xref: /openbmc/qemu/migration/migration.c (revision fe44dc91)
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/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"
39 #include "migration/colo.h"
40 
41 #define MAX_THROTTLE  (32 << 20)      /* Migration transfer speed throttling */
42 
43 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
44  * data. */
45 #define BUFFER_DELAY     100
46 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
47 
48 /* Time in milliseconds we are allowed to stop the source,
49  * for sending the last part */
50 #define DEFAULT_MIGRATE_SET_DOWNTIME 300
51 
52 /* Default compression thread count */
53 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
54 /* Default decompression thread count, usually decompression is at
55  * least 4 times as fast as compression.*/
56 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2
57 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */
58 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1
59 /* Define default autoconverge cpu throttle migration parameters */
60 #define DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL 20
61 #define DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT 10
62 
63 /* Migration XBZRLE default cache size */
64 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
65 
66 /* The delay time (in ms) between two COLO checkpoints
67  * Note: Please change this default value to 10000 when we support hybrid mode.
68  */
69 #define DEFAULT_MIGRATE_X_CHECKPOINT_DELAY 200
70 
71 static NotifierList migration_state_notifiers =
72     NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
73 
74 static bool deferred_incoming;
75 
76 /*
77  * Current state of incoming postcopy; note this is not part of
78  * MigrationIncomingState since it's state is used during cleanup
79  * at the end as MIS is being freed.
80  */
81 static PostcopyState incoming_postcopy_state;
82 
83 /* When we add fault tolerance, we could have several
84    migrations at once.  For now we don't need to add
85    dynamic creation of migration */
86 
87 /* For outgoing */
88 MigrationState *migrate_get_current(void)
89 {
90     static bool once;
91     static MigrationState current_migration = {
92         .state = MIGRATION_STATUS_NONE,
93         .xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE,
94         .mbps = -1,
95         .parameters = {
96             .compress_level = DEFAULT_MIGRATE_COMPRESS_LEVEL,
97             .compress_threads = DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT,
98             .decompress_threads = DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT,
99             .cpu_throttle_initial = DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL,
100             .cpu_throttle_increment = DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT,
101             .max_bandwidth = MAX_THROTTLE,
102             .downtime_limit = DEFAULT_MIGRATE_SET_DOWNTIME,
103             .x_checkpoint_delay = DEFAULT_MIGRATE_X_CHECKPOINT_DELAY,
104         },
105     };
106 
107     if (!once) {
108         qemu_mutex_init(&current_migration.src_page_req_mutex);
109         once = true;
110     }
111     return &current_migration;
112 }
113 
114 /* For incoming */
115 static MigrationIncomingState *mis_current;
116 
117 MigrationIncomingState *migration_incoming_get_current(void)
118 {
119     return mis_current;
120 }
121 
122 MigrationIncomingState *migration_incoming_state_new(QEMUFile* f)
123 {
124     mis_current = g_new0(MigrationIncomingState, 1);
125     mis_current->from_src_file = f;
126     mis_current->state = MIGRATION_STATUS_NONE;
127     QLIST_INIT(&mis_current->loadvm_handlers);
128     qemu_mutex_init(&mis_current->rp_mutex);
129     qemu_event_init(&mis_current->main_thread_load_event, false);
130 
131     return mis_current;
132 }
133 
134 void migration_incoming_state_destroy(void)
135 {
136     qemu_event_destroy(&mis_current->main_thread_load_event);
137     loadvm_free_handlers(mis_current);
138     g_free(mis_current);
139     mis_current = NULL;
140 }
141 
142 
143 typedef struct {
144     bool optional;
145     uint32_t size;
146     uint8_t runstate[100];
147     RunState state;
148     bool received;
149 } GlobalState;
150 
151 static GlobalState global_state;
152 
153 int global_state_store(void)
154 {
155     if (!runstate_store((char *)global_state.runstate,
156                         sizeof(global_state.runstate))) {
157         error_report("runstate name too big: %s", global_state.runstate);
158         trace_migrate_state_too_big();
159         return -EINVAL;
160     }
161     return 0;
162 }
163 
164 void global_state_store_running(void)
165 {
166     const char *state = RunState_lookup[RUN_STATE_RUNNING];
167     strncpy((char *)global_state.runstate,
168            state, sizeof(global_state.runstate));
169 }
170 
171 static bool global_state_received(void)
172 {
173     return global_state.received;
174 }
175 
176 static RunState global_state_get_runstate(void)
177 {
178     return global_state.state;
179 }
180 
181 void global_state_set_optional(void)
182 {
183     global_state.optional = true;
184 }
185 
186 static bool global_state_needed(void *opaque)
187 {
188     GlobalState *s = opaque;
189     char *runstate = (char *)s->runstate;
190 
191     /* If it is not optional, it is mandatory */
192 
193     if (s->optional == false) {
194         return true;
195     }
196 
197     /* If state is running or paused, it is not needed */
198 
199     if (strcmp(runstate, "running") == 0 ||
200         strcmp(runstate, "paused") == 0) {
201         return false;
202     }
203 
204     /* for any other state it is needed */
205     return true;
206 }
207 
208 static int global_state_post_load(void *opaque, int version_id)
209 {
210     GlobalState *s = opaque;
211     Error *local_err = NULL;
212     int r;
213     char *runstate = (char *)s->runstate;
214 
215     s->received = true;
216     trace_migrate_global_state_post_load(runstate);
217 
218     r = qapi_enum_parse(RunState_lookup, runstate, RUN_STATE__MAX,
219                                 -1, &local_err);
220 
221     if (r == -1) {
222         if (local_err) {
223             error_report_err(local_err);
224         }
225         return -EINVAL;
226     }
227     s->state = r;
228 
229     return 0;
230 }
231 
232 static void global_state_pre_save(void *opaque)
233 {
234     GlobalState *s = opaque;
235 
236     trace_migrate_global_state_pre_save((char *)s->runstate);
237     s->size = strlen((char *)s->runstate) + 1;
238 }
239 
240 static const VMStateDescription vmstate_globalstate = {
241     .name = "globalstate",
242     .version_id = 1,
243     .minimum_version_id = 1,
244     .post_load = global_state_post_load,
245     .pre_save = global_state_pre_save,
246     .needed = global_state_needed,
247     .fields = (VMStateField[]) {
248         VMSTATE_UINT32(size, GlobalState),
249         VMSTATE_BUFFER(runstate, GlobalState),
250         VMSTATE_END_OF_LIST()
251     },
252 };
253 
254 void register_global_state(void)
255 {
256     /* We would use it independently that we receive it */
257     strcpy((char *)&global_state.runstate, "");
258     global_state.received = false;
259     vmstate_register(NULL, 0, &vmstate_globalstate, &global_state);
260 }
261 
262 static void migrate_generate_event(int new_state)
263 {
264     if (migrate_use_events()) {
265         qapi_event_send_migration(new_state, &error_abort);
266     }
267 }
268 
269 /*
270  * Called on -incoming with a defer: uri.
271  * The migration can be started later after any parameters have been
272  * changed.
273  */
274 static void deferred_incoming_migration(Error **errp)
275 {
276     if (deferred_incoming) {
277         error_setg(errp, "Incoming migration already deferred");
278     }
279     deferred_incoming = true;
280 }
281 
282 /* Request a range of pages from the source VM at the given
283  * start address.
