migration/migration.c

   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-common.h"
  17 #include "qemu/error-report.h"
  18 #include "qemu/main-loop.h"
  19 #include "migration/migration.h"
  20 #include "migration/qemu-file.h"
  21 #include "sysemu/sysemu.h"
  22 #include "block/block.h"
  23 #include "qapi/qmp/qerror.h"
  24 #include "qemu/sockets.h"
  25 #include "qemu/rcu.h"
  26 #include "migration/block.h"
  27 #include "qemu/thread.h"
  28 #include "qmp-commands.h"
  29 #include "trace.h"
  30 #include "qapi/util.h"
  31 #include "qapi-event.h"
  32 #include "qom/cpu.h"
  33
  34 #define MAX_THROTTLE  (32 << 20)      /* Migration transfer speed throttling */
  35
  36 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
  37  * data. */
  38 #define BUFFER_DELAY     100
  39 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
  40
  41 /* Default compression thread count */
  42 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
  43 /* Default decompression thread count, usually decompression is at
  44  * least 4 times as fast as compression.*/
  45 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2
  46 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */
  47 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1
  48 /* Define default autoconverge cpu throttle migration parameters */
  49 #define DEFAULT_MIGRATE_X_CPU_THROTTLE_INITIAL 20
  50 #define DEFAULT_MIGRATE_X_CPU_THROTTLE_INCREMENT 10
  51
  52 /* Migration XBZRLE default cache size */
  53 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
  54
  55 static NotifierList migration_state_notifiers =
  56     NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
  57
  58 static bool deferred_incoming;
  59
  60 /* When we add fault tolerance, we could have several
  61    migrations at once.  For now we don't need to add
  62    dynamic creation of migration */
  63
  64 /* For outgoing */
  65 MigrationState *migrate_get_current(void)
  66 {
  67     static MigrationState current_migration = {
  68         .state = MIGRATION_STATUS_NONE,
  69         .bandwidth_limit = MAX_THROTTLE,
  70         .xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE,
  71         .mbps = -1,
  72         .parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL] =
  73                 DEFAULT_MIGRATE_COMPRESS_LEVEL,
  74         .parameters[MIGRATION_PARAMETER_COMPRESS_THREADS] =
  75                 DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT,
  76         .parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS] =
  77                 DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT,
  78         .parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INITIAL] =
  79                 DEFAULT_MIGRATE_X_CPU_THROTTLE_INITIAL,
  80         .parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INCREMENT] =
  81                 DEFAULT_MIGRATE_X_CPU_THROTTLE_INCREMENT,
  82     };
  83
  84     return &current_migration;
  85 }
  86
  87 /* For incoming */
  88 static MigrationIncomingState *mis_current;
  89
  90 MigrationIncomingState *migration_incoming_get_current(void)
  91 {
  92     return mis_current;
  93 }
  94
  95 MigrationIncomingState *migration_incoming_state_new(QEMUFile* f)
  96 {
  97     mis_current = g_new0(MigrationIncomingState, 1);
  98     mis_current->file = f;
  99     QLIST_INIT(&mis_current->loadvm_handlers);
 100
 101     return mis_current;
 102 }
 103
 104 void migration_incoming_state_destroy(void)
 105 {
 106     loadvm_free_handlers(mis_current);
 107     g_free(mis_current);
 108     mis_current = NULL;
 109 }
 110
 111
 112 typedef struct {
 113     bool optional;
 114     uint32_t size;
 115     uint8_t runstate[100];
 116     RunState state;
 117     bool received;
 118 } GlobalState;
 119
 120 static GlobalState global_state;
 121
 122 int global_state_store(void)
 123 {
 124     if (!runstate_store((char *)global_state.runstate,
 125                         sizeof(global_state.runstate))) {
 126         error_report("runstate name too big: %s", global_state.runstate);
 127         trace_migrate_state_too_big();
 128         return -EINVAL;
 129     }
 130     return 0;
 131 }
 132
 133 void global_state_store_running(void)
 134 {
 135     const char *state = RunState_lookup[RUN_STATE_RUNNING];
 136     strncpy((char *)global_state.runstate,
 137            state, sizeof(global_state.runstate));
 138 }
 139
 140 static bool global_state_received(void)
 141 {
 142     return global_state.received;
 143 }
 144
 145 static RunState global_state_get_runstate(void)
 146 {
 147     return global_state.state;
 148 }
 149
 150 void global_state_set_optional(void)
 151 {
 152     global_state.optional = true;
 153 }
 154
 155 static bool global_state_needed(void *opaque)
 156 {
 157     GlobalState *s = opaque;
