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     int64_t bdev_length;
  49     unsigned long *cow_bitmap;
  50     BdrvDirtyBitmap *dirty_bitmap;
  51     HBitmapIter hbi;
  52     uint8_t *buf;
  53     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
  54     int buf_free_count;
  55
  56     unsigned long *in_flight_bitmap;
  57     int in_flight;
  58     int sectors_in_flight;
  59     int ret;
  60 } MirrorBlockJob;
  61
  62 typedef struct MirrorOp {
  63     MirrorBlockJob *s;
  64     QEMUIOVector qiov;
  65     int64_t sector_num;
  66     int nb_sectors;
  67 } MirrorOp;
  68
  69 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  70                                             int error)
  71 {
  72     s->synced = false;
  73     if (read) {
  74         return block_job_error_action(&s->common, s->common.bs,
  75                                       s->on_source_error, true, error);
  76     } else {
  77         return block_job_error_action(&s->common, s->target,
  78                                       s->on_target_error, false, error);
  79     }
  80 }
  81
  82 static void mirror_iteration_done(MirrorOp *op, int ret)
  83 {
  84     MirrorBlockJob *s = op->s;
  85     struct iovec *iov;
  86     int64_t chunk_num;
  87     int i, nb_chunks, sectors_per_chunk;
  88
  89     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
  90
  91     s->in_flight--;
  92     s->sectors_in_flight -= op->nb_sectors;
  93     iov = op->qiov.iov;
  94     for (i = 0; i < op->qiov.niov; i++) {
  95         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
  96         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
  97         s->buf_free_count++;
  98     }
  99
 100     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 101     chunk_num = op->sector_num / sectors_per_chunk;
 102     nb_chunks = op->nb_sectors / sectors_per_chunk;
 103     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
 104     if (ret >= 0) {
 105         if (s->cow_bitmap) {
 106             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
 107         }
 108         s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
 109     }
 110
 111     qemu_iovec_destroy(&op->qiov);
 112     g_slice_free(MirrorOp, op);
 113
 114     /* Enter coroutine when it is not sleeping.  The coroutine sleeps to
 115      * rate-limit itself.  The coroutine will eventually resume since there is
 116      * a sleep timeout so don't wake it early.
 117      */
 118     if (s->common.busy) {
 119         qemu_coroutine_enter(s->common.co, NULL);
 120     }
 121 }
 122
 123 static void mirror_write_complete(void *opaque, int ret)
 124 {
 125     MirrorOp *op = opaque;
 126     MirrorBlockJob *s = op->s;
 127     if (ret < 0) {
 128         BlockDriverState *source = s->common.bs;
 129         BlockErrorAction action;
 130
 131         bdrv_set_dirty_bitmap(source, s->dirty_bitmap, op->sector_num,
 132                               op->nb_sectors);
 133         action = mirror_error_action(s, false, -ret);
 134         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 135             s->ret = ret;
 136         }
 137     }
 138     mirror_iteration_done(op, ret);
 139 }
 140
 141 static void mirror_read_complete(void *opaque, int ret)
 142 {
 143     MirrorOp *op = opaque;
 144     MirrorBlockJob *s = op->s;
 145     if (ret < 0) {
 146         BlockDriverState *source = s->common.bs;
 147         BlockErrorAction action;
 148
 149         bdrv_set_dirty_bitmap(source, s->dirty_bitmap, op->sector_num,
 150                               op->nb_sectors);
 151         action = mirror_error_action(s, true, -ret);
 152         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 153             s->ret = ret;
 154         }
 155
 156         mirror_iteration_done(op, ret);
 157         return;
 158     }
 159     bdrv_aio_writev(s->target, op->sector_num, &op->qiov, op->nb_sectors,
 160                     mirror_write_complete, op);
 161 }
 162
 163 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
 164 {
 165     BlockDriverState *source = s->common.bs;
 166     int nb_sectors, sectors_per_chunk, nb_chunks;
 167     int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector;
 168     uint64_t delay_ns = 0;
 169     MirrorOp *op;
 170
 171     s->sector_num = hbitmap_iter_next(&s->hbi);
 172     if (s->sector_num < 0) {
 173         bdrv_dirty_iter_init(source, s->dirty_bitmap, &s->hbi);
 174         s->sector_num = hbitmap_iter_next(&s->hbi);
 175         trace_mirror_restart_iter(s,
 176                                   bdrv_get_dirty_count(source, s->dirty_bitmap));
 177         assert(s->sector_num >= 0);
 178     }
 179
 180     hbitmap_next_sector = s->sector_num;
 181     sector_num = s->sector_num;
 182     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 183     end = s->bdev_length / BDRV_SECTOR_SIZE;
 184
 185     /* Extend the QEMUIOVector to include all adjacent blocks that will
 186      * be copied in this operation.