284  *   rbname: Name of the RAMBlock to request the page in, if NULL it's the same
285  *           as the last request (a name must have been given previously)
286  *   Start: Address offset within the RB
287  *   Len: Length in bytes required - must be a multiple of pagesize
288  */
289 void migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname,
290                                ram_addr_t start, size_t len)
291 {
292     uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
293     size_t msglen = 12; /* start + len */
294 
295     *(uint64_t *)bufc = cpu_to_be64((uint64_t)start);
296     *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len);
297 
298     if (rbname) {
299         int rbname_len = strlen(rbname);
300         assert(rbname_len < 256);
301 
302         bufc[msglen++] = rbname_len;
303         memcpy(bufc + msglen, rbname, rbname_len);
304         msglen += rbname_len;
305         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES_ID, msglen, bufc);
306     } else {
307         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES, msglen, bufc);
308     }
309 }
310 
311 void qemu_start_incoming_migration(const char *uri, Error **errp)
312 {
313     const char *p;
314 
315     qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort);
316     if (!strcmp(uri, "defer")) {
317         deferred_incoming_migration(errp);
318     } else if (strstart(uri, "tcp:", &p)) {
319         tcp_start_incoming_migration(p, errp);
320 #ifdef CONFIG_RDMA
321     } else if (strstart(uri, "rdma:", &p)) {
322         rdma_start_incoming_migration(p, errp);
323 #endif
324     } else if (strstart(uri, "exec:", &p)) {
325         exec_start_incoming_migration(p, errp);
326     } else if (strstart(uri, "unix:", &p)) {
327         unix_start_incoming_migration(p, errp);
328     } else if (strstart(uri, "fd:", &p)) {
329         fd_start_incoming_migration(p, errp);
330     } else {
331         error_setg(errp, "unknown migration protocol: %s", uri);
332     }
333 }
334 
335 static void process_incoming_migration_bh(void *opaque)
336 {
337     Error *local_err = NULL;
338     MigrationIncomingState *mis = opaque;
339 
340     /* Make sure all file formats flush their mutable metadata */
341     bdrv_invalidate_cache_all(&local_err);
342     if (local_err) {
343         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
344                           MIGRATION_STATUS_FAILED);
345         error_report_err(local_err);
346         migrate_decompress_threads_join();
347         exit(EXIT_FAILURE);
348     }
349 
350     /*
351      * This must happen after all error conditions are dealt with and
352      * we're sure the VM is going to be running on this host.
353      */
354     qemu_announce_self();
355 
356     /* If global state section was not received or we are in running
357        state, we need to obey autostart. Any other state is set with
358        runstate_set. */
359 
360     if (!global_state_received() ||
361         global_state_get_runstate() == RUN_STATE_RUNNING) {
362         if (autostart) {
363             vm_start();
364         } else {
365             runstate_set(RUN_STATE_PAUSED);
366         }
367     } else {
368         runstate_set(global_state_get_runstate());
369     }
370     migrate_decompress_threads_join();
371     /*
372      * This must happen after any state changes since as soon as an external
373      * observer sees this event they might start to prod at the VM assuming
374      * it's ready to use.
375      */
376     migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
377                       MIGRATION_STATUS_COMPLETED);
378     qemu_bh_delete(mis->bh);
379     migration_incoming_state_destroy();
380 }
381 
382 static void process_incoming_migration_co(void *opaque)
383 {
384     QEMUFile *f = opaque;
385     MigrationIncomingState *mis;
386     PostcopyState ps;
387     int ret;
388 
389     mis = migration_incoming_state_new(f);
390     postcopy_state_set(POSTCOPY_INCOMING_NONE);
391     migrate_set_state(&mis->state, MIGRATION_STATUS_NONE,
392                       MIGRATION_STATUS_ACTIVE);
393     ret = qemu_loadvm_state(f);
394 
395     ps = postcopy_state_get();
396     trace_process_incoming_migration_co_end(ret, ps);
397     if (ps != POSTCOPY_INCOMING_NONE) {
398         if (ps == POSTCOPY_INCOMING_ADVISE) {
399             /*
400              * Where a migration had postcopy enabled (and thus went to advise)
401              * but managed to complete within the precopy period, we can use
402              * the normal exit.
403              */
404             postcopy_ram_incoming_cleanup(mis);
405         } else if (ret >= 0) {
406             /*
407              * Postcopy was started, cleanup should happen at the end of the
408              * postcopy thread.
409              */
410             trace_process_incoming_migration_co_postcopy_end_main();
411             return;
412         }
413         /* Else if something went wrong then just fall out of the normal exit */
414     }
415 
416     /* we get COLO info, and know if we are in COLO mode */
417     if (!ret && migration_incoming_enable_colo()) {
418         mis->migration_incoming_co = qemu_coroutine_self();
419         qemu_thread_create(&mis->colo_incoming_thread, "COLO incoming",
420              colo_process_incoming_thread, mis, QEMU_THREAD_JOINABLE);
421         mis->have_colo_incoming_thread = true;
422         qemu_coroutine_yield();
423 
424         /* Wait checkpoint incoming thread exit before free resource */
425         qemu_thread_join(&mis->colo_incoming_thread);
426     }
427 
428     qemu_fclose(f);
429     free_xbzrle_decoded_buf();
430 
431     if (ret < 0) {
432         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
433                           MIGRATION_STATUS_FAILED);
434         error_report("load of migration failed: %s", strerror(-ret));
435         migrate_decompress_threads_join();
436         exit(EXIT_FAILURE);
437     }
438 
439     mis->bh = qemu_bh_new(process_incoming_migration_bh, mis);
440     qemu_bh_schedule(mis->bh);
441 }
442 
443 void migration_fd_process_incoming(QEMUFile *f)
444 {
445     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co, f);
446 
447     migrate_decompress_threads_create();
448     qemu_file_set_blocking(f, false);
449     qemu_coroutine_enter(co);
450 }
451 
452 
453 void migration_channel_process_incoming(MigrationState *s,
454                                         QIOChannel *ioc)
455 {
456     trace_migration_set_incoming_channel(
457         ioc, object_get_typename(OBJECT(ioc)));
458 
459     if (s->parameters.tls_creds &&
460         !object_dynamic_cast(OBJECT(ioc),
461                              TYPE_QIO_CHANNEL_TLS)) {
462         Error *local_err = NULL;
463         migration_tls_channel_process_incoming(s, ioc, &local_err);
464         if (local_err) {
465             error_report_err(local_err);
466         }
467     } else {
468         QEMUFile *f = qemu_fopen_channel_input(ioc);
469         migration_fd_process_incoming(f);
470     }
471 }
472 
473 
474 void migration_channel_connect(MigrationState *s,
475                                QIOChannel *ioc,
476                                const char *hostname)
477 {
478     trace_migration_set_outgoing_channel(
479         ioc, object_get_typename(OBJECT(ioc)), hostname);
480 
481     if (s->parameters.tls_creds &&
482         !object_dynamic_cast(OBJECT(ioc),
483                              TYPE_QIO_CHANNEL_TLS)) {
484         Error *local_err = NULL;
485         migration_tls_channel_connect(s, ioc, hostname, &local_err);
486         if (local_err) {
487             migrate_fd_error(s, local_err);
488             error_free(local_err);
489         }
490     } else {
491         QEMUFile *f = qemu_fopen_channel_output(ioc);
492 
493         s->to_dst_file = f;
494 
495         migrate_fd_connect(s);
496     }
497 }
498 
499 
500 /*
501  * Send a message on the return channel back to the source
502  * of the migration.
503  */
504 void migrate_send_rp_message(MigrationIncomingState *mis,
505                              enum mig_rp_message_type message_type,
506                              uint16_t len, void *data)
507 {
508     trace_migrate_send_rp_message((int)message_type, len);
509     qemu_mutex_lock(&mis->rp_mutex);
510     qemu_put_be16(mis->to_src_file, (unsigned int)message_type);
511     qemu_put_be16(mis->to_src_file, len);
512     qemu_put_buffer(mis->to_src_file, data, len);
513     qemu_fflush(mis->to_src_file);
514     qemu_mutex_unlock(&mis->rp_mutex);
515 }
516 
517 /*
518  * Send a 'SHUT' message on the return channel with the given value
519  * to indicate that we've finished with the RP.  Non-0 value indicates
520  * error.
