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 = -1,
  40     MIG_STATE_NONE,
  41     MIG_STATE_SETUP,
  42     MIG_STATE_CANCELLED,
  43     MIG_STATE_ACTIVE,
  44     MIG_STATE_COMPLETED,
  45 };
  46
  47 #define MAX_THROTTLE  (32 << 20)      /* Migration speed throttling */
  48
  49 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
  50  * data. */
  51 #define BUFFER_DELAY     100
  52 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
  53
  54 /* Migration XBZRLE default cache size */
  55 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
  56
  57 static NotifierList migration_state_notifiers =
  58     NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
  59
  60 /* When we add fault tolerance, we could have several
  61    migrations at once.  For now we don't need to add
  62    dynamic creation of migration */
  63
  64 MigrationState *migrate_get_current(void)
  65 {
  66     static MigrationState current_migration = {
  67         .state = MIG_STATE_NONE,
  68         .bandwidth_limit = MAX_THROTTLE,
  69         .xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE,
  70         .mbps = -1,
  71     };
  72
  73     return &current_migration;
  74 }
  75
  76 void qemu_start_incoming_migration(const char *uri, Error **errp)
  77 {
  78     const char *p;
  79
  80     if (strstart(uri, "tcp:", &p))
  81         tcp_start_incoming_migration(p, errp);
  82 #ifdef CONFIG_RDMA
  83     else if (strstart(uri, "x-rdma:", &p))
  84         rdma_start_incoming_migration(p, errp);
  85 #endif
  86 #if !defined(WIN32)
  87     else if (strstart(uri, "exec:", &p))
  88         exec_start_incoming_migration(p, errp);
  89     else if (strstart(uri, "unix:", &p))
  90         unix_start_incoming_migration(p, errp);
  91     else if (strstart(uri, "fd:", &p))
  92         fd_start_incoming_migration(p, errp);
  93 #endif
  94     else {
  95         error_setg(errp, "unknown migration protocol: %s", uri);
  96     }
  97 }
  98
  99 static void process_incoming_migration_co(void *opaque)
 100 {
 101     QEMUFile *f = opaque;
 102     int ret;
 103
 104     ret = qemu_loadvm_state(f);
 105     qemu_fclose(f);
 106     if (ret < 0) {
 107         fprintf(stderr, "load of migration failed\n");
 108         exit(EXIT_FAILURE);
 109     }
 110     qemu_announce_self();
 111     DPRINTF("successfully loaded vm state\n");
 112
 113     bdrv_clear_incoming_migration_all();
 114     /* Make sure all file formats flush their mutable metadata */
 115     bdrv_invalidate_cache_all();
 116
 117     if (autostart) {
 118         vm_start();
 119     } else {
 120         runstate_set(RUN_STATE_PAUSED);
 121     }
 122 }
 123
 124 void process_incoming_migration(QEMUFile *f)
 125 {
 126     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co);
 127     int fd = qemu_get_fd(f);
 128
 129     assert(fd != -1);
 130     qemu_set_nonblock(fd);
 131     qemu_coroutine_enter(co, f);
 132 }
 133
 134 /* amount of nanoseconds we are willing to wait for migration to be down.
