migration/migration.c

   1 /*
   2  * QEMU live migration
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *
   6  * Authors:
   7  *  Anthony Liguori   <[email protected]>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Contributions after 2012-01-13 are licensed under the terms of the
  13  * GNU GPL, version 2 or (at your option) any later version.
  14  */
  15
  16 #include "qemu/osdep.h"
  17 #include "qemu/cutils.h"
  18 #include "qemu/error-report.h"
  19 #include "migration/blocker.h"
  20 #include "exec.h"
  21 #include "fd.h"
  22 #include "socket.h"
  23 #include "rdma.h"
  24 #include "ram.h"
  25 #include "migration/global_state.h"
  26 #include "migration/misc.h"
  27 #include "migration.h"
  28 #include "savevm.h"
  29 #include "qemu-file-channel.h"
  30 #include "qemu-file.h"
  31 #include "migration/vmstate.h"
  32 #include "block/block.h"
  33 #include "qapi/qmp/qerror.h"
  34 #include "qemu/rcu.h"
  35 #include "block.h"
  36 #include "postcopy-ram.h"
  37 #include "qemu/thread.h"
  38 #include "qmp-commands.h"
  39 #include "trace.h"
  40 #include "qapi-event.h"
  41 #include "exec/target_page.h"
  42 #include "io/channel-buffer.h"
  43 #include "migration/colo.h"
  44 #include "hw/boards.h"
  45 #include "monitor/monitor.h"
  46
  47 #define MAX_THROTTLE  (32 << 20)      /* Migration transfer speed throttling */
  48
  49 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
  50  * data. */
  51 #define BUFFER_DELAY     100
  52 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
  53
  54 /* Time in milliseconds we are allowed to stop the source,
  55  * for sending the last part */
  56 #define DEFAULT_MIGRATE_SET_DOWNTIME 300
  57
  58 /* Maximum migrate downtime set to 2000 seconds */
  59 #define MAX_MIGRATE_DOWNTIME_SECONDS 2000
  60 #define MAX_MIGRATE_DOWNTIME (MAX_MIGRATE_DOWNTIME_SECONDS * 1000)
  61
  62 /* Default compression thread count */
  63 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
  64 /* Default decompression thread count, usually decompression is at
  65  * least 4 times as fast as compression.*/
  66 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2
  67 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */
  68 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1
  69 /* Define default autoconverge cpu throttle migration parameters */
  70 #define DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL 20
  71 #define DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT 10
  72
  73 /* Migration XBZRLE default cache size */
  74 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
  75
  76 /* The delay time (in ms) between two COLO checkpoints
  77  * Note: Please change this default value to 10000 when we support hybrid mode.
  78  */
  79 #define DEFAULT_MIGRATE_X_CHECKPOINT_DELAY 200
  80
  81 static NotifierList migration_state_notifiers =
  82     NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
  83
  84 static bool deferred_incoming;
  85
  86 /* Messages sent on the return path from destination to source */
  87 enum mig_rp_message_type {
  88     MIG_RP_MSG_INVALID = 0,  /* Must be 0 */
  89     MIG_RP_MSG_SHUT,         /* sibling will not send any more RP messages */
  90     MIG_RP_MSG_PONG,         /* Response to a PING; data (seq: be32 ) */
  91
  92     MIG_RP_MSG_REQ_PAGES_ID, /* data (start: be64, len: be32, id: string) */
  93     MIG_RP_MSG_REQ_PAGES,    /* data (start: be64, len: be32) */
  94
  95     MIG_RP_MSG_MAX
  96 };
  97
  98 /* When we add fault tolerance, we could have several
  99    migrations at once.  For now we don't need to add
 100    dynamic creation of migration */
 101
 102 static MigrationState *current_migration;
 103
 104 static bool migration_object_check(MigrationState *ms, Error **errp);
 105
 106 void migration_object_init(void)
 107 {
 108     MachineState *ms = MACHINE(qdev_get_machine());
 109     Error *err = NULL;
 110
 111     /* This can only be called once. */
 112     assert(!current_migration);
 113     current_migration = MIGRATION_OBJ(object_new(TYPE_MIGRATION));
 114
 115     if (!migration_object_check(current_migration, &err)) {
 116         error_report_err(err);
 117         exit(1);
 118     }
 119
 120     /*
 121      * We cannot really do this in migration_instance_init() since at
 122      * that time global properties are not yet applied, then this
 123      * value will be definitely replaced by something else.
 124      */
 125     if (ms->enforce_config_section) {
 126         current_migration->send_configuration = true;
 127     }
 128 }
 129
 130 /* For outgoing */
 131 MigrationState *migrate_get_current(void)
 132 {
 133     /* This can only be called after the object created. */
 134     assert(current_migration);
 135     return current_migration;
 136 }
 137
 138 MigrationIncomingState *migration_incoming_get_current(void)
 139 {
 140     static bool once;
 141     static MigrationIncomingState mis_current;
 142
 143     if (!once) {
 144         mis_current.state = MIGRATION_STATUS_NONE;
 145         memset(&mis_current, 0, sizeof(MigrationIncomingState));
 146         qemu_mutex_init(&mis_current.rp_mutex);
 147         qemu_event_init(&mis_current.main_thread_load_event, false);
 148         once = true;
 149     }
 150     return &mis_current;
 151 }
 152
 153 void migration_incoming_state_destroy(void)
 154 {
 155     struct MigrationIncomingState *mis = migration_incoming_get_current();
 156
 157     if (mis->to_src_file) {
 158         /* Tell source that we are done */
 159         migrate_send_rp_shut(mis, qemu_file_get_error(mis->from_src_file) != 0);
 160         qemu_fclose(mis->to_src_file);
 161         mis->to_src_file = NULL;
 162     }
 163
 164     if (mis->from_src_file) {
 165         qemu_fclose(mis->from_src_file);
 166         mis->from_src_file = NULL;
 167     }
 168
 169     qemu_event_destroy(&mis->main_thread_load_event);
 170 }
 171
 172 static void migrate_generate_event(int new_state)
 173 {
 174     if (migrate_use_events()) {
 175         qapi_event_send_migration(new_state, &error_abort);
 176     }
 177 }
 178
 179 /*
 180  * Called on -incoming with a defer: uri.
 181  * The migration can be started later after any parameters have been
 182  * changed.
 183  */
 184 static void deferred_incoming_migration(Error **errp)
 185 {
 186     if (deferred_incoming) {
 187         error_setg(errp, "Incoming migration already deferred");
 188     }
 189     deferred_incoming = true;
 190 }
 191
 192 /*
 193  * Send a message on the return channel back to the source
 194  * of the migration.
 195  */
 196 static void migrate_send_rp_message(MigrationIncomingState *mis,
 197                                     enum mig_rp_message_type message_type,
 198                                     uint16_t len, void *data)
 199 {
 200     trace_migrate_send_rp_message((int)message_type, len);
 201     qemu_mutex_lock(&mis->rp_mutex);
 202     qemu_put_be16(mis->to_src_file, (unsigned int)message_type);
 203     qemu_put_be16(mis->to_src_file, len);
 204     qemu_put_buffer(mis->to_src_file, data, len);
 205     qemu_fflush(mis->to_src_file);
 206     qemu_mutex_unlock(&mis->rp_mutex);
 207 }
 208
 209 /* Request a range of pages from the source VM at the given
 210  * start address.
 211  *   rbname: Name of the RAMBlock to request the page in, if NULL it's the same
 212  *           as the last request (a name must have been given previously)
 213  *   Start: Address offset within the RB
 214  *   Len: Length in bytes required - must be a multiple of pagesize
 215  */
 216 void migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname,
 217                                ram_addr_t start, size_t len)
 218 {
 219     uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
 220     size_t msglen = 12; /* start + len */
 221
 222     *(uint64_t *)bufc = cpu_to_be64((uint64_t)start);
 223     *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len);
 224
 225     if (rbname) {
 226         int rbname_len = strlen(rbname);
 227         assert(rbname_len < 256);
 228
 229         bufc[msglen++] = rbname_len;
 230         memcpy(bufc + msglen, rbname, rbname_len);
 231         msglen += rbname_len;
 232         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES_ID, msglen, bufc);
 233     } else {
 234         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES, msglen, bufc);
 235     }
 236 }
 237
 238 void qemu_start_incoming_migration(const char *uri, Error **errp)
 239 {
 240     const char *p;
 241
 242     qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort);
 243     if (!strcmp(uri, "defer")) {
 244         deferred_incoming_migration(errp);
 245     } else if (strstart(uri, "tcp:", &p)) {
 246         tcp_start_incoming_migration(p, errp);
 247 #ifdef CONFIG_RDMA
 248     } else if (strstart(uri, "rdma:", &p)) {
 249         rdma_start_incoming_migration(p, errp);
 250 #endif
 251     } else if (strstart(uri, "exec:", &p)) {
 252         exec_start_incoming_migration(p, errp);
 253     } else if (strstart(uri, "unix:", &p)) {
 254         unix_start_incoming_migration(p, errp);
 255     } else if (strstart(uri, "fd:", &p)) {
 256         fd_start_incoming_migration(p, errp);
 257     } else {
 258         error_setg(errp, "unknown migration protocol: %s", uri);
 259     }
 260 }
 261
 262 static void process_incoming_migration_bh(void *opaque)
 263 {
 264     Error *local_err = NULL;
 265     MigrationIncomingState *mis = opaque;
 266
 267     /* Make sure all file formats flush their mutable metadata.
 268      * If we get an error here, just don't restart the VM yet. */
 269     bdrv_invalidate_cache_all(&local_err);
 270     if (local_err) {
 271         error_report_err(local_err);
 272         local_err = NULL;
 273         autostart = false;
 274     }
 275
 276     /*
 277      * This must happen after all error conditions are dealt with and
 278      * we're sure the VM is going to be running on this host.