 158     char *runstate = (char *)s->runstate;
 159
 160     /* If it is not optional, it is mandatory */
 161
 162     if (s->optional == false) {
 163         return true;
 164     }
 165
 166     /* If state is running or paused, it is not needed */
 167
 168     if (strcmp(runstate, "running") == 0 ||
 169         strcmp(runstate, "paused") == 0) {
 170         return false;
 171     }
 172
 173     /* for any other state it is needed */
 174     return true;
 175 }
 176
 177 static int global_state_post_load(void *opaque, int version_id)
 178 {
 179     GlobalState *s = opaque;
 180     Error *local_err = NULL;
 181     int r;
 182     char *runstate = (char *)s->runstate;
 183
 184     s->received = true;
 185     trace_migrate_global_state_post_load(runstate);
 186
 187     r = qapi_enum_parse(RunState_lookup, runstate, RUN_STATE_MAX,
 188                                 -1, &local_err);
 189
 190     if (r == -1) {
 191         if (local_err) {
 192             error_report_err(local_err);
 193         }
 194         return -EINVAL;
 195     }
 196     s->state = r;
 197
 198     return 0;
 199 }
 200
 201 static void global_state_pre_save(void *opaque)
 202 {
 203     GlobalState *s = opaque;
 204
 205     trace_migrate_global_state_pre_save((char *)s->runstate);
 206     s->size = strlen((char *)s->runstate) + 1;
 207 }
 208
 209 static const VMStateDescription vmstate_globalstate = {
 210     .name = "globalstate",
 211     .version_id = 1,
 212     .minimum_version_id = 1,
 213     .post_load = global_state_post_load,
 214     .pre_save = global_state_pre_save,
 215     .needed = global_state_needed,
 216     .fields = (VMStateField[]) {
 217         VMSTATE_UINT32(size, GlobalState),
 218         VMSTATE_BUFFER(runstate, GlobalState),
 219         VMSTATE_END_OF_LIST()
 220     },
 221 };
 222
 223 void register_global_state(void)
 224 {
 225     /* We would use it independently that we receive it */
 226     strcpy((char *)&global_state.runstate, "");
 227     global_state.received = false;
 228     vmstate_register(NULL, 0, &vmstate_globalstate, &global_state);
 229 }
 230
 231 static void migrate_generate_event(int new_state)
 232 {
 233     if (migrate_use_events()) {
 234         qapi_event_send_migration(new_state, &error_abort);
 235     }
 236 }
 237
 238 /*
 239  * Called on -incoming with a defer: uri.
 240  * The migration can be started later after any parameters have been
 241  * changed.
 242  */
 243 static void deferred_incoming_migration(Error **errp)
 244 {
 245     if (deferred_incoming) {
 246         error_setg(errp, "Incoming migration already deferred");
 247     }
 248     deferred_incoming = true;
 249 }
 250
 251 void qemu_start_incoming_migration(const char *uri, Error **errp)
 252 {
 253     const char *p;
 254
 255     qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort);
 256     if (!strcmp(uri, "defer")) {
 257         deferred_incoming_migration(errp);
 258     } else if (strstart(uri, "tcp:", &p)) {
 259         tcp_start_incoming_migration(p, errp);
 260 #ifdef CONFIG_RDMA
 261     } else if (strstart(uri, "rdma:", &p)) {
 262         rdma_start_incoming_migration(p, errp);
 263 #endif
 264 #if !defined(WIN32)
 265     } else if (strstart(uri, "exec:", &p)) {
 266         exec_start_incoming_migration(p, errp);
 267     } else if (strstart(uri, "unix:", &p)) {
 268         unix_start_incoming_migration(p, errp);
 269     } else if (strstart(uri, "fd:", &p)) {
 270         fd_start_incoming_migration(p, errp);
 271 #endif
 272     } else {
 273         error_setg(errp, "unknown migration protocol: %s", uri);
 274     }
 275 }
 276
 277 static void process_incoming_migration_co(void *opaque)
 278 {
 279     QEMUFile *f = opaque;
 280     Error *local_err = NULL;
 281     int ret;
 282
 283     migration_incoming_state_new(f);
 284     migrate_generate_event(MIGRATION_STATUS_ACTIVE);
 285     ret = qemu_loadvm_state(f);
 286
 287     qemu_fclose(f);
 288     free_xbzrle_decoded_buf();
 289     migration_incoming_state_destroy();
 290
 291     if (ret < 0) {
 292         migrate_generate_event(MIGRATION_STATUS_FAILED);
 293         error_report("load of migration failed: %s", strerror(-ret));
 294         migrate_decompress_threads_join();
 295         exit(EXIT_FAILURE);
 296     }
 297
 298     /* Make sure all file formats flush their mutable metadata */
 299     bdrv_invalidate_cache_all(&local_err);
 300     if (local_err) {
 301         migrate_generate_event(MIGRATION_STATUS_FAILED);
 302         error_report_err(local_err);
 303         migrate_decompress_threads_join();
 304         exit(EXIT_FAILURE);
 305     }
 306
 307     /*
 308      * This must happen after all error conditions are dealt with and
 309      * we're sure the VM is going to be running on this host.
