migration/colo.c

   1 /*
   2  * COarse-grain LOck-stepping Virtual Machines for Non-stop Service (COLO)
   3  * (a.k.a. Fault Tolerance or Continuous Replication)
   4  *
   5  * Copyright (c) 2016 HUAWEI TECHNOLOGIES CO., LTD.
   6  * Copyright (c) 2016 FUJITSU LIMITED
   7  * Copyright (c) 2016 Intel Corporation
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or
  10  * later.  See the COPYING file in the top-level directory.
  11  */
  12
  13 #include "qemu/osdep.h"
  14 #include "sysemu/sysemu.h"
  15 #include "qapi/error.h"
  16 #include "qapi/qapi-commands-migration.h"
  17 #include "migration.h"
  18 #include "qemu-file.h"
  19 #include "savevm.h"
  20 #include "migration/colo.h"
  21 #include "block.h"
  22 #include "io/channel-buffer.h"
  23 #include "trace.h"
  24 #include "qemu/error-report.h"
  25 #include "qemu/main-loop.h"
  26 #include "qemu/rcu.h"
  27 #include "migration/failover.h"
  28 #include "migration/ram.h"
  29 #include "block/replication.h"
  30 #include "net/colo-compare.h"
  31 #include "net/colo.h"
  32 #include "block/block.h"
  33 #include "qapi/qapi-events-migration.h"
  34 #include "sysemu/cpus.h"
  35 #include "sysemu/runstate.h"
  36 #include "net/filter.h"
  37 #include "options.h"
  38
  39 static bool vmstate_loading;
  40 static Notifier packets_compare_notifier;
  41
  42 /* User need to know colo mode after COLO failover */
  43 static COLOMode last_colo_mode;
  44
  45 #define COLO_BUFFER_BASE_SIZE (4 * 1024 * 1024)
  46
  47 bool migration_in_colo_state(void)
  48 {
  49     MigrationState *s = migrate_get_current();
  50
  51     return (s->state == MIGRATION_STATUS_COLO);
  52 }
  53
  54 bool migration_incoming_in_colo_state(void)
  55 {
  56     MigrationIncomingState *mis = migration_incoming_get_current();
  57
  58     return mis && (mis->state == MIGRATION_STATUS_COLO);
  59 }
  60
  61 static bool colo_runstate_is_stopped(void)
  62 {
  63     return runstate_check(RUN_STATE_COLO) || !runstate_is_running();
  64 }
  65
  66 static void colo_checkpoint_notify(void *opaque)
  67 {
  68     MigrationState *s = opaque;
  69     int64_t next_notify_time;
  70
  71     qemu_event_set(&s->colo_checkpoint_event);
  72     s->colo_checkpoint_time = qemu_clock_get_ms(QEMU_CLOCK_HOST);
  73     next_notify_time = s->colo_checkpoint_time + migrate_checkpoint_delay();
  74     timer_mod(s->colo_delay_timer, next_notify_time);
  75 }
  76
  77 void colo_checkpoint_delay_set(void)
  78 {
  79     if (migration_in_colo_state()) {
  80         colo_checkpoint_notify(migrate_get_current());
  81     }
  82 }
  83
  84 static void secondary_vm_do_failover(void)
  85 {
  86 /* COLO needs enable block-replication */
  87     int old_state;
  88     MigrationIncomingState *mis = migration_incoming_get_current();
  89     Error *local_err = NULL;
  90
  91     /* Can not do failover during the process of VM's loading VMstate, Or
  92      * it will break the secondary VM.
  93      */
  94     if (vmstate_loading) {
  95         old_state = failover_set_state(FAILOVER_STATUS_ACTIVE,
  96                         FAILOVER_STATUS_RELAUNCH);
  97         if (old_state != FAILOVER_STATUS_ACTIVE) {
  98             error_report("Unknown error while do failover for secondary VM,"
  99                          "old_state: %s", FailoverStatus_str(old_state));
 100         }
 101         return;
 102     }
 103
 104     migrate_set_state(&mis->state, MIGRATION_STATUS_COLO,
 105                       MIGRATION_STATUS_COMPLETED);
 106
 107     replication_stop_all(true, &local_err);
 108     if (local_err) {
 109         error_report_err(local_err);
 110         local_err = NULL;
 111     }
 112
 113     /* Notify all filters of all NIC to do checkpoint */
 114     colo_notify_filters_event(COLO_EVENT_FAILOVER, &local_err);
 115     if (local_err) {
 116         error_report_err(local_err);
 117     }
 118
 119     if (!autostart) {
 120         error_report("\"-S\" qemu option will be ignored in secondary side");
 121         /* recover runstate to normal migration finish state */
 122         autostart = true;
 123     }
 124     /*
 125      * Make sure COLO incoming thread not block in recv or send,
 126      * If mis->from_src_file and mis->to_src_file use the same fd,
 127      * The second shutdown() will return -1, we ignore this value,
 128      * It is harmless.
