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_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
 108 void migration_object_init(void)
 109 {
 110     MachineState *ms = MACHINE(qdev_get_machine());
 111     Error *err = NULL;
 112
 113     /* This can only be called once. */
 114     assert(!current_migration);
 115     current_migration = MIGRATION_OBJ(object_new(TYPE_MIGRATION));
 116
 117     if (!migration_object_check(current_migration, &err)) {
 118         error_report_err(err);
 119         exit(1);
 120     }
 121
 122     /*
 123      * We cannot really do this in migration_instance_init() since at
 124      * that time global properties are not yet applied, then this
 125      * value will be definitely replaced by something else.
 126      */
 127     if (ms->enforce_config_section) {
 128         current_migration->send_configuration = true;
 129     }
 130 }
 131
 132 /* For outgoing */
 133 MigrationState *migrate_get_current(void)
 134 {
 135     /* This can only be called after the object created. */
 136     assert(current_migration);
 137     return current_migration;
 138 }
 139
 140 MigrationIncomingState *migration_incoming_get_current(void)
 141 {
 142     static bool once;
 143     static MigrationIncomingState mis_current;
 144
 145     if (!once) {
 146         mis_current.state = MIGRATION_STATUS_NONE;
 147         memset(&mis_current, 0, sizeof(MigrationIncomingState));
 148         qemu_mutex_init(&mis_current.rp_mutex);
 149         qemu_event_init(&mis_current.main_thread_load_event, false);
 150         once = true;
 151     }
 152     return &mis_current;
 153 }
 154
 155 void migration_incoming_state_destroy(void)
 156 {
 157     struct MigrationIncomingState *mis = migration_incoming_get_current();
 158
 159     if (mis->to_src_file) {
 160         /* Tell source that we are done */
 161         migrate_send_rp_shut(mis, qemu_file_get_error(mis->from_src_file) != 0);
 162         qemu_fclose(mis->to_src_file);
 163         mis->to_src_file = NULL;
 164     }
 165
 166     if (mis->from_src_file) {
 167         qemu_fclose(mis->from_src_file);
 168         mis->from_src_file = NULL;
 169     }
 170
 171     qemu_event_reset(&mis->main_thread_load_event);
 172 }
 173
 174 static void migrate_generate_event(int new_state)
 175 {
 176     if (migrate_use_events()) {
 177         qapi_event_send_migration(new_state, &error_abort);
 178     }
 179 }
 180
 181 /*
 182  * Called on -incoming with a defer: uri.
 183  * The migration can be started later after any parameters have been
 184  * changed.
 185  */
 186 static void deferred_incoming_migration(Error **errp)
 187 {
 188     if (deferred_incoming) {
 189         error_setg(errp, "Incoming migration already deferred");
 190     }
 191     deferred_incoming = true;
 192 }
 193
 194 /*
 195  * Send a message on the return channel back to the source
 196  * of the migration.
 197  */
 198 static void migrate_send_rp_message(MigrationIncomingState *mis,
 199                                     enum mig_rp_message_type message_type,
 200                                     uint16_t len, void *data)
 201 {
 202     trace_migrate_send_rp_message((int)message_type, len);
 203     qemu_mutex_lock(&mis->rp_mutex);
 204     qemu_put_be16(mis->to_src_file, (unsigned int)message_type);
 205     qemu_put_be16(mis->to_src_file, len);
 206     qemu_put_buffer(mis->to_src_file, data, len);
 207     qemu_fflush(mis->to_src_file);
 208     qemu_mutex_unlock(&mis->rp_mutex);
 209 }
 210
 211 /* Request a range of pages from the source VM at the given
 212  * start address.
 213  *   rbname: Name of the RAMBlock to request the page in, if NULL it's the same
 214  *           as the last request (a name must have been given previously)
 215  *   Start: Address offset within the RB
 216  *   Len: Length in bytes required - must be a multiple of pagesize
 217  */
 218 void migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname,
 219                                ram_addr_t start, size_t len)
 220 {
 221     uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
 222     size_t msglen = 12; /* start + len */
 223
 224     *(uint64_t *)bufc = cpu_to_be64((uint64_t)start);
 225     *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len);
 226
 227     if (rbname) {
 228         int rbname_len = strlen(rbname);
 229         assert(rbname_len < 256);
 230
 231         bufc[msglen++] = rbname_len;
 232         memcpy(bufc + msglen, rbname, rbname_len);
 233         msglen += rbname_len;
 234         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES_ID, msglen, bufc);
 235     } else {
 236         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES, msglen, bufc);
 237     }
 238 }
 239
 240 void qemu_start_incoming_migration(const char *uri, Error **errp)
 241 {
 242     const char *p;
 243
 244     qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort);
 245     if (!strcmp(uri, "defer")) {
 246         deferred_incoming_migration(errp);
 247     } else if (strstart(uri, "tcp:", &p)) {
 248         tcp_start_incoming_migration(p, errp);
 249 #ifdef CONFIG_RDMA
 250     } else if (strstart(uri, "rdma:", &p)) {
 251         rdma_start_incoming_migration(p, errp);
 252 #endif
 253     } else if (strstart(uri, "exec:", &p)) {
 254         exec_start_incoming_migration(p, errp);
 255     } else if (strstart(uri, "unix:", &p)) {
 256         unix_start_incoming_migration(p, errp);
 257     } else if (strstart(uri, "fd:", &p)) {
 258         fd_start_incoming_migration(p, errp);
 259     } else {
 260         error_setg(errp, "unknown migration protocol: %s", uri);
 261     }
 262 }
 263
 264 static void process_incoming_migration_bh(void *opaque)
 265 {
 266     Error *local_err = NULL;
 267     MigrationIncomingState *mis = opaque;
 268
 269     /* Make sure all file formats flush their mutable metadata.
 270      * If we get an error here, just don't restart the VM yet. */
 271     bdrv_invalidate_cache_all(&local_err);
 272     if (local_err) {
 273         error_report_err(local_err);
 274         local_err = NULL;
 275         autostart = false;
 276     }
 277
 278     /*
 279      * This must happen after all error conditions are dealt with and
 280      * we're sure the VM is going to be running on this host.
 281      */
 282     qemu_announce_self();
 283
 284     if (multifd_load_cleanup(&local_err) != 0) {
 285         error_report_err(local_err);
 286         autostart = false;
 287     }
 288     /* If global state section was not received or we are in running
 289        state, we need to obey autostart. Any other state is set with
 290        runstate_set. */
 291
 292     if (!global_state_received() ||
 293         global_state_get_runstate() == RUN_STATE_RUNNING) {
 294         if (autostart) {
 295             vm_start();
 296         } else {
 297             runstate_set(RUN_STATE_PAUSED);
 298         }
 299     } else {
 300         runstate_set(global_state_get_runstate());
 301     }
 302     /*
 303      * This must happen after any state changes since as soon as an external
 304      * observer sees this event they might start to prod at the VM assuming
 305      * it's ready to use.
 306      */
 307     migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 308                       MIGRATION_STATUS_COMPLETED);
 309     qemu_bh_delete(mis->bh);
 310     migration_incoming_state_destroy();
 311 }
 312
 313 static void process_incoming_migration_co(void *opaque)
 314 {
 315     MigrationIncomingState *mis = migration_incoming_get_current();
 316     PostcopyState ps;
 317     int ret;
 318
 319     assert(mis->from_src_file);
 320     mis->largest_page_size = qemu_ram_pagesize_largest();
 321     postcopy_state_set(POSTCOPY_INCOMING_NONE);
 322     migrate_set_state(&mis->state, MIGRATION_STATUS_NONE,
 323                       MIGRATION_STATUS_ACTIVE);
 324     ret = qemu_loadvm_state(mis->from_src_file);
 325
 326     ps = postcopy_state_get();
 327     trace_process_incoming_migration_co_end(ret, ps);
 328     if (ps != POSTCOPY_INCOMING_NONE) {
 329         if (ps == POSTCOPY_INCOMING_ADVISE) {
 330             /*
 331              * Where a migration had postcopy enabled (and thus went to advise)
 332              * but managed to complete within the precopy period, we can use
 333              * the normal exit.
 334              */
 335             postcopy_ram_incoming_cleanup(mis);
 336         } else if (ret >= 0) {
 337             /*
 338              * Postcopy was started, cleanup should happen at the end of the
 339              * postcopy thread.
 340              */
 341             trace_process_incoming_migration_co_postcopy_end_main();
 342             return;
 343         }
 344         /* Else if something went wrong then just fall out of the normal exit */
 345     }
 346
 347     /* we get COLO info, and know if we are in COLO mode */
 348     if (!ret && migration_incoming_enable_colo()) {
 349         mis->migration_incoming_co = qemu_coroutine_self();
 350         qemu_thread_create(&mis->colo_incoming_thread, "COLO incoming",
 351              colo_process_incoming_thread, mis, QEMU_THREAD_JOINABLE);
 352         mis->have_colo_incoming_thread = true;
 353         qemu_coroutine_yield();
 354
 355         /* Wait checkpoint incoming thread exit before free resource */
 356         qemu_thread_join(&mis->colo_incoming_thread);
 357     }
 358
 359     if (ret < 0) {
 360         Error *local_err = NULL;
 361
 362         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 363                           MIGRATION_STATUS_FAILED);
 364         error_report("load of migration failed: %s", strerror(-ret));
 365         qemu_fclose(mis->from_src_file);
 366         if (multifd_load_cleanup(&local_err) != 0) {
 367             error_report_err(local_err);
 368         }
 369         exit(EXIT_FAILURE);
 370     }
 371     mis->bh = qemu_bh_new(process_incoming_migration_bh, mis);
 372     qemu_bh_schedule(mis->bh);
 373 }
 374
 375 static void migration_incoming_setup(QEMUFile *f)
 376 {
 377     MigrationIncomingState *mis = migration_incoming_get_current();
 378
 379     if (multifd_load_setup() != 0) {
 380         /* We haven't been able to create multifd threads
 381            nothing better to do */
 382         exit(EXIT_FAILURE);
 383     }
 384
 385     if (!mis->from_src_file) {
 386         mis->from_src_file = f;
 387     }
 388     qemu_file_set_blocking(f, false);
 389 }
 390
 391 static void migration_incoming_process(void)
 392 {
 393     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co, NULL);
 394     qemu_coroutine_enter(co);
 395 }
 396
 397 void migration_fd_process_incoming(QEMUFile *f)
 398 {
 399     migration_incoming_setup(f);
 400     migration_incoming_process();
 401 }
 402
 403 void migration_ioc_process_incoming(QIOChannel *ioc)
 404 {
 405     MigrationIncomingState *mis = migration_incoming_get_current();
 406
 407     if (!mis->from_src_file) {
 408         QEMUFile *f = qemu_fopen_channel_input(ioc);
 409         migration_fd_process_incoming(f);
 410     }
 411     /* We still only have a single channel.  Nothing to do here yet */
 412 }
 413
 414 /**
 415  * @migration_has_all_channels: We have received all channels that we need
 416  *
 417  * Returns true when we have got connections to all the channels that
 418  * we need for migration.
