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->has_total_time = true;
 235         info->total_time = s->total_time;
 236         info->has_downtime = true;
 237         info->downtime = s->downtime;
 238         info->has_setup_time = true;
 239         info->setup_time = s->setup_time;
 240
 241         info->has_ram = true;
 242         info->ram = g_malloc0(sizeof(*info->ram));
 243         info->ram->transferred = ram_bytes_transferred();
 244         info->ram->remaining = 0;
 245         info->ram->total = ram_bytes_total();
 246         info->ram->duplicate = dup_mig_pages_transferred();
 247         info->ram->skipped = skipped_mig_pages_transferred();
 248         info->ram->normal = norm_mig_pages_transferred();
 249         info->ram->normal_bytes = norm_mig_bytes_transferred();
 250         info->ram->mbps = s->mbps;
 251         break;
 252     case MIG_STATE_ERROR:
 253         info->has_status = true;
 254         info->status = g_strdup("failed");
 255         break;
 256     case MIG_STATE_CANCELLED:
 257         info->has_status = true;
 258         info->status = g_strdup("cancelled");
 259         break;
 260     }
 261
 262     return info;
 263 }
 264
 265 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 266                                   Error **errp)
 267 {
 268     MigrationState *s = migrate_get_current();
 269     MigrationCapabilityStatusList *cap;
 270
 271     if (s->state == MIG_STATE_ACTIVE || s->state == MIG_STATE_SETUP) {
 272         error_set(errp, QERR_MIGRATION_ACTIVE);
 273         return;
 274     }
 275
 276     for (cap = params; cap; cap = cap->next) {
 277         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 278     }
 279 }
 280
 281 /* shared migration helpers */
 282
 283 static void migrate_fd_cleanup(void *opaque)
 284 {
 285     MigrationState *s = opaque;
 286
 287     qemu_bh_delete(s->cleanup_bh);
 288     s->cleanup_bh = NULL;
 289
 290     if (s->file) {
 291         DPRINTF("closing file\n");
 292         qemu_mutex_unlock_iothread();
 293         qemu_thread_join(&s->thread);
 294         qemu_mutex_lock_iothread();
 295
 296         qemu_fclose(s->file);
 297         s->file = NULL;
 298     }
 299
 300     assert(s->state != MIG_STATE_ACTIVE);
 301
 302     if (s->state != MIG_STATE_COMPLETED) {
 303         qemu_savevm_state_cancel();
 304     }
 305
 306     notifier_list_notify(&migration_state_notifiers, s);
 307 }
 308
 309 static void migrate_set_state(MigrationState *s, int old_state, int new_state)
 310 {
 311     if (atomic_cmpxchg(&s->state, old_state, new_state) == new_state) {
 312         trace_migrate_set_state(new_state);
 313     }
 314 }
 315
 316 void migrate_fd_error(MigrationState *s)
 317 {
 318     DPRINTF("setting error state\n");
 319     assert(s->file == NULL);
 320     s->state = MIG_STATE_ERROR;
 321     trace_migrate_set_state(MIG_STATE_ERROR);
 322     notifier_list_notify(&migration_state_notifiers, s);
 323 }
 324
 325 static void migrate_fd_cancel(MigrationState *s)
 326 {
 327     DPRINTF("cancelling migration\n");
 328
 329     migrate_set_state(s, s->state, MIG_STATE_CANCELLED);
 330 }
 331
 332 void add_migration_state_change_notifier(Notifier *notify)
 333 {
 334     notifier_list_add(&migration_state_notifiers, notify);
 335 }
 336
 337 void remove_migration_state_change_notifier(Notifier *notify)
 338 {
 339     notifier_remove(notify);
 340 }
 341
 342 bool migration_in_setup(MigrationState *s)
 343 {
 344     return s->state == MIG_STATE_SETUP;
 345 }
 346
 347 bool migration_has_finished(MigrationState *s)
 348 {
 349     return s->state == MIG_STATE_COMPLETED;
 350 }
 351
 352 bool migration_has_failed(MigrationState *s)
 353 {
 354     return (s->state == MIG_STATE_CANCELLED ||
 355             s->state == MIG_STATE_ERROR);
 356 }
 357
 358 static MigrationState *migrate_init(const MigrationParams *params)
 359 {
 360     MigrationState *s = migrate_get_current();
 361     int64_t bandwidth_limit = s->bandwidth_limit;
 362     bool enabled_capabilities[MIGRATION_CAPABILITY_MAX];
 363     int64_t xbzrle_cache_size = s->xbzrle_cache_size;
 364
 365     memcpy(enabled_capabilities, s->enabled_capabilities,
 366            sizeof(enabled_capabilities));
 367
 368     memset(s, 0, sizeof(*s));
 369     s->params = *params;
 370     memcpy(s->enabled_capabilities, enabled_capabilities,
 371            sizeof(enabled_capabilities));
 372     s->xbzrle_cache_size = xbzrle_cache_size;
 373
 374     s->bandwidth_limit = bandwidth_limit;
 375     s->state = MIG_STATE_SETUP;
 376     trace_migrate_set_state(MIG_STATE_SETUP);
 377
 378     s->total_time = qemu_get_clock_ms(rt_clock);
 379     return s;
 380 }
 381
 382 static GSList *migration_blockers;
