1 /* 2 * QEMU live migration 3 * 4 * Copyright IBM, Corp. 2008 5 * 6 * Authors: 7 * Anthony Liguori <aliguori@us.ibm.com> 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 #include "io/channel-tls.h" 39 #include "migration/colo.h" 40 41 #define MAX_THROTTLE (32 << 20) /* Migration transfer speed throttling */ 42 43 /* Amount of time to allocate to each "chunk" of bandwidth-throttled 44 * data. */ 45 #define BUFFER_DELAY 100 46 #define XFER_LIMIT_RATIO (1000 / BUFFER_DELAY) 47 48 /* Time in milliseconds we are allowed to stop the source, 49 * for sending the last part */ 50 #define DEFAULT_MIGRATE_SET_DOWNTIME 300 51 52 /* Default compression thread count */ 53 #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8 54 /* Default decompression thread count, usually decompression is at 55 * least 4 times as fast as compression.*/ 56 #define DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT 2 57 /*0: means nocompress, 1: best speed, ... 9: best compress ratio */ 58 #define DEFAULT_MIGRATE_COMPRESS_LEVEL 1 59 /* Define default autoconverge cpu throttle migration parameters */ 60 #define DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL 20 61 #define DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT 10 62 63 /* Migration XBZRLE default cache size */ 64 #define DEFAULT_MIGRATE_CACHE_SIZE (64 * 1024 * 1024) 65 66 /* The delay time (in ms) between two COLO checkpoints 67 * Note: Please change this default value to 10000 when we support hybrid mode. 68 */ 69 #define DEFAULT_MIGRATE_X_CHECKPOINT_DELAY 200 70 71 static NotifierList migration_state_notifiers = 72 NOTIFIER_LIST_INITIALIZER(migration_state_notifiers); 73 74 static bool deferred_incoming; 75 76 /* 77 * Current state of incoming postcopy; note this is not part of 78 * MigrationIncomingState since it's state is used during cleanup 79 * at the end as MIS is being freed. 80 */ 81 static PostcopyState incoming_postcopy_state; 82 83 /* When we add fault tolerance, we could have several 84 migrations at once. For now we don't need to add 85 dynamic creation of migration */ 86 87 /* For outgoing */ 88 MigrationState *migrate_get_current(void) 89 { 90 static bool once; 91 static MigrationState current_migration = { 92 .state = MIGRATION_STATUS_NONE, 93 .xbzrle_cache_size = DEFAULT_MIGRATE_CACHE_SIZE, 94 .mbps = -1, 95 .parameters = { 96 .compress_level = DEFAULT_MIGRATE_COMPRESS_LEVEL, 97 .compress_threads = DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT, 98 .decompress_threads = DEFAULT_MIGRATE_DECOMPRESS_THREAD_COUNT, 99 .cpu_throttle_initial = DEFAULT_MIGRATE_CPU_THROTTLE_INITIAL, 100 .cpu_throttle_increment = DEFAULT_MIGRATE_CPU_THROTTLE_INCREMENT, 101 .max_bandwidth = MAX_THROTTLE, 102 .downtime_limit = DEFAULT_MIGRATE_SET_DOWNTIME, 103 .x_checkpoint_delay = DEFAULT_MIGRATE_X_CHECKPOINT_DELAY, 104 }, 105 }; 106 107 if (!once) { 108 qemu_mutex_init(¤t_migration.src_page_req_mutex); 109 once = true; 110 } 111 return ¤t_migration; 112 } 113 114 /* For incoming */ 115 static MigrationIncomingState *mis_current; 116 117 MigrationIncomingState *migration_incoming_get_current(void) 118 { 119 return mis_current; 120 } 121 122 MigrationIncomingState *migration_incoming_state_new(QEMUFile* f) 123 { 124 mis_current = g_new0(MigrationIncomingState, 1); 125 mis_current->from_src_file = f; 126 mis_current->state = MIGRATION_STATUS_NONE; 127 QLIST_INIT(&mis_current->loadvm_handlers); 128 qemu_mutex_init(&mis_current->rp_mutex); 129 qemu_event_init(&mis_current->main_thread_load_event, false); 130 131 return mis_current; 132 } 133 134 void migration_incoming_state_destroy(void) 135 { 136 qemu_event_destroy(&mis_current->main_thread_load_event); 137 loadvm_free_handlers(mis_current); 138 g_free(mis_current); 139 mis_current = NULL; 140 } 141 142 143 typedef struct { 144 bool optional; 145 uint32_t size; 146 uint8_t runstate[100]; 147 RunState state; 148 bool received; 149 } GlobalState; 150 151 static GlobalState global_state; 152 153 int global_state_store(void) 154 { 155 if (!runstate_store((char *)global_state.runstate, 156 sizeof(global_state.runstate))) { 157 error_report("runstate name too big: %s", global_state.runstate); 158 trace_migrate_state_too_big(); 159 return -EINVAL; 160 } 161 return 0; 162 } 163 164 void global_state_store_running(void) 165 { 166 const char *state = RunState_lookup[RUN_STATE_RUNNING]; 167 strncpy((char *)global_state.runstate, 168 state, sizeof(global_state.runstate)); 169 } 170 171 static bool global_state_received(void) 172 { 173 return global_state.received; 174 } 175 176 static RunState global_state_get_runstate(void) 177 { 178 return global_state.state; 179 } 180 181 void global_state_set_optional(void) 182 { 183 global_state.optional = true; 184 } 185 186 static bool global_state_needed(void *opaque) 187 { 188 GlobalState *s = opaque; 189 char *runstate = (char *)s->runstate; 190 191 /* If it is not optional, it is mandatory */ 192 193 if (s->optional == false) { 194 return true; 195 } 196 197 /* If state is running or paused, it is not needed */ 198 199 if (strcmp(runstate, "running") == 0 || 200 strcmp(runstate, "paused") == 0) { 201 return false; 202 } 203 204 /* for any other state it is needed */ 205 return true; 206 } 207 208 static int global_state_post_load(void *opaque, int version_id) 209 { 210 GlobalState *s = opaque; 211 Error *local_err = NULL; 212 int r; 213 char *runstate = (char *)s->runstate; 214 215 s->received = true; 216 trace_migrate_global_state_post_load(runstate); 217 218 r = qapi_enum_parse(RunState_lookup, runstate, RUN_STATE__MAX, 219 -1, &local_err); 220 221 if (r == -1) { 222 if (local_err) { 223 error_report_err(local_err); 224 } 225 return -EINVAL; 226 } 227 s->state = r; 228 229 return 0; 230 } 231 232 static void global_state_pre_save(void *opaque) 233 { 234 GlobalState *s = opaque; 235 236 trace_migrate_global_state_pre_save((char *)s->runstate); 237 s->size = strlen((char *)s->runstate) + 1; 238 } 239 240 static const VMStateDescription vmstate_globalstate = { 241 .name = "globalstate", 242 .version_id = 1, 243 .minimum_version_id = 1, 244 .post_load = global_state_post_load, 245 .pre_save = global_state_pre_save, 246 .needed = global_state_needed, 247 .fields = (VMStateField[]) { 248 VMSTATE_UINT32(size, GlobalState), 249 VMSTATE_BUFFER(runstate, GlobalState), 250 VMSTATE_END_OF_LIST() 251 }, 252 }; 253 254 void register_global_state(void) 255 { 256 /* We would use it independently that we receive it */ 257 strcpy((char *)&global_state.runstate, ""); 258 global_state.received = false; 259 vmstate_register(NULL, 0, &vmstate_globalstate, &global_state); 260 } 261 262 static void migrate_generate_event(int new_state) 263 { 264 if (migrate_use_events()) { 265 qapi_event_send_migration(new_state, &error_abort); 266 } 267 } 268 269 /* 270 * Called on -incoming with a defer: uri. 271 * The migration can be started later after any parameters have been 272 * changed. 273 */ 274 static void deferred_incoming_migration(Error **errp) 275 { 276 if (deferred_incoming) { 277 error_setg(errp, "Incoming migration already deferred"); 278 } 279 deferred_incoming = true; 280 } 281 282 /* Request a range of pages from the source VM at the given 283 * start address. 