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/error.h"
  34 #include "qapi/qapi-commands-migration.h"
  35 #include "qapi/qapi-events-migration.h"
  36 #include "qapi/qmp/qerror.h"
  37 #include "qapi/qmp/qnull.h"
  38 #include "qemu/rcu.h"
  39 #include "block.h"
  40 #include "postcopy-ram.h"
  41 #include "qemu/thread.h"
  42 #include "trace.h"
  43 #include "exec/target_page.h"
  44 #include "io/channel-buffer.h"
  45 #include "migration/colo.h"
  46 #include "hw/boards.h"
  47 #include "monitor/monitor.h"
  48
  49 #define MAX_THROTTLE  (32 << 20)      /* Migration transfer speed throttling */
  50
  51 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
  52  * data. */
  53 #define BUFFER_DELAY     100
  54 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
  55
  56 /* Time in milliseconds we are allowed to stop the source,
  57  * for sending the last part */
  58 #define DEFAULT_MIGRATE_SET_DOWNTIME 300
  59
  60 /* Maximum migrate downtime set to 2000 seconds */
  61 #define MAX_MIGRATE_DOWNTIME_SECONDS 2000
  62 #define MAX_MIGRATE_DOWNTIME (MAX_MIGRATE_DOWNTIME_SECONDS * 1000)
  63
  64 /* Default compression thread count */
  65 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
  66 /* Default decompression thread count, usually decompression is at
  67  * least 4 times as fast as compression.*/
  68 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2
  69 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */
  70 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1
  71 /* Define default autoconverge cpu throttle migration parameters */
  72 #define DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL 20
  73 #define DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT 10
  74
  75 /* Migration XBZRLE default cache size */
  76 #define DEFAULT_MIGRATE_XBZRLE_CACHE_SIZE (64 * 1024 * 1024)
  77
  78 /* The delay time (in ms) between two COLO checkpoints
  79  * Note: Please change this default value to 10000 when we support hybrid mode.
  80  */
  81 #define DEFAULT_MIGRATE_X_CHECKPOINT_DELAY 200
  82 #define DEFAULT_MIGRATE_MULTIFD_CHANNELS 2
  83 #define DEFAULT_MIGRATE_MULTIFD_PAGE_COUNT 16
  84
  85 static NotifierList migration_state_notifiers =
  86     NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
  87
  88 static bool deferred_incoming;
  89
  90 /* Messages sent on the return path from destination to source */
  91 enum mig_rp_message_type {
  92     MIG_RP_MSG_INVALID = 0,  /* Must be 0 */
  93     MIG_RP_MSG_SHUT,         /* sibling will not send any more RP messages */
  94     MIG_RP_MSG_PONG,         /* Response to a PING; data (seq: be32 ) */
  95
  96     MIG_RP_MSG_REQ_PAGES_ID, /* data (start: be64, len: be32, id: string) */
  97     MIG_RP_MSG_REQ_PAGES,    /* data (start: be64, len: be32) */
  98
  99     MIG_RP_MSG_MAX
 100 };
 101
 102 /* When we add fault tolerance, we could have several
 103    migrations at once.  For now we don't need to add
 104    dynamic creation of migration */
 105
 106 static MigrationState *current_migration;
 107
 108 static bool migration_object_check(MigrationState *ms, Error **errp);
 109 static int migration_maybe_pause(MigrationState *s,
 110                                  int *current_active_state,
 111                                  int new_state);
 112
 113 void migration_object_init(void)
 114 {
 115     MachineState *ms = MACHINE(qdev_get_machine());
 116     Error *err = NULL;
 117
 118     /* This can only be called once. */
 119     assert(!current_migration);
 120     current_migration = MIGRATION_OBJ(object_new(TYPE_MIGRATION));
 121
 122     if (!migration_object_check(current_migration, &err)) {
 123         error_report_err(err);
 124         exit(1);
 125     }
 126
 127     /*
 128      * We cannot really do this in migration_instance_init() since at
 129      * that time global properties are not yet applied, then this
 130      * value will be definitely replaced by something else.
 131      */
 132     if (ms->enforce_config_section) {
 133         current_migration->send_configuration = true;
 134     }
 135 }
 136
 137 void migration_object_finalize(void)
 138 {
 139     object_unref(OBJECT(current_migration));
 140 }
 141
 142 /* For outgoing */
 143 MigrationState *migrate_get_current(void)
 144 {
 145     /* This can only be called after the object created. */
 146     assert(current_migration);
 147     return current_migration;
 148 }
 149
 150 MigrationIncomingState *migration_incoming_get_current(void)
 151 {
 152     static bool once;
 153     static MigrationIncomingState mis_current;
 154
 155     if (!once) {
 156         mis_current.state = MIGRATION_STATUS_NONE;
 157         memset(&mis_current, 0, sizeof(MigrationIncomingState));
 158         qemu_mutex_init(&mis_current.rp_mutex);
 159         qemu_event_init(&mis_current.main_thread_load_event, false);
 160         once = true;
 161     }
 162     return &mis_current;
 163 }
 164
 165 void migration_incoming_state_destroy(void)
 166 {
 167     struct MigrationIncomingState *mis = migration_incoming_get_current();
 168
 169     if (mis->to_src_file) {
 170         /* Tell source that we are done */
 171         migrate_send_rp_shut(mis, qemu_file_get_error(mis->from_src_file) != 0);
 172         qemu_fclose(mis->to_src_file);
 173         mis->to_src_file = NULL;
 174     }
 175
 176     if (mis->from_src_file) {
 177         qemu_fclose(mis->from_src_file);
 178         mis->from_src_file = NULL;
 179     }
 180
 181     qemu_event_reset(&mis->main_thread_load_event);
 182 }
 183
 184 static void migrate_generate_event(int new_state)
 185 {
 186     if (migrate_use_events()) {
 187         qapi_event_send_migration(new_state, &error_abort);
 188     }
 189 }
 190
 191 /*
 192  * Called on -incoming with a defer: uri.
 193  * The migration can be started later after any parameters have been
 194  * changed.
 195  */
 196 static void deferred_incoming_migration(Error **errp)
 197 {
 198     if (deferred_incoming) {
 199         error_setg(errp, "Incoming migration already deferred");
 200     }
 201     deferred_incoming = true;
 202 }
 203
 204 /*
 205  * Send a message on the return channel back to the source
 206  * of the migration.
 207  */
 208 static int migrate_send_rp_message(MigrationIncomingState *mis,
 209                                    enum mig_rp_message_type message_type,
 210                                    uint16_t len, void *data)
 211 {
 212     int ret = 0;
 213
 214     trace_migrate_send_rp_message((int)message_type, len);
 215     qemu_mutex_lock(&mis->rp_mutex);
 216
 217     /*
 218      * It's possible that the file handle got lost due to network
 219      * failures.
 220      */
 221     if (!mis->to_src_file) {
 222         ret = -EIO;
 223         goto error;
 224     }
 225
 226     qemu_put_be16(mis->to_src_file, (unsigned int)message_type);
 227     qemu_put_be16(mis->to_src_file, len);
 228     qemu_put_buffer(mis->to_src_file, data, len);
 229     qemu_fflush(mis->to_src_file);
 230
 231     /* It's possible that qemu file got error during sending */
 232     ret = qemu_file_get_error(mis->to_src_file);
 233
 234 error:
 235     qemu_mutex_unlock(&mis->rp_mutex);
 236     return ret;
 237 }
 238
 239 /* Request a range of pages from the source VM at the given
 240  * start address.
 241  *   rbname: Name of the RAMBlock to request the page in, if NULL it's the same
 242  *           as the last request (a name must have been given previously)
 243  *   Start: Address offset within the RB
 244  *   Len: Length in bytes required - must be a multiple of pagesize
 245  */
 246 int migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname,
 247                               ram_addr_t start, size_t len)
 248 {
 249     uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
 250     size_t msglen = 12; /* start + len */
 251     enum mig_rp_message_type msg_type;
 252
 253     *(uint64_t *)bufc = cpu_to_be64((uint64_t)start);
 254     *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len);
 255
 256     if (rbname) {
 257         int rbname_len = strlen(rbname);
 258         assert(rbname_len < 256);
 259
 260         bufc[msglen++] = rbname_len;
 261         memcpy(bufc + msglen, rbname, rbname_len);
 262         msglen += rbname_len;
 263         msg_type = MIG_RP_MSG_REQ_PAGES_ID;
 264     } else {
 265         msg_type = MIG_RP_MSG_REQ_PAGES;
 266     }
 267
 268     return migrate_send_rp_message(mis, msg_type, msglen, bufc);
 269 }
 270
 271 void qemu_start_incoming_migration(const char *uri, Error **errp)
 272 {
 273     const char *p;
 274
 275     qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort);
 276     if (!strcmp(uri, "defer")) {
 277         deferred_incoming_migration(errp);
 278     } else if (strstart(uri, "tcp:", &p)) {
 279         tcp_start_incoming_migration(p, errp);
 280 #ifdef CONFIG_RDMA
 281     } else if (strstart(uri, "rdma:", &p)) {
 282         rdma_start_incoming_migration(p, errp);
 283 #endif
 284     } else if (strstart(uri, "exec:", &p)) {
 285         exec_start_incoming_migration(p, errp);
 286     } else if (strstart(uri, "unix:", &p)) {
 287         unix_start_incoming_migration(p, errp);
 288     } else if (strstart(uri, "fd:", &p)) {
 289         fd_start_incoming_migration(p, errp);
 290     } else {
 291         error_setg(errp, "unknown migration protocol: %s", uri);
 292     }
 293 }
 294
 295 static void process_incoming_migration_bh(void *opaque)
 296 {
 297     Error *local_err = NULL;
 298     MigrationIncomingState *mis = opaque;
 299
 300     /* Make sure all file formats flush their mutable metadata.
 301      * If we get an error here, just don't restart the VM yet. */
 302     bdrv_invalidate_cache_all(&local_err);
 303     if (local_err) {
 304         error_report_err(local_err);
 305         local_err = NULL;
 306         autostart = false;
 307     }
 308
 309     /*
 310      * This must happen after all error conditions are dealt with and
 311      * we're sure the VM is going to be running on this host.
 312      */
 313     qemu_announce_self();
 314
 315     if (multifd_load_cleanup(&local_err) != 0) {
 316         error_report_err(local_err);
 317         autostart = false;
 318     }
 319     /* If global state section was not received or we are in running
 320        state, we need to obey autostart. Any other state is set with
 321        runstate_set. */
 322
 323     if (!global_state_received() ||
 324         global_state_get_runstate() == RUN_STATE_RUNNING) {
 325         if (autostart) {
 326             vm_start();
 327         } else {
 328             runstate_set(RUN_STATE_PAUSED);
 329         }
 330     } else {
 331         runstate_set(global_state_get_runstate());
 332     }
 333     /*
 334      * This must happen after any state changes since as soon as an external
 335      * observer sees this event they might start to prod at the VM assuming
 336      * it's ready to use.
