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