migration.c

   1 /*
   2  * QEMU live migration
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *
   6  * Authors:
   7  *  Anthony Liguori   <[email protected]>
   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/main-loop.h"
  18 #include "migration/migration.h"
  19 #include "monitor/monitor.h"
  20 #include "migration/qemu-file.h"
  21 #include "sysemu/sysemu.h"
  22 #include "block/block.h"
  23 #include "qemu/sockets.h"
  24 #include "migration/block.h"
  25 #include "qemu/thread.h"
  26 #include "qmp-commands.h"
  27 #include "trace.h"
  28
  29 //#define DEBUG_MIGRATION
  30
  31 #ifdef DEBUG_MIGRATION
  32 #define DPRINTF(fmt, ...) \
  33     do { printf("migration: " fmt, ## __VA_ARGS__); } while (0)
  34 #else
  35 #define DPRINTF(fmt, ...) \
  36     do { } while (0)
  37 #endif
  38
  39 enum {
  40     MIG_STATE_ERROR = -1,
  41     MIG_STATE_NONE,
  42     MIG_STATE_SETUP,
  43     MIG_STATE_CANCELLED,
  44     MIG_STATE_ACTIVE,
  45     MIG_STATE_COMPLETED,
  46 };
  47
  48 #define MAX_THROTTLE  (32 << 20)      /* Migration speed throttling */
  49
  50 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
  51  * data. */
  52 #define BUFFER_DELAY     100
  53 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
  54
  55 /* Migration XBZRLE default cache size */
  56 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
  57
  58 static NotifierList migration_state_notifiers =
  59     NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
  60
  61 /* When we add fault tolerance, we could have several
  62    migrations at once.  For now we don't need to add
  63    dynamic creation of migration */
  64
  65 MigrationState *migrate_get_current(void)
  66 {
  67     static MigrationState current_migration = {
  68         .state = MIG_STATE_NONE,
  69         .bandwidth_limit = MAX_THROTTLE,
  70         .xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE,
  71         .mbps = -1,
  72     };
  73
  74     return &current_migration;
  75 }
  76
  77 void qemu_start_incoming_migration(const char *uri, Error **errp)
  78 {
  79     const char *p;
  80
  81     if (strstart(uri, "tcp:", &p))
  82         tcp_start_incoming_migration(p, errp);
  83 #ifdef CONFIG_RDMA
  84     else if (strstart(uri, "x-rdma:", &p))
  85         rdma_start_incoming_migration(p, errp);
  86 #endif
  87 #if !defined(WIN32)
  88     else if (strstart(uri, "exec:", &p))
  89         exec_start_incoming_migration(p, errp);
  90     else if (strstart(uri, "unix:", &p))
  91         unix_start_incoming_migration(p, errp);
  92     else if (strstart(uri, "fd:", &p))
  93         fd_start_incoming_migration(p, errp);
  94 #endif
  95     else {
  96         error_setg(errp, "unknown migration protocol: %s", uri);
  97     }
  98 }
  99
 100 static void process_incoming_migration_co(void *opaque)
 101 {
 102     QEMUFile *f = opaque;
 103     int ret;
 104
 105     ret = qemu_loadvm_state(f);
 106     qemu_fclose(f);
 107     if (ret < 0) {
 108         fprintf(stderr, "load of migration failed\n");
 109         exit(EXIT_FAILURE);
 110     }
 111     qemu_announce_self();
 112     DPRINTF("successfully loaded vm state\n");
 113
 114     bdrv_clear_incoming_migration_all();
 115     /* Make sure all file formats flush their mutable metadata */
 116     bdrv_invalidate_cache_all();
 117
 118     if (autostart) {
 119         vm_start();
 120     } else {
 121         runstate_set(RUN_STATE_PAUSED);
 122     }
 123 }
 124
 125 void process_incoming_migration(QEMUFile *f)
 126 {
 127     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co);
 128     int fd = qemu_get_fd(f);
 129
 130     assert(fd != -1);
 131     qemu_set_nonblock(fd);
 132     qemu_coroutine_enter(co, f);
 133 }
 134
 135 /* amount of nanoseconds we are willing to wait for migration to be down.
