block/mirror.c

   1 /*
   2  * Image mirroring
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Paolo Bonzini  <[email protected]>
   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 "qapi/qmp/qerror.h"
  18 #include "qemu/ratelimit.h"
  19 #include "qemu/bitmap.h"
  20
  21 #define SLICE_TIME    100000000ULL /* ns */
  22 #define MAX_IN_FLIGHT 16
  23
  24 /* The mirroring buffer is a list of granularity-sized chunks.
  25  * Free chunks are organized in a list.
  26  */
  27 typedef struct MirrorBuffer {
  28     QSIMPLEQ_ENTRY(MirrorBuffer) next;
  29 } MirrorBuffer;
  30
  31 typedef struct MirrorBlockJob {
  32     BlockJob common;
  33     RateLimit limit;
  34     BlockDriverState *target;
  35     BlockDriverState *base;
  36     /* The name of the graph node to replace */
  37     char *replaces;
  38     /* The BDS to replace */
  39     BlockDriverState *to_replace;
  40     /* Used to block operations on the drive-mirror-replace target */
  41     Error *replace_blocker;
  42     bool is_none_mode;
  43     BlockdevOnError on_source_error, on_target_error;
  44     bool synced;
  45     bool should_complete;
  46     int64_t sector_num;
  47     int64_t granularity;
  48     size_t buf_size;
  49     int64_t bdev_length;
  50     unsigned long *cow_bitmap;
  51     BdrvDirtyBitmap *dirty_bitmap;
  52     HBitmapIter hbi;
  53     uint8_t *buf;
  54     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
  55     int buf_free_count;
  56
  57     unsigned long *in_flight_bitmap;
  58     int in_flight;
  59     int sectors_in_flight;
  60     int ret;
  61 } MirrorBlockJob;
  62
  63 typedef struct MirrorOp {
  64     MirrorBlockJob *s;
  65     QEMUIOVector qiov;
  66     int64_t sector_num;
  67     int nb_sectors;
  68 } MirrorOp;
  69
  70 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  71                                             int error)
  72 {
  73     s->synced = false;
  74     if (read) {
  75         return block_job_error_action(&s->common, s->common.bs,
  76                                       s->on_source_error, true, error);
  77     } else {
  78         return block_job_error_action(&s->common, s->target,
  79                                       s->on_target_error, false, error);
  80     }
  81 }
  82
  83 static void mirror_iteration_done(MirrorOp *op, int ret)
  84 {
  85     MirrorBlockJob *s = op->s;
  86     struct iovec *iov;
  87     int64_t chunk_num;
  88     int i, nb_chunks, sectors_per_chunk;
  89
  90     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
  91
  92     s->in_flight--;
  93     s->sectors_in_flight -= op->nb_sectors;
  94     iov = op->qiov.iov;
  95     for (i = 0; i < op->qiov.niov; i++) {
  96         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
  97         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
  98         s->buf_free_count++;
  99     }
 100
 101     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 102     chunk_num = op->sector_num / sectors_per_chunk;
 103     nb_chunks = op->nb_sectors / sectors_per_chunk;
 104     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
 105     if (ret >= 0) {
 106         if (s->cow_bitmap) {
 107             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
 108         }
 109         s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
 110     }
 111
 112     qemu_iovec_destroy(&op->qiov);
 113     g_slice_free(MirrorOp, op);
 114
 115     /* Enter coroutine when it is not sleeping.  The coroutine sleeps to
 116      * rate-limit itself.  The coroutine will eventually resume since there is
 117      * a sleep timeout so don't wake it early.
