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