 337      */
 338     migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 339                       MIGRATION_STATUS_COMPLETED);
 340     qemu_bh_delete(mis->bh);
 341     migration_incoming_state_destroy();
 342 }
 343
 344 static void process_incoming_migration_co(void *opaque)
 345 {
 346     MigrationIncomingState *mis = migration_incoming_get_current();
 347     PostcopyState ps;
 348     int ret;
 349
 350     assert(mis->from_src_file);
 351     mis->largest_page_size = qemu_ram_pagesize_largest();
 352     postcopy_state_set(POSTCOPY_INCOMING_NONE);
 353     migrate_set_state(&mis->state, MIGRATION_STATUS_NONE,
 354                       MIGRATION_STATUS_ACTIVE);
 355     ret = qemu_loadvm_state(mis->from_src_file);
 356
 357     ps = postcopy_state_get();
 358     trace_process_incoming_migration_co_end(ret, ps);
 359     if (ps != POSTCOPY_INCOMING_NONE) {
 360         if (ps == POSTCOPY_INCOMING_ADVISE) {
 361             /*
 362              * Where a migration had postcopy enabled (and thus went to advise)
 363              * but managed to complete within the precopy period, we can use
 364              * the normal exit.
 365              */
 366             postcopy_ram_incoming_cleanup(mis);
 367         } else if (ret >= 0) {
 368             /*
 369              * Postcopy was started, cleanup should happen at the end of the
 370              * postcopy thread.
 371              */
 372             trace_process_incoming_migration_co_postcopy_end_main();
 373             return;
 374         }
 375         /* Else if something went wrong then just fall out of the normal exit */
 376     }
 377
 378     /* we get COLO info, and know if we are in COLO mode */
 379     if (!ret && migration_incoming_enable_colo()) {
 380         mis->migration_incoming_co = qemu_coroutine_self();
 381         qemu_thread_create(&mis->colo_incoming_thread, "COLO incoming",
 382              colo_process_incoming_thread, mis, QEMU_THREAD_JOINABLE);
 383         mis->have_colo_incoming_thread = true;
 384         qemu_coroutine_yield();
 385
 386         /* Wait checkpoint incoming thread exit before free resource */
 387         qemu_thread_join(&mis->colo_incoming_thread);
 388     }
 389
 390     if (ret < 0) {
 391         Error *local_err = NULL;
 392
 393         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 394                           MIGRATION_STATUS_FAILED);
 395         error_report("load of migration failed: %s", strerror(-ret));
 396         qemu_fclose(mis->from_src_file);
 397         if (multifd_load_cleanup(&local_err) != 0) {
 398             error_report_err(local_err);
 399         }
 400         exit(EXIT_FAILURE);
 401     }
 402     mis->bh = qemu_bh_new(process_incoming_migration_bh, mis);
 403     qemu_bh_schedule(mis->bh);
 404 }
 405
 406 static void migration_incoming_setup(QEMUFile *f)
 407 {
 408     MigrationIncomingState *mis = migration_incoming_get_current();
 409
 410     if (multifd_load_setup() != 0) {
 411         /* We haven't been able to create multifd threads
 412            nothing better to do */
 413         exit(EXIT_FAILURE);
 414     }
 415
 416     if (!mis->from_src_file) {
 417         mis->from_src_file = f;
 418     }
 419     qemu_file_set_blocking(f, false);
 420 }
 421
 422 static void migration_incoming_process(void)
 423 {
 424     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co, NULL);
 425     qemu_coroutine_enter(co);
 426 }
 427
 428 void migration_fd_process_incoming(QEMUFile *f)
 429 {
 430     migration_incoming_setup(f);
 431     migration_incoming_process();
 432 }
 433
 434 void migration_ioc_process_incoming(QIOChannel *ioc)
 435 {
 436     MigrationIncomingState *mis = migration_incoming_get_current();
 437
 438     if (!mis->from_src_file) {
 439         QEMUFile *f = qemu_fopen_channel_input(ioc);
 440         migration_fd_process_incoming(f);
 441     }
 442     /* We still only have a single channel.  Nothing to do here yet */
 443 }
 444
 445 /**
 446  * @migration_has_all_channels: We have received all channels that we need
 447  *
 448  * Returns true when we have got connections to all the channels that
 449  * we need for migration.
 450  */
 451 bool migration_has_all_channels(void)
 452 {
 453     return true;
 454 }
 455
 456 /*
 457  * Send a 'SHUT' message on the return channel with the given value
 458  * to indicate that we've finished with the RP.  Non-0 value indicates
 459  * error.
 460  */
 461 void migrate_send_rp_shut(MigrationIncomingState *mis,
 462                           uint32_t value)
 463 {
 464     uint32_t buf;
 465
 466     buf = cpu_to_be32(value);
 467     migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf);
 468 }
 469
 470 /*
 471  * Send a 'PONG' message on the return channel with the given value
 472  * (normally in response to a 'PING')
 473  */
 474 void migrate_send_rp_pong(MigrationIncomingState *mis,
 475                           uint32_t value)
 476 {
 477     uint32_t buf;
 478
 479     buf = cpu_to_be32(value);
 480     migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf);
 481 }
 482
 483 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
 484 {
 485     MigrationCapabilityStatusList *head = NULL;
 486     MigrationCapabilityStatusList *caps;
 487     MigrationState *s = migrate_get_current();
 488     int i;
 489
 490     caps = NULL; /* silence compiler warning */
 491     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
 492 #ifndef CONFIG_LIVE_BLOCK_MIGRATION
 493         if (i == MIGRATION_CAPABILITY_BLOCK) {
 494             continue;
 495         }
 496 #endif
 497         if (head == NULL) {
 498             head = g_malloc0(sizeof(*caps));
 499             caps = head;
 500         } else {
 501             caps->next = g_malloc0(sizeof(*caps));
 502             caps = caps->next;
 503         }
 504         caps->value =
 505             g_malloc(sizeof(*caps->value));
 506         caps->value->capability = i;
 507         caps->value->state = s->enabled_capabilities[i];
 508     }
 509
 510     return head;
 511 }
 512
 513 MigrationParameters *qmp_query_migrate_parameters(Error **errp)
 514 {
 515     MigrationParameters *params;
 516     MigrationState *s = migrate_get_current();
 517
 518     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 519     params = g_malloc0(sizeof(*params));
 520     params->has_compress_level = true;
 521     params->compress_level = s->parameters.compress_level;
 522     params->has_compress_threads = true;
 523     params->compress_threads = s->parameters.compress_threads;
 524     params->has_decompress_threads = true;
 525     params->decompress_threads = s->parameters.decompress_threads;
 526     params->has_cpu_throttle_initial = true;
 527     params->cpu_throttle_initial = s->parameters.cpu_throttle_initial;
 528     params->has_cpu_throttle_increment = true;
 529     params->cpu_throttle_increment = s->parameters.cpu_throttle_increment;
 530     params->has_tls_creds = true;
 531     params->tls_creds = g_strdup(s->parameters.tls_creds);
 532     params->has_tls_hostname = true;
 533     params->tls_hostname = g_strdup(s->parameters.tls_hostname);
 534     params->has_max_bandwidth = true;
 535     params->max_bandwidth = s->parameters.max_bandwidth;
 536     params->has_downtime_limit = true;
 537     params->downtime_limit = s->parameters.downtime_limit;
 538     params->has_x_checkpoint_delay = true;
 539     params->x_checkpoint_delay = s->parameters.x_checkpoint_delay;
 540     params->has_block_incremental = true;
 541     params->block_incremental = s->parameters.block_incremental;
 542     params->has_x_multifd_channels = true;
 543     params->x_multifd_channels = s->parameters.x_multifd_channels;
 544     params->has_x_multifd_page_count = true;
 545     params->x_multifd_page_count = s->parameters.x_multifd_page_count;
 546     params->has_xbzrle_cache_size = true;
 547     params->xbzrle_cache_size = s->parameters.xbzrle_cache_size;
 548
 549     return params;
 550 }
 551
 552 /*
 553  * Return true if we're already in the middle of a migration
 554  * (i.e. any of the active or setup states)
 555  */
 556 static bool migration_is_setup_or_active(int state)
 557 {
 558     switch (state) {
 559     case MIGRATION_STATUS_ACTIVE:
 560     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 561     case MIGRATION_STATUS_SETUP:
 562     case MIGRATION_STATUS_PRE_SWITCHOVER:
 563     case MIGRATION_STATUS_DEVICE:
 564         return true;
 565
 566     default:
 567         return false;
 568
 569     }
 570 }
 571
 572 static void populate_ram_info(MigrationInfo *info, MigrationState *s)
 573 {
 574     info->has_ram = true;
 575     info->ram = g_malloc0(sizeof(*info->ram));
 576     info->ram->transferred = ram_counters.transferred;
 577     info->ram->total = ram_bytes_total();
 578     info->ram->duplicate = ram_counters.duplicate;
 579     /* legacy value.  It is not used anymore */
 580     info->ram->skipped = 0;
 581     info->ram->normal = ram_counters.normal;
 582     info->ram->normal_bytes = ram_counters.normal *
 583         qemu_target_page_size();
 584     info->ram->mbps = s->mbps;
 585     info->ram->dirty_sync_count = ram_counters.dirty_sync_count;
 586     info->ram->postcopy_requests = ram_counters.postcopy_requests;
 587     info->ram->page_size = qemu_target_page_size();
 588
 589     if (migrate_use_xbzrle()) {
 590         info->has_xbzrle_cache = true;
 591         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
 592         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
 593         info->xbzrle_cache->bytes = xbzrle_counters.bytes;
 594         info->xbzrle_cache->pages = xbzrle_counters.pages;
 595         info->xbzrle_cache->cache_miss = xbzrle_counters.cache_miss;
 596         info->xbzrle_cache->cache_miss_rate = xbzrle_counters.cache_miss_rate;
 597         info->xbzrle_cache->overflow = xbzrle_counters.overflow;
 598     }
 599
 600     if (cpu_throttle_active()) {
 601         info->has_cpu_throttle_percentage = true;
 602         info->cpu_throttle_percentage = cpu_throttle_get_percentage();
 603     }
 604
 605     if (s->state != MIGRATION_STATUS_COMPLETED) {
 606         info->ram->remaining = ram_bytes_remaining();
 607         info->ram->dirty_pages_rate = ram_counters.dirty_pages_rate;
 608     }
 609 }
 610
 611 static void populate_disk_info(MigrationInfo *info)
 612 {
 613     if (blk_mig_active()) {
 614         info->has_disk = true;
 615         info->disk = g_malloc0(sizeof(*info->disk));
 616         info->disk->transferred = blk_mig_bytes_transferred();
 617         info->disk->remaining = blk_mig_bytes_remaining();
 618         info->disk->total = blk_mig_bytes_total();
 619     }
 620 }
 621
 622 MigrationInfo *qmp_query_migrate(Error **errp)
 623 {
 624     MigrationInfo *info = g_malloc0(sizeof(*info));
 625     MigrationState *s = migrate_get_current();
 626
 627     switch (s->state) {
 628     case MIGRATION_STATUS_NONE:
 629         /* no migration has happened ever */
 630         break;
 631     case MIGRATION_STATUS_SETUP:
 632         info->has_status = true;
 633         info->has_total_time = false;
 634         break;
 635     case MIGRATION_STATUS_ACTIVE:
 636     case MIGRATION_STATUS_CANCELLING:
 637     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 638     case MIGRATION_STATUS_PRE_SWITCHOVER:
 639     case MIGRATION_STATUS_DEVICE:
 640          /* TODO add some postcopy stats */
 641         info->has_status = true;
 642         info->has_total_time = true;
 643         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 644             - s->start_time;
 645         info->has_expected_downtime = true;
 646         info->expected_downtime = s->expected_downtime;
 647         info->has_setup_time = true;
 648         info->setup_time = s->setup_time;
 649
 650         populate_ram_info(info, s);
 651         populate_disk_info(info);
 652         break;
 653     case MIGRATION_STATUS_COLO:
 654         info->has_status = true;
 655         /* TODO: display COLO specific information (checkpoint info etc.) */
 656         break;
 657     case MIGRATION_STATUS_COMPLETED:
 658         info->has_status = true;
 659         info->has_total_time = true;
 660         info->total_time = s->total_time;
 661         info->has_downtime = true;
 662         info->downtime = s->downtime;
 663         info->has_setup_time = true;
 664         info->setup_time = s->setup_time;
 665
 666         populate_ram_info(info, s);
 667         break;
 668     case MIGRATION_STATUS_FAILED:
 669         info->has_status = true;
 670         if (s->error) {
 671             info->has_error_desc = true;
 672             info->error_desc = g_strdup(error_get_pretty(s->error));
 673         }
 674         break;
 675     case MIGRATION_STATUS_CANCELLED:
 676         info->has_status = true;
 677         break;
 678     }
 679     info->status = s->state;
 680
 681     return info;
 682 }
 683
 684 /**
 685  * @migration_caps_check - check capability validity
 686  *
 687  * @cap_list: old capability list, array of bool
 688  * @params: new capabilities to be applied soon
 689  * @errp: set *errp if the check failed, with reason
 690  *
 691  * Returns true if check passed, otherwise false.
