migration/migration.c

   1 /*
   2  * QEMU live migration
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *
   6  * Authors:
   7  *  Anthony Liguori   <[email protected]>
   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/osdep.h"
  17 #include "qemu/cutils.h"
  18 #include "qemu/error-report.h"
  19 #include "migration/blocker.h"
  20 #include "exec.h"
  21 #include "fd.h"
  22 #include "socket.h"
  23 #include "rdma.h"
  24 #include "ram.h"
  25 #include "migration/global_state.h"
  26 #include "migration/misc.h"
  27 #include "migration.h"
  28 #include "savevm.h"
  29 #include "qemu-file-channel.h"
  30 #include "qemu-file.h"
  31 #include "migration/vmstate.h"
  32 #include "block/block.h"
  33 #include "qapi/qmp/qerror.h"
  34 #include "qemu/rcu.h"
  35 #include "block.h"
  36 #include "postcopy-ram.h"
  37 #include "qemu/thread.h"
  38 #include "qmp-commands.h"
  39 #include "trace.h"
  40 #include "qapi-event.h"
  41 #include "exec/target_page.h"
  42 #include "io/channel-buffer.h"
  43 #include "migration/colo.h"
  44 #include "hw/boards.h"
  45 #include "monitor/monitor.h"
  46
  47 #define MAX_THROTTLE  (32 << 20)      /* Migration transfer 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 /* Time in milliseconds we are allowed to stop the source,
  55  * for sending the last part */
  56 #define DEFAULT_MIGRATE_SET_DOWNTIME 300
  57
  58 /* Maximum migrate downtime set to 2000 seconds */
  59 #define MAX_MIGRATE_DOWNTIME_SECONDS 2000
  60 #define MAX_MIGRATE_DOWNTIME (MAX_MIGRATE_DOWNTIME_SECONDS * 1000)
  61
  62 /* Default compression thread count */
  63 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
  64 /* Default decompression thread count, usually decompression is at
  65  * least 4 times as fast as compression.*/
  66 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2
  67 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */
  68 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1
  69 /* Define default autoconverge cpu throttle migration parameters */
  70 #define DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL 20
  71 #define DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT 10
  72
  73 /* Migration XBZRLE default cache size */
  74 #define DEFAULT_MIGRATE_XBZRLE_CACHE_SIZE (64 * 1024 * 1024)
  75
  76 /* The delay time (in ms) between two COLO checkpoints
  77  * Note: Please change this default value to 10000 when we support hybrid mode.
  78  */
  79 #define DEFAULT_MIGRATE_X_CHECKPOINT_DELAY 200
  80 #define DEFAULT_MIGRATE_MULTIFD_CHANNELS 2
  81 #define DEFAULT_MIGRATE_MULTIFD_PAGE_COUNT 16
  82
  83 static NotifierList migration_state_notifiers =
  84     NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
  85
  86 static bool deferred_incoming;
  87
  88 /* Messages sent on the return path from destination to source */
  89 enum mig_rp_message_type {
  90     MIG_RP_MSG_INVALID = 0,  /* Must be 0 */
  91     MIG_RP_MSG_SHUT,         /* sibling will not send any more RP messages */
  92     MIG_RP_MSG_PONG,         /* Response to a PING; data (seq: be32 ) */
  93
  94     MIG_RP_MSG_REQ_PAGES_ID, /* data (start: be64, len: be32, id: string) */
  95     MIG_RP_MSG_REQ_PAGES,    /* data (start: be64, len: be32) */
  96
  97     MIG_RP_MSG_MAX
  98 };
  99
 100 /* When we add fault tolerance, we could have several
 101    migrations at once.  For now we don't need to add
 102    dynamic creation of migration */
 103
 104 static MigrationState *current_migration;
 105
 106 static bool migration_object_check(MigrationState *ms, Error **errp);
 107 static int migration_maybe_pause(MigrationState *s,
 108                                  int *current_active_state,
 109                                  int new_state);
 110
 111 void migration_object_init(void)
 112 {
 113     MachineState *ms = MACHINE(qdev_get_machine());
 114     Error *err = NULL;
 115
 116     /* This can only be called once. */
 117     assert(!current_migration);
 118     current_migration = MIGRATION_OBJ(object_new(TYPE_MIGRATION));
 119
 120     if (!migration_object_check(current_migration, &err)) {
 121         error_report_err(err);
 122         exit(1);
 123     }
 124
 125     /*
 126      * We cannot really do this in migration_instance_init() since at
 127      * that time global properties are not yet applied, then this
 128      * value will be definitely replaced by something else.
 129      */
 130     if (ms->enforce_config_section) {
 131         current_migration->send_configuration = true;
 132     }
 133 }
 134
 135 void migration_object_finalize(void)
 136 {
 137     object_unref(OBJECT(current_migration));
 138 }
 139
 140 /* For outgoing */
 141 MigrationState *migrate_get_current(void)
 142 {
 143     /* This can only be called after the object created. */
 144     assert(current_migration);
 145     return current_migration;
 146 }
 147
 148 MigrationIncomingState *migration_incoming_get_current(void)
 149 {
 150     static bool once;
 151     static MigrationIncomingState mis_current;
 152
 153     if (!once) {
 154         mis_current.state = MIGRATION_STATUS_NONE;
 155         memset(&mis_current, 0, sizeof(MigrationIncomingState));
 156         qemu_mutex_init(&mis_current.rp_mutex);
 157         qemu_event_init(&mis_current.main_thread_load_event, false);
 158         once = true;
 159     }
 160     return &mis_current;
 161 }
 162
 163 void migration_incoming_state_destroy(void)
 164 {
 165     struct MigrationIncomingState *mis = migration_incoming_get_current();
 166
 167     if (mis->to_src_file) {
 168         /* Tell source that we are done */
 169         migrate_send_rp_shut(mis, qemu_file_get_error(mis->from_src_file) != 0);
 170         qemu_fclose(mis->to_src_file);
 171         mis->to_src_file = NULL;
 172     }
 173
 174     if (mis->from_src_file) {
 175         qemu_fclose(mis->from_src_file);
 176         mis->from_src_file = NULL;
 177     }
 178
 179     qemu_event_reset(&mis->main_thread_load_event);
 180 }
 181
 182 static void migrate_generate_event(int new_state)
 183 {
 184     if (migrate_use_events()) {
 185         qapi_event_send_migration(new_state, &error_abort);
 186     }
 187 }
 188
 189 /*
 190  * Called on -incoming with a defer: uri.
 191  * The migration can be started later after any parameters have been
 192  * changed.
 193  */
 194 static void deferred_incoming_migration(Error **errp)
 195 {
 196     if (deferred_incoming) {
 197         error_setg(errp, "Incoming migration already deferred");
 198     }
 199     deferred_incoming = true;
 200 }
 201
 202 /*
 203  * Send a message on the return channel back to the source
 204  * of the migration.
 205  */
 206 static void migrate_send_rp_message(MigrationIncomingState *mis,
 207                                     enum mig_rp_message_type message_type,
 208                                     uint16_t len, void *data)
 209 {
 210     trace_migrate_send_rp_message((int)message_type, len);
 211     qemu_mutex_lock(&mis->rp_mutex);
 212     qemu_put_be16(mis->to_src_file, (unsigned int)message_type);
 213     qemu_put_be16(mis->to_src_file, len);
 214     qemu_put_buffer(mis->to_src_file, data, len);
 215     qemu_fflush(mis->to_src_file);
 216     qemu_mutex_unlock(&mis->rp_mutex);
 217 }
 218
 219 /* Request a range of pages from the source VM at the given
 220  * start address.
 221  *   rbname: Name of the RAMBlock to request the page in, if NULL it's the same
 222  *           as the last request (a name must have been given previously)
 223  *   Start: Address offset within the RB
 224  *   Len: Length in bytes required - must be a multiple of pagesize
 225  */
 226 void migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname,
 227                                ram_addr_t start, size_t len)
 228 {
 229     uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
 230     size_t msglen = 12; /* start + len */
 231
 232     *(uint64_t *)bufc = cpu_to_be64((uint64_t)start);
 233     *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len);
 234
 235     if (rbname) {
 236         int rbname_len = strlen(rbname);
 237         assert(rbname_len < 256);
 238
 239         bufc[msglen++] = rbname_len;
 240         memcpy(bufc + msglen, rbname, rbname_len);
 241         msglen += rbname_len;
 242         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES_ID, msglen, bufc);
 243     } else {
 244         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES, msglen, bufc);
 245     }
 246 }
 247
 248 void qemu_start_incoming_migration(const char *uri, Error **errp)
 249 {
 250     const char *p;
 251
 252     qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort);
 253     if (!strcmp(uri, "defer")) {
 254         deferred_incoming_migration(errp);
 255     } else if (strstart(uri, "tcp:", &p)) {
 256         tcp_start_incoming_migration(p, errp);
 257 #ifdef CONFIG_RDMA
 258     } else if (strstart(uri, "rdma:", &p)) {
 259         rdma_start_incoming_migration(p, errp);
 260 #endif
 261     } else if (strstart(uri, "exec:", &p)) {
 262         exec_start_incoming_migration(p, errp);
 263     } else if (strstart(uri, "unix:", &p)) {
 264         unix_start_incoming_migration(p, errp);
 265     } else if (strstart(uri, "fd:", &p)) {
 266         fd_start_incoming_migration(p, errp);
 267     } else {
 268         error_setg(errp, "unknown migration protocol: %s", uri);
 269     }
 270 }
 271
 272 static void process_incoming_migration_bh(void *opaque)
 273 {
 274     Error *local_err = NULL;
 275     MigrationIncomingState *mis = opaque;
 276
 277     /* Make sure all file formats flush their mutable metadata.
 278      * If we get an error here, just don't restart the VM yet. */
 279     bdrv_invalidate_cache_all(&local_err);
 280     if (local_err) {
 281         error_report_err(local_err);
 282         local_err = NULL;
 283         autostart = false;
 284     }
 285
 286     /*
 287      * This must happen after all error conditions are dealt with and
 288      * we're sure the VM is going to be running on this host.
 289      */
 290     qemu_announce_self();
 291
 292     if (multifd_load_cleanup(&local_err) != 0) {
 293         error_report_err(local_err);
 294         autostart = false;
 295     }
 296     /* If global state section was not received or we are in running
 297        state, we need to obey autostart. Any other state is set with
 298        runstate_set. */
 299
 300     if (!global_state_received() ||
 301         global_state_get_runstate() == RUN_STATE_RUNNING) {
 302         if (autostart) {
 303             vm_start();
 304         } else {
 305             runstate_set(RUN_STATE_PAUSED);
 306         }
 307     } else {
 308         runstate_set(global_state_get_runstate());
 309     }
 310     /*
 311      * This must happen after any state changes since as soon as an external
 312      * observer sees this event they might start to prod at the VM assuming
 313      * it's ready to use.
 314      */
 315     migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 316                       MIGRATION_STATUS_COMPLETED);
 317     qemu_bh_delete(mis->bh);
 318     migration_incoming_state_destroy();
 319 }
 320
 321 static void process_incoming_migration_co(void *opaque)
 322 {
 323     MigrationIncomingState *mis = migration_incoming_get_current();
 324     PostcopyState ps;
 325     int ret;
 326
 327     assert(mis->from_src_file);
 328     mis->largest_page_size = qemu_ram_pagesize_largest();
 329     postcopy_state_set(POSTCOPY_INCOMING_NONE);
 330     migrate_set_state(&mis->state, MIGRATION_STATUS_NONE,
 331                       MIGRATION_STATUS_ACTIVE);
 332     ret = qemu_loadvm_state(mis->from_src_file);
 333
 334     ps = postcopy_state_get();
 335     trace_process_incoming_migration_co_end(ret, ps);
 336     if (ps != POSTCOPY_INCOMING_NONE) {
 337         if (ps == POSTCOPY_INCOMING_ADVISE) {
 338             /*
 339              * Where a migration had postcopy enabled (and thus went to advise)
 340              * but managed to complete within the precopy period, we can use
 341              * the normal exit.