 279      */
 280     qemu_announce_self();
 281
 282     /* If global state section was not received or we are in running
 283        state, we need to obey autostart. Any other state is set with
 284        runstate_set. */
 285
 286     if (!global_state_received() ||
 287         global_state_get_runstate() == RUN_STATE_RUNNING) {
 288         if (autostart) {
 289             vm_start();
 290         } else {
 291             runstate_set(RUN_STATE_PAUSED);
 292         }
 293     } else {
 294         runstate_set(global_state_get_runstate());
 295     }
 296     /*
 297      * This must happen after any state changes since as soon as an external
 298      * observer sees this event they might start to prod at the VM assuming
 299      * it's ready to use.
 300      */
 301     migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 302                       MIGRATION_STATUS_COMPLETED);
 303     qemu_bh_delete(mis->bh);
 304     migration_incoming_state_destroy();
 305 }
 306
 307 static void process_incoming_migration_co(void *opaque)
 308 {
 309     QEMUFile *f = opaque;
 310     MigrationIncomingState *mis = migration_incoming_get_current();
 311     PostcopyState ps;
 312     int ret;
 313
 314     mis->from_src_file = f;
 315     mis->largest_page_size = qemu_ram_pagesize_largest();
 316     postcopy_state_set(POSTCOPY_INCOMING_NONE);
 317     migrate_set_state(&mis->state, MIGRATION_STATUS_NONE,
 318                       MIGRATION_STATUS_ACTIVE);
 319     ret = qemu_loadvm_state(f);
 320
 321     ps = postcopy_state_get();
 322     trace_process_incoming_migration_co_end(ret, ps);
 323     if (ps != POSTCOPY_INCOMING_NONE) {
 324         if (ps == POSTCOPY_INCOMING_ADVISE) {
 325             /*
 326              * Where a migration had postcopy enabled (and thus went to advise)
 327              * but managed to complete within the precopy period, we can use
 328              * the normal exit.
 329              */
 330             postcopy_ram_incoming_cleanup(mis);
 331         } else if (ret >= 0) {
 332             /*
 333              * Postcopy was started, cleanup should happen at the end of the
 334              * postcopy thread.
 335              */
 336             trace_process_incoming_migration_co_postcopy_end_main();
 337             return;
 338         }
 339         /* Else if something went wrong then just fall out of the normal exit */
 340     }
 341
 342     /* we get COLO info, and know if we are in COLO mode */
 343     if (!ret && migration_incoming_enable_colo()) {
 344         mis->migration_incoming_co = qemu_coroutine_self();
 345         qemu_thread_create(&mis->colo_incoming_thread, "COLO incoming",
 346              colo_process_incoming_thread, mis, QEMU_THREAD_JOINABLE);
 347         mis->have_colo_incoming_thread = true;
 348         qemu_coroutine_yield();
 349
 350         /* Wait checkpoint incoming thread exit before free resource */
 351         qemu_thread_join(&mis->colo_incoming_thread);
 352     }
 353
 354     if (ret < 0) {
 355         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 356                           MIGRATION_STATUS_FAILED);
 357         error_report("load of migration failed: %s", strerror(-ret));
 358         qemu_fclose(mis->from_src_file);
 359         exit(EXIT_FAILURE);
 360     }
 361     mis->bh = qemu_bh_new(process_incoming_migration_bh, mis);
 362     qemu_bh_schedule(mis->bh);
 363 }
 364
 365 void migration_fd_process_incoming(QEMUFile *f)
 366 {
 367     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co, f);
 368
 369     qemu_file_set_blocking(f, false);
 370     qemu_coroutine_enter(co);
 371 }
 372
 373 /*
 374  * Send a 'SHUT' message on the return channel with the given value
 375  * to indicate that we've finished with the RP.  Non-0 value indicates
 376  * error.
 377  */
 378 void migrate_send_rp_shut(MigrationIncomingState *mis,
 379                           uint32_t value)
 380 {
 381     uint32_t buf;
 382
 383     buf = cpu_to_be32(value);
 384     migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf);
 385 }
 386
 387 /*
 388  * Send a 'PONG' message on the return channel with the given value
 389  * (normally in response to a 'PING')
 390  */
 391 void migrate_send_rp_pong(MigrationIncomingState *mis,
 392                           uint32_t value)
 393 {
 394     uint32_t buf;
 395
 396     buf = cpu_to_be32(value);
 397     migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf);
 398 }
 399
 400 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
 401 {
 402     MigrationCapabilityStatusList *head = NULL;
 403     MigrationCapabilityStatusList *caps;
 404     MigrationState *s = migrate_get_current();
 405     int i;
 406
 407     caps = NULL; /* silence compiler warning */
 408     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
 409 #ifndef CONFIG_LIVE_BLOCK_MIGRATION
 410         if (i == MIGRATION_CAPABILITY_BLOCK) {
 411             continue;
 412         }
 413 #endif
 414         if (head == NULL) {
 415             head = g_malloc0(sizeof(*caps));
 416             caps = head;
 417         } else {
 418             caps->next = g_malloc0(sizeof(*caps));
 419             caps = caps->next;
 420         }
 421         caps->value =
 422             g_malloc(sizeof(*caps->value));
 423         caps->value->capability = i;
 424         caps->value->state = s->enabled_capabilities[i];
 425     }
 426
 427     return head;
 428 }
 429
 430 MigrationParameters *qmp_query_migrate_parameters(Error **errp)
 431 {
 432     MigrationParameters *params;
 433     MigrationState *s = migrate_get_current();
 434
 435     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 436     params = g_malloc0(sizeof(*params));
 437     params->has_compress_level = true;
 438     params->compress_level = s->parameters.compress_level;
 439     params->has_compress_threads = true;
 440     params->compress_threads = s->parameters.compress_threads;
 441     params->has_decompress_threads = true;
 442     params->decompress_threads = s->parameters.decompress_threads;
 443     params->has_cpu_throttle_initial = true;
 444     params->cpu_throttle_initial = s->parameters.cpu_throttle_initial;
 445     params->has_cpu_throttle_increment = true;
 446     params->cpu_throttle_increment = s->parameters.cpu_throttle_increment;
 447     params->has_tls_creds = true;
 448     params->tls_creds = g_strdup(s->parameters.tls_creds);
 449     params->has_tls_hostname = true;
 450     params->tls_hostname = g_strdup(s->parameters.tls_hostname);
 451     params->has_max_bandwidth = true;
 452     params->max_bandwidth = s->parameters.max_bandwidth;
 453     params->has_downtime_limit = true;
 454     params->downtime_limit = s->parameters.downtime_limit;
 455     params->has_x_checkpoint_delay = true;
 456     params->x_checkpoint_delay = s->parameters.x_checkpoint_delay;
 457     params->has_block_incremental = true;
 458     params->block_incremental = s->parameters.block_incremental;
 459
 460     return params;
 461 }
 462
 463 /*
 464  * Return true if we're already in the middle of a migration
 465  * (i.e. any of the active or setup states)
 466  */
 467 static bool migration_is_setup_or_active(int state)
 468 {
 469     switch (state) {
 470     case MIGRATION_STATUS_ACTIVE:
 471     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 472     case MIGRATION_STATUS_SETUP:
 473         return true;
 474
 475     default:
 476         return false;
 477
 478     }
 479 }
 480
 481 static void populate_ram_info(MigrationInfo *info, MigrationState *s)
 482 {
 483     info->has_ram = true;
 484     info->ram = g_malloc0(sizeof(*info->ram));
 485     info->ram->transferred = ram_counters.transferred;
 486     info->ram->total = ram_bytes_total();
 487     info->ram->duplicate = ram_counters.duplicate;
 488     /* legacy value.  It is not used anymore */
 489     info->ram->skipped = 0;
 490     info->ram->normal = ram_counters.normal;
 491     info->ram->normal_bytes = ram_counters.normal *
 492         qemu_target_page_size();
 493     info->ram->mbps = s->mbps;
 494     info->ram->dirty_sync_count = ram_counters.dirty_sync_count;
 495     info->ram->postcopy_requests = ram_counters.postcopy_requests;
 496     info->ram->page_size = qemu_target_page_size();
 497
 498     if (migrate_use_xbzrle()) {
 499         info->has_xbzrle_cache = true;
 500         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
 501         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
 502         info->xbzrle_cache->bytes = xbzrle_counters.bytes;
 503         info->xbzrle_cache->pages = xbzrle_counters.pages;
 504         info->xbzrle_cache->cache_miss = xbzrle_counters.cache_miss;
 505         info->xbzrle_cache->cache_miss_rate = xbzrle_counters.cache_miss_rate;
 506         info->xbzrle_cache->overflow = xbzrle_counters.overflow;
 507     }
 508
 509     if (cpu_throttle_active()) {
 510         info->has_cpu_throttle_percentage = true;
 511         info->cpu_throttle_percentage = cpu_throttle_get_percentage();
 512     }
 513
 514     if (s->state != MIGRATION_STATUS_COMPLETED) {
 515         info->ram->remaining = ram_bytes_remaining();
 516         info->ram->dirty_pages_rate = ram_counters.dirty_pages_rate;
 517     }
 518 }
 519
 520 static void populate_disk_info(MigrationInfo *info)
 521 {
 522     if (blk_mig_active()) {
 523         info->has_disk = true;
 524         info->disk = g_malloc0(sizeof(*info->disk));
 525         info->disk->transferred = blk_mig_bytes_transferred();
 526         info->disk->remaining = blk_mig_bytes_remaining();
 527         info->disk->total = blk_mig_bytes_total();
 528     }
 529 }
 530
 531 MigrationInfo *qmp_query_migrate(Error **errp)
 532 {
 533     MigrationInfo *info = g_malloc0(sizeof(*info));
 534     MigrationState *s = migrate_get_current();
 535
 536     switch (s->state) {
 537     case MIGRATION_STATUS_NONE:
 538         /* no migration has happened ever */
 539         break;
 540     case MIGRATION_STATUS_SETUP:
 541         info->has_status = true;
 542         info->has_total_time = false;
 543         break;
 544     case MIGRATION_STATUS_ACTIVE:
 545     case MIGRATION_STATUS_CANCELLING:
 546     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 547          /* TODO add some postcopy stats */
 548         info->has_status = true;
 549         info->has_total_time = true;
 550         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 551             - s->total_time;
 552         info->has_expected_downtime = true;
 553         info->expected_downtime = s->expected_downtime;
 554         info->has_setup_time = true;
 555         info->setup_time = s->setup_time;
 556
 557         populate_ram_info(info, s);
 558         populate_disk_info(info);
 559         break;
 560     case MIGRATION_STATUS_COLO:
 561         info->has_status = true;
 562         /* TODO: display COLO specific information (checkpoint info etc.) */
 563         break;
 564     case MIGRATION_STATUS_COMPLETED:
 565         info->has_status = true;
 566         info->has_total_time = true;
 567         info->total_time = s->total_time;
 568         info->has_downtime = true;
 569         info->downtime = s->downtime;
 570         info->has_setup_time = true;
 571         info->setup_time = s->setup_time;
 572
 573         populate_ram_info(info, s);
 574         break;
 575     case MIGRATION_STATUS_FAILED:
 576         info->has_status = true;
 577         if (s->error) {
 578             info->has_error_desc = true;
 579             info->error_desc = g_strdup(error_get_pretty(s->error));
 580         }
 581         break;
 582     case MIGRATION_STATUS_CANCELLED:
 583         info->has_status = true;
 584         break;
 585     }
 586     info->status = s->state;
 587
 588     return info;
 589 }
 590
 591 /**
 592  * @migration_caps_check - check capability validity
 593  *
 594  * @cap_list: old capability list, array of bool
 595  * @params: new capabilities to be applied soon
 596  * @errp: set *errp if the check failed, with reason
 597  *
 598  * Returns true if check passed, otherwise false.
