block/replication.c

   1 /*
   2  * Replication Block filter
   3  *
   4  * Copyright (c) 2016 HUAWEI TECHNOLOGIES CO., LTD.
   5  * Copyright (c) 2016 Intel Corporation
   6  * Copyright (c) 2016 FUJITSU LIMITED
   7  *
   8  * Author:
   9  *   Wen Congyang <wency@cn.fujitsu.com>
  10  *
  11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  12  * See the COPYING file in the top-level directory.
  13  */
  14
  15 #include "qemu/osdep.h"
  16 #include "qemu-common.h"
  17 #include "block/nbd.h"
  18 #include "block/blockjob.h"
  19 #include "block/block_int.h"
  20 #include "block/block_backup.h"
  21 #include "sysemu/block-backend.h"
  22 #include "qapi/error.h"
  23 #include "replication.h"
  24
  25 typedef struct BDRVReplicationState {
  26     ReplicationMode mode;
  27     int replication_state;
  28     BdrvChild *active_disk;
  29     BdrvChild *hidden_disk;
  30     BdrvChild *secondary_disk;
  31     char *top_id;
  32     ReplicationState *rs;
  33     Error *blocker;
  34     int orig_hidden_flags;
  35     int orig_secondary_flags;
  36     int error;
  37 } BDRVReplicationState;
  38
  39 enum {
  40     BLOCK_REPLICATION_NONE,             /* block replication is not started */
  41     BLOCK_REPLICATION_RUNNING,          /* block replication is running */
  42     BLOCK_REPLICATION_FAILOVER,         /* failover is running in background */
  43     BLOCK_REPLICATION_FAILOVER_FAILED,  /* failover failed */
  44     BLOCK_REPLICATION_DONE,             /* block replication is done */
  45 };
  46
  47 static void replication_start(ReplicationState *rs, ReplicationMode mode,
  48                               Error **errp);
  49 static void replication_do_checkpoint(ReplicationState *rs, Error **errp);
  50 static void replication_get_error(ReplicationState *rs, Error **errp);
  51 static void replication_stop(ReplicationState *rs, bool failover,
  52                              Error **errp);
  53
  54 #define REPLICATION_MODE        "mode"
  55 #define REPLICATION_TOP_ID      "top-id"
  56 static QemuOptsList replication_runtime_opts = {
  57     .name = "replication",
  58     .head = QTAILQ_HEAD_INITIALIZER(replication_runtime_opts.head),
  59     .desc = {
  60         {
  61             .name = REPLICATION_MODE,
  62             .type = QEMU_OPT_STRING,
  63         },
  64         {
  65             .name = REPLICATION_TOP_ID,
  66             .type = QEMU_OPT_STRING,
  67         },
  68         { /* end of list */ }
  69     },
  70 };
  71
  72 static ReplicationOps replication_ops = {
  73     .start = replication_start,
  74     .checkpoint = replication_do_checkpoint,
  75     .get_error = replication_get_error,
  76     .stop = replication_stop,
  77 };
  78
  79 static int replication_open(BlockDriverState *bs, QDict *options,
  80                             int flags, Error **errp)
  81 {
  82     int ret;
  83     BDRVReplicationState *s = bs->opaque;
  84     Error *local_err = NULL;
  85     QemuOpts *opts = NULL;
  86     const char *mode;
  87     const char *top_id;
  88
  89     ret = -EINVAL;
  90     opts = qemu_opts_create(&replication_runtime_opts, NULL, 0, &error_abort);
  91     qemu_opts_absorb_qdict(opts, options, &local_err);
  92     if (local_err) {
  93         goto fail;
  94     }
  95
  96     mode = qemu_opt_get(opts, REPLICATION_MODE);
  97     if (!mode) {
  98         error_setg(&local_err, "Missing the option mode");
  99         goto fail;
 100     }
 101
 102     if (!strcmp(mode, "primary")) {
 103         s->mode = REPLICATION_MODE_PRIMARY;
 104     } else if (!strcmp(mode, "secondary")) {
 105         s->mode = REPLICATION_MODE_SECONDARY;
 106         top_id = qemu_opt_get(opts, REPLICATION_TOP_ID);
 107         s->top_id = g_strdup(top_id);
 108         if (!s->top_id) {
 109             error_setg(&local_err, "Missing the option top-id");
 110             goto fail;
 111         }
 112     } else {
 113         error_setg(&local_err,
 114                    "The option mode's value should be primary or secondary");
 115         goto fail;
 116     }
 117
 118     s->rs = replication_new(bs, &replication_ops);
 119
 120     ret = 0;
 121
 122 fail:
 123     qemu_opts_del(opts);
