qemu/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 "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 #define DEFAULT_MIRROR_BUF_SIZE   (10 << 20)
  24
  25 /* The mirroring buffer is a list of granularity-sized chunks.
  26  * Free chunks are organized in a list.
  27  */
  28 typedef struct MirrorBuffer {
  29     QSIMPLEQ_ENTRY(MirrorBuffer) next;
  30 } MirrorBuffer;
  31
  32 typedef struct MirrorBlockJob {
  33     BlockJob common;
  34     RateLimit limit;
  35     BlockDriverState *target;
  36     BlockDriverState *base;
  37     /* The name of the graph node to replace */
  38     char *replaces;
  39     /* The BDS to replace */
  40     BlockDriverState *to_replace;
  41     /* Used to block operations on the drive-mirror-replace target */
  42     Error *replace_blocker;
  43     bool is_none_mode;
  44     BlockdevOnError on_source_error, on_target_error;
  45     bool synced;
  46     bool should_complete;
  47     int64_t sector_num;
  48     int64_t granularity;
  49     size_t buf_size;
  50     int64_t bdev_length;
  51     unsigned long *cow_bitmap;
  52     BdrvDirtyBitmap *dirty_bitmap;
  53     HBitmapIter hbi;
  54     uint8_t *buf;
  55     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
  56     int buf_free_count;
  57
  58     unsigned long *in_flight_bitmap;
  59     int in_flight;
  60     int sectors_in_flight;
  61     int ret;
  62     bool unmap;
  63 } MirrorBlockJob;
  64
  65 typedef struct MirrorOp {
  66     MirrorBlockJob *s;
  67     QEMUIOVector qiov;
  68     int64_t sector_num;
  69     int nb_sectors;
  70 } MirrorOp;
  71
  72 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  73                                             int error)
  74 {
  75     s->synced = false;
  76     if (read) {
  77         return block_job_error_action(&s->common, s->common.bs,
  78                                       s->on_source_error, true, error);
  79     } else {
  80         return block_job_error_action(&s->common, s->target,
  81                                       s->on_target_error, false, error);
  82     }
  83 }
  84
  85 static void mirror_iteration_done(MirrorOp *op, int ret)
  86 {
  87     MirrorBlockJob *s = op->s;
  88     struct iovec *iov;
  89     int64_t chunk_num;
  90     int i, nb_chunks, sectors_per_chunk;
  91
  92     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
  93
  94     s->in_flight--;
  95     s->sectors_in_flight -= op->nb_sectors;
  96     iov = op->qiov.iov;
  97     for (i = 0; i < op->qiov.niov; i++) {
  98         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
  99         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
 100         s->buf_free_count++;
 101     }
 102
 103     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 104     chunk_num = op->sector_num / sectors_per_chunk;
 105     nb_chunks = op->nb_sectors / sectors_per_chunk;
 106     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
 107     if (ret >= 0) {
 108         if (s->cow_bitmap) {
 109             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
 110         }
 111         s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
 112     }
 113
 114     qemu_iovec_destroy(&op->qiov);
 115     g_slice_free(MirrorOp, op);
 116
 117     /* Enter coroutine when it is not sleeping.  The coroutine sleeps to
 118      * rate-limit itself.  The coroutine will eventually resume since there is
 119      * a sleep timeout so don't wake it early.