 419  */
 420 bool migration_has_all_channels(void)
 421 {
 422     return true;
 423 }
 424
 425 /*
 426  * Send a 'SHUT' message on the return channel with the given value
 427  * to indicate that we've finished with the RP.  Non-0 value indicates
 428  * error.
 429  */
 430 void migrate_send_rp_shut(MigrationIncomingState *mis,
 431                           uint32_t value)
 432 {
 433     uint32_t buf;
 434
 435     buf = cpu_to_be32(value);
 436     migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf);
 437 }
 438
 439 /*
 440  * Send a 'PONG' message on the return channel with the given value
 441  * (normally in response to a 'PING')
 442  */
 443 void migrate_send_rp_pong(MigrationIncomingState *mis,
 444                           uint32_t value)
 445 {
 446     uint32_t buf;
 447
 448     buf = cpu_to_be32(value);
 449     migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf);
 450 }
 451
 452 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
 453 {
 454     MigrationCapabilityStatusList *head = NULL;
 455     MigrationCapabilityStatusList *caps;
 456     MigrationState *s = migrate_get_current();
 457     int i;
 458
 459     caps = NULL; /* silence compiler warning */
 460     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
 461 #ifndef CONFIG_LIVE_BLOCK_MIGRATION
 462         if (i == MIGRATION_CAPABILITY_BLOCK) {
 463             continue;
 464         }
 465 #endif
 466         if (head == NULL) {
 467             head = g_malloc0(sizeof(*caps));
 468             caps = head;
 469         } else {
 470             caps->next = g_malloc0(sizeof(*caps));
 471             caps = caps->next;
 472         }
 473         caps->value =
 474             g_malloc(sizeof(*caps->value));
 475         caps->value->capability = i;
 476         caps->value->state = s->enabled_capabilities[i];
 477     }
 478
 479     return head;
 480 }
 481
 482 MigrationParameters *qmp_query_migrate_parameters(Error **errp)
 483 {
 484     MigrationParameters *params;
 485     MigrationState *s = migrate_get_current();
 486
 487     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 488     params = g_malloc0(sizeof(*params));
 489     params->has_compress_level = true;
 490     params->compress_level = s->parameters.compress_level;
 491     params->has_compress_threads = true;
 492     params->compress_threads = s->parameters.compress_threads;
 493     params->has_decompress_threads = true;
 494     params->decompress_threads = s->parameters.decompress_threads;
 495     params->has_cpu_throttle_initial = true;
 496     params->cpu_throttle_initial = s->parameters.cpu_throttle_initial;
 497     params->has_cpu_throttle_increment = true;
 498     params->cpu_throttle_increment = s->parameters.cpu_throttle_increment;
 499     params->has_tls_creds = true;
 500     params->tls_creds = g_strdup(s->parameters.tls_creds);
 501     params->has_tls_hostname = true;
 502     params->tls_hostname = g_strdup(s->parameters.tls_hostname);
 503     params->has_max_bandwidth = true;
 504     params->max_bandwidth = s->parameters.max_bandwidth;
 505     params->has_downtime_limit = true;
 506     params->downtime_limit = s->parameters.downtime_limit;
 507     params->has_x_checkpoint_delay = true;
 508     params->x_checkpoint_delay = s->parameters.x_checkpoint_delay;
 509     params->has_block_incremental = true;
 510     params->block_incremental = s->parameters.block_incremental;
 511     params->has_x_multifd_channels = true;
 512     params->x_multifd_channels = s->parameters.x_multifd_channels;
 513     params->has_x_multifd_page_count = true;
 514     params->x_multifd_page_count = s->parameters.x_multifd_page_count;
 515
 516     return params;
 517 }
 518
 519 /*
 520  * Return true if we're already in the middle of a migration
 521  * (i.e. any of the active or setup states)
 522  */
 523 static bool migration_is_setup_or_active(int state)
 524 {
 525     switch (state) {
 526     case MIGRATION_STATUS_ACTIVE:
 527     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 528     case MIGRATION_STATUS_SETUP:
 529     case MIGRATION_STATUS_PRE_SWITCHOVER:
 530     case MIGRATION_STATUS_DEVICE:
 531         return true;
 532
 533     default:
 534         return false;
 535
 536     }
 537 }
 538
 539 static void populate_ram_info(MigrationInfo *info, MigrationState *s)
 540 {
 541     info->has_ram = true;
 542     info->ram = g_malloc0(sizeof(*info->ram));
 543     info->ram->transferred = ram_counters.transferred;
 544     info->ram->total = ram_bytes_total();
 545     info->ram->duplicate = ram_counters.duplicate;
 546     /* legacy value.  It is not used anymore */
 547     info->ram->skipped = 0;
 548     info->ram->normal = ram_counters.normal;
 549     info->ram->normal_bytes = ram_counters.normal *
 550         qemu_target_page_size();
 551     info->ram->mbps = s->mbps;
 552     info->ram->dirty_sync_count = ram_counters.dirty_sync_count;
 553     info->ram->postcopy_requests = ram_counters.postcopy_requests;
 554     info->ram->page_size = qemu_target_page_size();
 555
 556     if (migrate_use_xbzrle()) {
 557         info->has_xbzrle_cache = true;
 558         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
 559         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
 560         info->xbzrle_cache->bytes = xbzrle_counters.bytes;
 561         info->xbzrle_cache->pages = xbzrle_counters.pages;
 562         info->xbzrle_cache->cache_miss = xbzrle_counters.cache_miss;
 563         info->xbzrle_cache->cache_miss_rate = xbzrle_counters.cache_miss_rate;
 564         info->xbzrle_cache->overflow = xbzrle_counters.overflow;
 565     }
 566
 567     if (cpu_throttle_active()) {
 568         info->has_cpu_throttle_percentage = true;
 569         info->cpu_throttle_percentage = cpu_throttle_get_percentage();
 570     }
 571
 572     if (s->state != MIGRATION_STATUS_COMPLETED) {
 573         info->ram->remaining = ram_bytes_remaining();
 574         info->ram->dirty_pages_rate = ram_counters.dirty_pages_rate;
 575     }
 576 }
 577
 578 static void populate_disk_info(MigrationInfo *info)
 579 {
 580     if (blk_mig_active()) {
 581         info->has_disk = true;
 582         info->disk = g_malloc0(sizeof(*info->disk));
 583         info->disk->transferred = blk_mig_bytes_transferred();
 584         info->disk->remaining = blk_mig_bytes_remaining();
 585         info->disk->total = blk_mig_bytes_total();
 586     }
 587 }
 588
 589 MigrationInfo *qmp_query_migrate(Error **errp)
 590 {
 591     MigrationInfo *info = g_malloc0(sizeof(*info));
 592     MigrationState *s = migrate_get_current();
 593
 594     switch (s->state) {
 595     case MIGRATION_STATUS_NONE:
 596         /* no migration has happened ever */
 597         break;
 598     case MIGRATION_STATUS_SETUP:
 599         info->has_status = true;
 600         info->has_total_time = false;
 601         break;
 602     case MIGRATION_STATUS_ACTIVE:
 603     case MIGRATION_STATUS_CANCELLING:
 604     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 605     case MIGRATION_STATUS_PRE_SWITCHOVER:
 606     case MIGRATION_STATUS_DEVICE:
 607          /* TODO add some postcopy stats */
 608         info->has_status = true;
 609         info->has_total_time = true;
 610         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 611             - s->total_time;
 612         info->has_expected_downtime = true;
 613         info->expected_downtime = s->expected_downtime;
 614         info->has_setup_time = true;
 615         info->setup_time = s->setup_time;
 616
 617         populate_ram_info(info, s);
 618         populate_disk_info(info);
 619         break;
 620     case MIGRATION_STATUS_COLO:
 621         info->has_status = true;
 622         /* TODO: display COLO specific information (checkpoint info etc.) */
 623         break;
 624     case MIGRATION_STATUS_COMPLETED:
 625         info->has_status = true;
 626         info->has_total_time = true;
 627         info->total_time = s->total_time;
 628         info->has_downtime = true;
 629         info->downtime = s->downtime;
 630         info->has_setup_time = true;
 631         info->setup_time = s->setup_time;
 632
 633         populate_ram_info(info, s);
 634         break;
 635     case MIGRATION_STATUS_FAILED:
 636         info->has_status = true;
 637         if (s->error) {
 638             info->has_error_desc = true;
 639             info->error_desc = g_strdup(error_get_pretty(s->error));
 640         }
 641         break;
 642     case MIGRATION_STATUS_CANCELLED:
 643         info->has_status = true;
 644         break;
 645     }
 646     info->status = s->state;
 647
 648     return info;
 649 }
 650
 651 /**
 652  * @migration_caps_check - check capability validity
 653  *
 654  * @cap_list: old capability list, array of bool
 655  * @params: new capabilities to be applied soon
 656  * @errp: set *errp if the check failed, with reason
 657  *
 658  * Returns true if check passed, otherwise false.
 659  */
 660 static bool migrate_caps_check(bool *cap_list,
 661                                MigrationCapabilityStatusList *params,
 662                                Error **errp)
 663 {
 664     MigrationCapabilityStatusList *cap;
 665     bool old_postcopy_cap;
 666     MigrationIncomingState *mis = migration_incoming_get_current();
 667
 668     old_postcopy_cap = cap_list[MIGRATION_CAPABILITY_POSTCOPY_RAM];
 669
 670     for (cap = params; cap; cap = cap->next) {
 671         cap_list[cap->value->capability] = cap->value->state;
 672     }
 673
 674 #ifndef CONFIG_LIVE_BLOCK_MIGRATION
 675     if (cap_list[MIGRATION_CAPABILITY_BLOCK]) {
 676         error_setg(errp, "QEMU compiled without old-style (blk/-b, inc/-i) "
 677                    "block migration");
 678         error_append_hint(errp, "Use drive_mirror+NBD instead.\n");
 679         return false;
 680     }
 681 #endif
 682
 683     if (cap_list[MIGRATION_CAPABILITY_POSTCOPY_RAM]) {
 684         if (cap_list[MIGRATION_CAPABILITY_COMPRESS]) {
 685             /* The decompression threads asynchronously write into RAM
 686              * rather than use the atomic copies needed to avoid
 687              * userfaulting.  It should be possible to fix the decompression
 688              * threads for compatibility in future.