 342              */
 343             postcopy_ram_incoming_cleanup(mis);
 344         } else if (ret >= 0) {
 345             /*
 346              * Postcopy was started, cleanup should happen at the end of the
 347              * postcopy thread.
 348              */
 349             trace_process_incoming_migration_co_postcopy_end_main();
 350             return;
 351         }
 352         /* Else if something went wrong then just fall out of the normal exit */
 353     }
 354
 355     /* we get COLO info, and know if we are in COLO mode */
 356     if (!ret && migration_incoming_enable_colo()) {
 357         mis->migration_incoming_co = qemu_coroutine_self();
 358         qemu_thread_create(&mis->colo_incoming_thread, "COLO incoming",
 359              colo_process_incoming_thread, mis, QEMU_THREAD_JOINABLE);
 360         mis->have_colo_incoming_thread = true;
 361         qemu_coroutine_yield();
 362
 363         /* Wait checkpoint incoming thread exit before free resource */
 364         qemu_thread_join(&mis->colo_incoming_thread);
 365     }
 366
 367     if (ret < 0) {
 368         Error *local_err = NULL;
 369
 370         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 371                           MIGRATION_STATUS_FAILED);
 372         error_report("load of migration failed: %s", strerror(-ret));
 373         qemu_fclose(mis->from_src_file);
 374         if (multifd_load_cleanup(&local_err) != 0) {
 375             error_report_err(local_err);
 376         }
 377         exit(EXIT_FAILURE);
 378     }
 379     mis->bh = qemu_bh_new(process_incoming_migration_bh, mis);
 380     qemu_bh_schedule(mis->bh);
 381 }
 382
 383 static void migration_incoming_setup(QEMUFile *f)
 384 {
 385     MigrationIncomingState *mis = migration_incoming_get_current();
 386
 387     if (multifd_load_setup() != 0) {
 388         /* We haven't been able to create multifd threads
 389            nothing better to do */
 390         exit(EXIT_FAILURE);
 391     }
 392
 393     if (!mis->from_src_file) {
 394         mis->from_src_file = f;
 395     }
 396     qemu_file_set_blocking(f, false);
 397 }
 398
 399 static void migration_incoming_process(void)
 400 {
 401     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co, NULL);
 402     qemu_coroutine_enter(co);
 403 }
 404
 405 void migration_fd_process_incoming(QEMUFile *f)
 406 {
 407     migration_incoming_setup(f);
 408     migration_incoming_process();
 409 }
 410
 411 void migration_ioc_process_incoming(QIOChannel *ioc)
 412 {
 413     MigrationIncomingState *mis = migration_incoming_get_current();
 414
 415     if (!mis->from_src_file) {
 416         QEMUFile *f = qemu_fopen_channel_input(ioc);
 417         migration_fd_process_incoming(f);
 418     }
 419     /* We still only have a single channel.  Nothing to do here yet */
 420 }
 421
 422 /**
 423  * @migration_has_all_channels: We have received all channels that we need
 424  *
 425  * Returns true when we have got connections to all the channels that
 426  * we need for migration.
 427  */
 428 bool migration_has_all_channels(void)
 429 {
 430     return true;
 431 }
 432
 433 /*
 434  * Send a 'SHUT' message on the return channel with the given value
 435  * to indicate that we've finished with the RP.  Non-0 value indicates
 436  * error.
 437  */
 438 void migrate_send_rp_shut(MigrationIncomingState *mis,
 439                           uint32_t value)
 440 {
 441     uint32_t buf;
 442
 443     buf = cpu_to_be32(value);
 444     migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf);
 445 }
 446
 447 /*
 448  * Send a 'PONG' message on the return channel with the given value
 449  * (normally in response to a 'PING')
 450  */
 451 void migrate_send_rp_pong(MigrationIncomingState *mis,
 452                           uint32_t value)
 453 {
 454     uint32_t buf;
 455
 456     buf = cpu_to_be32(value);
 457     migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf);
 458 }
 459
 460 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
 461 {
 462     MigrationCapabilityStatusList *head = NULL;
 463     MigrationCapabilityStatusList *caps;
 464     MigrationState *s = migrate_get_current();
 465     int i;
 466
 467     caps = NULL; /* silence compiler warning */
 468     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
 469 #ifndef CONFIG_LIVE_BLOCK_MIGRATION
 470         if (i == MIGRATION_CAPABILITY_BLOCK) {
 471             continue;
 472         }
 473 #endif
 474         if (head == NULL) {
 475             head = g_malloc0(sizeof(*caps));
 476             caps = head;
 477         } else {
 478             caps->next = g_malloc0(sizeof(*caps));
 479             caps = caps->next;
 480         }
 481         caps->value =
 482             g_malloc(sizeof(*caps->value));
 483         caps->value->capability = i;
 484         caps->value->state = s->enabled_capabilities[i];
 485     }
 486
 487     return head;
 488 }
 489
 490 MigrationParameters *qmp_query_migrate_parameters(Error **errp)
 491 {
 492     MigrationParameters *params;
 493     MigrationState *s = migrate_get_current();
 494
 495     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 496     params = g_malloc0(sizeof(*params));
 497     params->has_compress_level = true;
 498     params->compress_level = s->parameters.compress_level;
 499     params->has_compress_threads = true;
 500     params->compress_threads = s->parameters.compress_threads;
 501     params->has_decompress_threads = true;
 502     params->decompress_threads = s->parameters.decompress_threads;
 503     params->has_cpu_throttle_initial = true;
 504     params->cpu_throttle_initial = s->parameters.cpu_throttle_initial;
 505     params->has_cpu_throttle_increment = true;
 506     params->cpu_throttle_increment = s->parameters.cpu_throttle_increment;
 507     params->has_tls_creds = true;
 508     params->tls_creds = g_strdup(s->parameters.tls_creds);
 509     params->has_tls_hostname = true;
 510     params->tls_hostname = g_strdup(s->parameters.tls_hostname);
 511     params->has_max_bandwidth = true;
 512     params->max_bandwidth = s->parameters.max_bandwidth;
 513     params->has_downtime_limit = true;
 514     params->downtime_limit = s->parameters.downtime_limit;
 515     params->has_x_checkpoint_delay = true;
 516     params->x_checkpoint_delay = s->parameters.x_checkpoint_delay;
 517     params->has_block_incremental = true;
 518     params->block_incremental = s->parameters.block_incremental;
 519     params->has_x_multifd_channels = true;
 520     params->x_multifd_channels = s->parameters.x_multifd_channels;
 521     params->has_x_multifd_page_count = true;
 522     params->x_multifd_page_count = s->parameters.x_multifd_page_count;
 523     params->has_xbzrle_cache_size = true;
 524     params->xbzrle_cache_size = s->parameters.xbzrle_cache_size;
 525
 526     return params;
 527 }
 528
 529 /*
 530  * Return true if we're already in the middle of a migration
 531  * (i.e. any of the active or setup states)
 532  */
 533 static bool migration_is_setup_or_active(int state)
 534 {
 535     switch (state) {
 536     case MIGRATION_STATUS_ACTIVE:
 537     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 538     case MIGRATION_STATUS_SETUP:
 539     case MIGRATION_STATUS_PRE_SWITCHOVER:
 540     case MIGRATION_STATUS_DEVICE:
 541         return true;
 542
 543     default:
 544         return false;
 545
 546     }
 547 }
 548
 549 static void populate_ram_info(MigrationInfo *info, MigrationState *s)
 550 {
 551     info->has_ram = true;
 552     info->ram = g_malloc0(sizeof(*info->ram));
 553     info->ram->transferred = ram_counters.transferred;
 554     info->ram->total = ram_bytes_total();
 555     info->ram->duplicate = ram_counters.duplicate;
 556     /* legacy value.  It is not used anymore */
 557     info->ram->skipped = 0;
 558     info->ram->normal = ram_counters.normal;
 559     info->ram->normal_bytes = ram_counters.normal *
 560         qemu_target_page_size();
 561     info->ram->mbps = s->mbps;
 562     info->ram->dirty_sync_count = ram_counters.dirty_sync_count;
 563     info->ram->postcopy_requests = ram_counters.postcopy_requests;
 564     info->ram->page_size = qemu_target_page_size();
 565
 566     if (migrate_use_xbzrle()) {
 567         info->has_xbzrle_cache = true;
 568         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
 569         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
 570         info->xbzrle_cache->bytes = xbzrle_counters.bytes;
 571         info->xbzrle_cache->pages = xbzrle_counters.pages;
 572         info->xbzrle_cache->cache_miss = xbzrle_counters.cache_miss;
 573         info->xbzrle_cache->cache_miss_rate = xbzrle_counters.cache_miss_rate;
 574         info->xbzrle_cache->overflow = xbzrle_counters.overflow;
 575     }
 576
 577     if (cpu_throttle_active()) {
 578         info->has_cpu_throttle_percentage = true;
 579         info->cpu_throttle_percentage = cpu_throttle_get_percentage();
 580     }
 581
 582     if (s->state != MIGRATION_STATUS_COMPLETED) {
 583         info->ram->remaining = ram_bytes_remaining();
 584         info->ram->dirty_pages_rate = ram_counters.dirty_pages_rate;
 585     }
 586 }
 587
 588 static void populate_disk_info(MigrationInfo *info)
 589 {
 590     if (blk_mig_active()) {
 591         info->has_disk = true;
 592         info->disk = g_malloc0(sizeof(*info->disk));
 593         info->disk->transferred = blk_mig_bytes_transferred();
 594         info->disk->remaining = blk_mig_bytes_remaining();
 595         info->disk->total = blk_mig_bytes_total();
 596     }
 597 }
 598
 599 static void fill_source_migration_info(MigrationInfo *info)
 600 {
 601     MigrationState *s = migrate_get_current();
 602
 603     switch (s->state) {
 604     case MIGRATION_STATUS_NONE:
 605         /* no migration has happened ever */
 606         /* do not overwrite destination migration status */
 607         return;
 608         break;
 609     case MIGRATION_STATUS_SETUP:
 610         info->has_status = true;
 611         info->has_total_time = false;
 612         break;
 613     case MIGRATION_STATUS_ACTIVE:
 614     case MIGRATION_STATUS_CANCELLING:
 615     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 616     case MIGRATION_STATUS_PRE_SWITCHOVER:
 617     case MIGRATION_STATUS_DEVICE:
 618          /* TODO add some postcopy stats */
 619         info->has_status = true;
 620         info->has_total_time = true;
 621         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 622             - s->start_time;
 623         info->has_expected_downtime = true;
 624         info->expected_downtime = s->expected_downtime;
 625         info->has_setup_time = true;
 626         info->setup_time = s->setup_time;
 627
 628         populate_ram_info(info, s);
 629         populate_disk_info(info);
 630         break;
 631     case MIGRATION_STATUS_COLO:
 632         info->has_status = true;
 633         /* TODO: display COLO specific information (checkpoint info etc.) */
 634         break;
 635     case MIGRATION_STATUS_COMPLETED:
 636         info->has_status = true;
 637         info->has_total_time = true;
 638         info->total_time = s->total_time;
 639         info->has_downtime = true;
 640         info->downtime = s->downtime;
 641         info->has_setup_time = true;
 642         info->setup_time = s->setup_time;
 643
 644         populate_ram_info(info, s);
 645         break;
 646     case MIGRATION_STATUS_FAILED:
 647         info->has_status = true;
 648         if (s->error) {
 649             info->has_error_desc = true;
 650             info->error_desc = g_strdup(error_get_pretty(s->error));
 651         }
 652         break;
 653     case MIGRATION_STATUS_CANCELLED:
 654         info->has_status = true;
 655         break;
 656     }
 657     info->status = s->state;
 658 }
 659
 660 /**
 661  * @migration_caps_check - check capability validity
 662  *
 663  * @cap_list: old capability list, array of bool
 664  * @params: new capabilities to be applied soon
 665  * @errp: set *errp if the check failed, with reason
 666  *
 667  * Returns true if check passed, otherwise false.
