too_big();
148         return -EINVAL;
149     }
150     return 0;
151 }
152 
153 void global_state_store_running(void)
154 {
155     const char *state = RunState_lookup[RUN_STATE_RUNNING];
156     strncpy((char *)global_state.runstate,
157            state, sizeof(global_state.runstate));
158 }
159 
160 static bool global_state_received(void)
161 {
162     return global_state.received;
163 }
164 
165 static RunState global_state_get_runstate(void)
166 {
167     return global_state.state;
168 }
169 
170 void global_state_set_optional(void)
171 {
172     global_state.optional = true;
173 }
174 
175 static bool global_state_needed(void *opaque)
176 {
177     GlobalState *s = opaque;
178     char *runstate = (char *)s->runstate;
179 
180     /* If it is not optional, it is mandatory */
181 
182     if (s->optional == false) {
183         return true;
184     }
185 
186     /* If state is running or paused, it is not needed */
187 
188     if (strcmp(runstate, "running") == 0 ||
189         strcmp(runstate, "paused") == 0) {
190         return false;
191     }
192 
193     /* for any other state it is needed */
194     return true;
195 }
196 
197 static int global_state_post_load(void *opaque, int version_id)
198 {
199     GlobalState *s = opaque;
200     Error *local_err = NULL;
201     int r;
202     char *runstate = (char *)s->runstate;
203 
204     s->received = true;
205     trace_migrate_global_state_post_load(runstate);
206 
207     r = qapi_enum_parse(RunState_lookup, runstate, RUN_STATE__MAX,
208                                 -1, &local_err);
209 
210     if (r == -1) {
211         if (local_err) {
212             error_report_err(local_err);
213         }
214         return -EINVAL;
215     }
216     s->state = r;
217 
218     return 0;
219 }
220 
221 static void global_state_pre_save(void *opaque)
222 {
223     GlobalState *s = opaque;
224 
225     trace_migrate_global_state_pre_save((char *)s->runstate);
226     s->size = strlen((char *)s->runstate) + 1;
227 }
228 
229 static const VMStateDescription vmstate_globalstate = {
230     .name = "globalstate",
231     .version_id = 1,
232     .minimum_version_id = 1,
233     .post_load = global_state_post_load,
234     .pre_save = global_state_pre_save,
235     .needed = global_state_needed,
236     .fields = (VMStateField[]) {
237         VMSTATE_UINT32(size, GlobalState),
238         VMSTATE_BUFFER(runstate, GlobalState),
239         VMSTATE_END_OF_LIST()
240     },
241 };
242 
243 void register_global_state(void)
244 {
245     /* We would use it independently that we receive it */
246     strcpy((char *)&global_state.runstate, "");
247     global_state.received = false;
248     vmstate_register(NULL, 0, &vmstate_globalstate, &global_state);
249 }
250 
251 static void migrate_generate_event(int new_state)
252 {
253     if (migrate_use_events()) {
254         qapi_event_send_migration(new_state, &error_abort);
255     }
256 }
257 
258 /*
259  * Called on -incoming with a defer: uri.
260  * The migration can be started later after any parameters have been
261  * changed.
262  */
263 static void deferred_incoming_migration(Error **errp)
264 {
265     if (deferred_incoming) {
266         error_setg(errp, "Incoming migration already deferred");
267     }
268     deferred_incoming = true;
269 }
270 
271 /* Request a range of pages from the source VM at the given
272  * start address.
273  *   rbname: Name of the RAMBlock to request the page in, if NULL it's the same
274  *           as the last request (a name must have been given previously)
275  *   Start: Address offset within the RB
276  *   Len: Length in bytes required - must be a multiple of pagesize
277  */
278 void migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname,
279                                ram_addr_t start, size_t len)
280 {
281     uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname upto 256 */
282     size_t msglen = 12; /* start + len */
283 
284     *(uint64_t *)bufc = cpu_to_be64((uint64_t)start);
285     *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len);
286 
287     if (rbname) {
288         int rbname_len = strlen(rbname);
289         assert(rbname_len < 256);
290 
291         bufc[msglen++] = rbname_len;
292         memcpy(bufc + msglen, rbname, rbname_len);
293         msglen += rbname_len;
294         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES_ID, msglen, bufc);
295     } else {
296         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES, msglen, bufc);
297     }
298 }
299 
300 void qemu_start_incoming_migration(const char *uri, Error **errp)
301 {
302     const char *p;
303 
304     qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort);
305     if (!strcmp(uri, "defer")) {
306         deferred_incoming_migration(errp);
307     } else if (strstart(uri, "tcp:", &p)) {
308         tcp_start_incoming_migration(p, errp);
309 #ifdef CONFIG_RDMA
310     } else if (strstart(uri, "rdma:", &p)) {
311         rdma_start_incoming_migration(p, errp);
312 #endif
313 #if !defined(WIN32)
314     } else if (strstart(uri, "exec:", &p)) {
315         exec_start_incoming_migration(p, errp);
316     } else if (strstart(uri, "unix:", &p)) {
317         unix_start_incoming_migration(p, errp);
318     } else if (strstart(uri, "fd:", &p)) {
319         fd_start_incoming_migration(p, errp);
320 #endif
321     } else {
322         error_setg(errp, "unknown migration protocol: %s", uri);
323     }
324 }
325 
326 static void process_incoming_migration_bh(void *opaque)
327 {
328     Error *local_err = NULL;
329     MigrationIncomingState *mis = opaque;
330 
331     /* Make sure all file formats flush their mutable metadata */
332     bdrv_invalidate_cache_all(&local_err);
333     if (local_err) {
334         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
335                           MIGRATION_STATUS_FAILED);
336         error_report_err(local_err);
337         migrate_decompress_threads_join();
338         exit(EXIT_FAILURE);
339     }
340 
341     /*
342      * This must happen after all error conditions are dealt with and
343      * we're sure the VM is going to be running on this host.
344      */
345     qemu_announce_self();
346 
347     /* If global state section was not received or we are in running
348        state, we need to obey autostart. Any other state is set with
349        runstate_set. */
350 
351     if (!global_state_received() ||
352         global_state_get_runstate() == RUN_STATE_RUNNING) {
353         if (autostart) {
354             vm_start();
355         } else {
356             runstate_set(RUN_STATE_PAUSED);
357         }
358     } else {
359         runstate_set(global_state_get_runstate());
360     }
361     migrate_decompress_threads_join();
362     /*
363      * This must happen after any state changes since as soon as an external
364      * observer sees this event they might start to prod at the VM assuming
365      * it's ready to use.