 129      */
 130     if (mis->from_src_file) {
 131         qemu_file_shutdown(mis->from_src_file);
 132     }
 133     if (mis->to_src_file) {
 134         qemu_file_shutdown(mis->to_src_file);
 135     }
 136
 137     old_state = failover_set_state(FAILOVER_STATUS_ACTIVE,
 138                                    FAILOVER_STATUS_COMPLETED);
 139     if (old_state != FAILOVER_STATUS_ACTIVE) {
 140         error_report("Incorrect state (%s) while doing failover for "
 141                      "secondary VM", FailoverStatus_str(old_state));
 142         return;
 143     }
 144     /* Notify COLO incoming thread that failover work is finished */
 145     qemu_sem_post(&mis->colo_incoming_sem);
 146
 147     /* For Secondary VM, jump to incoming co */
 148     if (mis->colo_incoming_co) {
 149         qemu_coroutine_enter(mis->colo_incoming_co);
 150     }
 151 }
 152
 153 static void primary_vm_do_failover(void)
 154 {
 155     MigrationState *s = migrate_get_current();
 156     int old_state;
 157     Error *local_err = NULL;
 158
 159     migrate_set_state(&s->state, MIGRATION_STATUS_COLO,
 160                       MIGRATION_STATUS_COMPLETED);
 161     /*
 162      * kick COLO thread which might wait at
 163      * qemu_sem_wait(&s->colo_checkpoint_sem).
 164      */
 165     colo_checkpoint_notify(s);
 166
 167     /*
 168      * Wake up COLO thread which may blocked in recv() or send(),
 169      * The s->rp_state.from_dst_file and s->to_dst_file may use the
 170      * same fd, but we still shutdown the fd for twice, it is harmless.
 171      */
 172     if (s->to_dst_file) {
 173         qemu_file_shutdown(s->to_dst_file);
 174     }
 175     if (s->rp_state.from_dst_file) {
 176         qemu_file_shutdown(s->rp_state.from_dst_file);
 177     }
 178
 179     old_state = failover_set_state(FAILOVER_STATUS_ACTIVE,
 180                                    FAILOVER_STATUS_COMPLETED);
 181     if (old_state != FAILOVER_STATUS_ACTIVE) {
 182         error_report("Incorrect state (%s) while doing failover for Primary VM",
 183                      FailoverStatus_str(old_state));
 184         return;
 185     }
 186
 187     replication_stop_all(true, &local_err);
 188     if (local_err) {
 189         error_report_err(local_err);
 190         local_err = NULL;
 191     }
 192
 193     /* Notify COLO thread that failover work is finished */
 194     qemu_sem_post(&s->colo_exit_sem);
 195 }
 196
 197 COLOMode get_colo_mode(void)
 198 {
 199     if (migration_in_colo_state()) {
 200         return COLO_MODE_PRIMARY;
 201     } else if (migration_incoming_in_colo_state()) {
 202         return COLO_MODE_SECONDARY;
 203     } else {
 204         return COLO_MODE_NONE;
 205     }
 206 }
 207
 208 void colo_do_failover(void)
 209 {
 210     /* Make sure VM stopped while failover happened. */
 211     if (!colo_runstate_is_stopped()) {
 212         vm_stop_force_state(RUN_STATE_COLO);
 213     }
 214
 215     switch (last_colo_mode = get_colo_mode()) {
 216     case COLO_MODE_PRIMARY:
 217         primary_vm_do_failover();
 218         break;
 219     case COLO_MODE_SECONDARY:
 220         secondary_vm_do_failover();
 221         break;
 222     default:
 223         error_report("colo_do_failover failed because the colo mode"
 224                      " could not be obtained");