 689              */
 690             error_setg(errp, "Postcopy is not currently compatible "
 691                        "with compression");
 692             return false;
 693         }
 694
 695         /* This check is reasonably expensive, so only when it's being
 696          * set the first time, also it's only the destination that needs
 697          * special support.
 698          */
 699         if (!old_postcopy_cap && runstate_check(RUN_STATE_INMIGRATE) &&
 700             !postcopy_ram_supported_by_host(mis)) {
 701             /* postcopy_ram_supported_by_host will have emitted a more
 702              * detailed message
 703              */
 704             error_setg(errp, "Postcopy is not supported");
 705             return false;
 706         }
 707     }
 708
 709     return true;
 710 }
 711
 712 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 713                                   Error **errp)
 714 {
 715     MigrationState *s = migrate_get_current();
 716     MigrationCapabilityStatusList *cap;
 717
 718     if (migration_is_setup_or_active(s->state)) {
 719         error_setg(errp, QERR_MIGRATION_ACTIVE);
 720         return;
 721     }
 722
 723     if (!migrate_caps_check(s->enabled_capabilities, params, errp)) {
 724         return;
 725     }
 726
 727     for (cap = params; cap; cap = cap->next) {
 728         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 729     }
 730 }
 731
 732 /*
 733  * Check whether the parameters are valid. Error will be put into errp
 734  * (if provided). Return true if valid, otherwise false.
 735  */
 736 static bool migrate_params_check(MigrationParameters *params, Error **errp)
 737 {
 738     if (params->has_compress_level &&
 739         (params->compress_level < 0 || params->compress_level > 9)) {
 740         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level",
 741                    "is invalid, it should be in the range of 0 to 9");
 742         return false;
 743     }
 744
 745     if (params->has_compress_threads &&
 746         (params->compress_threads < 1 || params->compress_threads > 255)) {
 747         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 748                    "compress_threads",
 749                    "is invalid, it should be in the range of 1 to 255");
 750         return false;
 751     }
 752
 753     if (params->has_decompress_threads &&
 754         (params->decompress_threads < 1 || params->decompress_threads > 255)) {
 755         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 756                    "decompress_threads",
 757                    "is invalid, it should be in the range of 1 to 255");
 758         return false;
 759     }
 760
 761     if (params->has_cpu_throttle_initial &&
 762         (params->cpu_throttle_initial < 1 ||
 763          params->cpu_throttle_initial > 99)) {
 764         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 765                    "cpu_throttle_initial",
 766                    "an integer in the range of 1 to 99");
 767         return false;
 768     }
 769
 770     if (params->has_cpu_throttle_increment &&
 771         (params->cpu_throttle_increment < 1 ||
 772          params->cpu_throttle_increment > 99)) {
 773         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 774                    "cpu_throttle_increment",
 775                    "an integer in the range of 1 to 99");
 776         return false;
 777     }
 778
 779     if (params->has_max_bandwidth &&
 780         (params->max_bandwidth < 0 || params->max_bandwidth > SIZE_MAX)) {
 781         error_setg(errp, "Parameter 'max_bandwidth' expects an integer in the"
 782                          " range of 0 to %zu bytes/second", SIZE_MAX);
 783         return false;
 784     }
 785
 786     if (params->has_downtime_limit &&
 787         (params->downtime_limit < 0 ||
 788          params->downtime_limit > MAX_MIGRATE_DOWNTIME)) {
 789         error_setg(errp, "Parameter 'downtime_limit' expects an integer in "
 790                          "the range of 0 to %d milliseconds",
 791                          MAX_MIGRATE_DOWNTIME);
 792         return false;
 793     }
 794
 795     if (params->has_x_checkpoint_delay && (params->x_checkpoint_delay < 0)) {
 796         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 797                     "x_checkpoint_delay",
 798                     "is invalid, it should be positive");
 799         return false;
 800     }
 801     if (params->has_x_multifd_channels &&
 802         (params->x_multifd_channels < 1 || params->x_multifd_channels > 255)) {
 803         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 804                    "multifd_channels",
 805                    "is invalid, it should be in the range of 1 to 255");
 806         return false;
 807     }
 808     if (params->has_x_multifd_page_count &&
 809             (params->x_multifd_page_count < 1 ||
 810              params->x_multifd_page_count > 10000)) {
 811         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 812                    "multifd_page_count",
 813                    "is invalid, it should be in the range of 1 to 10000");
 814         return false;
 815     }
 816
 817     return true;
 818 }
 819
 820 static void migrate_params_test_apply(MigrateSetParameters *params,
 821                                       MigrationParameters *dest)
 822 {
 823     *dest = migrate_get_current()->parameters;
 824
 825     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 826
 827     if (params->has_compress_level) {
 828         dest->compress_level = params->compress_level;
 829     }
 830
 831     if (params->has_compress_threads) {
 832         dest->compress_threads = params->compress_threads;
 833     }
 834
 835     if (params->has_decompress_threads) {
 836         dest->decompress_threads = params->decompress_threads;
 837     }
 838
 839     if (params->has_cpu_throttle_initial) {
 840         dest->cpu_throttle_initial = params->cpu_throttle_initial;
 841     }
 842
 843     if (params->has_cpu_throttle_increment) {
 844         dest->cpu_throttle_increment = params->cpu_throttle_increment;
 845     }
 846
 847     if (params->has_tls_creds) {
 848         assert(params->tls_creds->type == QTYPE_QSTRING);
 849         dest->tls_creds = g_strdup(params->tls_creds->u.s);
 850     }
 851
 852     if (params->has_tls_hostname) {
 853         assert(params->tls_hostname->type == QTYPE_QSTRING);
 854         dest->tls_hostname = g_strdup(params->tls_hostname->u.s);
 855     }
 856
 857     if (params->has_max_bandwidth) {
 858         dest->max_bandwidth = params->max_bandwidth;
 859     }
 860
 861     if (params->has_downtime_limit) {
 862         dest->downtime_limit = params->downtime_limit;
 863     }
 864
 865     if (params->has_x_checkpoint_delay) {
 866         dest->x_checkpoint_delay = params->x_checkpoint_delay;
 867     }
 868
 869     if (params->has_block_incremental) {
 870         dest->block_incremental = params->block_incremental;
 871     }
 872     if (params->has_x_multifd_channels) {
 873         dest->x_multifd_channels = params->x_multifd_channels;
 874     }
 875     if (params->has_x_multifd_page_count) {
 876         dest->x_multifd_page_count = params->x_multifd_page_count;
 877     }
 878 }
 879
 880 static void migrate_params_apply(MigrateSetParameters *params)
 881 {
 882     MigrationState *s = migrate_get_current();
 883
 884     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 885
 886     if (params->has_compress_level) {
 887         s->parameters.compress_level = params->compress_level;
 888     }
 889
 890     if (params->has_compress_threads) {
 891         s->parameters.compress_threads = params->compress_threads;
 892     }
 893
 894     if (params->has_decompress_threads) {
 895         s->parameters.decompress_threads = params->decompress_threads;
 896     }
 897
 898     if (params->has_cpu_throttle_initial) {
 899         s->parameters.cpu_throttle_initial = params->cpu_throttle_initial;
 900     }
 901
 902     if (params->has_cpu_throttle_increment) {
 903         s->parameters.cpu_throttle_increment = params->cpu_throttle_increment;
 904     }
 905
 906     if (params->has_tls_creds) {
 907         g_free(s->parameters.tls_creds);
 908         assert(params->tls_creds->type == QTYPE_QSTRING);
 909         s->parameters.tls_creds = g_strdup(params->tls_creds->u.s);
 910     }
 911
 912     if (params->has_tls_hostname) {
 913         g_free(s->parameters.tls_hostname);
 914         assert(params->tls_hostname->type == QTYPE_QSTRING);
 915         s->parameters.tls_hostname = g_strdup(params->tls_hostname->u.s);
 916     }
 917
 918     if (params->has_max_bandwidth) {
 919         s->parameters.max_bandwidth = params->max_bandwidth;
 920         if (s->to_dst_file) {
 921             qemu_file_set_rate_limit(s->to_dst_file,
 922                                 s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
 923         }
 924     }
 925
 926     if (params->has_downtime_limit) {
 927         s->parameters.downtime_limit = params->downtime_limit;
 928     }
 929
 930     if (params->has_x_checkpoint_delay) {
 931         s->parameters.x_checkpoint_delay = params->x_checkpoint_delay;
 932         if (migration_in_colo_state()) {
 933             colo_checkpoint_notify(s);
 934         }
 935     }
 936
 937     if (params->has_block_incremental) {
 938         s->parameters.block_incremental = params->block_incremental;
 939     }
 940     if (params->has_x_multifd_channels) {
 941         s->parameters.x_multifd_channels = params->x_multifd_channels;
 942     }
 943     if (params->has_x_multifd_page_count) {
 944         s->parameters.x_multifd_page_count = params->x_multifd_page_count;
 945     }
 946 }
 947
 948 void qmp_migrate_set_parameters(MigrateSetParameters *params, Error **errp)
 949 {
 950     MigrationParameters tmp;
 951
 952     /* TODO Rewrite "" to null instead */
 953     if (params->has_tls_creds
 954         && params->tls_creds->type == QTYPE_QNULL) {
 955         QDECREF(params->tls_creds->u.n);
 956         params->tls_creds->type = QTYPE_QSTRING;
 957         params->tls_creds->u.s = strdup("");
 958     }
 959     /* TODO Rewrite "" to null instead */
 960     if (params->has_tls_hostname
 961         && params->tls_hostname->type == QTYPE_QNULL) {
 962         QDECREF(params->tls_hostname->u.n);
 963         params->tls_hostname->type = QTYPE_QSTRING;
 964         params->tls_hostname->u.s = strdup("");
 965     }
 966
 967     migrate_params_test_apply(params, &tmp);
 968
 969     if (!migrate_params_check(&tmp, errp)) {
 970         /* Invalid parameter */
 971         return;
 972     }
 973
 974     migrate_params_apply(params);
 975 }
 976
 977
 978 void qmp_migrate_start_postcopy(Error **errp)
 979 {
 980     MigrationState *s = migrate_get_current();
 981
 982     if (!migrate_postcopy_ram()) {
 983         error_setg(errp, "Enable postcopy with migrate_set_capability before"
 984                          " the start of migration");
 985         return;
 986     }
 987
 988     if (s->state == MIGRATION_STATUS_NONE) {
 989         error_setg(errp, "Postcopy must be started after migration has been"
 990                          " started");
 991         return;
 992     }
 993     /*
 994      * we don't error if migration has finished since that would be racy
 995      * with issuing this command.