366      */
367     migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
368                       MIGRATION_STATUS_COMPLETED);
369     qemu_bh_delete(mis->bh);
370     migration_incoming_state_destroy();
371 }
372 
373 static void process_incoming_migration_co(void *opaque)
374 {
375     QEMUFile *f = opaque;
376     MigrationIncomingState *mis;
377     PostcopyState ps;
378     int ret;
379 
380     mis = migration_incoming_state_new(f);
381     postcopy_state_set(POSTCOPY_INCOMING_NONE);
382     migrate_set_state(&mis->state, MIGRATION_STATUS_NONE,
383                       MIGRATION_STATUS_ACTIVE);
384     ret = qemu_loadvm_state(f);
385 
386     ps = postcopy_state_get();
387     trace_process_incoming_migration_co_end(ret, ps);
388     if (ps != POSTCOPY_INCOMING_NONE) {
389         if (ps == POSTCOPY_INCOMING_ADVISE) {
390             /*
391              * Where a migration had postcopy enabled (and thus went to advise)
392              * but managed to complete within the precopy period, we can use
393              * the normal exit.
394              */
395             postcopy_ram_incoming_cleanup(mis);
396         } else if (ret >= 0) {
397             /*
398              * Postcopy was started, cleanup should happen at the end of the
399              * postcopy thread.
400              */
401             trace_process_incoming_migration_co_postcopy_end_main();
402             return;
403         }
404         /* Else if something went wrong then just fall out of the normal exit */
405     }
406 
407     qemu_fclose(f);
408     free_xbzrle_decoded_buf();
409 
410     if (ret < 0) {
411         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
412                           MIGRATION_STATUS_FAILED);
413         error_report("load of migration failed: %s", strerror(-ret));
414         migrate_decompress_threads_join();
415         exit(EXIT_FAILURE);
416     }
417 
418     mis->bh = qemu_bh_new(process_incoming_migration_bh, mis);
419     qemu_bh_schedule(mis->bh);
420 }
421 
422 void process_incoming_migration(QEMUFile *f)
423 {
424     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co);
425     int fd = qemu_get_fd(f);
426 
427     assert(fd != -1);
428     migrate_decompress_threads_create();
429     qemu_set_nonblock(fd);
430     qemu_coroutine_enter(co, f);
431 }
432 
433 /*
434  * Send a message on the return channel back to the source
435  * of the migration.
436  */
437 void migrate_send_rp_message(MigrationIncomingState *mis,
438                              enum mig_rp_message_type message_type,
439                              uint16_t len, void *data)
440 {
441     trace_migrate_send_rp_message((int)message_type, len);
442     qemu_mutex_lock(&mis->rp_mutex);
443     qemu_put_be16(mis->to_src_file, (unsigned int)message_type);
444     qemu_put_be16(mis->to_src_file, len);
445     qemu_put_buffer(mis->to_src_file, data, len);
446     qemu_fflush(mis->to_src_file);
447     qemu_mutex_unlock(&mis->rp_mutex);
448 }
449 
450 /*
451  * Send a 'SHUT' message on the return channel with the given value
452  * to indicate that we've finished with the RP.  Non-0 value indicates
453  * error.
454  */
455 void migrate_send_rp_shut(MigrationIncomingState *mis,
456                           uint32_t value)
457 {
458     uint32_t buf;
459 
460     buf = cpu_to_be32(value);
461     migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf);
462 }
463 
464 /*
465  * Send a 'PONG' message on the return channel with the given value
466  * (normally in response to a 'PING')
467  */
468 void migrate_send_rp_pong(MigrationIncomingState *mis,
469                           uint32_t value)
470 {
471     uint32_t buf;
472 
473     buf = cpu_to_be32(value);
474     migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf);
475 }
476 
477 /* amount of nanoseconds we are willing to wait for migration to be down.
478  * the choice of nanoseconds is because it is the maximum resolution that
479  * get_clock() can achieve. It is an internal measure. All user-visible
480  * units must be in seconds */
481 static uint64_t max_downtime = 300000000;
482 
483 uint64_t migrate_max_downtime(void)
484 {
485     return max_downtime;
486 }
487 
488 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
489 {
490     MigrationCapabilityStatusList *head = NULL;
491     MigrationCapabilityStatusList *caps;
492     MigrationState *s = migrate_get_current();
493     int i;
494 
495     caps = NULL; /* silence compiler warning */
496     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
497         if (head == NULL) {
498             head = g_malloc0(sizeof(*caps));
499             caps = head;
500         } else {
501             caps->next = g_malloc0(sizeof(*caps));
502             caps = caps->next;
503         }
504         caps->value =
505             g_malloc(sizeof(*caps->value));
506         caps->value->capability = i;
507         caps->value->state = s->enabled_capabilities[i];
508     }
509 
510     return head;
511 }
512 
513 MigrationParameters *qmp_query_migrate_parameters(Error **errp)
514 {
515     MigrationParameters *params;
516     MigrationState *s = migrate_get_current();
517 
518     params = g_malloc0(sizeof(*params));
519     params->compress_level = s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL];
520     params->compress_threads =
521             s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS];
522     params->decompress_threads =
523             s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS];
524     params->x_cpu_throttle_initial =
525             s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INITIAL];
526     params->x_cpu_throttle_increment =
527             s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INCREMENT];
528 
529     return params;
530 }
531 
532 /*
533  * Return true if we're already in the middle of a migration
534  * (i.e. any of the active or setup states)
535  */
536 static bool migration_is_setup_or_active(int state)
537 {
538     switch (state) {
539     case MIGRATION_STATUS_ACTIVE:
540     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
541     case MIGRATION_STATUS_SETUP:
542         return true;
543 
544     default:
545         return false;
546 
547     }
548 }
549 
550 static void get_xbzrle_cache_stats(MigrationInfo *info)
551 {
552     if (migrate_use_xbzrle()) {
553         info->has_xbzrle_cache = true;
554         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
555         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
556         info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred();
557         info->xbzrle_cache->pages = xbzrle_mig_pages_transferred();
558         info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss();
559         info->xbzrle_cache->cache_miss_rate = xbzrle_mig_cache_miss_rate();
560         info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow();
561     }
562 }
563 
564 MigrationInfo *qmp_query_migrate(Error **errp)
565 {
566     MigrationInfo *info = g_malloc0(sizeof(*info));
567     MigrationState *s = migrate_get_current();
568 
569     switch (s->state) {
570     case MIGRATION_STATUS_NONE:
571         /* no migration has happened ever */
572         break;
573     case MIGRATION_STATUS_SETUP:
574         info->has_status = true;
575         info->has_total_time = false;
576         break;
577     case MIGRATION_STATUS_ACTIVE:
578     case MIGRATION_STATUS_CANCELLING:
579         info->has_status = true;
580         info->has_total_time = true;
581         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
582             - s->total_time;
583         info->has_expected_downtime = true;
584         info->expected_downtime = s->expected_downtime;
585         info->has_setup_time = true;
586         info->setup_time = s->setup_time;
587 
588         info->has_ram = true;
589         info->ram = g_malloc0(sizeof(*info->ram));
590         info->ram->transferred = ram_bytes_transferred();
591         info->ram->remaining = ram_bytes_remaining();
592         info->ram->total = ram_bytes_total();
593         info->ram->duplicate = dup_mig_pages_transferred();
594         info->ram->skipped = skipped_mig_pages_transferred();
595         info->ram->normal = norm_mig_pages_transferred();
596         info->ram->normal_bytes = norm_mig_bytes_transferred();
597         info->ram->dirty_pages_rate = s->dirty_pages_rate;
598         info->ram->mbps = s->mbps;
599         info->ram->dirty_sync_count = s->dirty_sync_count;