 135  * the choice of nanoseconds is because it is the maximum resolution that
 136  * get_clock() can achieve. It is an internal measure. All user-visible
 137  * units must be in seconds */
 138 static uint64_t max_downtime = 30000000;
 139
 140 uint64_t migrate_max_downtime(void)
 141 {
 142     return max_downtime;
 143 }
 144
 145 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
 146 {
 147     MigrationCapabilityStatusList *head = NULL;
 148     MigrationCapabilityStatusList *caps;
 149     MigrationState *s = migrate_get_current();
 150     int i;
 151
 152     for (i = 0; i < MIGRATION_CAPABILITY_MAX; i++) {
 153         if (head == NULL) {
 154             head = g_malloc0(sizeof(*caps));
 155             caps = head;
 156         } else {
 157             caps->next = g_malloc0(sizeof(*caps));
 158             caps = caps->next;
 159         }
 160         caps->value =
 161             g_malloc(sizeof(*caps->value));
 162         caps->value->capability = i;
 163         caps->value->state = s->enabled_capabilities[i];
 164     }
 165
 166     return head;
 167 }
 168
 169 static void get_xbzrle_cache_stats(MigrationInfo *info)
 170 {
 171     if (migrate_use_xbzrle()) {
 172         info->has_xbzrle_cache = true;
 173         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
 174         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
 175         info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred();
 176         info->xbzrle_cache->pages = xbzrle_mig_pages_transferred();
 177         info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss();
 178         info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow();
 179     }
 180 }
 181
 182 MigrationInfo *qmp_query_migrate(Error **errp)
 183 {
 184     MigrationInfo *info = g_malloc0(sizeof(*info));
 185     MigrationState *s = migrate_get_current();
 186
 187     switch (s->state) {
 188     case MIG_STATE_NONE:
 189         /* no migration has happened ever */
 190         break;
 191     case MIG_STATE_SETUP:
 192         info->has_status = true;
 193         info->status = g_strdup("setup");
 194         break;
 195     case MIG_STATE_ACTIVE:
 196         info->has_status = true;
 197         info->status = g_strdup("active");
 198         info->has_total_time = true;
 199         info->total_time = qemu_get_clock_ms(rt_clock)
 200             - s->total_time;
 201         info->has_expected_downtime = true;
 202         info->expected_downtime = s->expected_downtime;
 203
 204         info->has_ram = true;
 205         info->ram = g_malloc0(sizeof(*info->ram));
 206         info->ram->transferred = ram_bytes_transferred();
 207         info->ram->remaining = ram_bytes_remaining();
 208         info->ram->total = ram_bytes_total();
 209         info->ram->duplicate = dup_mig_pages_transferred();
 210         info->ram->skipped = skipped_mig_pages_transferred();
 211         info->ram->normal = norm_mig_pages_transferred();
 212         info->ram->normal_bytes = norm_mig_bytes_transferred();
 213         info->ram->dirty_pages_rate = s->dirty_pages_rate;
 214         info->ram->mbps = s->mbps;
 215
 216         if (blk_mig_active()) {
 217             info->has_disk = true;
 218             info->disk = g_malloc0(sizeof(*info->disk));
 219             info->disk->transferred = blk_mig_bytes_transferred();
 220             info->disk->remaining = blk_mig_bytes_remaining();
 221             info->disk->total = blk_mig_bytes_total();
 222         }
 223
 224         get_xbzrle_cache_stats(info);
 225         break;
 226     case MIG_STATE_COMPLETED:
 227         get_xbzrle_cache_stats(info);
 228
 229         info->has_status = true;
 230         info->status = g_strdup("completed");
 231         info->total_time = s->total_time;
 232         info->has_downtime = true;
 233         info->downtime = s->downtime;
 234
 235         info->has_ram = true;
 236         info->ram = g_malloc0(sizeof(*info->ram));
 237         info->ram->transferred = ram_bytes_transferred();
 238         info->ram->remaining = 0;
 239         info->ram->total = ram_bytes_total();
 240         info->ram->duplicate = dup_mig_pages_transferred();
 241         info->ram->skipped = skipped_mig_pages_transferred();
 242         info->ram->normal = norm_mig_pages_transferred();
 243         info->ram->normal_bytes = norm_mig_bytes_transferred();