 996      */
 997     atomic_set(&s->start_postcopy, true);
 998 }
 999
1000 /* shared migration helpers */
1001
1002 void migrate_set_state(int *state, int old_state, int new_state)
1003 {
1004     assert(new_state < MIGRATION_STATUS__MAX);
1005     if (atomic_cmpxchg(state, old_state, new_state) == old_state) {
1006         trace_migrate_set_state(MigrationStatus_str(new_state));
1007         migrate_generate_event(new_state);
1008     }
1009 }
1010
1011 static MigrationCapabilityStatusList *migrate_cap_add(
1012     MigrationCapabilityStatusList *list,
1013     MigrationCapability index,
1014     bool state)
1015 {
1016     MigrationCapabilityStatusList *cap;
1017
1018     cap = g_new0(MigrationCapabilityStatusList, 1);
1019     cap->value = g_new0(MigrationCapabilityStatus, 1);
1020     cap->value->capability = index;
1021     cap->value->state = state;
1022     cap->next = list;
1023
1024     return cap;
1025 }
1026
1027 void migrate_set_block_enabled(bool value, Error **errp)
1028 {
1029     MigrationCapabilityStatusList *cap;
1030
1031     cap = migrate_cap_add(NULL, MIGRATION_CAPABILITY_BLOCK, value);
1032     qmp_migrate_set_capabilities(cap, errp);
1033     qapi_free_MigrationCapabilityStatusList(cap);
1034 }
1035
1036 static void migrate_set_block_incremental(MigrationState *s, bool value)
1037 {
1038     s->parameters.block_incremental = value;
1039 }
1040
1041 static void block_cleanup_parameters(MigrationState *s)
1042 {
1043     if (s->must_remove_block_options) {
1044         /* setting to false can never fail */
1045         migrate_set_block_enabled(false, &error_abort);
1046         migrate_set_block_incremental(s, false);
1047         s->must_remove_block_options = false;
1048     }
1049 }
1050
1051 static void migrate_fd_cleanup(void *opaque)
1052 {
1053     MigrationState *s = opaque;
1054
1055     qemu_bh_delete(s->cleanup_bh);
1056     s->cleanup_bh = NULL;
1057
1058     if (s->to_dst_file) {
1059         Error *local_err = NULL;
1060
1061         trace_migrate_fd_cleanup();
1062         qemu_mutex_unlock_iothread();
1063         if (s->migration_thread_running) {
1064             qemu_thread_join(&s->thread);
1065             s->migration_thread_running = false;
1066         }
1067         qemu_mutex_lock_iothread();
1068
1069         if (multifd_save_cleanup(&local_err) != 0) {
1070             error_report_err(local_err);
1071         }
1072         qemu_fclose(s->to_dst_file);
1073         s->to_dst_file = NULL;
1074     }
1075
1076     assert((s->state != MIGRATION_STATUS_ACTIVE) &&
1077            (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE));
1078
1079     if (s->state == MIGRATION_STATUS_CANCELLING) {
1080         migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING,
1081                           MIGRATION_STATUS_CANCELLED);
1082     }
1083
1084     notifier_list_notify(&migration_state_notifiers, s);
1085     block_cleanup_parameters(s);
1086 }
1087
1088 void migrate_fd_error(MigrationState *s, const Error *error)
1089 {
1090     trace_migrate_fd_error(error_get_pretty(error));
1091     assert(s->to_dst_file == NULL);
1092     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1093                       MIGRATION_STATUS_FAILED);
1094     if (!s->error) {
1095         s->error = error_copy(error);
1096     }
1097     notifier_list_notify(&migration_state_notifiers, s);
1098     block_cleanup_parameters(s);
1099 }
1100
1101 static void migrate_fd_cancel(MigrationState *s)
1102 {
1103     int old_state ;
1104     QEMUFile *f = migrate_get_current()->to_dst_file;
1105     trace_migrate_fd_cancel();
1106
1107     if (s->rp_state.from_dst_file) {
1108         /* shutdown the rp socket, so causing the rp thread to shutdown */
1109         qemu_file_shutdown(s->rp_state.from_dst_file);
1110     }
1111
1112     do {
1113         old_state = s->state;
1114         if (!migration_is_setup_or_active(old_state)) {
1115             break;
1116         }
1117         migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING);
1118     } while (s->state != MIGRATION_STATUS_CANCELLING);
1119
1120     /*
1121      * If we're unlucky the migration code might be stuck somewhere in a
1122      * send/write while the network has failed and is waiting to timeout;
1123      * if we've got shutdown(2) available then we can force it to quit.
1124      * The outgoing qemu file gets closed in migrate_fd_cleanup that is
1125      * called in a bh, so there is no race against this cancel.
1126      */
1127     if (s->state == MIGRATION_STATUS_CANCELLING && f) {
1128         qemu_file_shutdown(f);
1129     }
1130     if (s->state == MIGRATION_STATUS_CANCELLING && s->block_inactive) {
1131         Error *local_err = NULL;
1132
1133         bdrv_invalidate_cache_all(&local_err);
1134         if (local_err) {
1135             error_report_err(local_err);
1136         } else {
1137             s->block_inactive = false;
1138         }
1139     }
1140     block_cleanup_parameters(s);
1141 }
1142
1143 void add_migration_state_change_notifier(Notifier *notify)
1144 {
1145     notifier_list_add(&migration_state_notifiers, notify);
1146 }
1147
1148 void remove_migration_state_change_notifier(Notifier *notify)
1149 {
1150     notifier_remove(notify);
1151 }
1152
1153 bool migration_in_setup(MigrationState *s)
1154 {
1155     return s->state == MIGRATION_STATUS_SETUP;
1156 }
1157
1158 bool migration_has_finished(MigrationState *s)
1159 {
1160     return s->state == MIGRATION_STATUS_COMPLETED;
1161 }
1162
1163 bool migration_has_failed(MigrationState *s)
1164 {
1165     return (s->state == MIGRATION_STATUS_CANCELLED ||
1166             s->state == MIGRATION_STATUS_FAILED);
1167 }
1168
1169 bool migration_in_postcopy(void)
1170 {
1171     MigrationState *s = migrate_get_current();
1172
1173     return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
1174 }
1175
1176 bool migration_in_postcopy_after_devices(MigrationState *s)
1177 {
1178     return migration_in_postcopy() && s->postcopy_after_devices;
1179 }
1180
1181 bool migration_is_idle(void)
1182 {
1183     MigrationState *s = migrate_get_current();
1184
1185     switch (s->state) {
1186     case MIGRATION_STATUS_NONE:
1187     case MIGRATION_STATUS_CANCELLED:
1188     case MIGRATION_STATUS_COMPLETED:
1189     case MIGRATION_STATUS_FAILED:
1190         return true;
1191     case MIGRATION_STATUS_SETUP:
1192     case MIGRATION_STATUS_CANCELLING:
1193     case MIGRATION_STATUS_ACTIVE:
1194     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
1195     case MIGRATION_STATUS_COLO:
1196     case MIGRATION_STATUS_PRE_SWITCHOVER:
1197     case MIGRATION_STATUS_DEVICE:
1198         return false;
1199     case MIGRATION_STATUS__MAX:
1200         g_assert_not_reached();
1201     }
1202
1203     return false;
1204 }
1205
1206 MigrationState *migrate_init(void)
1207 {
1208     MigrationState *s = migrate_get_current();
1209
1210     /*
1211      * Reinitialise all migration state, except
1212      * parameters/capabilities that the user set, and
1213      * locks.