521  */
522 void migrate_send_rp_shut(MigrationIncomingState *mis,
523                           uint32_t value)
524 {
525     uint32_t buf;
526 
527     buf = cpu_to_be32(value);
528     migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf);
529 }
530 
531 /*
532  * Send a 'PONG' message on the return channel with the given value
533  * (normally in response to a 'PING')
534  */
535 void migrate_send_rp_pong(MigrationIncomingState *mis,
536                           uint32_t value)
537 {
538     uint32_t buf;
539 
540     buf = cpu_to_be32(value);
541     migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf);
542 }
543 
544 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
545 {
546     MigrationCapabilityStatusList *head = NULL;
547     MigrationCapabilityStatusList *caps;
548     MigrationState *s = migrate_get_current();
549     int i;
550 
551     caps = NULL; /* silence compiler warning */
552     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
553         if (i == MIGRATION_CAPABILITY_X_COLO && !colo_supported()) {
554             continue;
555         }
556         if (head == NULL) {
557             head = g_malloc0(sizeof(*caps));
558             caps = head;
559         } else {
560             caps->next = g_malloc0(sizeof(*caps));
561             caps = caps->next;
562         }
563         caps->value =
564             g_malloc(sizeof(*caps->value));
565         caps->value->capability = i;
566         caps->value->state = s->enabled_capabilities[i];
567     }
568 
569     return head;
570 }
571 
572 MigrationParameters *qmp_query_migrate_parameters(Error **errp)
573 {
574     MigrationParameters *params;
575     MigrationState *s = migrate_get_current();
576 
577     params = g_malloc0(sizeof(*params));
578     params->has_compress_level = true;
579     params->compress_level = s->parameters.compress_level;
580     params->has_compress_threads = true;
581     params->compress_threads = s->parameters.compress_threads;
582     params->has_decompress_threads = true;
583     params->decompress_threads = s->parameters.decompress_threads;
584     params->has_cpu_throttle_initial = true;
585     params->cpu_throttle_initial = s->parameters.cpu_throttle_initial;
586     params->has_cpu_throttle_increment = true;
587     params->cpu_throttle_increment = s->parameters.cpu_throttle_increment;
588     params->has_tls_creds = !!s->parameters.tls_creds;
589     params->tls_creds = g_strdup(s->parameters.tls_creds);
590     params->has_tls_hostname = !!s->parameters.tls_hostname;
591     params->tls_hostname = g_strdup(s->parameters.tls_hostname);
592     params->has_max_bandwidth = true;
593     params->max_bandwidth = s->parameters.max_bandwidth;
594     params->has_downtime_limit = true;
595     params->downtime_limit = s->parameters.downtime_limit;
596     params->has_x_checkpoint_delay = true;
597     params->x_checkpoint_delay = s->parameters.x_checkpoint_delay;
598 
599     return params;
600 }
601 
602 /*
603  * Return true if we're already in the middle of a migration
604  * (i.e. any of the active or setup states)
605  */
606 static bool migration_is_setup_or_active(int state)
607 {
608     switch (state) {
609     case MIGRATION_STATUS_ACTIVE:
610     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
611     case MIGRATION_STATUS_SETUP:
612         return true;
613 
614     default:
615         return false;
616 
617     }
618 }
619 
620 static void get_xbzrle_cache_stats(MigrationInfo *info)
621 {
622     if (migrate_use_xbzrle()) {
623         info->has_xbzrle_cache = true;
624         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
625         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
626         info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred();
627         info->xbzrle_cache->pages = xbzrle_mig_pages_transferred();
628         info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss();
629         info->xbzrle_cache->cache_miss_rate = xbzrle_mig_cache_miss_rate();
630         info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow();
631     }
632 }
633 
634 static void populate_ram_info(MigrationInfo *info, MigrationState *s)
635 {
636     info->has_ram = true;
637     info->ram = g_malloc0(sizeof(*info->ram));
638     info->ram->transferred = ram_bytes_transferred();
639     info->ram->total = ram_bytes_total();
640     info->ram->duplicate = dup_mig_pages_transferred();
641     info->ram->skipped = skipped_mig_pages_transferred();
642     info->ram->normal = norm_mig_pages_transferred();
643     info->ram->normal_bytes = norm_mig_bytes_transferred();
644     info->ram->mbps = s->mbps;
645     info->ram->dirty_sync_count = s->dirty_sync_count;
646     info->ram->postcopy_requests = s->postcopy_requests;
647 
648     if (s->state != MIGRATION_STATUS_COMPLETED) {
649         info->ram->remaining = ram_bytes_remaining();
650         info->ram->dirty_pages_rate = s->dirty_pages_rate;
651     }
652 }
653 
654 MigrationInfo *qmp_query_migrate(Error **errp)
655 {
656     MigrationInfo *info = g_malloc0(sizeof(*info));
657     MigrationState *s = migrate_get_current();
658 
659     switch (s->state) {
660     case MIGRATION_STATUS_NONE:
661         /* no migration has happened ever */
662         break;
663     case MIGRATION_STATUS_SETUP:
664         info->has_status = true;
665         info->has_total_time = false;
666         break;
667     case MIGRATION_STATUS_ACTIVE:
668     case MIGRATION_STATUS_CANCELLING:
669         info->has_status = true;
670         info->has_total_time = true;
671         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
672             - s->total_time;
673         info->has_expected_downtime = true;
674         info->expected_downtime = s->expected_downtime;
675         info->has_setup_time = true;
676         info->setup_time = s->setup_time;
677 
678         populate_ram_info(info, s);
679 
680         if (blk_mig_active()) {
681             info->has_disk = true;
682             info->disk = g_malloc0(sizeof(*info->disk));
683             info->disk->transferred = blk_mig_bytes_transferred();
684             info->disk->remaining = blk_mig_bytes_remaining();
685             info->disk->total = blk_mig_bytes_total();
686         }
687 
688         if (cpu_throttle_active()) {
689             info->has_cpu_throttle_percentage = true;
690             info->cpu_throttle_percentage = cpu_throttle_get_percentage();
691         }
692 
693         get_xbzrle_cache_stats(info);
694         break;
695     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
696         /* Mostly the same as active; TODO add some postcopy stats */
697         info->has_status = true;
698         info->has_total_time = true;
699         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
700             - s->total_time;
701         info->has_expected_downtime = true;
702         info->expected_downtime = s->expected_downtime;
703         info->has_setup_time = true;
704         info->setup_time = s->setup_time;
705 
706         populate_ram_info(info, s);
707 
708         if (blk_mig_active()) {
709             info->has_disk = true;
710             info->disk = g_malloc0(sizeof(*info->disk));
711             info->disk->transferred = blk_mig_bytes_transferred();
712             info->disk->remaining = blk_mig_bytes_remaining();
713             info->disk->total = blk_mig_bytes_total();
714         }
715 
716         get_xbzrle_cache_stats(info);
717         break;
718     case MIGRATION_STATUS_COLO:
719         info->has_status = true;
720         /* TODO: display COLO specific information (checkpoint info etc.) */
721         break;
722     case MIGRATION_STATUS_COMPLETED:
723         get_xbzrle_cache_stats(info);
724 
725         info->has_status = true;
726         info->has_total_time = true;
727         info->total_time = s->total_time;
728         info->has_downtime = true;
729         info->downtime = s->downtime;
730         info->has_setup_time = true;
731         info->setup_time = s->setup_time;
732 
733         populate_ram_info(info, s);
734         break;
735     case MIGRATION_STATUS_FAILED:
736         info->has_status = true;
737         if (s->error) {
738             info->has_error_desc = true;
739             info->error_desc = g_strdup(error_get_pretty(s->error));
740         }
741         break;
742     case MIGRATION_STATUS_CANCELLED:
743         info->has_status = true;
744         break;
745     }
746     info->status = s->state;
747 
748     return info;
749 }
750 
751 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
752                                   Error **errp)
753 {
754     MigrationState *s = migrate_get_current();
755     MigrationCapabilityStatusList *cap;
756     bool old_postcopy_cap = migrate_postcopy_ram();
757 
758     if (migration_is_setup_or_active(s->state)) {
759         error_setg(errp, QERR_MIGRATION_ACTIVE);
760         return;
761     }
762 
763     for (cap = params; cap; cap = cap->next) {
764         if (cap->value->capability == MIGRATION_CAPABILITY_X_COLO) {
765             if (!colo_supported()) {
766                 error_setg(errp, "COLO is not currently supported, please"
767                              " configure with --enable-colo option in order to"
768                              " support COLO feature");
769                 continue;
770             }
771         }
772         s->enabled_capabilities[cap->value->capability] = cap->value->state;
773     }
774 
775     if (migrate_postcopy_ram()) {
776         if (migrate_use_compression()) {
777             /* The decompression threads asynchronously write into RAM
778              * rather than use the atomic copies needed to avoid
779              * userfaulting.  It should be possible to fix the decompression
780              * threads for compatibility in future.
781              */
782             error_report("Postcopy is not currently compatible with "
783                          "compression");
784             s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] =
785                 false;
786         }
787         /* This check is reasonably expensive, so only when it's being
788          * set the first time, also it's only the destination that needs
789          * special support.
