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;
 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         } else {
 251             delay_ns = 0;
 252         }
 253     } while (delay_ns == 0 && next_sector < end);
 254
 255     /* Allocate a MirrorOp that is used as an AIO callback.  */
 256     op = g_slice_new(MirrorOp);
 257     op->s = s;
 258     op->sector_num = sector_num;
 259     op->nb_sectors = nb_sectors;
 260
 261     /* Now make a QEMUIOVector taking enough granularity-sized chunks
 262      * from s->buf_free.
 263      */
 264     qemu_iovec_init(&op->qiov, nb_chunks);
 265     next_sector = sector_num;
 266     while (nb_chunks-- > 0) {
 267         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
 268         size_t remaining = (nb_sectors * BDRV_SECTOR_SIZE) - op->qiov.size;
 269
 270         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
 271         s->buf_free_count--;
 272         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
 273
 274         /* Advance the HBitmapIter in parallel, so that we do not examine
 275          * the same sector twice.
 276          */
 277         if (next_sector > hbitmap_next_sector
 278             && bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
 279             hbitmap_next_sector = hbitmap_iter_next(&s->hbi);
 280         }
 281
 282         next_sector += sectors_per_chunk;
 283     }
 284
 285     bdrv_reset_dirty(source, sector_num, nb_sectors);
 286
 287     /* Copy the dirty cluster.  */
 288     s->in_flight++;
 289     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 290     bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
 291                    mirror_read_complete, op);
 292     return delay_ns;
 293 }
 294
 295 static void mirror_free_init(MirrorBlockJob *s)
 296 {
 297     int granularity = s->granularity;
 298     size_t buf_size = s->buf_size;
 299     uint8_t *buf = s->buf;
 300
 301     assert(s->buf_free_count == 0);
 302     QSIMPLEQ_INIT(&s->buf_free);
 303     while (buf_size != 0) {
 304         MirrorBuffer *cur = (MirrorBuffer *)buf;
 305         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 306         s->buf_free_count++;
 307         buf_size -= granularity;
 308         buf += granularity;
 309     }
 310 }
 311
 312 static void mirror_drain(MirrorBlockJob *s)
 313 {
 314     while (s->in_flight > 0) {
 315         qemu_coroutine_yield();
 316     }
 317 }
 318
 319 static void coroutine_fn mirror_run(void *opaque)
 320 {
 321     MirrorBlockJob *s = opaque;
 322     BlockDriverState *bs = s->common.bs;
 323     int64_t sector_num, end, sectors_per_chunk, length;
 324     uint64_t last_pause_ns;
 325     BlockDriverInfo bdi;
 326     char backing_filename[1024];
 327     int ret = 0;
 328     int n;
 329
 330     if (block_job_is_cancelled(&s->common)) {
 331         goto immediate_exit;
 332     }
 333
 334     s->common.len = bdrv_getlength(bs);
 335     if (s->common.len < 0) {
 336         ret = s->common.len;
 337         goto immediate_exit;
 338     } else if (s->common.len == 0) {
 339         /* Report BLOCK_JOB_READY and wait for complete. */
 340         block_job_event_ready(&s->common);
 341         s->synced = true;
 342         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
 343             block_job_yield(&s->common);
 344         }
 345         s->common.cancelled = false;
 346         goto immediate_exit;
 347     }
 348
 349     length = DIV_ROUND_UP(s->common.len, s->granularity);
 350     s->in_flight_bitmap = bitmap_new(length);
 351
 352     /* If we have no backing file yet in the destination, we cannot let
 353      * the destination do COW.  Instead, we copy sectors around the
 354      * dirty data if needed.  We need a bitmap to do that.