600 
601         if (blk_mig_active()) {
602             info->has_disk = true;
603             info->disk = g_malloc0(sizeof(*info->disk));
604             info->disk->transferred = blk_mig_bytes_transferred();
605             info->disk->remaining = blk_mig_bytes_remaining();
606             info->disk->total = blk_mig_bytes_total();
607         }
608 
609         if (cpu_throttle_active()) {
610             info->has_x_cpu_throttle_percentage = true;
611             info->x_cpu_throttle_percentage = cpu_throttle_get_percentage();
612         }
613 
614         get_xbzrle_cache_stats(info);
615         break;
616     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
617         /* Mostly the same as active; TODO add some postcopy stats */
618         info->has_status = true;
619         info->has_total_time = true;
620         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
621             - s->total_time;
622         info->has_expected_downtime = true;
623         info->expected_downtime = s->expected_downtime;
624         info->has_setup_time = true;
625         info->setup_time = s->setup_time;
626 
627         info->has_ram = true;
628         info->ram = g_malloc0(sizeof(*info->ram));
629         info->ram->transferred = ram_bytes_transferred();
630         info->ram->remaining = ram_bytes_remaining();
631         info->ram->total = ram_bytes_total();
632         info->ram->duplicate = dup_mig_pages_transferred();
633         info->ram->skipped = skipped_mig_pages_transferred();
634         info->ram->normal = norm_mig_pages_transferred();
635         info->ram->normal_bytes = norm_mig_bytes_transferred();
636         info->ram->dirty_pages_rate = s->dirty_pages_rate;
637         info->ram->mbps = s->mbps;
638         info->ram->dirty_sync_count = s->dirty_sync_count;
639 
640         if (blk_mig_active()) {
641             info->has_disk = true;
642             info->disk = g_malloc0(sizeof(*info->disk));
643             info->disk->transferred = blk_mig_bytes_transferred();
644             info->disk->remaining = blk_mig_bytes_remaining();
645             info->disk->total = blk_mig_bytes_total();
646         }
647 
648         get_xbzrle_cache_stats(info);
649         break;
650     case MIGRATION_STATUS_COMPLETED:
651         get_xbzrle_cache_stats(info);
652 
653         info->has_status = true;
654         info->has_total_time = true;
655         info->total_time = s->total_time;
656         info->has_downtime = true;
657         info->downtime = s->downtime;
658         info->has_setup_time = true;
659         info->setup_time = s->setup_time;
660 
661         info->has_ram = true;
662         info->ram = g_malloc0(sizeof(*info->ram));
663         info->ram->transferred = ram_bytes_transferred();
664         info->ram->remaining = 0;
665         info->ram->total = ram_bytes_total();
666         info->ram->duplicate = dup_mig_pages_transferred();
667         info->ram->skipped = skipped_mig_pages_transferred();
668         info->ram->normal = norm_mig_pages_transferred();
669         info->ram->normal_bytes = norm_mig_bytes_transferred();
670         info->ram->mbps = s->mbps;
671         info->ram->dirty_sync_count = s->dirty_sync_count;
672         break;
673     case MIGRATION_STATUS_FAILED:
674         info->has_status = true;
675         break;
676     case MIGRATION_STATUS_CANCELLED:
677         info->has_status = true;
678         break;
679     }
680     info->status = s->state;
681 
682     return info;
683 }
684 
685 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
686                                   Error **errp)
687 {
688     MigrationState *s = migrate_get_current();
689     MigrationCapabilityStatusList *cap;
690 
691     if (migration_is_setup_or_active(s->state)) {
692         error_setg(errp, QERR_MIGRATION_ACTIVE);
693         return;
694     }
695 
696     for (cap = params; cap; cap = cap->next) {
697         s->enabled_capabilities[cap->value->capability] = cap->value->state;
698     }
699 
700     if (migrate_postcopy_ram()) {
701         if (migrate_use_compression()) {
702             /* The decompression threads asynchronously write into RAM
703              * rather than use the atomic copies needed to avoid
704              * userfaulting.  It should be possible to fix the decompression
705              * threads for compatibility in future.
706              */
707             error_report("Postcopy is not currently compatible with "
708                          "compression");
709             s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] =
710                 false;
711         }
712     }
713 }
714 
715 void qmp_migrate_set_parameters(bool has_compress_level,
716                                 int64_t compress_level,
717                                 bool has_compress_threads,
718                                 int64_t compress_threads,
719                                 bool has_decompress_threads,
720                                 int64_t decompress_threads,
721                                 bool has_x_cpu_throttle_initial,
722                                 int64_t x_cpu_throttle_initial,
723                                 bool has_x_cpu_throttle_increment,
724                                 int64_t x_cpu_throttle_increment, Error **errp)
725 {
726     MigrationState *s = migrate_get_current();
727 
728     if (has_compress_level && (compress_level < 0 || compress_level > 9)) {
729         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level",
730                    "is invalid, it should be in the range of 0 to 9");
731         return;
732     }
733     if (has_compress_threads &&
734             (compress_threads < 1 || compress_threads > 255)) {
735         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
736                    "compress_threads",
737                    "is invalid, it should be in the range of 1 to 255");
738         return;
739     }
740     if (has_decompress_threads &&
741             (decompress_threads < 1 || decompress_threads > 255)) {
742         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
743                    "decompress_threads",
744                    "is invalid, it should be in the range of 1 to 255");
745         return;
746     }
747     if (has_x_cpu_throttle_initial &&
748             (x_cpu_throttle_initial < 1 || x_cpu_throttle_initial > 99)) {
749         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
750                    "x_cpu_throttle_initial",
751                    "an integer in the range of 1 to 99");
752     }
753     if (has_x_cpu_throttle_increment &&
754             (x_cpu_throttle_increment < 1 || x_cpu_throttle_increment > 99)) {
755         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
756                    "x_cpu_throttle_increment",
757                    "an integer in the range of 1 to 99");
758     }
759 
760     if (has_compress_level) {
761         s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL] = compress_level;
762     }
763     if (has_compress_threads) {
764         s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS] = compress_threads;
765     }
766     if (has_decompress_threads) {
767         s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS] =
768                                                     decompress_threads;
769     }
770     if (has_x_cpu_throttle_initial) {
771         s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INITIAL] =
772                                                     x_cpu_throttle_initial;
773     }
774 
775     if (has_x_cpu_throttle_increment) {
776         s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INCREMENT] =
777                                                     x_cpu_throttle_increment;
778     }
779 }
780 
781 void qmp_migrate_start_postcopy(Error **errp)
782 {
783     MigrationState *s = migrate_get_current();
784 
785     if (!migrate_postcopy_ram()) {
786         error_setg(errp, "Enable postcopy with migrate_set_capability before"
787                          " the start of migration");
788         return;
789     }
790 
791     if (s->state == MIGRATION_STATUS_NONE) {
792         error_setg(errp, "Postcopy must be started after migration has been"
793                          " started");
794         return;
795     }
796     /*
797      * we don't error if migration has finished since that would be racy
798      * with issuing this command.
