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         top_id = qemu_opt_get(opts, REPLICATION_TOP_ID);
 105         if (top_id) {
 106             error_setg(&local_err, "The primary side does not support option top-id");
 107             goto fail;
 108         }
 109     } else if (!strcmp(mode, "secondary")) {
 110         s->mode = REPLICATION_MODE_SECONDARY;
 111         top_id = qemu_opt_get(opts, REPLICATION_TOP_ID);
 112         s->top_id = g_strdup(top_id);
 113         if (!s->top_id) {
 114             error_setg(&local_err, "Missing the option top-id");
 115             goto fail;
 116         }
 117     } else {
 118         error_setg(&local_err,
 119                    "The option mode's value should be primary or secondary");
 120         goto fail;
 121     }
 122
 123     s->rs = replication_new(bs, &replication_ops);
 124
 125     ret = 0;
 126
 127 fail:
 128     qemu_opts_del(opts);
 129     error_propagate(errp, local_err);
 130
 131     return ret;
 132 }
 133
 134 static void replication_close(BlockDriverState *bs)
 135 {
 136     BDRVReplicationState *s = bs->opaque;
 137
 138     if (s->replication_state == BLOCK_REPLICATION_RUNNING) {
 139         replication_stop(s->rs, false, NULL);
 140     }
 141
 142     if (s->mode == REPLICATION_MODE_SECONDARY) {
 143         g_free(s->top_id);
 144     }
 145
 146     replication_remove(s->rs);
 147 }
 148
 149 static int64_t replication_getlength(BlockDriverState *bs)
 150 {
 151     return bdrv_getlength(bs->file->bs);
 152 }
 153
 154 static int replication_get_io_status(BDRVReplicationState *s)
 155 {
 156     switch (s->replication_state) {
 157     case BLOCK_REPLICATION_NONE:
 158         return -EIO;
 159     case BLOCK_REPLICATION_RUNNING:
 160         return 0;
 161     case BLOCK_REPLICATION_FAILOVER:
 162         return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 0;
 163     case BLOCK_REPLICATION_FAILOVER_FAILED:
 164         return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 1;
 165     case BLOCK_REPLICATION_DONE:
 166         /*
 167          * active commit job completes, and active disk and secondary_disk
 168          * is swapped, so we can operate bs->file directly
 169          */
 170         return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 0;
 171     default:
 172         abort();
 173     }
 174 }
 175
 176 static int replication_return_value(BDRVReplicationState *s, int ret)
 177 {
 178     if (s->mode == REPLICATION_MODE_SECONDARY) {
 179         return ret;
 180     }
 181
 182     if (ret < 0) {
 183         s->error = ret;
 184         ret = 0;
 185     }
 186
 187     return ret;
 188 }
 189
 190 static coroutine_fn int replication_co_readv(BlockDriverState *bs,
 191                                              int64_t sector_num,
 192                                              int remaining_sectors,
 193                                              QEMUIOVector *qiov)
 194 {
 195     BDRVReplicationState *s = bs->opaque;
 196     BdrvChild *child = s->secondary_disk;
 197     BlockJob *job = NULL;
 198     CowRequest req;
 199     int ret;
 200
 201     if (s->mode == REPLICATION_MODE_PRIMARY) {
 202         /* We only use it to forward primary write requests */
 203         return -EIO;
 204     }
 205
 206     ret = replication_get_io_status(s);
 207     if (ret < 0) {
 208         return ret;
 209     }
 210
 211     if (child && child->bs) {
 212         job = child->bs->job;
 213     }
 214
 215     if (job) {
 216         backup_wait_for_overlapping_requests(child->bs->job, sector_num,
 217                                              remaining_sectors);
 218         backup_cow_request_begin(&req, child->bs->job, sector_num,
 219                                  remaining_sectors);
 220         ret = bdrv_co_readv(bs->file, sector_num, remaining_sectors,
 221                             qiov);
 222         backup_cow_request_end(&req);
 223         goto out;
 224     }
 225
 226     ret = bdrv_co_readv(bs->file, sector_num, remaining_sectors, qiov);
 227 out:
 228     return replication_return_value(s, ret);
 229 }
 230
 231 static coroutine_fn int replication_co_writev(BlockDriverState *bs,
 232                                               int64_t sector_num,
 233                                               int remaining_sectors,
 234                                               QEMUIOVector *qiov)
 235 {
 236     BDRVReplicationState *s = bs->opaque;
 237     QEMUIOVector hd_qiov;
 238     uint64_t bytes_done = 0;
 239     BdrvChild *top = bs->file;
 240     BdrvChild *base = s->secondary_disk;
 241     BdrvChild *target;
 242     int ret, n;
 243
 244     ret = replication_get_io_status(s);
 245     if (ret < 0) {
 246         goto out;
 247     }
 248
 249     if (ret == 0) {
 250         ret = bdrv_co_writev(top, sector_num,
 251                              remaining_sectors, qiov);
 252         return replication_return_value(s, ret);
 253     }
 254
 255     /*
 256      * Failover failed, only write to active disk if the sectors
 257      * have already been allocated in active disk/hidden disk.