 355      */
 356     bdrv_get_backing_filename(s->target, backing_filename,
 357                               sizeof(backing_filename));
 358     if (backing_filename[0] && !s->target->backing_hd) {
 359         ret = bdrv_get_info(s->target, &bdi);
 360         if (ret < 0) {
 361             goto immediate_exit;
 362         }
 363         if (s->granularity < bdi.cluster_size) {
 364             s->buf_size = MAX(s->buf_size, bdi.cluster_size);
 365             s->cow_bitmap = bitmap_new(length);
 366         }
 367     }
 368
 369     end = s->common.len >> BDRV_SECTOR_BITS;
 370     s->buf = qemu_blockalign(bs, s->buf_size);
 371     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 372     mirror_free_init(s);
 373
 374     if (!s->is_none_mode) {
 375         /* First part, loop on the sectors and initialize the dirty bitmap.  */
 376         BlockDriverState *base = s->base;
 377         for (sector_num = 0; sector_num < end; ) {
 378             int64_t next = (sector_num | (sectors_per_chunk - 1)) + 1;
 379             ret = bdrv_is_allocated_above(bs, base,
 380                                           sector_num, next - sector_num, &n);
 381
 382             if (ret < 0) {
 383                 goto immediate_exit;
 384             }
 385
 386             assert(n > 0);
 387             if (ret == 1) {
 388                 bdrv_set_dirty(bs, sector_num, n);
 389                 sector_num = next;
 390             } else {
 391                 sector_num += n;
 392             }
 393         }
 394     }
 395
 396     bdrv_dirty_iter_init(bs, s->dirty_bitmap, &s->hbi);
 397     last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 398     for (;;) {
 399         uint64_t delay_ns = 0;
 400         int64_t cnt;
 401         bool should_complete;
 402
 403         if (s->ret < 0) {
 404             ret = s->ret;
 405             goto immediate_exit;
 406         }
 407
 408         cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 409
 410         /* Note that even when no rate limit is applied we need to yield
 411          * periodically with no pending I/O so that qemu_aio_flush() returns.
 412          * We do so every SLICE_TIME nanoseconds, or when there is an error,
 413          * or when the source is clean, whichever comes first.
 414          */
 415         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
 416             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 417             if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 418                 (cnt == 0 && s->in_flight > 0)) {
 419                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
 420                 qemu_coroutine_yield();
 421                 continue;
 422             } else if (cnt != 0) {
 423                 delay_ns = mirror_iteration(s);
 424                 if (delay_ns == 0) {
 425                     continue;
 426                 }
 427             }
 428         }
 429
 430         should_complete = false;
 431         if (s->in_flight == 0 && cnt == 0) {
 432             trace_mirror_before_flush(s);
 433             ret = bdrv_flush(s->target);
 434             if (ret < 0) {
 435                 if (mirror_error_action(s, false, -ret) ==
 436                     BLOCK_ERROR_ACTION_REPORT) {
 437                     goto immediate_exit;
 438                 }
 439             } else {
 440                 /* We're out of the streaming phase.  From now on, if the job
 441                  * is cancelled we will actually complete all pending I/O and
 442                  * report completion.  This way, block-job-cancel will leave
 443                  * the target in a consistent state.
 444                  */
 445                 s->common.offset = end * BDRV_SECTOR_SIZE;
 446                 if (!s->synced) {
 447                     block_job_event_ready(&s->common);
 448                     s->synced = true;
 449                 }
 450
 451                 should_complete = s->should_complete ||
 452                     block_job_is_cancelled(&s->common);
 453                 cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 454             }
 455         }
 456
 457         if (cnt == 0 && should_complete) {
 458             /* The dirty bitmap is not updated while operations are pending.
 459              * If we're about to exit, wait for pending operations before
 460              * calling bdrv_get_dirty_count(bs), or we may exit while the
 461              * source has dirty data to copy!
 462              *
 463              * Note that I/O can be submitted by the guest while
 464              * mirror_populate runs.
 465              */
 466             trace_mirror_before_drain(s, cnt);
 467             bdrv_drain_all();
 468             cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 469         }
 470
 471         ret = 0;
 472         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
 473         if (!s->synced) {
 474             /* Publish progress */
 475             s->common.offset = (end - cnt) * BDRV_SECTOR_SIZE;
 476             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 477             if (block_job_is_cancelled(&s->common)) {
 478                 break;
 479             }
 480         } else if (!should_complete) {
 481             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 482             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 483         } else if (cnt == 0) {
 484             /* The two disks are in sync.  Exit and report successful
 485              * completion.
 486              */
 487             assert(QLIST_EMPTY(&bs->tracked_requests));
 488             s->common.cancelled = false;
 489             break;
 490         }
 491         last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 492     }
 493
 494 immediate_exit:
 495     if (s->in_flight > 0) {
 496         /* We get here only if something went wrong.  Either the job failed,
 497          * or it was cancelled prematurely so that we do not guarantee that
 498          * the target is a copy of the source.