799      */
800     atomic_set(&s->start_postcopy, true);
801 }
802 
803 /* shared migration helpers */
804 
805 void migrate_set_state(int *state, int old_state, int new_state)
806 {
807     if (atomic_cmpxchg(state, old_state, new_state) == old_state) {
808         trace_migrate_set_state(new_state);
809         migrate_generate_event(new_state);
810     }
811 }
812 
813 static void migrate_fd_cleanup(void *opaque)
814 {
815     MigrationState *s = opaque;
816 
817     qemu_bh_delete(s->cleanup_bh);
818     s->cleanup_bh = NULL;
819 
820     flush_page_queue(s);
821 
822     if (s->to_dst_file) {
823         trace_migrate_fd_cleanup();
824         qemu_mutex_unlock_iothread();
825         if (s->migration_thread_running) {
826             qemu_thread_join(&s->thread);
827             s->migration_thread_running = false;
828         }
829         qemu_mutex_lock_iothread();
830 
831         migrate_compress_threads_join();
832         qemu_fclose(s->to_dst_file);
833         s->to_dst_file = NULL;
834     }
835 
836     assert((s->state != MIGRATION_STATUS_ACTIVE) &&
837            (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE));
838 
839     if (s->state == MIGRATION_STATUS_CANCELLING) {
840         migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING,
841                           MIGRATION_STATUS_CANCELLED);
842     }
843 
844     notifier_list_notify(&migration_state_notifiers, s);
845 }
846 
847 void migrate_fd_error(MigrationState *s)
848 {
849     trace_migrate_fd_error();
850     assert(s->to_dst_file == NULL);
851     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
852                       MIGRATION_STATUS_FAILED);
853     notifier_list_notify(&migration_state_notifiers, s);
854 }
855 
856 static void migrate_fd_cancel(MigrationState *s)
857 {
858     int old_state ;
859     QEMUFile *f = migrate_get_current()->to_dst_file;
860     trace_migrate_fd_cancel();
861 
862     if (s->rp_state.from_dst_file) {
863         /* shutdown the rp socket, so causing the rp thread to shutdown */
864         qemu_file_shutdown(s->rp_state.from_dst_file);
865     }
866 
867     do {
868         old_state = s->state;
869         if (!migration_is_setup_or_active(old_state)) {
870             break;
871         }
872         migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING);
873     } while (s->state != MIGRATION_STATUS_CANCELLING);
874 
875     /*
876      * If we're unlucky the migration code might be stuck somewhere in a
877      * send/write while the network has failed and is waiting to timeout;
878      * if we've got shutdown(2) available then we can force it to quit.
879      * The outgoing qemu file gets closed in migrate_fd_cleanup that is
880      * called in a bh, so there is no race against this cancel.
881      */
882     if (s->state == MIGRATION_STATUS_CANCELLING && f) {
883         qemu_file_shutdown(f);
884     }
885 }
886 
887 void add_migration_state_change_notifier(Notifier *notify)
888 {
889     notifier_list_add(&migration_state_notifiers, notify);
890 }
891 
892 void remove_migration_state_change_notifier(Notifier *notify)
893 {
894     notifier_remove(notify);
895 }
896 
897 bool migration_in_setup(MigrationState *s)
898 {
899     return s->state == MIGRATION_STATUS_SETUP;
900 }
901 
902 bool migration_has_finished(MigrationState *s)
903 {
904     return s->state == MIGRATION_STATUS_COMPLETED;
905 }
906 
907 bool migration_has_failed(MigrationState *s)
908 {
909     return (s->state == MIGRATION_STATUS_CANCELLED ||
910             s->state == MIGRATION_STATUS_FAILED);
911 }
912 
913 bool migration_in_postcopy(MigrationState *s)
914 {
915     return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
916 }
917 
918 bool migration_in_postcopy_after_devices(MigrationState *s)
919 {
920     return migration_in_postcopy(s) && s->postcopy_after_devices;
921 }
922 
923 MigrationState *migrate_init(const MigrationParams *params)
924 {
925     MigrationState *s = migrate_get_current();
926 
927     /*
928      * Reinitialise all migration state, except
929      * parameters/capabilities that the user set, and
930      * locks.
931      */
932     s->bytes_xfer = 0;
933     s->xfer_limit = 0;
934     s->cleanup_bh = 0;
935     s->to_dst_file = NULL;
936     s->state = MIGRATION_STATUS_NONE;
937     s->params = *params;
938     s->rp_state.from_dst_file = NULL;
939     s->rp_state.error = false;
940     s->mbps = 0.0;
941     s->downtime = 0;
942     s->expected_downtime = 0;
943     s->dirty_pages_rate = 0;
944     s->dirty_bytes_rate = 0;
945     s->setup_time = 0;
946     s->dirty_sync_count = 0;
947     s->start_postcopy = false;
948     s->postcopy_after_devices = false;
949     s->migration_thread_running = false;
950     s->last_req_rb = NULL;
951 
952     migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP);
953 
954     QSIMPLEQ_INIT(&s->src_page_requests);
955 
956     s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
957     return s;
958 }
959 
960 static GSList *migration_blockers;
961 
962 void migrate_add_blocker(Error *reason)
963 {
964     migration_blockers = g_slist_prepend(migration_blockers, reason);
965 }
966 
967 void migrate_del_blocker(Error *reason)
968 {
969     migration_blockers = g_slist_remove(migration_blockers, reason);
970 }
971 
972 void qmp_migrate_incoming(const char *uri, Error **errp)
973 {
974     Error *local_err = NULL;
975     static bool once = true;
976 
977     if (!deferred_incoming) {
978         error_setg(errp, "For use with '-incoming defer'");
979         return;
980     }
981     if (!once) {
982         error_setg(errp, "The incoming migration has already been started");
983     }
984 
985     qemu_start_incoming_migration(uri, &local_err);
986 
987     if (local_err) {
988         error_propagate(errp, local_err);
989         return;
990     }
991 
992     once = false;
993 }
994 
995 void qmp_migrate(const char *uri, bool has_blk, bool blk,
996                  bool has_inc, bool inc, bool has_detach, bool detach,
997                  Error **errp)
998 {
999     Error *local_err = NULL;
1000     MigrationState *s = migrate_get_current();
1001     MigrationParams params;
1002     const char *p;
1003 
1004     params.blk = has_blk && blk;
1005     params.shared = has_inc && inc;
1006 
1007     if (migration_is_setup_or_active(s->state) ||
1008         s->state == MIGRATION_STATUS_CANCELLING) {
1009         error_setg(errp, QERR_MIGRATION_ACTIVE);
1010         return;