790          */
791         if (!old_postcopy_cap && runstate_check(RUN_STATE_INMIGRATE) &&
792             !postcopy_ram_supported_by_host()) {
793             /* postcopy_ram_supported_by_host will have emitted a more
794              * detailed message
795              */
796             error_report("Postcopy is not supported");
797             s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] =
798                 false;
799         }
800     }
801 }
802 
803 void qmp_migrate_set_parameters(MigrationParameters *params, Error **errp)
804 {
805     MigrationState *s = migrate_get_current();
806 
807     if (params->has_compress_level &&
808         (params->compress_level < 0 || params->compress_level > 9)) {
809         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level",
810                    "is invalid, it should be in the range of 0 to 9");
811         return;
812     }
813     if (params->has_compress_threads &&
814         (params->compress_threads < 1 || params->compress_threads > 255)) {
815         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
816                    "compress_threads",
817                    "is invalid, it should be in the range of 1 to 255");
818         return;
819     }
820     if (params->has_decompress_threads &&
821         (params->decompress_threads < 1 || params->decompress_threads > 255)) {
822         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
823                    "decompress_threads",
824                    "is invalid, it should be in the range of 1 to 255");
825         return;
826     }
827     if (params->has_cpu_throttle_initial &&
828         (params->cpu_throttle_initial < 1 ||
829          params->cpu_throttle_initial > 99)) {
830         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
831                    "cpu_throttle_initial",
832                    "an integer in the range of 1 to 99");
833         return;
834     }
835     if (params->has_cpu_throttle_increment &&
836         (params->cpu_throttle_increment < 1 ||
837          params->cpu_throttle_increment > 99)) {
838         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
839                    "cpu_throttle_increment",
840                    "an integer in the range of 1 to 99");
841         return;
842     }
843     if (params->has_max_bandwidth &&
844         (params->max_bandwidth < 0 || params->max_bandwidth > SIZE_MAX)) {
845         error_setg(errp, "Parameter 'max_bandwidth' expects an integer in the"
846                          " range of 0 to %zu bytes/second", SIZE_MAX);
847         return;
848     }
849     if (params->has_downtime_limit &&
850         (params->downtime_limit < 0 || params->downtime_limit > 2000000)) {
851         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
852                    "downtime_limit",
853                    "an integer in the range of 0 to 2000000 milliseconds");
854         return;
855     }
856     if (params->has_x_checkpoint_delay && (params->x_checkpoint_delay < 0)) {
857         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
858                     "x_checkpoint_delay",
859                     "is invalid, it should be positive");
860     }
861 
862     if (params->has_compress_level) {
863         s->parameters.compress_level = params->compress_level;
864     }
865     if (params->has_compress_threads) {
866         s->parameters.compress_threads = params->compress_threads;
867     }
868     if (params->has_decompress_threads) {
869         s->parameters.decompress_threads = params->decompress_threads;
870     }
871     if (params->has_cpu_throttle_initial) {
872         s->parameters.cpu_throttle_initial = params->cpu_throttle_initial;
873     }
874     if (params->has_cpu_throttle_increment) {
875         s->parameters.cpu_throttle_increment = params->cpu_throttle_increment;
876     }
877     if (params->has_tls_creds) {
878         g_free(s->parameters.tls_creds);
879         s->parameters.tls_creds = g_strdup(params->tls_creds);
880     }
881     if (params->has_tls_hostname) {
882         g_free(s->parameters.tls_hostname);
883         s->parameters.tls_hostname = g_strdup(params->tls_hostname);
884     }
885     if (params->has_max_bandwidth) {
886         s->parameters.max_bandwidth = params->max_bandwidth;
887         if (s->to_dst_file) {
888             qemu_file_set_rate_limit(s->to_dst_file,
889                                 s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
890         }
891     }
892     if (params->has_downtime_limit) {
893         s->parameters.downtime_limit = params->downtime_limit;
894     }
895 
896     if (params->has_x_checkpoint_delay) {
897         s->parameters.x_checkpoint_delay = params->x_checkpoint_delay;
898     }
899 }
900 
901 
902 void qmp_migrate_start_postcopy(Error **errp)
903 {
904     MigrationState *s = migrate_get_current();
905 
906     if (!migrate_postcopy_ram()) {
907         error_setg(errp, "Enable postcopy with migrate_set_capability before"
908                          " the start of migration");
909         return;
910     }
911 
912     if (s->state == MIGRATION_STATUS_NONE) {
913         error_setg(errp, "Postcopy must be started after migration has been"
914                          " started");
915         return;
916     }
917     /*
918      * we don't error if migration has finished since that would be racy
919      * with issuing this command.
920      */
921     atomic_set(&s->start_postcopy, true);
922 }
923 
924 /* shared migration helpers */
925 
926 void migrate_set_state(int *state, int old_state, int new_state)
927 {
928     if (atomic_cmpxchg(state, old_state, new_state) == old_state) {
929         trace_migrate_set_state(new_state);
930         migrate_generate_event(new_state);
931     }
932 }
933 
934 static void migrate_fd_cleanup(void *opaque)
935 {
936     MigrationState *s = opaque;
937 
938     qemu_bh_delete(s->cleanup_bh);
939     s->cleanup_bh = NULL;
940 
941     flush_page_queue(s);
942 
943     if (s->to_dst_file) {
944         trace_migrate_fd_cleanup();
945         qemu_mutex_unlock_iothread();
946         if (s->migration_thread_running) {
947             qemu_thread_join(&s->thread);
948             s->migration_thread_running = false;
949         }
950         qemu_mutex_lock_iothread();
951 
952         migrate_compress_threads_join();
953         qemu_fclose(s->to_dst_file);
954         s->to_dst_file = NULL;
955     }
956 
957     assert((s->state != MIGRATION_STATUS_ACTIVE) &&
958            (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE));
959 
960     if (s->state == MIGRATION_STATUS_CANCELLING) {
961         migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING,
962                           MIGRATION_STATUS_CANCELLED);
963     }
964 
965     notifier_list_notify(&migration_state_notifiers, s);
966 }
967 
968 void migrate_fd_error(MigrationState *s, const Error *error)
969 {
970     trace_migrate_fd_error(error_get_pretty(error));
971     assert(s->to_dst_file == NULL);
972     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
973                       MIGRATION_STATUS_FAILED);
974     if (!s->error) {
975         s->error = error_copy(error);
976     }
977     notifier_list_notify(&migration_state_notifiers, s);
978 }
979 
980 static void migrate_fd_cancel(MigrationState *s)
981 {
982     int old_state ;
983     QEMUFile *f = migrate_get_current()->to_dst_file;
984     trace_migrate_fd_cancel();
985 
986     if (s->rp_state.from_dst_file) {
987         /* shutdown the rp socket, so causing the rp thread to shutdown */
988         qemu_file_shutdown(s->rp_state.from_dst_file);
989     }
990 
991     do {
992         old_state = s->state;
993         if (!migration_is_setup_or_active(old_state)) {
994             break;
995         }
996         migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING);
997     } while (s->state != MIGRATION_STATUS_CANCELLING);
998 
999     /*
1000      * If we're unlucky the migration code might be stuck somewhere in a
1001      * send/write while the network has failed and is waiting to timeout;
1002      * if we've got shutdown(2) available then we can force it to quit.
1003      * The outgoing qemu file gets closed in migrate_fd_cleanup that is
1004      * called in a bh, so there is no race against this cancel.
