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 "qemu/main-loop.h"
  20 #include "migration/migration.h"
  21 #include "migration/qemu-file.h"
  22 #include "sysemu/sysemu.h"
  23 #include "block/block.h"
  24 #include "qapi/qmp/qerror.h"
  25 #include "qapi/util.h"
  26 #include "qemu/sockets.h"
  27 #include "qemu/rcu.h"
  28 #include "migration/block.h"
  29 #include "migration/postcopy-ram.h"
  30 #include "qemu/thread.h"
  31 #include "qmp-commands.h"
  32 #include "trace.h"
  33 #include "qapi-event.h"
  34 #include "qom/cpu.h"
  35 #include "exec/memory.h"
  36 #include "exec/address-spaces.h"
  37 #include "io/channel-buffer.h"
  38
  39 #define MAX_THROTTLE  (32 << 20)      /* Migration transfer speed throttling */
  40
  41 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
  42  * data. */
  43 #define BUFFER_DELAY     100
  44 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY)
  45
  46 /* Default compression thread count */
  47 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
  48 /* Default decompression thread count, usually decompression is at
  49  * least 4 times as fast as compression.*/
  50 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2
  51 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */
  52 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1
  53 /* Define default autoconverge cpu throttle migration parameters */
  54 #define DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL 20
  55 #define DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT 10
  56
  57 /* Migration XBZRLE default cache size */
  58 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024)
  59
  60 static NotifierList migration_state_notifiers =
  61     NOTIFIER_LIST_INITIALIZER(migration_state_notifiers);
  62
  63 static bool deferred_incoming;
  64
  65 /*
  66  * Current state of incoming postcopy; note this is not part of
  67  * MigrationIncomingState since it's state is used during cleanup
  68  * at the end as MIS is being freed.
  69  */
  70 static PostcopyState incoming_postcopy_state;
  71
  72 /* When we add fault tolerance, we could have several
  73    migrations at once.  For now we don't need to add
  74    dynamic creation of migration */
  75
  76 /* For outgoing */
  77 MigrationState *migrate_get_current(void)
  78 {
  79     static bool once;
  80     static MigrationState current_migration = {
  81         .state = MIGRATION_STATUS_NONE,
  82         .bandwidth_limit = MAX_THROTTLE,
  83         .xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE,
  84         .mbps = -1,
  85         .parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL] =
  86                 DEFAULT_MIGRATE_COMPRESS_LEVEL,
  87         .parameters[MIGRATION_PARAMETER_COMPRESS_THREADS] =
  88                 DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT,
  89         .parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS] =
  90                 DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT,
  91         .parameters[MIGRATION_PARAMETER_CPU_THROTTLE_INITIAL] =
  92                 DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL,
  93         .parameters[MIGRATION_PARAMETER_CPU_THROTTLE_INCREMENT] =
  94                 DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT,
  95     };
  96
  97     if (!once) {
  98         qemu_mutex_init(&current_migration.src_page_req_mutex);
  99         once = true;
 100     }
 101     return &current_migration;
 102 }
 103
 104 /* For incoming */
 105 static MigrationIncomingState *mis_current;
 106
 107 MigrationIncomingState *migration_incoming_get_current(void)
 108 {
 109     return mis_current;
 110 }
 111
 112 MigrationIncomingState *migration_incoming_state_new(QEMUFile* f)
 113 {
 114     mis_current = g_new0(MigrationIncomingState, 1);
 115     mis_current->from_src_file = f;
 116     mis_current->state = MIGRATION_STATUS_NONE;
 117     QLIST_INIT(&mis_current->loadvm_handlers);
 118     qemu_mutex_init(&mis_current->rp_mutex);
 119     qemu_event_init(&mis_current->main_thread_load_event, false);
 120
 121     return mis_current;
 122 }
 123
 124 void migration_incoming_state_destroy(void)
 125 {
 126     qemu_event_destroy(&mis_current->main_thread_load_event);
 127     loadvm_free_handlers(mis_current);
 128     g_free(mis_current);
 129     mis_current = NULL;
 130 }
 131
 132
 133 typedef struct {
 134     bool optional;
 135     uint32_t size;
 136     uint8_t runstate[100];
 137     RunState state;
 138     bool received;
 139 } GlobalState;
 140
 141 static GlobalState global_state;
 142
 143 int global_state_store(void)
 144 {
 145     if (!runstate_store((char *)global_state.runstate,
 146                         sizeof(global_state.runstate))) {
 147         error_report("runstate name too big: %s", global_state.runstate);
 148         trace_migrate_state_too_big();
 149         return -EINVAL;
 150     }
 151     return 0;
 152 }
 153
 154 void global_state_store_running(void)
 155 {
 156     const char *state = RunState_lookup[RUN_STATE_RUNNING];
 157     strncpy((char *)global_state.runstate,
 158            state, sizeof(global_state.runstate));
 159 }
 160
 161 static bool global_state_received(void)
 162 {
 163     return global_state.received;
 164 }
 165
 166 static RunState global_state_get_runstate(void)
 167 {
 168     return global_state.state;
 169 }
 170
 171 void global_state_set_optional(void)
 172 {
 173     global_state.optional = true;
 174 }
 175
 176 static bool global_state_needed(void *opaque)
 177 {
 178     GlobalState *s = opaque;
 179     char *runstate = (char *)s->runstate;
 180
 181     /* If it is not optional, it is mandatory */
 182
 183     if (s->optional == false) {
 184         return true;
 185     }
 186
 187     /* If state is running or paused, it is not needed */
 188
 189     if (strcmp(runstate, "running") == 0 ||
 190         strcmp(runstate, "paused") == 0) {
 191         return false;
 192     }
 193
 194     /* for any other state it is needed */
 195     return true;
 196 }
 197
 198 static int global_state_post_load(void *opaque, int version_id)
 199 {
 200     GlobalState *s = opaque;
 201     Error *local_err = NULL;
 202     int r;
 203     char *runstate = (char *)s->runstate;
 204
 205     s->received = true;
 206     trace_migrate_global_state_post_load(runstate);
 207
 208     r = qapi_enum_parse(RunState_lookup, runstate, RUN_STATE__MAX,
 209                                 -1, &local_err);
 210
 211     if (r == -1) {
 212         if (local_err) {
 213             error_report_err(local_err);
 214         }
 215         return -EINVAL;
 216     }
 217     s->state = r;
 218
 219     return 0;
 220 }
 221
 222 static void global_state_pre_save(void *opaque)
 223 {
 224     GlobalState *s = opaque;
 225
 226     trace_migrate_global_state_pre_save((char *)s->runstate);
 227     s->size = strlen((char *)s->runstate) + 1;
 228 }
 229
 230 static const VMStateDescription vmstate_globalstate = {
 231     .name = "globalstate",
 232     .version_id = 1,
 233     .minimum_version_id = 1,
 234     .post_load = global_state_post_load,
 235     .pre_save = global_state_pre_save,
 236     .needed = global_state_needed,
 237     .fields = (VMStateField[]) {
 238         VMSTATE_UINT32(size, GlobalState),
 239         VMSTATE_BUFFER(runstate, GlobalState),
 240         VMSTATE_END_OF_LIST()
 241     },
 242 };
 243
 244 void register_global_state(void)
 245 {
 246     /* We would use it independently that we receive it */
 247     strcpy((char *)&global_state.runstate, "");
 248     global_state.received = false;
 249     vmstate_register(NULL, 0, &vmstate_globalstate, &global_state);
 250 }
 251
 252 static void migrate_generate_event(int new_state)
 253 {
 254     if (migrate_use_events()) {
 255         qapi_event_send_migration(new_state, &error_abort);
 256     }
 257 }
 258
 259 /*
 260  * Called on -incoming with a defer: uri.
 261  * The migration can be started later after any parameters have been
 262  * changed.