 499          */
 500         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 501         mirror_drain(s);
 502     }
 503
 504     assert(s->in_flight == 0);
 505     qemu_vfree(s->buf);
 506     g_free(s->cow_bitmap);
 507     g_free(s->in_flight_bitmap);
 508     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
 509     bdrv_iostatus_disable(s->target);
 510     if (s->should_complete && ret == 0) {
 511         BlockDriverState *to_replace = s->common.bs;
 512         if (s->to_replace) {
 513             to_replace = s->to_replace;
 514         }
 515         if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
 516             bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
 517         }
 518         bdrv_swap(s->target, to_replace);
 519         if (s->common.driver->job_type == BLOCK_JOB_TYPE_COMMIT) {
 520             /* drop the bs loop chain formed by the swap: break the loop then
 521              * trigger the unref from the top one */
 522             BlockDriverState *p = s->base->backing_hd;
 523             bdrv_set_backing_hd(s->base, NULL);
 524             bdrv_unref(p);
 525         }
 526     }
 527     if (s->to_replace) {
 528         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
 529         error_free(s->replace_blocker);
 530         bdrv_unref(s->to_replace);
 531     }
 532     g_free(s->replaces);
 533     bdrv_unref(s->target);
 534     block_job_completed(&s->common, ret);
 535 }
 536
 537 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 538 {
 539     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 540
 541     if (speed < 0) {
 542         error_set(errp, QERR_INVALID_PARAMETER, "speed");
 543         return;
 544     }
 545     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 546 }
 547
 548 static void mirror_iostatus_reset(BlockJob *job)
 549 {
 550     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 551
 552     bdrv_iostatus_reset(s->target);
 553 }
 554
 555 static void mirror_complete(BlockJob *job, Error **errp)
 556 {
 557     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 558     Error *local_err = NULL;
 559     int ret;
 560
 561     ret = bdrv_open_backing_file(s->target, NULL, &local_err);
 562     if (ret < 0) {
 563         error_propagate(errp, local_err);
 564         return;
 565     }
 566     if (!s->synced) {
 567         error_set(errp, QERR_BLOCK_JOB_NOT_READY, job->bs->device_name);
 568         return;
 569     }
 570
 571     /* check the target bs is not blocked and block all operations on it */
 572     if (s->replaces) {
 573         s->to_replace = check_to_replace_node(s->replaces, &local_err);
 574         if (!s->to_replace) {
 575             error_propagate(errp, local_err);
 576             return;
 577         }
 578
 579         error_setg(&s->replace_blocker,
 580                    "block device is in use by block-job-complete");
 581         bdrv_op_block_all(s->to_replace, s->replace_blocker);
 582         bdrv_ref(s->to_replace);
 583     }
 584
 585     s->should_complete = true;
 586     block_job_resume(job);
 587 }
 588
 589 static const BlockJobDriver mirror_job_driver = {
 590     .instance_size = sizeof(MirrorBlockJob),
 591     .job_type      = BLOCK_JOB_TYPE_MIRROR,
 592     .set_speed     = mirror_set_speed,
 593     .iostatus_reset= mirror_iostatus_reset,
 594     .complete      = mirror_complete,
 595 };
 596
 597 static const BlockJobDriver commit_active_job_driver = {
 598     .instance_size = sizeof(MirrorBlockJob),
 599     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 600     .set_speed     = mirror_set_speed,
 601     .iostatus_reset
 602                    = mirror_iostatus_reset,
 603     .complete      = mirror_complete,
 604 };
 605
 606 static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
 607                              const char *replaces,
 608                              int64_t speed, int64_t granularity,
 609                              int64_t buf_size,
 610                              BlockdevOnError on_source_error,
 611                              BlockdevOnError on_target_error,
 612                              BlockDriverCompletionFunc *cb,
 613                              void *opaque, Error **errp,
 614                              const BlockJobDriver *driver,
 615                              bool is_none_mode, BlockDriverState *base)
 616 {
 617     MirrorBlockJob *s;
 618
 619     if (granularity == 0) {
 620         /* Choose the default granularity based on the target file's cluster
 621          * size, clamped between 4k and 64k.  */
