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 MigrationIncomingState *migration_incoming_get_current(void) 115 { 116 static bool once; 117 static MigrationIncomingState mis_current; 118 119 if (!once) { 120 mis_current.state = MIGRATION_STATUS_NONE; 121 memset(&mis_current, 0, sizeof(MigrationIncomingState)); 122 QLIST_INIT(&mis_current.loadvm_handlers); 123 qemu_mutex_init(&mis_current.rp_mutex); 124 qemu_event_init(&mis_current.main_thread_load_event, false); 125 once = true; 126 } 127 return &mis_current; 128 } 129 130 void migration_incoming_state_destroy(void) 131 { 132 struct MigrationIncomingState *mis = migration_incoming_get_current(); 133 134 qemu_event_destroy(&mis->main_thread_load_event); 135 loadvm_free_handlers(mis); 136 } 137 138 139 typedef struct { 140 bool optional; 141 uint32_t size; 142 uint8_t runstate[100]; 143 RunState state; 144 bool received; 145 } GlobalState; 146 147 static GlobalState global_state; 148 149 int global_state_store(void) 150 { 151 if (!runstate_store((char *)global_state.runstate, 152 sizeof(global_state.runstate))) { 153 error_report("runstate name too big: %s", global_state.runstate); 154 trace_migrate_state_too_big(); 155 return -EINVAL; 156 } 157 return 0; 158 } 159 160 void global_state_store_running(void) 161 { 162 const char *state = RunState_lookup[RUN_STATE_RUNNING]; 163 strncpy((char *)global_state.runstate, 164 state, sizeof(global_state.runstate)); 165 } 166 167 static bool global_state_received(void) 168 { 169 return global_state.received; 170 } 171 172 static RunState global_state_get_runstate(void) 173 { 174 return global_state.state; 175 } 176 177 void global_state_set_optional(void) 178 { 179 global_state.optional = true; 180 } 181 182 static bool global_state_needed(void *opaque) 183 { 184 GlobalState *s = opaque; 185 char *runstate = (char *)s->runstate; 186 187 /* If it is not optional, it is mandatory */ 188 189 if (s->optional == false) { 190 return true; 191 } 192 193 /* If state is running or paused, it is not needed */ 194 195 if (strcmp(runstate, "running") == 0 || 196 strcmp(runstate, "paused") == 0) { 197 return false; 198 } 199 200 /* for any other state it is needed */ 201 return true; 202 } 203 204 static int global_state_post_load(void *opaque, int version_id) 205 { 206 GlobalState *s = opaque; 207 Error *local_err = NULL; 208 int r; 209 char *runstate = (char *)s->runstate; 210 211 s->received = true; 212 trace_migrate_global_state_post_load(runstate); 213 214 r = qapi_enum_parse(RunState_lookup, runstate, RUN_STATE__MAX, 215 -1, &local_err); 216 217 if (r == -1) { 218 if (local_err) { 219 error_report_err(local_err); 220 } 221 return -EINVAL; 222 } 223 s->state = r; 224 225 return 0; 226 } 227 228 static void global_state_pre_save(void *opaque) 229 { 230 GlobalState *s = opaque; 231 232 trace_migrate_global_state_pre_save((char *)s->runstate); 233 s->size = strlen((char *)s->runstate) + 1; 234 } 235 236 static const VMStateDescription vmstate_globalstate = { 237 .name = "globalstate", 238 .version_id = 1, 239 .minimum_version_id = 1, 240 .post_load = global_state_post_load, 241 .pre_save = global_state_pre_save, 242 .needed = global_state_needed, 243 .fields = (VMStateField[]) { 244 VMSTATE_UINT32(size, GlobalState), 245 VMSTATE_BUFFER(runstate, GlobalState), 246 VMSTATE_END_OF_LIST() 247 }, 248 }; 249 250 void register_global_state(void) 251 { 252 /* We would use it independently that we receive it */ 253 strcpy((char *)&global_state.runstate, ""); 254 global_state.received = false; 255 vmstate_register(NULL, 0, &vmstate_globalstate, &global_state); 256 } 257 258 static void migrate_generate_event(int new_state) 259 { 260 if (migrate_use_events()) { 261 qapi_event_send_migration(new_state, &error_abort); 262 } 263 } 264 265 /* 266 * Called on -incoming with a defer: uri. 267 * The migration can be started later after any parameters have been 268 * changed. 269 */ 270 static void deferred_incoming_migration(Error **errp) 271 { 272 if (deferred_incoming) { 273 error_setg(errp, "Incoming migration already deferred"); 274 } 275 deferred_incoming = true; 276 } 277 278 /* Request a range of pages from the source VM at the given 279 * start address. 280 * rbname: Name of the RAMBlock to request the page in, if NULL it's the same 281 * as the last request (a name must have been given previously) 282 * Start: Address offset within the RB 283 * Len: Length in bytes required - must be a multiple of pagesize 284 */ 285 void migrate_send_rp_req_pages(MigrationIncomingState *mis, const char *rbname, 286 ram_addr_t start, size_t len) 287 { 288 uint8_t bufc[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */ 289 size_t msglen = 12; /* start + len */ 290 291 *(uint64_t *)bufc = cpu_to_be64((uint64_t)start); 292 *(uint32_t *)(bufc + 8) = cpu_to_be32((uint32_t)len); 293 294 if (rbname) { 295 int rbname_len = strlen(rbname); 296 assert(rbname_len < 256); 297 298 bufc[msglen++] = rbname_len; 299 memcpy(bufc + msglen, rbname, rbname_len); 300 msglen += rbname_len; 301 migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES_ID, msglen, bufc); 302 } else { 303 migrate_send_rp_message(mis, MIG_RP_MSG_REQ_PAGES, msglen, bufc); 304 } 305 } 306 307 void qemu_start_incoming_migration(const char *uri, Error **errp) 308 { 309 const char *p; 310 311 qapi_event_send_migration(MIGRATION_STATUS_SETUP, &error_abort); 312 if (!strcmp(uri, "defer")) { 313 deferred_incoming_migration(errp); 314 } else if (strstart(uri, "tcp:", &p)) { 315 tcp_start_incoming_migration(p, errp); 316 #ifdef CONFIG_RDMA 317 } else if (strstart(uri, "rdma:", &p)) { 318 rdma_start_incoming_migration(p, errp); 319 #endif 320 } else if (strstart(uri, "exec:", &p)) { 321 exec_start_incoming_migration(p, errp); 322 } else if (strstart(uri, "unix:", &p)) { 323 unix_start_incoming_migration(p, errp); 324 } else if (strstart(uri, "fd:", &p)) { 325 fd_start_incoming_migration(p, errp); 326 } else { 327 error_setg(errp, "unknown migration protocol: %s", uri); 328 } 329 } 330 331 static void process_incoming_migration_bh(void *opaque) 332 { 333 Error *local_err = NULL; 334 MigrationIncomingState *mis = opaque; 335 336 /* Make sure all file formats flush their mutable metadata */ 337 bdrv_invalidate_cache_all(&local_err); 338 if (local_err) { 339 migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE, 340 MIGRATION_STATUS_FAILED); 341 error_report_err(local_err); 342 migrate_decompress_threads_join(); 343 exit(EXIT_FAILURE); 344 } 345 346 /* 347 * This must happen after all error conditions are dealt with and 348 * we're sure the VM is going to be running on this host. 349 */ 350 qemu_announce_self(); 351 352 /* If global state section was not received or we are in running 353 state, we need to obey autostart. Any other state is set with 354 runstate_set. */ 355 356 if (!global_state_received() || 357 global_state_get_runstate() == RUN_STATE_RUNNING) { 358 if (autostart) { 359 vm_start(); 360 } else { 361 runstate_set(RUN_STATE_PAUSED); 362 } 363 } else { 364 runstate_set(global_state_get_runstate()); 365 } 366 migrate_decompress_threads_join(); 367 /* 368 * This must happen after any state changes since as soon as an external 369 * observer sees this event they might start to prod at the VM assuming 370 * it's ready to use. 371 */ 372 migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE, 373 MIGRATION_STATUS_COMPLETED); 374 qemu_bh_delete(mis->bh); 375 migration_incoming_state_destroy(); 376 } 377 378 static void process_incoming_migration_co(void *opaque) 379 { 380 QEMUFile *f = opaque; 381 MigrationIncomingState *mis = migration_incoming_get_current(); 382 PostcopyState ps; 383 int ret; 384 385 mis->from_src_file = f; 386 postcopy_state_set(POSTCOPY_INCOMING_NONE); 387 migrate_set_state(&mis->state, MIGRATION_STATUS_NONE, 388 MIGRATION_STATUS_ACTIVE); 389 ret = qemu_loadvm_state(f); 390 391 ps = postcopy_state_get(); 392 trace_process_incoming_migration_co_end(ret, ps); 393 if (ps != POSTCOPY_INCOMING_NONE) { 394 if (ps == POSTCOPY_INCOMING_ADVISE) { 395 /* 396 * Where a migration had postcopy enabled (and thus went to advise) 397 * but managed to complete within the precopy period, we can use 398 * the normal exit. 399 */ 400 postcopy_ram_incoming_cleanup(mis); 401 } else if (ret >= 0) { 402 /* 403 * Postcopy was started, cleanup should happen at the end of the 404 * postcopy thread. 405 */ 406 trace_process_incoming_migration_co_postcopy_end_main(); 407 return; 408 } 409 /* Else if something went wrong then just fall out of the normal exit */ 410 } 411 412 /* we get COLO info, and know if we are in COLO mode */ 413 if (!ret && migration_incoming_enable_colo()) { 414 mis->migration_incoming_co = qemu_coroutine_self(); 415 qemu_thread_create(&mis->colo_incoming_thread, "COLO incoming", 416 colo_process_incoming_thread, mis, QEMU_THREAD_JOINABLE); 417 mis->have_colo_incoming_thread = true; 418 qemu_coroutine_yield(); 419 420 /* Wait checkpoint incoming thread exit before free resource */ 421 qemu_thread_join(&mis->colo_incoming_thread); 422 } 423 424 qemu_fclose(f); 425 free_xbzrle_decoded_buf(); 426 427 if (ret < 0) { 428 migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE, 429 MIGRATION_STATUS_FAILED); 430 error_report("load of migration failed: %s", strerror(-ret)); 431 migrate_decompress_threads_join(); 432 exit(EXIT_FAILURE); 433 } 434 435 mis->bh = qemu_bh_new(process_incoming_migration_bh, mis); 436 qemu_bh_schedule(mis->bh); 437 } 438 439 void migration_fd_process_incoming(QEMUFile *f) 440 { 441 Coroutine *co = qemu_coroutine_create(process_incoming_migration_co, f); 442 443 migrate_decompress_threads_create(); 444 qemu_file_set_blocking(f, false); 445 qemu_coroutine_enter(co); 446 } 447 448 449 void migration_channel_process_incoming(MigrationState *s, 450 QIOChannel *ioc) 451 { 452 trace_migration_set_incoming_channel( 453 ioc, object_get_typename(OBJECT(ioc))); 454 455 if (s->parameters.tls_creds && 456 !object_dynamic_cast(OBJECT(ioc), 457 TYPE_QIO_CHANNEL_TLS)) { 458 Error *local_err = NULL; 459 migration_tls_channel_process_incoming(s, ioc, &local_err); 460 if (local_err) { 461 error_report_err(local_err); 462 } 463 } else { 464 QEMUFile *f = qemu_fopen_channel_input(ioc); 465 migration_fd_process_incoming(f); 466 } 467 } 468 469 470 void migration_channel_connect(MigrationState *s, 471 QIOChannel *ioc, 472 const char *hostname) 473 { 474 trace_migration_set_outgoing_channel( 475 ioc, object_get_typename(OBJECT(ioc)), hostname); 476 477 if (s->parameters.tls_creds && 478 !object_dynamic_cast(OBJECT(ioc), 479 TYPE_QIO_CHANNEL_TLS)) { 480 Error *local_err = NULL; 481 migration_tls_channel_connect(s, ioc, hostname, &local_err); 482 if (local_err) { 483 migrate_fd_error(s, local_err); 484 error_free(local_err); 485 } 486 } else { 487 QEMUFile *f = qemu_fopen_channel_output(ioc); 488 489 s->to_dst_file = f; 490 491 migrate_fd_connect(s); 492 } 493 } 494 495 496 /* 497 * Send a message on the return channel back to the source 498 * of the migration. 499 */ 500 void migrate_send_rp_message(MigrationIncomingState *mis, 501 enum mig_rp_message_type message_type, 502 uint16_t len, void *data) 503 { 504 trace_migrate_send_rp_message((int)message_type, len); 505 qemu_mutex_lock(&mis->rp_mutex); 506 qemu_put_be16(mis->to_src_file, (unsigned int)message_type); 507 qemu_put_be16(mis->to_src_file, len); 508 qemu_put_buffer(mis->to_src_file, data, len); 509 qemu_fflush(mis->to_src_file); 510 qemu_mutex_unlock(&mis->rp_mutex); 511 } 512 513 /* 514 * Send a 'SHUT' message on the return channel with the given value 515 * to indicate that we've finished with the RP. Non-0 value indicates 516 * error. 517 */ 518 void migrate_send_rp_shut(MigrationIncomingState *mis, 519 uint32_t value) 520 { 521 uint32_t buf; 522 523 buf = cpu_to_be32(value); 524 migrate_send_rp_message(mis, MIG_RP_MSG_SHUT, sizeof(buf), &buf); 525 } 526 527 /* 528 * Send a 'PONG' message on the return channel with the given value 529 * (normally in response to a 'PING') 530 */ 531 void migrate_send_rp_pong(MigrationIncomingState *mis, 532 uint32_t value) 533 { 534 uint32_t buf; 535 536 buf = cpu_to_be32(value); 537 migrate_send_rp_message(mis, MIG_RP_MSG_PONG, sizeof(buf), &buf); 538 } 539 540 MigrationCapabilityStatusList *qmp_query_migrate_capabilities(Error **errp) 541 { 542 MigrationCapabilityStatusList *head = NULL; 543 MigrationCapabilityStatusList *caps; 544 MigrationState *s = migrate_get_current(); 545 int i; 546 547 caps = NULL; /* silence compiler warning */ 548 for (i = 0; i < MIGRATION_CAPABILITY__MAX; i++) { 549 if (i == MIGRATION_CAPABILITY_X_COLO && !colo_supported()) { 550 continue; 551 } 552 if (head == NULL) { 553 head = g_malloc0(sizeof(*caps)); 554 caps = head; 555 } else { 556 caps->next = g_malloc0(sizeof(*caps)); 557 caps = caps->next; 558 } 559 caps->value = 560 g_malloc(sizeof(*caps->value)); 561 caps->value->capability = i; 562 caps->value->state = s->enabled_capabilities[i]; 563 } 564 565 return head; 566 } 567 568 MigrationParameters *qmp_query_migrate_parameters(Error **errp) 569 { 570 MigrationParameters *params; 571 MigrationState *s = migrate_get_current(); 572 573 params = g_malloc0(sizeof(*params)); 574 params->has_compress_level = true; 575 params->compress_level = s->parameters.compress_level; 576 params->has_compress_threads = true; 577 params->compress_threads = s->parameters.compress_threads; 578 params->has_decompress_threads = true; 579 params->decompress_threads = s->parameters.decompress_threads; 580 params->has_cpu_throttle_initial = true; 581 params->cpu_throttle_initial = s->parameters.cpu_throttle_initial; 582 params->has_cpu_throttle_increment = true; 583 params->cpu_throttle_increment = s->parameters.cpu_throttle_increment; 584 params->has_tls_creds = !!s->parameters.tls_creds; 585 params->tls_creds = g_strdup(s->parameters.tls_creds); 586 params->has_tls_hostname = !!s->parameters.tls_hostname; 587 params->tls_hostname = g_strdup(s->parameters.tls_hostname); 588 params->has_max_bandwidth = true; 589 params->max_bandwidth = s->parameters.max_bandwidth; 590 params->has_downtime_limit = true; 591 params->downtime_limit = s->parameters.downtime_limit; 592 params->has_x_checkpoint_delay = true; 593 params->x_checkpoint_delay = s->parameters.x_checkpoint_delay; 594 595 return params; 596 } 597 598 /* 599 * Return true if we're already in the middle of a migration 600 * (i.e. any of the active or setup states) 601 */ 602 static bool migration_is_setup_or_active(int state) 603 { 604 switch (state) { 605 case