block/mirror.c

   1 /*
   2  * Image mirroring
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Paolo Bonzini  <pbonzini@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
  10  * See the COPYING.LIB file in the top-level directory.
  11  *
  12  */
  13
  14 #include "trace.h"
  15 #include "block/blockjob.h"
  16 #include "block/block_int.h"
  17 #include "qemu/ratelimit.h"
  18 #include "qemu/bitmap.h"
  19
  20 #define SLICE_TIME    100000000ULL /* ns */
  21 #define MAX_IN_FLIGHT 16
  22
  23 /* The mirroring buffer is a list of granularity-sized chunks.
  24  * Free chunks are organized in a list.
  25  */
  26 typedef struct MirrorBuffer {
  27     QSIMPLEQ_ENTRY(MirrorBuffer) next;
  28 } MirrorBuffer;
  29
  30 typedef struct MirrorBlockJob {
  31     BlockJob common;
  32     RateLimit limit;
  33     BlockDriverState *target;
  34     BlockDriverState *base;
  35     /* The name of the graph node to replace */
  36     char *replaces;
  37     /* The BDS to replace */
  38     BlockDriverState *to_replace;
  39     /* Used to block operations on the drive-mirror-replace target */
  40     Error *replace_blocker;
  41     bool is_none_mode;
  42     BlockdevOnError on_source_error, on_target_error;
  43     bool synced;
  44     bool should_complete;
  45     int64_t sector_num;
  46     int64_t granularity;
  47     size_t buf_size;
  48     unsigned long *cow_bitmap;
  49     BdrvDirtyBitmap *dirty_bitmap;
  50     HBitmapIter hbi;
  51     uint8_t *buf;
  52     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
  53     int buf_free_count;
  54
  55     unsigned long *in_flight_bitmap;
  56     int in_flight;
  57     int ret;
  58 } MirrorBlockJob;
  59
  60 typedef struct MirrorOp {
  61     MirrorBlockJob *s;
  62     QEMUIOVector qiov;
  63     int64_t sector_num;
  64     int nb_sectors;
  65 } MirrorOp;
  66
  67 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  68                                             int error)
  69 {
  70     s->synced = false;
  71     if (read) {
  72         return block_job_error_action(&s->common, s->common.bs,
  73                                       s->on_source_error, true, error);
  74     } else {
  75         return block_job_error_action(&s->common, s->target,
  76                                       s->on_target_error, false, error);
  77     }
  78 }
  79
  80 static void mirror_iteration_done(MirrorOp *op, int ret)
  81 {
  82     MirrorBlockJob *s = op->s;
  83     struct iovec *iov;
  84     int64_t chunk_num;
  85     int i, nb_chunks, sectors_per_chunk;
  86
  87     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
  88
  89     s->in_flight--;
  90     iov = op->qiov.iov;
  91     for (i = 0; i < op->qiov.niov; i++) {
  92         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
  93         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
  94         s->buf_free_count++;
  95     }
  96
  97     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
  98     chunk_num = op->sector_num / sectors_per_chunk;
  99     nb_chunks = op->nb_sectors / sectors_per_chunk;
 100     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
 101     if (s->cow_bitmap && ret >= 0) {
 102         bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
 103     }
 104
 105     qemu_iovec_destroy(&op->qiov);
 106     g_slice_free(MirrorOp, op);
 107
 108     /* Enter coroutine when it is not sleeping.  The coroutine sleeps to
 109      * rate-limit itself.  The coroutine will eventually resume since there is
 110      * a sleep timeout so don't wake it early.