 124     error_propagate(errp, local_err);
 125
 126     return ret;
 127 }
 128
 129 static void replication_close(BlockDriverState *bs)
 130 {
 131     BDRVReplicationState *s = bs->opaque;
 132
 133     if (s->replication_state == BLOCK_REPLICATION_RUNNING) {
 134         replication_stop(s->rs, false, NULL);
 135     }
 136
 137     if (s->mode == REPLICATION_MODE_SECONDARY) {
 138         g_free(s->top_id);
 139     }
 140
 141     replication_remove(s->rs);
 142 }
 143
 144 static int64_t replication_getlength(BlockDriverState *bs)
 145 {
 146     return bdrv_getlength(bs->file->bs);
 147 }
 148
 149 static int replication_get_io_status(BDRVReplicationState *s)
 150 {
 151     switch (s->replication_state) {
 152     case BLOCK_REPLICATION_NONE:
 153         return -EIO;
 154     case BLOCK_REPLICATION_RUNNING:
 155         return 0;
 156     case BLOCK_REPLICATION_FAILOVER:
 157         return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 0;
 158     case BLOCK_REPLICATION_FAILOVER_FAILED:
 159         return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 1;
 160     case BLOCK_REPLICATION_DONE:
 161         /*
 162          * active commit job completes, and active disk and secondary_disk
 163          * is swapped, so we can operate bs->file directly
 164          */
 165         return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 0;
 166     default:
 167         abort();
 168     }
 169 }
 170
 171 static int replication_return_value(BDRVReplicationState *s, int ret)
 172 {
 173     if (s->mode == REPLICATION_MODE_SECONDARY) {
 174         return ret;
 175     }
 176
 177     if (ret < 0) {
 178         s->error = ret;
 179         ret = 0;
 180     }
 181
 182     return ret;
 183 }
 184
 185 static coroutine_fn int replication_co_readv(BlockDriverState *bs,
 186                                              int64_t sector_num,
 187                                              int remaining_sectors,
 188                                              QEMUIOVector *qiov)
 189 {
 190     BDRVReplicationState *s = bs->opaque;
 191     BdrvChild *child = s->secondary_disk;
 192     BlockJob *job = NULL;
 193     CowRequest req;
 194     int ret;
 195
 196     if (s->mode == REPLICATION_MODE_PRIMARY) {
 197         /* We only use it to forward primary write requests */
 198         return -EIO;
 199     }
 200
 201     ret = replication_get_io_status(s);
 202     if (ret < 0) {
 203         return ret;
 204     }
 205
 206     if (child && child->bs) {
 207         job = child->bs->job;
 208     }
 209
 210     if (job) {
 211         backup_wait_for_overlapping_requests(child->bs->job, sector_num,
 212                                              remaining_sectors);
 213         backup_cow_request_begin(&req, child->bs->job, sector_num,
 214                                  remaining_sectors);
 215         ret = bdrv_co_readv(bs->file, sector_num, remaining_sectors,
 216                             qiov);
 217         backup_cow_request_end(&req);
 218         goto out;
 219     }
 220
 221     ret = bdrv_co_readv(bs->file, sector_num, remaining_sectors, qiov);
 222 out:
 223     return replication_return_value(s, ret);
 224 }
 225
 226 static coroutine_fn int replication_co_writev(BlockDriverState *bs,
 227                                               int64_t sector_num,
 228                                               int remaining_sectors,
 229                                               QEMUIOVector *qiov)
 230 {
 231     BDRVReplicationState *s = bs->opaque;
 232     QEMUIOVector hd_qiov;
 233     uint64_t bytes_done = 0;
 234     BdrvChild *top = bs->file;
 235     BdrvChild *base = s->secondary_disk;
 236     BdrvChild *target;
 237     int ret, n;
 238
 239     ret = replication_get_io_status(s);
 240     if (ret < 0) {
 241         goto out;
 242     }
 243
 244     if (ret == 0) {
 245         ret = bdrv_co_writev(top, sector_num,
 246                              remaining_sectors, qiov);
 247         return replication_return_value(s, ret);
 248     }
 249
 250     /*
 251      * Failover failed, only write to active disk if the sectors
 252      * have already been allocated in active disk/hidden disk.