284 * rbname: Name of the RAMBlock to request the page in, if NULL it's the same 285 * as the last request (a name must have been given previously) 286 * Start: Address offset within the RB 287 * Len: Length in bytes required - must be a multiple of pagesize 288 */ 289 void migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname, 290 ram_addr_t start, size_t len) 291 { 292 uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */ 293 size_t msglen = 12; /* start + len */ 294 295 *(uint64_t *)bufc = cpu_to_be64((uint64_t)start); 296 *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len); 297 298 if (rbname) { 299 int rbname_len = strlen(rbname); 300 assert(rbname_len < 256); 301 302 bufc[msglen++] = rbname_len; 303 memcpy(bufc + msglen, rbname, rbname_len); 304 msglen += rbname_len; 305 migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES_ID, msglen, bufc); 306 } else { 307 migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES, msglen, bufc); 308 } 309 } 310 311 void qemu_start_incoming_migration(const char *uri, Error **errp) 312 { 313 const char *p; 314 315 qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort); 316 if (!strcmp(uri, "defer")) { 317 deferred_incoming_migration(errp); 318 } else if (strstart(uri, "tcp:", &p)) { 319 tcp_start_incoming_migration(p, errp); 320 #ifdef CONFIG_RDMA 321 } else if (strstart(uri, "rdma:", &p)) { 322 rdma_start_incoming_migration(p, errp); 323 #endif 324 } else if (strstart(uri, "exec:", &p)) { 325 exec_start_incoming_migration(p, errp); 326 } else if (strstart(uri, "unix:", &p)) { 327 unix_start_incoming_migration(p, errp); 328 } else if (strstart(uri, "fd:", &p)) { 329 fd_start_incoming_migration(p, errp); 330 } else { 331 error_setg(errp, "unknown migration protocol: %s", uri); 332 } 333 } 334 335 static void process_incoming_migration_bh(void *opaque) 336 { 337 Error *local_err = NULL; 338 MigrationIncomingState *mis = opaque; 339 340 /* Make sure all file formats flush their mutable metadata */ 341 bdrv_invalidate_cache_all(&local_err); 342 if (local_err) { 343 migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE, 344 MIGRATION_STATUS_FAILED); 345 error_report_err(local_err); 346 migrate_decompress_threads_join(); 347 exit(EXIT_FAILURE); 348 } 349 350 /* 351 * This must happen after all error conditions are dealt with and 352 * we're sure the VM is going to be running on this host. 353 */ 354 qemu_announce_self(); 355 356 /* If global state section was not received or we are in running 357 state, we need to obey autostart. Any other state is set with 358 runstate_set. */ 359 360 if (!global_state_received() || 361 global_state_get_runstate() == RUN_STATE_RUNNING) { 362 if (autostart) { 363 vm_start(); 364 } else { 365 runstate_set(RUN_STATE_PAUSED); 366 } 367 } else { 368 runstate_set(global_state_get_runstate()); 369 } 370 migrate_decompress_threads_join(); 371 /* 372 * This must happen after any state changes since as soon as an external 373 * observer sees this event they might start to prod at the VM assuming 374 * it's ready to use. 375 */ 376 migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE, 377 MIGRATION_STATUS_COMPLETED); 378 qemu_bh_delete(mis->bh); 379 migration_incoming_state_destroy(); 380 } 381 382 static void process_incoming_migration_co(void *opaque) 383 { 384 QEMUFile *f = opaque; 385 MigrationIncomingState *mis; 386 PostcopyState ps; 387 int ret; 388 389 mis = migration_incoming_state_new(f); 390 postcopy_state_set(POSTCOPY_INCOMING_NONE); 391 migrate_set_state(&mis->state, MIGRATION_STATUS_NONE, 392 MIGRATION_STATUS_ACTIVE); 393 ret = qemu_loadvm_state(f); 394 395 ps = postcopy_state_get(); 396 trace_process_incoming_migration_co_end(ret, ps); 397 if (ps != POSTCOPY_INCOMING_NONE) { 398 if (ps == POSTCOPY_INCOMING_ADVISE) { 399 /* 400 * Where a migration had postcopy enabled (and thus went to advise) 401 * but managed to complete within the precopy period, we can use 402 * the normal exit. 403 */ 404 postcopy_ram_incoming_cleanup(mis); 405 } else if (ret >= 0) { 406 /* 407 * Postcopy was started, cleanup should happen at the end of the 408 * postcopy thread. 409 */ 410 trace_process_incoming_migration_co_postcopy_end_main(); 411 return; 412 } 413 /* Else if something went wrong then just fall out of the normal exit */ 414 } 415 416 /* we get COLO info, and know if we are in COLO mode */ 417 if (!ret && migration_incoming_enable_colo()) { 418 mis->migration_incoming_co = qemu_coroutine_self(); 419 qemu_thread_create(&mis->colo_incoming_thread, "COLO incoming", 420 colo_process_incoming_thread, mis, QEMU_THREAD_JOINABLE); 421 mis->have_colo_incoming_thread = true; 422 qemu_coroutine_yield(); 423 424 /* Wait checkpoint incoming thread exit before free resource */ 425 qemu_thread_join(&mis->colo_incoming_thread); 426 } 427 428 qemu_fclose(f); 429 free_xbzrle_decoded_buf(); 430 431 if (ret < 0) { 432 migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE, 433 MIGRATION_STATUS_FAILED); 434 error_report("load of migration failed: %s", strerror(-ret)); 435 migrate_decompress_threads_join(); 436 exit(EXIT_FAILURE); 437 } 438 439 mis->bh = qemu_bh_new(process_incoming_migration_bh, mis); 440 qemu_bh_schedule(mis->bh); 441 } 442 443 void migration_fd_process_incoming(QEMUFile *f) 444 { 445 Coroutine *co = qemu_coroutine_create(process_incoming_migration_co, f); 446 447 migrate_decompress_threads_create(); 448 qemu_file_set_blocking(f, false); 449 qemu_coroutine_enter(co); 450 } 451 452 453 void migration_channel_process_incoming(MigrationState *s, 454 QIOChannel *ioc) 455 { 456 trace_migration_set_incoming_channel( 457 ioc, object_get_typename(OBJECT(ioc))); 458 459 if (s->parameters.tls_creds && 460 !object_dynamic_cast(OBJECT(ioc), 461 TYPE_QIO_CHANNEL_TLS)) { 462 Error *local_err = NULL; 463 migration_tls_channel_process_incoming(s, ioc, &local_err); 464 if (local_err) { 465 error_report_err(local_err); 466 } 467 } else { 468 QEMUFile *f = qemu_fopen_channel_input(ioc); 469 migration_fd_process_incoming(f); 470 } 471 } 472 473 474 void migration_channel_connect(MigrationState *s, 475 QIOChannel *ioc, 476 const char *hostname) 477 { 478 trace_migration_set_outgoing_channel( 479 ioc, object_get_typename(OBJECT(ioc)), hostname); 480 481 if (s->parameters.tls_creds && 482 !object_dynamic_cast(OBJECT(ioc), 483 TYPE_QIO_CHANNEL_TLS)) { 484 Error *local_err = NULL; 485 migration_tls_channel_connect(s, ioc, hostname, &local_err); 486 if (local_err) { 487 migrate_fd_error(s, local_err); 488 error_free(local_err); 489 } 490 } else { 491 QEMUFile *f = qemu_fopen_channel_output(ioc); 492 493 s->to_dst_file = f; 494 495 migrate_fd_connect(s); 496 } 497 } 498 499 500 /* 501 * Send a message on the return channel back to the source 502 * of the migration. 503 */ 504 void migrate_send_rp_message(MigrationIncomingState *mis, 505 enum mig_rp_message_type message_type, 506 uint16_t len, void *data) 507 { 508 trace_migrate_send_rp_message((int)message_type, len); 509 qemu_mutex_lock(&mis->rp_mutex); 510 qemu_put_be16(mis->to_src_file, (unsigned int)message_type); 511 qemu_put_be16(mis->to_src_file, len); 512 qemu_put_buffer(mis->to_src_file, data, len); 513 qemu_fflush(mis->to_src_file); 514 qemu_mutex_unlock(&mis->rp_mutex); 515 } 516 517 /* 518 * Send a 'SHUT' message on the return channel with the given value 519 * to indicate that we've finished with the RP. Non-0 value indicates 520 * error. 