1214      */
1215     s->bytes_xfer = 0;
1216     s->xfer_limit = 0;
1217     s->cleanup_bh = 0;
1218     s->to_dst_file = NULL;
1219     s->state = MIGRATION_STATUS_NONE;
1220     s->rp_state.from_dst_file = NULL;
1221     s->rp_state.error = false;
1222     s->mbps = 0.0;
1223     s->downtime = 0;
1224     s->expected_downtime = 0;
1225     s->setup_time = 0;
1226     s->start_postcopy = false;
1227     s->postcopy_after_devices = false;
1228     s->migration_thread_running = false;
1229     error_free(s->error);
1230     s->error = NULL;
1231
1232     migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP);
1233
1234     s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1235     return s;
1236 }
1237
1238 static GSList *migration_blockers;
1239
1240 int migrate_add_blocker(Error *reason, Error **errp)
1241 {
1242     if (migrate_get_current()->only_migratable) {
1243         error_propagate(errp, error_copy(reason));
1244         error_prepend(errp, "disallowing migration blocker "
1245                           "(--only_migratable) for: ");
1246         return -EACCES;
1247     }
1248
1249     if (migration_is_idle()) {
1250         migration_blockers = g_slist_prepend(migration_blockers, reason);
1251         return 0;
1252     }
1253
1254     error_propagate(errp, error_copy(reason));
1255     error_prepend(errp, "disallowing migration blocker (migration in "
1256                       "progress) for: ");
1257     return -EBUSY;
1258 }
1259
1260 void migrate_del_blocker(Error *reason)
1261 {
1262     migration_blockers = g_slist_remove(migration_blockers, reason);
1263 }
1264
1265 void qmp_migrate_incoming(const char *uri, Error **errp)
1266 {
1267     Error *local_err = NULL;
1268     static bool once = true;
1269
1270     if (!deferred_incoming) {
1271         error_setg(errp, "For use with '-incoming defer'");
1272         return;
1273     }
1274     if (!once) {
1275         error_setg(errp, "The incoming migration has already been started");
1276     }
1277
1278     qemu_start_incoming_migration(uri, &local_err);
1279
1280     if (local_err) {
1281         error_propagate(errp, local_err);
1282         return;
1283     }
1284
1285     once = false;
1286 }
1287
1288 bool migration_is_blocked(Error **errp)
1289 {
1290     if (qemu_savevm_state_blocked(errp)) {
1291         return true;
1292     }
1293
1294     if (migration_blockers) {
1295         error_propagate(errp, error_copy(migration_blockers->data));
1296         return true;
1297     }
1298
1299     return false;
1300 }
1301
1302 void qmp_migrate(const char *uri, bool has_blk, bool blk,
1303                  bool has_inc, bool inc, bool has_detach, bool detach,
1304                  Error **errp)
1305 {
1306     Error *local_err = NULL;
1307     MigrationState *s = migrate_get_current();
1308     const char *p;
1309
1310     if (migration_is_setup_or_active(s->state) ||
1311         s->state == MIGRATION_STATUS_CANCELLING ||
1312         s->state == MIGRATION_STATUS_COLO) {
1313         error_setg(errp, QERR_MIGRATION_ACTIVE);
1314         return;
1315     }
1316     if (runstate_check(RUN_STATE_INMIGRATE)) {
1317         error_setg(errp, "Guest is waiting for an incoming migration");
1318         return;
1319     }
1320
1321     if (migration_is_blocked(errp)) {
1322         return;
1323     }
1324
1325     if ((has_blk && blk) || (has_inc && inc)) {
1326         if (migrate_use_block() || migrate_use_block_incremental()) {
1327             error_setg(errp, "Command options are incompatible with "
1328                        "current migration capabilities");
1329             return;
1330         }
1331         migrate_set_block_enabled(true, &local_err);
1332         if (local_err) {
1333             error_propagate(errp, local_err);
1334             return;
1335         }
1336         s->must_remove_block_options = true;
1337     }
1338
1339     if (has_inc && inc) {
1340         migrate_set_block_incremental(s, true);
1341     }
1342
1343     s = migrate_init();
1344
1345     if (strstart(uri, "tcp:", &p)) {
1346         tcp_start_outgoing_migration(s, p, &local_err);
1347 #ifdef CONFIG_RDMA
1348     } else if (strstart(uri, "rdma:", &p)) {
1349         rdma_start_outgoing_migration(s, p, &local_err);
1350 #endif
1351     } else if (strstart(uri, "exec:", &p)) {
1352         exec_start_outgoing_migration(s, p, &local_err);
1353     } else if (strstart(uri, "unix:", &p)) {
1354         unix_start_outgoing_migration(s, p, &local_err);
1355     } else if (strstart(uri, "fd:", &p)) {
1356         fd_start_outgoing_migration(s, p, &local_err);
1357     } else {
1358         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri",
1359                    "a valid migration protocol");
1360         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1361                           MIGRATION_STATUS_FAILED);
1362         return;
1363     }
1364
1365     if (local_err) {
1366         migrate_fd_error(s, local_err);
1367         error_propagate(errp, local_err);
1368         return;
1369     }
1370 }
1371
1372 void qmp_migrate_cancel(Error **errp)
1373 {
1374     migrate_fd_cancel(migrate_get_current());
1375 }
1376
1377 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
1378 {
1379     MigrationState *s = migrate_get_current();
1380     int64_t new_size;
1381
1382     new_size = xbzrle_cache_resize(value, errp);
1383     if (new_size < 0) {
1384         return;
1385     }
1386
1387     s->xbzrle_cache_size = new_size;
1388 }
1389
1390 int64_t qmp_query_migrate_cache_size(Error **errp)
1391 {
1392     return migrate_xbzrle_cache_size();
1393 }
1394
1395 void qmp_migrate_set_speed(int64_t value, Error **errp)
1396 {
1397     MigrateSetParameters p = {
1398         .has_max_bandwidth = true,
1399         .max_bandwidth = value,
1400     };
1401
1402     qmp_migrate_set_parameters(&p, errp);
1403 }
1404
1405 void qmp_migrate_set_downtime(double value, Error **errp)
1406 {
1407     if (value < 0 || value > MAX_MIGRATE_DOWNTIME_SECONDS) {
1408         error_setg(errp, "Parameter 'downtime_limit' expects an integer in "
1409                          "the range of 0 to %d seconds",
1410                          MAX_MIGRATE_DOWNTIME_SECONDS);
1411         return;
1412     }
1413
1414     value *= 1000; /* Convert to milliseconds */
1415     value = MAX(0, MIN(INT64_MAX, value));
1416
1417     MigrateSetParameters p = {
1418         .has_downtime_limit = true,
1419         .downtime_limit = value,
1420     };
1421
1422     qmp_migrate_set_parameters(&p, errp);
1423 }
1424
1425 bool migrate_release_ram(void)
1426 {
1427     MigrationState *s;
1428
1429     s = migrate_get_current();
1430
1431     return s->enabled_capabilities[MIGRATION_CAPABILITY_RELEASE_RAM];
1432 }
1433
1434 bool migrate_postcopy_ram(void)
1435 {
1436     MigrationState *s;
1437
1438     s = migrate_get_current();
1439
1440     return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM];
1441 }
1442
1443 bool migrate_postcopy(void)
1444 {
1445     return migrate_postcopy_ram();
1446 }
1447
1448 bool migrate_auto_converge(void)
1449 {
1450     MigrationState *s;
1451
1452     s = migrate_get_current();
1453
1454     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
1455 }
1456
1457 bool migrate_zero_blocks(void)
1458 {
1459     MigrationState *s;
1460
1461     s = migrate_get_current();
1462
1463     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
1464 }
1465
1466 bool migrate_use_compression(void)
1467 {
1468     MigrationState *s;
1469
1470     s = migrate_get_current();
1471
1472     return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS];
1473 }
1474
1475 int migrate_compress_level(void)
1476 {
1477     MigrationState *s;
1478
1479     s = migrate_get_current();
1480
1481     return s->parameters.compress_level;
1482 }
1483
1484 int migrate_compress_threads(void)
1485 {
1486     MigrationState *s;
1487
1488     s = migrate_get_current();
1489
1490     return s->parameters.compress_threads;
1491 }
1492
1493 int migrate_decompress_threads(void)
1494 {
1495     MigrationState *s;
1496
1497     s = migrate_get_current();
1498
1499     return s->parameters.decompress_threads;
1500 }
1501
1502 bool migrate_use_events(void)
1503 {
1504     MigrationState *s;
1505
1506     s = migrate_get_current();
1507
1508     return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS];
1509 }
1510
1511 bool migrate_use_multifd(void)
1512 {
1513     MigrationState *s;
1514
1515     s = migrate_get_current();
1516
1517     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_MULTIFD];
1518 }
1519
1520 bool migrate_pause_before_switchover(void)
1521 {
1522     MigrationState *s;
1523
1524     s = migrate_get_current();
1525
1526     return s->enabled_capabilities[
1527         MIGRATION_CAPABILITY_PAUSE_BEFORE_SWITCHOVER];
1528 }
1529
1530 int migrate_multifd_channels(void)
1531 {
1532     MigrationState *s;
1533
1534     s = migrate_get_current();
1535
1536     return s->parameters.x_multifd_channels;
1537 }
1538
1539 int migrate_multifd_page_count(void)
1540 {
1541     MigrationState *s;
1542
1543     s = migrate_get_current();
1544
1545     return s->parameters.x_multifd_page_count;
1546 }
1547
1548 int migrate_use_xbzrle(void)
1549 {
1550     MigrationState *s;
1551
1552     s = migrate_get_current();
1553
1554     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
1555 }
1556
1557 int64_t migrate_xbzrle_cache_size(void)
1558 {
1559     MigrationState *s;
1560
1561     s = migrate_get_current();
1562
1563     return s->xbzrle_cache_size;
1564 }
1565
1566 bool migrate_use_block(void)
1567 {
1568     MigrationState *s;
1569
1570     s = migrate_get_current();
1571
1572     return s->enabled_capabilities[MIGRATION_CAPABILITY_BLOCK];
1573 }
1574
1575 bool migrate_use_return_path(void)
1576 {
1577     MigrationState *s;
1578
1579     s = migrate_get_current();
1580
1581     return s->enabled_capabilities[MIGRATION_CAPABILITY_RETURN_PATH];
1582 }
1583
1584 bool migrate_use_block_incremental(void)
1585 {
1586     MigrationState *s;
1587
1588     s = migrate_get_current();
1589
1590     return s->parameters.block_incremental;
1591 }
1592
1593 /* migration thread support */
1594 /*
1595  * Something bad happened to the RP stream, mark an error
1596  * The caller shall print or trace something to indicate why
1597  */
1598 static void mark_source_rp_bad(MigrationState *s)
1599 {
1600     s->rp_state.error = true;
1601 }
1602
1603 static struct rp_cmd_args {
1604     ssize_t     len; /* -1 = variable */
1605     const char *name;
1606 } rp_cmd_args[] = {
1607     [MIG_RP_MSG_INVALID]        = { .len = -1, .name = "INVALID" },
1608     [MIG_RP_MSG_SHUT]           = { .len =  4, .name = "SHUT" },
1609     [MIG_RP_MSG_PONG]           = { .len =  4, .name = "PONG" },
1610     [MIG_RP_MSG_REQ_PAGES]      = { .len = 12, .name = "REQ_PAGES" },
1611     [MIG_RP_MSG_REQ_PAGES_ID]   = { .len = -1, .name = "REQ_PAGES_ID" },
1612     [MIG_RP_MSG_MAX]            = { .len = -1, .name = "MAX" },
1613 };
1614
1615 /*
1616  * Process a request for pages received on the return path,
1617  * We're allowed to send more than requested (e.g. to round to our page size)
1618  * and we don't need to send pages that have already been sent.
1619  */
1620 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname,
1621                                        ram_addr_t start, size_t len)
1622 {
1623     long our_host_ps = getpagesize();
1624
1625     trace_migrate_handle_rp_req_pages(rbname, start, len);
1626
1627     /*
1628      * Since we currently insist on matching page sizes, just sanity check
1629      * we're being asked for whole host pages.