 668  */
 669 static bool migrate_caps_check(bool *cap_list,
 670                                MigrationCapabilityStatusList *params,
 671                                Error **errp)
 672 {
 673     MigrationCapabilityStatusList *cap;
 674     bool old_postcopy_cap;
 675     MigrationIncomingState *mis = migration_incoming_get_current();
 676
 677     old_postcopy_cap = cap_list[MIGRATION_CAPABILITY_POSTCOPY_RAM];
 678
 679     for (cap = params; cap; cap = cap->next) {
 680         cap_list[cap->value->capability] = cap->value->state;
 681     }
 682
 683 #ifndef CONFIG_LIVE_BLOCK_MIGRATION
 684     if (cap_list[MIGRATION_CAPABILITY_BLOCK]) {
 685         error_setg(errp, "QEMU compiled without old-style (blk/-b, inc/-i) "
 686                    "block migration");
 687         error_append_hint(errp, "Use drive_mirror+NBD instead.\n");
 688         return false;
 689     }
 690 #endif
 691
 692     if (cap_list[MIGRATION_CAPABILITY_POSTCOPY_RAM]) {
 693         if (cap_list[MIGRATION_CAPABILITY_COMPRESS]) {
 694             /* The decompression threads asynchronously write into RAM
 695              * rather than use the atomic copies needed to avoid
 696              * userfaulting.  It should be possible to fix the decompression
 697              * threads for compatibility in future.
 698              */
 699             error_setg(errp, "Postcopy is not currently compatible "
 700                        "with compression");
 701             return false;
 702         }
 703
 704         /* This check is reasonably expensive, so only when it's being
 705          * set the first time, also it's only the destination that needs
 706          * special support.
 707          */
 708         if (!old_postcopy_cap && runstate_check(RUN_STATE_INMIGRATE) &&
 709             !postcopy_ram_supported_by_host(mis)) {
 710             /* postcopy_ram_supported_by_host will have emitted a more
 711              * detailed message
 712              */
 713             error_setg(errp, "Postcopy is not supported");
 714             return false;
 715         }
 716     }
 717
 718     return true;
 719 }
 720
 721 static void fill_destination_migration_info(MigrationInfo *info)
 722 {
 723     MigrationIncomingState *mis = migration_incoming_get_current();
 724
 725     switch (mis->state) {
 726     case MIGRATION_STATUS_NONE:
 727         return;
 728         break;
 729     case MIGRATION_STATUS_SETUP:
 730     case MIGRATION_STATUS_CANCELLING:
 731     case MIGRATION_STATUS_CANCELLED:
 732     case MIGRATION_STATUS_ACTIVE:
 733     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 734     case MIGRATION_STATUS_FAILED:
 735     case MIGRATION_STATUS_COLO:
 736         info->has_status = true;
 737         break;
 738     case MIGRATION_STATUS_COMPLETED:
 739         info->has_status = true;
 740         fill_destination_postcopy_migration_info(info);
 741         break;
 742     }
 743     info->status = mis->state;
 744 }
 745
 746 MigrationInfo *qmp_query_migrate(Error **errp)
 747 {
 748     MigrationInfo *info = g_malloc0(sizeof(*info));
 749
 750     fill_destination_migration_info(info);
 751     fill_source_migration_info(info);
 752
 753     return info;
 754 }
 755
 756 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 757                                   Error **errp)
 758 {
 759     MigrationState *s = migrate_get_current();
 760     MigrationCapabilityStatusList *cap;
 761
 762     if (migration_is_setup_or_active(s->state)) {
 763         error_setg(errp, QERR_MIGRATION_ACTIVE);
 764         return;
 765     }
 766
 767     if (!migrate_caps_check(s->enabled_capabilities, params, errp)) {
 768         return;
 769     }
 770
 771     for (cap = params; cap; cap = cap->next) {
 772         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 773     }
 774 }
 775
 776 /*
 777  * Check whether the parameters are valid. Error will be put into errp
 778  * (if provided). Return true if valid, otherwise false.
 779  */
 780 static bool migrate_params_check(MigrationParameters *params, Error **errp)
 781 {
 782     if (params->has_compress_level &&
 783         (params->compress_level > 9)) {
 784         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level",
 785                    "is invalid, it should be in the range of 0 to 9");
 786         return false;
 787     }
 788
 789     if (params->has_compress_threads && (params->compress_threads < 1)) {
 790         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 791                    "compress_threads",
 792                    "is invalid, it should be in the range of 1 to 255");
 793         return false;
 794     }
 795
 796     if (params->has_decompress_threads && (params->decompress_threads < 1)) {
 797         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 798                    "decompress_threads",
 799                    "is invalid, it should be in the range of 1 to 255");
 800         return false;
 801     }
 802
 803     if (params->has_cpu_throttle_initial &&
 804         (params->cpu_throttle_initial < 1 ||
 805          params->cpu_throttle_initial > 99)) {
 806         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 807                    "cpu_throttle_initial",
 808                    "an integer in the range of 1 to 99");
 809         return false;
 810     }
 811
 812     if (params->has_cpu_throttle_increment &&
 813         (params->cpu_throttle_increment < 1 ||
 814          params->cpu_throttle_increment > 99)) {
 815         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 816                    "cpu_throttle_increment",
 817                    "an integer in the range of 1 to 99");
 818         return false;
 819     }
 820
 821     if (params->has_max_bandwidth && (params->max_bandwidth > SIZE_MAX)) {
 822         error_setg(errp, "Parameter 'max_bandwidth' expects an integer in the"
 823                          " range of 0 to %zu bytes/second", SIZE_MAX);
 824         return false;
 825     }
 826
 827     if (params->has_downtime_limit &&
 828         (params->downtime_limit > MAX_MIGRATE_DOWNTIME)) {
 829         error_setg(errp, "Parameter 'downtime_limit' expects an integer in "
 830                          "the range of 0 to %d milliseconds",
 831                          MAX_MIGRATE_DOWNTIME);
 832         return false;
 833     }
 834
 835     /* x_checkpoint_delay is now always positive */
 836
 837     if (params->has_x_multifd_channels && (params->x_multifd_channels < 1)) {
 838         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 839                    "multifd_channels",
 840                    "is invalid, it should be in the range of 1 to 255");
 841         return false;
 842     }
 843     if (params->has_x_multifd_page_count &&
 844         (params->x_multifd_page_count < 1 ||
 845          params->x_multifd_page_count > 10000)) {
 846         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 847                    "multifd_page_count",
 848                    "is invalid, it should be in the range of 1 to 10000");
 849         return false;
 850     }
 851
 852     if (params->has_xbzrle_cache_size &&
 853         (params->xbzrle_cache_size < qemu_target_page_size() ||
 854          !is_power_of_2(params->xbzrle_cache_size))) {
 855         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 856                    "xbzrle_cache_size",
 857                    "is invalid, it should be bigger than target page size"
 858                    " and a power of two");
 859         return false;
 860     }
 861
 862     return true;
 863 }
 864
 865 static void migrate_params_test_apply(MigrateSetParameters *params,
 866                                       MigrationParameters *dest)
 867 {
 868     *dest = migrate_get_current()->parameters;
 869
 870     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 871
 872     if (params->has_compress_level) {
 873         dest->compress_level = params->compress_level;
 874     }
 875
 876     if (params->has_compress_threads) {
 877         dest->compress_threads = params->compress_threads;
 878     }
 879
 880     if (params->has_decompress_threads) {
 881         dest->decompress_threads = params->decompress_threads;
 882     }
 883
 884     if (params->has_cpu_throttle_initial) {
 885         dest->cpu_throttle_initial = params->cpu_throttle_initial;
 886     }
 887
 888     if (params->has_cpu_throttle_increment) {
 889         dest->cpu_throttle_increment = params->cpu_throttle_increment;
 890     }
 891
 892     if (params->has_tls_creds) {
 893         assert(params->tls_creds->type == QTYPE_QSTRING);
 894         dest->tls_creds = g_strdup(params->tls_creds->u.s);
 895     }
 896
 897     if (params->has_tls_hostname) {
 898         assert(params->tls_hostname->type == QTYPE_QSTRING);
 899         dest->tls_hostname = g_strdup(params->tls_hostname->u.s);
 900     }
 901
 902     if (params->has_max_bandwidth) {
 903         dest->max_bandwidth = params->max_bandwidth;
 904     }
 905
 906     if (params->has_downtime_limit) {
 907         dest->downtime_limit = params->downtime_limit;
 908     }
 909
 910     if (params->has_x_checkpoint_delay) {
 911         dest->x_checkpoint_delay = params->x_checkpoint_delay;
 912     }
 913
 914     if (params->has_block_incremental) {
 915         dest->block_incremental = params->block_incremental;
 916     }
 917     if (params->has_x_multifd_channels) {
 918         dest->x_multifd_channels = params->x_multifd_channels;
 919     }
 920     if (params->has_x_multifd_page_count) {
 921         dest->x_multifd_page_count = params->x_multifd_page_count;
 922     }
 923     if (params->has_xbzrle_cache_size) {
 924         dest->xbzrle_cache_size = params->xbzrle_cache_size;
 925     }
 926 }
 927
 928 static void migrate_params_apply(MigrateSetParameters *params, Error **errp)
 929 {
 930     MigrationState *s = migrate_get_current();
 931
 932     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 933
 934     if (params->has_compress_level) {
 935         s->parameters.compress_level = params->compress_level;
 936     }
 937
 938     if (params->has_compress_threads) {
 939         s->parameters.compress_threads = params->compress_threads;
 940     }
 941
 942     if (params->has_decompress_threads) {
 943         s->parameters.decompress_threads = params->decompress_threads;
 944     }
 945
 946     if (params->has_cpu_throttle_initial) {
 947         s->parameters.cpu_throttle_initial = params->cpu_throttle_initial;
 948     }
 949
 950     if (params->has_cpu_throttle_increment) {
 951         s->parameters.cpu_throttle_increment = params->cpu_throttle_increment;
 952     }
 953
 954     if (params->has_tls_creds) {
 955         g_free(s->parameters.tls_creds);
 956         assert(params->tls_creds->type == QTYPE_QSTRING);
 957         s->parameters.tls_creds = g_strdup(params->tls_creds->u.s);
 958     }
 959
 960     if (params->has_tls_hostname) {
 961         g_free(s->parameters.tls_hostname);
 962         assert(params->tls_hostname->type == QTYPE_QSTRING);
 963         s->parameters.tls_hostname = g_strdup(params->tls_hostname->u.s);
 964     }
 965
 966     if (params->has_max_bandwidth) {
 967         s->parameters.max_bandwidth = params->max_bandwidth;
 968         if (s->to_dst_file) {
 969             qemu_file_set_rate_limit(s->to_dst_file,
 970                                 s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
 971         }
 972     }
 973
 974     if (params->has_downtime_limit) {
 975         s->parameters.downtime_limit = params->downtime_limit;
 976     }
 977
 978     if (params->has_x_checkpoint_delay) {
 979         s->parameters.x_checkpoint_delay = params->x_checkpoint_delay;
 980         if (migration_in_colo_state()) {
 981             colo_checkpoint_notify(s);
 982         }
 983     }
 984
 985     if (params->has_block_incremental) {
 986         s->parameters.block_incremental = params->block_incremental;
 987     }
 988     if (params->has_x_multifd_channels) {
 989         s->parameters.x_multifd_channels = params->x_multifd_channels;
 990     }
 991     if (params->has_x_multifd_page_count) {
 992         s->parameters.x_multifd_page_count = params->x_multifd_page_count;
 993     }
 994     if (params->has_xbzrle_cache_size) {
 995         s->parameters.xbzrle_cache_size = params->xbzrle_cache_size;
 996         xbzrle_cache_resize(params->xbzrle_cache_size, errp);
 997     }
 998 }
 999
1000 void qmp_migrate_set_parameters(MigrateSetParameters *params, Error **errp)
1001 {
1002     MigrationParameters tmp;
1003
1004     /* TODO Rewrite "" to null instead */
1005     if (params->has_tls_creds
1006         && params->tls_creds->type == QTYPE_QNULL) {
1007         QDECREF(params->tls_creds->u.n);
1008         params->tls_creds->type = QTYPE_QSTRING;
1009         params->tls_creds->u.s = strdup("");
1010     }
1011     /* TODO Rewrite "" to null instead */
1012     if (params->has_tls_hostname
1013         && params->tls_hostname->type == QTYPE_QNULL) {
1014         QDECREF(params->tls_hostname->u.n);
1015         params->tls_hostname->type = QTYPE_QSTRING;
1016         params->tls_hostname->u.s = strdup("");
1017     }
1018
1019     migrate_params_test_apply(params, &tmp);
1020
1021     if (!migrate_params_check(&tmp, errp)) {
1022         /* Invalid parameter */
1023         return;
1024     }
1025
1026     migrate_params_apply(params, errp);
1027 }
1028
1029
1030 void qmp_migrate_start_postcopy(Error **errp)
1031 {
1032     MigrationState *s = migrate_get_current();
1033
1034     if (!migrate_postcopy_ram()) {
1035         error_setg(errp, "Enable postcopy with migrate_set_capability before"
1036                          " the start of migration");
1037         return;
1038     }
1039
1040     if (s->state == MIGRATION_STATUS_NONE) {
1041         error_setg(errp, "Postcopy must be started after migration has been"
1042                          " started");
1043         return;
1044     }
1045     /*
1046      * we don't error if migration has finished since that would be racy
1047      * with issuing this command.