MIGRATION_STATUS_ACTIVE: 606 case MIGRATION_STATUS_POSTCOPY_ACTIVE: 607 case MIGRATION_STATUS_SETUP: 608 return true; 609 610 default: 611 return false; 612 613 } 614 } 615 616 static void get_xbzrle_cache_stats(MigrationInfo *info) 617 { 618 if (migrate_use_xbzrle()) { 619 info->has_xbzrle_cache = true; 620 info->xbzrle_cache = g_malloc0(sizeof(*info->xbzrle_cache)); 621 info->xbzrle_cache->cache_size = migrate_xbzrle_cache_size(); 622 info->xbzrle_cache->bytes = xbzrle_mig_bytes_transferred(); 623 info->xbzrle_cache->pages = xbzrle_mig_pages_transferred(); 624 info->xbzrle_cache->cache_miss = xbzrle_mig_pages_cache_miss(); 625 info->xbzrle_cache->cache_miss_rate = xbzrle_mig_cache_miss_rate(); 626 info->xbzrle_cache->overflow = xbzrle_mig_pages_overflow(); 627 } 628 } 629 630 static void populate_ram_info(MigrationInfo *info, MigrationState *s) 631 { 632 info->has_ram = true; 633 info->ram = g_malloc0(sizeof(*info->ram)); 634 info->ram->transferred = ram_bytes_transferred(); 635 info->ram->total = ram_bytes_total(); 636 info->ram->duplicate = dup_mig_pages_transferred(); 637 info->ram->skipped = skipped_mig_pages_transferred(); 638 info->ram->normal = norm_mig_pages_transferred(); 639 info->ram->normal_bytes = norm_mig_bytes_transferred(); 640 info->ram->mbps = s->mbps; 641 info->ram->dirty_sync_count = s->dirty_sync_count; 642 info->ram->postcopy_requests = s->postcopy_requests; 643 644 if (s->state != MIGRATION_STATUS_COMPLETED) { 645 info->ram->remaining = ram_bytes_remaining(); 646 info->ram->dirty_pages_rate = s->dirty_pages_rate; 647 } 648 } 649 650 MigrationInfo *qmp_query_migrate(Error **errp) 651 { 652 MigrationInfo *info = g_malloc0(sizeof(*info)); 653 MigrationState *s = migrate_get_current(); 654 655 switch (s->state) { 656 case MIGRATION_STATUS_NONE: 657 /* no migration has happened ever */ 658 break; 659 case MIGRATION_STATUS_SETUP: 660 info->has_status = true; 661 info->has_total_time = false; 662 break; 663 case MIGRATION_STATUS_ACTIVE: 664 case MIGRATION_STATUS_CANCELLING: 665 info->has_status = true; 666 info->has_total_time = true; 667 info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) 668 - s->total_time; 669 info->has_expected_downtime = true; 670 info->expected_downtime = s->expected_downtime; 671 info->has_setup_time = true; 672 info->setup_time = s->setup_time; 673 674 populate_ram_info(info, s); 675 676 if (blk_mig_active()) { 677 info->has_disk = true; 678 info->disk = g_malloc0(sizeof(*info->disk)); 679 info->disk->transferred = blk_mig_bytes_transferred(); 680 info->disk->remaining = blk_mig_bytes_remaining(); 681 info->disk->total = blk_mig_bytes_total(); 682 } 683 684 if (cpu_throttle_active()) { 685 info->has_cpu_throttle_percentage = true; 686 info->cpu_throttle_percentage = cpu_throttle_get_percentage(); 687 } 688 689 get_xbzrle_cache_stats(info); 690 break; 691 case MIGRATION_STATUS_POSTCOPY_ACTIVE: 692 /* Mostly the same as active; TODO add some postcopy stats */ 693 info->has_status = true; 694 info->has_total_time = true; 695 info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) 696 - s->total_time; 697 info->has_expected_downtime = true; 698 info->expected_downtime = s->expected_downtime; 699 info->has_setup_time = true; 700 info->setup_time = s->setup_time; 701 702 populate_ram_info(info, s); 703 704 if (blk_mig_active()) { 705 info->has_disk = true; 706 info->disk = g_malloc0(sizeof(*info->disk)); 707 info->disk->transferred = blk_mig_bytes_transferred(); 708 info->disk->remaining = blk_mig_bytes_remaining(); 709 info->disk->total = blk_mig_bytes_total(); 710 } 711 712 get_xbzrle_cache_stats(info); 713 break; 714 case MIGRATION_STATUS_COLO: 715 info->has_status = true; 716 /* TODO: display COLO specific information (checkpoint info etc.) */ 717 break; 718 case MIGRATION_STATUS_COMPLETED: 719 get_xbzrle_cache_stats(info); 720 721 info->has_status = true; 722 info->has_total_time = true; 723 info->total_time = s->total_time; 724 info->has_downtime = true; 725 info->downtime = s->downtime; 726 info->has_setup_time = true; 727 info->setup_time = s->setup_time; 728 729 populate_ram_info(info, s); 730 break; 731 case MIGRATION_STATUS_FAILED: 732 info->has_status = true; 733 if (s->error) { 734 info->has_error_desc = true; 735 info->error_desc = g_strdup(error_get_pretty(s->error)); 736 } 737 break; 738 case MIGRATION_STATUS_CANCELLED: 739 info->has_status = true; 740 break; 741 } 742 info->status = s->state; 743 744 return info; 745 } 746 747 void qmp_migrate_set_capabilities(MigrationCapabilityStatusList *params, 748 Error **errp) 749 { 750 MigrationState *s = migrate_get_current(); 751 MigrationCapabilityStatusList *cap; 752 bool old_postcopy_cap = migrate_postcopy_ram(); 753 754 if (migration_is_setup_or_active(s->state)) { 755 error_setg(errp, QERR_MIGRATION_ACTIVE); 756 return; 757 } 758 759 for (cap = params; cap; cap = cap->next) { 760 if (cap->value->capability == MIGRATION_CAPABILITY_X_COLO) { 761 if (!colo_supported()) { 762 error_setg(errp, "COLO is not currently supported, please" 763 " configure with --enable-colo option in order to" 764 " support COLO feature"); 765 continue; 766 } 767 } 768 s->enabled_capabilities[cap->value->capability] = cap->value->state; 769 } 770 771 if (migrate_postcopy_ram()) { 772 if (migrate_use_compression()) { 773 /* The decompression threads asynchronously write into RAM 774 * rather than use the atomic copies needed to avoid 775 * userfaulting. It should be possible to fix the decompression 776 * threads for compatibility in future. 777 */ 778 error_report("Postcopy is not currently compatible with " 779 "compression"); 780 s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] = 781 false; 782 } 783 /* This check is reasonably expensive, so only when it's being 784 * set the first time, also it's only the destination that needs 785 * special support. 786 */ 787 if (!old_postcopy_cap && runstate_check(RUN_STATE_INMIGRATE) && 788 !postcopy_ram_supported_by_host()) { 789 /* postcopy_ram_supported_by_host will have emitted a more 790 * detailed message 791 */ 792 error_report("Postcopy is not supported"); 793 s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM] = 794 false; 795 } 796 } 797 } 798 799 void qmp_migrate_set_parameters(MigrationParameters *params, Error **errp) 800 { 801 MigrationState *s = migrate_get_current(); 802 803 if (params->has_compress_level && 804 (params->compress_level < 0 || params->compress_level > 9)) { 805 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "compress_level", 806 "is invalid, it should be in the range of 0 to 9"); 807 return; 808 } 809 if (params->has_compress_threads && 810 (params->compress_threads < 1 || params->compress_threads > 255)) { 811 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 812 "compress_threads", 813 "is invalid, it should be in the range of 1 to 255"); 814 return; 815 } 816 if (params->has_decompress_threads && 817 (params->decompress_threads < 1 || params->decompress_threads > 255)) { 818 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 819 "decompress_threads", 820 "is invalid, it should be in the range of 1 to 255"); 821 return; 822 } 823 if (params->has_cpu_throttle_initial && 824 (params->cpu_throttle_initial < 1 || 825 params->cpu_throttle_initial > 