1011     }
1012     if (runstate_check(RUN_STATE_INMIGRATE)) {
1013         error_setg(errp, "Guest is waiting for an incoming migration");
1014         return;
1015     }
1016 
1017     if (qemu_savevm_state_blocked(errp)) {
1018         return;
1019     }
1020 
1021     if (migration_blockers) {
1022         *errp = error_copy(migration_blockers->data);
1023         return;
1024     }
1025 
1026     s = migrate_init(&params);
1027 
1028     if (strstart(uri, "tcp:", &p)) {
1029         tcp_start_outgoing_migration(s, p, &local_err);
1030 #ifdef CONFIG_RDMA
1031     } else if (strstart(uri, "rdma:", &p)) {
1032         rdma_start_outgoing_migration(s, p, &local_err);
1033 #endif
1034 #if !defined(WIN32)
1035     } else if (strstart(uri, "exec:", &p)) {
1036         exec_start_outgoing_migration(s, p, &local_err);
1037     } else if (strstart(uri, "unix:", &p)) {
1038         unix_start_outgoing_migration(s, p, &local_err);
1039     } else if (strstart(uri, "fd:", &p)) {
1040         fd_start_outgoing_migration(s, p, &local_err);
1041 #endif
1042     } else {
1043         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri",
1044                    "a valid migration protocol");
1045         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1046                           MIGRATION_STATUS_FAILED);
1047         return;
1048     }
1049 
1050     if (local_err) {
1051         migrate_fd_error(s);
1052         error_propagate(errp, local_err);
1053         return;
1054     }
1055 }
1056 
1057 void qmp_migrate_cancel(Error **errp)
1058 {
1059     migrate_fd_cancel(migrate_get_current());
1060 }
1061 
1062 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
1063 {
1064     MigrationState *s = migrate_get_current();
1065     int64_t new_size;
1066 
1067     /* Check for truncation */
1068     if (value != (size_t)value) {
1069         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1070                    "exceeding address space");
1071         return;
1072     }
1073 
1074     /* Cache should not be larger than guest ram size */
1075     if (value > ram_bytes_total()) {
1076         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1077                    "exceeds guest ram size ");
1078         return;
1079     }
1080 
1081     new_size = xbzrle_cache_resize(value);
1082     if (new_size < 0) {
1083         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1084                    "is smaller than page size");
1085         return;
1086     }
1087 
1088     s->xbzrle_cache_size = new_size;
1089 }
1090 
1091 int64_t qmp_query_migrate_cache_size(Error **errp)
1092 {
1093     return migrate_xbzrle_cache_size();
1094 }
1095 
1096 void qmp_migrate_set_speed(int64_t value, Error **errp)
1097 {
1098     MigrationState *s;
1099 
1100     if (value < 0) {
1101         value = 0;
1102     }
1103     if (value > SIZE_MAX) {
1104         value = SIZE_MAX;
1105     }
1106 
1107     s = migrate_get_current();
1108     s->bandwidth_limit = value;
1109     if (s->to_dst_file) {
1110         qemu_file_set_rate_limit(s->to_dst_file,
1111                                  s->bandwidth_limit / XFER_LIMIT_RATIO);
1112     }
1113 }
1114 
1115 void qmp_migrate_set_downtime(double value, Error **errp)
1116 {
1117     value *= 1e9;
1118     value = MAX(0, MIN(UINT64_MAX, value));
1119     max_downtime = (uint64_t)value;
1120 }
1121 
1122 bool migrate_postcopy_ram(void)
1123 {
1124     MigrationState *s;
1125 
1126     s = migrate_get_current();
1127 
1128     return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM];
1129 }
1130 
1131 bool migrate_auto_converge(void)
1132 {
1133     MigrationState *s;
1134 
1135     s = migrate_get_current();
1136 
1137     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
1138 }
1139 
1140 bool migrate_zero_blocks(void)
1141 {
1142     MigrationState *s;
1143 
1144     s = migrate_get_current();
1145 
1146     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
1147 }
1148 
1149 bool migrate_use_compression(void)
1150 {
1151     MigrationState *s;
1152 
1153     s = migrate_get_current();
1154 
1155     return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS];
1156 }
1157 
1158 int migrate_compress_level(void)
1159 {
1160     MigrationState *s;
1161 
1162     s = migrate_get_current();
1163 
1164     return s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL];
1165 }
1166 
1167 int migrate_compress_threads(void)
1168 {
1169     MigrationState *s;
1170 
1171     s = migrate_get_current();
1172 
1173     return s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS];
1174 }
1175 
1176 int migrate_decompress_threads(void)
1177 {
1178     MigrationState *s;
1179 
1180     s = migrate_get_current();
1181 
1182     return s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS];
1183 }
1184 
1185 bool migrate_use_events(void)
1186 {
1187     MigrationState *s;
1188 
1189     s = migrate_get_current();
1190 
1191     return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS];
1192 }
1193 
1194 int migrate_use_xbzrle(void)
1195 {
1196     MigrationState *s;
1197 
1198     s = migrate_get_current();
1199 
1200     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
1201 }
1202 
1203 int64_t migrate_xbzrle_cache_size(void)
1204 {
1205     MigrationState *s;
1206 
1207     s = migrate_get_current();
1208 
1209     return s->xbzrle_cache_size;
1210 }
1211 
1212 /* migration thread support */
1213 /*
1214  * Something bad happened to the RP stream, mark an error
1215  * The caller shall print or trace something to indicate why
1216  */
1217 static void mark_source_rp_bad(MigrationState *s)
1218 {
1219     s->rp_state.error = true;
1220 }
1221 
1222 static struct rp_cmd_args {
1223     ssize_t     len; /* -1 = variable */
1224     const char *name;
1225 } rp_cmd_args[] = {
1226     [MIG_RP_MSG_INVALID]        = { .len = -1, .name = "INVALID" },
1227     [MIG_RP_MSG_SHUT]           = { .len =  4, .name = "SHUT" },
1228     [MIG_RP_MSG_PONG]           = { .len =  4, .name = "PONG" },
1229     [MIG_RP_MSG_REQ_PAGES]      = { .len = 12, .name = "REQ_PAGES" },
1230     [MIG_RP_MSG_REQ_PAGES_ID]   = { .len = -1, .name = "REQ_PAGES_ID" },
1231     [MIG_RP_MSG_MAX]            = { .len = -1, .name = "MAX" },
1232 };
1233 
1234 /*
1235  * Process a request for pages received on the return path,
1236  * We're allowed to send more than requested (e.g. to round to our page size)
1237  * and we don't need to send pages that have already been sent.