 244         info->ram->mbps = s->mbps;
 245         break;
 246     case MIG_STATE_ERROR:
 247         info->has_status = true;
 248         info->status = g_strdup("failed");
 249         break;
 250     case MIG_STATE_CANCELLED:
 251         info->has_status = true;
 252         info->status = g_strdup("cancelled");
 253         break;
 254     }
 255
 256     return info;
 257 }
 258
 259 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 260                                   Error **errp)
 261 {
 262     MigrationState *s = migrate_get_current();
 263     MigrationCapabilityStatusList *cap;
 264
 265     if (s->state == MIG_STATE_ACTIVE || s->state == MIG_STATE_SETUP) {
 266         error_set(errp, QERR_MIGRATION_ACTIVE);
 267         return;
 268     }
 269
 270     for (cap = params; cap; cap = cap->next) {
 271         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 272     }
 273 }
 274
 275 /* shared migration helpers */
 276
 277 static void migrate_fd_cleanup(void *opaque)
 278 {
 279     MigrationState *s = opaque;
 280
 281     qemu_bh_delete(s->cleanup_bh);
 282     s->cleanup_bh = NULL;
 283
 284     if (s->file) {
 285         DPRINTF("closing file\n");
 286         qemu_mutex_unlock_iothread();
 287         qemu_thread_join(&s->thread);
 288         qemu_mutex_lock_iothread();
 289
 290         qemu_fclose(s->file);
 291         s->file = NULL;
 292     }
 293
 294     assert(s->state != MIG_STATE_ACTIVE);
 295
 296     if (s->state != MIG_STATE_COMPLETED) {
 297         qemu_savevm_state_cancel();
 298     }
 299
 300     notifier_list_notify(&migration_state_notifiers, s);
 301 }
 302
 303 static void migrate_set_state(MigrationState *s, int old_state, int new_state)
 304 {
 305     if (atomic_cmpxchg(&s->state, old_state, new_state) == new_state) {
 306         trace_migrate_set_state(new_state);
 307     }
 308 }
 309
 310 void migrate_fd_error(MigrationState *s)
 311 {
 312     DPRINTF("setting error state\n");
 313     assert(s->file == NULL);
 314     s->state = MIG_STATE_ERROR;
 315     trace_migrate_set_state(MIG_STATE_ERROR);
 316     notifier_list_notify(&migration_state_notifiers, s);
 317 }
 318
 319 static void migrate_fd_cancel(MigrationState *s)
 320 {
 321     DPRINTF("cancelling migration\n");
 322
 323     migrate_set_state(s, s->state, MIG_STATE_CANCELLED);
 324 }
 325
 326 void add_migration_state_change_notifier(Notifier *notify)
 327 {
 328     notifier_list_add(&migration_state_notifiers, notify);
 329 }
 330
 331 void remove_migration_state_change_notifier(Notifier *notify)
 332 {
 333     notifier_remove(notify);
 334 }
 335
 336 bool migration_is_active(MigrationState *s)
 337 {
 338     return s->state == MIG_STATE_ACTIVE;
 339 }
 340
 341 bool migration_has_finished(MigrationState *s)
 342 {
 343     return s->state == MIG_STATE_COMPLETED;
 344 }
 345
 346 bool migration_has_failed(MigrationState *s)
 347 {
 348     return (s->state == MIG_STATE_CANCELLED ||
 349             s->state == MIG_STATE_ERROR);
 350 }
 351
 352 static MigrationState *migrate_init(const MigrationParams *params)
 353 {
 354     MigrationState *s = migrate_get_current();
 355     int64_t bandwidth_limit = s->bandwidth_limit;
 356     bool enabled_capabilities[MIGRATION_CAPABILITY_MAX];
 357     int64_t xbzrle_cache_size = s->xbzrle_cache_size;
 358
 359     memcpy(enabled_capabilities, s->enabled_capabilities,
 360            sizeof(enabled_capabilities));
 361
 362     memset(s, 0, sizeof(*s));
 363     s->params = *params;
 364     memcpy(s->enabled_capabilities, enabled_capabilities,
 365            sizeof(enabled_capabilities));
 366     s->xbzrle_cache_size = xbzrle_cache_size;
 367
 368     s->bandwidth_limit = bandwidth_limit;
 369     s->state = MIG_STATE_SETUP;
 370     trace_migrate_set_state(MIG_STATE_SETUP);
 371
 372     s->total_time = qemu_get_clock_ms(rt_clock);
 373     return s;
 374 }
 375
 376 static GSList *migration_blockers;
 377
 378 void migrate_add_blocker(Error *reason)
 379 {
 380     migration_blockers = g_slist_prepend(migration_blockers, reason);
 381 }
 382
 383 void migrate_del_blocker(Error *reason)
 384 {
 385     migration_blockers = g_slist_remove(migration_blockers, reason);
 386 }
 387
 388 void qmp_migrate(const char *uri, bool has_blk, bool blk,
 389                  bool has_inc, bool inc, bool has_detach, bool detach,
 390                  Error **errp)