 258      */
 259     qemu_iovec_init(&hd_qiov, qiov->niov);
 260     while (remaining_sectors > 0) {
 261         ret = bdrv_is_allocated_above(top->bs, base->bs, sector_num,
 262                                       remaining_sectors, &n);
 263         if (ret < 0) {
 264             goto out1;
 265         }
 266
 267         qemu_iovec_reset(&hd_qiov);
 268         qemu_iovec_concat(&hd_qiov, qiov, bytes_done, n * BDRV_SECTOR_SIZE);
 269
 270         target = ret ? top : base;
 271         ret = bdrv_co_writev(target, sector_num, n, &hd_qiov);
 272         if (ret < 0) {
 273             goto out1;
 274         }
 275
 276         remaining_sectors -= n;
 277         sector_num += n;
 278         bytes_done += n * BDRV_SECTOR_SIZE;
 279     }
 280
 281 out1:
 282     qemu_iovec_destroy(&hd_qiov);
 283 out:
 284     return ret;
 285 }
 286
 287 static bool replication_recurse_is_first_non_filter(BlockDriverState *bs,
 288                                                     BlockDriverState *candidate)
 289 {
 290     return bdrv_recurse_is_first_non_filter(bs->file->bs, candidate);
 291 }
 292
 293 static void secondary_do_checkpoint(BDRVReplicationState *s, Error **errp)
 294 {
 295     Error *local_err = NULL;
 296     int ret;
 297
 298     if (!s->secondary_disk->bs->job) {
 299         error_setg(errp, "Backup job was cancelled unexpectedly");
 300         return;
 301     }
 302
 303     backup_do_checkpoint(s->secondary_disk->bs->job, &local_err);
 304     if (local_err) {
 305         error_propagate(errp, local_err);
 306         return;
 307     }
 308
 309     ret = s->active_disk->bs->drv->bdrv_make_empty(s->active_disk->bs);
 310     if (ret < 0) {
 311         error_setg(errp, "Cannot make active disk empty");
 312         return;
 313     }
 314
 315     ret = s->hidden_disk->bs->drv->bdrv_make_empty(s->hidden_disk->bs);
 316     if (ret < 0) {
 317         error_setg(errp, "Cannot make hidden disk empty");
 318         return;
 319     }
 320 }
 321
 322 static void reopen_backing_file(BDRVReplicationState *s, bool writable,
 323                                 Error **errp)
 324 {
 325     BlockReopenQueue *reopen_queue = NULL;
 326     int orig_hidden_flags, orig_secondary_flags;
 327     int new_hidden_flags, new_secondary_flags;
 328     Error *local_err = NULL;
 329
 330     if (writable) {
 331         orig_hidden_flags = s->orig_hidden_flags =
 332                                 bdrv_get_flags(s->hidden_disk->bs);
 333         new_hidden_flags = (orig_hidden_flags | BDRV_O_RDWR) &
 334                                                     ~BDRV_O_INACTIVE;
 335         orig_secondary_flags = s->orig_secondary_flags =
 336                                 bdrv_get_flags(s->secondary_disk->bs);
 337         new_secondary_flags = (orig_secondary_flags | BDRV_O_RDWR) &
 338                                                      ~BDRV_O_INACTIVE;
 339     } else {
 340         orig_hidden_flags = (s->orig_hidden_flags | BDRV_O_RDWR) &
 341                                                     ~BDRV_O_INACTIVE;
 342         new_hidden_flags = s->orig_hidden_flags;
 343         orig_secondary_flags = (s->orig_secondary_flags | BDRV_O_RDWR) &
 344                                                     ~BDRV_O_INACTIVE;
 345         new_secondary_flags = s->orig_secondary_flags;
 346     }
 347
 348     if (orig_hidden_flags != new_hidden_flags) {
 349         reopen_queue = bdrv_reopen_queue(reopen_queue, s->hidden_disk->bs, NULL,
 350                                          new_hidden_flags);
 351     }
 352
 353     if (!(orig_secondary_flags & BDRV_O_RDWR)) {