 263  */
 264 static void deferred_incoming_migration(Error **errp)
 265 {
 266     if (deferred_incoming) {
 267         error_setg(errp, "Incoming migration already deferred");
 268     }
 269     deferred_incoming = true;
 270 }
 271
 272 /* Request a range of pages from the source VM at the given
 273  * start address.
 274  *   rbname: Name of the RAMBlock to request the page in, if NULL it's the same
 275  *           as the last request (a name must have been given previously)
 276  *   Start: Address offset within the RB
 277  *   Len: Length in bytes required - must be a multiple of pagesize
 278  */
 279 void migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname,
 280                                ram_addr_t start, size_t len)
 281 {
 282     uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
 283     size_t msglen = 12; /* start + len */
 284
 285     *(uint64_t *)bufc = cpu_to_be64((uint64_t)start);
 286     *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len);
 287
 288     if (rbname) {
 289         int rbname_len = strlen(rbname);
 290         assert(rbname_len < 256);
 291
 292         bufc[msglen++] = rbname_len;
 293         memcpy(bufc + msglen, rbname, rbname_len);
 294         msglen += rbname_len;
 295         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES_ID, msglen, bufc);
 296     } else {
 297         migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES, msglen, bufc);
 298     }
 299 }
 300
 301 void qemu_start_incoming_migration(const char *uri, Error **errp)
 302 {
 303     const char *p;
 304
 305     qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort);
 306     if (!strcmp(uri, "defer")) {
 307         deferred_incoming_migration(errp);
 308     } else if (strstart(uri, "tcp:", &p)) {
 309         tcp_start_incoming_migration(p, errp);
 310 #ifdef CONFIG_RDMA
 311     } else if (strstart(uri, "rdma:", &p)) {
 312         rdma_start_incoming_migration(p, errp);
 313 #endif
 314 #if !defined(WIN32)
 315     } else if (strstart(uri, "exec:", &p)) {
 316         exec_start_incoming_migration(p, errp);
 317     } else if (strstart(uri, "unix:", &p)) {
 318         unix_start_incoming_migration(p, errp);
 319     } else if (strstart(uri, "fd:", &p)) {
 320         fd_start_incoming_migration(p, errp);
 321 #endif
 322     } else {
 323         error_setg(errp, "unknown migration protocol: %s", uri);
 324     }
 325 }
 326
 327 static void process_incoming_migration_bh(void *opaque)
 328 {
 329     Error *local_err = NULL;
 330     MigrationIncomingState *mis = opaque;
 331
 332     /* Make sure all file formats flush their mutable metadata */
 333     bdrv_invalidate_cache_all(&local_err);
 334     if (local_err) {
 335         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 336                           MIGRATION_STATUS_FAILED);
 337         error_report_err(local_err);
 338         migrate_decompress_threads_join();
 339         exit(EXIT_FAILURE);
 340     }
 341
 342     /*
 343      * This must happen after all error conditions are dealt with and
 344      * we're sure the VM is going to be running on this host.
 345      */
 346     qemu_announce_self();
 347
 348     /* If global state section was not received or we are in running
 349        state, we need to obey autostart. Any other state is set with
 350        runstate_set. */
 351
 352     if (!global_state_received() ||
 353         global_state_get_runstate() == RUN_STATE_RUNNING) {
 354         if (autostart) {
 355             vm_start();
 356         } else {
 357             runstate_set(RUN_STATE_PAUSED);
 358         }
 359     } else {
 360         runstate_set(global_state_get_runstate());
 361     }
 362     migrate_decompress_threads_join();
 363     /*
 364      * This must happen after any state changes since as soon as an external
 365      * observer sees this event they might start to prod at the VM assuming
 366      * it's ready to use.
 367      */
 368     migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 369                       MIGRATION_STATUS_COMPLETED);
 370     qemu_bh_delete(mis->bh);
 371     migration_incoming_state_destroy();
 372 }
 373
 374 static void process_incoming_migration_co(void *opaque)
 375 {
 376     QEMUFile *f = opaque;
 377     MigrationIncomingState *mis;
 378     PostcopyState ps;
 379     int ret;
 380
 381     mis = migration_incoming_state_new(f);
 382     postcopy_state_set(POSTCOPY_INCOMING_NONE);
 383     migrate_set_state(&mis->state, MIGRATION_STATUS_NONE,
 384                       MIGRATION_STATUS_ACTIVE);
 385     ret = qemu_loadvm_state(f);
 386
 387     ps = postcopy_state_get();
 388     trace_process_incoming_migration_co_end(ret, ps);
 389     if (ps != POSTCOPY_INCOMING_NONE) {
 390         if (ps == POSTCOPY_INCOMING_ADVISE) {
 391             /*
 392              * Where a migration had postcopy enabled (and thus went to advise)
 393              * but managed to complete within the precopy period, we can use
 394              * the normal exit.
 395              */
 396             postcopy_ram_incoming_cleanup(mis);
 397         } else if (ret >= 0) {
 398             /*
 399              * Postcopy was started, cleanup should happen at the end of the
 400              * postcopy thread.
 401              */
 402             trace_process_incoming_migration_co_postcopy_end_main();
 403             return;
 404         }
 405         /* Else if something went wrong then just fall out of the normal exit */
 406     }
 407
 408     qemu_fclose(f);
 409     free_xbzrle_decoded_buf();
 410
 411     if (ret < 0) {
 412         migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 413                           MIGRATION_STATUS_FAILED);
 414         error_report("load of migration failed: %s", strerror(-ret));
 415         migrate_decompress_threads_join();
 416         exit(EXIT_FAILURE);
 417     }
 418
 419     mis->bh = qemu_bh_new(process_incoming_migration_bh, mis);
 420     qemu_bh_schedule(mis->bh);
 421 }
 422
 423 void process_incoming_migration(QEMUFile *f)
 424 {
 425     Coroutine *co = qemu_coroutine_create(process_incoming_migration_co);
 426
 427     migrate_decompress_threads_create();
 428     qemu_file_set_blocking(f, false);
 429     qemu_coroutine_enter(co, f);
 430 }
 431
 432
 433 void migration_set_incoming_channel(MigrationState *s,
 434                                     QIOChannel *ioc)
 435 {
 436     QEMUFile *f = qemu_fopen_channel_input(ioc);
 437
 438     process_incoming_migration(f);
 439 }
 440
 441
 442 void migration_set_outgoing_channel(MigrationState *s,
 443                                     QIOChannel *ioc)
 444 {
 445     QEMUFile *f = qemu_fopen_channel_output(ioc);
 446
 447     s->to_dst_file = f;
 448
 449     migrate_fd_connect(s);
 450 }
 451
 452
 453 /*
 454  * Send a message on the return channel back to the source
 455  * of the migration.
 456  */
 457 void migrate_send_rp_message(MigrationIncomingState *mis,
 458                              enum mig_rp_message_type message_type,
 459                              uint16_t len, void *data)
 460 {
 461     trace_migrate_send_rp_message((int)message_type, len);
 462     qemu_mutex_lock(&mis->rp_mutex);
 463     qemu_put_be16(mis->to_src_file, (unsigned int)message_type);
 464     qemu_put_be16(mis->to_src_file, len);
 465     qemu_put_buffer(mis->to_src_file, data, len);
 466     qemu_fflush(mis->to_src_file);
 467     qemu_mutex_unlock(&mis->rp_mutex);
 468 }
 469
 470 /*
 471  * Send a 'SHUT' message on the return channel with the given value
 472  * to indicate that we've finished with the RP.  Non-0 value indicates
 473  * error.
 474  */
 475 void migrate_send_rp_shut(MigrationIncomingState *mis,
 476                           uint32_t value)
 477 {
 478     uint32_t buf;
 479
 480     buf = cpu_to_be32(value);
 481     migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf);
 482 }
 483
 484 /*
 485  * Send a 'PONG' message on the return channel with the given value
 486  * (normally in response to a 'PING')
 487  */
 488 void migrate_send_rp_pong(MigrationIncomingState *mis,
 489                           uint32_t value)
 490 {
 491     uint32_t buf;
 492
 493     buf = cpu_to_be32(value);
 494     migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf);
 495 }
 496
 497 /* amount of nanoseconds we are willing to wait for migration to be down.