 111      */
 112     if (s->common.busy) {
 113         qemu_coroutine_enter(s->common.co, NULL);
 114     }
 115 }
 116
 117 static void mirror_write_complete(void *opaque, int ret)
 118 {
 119     MirrorOp *op = opaque;
 120     MirrorBlockJob *s = op->s;
 121     if (ret < 0) {
 122         BlockDriverState *source = s->common.bs;
 123         BlockErrorAction action;
 124
 125         bdrv_set_dirty(source, op->sector_num, op->nb_sectors);
 126         action = mirror_error_action(s, false, -ret);
 127         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 128             s->ret = ret;
 129         }
 130     }
 131     mirror_iteration_done(op, ret);
 132 }
 133
 134 static void mirror_read_complete(void *opaque, int ret)
 135 {
 136     MirrorOp *op = opaque;
 137     MirrorBlockJob *s = op->s;
 138     if (ret < 0) {
 139         BlockDriverState *source = s->common.bs;
 140         BlockErrorAction action;
 141
 142         bdrv_set_dirty(source, op->sector_num, op->nb_sectors);
 143         action = mirror_error_action(s, true, -ret);
 144         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 145             s->ret = ret;
 146         }
 147
 148         mirror_iteration_done(op, ret);
 149         return;
 150     }
 151     bdrv_aio_writev(s->target, op->sector_num, &op->qiov, op->nb_sectors,
 152                     mirror_write_complete, op);
 153 }
 154
 155 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
 156 {
 157     BlockDriverState *source = s->common.bs;
 158     int nb_sectors, sectors_per_chunk, nb_chunks;
 159     int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector;
 160     uint64_t delay_ns = 0;
 161     MirrorOp *op;
 162
 163     s->sector_num = hbitmap_iter_next(&s->hbi);
 164     if (s->sector_num < 0) {
 165         bdrv_dirty_iter_init(source, s->dirty_bitmap, &s->hbi);
 166         s->sector_num = hbitmap_iter_next(&s->hbi);
 167         trace_mirror_restart_iter(s,
 168                                   bdrv_get_dirty_count(source, s->dirty_bitmap));
 169         assert(s->sector_num >= 0);
 170     }
 171
 172     hbitmap_next_sector = s->sector_num;
 173     sector_num = s->sector_num;
 174     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 175     end = s->common.len >> BDRV_SECTOR_BITS;
 176
 177     /* Extend the QEMUIOVector to include all adjacent blocks that will
 178      * be copied in this operation.
 179      *
 180      * We have to do this if we have no backing file yet in the destination,
 181      * and the cluster size is very large.  Then we need to do COW ourselves.
 182      * The first time a cluster is copied, copy it entirely.  Note that,
 183      * because both the granularity and the cluster size are powers of two,
 184      * the number of sectors to copy cannot exceed one cluster.
 185      *
 186      * We also want to extend the QEMUIOVector to include more adjacent
 187      * dirty blocks if possible, to limit the number of I/O operations and
 188      * run efficiently even with a small granularity.
 189      */
 190     nb_chunks = 0;
 191     nb_sectors = 0;
 192     next_sector = sector_num;
 193     next_chunk = sector_num / sectors_per_chunk;
 194
 195     /* Wait for I/O to this cluster (from a previous iteration) to be done.  */
 196     while (test_bit(next_chunk, s->in_flight_bitmap)) {
 197         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 198         qemu_coroutine_yield();
 199     }
 200
 201     do {
 202         int added_sectors, added_chunks;
 203
 204         if (!bdrv_get_dirty(source, s->dirty_bitmap, next_sector) ||
 205             test_bit(next_chunk, s->in_flight_bitmap)) {
 206             assert(nb_sectors > 0);
 207             break;
 208         }
 209
 210         added_sectors = sectors_per_chunk;
 211         if (s->cow_bitmap && !test_bit(next_chunk, s->cow_bitmap)) {
 212             bdrv_round_to_clusters(s->target,
 213                                    next_sector, added_sectors,
 214                                    &next_sector, &added_sectors);
 215
 216             /* On the first iteration, the rounding may make us copy
 217              * sectors before the first dirty one.
 218              */
 219             if (next_sector < sector_num) {
 220                 assert(nb_sectors == 0);
 221                 sector_num = next_sector;
 222                 next_chunk = next_sector / sectors_per_chunk;
 223             }
 224         }
 225
 226         added_sectors = MIN(added_sectors, end - (sector_num + nb_sectors));
 227         added_chunks = (added_sectors + sectors_per_chunk - 1) / sectors_per_chunk;
 228
 229         /* When doing COW, it may happen that there is not enough space for
 230          * a full cluster.  Wait if that is the case.