1048      */
1049     atomic_set(&s->start_postcopy, true);
1050 }
1051
1052 /* shared migration helpers */
1053
1054 void migrate_set_state(int *state, int old_state, int new_state)
1055 {
1056     assert(new_state < MIGRATION_STATUS__MAX);
1057     if (atomic_cmpxchg(state, old_state, new_state) == old_state) {
1058         trace_migrate_set_state(MigrationStatus_str(new_state));
1059         migrate_generate_event(new_state);
1060     }
1061 }
1062
1063 static MigrationCapabilityStatusList *migrate_cap_add(
1064     MigrationCapabilityStatusList *list,
1065     MigrationCapability index,
1066     bool state)
1067 {
1068     MigrationCapabilityStatusList *cap;
1069
1070     cap = g_new0(MigrationCapabilityStatusList, 1);
1071     cap->value = g_new0(MigrationCapabilityStatus, 1);
1072     cap->value->capability = index;
1073     cap->value->state = state;
1074     cap->next = list;
1075
1076     return cap;
1077 }
1078
1079 void migrate_set_block_enabled(bool value, Error **errp)
1080 {
1081     MigrationCapabilityStatusList *cap;
1082
1083     cap = migrate_cap_add(NULL, MIGRATION_CAPABILITY_BLOCK, value);
1084     qmp_migrate_set_capabilities(cap, errp);
1085     qapi_free_MigrationCapabilityStatusList(cap);
1086 }
1087
1088 static void migrate_set_block_incremental(MigrationState *s, bool value)
1089 {
1090     s->parameters.block_incremental = value;
1091 }
1092
1093 static void block_cleanup_parameters(MigrationState *s)
1094 {
1095     if (s->must_remove_block_options) {
1096         /* setting to false can never fail */
1097         migrate_set_block_enabled(false, &error_abort);
1098         migrate_set_block_incremental(s, false);
1099         s->must_remove_block_options = false;
1100     }
1101 }
1102
1103 static void migrate_fd_cleanup(void *opaque)
1104 {
1105     MigrationState *s = opaque;
1106
1107     qemu_bh_delete(s->cleanup_bh);
1108     s->cleanup_bh = NULL;
1109
1110     qemu_savevm_state_cleanup();
1111
1112     if (s->to_dst_file) {
1113         Error *local_err = NULL;
1114
1115         trace_migrate_fd_cleanup();
1116         qemu_mutex_unlock_iothread();
1117         if (s->migration_thread_running) {
1118             qemu_thread_join(&s->thread);
1119             s->migration_thread_running = false;
1120         }
1121         qemu_mutex_lock_iothread();
1122
1123         if (multifd_save_cleanup(&local_err) != 0) {
1124             error_report_err(local_err);
1125         }
1126         qemu_fclose(s->to_dst_file);
1127         s->to_dst_file = NULL;
1128     }
1129
1130     assert((s->state != MIGRATION_STATUS_ACTIVE) &&
1131            (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE));
1132
1133     if (s->state == MIGRATION_STATUS_CANCELLING) {
1134         migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING,
1135                           MIGRATION_STATUS_CANCELLED);
1136     }
1137
1138     if (s->error) {
1139         /* It is used on info migrate.  We can't free it */
1140         error_report_err(error_copy(s->error));
1141     }
1142     notifier_list_notify(&migration_state_notifiers, s);
1143     block_cleanup_parameters(s);
1144 }
1145
1146 void migrate_set_error(MigrationState *s, const Error *error)
1147 {
1148     qemu_mutex_lock(&s->error_mutex);
1149     if (!s->error) {
1150         s->error = error_copy(error);
1151     }
1152     qemu_mutex_unlock(&s->error_mutex);
1153 }
1154
1155 void migrate_fd_error(MigrationState *s, const Error *error)
1156 {
1157     trace_migrate_fd_error(error_get_pretty(error));
1158     assert(s->to_dst_file == NULL);
1159     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1160                       MIGRATION_STATUS_FAILED);
1161     migrate_set_error(s, error);
1162     notifier_list_notify(&migration_state_notifiers, s);
1163     block_cleanup_parameters(s);
1164 }
1165
1166 static void migrate_fd_cancel(MigrationState *s)
1167 {
1168     int old_state ;
1169     QEMUFile *f = migrate_get_current()->to_dst_file;
1170     trace_migrate_fd_cancel();
1171
1172     if (s->rp_state.from_dst_file) {
1173         /* shutdown the rp socket, so causing the rp thread to shutdown */
1174         qemu_file_shutdown(s->rp_state.from_dst_file);
1175     }
1176
1177     do {
1178         old_state = s->state;
1179         if (!migration_is_setup_or_active(old_state)) {
1180             break;
1181         }
1182         /* If the migration is paused, kick it out of the pause */
1183         if (old_state == MIGRATION_STATUS_PRE_SWITCHOVER) {
1184             qemu_sem_post(&s->pause_sem);
1185         }
1186         migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING);
1187     } while (s->state != MIGRATION_STATUS_CANCELLING);
1188
1189     /*
1190      * If we're unlucky the migration code might be stuck somewhere in a
1191      * send/write while the network has failed and is waiting to timeout;
1192      * if we've got shutdown(2) available then we can force it to quit.
1193      * The outgoing qemu file gets closed in migrate_fd_cleanup that is
1194      * called in a bh, so there is no race against this cancel.
1195      */
1196     if (s->state == MIGRATION_STATUS_CANCELLING && f) {
1197         qemu_file_shutdown(f);
1198     }
1199     if (s->state == MIGRATION_STATUS_CANCELLING && s->block_inactive) {
1200         Error *local_err = NULL;
1201
1202         bdrv_invalidate_cache_all(&local_err);
1203         if (local_err) {
1204             error_report_err(local_err);
1205         } else {
1206             s->block_inactive = false;
1207         }
1208     }
1209     block_cleanup_parameters(s);
1210 }
1211
1212 void add_migration_state_change_notifier(Notifier *notify)
1213 {
1214     notifier_list_add(&migration_state_notifiers, notify);
1215 }
1216
1217 void remove_migration_state_change_notifier(Notifier *notify)
1218 {
1219     notifier_remove(notify);
1220 }
1221
1222 bool migration_in_setup(MigrationState *s)
1223 {
1224     return s->state == MIGRATION_STATUS_SETUP;
1225 }
1226
1227 bool migration_has_finished(MigrationState *s)
1228 {
1229     return s->state == MIGRATION_STATUS_COMPLETED;
1230 }
1231
1232 bool migration_has_failed(MigrationState *s)
1233 {
1234     return (s->state == MIGRATION_STATUS_CANCELLED ||
1235             s->state == MIGRATION_STATUS_FAILED);
1236 }
1237
1238 bool migration_in_postcopy(void)
1239 {
1240     MigrationState *s = migrate_get_current();
1241
1242     return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
1243 }
1244
1245 bool migration_in_postcopy_after_devices(MigrationState *s)
1246 {
1247     return migration_in_postcopy() && s->postcopy_after_devices;
1248 }
1249
1250 bool migration_is_idle(void)
1251 {
1252     MigrationState *s = migrate_get_current();
1253
1254     switch (s->state) {
1255     case MIGRATION_STATUS_NONE:
1256     case MIGRATION_STATUS_CANCELLED:
1257     case MIGRATION_STATUS_COMPLETED:
1258     case MIGRATION_STATUS_FAILED:
1259         return true;
1260     case MIGRATION_STATUS_SETUP:
1261     case MIGRATION_STATUS_CANCELLING:
1262     case MIGRATION_STATUS_ACTIVE:
1263     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
1264     case MIGRATION_STATUS_COLO:
1265     case MIGRATION_STATUS_PRE_SWITCHOVER:
1266     case MIGRATION_STATUS_DEVICE:
1267         return false;
1268     case MIGRATION_STATUS__MAX:
1269         g_assert_not_reached();
1270     }
1271
1272     return false;
1273 }
1274
1275 MigrationState *migrate_init(void)
1276 {
1277     MigrationState *s = migrate_get_current();
1278
1279     /*
1280      * Reinitialise all migration state, except
1281      * parameters/capabilities that the user set, and
1282      * locks.