 136  * the choice of nanoseconds is because it is the maximum resolution that
 137  * get_clock() can achieve. It is an internal measure. All user-visible
 138  * units must be in seconds */
 139 static uint64_t max_downtime = 30000000;
 140
 141 uint64_t migrate_max_downtime(void)
 142 {
 143     return max_downtime;
 144 }
 145
 146 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
 147 {
 148     MigrationCapabilityStatusList *head = NULL;
 149     MigrationCapabilityStatusList *caps;
 150     MigrationState *s = migrate_get_current();
 151     int i;
 152
 153     for (i = 0; i < MIGRATION_CAPABILITY_MAX; i++) {
 154         if (head == NULL) {
 155             head = g_malloc0(sizeof(*caps));
 156             caps = head;
 157         } else {
 158             caps->next = g_malloc0(sizeof(*caps));
 159             caps = caps->next;
 160         }
 161         caps->value =
 162             g_malloc(sizeof(*caps->value));
 163         caps->value->capability = i;
 164         caps->value->state = s->enabled_capabilities[i];
 165     }
 166
 167     return head;
 168 }
 169
 170 static void get_xbzrle_cache_stats(MigrationInfo *info)
 171 {
 172     if (migrate_use_xbzrle()) {
 173         info->has_xbzrle_cache = true;
 174         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
 175         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
 176         info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred();
 177         info->xbzrle_cache->pages = xbzrle_mig_pages_transferred();
 178         info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss();
 179         info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow();
 180     }
 181 }
 182
 183 MigrationInfo *qmp_query_migrate(Error **errp)
 184 {
 185     MigrationInfo *info = g_malloc0(sizeof(*info));
 186     MigrationState *s = migrate_get_current();
 187
 188     switch (s->state) {
 189     case MIG_STATE_NONE:
 190         /* no migration has happened ever */
 191         break;
 192     case MIG_STATE_SETUP:
 193         info->has_status = true;
 194         info->status = g_strdup("setup");
 195         info->has_total_time = false;
 196         break;
 197     case MIG_STATE_ACTIVE:
 198         info->has_status = true;
 199         info->status = g_strdup("active");
 200         info->has_total_time = true;
 201         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 202             - s->total_time;
 203         info->has_expected_downtime = true;
 204         info->expected_downtime = s->expected_downtime;
 205         info->has_setup_time = true;
 206         info->setup_time = s->setup_time;
 207
 208         info->has_ram = true;
 209         info->ram = g_malloc0(sizeof(*info->ram));
 210         info->ram->transferred = ram_bytes_transferred();
 211         info->ram->remaining = ram_bytes_remaining();
 212         info->ram->total = ram_bytes_total();
 213         info->ram->duplicate = dup_mig_pages_transferred();
 214         info->ram->skipped = skipped_mig_pages_transferred();
 215         info->ram->normal = norm_mig_pages_transferred();
 216         info->ram->normal_bytes = norm_mig_bytes_transferred();
 217         info->ram->dirty_pages_rate = s->dirty_pages_rate;
 218         info->ram->mbps = s->mbps;
 219
 220         if (blk_mig_active()) {
 221             info->has_disk = true;
 222             info->disk = g_malloc0(sizeof(*info->disk));
 223             info->disk->transferred = blk_mig_bytes_transferred();
 224             info->disk->remaining = blk_mig_bytes_remaining();
 225             info->disk->total = blk_mig_bytes_total();
 226         }
 227
 228         get_xbzrle_cache_stats(info);
 229         break;
 230     case MIG_STATE_COMPLETED:
 231         get_xbzrle_cache_stats(info);
 232
 233         info->has_status = true;
 234         info->status = g_strdup("completed");
 235         info->has_total_time = true;
 236         info->total_time = s->total_time;
 237         info->has_downtime = true;
 238         info->downtime = s->downtime;
 239         info->has_setup_time = true;
 240         info->setup_time = s->setup_time;
 241
 242         info->has_ram = true;
 243         info->ram = g_malloc0(sizeof(*info->ram));
 244         info->ram->transferred = ram_bytes_transferred();
 245         info->ram->remaining = 0;
 246         info->ram->total = ram_bytes_total();
 247         info->ram->duplicate = dup_mig_pages_transferred();
 248         info->ram->skipped = skipped_mig_pages_transferred();