1238  */
1239 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname,
1240                                        ram_addr_t start, size_t len)
1241 {
1242     long our_host_ps = getpagesize();
1243 
1244     trace_migrate_handle_rp_req_pages(rbname, start, len);
1245 
1246     /*
1247      * Since we currently insist on matching page sizes, just sanity check
1248      * we're being asked for whole host pages.
1249      */
1250     if (start & (our_host_ps-1) ||
1251        (len & (our_host_ps-1))) {
1252         error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1253                      " len: %zd", __func__, start, len);
1254         mark_source_rp_bad(ms);
1255         return;
1256     }
1257 
1258     if (ram_save_queue_pages(ms, rbname, start, len)) {
1259         mark_source_rp_bad(ms);
1260     }
1261 }
1262 
1263 /*
1264  * Handles messages sent on the return path towards the source VM
1265  *
1266  */
1267 static void *source_return_path_thread(void *opaque)
1268 {
1269     MigrationState *ms = opaque;
1270     QEMUFile *rp = ms->rp_state.from_dst_file;
1271     uint16_t header_len, header_type;
1272     uint8_t buf[512];
1273     uint32_t tmp32, sibling_error;
1274     ram_addr_t start = 0; /* =0 to silence warning */
1275     size_t  len = 0, expected_len;
1276     int res;
1277 
1278     trace_source_return_path_thread_entry();
1279     while (!ms->rp_state.error && !qemu_file_get_error(rp) &&
1280            migration_is_setup_or_active(ms->state)) {
1281         trace_source_return_path_thread_loop_top();
1282         header_type = qemu_get_be16(rp);
1283         header_len = qemu_get_be16(rp);
1284 
1285         if (header_type >= MIG_RP_MSG_MAX ||
1286             header_type == MIG_RP_MSG_INVALID) {
1287             error_report("RP: Received invalid message 0x%04x length 0x%04x",
1288                     header_type, header_len);
1289             mark_source_rp_bad(ms);
1290             goto out;
1291         }
1292 
1293         if ((rp_cmd_args[header_type].len != -1 &&
1294             header_len != rp_cmd_args[header_type].len) ||
1295             header_len > sizeof(buf)) {
1296             error_report("RP: Received '%s' message (0x%04x) with"
1297                     "incorrect length %d expecting %zu",
1298                     rp_cmd_args[header_type].name, header_type, header_len,
1299                     (size_t)rp_cmd_args[header_type].len);
1300             mark_source_rp_bad(ms);
1301             goto out;
1302         }
1303 
1304         /* We know we've got a valid header by this point */
1305         res = qemu_get_buffer(rp, buf, header_len);
1306         if (res != header_len) {
1307             error_report("RP: Failed reading data for message 0x%04x"
1308                          " read %d expected %d",
1309                          header_type, res, header_len);
1310             mark_source_rp_bad(ms);
1311             goto out;
1312         }
1313 
1314         /* OK, we have the message and the data */
1315         switch (header_type) {
1316         case MIG_RP_MSG_SHUT:
1317             sibling_error = be32_to_cpup((uint32_t *)buf);
1318             trace_source_return_path_thread_shut(sibling_error);
1319             if (sibling_error) {
1320                 error_report("RP: Sibling indicated error %d", sibling_error);
1321                 mark_source_rp_bad(ms);
1322             }
1323             /*
1324              * We'll let the main thread deal with closing the RP
1325              * we could do a shutdown(2) on it, but we're the only user
1326              * anyway, so there's nothing gained.
1327              */
1328             goto out;
1329 
1330         case MIG_RP_MSG_PONG:
1331             tmp32 = be32_to_cpup((uint32_t *)buf);
1332             trace_source_return_path_thread_pong(tmp32);
1333             break;
1334 
1335         case MIG_RP_MSG_REQ_PAGES:
1336             start = be64_to_cpup((uint64_t *)buf);
1337             len = be32_to_cpup((uint32_t *)(buf + 8));
1338             migrate_handle_rp_req_pages(ms, NULL, start, len);
1339             break;
1340 
1341         case MIG_RP_MSG_REQ_PAGES_ID:
1342             expected_len = 12 + 1; /* header + termination */
1343 
1344             if (header_len >= expected_len) {
1345                 start = be64_to_cpup((uint64_t *)buf);
1346                 len = be32_to_cpup((uint32_t *)(buf + 8));
1347                 /* Now we expect an idstr */
1348                 tmp32 = buf[12]; /* Length of the following idstr */
1349                 buf[13 + tmp32] = '\0';
1350                 expected_len += tmp32;
1351             }
1352             if (header_len != expected_len) {
1353                 error_report("RP: Req_Page_id with length %d expecting %zd",
1354                         header_len, expected_len);
1355                 mark_source_rp_bad(ms);
1356                 goto out;
1357             }
1358             migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len);
1359             break;
1360 
1361         default:
1362             break;
1363         }
1364     }
1365     if (qemu_file_get_error(rp)) {
1366         trace_source_return_path_thread_bad_end();
1367         mark_source_rp_bad(ms);
1368     }
1369 
1370     trace_source_return_path_thread_end();
1371 out:
1372     ms->rp_state.from_dst_file = NULL;
1373     qemu_fclose(rp);
1374     return NULL;
1375 }
1376 
1377 static int open_return_path_on_source(MigrationState *ms)
1378 {
1379 
1380     ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file);
1381     if (!ms->rp_state.from_dst_file) {
1382         return -1;
1383     }
1384 
1385     trace_open_return_path_on_source();
1386     qemu_thread_create(&ms->rp_state.rp_thread, "return path",
1387                        source_return_path_thread, ms, QEMU_THREAD_JOINABLE);
1388 
1389     trace_open_return_path_on_source_continue();
1390 
1391     return 0;
1392 }
1393 
1394 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
1395 static int await_return_path_close_on_source(MigrationState *ms)
1396 {
1397     /*
1398      * If this is a normal exit then the destination will send a SHUT and the
1399      * rp_thread will exit, however if there's an error we need to cause
1400      * it to exit.
1401      */
1402     if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
1403         /*
1404          * shutdown(2), if we have it, will cause it to unblock if it's stuck
1405          * waiting for the destination.