1283      */
1284     s->bytes_xfer = 0;
1285     s->xfer_limit = 0;
1286     s->cleanup_bh = 0;
1287     s->to_dst_file = NULL;
1288     s->state = MIGRATION_STATUS_NONE;
1289     s->rp_state.from_dst_file = NULL;
1290     s->rp_state.error = false;
1291     s->mbps = 0.0;
1292     s->downtime = 0;
1293     s->expected_downtime = 0;
1294     s->setup_time = 0;
1295     s->start_postcopy = false;
1296     s->postcopy_after_devices = false;
1297     s->migration_thread_running = false;
1298     error_free(s->error);
1299     s->error = NULL;
1300
1301     migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP);
1302
1303     s->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1304     s->total_time = 0;
1305     return s;
1306 }
1307
1308 static GSList *migration_blockers;
1309
1310 int migrate_add_blocker(Error *reason, Error **errp)
1311 {
1312     if (migrate_get_current()->only_migratable) {
1313         error_propagate(errp, error_copy(reason));
1314         error_prepend(errp, "disallowing migration blocker "
1315                           "(--only_migratable) for: ");
1316         return -EACCES;
1317     }
1318
1319     if (migration_is_idle()) {
1320         migration_blockers = g_slist_prepend(migration_blockers, reason);
1321         return 0;
1322     }
1323
1324     error_propagate(errp, error_copy(reason));
1325     error_prepend(errp, "disallowing migration blocker (migration in "
1326                       "progress) for: ");
1327     return -EBUSY;
1328 }
1329
1330 void migrate_del_blocker(Error *reason)
1331 {
1332     migration_blockers = g_slist_remove(migration_blockers, reason);
1333 }
1334
1335 void qmp_migrate_incoming(const char *uri, Error **errp)
1336 {
1337     Error *local_err = NULL;
1338     static bool once = true;
1339
1340     if (!deferred_incoming) {
1341         error_setg(errp, "For use with '-incoming defer'");
1342         return;
1343     }
1344     if (!once) {
1345         error_setg(errp, "The incoming migration has already been started");
1346     }
1347
1348     qemu_start_incoming_migration(uri, &local_err);
1349
1350     if (local_err) {
1351         error_propagate(errp, local_err);
1352         return;
1353     }
1354
1355     once = false;
1356 }
1357
1358 bool migration_is_blocked(Error **errp)
1359 {
1360     if (qemu_savevm_state_blocked(errp)) {
1361         return true;
1362     }
1363
1364     if (migration_blockers) {
1365         error_propagate(errp, error_copy(migration_blockers->data));
1366         return true;
1367     }
1368
1369     return false;
1370 }
1371
1372 void qmp_migrate(const char *uri, bool has_blk, bool blk,
1373                  bool has_inc, bool inc, bool has_detach, bool detach,
1374                  Error **errp)
1375 {
1376     Error *local_err = NULL;
1377     MigrationState *s = migrate_get_current();
1378     const char *p;
1379
1380     if (migration_is_setup_or_active(s->state) ||
1381         s->state == MIGRATION_STATUS_CANCELLING ||
1382         s->state == MIGRATION_STATUS_COLO) {
1383         error_setg(errp, QERR_MIGRATION_ACTIVE);
1384         return;
1385     }
1386     if (runstate_check(RUN_STATE_INMIGRATE)) {
1387         error_setg(errp, "Guest is waiting for an incoming migration");
1388         return;
1389     }
1390
1391     if (migration_is_blocked(errp)) {
1392         return;
1393     }
1394
1395     if ((has_blk && blk) || (has_inc && inc)) {
1396         if (migrate_use_block() || migrate_use_block_incremental()) {
1397             error_setg(errp, "Command options are incompatible with "
1398                        "current migration capabilities");
1399             return;
1400         }
1401         migrate_set_block_enabled(true, &local_err);
1402         if (local_err) {
1403             error_propagate(errp, local_err);
1404             return;
1405         }
1406         s->must_remove_block_options = true;
1407     }
1408
1409     if (has_inc && inc) {
1410         migrate_set_block_incremental(s, true);
1411     }
1412
1413     s = migrate_init();
1414
1415     if (strstart(uri, "tcp:", &p)) {
1416         tcp_start_outgoing_migration(s, p, &local_err);
1417 #ifdef CONFIG_RDMA
1418     } else if (strstart(uri, "rdma:", &p)) {
1419         rdma_start_outgoing_migration(s, p, &local_err);
1420 #endif
1421     } else if (strstart(uri, "exec:", &p)) {
1422         exec_start_outgoing_migration(s, p, &local_err);
1423     } else if (strstart(uri, "unix:", &p)) {
1424         unix_start_outgoing_migration(s, p, &local_err);
1425     } else if (strstart(uri, "fd:", &p)) {
1426         fd_start_outgoing_migration(s, p, &local_err);
1427     } else {
1428         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri",
1429                    "a valid migration protocol");
1430         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1431                           MIGRATION_STATUS_FAILED);
1432         return;
1433     }
1434
1435     if (local_err) {
1436         migrate_fd_error(s, local_err);
1437         error_propagate(errp, local_err);
1438         return;
1439     }
1440 }
1441
1442 void qmp_migrate_cancel(Error **errp)
1443 {
1444     migrate_fd_cancel(migrate_get_current());
1445 }
1446
1447 void qmp_migrate_continue(MigrationStatus state, Error **errp)
1448 {
1449     MigrationState *s = migrate_get_current();
1450     if (s->state != state) {
1451         error_setg(errp,  "Migration not in expected state: %s",
1452                    MigrationStatus_str(s->state));
1453         return;
1454     }
1455     qemu_sem_post(&s->pause_sem);
1456 }
1457
1458 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
1459 {
1460     MigrateSetParameters p = {
1461         .has_xbzrle_cache_size = true,
1462         .xbzrle_cache_size = value,
1463     };
1464
1465     qmp_migrate_set_parameters(&p, errp);
1466 }
1467
1468 int64_t qmp_query_migrate_cache_size(Error **errp)
1469 {
1470     return migrate_xbzrle_cache_size();
1471 }
1472
1473 void qmp_migrate_set_speed(int64_t value, Error **errp)
1474 {
1475     MigrateSetParameters p = {
1476         .has_max_bandwidth = true,
1477         .max_bandwidth = value,
1478     };
1479
1480     qmp_migrate_set_parameters(&p, errp);
1481 }
1482
1483 void qmp_migrate_set_downtime(double value, Error **errp)
1484 {
1485     if (value < 0 || value > MAX_MIGRATE_DOWNTIME_SECONDS) {
1486         error_setg(errp, "Parameter 'downtime_limit' expects an integer in "
1487                          "the range of 0 to %d seconds",
1488                          MAX_MIGRATE_DOWNTIME_SECONDS);
1489         return;
1490     }
1491
1492     value *= 1000; /* Convert to milliseconds */
1493     value = MAX(0, MIN(INT64_MAX, value));
1494
1495     MigrateSetParameters p = {
1496         .has_downtime_limit = true,
1497         .downtime_limit = value,
1498     };
1499
1500     qmp_migrate_set_parameters(&p, errp);
1501 }
1502
1503 bool migrate_release_ram(void)
1504 {
1505     MigrationState *s;
1506
1507     s = migrate_get_current();
1508
1509     return s->enabled_capabilities[MIGRATION_CAPABILITY_RELEASE_RAM];
1510 }
1511
1512 bool migrate_postcopy_ram(void)
1513 {
1514     MigrationState *s;
1515
1516     s = migrate_get_current();
1517
1518     return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM];
1519 }
1520
1521 bool migrate_postcopy(void)
1522 {
1523     return migrate_postcopy_ram();
1524 }
1525
1526 bool migrate_auto_converge(void)
1527 {
1528     MigrationState *s;
1529
1530     s = migrate_get_current();
1531
1532     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
1533 }
1534
1535 bool migrate_zero_blocks(void)
1536 {
1537     MigrationState *s;
1538
1539     s = migrate_get_current();
1540
1541     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
1542 }
1543
1544 bool migrate_postcopy_blocktime(void)
1545 {
1546     MigrationState *s;
1547
1548     s = migrate_get_current();
1549
1550     return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_BLOCKTIME];
1551 }
1552
1553 bool migrate_use_compression(void)
1554 {
1555     MigrationState *s;
1556
1557     s = migrate_get_current();
1558
1559     return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS];
1560 }
1561
1562 int migrate_compress_level(void)
1563 {
1564     MigrationState *s;
1565
1566     s = migrate_get_current();
1567
1568     return s->parameters.compress_level;
1569 }
1570
1571 int migrate_compress_threads(void)
1572 {
1573     MigrationState *s;
1574
1575     s = migrate_get_current();
1576
1577     return s->parameters.compress_threads;
1578 }
1579
1580 int migrate_decompress_threads(void)
1581 {
1582     MigrationState *s;
1583
1584     s = migrate_get_current();
1585
1586     return s->parameters.decompress_threads;
1587 }
1588
1589 bool migrate_use_events(void)
1590 {
1591     MigrationState *s;
1592
1593     s = migrate_get_current();
1594
1595     return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS];
1596 }
1597
1598 bool migrate_use_multifd(void)
1599 {
1600     MigrationState *s;
1601
1602     s = migrate_get_current();
1603
1604     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_MULTIFD];
1605 }
1606
1607 bool migrate_pause_before_switchover(void)
1608 {
1609     MigrationState *s;
1610
1611     s = migrate_get_current();
1612
1613     return s->enabled_capabilities[
1614         MIGRATION_CAPABILITY_PAUSE_BEFORE_SWITCHOVER];
1615 }
1616
1617 int migrate_multifd_channels(void)
1618 {
1619     MigrationState *s;
1620
1621     s = migrate_get_current();
1622
1623     return s->parameters.x_multifd_channels;
1624 }
1625
1626 int migrate_multifd_page_count(void)
1627 {
1628     MigrationState *s;
1629
1630     s = migrate_get_current();
1631
1632     return s->parameters.x_multifd_page_count;
1633 }
1634
1635 int migrate_use_xbzrle(void)
1636 {
1637     MigrationState *s;
1638
1639     s = migrate_get_current();
1640
1641     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
1642 }
1643
1644 int64_t migrate_xbzrle_cache_size(void)
1645 {
1646     MigrationState *s;
1647
1648     s = migrate_get_current();
1649
1650     return s->parameters.xbzrle_cache_size;
1651 }
1652
1653 bool migrate_use_block(void)
1654 {
1655     MigrationState *s;
1656
1657     s = migrate_get_current();
1658
1659     return s->enabled_capabilities[MIGRATION_CAPABILITY_BLOCK];
1660 }
1661
1662 bool migrate_use_return_path(void)
1663 {
1664     MigrationState *s;
1665
1666     s = migrate_get_current();
1667
1668     return s->enabled_capabilities[MIGRATION_CAPABILITY_RETURN_PATH];
1669 }
1670
1671 bool migrate_use_block_incremental(void)
1672 {
1673     MigrationState *s;
1674
1675     s = migrate_get_current();
1676
1677     return s->parameters.block_incremental;
1678 }
1679
1680 /* migration thread support */
1681 /*
1682  * Something bad happened to the RP stream, mark an error
1683  * The caller shall print or trace something to indicate why
1684  */
1685 static void mark_source_rp_bad(MigrationState *s)
1686 {
1687     s->rp_state.error = true;
1688 }
1689
1690 static struct rp_cmd_args {
1691     ssize_t     len; /* -1 = variable */
1692     const char *name;
1693 } rp_cmd_args[] = {
1694     [MIG_RP_MSG_INVALID]        = { .len = -1, .name = "INVALID" },
1695     [MIG_RP_MSG_SHUT]           = { .len =  4, .name = "SHUT" },
1696     [MIG_RP_MSG_PONG]           = { .len =  4, .name = "PONG" },
1697     [MIG_RP_MSG_REQ_PAGES]      = { .len = 12, .name = "REQ_PAGES" },
1698     [MIG_RP_MSG_REQ_PAGES_ID]   = { .len = -1, .name = "REQ_PAGES_ID" },
1699     [MIG_RP_MSG_MAX]            = { .len = -1, .name = "MAX" },
1700 };
1701
1702 /*
1703  * Process a request for pages received on the return path,
1704  * We're allowed to send more than requested (e.g. to round to our page size)
1705  * and we don't need to send pages that have already been sent.