 498  * the choice of nanoseconds is because it is the maximum resolution that
 499  * get_clock() can achieve. It is an internal measure. All user-visible
 500  * units must be in seconds */
 501 static uint64_t max_downtime = 300000000;
 502
 503 uint64_t migrate_max_downtime(void)
 504 {
 505     return max_downtime;
 506 }
 507
 508 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp)
 509 {
 510     MigrationCapabilityStatusList *head = NULL;
 511     MigrationCapabilityStatusList *caps;
 512     MigrationState *s = migrate_get_current();
 513     int i;
 514
 515     caps = NULL; /* silence compiler warning */
 516     for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) {
 517         if (head == NULL) {
 518             head = g_malloc0(sizeof(*caps));
 519             caps = head;
 520         } else {
 521             caps->next = g_malloc0(sizeof(*caps));
 522             caps = caps->next;
 523         }
 524         caps->value =
 525             g_malloc(sizeof(*caps->value));
 526         caps->value->capability = i;
 527         caps->value->state = s->enabled_capabilities[i];
 528     }
 529
 530     return head;
 531 }
 532
 533 MigrationParameters *qmp_query_migrate_parameters(Error **errp)
 534 {
 535     MigrationParameters *params;
 536     MigrationState *s = migrate_get_current();
 537
 538     params = g_malloc0(sizeof(*params));
 539     params->compress_level = s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL];
 540     params->compress_threads =
 541             s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS];
 542     params->decompress_threads =
 543             s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS];
 544     params->cpu_throttle_initial =
 545             s->parameters[MIGRATION_PARAMETER_CPU_THROTTLE_INITIAL];
 546     params->cpu_throttle_increment =
 547             s->parameters[MIGRATION_PARAMETER_CPU_THROTTLE_INCREMENT];
 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         return true;
 563
 564     default:
 565         return false;
 566
 567     }
 568 }
 569
 570 static void get_xbzrle_cache_stats(MigrationInfo *info)
 571 {
 572     if (migrate_use_xbzrle()) {
 573         info->has_xbzrle_cache = true;
 574         info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache));
 575         info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size();
 576         info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred();
 577         info->xbzrle_cache->pages = xbzrle_mig_pages_transferred();
 578         info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss();
 579         info->xbzrle_cache->cache_miss_rate = xbzrle_mig_cache_miss_rate();
 580         info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow();
 581     }
 582 }
 583
 584 MigrationInfo *qmp_query_migrate(Error **errp)
 585 {
 586     MigrationInfo *info = g_malloc0(sizeof(*info));
 587     MigrationState *s = migrate_get_current();
 588
 589     switch (s->state) {
 590     case MIGRATION_STATUS_NONE:
 591         /* no migration has happened ever */
 592         break;
 593     case MIGRATION_STATUS_SETUP:
 594         info->has_status = true;
 595         info->has_total_time = false;
 596         break;
 597     case MIGRATION_STATUS_ACTIVE:
 598     case MIGRATION_STATUS_CANCELLING:
 599         info->has_status = true;
 600         info->has_total_time = true;
 601         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 602             - s->total_time;
 603         info->has_expected_downtime = true;
 604         info->expected_downtime = s->expected_downtime;
 605         info->has_setup_time = true;
 606         info->setup_time = s->setup_time;
 607
 608         info->has_ram = true;
 609         info->ram = g_malloc0(sizeof(*info->ram));
 610         info->ram->transferred = ram_bytes_transferred();
 611         info->ram->remaining = ram_bytes_remaining();
 612         info->ram->total = ram_bytes_total();
 613         info->ram->duplicate = dup_mig_pages_transferred();
 614         info->ram->skipped = skipped_mig_pages_transferred();
 615         info->ram->normal = norm_mig_pages_transferred();
 616         info->ram->normal_bytes = norm_mig_bytes_transferred();
 617         info->ram->dirty_pages_rate = s->dirty_pages_rate;
 618         info->ram->mbps = s->mbps;
 619         info->ram->dirty_sync_count = s->dirty_sync_count;
 620
 621         if (blk_mig_active()) {
 622             info->has_disk = true;
 623             info->disk = g_malloc0(sizeof(*info->disk));
 624             info->disk->transferred = blk_mig_bytes_transferred();
 625             info->disk->remaining = blk_mig_bytes_remaining();
 626             info->disk->total = blk_mig_bytes_total();
 627         }
 628
 629         if (cpu_throttle_active()) {
 630             info->has_cpu_throttle_percentage = true;
 631             info->cpu_throttle_percentage = cpu_throttle_get_percentage();
 632         }
 633
 634         get_xbzrle_cache_stats(info);
 635         break;
 636     case MIGRATION_STATUS_POSTCOPY_ACTIVE:
 637         /* Mostly the same as active; TODO add some postcopy stats */
 638         info->has_status = true;
 639         info->has_total_time = true;
 640         info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 641             - s->total_time;
 642         info->has_expected_downtime = true;
 643         info->expected_downtime = s->expected_downtime;
 644         info->has_setup_time = true;
 645         info->setup_time = s->setup_time;
 646
 647         info->has_ram = true;
 648         info->ram = g_malloc0(sizeof(*info->ram));
 649         info->ram->transferred = ram_bytes_transferred();
 650         info->ram->remaining = ram_bytes_remaining();
 651         info->ram->total = ram_bytes_total();
 652         info->ram->duplicate = dup_mig_pages_transferred();
 653         info->ram->skipped = skipped_mig_pages_transferred();
 654         info->ram->normal = norm_mig_pages_transferred();
 655         info->ram->normal_bytes = norm_mig_bytes_transferred();
 656         info->ram->dirty_pages_rate = s->dirty_pages_rate;
 657         info->ram->mbps = s->mbps;
 658         info->ram->dirty_sync_count = s->dirty_sync_count;
 659
 660         if (blk_mig_active()) {
 661             info->has_disk = true;
 662             info->disk = g_malloc0(sizeof(*info->disk));
 663             info->disk->transferred = blk_mig_bytes_transferred();
 664             info->disk->remaining = blk_mig_bytes_remaining();
 665             info->disk->total = blk_mig_bytes_total();
 666         }
 667
 668         get_xbzrle_cache_stats(info);
 669         break;
 670     case MIGRATION_STATUS_COMPLETED:
 671         get_xbzrle_cache_stats(info);
 672
 673         info->has_status = true;
 674         info->has_total_time = true;
 675         info->total_time = s->total_time;
 676         info->has_downtime = true;
 677         info->downtime = s->downtime;
 678         info->has_setup_time = true;
 679         info->setup_time = s->setup_time;
 680
 681         info->has_ram = true;
 682         info->ram = g_malloc0(sizeof(*info->ram));
 683         info->ram->transferred = ram_bytes_transferred();
 684         info->ram->remaining = 0;
 685         info->ram->total = ram_bytes_total();
 686         info->ram->duplicate = dup_mig_pages_transferred();
 687         info->ram->skipped = skipped_mig_pages_transferred();
 688         info->ram->normal = norm_mig_pages_transferred();
 689         info->ram->normal_bytes = norm_mig_bytes_transferred();
 690         info->ram->mbps = s->mbps;
 691         info->ram->dirty_sync_count = s->dirty_sync_count;
 692         break;
 693     case MIGRATION_STATUS_FAILED:
 694         info->has_status = true;
 695         if (s->error) {
 696             info->has_error_desc = true;
 697             info->error_desc = g_strdup(error_get_pretty(s->error));
 698         }
 699         break;
 700     case MIGRATION_STATUS_CANCELLED:
 701         info->has_status = true;
 702         break;
 703     }
 704     info->status = s->state;
 705
 706     return info;
 707 }
 708
 709 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params,
 710                                   Error **errp)
 711 {
 712     MigrationState *s = migrate_get_current();
 713     MigrationCapabilityStatusList *cap;
 714
 715     if (migration_is_setup_or_active(s->state)) {
 716         error_setg(errp, QERR_MIGRATION_ACTIVE);
 717         return;
 718     }
 719
 720     for (cap = params; cap; cap = cap->next) {
 721         s->enabled_capabilities[cap->value->capability] = cap->value->state;
 722     }
 723
 724     if (migrate_postcopy_ram()) {
 725         if (migrate_use_compression()) {
 726             /* The decompression threads asynchronously write into RAM
 727              * rather than use the atomic copies needed to avoid
 728              * userfaulting.  It should be possible to fix the decompression
 729              * threads for compatibility in future.