521 */ 522 void migrate_send_rp_shut(MigrationIncomingState *mis, 523 uint32_t value) 524 { 525 uint32_t buf; 526 527 buf = cpu_to_be32(value); 528 migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf); 529 } 530 531 /* 532 * Send a 'PONG' message on the return channel with the given value 533 * (normally in response to a 'PING') 534 */ 535 void migrate_send_rp_pong(MigrationIncomingState *mis, 536 uint32_t value) 537 { 538 uint32_t buf; 539 540 buf = cpu_to_be32(value); 541 migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf); 542 } 543 544 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp) 545 { 546 MigrationCapabilityStatusList *head = NULL; 547 MigrationCapabilityStatusList *caps; 548 MigrationState *s = migrate_get_current(); 549 int i; 550 551 caps = NULL; /* silence compiler warning */ 552 for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) { 553 if (i == MIGRATION_CAPABILITY_X_COLO && !colo_supported()) { 554 continue; 555 } 556 if (head == NULL) { 557 head = g_malloc0(sizeof(*caps)); 558 caps = head; 559 } else { 560 caps->next = g_malloc0(sizeof(*caps)); 561 caps = caps->next; 562 } 563 caps->value = 564 g_malloc(sizeof(*caps->value)); 565 caps->value->capability = i; 566 caps->value->state = s->enabled_capabilities[i]; 567 } 568 569 return head; 570 } 571 572 MigrationParameters *qmp_query_migrate_parameters(Error **errp) 573 { 574 MigrationParameters *params; 575 MigrationState *s = migrate_get_current(); 576 577 params = g_malloc0(sizeof(*params)); 578 params->has_compress_level = true; 579 params->compress_level = s->parameters.compress_level; 580 params->has_compress_threads = true; 581 params->compress_threads = s->parameters.compress_threads; 582 params->has_decompress_threads = true; 583 params->decompress_threads = s->parameters.decompress_threads; 584 params->has_cpu_throttle_initial = true; 585 params->cpu_throttle_initial = s->parameters.cpu_throttle_initial; 586 params->has_cpu_throttle_increment = true; 587 params->cpu_throttle_increment = s->parameters.cpu_throttle_increment; 588 params->has_tls_creds = !!s->parameters.tls_creds; 589 params->tls_creds = g_strdup(s->parameters.tls_creds); 590 params->has_tls_hostname = !!s->parameters.tls_hostname; 591 params->tls_hostname = g_strdup(s->parameters.tls_hostname); 592 params->has_max_bandwidth = true; 593 params->max_bandwidth = s->parameters.max_bandwidth; 594 params->has_downtime_limit = true; 595 params->downtime_limit = s->parameters.downtime_limit; 596 params->x_checkpoint_delay = s->parameters.x_checkpoint_delay; 597 598 return params; 599 } 600 601 /* 602 * Return true if we're already in the middle of a migration 603 * (i.e. any of the active or setup states) 604 */ 605 static bool migration_is_setup_or_active(int state) 606 { 607 switch (state) { 608 case MIGRATION_STATUS_ACTIVE: 609 case MIGRATION_STATUS_POSTCOPY_ACTIVE: 610 case MIGRATION_STATUS_SETUP: 611 return true; 612 613 default: 614 return false; 615 616 } 617 } 618 619 static void get_xbzrle_cache_stats(MigrationInfo *info) 620 { 621 if (migrate_use_xbzrle()) { 622 info->has_xbzrle_cache = true; 623 info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache)); 624 info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size(); 625 info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred(); 626 info->xbzrle_cache->pages = xbzrle_mig_pages_transferred(); 627 info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss(); 628 info->xbzrle_cache->cache_miss_rate = xbzrle_mig_cache_miss_rate(); 629 info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow(); 630 } 631 } 632 633 static void populate_ram_info(MigrationInfo *info, MigrationState *s) 634 { 635 info->has_ram = true; 636 info->ram = g_malloc0(sizeof(*info->ram)); 637 info->ram->transferred = ram_bytes_transferred(); 638 info->ram->total = ram_bytes_total(); 639 info->ram->duplicate = dup_mig_pages_transferred(); 640 info->ram->skipped = skipped_mig_pages_transferred(); 641 info->ram->normal = norm_mig_pages_transferred(); 642 info->ram->normal_bytes = norm_mig_bytes_transferred(); 643 info->ram->mbps = s->mbps; 644 info->ram->dirty_sync_count = s->dirty_sync_count; 645 info->ram->postcopy_requests = s->postcopy_requests; 646 647 if (s->state != MIGRATION_STATUS_COMPLETED) { 648 info->ram->remaining = ram_bytes_remaining(); 649 info->ram->dirty_pages_rate = s->dirty_pages_rate; 650 } 651 } 652 653 MigrationInfo *qmp_query_migrate(Error **errp) 654 { 655 MigrationInfo *info = g_malloc0(sizeof(*info)); 656 MigrationState *s = migrate_get_current(); 657 658 switch (s->state) { 659 case MIGRATION_STATUS_NONE: 660 /* no migration has happened ever */ 661 break; 662 case MIGRATION_STATUS_SETUP: 663 info->has_status = true; 664 info->has_total_time = false; 665 break; 666 case MIGRATION_STATUS_ACTIVE: 667 case MIGRATION_STATUS_CANCELLING: 668 info->has_status = true; 669 info->has_total_time = true; 670 info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) 671 - s->total_time; 672 info->has_expected_downtime = true; 673 info->expected_downtime = s->expected_downtime; 674 info->has_setup_time = true; 675 info->setup_time = s->setup_time; 676 677 populate_ram_info(info, s); 678 679 if (blk_mig_active()) { 680 info->has_disk = true; 681 info->disk = g_malloc0(sizeof(*info->disk)); 682 info->disk->transferred = blk_mig_bytes_transferred(); 683 info->disk->remaining = blk_mig_bytes_remaining(); 684 info->disk->total = blk_mig_bytes_total(); 685 } 686 687 if (cpu_throttle_active()) { 688 info->has_cpu_throttle_percentage = true; 689 info->cpu_throttle_percentage = cpu_throttle_get_percentage(); 690 } 691 692 get_xbzrle_cache_stats(info); 693 break; 694 case MIGRATION_STATUS_POSTCOPY_ACTIVE: 695 /* Mostly the same as active; TODO add some postcopy stats */ 696 info->has_status = true; 697 info->has_total_time = true; 698 info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) 699 - s->total_time; 700 info->has_expected_downtime = true; 701 info->expected_downtime = s->expected_downtime; 702 info->has_setup_time = true; 703 info->setup_time = s->setup_time; 704 705 populate_ram_info(info, s); 706 707 if (blk_mig_active()) { 708 info->has_disk = true; 709 info->disk = g_malloc0(sizeof(*info->disk)); 710 info->disk->transferred = blk_mig_bytes_transferred(); 711 info->disk->remaining = blk_mig_bytes_remaining(); 712 info->disk->total = blk_mig_bytes_total(); 713 } 714 715 get_xbzrle_cache_stats(info); 716 break; 717 case MIGRATION_STATUS_COLO: 718 info->has_status = true; 719 /* TODO: display COLO specific information (checkpoint info etc.) */ 720 break; 721 case MIGRATION_STATUS_COMPLETED: 722 get_xbzrle_cache_stats(info); 723 724 info->has_status = true; 725 info->has_total_time = true; 726 info->total_time = s->total_time; 727 info->has_downtime = true; 728 info->downtime = s->downtime; 729 info->has_setup_time = true; 730 info->setup_time = s->setup_time; 731 732 populate_ram_info(info, s); 733 break; 734 case MIGRATION_STATUS_FAILED: 735 info->has_status = true; 736 if (s->error) { 737 info->has_error_desc = true; 738 info->error_desc = g_strdup(error_get_pretty(s->error)); 739 } 740 break; 741 case MIGRATION_STATUS_CANCELLED: 742 info->has_status = true; 743 break; 744 } 745 info->status = s->state; 746 747 return info; 748 } 749 750 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params, 751 Error **errp) 752 { 753 MigrationState *s = migrate_get_current(); 754 MigrationCapabilityStatusList *cap; 755 bool old_postcopy_cap = migrate_postcopy_ram(); 756 757 if (migration_is_setup_or_active(s->state)) { 758 error_setg(errp, QERR_MIGRATION_ACTIVE); 759 return; 760 } 761 762 for (cap = params; cap; cap = cap->next) { 763 if (cap->value->capability == MIGRATION_CAPABILITY_X_COLO) { 764 if (!colo_supported()) { 765 error_setg(errp, "COLO is not currently supported, please" 766 " configure with --enable-colo option in order to" 767 " support COLO feature"); 768 continue; 769 } 770 } 771 s->enabled_capabilities[cap->value->capability] = cap->value->state; 772 } 773 774 if (migrate_postcopy_ram()) { 775 if (migrate_use_compression()) { 776 /* The decompression threads asynchronously write into RAM 777 * rather than use the atomic copies needed to avoid 778 * userfaulting. It should be possible to fix the decompression 779 * threads for compatibility in future. 