 599  */
 600 static bool migrate_caps_check(bool *cap_list,
 601                                MigrationCapabilityStatusList *params,
 602                                Error **errp)
 603 {
 604     MigrationCapabilityStatusList *cap;
 605     bool old_postcopy_cap;
 606
 607     old_postcopy_cap = cap_list[MIGRATION_CAPABILITY_POSTCOPY_RAM];
 608
 609     for (cap = params; cap; cap = cap->next) {
 610         cap_list[cap->value->capability] = cap->value->state;
 611     }
 612
 613 #ifndef CONFIG_LIVE_BLOCK_MIGRATION
 614     if (cap_list[MIGRATION_CAPABILITY_BLOCK]) {
 615         error_setg(errp, "QEMU compiled without old-style (blk/-b, inc/-i) "
 616                    "block migration");
 617         error_append_hint(errp, "Use drive_mirror+NBD instead.\n");
 618         return false;
 619     }
 620 #endif
 621
 622     if (cap_list[MIGRATION_CAPABILITY_POSTCOPY_RAM]) {
 623         if (cap_list[MIGRATION_CAPABILITY_COMPRESS]) {
 624             /* The decompression threads asynchronously write into RAM
 625              * rather than use the atomic copies needed to avoid
 626              * userfaulting.  It should be possible to fix the decompression
 627              * threads for compatibility in future.
 628              */
 629             error_setg(errp, "Postcopy is not currently compatible "
 630                        "with compression");
 631             return false;
 632         }
 633
 634         /* This check is reasonably expensive, so only when it's being
 635          * set the first time, also it's only the destination that needs
 636          * special support.
 637          */
 638         if (!old_postcopy_cap && runstate_check(RUN_STATE_INMIGRATE) &&
 639             !postcopy_ram_supported_by_host()) {
 640             /* postcopy_ram_supported_by_host will have emitted a more
 641              * detailed message
 642              */
 643             error_setg(errp, "Postcopy is not supported");
 644             return false;
 645         }
 646     }
 647
 648     return true;
 649 }
 650
 651 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 652                                   Error **errp)
 653 {
 654     MigrationState *s = migrate_get_current();
 655     MigrationCapabilityStatusList *cap;
 656
 657     if (migration_is_setup_or_active(s->state)) {
 658         error_setg(errp, QERR_MIGRATION_ACTIVE);
 659         return;
 660     }
 661
 662     if (!migrate_caps_check(s->enabled_capabilities, params, errp)) {
 663         return;
 664     }
 665
 666     for (cap = params; cap; cap = cap->next) {
 667         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 668     }
 669 }
 670
 671 /*
 672  * Check whether the parameters are valid. Error will be put into errp
 673  * (if provided). Return true if valid, otherwise false.
 674  */
 675 static bool migrate_params_check(MigrationParameters *params, Error **errp)
 676 {
 677     if (params->has_compress_level &&
 678         (params->compress_level < 0 || params->compress_level > 9)) {
 679         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level",
 680                    "is invalid, it should be in the range of 0 to 9");
 681         return false;
 682     }
 683
 684     if (params->has_compress_threads &&
 685         (params->compress_threads < 1 || params->compress_threads > 255)) {
 686         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 687                    "compress_threads",
 688                    "is invalid, it should be in the range of 1 to 255");
 689         return false;
 690     }
 691
 692     if (params->has_decompress_threads &&
 693         (params->decompress_threads < 1 || params->decompress_threads > 255)) {
 694         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 695                    "decompress_threads",
 696                    "is invalid, it should be in the range of 1 to 255");
 697         return false;
 698     }
 699
 700     if (params->has_cpu_throttle_initial &&
 701         (params->cpu_throttle_initial < 1 ||
 702          params->cpu_throttle_initial > 99)) {
 703         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 704                    "cpu_throttle_initial",
 705                    "an integer in the range of 1 to 99");
 706         return false;
 707     }
 708
 709     if (params->has_cpu_throttle_increment &&
 710         (params->cpu_throttle_increment < 1 ||
 711          params->cpu_throttle_increment > 99)) {
 712         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 713                    "cpu_throttle_increment",
 714                    "an integer in the range of 1 to 99");
 715         return false;
 716     }
 717
 718     if (params->has_max_bandwidth &&
 719         (params->max_bandwidth < 0 || params->max_bandwidth > SIZE_MAX)) {
 720         error_setg(errp, "Parameter 'max_bandwidth' expects an integer in the"
 721                          " range of 0 to %zu bytes/second", SIZE_MAX);
 722         return false;
 723     }
 724
 725     if (params->has_downtime_limit &&
 726         (params->downtime_limit < 0 ||
 727          params->downtime_limit > MAX_MIGRATE_DOWNTIME)) {
 728         error_setg(errp, "Parameter 'downtime_limit' expects an integer in "
 729                          "the range of 0 to %d milliseconds",
 730                          MAX_MIGRATE_DOWNTIME);
 731         return false;
 732     }
 733
 734     if (params->has_x_checkpoint_delay && (params->x_checkpoint_delay < 0)) {
 735         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 736                     "x_checkpoint_delay",
 737                     "is invalid, it should be positive");
 738         return false;
 739     }
 740
 741     return true;
 742 }
 743
 744 static void migrate_params_test_apply(MigrateSetParameters *params,
 745                                       MigrationParameters *dest)
 746 {
 747     *dest = migrate_get_current()->parameters;
 748
 749     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 750
 751     if (params->has_compress_level) {
 752         dest->compress_level = params->compress_level;
 753     }
 754
 755     if (params->has_compress_threads) {
 756         dest->compress_threads = params->compress_threads;
 757     }
 758
 759     if (params->has_decompress_threads) {
 760         dest->decompress_threads = params->decompress_threads;
 761     }
 762
 763     if (params->has_cpu_throttle_initial) {
 764         dest->cpu_throttle_initial = params->cpu_throttle_initial;
 765     }
 766
 767     if (params->has_cpu_throttle_increment) {
 768         dest->cpu_throttle_increment = params->cpu_throttle_increment;
 769     }
 770
 771     if (params->has_tls_creds) {
 772         assert(params->tls_creds->type == QTYPE_QSTRING);
 773         dest->tls_creds = g_strdup(params->tls_creds->u.s);
 774     }
 775
 776     if (params->has_tls_hostname) {
 777         assert(params->tls_hostname->type == QTYPE_QSTRING);
 778         dest->tls_hostname = g_strdup(params->tls_hostname->u.s);
 779     }
 780
 781     if (params->has_max_bandwidth) {
 782         dest->max_bandwidth = params->max_bandwidth;
 783     }
 784
 785     if (params->has_downtime_limit) {
 786         dest->downtime_limit = params->downtime_limit;
 787     }
 788
 789     if (params->has_x_checkpoint_delay) {
 790         dest->x_checkpoint_delay = params->x_checkpoint_delay;
 791     }
 792
 793     if (params->has_block_incremental) {
 794         dest->block_incremental = params->block_incremental;
 795     }
 796 }
 797
 798 static void migrate_params_apply(MigrateSetParameters *params)
 799 {
 800     MigrationState *s = migrate_get_current();
 801
 802     /* TODO use QAPI_CLONE() instead of duplicating it inline */
 803
 804     if (params->has_compress_level) {
 805         s->parameters.compress_level = params->compress_level;
 806     }
 807
 808     if (params->has_compress_threads) {
 809         s->parameters.compress_threads = params->compress_threads;
 810     }
 811
 812     if (params->has_decompress_threads) {
 813         s->parameters.decompress_threads = params->decompress_threads;
 814     }
 815
 816     if (params->has_cpu_throttle_initial) {
 817         s->parameters.cpu_throttle_initial = params->cpu_throttle_initial;
 818     }
 819
 820     if (params->has_cpu_throttle_increment) {
 821         s->parameters.cpu_throttle_increment = params->cpu_throttle_increment;
 822     }
 823
 824     if (params->has_tls_creds) {
 825         g_free(s->parameters.tls_creds);
 826         assert(params->tls_creds->type == QTYPE_QSTRING);
 827         s->parameters.tls_creds = g_strdup(params->tls_creds->u.s);
 828     }
 829
 830     if (params->has_tls_hostname) {
 831         g_free(s->parameters.tls_hostname);
 832         assert(params->tls_hostname->type == QTYPE_QSTRING);
 833         s->parameters.tls_hostname = g_strdup(params->tls_hostname->u.s);
 834     }
 835
 836     if (params->has_max_bandwidth) {
 837         s->parameters.max_bandwidth = params->max_bandwidth;
 838         if (s->to_dst_file) {
 839             qemu_file_set_rate_limit(s->to_dst_file,
 840                                 s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
 841         }
 842     }
 843
 844     if (params->has_downtime_limit) {
 845         s->parameters.downtime_limit = params->downtime_limit;
 846     }
 847
 848     if (params->has_x_checkpoint_delay) {
 849         s->parameters.x_checkpoint_delay = params->x_checkpoint_delay;
 850         if (migration_in_colo_state()) {
 851             colo_checkpoint_notify(s);
 852         }
 853     }
 854
 855     if (params->has_block_incremental) {
 856         s->parameters.block_incremental = params->block_incremental;
 857     }
 858 }
 859
 860 void qmp_migrate_set_parameters(MigrateSetParameters *params, Error **errp)
 861 {
 862     MigrationParameters tmp;
 863
 864     /* TODO Rewrite "" to null instead */
 865     if (params->has_tls_creds
 866         && params->tls_creds->type == QTYPE_QNULL) {
 867         QDECREF(params->tls_creds->u.n);
 868         params->tls_creds->type = QTYPE_QSTRING;
 869         params->tls_creds->u.s = strdup("");
 870     }
 871     /* TODO Rewrite "" to null instead */
 872     if (params->has_tls_hostname
 873         && params->tls_hostname->type == QTYPE_QNULL) {
 874         QDECREF(params->tls_hostname->u.n);
 875         params->tls_hostname->type = QTYPE_QSTRING;
 876         params->tls_hostname->u.s = strdup("");
 877     }
 878
 879     migrate_params_test_apply(params, &tmp);
 880
 881     if (!migrate_params_check(&tmp, errp)) {
 882         /* Invalid parameter */
 883         return;
 884     }
 885
 886     migrate_params_apply(params);
 887 }
 888
 889
 890 void qmp_migrate_start_postcopy(Error **errp)
 891 {
 892     MigrationState *s = migrate_get_current();
 893
 894     if (!migrate_postcopy_ram()) {
 895         error_setg(errp, "Enable postcopy with migrate_set_capability before"
 896                          " the start of migration");
 897         return;
 898     }
 899
 900     if (s->state == MIGRATION_STATUS_NONE) {
 901         error_setg(errp, "Postcopy must be started after migration has been"
 902                          " started");
 903         return;
 904     }
 905     /*
 906      * we don't error if migration has finished since that would be racy
 907      * with issuing this command.
