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