 120      */
 121     if (s->common.busy) {
 122         qemu_coroutine_enter(s->common.co, NULL);
 123     }
 124 }
 125
 126 static void mirror_write_complete(void *opaque, int ret)
 127 {
 128     MirrorOp *op = opaque;
 129     MirrorBlockJob *s = op->s;
 130     if (ret < 0) {
 131         BlockErrorAction action;
 132
 133         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
 134         action = mirror_error_action(s, false, -ret);
 135         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 136             s->ret = ret;
 137         }
 138     }
 139     mirror_iteration_done(op, ret);
 140 }
 141
 142 static void mirror_read_complete(void *opaque, int ret)
 143 {
 144     MirrorOp *op = opaque;
 145     MirrorBlockJob *s = op->s;
 146     if (ret < 0) {
 147         BlockErrorAction action;
 148
 149         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
 150         action = mirror_error_action(s, true, -ret);
 151         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 152             s->ret = ret;
 153         }
 154
 155         mirror_iteration_done(op, ret);
 156         return;
 157     }
 158     bdrv_aio_writev(s->target, op->sector_num, &op->qiov, op->nb_sectors,
 159                     mirror_write_complete, op);
 160 }
 161
 162 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
 163 {
 164     BlockDriverState *source = s->common.bs;
 165     int nb_sectors, sectors_per_chunk, nb_chunks;
 166     int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector;
 167     uint64_t delay_ns = 0;
 168     MirrorOp *op;
 169     int pnum;
 170     int64_t ret;
 171
 172     s->sector_num = hbitmap_iter_next(&s->hbi);
 173     if (s->sector_num < 0) {
 174         bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
 175         s->sector_num = hbitmap_iter_next(&s->hbi);
 176         trace_mirror_restart_iter(s, bdrv_get_dirty_count(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(s->dirty_bitmap, sector_num, nb_sectors);
 292
 293     /* Copy the dirty cluster.  */
 294     s->in_flight++;
 295     s->sectors_in_flight += nb_sectors;
 296     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 297
 298     ret = bdrv_get_block_status_above(source, NULL, sector_num,
 299                                       nb_sectors, &pnum);
 300     if (ret < 0 || pnum < nb_sectors ||
 301             (ret & BDRV_BLOCK_DATA && !(ret & BDRV_BLOCK_ZERO))) {
 302         bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
 303                        mirror_read_complete, op);
 304     } else if (ret & BDRV_BLOCK_ZERO) {
 305         bdrv_aio_write_zeroes(s->target, sector_num, op->nb_sectors,
 306                               s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
 307                               mirror_write_complete, op);
 308     } else {
 309         assert(!(ret & BDRV_BLOCK_DATA));
 310         bdrv_aio_discard(s->target, sector_num, op->nb_sectors,
 311                          mirror_write_complete, op);
 312     }
 313     return delay_ns;
 314 }
 315
 316 static void mirror_free_init(MirrorBlockJob *s)
 317 {
 318     int granularity = s->granularity;
 319     size_t buf_size = s->buf_size;
 320     uint8_t *buf = s->buf;
 321
 322     assert(s->buf_free_count == 0);
 323     QSIMPLEQ_INIT(&s->buf_free);
 324     while (buf_size != 0) {
 325         MirrorBuffer *cur = (MirrorBuffer *)buf;
 326         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 327         s->buf_free_count++;
 328         buf_size -= granularity;
 329         buf += granularity;
 330     }
 331 }
 332
 333 static void mirror_drain(MirrorBlockJob *s)
 334 {
 335     while (s->in_flight > 0) {
 336         qemu_coroutine_yield();
 337     }
 338 }
 339
 340 typedef struct {
 341     int ret;
 342 } MirrorExitData;
 343
 344 static void mirror_exit(BlockJob *job, void *opaque)
 345 {
 346     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 347     MirrorExitData *data = opaque;
 348     AioContext *replace_aio_context = NULL;
 349
 350     if (s->to_replace) {
 351         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 352         aio_context_acquire(replace_aio_context);
 353     }
 354
 355     if (s->should_complete && data->ret == 0) {
 356         BlockDriverState *to_replace = s->common.bs;
 357         if (s->to_replace) {
 358             to_replace = s->to_replace;
 359         }
 360         if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
 361             bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
 362         }
 363         bdrv_swap(s->target, to_replace);
 364         if (s->common.driver->job_type == BLOCK_JOB_TYPE_COMMIT) {
 365             /* drop the bs loop chain formed by the swap: break the loop then
 366              * trigger the unref from the top one */
 367             BlockDriverState *p = s->base->backing_hd;
 368             bdrv_set_backing_hd(s->base, NULL);
 369             bdrv_unref(p);
 370         }
 371     }
 372     if (s->to_replace) {
 373         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
 374         error_free(s->replace_blocker);
 375         bdrv_unref(s->to_replace);
 376     }
 377     if (replace_aio_context) {
 378         aio_context_release(replace_aio_context);
 379     }
 380     g_free(s->replaces);
 381     bdrv_unref(s->target);
 382     block_job_completed(&s->common, data->ret);
 383     g_free(data);
 384 }
 385
 386 static void coroutine_fn mirror_run(void *opaque)
 387 {
 388     MirrorBlockJob *s = opaque;
 389     MirrorExitData *data;
 390     BlockDriverState *bs = s->common.bs;
 391     int64_t sector_num, end, length;
 392     uint64_t last_pause_ns;
 393     BlockDriverInfo bdi;
 394     char backing_filename[2]; /* we only need 2 characters because we are only
 395                                  checking for a NULL string */
 396     int ret = 0;
 397     int n;
 398
 399     if (block_job_is_cancelled(&s->common)) {
 400         goto immediate_exit;
 401     }
 402
 403     s->bdev_length = bdrv_getlength(bs);
 404     if (s->bdev_length < 0) {
 405         ret = s->bdev_length;
 406         goto immediate_exit;
 407     } else if (s->bdev_length == 0) {
 408         /* Report BLOCK_JOB_READY and wait for complete. */
 409         block_job_event_ready(&s->common);
 410         s->synced = true;
 411         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
 412             block_job_yield(&s->common);
 413         }
 414         s->common.cancelled = false;
 415         goto immediate_exit;
 416     }
 417
 418     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
 419     s->in_flight_bitmap = bitmap_new(length);
 420
 421     /* If we have no backing file yet in the destination, we cannot let
 422      * the destination do COW.  Instead, we copy sectors around the
 423      * dirty data if needed.  We need a bitmap to do that.