 692  */
 693 static bool migrate_caps_check(bool *cap_list,
 694                                MigrationCapabilityStatusList *params,
 695                                Error **errp)
 696 {
 697     MigrationCapabilityStatusList *cap;
 698     bool old_postcopy_cap;
 699     MigrationIncomingState *mis = migration_incoming_get_current();
 700
 701     old_postcopy_cap = cap_list[MIGRATION_CAPABILITY_POSTCOPY_RAM];
 702
 703     for (cap = params; cap; cap = cap->next) {
 704         cap_list[cap->value->capability] = cap->value->state;
 705     }
 706
 707 #ifndef CONFIG_LIVE_BLOCK_MIGRATION
 708     if (cap_list[MIGRATION_CAPABILITY_BLOCK]) {
 709         error_setg(errp, "QEMU compiled without old-style (blk/-b, inc/-i) "
 710                    "block migration");
 711         error_append_hint(errp, "Use drive_mirror+NBD instead.\n");
 712         return false;
 713     }
 714 #endif
 715
 716     if (cap_list[MIGRATION_CAPABILITY_POSTCOPY_RAM]) {
 717         if (cap_list[MIGRATION_CAPABILITY_COMPRESS]) {
 718             /* The decompression threads asynchronously write into RAM
 719              * rather than use the atomic copies needed to avoid
 720              * userfaulting.  It should be possible to fix the decompression
 721              * threads for compatibility in future.
 722              */
 723             error_setg(errp, "Postcopy is not currently compatible "
 724                        "with compression");
 725             return false;
 726         }
 727
 728         /* This check is reasonably expensive, so only when it's being
 729          * set the first time, also it's only the destination that needs
 730          * special support.
 731          */
 732         if (!old_postcopy_cap && runstate_check(RUN_STATE_INMIGRATE) &&
 733             !postcopy_ram_supported_by_host(mis)) {
 734             /* postcopy_ram_supported_by_host will have emitted a more
 735              * detailed message
 736              */
 737             error_setg(errp, "Postcopy is not supported");
 738             return false;
 739         }
 740     }
 741
 742     return true;
 743 }
 744
 745 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 746                                   Error **errp)
 747 {
 748     MigrationState *s = migrate_get_current();
 749     MigrationCapabilityStatusList *cap;
 750     bool cap_list[MIGRATION_CAPABILITY__MAX];
 751
 752     if (migration_is_setup_or_active(s->state)) {
 753         error_setg(errp, QERR_MIGRATION_ACTIVE);
 754         return;
 755     }
 756
 757     memcpy(cap_list, s->enabled_capabilities, sizeof(cap_list));
 758     if (!migrate_caps_check(cap_list, params, errp)) {
 759         return;
 760     }
 761
 762     for (cap = params; cap; cap = cap->next) {
 763         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 764     }
 765 }
 766
 767 /*
 768  * Check whether the parameters are valid. Error will be put into errp
 769  * (if provided). Return true if valid, otherwise false.
 770  */
 771 static bool migrate_params_check(MigrationParameters *params, Error **errp)
 772 {
 773     if (params->has_compress_level &&
 774         (params->compress_level > 9)) {
 775         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level",
 776                    "is invalid, it should be in the range of 0 to 9");
 777         return false;
 778     }
 779
 780     if (params->has_compress_threads && (params->compress_threads < 1)) {
 781         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 782                    "compress_threads",
 783                    "is invalid, it should be in the range of 1 to 255");
 784         return false;
 785     }
 786
 787     if (params->has_decompress_threads && (params->decompress_threads < 1)) {
 788         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 789                    "decompress_threads",
 790                    "is invalid, it should be in the range of 1 to 255");
 791         return false;
 792     }
 793
 794     if (params->has_cpu_throttle_initial &&
 795         (params->cpu_throttle_initial < 1 ||
 796          params->cpu_throttle_initial > 99)) {
 797         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 798                    "cpu_throttle_initial",
 799                    "an integer in the range of 1 to 99");
 800         return false;
 801     }
 802
 803     if (params->has_cpu_throttle_increment &&
 804         (params->cpu_throttle_increment < 1 ||
 805          params->cpu_throttle_increment > 99)) {
 806         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 807                    "cpu_throttle_increment",
 808                    "an integer in the range of 1 to 99");
 809         return false;
 810     }
 811
 812     if (params->has_max_bandwidth && (params->max_bandwidth > SIZE_MAX)) {
 813         error_setg(errp, "Parameter 'max_bandwidth' expects an integer in the"
 814                          " range of 0 to %zu bytes/second", SIZE_MAX);
 815         return false;
 816     }
 817
 818     if (params->has_downtime_limit &&
 819         (params->downtime_limit > MAX_MIGRATE_DOWNTIME)) {
 820         error_setg(errp, "Parameter 'downtime_limit' expects an integer in "
 821                          "the range of 0 to %d milliseconds",
 822                          MAX_MIGRATE_DOWNTIME);
 823         return false;
 824     }
 825
 826     /* x_checkpoint_delay is now always positive */
 827
 828     if (params->has_x_multifd_channels && (params->x_multifd_channels < 1)) {
 829         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 830                    "multifd_channels",
 831                    "is invalid, it should be in the range of 1 to 255");
 832         return false;
 833     }
 834     if (params->has_x_multifd_page_count &&
 835         (params->x_multifd_page_count < 1 ||
 836          params->x_multifd_page_count > 10000)) {
 837         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 838                    "multifd_page_count",
 839                    "is invalid, it should be in the range of 1 to 10000");
 840         return false;
 841     }
 842
 843     if (params->has_xbzrle_cache_size &&
 844         (params->xbzrle_cache_size < qemu_target_page_size() ||
 845          !is_power_of_2(params->xbzrle_cache_size))) {
 846         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 847                    "xbzrle_cache_size",
 848                    "is invalid, it should be bigger than target page size"
 849                    " and a power of two");
 850         return false;
 851     }
 852
 853     return true;
 854 }
 855
 856 static void migrate_params_test_apply(MigrateSetParameters *params,
 857                                       MigrationParameters *dest)
 858 {
 859     *dest = migrate_get_current()->parameters;
 860
 861     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 862
 863     if (params->has_compress_level) {
 864         dest->compress_level = params->compress_level;
 865     }
 866
 867     if (params->has_compress_threads) {
 868         dest->compress_threads = params->compress_threads;
 869     }
 870
 871     if (params->has_decompress_threads) {
 872         dest->decompress_threads = params->decompress_threads;
 873     }
 874
 875     if (params->has_cpu_throttle_initial) {
 876         dest->cpu_throttle_initial = params->cpu_throttle_initial;
 877     }
 878
 879     if (params->has_cpu_throttle_increment) {
 880         dest->cpu_throttle_increment = params->cpu_throttle_increment;
 881     }
 882
 883     if (params->has_tls_creds) {
 884         assert(params->tls_creds->type == QTYPE_QSTRING);
 885         dest->tls_creds = g_strdup(params->tls_creds->u.s);
 886     }
 887
 888     if (params->has_tls_hostname) {
 889         assert(params->tls_hostname->type == QTYPE_QSTRING);
 890         dest->tls_hostname = g_strdup(params->tls_hostname->u.s);
 891     }
 892
 893     if (params->has_max_bandwidth) {
 894         dest->max_bandwidth = params->max_bandwidth;
 895     }
 896
 897     if (params->has_downtime_limit) {
 898         dest->downtime_limit = params->downtime_limit;
 899     }
 900
 901     if (params->has_x_checkpoint_delay) {
 902         dest->x_checkpoint_delay = params->x_checkpoint_delay;
 903     }
 904
 905     if (params->has_block_incremental) {
 906         dest->block_incremental = params->block_incremental;
 907     }
 908     if (params->has_x_multifd_channels) {
 909         dest->x_multifd_channels = params->x_multifd_channels;
 910     }
 911     if (params->has_x_multifd_page_count) {
 912         dest->x_multifd_page_count = params->x_multifd_page_count;
 913     }
 914     if (params->has_xbzrle_cache_size) {
 915         dest->xbzrle_cache_size = params->xbzrle_cache_size;
 916     }
 917 }
 918
 919 static void migrate_params_apply(MigrateSetParameters *params, Error **errp)
 920 {
 921     MigrationState *s = migrate_get_current();
 922
 923     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 924
 925     if (params->has_compress_level) {
 926         s->parameters.compress_level = params->compress_level;
 927     }
 928
 929     if (params->has_compress_threads) {
 930         s->parameters.compress_threads = params->compress_threads;
 931     }
 932
 933     if (params->has_decompress_threads) {
 934         s->parameters.decompress_threads = params->decompress_threads;
 935     }
 936
 937     if (params->has_cpu_throttle_initial) {
 938         s->parameters.cpu_throttle_initial = params->cpu_throttle_initial;
 939     }
 940
 941     if (params->has_cpu_throttle_increment) {
 942         s->parameters.cpu_throttle_increment = params->cpu_throttle_increment;
 943     }
 944
 945     if (params->has_tls_creds) {
 946         g_free(s->parameters.tls_creds);
 947         assert(params->tls_creds->type == QTYPE_QSTRING);
 948         s->parameters.tls_creds = g_strdup(params->tls_creds->u.s);
 949     }
 950
 951     if (params->has_tls_hostname) {
 952         g_free(s->parameters.tls_hostname);
 953         assert(params->tls_hostname->type == QTYPE_QSTRING);
 954         s->parameters.tls_hostname = g_strdup(params->tls_hostname->u.s);
 955     }
 956
 957     if (params->has_max_bandwidth) {
 958         s->parameters.max_bandwidth = params->max_bandwidth;
 959         if (s->to_dst_file) {
 960             qemu_file_set_rate_limit(s->to_dst_file,
 961                                 s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
 962         }
 963     }
 964
 965     if (params->has_downtime_limit) {
 966         s->parameters.downtime_limit = params->downtime_limit;
 967     }
 968
 969     if (params->has_x_checkpoint_delay) {
 970         s->parameters.x_checkpoint_delay = params->x_checkpoint_delay;
 971         if (migration_in_colo_state()) {
 972             colo_checkpoint_notify(s);
 973         }
 974     }
 975
 976     if (params->has_block_incremental) {
 977         s->parameters.block_incremental = params->block_incremental;
 978     }
 979     if (params->has_x_multifd_channels) {
 980         s->parameters.x_multifd_channels = params->x_multifd_channels;
 981     }
 982     if (params->has_x_multifd_page_count) {
 983         s->parameters.x_multifd_page_count = params->x_multifd_page_count;
 984     }
 985     if (params->has_xbzrle_cache_size) {
 986         s->parameters.xbzrle_cache_size = params->xbzrle_cache_size;
 987         xbzrle_cache_resize(params->xbzrle_cache_size, errp);
 988     }
 989 }
 990
 991 void qmp_migrate_set_parameters(MigrateSetParameters *params, Error **errp)
 992 {
 993     MigrationParameters tmp;
 994
 995     /* TODO Rewrite "" to null instead */
 996     if (params->has_tls_creds
 997         && params->tls_creds->type == QTYPE_QNULL) {
 998         QDECREF(params->tls_creds->u.n);
 999         params->tls_creds->type = QTYPE_QSTRING;
1000         params->tls_creds->u.s = strdup("");
1001     }
1002     /* TODO Rewrite "" to null instead */
1003     if (params->has_tls_hostname
1004         && params->tls_hostname->type == QTYPE_QNULL) {
1005         QDECREF(params->tls_hostname->u.n);
1006         params->tls_hostname->type = QTYPE_QSTRING;
1007         params->tls_hostname->u.s = strdup("");
1008     }
1009
1010     migrate_params_test_apply(params, &tmp);
1011
1012     if (!migrate_params_check(&tmp, errp)) {
1013         /* Invalid parameter */
1014         return;
1015     }
1016
1017     migrate_params_apply(params, errp);
1018 }
1019
1020
1021 void qmp_migrate_start_postcopy(Error **errp)
1022 {
1023     MigrationState *s = migrate_get_current();
1024
1025     if (!migrate_postcopy_ram()) {
1026         error_setg(errp, "Enable postcopy with migrate_set_capability before"
1027                          " the start of migration");
1028         return;
1029     }
1030
1031     if (s->state == MIGRATION_STATUS_NONE) {
1032         error_setg(errp, "Postcopy must be started after migration has been"
1033                          " started");
1034         return;
1035     }
1036     /*
1037      * we don't error if migration has finished since that would be racy
1038      * with issuing this command.