 908      */
 909     atomic_set(&s->start_postcopy, true);
 910 }
 911
 912 /* shared migration helpers */
 913
 914 void migrate_set_state(int *state, int old_state, int new_state)
 915 {
 916     if (atomic_cmpxchg(state, old_state, new_state) == old_state) {
 917         trace_migrate_set_state(new_state);
 918         migrate_generate_event(new_state);
 919     }
 920 }
 921
 922 static MigrationCapabilityStatusList *migrate_cap_add(
 923     MigrationCapabilityStatusList *list,
 924     MigrationCapability index,
 925     bool state)
 926 {
 927     MigrationCapabilityStatusList *cap;
 928
 929     cap = g_new0(MigrationCapabilityStatusList, 1);
 930     cap->value = g_new0(MigrationCapabilityStatus, 1);
 931     cap->value->capability = index;
 932     cap->value->state = state;
 933     cap->next = list;
 934
 935     return cap;
 936 }
 937
 938 void migrate_set_block_enabled(bool value, Error **errp)
 939 {
 940     MigrationCapabilityStatusList *cap;
 941
 942     cap = migrate_cap_add(NULL, MIGRATION_CAPABILITY_BLOCK, value);
 943     qmp_migrate_set_capabilities(cap, errp);
 944     qapi_free_MigrationCapabilityStatusList(cap);
 945 }
 946
 947 static void migrate_set_block_incremental(MigrationState *s, bool value)
 948 {
 949     s->parameters.block_incremental = value;
 950 }
 951
 952 static void block_cleanup_parameters(MigrationState *s)
 953 {
 954     if (s->must_remove_block_options) {
 955         /* setting to false can never fail */
 956         migrate_set_block_enabled(false, &error_abort);
 957         migrate_set_block_incremental(s, false);
 958         s->must_remove_block_options = false;
 959     }
 960 }
 961
 962 static void migrate_fd_cleanup(void *opaque)
 963 {
 964     MigrationState *s = opaque;
 965
 966     qemu_bh_delete(s->cleanup_bh);
 967     s->cleanup_bh = NULL;
 968
 969     if (s->to_dst_file) {
 970         trace_migrate_fd_cleanup();
 971         qemu_mutex_unlock_iothread();
 972         if (s->migration_thread_running) {
 973             qemu_thread_join(&s->thread);
 974             s->migration_thread_running = false;
 975         }
 976         qemu_mutex_lock_iothread();
 977
 978         qemu_fclose(s->to_dst_file);
 979         s->to_dst_file = NULL;
 980     }
 981
 982     assert((s->state != MIGRATION_STATUS_ACTIVE) &&
 983            (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE));
 984
 985     if (s->state == MIGRATION_STATUS_CANCELLING) {
 986         migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING,
 987                           MIGRATION_STATUS_CANCELLED);
 988     }
 989
 990     notifier_list_notify(&migration_state_notifiers, s);
 991     block_cleanup_parameters(s);
 992 }
 993
 994 void migrate_fd_error(MigrationState *s, const Error *error)
 995 {
 996     trace_migrate_fd_error(error_get_pretty(error));
 997     assert(s->to_dst_file == NULL);
 998     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
 999                       MIGRATION_STATUS_FAILED);
1000     if (!s->error) {
1001         s->error = error_copy(error);
1002     }
1003     notifier_list_notify(&migration_state_notifiers, s);
1004     block_cleanup_parameters(s);
1005 }
1006
1007 static void migrate_fd_cancel(MigrationState *s)
1008 {
1009     int old_state ;
1010     QEMUFile *f = migrate_get_current()->to_dst_file;
1011     trace_migrate_fd_cancel();
1012
1013     if (s->rp_state.from_dst_file) {
1014         /* shutdown the rp socket, so causing the rp thread to shutdown */
1015         qemu_file_shutdown(s->rp_state.from_dst_file);
1016     }
1017
1018     do {
1019         old_state = s->state;
1020         if (!migration_is_setup_or_active(old_state)) {
1021             break;
1022         }
1023         migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING);
1024     } while (s->state != MIGRATION_STATUS_CANCELLING);
1025
1026     /*
1027      * If we're unlucky the migration code might be stuck somewhere in a
1028      * send/write while the network has failed and is waiting to timeout;
1029      * if we've got shutdown(2) available then we can force it to quit.
1030      * The outgoing qemu file gets closed in migrate_fd_cleanup that is
1031      * called in a bh, so there is no race against this cancel.
1032      */
1033     if (s->state == MIGRATION_STATUS_CANCELLING && f) {
1034         qemu_file_shutdown(f);
1035     }
1036     if (s->state == MIGRATION_STATUS_CANCELLING && s->block_inactive) {
1037         Error *local_err = NULL;
1038
1039         bdrv_invalidate_cache_all(&local_err);
1040         if (local_err) {
1041             error_report_err(local_err);
1042         } else {
1043             s->block_inactive = false;
1044         }
1045     }
1046     block_cleanup_parameters(s);
1047 }
1048
1049 void add_migration_state_change_notifier(Notifier *notify)
1050 {
1051     notifier_list_add(&migration_state_notifiers, notify);
1052 }
1053
1054 void remove_migration_state_change_notifier(Notifier *notify)
1055 {
1056     notifier_remove(notify);
1057 }
1058
1059 bool migration_in_setup(MigrationState *s)
1060 {
1061     return s->state == MIGRATION_STATUS_SETUP;
1062 }
1063
1064 bool migration_has_finished(MigrationState *s)
1065 {
1066     return s->state == MIGRATION_STATUS_COMPLETED;
1067 }
1068
1069 bool migration_has_failed(MigrationState *s)
1070 {
1071     return (s->state == MIGRATION_STATUS_CANCELLED ||
1072             s->state == MIGRATION_STATUS_FAILED);
1073 }
1074
1075 bool migration_in_postcopy(void)
1076 {
1077     MigrationState *s = migrate_get_current();
1078
1079     return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
1080 }
1081
1082 bool migration_in_postcopy_after_devices(MigrationState *s)
1083 {
1084     return migration_in_postcopy() && s->postcopy_after_devices;
1085 }
1086
1087 bool migration_is_idle(void)
1088 {
1089     MigrationState *s = migrate_get_current();
1090
1091     switch (s->state) {
1092     case MIGRATION_STATUS_NONE:
1093     case MIGRATION_STATUS_CANCELLED:
1094     case MIGRATION_STATUS_COMPLETED:
1095     case MIGRATION_STATUS_FAILED:
1096         return true;
1097     case MIGRATION_STATUS_SETUP:
1098     case MIGRATION_STATUS_CANCELLING:
1099     case MIGRATION_STATUS_ACTIVE:
1100     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
1101     case MIGRATION_STATUS_COLO:
1102         return false;
1103     case MIGRATION_STATUS__MAX:
1104         g_assert_not_reached();
1105     }
1106
1107     return false;
1108 }
1109
1110 MigrationState *migrate_init(void)
1111 {
1112     MigrationState *s = migrate_get_current();
1113
1114     /*
1115      * Reinitialise all migration state, except
1116      * parameters/capabilities that the user set, and
1117      * locks.