 253      */
 254     qemu_iovec_init(&hd_qiov, qiov->niov);
 255     while (remaining_sectors > 0) {
 256         ret = bdrv_is_allocated_above(top->bs, base->bs, sector_num,
 257                                       remaining_sectors, &n);
 258         if (ret < 0) {
 259             goto out1;
 260         }
 261
 262         qemu_iovec_reset(&hd_qiov);
 263         qemu_iovec_concat(&hd_qiov, qiov, bytes_done, n * BDRV_SECTOR_SIZE);
 264
 265         target = ret ? top : base;
 266         ret = bdrv_co_writev(target, sector_num, n, &hd_qiov);
 267         if (ret < 0) {
 268             goto out1;
 269         }
 270
 271         remaining_sectors -= n;
 272         sector_num += n;
 273         bytes_done += n * BDRV_SECTOR_SIZE;
 274     }
 275
 276 out1:
 277     qemu_iovec_destroy(&hd_qiov);
 278 out:
 279     return ret;
 280 }
 281
 282 static bool replication_recurse_is_first_non_filter(BlockDriverState *bs,
 283                                                     BlockDriverState *candidate)
 284 {
 285     return bdrv_recurse_is_first_non_filter(bs->file->bs, candidate);
 286 }
 287
 288 static void secondary_do_checkpoint(BDRVReplicationState *s, Error **errp)
 289 {
 290     Error *local_err = NULL;
 291     int ret;
 292
 293     if (!s->secondary_disk->bs->job) {
 294         error_setg(errp, "Backup job was cancelled unexpectedly");
 295         return;
 296     }
 297
 298     backup_do_checkpoint(s->secondary_disk->bs->job, &local_err);
 299     if (local_err) {
 300         error_propagate(errp, local_err);
 301         return;
 302     }
 303
 304     ret = s->active_disk->bs->drv->bdrv_make_empty(s->active_disk->bs);
 305     if (ret < 0) {
 306         error_setg(errp, "Cannot make active disk empty");
 307         return;
 308     }
 309
 310     ret = s->hidden_disk->bs->drv->bdrv_make_empty(s->hidden_disk->bs);
 311     if (ret < 0) {
 312         error_setg(errp, "Cannot make hidden disk empty");
 313         return;
 314     }
 315 }
 316
 317 static void reopen_backing_file(BDRVReplicationState *s, bool writable,
 318                                 Error **errp)
 319 {
 320     BlockReopenQueue *reopen_queue = NULL;
 321     int orig_hidden_flags, orig_secondary_flags;
 322     int new_hidden_flags, new_secondary_flags;
 323     Error *local_err = NULL;
 324
 325     if (writable) {
 326         orig_hidden_flags = s->orig_hidden_flags =
 327                                 bdrv_get_flags(s->hidden_disk->bs);
 328         new_hidden_flags = (orig_hidden_flags | BDRV_O_RDWR) &
 329                                                     ~BDRV_O_INACTIVE;
 330         orig_secondary_flags = s->orig_secondary_flags =
 331                                 bdrv_get_flags(s->secondary_disk->bs);
 332         new_secondary_flags = (orig_secondary_flags | BDRV_O_RDWR) &
 333                                                      ~BDRV_O_INACTIVE;
 334     } else {
 335         orig_hidden_flags = (s->orig_hidden_flags | BDRV_O_RDWR) &
 336                                                     ~BDRV_O_INACTIVE;
 337         new_hidden_flags = s->orig_hidden_flags;
 338         orig_secondary_flags = (s->orig_secondary_flags | BDRV_O_RDWR) &
 339                                                     ~BDRV_O_INACTIVE;
 340         new_secondary_flags = s->orig_secondary_flags;
 341     }
 342
 343     if (orig_hidden_flags != new_hidden_flags) {
 344         reopen_queue = bdrv_reopen_queue(reopen_queue, s->hidden_disk->bs, NULL,
 345                                          new_hidden_flags);
 346     }
 347
 348     if (!(orig_secondary_flags & BDRV_O_RDWR)) {