 118      */
 119     if (s->common.busy) {
 120         qemu_coroutine_enter(s->common.co, NULL);
 121     }
 122 }
 123
 124 static void mirror_write_complete(void *opaque, int ret)
 125 {
 126     MirrorOp *op = opaque;
 127     MirrorBlockJob *s = op->s;
 128     if (ret < 0) {
 129         BlockErrorAction action;
 130
 131         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
 132         action = mirror_error_action(s, false, -ret);
 133         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 134             s->ret = ret;
 135         }
 136     }
 137     mirror_iteration_done(op, ret);
 138 }
 139
 140 static void mirror_read_complete(void *opaque, int ret)
 141 {
 142     MirrorOp *op = opaque;
 143     MirrorBlockJob *s = op->s;
 144     if (ret < 0) {
 145         BlockErrorAction action;
 146
 147         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
 148         action = mirror_error_action(s, true, -ret);
 149         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 150             s->ret = ret;
 151         }
 152
 153         mirror_iteration_done(op, ret);
 154         return;
 155     }
 156     bdrv_aio_writev(s->target, op->sector_num, &op->qiov, op->nb_sectors,
 157                     mirror_write_complete, op);
 158 }
 159
 160 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
 161 {
 162     BlockDriverState *source = s->common.bs;
 163     int nb_sectors, sectors_per_chunk, nb_chunks;
 164     int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector;
 165     uint64_t delay_ns = 0;
 166     MirrorOp *op;
 167
 168     s->sector_num = hbitmap_iter_next(&s->hbi);
 169     if (s->sector_num < 0) {
 170         bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
 171         s->sector_num = hbitmap_iter_next(&s->hbi);
 172         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
 173         assert(s->sector_num >= 0);
 174     }
 175
 176     hbitmap_next_sector = s->sector_num;
 177     sector_num = s->sector_num;
 178     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 179     end = s->bdev_length / BDRV_SECTOR_SIZE;
 180
 181     /* Extend the QEMUIOVector to include all adjacent blocks that will
 182      * be copied in this operation.
 183      *
 184      * We have to do this if we have no backing file yet in the destination,
 185      * and the cluster size is very large.  Then we need to do COW ourselves.
 186      * The first time a cluster is copied, copy it entirely.  Note that,
 187      * because both the granularity and the cluster size are powers of two,
 188      * the number of sectors to copy cannot exceed one cluster.
 189      *
 190      * We also want to extend the QEMUIOVector to include more adjacent
 191      * dirty blocks if possible, to limit the number of I/O operations and
 192      * run efficiently even with a small granularity.
 193      */
 194     nb_chunks = 0;
 195     nb_sectors = 0;
 196     next_sector = sector_num;
 197     next_chunk = sector_num / sectors_per_chunk;
 198
 199     /* Wait for I/O to this cluster (from a previous iteration) to be done.  */
 200     while (test_bit(next_chunk, s->in_flight_bitmap)) {
 201         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 202         qemu_coroutine_yield();
 203     }
 204
 205     do {
 206         int added_sectors, added_chunks;
 207
 208         if (!bdrv_get_dirty(source, s->dirty_bitmap, next_sector) ||
 209             test_bit(next_chunk, s->in_flight_bitmap)) {
 210             assert(nb_sectors > 0);
 211             break;
 212         }
 213
 214         added_sectors = sectors_per_chunk;
 215         if (s->cow_bitmap && !test_bit(next_chunk, s->cow_bitmap)) {
 216             bdrv_round_to_clusters(s->target,
 217                                    next_sector, added_sectors,
 218                                    &next_sector, &added_sectors);
 219
 220             /* On the first iteration, the rounding may make us copy
 221              * sectors before the first dirty one.
 222              */
 223             if (next_sector < sector_num) {
 224                 assert(nb_sectors == 0);
 225                 sector_num = next_sector;
 226                 next_chunk = next_sector / sectors_per_chunk;
 227             }
 228         }
 229
 230         added_sectors = MIN(added_sectors, end - (sector_num + nb_sectors));
 231         added_chunks = (added_sectors + sectors_per_chunk - 1) / sectors_per_chunk;
 232
 233         /* When doing COW, it may happen that there is not enough space for
 234          * a full cluster.  Wait if that is the case.
 235          */
 236         while (nb_chunks == 0 && s->buf_free_count < added_chunks) {
 237             trace_mirror_yield_buf_busy(s, nb_chunks, s->in_flight);
 238             qemu_coroutine_yield();
 239         }
 240         if (s->buf_free_count < nb_chunks + added_chunks) {
 241             trace_mirror_break_buf_busy(s, nb_chunks, s->in_flight);
 242             break;
 243         }
 244
 245         /* We have enough free space to copy these sectors.  */
 246         bitmap_set(s->in_flight_bitmap, next_chunk, added_chunks);
 247
 248         nb_sectors += added_sectors;
 249         nb_chunks += added_chunks;
 250         next_sector += added_sectors;
 251         next_chunk += added_chunks;
 252         if (!s->synced && s->common.speed) {
 253             delay_ns = ratelimit_calculate_delay(&s->limit, added_sectors);
 254         }
 255     } while (delay_ns == 0 && next_sector < end);
 256
 257     /* Allocate a MirrorOp that is used as an AIO callback.  */
 258     op = g_slice_new(MirrorOp);
 259     op->s = s;
 260     op->sector_num = sector_num;
 261     op->nb_sectors = nb_sectors;
 262
 263     /* Now make a QEMUIOVector taking enough granularity-sized chunks
 264      * from s->buf_free.