 231          */
 232         while (nb_chunks == 0 && s->buf_free_count < added_chunks) {
 233             trace_mirror_yield_buf_busy(s, nb_chunks, s->in_flight);
 234             qemu_coroutine_yield();
 235         }
 236         if (s->buf_free_count < nb_chunks + added_chunks) {
 237             trace_mirror_break_buf_busy(s, nb_chunks, s->in_flight);
 238             break;
 239         }
 240
 241         /* We have enough free space to copy these sectors.  */
 242         bitmap_set(s->in_flight_bitmap, next_chunk, added_chunks);
 243
 244         nb_sectors += added_sectors;
 245         nb_chunks += added_chunks;
 246         next_sector += added_sectors;
 247         next_chunk += added_chunks;
 248         if (!s->synced && s->common.speed) {
 249             delay_ns = ratelimit_calculate_delay(&s->limit, added_sectors);
 250         }
 251     } while (delay_ns == 0 && next_sector < end);
 252
 253     /* Allocate a MirrorOp that is used as an AIO callback.  */
 254     op = g_slice_new(MirrorOp);
 255     op->s = s;
 256     op->sector_num = sector_num;
 257     op->nb_sectors = nb_sectors;
 258
 259     /* Now make a QEMUIOVector taking enough granularity-sized chunks
 260      * from s->buf_free.
 261      */
 262     qemu_iovec_init(&op->qiov, nb_chunks);
 263     next_sector = sector_num;
 264     while (nb_chunks-- > 0) {
 265         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
 266         size_t remaining = (nb_sectors * BDRV_SECTOR_SIZE) - op->qiov.size;
 267
 268         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
 269         s->buf_free_count--;
 270         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
 271
 272         /* Advance the HBitmapIter in parallel, so that we do not examine
 273          * the same sector twice.
 274          */
 275         if (next_sector > hbitmap_next_sector
 276             && bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
 277             hbitmap_next_sector = hbitmap_iter_next(&s->hbi);
 278         }
 279
 280         next_sector += sectors_per_chunk;
 281     }
 282
 283     bdrv_reset_dirty(source, sector_num, nb_sectors);
 284
 285     /* Copy the dirty cluster.  */
 286     s->in_flight++;
 287     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 288     bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
 289                    mirror_read_complete, op);
 290     return delay_ns;
 291 }
 292
 293 static void mirror_free_init(MirrorBlockJob *s)
 294 {
 295     int granularity = s->granularity;
 296     size_t buf_size = s->buf_size;
 297     uint8_t *buf = s->buf;
 298
 299     assert(s->buf_free_count == 0);
 300     QSIMPLEQ_INIT(&s->buf_free);
 301     while (buf_size != 0) {
 302         MirrorBuffer *cur = (MirrorBuffer *)buf;
 303         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 304         s->buf_free_count++;
 305         buf_size -= granularity;
 306         buf += granularity;
 307     }
 308 }
 309
 310 static void mirror_drain(MirrorBlockJob *s)
 311 {
 312     while (s->in_flight > 0) {
 313         qemu_coroutine_yield();
 314     }
 315 }
 316
 317 static void coroutine_fn mirror_run(void *opaque)
 318 {
 319     MirrorBlockJob *s = opaque;
 320     BlockDriverState *bs = s->common.bs;
 321     int64_t sector_num, end, sectors_per_chunk, length;
 322     uint64_t last_pause_ns;
 323     BlockDriverInfo bdi;
 324     char backing_filename[1024];
 325     int ret = 0;
 326     int n;
 327
 328     if (block_job_is_cancelled(&s->common)) {
 329         goto immediate_exit;
 330     }
 331
 332     s->common.len = bdrv_getlength(bs);
 333     if (s->common.len < 0) {
 334         ret = s->common.len;
 335         goto immediate_exit;
 336     } else if (s->common.len == 0) {
 337         /* Report BLOCK_JOB_READY and wait for complete. */
 338         block_job_event_ready(&s->common);
 339         s->synced = true;
 340         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
 341             block_job_yield(&s->common);
 342         }
 343         s->common.cancelled = false;
 344         goto immediate_exit;
 345     }
 346
 347     length = DIV_ROUND_UP(s->common.len, s->granularity);
 348     s->in_flight_bitmap = bitmap_new(length);
 349
 350     /* If we have no backing file yet in the destination, we cannot let
 351      * the destination do COW.  Instead, we copy sectors around the
 352      * dirty data if needed.  We need a bitmap to do that.