 225     }
 226 }
 227
 228 void qmp_xen_set_replication(bool enable, bool primary,
 229                              bool has_failover, bool failover,
 230                              Error **errp)
 231 {
 232     ReplicationMode mode = primary ?
 233                            REPLICATION_MODE_PRIMARY :
 234                            REPLICATION_MODE_SECONDARY;
 235
 236     if (has_failover && enable) {
 237         error_setg(errp, "Parameter 'failover' is only for"
 238                    " stopping replication");
 239         return;
 240     }
 241
 242     if (enable) {
 243         replication_start_all(mode, errp);
 244     } else {
 245         if (!has_failover) {
 246             failover = NULL;
 247         }
 248         replication_stop_all(failover, failover ? NULL : errp);
 249     }
 250 }
 251
 252 ReplicationStatus *qmp_query_xen_replication_status(Error **errp)
 253 {
 254     Error *err = NULL;
 255     ReplicationStatus *s = g_new0(ReplicationStatus, 1);
 256
 257     replication_get_error_all(&err);
 258     if (err) {
 259         s->error = true;
 260         s->desc = g_strdup(error_get_pretty(err));
 261     } else {
 262         s->error = false;
 263     }
 264
 265     error_free(err);
 266     return s;
 267 }
 268
 269 void qmp_xen_colo_do_checkpoint(Error **errp)
 270 {
 271     Error *err = NULL;
 272
 273     replication_do_checkpoint_all(&err);
 274     if (err) {
 275         error_propagate(errp, err);
 276         return;
 277     }
 278     /* Notify all filters of all NIC to do checkpoint */
 279     colo_notify_filters_event(COLO_EVENT_CHECKPOINT, errp);
 280 }
 281
 282 COLOStatus *qmp_query_colo_status(Error **errp)
 283 {
 284     COLOStatus *s = g_new0(COLOStatus, 1);
 285
 286     s->mode = get_colo_mode();
 287     s->last_mode = last_colo_mode;
 288
 289     switch (failover_get_state()) {
 290     case FAILOVER_STATUS_NONE:
 291         s->reason = COLO_EXIT_REASON_NONE;
 292         break;
 293     case FAILOVER_STATUS_COMPLETED:
 294         s->reason = COLO_EXIT_REASON_REQUEST;
 295         break;
 296     default:
 297         if (migration_in_colo_state()) {
 298             s->reason = COLO_EXIT_REASON_PROCESSING;
 299         } else {
 300             s->reason = COLO_EXIT_REASON_ERROR;
 301         }
 302     }
 303
 304     return s;
 305 }
 306
 307 static void colo_send_message(QEMUFile *f, COLOMessage msg,
 308                               Error **errp)
 309 {
 310     int ret;
 311
 312     if (msg >= COLO_MESSAGE__MAX) {
 313         error_setg(errp, "%s: Invalid message", __func__);
 314         return;
 315     }
 316     qemu_put_be32(f, msg);
 317     ret = qemu_fflush(f);
 318     if (ret < 0) {
 319         error_setg_errno(errp, -ret, "Can't send COLO message");
 320     }
 321     trace_colo_send_message(COLOMessage_str(msg));
 322 }
 323
 324 static void colo_send_message_value(QEMUFile *f, COLOMessage msg,
 325                                     uint64_t value, Error **errp)
 326 {
 327     Error *local_err = NULL;
 328     int ret;
 329
 330     colo_send_message(f, msg, &local_err);
 331     if (local_err) {
 332         error_propagate(errp, local_err);
 333         return;
 334     }
 335     qemu_put_be64(f, value);
 336     ret = qemu_fflush(f);
 337     if (ret < 0) {
 338         error_setg_errno(errp, -ret, "Failed to send value for message:%s",
 339                          COLOMessage_str(msg));
 340     }
 341 }
 342
 343 static COLOMessage colo_receive_message(QEMUFile *f, Error **errp)
 344 {
 345     COLOMessage msg;
 346     int ret;
 347
 348     msg = qemu_get_be32(f);
 349     ret = qemu_file_get_error(f);
 350     if (ret < 0) {
 351         error_setg_errno(errp, -ret, "Can't receive COLO message");
 352         return msg;
 353     }