 730              */
 731             error_report("Postcopy is not currently compatible with "
 732                          "compression");
 733             s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] =
 734                 false;
 735         }
 736     }
 737 }
 738
 739 void qmp_migrate_set_parameters(bool has_compress_level,
 740                                 int64_t compress_level,
 741                                 bool has_compress_threads,
 742                                 int64_t compress_threads,
 743                                 bool has_decompress_threads,
 744                                 int64_t decompress_threads,
 745                                 bool has_cpu_throttle_initial,
 746                                 int64_t cpu_throttle_initial,
 747                                 bool has_cpu_throttle_increment,
 748                                 int64_t cpu_throttle_increment, Error **errp)
 749 {
 750     MigrationState *s = migrate_get_current();
 751
 752     if (has_compress_level && (compress_level < 0 || compress_level > 9)) {
 753         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level",
 754                    "is invalid, it should be in the range of 0 to 9");
 755         return;
 756     }
 757     if (has_compress_threads &&
 758             (compress_threads < 1 || compress_threads > 255)) {
 759         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 760                    "compress_threads",
 761                    "is invalid, it should be in the range of 1 to 255");
 762         return;
 763     }
 764     if (has_decompress_threads &&
 765             (decompress_threads < 1 || decompress_threads > 255)) {
 766         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 767                    "decompress_threads",
 768                    "is invalid, it should be in the range of 1 to 255");
 769         return;
 770     }
 771     if (has_cpu_throttle_initial &&
 772             (cpu_throttle_initial < 1 || cpu_throttle_initial > 99)) {
 773         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 774                    "cpu_throttle_initial",
 775                    "an integer in the range of 1 to 99");
 776     }
 777     if (has_cpu_throttle_increment &&
 778             (cpu_throttle_increment < 1 || cpu_throttle_increment > 99)) {
 779         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 780                    "cpu_throttle_increment",
 781                    "an integer in the range of 1 to 99");
 782     }
 783
 784     if (has_compress_level) {
 785         s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL] = compress_level;
 786     }
 787     if (has_compress_threads) {
 788         s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS] = compress_threads;
 789     }
 790     if (has_decompress_threads) {
 791         s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS] =
 792                                                     decompress_threads;
 793     }
 794     if (has_cpu_throttle_initial) {
 795         s->parameters[MIGRATION_PARAMETER_CPU_THROTTLE_INITIAL] =
 796                                                     cpu_throttle_initial;
 797     }
 798
 799     if (has_cpu_throttle_increment) {
 800         s->parameters[MIGRATION_PARAMETER_CPU_THROTTLE_INCREMENT] =
 801                                                     cpu_throttle_increment;
 802     }
 803 }
 804
 805 void qmp_migrate_start_postcopy(Error **errp)
 806 {
 807     MigrationState *s = migrate_get_current();
 808
 809     if (!migrate_postcopy_ram()) {
 810         error_setg(errp, "Enable postcopy with migrate_set_capability before"
 811                          " the start of migration");
 812         return;
 813     }
 814
 815     if (s->state == MIGRATION_STATUS_NONE) {
 816         error_setg(errp, "Postcopy must be started after migration has been"
 817                          " started");
 818         return;
 819     }
 820     /*
 821      * we don't error if migration has finished since that would be racy
 822      * with issuing this command.
 823      */
 824     atomic_set(&s->start_postcopy, true);
 825 }
 826
 827 /* shared migration helpers */
 828
 829 void migrate_set_state(int *state, int old_state, int new_state)
 830 {
 831     if (atomic_cmpxchg(state, old_state, new_state) == old_state) {
 832         trace_migrate_set_state(new_state);
 833         migrate_generate_event(new_state);
 834     }
 835 }
 836
 837 static void migrate_fd_cleanup(void *opaque)
 838 {
 839     MigrationState *s = opaque;
 840
 841     qemu_bh_delete(s->cleanup_bh);
 842     s->cleanup_bh = NULL;
 843
 844     flush_page_queue(s);
 845
 846     if (s->to_dst_file) {
 847         trace_migrate_fd_cleanup();
 848         qemu_mutex_unlock_iothread();
 849         if (s->migration_thread_running) {
 850             qemu_thread_join(&s->thread);
 851             s->migration_thread_running = false;
 852         }
 853         qemu_mutex_lock_iothread();
 854
 855         migrate_compress_threads_join();
 856         qemu_fclose(s->to_dst_file);
 857         s->to_dst_file = NULL;
 858     }
 859
 860     assert((s->state != MIGRATION_STATUS_ACTIVE) &&
 861            (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE));
 862
 863     if (s->state == MIGRATION_STATUS_CANCELLING) {
 864         migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING,
 865                           MIGRATION_STATUS_CANCELLED);
 866     }
 867
 868     notifier_list_notify(&migration_state_notifiers, s);
 869 }
 870
 871 void migrate_fd_error(MigrationState *s, const Error *error)
 872 {
 873     trace_migrate_fd_error(error ? error_get_pretty(error) : "");
 874     assert(s->to_dst_file == NULL);
 875     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
 876                       MIGRATION_STATUS_FAILED);
 877     if (!s->error) {
 878         s->error = error_copy(error);
 879     }
 880     notifier_list_notify(&migration_state_notifiers, s);
 881 }
 882
 883 static void migrate_fd_cancel(MigrationState *s)
 884 {
 885     int old_state ;
 886     QEMUFile *f = migrate_get_current()->to_dst_file;
 887     trace_migrate_fd_cancel();
 888
 889     if (s->rp_state.from_dst_file) {
 890         /* shutdown the rp socket, so causing the rp thread to shutdown */
 891         qemu_file_shutdown(s->rp_state.from_dst_file);
 892     }
 893
 894     do {
 895         old_state = s->state;
 896         if (!migration_is_setup_or_active(old_state)) {
 897             break;
 898         }
 899         migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING);
 900     } while (s->state != MIGRATION_STATUS_CANCELLING);
 901
 902     /*
 903      * If we're unlucky the migration code might be stuck somewhere in a
 904      * send/write while the network has failed and is waiting to timeout;
 905      * if we've got shutdown(2) available then we can force it to quit.
 906      * The outgoing qemu file gets closed in migrate_fd_cleanup that is
 907      * called in a bh, so there is no race against this cancel.