780 */ 781 error_report("Postcopy is not currently compatible with " 782 "compression"); 783 s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] = 784 false; 785 } 786 /* This check is reasonably expensive, so only when it's being 787 * set the first time, also it's only the destination that needs 788 * special support. 789 */ 790 if (!old_postcopy_cap && runstate_check(RUN_STATE_INMIGRATE) && 791 !postcopy_ram_supported_by_host()) { 792 /* postcopy_ram_supported_by_host will have emitted a more 793 * detailed message 794 */ 795 error_report("Postcopy is not supported"); 796 s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] = 797 false; 798 } 799 } 800 } 801 802 void qmp_migrate_set_parameters(MigrationParameters *params, Error **errp) 803 { 804 MigrationState *s = migrate_get_current(); 805 806 if (params->has_compress_level && 807 (params->compress_level < 0 || params->compress_level > 9)) { 808 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level", 809 "is invalid, it should be in the range of 0 to 9"); 810 return; 811 } 812 if (params->has_compress_threads && 813 (params->compress_threads < 1 || params->compress_threads > 255)) { 814 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 815 "compress_threads", 816 "is invalid, it should be in the range of 1 to 255"); 817 return; 818 } 819 if (params->has_decompress_threads && 820 (params->decompress_threads < 1 || params->decompress_threads > 255)) { 821 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 822 "decompress_threads", 823 "is invalid, it should be in the range of 1 to 255"); 824 return; 825 } 826 if (params->has_cpu_throttle_initial && 827 (params->cpu_throttle_initial < 1 || 828 params->cpu_throttle_initial > 99)) { 829 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 830 "cpu_throttle_initial", 831 "an integer in the range of 1 to 99"); 832 return; 833 } 834 if (params->has_cpu_throttle_increment && 835 (params->cpu_throttle_increment < 1 || 836 params->cpu_throttle_increment > 99)) { 837 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 838 "cpu_throttle_increment", 839 "an integer in the range of 1 to 99"); 840 return; 841 } 842 if (params->has_max_bandwidth && 843 (params->max_bandwidth < 0 || params->max_bandwidth > SIZE_MAX)) { 844 error_setg(errp, "Parameter 'max_bandwidth' expects an integer in the" 845 " range of 0 to %zu bytes/second", SIZE_MAX); 846 return; 847 } 848 if (params->has_downtime_limit && 849 (params->downtime_limit < 0 || params->downtime_limit > 2000000)) { 850 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 851 "downtime_limit", 852 "an integer in the range of 0 to 2000000 milliseconds"); 853 return; 854 } 855 if (params->has_x_checkpoint_delay && (params->x_checkpoint_delay < 0)) { 856 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 857 "x_checkpoint_delay", 858 "is invalid, it should be positive"); 859 } 860 861 if (params->has_compress_level) { 862 s->parameters.compress_level = params->compress_level; 863 } 864 if (params->has_compress_threads) { 865 s->parameters.compress_threads = params->compress_threads; 866 } 867 if (params->has_decompress_threads) { 868 s->parameters.decompress_threads = params->decompress_threads; 869 } 870 if (params->has_cpu_throttle_initial) { 871 s->parameters.cpu_throttle_initial = params->cpu_throttle_initial; 872 } 873 if (params->has_cpu_throttle_increment) { 874 s->parameters.cpu_throttle_increment = params->cpu_throttle_increment; 875 } 876 if (params->has_tls_creds) { 877 g_free(s->parameters.tls_creds); 878 s->parameters.tls_creds = g_strdup(params->tls_creds); 879 } 880 if (params->has_tls_hostname) { 881 g_free(s->parameters.tls_hostname); 882 s->parameters.tls_hostname = g_strdup(params->tls_hostname); 883 } 884 if (params->has_max_bandwidth) { 885 s->parameters.max_bandwidth = params->max_bandwidth; 886 if (s->to_dst_file) { 887 qemu_file_set_rate_limit(s->to_dst_file, 888 s->parameters.max_bandwidth / XFER_LIMIT_RATIO); 889 } 890 } 891 if (params->has_downtime_limit) { 892 s->parameters.downtime_limit = params->downtime_limit; 893 } 894 895 if (params->has_x_checkpoint_delay) { 896 s->parameters.x_checkpoint_delay = params->x_checkpoint_delay; 897 } 898 } 899 900 901 void qmp_migrate_start_postcopy(Error **errp) 902 { 903 MigrationState *s = migrate_get_current(); 904 905 if (!migrate_postcopy_ram()) { 906 error_setg(errp, "Enable postcopy with migrate_set_capability before" 907 " the start of migration"); 908 return; 909 } 910 911 if (s->state == MIGRATION_STATUS_NONE) { 912 error_setg(errp, "Postcopy must be started after migration has been" 913 " started"); 914 return; 915 } 916 /* 917 * we don't error if migration has finished since that would be racy 918 * with issuing this command. 919 */ 920 atomic_set(&s->start_postcopy, true); 921 } 922 923 /* shared migration helpers */ 924 925 void migrate_set_state(int *state, int old_state, int new_state) 926 { 927 if (atomic_cmpxchg(state, old_state, new_state) == old_state) { 928 trace_migrate_set_state(new_state); 929 migrate_generate_event(new_state); 930 } 931 } 932 933 static void migrate_fd_cleanup(void *opaque) 934 { 935 MigrationState *s = opaque; 936 937 qemu_bh_delete(s->cleanup_bh); 938 s->cleanup_bh = NULL; 939 940 flush_page_queue(s); 941 942 if (s->to_dst_file) { 943 trace_migrate_fd_cleanup(); 944 qemu_mutex_unlock_iothread(); 945 if (s->migration_thread_running) { 946 qemu_thread_join(&s->thread); 947 s->migration_thread_running = false; 948 } 949 qemu_mutex_lock_iothread(); 950 951 migrate_compress_threads_join(); 952 qemu_fclose(s->to_dst_file); 953 s->to_dst_file = NULL; 954 } 955 956 assert((s->state != MIGRATION_STATUS_ACTIVE) && 957 (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE)); 958 959 if (s->state == MIGRATION_STATUS_CANCELLING) { 960 migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING, 961 MIGRATION_STATUS_CANCELLED); 962 } 963 964 notifier_list_notify(&migration_state_notifiers, s); 965 } 966 967 void migrate_fd_error(MigrationState *s, const Error *error) 968 { 969 trace_migrate_fd_error(error_get_pretty(error)); 970 assert(s->to_dst_file == NULL); 971 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP, 972 MIGRATION_STATUS_FAILED); 973 if (!s->error) { 974 s->error = error_copy(error); 975 } 976 notifier_list_notify(&migration_state_notifiers, s); 977 } 978 979 static void migrate_fd_cancel(MigrationState *s) 980 { 981 int old_state ; 982 QEMUFile *f = migrate_get_current()->to_dst_file; 983 trace_migrate_fd_cancel(); 984 985 if (s->rp_state.from_dst_file) { 986 /* shutdown the rp socket, so causing the rp thread to shutdown */ 987 qemu_file_shutdown(s->rp_state.from_dst_file); 988 } 989 990 do { 991 old_state = s->state; 992 if (!migration_is_setup_or_active(old_state)) { 993 break; 994 } 995 migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING); 996 } while (s->state != MIGRATION_STATUS_CANCELLING); 997 998 /* 999 * If we're unlucky the migration code might be stuck somewhere in a 1000 * send/write while the network has failed and is waiting to timeout; 1001 * if we've got shutdown(2) available then we can force it to quit. 1002 * The outgoing qemu file gets closed in migrate_fd_cleanup that is 1003 * called in a bh, so there is no race against this cancel. 