 354     if (msg >= COLO_MESSAGE__MAX) {
 355         error_setg(errp, "%s: Invalid message", __func__);
 356         return msg;
 357     }
 358     trace_colo_receive_message(COLOMessage_str(msg));
 359     return msg;
 360 }
 361
 362 static void colo_receive_check_message(QEMUFile *f, COLOMessage expect_msg,
 363                                        Error **errp)
 364 {
 365     COLOMessage msg;
 366     Error *local_err = NULL;
 367
 368     msg = colo_receive_message(f, &local_err);
 369     if (local_err) {
 370         error_propagate(errp, local_err);
 371         return;
 372     }
 373     if (msg != expect_msg) {
 374         error_setg(errp, "Unexpected COLO message %d, expected %d",
 375                           msg, expect_msg);
 376     }
 377 }
 378
 379 static uint64_t colo_receive_message_value(QEMUFile *f, uint32_t expect_msg,
 380                                            Error **errp)
 381 {
 382     Error *local_err = NULL;
 383     uint64_t value;
 384     int ret;
 385
 386     colo_receive_check_message(f, expect_msg, &local_err);
 387     if (local_err) {
 388         error_propagate(errp, local_err);
 389         return 0;
 390     }
 391
 392     value = qemu_get_be64(f);
 393     ret = qemu_file_get_error(f);
 394     if (ret < 0) {
 395         error_setg_errno(errp, -ret, "Failed to get value for COLO message: %s",
 396                          COLOMessage_str(expect_msg));
 397     }
 398     return value;
 399 }
 400
 401 static int colo_do_checkpoint_transaction(MigrationState *s,
 402                                           QIOChannelBuffer *bioc,
 403                                           QEMUFile *fb)
 404 {
 405     Error *local_err = NULL;
 406     int ret = -1;
 407
 408     colo_send_message(s->to_dst_file, COLO_MESSAGE_CHECKPOINT_REQUEST,
 409                       &local_err);
 410     if (local_err) {
 411         goto out;
 412     }
 413
 414     colo_receive_check_message(s->rp_state.from_dst_file,
 415                     COLO_MESSAGE_CHECKPOINT_REPLY, &local_err);
 416     if (local_err) {
 417         goto out;
 418     }
 419     /* Reset channel-buffer directly */
 420     qio_channel_io_seek(QIO_CHANNEL(bioc), 0, 0, NULL);
 421     bioc->usage = 0;
 422
 423     bql_lock();
 424     if (failover_get_state() != FAILOVER_STATUS_NONE) {
 425         bql_unlock();
 426         goto out;
 427     }
 428     vm_stop_force_state(RUN_STATE_COLO);
 429     bql_unlock();
 430     trace_colo_vm_state_change("run", "stop");
 431     /*
 432      * Failover request bh could be called after vm_stop_force_state(),
 433      * So we need check failover_request_is_active() again.
 434      */
 435     if (failover_get_state() != FAILOVER_STATUS_NONE) {
 436         goto out;
 437     }
 438     bql_lock();
 439
 440     replication_do_checkpoint_all(&local_err);
 441     if (local_err) {
 442         bql_unlock();
 443         goto out;
 444     }
 445
 446     colo_send_message(s->to_dst_file, COLO_MESSAGE_VMSTATE_SEND, &local_err);
 447     if (local_err) {
 448         bql_unlock();
 449         goto out;
 450     }
 451     /* Note: device state is saved into buffer */
 452     ret = qemu_save_device_state(fb);
 453
 454     bql_unlock();
 455     if (ret < 0) {
 456         goto out;
 457     }
 458
 459     if (migrate_auto_converge()) {
 460         mig_throttle_counter_reset();
 461     }
 462     /*
 463      * Only save VM's live state, which not including device state.
 464      * TODO: We may need a timeout mechanism to prevent COLO process
 465      * to be blocked here.
 466      */
 467     qemu_savevm_live_state(s->to_dst_file);
 468
 469     qemu_fflush(fb);
 470
 471     /*
 472      * We need the size of the VMstate data in Secondary side,
 473      * With which we can decide how much data should be read.