99)) { 826 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 827 "cpu_throttle_initial", 828 "an integer in the range of 1 to 99"); 829 return; 830 } 831 if (params->has_cpu_throttle_increment && 832 (params->cpu_throttle_increment < 1 || 833 params->cpu_throttle_increment > 99)) { 834 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 835 "cpu_throttle_increment", 836 "an integer in the range of 1 to 99"); 837 return; 838 } 839 if (params->has_max_bandwidth && 840 (params->max_bandwidth < 0 || params->max_bandwidth > SIZE_MAX)) { 841 error_setg(errp, "Parameter 'max_bandwidth' expects an integer in the" 842 " range of 0 to %zu bytes/second", SIZE_MAX); 843 return; 844 } 845 if (params->has_downtime_limit && 846 (params->downtime_limit < 0 || params->downtime_limit > 2000000)) { 847 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 848 "downtime_limit", 849 "an integer in the range of 0 to 2000000 milliseconds"); 850 return; 851 } 852 if (params->has_x_checkpoint_delay && (params->x_checkpoint_delay < 0)) { 853 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, 854 "x_checkpoint_delay", 855 "is invalid, it should be positive"); 856 } 857 858 if (params->has_compress_level) { 859 s->parameters.compress_level = params->compress_level; 860 } 861 if (params->has_compress_threads) { 862 s->parameters.compress_threads = params->compress_threads; 863 } 864 if (params->has_decompress_threads) { 865 s->parameters.decompress_threads = params->decompress_threads; 866 } 867 if (params->has_cpu_throttle_initial) { 868 s->parameters.cpu_throttle_initial = params->cpu_throttle_initial; 869 } 870 if (params->has_cpu_throttle_increment) { 871 s->parameters.cpu_throttle_increment = params->cpu_throttle_increment; 872 } 873 if (params->has_tls_creds) { 874 g_free(s->parameters.tls_creds); 875 s->parameters.tls_creds = g_strdup(params->tls_creds); 876 } 877 if (params->has_tls_hostname) { 878 g_free(s->parameters.tls_hostname); 879 s->parameters.tls_hostname = g_strdup(params->tls_hostname); 880 } 881 if (params->has_max_bandwidth) { 882 s->parameters.max_bandwidth = params->max_bandwidth; 883 if (s->to_dst_file) { 884 qemu_file_set_rate_limit(s->to_dst_file, 885 s->parameters.max_bandwidth / XFER_LIMIT_RATIO); 886 } 887 } 888 if (params->has_downtime_limit) { 889 s->parameters.downtime_limit = params->downtime_limit; 890 } 891 892 if (params->has_x_checkpoint_delay) { 893 s->parameters.x_checkpoint_delay = params->x_checkpoint_delay; 894 } 895 } 896 897 898 void qmp_migrate_start_postcopy(Error **errp) 899 { 900 MigrationState *s = migrate_get_current(); 901 902 if (!migrate_postcopy_ram()) { 903 error_setg(errp, "Enable postcopy with migrate_set_capability before" 904 " the start of migration"); 905 return; 906 } 907 908 if (s->state == MIGRATION_STATUS_NONE) { 909 error_setg(errp, "Postcopy must be started after migration has been" 910 " started"); 911 return; 912 } 913 /* 914 * we don't error if migration has finished since that would be racy 915 * with issuing this command. 916 */ 917 atomic_set(&s->start_postcopy, true); 918 } 919 920 /* shared migration helpers */ 921 922 void migrate_set_state(int *state, int old_state, int new_state) 923 { 924 if (atomic_cmpxchg(state, old_state, new_state) == old_state) { 925 trace_migrate_set_state(new_state); 926 migrate_generate_event(new_state); 927 } 928 } 929 930 static void migrate_fd_cleanup(void *opaque) 931 { 932 MigrationState *s = opaque; 933 934 qemu_bh_delete(s->cleanup_bh); 935 s->cleanup_bh = NULL; 936 937 flush_page_queue(s); 938 939 if (s->to_dst_file) { 940 trace_migrate_fd_cleanup(); 941 qemu_mutex_unlock_iothread(); 942 if (s->migration_thread_running) { 943 qemu_thread_join(&s->thread); 944 s->migration_thread_running = false; 945 } 946 qemu_mutex_lock_iothread(); 947 948 migrate_compress_threads_join(); 949 qemu_fclose(s->to_dst_file); 950 s->to_dst_file = NULL; 951 } 952 953 assert((s->state != MIGRATION_STATUS_ACTIVE) && 954 (s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE)); 955 956 if (s->state == MIGRATION_STATUS_CANCELLING) { 957 migrate_set_state(&s->state, MIGRATION_STATUS_CANCELLING, 958 MIGRATION_STATUS_CANCELLED); 959 } 960 961 notifier_list_notify(&migration_state_notifiers, s); 962 } 963 964 void migrate_fd_error(MigrationState *s, const Error *error) 965 { 966 trace_migrate_fd_error(error_get_pretty(error)); 967 assert(s->to_dst_file == NULL); 968 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP, 969 MIGRATION_STATUS_FAILED); 970 if (!s->error) { 971 s->error = error_copy(error); 972 } 973 notifier_list_notify(&migration_state_notifiers, s); 974 } 975 976 static void migrate_fd_cancel(MigrationState *s) 977 { 978 int old_state ; 979 QEMUFile *f = migrate_get_current()->to_dst_file; 980 trace_migrate_fd_cancel(); 981 982 if (s->rp_state.from_dst_file) { 983 /* shutdown the rp socket, so causing the rp thread to shutdown */ 984 qemu_file_shutdown(s->rp_state.from_dst_file); 985 } 986 987 do { 988 old_state = s->state; 989 if (!migration_is_setup_or_active(old_state)) { 990 break; 991 } 992 migrate_set_state(&s->state, old_state, MIGRATION_STATUS_CANCELLING); 993 } while (s->state != MIGRATION_STATUS_CANCELLING); 994 995 /* 996 * If we're unlucky the migration code might be stuck somewhere in a 997 * send/write while the network has failed and is waiting to timeout; 998 * if we've got shutdown(2) available then we can force it to quit. 999 * The outgoing qemu file gets closed in migrate_fd_cleanup that is 1000 * called in a bh, so there is no race against this cancel. 1001 */ 1002 if (s->state == MIGRATION_STATUS_CANCELLING && f) { 1003 qemu_file_shutdown(f); 1004 } 1005 if (s->state == MIGRATION_STATUS_CANCELLING && s->block_inactive) { 1006 Error *local_err = NULL; 1007 1008 bdrv_invalidate_cache_all(&local_err); 1009 if (local_err) { 1010 error_report_err(local_err); 1011 } else { 1012 s->block_inactive = false; 1013 } 1014 } 1015 } 1016 1017 void add_migration_state_change_notifier(Notifier *notify) 1018 { 1019 notifier_list_add(&migration_state_notifiers, notify); 1020 } 1021 1022 void remove_migration_state_change_notifier(Notifier *notify) 1023 { 1024 notifier_remove(notify); 1025 } 1026 1027 bool migration_in_setup(MigrationState *s) 1028 { 1029 return s->state == MIGRATION_STATUS_SETUP; 1030 } 1031 1032 bool migration_has_finished(MigrationState *s) 1033 { 1034 return s->state == MIGRATION_STATUS_COMPLETED; 1035 } 1036 1037 bool migration_has_failed(MigrationState *s) 1038 { 1039 return (s->state == MIGRATION_STATUS_CANCELLED || 1040 s->state == MIGRATION_STATUS_FAILED); 1041 } 1042 1043 bool migration_in_postcopy(MigrationState *s) 1044 { 1045 return (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE); 1046 } 1047 1048 bool migration_in_postcopy_after_devices(MigrationState *s) 1049 { 1050 return migration_in_postcopy(s) && s->postcopy_after_devices; 1051 } 1052 1053 bool migration_is_idle(MigrationState *s) 1054 { 1055 if (!s) { 1056 s = migrate_get_current(); 1057 } 1058 1059 switch (s->state) { 1060 case MIGRATION_STATUS_NONE: 1061 case MIGRATION_STATUS_CANCELLED: 1062 case MIGRATION_STATUS_COMPLETED: 1063 case MIGRATION_STATUS_FAILED: 1064 return true; 1065 case MIGRATION_STATUS_SETUP: 1066 case MIGRATION_STATUS_CANCELLING: 1067 case MIGRATION_STATUS_ACTIVE: 1068 case MIGRATION_STATUS_POSTCOPY_ACTIVE: 1069 case MIGRATION_STATUS_COLO: 1070 return false; 1071 case MIGRATION_STATUS__MAX: 1072 g_assert_not_reached(); 1073 } 1074 1075 return false; 1076 } 1077 1078 MigrationState *migrate_init(const MigrationParams *params) 1079 { 1080 MigrationState *s = migrate_get_current(); 1081 1082 /* 1083 * Reinitialise all migration state, except 1084 * parameters/capabilities that the user set, and 1085 * locks. 