 349         reopen_queue = bdrv_reopen_queue(reopen_queue, s->secondary_disk->bs,
 350                                          NULL, new_secondary_flags);
 351     }
 352
 353     if (reopen_queue) {
 354         bdrv_reopen_multiple(reopen_queue, &local_err);
 355         error_propagate(errp, local_err);
 356     }
 357 }
 358
 359 static void backup_job_cleanup(BDRVReplicationState *s)
 360 {
 361     BlockDriverState *top_bs;
 362
 363     top_bs = bdrv_lookup_bs(s->top_id, s->top_id, NULL);
 364     if (!top_bs) {
 365         return;
 366     }
 367     bdrv_op_unblock_all(top_bs, s->blocker);
 368     error_free(s->blocker);
 369     reopen_backing_file(s, false, NULL);
 370 }
 371
 372 static void backup_job_completed(void *opaque, int ret)
 373 {
 374     BDRVReplicationState *s = opaque;
 375
 376     if (s->replication_state != BLOCK_REPLICATION_FAILOVER) {
 377         /* The backup job is cancelled unexpectedly */
 378         s->error = -EIO;
 379     }
 380
 381     backup_job_cleanup(s);
 382 }
 383
 384 static bool check_top_bs(BlockDriverState *top_bs, BlockDriverState *bs)
 385 {
 386     BdrvChild *child;
 387
 388     /* The bs itself is the top_bs */
 389     if (top_bs == bs) {
 390         return true;
 391     }
 392
 393     /* Iterate over top_bs's children */
 394     QLIST_FOREACH(child, &top_bs->children, next) {
 395         if (child->bs == bs || check_top_bs(child->bs, bs)) {
 396             return true;
 397         }
 398     }
 399
 400     return false;
 401 }
 402
 403 static void replication_start(ReplicationState *rs, ReplicationMode mode,
 404                               Error **errp)
 405 {
 406     BlockDriverState *bs = rs->opaque;
 407     BDRVReplicationState *s;
 408     BlockDriverState *top_bs;
 409     int64_t active_length, hidden_length, disk_length;
 410     AioContext *aio_context;
 411     Error *local_err = NULL;
 412
 413     aio_context = bdrv_get_aio_context(bs);
 414     aio_context_acquire(aio_context);
 415     s = bs->opaque;
 416
 417     if (s->replication_state != BLOCK_REPLICATION_NONE) {
 418         error_setg(errp, "Block replication is running or done");
 419         aio_context_release(aio_context);
 420         return;
 421     }
 422
 423     if (s->mode != mode) {
 424         error_setg(errp, "The parameter mode's value is invalid, needs %d,"
 425                    " but got %d", s->mode, mode);
 426         aio_context_release(aio_context);
 427         return;
 428     }
 429
 430     switch (s->mode) {
 431     case REPLICATION_MODE_PRIMARY:
 432         break;
 433     case REPLICATION_MODE_SECONDARY:
 434         s->active_disk = bs->file;
 435         if (!s->active_disk || !s->active_disk->bs ||
 436                                     !s->active_disk->bs->backing) {
 437             error_setg(errp, "Active disk doesn't have backing file");
 438             aio_context_release(aio_context);
 439             return;
 440         }
 441
 442         s->hidden_disk = s->active_disk->bs->backing;
 443         if (!s->hidden_disk->bs || !s->hidden_disk->bs->backing) {
 444             error_setg(errp, "Hidden disk doesn't have backing file");
 445             aio_context_release(aio_context);
 446             return;
 447         }
 448
 449         s->secondary_disk = s->hidden_disk->bs->backing;
 450         if (!s->secondary_disk->bs || !bdrv_has_blk(s->secondary_disk->bs)) {
 451             error_setg(errp, "The secondary disk doesn't have block backend");
 452             aio_context_release(aio_context);
 453             return;
 454         }
 455
 456         /* verify the length */
 457         active_length = bdrv_getlength(s->active_disk->bs);