 310      */
 311     qemu_announce_self();
 312
 313     /* If global state section was not received or we are in running
 314        state, we need to obey autostart. Any other state is set with
 315        runstate_set. */
 316
 317     if (!global_state_received() ||
 318         global_state_get_runstate() == RUN_STATE_RUNNING) {
 319         if (autostart) {
 320             vm_start();
 321         } else {
 322             runstate_set(RUN_STATE_PAUSED);
 323         }
 324     } else {
 325         runstate_set(global_state_get_runstate());
 326     }
 327     migrate_decompress_threads_join();
 328     /*
 329      * This must happen after any state changes since as soon as an external
 330      * observer sees this event they might start to prod at the VM assuming
 331      * it's ready to use.
 332      */
 333     migrate_generate_event(MIGRATION_STATUS_COMPLETED);
 334 }
 335
 336 void process_incoming_migration(QEMUFile *f)
 337 {
 338     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co);
 339     int fd = qemu_get_fd(f);
 340
 341     assert(fd != -1);
 342     migrate_decompress_threads_create();
 343     qemu_set_nonblock(fd);
 344     qemu_coroutine_enter(co, f);
 345 }
 346
 347 /* amount of nanoseconds we are willing to wait for migration to be down.
 348  * the choice of nanoseconds is because it is the maximum resolution that
 349  * get_clock() can achieve. It is an internal measure. All user-visible
 350  * units must be in seconds */
 351 static uint64_t max_downtime = 300000000;
 352
 353 uint64_t migrate_max_downtime(void)
 354 {
 355     return max_downtime;
 356 }
 357
 358 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
 359 {
 360     MigrationCapabilityStatusList *head = NULL;
 361     MigrationCapabilityStatusList *caps;
 362     MigrationState *s = migrate_get_current();
 363     int i;
 364
 365     caps = NULL; /* silence compiler warning */
 366     for (i = 0; i < MIGRATION_CAPABILITY_MAX; i++) {
 367         if (head == NULL) {
 368             head = g_malloc0(sizeof(*caps));
 369             caps = head;
 370         } else {
 371             caps->next = g_malloc0(sizeof(*caps));
 372             caps = caps->next;
 373         }
 374         caps->value =
 375             g_malloc(sizeof(*caps->value));
 376         caps->value->capability = i;
 377         caps->value->state = s->enabled_capabilities[i];
 378     }
 379
 380     return head;
 381 }
 382
 383 MigrationParameters *qmp_query_migrate_parameters(Error **errp)
 384 {
 385     MigrationParameters *params;
 386     MigrationState *s = migrate_get_current();
 387
 388     params = g_malloc0(sizeof(*params));
 389     params->compress_level = s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL];
 390     params->compress_threads =
 391             s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS];
 392     params->decompress_threads =
 393             s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS];
 394     params->x_cpu_throttle_initial =
 395             s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INITIAL];
 396     params->x_cpu_throttle_increment =
 397             s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INCREMENT];
 398
 399     return params;
 400 }
 401
 402 static void get_xbzrle_cache_stats(MigrationInfo *info)
 403 {
 404     if (migrate_use_xbzrle()) {
 405         info->has_xbzrle_cache = true;
 406         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
 407         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
 408         info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred();
 409         info->xbzrle_cache->pages = xbzrle_mig_pages_transferred();
 410         info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss();
 411         info->xbzrle_cache->cache_miss_rate = xbzrle_mig_cache_miss_rate();
 412         info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow();
 413     }
 414 }
 415
 416 MigrationInfo *qmp_query_migrate(Error **errp)
 417 {
 418     MigrationInfo *info = g_malloc0(sizeof(*info));
 419     MigrationState *s = migrate_get_current();
 420
 421     switch (s->state) {
 422     case MIGRATION_STATUS_NONE:
 423         /* no migration has happened ever */
 424         break;
 425     case MIGRATION_STATUS_SETUP:
 426         info->has_status = true;
 427         info->has_total_time = false;
 428         break;
 429     case MIGRATION_STATUS_ACTIVE:
 430     case MIGRATION_STATUS_CANCELLING:
 431         info->has_status = true;
 432         info->has_total_time = true;