 474      */
 475     colo_send_message_value(s->to_dst_file, COLO_MESSAGE_VMSTATE_SIZE,
 476                             bioc->usage, &local_err);
 477     if (local_err) {
 478         goto out;
 479     }
 480
 481     qemu_put_buffer(s->to_dst_file, bioc->data, bioc->usage);
 482     ret = qemu_fflush(s->to_dst_file);
 483     if (ret < 0) {
 484         goto out;
 485     }
 486
 487     colo_receive_check_message(s->rp_state.from_dst_file,
 488                        COLO_MESSAGE_VMSTATE_RECEIVED, &local_err);
 489     if (local_err) {
 490         goto out;
 491     }
 492
 493     qemu_event_reset(&s->colo_checkpoint_event);
 494     colo_notify_compares_event(NULL, COLO_EVENT_CHECKPOINT, &local_err);
 495     if (local_err) {
 496         goto out;
 497     }
 498
 499     colo_receive_check_message(s->rp_state.from_dst_file,
 500                        COLO_MESSAGE_VMSTATE_LOADED, &local_err);
 501     if (local_err) {
 502         goto out;
 503     }
 504
 505     ret = 0;
 506
 507     bql_lock();
 508     vm_start();
 509     bql_unlock();
 510     trace_colo_vm_state_change("stop", "run");
 511
 512 out:
 513     if (local_err) {
 514         error_report_err(local_err);
 515     }
 516     return ret;
 517 }
 518
 519 static void colo_compare_notify_checkpoint(Notifier *notifier, void *data)
 520 {
 521     colo_checkpoint_notify(data);
 522 }
 523
 524 static void colo_process_checkpoint(MigrationState *s)
 525 {
 526     QIOChannelBuffer *bioc;
 527     QEMUFile *fb = NULL;
 528     Error *local_err = NULL;
 529     int ret;
 530
 531     if (get_colo_mode() != COLO_MODE_PRIMARY) {
 532         error_report("COLO mode must be COLO_MODE_PRIMARY");
 533         return;
 534     }
 535
 536     failover_init_state();
 537
 538     s->rp_state.from_dst_file = qemu_file_get_return_path(s->to_dst_file);
 539     if (!s->rp_state.from_dst_file) {
 540         error_report("Open QEMUFile from_dst_file failed");
 541         goto out;
 542     }
 543
 544     packets_compare_notifier.notify = colo_compare_notify_checkpoint;
 545     colo_compare_register_notifier(&packets_compare_notifier);
 546
 547     /*
 548      * Wait for Secondary finish loading VM states and enter COLO
 549      * restore.
 550      */
 551     colo_receive_check_message(s->rp_state.from_dst_file,
 552                        COLO_MESSAGE_CHECKPOINT_READY, &local_err);
 553     if (local_err) {
 554         goto out;
 555     }
 556     bioc = qio_channel_buffer_new(COLO_BUFFER_BASE_SIZE);
 557     fb = qemu_file_new_output(QIO_CHANNEL(bioc));
 558     object_unref(OBJECT(bioc));
 559
 560     bql_lock();
 561     replication_start_all(REPLICATION_MODE_PRIMARY, &local_err);
 562     if (local_err) {
 563         bql_unlock();
 564         goto out;
 565     }
 566
 567     vm_start();
 568     bql_unlock();
 569     trace_colo_vm_state_change("stop", "run");
 570
 571     timer_mod(s->colo_delay_timer, qemu_clock_get_ms(QEMU_CLOCK_HOST) +
 572               migrate_checkpoint_delay());
 573
 574     while (s->state == MIGRATION_STATUS_COLO) {
 575         if (failover_get_state() != FAILOVER_STATUS_NONE) {
 576             error_report("failover request");
 577             goto out;
 578         }
 579
 580         qemu_event_wait(&s->colo_checkpoint_event);
 581
 582         if (s->state != MIGRATION_STATUS_COLO) {
 583             goto out;
 584         }
 585         ret = colo_do_checkpoint_transaction(s, bioc, fb);
 586         if (ret < 0) {
 587             goto out;
 588         }
 589     }
 590
 591 out:
 592     /* Throw the unreported error message after exited from loop */
 593     if (local_err) {
 594         error_report_err(local_err);
 595     }
 596
 597     if (fb) {
 598         qemu_fclose(fb);
 599     }
 600
 601     /*
 602      * There are only two reasons we can get here, some error happened
 603      * or the user triggered failover.