 265      */
 266     qemu_iovec_init(&op->qiov, nb_chunks);
 267     next_sector = sector_num;
 268     while (nb_chunks-- > 0) {
 269         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
 270         size_t remaining = (nb_sectors * BDRV_SECTOR_SIZE) - op->qiov.size;
 271
 272         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
 273         s->buf_free_count--;
 274         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
 275
 276         /* Advance the HBitmapIter in parallel, so that we do not examine
 277          * the same sector twice.
 278          */
 279         if (next_sector > hbitmap_next_sector
 280             && bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
 281             hbitmap_next_sector = hbitmap_iter_next(&s->hbi);
 282         }
 283
 284         next_sector += sectors_per_chunk;
 285     }
 286
 287     bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num, nb_sectors);
 288
 289     /* Copy the dirty cluster.  */
 290     s->in_flight++;
 291     s->sectors_in_flight += nb_sectors;
 292     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 293     bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
 294                    mirror_read_complete, op);
 295     return delay_ns;
 296 }
 297
 298 static void mirror_free_init(MirrorBlockJob *s)
 299 {
 300     int granularity = s->granularity;
 301     size_t buf_size = s->buf_size;
 302     uint8_t *buf = s->buf;
 303
 304     assert(s->buf_free_count == 0);
 305     QSIMPLEQ_INIT(&s->buf_free);
 306     while (buf_size != 0) {
 307         MirrorBuffer *cur = (MirrorBuffer *)buf;
 308         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 309         s->buf_free_count++;
 310         buf_size -= granularity;
 311         buf += granularity;
 312     }
 313 }
 314
 315 static void mirror_drain(MirrorBlockJob *s)
 316 {
 317     while (s->in_flight > 0) {
 318         qemu_coroutine_yield();
 319     }
 320 }
 321
 322 typedef struct {
 323     int ret;
 324 } MirrorExitData;
 325
 326 static void mirror_exit(BlockJob *job, void *opaque)
 327 {
 328     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 329     MirrorExitData *data = opaque;
 330     AioContext *replace_aio_context = NULL;
 331
 332     if (s->to_replace) {
 333         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 334         aio_context_acquire(replace_aio_context);
 335     }
 336
 337     if (s->should_complete && data->ret == 0) {
 338         BlockDriverState *to_replace = s->common.bs;
 339         if (s->to_replace) {
 340             to_replace = s->to_replace;
 341         }
 342         if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
 343             bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
 344         }
 345         bdrv_swap(s->target, to_replace);
 346         if (s->common.driver->job_type == BLOCK_JOB_TYPE_COMMIT) {
 347             /* drop the bs loop chain formed by the swap: break the loop then
 348              * trigger the unref from the top one */
 349             BlockDriverState *p = s->base->backing_hd;
 350             bdrv_set_backing_hd(s->base, NULL);
 351             bdrv_unref(p);
 352         }
 353     }
 354     if (s->to_replace) {
 355         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
 356         error_free(s->replace_blocker);
 357         bdrv_unref(s->to_replace);
 358     }
 359     if (replace_aio_context) {
 360         aio_context_release(replace_aio_context);
 361     }
 362     g_free(s->replaces);
 363     bdrv_unref(s->target);
 364     block_job_completed(&s->common, data->ret);
 365     g_free(data);
 366 }
 367
 368 static void coroutine_fn mirror_run(void *opaque)
 369 {
 370     MirrorBlockJob *s = opaque;
 371     MirrorExitData *data;
 372     BlockDriverState *bs = s->common.bs;
 373     int64_t sector_num, end, sectors_per_chunk, length;
 374     uint64_t last_pause_ns;
 375     BlockDriverInfo bdi;
 376     char backing_filename[2]; /* we only need 2 characters because we are only
 377                                  checking for a NULL string */
 378     int ret = 0;
 379     int n;
 380
 381     if (block_job_is_cancelled(&s->common)) {
 382         goto immediate_exit;
 383     }
 384
 385     s->bdev_length = bdrv_getlength(bs);
 386     if (s->bdev_length < 0) {
 387         ret = s->bdev_length;
 388         goto immediate_exit;
 389     } else if (s->bdev_length == 0) {
 390         /* Report BLOCK_JOB_READY and wait for complete. */
 391         block_job_event_ready(&s->common);
 392         s->synced = true;
 393         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
 394             block_job_yield(&s->common);
 395         }
 396         s->common.cancelled = false;
 397         goto immediate_exit;
 398     }
 399
 400     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
 401     s->in_flight_bitmap = bitmap_new(length);
 402
 403     /* If we have no backing file yet in the destination, we cannot let
 404      * the destination do COW.  Instead, we copy sectors around the
 405      * dirty data if needed.  We need a bitmap to do that.