1039      */
1040     atomic_set(&s->start_postcopy, true);
1041 }
1042
1043 /* shared migration helpers */
1044
1045 void migrate_set_state(int *state, int old_state, int new_state)
1046 {
1047     assert(new_state < MIGRATION_STATUS__MAX);
1048     if (atomic_cmpxchg(state, old_state, new_state) == old_state) {
1049         trace_migrate_set_state(MigrationStatus_str(new_state));
1050         migrate_generate_event(new_state);
1051     }
1052 }
1053
1054 static MigrationCapabilityStatusList *migrate_cap_add(
1055     MigrationCapabilityStatusList *list,
1056     MigrationCapability index,
1057     bool state)
1058 {
1059     MigrationCapabilityStatusList *cap;
1060
1061     cap = g_new0(MigrationCapabilityStatusList, 1);
1062     cap->value = g_new0(MigrationCapabilityStatus, 1);
1063     cap->value->capability = index;
1064     cap->value->state = state;
1065     cap->next = list;
1066
1067     return cap;
1068 }
1069
1070 void migrate_set_block_enabled(bool value, Error **errp)
1071 {
1072     MigrationCapabilityStatusList *cap;
1073
1074     cap = migrate_cap_add(NULL, MIGRATION_CAPABILITY_BLOCK, value);
1075     qmp_migrate_set_capabilities(cap, errp);
1076     qapi_free_MigrationCapabilityStatusList(cap);
1077 }
1078
1079 static void migrate_set_block_incremental(MigrationState *s, bool value)
1080 {
1081     s->parameters.block_incremental = value;
1082 }
1083
1084 static void block_cleanup_parameters(MigrationState *s)
1085 {
1086     if (s->must_remove_block_options) {
1087         /* setting to false can never fail */
1088         migrate_set_block_enabled(false, &error_abort);
1089         migrate_set_block_incremental(s, false);
1090         s->must_remove_block_options = false;
1091     }
1092 }
1093
1094 static void migrate_fd_cleanup(void *opaque)
1095 {
1096     MigrationState *s = opaque;
1097
1098     qemu_bh_delete(s->cleanup_bh);
1099     s->cleanup_bh = NULL;
1100
1101     qemu_savevm_state_cleanup();
1102
1103     if (s->to_dst_file) {
1104         Error *local_err = NULL;
1105
1106         trace_migrate_fd_cleanup();
1107         qemu_mutex_unlock_iothread();
1108         if (s->migration_thread_running) {
1109             qemu_thread_join(&s->thread);
1110             s->migration_thread_running = false;
1111         }
1112         qemu_mutex_lock_iothread();
1113
1114         if (multifd_save_cleanup(&local_err) != 0) {
1115             error_report_err(local_err);
1116         }
1117         qemu_fclose(s->to_dst_file);
1118         s->to_dst_file = NULL;
1119     }
1120
1121     assert((s->state != MIGRATION_STATUS_ACTIVE) &&
1122            (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE));
1123
1124     if (s->state == MIGRATION_STATUS_CANCELLING) {
1125         migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING,
1126                           MIGRATION_STATUS_CANCELLED);
1127     }
1128
1129     if (s->error) {
1130         /* It is used on info migrate.  We can't free it */
1131         error_report_err(error_copy(s->error));
1132     }
1133     notifier_list_notify(&migration_state_notifiers, s);
1134     block_cleanup_parameters(s);
1135 }
1136
1137 void migrate_set_error(MigrationState *s, const Error *error)
1138 {
1139     qemu_mutex_lock(&s->error_mutex);
1140     if (!s->error) {
1141         s->error = error_copy(error);
1142     }
1143     qemu_mutex_unlock(&s->error_mutex);
1144 }
1145
1146 void migrate_fd_error(MigrationState *s, const Error *error)
1147 {
1148     trace_migrate_fd_error(error_get_pretty(error));
1149     assert(s->to_dst_file == NULL);
1150     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1151                       MIGRATION_STATUS_FAILED);
1152     migrate_set_error(s, error);
1153 }
1154
1155 static void migrate_fd_cancel(MigrationState *s)
1156 {
1157     int old_state ;
1158     QEMUFile *f = migrate_get_current()->to_dst_file;
1159     trace_migrate_fd_cancel();
1160
1161     if (s->rp_state.from_dst_file) {
1162         /* shutdown the rp socket, so causing the rp thread to shutdown */
1163         qemu_file_shutdown(s->rp_state.from_dst_file);
1164     }
1165
1166     do {
1167         old_state = s->state;
1168         if (!migration_is_setup_or_active(old_state)) {
1169             break;
1170         }
1171         /* If the migration is paused, kick it out of the pause */
1172         if (old_state == MIGRATION_STATUS_PRE_SWITCHOVER) {
1173             qemu_sem_post(&s->pause_sem);
1174         }
1175         migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING);
1176     } while (s->state != MIGRATION_STATUS_CANCELLING);
1177
1178     /*
1179      * If we're unlucky the migration code might be stuck somewhere in a
1180      * send/write while the network has failed and is waiting to timeout;
1181      * if we've got shutdown(2) available then we can force it to quit.
1182      * The outgoing qemu file gets closed in migrate_fd_cleanup that is
1183      * called in a bh, so there is no race against this cancel.
1184      */
1185     if (s->state == MIGRATION_STATUS_CANCELLING && f) {
1186         qemu_file_shutdown(f);
1187     }
1188     if (s->state == MIGRATION_STATUS_CANCELLING && s->block_inactive) {
1189         Error *local_err = NULL;
1190
1191         bdrv_invalidate_cache_all(&local_err);
1192         if (local_err) {
1193             error_report_err(local_err);
1194         } else {
1195             s->block_inactive = false;
1196         }
1197     }
1198 }
1199
1200 void add_migration_state_change_notifier(Notifier *notify)
1201 {
1202     notifier_list_add(&migration_state_notifiers, notify);
1203 }
1204
1205 void remove_migration_state_change_notifier(Notifier *notify)
1206 {
1207     notifier_remove(notify);
1208 }
1209
1210 bool migration_in_setup(MigrationState *s)
1211 {
1212     return s->state == MIGRATION_STATUS_SETUP;
1213 }
1214
1215 bool migration_has_finished(MigrationState *s)
1216 {
1217     return s->state == MIGRATION_STATUS_COMPLETED;
1218 }
1219
1220 bool migration_has_failed(MigrationState *s)
1221 {
1222     return (s->state == MIGRATION_STATUS_CANCELLED ||
1223             s->state == MIGRATION_STATUS_FAILED);
1224 }
1225
1226 bool migration_in_postcopy(void)
1227 {
1228     MigrationState *s = migrate_get_current();
1229
1230     return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
1231 }
1232
1233 bool migration_in_postcopy_after_devices(MigrationState *s)
1234 {
1235     return migration_in_postcopy() && s->postcopy_after_devices;
1236 }
1237
1238 bool migration_is_idle(void)
1239 {
1240     MigrationState *s = migrate_get_current();
1241
1242     switch (s->state) {
1243     case MIGRATION_STATUS_NONE:
1244     case MIGRATION_STATUS_CANCELLED:
1245     case MIGRATION_STATUS_COMPLETED:
1246     case MIGRATION_STATUS_FAILED:
1247         return true;
1248     case MIGRATION_STATUS_SETUP:
1249     case MIGRATION_STATUS_CANCELLING:
1250     case MIGRATION_STATUS_ACTIVE:
1251     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
1252     case MIGRATION_STATUS_COLO:
1253     case MIGRATION_STATUS_PRE_SWITCHOVER:
1254     case MIGRATION_STATUS_DEVICE:
1255         return false;
1256     case MIGRATION_STATUS__MAX:
1257         g_assert_not_reached();
1258     }
1259
1260     return false;
1261 }
1262
1263 void migrate_init(MigrationState *s)
1264 {
1265     /*
1266      * Reinitialise all migration state, except
1267      * parameters/capabilities that the user set, and
1268      * locks.