 604      */
 605     switch (failover_get_state()) {
 606     case FAILOVER_STATUS_COMPLETED:
 607         qapi_event_send_colo_exit(COLO_MODE_PRIMARY,
 608                                   COLO_EXIT_REASON_REQUEST);
 609         break;
 610     default:
 611         qapi_event_send_colo_exit(COLO_MODE_PRIMARY,
 612                                   COLO_EXIT_REASON_ERROR);
 613     }
 614
 615     /* Hope this not to be too long to wait here */
 616     qemu_sem_wait(&s->colo_exit_sem);
 617     qemu_sem_destroy(&s->colo_exit_sem);
 618
 619     /*
 620      * It is safe to unregister notifier after failover finished.
 621      * Besides, colo_delay_timer and colo_checkpoint_sem can't be
 622      * released before unregister notifier, or there will be use-after-free
 623      * error.
 624      */
 625     colo_compare_unregister_notifier(&packets_compare_notifier);
 626     timer_free(s->colo_delay_timer);
 627     qemu_event_destroy(&s->colo_checkpoint_event);
 628
 629     /*
 630      * Must be called after failover BH is completed,
 631      * Or the failover BH may shutdown the wrong fd that
 632      * re-used by other threads after we release here.
 633      */
 634     if (s->rp_state.from_dst_file) {
 635         qemu_fclose(s->rp_state.from_dst_file);
 636         s->rp_state.from_dst_file = NULL;
 637     }
 638 }
 639
 640 void migrate_start_colo_process(MigrationState *s)
 641 {
 642     bql_unlock();
 643     qemu_event_init(&s->colo_checkpoint_event, false);
 644     s->colo_delay_timer =  timer_new_ms(QEMU_CLOCK_HOST,
 645                                 colo_checkpoint_notify, s);
 646
 647     qemu_sem_init(&s->colo_exit_sem, 0);
 648     colo_process_checkpoint(s);
 649     bql_lock();
 650 }
 651
 652 static void colo_incoming_process_checkpoint(MigrationIncomingState *mis,
 653                       QEMUFile *fb, QIOChannelBuffer *bioc, Error **errp)
 654 {
 655     uint64_t total_size;
 656     uint64_t value;
 657     Error *local_err = NULL;
 658     int ret;
 659
 660     bql_lock();
 661     vm_stop_force_state(RUN_STATE_COLO);
 662     bql_unlock();
 663     trace_colo_vm_state_change("run", "stop");
 664
 665     /* FIXME: This is unnecessary for periodic checkpoint mode */
 666     colo_send_message(mis->to_src_file, COLO_MESSAGE_CHECKPOINT_REPLY,
 667                  &local_err);
 668     if (local_err) {
 669         error_propagate(errp, local_err);
 670         return;
 671     }
 672
 673     colo_receive_check_message(mis->from_src_file,
 674                        COLO_MESSAGE_VMSTATE_SEND, &local_err);
 675     if (local_err) {
 676         error_propagate(errp, local_err);
 677         return;
 678     }
 679
 680     bql_lock();
 681     cpu_synchronize_all_states();
 682     ret = qemu_loadvm_state_main(mis->from_src_file, mis);
 683     bql_unlock();
 684
 685     if (ret < 0) {
 686         error_setg(errp, "Load VM's live state (ram) error");
 687         return;
 688     }
 689
 690     value = colo_receive_message_value(mis->from_src_file,
 691                              COLO_MESSAGE_VMSTATE_SIZE, &local_err);
 692     if (local_err) {
 693         error_propagate(errp, local_err);
 694         return;
 695     }
 696
 697     /*
 698      * Read VM device state data into channel buffer,
 699      * It's better to re-use the memory allocated.
 700      * Here we need to handle the channel buffer directly.