1086 */ 1087 s->bytes_xfer = 0; 1088 s->xfer_limit = 0; 1089 s->cleanup_bh = 0; 1090 s->to_dst_file = NULL; 1091 s->state = MIGRATION_STATUS_NONE; 1092 s->params = *params; 1093 s->rp_state.from_dst_file = NULL; 1094 s->rp_state.error = false; 1095 s->mbps = 0.0; 1096 s->downtime = 0; 1097 s->expected_downtime = 0; 1098 s->dirty_pages_rate = 0; 1099 s->dirty_bytes_rate = 0; 1100 s->setup_time = 0; 1101 s->dirty_sync_count = 0; 1102 s->start_postcopy = false; 1103 s->postcopy_after_devices = false; 1104 s->postcopy_requests = 0; 1105 s->migration_thread_running = false; 1106 s->last_req_rb = NULL; 1107 error_free(s->error); 1108 s->error = NULL; 1109 1110 migrate_set_state(&s->state, MIGRATION_STATUS_NONE, MIGRATION_STATUS_SETUP); 1111 1112 QSIMPLEQ_INIT(&s->src_page_requests); 1113 1114 s->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1115 return s; 1116 } 1117 1118 static GSList *migration_blockers; 1119 1120 int migrate_add_blocker(Error *reason, Error **errp) 1121 { 1122 if (only_migratable) { 1123 error_propagate(errp, error_copy(reason)); 1124 error_prepend(errp, "disallowing migration blocker " 1125 "(--only_migratable) for: "); 1126 return -EACCES; 1127 } 1128 1129 if (migration_is_idle(NULL)) { 1130 migration_blockers = g_slist_prepend(migration_blockers, reason); 1131 return 0; 1132 } 1133 1134 error_propagate(errp, error_copy(reason)); 1135 error_prepend(errp, "disallowing migration blocker (migration in " 1136 "progress) for: "); 1137 return -EBUSY; 1138 } 1139 1140 void migrate_del_blocker(Error *reason) 1141 { 1142 migration_blockers = g_slist_remove(migration_blockers, reason); 1143 } 1144 1145 void qmp_migrate_incoming(const char *uri, Error **errp) 1146 { 1147 Error *local_err = NULL; 1148 static bool once = true; 1149 1150 if (!deferred_incoming) { 1151 error_setg(errp, "For use with '-incoming defer'"); 1152 return; 1153 } 1154 if (!once) { 1155 error_setg(errp, "The incoming migration has already been started"); 1156 } 1157 1158 qemu_start_incoming_migration(uri, &local_err); 1159 1160 if (local_err) { 1161 error_propagate(errp, local_err); 1162 return; 1163 } 1164 1165 once = false; 1166 } 1167 1168 bool migration_is_blocked(Error **errp) 1169 { 1170 if (qemu_savevm_state_blocked(errp)) { 1171 return true; 1172 } 1173 1174 if (migration_blockers) { 1175 *errp = error_copy(migration_blockers->data); 1176 return true; 1177 } 1178 1179 return false; 1180 } 1181 1182 void qmp_migrate(const char *uri, bool has_blk, bool blk, 1183 bool has_inc, bool inc, bool has_detach, bool detach, 1184 Error **errp) 1185 { 1186 Error *local_err = NULL; 1187 MigrationState *s = migrate_get_current(); 1188 MigrationParams params; 1189 const char *p; 1190 1191 params.blk = has_blk && blk; 1192 params.shared = has_inc && inc; 1193 1194 if (migration_is_setup_or_active(s->state) || 1195 s->state == MIGRATION_STATUS_CANCELLING || 1196 s->state == MIGRATION_STATUS_COLO) { 1197 error_setg(errp, QERR_MIGRATION_ACTIVE); 1198 return; 1199 } 1200 if (runstate_check(RUN_STATE_INMIGRATE)) { 1201 error_setg(errp, "Guest is waiting for an incoming migration"); 1202 return; 1203 } 1204 1205 if (migration_is_blocked(errp)) { 1206 return; 1207 } 1208 1209 s = migrate_init(¶ms); 1210 1211 if (strstart(uri, "tcp:", &p)) { 1212 tcp_start_outgoing_migration(s, p, &local_err); 1213 #ifdef CONFIG_RDMA 1214 } else if (strstart(uri, "rdma:", &p)) { 1215 rdma_start_outgoing_migration(s, p, &local_err); 1216 #endif 1217 } else if (strstart(uri, "exec:", &p)) { 1218 exec_start_outgoing_migration(s, p, &local_err); 1219 } else if (strstart(uri, "unix:", &p)) { 1220 unix_start_outgoing_migration(s, p, &local_err); 1221 } else if (strstart(uri, "fd:", &p)) { 1222 fd_start_outgoing_migration(s, p, &local_err); 1223 } else { 1224 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "uri", 1225 "a valid migration protocol"); 1226 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP, 1227 MIGRATION_STATUS_FAILED); 1228 return; 1229 } 1230 1231 if (local_err) { 1232 migrate_fd_error(s, local_err); 1233 error_propagate(errp, local_err); 1234 return; 1235 } 1236 } 1237 1238 void qmp_migrate_cancel(Error **errp) 1239 { 1240 migrate_fd_cancel(migrate_get_current()); 1241 } 1242 1243 void qmp_migrate_set_cache_size(int64_t value, Error **errp) 1244 { 1245 MigrationState *s = migrate_get_current(); 1246 int64_t new_size; 1247 1248 /* Check for truncation */ 1249 if (value != (size_t)value) { 1250 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", 1251 "exceeding address space"); 1252 return; 1253 } 1254 1255 /* Cache should not be larger than guest ram size */ 1256 if (value > ram_bytes_total()) { 1257 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", 1258 "exceeds guest ram size "); 1259 return; 1260 } 1261 1262 new_size = xbzrle_cache_resize(value); 1263 if (new_size < 0) { 1264 error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", 1265 "is smaller than page size"); 1266 return; 1267 } 1268 1269 s->xbzrle_cache_size = new_size; 1270 } 1271 1272 int64_t qmp_query_migrate_cache_size(Error **errp) 1273 { 1274 return migrate_xbzrle_cache_size(); 1275 } 1276 1277 void qmp_migrate_set_speed(int64_t value, Error **errp) 1278 { 1279 MigrationParameters p = { 1280 .has_max_bandwidth = true, 1281 .max_bandwidth = value, 1282 }; 1283 1284 qmp_migrate_set_parameters(&p, errp); 1285 } 1286 1287 void qmp_migrate_set_downtime(double value, Error **errp) 1288 { 1289 value *= 1000; /* Convert to milliseconds */ 1290 value = MAX(0, MIN(INT64_MAX, value)); 1291 1292 MigrationParameters p = { 1293 .has_downtime_limit = true, 1294 .downtime_limit = value, 1295 }; 1296 1297 qmp_migrate_set_parameters(&p, errp); 1298 } 1299 1300 bool migrate_postcopy_ram(void) 1301 { 1302 MigrationState *s; 1303 1304 s = migrate_get_current(); 1305 1306 return s->enabled_capabilities[MIGRATION_CAPABILITY_POSTCOPY_RAM]; 1307 } 1308 1309 bool migrate_auto_converge(void) 1310 { 1311 MigrationState *s; 1312 1313 s = migrate_get_current(); 1314 1315 return s->enabled_capabilities[MIGRATION_CAPABILITY_AUTO_CONVERGE]; 1316 } 1317 1318 bool migrate_zero_blocks(void) 1319 { 1320 MigrationState *s; 1321 1322 s = migrate_get_current(); 1323 1324 return s->enabled_capabilities[MIGRATION_CAPABILITY_ZERO_BLOCKS]; 1325 } 1326 1327 bool migrate_use_compression(void) 1328 { 1329 MigrationState *s; 1330 1331 s = migrate_get_current(); 1332 1333 return s->enabled_capabilities[MIGRATION_CAPABILITY_COMPRESS]; 1334 } 1335 1336 int migrate_compress_level(void) 1337 { 1338 MigrationState *s; 1339 1340 s = migrate_get_current(); 1341 1342 return s->parameters.compress_level; 1343 } 1344 1345 int migrate_compress_threads(void) 1346 { 1347 MigrationState *s; 1348 1349 s = migrate_get_current(); 1350 1351 return s->parameters.compress_threads; 