1269      */
1270     s->bytes_xfer = 0;
1271     s->xfer_limit = 0;
1272     s->cleanup_bh = 0;
1273     s->to_dst_file = NULL;
1274     s->state = MIGRATION_STATUS_NONE;
1275     s->rp_state.from_dst_file = NULL;
1276     s->rp_state.error = false;
1277     s->mbps = 0.0;
1278     s->downtime = 0;
1279     s->expected_downtime = 0;
1280     s->setup_time = 0;
1281     s->start_postcopy = false;
1282     s->postcopy_after_devices = false;
1283     s->migration_thread_running = false;
1284     error_free(s->error);
1285     s->error = NULL;
1286
1287     migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP);
1288
1289     s->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1290     s->total_time = 0;
1291     s->vm_was_running = false;
1292     s->iteration_initial_bytes = 0;
1293     s->threshold_size = 0;
1294 }
1295
1296 static GSList *migration_blockers;
1297
1298 int migrate_add_blocker(Error *reason, Error **errp)
1299 {
1300     if (migrate_get_current()->only_migratable) {
1301         error_propagate(errp, error_copy(reason));
1302         error_prepend(errp, "disallowing migration blocker "
1303                           "(--only_migratable) for: ");
1304         return -EACCES;
1305     }
1306
1307     if (migration_is_idle()) {
1308         migration_blockers = g_slist_prepend(migration_blockers, reason);
1309         return 0;
1310     }
1311
1312     error_propagate(errp, error_copy(reason));
1313     error_prepend(errp, "disallowing migration blocker (migration in "
1314                       "progress) for: ");
1315     return -EBUSY;
1316 }
1317
1318 void migrate_del_blocker(Error *reason)
1319 {
1320     migration_blockers = g_slist_remove(migration_blockers, reason);
1321 }
1322
1323 void qmp_migrate_incoming(const char *uri, Error **errp)
1324 {
1325     Error *local_err = NULL;
1326     static bool once = true;
1327
1328     if (!deferred_incoming) {
1329         error_setg(errp, "For use with '-incoming defer'");
1330         return;
1331     }
1332     if (!once) {
1333         error_setg(errp, "The incoming migration has already been started");
1334     }
1335
1336     qemu_start_incoming_migration(uri, &local_err);
1337
1338     if (local_err) {
1339         error_propagate(errp, local_err);
1340         return;
1341     }
1342
1343     once = false;
1344 }
1345
1346 bool migration_is_blocked(Error **errp)
1347 {
1348     if (qemu_savevm_state_blocked(errp)) {
1349         return true;
1350     }
1351
1352     if (migration_blockers) {
1353         error_propagate(errp, error_copy(migration_blockers->data));
1354         return true;
1355     }
1356
1357     return false;
1358 }
1359
1360 void qmp_migrate(const char *uri, bool has_blk, bool blk,
1361                  bool has_inc, bool inc, bool has_detach, bool detach,
1362                  Error **errp)
1363 {
1364     Error *local_err = NULL;
1365     MigrationState *s = migrate_get_current();
1366     const char *p;
1367
1368     if (migration_is_setup_or_active(s->state) ||
1369         s->state == MIGRATION_STATUS_CANCELLING ||
1370         s->state == MIGRATION_STATUS_COLO) {
1371         error_setg(errp, QERR_MIGRATION_ACTIVE);
1372         return;
1373     }
1374     if (runstate_check(RUN_STATE_INMIGRATE)) {
1375         error_setg(errp, "Guest is waiting for an incoming migration");
1376         return;
1377     }
1378
1379     if (migration_is_blocked(errp)) {
1380         return;
1381     }
1382
1383     if ((has_blk && blk) || (has_inc && inc)) {
1384         if (migrate_use_block() || migrate_use_block_incremental()) {
1385             error_setg(errp, "Command options are incompatible with "
1386                        "current migration capabilities");
1387             return;
1388         }
1389         migrate_set_block_enabled(true, &local_err);
1390         if (local_err) {
1391             error_propagate(errp, local_err);
1392             return;
1393         }
1394         s->must_remove_block_options = true;
1395     }
1396
1397     if (has_inc && inc) {
1398         migrate_set_block_incremental(s, true);
1399     }
1400
1401     migrate_init(s);
1402
1403     if (strstart(uri, "tcp:", &p)) {
1404         tcp_start_outgoing_migration(s, p, &local_err);
1405 #ifdef CONFIG_RDMA
1406     } else if (strstart(uri, "rdma:", &p)) {
1407         rdma_start_outgoing_migration(s, p, &local_err);
1408 #endif
1409     } else if (strstart(uri, "exec:", &p)) {
1410         exec_start_outgoing_migration(s, p, &local_err);
1411     } else if (strstart(uri, "unix:", &p)) {
1412         unix_start_outgoing_migration(s, p, &local_err);
1413     } else if (strstart(uri, "fd:", &p)) {
1414         fd_start_outgoing_migration(s, p, &local_err);
1415     } else {
1416         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri",
1417                    "a valid migration protocol");
1418         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1419                           MIGRATION_STATUS_FAILED);
1420         return;
1421     }
1422
1423     if (local_err) {
1424         migrate_fd_error(s, local_err);
1425         error_propagate(errp, local_err);
1426         return;
1427     }
1428 }
1429
1430 void qmp_migrate_cancel(Error **errp)
1431 {
1432     migrate_fd_cancel(migrate_get_current());
1433 }
1434
1435 void qmp_migrate_continue(MigrationStatus state, Error **errp)
1436 {
1437     MigrationState *s = migrate_get_current();
1438     if (s->state != state) {
1439         error_setg(errp,  "Migration not in expected state: %s",
1440                    MigrationStatus_str(s->state));
1441         return;
1442     }
1443     qemu_sem_post(&s->pause_sem);
1444 }
1445
1446 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
1447 {
1448     MigrateSetParameters p = {
1449         .has_xbzrle_cache_size = true,
1450         .xbzrle_cache_size = value,
1451     };
1452
1453     qmp_migrate_set_parameters(&p, errp);
1454 }
1455
1456 int64_t qmp_query_migrate_cache_size(Error **errp)
1457 {
1458     return migrate_xbzrle_cache_size();
1459 }
1460
1461 void qmp_migrate_set_speed(int64_t value, Error **errp)
1462 {
1463     MigrateSetParameters p = {
1464         .has_max_bandwidth = true,
1465         .max_bandwidth = value,
1466     };
1467
1468     qmp_migrate_set_parameters(&p, errp);
1469 }
1470
1471 void qmp_migrate_set_downtime(double value, Error **errp)
1472 {
1473     if (value < 0 || value > MAX_MIGRATE_DOWNTIME_SECONDS) {
1474         error_setg(errp, "Parameter 'downtime_limit' expects an integer in "
1475                          "the range of 0 to %d seconds",
1476                          MAX_MIGRATE_DOWNTIME_SECONDS);
1477         return;
1478     }
1479
1480     value *= 1000; /* Convert to milliseconds */
1481     value = MAX(0, MIN(INT64_MAX, value));
1482
1483     MigrateSetParameters p = {
1484         .has_downtime_limit = true,
1485         .downtime_limit = value,
1486     };
1487
1488     qmp_migrate_set_parameters(&p, errp);
1489 }
1490
1491 bool migrate_release_ram(void)
1492 {
1493     MigrationState *s;
1494
1495     s = migrate_get_current();
1496
1497     return s->enabled_capabilities[MIGRATION_CAPABILITY_RELEASE_RAM];
1498 }
1499
1500 bool migrate_postcopy_ram(void)
1501 {
1502     MigrationState *s;
1503
1504     s = migrate_get_current();
1505
1506     return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM];
1507 }
1508
1509 bool migrate_postcopy(void)
1510 {
1511     return migrate_postcopy_ram();
1512 }
1513
1514 bool migrate_auto_converge(void)
1515 {
1516     MigrationState *s;
1517
1518     s = migrate_get_current();
1519
1520     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
1521 }
1522
1523 bool migrate_zero_blocks(void)
1524 {
1525     MigrationState *s;
1526
1527     s = migrate_get_current();
1528
1529     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
1530 }
1531
1532 bool migrate_use_compression(void)
1533 {
1534     MigrationState *s;
1535
1536     s = migrate_get_current();
1537
1538     return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS];
1539 }
1540
1541 int migrate_compress_level(void)
1542 {
1543     MigrationState *s;
1544
1545     s = migrate_get_current();
1546
1547     return s->parameters.compress_level;
1548 }
1549
1550 int migrate_compress_threads(void)
1551 {
1552     MigrationState *s;
1553
1554     s = migrate_get_current();
1555
1556     return s->parameters.compress_threads;
1557 }
1558
1559 int migrate_decompress_threads(void)
1560 {
1561     MigrationState *s;
1562
1563     s = migrate_get_current();
1564
1565     return s->parameters.decompress_threads;
1566 }
1567
1568 bool migrate_use_events(void)
1569 {
1570     MigrationState *s;
1571
1572     s = migrate_get_current();
1573
1574     return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS];
1575 }
1576
1577 bool migrate_use_multifd(void)
1578 {
1579     MigrationState *s;
1580
1581     s = migrate_get_current();
1582
1583     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_MULTIFD];
1584 }
1585
1586 bool migrate_pause_before_switchover(void)
1587 {
1588     MigrationState *s;
1589
1590     s = migrate_get_current();
1591
1592     return s->enabled_capabilities[
1593         MIGRATION_CAPABILITY_PAUSE_BEFORE_SWITCHOVER];
1594 }
1595
1596 int migrate_multifd_channels(void)
1597 {
1598     MigrationState *s;
1599
1600     s = migrate_get_current();
1601
1602     return s->parameters.x_multifd_channels;
1603 }
1604
1605 int migrate_multifd_page_count(void)
1606 {
1607     MigrationState *s;
1608
1609     s = migrate_get_current();
1610
1611     return s->parameters.x_multifd_page_count;
1612 }
1613
1614 int migrate_use_xbzrle(void)
1615 {
1616     MigrationState *s;
1617
1618     s = migrate_get_current();
1619
1620     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
1621 }
1622
1623 int64_t migrate_xbzrle_cache_size(void)
1624 {
1625     MigrationState *s;
1626
1627     s = migrate_get_current();
1628
1629     return s->parameters.xbzrle_cache_size;
1630 }
1631
1632 bool migrate_use_block(void)
1633 {
1634     MigrationState *s;
1635
1636     s = migrate_get_current();
1637
1638     return s->enabled_capabilities[MIGRATION_CAPABILITY_BLOCK];
1639 }
1640
1641 bool migrate_use_return_path(void)
1642 {
1643     MigrationState *s;
1644
1645     s = migrate_get_current();
1646
1647     return s->enabled_capabilities[MIGRATION_CAPABILITY_RETURN_PATH];
1648 }
1649
1650 bool migrate_use_block_incremental(void)
1651 {
1652     MigrationState *s;
1653
1654     s = migrate_get_current();
1655
1656     return s->parameters.block_incremental;
1657 }
1658
1659 /* migration thread support */
1660 /*
1661  * Something bad happened to the RP stream, mark an error
1662  * The caller shall print or trace something to indicate why
1663  */
1664 static void mark_source_rp_bad(MigrationState *s)
1665 {
1666     s->rp_state.error = true;
1667 }
1668
1669 static struct rp_cmd_args {
1670     ssize_t     len; /* -1 = variable */
1671     const char *name;
1672 } rp_cmd_args[] = {
1673     [MIG_RP_MSG_INVALID]        = { .len = -1, .name = "INVALID" },
1674     [MIG_RP_MSG_SHUT]           = { .len =  4, .name = "SHUT" },
1675     [MIG_RP_MSG_PONG]           = { .len =  4, .name = "PONG" },
1676     [MIG_RP_MSG_REQ_PAGES]      = { .len = 12, .name = "REQ_PAGES" },
1677     [MIG_RP_MSG_REQ_PAGES_ID]   = { .len = -1, .name = "REQ_PAGES_ID" },
1678     [MIG_RP_MSG_MAX]            = { .len = -1, .name = "MAX" },
1679 };
1680
1681 /*
1682  * Process a request for pages received on the return path,
1683  * We're allowed to send more than requested (e.g. to round to our page size)
1684  * and we don't need to send pages that have already been sent.
1685  */
1686 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname,
1687                                        ram_addr_t start, size_t len)
1688 {
1689     long our_host_ps = getpagesize();
1690
1691     trace_migrate_handle_rp_req_pages(rbname, start, len);
1692
1693     /*
1694      * Since we currently insist on matching page sizes, just sanity check
1695      * we're being asked for whole host pages.