 458         hidden_length = bdrv_getlength(s->hidden_disk->bs);
 459         disk_length = bdrv_getlength(s->secondary_disk->bs);
 460         if (active_length < 0 || hidden_length < 0 || disk_length < 0 ||
 461             active_length != hidden_length || hidden_length != disk_length) {
 462             error_setg(errp, "Active disk, hidden disk, secondary disk's length"
 463                        " are not the same");
 464             aio_context_release(aio_context);
 465             return;
 466         }
 467
 468         if (!s->active_disk->bs->drv->bdrv_make_empty ||
 469             !s->hidden_disk->bs->drv->bdrv_make_empty) {
 470             error_setg(errp,
 471                        "Active disk or hidden disk doesn't support make_empty");
 472             aio_context_release(aio_context);
 473             return;
 474         }
 475
 476         /* reopen the backing file in r/w mode */
 477         reopen_backing_file(s, true, &local_err);
 478         if (local_err) {
 479             error_propagate(errp, local_err);
 480             aio_context_release(aio_context);
 481             return;
 482         }
 483
 484         /* start backup job now */
 485         error_setg(&s->blocker,
 486                    "Block device is in use by internal backup job");
 487
 488         top_bs = bdrv_lookup_bs(s->top_id, s->top_id, NULL);
 489         if (!top_bs || !bdrv_is_root_node(top_bs) ||
 490             !check_top_bs(top_bs, bs)) {
 491             error_setg(errp, "No top_bs or it is invalid");
 492             reopen_backing_file(s, false, NULL);
 493             aio_context_release(aio_context);
 494             return;
 495         }
 496         bdrv_op_block_all(top_bs, s->blocker);
 497         bdrv_op_unblock(top_bs, BLOCK_OP_TYPE_DATAPLANE, s->blocker);
 498
 499         backup_start("replication-backup", s->secondary_disk->bs,
 500                      s->hidden_disk->bs, 0, MIRROR_SYNC_MODE_NONE, NULL, false,
 501                      BLOCKDEV_ON_ERROR_REPORT, BLOCKDEV_ON_ERROR_REPORT,
 502                      backup_job_completed, s, NULL, &local_err);
 503         if (local_err) {
 504             error_propagate(errp, local_err);
 505             backup_job_cleanup(s);
 506             aio_context_release(aio_context);
 507             return;
 508         }
 509         break;
 510     default:
 511         aio_context_release(aio_context);
 512         abort();
 513     }
 514
 515     s->replication_state = BLOCK_REPLICATION_RUNNING;
 516
 517     if (s->mode == REPLICATION_MODE_SECONDARY) {
 518         secondary_do_checkpoint(s, errp);
 519     }
 520
 521     s->error = 0;
 522     aio_context_release(aio_context);
 523 }
 524
 525 static void replication_do_checkpoint(ReplicationState *rs, Error **errp)
 526 {
 527     BlockDriverState *bs = rs->opaque;
 528     BDRVReplicationState *s;
 529     AioContext *aio_context;
 530
 531     aio_context = bdrv_get_aio_context(bs);
 532     aio_context_acquire(aio_context);
 533     s = bs->opaque;
 534
 535     if (s->mode == REPLICATION_MODE_SECONDARY) {
 536         secondary_do_checkpoint(s, errp);
 537     }
 538     aio_context_release(aio_context);
 539 }
 540
 541 static void replication_get_error(ReplicationState *rs, Error **errp)
 542 {
 543     BlockDriverState *bs = rs->opaque;
 544     BDRVReplicationState *s;
 545     AioContext *aio_context;
 546
 547     aio_context = bdrv_get_aio_context(bs);
 548     aio_context_acquire(aio_context);
 549     s = bs->opaque;
 550
 551     if (s->replication_state != BLOCK_REPLICATION_RUNNING) {
 552         error_setg(errp, "Block replication is not running");