 249         info->ram->normal = norm_mig_pages_transferred();
 250         info->ram->normal_bytes = norm_mig_bytes_transferred();
 251         info->ram->mbps = s->mbps;
 252         break;
 253     case MIG_STATE_ERROR:
 254         info->has_status = true;
 255         info->status = g_strdup("failed");
 256         break;
 257     case MIG_STATE_CANCELLED:
 258         info->has_status = true;
 259         info->status = g_strdup("cancelled");
 260         break;
 261     }
 262
 263     return info;
 264 }
 265
 266 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 267                                   Error **errp)
 268 {
 269     MigrationState *s = migrate_get_current();
 270     MigrationCapabilityStatusList *cap;
 271
 272     if (s->state == MIG_STATE_ACTIVE || s->state == MIG_STATE_SETUP) {
 273         error_set(errp, QERR_MIGRATION_ACTIVE);
 274         return;
 275     }
 276
 277     for (cap = params; cap; cap = cap->next) {
 278         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 279     }
 280 }
 281
 282 /* shared migration helpers */
 283
 284 static void migrate_fd_cleanup(void *opaque)
 285 {
 286     MigrationState *s = opaque;
 287
 288     qemu_bh_delete(s->cleanup_bh);
 289     s->cleanup_bh = NULL;
 290
 291     if (s->file) {
 292         DPRINTF("closing file\n");
 293         qemu_mutex_unlock_iothread();
 294         qemu_thread_join(&s->thread);
 295         qemu_mutex_lock_iothread();
 296
 297         qemu_fclose(s->file);
 298         s->file = NULL;
 299     }
 300
 301     assert(s->state != MIG_STATE_ACTIVE);
 302
 303     if (s->state != MIG_STATE_COMPLETED) {
 304         qemu_savevm_state_cancel();
 305     }
 306
 307     notifier_list_notify(&migration_state_notifiers, s);
 308 }
 309
 310 static void migrate_set_state(MigrationState *s, int old_state, int new_state)
 311 {
 312     if (atomic_cmpxchg(&s->state, old_state, new_state) == new_state) {
 313         trace_migrate_set_state(new_state);
 314     }
 315 }
 316
 317 void migrate_fd_error(MigrationState *s)
 318 {
 319     DPRINTF("setting error state\n");
 320     assert(s->file == NULL);
 321     s->state = MIG_STATE_ERROR;
 322     trace_migrate_set_state(MIG_STATE_ERROR);
 323     notifier_list_notify(&migration_state_notifiers, s);
 324 }
 325
 326 static void migrate_fd_cancel(MigrationState *s)
 327 {
 328     DPRINTF("cancelling migration\n");
 329
 330     migrate_set_state(s, s->state, MIG_STATE_CANCELLED);
 331 }
 332
 333 void add_migration_state_change_notifier(Notifier *notify)
 334 {
 335     notifier_list_add(&migration_state_notifiers, notify);
 336 }
 337
 338 void remove_migration_state_change_notifier(Notifier *notify)
 339 {
 340     notifier_remove(notify);
 341 }
 342
 343 bool migration_in_setup(MigrationState *s)
 344 {
 345     return s->state == MIG_STATE_SETUP;
 346 }
 347
 348 bool migration_has_finished(MigrationState *s)
 349 {
 350     return s->state == MIG_STATE_COMPLETED;
 351 }
 352
 353 bool migration_has_failed(MigrationState *s)
 354 {
 355     return (s->state == MIG_STATE_CANCELLED ||
 356             s->state == MIG_STATE_ERROR);
 357 }
 358
 359 static MigrationState *migrate_init(const MigrationParams *params)
 360 {
 361     MigrationState *s = migrate_get_current();
 362     int64_t bandwidth_limit = s->bandwidth_limit;
 363     bool enabled_capabilities[MIGRATION_CAPABILITY_MAX];
 364     int64_t xbzrle_cache_size = s->xbzrle_cache_size;
 365
 366     memcpy(enabled_capabilities, s->enabled_capabilities,
 367            sizeof(enabled_capabilities));
 368
 369     memset(s, 0, sizeof(*s));
 370     s->params = *params;
 371     memcpy(s->enabled_capabilities, enabled_capabilities,
 372            sizeof(enabled_capabilities));
 373     s->xbzrle_cache_size = xbzrle_cache_size;
 374
 375     s->bandwidth_limit = bandwidth_limit;
 376     s->state = MIG_STATE_SETUP;
 377     trace_migrate_set_state(MIG_STATE_SETUP);
 378
 379     s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 380     return s;
 381 }
 382
 383 static GSList *migration_blockers;
 384
 385 void migrate_add_blocker(Error *reason)
 386 {
 387     migration_blockers = g_slist_prepend(migration_blockers, reason);
 388 }
 389
 390 void migrate_del_blocker(Error *reason)
 391 {
 392     migration_blockers = g_slist_remove(migration_blockers, reason);
 393 }
 394
 395 void qmp_migrate(const char *uri, bool has_blk, bool blk,
 396                  bool has_inc, bool inc, bool has_detach, bool detach,
 397                  Error **errp)
 398 {