1706  */
1707 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname,
1708                                        ram_addr_t start, size_t len)
1709 {
1710     long our_host_ps = getpagesize();
1711
1712     trace_migrate_handle_rp_req_pages(rbname, start, len);
1713
1714     /*
1715      * Since we currently insist on matching page sizes, just sanity check
1716      * we're being asked for whole host pages.
1717      */
1718     if (start & (our_host_ps-1) ||
1719        (len & (our_host_ps-1))) {
1720         error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1721                      " len: %zd", __func__, start, len);
1722         mark_source_rp_bad(ms);
1723         return;
1724     }
1725
1726     if (ram_save_queue_pages(rbname, start, len)) {
1727         mark_source_rp_bad(ms);
1728     }
1729 }
1730
1731 /*
1732  * Handles messages sent on the return path towards the source VM
1733  *
1734  */
1735 static void *source_return_path_thread(void *opaque)
1736 {
1737     MigrationState *ms = opaque;
1738     QEMUFile *rp = ms->rp_state.from_dst_file;
1739     uint16_t header_len, header_type;
1740     uint8_t buf[512];
1741     uint32_t tmp32, sibling_error;
1742     ram_addr_t start = 0; /* =0 to silence warning */
1743     size_t  len = 0, expected_len;
1744     int res;
1745
1746     trace_source_return_path_thread_entry();
1747     while (!ms->rp_state.error && !qemu_file_get_error(rp) &&
1748            migration_is_setup_or_active(ms->state)) {
1749         trace_source_return_path_thread_loop_top();
1750         header_type = qemu_get_be16(rp);
1751         header_len = qemu_get_be16(rp);
1752
1753         if (header_type >= MIG_RP_MSG_MAX ||
1754             header_type == MIG_RP_MSG_INVALID) {
1755             error_report("RP: Received invalid message 0x%04x length 0x%04x",
1756                     header_type, header_len);
1757             mark_source_rp_bad(ms);
1758             goto out;
1759         }
1760
1761         if ((rp_cmd_args[header_type].len != -1 &&
1762             header_len != rp_cmd_args[header_type].len) ||
1763             header_len > sizeof(buf)) {
1764             error_report("RP: Received '%s' message (0x%04x) with"
1765                     "incorrect length %d expecting %zu",
1766                     rp_cmd_args[header_type].name, header_type, header_len,
1767                     (size_t)rp_cmd_args[header_type].len);
1768             mark_source_rp_bad(ms);
1769             goto out;
1770         }
1771
1772         /* We know we've got a valid header by this point */
1773         res = qemu_get_buffer(rp, buf, header_len);
1774         if (res != header_len) {
1775             error_report("RP: Failed reading data for message 0x%04x"
1776                          " read %d expected %d",
1777                          header_type, res, header_len);
1778             mark_source_rp_bad(ms);
1779             goto out;
1780         }
1781
1782         /* OK, we have the message and the data */
1783         switch (header_type) {
1784         case MIG_RP_MSG_SHUT:
1785             sibling_error = ldl_be_p(buf);
1786             trace_source_return_path_thread_shut(sibling_error);
1787             if (sibling_error) {
1788                 error_report("RP: Sibling indicated error %d", sibling_error);
1789                 mark_source_rp_bad(ms);
1790             }
1791             /*
1792              * We'll let the main thread deal with closing the RP
1793              * we could do a shutdown(2) on it, but we're the only user
1794              * anyway, so there's nothing gained.
1795              */
1796             goto out;
1797
1798         case MIG_RP_MSG_PONG:
1799             tmp32 = ldl_be_p(buf);
1800             trace_source_return_path_thread_pong(tmp32);
1801             break;
1802
1803         case MIG_RP_MSG_REQ_PAGES:
1804             start = ldq_be_p(buf);
1805             len = ldl_be_p(buf + 8);
1806             migrate_handle_rp_req_pages(ms, NULL, start, len);
1807             break;
1808
1809         case MIG_RP_MSG_REQ_PAGES_ID:
1810             expected_len = 12 + 1; /* header + termination */
1811
1812             if (header_len >= expected_len) {
1813                 start = ldq_be_p(buf);
1814                 len = ldl_be_p(buf + 8);
1815                 /* Now we expect an idstr */
1816                 tmp32 = buf[12]; /* Length of the following idstr */
1817                 buf[13 + tmp32] = '\0';
1818                 expected_len += tmp32;
1819             }
1820             if (header_len != expected_len) {
1821                 error_report("RP: Req_Page_id with length %d expecting %zd",
1822                         header_len, expected_len);
1823                 mark_source_rp_bad(ms);
1824                 goto out;
1825             }
1826             migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len);
1827             break;
1828
1829         default:
1830             break;
1831         }
1832     }
1833     if (qemu_file_get_error(rp)) {
1834         trace_source_return_path_thread_bad_end();
1835         mark_source_rp_bad(ms);
1836     }
1837
1838     trace_source_return_path_thread_end();
1839 out:
1840     ms->rp_state.from_dst_file = NULL;
1841     qemu_fclose(rp);
1842     return NULL;
1843 }
1844
1845 static int open_return_path_on_source(MigrationState *ms)
1846 {
1847
1848     ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file);
1849     if (!ms->rp_state.from_dst_file) {
1850         return -1;
1851     }
1852
1853     trace_open_return_path_on_source();
1854     qemu_thread_create(&ms->rp_state.rp_thread, "return path",
1855                        source_return_path_thread, ms, QEMU_THREAD_JOINABLE);
1856
1857     trace_open_return_path_on_source_continue();
1858
1859     return 0;
1860 }
1861
1862 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
1863 static int await_return_path_close_on_source(MigrationState *ms)
1864 {
1865     /*
1866      * If this is a normal exit then the destination will send a SHUT and the
1867      * rp_thread will exit, however if there's an error we need to cause
1868      * it to exit.
1869      */
1870     if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
1871         /*
1872          * shutdown(2), if we have it, will cause it to unblock if it's stuck
1873          * waiting for the destination.
1874          */
1875         qemu_file_shutdown(ms->rp_state.from_dst_file);
1876         mark_source_rp_bad(ms);
1877     }
1878     trace_await_return_path_close_on_source_joining();
1879     qemu_thread_join(&ms->rp_state.rp_thread);
1880     trace_await_return_path_close_on_source_close();
1881     return ms->rp_state.error;
1882 }
1883
1884 /*
1885  * Switch from normal iteration to postcopy
1886  * Returns non-0 on error
1887  */
1888 static int postcopy_start(MigrationState *ms, bool *old_vm_running)
1889 {
1890     int ret;
1891     QIOChannelBuffer *bioc;
1892     QEMUFile *fb;
1893     int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1894     bool restart_block = false;
1895     int cur_state = MIGRATION_STATUS_ACTIVE;
1896     if (!migrate_pause_before_switchover()) {
1897         migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE,
1898                           MIGRATION_STATUS_POSTCOPY_ACTIVE);
1899     }
1900
1901     trace_postcopy_start();
1902     qemu_mutex_lock_iothread();
1903     trace_postcopy_start_set_run();
1904
1905     qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1906     *old_vm_running = runstate_is_running();
1907     global_state_store();
1908     ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1909     if (ret < 0) {
1910         goto fail;
1911     }
1912
1913     ret = migration_maybe_pause(ms, &cur_state,
1914                                 MIGRATION_STATUS_POSTCOPY_ACTIVE);
1915     if (ret < 0) {
1916         goto fail;
1917     }
1918
1919     ret = bdrv_inactivate_all();
1920     if (ret < 0) {
1921         goto fail;
1922     }
1923     restart_block = true;
1924
1925     /*
1926      * Cause any non-postcopiable, but iterative devices to
1927      * send out their final data.
1928      */
1929     qemu_savevm_state_complete_precopy(ms->to_dst_file, true, false);
1930
1931     /*
1932      * in Finish migrate and with the io-lock held everything should
1933      * be quiet, but we've potentially still got dirty pages and we
1934      * need to tell the destination to throw any pages it's already received
1935      * that are dirty
1936      */
1937     if (migrate_postcopy_ram()) {
1938         if (ram_postcopy_send_discard_bitmap(ms)) {
1939             error_report("postcopy send discard bitmap failed");
1940             goto fail;
1941         }
1942     }
1943
1944     /*
1945      * send rest of state - note things that are doing postcopy
1946      * will notice we're in POSTCOPY_ACTIVE and not actually
1947      * wrap their state up here
1948      */
1949     qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX);
1950     if (migrate_postcopy_ram()) {
1951         /* Ping just for debugging, helps line traces up */
1952         qemu_savevm_send_ping(ms->to_dst_file, 2);
1953     }
1954
1955     /*
1956      * While loading the device state we may trigger page transfer
1957      * requests and the fd must be free to process those, and thus
1958      * the destination must read the whole device state off the fd before
1959      * it starts processing it.  Unfortunately the ad-hoc migration format
1960      * doesn't allow the destination to know the size to read without fully
1961      * parsing it through each devices load-state code (especially the open
1962      * coded devices that use get/put).
1963      * So we wrap the device state up in a package with a length at the start;
1964      * to do this we use a qemu_buf to hold the whole of the device state.
1965      */
1966     bioc = qio_channel_buffer_new(4096);
1967     qio_channel_set_name(QIO_CHANNEL(bioc), "migration-postcopy-buffer");
1968     fb = qemu_fopen_channel_output(QIO_CHANNEL(bioc));
1969     object_unref(OBJECT(bioc));
1970
1971     /*
1972      * Make sure the receiver can get incoming pages before we send the rest
1973      * of the state
1974      */
1975     qemu_savevm_send_postcopy_listen(fb);
1976
1977     qemu_savevm_state_complete_precopy(fb, false, false);
1978     if (migrate_postcopy_ram()) {
1979         qemu_savevm_send_ping(fb, 3);
1980     }
1981
1982     qemu_savevm_send_postcopy_run(fb);
1983
1984     /* <><> end of stuff going into the package */
1985
1986     /* Last point of recovery; as soon as we send the package the destination
1987      * can open devices and potentially start running.
1988      * Lets just check again we've not got any errors.
1989      */
1990     ret = qemu_file_get_error(ms->to_dst_file);
1991     if (ret) {
1992         error_report("postcopy_start: Migration stream errored (pre package)");
1993         goto fail_closefb;
1994     }
1995
1996     restart_block = false;
1997
1998     /* Now send that blob */
1999     if (qemu_savevm_send_packaged(ms->to_dst_file, bioc->data, bioc->usage)) {
2000         goto fail_closefb;
2001     }
2002     qemu_fclose(fb);
2003
2004     /* Send a notify to give a chance for anything that needs to happen
2005      * at the transition to postcopy and after the device state; in particular
2006      * spice needs to trigger a transition now
2007      */
2008     ms->postcopy_after_devices = true;
2009     notifier_list_notify(&migration_state_notifiers, ms);
2010
2011     ms->downtime =  qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop;
2012
2013     qemu_mutex_unlock_iothread();
2014
2015     if (migrate_postcopy_ram()) {
2016         /*
2017          * Although this ping is just for debug, it could potentially be
2018          * used for getting a better measurement of downtime at the source.