1696      */
1697     if (start & (our_host_ps-1) ||
1698        (len & (our_host_ps-1))) {
1699         error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1700                      " len: %zd", __func__, start, len);
1701         mark_source_rp_bad(ms);
1702         return;
1703     }
1704
1705     if (ram_save_queue_pages(rbname, start, len)) {
1706         mark_source_rp_bad(ms);
1707     }
1708 }
1709
1710 /*
1711  * Handles messages sent on the return path towards the source VM
1712  *
1713  */
1714 static void *source_return_path_thread(void *opaque)
1715 {
1716     MigrationState *ms = opaque;
1717     QEMUFile *rp = ms->rp_state.from_dst_file;
1718     uint16_t header_len, header_type;
1719     uint8_t buf[512];
1720     uint32_t tmp32, sibling_error;
1721     ram_addr_t start = 0; /* =0 to silence warning */
1722     size_t  len = 0, expected_len;
1723     int res;
1724
1725     trace_source_return_path_thread_entry();
1726     while (!ms->rp_state.error && !qemu_file_get_error(rp) &&
1727            migration_is_setup_or_active(ms->state)) {
1728         trace_source_return_path_thread_loop_top();
1729         header_type = qemu_get_be16(rp);
1730         header_len = qemu_get_be16(rp);
1731
1732         if (qemu_file_get_error(rp)) {
1733             mark_source_rp_bad(ms);
1734             goto out;
1735         }
1736
1737         if (header_type >= MIG_RP_MSG_MAX ||
1738             header_type == MIG_RP_MSG_INVALID) {
1739             error_report("RP: Received invalid message 0x%04x length 0x%04x",
1740                     header_type, header_len);
1741             mark_source_rp_bad(ms);
1742             goto out;
1743         }
1744
1745         if ((rp_cmd_args[header_type].len != -1 &&
1746             header_len != rp_cmd_args[header_type].len) ||
1747             header_len > sizeof(buf)) {
1748             error_report("RP: Received '%s' message (0x%04x) with"
1749                     "incorrect length %d expecting %zu",
1750                     rp_cmd_args[header_type].name, header_type, header_len,
1751                     (size_t)rp_cmd_args[header_type].len);
1752             mark_source_rp_bad(ms);
1753             goto out;
1754         }
1755
1756         /* We know we've got a valid header by this point */
1757         res = qemu_get_buffer(rp, buf, header_len);
1758         if (res != header_len) {
1759             error_report("RP: Failed reading data for message 0x%04x"
1760                          " read %d expected %d",
1761                          header_type, res, header_len);
1762             mark_source_rp_bad(ms);
1763             goto out;
1764         }
1765
1766         /* OK, we have the message and the data */
1767         switch (header_type) {
1768         case MIG_RP_MSG_SHUT:
1769             sibling_error = ldl_be_p(buf);
1770             trace_source_return_path_thread_shut(sibling_error);
1771             if (sibling_error) {
1772                 error_report("RP: Sibling indicated error %d", sibling_error);
1773                 mark_source_rp_bad(ms);
1774             }
1775             /*
1776              * We'll let the main thread deal with closing the RP
1777              * we could do a shutdown(2) on it, but we're the only user
1778              * anyway, so there's nothing gained.
1779              */
1780             goto out;
1781
1782         case MIG_RP_MSG_PONG:
1783             tmp32 = ldl_be_p(buf);
1784             trace_source_return_path_thread_pong(tmp32);
1785             break;
1786
1787         case MIG_RP_MSG_REQ_PAGES:
1788             start = ldq_be_p(buf);
1789             len = ldl_be_p(buf + 8);
1790             migrate_handle_rp_req_pages(ms, NULL, start, len);
1791             break;
1792
1793         case MIG_RP_MSG_REQ_PAGES_ID:
1794             expected_len = 12 + 1; /* header + termination */
1795
1796             if (header_len >= expected_len) {
1797                 start = ldq_be_p(buf);
1798                 len = ldl_be_p(buf + 8);
1799                 /* Now we expect an idstr */
1800                 tmp32 = buf[12]; /* Length of the following idstr */
1801                 buf[13 + tmp32] = '\0';
1802                 expected_len += tmp32;
1803             }
1804             if (header_len != expected_len) {
1805                 error_report("RP: Req_Page_id with length %d expecting %zd",
1806                         header_len, expected_len);
1807                 mark_source_rp_bad(ms);
1808                 goto out;
1809             }
1810             migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len);
1811             break;
1812
1813         default:
1814             break;
1815         }
1816     }
1817     if (qemu_file_get_error(rp)) {
1818         trace_source_return_path_thread_bad_end();
1819         mark_source_rp_bad(ms);
1820     }
1821
1822     trace_source_return_path_thread_end();
1823 out:
1824     ms->rp_state.from_dst_file = NULL;
1825     qemu_fclose(rp);
1826     return NULL;
1827 }
1828
1829 static int open_return_path_on_source(MigrationState *ms)
1830 {
1831
1832     ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file);
1833     if (!ms->rp_state.from_dst_file) {
1834         return -1;
1835     }
1836
1837     trace_open_return_path_on_source();
1838     qemu_thread_create(&ms->rp_state.rp_thread, "return path",
1839                        source_return_path_thread, ms, QEMU_THREAD_JOINABLE);
1840
1841     trace_open_return_path_on_source_continue();
1842
1843     return 0;
1844 }
1845
1846 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
1847 static int await_return_path_close_on_source(MigrationState *ms)
1848 {
1849     /*
1850      * If this is a normal exit then the destination will send a SHUT and the
1851      * rp_thread will exit, however if there's an error we need to cause
1852      * it to exit.
1853      */
1854     if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
1855         /*
1856          * shutdown(2), if we have it, will cause it to unblock if it's stuck
1857          * waiting for the destination.
1858          */
1859         qemu_file_shutdown(ms->rp_state.from_dst_file);
1860         mark_source_rp_bad(ms);
1861     }
1862     trace_await_return_path_close_on_source_joining();
1863     qemu_thread_join(&ms->rp_state.rp_thread);
1864     trace_await_return_path_close_on_source_close();
1865     return ms->rp_state.error;
1866 }
1867
1868 /*
1869  * Switch from normal iteration to postcopy
1870  * Returns non-0 on error
1871  */
1872 static int postcopy_start(MigrationState *ms)
1873 {
1874     int ret;
1875     QIOChannelBuffer *bioc;
1876     QEMUFile *fb;
1877     int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1878     bool restart_block = false;
1879     int cur_state = MIGRATION_STATUS_ACTIVE;
1880     if (!migrate_pause_before_switchover()) {
1881         migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE,
1882                           MIGRATION_STATUS_POSTCOPY_ACTIVE);
1883     }
1884
1885     trace_postcopy_start();
1886     qemu_mutex_lock_iothread();
1887     trace_postcopy_start_set_run();
1888
1889     qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1890     global_state_store();
1891     ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1892     if (ret < 0) {
1893         goto fail;
1894     }
1895
1896     ret = migration_maybe_pause(ms, &cur_state,
1897                                 MIGRATION_STATUS_POSTCOPY_ACTIVE);
1898     if (ret < 0) {
1899         goto fail;
1900     }
1901
1902     ret = bdrv_inactivate_all();
1903     if (ret < 0) {
1904         goto fail;
1905     }
1906     restart_block = true;
1907
1908     /*
1909      * Cause any non-postcopiable, but iterative devices to
1910      * send out their final data.
1911      */
1912     qemu_savevm_state_complete_precopy(ms->to_dst_file, true, false);
1913
1914     /*
1915      * in Finish migrate and with the io-lock held everything should
1916      * be quiet, but we've potentially still got dirty pages and we
1917      * need to tell the destination to throw any pages it's already received
1918      * that are dirty
1919      */
1920     if (migrate_postcopy_ram()) {
1921         if (ram_postcopy_send_discard_bitmap(ms)) {
1922             error_report("postcopy send discard bitmap failed");
1923             goto fail;
1924         }
1925     }
1926
1927     /*
1928      * send rest of state - note things that are doing postcopy
1929      * will notice we're in POSTCOPY_ACTIVE and not actually
1930      * wrap their state up here
1931      */
1932     qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX);
1933     if (migrate_postcopy_ram()) {
1934         /* Ping just for debugging, helps line traces up */
1935         qemu_savevm_send_ping(ms->to_dst_file, 2);
1936     }
1937
1938     /*
1939      * While loading the device state we may trigger page transfer
1940      * requests and the fd must be free to process those, and thus
1941      * the destination must read the whole device state off the fd before
1942      * it starts processing it.  Unfortunately the ad-hoc migration format
1943      * doesn't allow the destination to know the size to read without fully
1944      * parsing it through each devices load-state code (especially the open
1945      * coded devices that use get/put).
1946      * So we wrap the device state up in a package with a length at the start;
1947      * to do this we use a qemu_buf to hold the whole of the device state.
1948      */
1949     bioc = qio_channel_buffer_new(4096);
1950     qio_channel_set_name(QIO_CHANNEL(bioc), "migration-postcopy-buffer");
1951     fb = qemu_fopen_channel_output(QIO_CHANNEL(bioc));
1952     object_unref(OBJECT(bioc));
1953
1954     /*
1955      * Make sure the receiver can get incoming pages before we send the rest
1956      * of the state
1957      */
1958     qemu_savevm_send_postcopy_listen(fb);
1959
1960     qemu_savevm_state_complete_precopy(fb, false, false);
1961     if (migrate_postcopy_ram()) {
1962         qemu_savevm_send_ping(fb, 3);
1963     }
1964
1965     qemu_savevm_send_postcopy_run(fb);
1966
1967     /* <><> end of stuff going into the package */
1968
1969     /* Last point of recovery; as soon as we send the package the destination
1970      * can open devices and potentially start running.
1971      * Lets just check again we've not got any errors.
1972      */
1973     ret = qemu_file_get_error(ms->to_dst_file);
1974     if (ret) {
1975         error_report("postcopy_start: Migration stream errored (pre package)");
1976         goto fail_closefb;
1977     }
1978
1979     restart_block = false;
1980
1981     /* Now send that blob */
1982     if (qemu_savevm_send_packaged(ms->to_dst_file, bioc->data, bioc->usage)) {
1983         goto fail_closefb;
1984     }
1985     qemu_fclose(fb);
1986
1987     /* Send a notify to give a chance for anything that needs to happen
1988      * at the transition to postcopy and after the device state; in particular
1989      * spice needs to trigger a transition now
1990      */
1991     ms->postcopy_after_devices = true;
1992     notifier_list_notify(&migration_state_notifiers, ms);
1993
1994     ms->downtime =  qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop;
1995
1996     qemu_mutex_unlock_iothread();
1997
1998     if (migrate_postcopy_ram()) {
1999         /*
2000          * Although this ping is just for debug, it could potentially be
2001          * used for getting a better measurement of downtime at the source.
2002          */
2003         qemu_savevm_send_ping(ms->to_dst_file, 4);
2004     }
2005
2006     if (migrate_release_ram()) {
2007         ram_postcopy_migrated_memory_release(ms);
2008     }
2009
2010     ret = qemu_file_get_error(ms->to_dst_file);
2011     if (ret) {
2012         error_report("postcopy_start: Migration stream errored");
2013         migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
2014                               MIGRATION_STATUS_FAILED);
2015     }
2016
2017     return ret;
2018
2019 fail_closefb:
2020     qemu_fclose(fb);
2021 fail:
2022     migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
2023                           MIGRATION_STATUS_FAILED);
2024     if (restart_block) {
2025         /* A failure happened early enough that we know the destination hasn't
2026          * accessed block devices, so we're safe to recover.
