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