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 #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 (atomic_cmpxchg(&s->state, MIG_STATE_ACTIVE, new_state) == new_state) {
 297         trace_migrate_set_state(new_state);
 298     }
 299 }
 300
 301 void migrate_fd_error(MigrationState *s)
 302 {
 303     DPRINTF("setting error state\n");
 304     assert(s->file == NULL);
 305     s->state = MIG_STATE_ERROR;
 306     trace_migrate_set_state(MIG_STATE_ERROR);
 307     notifier_list_notify(&migration_state_notifiers, s);
 308 }
 309
 310 static void migrate_fd_cancel(MigrationState *s)
 311 {
 312     DPRINTF("cancelling migration\n");
 313
 314     migrate_finish_set_state(s, MIG_STATE_CANCELLED);
 315 }
 316
 317 void add_migration_state_change_notifier(Notifier *notify)
 318 {
 319     notifier_list_add(&migration_state_notifiers, notify);
 320 }
 321
 322 void remove_migration_state_change_notifier(Notifier *notify)
 323 {
 324     notifier_remove(notify);
 325 }
 326
 327 bool migration_is_active(MigrationState *s)
 328 {
 329     return s->state == MIG_STATE_ACTIVE;
 330 }
 331
 332 bool migration_has_finished(MigrationState *s)
 333 {
 334     return s->state == MIG_STATE_COMPLETED;
 335 }
 336
 337 bool migration_has_failed(MigrationState *s)
 338 {
 339     return (s->state == MIG_STATE_CANCELLED ||
 340             s->state == MIG_STATE_ERROR);
 341 }
 342
 343 static MigrationState *migrate_init(const MigrationParams *params)
 344 {
 345     MigrationState *s = migrate_get_current();
 346     int64_t bandwidth_limit = s->bandwidth_limit;
 347     bool enabled_capabilities[MIGRATION_CAPABILITY_MAX];
 348     int64_t xbzrle_cache_size = s->xbzrle_cache_size;
 349
 350     memcpy(enabled_capabilities, s->enabled_capabilities,
 351            sizeof(enabled_capabilities));
 352
 353     memset(s, 0, sizeof(*s));
 354     s->params = *params;
 355     memcpy(s->enabled_capabilities, enabled_capabilities,
 356            sizeof(enabled_capabilities));
 357     s->xbzrle_cache_size = xbzrle_cache_size;
 358
 359     s->bandwidth_limit = bandwidth_limit;
 360     s->state = MIG_STATE_SETUP;
 361     trace_migrate_set_state(MIG_STATE_SETUP);
 362
 363     s->total_time = qemu_get_clock_ms(rt_clock);
 364     return s;
 365 }
 366
 367 static GSList *migration_blockers;
 368
 369 void migrate_add_blocker(Error *reason)
 370 {
 371     migration_blockers = g_slist_prepend(migration_blockers, reason);
 372 }
 373
 374 void migrate_del_blocker(Error *reason)
 375 {
 376     migration_blockers = g_slist_remove(migration_blockers, reason);
 377 }
 378
 379 void qmp_migrate(const char *uri, bool has_blk, bool blk,
 380                  bool has_inc, bool inc, bool has_detach, bool detach,
 381                  Error **errp)
 382 {
 383     Error *local_err = NULL;
 384     MigrationState *s = migrate_get_current();
 385     MigrationParams params;
 386     const char *p;
 387
 388     params.blk = blk;
 389     params.shared = inc;
 390
 391     if (s->state == MIG_STATE_ACTIVE) {
 392         error_set(errp, QERR_MIGRATION_ACTIVE);
 393         return;
 394     }
 395
 396     if (qemu_savevm_state_blocked(errp)) {
 397         return;
 398     }
 399
 400     if (migration_blockers) {
 401         *errp = error_copy(migration_blockers->data);
 402         return;
 403     }
 404
 405     s = migrate_init(&params);
 406
 407     if (strstart(uri, "tcp:", &p)) {
 408         tcp_start_outgoing_migration(s, p, &local_err);
 409 #if !defined(WIN32)
 410     } else if (strstart(uri, "exec:", &p)) {
 411         exec_start_outgoing_migration(s, p, &local_err);
 412     } else if (strstart(uri, "unix:", &p)) {
 413         unix_start_outgoing_migration(s, p, &local_err);
 414     } else if (strstart(uri, "fd:", &p)) {
 415         fd_start_outgoing_migration(s, p, &local_err);
 416 #endif
 417     } else {
 418         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "uri", "a valid migration protocol");
 419         return;
 420     }
 421
 422     if (local_err) {
 423         migrate_fd_error(s);
 424         error_propagate(errp, local_err);
 425         return;
 426     }
 427 }
 428
 429 void qmp_migrate_cancel(Error **errp)
 430 {
 431     migrate_fd_cancel(migrate_get_current());
 432 }
 433
 434 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
 435 {
 436     MigrationState *s = migrate_get_current();
 437
 438     /* Check for truncation */
 439     if (value != (size_t)value) {
 440         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
 441                   "exceeding address space");
 442         return;
 443     }
 444
 445     s->xbzrle_cache_size = xbzrle_cache_resize(value);
 446 }
 447
 448 int64_t qmp_query_migrate_cache_size(Error **errp)
 449 {
 450     return migrate_xbzrle_cache_size();
 451 }
 452
 453 void qmp_migrate_set_speed(int64_t value, Error **errp)
 454 {
 455     MigrationState *s;
 456
 457     if (value < 0) {
 458         value = 0;
 459     }
 460     if (value > SIZE_MAX) {
 461         value = SIZE_MAX;
 462     }
 463
 464     s = migrate_get_current();
 465     s->bandwidth_limit = value;
 466     if (s->file) {
 467         qemu_file_set_rate_limit(s->file, s->bandwidth_limit / XFER_LIMIT_RATIO);
 468     }
 469 }
 470