 424      */
 425     bdrv_get_backing_filename(s->target, backing_filename,
 426                               sizeof(backing_filename));
 427     if (backing_filename[0] && !s->target->backing_hd) {
 428         ret = bdrv_get_info(s->target, &bdi);
 429         if (ret < 0) {
 430             goto immediate_exit;
 431         }
 432         if (s->granularity < bdi.cluster_size) {
 433             s->buf_size = MAX(s->buf_size, bdi.cluster_size);
 434             s->cow_bitmap = bitmap_new(length);
 435         }
 436     }
 437
 438     end = s->bdev_length / BDRV_SECTOR_SIZE;
 439     s->buf = qemu_try_blockalign(bs, s->buf_size);
 440     if (s->buf == NULL) {
 441         ret = -ENOMEM;
 442         goto immediate_exit;
 443     }
 444
 445     mirror_free_init(s);
 446
 447     last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 448     if (!s->is_none_mode) {
 449         /* First part, loop on the sectors and initialize the dirty bitmap.  */
 450         BlockDriverState *base = s->base;
 451         for (sector_num = 0; sector_num < end; ) {
 452             /* Just to make sure we are not exceeding int limit. */
 453             int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS,
 454                                  end - sector_num);
 455             int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 456
 457             if (now - last_pause_ns > SLICE_TIME) {
 458                 last_pause_ns = now;
 459                 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
 460             }
 461
 462             if (block_job_is_cancelled(&s->common)) {
 463                 goto immediate_exit;
 464             }
 465
 466             ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n);
 467
 468             if (ret < 0) {
 469                 goto immediate_exit;
 470             }
 471
 472             assert(n > 0);
 473             if (ret == 1) {
 474                 bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
 475             }
 476             sector_num += n;
 477         }
 478     }
 479
 480     bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
 481     for (;;) {
 482         uint64_t delay_ns = 0;
 483         int64_t cnt;
 484         bool should_complete;
 485
 486         if (s->ret < 0) {
 487             ret = s->ret;
 488             goto immediate_exit;
 489         }
 490
 491         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 492         /* s->common.offset contains the number of bytes already processed so
 493          * far, cnt is the number of dirty sectors remaining and
 494          * s->sectors_in_flight is the number of sectors currently being
 495          * processed; together those are the current total operation length */
 496         s->common.len = s->common.offset +
 497                         (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
 498
 499         /* Note that even when no rate limit is applied we need to yield
 500          * periodically with no pending I/O so that bdrv_drain_all() returns.
 501          * We do so every SLICE_TIME nanoseconds, or when there is an error,
 502          * or when the source is clean, whichever comes first.
 503          */
 504         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
 505             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 506             if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 507                 (cnt == 0 && s->in_flight > 0)) {
 508                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
 509                 qemu_coroutine_yield();
 510                 continue;
 511             } else if (cnt != 0) {
 512                 delay_ns = mirror_iteration(s);
 513             }
 514         }
 515
 516         should_complete = false;
 517         if (s->in_flight == 0 && cnt == 0) {
 518             trace_mirror_before_flush(s);
 519             ret = bdrv_flush(s->target);
 520             if (ret < 0) {
 521                 if (mirror_error_action(s, false, -ret) ==
 522                     BLOCK_ERROR_ACTION_REPORT) {
 523                     goto immediate_exit;
 524                 }
 525             } else {
 526                 /* We're out of the streaming phase.  From now on, if the job
 527                  * is cancelled we will actually complete all pending I/O and
 528                  * report completion.  This way, block-job-cancel will leave
 529                  * the target in a consistent state.
 530                  */
 531                 if (!s->synced) {
 532                     block_job_event_ready(&s->common);
 533                     s->synced = true;
 534                 }
 535
 536                 should_complete = s->should_complete ||
 537                     block_job_is_cancelled(&s->common);
 538                 cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 539             }
 540         }
 541
 542         if (cnt == 0 && should_complete) {
 543             /* The dirty bitmap is not updated while operations are pending.