 354         reopen_queue = bdrv_reopen_queue(reopen_queue, s->secondary_disk->bs,
 355                                          NULL, new_secondary_flags);
 356     }
 357
 358     if (reopen_queue) {
 359         bdrv_reopen_multiple(reopen_queue, &local_err);
 360         error_propagate(errp, local_err);
 361     }
 362 }
 363
 364 static void backup_job_cleanup(BDRVReplicationState *s)
 365 {
 366     BlockDriverState *top_bs;
 367
 368     top_bs = bdrv_lookup_bs(s->top_id, s->top_id, NULL);
 369     if (!top_bs) {
 370         return;
 371     }
 372     bdrv_op_unblock_all(top_bs, s->blocker);
 373     error_free(s->blocker);
 374     reopen_backing_file(s, false, NULL);
 375 }
 376
 377 static void backup_job_completed(void *opaque, int ret)
 378 {
 379     BDRVReplicationState *s = opaque;
 380
 381     if (s->replication_state != BLOCK_REPLICATION_FAILOVER) {
 382         /* The backup job is cancelled unexpectedly */
 383         s->error = -EIO;
 384     }
 385
 386     backup_job_cleanup(s);
 387 }
 388
 389 static bool check_top_bs(BlockDriverState *top_bs, BlockDriverState *bs)
 390 {
 391     BdrvChild *child;
 392
 393     /* The bs itself is the top_bs */
 394     if (top_bs == bs) {
 395         return true;
 396     }
 397
 398     /* Iterate over top_bs's children */
 399     QLIST_FOREACH(child, &top_bs->children, next) {
 400         if (child->bs == bs || check_top_bs(child->bs, bs)) {
 401             return true;
 402         }
 403     }
 404
 405     return false;
 406 }
 407
 408 static void replication_start(ReplicationState *rs, ReplicationMode mode,
 409                               Error **errp)
 410 {
 411     BlockDriverState *bs = rs->opaque;
 412     BDRVReplicationState *s;
 413     BlockDriverState *top_bs;
 414     int64_t active_length, hidden_length, disk_length;
 415     AioContext *aio_context;
 416     Error *local_err = NULL;
 417
 418     aio_context = bdrv_get_aio_context(bs);
 419     aio_context_acquire(aio_context);
 420     s = bs->opaque;
 421
 422     if (s->replication_state != BLOCK_REPLICATION_NONE) {
 423         error_setg(errp, "Block replication is running or done");
 424         aio_context_release(aio_context);
 425         return;
 426     }
 427
 428     if (s->mode != mode) {
 429         error_setg(errp, "The parameter mode's value is invalid, needs %d,"
 430                    " but got %d", s->mode, mode);
 431         aio_context_release(aio_context);
 432         return;
 433     }
 434
 435     switch (s->mode) {
 436     case REPLICATION_MODE_PRIMARY:
 437         break;
 438     case REPLICATION_MODE_SECONDARY:
 439         s->active_disk = bs->file;
 440         if (!s->active_disk || !s->active_disk->bs ||
 441                                     !s->active_disk->bs->backing) {
 442             error_setg(errp, "Active disk doesn't have backing file");
 443             aio_context_release(aio_context);
 444             return;
 445         }
 446
 447         s->hidden_disk = s->active_disk->bs->backing;
 448         if (!s->hidden_disk->bs || !s->hidden_disk->bs->backing) {
 449             error_setg(errp, "Hidden disk doesn't have backing file");
 450             aio_context_release(aio_context);
 451             return;
 452         }
 453
 454         s->secondary_disk = s->hidden_disk->bs->backing;
 455         if (!s->secondary_disk->bs || !bdrv_has_blk(s->secondary_disk->bs)) {
 456             error_setg(errp, "The secondary disk doesn't have block backend");
 457             aio_context_release(aio_context);
 458             return;
 459         }
 460
 461         /* verify the length */
 462         active_length = bdrv_getlength(s->active_disk->bs);