1004 */ 1005 if (s->state == MIGRATION_STATUS_CANCELLING && f) { 1006 qemu_file_shutdown(f); 1007 } 1008 } 1009 1010 void add_migration_state_change_notifier(Notifier *notify) 1011 { 1012 notifier_list_add(&migration_state_notifiers, notify); 1013 } 1014 1015 void remove_migration_state_change_notifier(Notifier *notify) 1016 { 1017 notifier_remove(notify); 1018 } 1019 1020 bool migration_in_setup(MigrationState *s) 1021 { 1022 return s->state == MIGRATION_STATUS_SETUP; 1023 } 1024 1025 bool migration_has_finished(MigrationState *s) 1026 { 1027 return s->state == MIGRATION_STATUS_COMPLETED; 1028 } 1029 1030 bool migration_has_failed(MigrationState *s) 1031 { 1032 return (s->state == MIGRATION_STATUS_CANCELLED || 1033 s->state == MIGRATION_STATUS_FAILED); 1034 } 1035 1036 bool migration_in_postcopy(MigrationState *s) 1037 { 1038 return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE); 1039 } 1040 1041 bool migration_in_postcopy_after_devices(MigrationState *s) 1042 { 1043 return migration_in_postcopy(s) && s->postcopy_after_devices; 1044 } 1045 1046 MigrationState *migrate_init(const MigrationParams *params) 1047 { 1048 MigrationState *s = migrate_get_current(); 1049 1050 /* 1051 * Reinitialise all migration state, except 1052 * parameters/capabilities that the user set, and 1053 * locks. 1054 */ 1055 s->bytes_xfer = 0; 1056 s->xfer_limit = 0; 1057 s->cleanup_bh = 0; 1058 s->to_dst_file = NULL; 1059 s->state = MIGRATION_STATUS_NONE; 1060 s->params = *params; 1061 s->rp_state.from_dst_file = NULL; 1062 s->rp_state.error = false; 1063 s->mbps = 0.0; 1064 s->downtime = 0; 1065 s->expected_downtime = 0; 1066 s->dirty_pages_rate = 0; 1067 s->dirty_bytes_rate = 0; 1068 s->setup_time = 0; 1069 s->dirty_sync_count = 0; 1070 s->start_postcopy = false; 1071 s->postcopy_after_devices = false; 1072 s->postcopy_requests = 0; 1073 s->migration_thread_running = false; 1074 s->last_req_rb = NULL; 1075 error_free(s->error); 1076 s->error = NULL; 1077 1078 migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP); 1079 1080 QSIMPLEQ_INIT(&s->src_page_requests); 1081 1082 s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1083 return s; 1084 } 1085 1086 static GSList *migration_blockers; 1087 1088 void migrate_add_blocker(Error *reason) 1089 { 1090 migration_blockers = g_slist_prepend(migration_blockers, reason); 1091 } 1092 1093 void migrate_del_blocker(Error *reason) 1094 { 1095 migration_blockers = g_slist_remove(migration_blockers, reason); 1096 } 1097 1098 void qmp_migrate_incoming(const char *uri, Error **errp) 1099 { 1100 Error *local_err = NULL; 1101 static bool once = true; 1102 1103 if (!deferred_incoming) { 1104 error_setg(errp, "For use with '-incoming defer'"); 1105 return; 1106 } 1107 if (!once) { 1108 error_setg(errp, "The incoming migration has already been started"); 1109 } 1110 1111 qemu_start_incoming_migration(uri, &local_err); 1112 1113 if (local_err) { 1114 error_propagate(errp, local_err); 1115 return; 1116 } 1117 1118 once = false; 1119 } 1120 1121 bool migration_is_blocked(Error **errp) 1122 { 1123 if (qemu_savevm_state_blocked(errp)) { 1124 return true; 1125 } 1126 1127 if (migration_blockers) { 1128 *errp = error_copy(migration_blockers->data); 1129 return true; 1130 } 1131 1132 return false; 1133 } 1134 1135 void qmp_migrate(const char *uri, bool has_blk, bool blk, 1136 bool has_inc, bool inc, bool has_detach, bool detach, 1137 Error **errp) 1138 { 1139 Error *local_err = NULL; 1140 MigrationState *s = migrate_get_current(); 1141 MigrationParams params; 1142 const char *p; 1143 1144 params.blk = has_blk && blk; 1145 params.shared = has_inc && inc; 1146 1147 if (migration_is_setup_or_active(s->state) || 1148 s->state == MIGRATION_STATUS_CANCELLING || 1149 s->state == MIGRATION_STATUS_COLO) { 1150 error_setg(errp, QERR_MIGRATION_ACTIVE); 1151 return; 1152 } 1153 if (runstate_check(RUN_STATE_INMIGRATE)) { 1154 error_setg(errp, "Guest is waiting for an incoming migration"); 1155 return; 1156 } 1157 1158 if (migration_is_blocked(errp)) { 1159 return; 1160 } 1161 1162 s = migrate_init(¶ms); 1163 1164 if (strstart(uri, "tcp:", &p)) { 1165 tcp_start_outgoing_migration(s, p, &local_err); 1166 #ifdef CONFIG_RDMA 1167 } else if (strstart(uri, "rdma:", &p)) { 1168 rdma_start_outgoing_migration(s, p, &local_err); 1169 #endif 1170 } else if (strstart(uri, "exec:", &p)) { 1171 exec_start_outgoing_migration(s, p, &local_err); 1172 } else if (strstart(uri, "unix:", &p)) { 1173 unix_start_outgoing_migration(s, p, &local_err); 1174 } else if (strstart(uri, "fd:", &p)) { 1175 fd_start_outgoing_migration(s, p, &local_err); 1176 } else { 1177 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri", 1178 "a valid migration protocol"); 1179 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP, 1180 MIGRATION_STATUS_FAILED); 1181 return; 1182 } 1183 1184 if (local_err) { 1185 migrate_fd_error(s, local_err); 1186 error_propagate(errp, local_err); 1187 return; 1188 } 1189 } 1190 1191 void qmp_migrate_cancel(Error **errp) 1192 { 1193 migrate_fd_cancel(migrate_get_current()); 1194 } 1195 1196 void qmp_migrate_set_cache_size(int64_t value, Error **errp) 1197 { 1198 MigrationState *s = migrate_get_current(); 1199 int64_t new_size; 1200 1201 /* Check for truncation */ 1202 if (value != (size_t)value) { 1203 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", 1204 "exceeding address space"); 1205 return; 1206 } 1207 1208 /* Cache should not be larger than guest ram size */ 1209 if (value > ram_bytes_total()) { 1210 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", 1211 "exceeds guest ram size "); 1212 return; 1213 } 1214 1215 new_size = xbzrle_cache_resize(value); 1216 if (new_size < 0) { 1217 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", 1218 "is smaller than page size"); 1219 return; 1220 } 1221 1222 s->xbzrle_cache_size = new_size; 1223 } 1224 1225 int64_t qmp_query_migrate_cache_size(Error **errp) 1226 { 1227 return migrate_xbzrle_cache_size(); 1228 } 1229 1230 void qmp_migrate_set_speed(int64_t value, Error **errp) 1231 { 1232 MigrationParameters p = { 1233 .has_max_bandwidth = true, 1234 .max_bandwidth = value, 1235 }; 1236 1237 qmp_migrate_set_parameters(&p, errp); 1238 } 1239 1240 void qmp_migrate_set_downtime(double value, Error **errp) 1241 { 1242 value *= 1000; /* Convert to milliseconds */ 1243 value = MAX(0, MIN(INT64_MAX, value)); 1244 1245 MigrationParameters p = { 1246 .has_downtime_limit = true, 1247 .downtime_limit = value, 1248 }; 1249 1250 qmp_migrate_set_parameters(&p, errp); 1251 } 1252 1253 bool migrate_postcopy_ram(void) 1254 { 1255 MigrationState *s; 1256 1257 s = migrate_get_current(); 1258 1259 return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM]; 1260 } 1261 1262 bool migrate_auto_converge(void) 1263 { 1264 MigrationState *s; 1265 1266 s = migrate_get_current(); 1267 1268 return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE]; 1269 } 1270 