 471 void qmp_migrate_set_downtime(double value, Error **errp)
 472 {
 473     value *= 1e9;
 474     value = MAX(0, MIN(UINT64_MAX, value));
 475     max_downtime = (uint64_t)value;
 476 }
 477
 478 bool migrate_rdma_pin_all(void)
 479 {
 480     MigrationState *s;
 481
 482     s = migrate_get_current();
 483
 484     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_RDMA_PIN_ALL];
 485 }
 486
 487 int migrate_use_xbzrle(void)
 488 {
 489     MigrationState *s;
 490
 491     s = migrate_get_current();
 492
 493     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
 494 }
 495
 496 int64_t migrate_xbzrle_cache_size(void)
 497 {
 498     MigrationState *s;
 499
 500     s = migrate_get_current();
 501
 502     return s->xbzrle_cache_size;
 503 }
 504
 505 /* migration thread support */
 506
 507 static void *migration_thread(void *opaque)
 508 {
 509     MigrationState *s = opaque;
 510     int64_t initial_time = qemu_get_clock_ms(rt_clock);
 511     int64_t initial_bytes = 0;
 512     int64_t max_size = 0;
 513     int64_t start_time = initial_time;
 514     bool old_vm_running = false;
 515
 516     DPRINTF("beginning savevm\n");
 517     qemu_savevm_state_begin(s->file, &s->params);
 518
 519     while (s->state == MIG_STATE_ACTIVE) {
 520         int64_t current_time;
 521         uint64_t pending_size;
 522
 523         if (!qemu_file_rate_limit(s->file)) {
 524             DPRINTF("iterate\n");
 525             pending_size = qemu_savevm_state_pending(s->file, max_size);
 526             DPRINTF("pending size %lu max %lu\n", pending_size, max_size);
 527             if (pending_size && pending_size >= max_size) {
 528                 qemu_savevm_state_iterate(s->file);
 529             } else {
 530                 DPRINTF("done iterating\n");
 531                 qemu_mutex_lock_iothread();
 532                 start_time = qemu_get_clock_ms(rt_clock);
 533                 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
 534                 old_vm_running = runstate_is_running();
 535                 vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
 536                 qemu_file_set_rate_limit(s->file, INT_MAX);
 537                 qemu_savevm_state_complete(s->file);
 538                 qemu_mutex_unlock_iothread();
 539                 if (!qemu_file_get_error(s->file)) {
 540                     migrate_finish_set_state(s, MIG_STATE_COMPLETED);
 541                     break;
 542                 }
 543             }
 544         }
 545
 546         if (qemu_file_get_error(s->file)) {
 547             migrate_finish_set_state(s, MIG_STATE_ERROR);
 548             break;
 549         }
 550         current_time = qemu_get_clock_ms(rt_clock);
 551         if (current_time >= initial_time + BUFFER_DELAY) {
 552             uint64_t transferred_bytes = qemu_ftell(s->file) - initial_bytes;
 553             uint64_t time_spent = current_time - initial_time;
 554             double bandwidth = transferred_bytes / time_spent;
 555             max_size = bandwidth * migrate_max_downtime() / 1000000;
 556
 557             s->mbps = time_spent ? (((double) transferred_bytes * 8.0) /
 558                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0 : -1;
 559
 560             DPRINTF("transferred %" PRIu64 " time_spent %" PRIu64
 561                     " bandwidth %g max_size %" PRId64 "\n",
 562                     transferred_bytes, time_spent, bandwidth, max_size);
 563             /* if we haven't sent anything, we don't want to recalculate
 564                10000 is a small enough number for our purposes */
 565             if (s->dirty_bytes_rate && transferred_bytes > 10000) {
 566                 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
 567             }
 568
 569             qemu_file_reset_rate_limit(s->file);
 570             initial_time = current_time;
 571             initial_bytes = qemu_ftell(s->file);
 572         }
 573         if (qemu_file_rate_limit(s->file)) {
 574             /* usleep expects microseconds */
 575             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
 576         }
 577     }
 578
 579     qemu_mutex_lock_iothread();
 580     if (s->state == MIG_STATE_COMPLETED) {
 581         int64_t end_time = qemu_get_clock_ms(rt_clock);
 582         s->total_time = end_time - s->total_time;
 583         s->downtime = end_time - start_time;
 584         runstate_set(RUN_STATE_POSTMIGRATE);
 585     } else {
 586         if (old_vm_running) {
 587             vm_start();
 588         }
 589     }
 590     qemu_bh_schedule(s->cleanup_bh);
 591     qemu_mutex_unlock_iothread();
 592
 593     return NULL;
 594 }
 595
 596 void migrate_fd_connect(MigrationState *s)
 597 {
 598     s->state = MIG_STATE_ACTIVE;
 599     trace_migrate_set_state(MIG_STATE_ACTIVE);
 600
 601     /* This is a best 1st approximation. ns to ms */
 602     s->expected_downtime = max_downtime/1000000;
 603     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
 604
 605     qemu_file_set_rate_limit(s->file,
 606                              s->bandwidth_limit / XFER_LIMIT_RATIO);
 607
 608     qemu_thread_create(&s->thread, migration_thread, s,
 609                        QEMU_THREAD_JOINABLE);
 610     notifier_list_notify(&migration_state_notifiers, s);
 611 }