1005      */
1006     if (s->state == MIGRATION_STATUS_CANCELLING && f) {
1007         qemu_file_shutdown(f);
1008     }
1009 }
1010 
1011 void add_migration_state_change_notifier(Notifier *notify)
1012 {
1013     notifier_list_add(&migration_state_notifiers, notify);
1014 }
1015 
1016 void remove_migration_state_change_notifier(Notifier *notify)
1017 {
1018     notifier_remove(notify);
1019 }
1020 
1021 bool migration_in_setup(MigrationState *s)
1022 {
1023     return s->state == MIGRATION_STATUS_SETUP;
1024 }
1025 
1026 bool migration_has_finished(MigrationState *s)
1027 {
1028     return s->state == MIGRATION_STATUS_COMPLETED;
1029 }
1030 
1031 bool migration_has_failed(MigrationState *s)
1032 {
1033     return (s->state == MIGRATION_STATUS_CANCELLED ||
1034             s->state == MIGRATION_STATUS_FAILED);
1035 }
1036 
1037 bool migration_in_postcopy(MigrationState *s)
1038 {
1039     return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
1040 }
1041 
1042 bool migration_in_postcopy_after_devices(MigrationState *s)
1043 {
1044     return migration_in_postcopy(s) && s->postcopy_after_devices;
1045 }
1046 
1047 bool migration_is_idle(MigrationState *s)
1048 {
1049     if (!s) {
1050         s = migrate_get_current();
1051     }
1052 
1053     switch (s->state) {
1054     case MIGRATION_STATUS_NONE:
1055     case MIGRATION_STATUS_CANCELLED:
1056     case MIGRATION_STATUS_COMPLETED:
1057     case MIGRATION_STATUS_FAILED:
1058         return true;
1059     case MIGRATION_STATUS_SETUP:
1060     case MIGRATION_STATUS_CANCELLING:
1061     case MIGRATION_STATUS_ACTIVE:
1062     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
1063     case MIGRATION_STATUS_COLO:
1064         return false;
1065     case MIGRATION_STATUS__MAX:
1066         g_assert_not_reached();
1067     }
1068 
1069     return false;
1070 }
1071 
1072 MigrationState *migrate_init(const MigrationParams *params)
1073 {
1074     MigrationState *s = migrate_get_current();
1075 
1076     /*
1077      * Reinitialise all migration state, except
1078      * parameters/capabilities that the user set, and
1079      * locks.
1080      */
1081     s->bytes_xfer = 0;
1082     s->xfer_limit = 0;
1083     s->cleanup_bh = 0;
1084     s->to_dst_file = NULL;
1085     s->state = MIGRATION_STATUS_NONE;
1086     s->params = *params;
1087     s->rp_state.from_dst_file = NULL;
1088     s->rp_state.error = false;
1089     s->mbps = 0.0;
1090     s->downtime = 0;
1091     s->expected_downtime = 0;
1092     s->dirty_pages_rate = 0;
1093     s->dirty_bytes_rate = 0;
1094     s->setup_time = 0;
1095     s->dirty_sync_count = 0;
1096     s->start_postcopy = false;
1097     s->postcopy_after_devices = false;
1098     s->postcopy_requests = 0;
1099     s->migration_thread_running = false;
1100     s->last_req_rb = NULL;
1101     error_free(s->error);
1102     s->error = NULL;
1103 
1104     migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP);
1105 
1106     QSIMPLEQ_INIT(&s->src_page_requests);
1107 
1108     s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1109     return s;
1110 }
1111 
1112 static GSList *migration_blockers;
1113 
1114 int migrate_add_blocker(Error *reason, Error **errp)
1115 {
1116     if (migration_is_idle(NULL)) {
1117         migration_blockers = g_slist_prepend(migration_blockers, reason);
1118         return 0;
1119     }
1120 
1121     error_propagate(errp, error_copy(reason));
1122     error_prepend(errp, "disallowing migration blocker (migration in "
1123                       "progress) for: ");
1124     return -EBUSY;
1125 }
1126 
1127 void migrate_del_blocker(Error *reason)
1128 {
1129     migration_blockers = g_slist_remove(migration_blockers, reason);
1130 }
1131 
1132 void qmp_migrate_incoming(const char *uri, Error **errp)
1133 {
1134     Error *local_err = NULL;
1135     static bool once = true;
1136 
1137     if (!deferred_incoming) {
1138         error_setg(errp, "For use with '-incoming defer'");
1139         return;
1140     }
1141     if (!once) {
1142         error_setg(errp, "The incoming migration has already been started");
1143     }
1144 
1145     qemu_start_incoming_migration(uri, &local_err);
1146 
1147     if (local_err) {
1148         error_propagate(errp, local_err);
1149         return;
1150     }
1151 
1152     once = false;
1153 }
1154 
1155 bool migration_is_blocked(Error **errp)
1156 {
1157     if (qemu_savevm_state_blocked(errp)) {
1158         return true;
1159     }
1160 
1161     if (migration_blockers) {
1162         *errp = error_copy(migration_blockers->data);
1163         return true;
1164     }
1165 
1166     return false;
1167 }
1168 
1169 void qmp_migrate(const char *uri, bool has_blk, bool blk,
1170                  bool has_inc, bool inc, bool has_detach, bool detach,
1171                  Error **errp)
1172 {
1173     Error *local_err = NULL;
1174     MigrationState *s = migrate_get_current();
1175     MigrationParams params;
1176     const char *p;
1177 
1178     params.blk = has_blk && blk;
1179     params.shared = has_inc && inc;
1180 
1181     if (migration_is_setup_or_active(s->state) ||
1182         s->state == MIGRATION_STATUS_CANCELLING ||
1183         s->state == MIGRATION_STATUS_COLO) {
1184         error_setg(errp, QERR_MIGRATION_ACTIVE);
1185         return;
1186     }
1187     if (runstate_check(RUN_STATE_INMIGRATE)) {
1188         error_setg(errp, "Guest is waiting for an incoming migration");
1189         return;
1190     }
1191 
1192     if (migration_is_blocked(errp)) {
1193         return;
1194     }
1195 
1196     s = migrate_init(&params);
1197 
1198     if (strstart(uri, "tcp:", &p)) {
1199         tcp_start_outgoing_migration(s, p, &local_err);
1200 #ifdef CONFIG_RDMA
1201     } else if (strstart(uri, "rdma:", &p)) {
1202         rdma_start_outgoing_migration(s, p, &local_err);
1203 #endif
1204     } else if (strstart(uri, "exec:", &p)) {
1205         exec_start_outgoing_migration(s, p, &local_err);
1206     } else if (strstart(uri, "unix:", &p)) {
1207         unix_start_outgoing_migration(s, p, &local_err);
1208     } else if (strstart(uri, "fd:", &p)) {
1209         fd_start_outgoing_migration(s, p, &local_err);
1210     } else {
1211         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri",
1212                    "a valid migration protocol");
1213         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1214                           MIGRATION_STATUS_FAILED);
1215         return;
1216     }
1217 
1218     if (local_err) {
1219         migrate_fd_error(s, local_err);
1220         error_propagate(errp, local_err);
1221         return;
1222     }
1223 }
1224 
1225 void qmp_migrate_cancel(Error **errp)
1226 {
1227     migrate_fd_cancel(migrate_get_current());
1228 }
1229 
1230 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
1231 {
1232     MigrationState *s = migrate_get_current();
1233     int64_t new_size;
1234 
1235     /* Check for truncation */
1236     if (value != (size_t)value) {
1237         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1238                    "exceeding address space");
1239         return;
1240     }
1241 
1242     /* Cache should not be larger than guest ram size */
1243     if (value > ram_bytes_total()) {
1244         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1245                    "exceeds guest ram size ");
1246         return;
1247     }
1248 
1249     new_size = xbzrle_cache_resize(value);
1250     if (new_size < 0) {
1251         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1252                    "is smaller than page size");
1253         return;
1254     }
1255 
1256     s->xbzrle_cache_size = new_size;
1257 }
1258 
1259 int64_t qmp_query_migrate_cache_size(Error **errp)
1260 {
1261     return migrate_xbzrle_cache_size();
1262 }
1263 
1264 void qmp_migrate_set_speed(int64_t value, Error **errp)
1265 {
1266     MigrationParameters p = {
1267         .has_max_bandwidth = true,
1268         .max_bandwidth = value,
1269     };
1270 
1271     qmp_migrate_set_parameters(&p, errp);
1272 }
1273 
1274 void qmp_migrate_set_downtime(double value, Error **errp)
1275 {
1276     value *= 1000; /* Convert to milliseconds */
1277     value = MAX(0, MIN(INT64_MAX, value));
1278 
1279     MigrationParameters p = {
1280         .has_downtime_limit = true,
1281         .downtime_limit = value,
1282     };
1283 
1284     qmp_migrate_set_parameters(&p, errp);
1285 }
1286 
1287 bool migrate_postcopy_ram(void)
1288 {
1289     MigrationState *s;
1290 
1291     s = migrate_get_current();
1292 
1293     return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM];
1294 }
1295 
1296 bool migrate_auto_converge(void)
1297 {
1298     MigrationState *s;
1299 
1300     s = migrate_get_current();
1301 
1302     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