 908      */
 909     if (s->state == MIGRATION_STATUS_CANCELLING && f) {
 910         qemu_file_shutdown(f);
 911     }
 912 }
 913
 914 void add_migration_state_change_notifier(Notifier *notify)
 915 {
 916     notifier_list_add(&migration_state_notifiers, notify);
 917 }
 918
 919 void remove_migration_state_change_notifier(Notifier *notify)
 920 {
 921     notifier_remove(notify);
 922 }
 923
 924 bool migration_in_setup(MigrationState *s)
 925 {
 926     return s->state == MIGRATION_STATUS_SETUP;
 927 }
 928
 929 bool migration_has_finished(MigrationState *s)
 930 {
 931     return s->state == MIGRATION_STATUS_COMPLETED;
 932 }
 933
 934 bool migration_has_failed(MigrationState *s)
 935 {
 936     return (s->state == MIGRATION_STATUS_CANCELLED ||
 937             s->state == MIGRATION_STATUS_FAILED);
 938 }
 939
 940 bool migration_in_postcopy(MigrationState *s)
 941 {
 942     return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
 943 }
 944
 945 bool migration_in_postcopy_after_devices(MigrationState *s)
 946 {
 947     return migration_in_postcopy(s) && s->postcopy_after_devices;
 948 }
 949
 950 MigrationState *migrate_init(const MigrationParams *params)
 951 {
 952     MigrationState *s = migrate_get_current();
 953
 954     /*
 955      * Reinitialise all migration state, except
 956      * parameters/capabilities that the user set, and
 957      * locks.
 958      */
 959     s->bytes_xfer = 0;
 960     s->xfer_limit = 0;
 961     s->cleanup_bh = 0;
 962     s->to_dst_file = NULL;
 963     s->state = MIGRATION_STATUS_NONE;
 964     s->params = *params;
 965     s->rp_state.from_dst_file = NULL;
 966     s->rp_state.error = false;
 967     s->mbps = 0.0;
 968     s->downtime = 0;
 969     s->expected_downtime = 0;
 970     s->dirty_pages_rate = 0;
 971     s->dirty_bytes_rate = 0;
 972     s->setup_time = 0;
 973     s->dirty_sync_count = 0;
 974     s->start_postcopy = false;
 975     s->postcopy_after_devices = false;
 976     s->migration_thread_running = false;
 977     s->last_req_rb = NULL;
 978     error_free(s->error);
 979     s->error = NULL;
 980
 981     migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP);
 982
 983     QSIMPLEQ_INIT(&s->src_page_requests);
 984
 985     s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 986     return s;
 987 }
 988
 989 static GSList *migration_blockers;
 990
 991 void migrate_add_blocker(Error *reason)
 992 {
 993     migration_blockers = g_slist_prepend(migration_blockers, reason);
 994 }
 995
 996 void migrate_del_blocker(Error *reason)
 997 {
 998     migration_blockers = g_slist_remove(migration_blockers, reason);
 999 }
1000
1001 void qmp_migrate_incoming(const char *uri, Error **errp)
1002 {
1003     Error *local_err = NULL;
1004     static bool once = true;
1005
1006     if (!deferred_incoming) {
1007         error_setg(errp, "For use with '-incoming defer'");
1008         return;
1009     }
1010     if (!once) {
1011         error_setg(errp, "The incoming migration has already been started");
1012     }
1013
1014     qemu_start_incoming_migration(uri, &local_err);
1015
1016     if (local_err) {
1017         error_propagate(errp, local_err);
1018         return;
1019     }
1020
1021     once = false;
1022 }
1023
1024 bool migration_is_blocked(Error **errp)
1025 {
1026     if (qemu_savevm_state_blocked(errp)) {
1027         return true;
1028     }
1029
1030     if (migration_blockers) {
1031         *errp = error_copy(migration_blockers->data);
1032         return true;
1033     }
1034
1035     return false;
1036 }
1037
1038 void qmp_migrate(const char *uri, bool has_blk, bool blk,
1039                  bool has_inc, bool inc, bool has_detach, bool detach,
1040                  Error **errp)
1041 {
1042     Error *local_err = NULL;
1043     MigrationState *s = migrate_get_current();
1044     MigrationParams params;
1045     const char *p;
1046
1047     params.blk = has_blk && blk;
1048     params.shared = has_inc && inc;
1049
1050     if (migration_is_setup_or_active(s->state) ||
1051         s->state == MIGRATION_STATUS_CANCELLING) {
1052         error_setg(errp, QERR_MIGRATION_ACTIVE);
1053         return;
1054     }
1055     if (runstate_check(RUN_STATE_INMIGRATE)) {
1056         error_setg(errp, "Guest is waiting for an incoming migration");
1057         return;
1058     }
1059
1060     if (migration_is_blocked(errp)) {
1061         return;
1062     }
1063
1064     s = migrate_init(&params);
1065
1066     if (strstart(uri, "tcp:", &p)) {
1067         tcp_start_outgoing_migration(s, p, &local_err);
1068 #ifdef CONFIG_RDMA
1069     } else if (strstart(uri, "rdma:", &p)) {
1070         rdma_start_outgoing_migration(s, p, &local_err);
1071 #endif
1072 #if !defined(WIN32)
1073     } else if (strstart(uri, "exec:", &p)) {
1074         exec_start_outgoing_migration(s, p, &local_err);
1075     } else if (strstart(uri, "unix:", &p)) {
1076         unix_start_outgoing_migration(s, p, &local_err);
1077     } else if (strstart(uri, "fd:", &p)) {
1078         fd_start_outgoing_migration(s, p, &local_err);
1079 #endif
1080     } else {
1081         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri",
1082                    "a valid migration protocol");
1083         migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1084                           MIGRATION_STATUS_FAILED);
1085         return;
1086     }
1087
1088     if (local_err) {
1089         migrate_fd_error(s, local_err);
1090         error_propagate(errp, local_err);
1091         return;
1092     }
1093 }
1094
1095 void qmp_migrate_cancel(Error **errp)
1096 {
1097     migrate_fd_cancel(migrate_get_current());
1098 }
1099
1100 void qmp_migrate_set_cache_size(int64_t value, Error **errp)
1101 {
1102     MigrationState *s = migrate_get_current();
1103     int64_t new_size;
1104
1105     /* Check for truncation */
1106     if (value != (size_t)value) {
1107         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1108                    "exceeding address space");
1109         return;
1110     }
1111
1112     /* Cache should not be larger than guest ram size */
1113     if (value > ram_bytes_total()) {
1114         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1115                    "exceeds guest ram size ");
1116         return;
1117     }
1118
1119     new_size = xbzrle_cache_resize(value);
1120     if (new_size < 0) {
1121         error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size",
1122                    "is smaller than page size");
1123         return;
1124     }
1125
1126     s->xbzrle_cache_size = new_size;
1127 }
1128
1129 int64_t qmp_query_migrate_cache_size(Error **errp)
1130 {
1131     return migrate_xbzrle_cache_size();
1132 }
1133
1134 void qmp_migrate_set_speed(int64_t value, Error **errp)
1135 {
1136     MigrationState *s;
1137
1138     if (value < 0) {
1139         value = 0;
1140     }
1141     if (value > SIZE_MAX) {
1142         value = SIZE_MAX;
1143     }
1144
1145     s = migrate_get_current();
1146     s->bandwidth_limit = value;
1147     if (s->to_dst_file) {
1148         qemu_file_set_rate_limit(s->to_dst_file,
1149                                  s->bandwidth_limit / XFER_LIMIT_RATIO);
1150     }
1151 }
1152
1153 void qmp_migrate_set_downtime(double value, Error **errp)
1154 {
1155     value *= 1e9;
1156     value = MAX(0, MIN(UINT64_MAX, value));
1157     max_downtime = (uint64_t)value;
1158 }
1159
1160 bool migrate_postcopy_ram(void)
1161 {
1162     MigrationState *s;
1163
1164     s = migrate_get_current();
1165
1166     return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM];
1167 }
1168
1169 bool migrate_auto_converge(void)
1170 {
1171     MigrationState *s;
1172
1173     s = migrate_get_current();
1174
1175     return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE];
1176 }
1177
1178 bool migrate_zero_blocks(void)
1179 {
1180     MigrationState *s;
1181
1182     s = migrate_get_current();
1183
1184     return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS];
1185 }
1186
1187 bool migrate_use_compression(void)
1188 {
1189     MigrationState *s;
1190
1191     s = migrate_get_current();
1192
1193     return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS];
1194 }
1195
1196 int migrate_compress_level(void)
1197 {
1198     MigrationState *s;
1199
1200     s = migrate_get_current();
1201