 463         hidden_length = bdrv_getlength(s->hidden_disk->bs);
 464         disk_length = bdrv_getlength(s->secondary_disk->bs);
 465         if (active_length < 0 || hidden_length < 0 || disk_length < 0 ||
 466             active_length != hidden_length || hidden_length != disk_length) {
 467             error_setg(errp, "Active disk, hidden disk, secondary disk's length"
 468                        " are not the same");
 469             aio_context_release(aio_context);
 470             return;
 471         }
 472
 473         if (!s->active_disk->bs->drv->bdrv_make_empty ||
 474             !s->hidden_disk->bs->drv->bdrv_make_empty) {
 475             error_setg(errp,
 476                        "Active disk or hidden disk doesn't support make_empty");
 477             aio_context_release(aio_context);
 478             return;
 479         }
 480
 481         /* reopen the backing file in r/w mode */
 482         reopen_backing_file(s, true, &local_err);
 483         if (local_err) {
 484             error_propagate(errp, local_err);
 485             aio_context_release(aio_context);
 486             return;
 487         }
 488
 489         /* start backup job now */
 490         error_setg(&s->blocker,
 491                    "Block device is in use by internal backup job");
 492
 493         top_bs = bdrv_lookup_bs(s->top_id, s->top_id, NULL);
 494         if (!top_bs || !bdrv_is_root_node(top_bs) ||
 495             !check_top_bs(top_bs, bs)) {
 496             error_setg(errp, "No top_bs or it is invalid");
 497             reopen_backing_file(s, false, NULL);
 498             aio_context_release(aio_context);
 499             return;
 500         }
 501         bdrv_op_block_all(top_bs, s->blocker);
 502         bdrv_op_unblock(top_bs, BLOCK_OP_TYPE_DATAPLANE, s->blocker);
 503
 504         backup_start("replication-backup", s->secondary_disk->bs,
 505                      s->hidden_disk->bs, 0, MIRROR_SYNC_MODE_NONE, NULL, false,
 506                      BLOCKDEV_ON_ERROR_REPORT, BLOCKDEV_ON_ERROR_REPORT,
 507                      backup_job_completed, s, NULL, &local_err);
 508         if (local_err) {
 509             error_propagate(errp, local_err);
 510             backup_job_cleanup(s);
 511             aio_context_release(aio_context);
 512             return;
 513         }
 514         break;
 515     default:
 516         aio_context_release(aio_context);
 517         abort();
 518     }
 519
 520     s->replication_state = BLOCK_REPLICATION_RUNNING;
 521
 522     if (s->mode == REPLICATION_MODE_SECONDARY) {
 523         secondary_do_checkpoint(s, errp);
 524     }
 525
 526     s->error = 0;
 527     aio_context_release(aio_context);
 528 }
 529
 530 static void replication_do_checkpoint(ReplicationState *rs, Error **errp)
 531 {
 532     BlockDriverState *bs = rs->opaque;
 533     BDRVReplicationState *s;
 534     AioContext *aio_context;
 535
 536     aio_context = bdrv_get_aio_context(bs);
 537     aio_context_acquire(aio_context);
 538     s = bs->opaque;
 539
 540     if (s->mode == REPLICATION_MODE_SECONDARY) {
 541         secondary_do_checkpoint(s, errp);
 542     }
 543     aio_context_release(aio_context);
 544 }
 545
 546 static void replication_get_error(ReplicationState *rs, Error **errp)
 547 {
 548     BlockDriverState *bs = rs->opaque;
 549     BDRVReplicationState *s;
 550     AioContext *aio_context;
 551
 552     aio_context = bdrv_get_aio_context(bs);
 553     aio_context_acquire(aio_context);
 554     s = bs->opaque;
 555
 556     if (s->replication_state != BLOCK_REPLICATION_RUNNING) {
 557         error_setg(errp, "Block replication is not running");