 399     Error *local_err = NULL;
 400     MigrationState *s = migrate_get_current();
 401     MigrationParams params;
 402     const char *p;
 403
 404     params.blk = has_blk && blk;
 405     params.shared = has_inc && inc;
 406
 407     if (s->state == MIG_STATE_ACTIVE || s->state == MIG_STATE_SETUP) {
 408         error_set(errp, QERR_MIGRATION_ACTIVE);
 409         return;
 410     }
 411
 412     if (qemu_savevm_state_blocked(errp)) {
 413         return;
 414     }
 415
 416     if (migration_blockers) {
 417         *errp = error_copy(migration_blockers->data);
 418         return;
 419     }
 420
 421     s = migrate_init(&params);
 422
 423     if (strstart(uri, "tcp:", &p)) {
 424         tcp_start_outgoing_migration(s, p, &local_err);
 425 #ifdef CONFIG_RDMA
 426     } else if (strstart(uri, "x-rdma:", &p)) {
 427         rdma_start_outgoing_migration(s, p, &local_err);
 428 #endif
 429 #if !defined(WIN32)
 430     } else if (strstart(uri, "exec:", &p)) {
 431         exec_start_outgoing_migration(s, p, &local_err);
 432     } else if (strstart(uri, "unix:", &p)) {
 433         unix_start_outgoing_migration(s, p, &local_err);
 434     } else if (strstart(uri, "fd:", &p)) {
 435         fd_start_outgoing_migration(s, p, &local_err);
 436 #endif
 437     } else {
 438         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "uri", "a valid migration protocol");
 439         return;
 440     }
 441
 442     if (local_err) {
 443         migrate_fd_error(s);
 444         error_propagate(errp, local_err);
 445         return;
 446     }
 447 }
 448
 449 void qmp_migrate_cancel(Error **errp)
 450 {
 451     migrate_fd_cancel(migrate_get_current());
 452 }
 453
 454 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
 455 {
 456     MigrationState *s = migrate_get_current();
 457
 458     /* Check for truncation */
 459     if (value != (size_t)value) {
 460         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
 461                   "exceeding address space");
 462         return;
 463     }
 464
 465     s->xbzrle_cache_size = xbzrle_cache_resize(value);
 466 }
 467
 468 int64_t qmp_query_migrate_cache_size(Error **errp)
 469 {
 470     return migrate_xbzrle_cache_size();
 471 }
 472
 473 void qmp_migrate_set_speed(int64_t value, Error **errp)
 474 {
 475     MigrationState *s;
 476
 477     if (value < 0) {
 478         value = 0;
 479     }
 480     if (value > SIZE_MAX) {
 481         value = SIZE_MAX;
 482     }
 483
 484     s = migrate_get_current();
 485     s->bandwidth_limit = value;
 486     if (s->file) {
 487         qemu_file_set_rate_limit(s->file, s->bandwidth_limit / XFER_LIMIT_RATIO);
 488     }
 489 }
 490
 491 void qmp_migrate_set_downtime(double value, Error **errp)
 492 {
 493     value *= 1e9;
 494     value = MAX(0, MIN(UINT64_MAX, value));
 495     max_downtime = (uint64_t)value;
 496 }
 497
 498 bool migrate_rdma_pin_all(void)
 499 {
 500     MigrationState *s;
 501
 502     s = migrate_get_current();
 503
 504     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_RDMA_PIN_ALL];
 505 }
 506
 507 bool migrate_auto_converge(void)
 508 {
 509     MigrationState *s;
 510
 511     s = migrate_get_current();
 512
 513     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
 514 }
 515
 516 bool migrate_zero_blocks(void)
 517 {
 518     MigrationState *s;
 519
 520     s = migrate_get_current();
 521
 522     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
 523 }
 524
 525 int migrate_use_xbzrle(void)
 526 {
 527     MigrationState *s;
 528
 529     s = migrate_get_current();
 530
 531     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
 532 }
 533
 534 int64_t migrate_xbzrle_cache_size(void)
 535 {
 536     MigrationState *s;
 537
 538     s = migrate_get_current();
 539
 540     return s->xbzrle_cache_size;
 541 }
 542
 543 /* migration thread support */
 544
 545 static void *migration_thread(void *opaque)
 546 {
 547     MigrationState *s = opaque;
 548     int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 549     int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
 550     int64_t initial_bytes = 0;
 551     int64_t max_size = 0;
 552     int64_t start_time = initial_time;
 553     bool old_vm_running = false;
 554
 555     DPRINTF("beginning savevm\n");
 556     qemu_savevm_state_begin(s->file, &s->params);
 557
 558     s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
 559     migrate_set_state(s, MIG_STATE_SETUP, MIG_STATE_ACTIVE);
 560
 561     DPRINTF("setup complete\n");
 562