 433         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 434             - s->total_time;
 435         info->has_expected_downtime = true;
 436         info->expected_downtime = s->expected_downtime;
 437         info->has_setup_time = true;
 438         info->setup_time = s->setup_time;
 439
 440         info->has_ram = true;
 441         info->ram = g_malloc0(sizeof(*info->ram));
 442         info->ram->transferred = ram_bytes_transferred();
 443         info->ram->remaining = ram_bytes_remaining();
 444         info->ram->total = ram_bytes_total();
 445         info->ram->duplicate = dup_mig_pages_transferred();
 446         info->ram->skipped = skipped_mig_pages_transferred();
 447         info->ram->normal = norm_mig_pages_transferred();
 448         info->ram->normal_bytes = norm_mig_bytes_transferred();
 449         info->ram->dirty_pages_rate = s->dirty_pages_rate;
 450         info->ram->mbps = s->mbps;
 451         info->ram->dirty_sync_count = s->dirty_sync_count;
 452
 453         if (blk_mig_active()) {
 454             info->has_disk = true;
 455             info->disk = g_malloc0(sizeof(*info->disk));
 456             info->disk->transferred = blk_mig_bytes_transferred();
 457             info->disk->remaining = blk_mig_bytes_remaining();
 458             info->disk->total = blk_mig_bytes_total();
 459         }
 460
 461         if (cpu_throttle_active()) {
 462             info->has_x_cpu_throttle_percentage = true;
 463             info->x_cpu_throttle_percentage = cpu_throttle_get_percentage();
 464         }
 465
 466         get_xbzrle_cache_stats(info);
 467         break;
 468     case MIGRATION_STATUS_COMPLETED:
 469         get_xbzrle_cache_stats(info);
 470
 471         info->has_status = true;
 472         info->has_total_time = true;
 473         info->total_time = s->total_time;
 474         info->has_downtime = true;
 475         info->downtime = s->downtime;
 476         info->has_setup_time = true;
 477         info->setup_time = s->setup_time;
 478
 479         info->has_ram = true;
 480         info->ram = g_malloc0(sizeof(*info->ram));
 481         info->ram->transferred = ram_bytes_transferred();
 482         info->ram->remaining = 0;
 483         info->ram->total = ram_bytes_total();
 484         info->ram->duplicate = dup_mig_pages_transferred();
 485         info->ram->skipped = skipped_mig_pages_transferred();
 486         info->ram->normal = norm_mig_pages_transferred();
 487         info->ram->normal_bytes = norm_mig_bytes_transferred();
 488         info->ram->mbps = s->mbps;
 489         info->ram->dirty_sync_count = s->dirty_sync_count;
 490         break;
 491     case MIGRATION_STATUS_FAILED:
 492         info->has_status = true;
 493         break;
 494     case MIGRATION_STATUS_CANCELLED:
 495         info->has_status = true;
 496         break;
 497     }
 498     info->status = s->state;
 499
 500     return info;
 501 }
 502
 503 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 504                                   Error **errp)
 505 {
 506     MigrationState *s = migrate_get_current();
 507     MigrationCapabilityStatusList *cap;
 508
 509     if (s->state == MIGRATION_STATUS_ACTIVE ||
 510         s->state == MIGRATION_STATUS_SETUP) {
 511         error_setg(errp, QERR_MIGRATION_ACTIVE);
 512         return;
 513     }
 514
 515     for (cap = params; cap; cap = cap->next) {
 516         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 517     }
 518 }
 519
 520 void qmp_migrate_set_parameters(bool has_compress_level,
 521                                 int64_t compress_level,
 522                                 bool has_compress_threads,
 523                                 int64_t compress_threads,
 524                                 bool has_decompress_threads,
 525                                 int64_t decompress_threads,
 526                                 bool has_x_cpu_throttle_initial,
 527                                 int64_t x_cpu_throttle_initial,
 528                                 bool has_x_cpu_throttle_increment,
 529                                 int64_t x_cpu_throttle_increment, Error **errp)
 530 {
 531     MigrationState *s = migrate_get_current();
 532
 533     if (has_compress_level && (compress_level < 0 || compress_level > 9)) {
 534         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level",
 535                    "is invalid, it should be in the range of 0 to 9");
 536         return;
 537     }
 538     if (has_compress_threads &&
 539             (compress_threads < 1 || compress_threads > 255)) {