 558         aio_context_release(aio_context);
 559         return;
 560     }
 561
 562     if (s->error) {
 563         error_setg(errp, "I/O error occurred");
 564         aio_context_release(aio_context);
 565         return;
 566     }
 567     aio_context_release(aio_context);
 568 }
 569
 570 static void replication_done(void *opaque, int ret)
 571 {
 572     BlockDriverState *bs = opaque;
 573     BDRVReplicationState *s = bs->opaque;
 574
 575     if (ret == 0) {
 576         s->replication_state = BLOCK_REPLICATION_DONE;
 577
 578         /* refresh top bs's filename */
 579         bdrv_refresh_filename(bs);
 580         s->active_disk = NULL;
 581         s->secondary_disk = NULL;
 582         s->hidden_disk = NULL;
 583         s->error = 0;
 584     } else {
 585         s->replication_state = BLOCK_REPLICATION_FAILOVER_FAILED;
 586         s->error = -EIO;
 587     }
 588 }
 589
 590 static void replication_stop(ReplicationState *rs, bool failover, Error **errp)
 591 {
 592     BlockDriverState *bs = rs->opaque;
 593     BDRVReplicationState *s;
 594     AioContext *aio_context;
 595
 596     aio_context = bdrv_get_aio_context(bs);
 597     aio_context_acquire(aio_context);
 598     s = bs->opaque;
 599
 600     if (s->replication_state != BLOCK_REPLICATION_RUNNING) {
 601         error_setg(errp, "Block replication is not running");
 602         aio_context_release(aio_context);
 603         return;
 604     }
 605
 606     switch (s->mode) {
 607     case REPLICATION_MODE_PRIMARY:
 608         s->replication_state = BLOCK_REPLICATION_DONE;
 609         s->error = 0;
 610         break;
 611     case REPLICATION_MODE_SECONDARY:
 612         /*
 613          * This BDS will be closed, and the job should be completed
 614          * before the BDS is closed, because we will access hidden
 615          * disk, secondary disk in backup_job_completed().
 616          */
 617         if (s->secondary_disk->bs->job) {
 618             block_job_cancel_sync(s->secondary_disk->bs->job);
 619         }
 620
 621         if (!failover) {
 622             secondary_do_checkpoint(s, errp);
 623             s->replication_state = BLOCK_REPLICATION_DONE;
 624             aio_context_release(aio_context);
 625             return;
 626         }
 627
 628         s->replication_state = BLOCK_REPLICATION_FAILOVER;
 629         commit_active_start("replication-commit", s->active_disk->bs,
 630                             s->secondary_disk->bs, 0, BLOCKDEV_ON_ERROR_REPORT,
 631                             replication_done,
 632                             bs, errp, true);
 633         break;
 634     default:
 635         aio_context_release(aio_context);
 636         abort();
 637     }
 638     aio_context_release(aio_context);
 639 }
 640
 641 BlockDriver bdrv_replication = {
 642     .format_name                = "replication",
 643     .protocol_name              = "replication",
 644     .instance_size              = sizeof(BDRVReplicationState),
 645
 646     .bdrv_open                  = replication_open,
 647     .bdrv_close                 = replication_close,
 648
 649     .bdrv_getlength             = replication_getlength,
 650     .bdrv_co_readv              = replication_co_readv,
 651     .bdrv_co_writev             = replication_co_writev,
 652
 653     .is_filter                  = true,
 654     .bdrv_recurse_is_first_non_filter = replication_recurse_is_first_non_filter,
 655
 656     .has_variable_length        = true,
 657 };
 658
 659 static void bdrv_replication_init(void)
 660 {
 661     bdrv_register(&bdrv_replication);
 662 }
 663
 664 block_init(bdrv_replication_init);