 563     while (s->state == MIG_STATE_ACTIVE) {
 564         int64_t current_time;
 565         uint64_t pending_size;
 566
 567         if (!qemu_file_rate_limit(s->file)) {
 568             DPRINTF("iterate\n");
 569             pending_size = qemu_savevm_state_pending(s->file, max_size);
 570             DPRINTF("pending size %lu max %lu\n", pending_size, max_size);
 571             if (pending_size && pending_size >= max_size) {
 572                 qemu_savevm_state_iterate(s->file);
 573             } else {
 574                 int ret;
 575
 576                 DPRINTF("done iterating\n");
 577                 qemu_mutex_lock_iothread();
 578                 start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 579                 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
 580                 old_vm_running = runstate_is_running();
 581
 582                 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
 583                 if (ret >= 0) {
 584                     qemu_file_set_rate_limit(s->file, INT_MAX);
 585                     qemu_savevm_state_complete(s->file);
 586                 }
 587                 qemu_mutex_unlock_iothread();
 588
 589                 if (ret < 0) {
 590                     migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_ERROR);
 591                     break;
 592                 }
 593
 594                 if (!qemu_file_get_error(s->file)) {
 595                     migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_COMPLETED);
 596                     break;
 597                 }
 598             }
 599         }
 600
 601         if (qemu_file_get_error(s->file)) {
 602             migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_ERROR);
 603             break;
 604         }
 605         current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 606         if (current_time >= initial_time + BUFFER_DELAY) {
 607             uint64_t transferred_bytes = qemu_ftell(s->file) - initial_bytes;
 608             uint64_t time_spent = current_time - initial_time;
 609             double bandwidth = transferred_bytes / time_spent;
 610             max_size = bandwidth * migrate_max_downtime() / 1000000;
 611
 612             s->mbps = time_spent ? (((double) transferred_bytes * 8.0) /
 613                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0 : -1;
 614
 615             DPRINTF("transferred %" PRIu64 " time_spent %" PRIu64
 616                     " bandwidth %g max_size %" PRId64 "\n",
 617                     transferred_bytes, time_spent, bandwidth, max_size);
 618             /* if we haven't sent anything, we don't want to recalculate
 619                10000 is a small enough number for our purposes */
 620             if (s->dirty_bytes_rate && transferred_bytes > 10000) {
 621                 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
 622             }
 623
 624             qemu_file_reset_rate_limit(s->file);
 625             initial_time = current_time;
 626             initial_bytes = qemu_ftell(s->file);
 627         }
 628         if (qemu_file_rate_limit(s->file)) {
 629             /* usleep expects microseconds */
 630             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
 631         }
 632     }
 633
 634     qemu_mutex_lock_iothread();
 635     if (s->state == MIG_STATE_COMPLETED) {
 636         int64_t end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 637         s->total_time = end_time - s->total_time;
 638         s->downtime = end_time - start_time;
 639         runstate_set(RUN_STATE_POSTMIGRATE);
 640     } else {
 641         if (old_vm_running) {
 642             vm_start();
 643         }
 644     }
 645     qemu_bh_schedule(s->cleanup_bh);
 646     qemu_mutex_unlock_iothread();
 647
 648     return NULL;
 649 }
 650
 651 void migrate_fd_connect(MigrationState *s)
 652 {
 653     s->state = MIG_STATE_SETUP;
 654     trace_migrate_set_state(MIG_STATE_SETUP);
 655
 656     /* This is a best 1st approximation. ns to ms */
 657     s->expected_downtime = max_downtime/1000000;
 658     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
 659
 660     qemu_file_set_rate_limit(s->file,
 661                              s->bandwidth_limit / XFER_LIMIT_RATIO);
 662
 663     /* Notify before starting migration thread */
 664     notifier_list_notify(&migration_state_notifiers, s);
 665
 666     qemu_thread_create(&s->thread, migration_thread, s,
 667                        QEMU_THREAD_JOINABLE);
 668 }