 187      *
 188      * We have to do this if we have no backing file yet in the destination,
 189      * and the cluster size is very large.  Then we need to do COW ourselves.
 190      * The first time a cluster is copied, copy it entirely.  Note that,
 191      * because both the granularity and the cluster size are powers of two,
 192      * the number of sectors to copy cannot exceed one cluster.
 193      *
 194      * We also want to extend the QEMUIOVector to include more adjacent
 195      * dirty blocks if possible, to limit the number of I/O operations and
 196      * run efficiently even with a small granularity.
 197      */
 198     nb_chunks = 0;
 199     nb_sectors = 0;
 200     next_sector = sector_num;
 201     next_chunk = sector_num / sectors_per_chunk;
 202
 203     /* Wait for I/O to this cluster (from a previous iteration) to be done.  */
 204     while (test_bit(next_chunk, s->in_flight_bitmap)) {
 205         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 206         qemu_coroutine_yield();
 207     }
 208
 209     do {
 210         int added_sectors, added_chunks;
 211
 212         if (!bdrv_get_dirty(source, s->dirty_bitmap, next_sector) ||
 213             test_bit(next_chunk, s->in_flight_bitmap)) {
 214             assert(nb_sectors > 0);
 215             break;
 216         }
 217
 218         added_sectors = sectors_per_chunk;
 219         if (s->cow_bitmap && !test_bit(next_chunk, s->cow_bitmap)) {
 220             bdrv_round_to_clusters(s->target,
 221                                    next_sector, added_sectors,
 222                                    &next_sector, &added_sectors);
 223
 224             /* On the first iteration, the rounding may make us copy
 225              * sectors before the first dirty one.
 226              */
 227             if (next_sector < sector_num) {
 228                 assert(nb_sectors == 0);
 229                 sector_num = next_sector;
 230                 next_chunk = next_sector / sectors_per_chunk;
 231             }
 232         }
 233
 234         added_sectors = MIN(added_sectors, end - (sector_num + nb_sectors));
 235         added_chunks = (added_sectors + sectors_per_chunk - 1) / sectors_per_chunk;
 236
 237         /* When doing COW, it may happen that there is not enough space for
 238          * a full cluster.  Wait if that is the case.
 239          */
 240         while (nb_chunks == 0 && s->buf_free_count < added_chunks) {
 241             trace_mirror_yield_buf_busy(s, nb_chunks, s->in_flight);
 242             qemu_coroutine_yield();
 243         }
 244         if (s->buf_free_count < nb_chunks + added_chunks) {
 245             trace_mirror_break_buf_busy(s, nb_chunks, s->in_flight);
 246             break;
 247         }
 248
 249         /* We have enough free space to copy these sectors.  */
 250         bitmap_set(s->in_flight_bitmap, next_chunk, added_chunks);
 251
 252         nb_sectors += added_sectors;
 253         nb_chunks += added_chunks;
 254         next_sector += added_sectors;
 255         next_chunk += added_chunks;
 256         if (!s->synced && s->common.speed) {
 257             delay_ns = ratelimit_calculate_delay(&s->limit, added_sectors);
 258         }
 259     } while (delay_ns == 0 && next_sector < end);
 260
 261     /* Allocate a MirrorOp that is used as an AIO callback.  */
 262     op = g_slice_new(MirrorOp);
 263     op->s = s;
 264     op->sector_num = sector_num;
 265     op->nb_sectors = nb_sectors;
 266
 267     /* Now make a QEMUIOVector taking enough granularity-sized chunks
 268      * from s->buf_free.
 269      */
 270     qemu_iovec_init(&op->qiov, nb_chunks);
 271     next_sector = sector_num;
 272     while (nb_chunks-- > 0) {
 273         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
 274         size_t remaining = (nb_sectors * BDRV_SECTOR_SIZE) - op->qiov.size;
 275
 276         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
 277         s->buf_free_count--;
 278         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
 279
 280         /* Advance the HBitmapIter in parallel, so that we do not examine
 281          * the same sector twice.