1271 bool migrate_zero_blocks(void) 1272 { 1273 MigrationState *s; 1274 1275 s = migrate_get_current(); 1276 1277 return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS]; 1278 } 1279 1280 bool migrate_use_compression(void) 1281 { 1282 MigrationState *s; 1283 1284 s = migrate_get_current(); 1285 1286 return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS]; 1287 } 1288 1289 int migrate_compress_level(void) 1290 { 1291 MigrationState *s; 1292 1293 s = migrate_get_current(); 1294 1295 return s->parameters.compress_level; 1296 } 1297 1298 int migrate_compress_threads(void) 1299 { 1300 MigrationState *s; 1301 1302 s = migrate_get_current(); 1303 1304 return s->parameters.compress_threads; 1305 } 1306 1307 int migrate_decompress_threads(void) 1308 { 1309 MigrationState *s; 1310 1311 s = migrate_get_current(); 1312 1313 return s->parameters.decompress_threads; 1314 } 1315 1316 bool migrate_use_events(void) 1317 { 1318 MigrationState *s; 1319 1320 s = migrate_get_current(); 1321 1322 return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS]; 1323 } 1324 1325 int migrate_use_xbzrle(void) 1326 { 1327 MigrationState *s; 1328 1329 s = migrate_get_current(); 1330 1331 return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE]; 1332 } 1333 1334 int64_t migrate_xbzrle_cache_size(void) 1335 { 1336 MigrationState *s; 1337 1338 s = migrate_get_current(); 1339 1340 return s->xbzrle_cache_size; 1341 } 1342 1343 /* migration thread support */ 1344 /* 1345 * Something bad happened to the RP stream, mark an error 1346 * The caller shall print or trace something to indicate why 1347 */ 1348 static void mark_source_rp_bad(MigrationState *s) 1349 { 1350 s->rp_state.error = true; 1351 } 1352 1353 static struct rp_cmd_args { 1354 ssize_t len; /* -1 = variable */ 1355 const char *name; 1356 } rp_cmd_args[] = { 1357 [MIG_RP_MSG_INVALID] = { .len = -1, .name = "INVALID" }, 1358 [MIG_RP_MSG_SHUT] = { .len = 4, .name = "SHUT" }, 1359 [MIG_RP_MSG_PONG] = { .len = 4, .name = "PONG" }, 1360 [MIG_RP_MSG_REQ_PAGES] = { .len = 12, .name = "REQ_PAGES" }, 1361 [MIG_RP_MSG_REQ_PAGES_ID] = { .len = -1, .name = "REQ_PAGES_ID" }, 1362 [MIG_RP_MSG_MAX] = { .len = -1, .name = "MAX" }, 1363 }; 1364 1365 /* 1366 * Process a request for pages received on the return path, 1367 * We're allowed to send more than requested (e.g. to round to our page size) 1368 * and we don't need to send pages that have already been sent. 1369 */ 1370 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname, 1371 ram_addr_t start, size_t len) 1372 { 1373 long our_host_ps = getpagesize(); 1374 1375 trace_migrate_handle_rp_req_pages(rbname, start, len); 1376 1377 /* 1378 * Since we currently insist on matching page sizes, just sanity check 1379 * we're being asked for whole host pages. 1380 */ 1381 if (start & (our_host_ps-1) || 1382 (len & (our_host_ps-1))) { 1383 error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT 1384 " len: %zd", __func__, start, len); 1385 mark_source_rp_bad(ms); 1386 return; 1387 } 1388 1389 if (ram_save_queue_pages(ms, rbname, start, len)) { 1390 mark_source_rp_bad(ms); 1391 } 1392 } 1393 1394 /* 1395 * Handles messages sent on the return path towards the source VM 1396 * 1397 */ 1398 static void *source_return_path_thread(void *opaque) 1399 { 1400 MigrationState *ms = opaque; 1401 QEMUFile *rp = ms->rp_state.from_dst_file; 1402 uint16_t header_len, header_type; 1403 uint8_t buf[512]; 1404 uint32_t tmp32, sibling_error; 1405 ram_addr_t start = 0; /* =0 to silence warning */ 1406 size_t len = 0, expected_len; 1407 int res; 1408 1409 trace_source_return_path_thread_entry(); 1410 while (!ms->rp_state.error && !qemu_file_get_error(rp) && 1411 migration_is_setup_or_active(ms->state)) { 1412 trace_source_return_path_thread_loop_top(); 1413 header_type = qemu_get_be16(rp); 1414 header_len = qemu_get_be16(rp); 1415 1416 if (header_type >= MIG_RP_MSG_MAX || 1417 header_type == MIG_RP_MSG_INVALID) { 1418 error_report("RP: Received invalid message 0x%04x length 0x%04x", 1419 header_type, header_len); 1420 mark_source_rp_bad(ms); 1421 goto out; 1422 } 1423 1424 if ((rp_cmd_args[header_type].len != -1 && 1425 header_len != rp_cmd_args[header_type].len) || 1426 header_len > sizeof(buf)) { 1427 error_report("RP: Received '%s' message (0x%04x) with" 1428 "incorrect length %d expecting %zu", 1429 rp_cmd_args[header_type].name, header_type, header_len, 1430 (size_t)rp_cmd_args[header_type].len); 1431 mark_source_rp_bad(ms); 1432 goto out; 1433 } 1434 1435 /* We know we've got a valid header by this point */ 1436 res = qemu_get_buffer(rp, buf, header_len); 1437 if (res != header_len) { 1438 error_report("RP: Failed reading data for message 0x%04x" 1439 " read %d expected %d", 1440 header_type, res, header_len); 1441 mark_source_rp_bad(ms); 1442 goto out; 1443 } 1444 1445 /* OK, we have the message and the data */ 1446 switch (header_type) { 1447 case MIG_RP_MSG_SHUT: 1448 sibling_error = ldl_be_p(buf); 1449 trace_source_return_path_thread_shut(sibling_error); 1450 if (sibling_error) { 1451 error_report("RP: Sibling indicated error %d", sibling_error); 1452 mark_source_rp_bad(ms); 1453 } 1454 /* 1455 * We'll let the main thread deal with closing the RP 1456 * we could do a shutdown(2) on it, but we're the only user 1457 * anyway, so there's nothing gained. 1458 */ 1459 goto out; 1460 1461 case MIG_RP_MSG_PONG: 1462 tmp32 = ldl_be_p(buf); 1463 trace_source_return_path_thread_pong(tmp32); 1464 break; 1465 1466 case MIG_RP_MSG_REQ_PAGES: 1467 start = ldq_be_p(buf); 1468 len = ldl_be_p(buf + 8); 1469 migrate_handle_rp_req_pages(ms, NULL, start, len); 1470 break; 1471 1472 case MIG_RP_MSG_REQ_PAGES_ID: 1473 expected_len = 12 + 1; /* header + termination */ 1474 1475 if (header_len >= expected_len) { 1476 start = ldq_be_p(buf); 1477 len = ldl_be_p(buf + 8); 1478 /* Now we expect an idstr */ 1479 tmp32 = buf[12]; /* Length of the following idstr */ 1480 buf[13 + tmp32] = '\0'; 1481 expected_len += tmp32; 1482 } 1483 if (header_len != expected_len) { 1484 error_report("RP: Req_Page_id with length %d expecting %zd", 1485 header_len, expected_len); 1486 mark_source_rp_bad(ms); 1487 goto out; 1488 } 1489 migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len); 1490 break; 1491 1492 default: 1493 break; 1494 } 1495 } 1496 if (qemu_file_get_error(rp)) { 1497 trace_source_return_path_thread_bad_end(); 1498 mark_source_rp_bad(ms); 1499 } 1500 1501 trace_source_return_path_thread_end(); 1502 out: 1503 ms->rp_state.from_dst_file = NULL; 1504 qemu_fclose(rp); 1505 return NULL; 1506 } 1507 1508 static int open_return_path_on_source(MigrationState *ms) 1509 { 1510 1511 ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file); 1512 if (!ms->rp_state.from_dst_file) { 1513 return -1; 1514 } 1515 1516 trace_open_return_path_on_source(); 1517 qemu_thread_create(&ms->rp_state.rp_thread, "return path", 1518 source_return_path_thread, ms, QEMU_THREAD_JOINABLE); 1519 1520 trace_open_return_path_on_source_continue(); 1521 1522 return 0; 1523 } 1524 1525 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */ 1526 static int await_return_path_close_on_source(MigrationState *ms) 1527 { 1528 /* 1529 * If this is a normal exit then the destination will send a SHUT and the 1530 * rp_thread will exit, however if there's an error we need to cause 1531 * it to exit. 1532 */ 1533 if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) { 1534 /* 1535 * shutdown(2), if we have it, will cause it to unblock if it's stuck 1536 * waiting for the destination. 