1630      */
1631     if (start & (our_host_ps-1) ||
1632        (len & (our_host_ps-1))) {
1633         error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1634                      " len: %zd", __func__, start, len);
1635         mark_source_rp_bad(ms);
1636         return;
1637     }
1638
1639     if (ram_save_queue_pages(rbname, start, len)) {
1640         mark_source_rp_bad(ms);
1641     }
1642 }
1643
1644 /*
1645  * Handles messages sent on the return path towards the source VM
1646  *
1647  */
1648 static void *source_return_path_thread(void *opaque)
1649 {
1650     MigrationState *ms = opaque;
1651     QEMUFile *rp = ms->rp_state.from_dst_file;
1652     uint16_t header_len, header_type;
1653     uint8_t buf[512];
1654     uint32_t tmp32, sibling_error;
1655     ram_addr_t start = 0; /* =0 to silence warning */
1656     size_t  len = 0, expected_len;
1657     int res;
1658
1659     trace_source_return_path_thread_entry();
1660     while (!ms->rp_state.error && !qemu_file_get_error(rp) &&
1661            migration_is_setup_or_active(ms->state)) {
1662         trace_source_return_path_thread_loop_top();
1663         header_type = qemu_get_be16(rp);
1664         header_len = qemu_get_be16(rp);
1665
1666         if (header_type >= MIG_RP_MSG_MAX ||
1667             header_type == MIG_RP_MSG_INVALID) {
1668             error_report("RP: Received invalid message 0x%04x length 0x%04x",
1669                     header_type, header_len);
1670             mark_source_rp_bad(ms);
1671             goto out;
1672         }
1673
1674         if ((rp_cmd_args[header_type].len != -1 &&
1675             header_len != rp_cmd_args[header_type].len) ||
1676             header_len > sizeof(buf)) {
1677             error_report("RP: Received '%s' message (0x%04x) with"
1678                     "incorrect length %d expecting %zu",
1679                     rp_cmd_args[header_type].name, header_type, header_len,
1680                     (size_t)rp_cmd_args[header_type].len);
1681             mark_source_rp_bad(ms);
1682             goto out;
1683         }
1684
1685         /* We know we've got a valid header by this point */
1686         res = qemu_get_buffer(rp, buf, header_len);
1687         if (res != header_len) {
1688             error_report("RP: Failed reading data for message 0x%04x"
1689                          " read %d expected %d",
1690                          header_type, res, header_len);
1691             mark_source_rp_bad(ms);
1692             goto out;
1693         }
1694
1695         /* OK, we have the message and the data */
1696         switch (header_type) {
1697         case MIG_RP_MSG_SHUT:
1698             sibling_error = ldl_be_p(buf);
1699             trace_source_return_path_thread_shut(sibling_error);
1700             if (sibling_error) {
1701                 error_report("RP: Sibling indicated error %d", sibling_error);
1702                 mark_source_rp_bad(ms);
1703             }
1704             /*
1705              * We'll let the main thread deal with closing the RP
1706              * we could do a shutdown(2) on it, but we're the only user
1707              * anyway, so there's nothing gained.
1708              */
1709             goto out;
1710
1711         case MIG_RP_MSG_PONG:
1712             tmp32 = ldl_be_p(buf);
1713             trace_source_return_path_thread_pong(tmp32);
1714             break;
1715
1716         case MIG_RP_MSG_REQ_PAGES:
1717             start = ldq_be_p(buf);
1718             len = ldl_be_p(buf + 8);
1719             migrate_handle_rp_req_pages(ms, NULL, start, len);
1720             break;
1721
1722         case MIG_RP_MSG_REQ_PAGES_ID:
1723             expected_len = 12 + 1; /* header + termination */
1724
1725             if (header_len >= expected_len) {
1726                 start = ldq_be_p(buf);
1727                 len = ldl_be_p(buf + 8);
1728                 /* Now we expect an idstr */
1729                 tmp32 = buf[12]; /* Length of the following idstr */
1730                 buf[13 + tmp32] = '\0';
1731                 expected_len += tmp32;
1732             }
1733             if (header_len != expected_len) {
1734                 error_report("RP: Req_Page_id with length %d expecting %zd",
1735                         header_len, expected_len);
1736                 mark_source_rp_bad(ms);
1737                 goto out;
1738             }
1739             migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len);
1740             break;
1741
1742         default:
1743             break;
1744         }
1745     }
1746     if (qemu_file_get_error(rp)) {
1747         trace_source_return_path_thread_bad_end();
1748         mark_source_rp_bad(ms);
1749     }
1750
1751     trace_source_return_path_thread_end();
1752 out:
1753     ms->rp_state.from_dst_file = NULL;
1754     qemu_fclose(rp);
1755     return NULL;
1756 }
1757
1758 static int open_return_path_on_source(MigrationState *ms)
1759 {
1760
1761     ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file);
1762     if (!ms->rp_state.from_dst_file) {
1763         return -1;
1764     }
1765
1766     trace_open_return_path_on_source();
1767     qemu_thread_create(&ms->rp_state.rp_thread, "return path",
1768                        source_return_path_thread, ms, QEMU_THREAD_JOINABLE);
1769
1770     trace_open_return_path_on_source_continue();
1771
1772     return 0;
1773 }
1774
1775 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
1776 static int await_return_path_close_on_source(MigrationState *ms)
1777 {
1778     /*
1779      * If this is a normal exit then the destination will send a SHUT and the
1780      * rp_thread will exit, however if there's an error we need to cause
1781      * it to exit.
1782      */
1783     if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
1784         /*
1785          * shutdown(2), if we have it, will cause it to unblock if it's stuck
1786          * waiting for the destination.
1787          */
1788         qemu_file_shutdown(ms->rp_state.from_dst_file);
1789         mark_source_rp_bad(ms);
1790     }
1791     trace_await_return_path_close_on_source_joining();
1792     qemu_thread_join(&ms->rp_state.rp_thread);
1793     trace_await_return_path_close_on_source_close();
1794     return ms->rp_state.error;
1795 }
1796
1797 /*
1798  * Switch from normal iteration to postcopy
1799  * Returns non-0 on error
1800  */
1801 static int postcopy_start(MigrationState *ms, bool *old_vm_running)
1802 {
1803     int ret;
1804     QIOChannelBuffer *bioc;
1805     QEMUFile *fb;
1806     int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1807     bool restart_block = false;
1808     migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE,
1809                       MIGRATION_STATUS_POSTCOPY_ACTIVE);
1810
1811     trace_postcopy_start();
1812     qemu_mutex_lock_iothread();
1813     trace_postcopy_start_set_run();
1814
1815     qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1816     *old_vm_running = runstate_is_running();
1817     global_state_store();
1818     ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1819     if (ret < 0) {
1820         goto fail;
1821     }
1822
1823     ret = bdrv_inactivate_all();
1824     if (ret < 0) {
1825         goto fail;
1826     }
1827     restart_block = true;
1828
1829     /*
1830      * Cause any non-postcopiable, but iterative devices to
1831      * send out their final data.
1832      */
1833     qemu_savevm_state_complete_precopy(ms->to_dst_file, true, false);
1834
1835     /*
1836      * in Finish migrate and with the io-lock held everything should
1837      * be quiet, but we've potentially still got dirty pages and we
1838      * need to tell the destination to throw any pages it's already received
1839      * that are dirty
1840      */
1841     if (migrate_postcopy_ram()) {
1842         if (ram_postcopy_send_discard_bitmap(ms)) {
1843             error_report("postcopy send discard bitmap failed");
1844             goto fail;
1845         }
1846     }
1847
1848     /*
1849      * send rest of state - note things that are doing postcopy
1850      * will notice we're in POSTCOPY_ACTIVE and not actually
1851      * wrap their state up here
1852      */
1853     qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX);
1854     if (migrate_postcopy_ram()) {
1855         /* Ping just for debugging, helps line traces up */
1856         qemu_savevm_send_ping(ms->to_dst_file, 2);
1857     }
1858
1859     /*
1860      * While loading the device state we may trigger page transfer
1861      * requests and the fd must be free to process those, and thus
1862      * the destination must read the whole device state off the fd before
1863      * it starts processing it.  Unfortunately the ad-hoc migration format
1864      * doesn't allow the destination to know the size to read without fully
1865      * parsing it through each devices load-state code (especially the open
1866      * coded devices that use get/put).
1867      * So we wrap the device state up in a package with a length at the start;
1868      * to do this we use a qemu_buf to hold the whole of the device state.
1869      */
1870     bioc = qio_channel_buffer_new(4096);
1871     qio_channel_set_name(QIO_CHANNEL(bioc), "migration-postcopy-buffer");
1872     fb = qemu_fopen_channel_output(QIO_CHANNEL(bioc));
1873     object_unref(OBJECT(bioc));
1874
1875     /*
1876      * Make sure the receiver can get incoming pages before we send the rest
1877      * of the state
1878      */
1879     qemu_savevm_send_postcopy_listen(fb);
1880
1881     qemu_savevm_state_complete_precopy(fb, false, false);
1882     if (migrate_postcopy_ram()) {
1883         qemu_savevm_send_ping(fb, 3);
1884     }
1885
1886     qemu_savevm_send_postcopy_run(fb);
1887
1888     /* <><> end of stuff going into the package */
1889
1890     /* Last point of recovery; as soon as we send the package the destination
1891      * can open devices and potentially start running.
1892      * Lets just check again we've not got any errors.
1893      */
1894     ret = qemu_file_get_error(ms->to_dst_file);
1895     if (ret) {
1896         error_report("postcopy_start: Migration stream errored (pre package)");
1897         goto fail_closefb;
1898     }
1899
1900     restart_block = false;
1901
1902     /* Now send that blob */
1903     if (qemu_savevm_send_packaged(ms->to_dst_file, bioc->data, bioc->usage)) {
1904         goto fail_closefb;
1905     }
1906     qemu_fclose(fb);
1907
1908     /* Send a notify to give a chance for anything that needs to happen
1909      * at the transition to postcopy and after the device state; in particular
1910      * spice needs to trigger a transition now
1911      */
1912     ms->postcopy_after_devices = true;
1913     notifier_list_notify(&migration_state_notifiers, ms);
1914
1915     ms->downtime =  qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop;
1916
1917     qemu_mutex_unlock_iothread();
1918
1919     if (migrate_postcopy_ram()) {
1920         /*
1921          * Although this ping is just for debug, it could potentially be
1922          * used for getting a better measurement of downtime at the source.
1923          */
1924         qemu_savevm_send_ping(ms->to_dst_file, 4);
1925     }
1926
1927     if (migrate_release_ram()) {
1928         ram_postcopy_migrated_memory_release(ms);
1929     }
1930
1931     ret = qemu_file_get_error(ms->to_dst_file);
1932     if (ret) {
1933         error_report("postcopy_start: Migration stream errored");
1934         migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1935                               MIGRATION_STATUS_FAILED);
1936     }
1937
1938     return ret;
1939
1940 fail_closefb:
1941     qemu_fclose(fb);
1942 fail:
1943     migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1944                           MIGRATION_STATUS_FAILED);
1945     if (restart_block) {
1946         /* A failure happened early enough that we know the destination hasn't
1947          * accessed block devices, so we're safe to recover.