1352 } 1353 1354 int migrate_decompress_threads(void) 1355 { 1356 MigrationState *s; 1357 1358 s = migrate_get_current(); 1359 1360 return s->parameters.decompress_threads; 1361 } 1362 1363 bool migrate_use_events(void) 1364 { 1365 MigrationState *s; 1366 1367 s = migrate_get_current(); 1368 1369 return s->enabled_capabilities[MIGRATION_CAPABILITY_EVENTS]; 1370 } 1371 1372 int migrate_use_xbzrle(void) 1373 { 1374 MigrationState *s; 1375 1376 s = migrate_get_current(); 1377 1378 return s->enabled_capabilities[MIGRATION_CAPABILITY_XBZRLE]; 1379 } 1380 1381 int64_t migrate_xbzrle_cache_size(void) 1382 { 1383 MigrationState *s; 1384 1385 s = migrate_get_current(); 1386 1387 return s->xbzrle_cache_size; 1388 } 1389 1390 /* migration thread support */ 1391 /* 1392 * Something bad happened to the RP stream, mark an error 1393 * The caller shall print or trace something to indicate why 1394 */ 1395 static void mark_source_rp_bad(MigrationState *s) 1396 { 1397 s->rp_state.error = true; 1398 } 1399 1400 static struct rp_cmd_args { 1401 ssize_t len; /* -1 = variable */ 1402 const char *name; 1403 } rp_cmd_args[] = { 1404 [MIG_RP_MSG_INVALID] = { .len = -1, .name = "INVALID" }, 1405 [MIG_RP_MSG_SHUT] = { .len = 4, .name = "SHUT" }, 1406 [MIG_RP_MSG_PONG] = { .len = 4, .name = "PONG" }, 1407 [MIG_RP_MSG_REQ_PAGES] = { .len = 12, .name = "REQ_PAGES" }, 1408 [MIG_RP_MSG_REQ_PAGES_ID] = { .len = -1, .name = "REQ_PAGES_ID" }, 1409 [MIG_RP_MSG_MAX] = { .len = -1, .name = "MAX" }, 1410 }; 1411 1412 /* 1413 * Process a request for pages received on the return path, 1414 * We're allowed to send more than requested (e.g. to round to our page size) 1415 * and we don't need to send pages that have already been sent. 1416 */ 1417 static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname, 1418 ram_addr_t start, size_t len) 1419 { 1420 long our_host_ps = getpagesize(); 1421 1422 trace_migrate_handle_rp_req_pages(rbname, start, len); 1423 1424 /* 1425 * Since we currently insist on matching page sizes, just sanity check 1426 * we're being asked for whole host pages. 1427 */ 1428 if (start & (our_host_ps-1) || 1429 (len & (our_host_ps-1))) { 1430 error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT 1431 " len: %zd", __func__, start, len); 1432 mark_source_rp_bad(ms); 1433 return; 1434 } 1435 1436 if (ram_save_queue_pages(ms, rbname, start, len)) { 1437 mark_source_rp_bad(ms); 1438 } 1439 } 1440 1441 /* 1442 * Handles messages sent on the return path towards the source VM 1443 * 1444 */ 1445 static void *source_return_path_thread(void *opaque) 1446 { 1447 MigrationState *ms = opaque; 1448 QEMUFile *rp = ms->rp_state.from_dst_file; 1449 uint16_t header_len, header_type; 1450 uint8_t buf[512]; 1451 uint32_t tmp32, sibling_error; 1452 ram_addr_t start = 0; /* =0 to silence warning */ 1453 size_t len = 0, expected_len; 1454 int res; 1455 1456 trace_source_return_path_thread_entry(); 1457 while (!ms->rp_state.error && !qemu_file_get_error(rp) && 1458 migration_is_setup_or_active(ms->state)) { 1459 trace_source_return_path_thread_loop_top(); 1460 header_type = qemu_get_be16(rp); 1461 header_len = qemu_get_be16(rp); 1462 1463 if (header_type >= MIG_RP_MSG_MAX || 1464 header_type == MIG_RP_MSG_INVALID) { 1465 error_report("RP: Received invalid message 0x%04x length 0x%04x", 1466 header_type, header_len); 1467 mark_source_rp_bad(ms); 1468 goto out; 1469 } 1470 1471 if ((rp_cmd_args[header_type].len != -1 && 1472 header_len != rp_cmd_args[header_type].len) || 1473 header_len > sizeof(buf)) { 1474 error_report("RP: Received '%s' message (0x%04x) with" 1475 "incorrect length %d expecting %zu", 1476 rp_cmd_args[header_type].name, header_type, header_len, 1477 (size_t)rp_cmd_args[header_type].len); 1478 mark_source_rp_bad(ms); 1479 goto out; 1480 } 1481 1482 /* We know we've got a valid header by this point */ 1483 res = qemu_get_buffer(rp, buf, header_len); 1484 if (res != header_len) { 1485 error_report("RP: Failed reading data for message 0x%04x" 1486 " read %d expected %d", 1487 header_type, res, header_len); 1488 mark_source_rp_bad(ms); 1489 goto out; 1490 } 1491 1492 /* OK, we have the message and the data */ 1493 switch (header_type) { 1494 case MIG_RP_MSG_SHUT: 1495 sibling_error = ldl_be_p(buf); 1496 trace_source_return_path_thread_shut(sibling_error); 1497 if (sibling_error) { 1498 error_report("RP: Sibling indicated error %d", sibling_error); 1499 mark_source_rp_bad(ms); 1500 } 1501 /* 1502 * We'll let the main thread deal with closing the RP 1503 * we could do a shutdown(2) on it, but we're the only user 1504 * anyway, so there's nothing gained. 1505 */ 1506 goto out; 1507 1508 case MIG_RP_MSG_PONG: 1509 tmp32 = ldl_be_p(buf); 1510 trace_source_return_path_thread_pong(tmp32); 1511 break; 1512 1513 case MIG_RP_MSG_REQ_PAGES: 1514 start = ldq_be_p(buf); 1515 len = ldl_be_p(buf + 8); 1516 migrate_handle_rp_req_pages(ms, NULL, start, len); 1517 break; 1518 1519 case MIG_RP_MSG_REQ_PAGES_ID: 1520 expected_len = 12 + 1; /* header + termination */ 1521 1522 if (header_len >= expected_len) { 1523 start = ldq_be_p(buf); 1524 len = ldl_be_p(buf + 8); 1525 /* Now we expect an idstr */ 1526 tmp32 = buf[12]; /* Length of the following idstr */ 1527 buf[13 + tmp32] = '\0'; 1528 expected_len += tmp32; 1529 } 1530 if (header_len != expected_len) { 1531 error_report("RP: Req_Page_id with length %d expecting %zd", 1532 header_len, expected_len); 1533 mark_source_rp_bad(ms); 1534 goto out; 1535 } 1536 migrate_handle_rp_req_pages(ms, (char *)&buf[13], start, len); 1537 break; 1538 1539 default: 1540 break; 1541 } 1542 } 1543 if (qemu_file_get_error(rp)) { 1544 trace_source_return_path_thread_bad_end(); 1545 mark_source_rp_bad(ms); 1546 } 1547 1548 trace_source_return_path_thread_end(); 1549 out: 1550 ms->rp_state.from_dst_file = NULL; 1551 qemu_fclose(rp); 1552 return NULL; 1553 } 1554 1555 static int open_return_path_on_source(MigrationState *ms) 1556 { 1557 1558 ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file); 1559 if (!ms->rp_state.from_dst_file) { 1560 return -1; 1561 } 1562 1563 trace_open_return_path_on_source(); 1564 qemu_thread_create(&ms->rp_state.rp_thread, "return path", 1565 source_return_path_thread, ms, QEMU_THREAD_JOINABLE); 1566 1567 trace_open_return_path_on_source_continue(); 1568 1569 return 0; 1570 } 1571 1572 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */ 1573 static int await_return_path_close_on_source(MigrationState *ms) 1574 { 1575 /* 1576 * If this is a normal exit then the destination will send a SHUT and the 1577 * rp_thread will exit, however if there's an error we need to cause 1578 * it to exit. 1579 */ 1580 if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) { 1581 /* 1582 * shutdown(2), if we have it, will cause it to unblock if it's stuck 1583 * waiting for the destination. 