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(EXIT_FAILURE);
 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     qemu_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->skipped = skipped_mig_pages_transferred();
 201         info->ram->normal = norm_mig_pages_transferred();
 202         info->ram->normal_bytes = norm_mig_bytes_transferred();
 203         info->ram->dirty_pages_rate = s->dirty_pages_rate;
 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->skipped = skipped_mig_pages_transferred();
 231         info->ram->normal = norm_mig_pages_transferred();
 232         info->ram->normal_bytes = norm_mig_bytes_transferred();
 233         break;
 234     case MIG_STATE_ERROR:
 235         info->has_status = true;
 236         info->status = g_strdup("failed");
 237         break;
 238     case MIG_STATE_CANCELLED:
 239         info->has_status = true;
 240         info->status = g_strdup("cancelled");
 241         break;
 242     }
 243
 244     return info;
 245 }
 246
 247 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 248                                   Error **errp)
 249 {
 250     MigrationState *s = migrate_get_current();
 251     MigrationCapabilityStatusList *cap;
 252
 253     if (s->state == MIG_STATE_ACTIVE) {
 254         error_set(errp, QERR_MIGRATION_ACTIVE);
 255         return;
 256     }
 257
 258     for (cap = params; cap; cap = cap->next) {
 259         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 260     }
 261 }
 262
 263 /* shared migration helpers */
 264
 265 static void migrate_fd_cleanup(void *opaque)
 266 {
 267     MigrationState *s = opaque;
 268
 269     qemu_bh_delete(s->cleanup_bh);
 270     s->cleanup_bh = NULL;
 271
 272     if (s->file) {
 273         DPRINTF("closing file\n");
 274         qemu_mutex_unlock_iothread();
 275         qemu_thread_join(&s->thread);
 276         qemu_mutex_lock_iothread();
 277
 278         qemu_fclose(s->file);
 279         s->file = NULL;
 280     }
 281
 282     assert(s->state != MIG_STATE_ACTIVE);
 283
 284     if (s->state != MIG_STATE_COMPLETED) {
 285         qemu_savevm_state_cancel();
 286     }
 287
 288     notifier_list_notify(&migration_state_notifiers, s);
 289 }
 290
 291 static void migrate_finish_set_state(MigrationState *s, int new_state)
 292 {
 293     if (__sync_val_compare_and_swap(&s->state, MIG_STATE_ACTIVE,
 294                                     new_state) == new_state) {
 295         trace_migrate_set_state(new_state);
 296     }
 297 }
 298
 299 void migrate_fd_error(MigrationState *s)
 300 {
 301     DPRINTF("setting error state\n");
 302     assert(s->file == NULL);
 303     s->state = MIG_STATE_ERROR;
 304     trace_migrate_set_state(MIG_STATE_ERROR);
 305     notifier_list_notify(&migration_state_notifiers, s);
 306 }
 307
 308 static void migrate_fd_cancel(MigrationState *s)
 309 {
 310     DPRINTF("cancelling migration\n");
 311
 312     migrate_finish_set_state(s, MIG_STATE_CANCELLED);
 313 }
 314
 315 void add_migration_state_change_notifier(Notifier *notify)
 316 {
 317     notifier_list_add(&migration_state_notifiers, notify);
 318 }
 319
 320 void remove_migration_state_change_notifier(Notifier *notify)
 321 {
 322     notifier_remove(notify);
 323 }
 324
 325 bool migration_is_active(MigrationState *s)
 326 {
 327     return s->state == MIG_STATE_ACTIVE;
 328 }
 329
 330 bool migration_has_finished(MigrationState *s)
 331 {
 332     return s->state == MIG_STATE_COMPLETED;
 333 }
 334
 335 bool migration_has_failed(MigrationState *s)
 336 {
 337     return (s->state == MIG_STATE_CANCELLED ||
 338             s->state == MIG_STATE_ERROR);
 339 }
 340
 341 static MigrationState *migrate_init(const MigrationParams *params)
 342 {
 343     MigrationState *s = migrate_get_current();
 344     int64_t bandwidth_limit = s->bandwidth_limit;
 345     bool enabled_capabilities[MIGRATION_CAPABILITY_MAX];
 346     int64_t xbzrle_cache_size = s->xbzrle_cache_size;
 347
 348     memcpy(enabled_capabilities, s->enabled_capabilities,
 349            sizeof(enabled_capabilities));
 350
 351     memset(s, 0, sizeof(*s));
 352     s->bandwidth_limit = bandwidth_limit;
 353     s->params = *params;
 354     memcpy(s->enabled_capabilities, enabled_capabilities,
 355            sizeof(enabled_capabilities));
 356     s->xbzrle_cache_size = xbzrle_cache_size;
 357
 358     s->bandwidth_limit = bandwidth_limit;
 359     s->state = MIG_STATE_SETUP;
 360     trace_migrate_set_state(MIG_STATE_SETUP);
 361
 362     s->total_time = qemu_get_clock_ms(rt_clock);
 363     return s;
 364 }
 365
 366 static GSList *migration_blockers;
 367
 368 void migrate_add_blocker(Error *reason)
 369 {
 370     migration_blockers = g_slist_prepend(migration_blockers, reason);
 371 }
 372
 373 void migrate_del_blocker(Error *reason)
 374 {
 375     migration_blockers = g_slist_remove(migration_blockers, reason);
 376 }
 377
 378 void qmp_migrate(const char *uri, bool has_blk, bool blk,