2027          */
2028         Error *local_err = NULL;
2029
2030         bdrv_invalidate_cache_all(&local_err);
2031         if (local_err) {
2032             error_report_err(local_err);
2033         }
2034     }
2035     qemu_mutex_unlock_iothread();
2036     return -1;
2037 }
2038
2039 /**
2040  * migration_maybe_pause: Pause if required to by
2041  * migrate_pause_before_switchover called with the iothread locked
2042  * Returns: 0 on success
2043  */
2044 static int migration_maybe_pause(MigrationState *s,
2045                                  int *current_active_state,
2046                                  int new_state)
2047 {
2048     if (!migrate_pause_before_switchover()) {
2049         return 0;
2050     }
2051
2052     /* Since leaving this state is not atomic with posting the semaphore
2053      * it's possible that someone could have issued multiple migrate_continue
2054      * and the semaphore is incorrectly positive at this point;
2055      * the docs say it's undefined to reinit a semaphore that's already
2056      * init'd, so use timedwait to eat up any existing posts.
2057      */
2058     while (qemu_sem_timedwait(&s->pause_sem, 1) == 0) {
2059         /* This block intentionally left blank */
2060     }
2061
2062     qemu_mutex_unlock_iothread();
2063     migrate_set_state(&s->state, *current_active_state,
2064                       MIGRATION_STATUS_PRE_SWITCHOVER);
2065     qemu_sem_wait(&s->pause_sem);
2066     migrate_set_state(&s->state, MIGRATION_STATUS_PRE_SWITCHOVER,
2067                       new_state);
2068     *current_active_state = new_state;
2069     qemu_mutex_lock_iothread();
2070
2071     return s->state == new_state ? 0 : -EINVAL;
2072 }
2073
2074 /**
2075  * migration_completion: Used by migration_thread when there's not much left.
2076  *   The caller 'breaks' the loop when this returns.
2077  *
2078  * @s: Current migration state
2079  */
2080 static void migration_completion(MigrationState *s)
2081 {
2082     int ret;
2083     int current_active_state = s->state;
2084
2085     if (s->state == MIGRATION_STATUS_ACTIVE) {
2086         qemu_mutex_lock_iothread();
2087         s->downtime_start = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2088         qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
2089         s->vm_was_running = runstate_is_running();
2090         ret = global_state_store();
2091
2092         if (!ret) {
2093             bool inactivate = !migrate_colo_enabled();
2094             ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
2095             if (ret >= 0) {
2096                 ret = migration_maybe_pause(s, &current_active_state,
2097                                             MIGRATION_STATUS_DEVICE);
2098             }
2099             if (ret >= 0) {
2100                 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX);
2101                 ret = qemu_savevm_state_complete_precopy(s->to_dst_file, false,
2102                                                          inactivate);
2103             }
2104             if (inactivate && ret >= 0) {
2105                 s->block_inactive = true;
2106             }
2107         }
2108         qemu_mutex_unlock_iothread();
2109
2110         if (ret < 0) {
2111             goto fail;
2112         }
2113     } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
2114         trace_migration_completion_postcopy_end();
2115
2116         qemu_savevm_state_complete_postcopy(s->to_dst_file);
2117         trace_migration_completion_postcopy_end_after_complete();
2118     }
2119
2120     /*
2121      * If rp was opened we must clean up the thread before
2122      * cleaning everything else up (since if there are no failures
2123      * it will wait for the destination to send it's status in
2124      * a SHUT command).
2125      */
2126     if (s->rp_state.from_dst_file) {
2127         int rp_error;
2128         trace_migration_return_path_end_before();
2129         rp_error = await_return_path_close_on_source(s);
2130         trace_migration_return_path_end_after(rp_error);
2131         if (rp_error) {
2132             goto fail_invalidate;
2133         }
2134     }
2135
2136     if (qemu_file_get_error(s->to_dst_file)) {
2137         trace_migration_completion_file_err();
2138         goto fail_invalidate;
2139     }
2140
2141     if (!migrate_colo_enabled()) {
2142         migrate_set_state(&s->state, current_active_state,
2143                           MIGRATION_STATUS_COMPLETED);
2144     }
2145
2146     return;
2147
2148 fail_invalidate:
2149     /* If not doing postcopy, vm_start() will be called: let's regain
2150      * control on images.
2151      */
2152     if (s->state == MIGRATION_STATUS_ACTIVE ||
2153         s->state == MIGRATION_STATUS_DEVICE) {
2154         Error *local_err = NULL;
2155
2156         qemu_mutex_lock_iothread();
2157         bdrv_invalidate_cache_all(&local_err);
2158         if (local_err) {
2159             error_report_err(local_err);
2160         } else {
2161             s->block_inactive = false;
2162         }
2163         qemu_mutex_unlock_iothread();
2164     }
2165
2166 fail:
2167     migrate_set_state(&s->state, current_active_state,
2168                       MIGRATION_STATUS_FAILED);
2169 }
2170
2171 bool migrate_colo_enabled(void)
2172 {
2173     MigrationState *s = migrate_get_current();
2174     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_COLO];
2175 }
2176
2177 static void migration_calculate_complete(MigrationState *s)
2178 {
2179     uint64_t bytes = qemu_ftell(s->to_dst_file);
2180     int64_t end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2181
2182     s->total_time = end_time - s->start_time;
2183     if (!s->downtime) {
2184         /*
2185          * It's still not set, so we are precopy migration.  For
2186          * postcopy, downtime is calculated during postcopy_start().
2187          */
2188         s->downtime = end_time - s->downtime_start;
2189     }
2190
2191     if (s->total_time) {
2192         s->mbps = ((double) bytes * 8.0) / s->total_time / 1000;
2193     }
2194 }
2195
2196 static void migration_update_counters(MigrationState *s,
2197                                       int64_t current_time)
2198 {
2199     uint64_t transferred, time_spent;
2200     double bandwidth;
2201
2202     if (current_time < s->iteration_start_time + BUFFER_DELAY) {
2203         return;
2204     }
2205
2206     transferred = qemu_ftell(s->to_dst_file) - s->iteration_initial_bytes;
2207     time_spent = current_time - s->iteration_start_time;
2208     bandwidth = (double)transferred / time_spent;
2209     s->threshold_size = bandwidth * s->parameters.downtime_limit;
2210
2211     s->mbps = (((double) transferred * 8.0) /
2212                ((double) time_spent / 1000.0)) / 1000.0 / 1000.0;
2213
2214     /*
2215      * if we haven't sent anything, we don't want to
2216      * recalculate. 10000 is a small enough number for our purposes
2217      */
2218     if (ram_counters.dirty_pages_rate && transferred > 10000) {
2219         s->expected_downtime = ram_counters.dirty_pages_rate *
2220             qemu_target_page_size() / bandwidth;
2221     }
2222
2223     qemu_file_reset_rate_limit(s->to_dst_file);
2224
2225     s->iteration_start_time = current_time;
2226     s->iteration_initial_bytes = qemu_ftell(s->to_dst_file);
2227
2228     trace_migrate_transferred(transferred, time_spent,
2229                               bandwidth, s->threshold_size);
2230 }
2231
2232 /* Migration thread iteration status */
2233 typedef enum {
2234     MIG_ITERATE_RESUME,         /* Resume current iteration */
2235     MIG_ITERATE_SKIP,           /* Skip current iteration */
2236     MIG_ITERATE_BREAK,          /* Break the loop */
2237 } MigIterateState;
2238
2239 /*
2240  * Return true if continue to the next iteration directly, false
2241  * otherwise.
2242  */
2243 static MigIterateState migration_iteration_run(MigrationState *s)
2244 {
2245     uint64_t pending_size, pend_post, pend_nonpost;
2246     bool in_postcopy = s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE;
2247
2248     qemu_savevm_state_pending(s->to_dst_file, s->threshold_size,
2249                               &pend_nonpost, &pend_post);
2250     pending_size = pend_nonpost + pend_post;
2251
2252     trace_migrate_pending(pending_size, s->threshold_size,
2253                           pend_post, pend_nonpost);
2254
2255     if (pending_size && pending_size >= s->threshold_size) {
2256         /* Still a significant amount to transfer */
2257         if (migrate_postcopy() && !in_postcopy &&
2258             pend_nonpost <= s->threshold_size &&
2259             atomic_read(&s->start_postcopy)) {
2260             if (postcopy_start(s)) {
2261                 error_report("%s: postcopy failed to start", __func__);
2262             }
2263             return MIG_ITERATE_SKIP;
2264         }
2265         /* Just another iteration step */
2266         qemu_savevm_state_iterate(s->to_dst_file,
2267             s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
2268     } else {
2269         trace_migration_thread_low_pending(pending_size);
2270         migration_completion(s);
2271         return MIG_ITERATE_BREAK;
2272     }
2273
2274     return MIG_ITERATE_RESUME;
2275 }
2276
2277 static void migration_iteration_finish(MigrationState *s)
2278 {
2279     /* If we enabled cpu throttling for auto-converge, turn it off. */
2280     cpu_throttle_stop();
2281
2282     qemu_mutex_lock_iothread();
2283     switch (s->state) {
2284     case MIGRATION_STATUS_COMPLETED:
2285         migration_calculate_complete(s);
2286         runstate_set(RUN_STATE_POSTMIGRATE);
2287         break;
2288
2289     case MIGRATION_STATUS_ACTIVE:
2290         /*
2291          * We should really assert here, but since it's during
2292          * migration, let's try to reduce the usage of assertions.
2293          */
2294         if (!migrate_colo_enabled()) {
2295             error_report("%s: critical error: calling COLO code without "
2296                          "COLO enabled", __func__);
2297         }
2298         migrate_start_colo_process(s);
2299         /*
2300          * Fixme: we will run VM in COLO no matter its old running state.
2301          * After exited COLO, we will keep running.
2302          */
2303         s->vm_was_running = true;
2304         /* Fallthrough */
2305     case MIGRATION_STATUS_FAILED:
2306     case MIGRATION_STATUS_CANCELLED:
2307         if (s->vm_was_running) {
2308             vm_start();
2309         } else {
2310             if (runstate_check(RUN_STATE_FINISH_MIGRATE)) {
2311                 runstate_set(RUN_STATE_POSTMIGRATE);
2312             }
2313         }
2314         break;
2315
2316     default:
2317         /* Should not reach here, but if so, forgive the VM. */
2318         error_report("%s: Unknown ending state %d", __func__, s->state);
2319         break;
2320     }
2321     qemu_bh_schedule(s->cleanup_bh);
2322     qemu_mutex_unlock_iothread();
2323 }
2324
2325 /*
2326  * Master migration thread on the source VM.
2327  * It drives the migration and pumps the data down the outgoing channel.
2328  */
2329 static void *migration_thread(void *opaque)
2330 {
2331     MigrationState *s = opaque;
2332     int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
2333
2334     rcu_register_thread();
2335
2336     s->iteration_start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2337
2338     qemu_savevm_state_header(s->to_dst_file);
2339
2340     /*
2341      * If we opened the return path, we need to make sure dst has it
2342      * opened as well.
2343      */
2344     if (s->rp_state.from_dst_file) {
2345         /* Now tell the dest that it should open its end so it can reply */
2346         qemu_savevm_send_open_return_path(s->to_dst_file);
2347
2348         /* And do a ping that will make stuff easier to debug */
2349         qemu_savevm_send_ping(s->to_dst_file, 1);
2350     }
2351
2352     if (migrate_postcopy()) {
2353         /*
2354          * Tell the destination that we *might* want to do postcopy later;
2355          * if the other end can't do postcopy it should fail now, nice and
2356          * early.