1202     return s->parameters[MIGRATION_PARAMETER_COMPRESS_LEVEL];
1203 }
1204
1205 int migrate_compress_threads(void)
1206 {
1207     MigrationState *s;
1208
1209     s = migrate_get_current();
1210
1211     return s->parameters[MIGRATION_PARAMETER_COMPRESS_THREADS];
1212 }
1213
1214 int migrate_decompress_threads(void)
1215 {
1216     MigrationState *s;
1217
1218     s = migrate_get_current();
1219
1220     return s->parameters[MIGRATION_PARAMETER_DECOMPRESS_THREADS];
1221 }
1222
1223 bool migrate_use_events(void)
1224 {
1225     MigrationState *s;
1226
1227     s = migrate_get_current();
1228
1229     return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS];
1230 }
1231
1232 int migrate_use_xbzrle(void)
1233 {
1234     MigrationState *s;
1235
1236     s = migrate_get_current();
1237
1238     return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE];
1239 }
1240
1241 int64_t migrate_xbzrle_cache_size(void)
1242 {
1243     MigrationState *s;
1244
1245     s = migrate_get_current();
1246
1247     return s->xbzrle_cache_size;
1248 }
1249
1250 /* migration thread support */
1251 /*
1252  * Something bad happened to the RP stream, mark an error
1253  * The caller shall print or trace something to indicate why
1254  */
1255 static void mark_source_rp_bad(MigrationState *s)
1256 {
1257     s->rp_state.error = true;
1258 }
1259
1260 static struct rp_cmd_args {
1261     ssize_t     len; /* -1 = variable */
1262     const char *name;
1263 } rp_cmd_args[] = {
1264     [MIG_RP_MSG_INVALID]        = { .len = -1, .name = "INVALID" },
1265     [MIG_RP_MSG_SHUT]           = { .len =  4, .name = "SHUT" },
1266     [MIG_RP_MSG_PONG]           = { .len =  4, .name = "PONG" },
1267     [MIG_RP_MSG_REQ_PAGES]      = { .len = 12, .name = "REQ_PAGES" },
1268     [MIG_RP_MSG_REQ_PAGES_ID]   = { .len = -1, .name = "REQ_PAGES_ID" },
1269     [MIG_RP_MSG_MAX]            = { .len = -1, .name = "MAX" },
1270 };
1271
1272 /*
1273  * Process a request for pages received on the return path,
1274  * We're allowed to send more than requested (e.g. to round to our page size)
1275  * and we don't need to send pages that have already been sent.
1276  */
1277 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname,
1278                                        ram_addr_t start, size_t len)
1279 {
1280     long our_host_ps = getpagesize();
1281
1282     trace_migrate_handle_rp_req_pages(rbname, start, len);
1283
1284     /*
1285      * Since we currently insist on matching page sizes, just sanity check
1286      * we're being asked for whole host pages.
1287      */
1288     if (start & (our_host_ps-1) ||
1289        (len & (our_host_ps-1))) {
1290         error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1291                      " len: %zd", __func__, start, len);
1292         mark_source_rp_bad(ms);
1293         return;
1294     }
1295
1296     if (ram_save_queue_pages(ms, rbname, start, len)) {
1297         mark_source_rp_bad(ms);
1298     }
1299 }
1300
1301 /*
1302  * Handles messages sent on the return path towards the source VM
1303  *
1304  */
1305 static void *source_return_path_thread(void *opaque)
1306 {
1307     MigrationState *ms = opaque;
1308     QEMUFile *rp = ms->rp_state.from_dst_file;
1309     uint16_t header_len, header_type;
1310     uint8_t buf[512];
1311     uint32_t tmp32, sibling_error;
1312     ram_addr_t start = 0; /* =0 to silence warning */
1313     size_t  len = 0, expected_len;
1314     int res;
1315
1316     trace_source_return_path_thread_entry();
1317     while (!ms->rp_state.error && !qemu_file_get_error(rp) &&
1318            migration_is_setup_or_active(ms->state)) {
1319         trace_source_return_path_thread_loop_top();
1320         header_type = qemu_get_be16(rp);
1321         header_len = qemu_get_be16(rp);
1322
1323         if (header_type >= MIG_RP_MSG_MAX ||
1324             header_type == MIG_RP_MSG_INVALID) {
1325             error_report("RP: Received invalid message 0x%04x length 0x%04x",
1326                     header_type, header_len);
1327             mark_source_rp_bad(ms);
1328             goto out;
1329         }
1330
1331         if ((rp_cmd_args[header_type].len != -1 &&
1332             header_len != rp_cmd_args[header_type].len) ||
1333             header_len > sizeof(buf)) {
1334             error_report("RP: Received '%s' message (0x%04x) with"
1335                     "incorrect length %d expecting %zu",
1336                     rp_cmd_args[header_type].name, header_type, header_len,
1337                     (size_t)rp_cmd_args[header_type].len);
1338             mark_source_rp_bad(ms);
1339             goto out;
1340         }
1341
1342         /* We know we've got a valid header by this point */
1343         res = qemu_get_buffer(rp, buf, header_len);
1344         if (res != header_len) {
1345             error_report("RP: Failed reading data for message 0x%04x"
1346                          " read %d expected %d",
1347                          header_type, res, header_len);
1348             mark_source_rp_bad(ms);
1349             goto out;
1350         }
1351
1352         /* OK, we have the message and the data */
1353         switch (header_type) {
1354         case MIG_RP_MSG_SHUT:
1355             sibling_error = be32_to_cpup((uint32_t *)buf);
1356             trace_source_return_path_thread_shut(sibling_error);
1357             if (sibling_error) {
1358                 error_report("RP: Sibling indicated error %d", sibling_error);
1359                 mark_source_rp_bad(ms);
1360             }
1361             /*
1362              * We'll let the main thread deal with closing the RP
1363              * we could do a shutdown(2) on it, but we're the only user
1364              * anyway, so there's nothing gained.
1365              */
1366             goto out;
1367
1368         case MIG_RP_MSG_PONG:
1369             tmp32 = be32_to_cpup((uint32_t *)buf);
1370             trace_source_return_path_thread_pong(tmp32);
1371             break;
1372
1373         case MIG_RP_MSG_REQ_PAGES:
1374             start = be64_to_cpup((uint64_t *)buf);
1375             len = be32_to_cpup((uint32_t *)(buf + 8));
1376             migrate_handle_rp_req_pages(ms, NULL, start, len);
1377             break;
1378
1379         case MIG_RP_MSG_REQ_PAGES_ID:
1380             expected_len = 12 + 1; /* header + termination */
1381
1382             if (header_len >= expected_len) {
1383                 start = be64_to_cpup((uint64_t *)buf);
1384                 len = be32_to_cpup((uint32_t *)(buf + 8));
1385                 /* Now we expect an idstr */
1386                 tmp32 = buf[12]; /* Length of the following idstr */
1387                 buf[13 + tmp32] = '\0';
1388                 expected_len += tmp32;
1389             }
1390             if (header_len != expected_len) {
1391                 error_report("RP: Req_Page_id with length %d expecting %zd",
1392                         header_len, expected_len);
1393                 mark_source_rp_bad(ms);
1394                 goto out;
1395             }
1396             migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len);
1397             break;
1398
1399         default:
1400             break;
1401         }
1402     }
1403     if (qemu_file_get_error(rp)) {
1404         trace_source_return_path_thread_bad_end();
1405         mark_source_rp_bad(ms);
1406     }
1407
1408     trace_source_return_path_thread_end();
1409 out:
1410     ms->rp_state.from_dst_file = NULL;
1411     qemu_fclose(rp);
1412     return NULL;
1413 }
1414
1415 static int open_return_path_on_source(MigrationState *ms)
1416 {
1417
1418     ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file);
1419     if (!ms->rp_state.from_dst_file) {
1420         return -1;
1421     }
1422
1423     trace_open_return_path_on_source();
1424     qemu_thread_create(&ms->rp_state.rp_thread, "return path",
1425                        source_return_path_thread, ms, QEMU_THREAD_JOINABLE);
1426
1427     trace_open_return_path_on_source_continue();
1428
1429     return 0;
1430 }
1431
1432 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
1433 static int await_return_path_close_on_source(MigrationState *ms)
1434 {
1435     /*
1436      * If this is a normal exit then the destination will send a SHUT and the
1437      * rp_thread will exit, however if there's an error we need to cause
1438      * it to exit.