 701      */
 702     if (value > bioc->capacity) {
 703         bioc->capacity = value;
 704         bioc->data = g_realloc(bioc->data, bioc->capacity);
 705     }
 706     total_size = qemu_get_buffer(mis->from_src_file, bioc->data, value);
 707     if (total_size != value) {
 708         error_setg(errp, "Got %" PRIu64 " VMState data, less than expected"
 709                     " %" PRIu64, total_size, value);
 710         return;
 711     }
 712     bioc->usage = total_size;
 713     qio_channel_io_seek(QIO_CHANNEL(bioc), 0, 0, NULL);
 714
 715     colo_send_message(mis->to_src_file, COLO_MESSAGE_VMSTATE_RECEIVED,
 716                  &local_err);
 717     if (local_err) {
 718         error_propagate(errp, local_err);
 719         return;
 720     }
 721
 722     bql_lock();
 723     vmstate_loading = true;
 724     colo_flush_ram_cache();
 725     ret = qemu_load_device_state(fb);
 726     if (ret < 0) {
 727         error_setg(errp, "COLO: load device state failed");
 728         vmstate_loading = false;
 729         bql_unlock();
 730         return;
 731     }
 732
 733     replication_get_error_all(&local_err);
 734     if (local_err) {
 735         error_propagate(errp, local_err);
 736         vmstate_loading = false;
 737         bql_unlock();
 738         return;
 739     }
 740
 741     /* discard colo disk buffer */
 742     replication_do_checkpoint_all(&local_err);
 743     if (local_err) {
 744         error_propagate(errp, local_err);
 745         vmstate_loading = false;
 746         bql_unlock();
 747         return;
 748     }
 749     /* Notify all filters of all NIC to do checkpoint */
 750     colo_notify_filters_event(COLO_EVENT_CHECKPOINT, &local_err);
 751
 752     if (local_err) {
 753         error_propagate(errp, local_err);
 754         vmstate_loading = false;
 755         bql_unlock();
 756         return;
 757     }
 758
 759     vmstate_loading = false;
 760     vm_start();
 761     bql_unlock();
 762     trace_colo_vm_state_change("stop", "run");
 763
 764     if (failover_get_state() == FAILOVER_STATUS_RELAUNCH) {
 765         return;
 766     }
 767
 768     colo_send_message(mis->to_src_file, COLO_MESSAGE_VMSTATE_LOADED,
 769                  &local_err);
 770     error_propagate(errp, local_err);
 771 }
 772
 773 static void colo_wait_handle_message(MigrationIncomingState *mis,
 774                 QEMUFile *fb, QIOChannelBuffer *bioc, Error **errp)
 775 {
 776     COLOMessage msg;
 777     Error *local_err = NULL;
 778
 779     msg = colo_receive_message(mis->from_src_file, &local_err);
 780     if (local_err) {
 781         error_propagate(errp, local_err);
 782         return;
 783     }
 784
 785     switch (msg) {
 786     case COLO_MESSAGE_CHECKPOINT_REQUEST:
 787         colo_incoming_process_checkpoint(mis, fb, bioc, errp);
 788         break;
 789     default:
 790         error_setg(errp, "Got unknown COLO message: %d", msg);
 791         break;
 792     }
 793 }
 794
 795 void colo_shutdown(void)
 796 {
 797     MigrationIncomingState *mis = NULL;
 798     MigrationState *s = NULL;
 799
 800     switch (get_colo_mode()) {
 801     case COLO_MODE_PRIMARY:
 802         s = migrate_get_current();
 803         qemu_event_set(&s->colo_checkpoint_event);
 804         qemu_sem_post(&s->colo_exit_sem);
 805         break;
 806     case COLO_MODE_SECONDARY:
 807         mis = migration_incoming_get_current();
 808         qemu_sem_post(&mis->colo_incoming_sem);
 809         break;
 810     default:
 811         break;
 812     }
 813 }
 814
 815 static void *colo_process_incoming_thread(void *opaque)
 816 {
 817     MigrationIncomingState *mis = opaque;
 818     QEMUFile *fb = NULL;
 819     QIOChannelBuffer *bioc = NULL; /* Cache incoming device state */
 820     Error *local_err = NULL;
 821
 822     rcu_register_thread();
 823     qemu_sem_init(&mis->colo_incoming_sem, 0);
 824
 825     migrate_set_state(&mis->state, MIGRATION_STATUS_ACTIVE,
 826                       MIGRATION_STATUS_COLO);
 827
 828     if (get_colo_mode() != COLO_MODE_SECONDARY) {
 829         error_report("COLO mode must be COLO_MODE_SECONDARY");
 830         return NULL;
 831     }
 832
 833     failover_init_state();
 834
 835     mis->to_src_file = qemu_file_get_return_path(mis->from_src_file);
 836     if (!mis->to_src_file) {
 837         error_report("COLO incoming thread: Open QEMUFile to_src_file failed");
 838         goto out;
 839     }
 840     /*
 841      * Note: the communication between Primary side and Secondary side
 842      * should be sequential, we set the fd to unblocked in migration incoming
 843      * coroutine, and here we are in the COLO incoming thread, so it is ok to
 844      * set the fd back to blocked.