 353      */
 354     bdrv_get_backing_filename(s->target, backing_filename,
 355                               sizeof(backing_filename));
 356     if (backing_filename[0] && !s->target->backing_hd) {
 357         ret = bdrv_get_info(s->target, &bdi);
 358         if (ret < 0) {
 359             goto immediate_exit;
 360         }
 361         if (s->granularity < bdi.cluster_size) {
 362             s->buf_size = MAX(s->buf_size, bdi.cluster_size);
 363             s->cow_bitmap = bitmap_new(length);
 364         }
 365     }
 366
 367     end = s->common.len >> BDRV_SECTOR_BITS;
 368     s->buf = qemu_try_blockalign(bs, s->buf_size);
 369     if (s->buf == NULL) {
 370         ret = -ENOMEM;
 371         goto immediate_exit;
 372     }
 373
 374     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 375     mirror_free_init(s);
 376
 377     if (!s->is_none_mode) {
 378         /* First part, loop on the sectors and initialize the dirty bitmap.  */
 379         BlockDriverState *base = s->base;
 380         for (sector_num = 0; sector_num < end; ) {
 381             int64_t next = (sector_num | (sectors_per_chunk - 1)) + 1;
 382             ret = bdrv_is_allocated_above(bs, base,
 383                                           sector_num, next - sector_num, &n);
 384
 385             if (ret < 0) {
 386                 goto immediate_exit;
 387             }
 388
 389             assert(n > 0);
 390             if (ret == 1) {
 391                 bdrv_set_dirty(bs, sector_num, n);
 392                 sector_num = next;
 393             } else {
 394                 sector_num += n;
 395             }
 396         }
 397     }
 398
 399     bdrv_dirty_iter_init(bs, s->dirty_bitmap, &s->hbi);
 400     last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 401     for (;;) {
 402         uint64_t delay_ns = 0;
 403         int64_t cnt;
 404         bool should_complete;
 405
 406         if (s->ret < 0) {
 407             ret = s->ret;
 408             goto immediate_exit;
 409         }
 410
 411         cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 412
 413         /* Note that even when no rate limit is applied we need to yield
 414          * periodically with no pending I/O so that qemu_aio_flush() returns.
 415          * We do so every SLICE_TIME nanoseconds, or when there is an error,
 416          * or when the source is clean, whichever comes first.
 417          */
 418         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
 419             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 420             if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 421                 (cnt == 0 && s->in_flight > 0)) {
 422                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
 423                 qemu_coroutine_yield();
 424                 continue;
 425             } else if (cnt != 0) {
 426                 delay_ns = mirror_iteration(s);
 427                 if (delay_ns == 0) {
 428                     continue;
 429                 }
 430             }
 431         }
 432
 433         should_complete = false;
 434         if (s->in_flight == 0 && cnt == 0) {
 435             trace_mirror_before_flush(s);
 436             ret = bdrv_flush(s->target);
 437             if (ret < 0) {
 438                 if (mirror_error_action(s, false, -ret) ==
 439                     BLOCK_ERROR_ACTION_REPORT) {
 440                     goto immediate_exit;
 441                 }
 442             } else {
 443                 /* We're out of the streaming phase.  From now on, if the job
 444                  * is cancelled we will actually complete all pending I/O and
 445                  * report completion.  This way, block-job-cancel will leave
 446                  * the target in a consistent state.
 447                  */
 448                 s->common.offset = end * BDRV_SECTOR_SIZE;
 449                 if (!s->synced) {
 450                     block_job_event_ready(&s->common);
 451                     s->synced = true;
 452                 }
 453
 454                 should_complete = s->should_complete ||
 455                     block_job_is_cancelled(&s->common);
 456                 cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 457             }
 458         }
 459
 460         if (cnt == 0 && should_complete) {
 461             /* The dirty bitmap is not updated while operations are pending.
 462              * If we're about to exit, wait for pending operations before
 463              * calling bdrv_get_dirty_count(bs), or we may exit while the
 464              * source has dirty data to copy!
 465              *
 466              * Note that I/O can be submitted by the guest while
 467              * mirror_populate runs.
 468              */
 469             trace_mirror_before_drain(s, cnt);
 470             bdrv_drain_all();
 471             cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 472         }
 473
 474         ret = 0;
 475         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
 476         if (!s->synced) {
 477             /* Publish progress */
 478             s->common.offset = (end - cnt) * BDRV_SECTOR_SIZE;
 479             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 480             if (block_job_is_cancelled(&s->common)) {
 481                 break;
 482             }
 483         } else if (!should_complete) {
 484             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 485             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 486         } else if (cnt == 0) {
 487             /* The two disks are in sync.  Exit and report successful
 488              * completion.