1439      */
1440     if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
1441         /*
1442          * shutdown(2), if we have it, will cause it to unblock if it's stuck
1443          * waiting for the destination.
1444          */
1445         qemu_file_shutdown(ms->rp_state.from_dst_file);
1446         mark_source_rp_bad(ms);
1447     }
1448     trace_await_return_path_close_on_source_joining();
1449     qemu_thread_join(&ms->rp_state.rp_thread);
1450     trace_await_return_path_close_on_source_close();
1451     return ms->rp_state.error;
1452 }
1453
1454 /*
1455  * Switch from normal iteration to postcopy
1456  * Returns non-0 on error
1457  */
1458 static int postcopy_start(MigrationState *ms, bool *old_vm_running)
1459 {
1460     int ret;
1461     QIOChannelBuffer *bioc;
1462     QEMUFile *fb;
1463     int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1464     migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE,
1465                       MIGRATION_STATUS_POSTCOPY_ACTIVE);
1466
1467     trace_postcopy_start();
1468     qemu_mutex_lock_iothread();
1469     trace_postcopy_start_set_run();
1470
1471     qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1472     *old_vm_running = runstate_is_running();
1473     global_state_store();
1474     ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1475     if (ret < 0) {
1476         goto fail;
1477     }
1478
1479     ret = bdrv_inactivate_all();
1480     if (ret < 0) {
1481         goto fail;
1482     }
1483
1484     /*
1485      * Cause any non-postcopiable, but iterative devices to
1486      * send out their final data.
1487      */
1488     qemu_savevm_state_complete_precopy(ms->to_dst_file, true);
1489
1490     /*
1491      * in Finish migrate and with the io-lock held everything should
1492      * be quiet, but we've potentially still got dirty pages and we
1493      * need to tell the destination to throw any pages it's already received
1494      * that are dirty
1495      */
1496     if (ram_postcopy_send_discard_bitmap(ms)) {
1497         error_report("postcopy send discard bitmap failed");
1498         goto fail;
1499     }
1500
1501     /*
1502      * send rest of state - note things that are doing postcopy
1503      * will notice we're in POSTCOPY_ACTIVE and not actually
1504      * wrap their state up here
1505      */
1506     qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX);
1507     /* Ping just for debugging, helps line traces up */
1508     qemu_savevm_send_ping(ms->to_dst_file, 2);
1509
1510     /*
1511      * While loading the device state we may trigger page transfer
1512      * requests and the fd must be free to process those, and thus
1513      * the destination must read the whole device state off the fd before
1514      * it starts processing it.  Unfortunately the ad-hoc migration format
1515      * doesn't allow the destination to know the size to read without fully
1516      * parsing it through each devices load-state code (especially the open
1517      * coded devices that use get/put).
1518      * So we wrap the device state up in a package with a length at the start;
1519      * to do this we use a qemu_buf to hold the whole of the device state.
1520      */
1521     bioc = qio_channel_buffer_new(4096);
1522     fb = qemu_fopen_channel_output(QIO_CHANNEL(bioc));
1523     object_unref(OBJECT(bioc));
1524
1525     /*
1526      * Make sure the receiver can get incoming pages before we send the rest
1527      * of the state
1528      */
1529     qemu_savevm_send_postcopy_listen(fb);
1530
1531     qemu_savevm_state_complete_precopy(fb, false);
1532     qemu_savevm_send_ping(fb, 3);
1533
1534     qemu_savevm_send_postcopy_run(fb);
1535
1536     /* <><> end of stuff going into the package */
1537
1538     /* Now send that blob */
1539     if (qemu_savevm_send_packaged(ms->to_dst_file, bioc->data, bioc->usage)) {
1540         goto fail_closefb;
1541     }
1542     qemu_fclose(fb);
1543
1544     /* Send a notify to give a chance for anything that needs to happen
1545      * at the transition to postcopy and after the device state; in particular
1546      * spice needs to trigger a transition now
1547      */
1548     ms->postcopy_after_devices = true;
1549     notifier_list_notify(&migration_state_notifiers, ms);
1550
1551     ms->downtime =  qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop;
1552
1553     qemu_mutex_unlock_iothread();
1554
1555     /*
1556      * Although this ping is just for debug, it could potentially be
1557      * used for getting a better measurement of downtime at the source.
1558      */
1559     qemu_savevm_send_ping(ms->to_dst_file, 4);
1560
1561     ret = qemu_file_get_error(ms->to_dst_file);
1562     if (ret) {
1563         error_report("postcopy_start: Migration stream errored");
1564         migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1565                               MIGRATION_STATUS_FAILED);
1566     }
1567
1568     return ret;
1569
1570 fail_closefb:
1571     qemu_fclose(fb);
1572 fail:
1573     migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE,
1574                           MIGRATION_STATUS_FAILED);
1575     qemu_mutex_unlock_iothread();
1576     return -1;
1577 }
1578
1579 /**
1580  * migration_completion: Used by migration_thread when there's not much left.
1581  *   The caller 'breaks' the loop when this returns.
1582  *
1583  * @s: Current migration state
1584  * @current_active_state: The migration state we expect to be in
1585  * @*old_vm_running: Pointer to old_vm_running flag
1586  * @*start_time: Pointer to time to update
1587  */
1588 static void migration_completion(MigrationState *s, int current_active_state,
1589                                  bool *old_vm_running,
1590                                  int64_t *start_time)
1591 {
1592     int ret;
1593
1594     if (s->state == MIGRATION_STATUS_ACTIVE) {
1595         qemu_mutex_lock_iothread();
1596         *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1597         qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
1598         *old_vm_running = runstate_is_running();
1599         ret = global_state_store();
1600
1601         if (!ret) {
1602             ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
1603             if (ret >= 0) {
1604                 ret = bdrv_inactivate_all();
1605             }
1606             if (ret >= 0) {
1607                 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX);
1608                 qemu_savevm_state_complete_precopy(s->to_dst_file, false);
1609             }
1610         }
1611         qemu_mutex_unlock_iothread();
1612
1613         if (ret < 0) {
1614             goto fail;
1615         }
1616     } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1617         trace_migration_completion_postcopy_end();
1618
1619         qemu_savevm_state_complete_postcopy(s->to_dst_file);
1620         trace_migration_completion_postcopy_end_after_complete();
1621     }
1622
1623     /*
1624      * If rp was opened we must clean up the thread before
1625      * cleaning everything else up (since if there are no failures
1626      * it will wait for the destination to send it's status in
1627      * a SHUT command).
1628      * Postcopy opens rp if enabled (even if it's not avtivated)
1629      */
1630     if (migrate_postcopy_ram()) {
1631         int rp_error;
1632         trace_migration_completion_postcopy_end_before_rp();
1633         rp_error = await_return_path_close_on_source(s);
1634         trace_migration_completion_postcopy_end_after_rp(rp_error);
1635         if (rp_error) {
1636             goto fail_invalidate;
1637         }
1638     }
1639
1640     if (qemu_file_get_error(s->to_dst_file)) {
1641         trace_migration_completion_file_err();
1642         goto fail_invalidate;
1643     }
1644
1645     migrate_set_state(&s->state, current_active_state,
1646                       MIGRATION_STATUS_COMPLETED);
1647     return;
1648
1649 fail_invalidate:
1650     /* If not doing postcopy, vm_start() will be called: let's regain
1651      * control on images.