 540         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 541                    "compress_threads",
 542                    "is invalid, it should be in the range of 1 to 255");
 543         return;
 544     }
 545     if (has_decompress_threads &&
 546             (decompress_threads < 1 || decompress_threads > 255)) {
 547         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 548                    "decompress_threads",
 549                    "is invalid, it should be in the range of 1 to 255");
 550         return;
 551     }
 552     if (has_x_cpu_throttle_initial &&
 553             (x_cpu_throttle_initial < 1 || x_cpu_throttle_initial > 99)) {
 554         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 555                    "x_cpu_throttle_initial",
 556                    "an integer in the range of 1 to 99");
 557     }
 558     if (has_x_cpu_throttle_increment &&
 559             (x_cpu_throttle_increment < 1 || x_cpu_throttle_increment > 99)) {
 560         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 561                    "x_cpu_throttle_increment",
 562                    "an integer in the range of 1 to 99");
 563     }
 564
 565     if (has_compress_level) {
 566         s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL] = compress_level;
 567     }
 568     if (has_compress_threads) {
 569         s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS] = compress_threads;
 570     }
 571     if (has_decompress_threads) {
 572         s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS] =
 573                                                     decompress_threads;
 574     }
 575     if (has_x_cpu_throttle_initial) {
 576         s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INITIAL] =
 577                                                     x_cpu_throttle_initial;
 578     }
 579
 580     if (has_x_cpu_throttle_increment) {
 581         s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INCREMENT] =
 582                                                     x_cpu_throttle_increment;
 583     }
 584 }
 585
 586 /* shared migration helpers */
 587
 588 static void migrate_set_state(MigrationState *s, int old_state, int new_state)
 589 {
 590     if (atomic_cmpxchg(&s->state, old_state, new_state) == old_state) {
 591         trace_migrate_set_state(new_state);
 592         migrate_generate_event(new_state);
 593     }
 594 }
 595
 596 static void migrate_fd_cleanup(void *opaque)
 597 {
 598     MigrationState *s = opaque;
 599
 600     qemu_bh_delete(s->cleanup_bh);
 601     s->cleanup_bh = NULL;
 602
 603     if (s->file) {
 604         trace_migrate_fd_cleanup();
 605         qemu_mutex_unlock_iothread();
 606         qemu_thread_join(&s->thread);
 607         qemu_mutex_lock_iothread();
 608
 609         migrate_compress_threads_join();
 610         qemu_fclose(s->file);
 611         s->file = NULL;
 612     }
 613
 614     assert(s->state != MIGRATION_STATUS_ACTIVE);
 615
 616     if (s->state != MIGRATION_STATUS_COMPLETED) {
 617         qemu_savevm_state_cancel();
 618         if (s->state == MIGRATION_STATUS_CANCELLING) {
 619             migrate_set_state(s, MIGRATION_STATUS_CANCELLING,
 620                               MIGRATION_STATUS_CANCELLED);
 621         }
 622     }
 623
 624     notifier_list_notify(&migration_state_notifiers, s);
 625 }
 626
 627 void migrate_fd_error(MigrationState *s)
 628 {
 629     trace_migrate_fd_error();
 630     assert(s->file == NULL);
 631     migrate_set_state(s, MIGRATION_STATUS_SETUP, MIGRATION_STATUS_FAILED);
 632     notifier_list_notify(&migration_state_notifiers, s);
 633 }
 634
 635 static void migrate_fd_cancel(MigrationState *s)
 636 {
 637     int old_state ;
 638     QEMUFile *f = migrate_get_current()->file;
 639     trace_migrate_fd_cancel();
 640
 641     do {
 642         old_state = s->state;
 643         if (old_state != MIGRATION_STATUS_SETUP &&
 644             old_state != MIGRATION_STATUS_ACTIVE) {
 645             break;
 646         }
 647         migrate_set_state(s, old_state, MIGRATION_STATUS_CANCELLING);
 648     } while (s->state != MIGRATION_STATUS_CANCELLING);
 649
 650     /*
 651      * If we're unlucky the migration code might be stuck somewhere in a
 652      * send/write while the network has failed and is waiting to timeout;
 653      * if we've got shutdown(2) available then we can force it to quit.
 654      * The outgoing qemu file gets closed in migrate_fd_cleanup that is
 655      * called in a bh, so there is no race against this cancel.