1118      */
1119     s->bytes_xfer = 0;
1120     s->xfer_limit = 0;
1121     s->cleanup_bh = 0;
1122     s->to_dst_file = NULL;
1123     s->state = MIGRATION_STATUS_NONE;
1124     s->rp_state.from_dst_file = NULL;
1125     s->rp_state.error = false;
1126     s->mbps = 0.0;
1127     s->downtime = 0;
1128     s->expected_downtime = 0;
1129     s->setup_time = 0;
1130     s->start_postcopy = false;
1131     s->postcopy_after_devices = false;
1132     s->migration_thread_running = false;
1133     error_free(s->error);
1134     s->error = NULL;
1135
1136     migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP);
1137
1138     s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1139     return s;
1140 }
1141
1142 static GSList *migration_blockers;
1143
1144 int migrate_add_blocker(Error *reason, Error **errp)
1145 {
1146     if (migrate_get_current()->only_migratable) {
1147         error_propagate(errp, error_copy(reason));
1148         error_prepend(errp, "disallowing migration blocker "
1149                           "(--only_migratable) for: ");
1150         return -EACCES;
1151     }
1152
1153     if (migration_is_idle()) {
1154         migration_blockers = g_slist_prepend(migration_blockers, reason);
1155         return 0;
1156     }
1157
1158     error_propagate(errp, error_copy(reason));
1159     error_prepend(errp, "disallowing migration blocker (migration in "
1160                       "progress) for: ");
1161     return -EBUSY;
1162 }
1163
1164 void migrate_del_blocker(Error *reason)
1165 {
1166     migration_blockers = g_slist_remove(migration_blockers, reason);
1167 }
1168
1169 void qmp_migrate_incoming(const char *uri, Error **errp)
1170 {
1171     Error *local_err = NULL;
1172     static bool once = true;
1173
1174     if (!deferred_incoming) {
1175         error_setg(errp, "For use with '-incoming defer'");
1176         return;
1177     }
1178     if (!once) {
1179         error_setg(errp, "The incoming migration has already been started");
1180     }
1181
1182     qemu_start_incoming_migration(uri, &local_err);
1183
1184     if (local_err) {
1185         error_propagate(errp, local_err);
1186         return;
1187     }
1188
1189     once = false;
1190 }
1191
1192 bool migration_is_blocked(Error **errp)
1193 {
1194     if (qemu_savevm_state_blocked(errp)) {
1195         return true;
1196     }
1197
1198     if (migration_blockers) {
1199         error_propagate(errp, error_copy(migration_blockers->data));
1200         return true;
1201     }
1202
1203     return false;
1204 }
1205
1206 void qmp_migrate(const char *uri, bool has_blk, bool blk,
1207                  bool has_inc, bool inc, bool has_detach, bool detach,
1208                  Error **errp)
1209 {
1210     Error *local_err = NULL;
1211     MigrationState *s = migrate_get_current();
1212     const char *p;
1213
1214     if (migration_is_setup_or_active(s->state) ||
1215         s->state == MIGRATION_STATUS_CANCELLING ||
1216         s->state == MIGRATION_STATUS_COLO) {
1217         error_setg(errp, QERR_MIGRATION_ACTIVE);
1218         return;
1219     }
1220     if (runstate_check(RUN_STATE_INMIGRATE)) {
1221         error_setg(errp, "Guest is waiting for an incoming migration");
1222         return;
1223     }
1224
1225     if (migration_is_blocked(errp)) {
1226         return;
1227     }
1228
1229     if ((has_blk && blk) || (has_inc && inc)) {
1230         if (migrate_use_block() || migrate_use_block_incremental()) {
1231             error_setg(errp, "Command options are incompatible with "
1232                        "current migration capabilities");
1233             return;
1234         }
1235         migrate_set_block_enabled(true, &local_err);
1236         if (local_err) {
1237             error_propagate(errp, local_err);
1238             return;
1239         }
1240         s->must_remove_block_options = true;
1241     }
1242
1243     if (has_inc && inc) {
1244         migrate_set_block_incremental(s, true);
1245     }
1246
1247     s = migrate_init();
1248
1249     if (strstart(uri, "tcp:", &p)) {
1250         tcp_start_outgoing_migration(s, p, &local_err);
1251 #ifdef CONFIG_RDMA
1252     } else if (strstart(uri, "rdma:", &p)) {
1253         rdma_start_outgoing_migration(s, p, &local_err);
1254 #endif
1255     } else if (strstart(uri, "exec:", &p)) {
1256         exec_start_outgoing_migration(s, p, &local_err);
1257     } else if (strstart(uri, "unix:", &p)) {
1258         unix_start_outgoing_migration(s, p, &local_err);
1259     } else if (strstart(uri, "fd:", &p)) {
1260         fd_start_outgoing_migration(s, p, &local_err);
1261     } else {
1262         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri",
1263                    "a valid migration protocol");
1264         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1265                           MIGRATION_STATUS_FAILED);
1266         return;
1267     }
1268
1269     if (local_err) {
1270         migrate_fd_error(s, local_err);
1271         error_propagate(errp, local_err);
1272         return;
1273     }
1274 }
1275
1276 void qmp_migrate_cancel(Error **errp)
1277 {
1278     migrate_fd_cancel(migrate_get_current());
1279 }
1280
1281 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
1282 {
1283     MigrationState *s = migrate_get_current();
1284     int64_t new_size;
1285
1286     /* Check for truncation */
1287     if (value != (size_t)value) {
1288         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1289                    "exceeding address space");
1290         return;
1291     }
1292
1293     /* Cache should not be larger than guest ram size */
1294     if (value > ram_bytes_total()) {
1295         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1296                    "exceeds guest ram size ");
1297         return;
1298     }
1299
1300     new_size = xbzrle_cache_resize(value);
1301     if (new_size < 0) {
1302         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1303                    "is smaller than page size");
1304         return;
1305     }
1306
1307     s->xbzrle_cache_size = new_size;
1308 }
1309
1310 int64_t qmp_query_migrate_cache_size(Error **errp)
1311 {
1312     return migrate_xbzrle_cache_size();
1313 }
1314
1315 void qmp_migrate_set_speed(int64_t value, Error **errp)
1316 {
1317     MigrateSetParameters p = {
1318         .has_max_bandwidth = true,
1319         .max_bandwidth = value,
1320     };
1321
1322     qmp_migrate_set_parameters(&p, errp);
1323 }
1324
1325 void qmp_migrate_set_downtime(double value, Error **errp)
1326 {
1327     if (value < 0 || value > MAX_MIGRATE_DOWNTIME_SECONDS) {
1328         error_setg(errp, "Parameter 'downtime_limit' expects an integer in "
1329                          "the range of 0 to %d seconds",
1330                          MAX_MIGRATE_DOWNTIME_SECONDS);
1331         return;
1332     }
1333
1334     value *= 1000; /* Convert to milliseconds */
1335     value = MAX(0, MIN(INT64_MAX, value));
1336
1337     MigrateSetParameters p = {
1338         .has_downtime_limit = true,
1339         .downtime_limit = value,
1340     };
1341
1342     qmp_migrate_set_parameters(&p, errp);
1343 }
1344
1345 bool migrate_release_ram(void)
1346 {
1347     MigrationState *s;
1348
1349     s = migrate_get_current();
1350
1351     return s->enabled_capabilities[MIGRATION_CAPABILITY_RELEASE_RAM];
1352 }
1353
1354 bool migrate_postcopy_ram(void)
1355 {
1356     MigrationState *s;
1357
1358     s = migrate_get_current();
1359
1360     return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM];
1361 }
1362
1363 bool migrate_auto_converge(void)
1364 {
1365     MigrationState *s;
1366
1367     s = migrate_get_current();
1368
1369     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
1370 }
1371
1372 bool migrate_zero_blocks(void)
1373 {
1374     MigrationState *s;
1375
1376     s = migrate_get_current();
1377
1378     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
1379 }
1380
1381 bool migrate_use_compression(void)
1382 {
1383     MigrationState *s;
1384
1385     s = migrate_get_current();
1386
1387     return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS];
1388 }
1389
1390 int migrate_compress_level(void)
1391 {
1392     MigrationState *s;
1393
1394     s = migrate_get_current();
1395
1396     return s->parameters.compress_level;
1397 }
1398
1399 int migrate_compress_threads(void)
1400 {
1401     MigrationState *s;
1402
1403     s = migrate_get_current();
1404
1405     return s->parameters.compress_threads;
1406 }
1407
1408 int migrate_decompress_threads(void)
1409 {
1410     MigrationState *s;
1411
1412     s = migrate_get_current();
1413
1414     return s->parameters.decompress_threads;
1415 }
1416
1417 bool migrate_use_events(void)
1418 {
1419     MigrationState *s;
1420
1421     s = migrate_get_current();
1422
1423     return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS];
1424 }
1425
1426 int migrate_use_xbzrle(void)
1427 {
1428     MigrationState *s;
1429
1430     s = migrate_get_current();
1431
1432     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
1433 }
1434
1435 int64_t migrate_xbzrle_cache_size(void)
1436 {
1437     MigrationState *s;
1438
1439     s = migrate_get_current();
1440
1441     return s->xbzrle_cache_size;
1442 }
1443
1444 bool migrate_use_block(void)
1445 {
1446     MigrationState *s;
1447
1448     s = migrate_get_current();
1449
1450     return s->enabled_capabilities[MIGRATION_CAPABILITY_BLOCK];
1451 }
1452
1453 bool migrate_use_return_path(void)
1454 {
1455     MigrationState *s;
1456
1457     s = migrate_get_current();
1458
1459     return s->enabled_capabilities[MIGRATION_CAPABILITY_RETURN_PATH];
1460 }
1461
1462 bool migrate_use_block_incremental(void)
1463 {
1464     MigrationState *s;
1465
1466     s = migrate_get_current();
1467
1468     return s->parameters.block_incremental;
1469 }
1470
1471 /* migration thread support */
1472 /*
1473  * Something bad happened to the RP stream, mark an error
1474  * The caller shall print or trace something to indicate why
1475  */
1476 static void mark_source_rp_bad(MigrationState *s)
1477 {
1478     s->rp_state.error = true;
1479 }
1480
1481 static struct rp_cmd_args {
1482     ssize_t     len; /* -1 = variable */
1483     const char *name;
1484 } rp_cmd_args[] = {
1485     [MIG_RP_MSG_INVALID]        = { .len = -1, .name = "INVALID" },
1486     [MIG_RP_MSG_SHUT]           = { .len =  4, .name = "SHUT" },
1487     [MIG_RP_MSG_PONG]           = { .len =  4, .name = "PONG" },
1488     [MIG_RP_MSG_REQ_PAGES]      = { .len = 12, .name = "REQ_PAGES" },
1489     [MIG_RP_MSG_REQ_PAGES_ID]   = { .len = -1, .name = "REQ_PAGES_ID" },
1490     [MIG_RP_MSG_MAX]            = { .len = -1, .name = "MAX" },
1491 };
1492
1493 /*
1494  * Process a request for pages received on the return path,
1495  * We're allowed to send more than requested (e.g. to round to our page size)
1496  * and we don't need to send pages that have already been sent.