1652      */
1653     if (s->state == MIGRATION_STATUS_ACTIVE) {
1654         Error *local_err = NULL;
1655
1656         bdrv_invalidate_cache_all(&local_err);
1657         if (local_err) {
1658             error_report_err(local_err);
1659         }
1660     }
1661
1662 fail:
1663     migrate_set_state(&s->state, current_active_state,
1664                       MIGRATION_STATUS_FAILED);
1665 }
1666
1667 /*
1668  * Master migration thread on the source VM.
1669  * It drives the migration and pumps the data down the outgoing channel.
1670  */
1671 static void *migration_thread(void *opaque)
1672 {
1673     MigrationState *s = opaque;
1674     /* Used by the bandwidth calcs, updated later */
1675     int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1676     int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST);
1677     int64_t initial_bytes = 0;
1678     int64_t max_size = 0;
1679     int64_t start_time = initial_time;
1680     int64_t end_time;
1681     bool old_vm_running = false;
1682     bool entered_postcopy = false;
1683     /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */
1684     enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE;
1685
1686     rcu_register_thread();
1687
1688     qemu_savevm_state_header(s->to_dst_file);
1689
1690     if (migrate_postcopy_ram()) {
1691         /* Now tell the dest that it should open its end so it can reply */
1692         qemu_savevm_send_open_return_path(s->to_dst_file);
1693
1694         /* And do a ping that will make stuff easier to debug */
1695         qemu_savevm_send_ping(s->to_dst_file, 1);
1696
1697         /*
1698          * Tell the destination that we *might* want to do postcopy later;
1699          * if the other end can't do postcopy it should fail now, nice and
1700          * early.
1701          */
1702         qemu_savevm_send_postcopy_advise(s->to_dst_file);
1703     }
1704
1705     qemu_savevm_state_begin(s->to_dst_file, &s->params);
1706
1707     s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start;
1708     current_active_state = MIGRATION_STATUS_ACTIVE;
1709     migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1710                       MIGRATION_STATUS_ACTIVE);
1711
1712     trace_migration_thread_setup_complete();
1713
1714     while (s->state == MIGRATION_STATUS_ACTIVE ||
1715            s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
1716         int64_t current_time;
1717         uint64_t pending_size;
1718
1719         if (!qemu_file_rate_limit(s->to_dst_file)) {
1720             uint64_t pend_post, pend_nonpost;
1721
1722             qemu_savevm_state_pending(s->to_dst_file, max_size, &pend_nonpost,
1723                                       &pend_post);
1724             pending_size = pend_nonpost + pend_post;
1725             trace_migrate_pending(pending_size, max_size,
1726                                   pend_post, pend_nonpost);
1727             if (pending_size && pending_size >= max_size) {
1728                 /* Still a significant amount to transfer */
1729
1730                 if (migrate_postcopy_ram() &&
1731                     s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE &&
1732                     pend_nonpost <= max_size &&
1733                     atomic_read(&s->start_postcopy)) {
1734
1735                     if (!postcopy_start(s, &old_vm_running)) {
1736                         current_active_state = MIGRATION_STATUS_POSTCOPY_ACTIVE;
1737                         entered_postcopy = true;
1738                     }
1739
1740                     continue;
1741                 }
1742                 /* Just another iteration step */
1743                 qemu_savevm_state_iterate(s->to_dst_file, entered_postcopy);
1744             } else {
1745                 trace_migration_thread_low_pending(pending_size);
1746                 migration_completion(s, current_active_state,
1747                                      &old_vm_running, &start_time);
1748                 break;
1749             }
1750         }
1751
1752         if (qemu_file_get_error(s->to_dst_file)) {
1753             migrate_set_state(&s->state, current_active_state,
1754                               MIGRATION_STATUS_FAILED);
1755             trace_migration_thread_file_err();
1756             break;
1757         }
1758         current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1759         if (current_time >= initial_time + BUFFER_DELAY) {
1760             uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) -
1761                                          initial_bytes;
1762             uint64_t time_spent = current_time - initial_time;
1763             double bandwidth = (double)transferred_bytes / time_spent;
1764             max_size = bandwidth * migrate_max_downtime() / 1000000;
1765
1766             s->mbps = (((double) transferred_bytes * 8.0) /
1767                     ((double) time_spent / 1000.0)) / 1000.0 / 1000.0;
1768
1769             trace_migrate_transferred(transferred_bytes, time_spent,
1770                                       bandwidth, max_size);
1771             /* if we haven't sent anything, we don't want to recalculate
1772                10000 is a small enough number for our purposes */
1773             if (s->dirty_bytes_rate && transferred_bytes > 10000) {
1774                 s->expected_downtime = s->dirty_bytes_rate / bandwidth;
1775             }
1776
1777             qemu_file_reset_rate_limit(s->to_dst_file);
1778             initial_time = current_time;
1779             initial_bytes = qemu_ftell(s->to_dst_file);
1780         }
1781         if (qemu_file_rate_limit(s->to_dst_file)) {
1782             /* usleep expects microseconds */
1783             g_usleep((initial_time + BUFFER_DELAY - current_time)*1000);
1784         }
1785     }
1786
1787     trace_migration_thread_after_loop();
1788     /* If we enabled cpu throttling for auto-converge, turn it off. */
1789     cpu_throttle_stop();
1790     end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
1791
1792     qemu_mutex_lock_iothread();
1793     qemu_savevm_state_cleanup();
1794     if (s->state == MIGRATION_STATUS_COMPLETED) {
1795         uint64_t transferred_bytes = qemu_ftell(s->to_dst_file);
1796         s->total_time = end_time - s->total_time;
1797         if (!entered_postcopy) {
1798             s->downtime = end_time - start_time;
1799         }
1800         if (s->total_time) {
1801             s->mbps = (((double) transferred_bytes * 8.0) /
1802                        ((double) s->total_time)) / 1000;
1803         }
1804         runstate_set(RUN_STATE_POSTMIGRATE);
1805     } else {
1806         if (old_vm_running && !entered_postcopy) {
1807             vm_start();
1808         }
1809     }
1810     qemu_bh_schedule(s->cleanup_bh);
1811     qemu_mutex_unlock_iothread();
1812
1813     rcu_unregister_thread();
1814     return NULL;
1815 }
1816
1817 void migrate_fd_connect(MigrationState *s)
1818 {
1819     /* This is a best 1st approximation. ns to ms */
1820     s->expected_downtime = max_downtime/1000000;
1821     s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s);
1822
1823     qemu_file_set_blocking(s->to_dst_file, true);
1824     qemu_file_set_rate_limit(s->to_dst_file,
1825                              s->bandwidth_limit / XFER_LIMIT_RATIO);
1826
1827     /* Notify before starting migration thread */
1828     notifier_list_notify(&migration_state_notifiers, s);
1829
1830     /*
1831      * Open the return path; currently for postcopy but other things might
1832      * also want it.
1833      */
1834     if (migrate_postcopy_ram()) {
1835         if (open_return_path_on_source(s)) {
1836             error_report("Unable to open return-path for postcopy");
1837             migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
1838                               MIGRATION_STATUS_FAILED);
1839             migrate_fd_cleanup(s);
1840             return;
1841         }
1842     }
1843
1844     migrate_compress_threads_create();
1845     qemu_thread_create(&s->thread, "migration", migration_thread, s,
1846                        QEMU_THREAD_JOINABLE);
1847     s->migration_thread_running = true;
1848 }
1849
1850 PostcopyState  postcopy_state_get(void)
1851 {
1852     return atomic_mb_read(&incoming_postcopy_state);
1853 }
1854
1855 /* Set the state and return the old state */
1856 PostcopyState postcopy_state_set(PostcopyState new_state)
1857 {
1858     return atomic_xchg(&incoming_postcopy_state, new_state);
1859 }
1860