1948          */
1949         Error *local_err = NULL;
1950
1951         bdrv_invalidate_cache_all(&local_err);
1952         if (local_err) {
1953             error_report_err(local_err);
1954         }
1955     }
1956     qemu_mutex_unlock_iothread();
1957     return -1;
1958 }
1959
1960 /**
1961  * migration_maybe_pause: Pause if required to by
1962  * migrate_pause_before_switchover called with the iothread locked
1963  * Returns: 0 on success
1964  */
1965 static int migration_maybe_pause(MigrationState *s, int *current_active_state)
1966 {
1967     if (!migrate_pause_before_switchover()) {
1968         return 0;
1969     }
1970
1971     /* Since leaving this state is not atomic with posting the semaphore
1972      * it's possible that someone could have issued multiple migrate_continue
1973      * and the semaphore is incorrectly positive at this point;
1974      * the docs say it's undefined to reinit a semaphore that's already
1975      * init'd, so use timedwait to eat up any existing posts.
1976      */
1977     while (qemu_sem_timedwait(&s->pause_sem, 1) == 0) {
1978         /* This block intentionally left blank */
1979     }
1980
1981     qemu_mutex_unlock_iothread();
1982     migrate_set_state(&s->state, *current_active_state,
1983                       MIGRATION_STATUS_PRE_SWITCHOVER);
1984     qemu_sem_wait(&s->pause_sem);
1985     migrate_set_state(&s->state, MIGRATION_STATUS_PRE_SWITCHOVER,
1986                       MIGRATION_STATUS_DEVICE);
1987     *current_active_state = MIGRATION_STATUS_DEVICE;
1988     qemu_mutex_lock_iothread();
1989
1990     return s->state == MIGRATION_STATUS_DEVICE ? 0 : -EINVAL;
1991 }
1992
1993 /**
1994  * migration_completion: Used by migration_thread when there's not much left.
1995  *   The caller 'breaks' the loop when this returns.
1996  *
1997  * @s: Current migration state
1998  * @current_active_state: The migration state we expect to be in
1999  * @*old_vm_running: Pointer to old_vm_running flag
2000  * @*start_time: Pointer to time to update
2001  */
2002 static void migration_completion(MigrationState *s, int current_active_state,
2003                                  bool *old_vm_running,
2004                                  int64_t *start_time)
2005 {
2006     int ret;
2007
2008     if (s->state == MIGRATION_STATUS_ACTIVE) {
2009         qemu_mutex_lock_iothread();
2010         *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2011         qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
2012         *old_vm_running = runstate_is_running();
2013         ret = global_state_store();
2014
2015         if (!ret) {
2016             bool inactivate = !migrate_colo_enabled();
2017             ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
2018             if (ret >= 0) {
2019                 ret = migration_maybe_pause(s, &current_active_state);
2020             }
2021             if (ret >= 0) {
2022                 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX);
2023                 ret = qemu_savevm_state_complete_precopy(s->to_dst_file, false,
2024                                                          inactivate);
2025             }
2026             if (inactivate && ret >= 0) {
2027                 s->block_inactive = true;
2028             }
2029         }
2030         qemu_mutex_unlock_iothread();
2031
2032         if (ret < 0) {
2033             goto fail;
2034         }
2035     } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
2036         trace_migration_completion_postcopy_end();
2037
2038         qemu_savevm_state_complete_postcopy(s->to_dst_file);
2039         trace_migration_completion_postcopy_end_after_complete();
2040     }
2041
2042     /*
2043      * If rp was opened we must clean up the thread before
2044      * cleaning everything else up (since if there are no failures
2045      * it will wait for the destination to send it's status in
2046      * a SHUT command).
2047      */
2048     if (s->rp_state.from_dst_file) {
2049         int rp_error;
2050         trace_migration_return_path_end_before();
2051         rp_error = await_return_path_close_on_source(s);
2052         trace_migration_return_path_end_after(rp_error);
2053         if (rp_error) {
2054             goto fail_invalidate;
2055         }
2056     }
2057
2058     if (qemu_file_get_error(s->to_dst_file)) {
2059         trace_migration_completion_file_err();
2060         goto fail_invalidate;
2061     }
2062
2063     if (!migrate_colo_enabled()) {
2064         migrate_set_state(&s->state, current_active_state,
2065                           MIGRATION_STATUS_COMPLETED);
2066     }
2067
2068     return;
2069
2070 fail_invalidate:
2071     /* If not doing postcopy, vm_start() will be called: let's regain
2072      * control on images.
2073      */
2074     if (s->state == MIGRATION_STATUS_ACTIVE) {
2075         Error *local_err = NULL;
2076
2077         qemu_mutex_lock_iothread();
2078         bdrv_invalidate_cache_all(&local_err);
2079         if (local_err) {
2080             error_report_err(local_err);
2081         } else {
2082             s->block_inactive = false;
2083         }
2084         qemu_mutex_unlock_iothread();
2085     }
2086
2087 fail:
2088     migrate_set_state(&s->state, current_active_state,
2089                       MIGRATION_STATUS_FAILED);
2090 }
2091
2092 bool migrate_colo_enabled(void)
2093 {
2094     MigrationState *s = migrate_get_current();
2095     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_COLO];
2096 }
2097
2098 /*
2099  * Master migration thread on the source VM.
2100  * It drives the migration and pumps the data down the outgoing channel.
2101  */
2102 static void *migration_thread(void *opaque)
2103 {
2104     MigrationState *s = opaque;
2105     /* Used by the bandwidth calcs, updated later */
2106     int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2107     int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
2108     int64_t initial_bytes = 0;
2109     /*
2110      * The final stage happens when the remaining data is smaller than
2111      * this threshold; it's calculated from the requested downtime and
2112      * measured bandwidth
2113      */
2114     int64_t threshold_size = 0;
2115     int64_t start_time = initial_time;
2116     int64_t end_time;
2117     bool old_vm_running = false;
2118     bool entered_postcopy = false;
2119     /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */
2120     enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE;
2121     bool enable_colo = migrate_colo_enabled();
2122
2123     rcu_register_thread();
2124
2125     qemu_savevm_state_header(s->to_dst_file);
2126
2127     /*
2128      * If we opened the return path, we need to make sure dst has it
2129      * opened as well.
2130      */
2131     if (s->rp_state.from_dst_file) {
2132         /* Now tell the dest that it should open its end so it can reply */
2133         qemu_savevm_send_open_return_path(s->to_dst_file);
2134
2135         /* And do a ping that will make stuff easier to debug */
2136         qemu_savevm_send_ping(s->to_dst_file, 1);
2137     }
2138
2139     if (migrate_postcopy()) {
2140         /*
2141          * Tell the destination that we *might* want to do postcopy later;
2142          * if the other end can't do postcopy it should fail now, nice and
2143          * early.
2144          */
2145         qemu_savevm_send_postcopy_advise(s->to_dst_file);
2146     }
2147
2148     qemu_savevm_state_setup(s->to_dst_file);
2149
2150     s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
2151     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2152                       MIGRATION_STATUS_ACTIVE);
2153
2154     trace_migration_thread_setup_complete();
2155
2156     while (s->state == MIGRATION_STATUS_ACTIVE ||
2157            s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
2158         int64_t current_time;
2159         uint64_t pending_size;
2160
2161         if (!qemu_file_rate_limit(s->to_dst_file)) {
2162             uint64_t pend_post, pend_nonpost;
2163
2164             qemu_savevm_state_pending(s->to_dst_file, threshold_size,
2165                                       &pend_nonpost, &pend_post);
2166             pending_size = pend_nonpost + pend_post;
2167             trace_migrate_pending(pending_size, threshold_size,
2168                                   pend_post, pend_nonpost);
2169             if (pending_size && pending_size >= threshold_size) {
2170                 /* Still a significant amount to transfer */
2171
2172                 if (migrate_postcopy() &&
2173                     s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE &&
2174                     pend_nonpost <= threshold_size &&
2175                     atomic_read(&s->start_postcopy)) {
2176
2177                     if (!postcopy_start(s, &old_vm_running)) {
2178                         current_active_state = MIGRATION_STATUS_POSTCOPY_ACTIVE;
2179                         entered_postcopy = true;
2180                     }
2181
2182                     continue;
2183                 }
2184                 /* Just another iteration step */
2185                 qemu_savevm_state_iterate(s->to_dst_file, entered_postcopy);
2186             } else {
2187                 trace_migration_thread_low_pending(pending_size);
2188                 migration_completion(s, current_active_state,
2189                                      &old_vm_running, &start_time);
2190                 break;
2191             }
2192         }
2193
2194         if (qemu_file_get_error(s->to_dst_file)) {
2195             migrate_set_state(&s->state, current_active_state,
2196                               MIGRATION_STATUS_FAILED);
2197             trace_migration_thread_file_err();
2198             break;
2199         }
2200         current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2201         if (current_time >= initial_time + BUFFER_DELAY) {
2202             uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) -
2203                                          initial_bytes;
2204             uint64_t time_spent = current_time - initial_time;
2205             double bandwidth = (double)transferred_bytes / time_spent;
2206             threshold_size = bandwidth * s->parameters.downtime_limit;
2207
2208             s->mbps = (((double) transferred_bytes * 8.0) /
2209                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0;
2210
2211             trace_migrate_transferred(transferred_bytes, time_spent,
2212                                       bandwidth, threshold_size);
2213             /* if we haven't sent anything, we don't want to recalculate
2214                10000 is a small enough number for our purposes */
2215             if (ram_counters.dirty_pages_rate && transferred_bytes > 10000) {
2216                 s->expected_downtime = ram_counters.dirty_pages_rate *
2217                     qemu_target_page_size() / bandwidth;
2218             }
2219
2220             qemu_file_reset_rate_limit(s->to_dst_file);
2221             initial_time = current_time;
2222             initial_bytes = qemu_ftell(s->to_dst_file);
2223         }
2224         if (qemu_file_rate_limit(s->to_dst_file)) {
2225             /* usleep expects microseconds */
2226             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
2227         }
2228     }
2229
2230     trace_migration_thread_after_loop();
2231     /* If we enabled cpu throttling for auto-converge, turn it off. */
2232     cpu_throttle_stop();
2233     end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2234
2235     qemu_mutex_lock_iothread();
2236     /*
2237      * The resource has been allocated by migration will be reused in COLO
2238      * process, so don't release them.