 656      */
 657     if (s->state == MIGRATION_STATUS_CANCELLING && f) {
 658         qemu_file_shutdown(f);
 659     }
 660 }
 661
 662 void add_migration_state_change_notifier(Notifier *notify)
 663 {
 664     notifier_list_add(&migration_state_notifiers, notify);
 665 }
 666
 667 void remove_migration_state_change_notifier(Notifier *notify)
 668 {
 669     notifier_remove(notify);
 670 }
 671
 672 bool migration_in_setup(MigrationState *s)
 673 {
 674     return s->state == MIGRATION_STATUS_SETUP;
 675 }
 676
 677 bool migration_has_finished(MigrationState *s)
 678 {
 679     return s->state == MIGRATION_STATUS_COMPLETED;
 680 }
 681
 682 bool migration_has_failed(MigrationState *s)
 683 {
 684     return (s->state == MIGRATION_STATUS_CANCELLED ||
 685             s->state == MIGRATION_STATUS_FAILED);
 686 }
 687
 688 static MigrationState *migrate_init(const MigrationParams *params)
 689 {
 690     MigrationState *s = migrate_get_current();
 691     int64_t bandwidth_limit = s->bandwidth_limit;
 692     bool enabled_capabilities[MIGRATION_CAPABILITY_MAX];
 693     int64_t xbzrle_cache_size = s->xbzrle_cache_size;
 694     int compress_level = s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL];
 695     int compress_thread_count =
 696             s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS];
 697     int decompress_thread_count =
 698             s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS];
 699     int x_cpu_throttle_initial =
 700             s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INITIAL];
 701     int x_cpu_throttle_increment =
 702             s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INCREMENT];
 703
 704     memcpy(enabled_capabilities, s->enabled_capabilities,
 705            sizeof(enabled_capabilities));
 706
 707     memset(s, 0, sizeof(*s));
 708     s->params = *params;
 709     memcpy(s->enabled_capabilities, enabled_capabilities,
 710            sizeof(enabled_capabilities));
 711     s->xbzrle_cache_size = xbzrle_cache_size;
 712
 713     s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL] = compress_level;
 714     s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS] =
 715                compress_thread_count;
 716     s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS] =
 717                decompress_thread_count;
 718     s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INITIAL] =
 719                 x_cpu_throttle_initial;
 720     s->parameters[MIGRATION_PARAMETER_X_CPU_THROTTLE_INCREMENT] =
 721                 x_cpu_throttle_increment;
 722     s->bandwidth_limit = bandwidth_limit;
 723     migrate_set_state(s, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP);
 724
 725     s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 726     return s;
 727 }
 728
 729 static GSList *migration_blockers;
 730
 731 void migrate_add_blocker(Error *reason)
 732 {
 733     migration_blockers = g_slist_prepend(migration_blockers, reason);
 734 }
 735
 736 void migrate_del_blocker(Error *reason)
 737 {
 738     migration_blockers = g_slist_remove(migration_blockers, reason);
 739 }
 740
 741 void qmp_migrate_incoming(const char *uri, Error **errp)
 742 {
 743     Error *local_err = NULL;
 744     static bool once = true;
 745
 746     if (!deferred_incoming) {
 747         error_setg(errp, "For use with '-incoming defer'");
 748         return;
 749     }
 750     if (!once) {
 751         error_setg(errp, "The incoming migration has already been started");
 752     }
 753
 754     qemu_start_incoming_migration(uri, &local_err);
 755
 756     if (local_err) {
 757         error_propagate(errp, local_err);
 758         return;
 759     }
 760
 761     once = false;
 762 }
 763
 764 void qmp_migrate(const char *uri, bool has_blk, bool blk,
 765                  bool has_inc, bool inc, bool has_detach, bool detach,
 766                  Error **errp)
 767 {
 768     Error *local_err = NULL;
 769     MigrationState *s = migrate_get_current();
 770     MigrationParams params;
 771     const char *p;
 772
 773     params.blk = has_blk && blk;
 774     params.shared = has_inc && inc;
 775
 776     if (s->state == MIGRATION_STATUS_ACTIVE ||
 777         s->state == MIGRATION_STATUS_SETUP ||
 778         s->state == MIGRATION_STATUS_CANCELLING) {
 779         error_setg(errp, QERR_MIGRATION_ACTIVE);
 780         return;
 781     }
 782     if (runstate_check(RUN_STATE_INMIGRATE)) {
 783         error_setg(errp, "Guest is waiting for an incoming migration");
 784         return;
 785     }
 786
 787     if (qemu_savevm_state_blocked(errp)) {
 788         return;
 789     }
 790
 791     if (migration_blockers) {
 792         *errp = error_copy(migration_blockers->data);
 793         return;
 794     }
 795
 796     /* We are starting a new migration, so we want to start in a clean
 797        state.  This change is only needed if previous migration