1497  */
1498 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname,
1499                                        ram_addr_t start, size_t len)
1500 {
1501     long our_host_ps = getpagesize();
1502
1503     trace_migrate_handle_rp_req_pages(rbname, start, len);
1504
1505     /*
1506      * Since we currently insist on matching page sizes, just sanity check
1507      * we're being asked for whole host pages.
1508      */
1509     if (start & (our_host_ps-1) ||
1510        (len & (our_host_ps-1))) {
1511         error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1512                      " len: %zd", __func__, start, len);
1513         mark_source_rp_bad(ms);
1514         return;
1515     }
1516
1517     if (ram_save_queue_pages(rbname, start, len)) {
1518         mark_source_rp_bad(ms);
1519     }
1520 }
1521
1522 /*
1523  * Handles messages sent on the return path towards the source VM
1524  *
1525  */
1526 static void *source_return_path_thread(void *opaque)
1527 {
1528     MigrationState *ms = opaque;
1529     QEMUFile *rp = ms->rp_state.from_dst_file;
1530     uint16_t header_len, header_type;
1531     uint8_t buf[512];
1532     uint32_t tmp32, sibling_error;
1533     ram_addr_t start = 0; /* =0 to silence warning */
1534     size_t  len = 0, expected_len;
1535     int res;
1536
1537     trace_source_return_path_thread_entry();
1538     while (!ms->rp_state.error && !qemu_file_get_error(rp) &&
1539            migration_is_setup_or_active(ms->state)) {
1540         trace_source_return_path_thread_loop_top();
1541         header_type = qemu_get_be16(rp);
1542         header_len = qemu_get_be16(rp);
1543
1544         if (header_type >= MIG_RP_MSG_MAX ||
1545             header_type == MIG_RP_MSG_INVALID) {
1546             error_report("RP: Received invalid message 0x%04x length 0x%04x",
1547                     header_type, header_len);
1548             mark_source_rp_bad(ms);
1549             goto out;
1550         }
1551
1552         if ((rp_cmd_args[header_type].len != -1 &&
1553             header_len != rp_cmd_args[header_type].len) ||
1554             header_len > sizeof(buf)) {
1555             error_report("RP: Received '%s' message (0x%04x) with"
1556                     "incorrect length %d expecting %zu",
1557                     rp_cmd_args[header_type].name, header_type, header_len,
1558                     (size_t)rp_cmd_args[header_type].len);
1559             mark_source_rp_bad(ms);
1560             goto out;
1561         }
1562
1563         /* We know we've got a valid header by this point */
1564         res = qemu_get_buffer(rp, buf, header_len);
1565         if (res != header_len) {
1566             error_report("RP: Failed reading data for message 0x%04x"
1567                          " read %d expected %d",
1568                          header_type, res, header_len);
1569             mark_source_rp_bad(ms);
1570             goto out;
1571         }
1572
1573         /* OK, we have the message and the data */
1574         switch (header_type) {
1575         case MIG_RP_MSG_SHUT:
1576             sibling_error = ldl_be_p(buf);
1577             trace_source_return_path_thread_shut(sibling_error);
1578             if (sibling_error) {
1579                 error_report("RP: Sibling indicated error %d", sibling_error);
1580                 mark_source_rp_bad(ms);
1581             }
1582             /*
1583              * We'll let the main thread deal with closing the RP
1584              * we could do a shutdown(2) on it, but we're the only user
1585              * anyway, so there's nothing gained.
1586              */
1587             goto out;
1588
1589         case MIG_RP_MSG_PONG:
1590             tmp32 = ldl_be_p(buf);
1591             trace_source_return_path_thread_pong(tmp32);
1592             break;
1593
1594         case MIG_RP_MSG_REQ_PAGES:
1595             start = ldq_be_p(buf);
1596             len = ldl_be_p(buf + 8);
1597             migrate_handle_rp_req_pages(ms, NULL, start, len);
1598             break;
1599
1600         case MIG_RP_MSG_REQ_PAGES_ID:
1601             expected_len = 12 + 1; /* header + termination */
1602
1603             if (header_len >= expected_len) {
1604                 start = ldq_be_p(buf);
1605                 len = ldl_be_p(buf + 8);
1606                 /* Now we expect an idstr */
1607                 tmp32 = buf[12]; /* Length of the following idstr */
1608                 buf[13 + tmp32] = '\0';
1609                 expected_len += tmp32;
1610             }
1611             if (header_len != expected_len) {
1612                 error_report("RP: Req_Page_id with length %d expecting %zd",
1613                         header_len, expected_len);
1614                 mark_source_rp_bad(ms);
1615                 goto out;
1616             }
1617             migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len);
1618             break;
1619
1620         default:
1621             break;
1622         }
1623     }
1624     if (qemu_file_get_error(rp)) {
1625         trace_source_return_path_thread_bad_end();
1626         mark_source_rp_bad(ms);
1627     }
1628
1629     trace_source_return_path_thread_end();
1630 out:
1631     ms->rp_state.from_dst_file = NULL;
1632     qemu_fclose(rp);
1633     return NULL;
1634 }
1635
1636 static int open_return_path_on_source(MigrationState *ms)
1637 {
1638
1639     ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file);
1640     if (!ms->rp_state.from_dst_file) {
1641         return -1;
1642     }
1643
1644     trace_open_return_path_on_source();
1645     qemu_thread_create(&ms->rp_state.rp_thread, "return path",
1646                        source_return_path_thread, ms, QEMU_THREAD_JOINABLE);
1647
1648     trace_open_return_path_on_source_continue();
1649
1650     return 0;
1651 }
1652
1653 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
1654 static int await_return_path_close_on_source(MigrationState *ms)
1655 {
1656     /*
1657      * If this is a normal exit then the destination will send a SHUT and the
1658      * rp_thread will exit, however if there's an error we need to cause
1659      * it to exit.
1660      */
1661     if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
1662         /*
1663          * shutdown(2), if we have it, will cause it to unblock if it's stuck
1664          * waiting for the destination.
1665          */
1666         qemu_file_shutdown(ms->rp_state.from_dst_file);
1667         mark_source_rp_bad(ms);
1668     }
1669     trace_await_return_path_close_on_source_joining();
1670     qemu_thread_join(&ms->rp_state.rp_thread);
1671     trace_await_return_path_close_on_source_close();
1672     return ms->rp_state.error;
1673 }
1674
1675 /*
1676  * Switch from normal iteration to postcopy
1677  * Returns non-0 on error
1678  */
1679 static int postcopy_start(MigrationState *ms, bool *old_vm_running)
1680 {
1681     int ret;
1682     QIOChannelBuffer *bioc;
1683     QEMUFile *fb;
1684     int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1685     bool restart_block = false;
1686     migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE,
1687                       MIGRATION_STATUS_POSTCOPY_ACTIVE);
1688
1689     trace_postcopy_start();
1690     qemu_mutex_lock_iothread();
1691     trace_postcopy_start_set_run();
1692
1693     qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1694     *old_vm_running = runstate_is_running();
1695     global_state_store();
1696     ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1697     if (ret < 0) {
1698         goto fail;
1699     }
1700
1701     ret = bdrv_inactivate_all();
1702     if (ret < 0) {
1703         goto fail;
1704     }
1705     restart_block = true;
1706
1707     /*
1708      * Cause any non-postcopiable, but iterative devices to
1709      * send out their final data.
1710      */
1711     qemu_savevm_state_complete_precopy(ms->to_dst_file, true, false);
1712
1713     /*
1714      * in Finish migrate and with the io-lock held everything should
1715      * be quiet, but we've potentially still got dirty pages and we
1716      * need to tell the destination to throw any pages it's already received
1717      * that are dirty
1718      */
1719     if (ram_postcopy_send_discard_bitmap(ms)) {
1720         error_report("postcopy send discard bitmap failed");
1721         goto fail;
1722     }
1723
1724     /*
1725      * send rest of state - note things that are doing postcopy
1726      * will notice we're in POSTCOPY_ACTIVE and not actually
1727      * wrap their state up here
1728      */
1729     qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX);
1730     /* Ping just for debugging, helps line traces up */
1731     qemu_savevm_send_ping(ms->to_dst_file, 2);
1732
1733     /*
1734      * While loading the device state we may trigger page transfer
1735      * requests and the fd must be free to process those, and thus
1736      * the destination must read the whole device state off the fd before
1737      * it starts processing it.  Unfortunately the ad-hoc migration format
1738      * doesn't allow the destination to know the size to read without fully
1739      * parsing it through each devices load-state code (especially the open
1740      * coded devices that use get/put).
1741      * So we wrap the device state up in a package with a length at the start;
1742      * to do this we use a qemu_buf to hold the whole of the device state.
1743      */
1744     bioc = qio_channel_buffer_new(4096);
1745     qio_channel_set_name(QIO_CHANNEL(bioc), "migration-postcopy-buffer");
1746     fb = qemu_fopen_channel_output(QIO_CHANNEL(bioc));
1747     object_unref(OBJECT(bioc));
1748
1749     /*
1750      * Make sure the receiver can get incoming pages before we send the rest
1751      * of the state
1752      */
1753     qemu_savevm_send_postcopy_listen(fb);
1754
1755     qemu_savevm_state_complete_precopy(fb, false, false);
1756     qemu_savevm_send_ping(fb, 3);
1757
1758     qemu_savevm_send_postcopy_run(fb);
1759
1760     /* <><> end of stuff going into the package */
1761
1762     /* Last point of recovery; as soon as we send the package the destination
1763      * can open devices and potentially start running.
1764      * Lets just check again we've not got any errors.