 282          */
 283         if (next_sector > hbitmap_next_sector
 284             && bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
 285             hbitmap_next_sector = hbitmap_iter_next(&s->hbi);
 286         }
 287
 288         next_sector += sectors_per_chunk;
 289     }
 290
 291     bdrv_reset_dirty_bitmap(source, s->dirty_bitmap, sector_num,
 292                             nb_sectors);
 293
 294     /* Copy the dirty cluster.  */
 295     s->in_flight++;
 296     s->sectors_in_flight += nb_sectors;
 297     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 298     bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
 299                    mirror_read_complete, op);
 300     return delay_ns;
 301 }
 302
 303 static void mirror_free_init(MirrorBlockJob *s)
 304 {
 305     int granularity = s->granularity;
 306     size_t buf_size = s->buf_size;
 307     uint8_t *buf = s->buf;
 308
 309     assert(s->buf_free_count == 0);
 310     QSIMPLEQ_INIT(&s->buf_free);
 311     while (buf_size != 0) {
 312         MirrorBuffer *cur = (MirrorBuffer *)buf;
 313         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 314         s->buf_free_count++;
 315         buf_size -= granularity;
 316         buf += granularity;
 317     }
 318 }
 319
 320 static void mirror_drain(MirrorBlockJob *s)
 321 {
 322     while (s->in_flight > 0) {
 323         qemu_coroutine_yield();
 324     }
 325 }
 326
 327 typedef struct {
 328     int ret;
 329 } MirrorExitData;
 330
 331 static void mirror_exit(BlockJob *job, void *opaque)
 332 {
 333     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 334     MirrorExitData *data = opaque;
 335     AioContext *replace_aio_context = NULL;
 336
 337     if (s->to_replace) {
 338         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 339         aio_context_acquire(replace_aio_context);
 340     }
 341
 342     if (s->should_complete && data->ret == 0) {
 343         BlockDriverState *to_replace = s->common.bs;
 344         if (s->to_replace) {
 345             to_replace = s->to_replace;
 346         }
 347         if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
 348             bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
 349         }
 350         bdrv_swap(s->target, to_replace);
 351         if (s->common.driver->job_type == BLOCK_JOB_TYPE_COMMIT) {
 352             /* drop the bs loop chain formed by the swap: break the loop then
 353              * trigger the unref from the top one */
 354             BlockDriverState *p = s->base->backing_hd;
 355             bdrv_set_backing_hd(s->base, NULL);
 356             bdrv_unref(p);
 357         }
 358     }
 359     if (s->to_replace) {
 360         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
 361         error_free(s->replace_blocker);
 362         bdrv_unref(s->to_replace);
 363     }
 364     if (replace_aio_context) {
 365         aio_context_release(replace_aio_context);
 366     }
 367     g_free(s->replaces);
 368     bdrv_unref(s->target);
 369     block_job_completed(&s->common, data->ret);
 370     g_free(data);
 371 }
 372
 373 static void coroutine_fn mirror_run(void *opaque)
 374 {
 375     MirrorBlockJob *s = opaque;
 376     MirrorExitData *data;
 377     BlockDriverState *bs = s->common.bs;
 378     int64_t sector_num, end, sectors_per_chunk, length;
 379     uint64_t last_pause_ns;
 380     BlockDriverInfo bdi;
 381     char backing_filename[2]; /* we only need 2 characters because we are only
 382                                  checking for a NULL string */
 383     int ret = 0;
 384     int n;
 385
 386     if (block_job_is_cancelled(&s->common)) {
 387         goto immediate_exit;
 388     }
 389
 390     s->bdev_length = bdrv_getlength(bs);
 391     if (s->bdev_length < 0) {
 392         ret = s->bdev_length;
 393         goto immediate_exit;
 394     } else if (s->bdev_length == 0) {
 395         /* Report BLOCK_JOB_READY and wait for complete. */
 396         block_job_event_ready(&s->common);
 397         s->synced = true;
 398         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
 399             block_job_yield(&s->common);
 400         }
 401         s->common.cancelled = false;
 402         goto immediate_exit;
 403     }
 404
 405     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
 406     s->in_flight_bitmap = bitmap_new(length);
 407
 408     /* If we have no backing file yet in the destination, we cannot let
 409      * the destination do COW.  Instead, we copy sectors around the
 410      * dirty data if needed.  We need a bitmap to do that.