 798        failed/was cancelled.  We don't use migrate_set_state() because
 799        we are setting the initial state, not changing it. */
 800     s->state = MIGRATION_STATUS_NONE;
 801
 802     s = migrate_init(&params);
 803
 804     if (strstart(uri, "tcp:", &p)) {
 805         tcp_start_outgoing_migration(s, p, &local_err);
 806 #ifdef CONFIG_RDMA
 807     } else if (strstart(uri, "rdma:", &p)) {
 808         rdma_start_outgoing_migration(s, p, &local_err);
 809 #endif
 810 #if !defined(WIN32)
 811     } else if (strstart(uri, "exec:", &p)) {
 812         exec_start_outgoing_migration(s, p, &local_err);
 813     } else if (strstart(uri, "unix:", &p)) {
 814         unix_start_outgoing_migration(s, p, &local_err);
 815     } else if (strstart(uri, "fd:", &p)) {
 816         fd_start_outgoing_migration(s, p, &local_err);
 817 #endif
 818     } else {
 819         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri",
 820                    "a valid migration protocol");
 821         migrate_set_state(s, MIGRATION_STATUS_SETUP, MIGRATION_STATUS_FAILED);
 822         return;
 823     }
 824
 825     if (local_err) {
 826         migrate_fd_error(s);
 827         error_propagate(errp, local_err);
 828         return;
 829     }
 830 }
 831
 832 void qmp_migrate_cancel(Error **errp)
 833 {
 834     migrate_fd_cancel(migrate_get_current());
 835 }
 836
 837 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
 838 {
 839     MigrationState *s = migrate_get_current();
 840     int64_t new_size;
 841
 842     /* Check for truncation */
 843     if (value != (size_t)value) {
 844         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
 845                    "exceeding address space");
 846         return;
 847     }
 848
 849     /* Cache should not be larger than guest ram size */
 850     if (value > ram_bytes_total()) {
 851         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
 852                    "exceeds guest ram size ");
 853         return;
 854     }
 855
 856     new_size = xbzrle_cache_resize(value);
 857     if (new_size < 0) {
 858         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
 859                    "is smaller than page size");
 860         return;
 861     }
 862
 863     s->xbzrle_cache_size = new_size;
 864 }
 865
 866 int64_t qmp_query_migrate_cache_size(Error **errp)
 867 {
 868     return migrate_xbzrle_cache_size();
 869 }
 870
 871 void qmp_migrate_set_speed(int64_t value, Error **errp)
 872 {
 873     MigrationState *s;
 874
 875     if (value < 0) {
 876         value = 0;
 877     }
 878     if (value > SIZE_MAX) {
 879         value = SIZE_MAX;
 880     }
 881
 882     s = migrate_get_current();
 883     s->bandwidth_limit = value;
 884     if (s->file) {
 885         qemu_file_set_rate_limit(s->file, s->bandwidth_limit / XFER_LIMIT_RATIO);
 886     }
 887 }
 888
 889 void qmp_migrate_set_downtime(double value, Error **errp)
 890 {
 891     value *= 1e9;
 892     value = MAX(0, MIN(UINT64_MAX, value));
 893     max_downtime = (uint64_t)value;
 894 }
 895
 896 bool migrate_auto_converge(void)
 897 {
 898     MigrationState *s;
 899
 900     s = migrate_get_current();
 901
 902     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
 903 }
 904
 905 bool migrate_zero_blocks(void)
 906 {
 907     MigrationState *s;
 908
 909     s = migrate_get_current();
 910
 911     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
 912 }
 913
 914 bool migrate_use_compression(void)
 915 {
 916     MigrationState *s;
 917
 918     s = migrate_get_current();
 919
 920     return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS];
 921 }
 922
 923 int migrate_compress_level(void)
 924 {
 925     MigrationState *s;
 926
 927     s = migrate_get_current();
 928
 929     return s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL];
 930 }
 931
 932 int migrate_compress_threads(void)
 933 {
 934     MigrationState *s;
 935
 936     s = migrate_get_current();
 937
 938     return s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS];
 939 }
 940
 941 int migrate_decompress_threads(void)
 942 {
 943     MigrationState *s;
 944
 945     s = migrate_get_current();
 946
 947     return s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS];
 948 }
 949
 950 bool migrate_use_events(void)
 951 {
 952     MigrationState *s;
 953
 954     s = migrate_get_current();
 955
 956     return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS];
 957 }
 958
 959 int migrate_use_xbzrle(void)
 960 {
 961     MigrationState *s;
 962
 963     s = migrate_get_current();
 964
 965     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
 966 }
 967
 968 int64_t migrate_xbzrle_cache_size(void)
 969 {
 970     MigrationState *s;
 971
 972     s = migrate_get_current();
 973
 974     return s->xbzrle_cache_size;
 975 }
 976
 977 /**
 978  * migration_completion: Used by migration_thread when there's not much left.
 979  *   The caller 'breaks' the loop when this returns.