1765      */
1766     ret = qemu_file_get_error(ms->to_dst_file);
1767     if (ret) {
1768         error_report("postcopy_start: Migration stream errored (pre package)");
1769         goto fail_closefb;
1770     }
1771
1772     restart_block = false;
1773
1774     /* Now send that blob */
1775     if (qemu_savevm_send_packaged(ms->to_dst_file, bioc->data, bioc->usage)) {
1776         goto fail_closefb;
1777     }
1778     qemu_fclose(fb);
1779
1780     /* Send a notify to give a chance for anything that needs to happen
1781      * at the transition to postcopy and after the device state; in particular
1782      * spice needs to trigger a transition now
1783      */
1784     ms->postcopy_after_devices = true;
1785     notifier_list_notify(&migration_state_notifiers, ms);
1786
1787     ms->downtime =  qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop;
1788
1789     qemu_mutex_unlock_iothread();
1790
1791     /*
1792      * Although this ping is just for debug, it could potentially be
1793      * used for getting a better measurement of downtime at the source.
1794      */
1795     qemu_savevm_send_ping(ms->to_dst_file, 4);
1796
1797     if (migrate_release_ram()) {
1798         ram_postcopy_migrated_memory_release(ms);
1799     }
1800
1801     ret = qemu_file_get_error(ms->to_dst_file);
1802     if (ret) {
1803         error_report("postcopy_start: Migration stream errored");
1804         migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1805                               MIGRATION_STATUS_FAILED);
1806     }
1807
1808     return ret;
1809
1810 fail_closefb:
1811     qemu_fclose(fb);
1812 fail:
1813     migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1814                           MIGRATION_STATUS_FAILED);
1815     if (restart_block) {
1816         /* A failure happened early enough that we know the destination hasn't
1817          * accessed block devices, so we're safe to recover.
1818          */
1819         Error *local_err = NULL;
1820
1821         bdrv_invalidate_cache_all(&local_err);
1822         if (local_err) {
1823             error_report_err(local_err);
1824         }
1825     }
1826     qemu_mutex_unlock_iothread();
1827     return -1;
1828 }
1829
1830 /**
1831  * migration_completion: Used by migration_thread when there's not much left.
1832  *   The caller 'breaks' the loop when this returns.
1833  *
1834  * @s: Current migration state
1835  * @current_active_state: The migration state we expect to be in
1836  * @*old_vm_running: Pointer to old_vm_running flag
1837  * @*start_time: Pointer to time to update
1838  */
1839 static void migration_completion(MigrationState *s, int current_active_state,
1840                                  bool *old_vm_running,
1841                                  int64_t *start_time)
1842 {
1843     int ret;
1844
1845     if (s->state == MIGRATION_STATUS_ACTIVE) {
1846         qemu_mutex_lock_iothread();
1847         *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1848         qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1849         *old_vm_running = runstate_is_running();
1850         ret = global_state_store();
1851
1852         if (!ret) {
1853             bool inactivate = !migrate_colo_enabled();
1854             ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1855             if (ret >= 0) {
1856                 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX);
1857                 ret = qemu_savevm_state_complete_precopy(s->to_dst_file, false,
1858                                                          inactivate);
1859             }
1860             if (inactivate && ret >= 0) {
1861                 s->block_inactive = true;
1862             }
1863         }
1864         qemu_mutex_unlock_iothread();
1865
1866         if (ret < 0) {
1867             goto fail;
1868         }
1869     } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1870         trace_migration_completion_postcopy_end();
1871
1872         qemu_savevm_state_complete_postcopy(s->to_dst_file);
1873         trace_migration_completion_postcopy_end_after_complete();
1874     }
1875
1876     /*
1877      * If rp was opened we must clean up the thread before
1878      * cleaning everything else up (since if there are no failures
1879      * it will wait for the destination to send it's status in
1880      * a SHUT command).
1881      */
1882     if (s->rp_state.from_dst_file) {
1883         int rp_error;
1884         trace_migration_return_path_end_before();
1885         rp_error = await_return_path_close_on_source(s);
1886         trace_migration_return_path_end_after(rp_error);
1887         if (rp_error) {
1888             goto fail_invalidate;
1889         }
1890     }
1891
1892     if (qemu_file_get_error(s->to_dst_file)) {
1893         trace_migration_completion_file_err();
1894         goto fail_invalidate;
1895     }
1896
1897     if (!migrate_colo_enabled()) {
1898         migrate_set_state(&s->state, current_active_state,
1899                           MIGRATION_STATUS_COMPLETED);
1900     }
1901
1902     return;
1903
1904 fail_invalidate:
1905     /* If not doing postcopy, vm_start() will be called: let's regain
1906      * control on images.
1907      */
1908     if (s->state == MIGRATION_STATUS_ACTIVE) {
1909         Error *local_err = NULL;
1910
1911         qemu_mutex_lock_iothread();
1912         bdrv_invalidate_cache_all(&local_err);
1913         if (local_err) {
1914             error_report_err(local_err);
1915         } else {
1916             s->block_inactive = false;
1917         }
1918         qemu_mutex_unlock_iothread();
1919     }
1920
1921 fail:
1922     migrate_set_state(&s->state, current_active_state,
1923                       MIGRATION_STATUS_FAILED);
1924 }
1925
1926 bool migrate_colo_enabled(void)
1927 {
1928     MigrationState *s = migrate_get_current();
1929     return s->enabled_capabilities[MIGRATION_CAPABILITY_X_COLO];
1930 }
1931
1932 /*
1933  * Master migration thread on the source VM.
1934  * It drives the migration and pumps the data down the outgoing channel.
1935  */
1936 static void *migration_thread(void *opaque)
1937 {
1938     MigrationState *s = opaque;
1939     /* Used by the bandwidth calcs, updated later */
1940     int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1941     int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
1942     int64_t initial_bytes = 0;
1943     /*
1944      * The final stage happens when the remaining data is smaller than
1945      * this threshold; it's calculated from the requested downtime and
1946      * measured bandwidth
1947      */
1948     int64_t threshold_size = 0;
1949     int64_t start_time = initial_time;
1950     int64_t end_time;
1951     bool old_vm_running = false;
1952     bool entered_postcopy = false;
1953     /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */
1954     enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE;
1955     bool enable_colo = migrate_colo_enabled();
1956
1957     rcu_register_thread();
1958
1959     qemu_savevm_state_header(s->to_dst_file);
1960
1961     /*
1962      * If we opened the return path, we need to make sure dst has it
1963      * opened as well.
1964      */
1965     if (s->rp_state.from_dst_file) {
1966         /* Now tell the dest that it should open its end so it can reply */
1967         qemu_savevm_send_open_return_path(s->to_dst_file);
1968
1969         /* And do a ping that will make stuff easier to debug */
1970         qemu_savevm_send_ping(s->to_dst_file, 1);
1971     }
1972
1973     if (migrate_postcopy_ram()) {
1974         /*
1975          * Tell the destination that we *might* want to do postcopy later;
1976          * if the other end can't do postcopy it should fail now, nice and
1977          * early.
1978          */
1979         qemu_savevm_send_postcopy_advise(s->to_dst_file);
1980     }
1981
1982     qemu_savevm_state_setup(s->to_dst_file);
1983
1984     s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
1985     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1986                       MIGRATION_STATUS_ACTIVE);
1987
1988     trace_migration_thread_setup_complete();
1989
1990     while (s->state == MIGRATION_STATUS_ACTIVE ||
1991            s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1992         int64_t current_time;
1993         uint64_t pending_size;
1994
1995         if (!qemu_file_rate_limit(s->to_dst_file)) {
1996             uint64_t pend_post, pend_nonpost;
1997
1998             qemu_savevm_state_pending(s->to_dst_file, threshold_size,
1999                                       &pend_nonpost, &pend_post);
2000             pending_size = pend_nonpost + pend_post;
2001             trace_migrate_pending(pending_size, threshold_size,
2002                                   pend_post, pend_nonpost);
2003             if (pending_size && pending_size >= threshold_size) {
2004                 /* Still a significant amount to transfer */
2005
2006                 if (migrate_postcopy_ram() &&
2007                     s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE &&
2008                     pend_nonpost <= threshold_size &&
2009                     atomic_read(&s->start_postcopy)) {
2010
2011                     if (!postcopy_start(s, &old_vm_running)) {
2012                         current_active_state = MIGRATION_STATUS_POSTCOPY_ACTIVE;
2013                         entered_postcopy = true;
2014                     }
2015
2016                     continue;
2017                 }
2018                 /* Just another iteration step */
2019                 qemu_savevm_state_iterate(s->to_dst_file, entered_postcopy);
2020             } else {
2021                 trace_migration_thread_low_pending(pending_size);
2022                 migration_completion(s, current_active_state,
2023                                      &old_vm_running, &start_time);
2024                 break;
2025             }
2026         }
2027
2028         if (qemu_file_get_error(s->to_dst_file)) {
2029             migrate_set_state(&s->state, current_active_state,
2030                               MIGRATION_STATUS_FAILED);
2031             trace_migration_thread_file_err();
2032             break;
2033         }
2034         current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2035         if (current_time >= initial_time + BUFFER_DELAY) {
2036             uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) -
2037                                          initial_bytes;
2038             uint64_t time_spent = current_time - initial_time;
2039             double bandwidth = (double)transferred_bytes / time_spent;
2040             threshold_size = bandwidth * s->parameters.downtime_limit;
2041
2042             s->mbps = (((double) transferred_bytes * 8.0) /
2043                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0;
2044
2045             trace_migrate_transferred(transferred_bytes, time_spent,
2046                                       bandwidth, threshold_size);
2047             /* if we haven't sent anything, we don't want to recalculate
2048                10000 is a small enough number for our purposes */
2049             if (ram_counters.dirty_pages_rate && transferred_bytes > 10000) {
2050                 s->expected_downtime = ram_counters.dirty_pages_rate *
2051                     qemu_target_page_size() / bandwidth;
2052             }
2053
2054             qemu_file_reset_rate_limit(s->to_dst_file);
2055             initial_time = current_time;
2056             initial_bytes = qemu_ftell(s->to_dst_file);
2057         }
2058         if (qemu_file_rate_limit(s->to_dst_file)) {
2059             /* usleep expects microseconds */
2060             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
2061         }
2062     }
2063
2064     trace_migration_thread_after_loop();
2065     /* If we enabled cpu throttling for auto-converge, turn it off. */
2066     cpu_throttle_stop();
2067     end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
2068
2069     qemu_mutex_lock_iothread();
2070     /*
2071      * The resource has been allocated by migration will be reused in COLO
2072      * process, so don't release them.