 489              */
 490             assert(QLIST_EMPTY(&bs->tracked_requests));
 491             s->common.cancelled = false;
 492             break;
 493         }
 494         last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 495     }
 496
 497 immediate_exit:
 498     if (s->in_flight > 0) {
 499         /* We get here only if something went wrong.  Either the job failed,
 500          * or it was cancelled prematurely so that we do not guarantee that
 501          * the target is a copy of the source.
 502          */
 503         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 504         mirror_drain(s);
 505     }
 506
 507     assert(s->in_flight == 0);
 508     qemu_vfree(s->buf);
 509     g_free(s->cow_bitmap);
 510     g_free(s->in_flight_bitmap);
 511     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
 512     bdrv_iostatus_disable(s->target);
 513     if (s->should_complete && ret == 0) {
 514         BlockDriverState *to_replace = s->common.bs;
 515         if (s->to_replace) {
 516             to_replace = s->to_replace;
 517         }
 518         if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
 519             bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
 520         }
 521         bdrv_swap(s->target, to_replace);
 522         if (s->common.driver->job_type == BLOCK_JOB_TYPE_COMMIT) {
 523             /* drop the bs loop chain formed by the swap: break the loop then
 524              * trigger the unref from the top one */
 525             BlockDriverState *p = s->base->backing_hd;
 526             bdrv_set_backing_hd(s->base, NULL);
 527             bdrv_unref(p);
 528         }
 529     }
 530     if (s->to_replace) {
 531         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
 532         error_free(s->replace_blocker);
 533         bdrv_unref(s->to_replace);
 534     }
 535     g_free(s->replaces);
 536     bdrv_unref(s->target);
 537     block_job_completed(&s->common, ret);
 538 }
 539
 540 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 541 {
 542     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 543
 544     if (speed < 0) {
 545         error_set(errp, QERR_INVALID_PARAMETER, "speed");
 546         return;
 547     }
 548     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 549 }
 550
 551 static void mirror_iostatus_reset(BlockJob *job)
 552 {
 553     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 554
 555     bdrv_iostatus_reset(s->target);
 556 }
 557
 558 static void mirror_complete(BlockJob *job, Error **errp)
 559 {
 560     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 561     Error *local_err = NULL;
 562     int ret;
 563
 564     ret = bdrv_open_backing_file(s->target, NULL, &local_err);
 565     if (ret < 0) {
 566         error_propagate(errp, local_err);
 567         return;
 568     }
 569     if (!s->synced) {
 570         error_set(errp, QERR_BLOCK_JOB_NOT_READY,
 571                   bdrv_get_device_name(job->bs));
 572         return;
 573     }
 574
 575     /* check the target bs is not blocked and block all operations on it */
 576     if (s->replaces) {
 577         s->to_replace = check_to_replace_node(s->replaces, &local_err);
 578         if (!s->to_replace) {
 579             error_propagate(errp, local_err);
 580             return;
 581         }
 582
 583         error_setg(&s->replace_blocker,
 584                    "block device is in use by block-job-complete");
 585         bdrv_op_block_all(s->to_replace, s->replace_blocker);
 586         bdrv_ref(s->to_replace);
 587     }
 588
 589     s->should_complete = true;
 590     block_job_resume(job);
 591 }
 592
 593 static const BlockJobDriver mirror_job_driver = {
 594     .instance_size = sizeof(MirrorBlockJob),
 595     .job_type      = BLOCK_JOB_TYPE_MIRROR,
 596     .set_speed     = mirror_set_speed,
 597     .iostatus_reset= mirror_iostatus_reset,
 598     .complete      = mirror_complete,
 599 };
 600
 601 static const BlockJobDriver commit_active_job_driver = {
 602     .instance_size = sizeof(MirrorBlockJob),
 603     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 604     .set_speed     = mirror_set_speed,
 605     .iostatus_reset
 606                    = mirror_iostatus_reset,
 607     .complete      = mirror_complete,
 608 };
 609
 610 static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
 611                              const char *replaces,
 612                              int64_t speed, int64_t granularity,
 613                              int64_t buf_size,
 614                              BlockdevOnError on_source_error,
 615                              BlockdevOnError on_target_error,
 616                              BlockCompletionFunc *cb,
 617                              void *opaque, Error **errp,
 618                              const BlockJobDriver *driver,