 406      */
 407     bdrv_get_backing_filename(s->target, backing_filename,
 408                               sizeof(backing_filename));
 409     if (backing_filename[0] && !s->target->backing_hd) {
 410         ret = bdrv_get_info(s->target, &bdi);
 411         if (ret < 0) {
 412             goto immediate_exit;
 413         }
 414         if (s->granularity < bdi.cluster_size) {
 415             s->buf_size = MAX(s->buf_size, bdi.cluster_size);
 416             s->cow_bitmap = bitmap_new(length);
 417         }
 418     }
 419
 420     end = s->bdev_length / BDRV_SECTOR_SIZE;
 421     s->buf = qemu_try_blockalign(bs, s->buf_size);
 422     if (s->buf == NULL) {
 423         ret = -ENOMEM;
 424         goto immediate_exit;
 425     }
 426
 427     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 428     mirror_free_init(s);
 429
 430     if (!s->is_none_mode) {
 431         /* First part, loop on the sectors and initialize the dirty bitmap.  */
 432         BlockDriverState *base = s->base;
 433         for (sector_num = 0; sector_num < end; ) {
 434             int64_t next = (sector_num | (sectors_per_chunk - 1)) + 1;
 435             ret = bdrv_is_allocated_above(bs, base,
 436                                           sector_num, next - sector_num, &n);
 437
 438             if (ret < 0) {
 439                 goto immediate_exit;
 440             }
 441
 442             assert(n > 0);
 443             if (ret == 1) {
 444                 bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
 445                 sector_num = next;
 446             } else {
 447                 sector_num += n;
 448             }
 449         }
 450     }
 451
 452     bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
 453     last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 454     for (;;) {
 455         uint64_t delay_ns = 0;
 456         int64_t cnt;
 457         bool should_complete;
 458
 459         if (s->ret < 0) {
 460             ret = s->ret;
 461             goto immediate_exit;
 462         }
 463
 464         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 465         /* s->common.offset contains the number of bytes already processed so
 466          * far, cnt is the number of dirty sectors remaining and
 467          * s->sectors_in_flight is the number of sectors currently being
 468          * processed; together those are the current total operation length */
 469         s->common.len = s->common.offset +
 470                         (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
 471
 472         /* Note that even when no rate limit is applied we need to yield
 473          * periodically with no pending I/O so that bdrv_drain_all() returns.
 474          * We do so every SLICE_TIME nanoseconds, or when there is an error,
 475          * or when the source is clean, whichever comes first.
 476          */
 477         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
 478             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 479             if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 480                 (cnt == 0 && s->in_flight > 0)) {
 481                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
 482                 qemu_coroutine_yield();
 483                 continue;
 484             } else if (cnt != 0) {
 485                 delay_ns = mirror_iteration(s);
 486             }
 487         }
 488
 489         should_complete = false;
 490         if (s->in_flight == 0 && cnt == 0) {
 491             trace_mirror_before_flush(s);
 492             ret = bdrv_flush(s->target);
 493             if (ret < 0) {
 494                 if (mirror_error_action(s, false, -ret) ==
 495                     BLOCK_ERROR_ACTION_REPORT) {
 496                     goto immediate_exit;
 497                 }
 498             } else {
 499                 /* We're out of the streaming phase.  From now on, if the job
 500                  * is cancelled we will actually complete all pending I/O and
 501                  * report completion.  This way, block-job-cancel will leave
 502                  * the target in a consistent state.
 503                  */
 504                 if (!s->synced) {
 505                     block_job_event_ready(&s->common);
 506                     s->synced = true;
 507                 }
 508
 509                 should_complete = s->should_complete ||
 510                     block_job_is_cancelled(&s->common);
 511                 cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 512             }
 513         }
 514
 515         if (cnt == 0 && should_complete) {
 516             /* The dirty bitmap is not updated while operations are pending.