 544              * If we're about to exit, wait for pending operations before
 545              * calling bdrv_get_dirty_count(bs), or we may exit while the
 546              * source has dirty data to copy!
 547              *
 548              * Note that I/O can be submitted by the guest while
 549              * mirror_populate runs.
 550              */
 551             trace_mirror_before_drain(s, cnt);
 552             bdrv_drain(bs);
 553             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 554         }
 555
 556         ret = 0;
 557         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
 558         if (!s->synced) {
 559             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 560             if (block_job_is_cancelled(&s->common)) {
 561                 break;
 562             }
 563         } else if (!should_complete) {
 564             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 565             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 566         } else if (cnt == 0) {
 567             /* The two disks are in sync.  Exit and report successful
 568              * completion.
 569              */
 570             assert(QLIST_EMPTY(&bs->tracked_requests));
 571             s->common.cancelled = false;
 572             break;
 573         }
 574         last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 575     }
 576
 577 immediate_exit:
 578     if (s->in_flight > 0) {
 579         /* We get here only if something went wrong.  Either the job failed,
 580          * or it was cancelled prematurely so that we do not guarantee that
 581          * the target is a copy of the source.
 582          */
 583         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 584         mirror_drain(s);
 585     }
 586
 587     assert(s->in_flight == 0);
 588     qemu_vfree(s->buf);
 589     g_free(s->cow_bitmap);
 590     g_free(s->in_flight_bitmap);
 591     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
 592     bdrv_iostatus_disable(s->target);
 593
 594     data = g_malloc(sizeof(*data));
 595     data->ret = ret;
 596     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
 597 }
 598
 599 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 600 {
 601     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 602
 603     if (speed < 0) {
 604         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 605         return;
 606     }
 607     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 608 }
 609
 610 static void mirror_iostatus_reset(BlockJob *job)
 611 {
 612     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 613
 614     bdrv_iostatus_reset(s->target);
 615 }
 616
 617 static void mirror_complete(BlockJob *job, Error **errp)
 618 {
 619     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 620     Error *local_err = NULL;
 621     int ret;
 622
 623     ret = bdrv_open_backing_file(s->target, NULL, &local_err);
 624     if (ret < 0) {
 625         error_propagate(errp, local_err);
 626         return;
 627     }
 628     if (!s->synced) {
 629         error_setg(errp, QERR_BLOCK_JOB_NOT_READY,
 630                    bdrv_get_device_name(job->bs));
 631         return;
 632     }
 633
 634     /* check the target bs is not blocked and block all operations on it */
 635     if (s->replaces) {
 636         AioContext *replace_aio_context;
 637
 638         s->to_replace = check_to_replace_node(s->replaces, &local_err);
 639         if (!s->to_replace) {
 640             error_propagate(errp, local_err);
 641             return;
 642         }
 643
 644         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 645         aio_context_acquire(replace_aio_context);
 646
 647         error_setg(&s->replace_blocker,
 648                    "block device is in use by block-job-complete");
 649         bdrv_op_block_all(s->to_replace, s->replace_blocker);
 650         bdrv_ref(s->to_replace);
 651
 652         aio_context_release(replace_aio_context);
 653     }
 654
 655     s->should_complete = true;
 656     block_job_enter(&s->common);
 657 }
 658
 659 static const BlockJobDriver mirror_job_driver = {
 660     .instance_size = sizeof(MirrorBlockJob),
 661     .job_type      = BLOCK_JOB_TYPE_MIRROR,
 662     .set_speed     = mirror_set_speed,
 663     .iostatus_reset= mirror_iostatus_reset,
 664     .complete      = mirror_complete,
 665 };
 666
 667 static const BlockJobDriver commit_active_job_driver = {
 668     .instance_size = sizeof(MirrorBlockJob),
 669     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 670     .set_speed     = mirror_set_speed,
 671     .iostatus_reset
 672                    = mirror_iostatus_reset,
 673     .complete      = mirror_complete,
 674 };
 675
 676 static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
 677                              const char *replaces,
 678                              int64_t speed, uint32_t granularity,
 679                              int64_t buf_size,
 680                              BlockdevOnError on_source_error,
 681                              BlockdevOnError on_target_error,
 682                              bool unmap,