 411      */
 412     bdrv_get_backing_filename(s->target, backing_filename,
 413                               sizeof(backing_filename));
 414     if (backing_filename[0] && !s->target->backing_hd) {
 415         ret = bdrv_get_info(s->target, &bdi);
 416         if (ret < 0) {
 417             goto immediate_exit;
 418         }
 419         if (s->granularity < bdi.cluster_size) {
 420             s->buf_size = MAX(s->buf_size, bdi.cluster_size);
 421             s->cow_bitmap = bitmap_new(length);
 422         }
 423     }
 424
 425     end = s->bdev_length / BDRV_SECTOR_SIZE;
 426     s->buf = qemu_try_blockalign(bs, s->buf_size);
 427     if (s->buf == NULL) {
 428         ret = -ENOMEM;
 429         goto immediate_exit;
 430     }
 431
 432     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 433     mirror_free_init(s);
 434
 435     if (!s->is_none_mode) {
 436         /* First part, loop on the sectors and initialize the dirty bitmap.  */
 437         BlockDriverState *base = s->base;
 438         for (sector_num = 0; sector_num < end; ) {
 439             int64_t next = (sector_num | (sectors_per_chunk - 1)) + 1;
 440             ret = bdrv_is_allocated_above(bs, base,
 441                                           sector_num, next - sector_num, &n);
 442
 443             if (ret < 0) {
 444                 goto immediate_exit;
 445             }
 446
 447             assert(n > 0);
 448             if (ret == 1) {
 449                 bdrv_set_dirty_bitmap(bs, s->dirty_bitmap, sector_num, n);
 450                 sector_num = next;
 451             } else {
 452                 sector_num += n;
 453             }
 454         }
 455     }
 456
 457     bdrv_dirty_iter_init(bs, s->dirty_bitmap, &s->hbi);
 458     last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 459     for (;;) {
 460         uint64_t delay_ns = 0;
 461         int64_t cnt;
 462         bool should_complete;
 463
 464         if (s->ret < 0) {
 465             ret = s->ret;
 466             goto immediate_exit;
 467         }
 468
 469         cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 470         /* s->common.offset contains the number of bytes already processed so
 471          * far, cnt is the number of dirty sectors remaining and
 472          * s->sectors_in_flight is the number of sectors currently being
 473          * processed; together those are the current total operation length */
 474         s->common.len = s->common.offset +
 475                         (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
 476
 477         /* Note that even when no rate limit is applied we need to yield
 478          * periodically with no pending I/O so that qemu_aio_flush() returns.
 479          * We do so every SLICE_TIME nanoseconds, or when there is an error,
 480          * or when the source is clean, whichever comes first.
 481          */
 482         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
 483             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 484             if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 485                 (cnt == 0 && s->in_flight > 0)) {
 486                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
 487                 qemu_coroutine_yield();
 488                 continue;
 489             } else if (cnt != 0) {
 490                 delay_ns = mirror_iteration(s);
 491                 if (delay_ns == 0) {
 492                     continue;
 493                 }
 494             }
 495         }
 496
 497         should_complete = false;
 498         if (s->in_flight == 0 && cnt == 0) {
 499             trace_mirror_before_flush(s);
 500             ret = bdrv_flush(s->target);
 501             if (ret < 0) {
 502                 if (mirror_error_action(s, false, -ret) ==
 503                     BLOCK_ERROR_ACTION_REPORT) {
 504                     goto immediate_exit;
 505                 }
 506             } else {
 507                 /* We're out of the streaming phase.  From now on, if the job
 508                  * is cancelled we will actually complete all pending I/O and
 509                  * report completion.  This way, block-job-cancel will leave
 510                  * the target in a consistent state.
 511                  */
 512                 if (!s->synced) {
 513                     block_job_event_ready(&s->common);
 514                     s->synced = true;
 515                 }
 516
 517                 should_complete = s->should_complete ||
 518                     block_job_is_cancelled(&s->common);
 519                 cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 520             }
 521         }
 522
 523         if (cnt == 0 && should_complete) {
 524             /* The dirty bitmap is not updated while operations are pending.
 525              * If we're about to exit, wait for pending operations before
 526              * calling bdrv_get_dirty_count(bs), or we may exit while the
 527              * source has dirty data to copy!
 528              *
 529              * Note that I/O can be submitted by the guest while
 530              * mirror_populate runs.