 383
 384 void migrate_add_blocker(Error *reason)
 385 {
 386     migration_blockers = g_slist_prepend(migration_blockers, reason);
 387 }
 388
 389 void migrate_del_blocker(Error *reason)
 390 {
 391     migration_blockers = g_slist_remove(migration_blockers, reason);
 392 }
 393
 394 void qmp_migrate(const char *uri, bool has_blk, bool blk,
 395                  bool has_inc, bool inc, bool has_detach, bool detach,
 396                  Error **errp)
 397 {
 398     Error *local_err = NULL;
 399     MigrationState *s = migrate_get_current();
 400     MigrationParams params;
 401     const char *p;
 402
 403     params.blk = has_blk && blk;
 404     params.shared = has_inc && inc;
 405
 406     if (s->state == MIG_STATE_ACTIVE || s->state == MIG_STATE_SETUP) {
 407         error_set(errp, QERR_MIGRATION_ACTIVE);
 408         return;
 409     }
 410
 411     if (qemu_savevm_state_blocked(errp)) {
 412         return;
 413     }
 414
 415     if (migration_blockers) {
 416         *errp = error_copy(migration_blockers->data);
 417         return;
 418     }
 419
 420     s = migrate_init(&params);
 421
 422     if (strstart(uri, "tcp:", &p)) {
 423         tcp_start_outgoing_migration(s, p, &local_err);
 424 #ifdef CONFIG_RDMA
 425     } else if (strstart(uri, "x-rdma:", &p)) {
 426         rdma_start_outgoing_migration(s, p, &local_err);
 427 #endif
 428 #if !defined(WIN32)
 429     } else if (strstart(uri, "exec:", &p)) {
 430         exec_start_outgoing_migration(s, p, &local_err);
 431     } else if (strstart(uri, "unix:", &p)) {
 432         unix_start_outgoing_migration(s, p, &local_err);
 433     } else if (strstart(uri, "fd:", &p)) {
 434         fd_start_outgoing_migration(s, p, &local_err);
 435 #endif
 436     } else {
 437         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "uri", "a valid migration protocol");
 438         return;
 439     }
 440
 441     if (local_err) {
 442         migrate_fd_error(s);
 443         error_propagate(errp, local_err);
 444         return;
 445     }
 446 }
 447
 448 void qmp_migrate_cancel(Error **errp)
 449 {
 450     migrate_fd_cancel(migrate_get_current());
 451 }
 452
 453 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
 454 {
 455     MigrationState *s = migrate_get_current();
 456
 457     /* Check for truncation */
 458     if (value != (size_t)value) {
 459         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
 460                   "exceeding address space");
 461         return;
 462     }
 463
 464     s->xbzrle_cache_size = xbzrle_cache_resize(value);
 465 }
 466
 467 int64_t qmp_query_migrate_cache_size(Error **errp)
 468 {
 469     return migrate_xbzrle_cache_size();
 470 }
 471
 472 void qmp_migrate_set_speed(int64_t value, Error **errp)
 473 {
 474     MigrationState *s;
 475
 476     if (value < 0) {
 477         value = 0;
 478     }
 479     if (value > SIZE_MAX) {
 480         value = SIZE_MAX;
 481     }
 482
 483     s = migrate_get_current();
 484     s->bandwidth_limit = value;
 485     if (s->file) {
 486         qemu_file_set_rate_limit(s->file, s->bandwidth_limit / XFER_LIMIT_RATIO);
 487     }
 488 }
 489
 490 void qmp_migrate_set_downtime(double value, Error **errp)
 491 {
 492     value *= 1e9;
 493     value = MAX(0, MIN(UINT64_MAX, value));
 494     max_downtime = (uint64_t)value;
 495 }
 496
 497 bool migrate_rdma_pin_all(void)
 498 {
 499     MigrationState *s;
 500
 501     s = migrate_get_current();
 502
 503     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_RDMA_PIN_ALL];
 504 }
 505
 506 bool migrate_auto_converge(void)
 507 {
 508     MigrationState *s;
 509
 510     s = migrate_get_current();
 511
 512     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
 513 }
 514
 515 bool migrate_zero_blocks(void)
 516 {
 517     MigrationState *s;
 518
 519     s = migrate_get_current();
 520
 521     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
 522 }
 523
 524 int migrate_use_xbzrle(void)
 525 {
 526     MigrationState *s;
 527
 528     s = migrate_get_current();
 529
 530     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
 531 }
 532
 533 int64_t migrate_xbzrle_cache_size(void)
 534 {
 535     MigrationState *s;
 536
 537     s = migrate_get_current();
 538
 539     return s->xbzrle_cache_size;
 540 }
 541
 542 /* migration thread support */
 543
 544 static void *migration_thread(void *opaque)
 545 {
 546     MigrationState *s = opaque;
 547     int64_t initial_time = qemu_get_clock_ms(rt_clock);
 548     int64_t setup_start = qemu_get_clock_ms(host_clock);
 549     int64_t initial_bytes = 0;
 550     int64_t max_size = 0;
 551     int64_t start_time = initial_time;
 552     bool old_vm_running = false;
 553
 554     DPRINTF("beginning savevm\n");