 391 {
 392     Error *local_err = NULL;
 393     MigrationState *s = migrate_get_current();
 394     MigrationParams params;
 395     const char *p;
 396
 397     params.blk = blk;
 398     params.shared = inc;
 399
 400     if (s->state == MIG_STATE_ACTIVE || s->state == MIG_STATE_SETUP) {
 401         error_set(errp, QERR_MIGRATION_ACTIVE);
 402         return;
 403     }
 404
 405     if (qemu_savevm_state_blocked(errp)) {
 406         return;
 407     }
 408
 409     if (migration_blockers) {
 410         *errp = error_copy(migration_blockers->data);
 411         return;
 412     }
 413
 414     s = migrate_init(&params);
 415
 416     if (strstart(uri, "tcp:", &p)) {
 417         tcp_start_outgoing_migration(s, p, &local_err);
 418 #ifdef CONFIG_RDMA
 419     } else if (strstart(uri, "x-rdma:", &p)) {
 420         rdma_start_outgoing_migration(s, p, &local_err);
 421 #endif
 422 #if !defined(WIN32)
 423     } else if (strstart(uri, "exec:", &p)) {
 424         exec_start_outgoing_migration(s, p, &local_err);
 425     } else if (strstart(uri, "unix:", &p)) {
 426         unix_start_outgoing_migration(s, p, &local_err);
 427     } else if (strstart(uri, "fd:", &p)) {
 428         fd_start_outgoing_migration(s, p, &local_err);
 429 #endif
 430     } else {
 431         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "uri", "a valid migration protocol");
 432         return;
 433     }
 434
 435     if (local_err) {
 436         migrate_fd_error(s);
 437         error_propagate(errp, local_err);
 438         return;
 439     }
 440 }
 441
 442 void qmp_migrate_cancel(Error **errp)
 443 {
 444     migrate_fd_cancel(migrate_get_current());
 445 }
 446
 447 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
 448 {
 449     MigrationState *s = migrate_get_current();
 450
 451     /* Check for truncation */
 452     if (value != (size_t)value) {
 453         error_set(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
 454                   "exceeding address space");
 455         return;
 456     }
 457
 458     s->xbzrle_cache_size = xbzrle_cache_resize(value);
 459 }
 460
 461 int64_t qmp_query_migrate_cache_size(Error **errp)
 462 {
 463     return migrate_xbzrle_cache_size();
 464 }
 465
 466 void qmp_migrate_set_speed(int64_t value, Error **errp)
 467 {
 468     MigrationState *s;
 469
 470     if (value < 0) {
 471         value = 0;
 472     }
 473     if (value > SIZE_MAX) {
 474         value = SIZE_MAX;
 475     }
 476
 477     s = migrate_get_current();
 478     s->bandwidth_limit = value;
 479     if (s->file) {
 480         qemu_file_set_rate_limit(s->file, s->bandwidth_limit / XFER_LIMIT_RATIO);
 481     }
 482 }
 483
 484 void qmp_migrate_set_downtime(double value, Error **errp)
 485 {
 486     value *= 1e9;
 487     value = MAX(0, MIN(UINT64_MAX, value));
 488     max_downtime = (uint64_t)value;
 489 }
 490
 491 bool migrate_rdma_pin_all(void)
 492 {
 493     MigrationState *s;
 494
 495     s = migrate_get_current();
 496
 497     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_RDMA_PIN_ALL];
 498 }
 499
 500 bool migrate_auto_converge(void)
 501 {
 502     MigrationState *s;
 503
 504     s = migrate_get_current();
 505
 506     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
 507 }
 508
 509 bool migrate_zero_blocks(void)
 510 {
 511     MigrationState *s;
 512
 513     s = migrate_get_current();
 514
 515     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
 516 }
 517
 518 int migrate_use_xbzrle(void)
 519 {
 520     MigrationState *s;
 521
 522     s = migrate_get_current();
 523
 524     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
 525 }
 526
 527 int64_t migrate_xbzrle_cache_size(void)
 528 {
 529     MigrationState *s;
 530
 531     s = migrate_get_current();
 532
 533     return s->xbzrle_cache_size;
 534 }
 535
 536 /* migration thread support */
 537
 538 static void *migration_thread(void *opaque)
 539 {
 540     MigrationState *s = opaque;
 541     int64_t initial_time = qemu_get_clock_ms(rt_clock);
 542     int64_t initial_bytes = 0;
 543     int64_t max_size = 0;
 544     int64_t start_time = initial_time;
 545     bool old_vm_running = false;
 546
 547     DPRINTF("beginning savevm\n");
 548     qemu_savevm_state_begin(s->file, &s->params);
 549
 550     migrate_set_state(s, MIG_STATE_SETUP, MIG_STATE_ACTIVE);
 551
 552     while (s->state == MIG_STATE_ACTIVE) {