 980  *
 981  * @s: Current migration state
 982  * @*old_vm_running: Pointer to old_vm_running flag
 983  * @*start_time: Pointer to time to update
 984  */
 985 static void migration_completion(MigrationState *s, bool *old_vm_running,
 986                                  int64_t *start_time)
 987 {
 988     int ret;
 989
 990     qemu_mutex_lock_iothread();
 991     *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 992     qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
 993     *old_vm_running = runstate_is_running();
 994
 995     ret = global_state_store();
 996     if (!ret) {
 997         ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
 998         if (ret >= 0) {
 999             qemu_file_set_rate_limit(s->file, INT64_MAX);
1000             qemu_savevm_state_complete(s->file);
1001         }
1002     }
1003     qemu_mutex_unlock_iothread();
1004
1005     if (ret < 0) {
1006         goto fail;
1007     }
1008
1009     if (qemu_file_get_error(s->file)) {
1010         trace_migration_completion_file_err();
1011         goto fail;
1012     }
1013
1014     migrate_set_state(s, MIGRATION_STATUS_ACTIVE, MIGRATION_STATUS_COMPLETED);
1015     return;
1016
1017 fail:
1018     migrate_set_state(s, MIGRATION_STATUS_ACTIVE, MIGRATION_STATUS_FAILED);
1019 }
1020
1021 /* migration thread support */
1022
1023 static void *migration_thread(void *opaque)
1024 {
1025     MigrationState *s = opaque;
1026     int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1027     int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
1028     int64_t initial_bytes = 0;
1029     int64_t max_size = 0;
1030     int64_t start_time = initial_time;
1031     bool old_vm_running = false;
1032
1033     rcu_register_thread();
1034
1035     qemu_savevm_state_header(s->file);
1036     qemu_savevm_state_begin(s->file, &s->params);
1037
1038     s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
1039     migrate_set_state(s, MIGRATION_STATUS_SETUP, MIGRATION_STATUS_ACTIVE);
1040
1041     while (s->state == MIGRATION_STATUS_ACTIVE) {
1042         int64_t current_time;
1043         uint64_t pending_size;
1044
1045         if (!qemu_file_rate_limit(s->file)) {
1046             pending_size = qemu_savevm_state_pending(s->file, max_size);
1047             trace_migrate_pending(pending_size, max_size);
1048             if (pending_size && pending_size >= max_size) {
1049                 qemu_savevm_state_iterate(s->file);
1050             } else {
1051                 trace_migration_thread_low_pending(pending_size);
1052                 migration_completion(s, &old_vm_running, &start_time);
1053                 break;
1054             }
1055         }
1056
1057         if (qemu_file_get_error(s->file)) {
1058             migrate_set_state(s, MIGRATION_STATUS_ACTIVE,
1059                               MIGRATION_STATUS_FAILED);
1060             break;
1061         }
1062         current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1063         if (current_time >= initial_time + BUFFER_DELAY) {
1064             uint64_t transferred_bytes = qemu_ftell(s->file) - initial_bytes;
1065             uint64_t time_spent = current_time - initial_time;
1066             double bandwidth = transferred_bytes / time_spent;
1067             max_size = bandwidth * migrate_max_downtime() / 1000000;
1068
1069             s->mbps = time_spent ? (((double) transferred_bytes * 8.0) /
1070                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0 : -1;
1071
1072             trace_migrate_transferred(transferred_bytes, time_spent,
1073                                       bandwidth, max_size);
1074             /* if we haven't sent anything, we don't want to recalculate
1075                10000 is a small enough number for our purposes */
1076             if (s->dirty_bytes_rate && transferred_bytes > 10000) {
1077                 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
1078             }
1079
1080             qemu_file_reset_rate_limit(s->file);
1081             initial_time = current_time;
1082             initial_bytes = qemu_ftell(s->file);
1083         }
1084         if (qemu_file_rate_limit(s->file)) {
1085             /* usleep expects microseconds */
1086             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
1087         }
1088     }
1089
1090     /* If we enabled cpu throttling for auto-converge, turn it off. */
1091     cpu_throttle_stop();
1092
1093     qemu_mutex_lock_iothread();
1094     if (s->state == MIGRATION_STATUS_COMPLETED) {
1095         int64_t end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1096         uint64_t transferred_bytes = qemu_ftell(s->file);
1097         s->total_time = end_time - s->total_time;
1098         s->downtime = end_time - start_time;
1099         if (s->total_time) {
1100             s->mbps = (((double) transferred_bytes * 8.0) /
1101                        ((double) s->total_time)) / 1000;
1102         }
1103         runstate_set(RUN_STATE_POSTMIGRATE);
1104     } else {
1105         if (old_vm_running) {
1106             vm_start();
1107         }
1108     }
1109     qemu_bh_schedule(s->cleanup_bh);
1110     qemu_mutex_unlock_iothread();
1111
1112     rcu_unregister_thread();
1113     return NULL;
1114 }
1115
1116 void migrate_fd_connect(MigrationState *s)
1117 {
1118     /* This is a best 1st approximation. ns to ms */
1119     s->expected_downtime = max_downtime/1000000;
1120     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
1121
1122     qemu_file_set_rate_limit(s->file,
1123                              s->bandwidth_limit / XFER_LIMIT_RATIO);
1124
1125     /* Notify before starting migration thread */
1126     notifier_list_notify(&migration_state_notifiers, s);
1127
1128     migrate_compress_threads_create();
1129     qemu_thread_create(&s->thread, "migration", migration_thread, s,
1130                        QEMU_THREAD_JOINABLE);
1131 }