 517              * If we're about to exit, wait for pending operations before
 518              * calling bdrv_get_dirty_count(bs), or we may exit while the
 519              * source has dirty data to copy!
 520              *
 521              * Note that I/O can be submitted by the guest while
 522              * mirror_populate runs.
 523              */
 524             trace_mirror_before_drain(s, cnt);
 525             bdrv_drain(bs);
 526             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 527         }
 528
 529         ret = 0;
 530         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
 531         if (!s->synced) {
 532             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 533             if (block_job_is_cancelled(&s->common)) {
 534                 break;
 535             }
 536         } else if (!should_complete) {
 537             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 538             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 539         } else if (cnt == 0) {
 540             /* The two disks are in sync.  Exit and report successful
 541              * completion.
 542              */
 543             assert(QLIST_EMPTY(&bs->tracked_requests));
 544             s->common.cancelled = false;
 545             break;
 546         }
 547         last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 548     }
 549
 550 immediate_exit:
 551     if (s->in_flight > 0) {
 552         /* We get here only if something went wrong.  Either the job failed,
 553          * or it was cancelled prematurely so that we do not guarantee that
 554          * the target is a copy of the source.
 555          */
 556         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 557         mirror_drain(s);
 558     }
 559
 560     assert(s->in_flight == 0);
 561     qemu_vfree(s->buf);
 562     g_free(s->cow_bitmap);
 563     g_free(s->in_flight_bitmap);
 564     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
 565     bdrv_iostatus_disable(s->target);
 566
 567     data = g_malloc(sizeof(*data));
 568     data->ret = ret;
 569     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
 570 }
 571
 572 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 573 {
 574     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 575
 576     if (speed < 0) {
 577         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 578         return;
 579     }
 580     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 581 }
 582
 583 static void mirror_iostatus_reset(BlockJob *job)
 584 {
 585     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 586
 587     bdrv_iostatus_reset(s->target);
 588 }
 589
 590 static void mirror_complete(BlockJob *job, Error **errp)
 591 {
 592     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 593     Error *local_err = NULL;
 594     int ret;
 595
 596     ret = bdrv_open_backing_file(s->target, NULL, &local_err);
 597     if (ret < 0) {
 598         error_propagate(errp, local_err);
 599         return;
 600     }
 601     if (!s->synced) {
 602         error_setg(errp, QERR_BLOCK_JOB_NOT_READY,
 603                    bdrv_get_device_name(job->bs));
 604         return;
 605     }
 606
 607     /* check the target bs is not blocked and block all operations on it */
 608     if (s->replaces) {
 609         AioContext *replace_aio_context;
 610
 611         s->to_replace = check_to_replace_node(s->replaces, &local_err);
 612         if (!s->to_replace) {
 613             error_propagate(errp, local_err);
 614             return;
 615         }
 616
 617         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 618         aio_context_acquire(replace_aio_context);
 619
 620         error_setg(&s->replace_blocker,
 621                    "block device is in use by block-job-complete");
 622         bdrv_op_block_all(s->to_replace, s->replace_blocker);
 623         bdrv_ref(s->to_replace);
 624
 625         aio_context_release(replace_aio_context);
 626     }
 627
 628     s->should_complete = true;
 629     block_job_enter(&s->common);
 630 }
 631
 632 static const BlockJobDriver mirror_job_driver = {
 633     .instance_size = sizeof(MirrorBlockJob),
 634     .job_type      = BLOCK_JOB_TYPE_MIRROR,
 635     .set_speed     = mirror_set_speed,
 636     .iostatus_reset= mirror_iostatus_reset,
 637     .complete      = mirror_complete,
 638 };
 639
 640 static const BlockJobDriver commit_active_job_driver = {
 641     .instance_size = sizeof(MirrorBlockJob),
 642     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 643     .set_speed     = mirror_set_speed,
 644     .iostatus_reset
 645                    = mirror_iostatus_reset,
 646     .complete      = mirror_complete,
 647 };
 648
 649 static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
 650                              const char *replaces,
 651                              int64_t speed, uint32_t granularity,
 652                              int64_t buf_size,