 555     qemu_savevm_state_begin(s->file, &s->params);
 556
 557     s->setup_time = qemu_get_clock_ms(host_clock) - setup_start;
 558     migrate_set_state(s, MIG_STATE_SETUP, MIG_STATE_ACTIVE);
 559
 560     DPRINTF("setup complete\n");
 561
 562     while (s->state == MIG_STATE_ACTIVE) {
 563         int64_t current_time;
 564         uint64_t pending_size;
 565
 566         if (!qemu_file_rate_limit(s->file)) {
 567             DPRINTF("iterate\n");
 568             pending_size = qemu_savevm_state_pending(s->file, max_size);
 569             DPRINTF("pending size %lu max %lu\n", pending_size, max_size);
 570             if (pending_size && pending_size >= max_size) {
 571                 qemu_savevm_state_iterate(s->file);
 572             } else {
 573                 int ret;
 574
 575                 DPRINTF("done iterating\n");
 576                 qemu_mutex_lock_iothread();
 577                 start_time = qemu_get_clock_ms(rt_clock);
 578                 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
 579                 old_vm_running = runstate_is_running();
 580
 581                 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
 582                 if (ret >= 0) {
 583                     qemu_file_set_rate_limit(s->file, INT_MAX);
 584                     qemu_savevm_state_complete(s->file);
 585                 }
 586                 qemu_mutex_unlock_iothread();
 587
 588                 if (ret < 0) {
 589                     migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_ERROR);
 590                     break;
 591                 }
 592
 593                 if (!qemu_file_get_error(s->file)) {
 594                     migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_COMPLETED);
 595                     break;
 596                 }
 597             }
 598         }
 599
 600         if (qemu_file_get_error(s->file)) {
 601             migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_ERROR);
 602             break;
 603         }
 604         current_time = qemu_get_clock_ms(rt_clock);
 605         if (current_time >= initial_time + BUFFER_DELAY) {
 606             uint64_t transferred_bytes = qemu_ftell(s->file) - initial_bytes;
 607             uint64_t time_spent = current_time - initial_time;
 608             double bandwidth = transferred_bytes / time_spent;
 609             max_size = bandwidth * migrate_max_downtime() / 1000000;
 610
 611             s->mbps = time_spent ? (((double) transferred_bytes * 8.0) /
 612                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0 : -1;
 613
 614             DPRINTF("transferred %" PRIu64 " time_spent %" PRIu64
 615                     " bandwidth %g max_size %" PRId64 "\n",
 616                     transferred_bytes, time_spent, bandwidth, max_size);
 617             /* if we haven't sent anything, we don't want to recalculate
 618                10000 is a small enough number for our purposes */
 619             if (s->dirty_bytes_rate && transferred_bytes > 10000) {
 620                 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
 621             }
 622
 623             qemu_file_reset_rate_limit(s->file);
 624             initial_time = current_time;
 625             initial_bytes = qemu_ftell(s->file);
 626         }
 627         if (qemu_file_rate_limit(s->file)) {
 628             /* usleep expects microseconds */
 629             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
 630         }
 631     }
 632
 633     qemu_mutex_lock_iothread();
 634     if (s->state == MIG_STATE_COMPLETED) {
 635         int64_t end_time = qemu_get_clock_ms(rt_clock);
 636         s->total_time = end_time - s->total_time;
 637         s->downtime = end_time - start_time;
 638         runstate_set(RUN_STATE_POSTMIGRATE);
 639     } else {
 640         if (old_vm_running) {
 641             vm_start();
 642         }
 643     }
 644     qemu_bh_schedule(s->cleanup_bh);
 645     qemu_mutex_unlock_iothread();
 646
 647     return NULL;
 648 }
 649
 650 void migrate_fd_connect(MigrationState *s)
 651 {
 652     s->state = MIG_STATE_SETUP;
 653     trace_migrate_set_state(MIG_STATE_SETUP);
 654
 655     /* This is a best 1st approximation. ns to ms */
 656     s->expected_downtime = max_downtime/1000000;
 657     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
 658
 659     qemu_file_set_rate_limit(s->file,
 660                              s->bandwidth_limit / XFER_LIMIT_RATIO);
 661
 662     /* Notify before starting migration thread */
 663     notifier_list_notify(&migration_state_notifiers, s);
 664
 665     qemu_thread_create(&s->thread, migration_thread, s,
 666                        QEMU_THREAD_JOINABLE);
 667 }