 622         BlockDriverInfo bdi;
 623         if (bdrv_get_info(target, &bdi) >= 0 && bdi.cluster_size != 0) {
 624             granularity = MAX(4096, bdi.cluster_size);
 625             granularity = MIN(65536, granularity);
 626         } else {
 627             granularity = 65536;
 628         }
 629     }
 630
 631     assert ((granularity & (granularity - 1)) == 0);
 632
 633     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 634          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 635         !bdrv_iostatus_is_enabled(bs)) {
 636         error_set(errp, QERR_INVALID_PARAMETER, "on-source-error");
 637         return;
 638     }
 639
 640
 641     s = block_job_create(driver, bs, speed, cb, opaque, errp);
 642     if (!s) {
 643         return;
 644     }
 645
 646     s->replaces = g_strdup(replaces);
 647     s->on_source_error = on_source_error;
 648     s->on_target_error = on_target_error;
 649     s->target = target;
 650     s->is_none_mode = is_none_mode;
 651     s->base = base;
 652     s->granularity = granularity;
 653     s->buf_size = MAX(buf_size, granularity);
 654
 655     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, errp);
 656     if (!s->dirty_bitmap) {
 657         return;
 658     }
 659     bdrv_set_enable_write_cache(s->target, true);
 660     bdrv_set_on_error(s->target, on_target_error, on_target_error);
 661     bdrv_iostatus_enable(s->target);
 662     s->common.co = qemu_coroutine_create(mirror_run);
 663     trace_mirror_start(bs, s, s->common.co, opaque);
 664     qemu_coroutine_enter(s->common.co, s);
 665 }
 666
 667 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
 668                   const char *replaces,
 669                   int64_t speed, int64_t granularity, int64_t buf_size,
 670                   MirrorSyncMode mode, BlockdevOnError on_source_error,
 671                   BlockdevOnError on_target_error,
 672                   BlockDriverCompletionFunc *cb,
 673                   void *opaque, Error **errp)
 674 {
 675     bool is_none_mode;
 676     BlockDriverState *base;
 677
 678     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
 679     base = mode == MIRROR_SYNC_MODE_TOP ? bs->backing_hd : NULL;
 680     mirror_start_job(bs, target, replaces,
 681                      speed, granularity, buf_size,
 682                      on_source_error, on_target_error, cb, opaque, errp,
 683                      &mirror_job_driver, is_none_mode, base);
 684 }
 685
 686 void commit_active_start(BlockDriverState *bs, BlockDriverState *base,
 687                          int64_t speed,
 688                          BlockdevOnError on_error,
 689                          BlockDriverCompletionFunc *cb,
 690                          void *opaque, Error **errp)
 691 {
 692     int64_t length, base_length;
 693     int orig_base_flags;
 694     int ret;
 695     Error *local_err = NULL;
 696
 697     orig_base_flags = bdrv_get_flags(base);
 698
 699     if (bdrv_reopen(base, bs->open_flags, errp)) {
 700         return;
 701     }
 702
 703     length = bdrv_getlength(bs);
 704     if (length < 0) {
 705         error_setg_errno(errp, -length,
 706                          "Unable to determine length of %s", bs->filename);
 707         goto error_restore_flags;
 708     }
 709
 710     base_length = bdrv_getlength(base);
 711     if (base_length < 0) {
 712         error_setg_errno(errp, -base_length,
 713                          "Unable to determine length of %s", base->filename);
 714         goto error_restore_flags;
 715     }
 716
 717     if (length > base_length) {
 718         ret = bdrv_truncate(base, length);
 719         if (ret < 0) {
 720             error_setg_errno(errp, -ret,
 721                             "Top image %s is larger than base image %s, and "
 722                              "resize of base image failed",
 723                              bs->filename, base->filename);
 724             goto error_restore_flags;
 725         }
 726     }
 727
 728     bdrv_ref(base);
 729     mirror_start_job(bs, base, NULL, speed, 0, 0,
 730                      on_error, on_error, cb, opaque, &local_err,
 731                      &commit_active_job_driver, false, base);
 732     if (local_err) {
 733         error_propagate(errp, local_err);
 734         goto error_restore_flags;
 735     }
 736
 737     return;
 738
 739 error_restore_flags:
 740     /* ignore error and errp for bdrv_reopen, because we want to propagate
 741      * the original error */
 742     bdrv_reopen(base, orig_base_flags, NULL);
 743     return;
 744 }