2357          */
2358         qemu_savevm_send_postcopy_advise(s->to_dst_file);
2359     }
2360
2361     qemu_savevm_state_setup(s->to_dst_file);
2362
2363     s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
2364     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2365                       MIGRATION_STATUS_ACTIVE);
2366
2367     trace_migration_thread_setup_complete();
2368
2369     while (s->state == MIGRATION_STATUS_ACTIVE ||
2370            s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
2371         int64_t current_time;
2372
2373         if (!qemu_file_rate_limit(s->to_dst_file)) {
2374             MigIterateState iter_state = migration_iteration_run(s);
2375             if (iter_state == MIG_ITERATE_SKIP) {
2376                 continue;
2377             } else if (iter_state == MIG_ITERATE_BREAK) {
2378                 break;
2379             }
2380         }
2381
2382         if (qemu_file_get_error(s->to_dst_file)) {
2383             if (migration_is_setup_or_active(s->state)) {
2384                 migrate_set_state(&s->state, s->state,
2385                                   MIGRATION_STATUS_FAILED);
2386             }
2387             trace_migration_thread_file_err();
2388             break;
2389         }
2390
2391         current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2392
2393         migration_update_counters(s, current_time);
2394
2395         if (qemu_file_rate_limit(s->to_dst_file)) {
2396             /* usleep expects microseconds */
2397             g_usleep((s->iteration_start_time + BUFFER_DELAY -
2398                       current_time) * 1000);
2399         }
2400     }
2401
2402     trace_migration_thread_after_loop();
2403     migration_iteration_finish(s);
2404     rcu_unregister_thread();
2405     return NULL;
2406 }
2407
2408 void migrate_fd_connect(MigrationState *s, Error *error_in)
2409 {
2410     s->expected_downtime = s->parameters.downtime_limit;
2411     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
2412     if (error_in) {
2413         migrate_fd_error(s, error_in);
2414         migrate_fd_cleanup(s);
2415         return;
2416     }
2417
2418     qemu_file_set_blocking(s->to_dst_file, true);
2419     qemu_file_set_rate_limit(s->to_dst_file,
2420                              s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
2421
2422     /* Notify before starting migration thread */
2423     notifier_list_notify(&migration_state_notifiers, s);
2424
2425     /*
2426      * Open the return path. For postcopy, it is used exclusively. For
2427      * precopy, only if user specified "return-path" capability would
2428      * QEMU uses the return path.
2429      */
2430     if (migrate_postcopy_ram() || migrate_use_return_path()) {
2431         if (open_return_path_on_source(s)) {
2432             error_report("Unable to open return-path for postcopy");
2433             migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2434                               MIGRATION_STATUS_FAILED);
2435             migrate_fd_cleanup(s);
2436             return;
2437         }
2438     }
2439
2440     if (multifd_save_setup() != 0) {
2441         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2442                           MIGRATION_STATUS_FAILED);
2443         migrate_fd_cleanup(s);
2444         return;
2445     }
2446     qemu_thread_create(&s->thread, "live_migration", migration_thread, s,
2447                        QEMU_THREAD_JOINABLE);
2448     s->migration_thread_running = true;
2449 }
2450
2451 void migration_global_dump(Monitor *mon)
2452 {
2453     MigrationState *ms = migrate_get_current();
2454
2455     monitor_printf(mon, "globals:\n");
2456     monitor_printf(mon, "store-global-state: %s\n",
2457                    ms->store_global_state ? "on" : "off");
2458     monitor_printf(mon, "only-migratable: %s\n",
2459                    ms->only_migratable ? "on" : "off");
2460     monitor_printf(mon, "send-configuration: %s\n",
2461                    ms->send_configuration ? "on" : "off");
2462     monitor_printf(mon, "send-section-footer: %s\n",
2463                    ms->send_section_footer ? "on" : "off");
2464 }
2465
2466 #define DEFINE_PROP_MIG_CAP(name, x)             \
2467     DEFINE_PROP_BOOL(name, MigrationState, enabled_capabilities[x], false)
2468
2469 static Property migration_properties[] = {
2470     DEFINE_PROP_BOOL("store-global-state", MigrationState,
2471                      store_global_state, true),
2472     DEFINE_PROP_BOOL("only-migratable", MigrationState, only_migratable, false),
2473     DEFINE_PROP_BOOL("send-configuration", MigrationState,
2474                      send_configuration, true),
2475     DEFINE_PROP_BOOL("send-section-footer", MigrationState,
2476                      send_section_footer, true),
2477
2478     /* Migration parameters */
2479     DEFINE_PROP_UINT8("x-compress-level", MigrationState,
2480                       parameters.compress_level,
2481                       DEFAULT_MIGRATE_COMPRESS_LEVEL),
2482     DEFINE_PROP_UINT8("x-compress-threads", MigrationState,
2483                       parameters.compress_threads,
2484                       DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT),
2485     DEFINE_PROP_UINT8("x-decompress-threads", MigrationState,
2486                       parameters.decompress_threads,
2487                       DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT),
2488     DEFINE_PROP_UINT8("x-cpu-throttle-initial", MigrationState,
2489                       parameters.cpu_throttle_initial,
2490                       DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL),
2491     DEFINE_PROP_UINT8("x-cpu-throttle-increment", MigrationState,
2492                       parameters.cpu_throttle_increment,
2493                       DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT),
2494     DEFINE_PROP_SIZE("x-max-bandwidth", MigrationState,
2495                       parameters.max_bandwidth, MAX_THROTTLE),
2496     DEFINE_PROP_UINT64("x-downtime-limit", MigrationState,
2497                       parameters.downtime_limit,
2498                       DEFAULT_MIGRATE_SET_DOWNTIME),
2499     DEFINE_PROP_UINT32("x-checkpoint-delay", MigrationState,
2500                       parameters.x_checkpoint_delay,
2501                       DEFAULT_MIGRATE_X_CHECKPOINT_DELAY),
2502     DEFINE_PROP_UINT8("x-multifd-channels", MigrationState,
2503                       parameters.x_multifd_channels,
2504                       DEFAULT_MIGRATE_MULTIFD_CHANNELS),
2505     DEFINE_PROP_UINT32("x-multifd-page-count", MigrationState,
2506                       parameters.x_multifd_page_count,
2507                       DEFAULT_MIGRATE_MULTIFD_PAGE_COUNT),
2508     DEFINE_PROP_SIZE("xbzrle-cache-size", MigrationState,
2509                       parameters.xbzrle_cache_size,
2510                       DEFAULT_MIGRATE_XBZRLE_CACHE_SIZE),
2511
2512     /* Migration capabilities */
2513     DEFINE_PROP_MIG_CAP("x-xbzrle", MIGRATION_CAPABILITY_XBZRLE),
2514     DEFINE_PROP_MIG_CAP("x-rdma-pin-all", MIGRATION_CAPABILITY_RDMA_PIN_ALL),
2515     DEFINE_PROP_MIG_CAP("x-auto-converge", MIGRATION_CAPABILITY_AUTO_CONVERGE),
2516     DEFINE_PROP_MIG_CAP("x-zero-blocks", MIGRATION_CAPABILITY_ZERO_BLOCKS),
2517     DEFINE_PROP_MIG_CAP("x-compress", MIGRATION_CAPABILITY_COMPRESS),
2518     DEFINE_PROP_MIG_CAP("x-events", MIGRATION_CAPABILITY_EVENTS),
2519     DEFINE_PROP_MIG_CAP("x-postcopy-ram", MIGRATION_CAPABILITY_POSTCOPY_RAM),
2520     DEFINE_PROP_MIG_CAP("x-colo", MIGRATION_CAPABILITY_X_COLO),
2521     DEFINE_PROP_MIG_CAP("x-release-ram", MIGRATION_CAPABILITY_RELEASE_RAM),
2522     DEFINE_PROP_MIG_CAP("x-block", MIGRATION_CAPABILITY_BLOCK),
2523     DEFINE_PROP_MIG_CAP("x-return-path", MIGRATION_CAPABILITY_RETURN_PATH),
2524     DEFINE_PROP_MIG_CAP("x-multifd", MIGRATION_CAPABILITY_X_MULTIFD),
2525
2526     DEFINE_PROP_END_OF_LIST(),
2527 };
2528
2529 static void migration_class_init(ObjectClass *klass, void *data)
2530 {
2531     DeviceClass *dc = DEVICE_CLASS(klass);
2532
2533     dc->user_creatable = false;
2534     dc->props = migration_properties;
2535 }
2536
2537 static void migration_instance_finalize(Object *obj)
2538 {
2539     MigrationState *ms = MIGRATION_OBJ(obj);
2540     MigrationParameters *params = &ms->parameters;
2541
2542     qemu_mutex_destroy(&ms->error_mutex);
2543     g_free(params->tls_hostname);
2544     g_free(params->tls_creds);
2545     qemu_sem_destroy(&ms->pause_sem);
2546     error_free(ms->error);
2547 }
2548
2549 static void migration_instance_init(Object *obj)
2550 {
2551     MigrationState *ms = MIGRATION_OBJ(obj);
2552     MigrationParameters *params = &ms->parameters;
2553
2554     ms->state = MIGRATION_STATUS_NONE;
2555     ms->mbps = -1;
2556     qemu_sem_init(&ms->pause_sem, 0);
2557     qemu_mutex_init(&ms->error_mutex);
2558
2559     params->tls_hostname = g_strdup("");
2560     params->tls_creds = g_strdup("");
2561
2562     /* Set has_* up only for parameter checks */
2563     params->has_compress_level = true;
2564     params->has_compress_threads = true;
2565     params->has_decompress_threads = true;
2566     params->has_cpu_throttle_initial = true;
2567     params->has_cpu_throttle_increment = true;
2568     params->has_max_bandwidth = true;
2569     params->has_downtime_limit = true;
2570     params->has_x_checkpoint_delay = true;
2571     params->has_block_incremental = true;
2572     params->has_x_multifd_channels = true;
2573     params->has_x_multifd_page_count = true;
2574     params->has_xbzrle_cache_size = true;
2575 }
2576
2577 /*
2578  * Return true if check pass, false otherwise. Error will be put
2579  * inside errp if provided.
2580  */
2581 static bool migration_object_check(MigrationState *ms, Error **errp)
2582 {
2583     MigrationCapabilityStatusList *head = NULL;
2584     /* Assuming all off */
2585     bool cap_list[MIGRATION_CAPABILITY__MAX] = { 0 }, ret;
2586     int i;
2587
2588     if (!migrate_params_check(&ms->parameters, errp)) {
2589         return false;
2590     }
2591
2592     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
2593         if (ms->enabled_capabilities[i]) {
2594             head = migrate_cap_add(head, i, true);
2595         }
2596     }
2597
2598     ret = migrate_caps_check(cap_list, head, errp);
2599
2600     /* It works with head == NULL */
2601     qapi_free_MigrationCapabilityStatusList(head);
2602
2603     return ret;
2604 }
2605
2606 static const TypeInfo migration_type = {
2607     .name = TYPE_MIGRATION,
2608     /*
2609      * NOTE: TYPE_MIGRATION is not really a device, as the object is
2610      * not created using qdev_create(), it is not attached to the qdev
2611      * device tree, and it is never realized.
2612      *
2613      * TODO: Make this TYPE_OBJECT once QOM provides something like
2614      * TYPE_DEVICE's "-global" properties.
2615      */
2616     .parent = TYPE_DEVICE,
2617     .class_init = migration_class_init,
2618     .class_size = sizeof(MigrationClass),
2619     .instance_size = sizeof(MigrationState),
2620     .instance_init = migration_instance_init,
2621     .instance_finalize = migration_instance_finalize,
2622 };
2623
2624 static void register_migration_types(void)
2625 {
2626     type_register_static(&migration_type);
2627 }
2628
2629 type_init(register_migration_types);