2019          */
2020         qemu_savevm_send_ping(ms->to_dst_file, 4);
2021     }
2022
2023     if (migrate_release_ram()) {
2024         ram_postcopy_migrated_memory_release(ms);
2025     }
2026
2027     ret = qemu_file_get_error(ms->to_dst_file);
2028     if (ret) {
2029         error_report("postcopy_start: Migration stream errored");
2030         migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
2031                               MIGRATION_STATUS_FAILED);
2032     }
2033
2034     return ret;
2035
2036 fail_closefb:
2037     qemu_fclose(fb);
2038 fail:
2039     migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
2040                           MIGRATION_STATUS_FAILED);
2041     if (restart_block) {
2042         /* A failure happened early enough that we know the destination hasn't
2043          * accessed block devices, so we're safe to recover.
2044          */
2045         Error *local_err = NULL;
2046
2047         bdrv_invalidate_cache_all(&local_err);
2048         if (local_err) {
2049             error_report_err(local_err);
2050         }
2051     }
2052     qemu_mutex_unlock_iothread();
2053     return -1;
2054 }
2055
2056 /**
2057  * migration_maybe_pause: Pause if required to by
2058  * migrate_pause_before_switchover called with the iothread locked
2059  * Returns: 0 on success
2060  */
2061 static int migration_maybe_pause(MigrationState *s,
2062                                  int *current_active_state,
2063                                  int new_state)
2064 {
2065     if (!migrate_pause_before_switchover()) {
2066         return 0;
2067     }
2068
2069     /* Since leaving this state is not atomic with posting the semaphore
2070      * it's possible that someone could have issued multiple migrate_continue
2071      * and the semaphore is incorrectly positive at this point;
2072      * the docs say it's undefined to reinit a semaphore that's already
2073      * init'd, so use timedwait to eat up any existing posts.
2074      */
2075     while (qemu_sem_timedwait(&s->pause_sem, 1) == 0) {
2076         /* This block intentionally left blank */
2077     }
2078
2079     qemu_mutex_unlock_iothread();
2080     migrate_set_state(&s->state, *current_active_state,
2081                       MIGRATION_STATUS_PRE_SWITCHOVER);
2082     qemu_sem_wait(&s->pause_sem);
2083     migrate_set_state(&s->state, MIGRATION_STATUS_PRE_SWITCHOVER,
2084                       new_state);
2085     *current_active_state = new_state;
2086     qemu_mutex_lock_iothread();
2087
2088     return s->state == new_state ? 0 : -EINVAL;
2089 }
2090
2091 /**
2092  * migration_completion: Used by migration_thread when there's not much left.
2093  *   The caller 'breaks' the loop when this returns.
2094  *
2095  * @s: Current migration state
2096  * @current_active_state: The migration state we expect to be in
2097  * @*old_vm_running: Pointer to old_vm_running flag
2098  * @*start_time: Pointer to time to update
2099  */
2100 static void migration_completion(MigrationState *s, int current_active_state,
2101                                  bool *old_vm_running,
2102                                  int64_t *start_time)
2103 {
2104     int ret;
2105
2106     if (s->state == MIGRATION_STATUS_ACTIVE) {
2107         qemu_mutex_lock_iothread();
2108         *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2109         qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
2110         *old_vm_running = runstate_is_running();
2111         ret = global_state_store();
2112
2113         if (!ret) {
2114             bool inactivate = !migrate_colo_enabled();
2115             ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
2116             if (ret >= 0) {
2117                 ret = migration_maybe_pause(s, &current_active_state,
2118                                             MIGRATION_STATUS_DEVICE);
2119             }
2120             if (ret >= 0) {
2121                 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX);
2122                 ret = qemu_savevm_state_complete_precopy(s->to_dst_file, false,
2123                                                          inactivate);
2124             }
2125             if (inactivate && ret >= 0) {
2126                 s->block_inactive = true;
2127             }
2128         }
2129         qemu_mutex_unlock_iothread();
2130
2131         if (ret < 0) {
2132             goto fail;
2133         }
2134     } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
2135         trace_migration_completion_postcopy_end();
2136
2137         qemu_savevm_state_complete_postcopy(s->to_dst_file);
2138         trace_migration_completion_postcopy_end_after_complete();
2139     }
2140
2141     /*
2142      * If rp was opened we must clean up the thread before
2143      * cleaning everything else up (since if there are no failures
2144      * it will wait for the destination to send it's status in
2145      * a SHUT command).
2146      */
2147     if (s->rp_state.from_dst_file) {
2148         int rp_error;
2149         trace_migration_return_path_end_before();
2150         rp_error = await_return_path_close_on_source(s);
2151         trace_migration_return_path_end_after(rp_error);
2152         if (rp_error) {
2153             goto fail_invalidate;
2154         }
2155     }
2156
2157     if (qemu_file_get_error(s->to_dst_file)) {
2158         trace_migration_completion_file_err();
2159         goto fail_invalidate;
2160     }
2161
2162     if (!migrate_colo_enabled()) {
2163         migrate_set_state(&s->state, current_active_state,
2164                           MIGRATION_STATUS_COMPLETED);
2165     }
2166
2167     return;
2168
2169 fail_invalidate:
2170     /* If not doing postcopy, vm_start() will be called: let's regain
2171      * control on images.
2172      */
2173     if (s->state == MIGRATION_STATUS_ACTIVE) {
2174         Error *local_err = NULL;
2175
2176         qemu_mutex_lock_iothread();
2177         bdrv_invalidate_cache_all(&local_err);
2178         if (local_err) {
2179             error_report_err(local_err);
2180         } else {
2181             s->block_inactive = false;
2182         }
2183         qemu_mutex_unlock_iothread();
2184     }
2185
2186 fail:
2187     migrate_set_state(&s->state, current_active_state,
2188                       MIGRATION_STATUS_FAILED);
2189 }
2190
2191 bool migrate_colo_enabled(void)
2192 {
2193     MigrationState *s = migrate_get_current();
2194     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_COLO];
2195 }
2196
2197 /*
2198  * Master migration thread on the source VM.
2199  * It drives the migration and pumps the data down the outgoing channel.
2200  */
2201 static void *migration_thread(void *opaque)
2202 {
2203     MigrationState *s = opaque;
2204     /* Used by the bandwidth calcs, updated later */
2205     int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2206     int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
2207     int64_t initial_bytes = 0;
2208     /*
2209      * The final stage happens when the remaining data is smaller than
2210      * this threshold; it's calculated from the requested downtime and
2211      * measured bandwidth
2212      */
2213     int64_t threshold_size = 0;
2214     int64_t start_time = initial_time;
2215     int64_t end_time;
2216     bool old_vm_running = false;
2217     bool entered_postcopy = false;
2218     /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */
2219     enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE;
2220
2221     rcu_register_thread();
2222
2223     qemu_savevm_state_header(s->to_dst_file);
2224
2225     /*
2226      * If we opened the return path, we need to make sure dst has it
2227      * opened as well.
2228      */
2229     if (s->rp_state.from_dst_file) {
2230         /* Now tell the dest that it should open its end so it can reply */
2231         qemu_savevm_send_open_return_path(s->to_dst_file);
2232
2233         /* And do a ping that will make stuff easier to debug */
2234         qemu_savevm_send_ping(s->to_dst_file, 1);
2235     }
2236
2237     if (migrate_postcopy()) {
2238         /*
2239          * Tell the destination that we *might* want to do postcopy later;
2240          * if the other end can't do postcopy it should fail now, nice and
2241          * early.
2242          */
2243         qemu_savevm_send_postcopy_advise(s->to_dst_file);
2244     }
2245
2246     qemu_savevm_state_setup(s->to_dst_file);
2247
2248     s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
2249     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2250                       MIGRATION_STATUS_ACTIVE);
2251
2252     trace_migration_thread_setup_complete();
2253
2254     while (s->state == MIGRATION_STATUS_ACTIVE ||
2255            s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
2256         int64_t current_time;
2257         uint64_t pending_size;
2258
2259         if (!qemu_file_rate_limit(s->to_dst_file)) {
2260             uint64_t pend_post, pend_nonpost;
2261
2262             qemu_savevm_state_pending(s->to_dst_file, threshold_size,
2263                                       &pend_nonpost, &pend_post);
2264             pending_size = pend_nonpost + pend_post;
2265             trace_migrate_pending(pending_size, threshold_size,
2266                                   pend_post, pend_nonpost);
2267             if (pending_size && pending_size >= threshold_size) {
2268                 /* Still a significant amount to transfer */
2269
2270                 if (migrate_postcopy() &&
2271                     s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE &&
2272                     pend_nonpost <= threshold_size &&
2273                     atomic_read(&s->start_postcopy)) {
2274
2275                     if (!postcopy_start(s, &old_vm_running)) {
2276                         current_active_state = MIGRATION_STATUS_POSTCOPY_ACTIVE;
2277                         entered_postcopy = true;
2278                     }
2279
2280                     continue;
2281                 }
2282                 /* Just another iteration step */
2283                 qemu_savevm_state_iterate(s->to_dst_file, entered_postcopy);
2284             } else {
2285                 trace_migration_thread_low_pending(pending_size);
2286                 migration_completion(s, current_active_state,
2287                                      &old_vm_running, &start_time);
2288                 break;
2289             }
2290         }
2291
2292         if (qemu_file_get_error(s->to_dst_file)) {
2293             migrate_set_state(&s->state, current_active_state,
2294                               MIGRATION_STATUS_FAILED);
2295             trace_migration_thread_file_err();
2296             break;
2297         }
2298         current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2299         if (current_time >= initial_time + BUFFER_DELAY) {
2300             uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) -
2301                                          initial_bytes;
2302             uint64_t time_spent = current_time - initial_time;
2303             double bandwidth = (double)transferred_bytes / time_spent;
2304             threshold_size = bandwidth * s->parameters.downtime_limit;
2305
2306             s->mbps = (((double) transferred_bytes * 8.0) /
2307                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0;
2308
2309             trace_migrate_transferred(transferred_bytes, time_spent,
2310                                       bandwidth, threshold_size);
2311             /* if we haven't sent anything, we don't want to recalculate
2312                10000 is a small enough number for our purposes */
2313             if (ram_counters.dirty_pages_rate && transferred_bytes > 10000) {
2314                 s->expected_downtime = ram_counters.dirty_pages_rate *
2315                     qemu_target_page_size() / bandwidth;
2316             }
2317
2318             qemu_file_reset_rate_limit(s->to_dst_file);
2319             initial_time = current_time;
2320             initial_bytes = qemu_ftell(s->to_dst_file);
2321         }
2322         if (qemu_file_rate_limit(s->to_dst_file)) {
2323             /* usleep expects microseconds */
2324             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
2325         }
2326     }
2327
2328     trace_migration_thread_after_loop();
2329     /* If we enabled cpu throttling for auto-converge, turn it off. */
2330     cpu_throttle_stop();
2331     end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2332
2333     qemu_mutex_lock_iothread();
2334     if (s->state == MIGRATION_STATUS_COMPLETED) {
2335         uint64_t transferred_bytes = qemu_ftell(s->to_dst_file);
2336         s->total_time = end_time - s->start_time;
2337         if (!entered_postcopy) {
2338             s->downtime = end_time - start_time;
2339         }
2340         if (s->total_time) {
2341             s->mbps = (((double) transferred_bytes * 8.0) /
2342                        ((double) s->total_time)) / 1000;
2343         }
2344         runstate_set(RUN_STATE_POSTMIGRATE);
2345     } else {
2346         if (s->state == MIGRATION_STATUS_ACTIVE) {
2347             /*
2348              * We should really assert here, but since it's during
2349              * migration, let's try to reduce the usage of assertions.