1537 */ 1538 qemu_file_shutdown(ms->rp_state.from_dst_file); 1539 mark_source_rp_bad(ms); 1540 } 1541 trace_await_return_path_close_on_source_joining(); 1542 qemu_thread_join(&ms->rp_state.rp_thread); 1543 trace_await_return_path_close_on_source_close(); 1544 return ms->rp_state.error; 1545 } 1546 1547 /* 1548 * Switch from normal iteration to postcopy 1549 * Returns non-0 on error 1550 */ 1551 static int postcopy_start(MigrationState *ms, bool *old_vm_running) 1552 { 1553 int ret; 1554 QIOChannelBuffer *bioc; 1555 QEMUFile *fb; 1556 int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1557 migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE, 1558 MIGRATION_STATUS_POSTCOPY_ACTIVE); 1559 1560 trace_postcopy_start(); 1561 qemu_mutex_lock_iothread(); 1562 trace_postcopy_start_set_run(); 1563 1564 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER); 1565 *old_vm_running = runstate_is_running(); 1566 global_state_store(); 1567 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE); 1568 if (ret < 0) { 1569 goto fail; 1570 } 1571 1572 ret = bdrv_inactivate_all(); 1573 if (ret < 0) { 1574 goto fail; 1575 } 1576 1577 /* 1578 * Cause any non-postcopiable, but iterative devices to 1579 * send out their final data. 1580 */ 1581 qemu_savevm_state_complete_precopy(ms->to_dst_file, true); 1582 1583 /* 1584 * in Finish migrate and with the io-lock held everything should 1585 * be quiet, but we've potentially still got dirty pages and we 1586 * need to tell the destination to throw any pages it's already received 1587 * that are dirty 1588 */ 1589 if (ram_postcopy_send_discard_bitmap(ms)) { 1590 error_report("postcopy send discard bitmap failed"); 1591 goto fail; 1592 } 1593 1594 /* 1595 * send rest of state - note things that are doing postcopy 1596 * will notice we're in POSTCOPY_ACTIVE and not actually 1597 * wrap their state up here 1598 */ 1599 qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX); 1600 /* Ping just for debugging, helps line traces up */ 1601 qemu_savevm_send_ping(ms->to_dst_file, 2); 1602 1603 /* 1604 * While loading the device state we may trigger page transfer 1605 * requests and the fd must be free to process those, and thus 1606 * the destination must read the whole device state off the fd before 1607 * it starts processing it. Unfortunately the ad-hoc migration format 1608 * doesn't allow the destination to know the size to read without fully 1609 * parsing it through each devices load-state code (especially the open 1610 * coded devices that use get/put). 1611 * So we wrap the device state up in a package with a length at the start; 1612 * to do this we use a qemu_buf to hold the whole of the device state. 1613 */ 1614 bioc = qio_channel_buffer_new(4096); 1615 qio_channel_set_name(QIO_CHANNEL(bioc), "migration-postcopy-buffer"); 1616 fb = qemu_fopen_channel_output(QIO_CHANNEL(bioc)); 1617 object_unref(OBJECT(bioc)); 1618 1619 /* 1620 * Make sure the receiver can get incoming pages before we send the rest 1621 * of the state 1622 */ 1623 qemu_savevm_send_postcopy_listen(fb); 1624 1625 qemu_savevm_state_complete_precopy(fb, false); 1626 qemu_savevm_send_ping(fb, 3); 1627 1628 qemu_savevm_send_postcopy_run(fb); 1629 1630 /* <><> end of stuff going into the package */ 1631 1632 /* Now send that blob */ 1633 if (qemu_savevm_send_packaged(ms->to_dst_file, bioc->data, bioc->usage)) { 1634 goto fail_closefb; 1635 } 1636 qemu_fclose(fb); 1637 1638 /* Send a notify to give a chance for anything that needs to happen 1639 * at the transition to postcopy and after the device state; in particular 1640 * spice needs to trigger a transition now 1641 */ 1642 ms->postcopy_after_devices = true; 1643 notifier_list_notify(&migration_state_notifiers, ms); 1644 1645 ms->downtime = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop; 1646 1647 qemu_mutex_unlock_iothread(); 1648 1649 /* 1650 * Although this ping is just for debug, it could potentially be 1651 * used for getting a better measurement of downtime at the source. 1652 */ 1653 qemu_savevm_send_ping(ms->to_dst_file, 4); 1654 1655 ret = qemu_file_get_error(ms->to_dst_file); 1656 if (ret) { 1657 error_report("postcopy_start: Migration stream errored"); 1658 migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE, 1659 MIGRATION_STATUS_FAILED); 1660 } 1661 1662 return ret; 1663 1664 fail_closefb: 1665 qemu_fclose(fb); 1666 fail: 1667 migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE, 1668 MIGRATION_STATUS_FAILED); 1669 qemu_mutex_unlock_iothread(); 1670 return -1; 1671 } 1672 1673 /** 1674 * migration_completion: Used by migration_thread when there's not much left. 1675 * The caller 'breaks' the loop when this returns. 1676 * 1677 * @s: Current migration state 1678 * @current_active_state: The migration state we expect to be in 1679 * @*old_vm_running: Pointer to old_vm_running flag 1680 * @*start_time: Pointer to time to update 1681 */ 1682 static void migration_completion(MigrationState *s, int current_active_state, 1683 bool *old_vm_running, 1684 int64_t *start_time) 1685 { 1686 int ret; 1687 1688 if (s->state == MIGRATION_STATUS_ACTIVE) { 1689 qemu_mutex_lock_iothread(); 1690 *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1691 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER); 1692 *old_vm_running = runstate_is_running(); 1693 ret = global_state_store(); 1694 1695 if (!ret) { 1696 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE); 1697 /* 1698 * Don't mark the image with BDRV_O_INACTIVE flag if 1699 * we will go into COLO stage later. 1700 */ 1701 if (ret >= 0 && !migrate_colo_enabled()) { 1702 ret = bdrv_inactivate_all(); 1703 } 1704 if (ret >= 0) { 1705 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX); 1706 qemu_savevm_state_complete_precopy(s->to_dst_file, false); 1707 } 1708 } 1709 qemu_mutex_unlock_iothread(); 1710 1711 if (ret < 0) { 1712 goto fail; 1713 } 1714 } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) { 1715 trace_migration_completion_postcopy_end(); 1716 1717 qemu_savevm_state_complete_postcopy(s->to_dst_file); 1718 trace_migration_completion_postcopy_end_after_complete(); 1719 } 1720 1721 /* 1722 * If rp was opened we must clean up the thread before 1723 * cleaning everything else up (since if there are no failures 1724 * it will wait for the destination to send it's status in 1725 * a SHUT command). 1726 * Postcopy opens rp if enabled (even if it's not avtivated) 1727 */ 1728 if (migrate_postcopy_ram()) { 1729 int rp_error; 1730 trace_migration_completion_postcopy_end_before_rp(); 1731 rp_error = await_return_path_close_on_source(s); 1732 trace_migration_completion_postcopy_end_after_rp(rp_error); 1733 if (rp_error) { 1734 goto fail_invalidate; 1735 } 1736 } 1737 1738 if (qemu_file_get_error(s->to_dst_file)) { 1739 trace_migration_completion_file_err(); 1740 goto fail_invalidate; 1741 } 1742 1743 if (!migrate_colo_enabled()) { 1744 migrate_set_state(&s->state, current_active_state, 1745 MIGRATION_STATUS_COMPLETED); 1746 } 1747 1748 return; 1749 1750 fail_invalidate: 1751 /* If not doing postcopy, vm_start() will be called: let's regain 1752 * control on images. 