1406          */
1407         qemu_file_shutdown(ms->rp_state.from_dst_file);
1408         mark_source_rp_bad(ms);
1409     }
1410     trace_await_return_path_close_on_source_joining();
1411     qemu_thread_join(&ms->rp_state.rp_thread);
1412     trace_await_return_path_close_on_source_close();
1413     return ms->rp_state.error;
1414 }
1415 
1416 /*
1417  * Switch from normal iteration to postcopy
1418  * Returns non-0 on error
1419  */
1420 static int postcopy_start(MigrationState *ms, bool *old_vm_running)
1421 {
1422     int ret;
1423     const QEMUSizedBuffer *qsb;
1424     int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1425     migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE,
1426                       MIGRATION_STATUS_POSTCOPY_ACTIVE);
1427 
1428     trace_postcopy_start();
1429     qemu_mutex_lock_iothread();
1430     trace_postcopy_start_set_run();
1431 
1432     qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1433     *old_vm_running = runstate_is_running();
1434     global_state_store();
1435     ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1436     if (ret < 0) {
1437         goto fail;
1438     }
1439 
1440     ret = bdrv_inactivate_all();
1441     if (ret < 0) {
1442         goto fail;
1443     }
1444 
1445     /*
1446      * Cause any non-postcopiable, but iterative devices to
1447      * send out their final data.
1448      */
1449     qemu_savevm_state_complete_precopy(ms->to_dst_file, true);
1450 
1451     /*
1452      * in Finish migrate and with the io-lock held everything should
1453      * be quiet, but we've potentially still got dirty pages and we
1454      * need to tell the destination to throw any pages it's already received
1455      * that are dirty
1456      */
1457     if (ram_postcopy_send_discard_bitmap(ms)) {
1458         error_report("postcopy send discard bitmap failed");
1459         goto fail;
1460     }
1461 
1462     /*
1463      * send rest of state - note things that are doing postcopy
1464      * will notice we're in POSTCOPY_ACTIVE and not actually
1465      * wrap their state up here
1466      */
1467     qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX);
1468     /* Ping just for debugging, helps line traces up */
1469     qemu_savevm_send_ping(ms->to_dst_file, 2);
1470 
1471     /*
1472      * While loading the device state we may trigger page transfer
1473      * requests and the fd must be free to process those, and thus
1474      * the destination must read the whole device state off the fd before
1475      * it starts processing it.  Unfortunately the ad-hoc migration format
1476      * doesn't allow the destination to know the size to read without fully
1477      * parsing it through each devices load-state code (especially the open
1478      * coded devices that use get/put).
1479      * So we wrap the device state up in a package with a length at the start;
1480      * to do this we use a qemu_buf to hold the whole of the device state.
1481      */
1482     QEMUFile *fb = qemu_bufopen("w", NULL);
1483     if (!fb) {
1484         error_report("Failed to create buffered file");
1485         goto fail;
1486     }
1487 
1488     /*
1489      * Make sure the receiver can get incoming pages before we send the rest
1490      * of the state
1491      */
1492     qemu_savevm_send_postcopy_listen(fb);
1493 
1494     qemu_savevm_state_complete_precopy(fb, false);
1495     qemu_savevm_send_ping(fb, 3);
1496 
1497     qemu_savevm_send_postcopy_run(fb);
1498 
1499     /* <><> end of stuff going into the package */
1500     qsb = qemu_buf_get(fb);
1501 
1502     /* Now send that blob */
1503     if (qemu_savevm_send_packaged(ms->to_dst_file, qsb)) {
1504         goto fail_closefb;
1505     }
1506     qemu_fclose(fb);
1507 
1508     /* Send a notify to give a chance for anything that needs to happen
1509      * at the transition to postcopy and after the device state; in particular
1510      * spice needs to trigger a transition now
1511      */
1512     ms->postcopy_after_devices = true;
1513     notifier_list_notify(&migration_state_notifiers, ms);
1514 
1515     ms->downtime =  qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop;
1516 
1517     qemu_mutex_unlock_iothread();
1518 
1519     /*
1520      * Although this ping is just for debug, it could potentially be
1521      * used for getting a better measurement of downtime at the source.
1522      */
1523     qemu_savevm_send_ping(ms->to_dst_file, 4);
1524 
1525     ret = qemu_file_get_error(ms->to_dst_file);
1526     if (ret) {
1527         error_report("postcopy_start: Migration stream errored");
1528         migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1529                               MIGRATION_STATUS_FAILED);
1530     }
1531 
1532     return ret;
1533 
1534 fail_closefb:
1535     qemu_fclose(fb);
1536 fail:
1537     migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1538                           MIGRATION_STATUS_FAILED);
1539     qemu_mutex_unlock_iothread();
1540     return -1;
1541 }
1542 
1543 /**
1544  * migration_completion: Used by migration_thread when there's not much left.
1545  *   The caller 'breaks' the loop when this returns.
1546  *
1547  * @s: Current migration state
1548  * @current_active_state: The migration state we expect to be in
1549  * @*old_vm_running: Pointer to old_vm_running flag
1550  * @*start_time: Pointer to time to update
1551  */
1552 static void migration_completion(MigrationState *s, int current_active_state,
1553                                  bool *old_vm_running,
1554                                  int64_t *start_time)
1555 {
1556     int ret;
1557 
1558     if (s->state == MIGRATION_STATUS_ACTIVE) {
1559         qemu_mutex_lock_iothread();
1560         *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1561         qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1562         *old_vm_running = runstate_is_running();
1563         ret = global_state_store();
1564 
1565         if (!ret) {
1566             ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1567             if (ret >= 0) {
1568                 ret = bdrv_inactivate_all();
1569             }
1570             if (ret >= 0) {
1571                 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX);
1572                 qemu_savevm_state_complete_precopy(s->to_dst_file, false);
1573             }
1574         }
1575         qemu_mutex_unlock_iothread();
1576 
1577         if (ret < 0) {
1578             goto fail;
1579         }
1580     } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1581         trace_migration_completion_postcopy_end();
1582 
1583         qemu_savevm_state_complete_postcopy(s->to_dst_file);
1584         trace_migration_completion_postcopy_end_after_complete();
1585     }
1586 
1587     /*
1588      * If rp was opened we must clean up the thread before
1589      * cleaning everything else up (since if there are no failures
1590      * it will wait for the destination to send it's status in
1591      * a SHUT command).
1592      * Postcopy opens rp if enabled (even if it's not avtivated)
1593      */
1594     if (migrate_postcopy_ram()) {
1595         int rp_error;
1596         trace_migration_completion_postcopy_end_before_rp();
1597         rp_error = await_return_path_close_on_source(s);
1598         trace_migration_completion_postcopy_end_after_rp(rp_error);
1599         if (rp_error) {
1600             goto fail;
1601         }
1602     }
1603 
1604     if (qemu_file_get_error(s->to_dst_file)) {
1605         trace_migration_completion_file_err();
1606         goto fail;
1607     }
1608 
1609     migrate_set_state(&s->state, current_active_state,
1610                       MIGRATION_STATUS_COMPLETED);
1611     return;
1612 
1613 fail:
1614     migrate_set_state(&s->state, current_active_state,
1615                       MIGRATION_STATUS_FAILED);
1616 }
1617 
1618 /*
1619  * Master migration thread on the source VM.
1620  * It drives the migration and pumps the data down the outgoing channel.