2239      */
2240     if (!enable_colo) {
2241         qemu_savevm_state_cleanup();
2242     }
2243     if (s->state == MIGRATION_STATUS_COMPLETED) {
2244         uint64_t transferred_bytes = qemu_ftell(s->to_dst_file);
2245         s->total_time = end_time - s->total_time;
2246         if (!entered_postcopy) {
2247             s->downtime = end_time - start_time;
2248         }
2249         if (s->total_time) {
2250             s->mbps = (((double) transferred_bytes * 8.0) /
2251                        ((double) s->total_time)) / 1000;
2252         }
2253         runstate_set(RUN_STATE_POSTMIGRATE);
2254     } else {
2255         if (s->state == MIGRATION_STATUS_ACTIVE && enable_colo) {
2256             migrate_start_colo_process(s);
2257             qemu_savevm_state_cleanup();
2258             /*
2259             * Fixme: we will run VM in COLO no matter its old running state.
2260             * After exited COLO, we will keep running.
2261             */
2262             old_vm_running = true;
2263         }
2264         if (old_vm_running && !entered_postcopy) {
2265             vm_start();
2266         } else {
2267             if (runstate_check(RUN_STATE_FINISH_MIGRATE)) {
2268                 runstate_set(RUN_STATE_POSTMIGRATE);
2269             }
2270         }
2271     }
2272     qemu_bh_schedule(s->cleanup_bh);
2273     qemu_mutex_unlock_iothread();
2274
2275     rcu_unregister_thread();
2276     return NULL;
2277 }
2278
2279 void migrate_fd_connect(MigrationState *s)
2280 {
2281     s->expected_downtime = s->parameters.downtime_limit;
2282     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
2283
2284     qemu_file_set_blocking(s->to_dst_file, true);
2285     qemu_file_set_rate_limit(s->to_dst_file,
2286                              s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
2287
2288     /* Notify before starting migration thread */
2289     notifier_list_notify(&migration_state_notifiers, s);
2290
2291     /*
2292      * Open the return path. For postcopy, it is used exclusively. For
2293      * precopy, only if user specified "return-path" capability would
2294      * QEMU uses the return path.
2295      */
2296     if (migrate_postcopy_ram() || migrate_use_return_path()) {
2297         if (open_return_path_on_source(s)) {
2298             error_report("Unable to open return-path for postcopy");
2299             migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2300                               MIGRATION_STATUS_FAILED);
2301             migrate_fd_cleanup(s);
2302             return;
2303         }
2304     }
2305
2306     if (multifd_save_setup() != 0) {
2307         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2308                           MIGRATION_STATUS_FAILED);
2309         migrate_fd_cleanup(s);
2310         return;
2311     }
2312     qemu_thread_create(&s->thread, "live_migration", migration_thread, s,
2313                        QEMU_THREAD_JOINABLE);
2314     s->migration_thread_running = true;
2315 }
2316
2317 void migration_global_dump(Monitor *mon)
2318 {
2319     MigrationState *ms = migrate_get_current();
2320
2321     monitor_printf(mon, "globals: store-global-state=%d, only_migratable=%d, "
2322                    "send-configuration=%d, send-section-footer=%d\n",
2323                    ms->store_global_state, ms->only_migratable,
2324                    ms->send_configuration, ms->send_section_footer);
2325 }
2326
2327 #define DEFINE_PROP_MIG_CAP(name, x)             \
2328     DEFINE_PROP_BOOL(name, MigrationState, enabled_capabilities[x], false)
2329
2330 static Property migration_properties[] = {
2331     DEFINE_PROP_BOOL("store-global-state", MigrationState,
2332                      store_global_state, true),
2333     DEFINE_PROP_BOOL("only-migratable", MigrationState, only_migratable, false),
2334     DEFINE_PROP_BOOL("send-configuration", MigrationState,
2335                      send_configuration, true),
2336     DEFINE_PROP_BOOL("send-section-footer", MigrationState,
2337                      send_section_footer, true),
2338
2339     /* Migration parameters */
2340     DEFINE_PROP_INT64("x-compress-level", MigrationState,
2341                       parameters.compress_level,
2342                       DEFAULT_MIGRATE_COMPRESS_LEVEL),
2343     DEFINE_PROP_INT64("x-compress-threads", MigrationState,
2344                       parameters.compress_threads,
2345                       DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT),
2346     DEFINE_PROP_INT64("x-decompress-threads", MigrationState,
2347                       parameters.decompress_threads,
2348                       DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT),
2349     DEFINE_PROP_INT64("x-cpu-throttle-initial", MigrationState,
2350                       parameters.cpu_throttle_initial,
2351                       DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL),
2352     DEFINE_PROP_INT64("x-cpu-throttle-increment", MigrationState,
2353                       parameters.cpu_throttle_increment,
2354                       DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT),
2355     DEFINE_PROP_INT64("x-max-bandwidth", MigrationState,
2356                       parameters.max_bandwidth, MAX_THROTTLE),
2357     DEFINE_PROP_INT64("x-downtime-limit", MigrationState,
2358                       parameters.downtime_limit,
2359                       DEFAULT_MIGRATE_SET_DOWNTIME),
2360     DEFINE_PROP_INT64("x-checkpoint-delay", MigrationState,
2361                       parameters.x_checkpoint_delay,
2362                       DEFAULT_MIGRATE_X_CHECKPOINT_DELAY),
2363     DEFINE_PROP_INT64("x-multifd-channels", MigrationState,
2364                       parameters.x_multifd_channels,
2365                       DEFAULT_MIGRATE_MULTIFD_CHANNELS),
2366     DEFINE_PROP_INT64("x-multifd-page-count", MigrationState,
2367                       parameters.x_multifd_page_count,
2368                       DEFAULT_MIGRATE_MULTIFD_PAGE_COUNT),
2369
2370     /* Migration capabilities */
2371     DEFINE_PROP_MIG_CAP("x-xbzrle", MIGRATION_CAPABILITY_XBZRLE),
2372     DEFINE_PROP_MIG_CAP("x-rdma-pin-all", MIGRATION_CAPABILITY_RDMA_PIN_ALL),
2373     DEFINE_PROP_MIG_CAP("x-auto-converge", MIGRATION_CAPABILITY_AUTO_CONVERGE),
2374     DEFINE_PROP_MIG_CAP("x-zero-blocks", MIGRATION_CAPABILITY_ZERO_BLOCKS),
2375     DEFINE_PROP_MIG_CAP("x-compress", MIGRATION_CAPABILITY_COMPRESS),
2376     DEFINE_PROP_MIG_CAP("x-events", MIGRATION_CAPABILITY_EVENTS),
2377     DEFINE_PROP_MIG_CAP("x-postcopy-ram", MIGRATION_CAPABILITY_POSTCOPY_RAM),
2378     DEFINE_PROP_MIG_CAP("x-colo", MIGRATION_CAPABILITY_X_COLO),
2379     DEFINE_PROP_MIG_CAP("x-release-ram", MIGRATION_CAPABILITY_RELEASE_RAM),
2380     DEFINE_PROP_MIG_CAP("x-block", MIGRATION_CAPABILITY_BLOCK),
2381     DEFINE_PROP_MIG_CAP("x-return-path", MIGRATION_CAPABILITY_RETURN_PATH),
2382     DEFINE_PROP_MIG_CAP("x-multifd", MIGRATION_CAPABILITY_X_MULTIFD),
2383
2384     DEFINE_PROP_END_OF_LIST(),
2385 };
2386
2387 static void migration_class_init(ObjectClass *klass, void *data)
2388 {
2389     DeviceClass *dc = DEVICE_CLASS(klass);
2390
2391     dc->user_creatable = false;
2392     dc->props = migration_properties;
2393 }
2394
2395 static void migration_instance_finalize(Object *obj)
2396 {
2397     MigrationState *ms = MIGRATION_OBJ(obj);
2398     MigrationParameters *params = &ms->parameters;
2399
2400     g_free(params->tls_hostname);
2401     g_free(params->tls_creds);
2402     qemu_sem_destroy(&ms->pause_sem);
2403 }
2404
2405 static void migration_instance_init(Object *obj)
2406 {
2407     MigrationState *ms = MIGRATION_OBJ(obj);
2408     MigrationParameters *params = &ms->parameters;
2409
2410     ms->state = MIGRATION_STATUS_NONE;
2411     ms->xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE;
2412     ms->mbps = -1;
2413     qemu_sem_init(&ms->pause_sem, 0);
2414
2415     params->tls_hostname = g_strdup("");
2416     params->tls_creds = g_strdup("");
2417
2418     /* Set has_* up only for parameter checks */
2419     params->has_compress_level = true;
2420     params->has_compress_threads = true;
2421     params->has_decompress_threads = true;
2422     params->has_cpu_throttle_initial = true;
2423     params->has_cpu_throttle_increment = true;
2424     params->has_max_bandwidth = true;
2425     params->has_downtime_limit = true;
2426     params->has_x_checkpoint_delay = true;
2427     params->has_block_incremental = true;
2428     params->has_x_multifd_channels = true;
2429     params->has_x_multifd_page_count = true;
2430 }
2431
2432 /*
2433  * Return true if check pass, false otherwise. Error will be put
2434  * inside errp if provided.
2435  */
2436 static bool migration_object_check(MigrationState *ms, Error **errp)
2437 {
2438     MigrationCapabilityStatusList *head = NULL;
2439     /* Assuming all off */
2440     bool cap_list[MIGRATION_CAPABILITY__MAX] = { 0 }, ret;
2441     int i;
2442
2443     if (!migrate_params_check(&ms->parameters, errp)) {
2444         return false;
2445     }
2446
2447     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
2448         if (ms->enabled_capabilities[i]) {
2449             head = migrate_cap_add(head, i, true);
2450         }
2451     }
2452
2453     ret = migrate_caps_check(cap_list, head, errp);
2454
2455     /* It works with head == NULL */
2456     qapi_free_MigrationCapabilityStatusList(head);
2457
2458     return ret;
2459 }
2460
2461 static const TypeInfo migration_type = {
2462     .name = TYPE_MIGRATION,
2463     /*
2464      * NOTE: TYPE_MIGRATION is not really a device, as the object is
2465      * not created using qdev_create(), it is not attached to the qdev
2466      * device tree, and it is never realized.
2467      *
2468      * TODO: Make this TYPE_OBJECT once QOM provides something like
2469      * TYPE_DEVICE's "-global" properties.
2470      */
2471     .parent = TYPE_DEVICE,
2472     .class_init = migration_class_init,
2473     .class_size = sizeof(MigrationClass),
2474     .instance_size = sizeof(MigrationState),
2475     .instance_init = migration_instance_init,
2476     .instance_finalize = migration_instance_finalize,
2477 };
2478
2479 static void register_migration_types(void)
2480 {
2481     type_register_static(&migration_type);
2482 }
2483
2484 type_init(register_migration_types);