 531              */
 532             trace_mirror_before_drain(s, cnt);
 533             bdrv_drain(bs);
 534             cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);
 535         }
 536
 537         ret = 0;
 538         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
 539         if (!s->synced) {
 540             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 541             if (block_job_is_cancelled(&s->common)) {
 542                 break;
 543             }
 544         } else if (!should_complete) {
 545             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 546             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 547         } else if (cnt == 0) {
 548             /* The two disks are in sync.  Exit and report successful
 549              * completion.
 550              */
 551             assert(QLIST_EMPTY(&bs->tracked_requests));
 552             s->common.cancelled = false;
 553             break;
 554         }
 555         last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 556     }
 557
 558 immediate_exit:
 559     if (s->in_flight > 0) {
 560         /* We get here only if something went wrong.  Either the job failed,
 561          * or it was cancelled prematurely so that we do not guarantee that
 562          * the target is a copy of the source.
 563          */
 564         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 565         mirror_drain(s);
 566     }
 567
 568     assert(s->in_flight == 0);
 569     qemu_vfree(s->buf);
 570     g_free(s->cow_bitmap);
 571     g_free(s->in_flight_bitmap);
 572     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
 573     bdrv_iostatus_disable(s->target);
 574
 575     data = g_malloc(sizeof(*data));
 576     data->ret = ret;
 577     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
 578 }
 579
 580 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 581 {
 582     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 583
 584     if (speed < 0) {
 585         error_set(errp, QERR_INVALID_PARAMETER, "speed");
 586         return;
 587     }
 588     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 589 }
 590
 591 static void mirror_iostatus_reset(BlockJob *job)
 592 {
 593     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 594
 595     bdrv_iostatus_reset(s->target);
 596 }
 597
 598 static void mirror_complete(BlockJob *job, Error **errp)
 599 {
 600     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 601     Error *local_err = NULL;
 602     int ret;
 603
 604     ret = bdrv_open_backing_file(s->target, NULL, &local_err);
 605     if (ret < 0) {
 606         error_propagate(errp, local_err);
 607         return;
 608     }
 609     if (!s->synced) {
 610         error_set(errp, QERR_BLOCK_JOB_NOT_READY,
 611                   bdrv_get_device_name(job->bs));
 612         return;
 613     }
 614
 615     /* check the target bs is not blocked and block all operations on it */
 616     if (s->replaces) {
 617         AioContext *replace_aio_context;
 618
 619         s->to_replace = check_to_replace_node(s->replaces, &local_err);
 620         if (!s->to_replace) {
 621             error_propagate(errp, local_err);
 622             return;
 623         }
 624
 625         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 626         aio_context_acquire(replace_aio_context);
 627
 628         error_setg(&s->replace_blocker,
 629                    "block device is in use by block-job-complete");
 630         bdrv_op_block_all(s->to_replace, s->replace_blocker);
 631         bdrv_ref(s->to_replace);
 632
 633         aio_context_release(replace_aio_context);
 634     }
 635
 636     s->should_complete = true;
 637     block_job_resume(job);
 638 }
 639
 640 static const BlockJobDriver mirror_job_driver = {
 641     .instance_size = sizeof(MirrorBlockJob),
 642     .job_type      = BLOCK_JOB_TYPE_MIRROR,
 643     .set_speed     = mirror_set_speed,
 644     .iostatus_reset= mirror_iostatus_reset,
 645     .complete      = mirror_complete,
 646 };
 647
 648 static const BlockJobDriver commit_active_job_driver = {
 649     .instance_size = sizeof(MirrorBlockJob),
 650     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 651     .set_speed     = mirror_set_speed,
 652     .iostatus_reset
 653                    = mirror_iostatus_reset,
 654     .complete      = mirror_complete,
 655 };
 656
 657 static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
 658                              const char *replaces,
 659                              int64_t speed, int64_t granularity,
 660                              int64_t buf_size,
 661                              BlockdevOnError on_source_error,
 662                              BlockdevOnError on_target_error,
 663                              BlockCompletionFunc *cb,
 664                              void *opaque, Error **errp,