1303 }
1304 
1305 bool migrate_zero_blocks(void)
1306 {
1307     MigrationState *s;
1308 
1309     s = migrate_get_current();
1310 
1311     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
1312 }
1313 
1314 bool migrate_use_compression(void)
1315 {
1316     MigrationState *s;
1317 
1318     s = migrate_get_current();
1319 
1320     return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS];
1321 }
1322 
1323 int migrate_compress_level(void)
1324 {
1325     MigrationState *s;
1326 
1327     s = migrate_get_current();
1328 
1329     return s->parameters.compress_level;
1330 }
1331 
1332 int migrate_compress_threads(void)
1333 {
1334     MigrationState *s;
1335 
1336     s = migrate_get_current();
1337 
1338     return s->parameters.compress_threads;
1339 }
1340 
1341 int migrate_decompress_threads(void)
1342 {
1343     MigrationState *s;
1344 
1345     s = migrate_get_current();
1346 
1347     return s->parameters.decompress_threads;
1348 }
1349 
1350 bool migrate_use_events(void)
1351 {
1352     MigrationState *s;
1353 
1354     s = migrate_get_current();
1355 
1356     return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS];
1357 }
1358 
1359 int migrate_use_xbzrle(void)
1360 {
1361     MigrationState *s;
1362 
1363     s = migrate_get_current();
1364 
1365     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
1366 }
1367 
1368 int64_t migrate_xbzrle_cache_size(void)
1369 {
1370     MigrationState *s;
1371 
1372     s = migrate_get_current();
1373 
1374     return s->xbzrle_cache_size;
1375 }
1376 
1377 /* migration thread support */
1378 /*
1379  * Something bad happened to the RP stream, mark an error
1380  * The caller shall print or trace something to indicate why
1381  */
1382 static void mark_source_rp_bad(MigrationState *s)
1383 {
1384     s->rp_state.error = true;
1385 }
1386 
1387 static struct rp_cmd_args {
1388     ssize_t     len; /* -1 = variable */
1389     const char *name;
1390 } rp_cmd_args[] = {
1391     [MIG_RP_MSG_INVALID]        = { .len = -1, .name = "INVALID" },
1392     [MIG_RP_MSG_SHUT]           = { .len =  4, .name = "SHUT" },
1393     [MIG_RP_MSG_PONG]           = { .len =  4, .name = "PONG" },
1394     [MIG_RP_MSG_REQ_PAGES]      = { .len = 12, .name = "REQ_PAGES" },
1395     [MIG_RP_MSG_REQ_PAGES_ID]   = { .len = -1, .name = "REQ_PAGES_ID" },
1396     [MIG_RP_MSG_MAX]            = { .len = -1, .name = "MAX" },
1397 };
1398 
1399 /*
1400  * Process a request for pages received on the return path,
1401  * We're allowed to send more than requested (e.g. to round to our page size)
1402  * and we don't need to send pages that have already been sent.
1403  */
1404 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname,
1405                                        ram_addr_t start, size_t len)
1406 {
1407     long our_host_ps = getpagesize();
1408 
1409     trace_migrate_handle_rp_req_pages(rbname, start, len);
1410 
1411     /*
1412      * Since we currently insist on matching page sizes, just sanity check
1413      * we're being asked for whole host pages.
1414      */
1415     if (start & (our_host_ps-1) ||
1416        (len & (our_host_ps-1))) {
1417         error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1418                      " len: %zd", __func__, start, len);
1419         mark_source_rp_bad(ms);
1420         return;
1421     }
1422 
1423     if (ram_save_queue_pages(ms, rbname, start, len)) {
1424         mark_source_rp_bad(ms);
1425     }
1426 }
1427 
1428 /*
1429  * Handles messages sent on the return path towards the source VM
1430  *
1431  */
1432 static void *source_return_path_thread(void *opaque)
1433 {
1434     MigrationState *ms = opaque;
1435     QEMUFile *rp = ms->rp_state.from_dst_file;
1436     uint16_t header_len, header_type;
1437     uint8_t buf[512];
1438     uint32_t tmp32, sibling_error;
1439     ram_addr_t start = 0; /* =0 to silence warning */
1440     size_t  len = 0, expected_len;
1441     int res;
1442 
1443     trace_source_return_path_thread_entry();
1444     while (!ms->rp_state.error && !qemu_file_get_error(rp) &&
1445            migration_is_setup_or_active(ms->state)) {
1446         trace_source_return_path_thread_loop_top();
1447         header_type = qemu_get_be16(rp);
1448         header_len = qemu_get_be16(rp);
1449 
1450         if (header_type >= MIG_RP_MSG_MAX ||
1451             header_type == MIG_RP_MSG_INVALID) {
1452             error_report("RP: Received invalid message 0x%04x length 0x%04x",
1453                     header_type, header_len);
1454             mark_source_rp_bad(ms);
1455             goto out;
1456         }
1457 
1458         if ((rp_cmd_args[header_type].len != -1 &&
1459             header_len != rp_cmd_args[header_type].len) ||
1460             header_len > sizeof(buf)) {
1461             error_report("RP: Received '%s' message (0x%04x) with"
1462                     "incorrect length %d expecting %zu",
1463                     rp_cmd_args[header_type].name, header_type, header_len,
1464                     (size_t)rp_cmd_args[header_type].len);
1465             mark_source_rp_bad(ms);
1466             goto out;
1467         }
1468 
1469         /* We know we've got a valid header by this point */
1470         res = qemu_get_buffer(rp, buf, header_len);
1471         if (res != header_len) {
1472             error_report("RP: Failed reading data for message 0x%04x"
1473                          " read %d expected %d",
1474                          header_type, res, header_len);
1475             mark_source_rp_bad(ms);
1476             goto out;
1477         }
1478 
1479         /* OK, we have the message and the data */
1480         switch (header_type) {
1481         case MIG_RP_MSG_SHUT:
1482             sibling_error = ldl_be_p(buf);
1483             trace_source_return_path_thread_shut(sibling_error);
1484             if (sibling_error) {
1485                 error_report("RP: Sibling indicated error %d", sibling_error);
1486                 mark_source_rp_bad(ms);
1487             }
1488             /*
1489              * We'll let the main thread deal with closing the RP
1490              * we could do a shutdown(2) on it, but we're the only user
1491              * anyway, so there's nothing gained.
1492              */
1493             goto out;
1494 
1495         case MIG_RP_MSG_PONG:
1496             tmp32 = ldl_be_p(buf);
1497             trace_source_return_path_thread_pong(tmp32);
1498             break;
1499 
1500         case MIG_RP_MSG_REQ_PAGES:
1501             start = ldq_be_p(buf);
1502             len = ldl_be_p(buf + 8);
1503             migrate_handle_rp_req_pages(ms, NULL, start, len);
1504             break;
1505 
1506         case MIG_RP_MSG_REQ_PAGES_ID:
1507             expected_len = 12 + 1; /* header + termination */
1508 
1509             if (header_len >= expected_len) {
1510                 start = ldq_be_p(buf);
1511                 len = ldl_be_p(buf + 8);
1512                 /* Now we expect an idstr */
1513                 tmp32 = buf[12]; /* Length of the following idstr */
1514                 buf[13 + tmp32] = '\0';
1515                 expected_len += tmp32;
1516             }
1517             if (header_len != expected_len) {
1518                 error_report("RP: Req_Page_id with length %d expecting %zd",
1519                         header_len, expected_len);
1520                 mark_source_rp_bad(ms);
1521                 goto out;
1522             }
1523             migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len);
1524             break;
1525 
1526         default:
1527             break;
1528         }
1529     }
1530     if (qemu_file_get_error(rp)) {
1531         trace_source_return_path_thread_bad_end();
1532         mark_source_rp_bad(ms);
1533     }
1534 
1535     trace_source_return_path_thread_end();
1536 out:
1537     ms->rp_state.from_dst_file = NULL;
1538     qemu_fclose(rp);
1539     return NULL;
1540 }
1541 
1542 static int open_return_path_on_source(MigrationState *ms)
1543 {
1544 
1545     ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file);
1546     if (!ms->rp_state.from_dst_file) {
1547         return -1;
1548     }
1549 
1550     trace_open_return_path_on_source();
1551     qemu_thread_create(&ms->rp_state.rp_thread, "return path",
1552                        source_return_path_thread, ms, QEMU_THREAD_JOINABLE);
1553 
1554     trace_open_return_path_on_source_continue();
1555 
1556     return 0;
1557 }
1558 
1559 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
1560 static int await_return_path_close_on_source(MigrationState *ms)
1561 {
1562     /*
1563      * If this is a normal exit then the destination will send a SHUT and the
1564      * rp_thread will exit, however if there's an error we need to cause
1565      * it to exit.