 553         int64_t current_time;
 554         uint64_t pending_size;
 555
 556         if (!qemu_file_rate_limit(s->file)) {
 557             DPRINTF("iterate\n");
 558             pending_size = qemu_savevm_state_pending(s->file, max_size);
 559             DPRINTF("pending size %lu max %lu\n", pending_size, max_size);
 560             if (pending_size && pending_size >= max_size) {
 561                 qemu_savevm_state_iterate(s->file);
 562             } else {
 563                 int ret;
 564
 565                 DPRINTF("done iterating\n");
 566                 qemu_mutex_lock_iothread();
 567                 start_time = qemu_get_clock_ms(rt_clock);
 568                 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
 569                 old_vm_running = runstate_is_running();
 570
 571                 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
 572                 if (ret >= 0) {
 573                     qemu_file_set_rate_limit(s->file, INT_MAX);
 574                     qemu_savevm_state_complete(s->file);
 575                 }
 576                 qemu_mutex_unlock_iothread();
 577
 578                 if (ret < 0) {
 579                     migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_ERROR);
 580                     break;
 581                 }
 582
 583                 if (!qemu_file_get_error(s->file)) {
 584                     migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_COMPLETED);
 585                     break;
 586                 }
 587             }
 588         }
 589
 590         if (qemu_file_get_error(s->file)) {
 591             migrate_set_state(s, MIG_STATE_ACTIVE, MIG_STATE_ERROR);
 592             break;
 593         }
 594         current_time = qemu_get_clock_ms(rt_clock);
 595         if (current_time >= initial_time + BUFFER_DELAY) {
 596             uint64_t transferred_bytes = qemu_ftell(s->file) - initial_bytes;
 597             uint64_t time_spent = current_time - initial_time;
 598             double bandwidth = transferred_bytes / time_spent;
 599             max_size = bandwidth * migrate_max_downtime() / 1000000;
 600
 601             s->mbps = time_spent ? (((double) transferred_bytes * 8.0) /
 602                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0 : -1;
 603
 604             DPRINTF("transferred %" PRIu64 " time_spent %" PRIu64
 605                     " bandwidth %g max_size %" PRId64 "\n",
 606                     transferred_bytes, time_spent, bandwidth, max_size);
 607             /* if we haven't sent anything, we don't want to recalculate
 608                10000 is a small enough number for our purposes */
 609             if (s->dirty_bytes_rate && transferred_bytes > 10000) {
 610                 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
 611             }
 612
 613             qemu_file_reset_rate_limit(s->file);
 614             initial_time = current_time;
 615             initial_bytes = qemu_ftell(s->file);
 616         }
 617         if (qemu_file_rate_limit(s->file)) {
 618             /* usleep expects microseconds */
 619             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
 620         }
 621     }
 622
 623     qemu_mutex_lock_iothread();
 624     if (s->state == MIG_STATE_COMPLETED) {
 625         int64_t end_time = qemu_get_clock_ms(rt_clock);
 626         s->total_time = end_time - s->total_time;
 627         s->downtime = end_time - start_time;
 628         runstate_set(RUN_STATE_POSTMIGRATE);
 629     } else {
 630         if (old_vm_running) {
 631             vm_start();
 632         }
 633     }
 634     qemu_bh_schedule(s->cleanup_bh);
 635     qemu_mutex_unlock_iothread();
 636
 637     return NULL;
 638 }
 639
 640 void migrate_fd_connect(MigrationState *s)
 641 {
 642     s->state = MIG_STATE_SETUP;
 643     trace_migrate_set_state(MIG_STATE_SETUP);
 644
 645     /* This is a best 1st approximation. ns to ms */
 646     s->expected_downtime = max_downtime/1000000;
 647     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
 648
 649     qemu_file_set_rate_limit(s->file,
 650                              s->bandwidth_limit / XFER_LIMIT_RATIO);
 651
 652     qemu_thread_create(&s->thread, migration_thread, s,
 653                        QEMU_THREAD_JOINABLE);
 654     notifier_list_notify(&migration_state_notifiers, s);
 655 }