2073      */
2074     if (!enable_colo) {
2075         qemu_savevm_state_cleanup();
2076     }
2077     if (s->state == MIGRATION_STATUS_COMPLETED) {
2078         uint64_t transferred_bytes = qemu_ftell(s->to_dst_file);
2079         s->total_time = end_time - s->total_time;
2080         if (!entered_postcopy) {
2081             s->downtime = end_time - start_time;
2082         }
2083         if (s->total_time) {
2084             s->mbps = (((double) transferred_bytes * 8.0) /
2085                        ((double) s->total_time)) / 1000;
2086         }
2087         runstate_set(RUN_STATE_POSTMIGRATE);
2088     } else {
2089         if (s->state == MIGRATION_STATUS_ACTIVE && enable_colo) {
2090             migrate_start_colo_process(s);
2091             qemu_savevm_state_cleanup();
2092             /*
2093             * Fixme: we will run VM in COLO no matter its old running state.
2094             * After exited COLO, we will keep running.
2095             */
2096             old_vm_running = true;
2097         }
2098         if (old_vm_running && !entered_postcopy) {
2099             vm_start();
2100         } else {
2101             if (runstate_check(RUN_STATE_FINISH_MIGRATE)) {
2102                 runstate_set(RUN_STATE_POSTMIGRATE);
2103             }
2104         }
2105     }
2106     qemu_bh_schedule(s->cleanup_bh);
2107     qemu_mutex_unlock_iothread();
2108
2109     rcu_unregister_thread();
2110     return NULL;
2111 }
2112
2113 void migrate_fd_connect(MigrationState *s)
2114 {
2115     s->expected_downtime = s->parameters.downtime_limit;
2116     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
2117
2118     qemu_file_set_blocking(s->to_dst_file, true);
2119     qemu_file_set_rate_limit(s->to_dst_file,
2120                              s->parameters.max_bandwidth / XFER_LIMIT_RATIO);
2121
2122     /* Notify before starting migration thread */
2123     notifier_list_notify(&migration_state_notifiers, s);
2124
2125     /*
2126      * Open the return path. For postcopy, it is used exclusively. For
2127      * precopy, only if user specified "return-path" capability would
2128      * QEMU uses the return path.
2129      */
2130     if (migrate_postcopy_ram() || migrate_use_return_path()) {
2131         if (open_return_path_on_source(s)) {
2132             error_report("Unable to open return-path for postcopy");
2133             migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
2134                               MIGRATION_STATUS_FAILED);
2135             migrate_fd_cleanup(s);
2136             return;
2137         }
2138     }
2139
2140     qemu_thread_create(&s->thread, "live_migration", migration_thread, s,
2141                        QEMU_THREAD_JOINABLE);
2142     s->migration_thread_running = true;
2143 }
2144
2145 void migration_global_dump(Monitor *mon)
2146 {
2147     MigrationState *ms = migrate_get_current();
2148
2149     monitor_printf(mon, "globals: store-global-state=%d, only_migratable=%d, "
2150                    "send-configuration=%d, send-section-footer=%d\n",
2151                    ms->store_global_state, ms->only_migratable,
2152                    ms->send_configuration, ms->send_section_footer);
2153 }
2154
2155 #define DEFINE_PROP_MIG_CAP(name, x)             \
2156     DEFINE_PROP_BOOL(name, MigrationState, enabled_capabilities[x], false)
2157
2158 static Property migration_properties[] = {
2159     DEFINE_PROP_BOOL("store-global-state", MigrationState,
2160                      store_global_state, true),
2161     DEFINE_PROP_BOOL("only-migratable", MigrationState, only_migratable, false),
2162     DEFINE_PROP_BOOL("send-configuration", MigrationState,
2163                      send_configuration, true),
2164     DEFINE_PROP_BOOL("send-section-footer", MigrationState,
2165                      send_section_footer, true),
2166
2167     /* Migration parameters */
2168     DEFINE_PROP_INT64("x-compress-level", MigrationState,
2169                       parameters.compress_level,
2170                       DEFAULT_MIGRATE_COMPRESS_LEVEL),
2171     DEFINE_PROP_INT64("x-compress-threads", MigrationState,
2172                       parameters.compress_threads,
2173                       DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT),
2174     DEFINE_PROP_INT64("x-decompress-threads", MigrationState,
2175                       parameters.decompress_threads,
2176                       DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT),
2177     DEFINE_PROP_INT64("x-cpu-throttle-initial", MigrationState,
2178                       parameters.cpu_throttle_initial,
2179                       DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL),
2180     DEFINE_PROP_INT64("x-cpu-throttle-increment", MigrationState,
2181                       parameters.cpu_throttle_increment,
2182                       DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT),
2183     DEFINE_PROP_INT64("x-max-bandwidth", MigrationState,
2184                       parameters.max_bandwidth, MAX_THROTTLE),
2185     DEFINE_PROP_INT64("x-downtime-limit", MigrationState,
2186                       parameters.downtime_limit,
2187                       DEFAULT_MIGRATE_SET_DOWNTIME),
2188     DEFINE_PROP_INT64("x-checkpoint-delay", MigrationState,
2189                       parameters.x_checkpoint_delay,
2190                       DEFAULT_MIGRATE_X_CHECKPOINT_DELAY),
2191
2192     /* Migration capabilities */
2193     DEFINE_PROP_MIG_CAP("x-xbzrle", MIGRATION_CAPABILITY_XBZRLE),
2194     DEFINE_PROP_MIG_CAP("x-rdma-pin-all", MIGRATION_CAPABILITY_RDMA_PIN_ALL),
2195     DEFINE_PROP_MIG_CAP("x-auto-converge", MIGRATION_CAPABILITY_AUTO_CONVERGE),
2196     DEFINE_PROP_MIG_CAP("x-zero-blocks", MIGRATION_CAPABILITY_ZERO_BLOCKS),
2197     DEFINE_PROP_MIG_CAP("x-compress", MIGRATION_CAPABILITY_COMPRESS),
2198     DEFINE_PROP_MIG_CAP("x-events", MIGRATION_CAPABILITY_EVENTS),
2199     DEFINE_PROP_MIG_CAP("x-postcopy-ram", MIGRATION_CAPABILITY_POSTCOPY_RAM),
2200     DEFINE_PROP_MIG_CAP("x-colo", MIGRATION_CAPABILITY_X_COLO),
2201     DEFINE_PROP_MIG_CAP("x-release-ram", MIGRATION_CAPABILITY_RELEASE_RAM),
2202     DEFINE_PROP_MIG_CAP("x-block", MIGRATION_CAPABILITY_BLOCK),
2203     DEFINE_PROP_MIG_CAP("x-return-path", MIGRATION_CAPABILITY_RETURN_PATH),
2204
2205     DEFINE_PROP_END_OF_LIST(),
2206 };
2207
2208 static void migration_class_init(ObjectClass *klass, void *data)
2209 {
2210     DeviceClass *dc = DEVICE_CLASS(klass);
2211
2212     dc->user_creatable = false;
2213     dc->props = migration_properties;
2214 }
2215
2216 static void migration_instance_finalize(Object *obj)
2217 {
2218     MigrationState *ms = MIGRATION_OBJ(obj);
2219     MigrationParameters *params = &ms->parameters;
2220
2221     g_free(params->tls_hostname);
2222     g_free(params->tls_creds);
2223 }
2224
2225 static void migration_instance_init(Object *obj)
2226 {
2227     MigrationState *ms = MIGRATION_OBJ(obj);
2228     MigrationParameters *params = &ms->parameters;
2229
2230     ms->state = MIGRATION_STATUS_NONE;
2231     ms->xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE;
2232     ms->mbps = -1;
2233
2234     params->tls_hostname = g_strdup("");
2235     params->tls_creds = g_strdup("");
2236
2237     /* Set has_* up only for parameter checks */
2238     params->has_compress_level = true;
2239     params->has_compress_threads = true;
2240     params->has_decompress_threads = true;
2241     params->has_cpu_throttle_initial = true;
2242     params->has_cpu_throttle_increment = true;
2243     params->has_max_bandwidth = true;
2244     params->has_downtime_limit = true;
2245     params->has_x_checkpoint_delay = true;
2246     params->has_block_incremental = true;
2247 }
2248
2249 /*
2250  * Return true if check pass, false otherwise. Error will be put
2251  * inside errp if provided.
2252  */
2253 static bool migration_object_check(MigrationState *ms, Error **errp)
2254 {
2255     MigrationCapabilityStatusList *head = NULL;
2256     /* Assuming all off */
2257     bool cap_list[MIGRATION_CAPABILITY__MAX] = { 0 }, ret;
2258     int i;
2259
2260     if (!migrate_params_check(&ms->parameters, errp)) {
2261         return false;
2262     }
2263
2264     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
2265         if (ms->enabled_capabilities[i]) {
2266             head = migrate_cap_add(head, i, true);
2267         }
2268     }
2269
2270     ret = migrate_caps_check(cap_list, head, errp);
2271
2272     /* It works with head == NULL */
2273     qapi_free_MigrationCapabilityStatusList(head);
2274
2275     return ret;
2276 }
2277
2278 static const TypeInfo migration_type = {
2279     .name = TYPE_MIGRATION,
2280     /*
2281      * NOTE: TYPE_MIGRATION is not really a device, as the object is
2282      * not created using qdev_create(), it is not attached to the qdev
2283      * device tree, and it is never realized.
2284      *
2285      * TODO: Make this TYPE_OBJECT once QOM provides something like
2286      * TYPE_DEVICE's "-global" properties.
2287      */
2288     .parent = TYPE_DEVICE,
2289     .class_init = migration_class_init,
2290     .class_size = sizeof(MigrationClass),
2291     .instance_size = sizeof(MigrationState),
2292     .instance_init = migration_instance_init,
2293     .instance_finalize = migration_instance_finalize,
2294 };
2295
2296 static void register_migration_types(void)
2297 {
2298     type_register_static(&migration_type);
2299 }
2300
2301 type_init(register_migration_types);