 845      */
 846     qemu_file_set_blocking(mis->from_src_file, true);
 847
 848     colo_incoming_start_dirty_log();
 849
 850     bioc = qio_channel_buffer_new(COLO_BUFFER_BASE_SIZE);
 851     fb = qemu_file_new_input(QIO_CHANNEL(bioc));
 852     object_unref(OBJECT(bioc));
 853
 854     bql_lock();
 855     replication_start_all(REPLICATION_MODE_SECONDARY, &local_err);
 856     if (local_err) {
 857         bql_unlock();
 858         goto out;
 859     }
 860     vm_start();
 861     bql_unlock();
 862     trace_colo_vm_state_change("stop", "run");
 863
 864     colo_send_message(mis->to_src_file, COLO_MESSAGE_CHECKPOINT_READY,
 865                       &local_err);
 866     if (local_err) {
 867         goto out;
 868     }
 869
 870     while (mis->state == MIGRATION_STATUS_COLO) {
 871         colo_wait_handle_message(mis, fb, bioc, &local_err);
 872         if (local_err) {
 873             error_report_err(local_err);
 874             break;
 875         }
 876
 877         if (failover_get_state() == FAILOVER_STATUS_RELAUNCH) {
 878             failover_set_state(FAILOVER_STATUS_RELAUNCH,
 879                             FAILOVER_STATUS_NONE);
 880             failover_request_active(NULL);
 881             break;
 882         }
 883
 884         if (failover_get_state() != FAILOVER_STATUS_NONE) {
 885             error_report("failover request");
 886             break;
 887         }
 888     }
 889
 890 out:
 891     /*
 892      * There are only two reasons we can get here, some error happened
 893      * or the user triggered failover.
 894      */
 895     switch (failover_get_state()) {
 896     case FAILOVER_STATUS_COMPLETED:
 897         qapi_event_send_colo_exit(COLO_MODE_SECONDARY,
 898                                   COLO_EXIT_REASON_REQUEST);
 899         break;
 900     default:
 901         qapi_event_send_colo_exit(COLO_MODE_SECONDARY,
 902                                   COLO_EXIT_REASON_ERROR);
 903     }
 904
 905     if (fb) {
 906         qemu_fclose(fb);
 907     }
 908
 909     /* Hope this not to be too long to loop here */
 910     qemu_sem_wait(&mis->colo_incoming_sem);
 911     qemu_sem_destroy(&mis->colo_incoming_sem);
 912
 913     rcu_unregister_thread();
 914     return NULL;
 915 }
 916
 917 int coroutine_fn colo_incoming_co(void)
 918 {
 919     MigrationIncomingState *mis = migration_incoming_get_current();
 920     Error *local_err = NULL;
 921     QemuThread th;
 922
 923     assert(bql_locked());
 924
 925     if (!migration_incoming_colo_enabled()) {
 926         return 0;
 927     }
 928
 929     /* Make sure all file formats throw away their mutable metadata */
 930     bdrv_activate_all(&local_err);
 931     if (local_err) {
 932         error_report_err(local_err);
 933         return -EINVAL;
 934     }
 935
 936     qemu_thread_create(&th, "COLO incoming", colo_process_incoming_thread,
 937                        mis, QEMU_THREAD_JOINABLE);
 938
 939     mis->colo_incoming_co = qemu_coroutine_self();
 940     qemu_coroutine_yield();
 941     mis->colo_incoming_co = NULL;
 942
 943     bql_unlock();
 944     /* Wait checkpoint incoming thread exit before free resource */
 945     qemu_thread_join(&th);
 946     bql_lock();
 947
 948     /* We hold the global BQL, so it is safe here */
 949     colo_release_ram_cache();
 950
 951     return 0;
 952 }