 653                              BlockdevOnError on_source_error,
 654                              BlockdevOnError on_target_error,
 655                              BlockCompletionFunc *cb,
 656                              void *opaque, Error **errp,
 657                              const BlockJobDriver *driver,
 658                              bool is_none_mode, BlockDriverState *base)
 659 {
 660     MirrorBlockJob *s;
 661
 662     if (granularity == 0) {
 663         granularity = bdrv_get_default_bitmap_granularity(target);
 664     }
 665
 666     assert ((granularity & (granularity - 1)) == 0);
 667
 668     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 669          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 670         !bdrv_iostatus_is_enabled(bs)) {
 671         error_setg(errp, QERR_INVALID_PARAMETER, "on-source-error");
 672         return;
 673     }
 674
 675
 676     s = block_job_create(driver, bs, speed, cb, opaque, errp);
 677     if (!s) {
 678         return;
 679     }
 680
 681     s->replaces = g_strdup(replaces);
 682     s->on_source_error = on_source_error;
 683     s->on_target_error = on_target_error;
 684     s->target = target;
 685     s->is_none_mode = is_none_mode;
 686     s->base = base;
 687     s->granularity = granularity;
 688     s->buf_size = MAX(buf_size, granularity);
 689
 690     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
 691     if (!s->dirty_bitmap) {
 692         return;
 693     }
 694     bdrv_set_enable_write_cache(s->target, true);
 695     bdrv_set_on_error(s->target, on_target_error, on_target_error);
 696     bdrv_iostatus_enable(s->target);
 697     s->common.co = qemu_coroutine_create(mirror_run);
 698     trace_mirror_start(bs, s, s->common.co, opaque);
 699     qemu_coroutine_enter(s->common.co, s);
 700 }
 701
 702 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
 703                   const char *replaces,
 704                   int64_t speed, uint32_t granularity, int64_t buf_size,
 705                   MirrorSyncMode mode, BlockdevOnError on_source_error,
 706                   BlockdevOnError on_target_error,
 707                   BlockCompletionFunc *cb,
 708                   void *opaque, Error **errp)
 709 {
 710     bool is_none_mode;
 711     BlockDriverState *base;
 712
 713     if (mode == MIRROR_SYNC_MODE_DIRTY_BITMAP) {
 714         error_setg(errp, "Sync mode 'dirty-bitmap' not supported");
 715         return;
 716     }
 717     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
 718     base = mode == MIRROR_SYNC_MODE_TOP ? bs->backing_hd : NULL;
 719     mirror_start_job(bs, target, replaces,
 720                      speed, granularity, buf_size,
 721                      on_source_error, on_target_error, cb, opaque, errp,
 722                      &mirror_job_driver, is_none_mode, base);
 723 }
 724
 725 void commit_active_start(BlockDriverState *bs, BlockDriverState *base,
 726                          int64_t speed,
 727                          BlockdevOnError on_error,
 728                          BlockCompletionFunc *cb,
 729                          void *opaque, Error **errp)
 730 {
 731     int64_t length, base_length;
 732     int orig_base_flags;
 733     int ret;
 734     Error *local_err = NULL;
 735
 736     orig_base_flags = bdrv_get_flags(base);
 737
 738     if (bdrv_reopen(base, bs->open_flags, errp)) {
 739         return;
 740     }
 741
 742     length = bdrv_getlength(bs);
 743     if (length < 0) {
 744         error_setg_errno(errp, -length,
 745                          "Unable to determine length of %s", bs->filename);
 746         goto error_restore_flags;
 747     }
 748
 749     base_length = bdrv_getlength(base);
 750     if (base_length < 0) {
 751         error_setg_errno(errp, -base_length,
 752                          "Unable to determine length of %s", base->filename);
 753         goto error_restore_flags;
 754     }
 755
 756     if (length > base_length) {
 757         ret = bdrv_truncate(base, length);
 758         if (ret < 0) {
 759             error_setg_errno(errp, -ret,
 760                             "Top image %s is larger than base image %s, and "
 761                              "resize of base image failed",
 762                              bs->filename, base->filename);
 763             goto error_restore_flags;
 764         }
 765     }
 766
 767     bdrv_ref(base);
 768     mirror_start_job(bs, base, NULL, speed, 0, 0,
 769                      on_error, on_error, cb, opaque, &local_err,
 770                      &commit_active_job_driver, false, base);
 771     if (local_err) {
 772         error_propagate(errp, local_err);
 773         goto error_restore_flags;
 774     }
 775
 776     return;
 777
 778 error_restore_flags:
 779     /* ignore error and errp for bdrv_reopen, because we want to propagate
 780      * the original error */
 781     bdrv_reopen(base, orig_base_flags, NULL);
 782     return;
 783 }