1566      */
1567     if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
1568         /*
1569          * shutdown(2), if we have it, will cause it to unblock if it's stuck
1570          * waiting for the destination.
1571          */
1572         qemu_file_shutdown(ms->rp_state.from_dst_file);
1573         mark_source_rp_bad(ms);
1574     }
1575     trace_await_return_path_close_on_source_joining();
1576     qemu_thread_join(&ms->rp_state.rp_thread);
1577     trace_await_return_path_close_on_source_close();
1578     return ms->rp_state.error;
1579 }
1580 
1581 /*
1582  * Switch from normal iteration to postcopy
1583  * Returns non-0 on error
1584  */
1585 static int postcopy_start(MigrationState *ms, bool *old_vm_running)
1586 {
1587     int ret;
1588     QIOChannelBuffer *bioc;
1589     QEMUFile *fb;
1590     int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1591     migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE,
1592                       MIGRATION_STATUS_POSTCOPY_ACTIVE);
1593 
1594     trace_postcopy_start();
1595     qemu_mutex_lock_iothread();
1596     trace_postcopy_start_set_run();
1597 
1598     qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1599     *old_vm_running = runstate_is_running();
1600     global_state_store();
1601     ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1602     if (ret < 0) {
1603         goto fail;
1604     }
1605 
1606     ret = bdrv_inactivate_all();
1607     if (ret < 0) {
1608         goto fail;
1609     }
1610 
1611     /*
1612      * Cause any non-postcopiable, but iterative devices to
1613      * send out their final data.
1614      */
1615     qemu_savevm_state_complete_precopy(ms->to_dst_file, true);
1616 
1617     /*
1618      * in Finish migrate and with the io-lock held everything should
1619      * be quiet, but we've potentially still got dirty pages and we
1620      * need to tell the destination to throw any pages it's already received
1621      * that are dirty
1622      */
1623     if (ram_postcopy_send_discard_bitmap(ms)) {
1624         error_report("postcopy send discard bitmap failed");
1625         goto fail;
1626     }
1627 
1628     /*
1629      * send rest of state - note things that are doing postcopy
1630      * will notice we're in POSTCOPY_ACTIVE and not actually
1631      * wrap their state up here
1632      */
1633     qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX);
1634     /* Ping just for debugging, helps line traces up */
1635     qemu_savevm_send_ping(ms->to_dst_file, 2);
1636 
1637     /*
1638      * While loading the device state we may trigger page transfer
1639      * requests and the fd must be free to process those, and thus
1640      * the destination must read the whole device state off the fd before
1641      * it starts processing it.  Unfortunately the ad-hoc migration format
1642      * doesn't allow the destination to know the size to read without fully
1643      * parsing it through each devices load-state code (especially the open
1644      * coded devices that use get/put).
1645      * So we wrap the device state up in a package with a length at the start;
1646      * to do this we use a qemu_buf to hold the whole of the device state.
1647      */
1648     bioc = qio_channel_buffer_new(4096);
1649     qio_channel_set_name(QIO_CHANNEL(bioc), "migration-postcopy-buffer");
1650     fb = qemu_fopen_channel_output(QIO_CHANNEL(bioc));
1651     object_unref(OBJECT(bioc));
1652 
1653     /*
1654      * Make sure the receiver can get incoming pages before we send the rest
1655      * of the state
1656      */
1657     qemu_savevm_send_postcopy_listen(fb);
1658 
1659     qemu_savevm_state_complete_precopy(fb, false);
1660     qemu_savevm_send_ping(fb, 3);
1661 
1662     qemu_savevm_send_postcopy_run(fb);
1663 
1664     /* <><> end of stuff going into the package */
1665 
1666     /* Now send that blob */
1667     if (qemu_savevm_send_packaged(ms->to_dst_file, bioc->data, bioc->usage)) {
1668         goto fail_closefb;
1669     }
1670     qemu_fclose(fb);
1671 
1672     /* Send a notify to give a chance for anything that needs to happen
1673      * at the transition to postcopy and after the device state; in particular
1674      * spice needs to trigger a transition now
1675      */
1676     ms->postcopy_after_devices = true;
1677     notifier_list_notify(&migration_state_notifiers, ms);
1678 
1679     ms->downtime =  qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop;
1680 
1681     qemu_mutex_unlock_iothread();
1682 
1683     /*
1684      * Although this ping is just for debug, it could potentially be
1685      * used for getting a better measurement of downtime at the source.
1686      */
1687     qemu_savevm_send_ping(ms->to_dst_file, 4);
1688 
1689     ret = qemu_file_get_error(ms->to_dst_file);
1690     if (ret) {
1691         error_report("postcopy_start: Migration stream errored");
1692         migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1693                               MIGRATION_STATUS_FAILED);
1694     }
1695 
1696     return ret;
1697 
1698 fail_closefb:
1699     qemu_fclose(fb);
1700 fail:
1701     migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1702                           MIGRATION_STATUS_FAILED);
1703     qemu_mutex_unlock_iothread();
1704     return -1;
1705 }
1706 
1707 /**
1708  * migration_completion: Used by migration_thread when there's not much left.
1709  *   The caller 'breaks' the loop when this returns.
1710  *
1711  * @s: Current migration state
1712  * @current_active_state: The migration state we expect to be in
1713  * @*old_vm_running: Pointer to old_vm_running flag
1714  * @*start_time: Pointer to time to update
1715  */
1716 static void migration_completion(MigrationState *s, int current_active_state,
1717                                  bool *old_vm_running,
1718                                  int64_t *start_time)
1719 {
1720     int ret;
1721 
1722     if (s->state == MIGRATION_STATUS_ACTIVE) {
1723         qemu_mutex_lock_iothread();
1724         *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1725         qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1726         *old_vm_running = runstate_is_running();
1727         ret = global_state_store();
1728 
1729         if (!ret) {
1730             ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1731             /*
1732              * Don't mark the image with BDRV_O_INACTIVE flag if
1733              * we will go into COLO stage later.
1734              */
1735             if (ret >= 0 && !migrate_colo_enabled()) {
1736                 ret = bdrv_inactivate_all();
1737             }
1738             if (ret >= 0) {
1739                 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX);
1740                 qemu_savevm_state_complete_precopy(s->to_dst_file, false);
1741             }
1742         }
1743         qemu_mutex_unlock_iothread();
1744 
1745         if (ret < 0) {
1746             goto fail;
1747         }
1748     } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1749         trace_migration_completion_postcopy_end();
1750 
1751         qemu_savevm_state_complete_postcopy(s->to_dst_file);
1752         trace_migration_completion_postcopy_end_after_complete();
1753     }
1754 
1755     /*
1756      * If rp was opened we must clean up the thread before
1757      * cleaning everything else up (since if there are no failures
1758      * it will wait for the destination to send it's status in
1759      * a SHUT command).
1760      * Postcopy opens rp if enabled (even if it's not avtivated)
1761      */
1762     if (migrate_postcopy_ram()) {
1763         int rp_error;
1764         trace_migration_completion_postcopy_end_before_rp();
1765         rp_error = await_return_path_close_on_source(s);
1766         trace_migration_completion_postcopy_end_after_rp(rp_error);
1767         if (rp_error) {
1768             goto fail_invalidate;
1769         }
1770     }
1771 
1772     if (qemu_file_get_error(s->to_dst_file)) {
1773         trace_migration_completion_file_err();
1774         goto fail_invalidate;
1775     }
1776 
1777     if (!migrate_colo_enabled()) {
1778         migrate_set_state(&s->state, current_active_state,
1779                           MIGRATION_STATUS_COMPLETED);
1780     }
1781 
1782     return;
1783 
1784 fail_invalidate:
1785     /* If not doing postcopy, vm_start() will be called: let's regain
1786      * control on images.