1584 */ 1585 qemu_file_shutdown(ms->rp_state.from_dst_file); 1586 mark_source_rp_bad(ms); 1587 } 1588 trace_await_return_path_close_on_source_joining(); 1589 qemu_thread_join(&ms->rp_state.rp_thread); 1590 trace_await_return_path_close_on_source_close(); 1591 return ms->rp_state.error; 1592 } 1593 1594 /* 1595 * Switch from normal iteration to postcopy 1596 * Returns non-0 on error 1597 */ 1598 static int postcopy_start(MigrationState *ms, bool *old_vm_running) 1599 { 1600 int ret; 1601 QIOChannelBuffer *bioc; 1602 QEMUFile *fb; 1603 int64_t time_at_stop = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1604 migrate_set_state(&ms->state, MIGRATION_STATUS_ACTIVE, 1605 MIGRATION_STATUS_POSTCOPY_ACTIVE); 1606 1607 trace_postcopy_start(); 1608 qemu_mutex_lock_iothread(); 1609 trace_postcopy_start_set_run(); 1610 1611 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER); 1612 *old_vm_running = runstate_is_running(); 1613 global_state_store(); 1614 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE); 1615 if (ret < 0) { 1616 goto fail; 1617 } 1618 1619 ret = bdrv_inactivate_all(); 1620 if (ret < 0) { 1621 goto fail; 1622 } 1623 1624 /* 1625 * Cause any non-postcopiable, but iterative devices to 1626 * send out their final data. 1627 */ 1628 qemu_savevm_state_complete_precopy(ms->to_dst_file, true); 1629 1630 /* 1631 * in Finish migrate and with the io-lock held everything should 1632 * be quiet, but we've potentially still got dirty pages and we 1633 * need to tell the destination to throw any pages it's already received 1634 * that are dirty 1635 */ 1636 if (ram_postcopy_send_discard_bitmap(ms)) { 1637 error_report("postcopy send discard bitmap failed"); 1638 goto fail; 1639 } 1640 1641 /* 1642 * send rest of state - note things that are doing postcopy 1643 * will notice we're in POSTCOPY_ACTIVE and not actually 1644 * wrap their state up here 1645 */ 1646 qemu_file_set_rate_limit(ms->to_dst_file, INT64_MAX); 1647 /* Ping just for debugging, helps line traces up */ 1648 qemu_savevm_send_ping(ms->to_dst_file, 2); 1649 1650 /* 1651 * While loading the device state we may trigger page transfer 1652 * requests and the fd must be free to process those, and thus 1653 * the destination must read the whole device state off the fd before 1654 * it starts processing it. Unfortunately the ad-hoc migration format 1655 * doesn't allow the destination to know the size to read without fully 1656 * parsing it through each devices load-state code (especially the open 1657 * coded devices that use get/put). 1658 * So we wrap the device state up in a package with a length at the start; 1659 * to do this we use a qemu_buf to hold the whole of the device state. 1660 */ 1661 bioc = qio_channel_buffer_new(4096); 1662 qio_channel_set_name(QIO_CHANNEL(bioc), "migration-postcopy-buffer"); 1663 fb = qemu_fopen_channel_output(QIO_CHANNEL(bioc)); 1664 object_unref(OBJECT(bioc)); 1665 1666 /* 1667 * Make sure the receiver can get incoming pages before we send the rest 1668 * of the state 1669 */ 1670 qemu_savevm_send_postcopy_listen(fb); 1671 1672 qemu_savevm_state_complete_precopy(fb, false); 1673 qemu_savevm_send_ping(fb, 3); 1674 1675 qemu_savevm_send_postcopy_run(fb); 1676 1677 /* <><> end of stuff going into the package */ 1678 1679 /* Now send that blob */ 1680 if (qemu_savevm_send_packaged(ms->to_dst_file, bioc->data, bioc->usage)) { 1681 goto fail_closefb; 1682 } 1683 qemu_fclose(fb); 1684 1685 /* Send a notify to give a chance for anything that needs to happen 1686 * at the transition to postcopy and after the device state; in particular 1687 * spice needs to trigger a transition now 1688 */ 1689 ms->postcopy_after_devices = true; 1690 notifier_list_notify(&migration_state_notifiers, ms); 1691 1692 ms->downtime = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop; 1693 1694 qemu_mutex_unlock_iothread(); 1695 1696 /* 1697 * Although this ping is just for debug, it could potentially be 1698 * used for getting a better measurement of downtime at the source. 1699 */ 1700 qemu_savevm_send_ping(ms->to_dst_file, 4); 1701 1702 ret = qemu_file_get_error(ms->to_dst_file); 1703 if (ret) { 1704 error_report("postcopy_start: Migration stream errored"); 1705 migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE, 1706 MIGRATION_STATUS_FAILED); 1707 } 1708 1709 return ret; 1710 1711 fail_closefb: 1712 qemu_fclose(fb); 1713 fail: 1714 migrate_set_state(&ms->state, MIGRATION_STATUS_POSTCOPY_ACTIVE, 1715 MIGRATION_STATUS_FAILED); 1716 qemu_mutex_unlock_iothread(); 1717 return -1; 1718 } 1719 1720 /** 1721 * migration_completion: Used by migration_thread when there's not much left. 1722 * The caller 'breaks' the loop when this returns. 1723 * 1724 * @s: Current migration state 1725 * @current_active_state: The migration state we expect to be in 1726 * @*old_vm_running: Pointer to old_vm_running flag 1727 * @*start_time: Pointer to time to update 1728 */ 1729 static void migration_completion(MigrationState *s, int current_active_state, 1730 bool *old_vm_running, 1731 int64_t *start_time) 1732 { 1733 int ret; 1734 1735 if (s->state == MIGRATION_STATUS_ACTIVE) { 1736 qemu_mutex_lock_iothread(); 1737 *start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1738 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER); 1739 *old_vm_running = runstate_is_running(); 1740 ret = global_state_store(); 1741 1742 if (!ret) { 1743 ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE); 1744 /* 1745 * Don't mark the image with BDRV_O_INACTIVE flag if 1746 * we will go into COLO stage later. 1747 */ 1748 if (ret >= 0 && !migrate_colo_enabled()) { 1749 ret = bdrv_inactivate_all(); 1750 } 1751 if (ret >= 0) { 1752 qemu_file_set_rate_limit(s->to_dst_file, INT64_MAX); 1753 qemu_savevm_state_complete_precopy(s->to_dst_file, false); 1754 s->block_inactive = true; 1755 } 1756 } 1757 qemu_mutex_unlock_iothread(); 1758 1759 if (ret < 0) { 1760 goto fail; 1761 } 1762 } else if (s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) { 1763 trace_migration_completion_postcopy_end(); 1764 1765 qemu_savevm_state_complete_postcopy(s->to_dst_file); 1766 trace_migration_completion_postcopy_end_after_complete(); 1767 } 1768 1769 /* 1770 * If rp was opened we must clean up the thread before 1771 * cleaning everything else up (since if there are no failures 1772 * it will wait for the destination to send it's status in 1773 * a SHUT command). 1774 * Postcopy opens rp if enabled (even if it's not avtivated) 1775 */ 1776 if (migrate_postcopy_ram()) { 1777 int rp_error; 1778 trace_migration_completion_postcopy_end_before_rp(); 1779 rp_error = await_return_path_close_on_source(s); 1780 trace_migration_completion_postcopy_end_after_rp(rp_error); 1781 if (rp_error) { 1782 goto fail_invalidate; 1783 } 1784 } 1785 1786 if (qemu_file_get_error(s->to_dst_file)) { 1787 trace_migration_completion_file_err(); 1788 goto fail_invalidate; 1789 } 1790 1791 if (!migrate_colo_enabled()) { 1792 migrate_set_state(&s->state, current_active_state, 1793 MIGRATION_STATUS_COMPLETED); 1794 } 1795 1796 return; 1797 1798 fail_invalidate: 1799 /* If not doing postcopy, vm_start() will be called: let's regain 1800 * control on images. 1801 */ 1802 if (s->state == MIGRATION_STATUS_ACTIVE) { 1803 Error *local_err = NULL; 1804 1805 qemu_mutex_lock_iothread(); 1806 bdrv_invalidate_cache_all(&local_err); 1807 if (local_err) { 1808 error_report_err(local_err); 1809 } else { 1810 s->block_inactive = false; 1811 } 1812 qemu_mutex_unlock_iothread(); 1813 } 1814 1815 fail: 1816 migrate_set_state(&s->state, current_active_state, 1817 MIGRATION_STATUS_FAILED); 1818 } 1819 1820 bool migrate_colo_enabled(void) 1821 { 1822 MigrationState *s = migrate_get_current(); 1823 return s->enabled_capabilities[MIGRATION_CAPABILITY_X_COLO]; 1824 } 1825 1826 /* 1827 * Master migration thread on the source VM. 1828 * It drives the migration and pumps the data down the outgoing channel. 