1753 */ 1754 if (s->state == MIGRATION_STATUS_ACTIVE) { 1755 Error *local_err = NULL; 1756 1757 bdrv_invalidate_cache_all(&local_err); 1758 if (local_err) { 1759 error_report_err(local_err); 1760 } 1761 } 1762 1763 fail: 1764 migrate_set_state(&s->state, current_active_state, 1765 MIGRATION_STATUS_FAILED); 1766 } 1767 1768 bool migrate_colo_enabled(void) 1769 { 1770 MigrationState *s = migrate_get_current(); 1771 return s->enabled_capabilities[MIGRATION_CAPABILITY_X_COLO]; 1772 } 1773 1774 /* 1775 * Master migration thread on the source VM. 1776 * It drives the migration and pumps the data down the outgoing channel. 1777 */ 1778 static void *migration_thread(void *opaque) 1779 { 1780 MigrationState *s = opaque; 1781 /* Used by the bandwidth calcs, updated later */ 1782 int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1783 int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST); 1784 int64_t initial_bytes = 0; 1785 int64_t max_size = 0; 1786 int64_t start_time = initial_time; 1787 int64_t end_time; 1788 bool old_vm_running = false; 1789 bool entered_postcopy = false; 1790 /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */ 1791 enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE; 1792 bool enable_colo = migrate_colo_enabled(); 1793 1794 rcu_register_thread(); 1795 1796 qemu_savevm_state_header(s->to_dst_file); 1797 1798 if (migrate_postcopy_ram()) { 1799 /* Now tell the dest that it should open its end so it can reply */ 1800 qemu_savevm_send_open_return_path(s->to_dst_file); 1801 1802 /* And do a ping that will make stuff easier to debug */ 1803 qemu_savevm_send_ping(s->to_dst_file, 1); 1804 1805 /* 1806 * Tell the destination that we *might* want to do postcopy later; 1807 * if the other end can't do postcopy it should fail now, nice and 1808 * early. 1809 */ 1810 qemu_savevm_send_postcopy_advise(s->to_dst_file); 1811 } 1812 1813 qemu_savevm_state_begin(s->to_dst_file, &s->params); 1814 1815 s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start; 1816 current_active_state = MIGRATION_STATUS_ACTIVE; 1817 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP, 1818 MIGRATION_STATUS_ACTIVE); 1819 1820 trace_migration_thread_setup_complete(); 1821 1822 while (s->state == MIGRATION_STATUS_ACTIVE || 1823 s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) { 1824 int64_t current_time; 1825 uint64_t pending_size; 1826 1827 if (!qemu_file_rate_limit(s->to_dst_file)) { 1828 uint64_t pend_post, pend_nonpost; 1829 1830 qemu_savevm_state_pending(s->to_dst_file, max_size, &pend_nonpost, 1831 &pend_post); 1832 pending_size = pend_nonpost + pend_post; 1833 trace_migrate_pending(pending_size, max_size, 1834 pend_post, pend_nonpost); 1835 if (pending_size && pending_size >= max_size) { 1836 /* Still a significant amount to transfer */ 1837 1838 if (migrate_postcopy_ram() && 1839 s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE && 1840 pend_nonpost <= max_size && 1841 atomic_read(&s->start_postcopy)) { 1842 1843 if (!postcopy_start(s, &old_vm_running)) { 1844 current_active_state = MIGRATION_STATUS_POSTCOPY_ACTIVE; 1845 entered_postcopy = true; 1846 } 1847 1848 continue; 1849 } 1850 /* Just another iteration step */ 1851 qemu_savevm_state_iterate(s->to_dst_file, entered_postcopy); 1852 } else { 1853 trace_migration_thread_low_pending(pending_size); 1854 migration_completion(s, current_active_state, 1855 &old_vm_running, &start_time); 1856 break; 1857 } 1858 } 1859 1860 if (qemu_file_get_error(s->to_dst_file)) { 1861 migrate_set_state(&s->state, current_active_state, 1862 MIGRATION_STATUS_FAILED); 1863 trace_migration_thread_file_err(); 1864 break; 1865 } 1866 current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1867 if (current_time >= initial_time + BUFFER_DELAY) { 1868 uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) - 1869 initial_bytes; 1870 uint64_t time_spent = current_time - initial_time; 1871 double bandwidth = (double)transferred_bytes / time_spent; 1872 max_size = bandwidth * s->parameters.downtime_limit; 1873 1874 s->mbps = (((double) transferred_bytes * 8.0) / 1875 ((double) time_spent / 1000.0)) / 1000.0 / 1000.0; 1876 1877 trace_migrate_transferred(transferred_bytes, time_spent, 1878 bandwidth, max_size); 1879 /* if we haven't sent anything, we don't want to recalculate 1880 10000 is a small enough number for our purposes */ 1881 if (s->dirty_bytes_rate && transferred_bytes > 10000) { 1882 s->expected_downtime = s->dirty_bytes_rate / bandwidth; 1883 } 1884 1885 qemu_file_reset_rate_limit(s->to_dst_file); 1886 initial_time = current_time; 1887 initial_bytes = qemu_ftell(s->to_dst_file); 1888 } 1889 if (qemu_file_rate_limit(s->to_dst_file)) { 1890 /* usleep expects microseconds */ 1891 g_usleep((initial_time + BUFFER_DELAY - current_time)*1000); 1892 } 1893 } 1894 1895 trace_migration_thread_after_loop(); 1896 /* If we enabled cpu throttling for auto-converge, turn it off. */ 1897 cpu_throttle_stop(); 1898 end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1899 1900 qemu_mutex_lock_iothread(); 1901 /* 1902 * The resource has been allocated by migration will be reused in COLO 1903 * process, so don't release them. 1904 */ 1905 if (!enable_colo) { 1906 qemu_savevm_state_cleanup(); 1907 } 1908 if (s->state == MIGRATION_STATUS_COMPLETED) { 1909 uint64_t transferred_bytes = qemu_ftell(s->to_dst_file); 1910 s->total_time = end_time - s->total_time; 1911 if (!entered_postcopy) { 1912 s->downtime = end_time - start_time; 1913 } 1914 if (s->total_time) { 1915 s->mbps = (((double) transferred_bytes * 8.0) / 1916 ((double) s->total_time)) / 1000; 1917 } 1918 runstate_set(RUN_STATE_POSTMIGRATE); 1919 } else { 1920 if (s->state == MIGRATION_STATUS_ACTIVE && enable_colo) { 1921 migrate_start_colo_process(s); 1922 qemu_savevm_state_cleanup(); 1923 /* 1924 * Fixme: we will run VM in COLO no matter its old running state. 1925 * After exited COLO, we will keep running. 1926 */ 1927 old_vm_running = true; 1928 } 1929 if (old_vm_running && !entered_postcopy) { 1930 vm_start(); 1931 } else { 1932 if (runstate_check(RUN_STATE_FINISH_MIGRATE)) { 1933 runstate_set(RUN_STATE_POSTMIGRATE); 1934 } 1935 } 1936 } 1937 qemu_bh_schedule(s->cleanup_bh); 1938 qemu_mutex_unlock_iothread(); 1939 1940 rcu_unregister_thread(); 1941 return NULL; 1942 } 1943 1944 void migrate_fd_connect(MigrationState *s) 1945 { 1946 s->expected_downtime = s->parameters.downtime_limit; 1947 s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s); 1948 1949 qemu_file_set_blocking(s->to_dst_file, true); 1950 qemu_file_set_rate_limit(s->to_dst_file, 1951 s->parameters.max_bandwidth / XFER_LIMIT_RATIO); 1952 1953 /* Notify before starting migration thread */ 1954 notifier_list_notify(&migration_state_notifiers, s); 1955 1956 /* 1957 * Open the return path; currently for postcopy but other things might 1958 * also want it. 1959 */ 1960 if (migrate_postcopy_ram()) { 1961 if (open_return_path_on_source(s)) { 1962 error_report("Unable to open return-path for postcopy"); 1963 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP, 1964 MIGRATION_STATUS_FAILED); 1965 migrate_fd_cleanup(s); 1966 return; 1967 } 1968 } 1969 1970 migrate_compress_threads_create(); 1971 qemu_thread_create(&s->thread, "migration", migration_thread, s, 1972 QEMU_THREAD_JOINABLE); 1973 s->migration_thread_running = true; 1974 } 1975 1976 PostcopyState postcopy_state_get(void) 1977 { 1978 return atomic_mb_read(&incoming_postcopy_state); 1979 } 1980 1981 /* Set the state and return the old state */ 1982 PostcopyState postcopy_state_set(PostcopyState new_state) 1983 { 1984 return atomic_xchg(&incoming_postcopy_state, new_state); 1985 } 1986 1987