1787      */
1788     if (s->state == MIGRATION_STATUS_ACTIVE) {
1789         Error *local_err = NULL;
1790 
1791         bdrv_invalidate_cache_all(&local_err);
1792         if (local_err) {
1793             error_report_err(local_err);
1794         }
1795     }
1796 
1797 fail:
1798     migrate_set_state(&s->state, current_active_state,
1799                       MIGRATION_STATUS_FAILED);
1800 }
1801 
1802 bool migrate_colo_enabled(void)
1803 {
1804     MigrationState *s = migrate_get_current();
1805     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_COLO];
1806 }
1807 
1808 /*
1809  * Master migration thread on the source VM.
1810  * It drives the migration and pumps the data down the outgoing channel.
1811  */
1812 static void *migration_thread(void *opaque)
1813 {
1814     MigrationState *s = opaque;
1815     /* Used by the bandwidth calcs, updated later */
1816     int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1817     int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
1818     int64_t initial_bytes = 0;
1819     int64_t max_size = 0;
1820     int64_t start_time = initial_time;
1821     int64_t end_time;
1822     bool old_vm_running = false;
1823     bool entered_postcopy = false;
1824     /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */
1825     enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE;
1826     bool enable_colo = migrate_colo_enabled();
1827 
1828     rcu_register_thread();
1829 
1830     qemu_savevm_state_header(s->to_dst_file);
1831 
1832     if (migrate_postcopy_ram()) {
1833         /* Now tell the dest that it should open its end so it can reply */
1834         qemu_savevm_send_open_return_path(s->to_dst_file);
1835 
1836         /* And do a ping that will make stuff easier to debug */
1837         qemu_savevm_send_ping(s->to_dst_file, 1);
1838 
1839         /*
1840          * Tell the destination that we *might* want to do postcopy later;
1841          * if the other end can't do postcopy it should fail now, nice and
1842          * early.
1843          */
1844         qemu_savevm_send_postcopy_advise(s->to_dst_file);
1845     }
1846 
1847     qemu_savevm_state_begin(s->to_dst_file, &s->params);
1848 
1849     s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
1850     current_active_state = MIGRATION_STATUS_ACTIVE;
1851     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1852                       MIGRATION_STATUS_ACTIVE);
1853 
1854     trace_migration_thread_setup_complete();
1855 
1856     while (s->state == MIGRATION_STATUS_ACTIVE ||
1857            s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1858         int64_t current_time;
1859         uint64_t pending_size;
1860 
1861         if (!qemu_file_rate_limit(s->to_dst_file)) {
1862             uint64_t pend_post, pend_nonpost;
1863 
1864             qemu_savevm_state_pending(s->to_dst_file, max_size, &pend_nonpost,
1865                                       &pend_post);
1866             pending_size = pend_nonpost + pend_post;
1867             trace_migrate_pending(pending_size, max_size,
1868                                   pend_post, pend_nonpost);
1869             if (pending_size && pending_size >= max_size) {
1870                 /* Still a significant amount to transfer */
1871 
1872                 if (migrate_postcopy_ram() &&
1873                     s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE &&
1874                     pend_nonpost <= max_size &&
1875                     atomic_read(&s->start_postcopy)) {
1876 
1877                     if (!postcopy_start(s, &old_vm_running)) {
1878                         current_active_state = MIGRATION_STATUS_POSTCOPY_ACTIVE;
1879                         entered_postcopy = true;
1880                     }
1881 
1882                     continue;
1883                 }
1884                 /* Just another iteration step */
1885                 qemu_savevm_state_iterate(s->to_dst_file, entered_postcopy);
1886             } else {
1887                 trace_migration_thread_low_pending(pending_size);
1888                 migration_completion(s, current_active_state,
1889                                      &old_vm_running, &start_time);
1890                 break;
1891             }
1892         }
1893 
1894         if (qemu_file_get_error(s->to_dst_file)) {
1895             migrate_set_state(&s->state, current_active_state,
1896                               MIGRATION_STATUS_FAILED);
1897             trace_migration_thread_file_err();
1898             break;
1899         }
1900         current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1901         if (current_time >= initial_time + BUFFER_DELAY) {
1902             uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) -
1903                                          initial_bytes;
1904             uint64_t time_spent = current_time - initial_time;
1905             double bandwidth = (double)transferred_bytes / time_spent;
1906             max_size = bandwidth * s->parameters.downtime_limit;
1907 
1908             s->mbps = (((double) transferred_bytes * 8.0) /
1909                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0;
1910 
1911             trace_migrate_transferred(transferred_bytes, time_spent,
1912                                       bandwidth, max_size);
1913             /* if we haven't sent anything, we don't want to recalculate
1914                10000 is a small enough number for our purposes */
1915             if (s->dirty_bytes_rate && transferred_bytes > 10000) {
1916                 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
1917             }
1918 
1919             qemu_file_reset_rate_limit(s->to_dst_file);
1920             initial_time = current_time;
1921             initial_bytes = qemu_ftell(s->to_dst_file);
1922         }
1923         if (qemu_file_rate_limit(s->to_dst_file)) {
1924             /* usleep expects microseconds */
1925             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
1926         }
1927     }
1928 
1929     trace_migration_thread_after_loop();
1930     /* If we enabled cpu throttling for auto-converge, turn it off. */
1931     cpu_throttle_stop();
1932     end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1933 
1934     qemu_mutex_lock_iothread();
1935     /*
1936      * The resource has been allocated by migration will be reused in COLO
1937      * process, so don't release them.
1938      */
1939     if (!enable_colo) {
1940         qemu_savevm_state_cleanup();
1941     }
1942     if (s->state == MIGRATION_STATUS_COMPLETED) {
1943         uint64_t transferred_bytes = qemu_ftell(s->to_dst_file);
1944         s->total_time = end_time - s->total_time;
1945         if (!entered_postcopy) {
1946             s->downtime = end_time - start_time;
1947         }
1948         if (s->total_time) {
1949             s->mbps = (((double) transferred_bytes * 8.0) /
1950                        ((double) s->total_time)) / 1000;
1951         }
1952         runstate_set(RUN_STATE_POSTMIGRATE);
1953     } else {
1954         if (s->state == MIGRATION_STATUS_ACTIVE && enable_colo) {
1955             migrate_start_colo_process(s);
1956             qemu_savevm_state_cleanup();
1957             /*
1958             * Fixme: we will run VM in COLO no matter its old running state.
1959             * After exited COLO, we will keep running.
1960             */
1961             old_vm_running = true;
1962         }
1963         if (old_vm_running && !entered_postcopy) {
1964             vm_start();
1965         } else {
1966             if (runstate_check(RUN_STATE_FINISH_MIGRATE)) {
1967                 runstate_set(RUN_STATE_POSTMIGRATE);
1968             }
1969         }
1970     }
1971     qemu_bh_schedule(s->cleanup_bh);
1972     qemu_mutex_unlock_iothread();
1973 
1974     rcu_unregister_thread();
1975     return NULL;
1976 }
1977 
1978 void migrate_fd_connect(MigrationState *s)
1979 {
1980     s->expected_downtime = s->parameters.downtime_limit;
1981     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
1982 
1983     qemu_file_set_blocking(s->to_dst_file, true);
1984     qemu_file_set_rate_limit(s->to_dst_file,
1985                              s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
1986 
1987     /* Notify before starting migration thread */
1988     notifier_list_notify(&migration_state_notifiers, s);
1989 
1990     /*
1991      * Open the return path; currently for postcopy but other things might
1992      * also want it.
1993      */
1994     if (migrate_postcopy_ram()) {
1995         if (open_return_path_on_source(s)) {
1996             error_report("Unable to open return-path for postcopy");
1997             migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1998                               MIGRATION_STATUS_FAILED);
1999             migrate_fd_cleanup(s);
2000             return;
2001         }
2002     }
2003 
2004     migrate_compress_threads_create();
2005     qemu_thread_create(&s->thread, "migration", migration_thread, s,
2006                        QEMU_THREAD_JOINABLE);
2007     s->migration_thread_running = true;
2008 }
2009 
2010 PostcopyState  postcopy_state_get(void)
2011 {
2012     return atomic_mb_read(&incoming_postcopy_state);
2013 }
2014 
2015 /* Set the state and return the old state */
2016 PostcopyState postcopy_state_set(PostcopyState new_state)
2017 {
2018     return atomic_xchg(&incoming_postcopy_state, new_state);
2019 }
2020 
2021