 683                              BlockCompletionFunc *cb,
 684                              void *opaque, Error **errp,
 685                              const BlockJobDriver *driver,
 686                              bool is_none_mode, BlockDriverState *base)
 687 {
 688     MirrorBlockJob *s;
 689
 690     if (granularity == 0) {
 691         granularity = bdrv_get_default_bitmap_granularity(target);
 692     }
 693
 694     assert ((granularity & (granularity - 1)) == 0);
 695
 696     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 697          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 698         !bdrv_iostatus_is_enabled(bs)) {
 699         error_setg(errp, QERR_INVALID_PARAMETER, "on-source-error");
 700         return;
 701     }
 702
 703     if (buf_size < 0) {
 704         error_setg(errp, "Invalid parameter 'buf-size'");
 705         return;
 706     }
 707
 708     if (buf_size == 0) {
 709         buf_size = DEFAULT_MIRROR_BUF_SIZE;
 710     }
 711
 712     s = block_job_create(driver, bs, speed, cb, opaque, errp);
 713     if (!s) {
 714         return;
 715     }
 716
 717     s->replaces = g_strdup(replaces);
 718     s->on_source_error = on_source_error;
 719     s->on_target_error = on_target_error;
 720     s->target = target;
 721     s->is_none_mode = is_none_mode;
 722     s->base = base;
 723     s->granularity = granularity;
 724     s->buf_size = ROUND_UP(buf_size, granularity);
 725     s->unmap = unmap;
 726
 727     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
 728     if (!s->dirty_bitmap) {
 729         g_free(s->replaces);
 730         block_job_release(bs);
 731         return;
 732     }
 733     bdrv_set_enable_write_cache(s->target, true);
 734     bdrv_set_on_error(s->target, on_target_error, on_target_error);
 735     bdrv_iostatus_enable(s->target);
 736     s->common.co = qemu_coroutine_create(mirror_run);
 737     trace_mirror_start(bs, s, s->common.co, opaque);
 738     qemu_coroutine_enter(s->common.co, s);
 739 }
 740
 741 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
 742                   const char *replaces,
 743                   int64_t speed, uint32_t granularity, int64_t buf_size,
 744                   MirrorSyncMode mode, BlockdevOnError on_source_error,
 745                   BlockdevOnError on_target_error,
 746                   bool unmap,
 747                   BlockCompletionFunc *cb,
 748                   void *opaque, Error **errp)
 749 {
 750     bool is_none_mode;
 751     BlockDriverState *base;
 752
 753     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 754         error_setg(errp, "Sync mode 'incremental' not supported");
 755         return;
 756     }
 757     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
 758     base = mode == MIRROR_SYNC_MODE_TOP ? bs->backing_hd : NULL;
 759     mirror_start_job(bs, target, replaces,
 760                      speed, granularity, buf_size,
 761                      on_source_error, on_target_error, unmap, cb, opaque, errp,
 762                      &mirror_job_driver, is_none_mode, base);
 763 }
 764
 765 void commit_active_start(BlockDriverState *bs, BlockDriverState *base,
 766                          int64_t speed,
 767                          BlockdevOnError on_error,
 768                          BlockCompletionFunc *cb,
 769                          void *opaque, Error **errp)
 770 {
 771     int64_t length, base_length;
 772     int orig_base_flags;
 773     int ret;
 774     Error *local_err = NULL;
 775
 776     orig_base_flags = bdrv_get_flags(base);
 777
 778     if (bdrv_reopen(base, bs->open_flags, errp)) {
 779         return;
 780     }
 781
 782     length = bdrv_getlength(bs);
 783     if (length < 0) {
 784         error_setg_errno(errp, -length,
 785                          "Unable to determine length of %s", bs->filename);
 786         goto error_restore_flags;
 787     }
 788
 789     base_length = bdrv_getlength(base);
 790     if (base_length < 0) {
 791         error_setg_errno(errp, -base_length,
 792                          "Unable to determine length of %s", base->filename);
 793         goto error_restore_flags;
 794     }
 795
 796     if (length > base_length) {
 797         ret = bdrv_truncate(base, length);
 798         if (ret < 0) {
 799             error_setg_errno(errp, -ret,
 800                             "Top image %s is larger than base image %s, and "
 801                              "resize of base image failed",
 802                              bs->filename, base->filename);
 803             goto error_restore_flags;
 804         }
 805     }
 806
 807     bdrv_ref(base);
 808     mirror_start_job(bs, base, NULL, speed, 0, 0,
 809                      on_error, on_error, false, cb, opaque, &local_err,
 810                      &commit_active_job_driver, false, base);
 811     if (local_err) {
 812         error_propagate(errp, local_err);
 813         goto error_restore_flags;
 814     }
 815
 816     return;
 817
 818 error_restore_flags:
 819     /* ignore error and errp for bdrv_reopen, because we want to propagate
 820      * the original error */
 821     bdrv_reopen(base, orig_base_flags, NULL);
 822     return;
 823 }