 619                              bool is_none_mode, BlockDriverState *base)
 620 {
 621     MirrorBlockJob *s;
 622
 623     if (granularity == 0) {
 624         /* Choose the default granularity based on the target file's cluster
 625          * size, clamped between 4k and 64k.  */
 626         BlockDriverInfo bdi;
 627         if (bdrv_get_info(target, &bdi) >= 0 && bdi.cluster_size != 0) {
 628             granularity = MAX(4096, bdi.cluster_size);
 629             granularity = MIN(65536, granularity);
 630         } else {
 631             granularity = 65536;
 632         }
 633     }
 634
 635     assert ((granularity & (granularity - 1)) == 0);
 636
 637     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 638          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 639         !bdrv_iostatus_is_enabled(bs)) {
 640         error_set(errp, QERR_INVALID_PARAMETER, "on-source-error");
 641         return;
 642     }
 643
 644
 645     s = block_job_create(driver, bs, speed, cb, opaque, errp);
 646     if (!s) {
 647         return;
 648     }
 649
 650     s->replaces = g_strdup(replaces);
 651     s->on_source_error = on_source_error;
 652     s->on_target_error = on_target_error;
 653     s->target = target;
 654     s->is_none_mode = is_none_mode;
 655     s->base = base;
 656     s->granularity = granularity;
 657     s->buf_size = MAX(buf_size, granularity);
 658
 659     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, errp);
 660     if (!s->dirty_bitmap) {
 661         return;
 662     }
 663     bdrv_set_enable_write_cache(s->target, true);
 664     bdrv_set_on_error(s->target, on_target_error, on_target_error);
 665     bdrv_iostatus_enable(s->target);
 666     s->common.co = qemu_coroutine_create(mirror_run);
 667     trace_mirror_start(bs, s, s->common.co, opaque);
 668     qemu_coroutine_enter(s->common.co, s);
 669 }
 670
 671 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
 672                   const char *replaces,
 673                   int64_t speed, int64_t granularity, int64_t buf_size,
 674                   MirrorSyncMode mode, BlockdevOnError on_source_error,
 675                   BlockdevOnError on_target_error,
 676                   BlockCompletionFunc *cb,
 677                   void *opaque, Error **errp)
 678 {
 679     bool is_none_mode;
 680     BlockDriverState *base;
 681
 682     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
 683     base = mode == MIRROR_SYNC_MODE_TOP ? bs->backing_hd : NULL;
 684     mirror_start_job(bs, target, replaces,
 685                      speed, granularity, buf_size,
 686                      on_source_error, on_target_error, cb, opaque, errp,
 687                      &mirror_job_driver, is_none_mode, base);
 688 }
 689
 690 void commit_active_start(BlockDriverState *bs, BlockDriverState *base,
 691                          int64_t speed,
 692                          BlockdevOnError on_error,
 693                          BlockCompletionFunc *cb,
 694                          void *opaque, Error **errp)
 695 {
 696     int64_t length, base_length;
 697     int orig_base_flags;
 698     int ret;
 699     Error *local_err = NULL;
 700
 701     orig_base_flags = bdrv_get_flags(base);
 702
 703     if (bdrv_reopen(base, bs->open_flags, errp)) {
 704         return;
 705     }
 706
 707     length = bdrv_getlength(bs);
 708     if (length < 0) {
 709         error_setg_errno(errp, -length,
 710                          "Unable to determine length of %s", bs->filename);
 711         goto error_restore_flags;
 712     }
 713
 714     base_length = bdrv_getlength(base);
 715     if (base_length < 0) {
 716         error_setg_errno(errp, -base_length,
 717                          "Unable to determine length of %s", base->filename);
 718         goto error_restore_flags;
 719     }
 720
 721     if (length > base_length) {
 722         ret = bdrv_truncate(base, length);
 723         if (ret < 0) {
 724             error_setg_errno(errp, -ret,
 725                             "Top image %s is larger than base image %s, and "
 726                              "resize of base image failed",
 727                              bs->filename, base->filename);
 728             goto error_restore_flags;
 729         }
 730     }
 731
 732     bdrv_ref(base);
 733     mirror_start_job(bs, base, NULL, speed, 0, 0,
 734                      on_error, on_error, cb, opaque, &local_err,
 735                      &commit_active_job_driver, false, base);
 736     if (local_err) {
 737         error_propagate(errp, local_err);
 738         goto error_restore_flags;
 739     }
 740
 741     return;
 742
 743 error_restore_flags:
 744     /* ignore error and errp for bdrv_reopen, because we want to propagate
 745      * the original error */
 746     bdrv_reopen(base, orig_base_flags, NULL);
 747     return;
 748 }