 379                  bool has_inc, bool inc, bool has_detach, bool detach,
 380                  Error **errp)
 381 {
 382     Error *local_err = NULL;
 383     MigrationState *s = migrate_get_current();
 384     MigrationParams params;
 385     const char *p;
 386
 387     params.blk = blk;
 388     params.shared = inc;
 389
 390     if (s->state == MIG_STATE_ACTIVE) {
 391         error_set(errp, QERR_MIGRATION_ACTIVE);
 392         return;
 393     }
 394
 395     if (qemu_savevm_state_blocked(errp)) {
 396         return;
 397     }
 398
 399     if (migration_blockers) {
 400         *errp = error_copy(migration_blockers->data);
 401         return;
 402     }
 403
 404     s = migrate_init(&params);
 405
 406     if (strstart(uri, "tcp:", &p)) {
 407         tcp_start_outgoing_migration(s, p, &local_err);
 408 #if !defined(WIN32)
 409     } else if (strstart(uri, "exec:", &p)) {
 410         exec_start_outgoing_migration(s, p, &local_err);
 411     } else if (strstart(uri, "unix:", &p)) {
 412         unix_start_outgoing_migration(s, p, &local_err);
 413     } else if (strstart(uri, "fd:", &p)) {
 414         fd_start_outgoing_migration(s, p, &local_err);
 415 #endif
 416     } else {
 417         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "uri", "a valid migration protocol");
 418         return;
 419     }
 420
 421     if (local_err) {
 422         migrate_fd_error(s);
 423         error_propagate(errp, local_err);
 424         return;
 425     }
 426 }
 427
 428 void qmp_migrate_cancel(Error **errp)
 429 {
 430     migrate_fd_cancel(migrate_get_current());
 431 }
 432
 433 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
 434 {
 435     MigrationState *s = migrate_get_current();
 436
 437     /* Check for truncation */
 438     if (value != (size_t)value) {
 439         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
 440                   "exceeding address space");
 441         return;
 442     }
 443
 444     s->xbzrle_cache_size = xbzrle_cache_resize(value);
 445 }
 446
 447 int64_t qmp_query_migrate_cache_size(Error **errp)
 448 {
 449     return migrate_xbzrle_cache_size();
 450 }
 451
 452 void qmp_migrate_set_speed(int64_t value, Error **errp)
 453 {
 454     MigrationState *s;
 455
 456     if (value < 0) {
 457         value = 0;
 458     }
 459     if (value > SIZE_MAX) {
 460         value = SIZE_MAX;
 461     }
 462
 463     s = migrate_get_current();
 464     s->bandwidth_limit = value;
 465     if (s->file) {
 466         qemu_file_set_rate_limit(s->file, s->bandwidth_limit / XFER_LIMIT_RATIO);
 467     }
 468 }
 469
 470 void qmp_migrate_set_downtime(double value, Error **errp)
 471 {
 472     value *= 1e9;
 473     value = MAX(0, MIN(UINT64_MAX, value));
 474     max_downtime = (uint64_t)value;
 475 }
 476
 477 int migrate_use_xbzrle(void)
 478 {
 479     MigrationState *s;
 480
 481     s = migrate_get_current();
 482
 483     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
 484 }
 485
 486 int64_t migrate_xbzrle_cache_size(void)
 487 {
 488     MigrationState *s;
 489
 490     s = migrate_get_current();
 491
 492     return s->xbzrle_cache_size;
 493 }
 494
 495 /* migration thread support */
 496
 497 static void *migration_thread(void *opaque)
 498 {
 499     MigrationState *s = opaque;
 500     int64_t initial_time = qemu_get_clock_ms(rt_clock);
 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;
 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             initial_time = current_time;
 558             initial_bytes = qemu_ftell(s->file);
 559         }
 560         if (qemu_file_rate_limit(s->file)) {
 561             /* usleep expects microseconds */
 562             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
 563         }
 564     }
 565
 566     qemu_mutex_lock_iothread();
 567     if (s->state == MIG_STATE_COMPLETED) {
 568         int64_t end_time = qemu_get_clock_ms(rt_clock);
 569         s->total_time = end_time - s->total_time;
 570         s->downtime = end_time - start_time;
 571         runstate_set(RUN_STATE_POSTMIGRATE);
 572     } else {
 573         if (old_vm_running) {
 574             vm_start();
 575         }
 576     }
 577     qemu_bh_schedule(s->cleanup_bh);
 578     qemu_mutex_unlock_iothread();
 579
 580     return NULL;
 581 }
 582
 583 void migrate_fd_connect(MigrationState *s)
 584 {
 585     s->state = MIG_STATE_ACTIVE;
 586     trace_migrate_set_state(MIG_STATE_ACTIVE);
 587
 588     /* This is a best 1st approximation. ns to ms */
 589     s->expected_downtime = max_downtime/1000000;
 590     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
 591
 592     qemu_file_set_rate_limit(s->file,
 593                              s->bandwidth_limit / XFER_LIMIT_RATIO);
 594
 595     qemu_thread_create(&s->thread, migration_thread, s,
 596                        QEMU_THREAD_JOINABLE);
 597     notifier_list_notify(&migration_state_notifiers, s);
 598 }