2350              */
2351             if (!migrate_colo_enabled()) {
2352                 error_report("%s: critical error: calling COLO code without "
2353                              "COLO enabled", __func__);
2354             }
2355             migrate_start_colo_process(s);
2356             /*
2357             * Fixme: we will run VM in COLO no matter its old running state.
2358             * After exited COLO, we will keep running.
2359             */
2360             old_vm_running = true;
2361         }
2362         if (old_vm_running && !entered_postcopy) {
2363             vm_start();
2364         } else {
2365             if (runstate_check(RUN_STATE_FINISH_MIGRATE)) {
2366                 runstate_set(RUN_STATE_POSTMIGRATE);
2367             }
2368         }
2369     }
2370     qemu_bh_schedule(s->cleanup_bh);
2371     qemu_mutex_unlock_iothread();
2372
2373     rcu_unregister_thread();
2374     return NULL;
2375 }
2376
2377 void migrate_fd_connect(MigrationState *s)
2378 {
2379     s->expected_downtime = s->parameters.downtime_limit;
2380     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
2381
2382     qemu_file_set_blocking(s->to_dst_file, true);
2383     qemu_file_set_rate_limit(s->to_dst_file,
2384                              s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
2385
2386     /* Notify before starting migration thread */
2387     notifier_list_notify(&migration_state_notifiers, s);
2388
2389     /*
2390      * Open the return path. For postcopy, it is used exclusively. For
2391      * precopy, only if user specified "return-path" capability would
2392      * QEMU uses the return path.
2393      */
2394     if (migrate_postcopy_ram() || migrate_use_return_path()) {
2395         if (open_return_path_on_source(s)) {
2396             error_report("Unable to open return-path for postcopy");
2397             migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2398                               MIGRATION_STATUS_FAILED);
2399             migrate_fd_cleanup(s);
2400             return;
2401         }
2402     }
2403
2404     if (multifd_save_setup() != 0) {
2405         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2406                           MIGRATION_STATUS_FAILED);
2407         migrate_fd_cleanup(s);
2408         return;
2409     }
2410     qemu_thread_create(&s->thread, "live_migration", migration_thread, s,
2411                        QEMU_THREAD_JOINABLE);
2412     s->migration_thread_running = true;
2413 }
2414
2415 void migration_global_dump(Monitor *mon)
2416 {
2417     MigrationState *ms = migrate_get_current();
2418
2419     monitor_printf(mon, "globals:\n");
2420     monitor_printf(mon, "store-global-state: %s\n",
2421                    ms->store_global_state ? "on" : "off");
2422     monitor_printf(mon, "only-migratable: %s\n",
2423                    ms->only_migratable ? "on" : "off");
2424     monitor_printf(mon, "send-configuration: %s\n",
2425                    ms->send_configuration ? "on" : "off");
2426     monitor_printf(mon, "send-section-footer: %s\n",
2427                    ms->send_section_footer ? "on" : "off");
2428 }
2429
2430 #define DEFINE_PROP_MIG_CAP(name, x)             \
2431     DEFINE_PROP_BOOL(name, MigrationState, enabled_capabilities[x], false)
2432
2433 static Property migration_properties[] = {
2434     DEFINE_PROP_BOOL("store-global-state", MigrationState,
2435                      store_global_state, true),
2436     DEFINE_PROP_BOOL("only-migratable", MigrationState, only_migratable, false),
2437     DEFINE_PROP_BOOL("send-configuration", MigrationState,
2438                      send_configuration, true),
2439     DEFINE_PROP_BOOL("send-section-footer", MigrationState,
2440                      send_section_footer, true),
2441
2442     /* Migration parameters */
2443     DEFINE_PROP_UINT8("x-compress-level", MigrationState,
2444                       parameters.compress_level,
2445                       DEFAULT_MIGRATE_COMPRESS_LEVEL),
2446     DEFINE_PROP_UINT8("x-compress-threads", MigrationState,
2447                       parameters.compress_threads,
2448                       DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT),
2449     DEFINE_PROP_UINT8("x-decompress-threads", MigrationState,
2450                       parameters.decompress_threads,
2451                       DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT),
2452     DEFINE_PROP_UINT8("x-cpu-throttle-initial", MigrationState,
2453                       parameters.cpu_throttle_initial,
2454                       DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL),
2455     DEFINE_PROP_UINT8("x-cpu-throttle-increment", MigrationState,
2456                       parameters.cpu_throttle_increment,
2457                       DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT),
2458     DEFINE_PROP_SIZE("x-max-bandwidth", MigrationState,
2459                       parameters.max_bandwidth, MAX_THROTTLE),
2460     DEFINE_PROP_UINT64("x-downtime-limit", MigrationState,
2461                       parameters.downtime_limit,
2462                       DEFAULT_MIGRATE_SET_DOWNTIME),
2463     DEFINE_PROP_UINT32("x-checkpoint-delay", MigrationState,
2464                       parameters.x_checkpoint_delay,
2465                       DEFAULT_MIGRATE_X_CHECKPOINT_DELAY),
2466     DEFINE_PROP_UINT8("x-multifd-channels", MigrationState,
2467                       parameters.x_multifd_channels,
2468                       DEFAULT_MIGRATE_MULTIFD_CHANNELS),
2469     DEFINE_PROP_UINT32("x-multifd-page-count", MigrationState,
2470                       parameters.x_multifd_page_count,
2471                       DEFAULT_MIGRATE_MULTIFD_PAGE_COUNT),
2472     DEFINE_PROP_SIZE("xbzrle-cache-size", MigrationState,
2473                       parameters.xbzrle_cache_size,
2474                       DEFAULT_MIGRATE_XBZRLE_CACHE_SIZE),
2475
2476     /* Migration capabilities */
2477     DEFINE_PROP_MIG_CAP("x-xbzrle", MIGRATION_CAPABILITY_XBZRLE),
2478     DEFINE_PROP_MIG_CAP("x-rdma-pin-all", MIGRATION_CAPABILITY_RDMA_PIN_ALL),
2479     DEFINE_PROP_MIG_CAP("x-auto-converge", MIGRATION_CAPABILITY_AUTO_CONVERGE),
2480     DEFINE_PROP_MIG_CAP("x-zero-blocks", MIGRATION_CAPABILITY_ZERO_BLOCKS),
2481     DEFINE_PROP_MIG_CAP("x-compress", MIGRATION_CAPABILITY_COMPRESS),
2482     DEFINE_PROP_MIG_CAP("x-events", MIGRATION_CAPABILITY_EVENTS),
2483     DEFINE_PROP_MIG_CAP("x-postcopy-ram", MIGRATION_CAPABILITY_POSTCOPY_RAM),
2484     DEFINE_PROP_MIG_CAP("x-colo", MIGRATION_CAPABILITY_X_COLO),
2485     DEFINE_PROP_MIG_CAP("x-release-ram", MIGRATION_CAPABILITY_RELEASE_RAM),
2486     DEFINE_PROP_MIG_CAP("x-block", MIGRATION_CAPABILITY_BLOCK),
2487     DEFINE_PROP_MIG_CAP("x-return-path", MIGRATION_CAPABILITY_RETURN_PATH),
2488     DEFINE_PROP_MIG_CAP("x-multifd", MIGRATION_CAPABILITY_X_MULTIFD),
2489
2490     DEFINE_PROP_END_OF_LIST(),
2491 };
2492
2493 static void migration_class_init(ObjectClass *klass, void *data)
2494 {
2495     DeviceClass *dc = DEVICE_CLASS(klass);
2496
2497     dc->user_creatable = false;
2498     dc->props = migration_properties;
2499 }
2500
2501 static void migration_instance_finalize(Object *obj)
2502 {
2503     MigrationState *ms = MIGRATION_OBJ(obj);
2504     MigrationParameters *params = &ms->parameters;
2505
2506     qemu_mutex_destroy(&ms->error_mutex);
2507     g_free(params->tls_hostname);
2508     g_free(params->tls_creds);
2509     qemu_sem_destroy(&ms->pause_sem);
2510 }
2511
2512 static void migration_instance_init(Object *obj)
2513 {
2514     MigrationState *ms = MIGRATION_OBJ(obj);
2515     MigrationParameters *params = &ms->parameters;
2516
2517     ms->state = MIGRATION_STATUS_NONE;
2518     ms->mbps = -1;
2519     qemu_sem_init(&ms->pause_sem, 0);
2520     qemu_mutex_init(&ms->error_mutex);
2521
2522     params->tls_hostname = g_strdup("");
2523     params->tls_creds = g_strdup("");
2524
2525     /* Set has_* up only for parameter checks */
2526     params->has_compress_level = true;
2527     params->has_compress_threads = true;
2528     params->has_decompress_threads = true;
2529     params->has_cpu_throttle_initial = true;
2530     params->has_cpu_throttle_increment = true;
2531     params->has_max_bandwidth = true;
2532     params->has_downtime_limit = true;
2533     params->has_x_checkpoint_delay = true;
2534     params->has_block_incremental = true;
2535     params->has_x_multifd_channels = true;
2536     params->has_x_multifd_page_count = true;
2537     params->has_xbzrle_cache_size = true;
2538 }
2539
2540 /*
2541  * Return true if check pass, false otherwise. Error will be put
2542  * inside errp if provided.
2543  */
2544 static bool migration_object_check(MigrationState *ms, Error **errp)
2545 {
2546     MigrationCapabilityStatusList *head = NULL;
2547     /* Assuming all off */
2548     bool cap_list[MIGRATION_CAPABILITY__MAX] = { 0 }, ret;
2549     int i;
2550
2551     if (!migrate_params_check(&ms->parameters, errp)) {
2552         return false;
2553     }
2554
2555     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
2556         if (ms->enabled_capabilities[i]) {
2557             head = migrate_cap_add(head, i, true);
2558         }
2559     }
2560
2561     ret = migrate_caps_check(cap_list, head, errp);
2562
2563     /* It works with head == NULL */
2564     qapi_free_MigrationCapabilityStatusList(head);
2565
2566     return ret;
2567 }
2568
2569 static const TypeInfo migration_type = {
2570     .name = TYPE_MIGRATION,
2571     /*
2572      * NOTE: TYPE_MIGRATION is not really a device, as the object is
2573      * not created using qdev_create(), it is not attached to the qdev
2574      * device tree, and it is never realized.
2575      *
2576      * TODO: Make this TYPE_OBJECT once QOM provides something like
2577      * TYPE_DEVICE's "-global" properties.
2578      */
2579     .parent = TYPE_DEVICE,
2580     .class_init = migration_class_init,
2581     .class_size = sizeof(MigrationClass),
2582     .instance_size = sizeof(MigrationState),
2583     .instance_init = migration_instance_init,
2584     .instance_finalize = migration_instance_finalize,
2585 };
2586
2587 static void register_migration_types(void)
2588 {
2589     type_register_static(&migration_type);
2590 }
2591
2592 type_init(register_migration_types);