1621  */
1622 static void *migration_thread(void *opaque)
1623 {
1624     MigrationState *s = opaque;
1625     /* Used by the bandwidth calcs, updated later */
1626     int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1627     int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
1628     int64_t initial_bytes = 0;
1629     int64_t max_size = 0;
1630     int64_t start_time = initial_time;
1631     int64_t end_time;
1632     bool old_vm_running = false;
1633     bool entered_postcopy = false;
1634     /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */
1635     enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE;
1636 
1637     rcu_register_thread();
1638 
1639     qemu_savevm_state_header(s->to_dst_file);
1640 
1641     if (migrate_postcopy_ram()) {
1642         /* Now tell the dest that it should open its end so it can reply */
1643         qemu_savevm_send_open_return_path(s->to_dst_file);
1644 
1645         /* And do a ping that will make stuff easier to debug */
1646         qemu_savevm_send_ping(s->to_dst_file, 1);
1647 
1648         /*
1649          * Tell the destination that we *might* want to do postcopy later;
1650          * if the other end can't do postcopy it should fail now, nice and
1651          * early.
1652          */
1653         qemu_savevm_send_postcopy_advise(s->to_dst_file);
1654     }
1655 
1656     qemu_savevm_state_begin(s->to_dst_file, &s->params);
1657 
1658     s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
1659     current_active_state = MIGRATION_STATUS_ACTIVE;
1660     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1661                       MIGRATION_STATUS_ACTIVE);
1662 
1663     trace_migration_thread_setup_complete();
1664 
1665     while (s->state == MIGRATION_STATUS_ACTIVE ||
1666            s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1667         int64_t current_time;
1668         uint64_t pending_size;
1669 
1670         if (!qemu_file_rate_limit(s->to_dst_file)) {
1671             uint64_t pend_post, pend_nonpost;
1672 
1673             qemu_savevm_state_pending(s->to_dst_file, max_size, &pend_nonpost,
1674                                       &pend_post);
1675             pending_size = pend_nonpost + pend_post;
1676             trace_migrate_pending(pending_size, max_size,
1677                                   pend_post, pend_nonpost);
1678             if (pending_size && pending_size >= max_size) {
1679                 /* Still a significant amount to transfer */
1680 
1681                 if (migrate_postcopy_ram() &&
1682                     s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE &&
1683                     pend_nonpost <= max_size &&
1684                     atomic_read(&s->start_postcopy)) {
1685 
1686                     if (!postcopy_start(s, &old_vm_running)) {
1687                         current_active_state = MIGRATION_STATUS_POSTCOPY_ACTIVE;
1688                         entered_postcopy = true;
1689                     }
1690 
1691                     continue;
1692                 }
1693                 /* Just another iteration step */
1694                 qemu_savevm_state_iterate(s->to_dst_file, entered_postcopy);
1695             } else {
1696                 trace_migration_thread_low_pending(pending_size);
1697                 migration_completion(s, current_active_state,
1698                                      &old_vm_running, &start_time);
1699                 break;
1700             }
1701         }
1702 
1703         if (qemu_file_get_error(s->to_dst_file)) {
1704             migrate_set_state(&s->state, current_active_state,
1705                               MIGRATION_STATUS_FAILED);
1706             trace_migration_thread_file_err();
1707             break;
1708         }
1709         current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1710         if (current_time >= initial_time + BUFFER_DELAY) {
1711             uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) -
1712                                          initial_bytes;
1713             uint64_t time_spent = current_time - initial_time;
1714             double bandwidth = (double)transferred_bytes / time_spent;
1715             max_size = bandwidth * migrate_max_downtime() / 1000000;
1716 
1717             s->mbps = (((double) transferred_bytes * 8.0) /
1718                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0;
1719 
1720             trace_migrate_transferred(transferred_bytes, time_spent,
1721                                       bandwidth, max_size);
1722             /* if we haven't sent anything, we don't want to recalculate
1723                10000 is a small enough number for our purposes */
1724             if (s->dirty_bytes_rate && transferred_bytes > 10000) {
1725                 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
1726             }
1727 
1728             qemu_file_reset_rate_limit(s->to_dst_file);
1729             initial_time = current_time;
1730             initial_bytes = qemu_ftell(s->to_dst_file);
1731         }
1732         if (qemu_file_rate_limit(s->to_dst_file)) {
1733             /* usleep expects microseconds */
1734             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
1735         }
1736     }
1737 
1738     trace_migration_thread_after_loop();
1739     /* If we enabled cpu throttling for auto-converge, turn it off. */
1740     cpu_throttle_stop();
1741     end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1742 
1743     qemu_mutex_lock_iothread();
1744     qemu_savevm_state_cleanup();
1745     if (s->state == MIGRATION_STATUS_COMPLETED) {
1746         uint64_t transferred_bytes = qemu_ftell(s->to_dst_file);
1747         s->total_time = end_time - s->total_time;
1748         if (!entered_postcopy) {
1749             s->downtime = end_time - start_time;
1750         }
1751         if (s->total_time) {
1752             s->mbps = (((double) transferred_bytes * 8.0) /
1753                        ((double) s->total_time)) / 1000;
1754         }
1755         runstate_set(RUN_STATE_POSTMIGRATE);
1756     } else {
1757         if (old_vm_running && !entered_postcopy) {
1758             vm_start();
1759         }
1760     }
1761     qemu_bh_schedule(s->cleanup_bh);
1762     qemu_mutex_unlock_iothread();
1763 
1764     rcu_unregister_thread();
1765     return NULL;
1766 }
1767 
1768 void migrate_fd_connect(MigrationState *s)
1769 {
1770     /* This is a best 1st approximation. ns to ms */
1771     s->expected_downtime = max_downtime/1000000;
1772     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
1773 
1774     qemu_file_set_rate_limit(s->to_dst_file,
1775                              s->bandwidth_limit / XFER_LIMIT_RATIO);
1776 
1777     /* Notify before starting migration thread */
1778     notifier_list_notify(&migration_state_notifiers, s);
1779 
1780     /*
1781      * Open the return path; currently for postcopy but other things might
1782      * also want it.
1783      */
1784     if (migrate_postcopy_ram()) {
1785         if (open_return_path_on_source(s)) {
1786             error_report("Unable to open return-path for postcopy");
1787             migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1788                               MIGRATION_STATUS_FAILED);
1789             migrate_fd_cleanup(s);
1790             return;
1791         }
1792     }
1793 
1794     migrate_compress_threads_create();
1795     qemu_thread_create(&s->thread, "migration", migration_thread, s,
1796                        QEMU_THREAD_JOINABLE);
1797     s->migration_thread_running = true;
1798 }
1799 
1800 PostcopyState  postcopy_state_get(void)
1801 {
1802     return atomic_mb_read(&incoming_postcopy_state);
1803 }
1804 
1805 /* Set the state and return the old state */
1806 PostcopyState postcopy_state_set(PostcopyState new_state)
1807 {
1808     return atomic_xchg(&incoming_postcopy_state, new_state);
1809 }
1810 
1811