1829 */ 1830 static void *migration_thread(void *opaque) 1831 { 1832 MigrationState *s = opaque; 1833 /* Used by the bandwidth calcs, updated later */ 1834 int64_t initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1835 int64_t setup_start = qemu_clock_get_ms(QEMU_CLOCK_HOST); 1836 int64_t initial_bytes = 0; 1837 int64_t max_size = 0; 1838 int64_t start_time = initial_time; 1839 int64_t end_time; 1840 bool old_vm_running = false; 1841 bool entered_postcopy = false; 1842 /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */ 1843 enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE; 1844 bool enable_colo = migrate_colo_enabled(); 1845 1846 rcu_register_thread(); 1847 1848 qemu_savevm_state_header(s->to_dst_file); 1849 1850 if (migrate_postcopy_ram()) { 1851 /* Now tell the dest that it should open its end so it can reply */ 1852 qemu_savevm_send_open_return_path(s->to_dst_file); 1853 1854 /* And do a ping that will make stuff easier to debug */ 1855 qemu_savevm_send_ping(s->to_dst_file, 1); 1856 1857 /* 1858 * Tell the destination that we *might* want to do postcopy later; 1859 * if the other end can't do postcopy it should fail now, nice and 1860 * early. 1861 */ 1862 qemu_savevm_send_postcopy_advise(s->to_dst_file); 1863 } 1864 1865 qemu_savevm_state_begin(s->to_dst_file, &s->params); 1866 1867 s->setup_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) - setup_start; 1868 current_active_state = MIGRATION_STATUS_ACTIVE; 1869 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP, 1870 MIGRATION_STATUS_ACTIVE); 1871 1872 trace_migration_thread_setup_complete(); 1873 1874 while (s->state == MIGRATION_STATUS_ACTIVE || 1875 s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) { 1876 int64_t current_time; 1877 uint64_t pending_size; 1878 1879 if (!qemu_file_rate_limit(s->to_dst_file)) { 1880 uint64_t pend_post, pend_nonpost; 1881 1882 qemu_savevm_state_pending(s->to_dst_file, max_size, &pend_nonpost, 1883 &pend_post); 1884 pending_size = pend_nonpost + pend_post; 1885 trace_migrate_pending(pending_size, max_size, 1886 pend_post, pend_nonpost); 1887 if (pending_size && pending_size >= max_size) { 1888 /* Still a significant amount to transfer */ 1889 1890 if (migrate_postcopy_ram() && 1891 s->state != MIGRATION_STATUS_POSTCOPY_ACTIVE && 1892 pend_nonpost <= max_size && 1893 atomic_read(&s->start_postcopy)) { 1894 1895 if (!postcopy_start(s, &old_vm_running)) { 1896 current_active_state = MIGRATION_STATUS_POSTCOPY_ACTIVE; 1897 entered_postcopy = true; 1898 } 1899 1900 continue; 1901 } 1902 /* Just another iteration step */ 1903 qemu_savevm_state_iterate(s->to_dst_file, entered_postcopy); 1904 } else { 1905 trace_migration_thread_low_pending(pending_size); 1906 migration_completion(s, current_active_state, 1907 &old_vm_running, &start_time); 1908 break; 1909 } 1910 } 1911 1912 if (qemu_file_get_error(s->to_dst_file)) { 1913 migrate_set_state(&s->state, current_active_state, 1914 MIGRATION_STATUS_FAILED); 1915 trace_migration_thread_file_err(); 1916 break; 1917 } 1918 current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1919 if (current_time >= initial_time + BUFFER_DELAY) { 1920 uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) - 1921 initial_bytes; 1922 uint64_t time_spent = current_time - initial_time; 1923 double bandwidth = (double)transferred_bytes / time_spent; 1924 max_size = bandwidth * s->parameters.downtime_limit; 1925 1926 s->mbps = (((double) transferred_bytes * 8.0) / 1927 ((double) time_spent / 1000.0)) / 1000.0 / 1000.0; 1928 1929 trace_migrate_transferred(transferred_bytes, time_spent, 1930 bandwidth, max_size); 1931 /* if we haven't sent anything, we don't want to recalculate 1932 10000 is a small enough number for our purposes */ 1933 if (s->dirty_bytes_rate && transferred_bytes > 10000) { 1934 s->expected_downtime = s->dirty_bytes_rate / bandwidth; 1935 } 1936 1937 qemu_file_reset_rate_limit(s->to_dst_file); 1938 initial_time = current_time; 1939 initial_bytes = qemu_ftell(s->to_dst_file); 1940 } 1941 if (qemu_file_rate_limit(s->to_dst_file)) { 1942 /* usleep expects microseconds */ 1943 g_usleep((initial_time + BUFFER_DELAY - current_time)*1000); 1944 } 1945 } 1946 1947 trace_migration_thread_after_loop(); 1948 /* If we enabled cpu throttling for auto-converge, turn it off. */ 1949 cpu_throttle_stop(); 1950 end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 1951 1952 qemu_mutex_lock_iothread(); 1953 /* 1954 * The resource has been allocated by migration will be reused in COLO 1955 * process, so don't release them. 1956 */ 1957 if (!enable_colo) { 1958 qemu_savevm_state_cleanup(); 1959 } 1960 if (s->state == MIGRATION_STATUS_COMPLETED) { 1961 uint64_t transferred_bytes = qemu_ftell(s->to_dst_file); 1962 s->total_time = end_time - s->total_time; 1963 if (!entered_postcopy) { 1964 s->downtime = end_time - start_time; 1965 } 1966 if (s->total_time) { 1967 s->mbps = (((double) transferred_bytes * 8.0) / 1968 ((double) s->total_time)) / 1000; 1969 } 1970 runstate_set(RUN_STATE_POSTMIGRATE); 1971 } else { 1972 if (s->state == MIGRATION_STATUS_ACTIVE && enable_colo) { 1973 migrate_start_colo_process(s); 1974 qemu_savevm_state_cleanup(); 1975 /* 1976 * Fixme: we will run VM in COLO no matter its old running state. 1977 * After exited COLO, we will keep running. 1978 */ 1979 old_vm_running = true; 1980 } 1981 if (old_vm_running && !entered_postcopy) { 1982 vm_start(); 1983 } else { 1984 if (runstate_check(RUN_STATE_FINISH_MIGRATE)) { 1985 runstate_set(RUN_STATE_POSTMIGRATE); 1986 } 1987 } 1988 } 1989 qemu_bh_schedule(s->cleanup_bh); 1990 qemu_mutex_unlock_iothread(); 1991 1992 rcu_unregister_thread(); 1993 return NULL; 1994 } 1995 1996 void migrate_fd_connect(MigrationState *s) 1997 { 1998 s->expected_downtime = s->parameters.downtime_limit; 1999 s->cleanup_bh = qemu_bh_new(migrate_fd_cleanup, s); 2000 2001 qemu_file_set_blocking(s->to_dst_file, true); 2002 qemu_file_set_rate_limit(s->to_dst_file, 2003 s->parameters.max_bandwidth / XFER_LIMIT_RATIO); 2004 2005 /* Notify before starting migration thread */ 2006 notifier_list_notify(&migration_state_notifiers, s); 2007 2008 /* 2009 * Open the return path; currently for postcopy but other things might 2010 * also want it. 2011 */ 2012 if (migrate_postcopy_ram()) { 2013 if (open_return_path_on_source(s)) { 2014 error_report("Unable to open return-path for postcopy"); 2015 migrate_set_state(&s->state, MIGRATION_STATUS_SETUP, 2016 MIGRATION_STATUS_FAILED); 2017 migrate_fd_cleanup(s); 2018 return; 2019 } 2020 } 2021 2022 migrate_compress_threads_create(); 2023 qemu_thread_create(&s->thread, "live_migration", migration_thread, s, 2024 QEMU_THREAD_JOINABLE); 2025 s->migration_thread_running = true; 2026 } 2027 2028 PostcopyState postcopy_state_get(void) 2029 { 2030 return atomic_mb_read(&incoming_postcopy_state); 2031 } 2032 2033 /* Set the state and return the old state */ 2034 PostcopyState postcopy_state_set(PostcopyState new_state) 2035 { 2036 return atomic_xchg(&incoming_postcopy_state, new_state); 2037 } 2038 2039