 665                              const BlockJobDriver *driver,
 666                              bool is_none_mode, BlockDriverState *base)
 667 {
 668     MirrorBlockJob *s;
 669
 670     if (granularity == 0) {
 671         /* Choose the default granularity based on the target file's cluster
 672          * size, clamped between 4k and 64k.  */
 673         BlockDriverInfo bdi;
 674         if (bdrv_get_info(target, &bdi) >= 0 && bdi.cluster_size != 0) {
 675             granularity = MAX(4096, bdi.cluster_size);
 676             granularity = MIN(65536, granularity);
 677         } else {
 678             granularity = 65536;
 679         }
 680     }
 681
 682     assert ((granularity & (granularity - 1)) == 0);
 683
 684     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 685          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 686         !bdrv_iostatus_is_enabled(bs)) {
 687         error_set(errp, QERR_INVALID_PARAMETER, "on-source-error");
 688         return;
 689     }
 690
 691
 692     s = block_job_create(driver, bs, speed, cb, opaque, errp);
 693     if (!s) {
 694         return;
 695     }
 696
 697     s->replaces = g_strdup(replaces);
 698     s->on_source_error = on_source_error;
 699     s->on_target_error = on_target_error;
 700     s->target = target;
 701     s->is_none_mode = is_none_mode;
 702     s->base = base;
 703     s->granularity = granularity;
 704     s->buf_size = MAX(buf_size, granularity);
 705
 706     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, errp);
 707     if (!s->dirty_bitmap) {
 708         return;
 709     }
 710     bdrv_set_enable_write_cache(s->target, true);
 711     bdrv_set_on_error(s->target, on_target_error, on_target_error);
 712     bdrv_iostatus_enable(s->target);
 713     s->common.co = qemu_coroutine_create(mirror_run);
 714     trace_mirror_start(bs, s, s->common.co, opaque);
 715     qemu_coroutine_enter(s->common.co, s);
 716 }
 717
 718 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
 719                   const char *replaces,
 720                   int64_t speed, int64_t granularity, int64_t buf_size,
 721                   MirrorSyncMode mode, BlockdevOnError on_source_error,
 722                   BlockdevOnError on_target_error,
 723                   BlockCompletionFunc *cb,
 724                   void *opaque, Error **errp)
 725 {
 726     bool is_none_mode;
 727     BlockDriverState *base;
 728
 729     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
 730     base = mode == MIRROR_SYNC_MODE_TOP ? bs->backing_hd : NULL;
 731     mirror_start_job(bs, target, replaces,
 732                      speed, granularity, buf_size,
 733                      on_source_error, on_target_error, cb, opaque, errp,
 734                      &mirror_job_driver, is_none_mode, base);
 735 }
 736
 737 void commit_active_start(BlockDriverState *bs, BlockDriverState *base,
 738                          int64_t speed,
 739                          BlockdevOnError on_error,
 740                          BlockCompletionFunc *cb,
 741                          void *opaque, Error **errp)
 742 {
 743     int64_t length, base_length;
 744     int orig_base_flags;
 745     int ret;
 746     Error *local_err = NULL;
 747
 748     orig_base_flags = bdrv_get_flags(base);
 749
 750     if (bdrv_reopen(base, bs->open_flags, errp)) {
 751         return;
 752     }
 753
 754     length = bdrv_getlength(bs);
 755     if (length < 0) {
 756         error_setg_errno(errp, -length,
 757                          "Unable to determine length of %s", bs->filename);
 758         goto error_restore_flags;
 759     }
 760
 761     base_length = bdrv_getlength(base);
 762     if (base_length < 0) {
 763         error_setg_errno(errp, -base_length,
 764                          "Unable to determine length of %s", base->filename);
 765         goto error_restore_flags;
 766     }
 767
 768     if (length > base_length) {
 769         ret = bdrv_truncate(base, length);
 770         if (ret < 0) {
 771             error_setg_errno(errp, -ret,
 772                             "Top image %s is larger than base image %s, and "
 773                              "resize of base image failed",
 774                              bs->filename, base->filename);
 775             goto error_restore_flags;
 776         }
 777     }
 778
 779     bdrv_ref(base);
 780     mirror_start_job(bs, base, NULL, speed, 0, 0,
 781                      on_error, on_error, cb, opaque, &local_err,
 782                      &commit_active_job_driver, false, base);
 783     if (local_err) {
 784         error_propagate(errp, local_err);
 785         goto error_restore_flags;
 786     }
 787
 788     return;
 789
 790 error_restore_flags:
 791     /* ignore error and errp for bdrv_reopen, because we want to propagate
 792      * the original error */
 793     bdrv_reopen(base, orig_base_flags, NULL);
 794     return;
 795 }