 553         aio_context_release(aio_context);
 554         return;
 555     }
 556
 557     if (s->error) {
 558         error_setg(errp, "I/O error occurred");
 559         aio_context_release(aio_context);
 560         return;
 561     }
 562     aio_context_release(aio_context);
 563 }
 564
 565 static void replication_done(void *opaque, int ret)
 566 {
 567     BlockDriverState *bs = opaque;
 568     BDRVReplicationState *s = bs->opaque;
 569
 570     if (ret == 0) {
 571         s->replication_state = BLOCK_REPLICATION_DONE;
 572
 573         /* refresh top bs's filename */
 574         bdrv_refresh_filename(bs);
 575         s->active_disk = NULL;
 576         s->secondary_disk = NULL;
 577         s->hidden_disk = NULL;
 578         s->error = 0;
 579     } else {
 580         s->replication_state = BLOCK_REPLICATION_FAILOVER_FAILED;
 581         s->error = -EIO;
 582     }
 583 }
 584
 585 static void replication_stop(ReplicationState *rs, bool failover, Error **errp)
 586 {
 587     BlockDriverState *bs = rs->opaque;
 588     BDRVReplicationState *s;
 589     AioContext *aio_context;
 590
 591     aio_context = bdrv_get_aio_context(bs);
 592     aio_context_acquire(aio_context);
 593     s = bs->opaque;
 594
 595     if (s->replication_state != BLOCK_REPLICATION_RUNNING) {
 596         error_setg(errp, "Block replication is not running");
 597         aio_context_release(aio_context);
 598         return;
 599     }
 600
 601     switch (s->mode) {
 602     case REPLICATION_MODE_PRIMARY:
 603         s->replication_state = BLOCK_REPLICATION_DONE;
 604         s->error = 0;
 605         break;
 606     case REPLICATION_MODE_SECONDARY:
 607         /*
 608          * This BDS will be closed, and the job should be completed
 609          * before the BDS is closed, because we will access hidden
 610          * disk, secondary disk in backup_job_completed().
 611          */
 612         if (s->secondary_disk->bs->job) {
 613             block_job_cancel_sync(s->secondary_disk->bs->job);
 614         }
 615
 616         if (!failover) {
 617             secondary_do_checkpoint(s, errp);
 618             s->replication_state = BLOCK_REPLICATION_DONE;
 619             aio_context_release(aio_context);
 620             return;
 621         }
 622
 623         s->replication_state = BLOCK_REPLICATION_FAILOVER;
 624         commit_active_start("replication-commit", s->active_disk->bs,
 625                             s->secondary_disk->bs, 0, BLOCKDEV_ON_ERROR_REPORT,
 626                             replication_done,
 627                             bs, errp, true);
 628         break;
 629     default:
 630         aio_context_release(aio_context);
 631         abort();
 632     }
 633     aio_context_release(aio_context);
 634 }
 635
 636 BlockDriver bdrv_replication = {
 637     .format_name                = "replication",
 638     .protocol_name              = "replication",
 639     .instance_size              = sizeof(BDRVReplicationState),
 640
 641     .bdrv_open                  = replication_open,
 642     .bdrv_close                 = replication_close,
 643
 644     .bdrv_getlength             = replication_getlength,
 645     .bdrv_co_readv              = replication_co_readv,
 646     .bdrv_co_writev             = replication_co_writev,
 647
 648     .is_filter                  = true,
 649     .bdrv_recurse_is_first_non_filter = replication_recurse_is_first_non_filter,
 650
 651     .has_variable_length        = true,
 652 };
 653
 654 static void bdrv_replication_init(void)
 655 {
 656     bdrv_register(&bdrv_replication);
 657 }
 658
 659 block_init(bdrv_replication_init);