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 "sysemu/block-backend.h"
  18 #include "qapi/qmp/qerror.h"
  19 #include "qemu/ratelimit.h"
  20 #include "qemu/bitmap.h"
  21
  22 #define SLICE_TIME    100000000ULL /* ns */
  23 #define MAX_IN_FLIGHT 16
  24 #define DEFAULT_MIRROR_BUF_SIZE   (10 << 20)
  25
  26 /* The mirroring buffer is a list of granularity-sized chunks.
  27  * Free chunks are organized in a list.
  28  */
  29 typedef struct MirrorBuffer {
  30     QSIMPLEQ_ENTRY(MirrorBuffer) next;
  31 } MirrorBuffer;
  32
  33 typedef struct MirrorBlockJob {
  34     BlockJob common;
  35     RateLimit limit;
  36     BlockDriverState *target;
  37     BlockDriverState *base;
  38     /* The name of the graph node to replace */
  39     char *replaces;
  40     /* The BDS to replace */
  41     BlockDriverState *to_replace;
  42     /* Used to block operations on the drive-mirror-replace target */
  43     Error *replace_blocker;
  44     bool is_none_mode;
  45     BlockdevOnError on_source_error, on_target_error;
  46     bool synced;
  47     bool should_complete;
  48     int64_t sector_num;
  49     int64_t granularity;
  50     size_t buf_size;
  51     int64_t bdev_length;
  52     unsigned long *cow_bitmap;
  53     BdrvDirtyBitmap *dirty_bitmap;
  54     HBitmapIter hbi;
  55     uint8_t *buf;
  56     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
  57     int buf_free_count;
  58
  59     unsigned long *in_flight_bitmap;
  60     int in_flight;
  61     int sectors_in_flight;
  62     int ret;
  63     bool unmap;
  64     bool waiting_for_io;
  65 } MirrorBlockJob;
  66
  67 typedef struct MirrorOp {
  68     MirrorBlockJob *s;
  69     QEMUIOVector qiov;
  70     int64_t sector_num;
  71     int nb_sectors;
  72 } MirrorOp;
  73
  74 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  75                                             int error)
  76 {
  77     s->synced = false;
  78     if (read) {
  79         return block_job_error_action(&s->common, s->common.bs,
  80                                       s->on_source_error, true, error);
  81     } else {
  82         return block_job_error_action(&s->common, s->target,
  83                                       s->on_target_error, false, error);
  84     }
  85 }
  86
  87 static void mirror_iteration_done(MirrorOp *op, int ret)
  88 {
  89     MirrorBlockJob *s = op->s;
  90     struct iovec *iov;
  91     int64_t chunk_num;
  92     int i, nb_chunks, sectors_per_chunk;
  93
  94     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
  95
  96     s->in_flight--;
  97     s->sectors_in_flight -= op->nb_sectors;
  98     iov = op->qiov.iov;
  99     for (i = 0; i < op->qiov.niov; i++) {
 100         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
 101         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
 102         s->buf_free_count++;
 103     }
 104
 105     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 106     chunk_num = op->sector_num / sectors_per_chunk;
 107     nb_chunks = op->nb_sectors / sectors_per_chunk;
 108     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
 109     if (ret >= 0) {
 110         if (s->cow_bitmap) {
 111             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
 112         }
 113         s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
 114     }
 115
 116     qemu_iovec_destroy(&op->qiov);
 117     g_free(op);
 118
 119     if (s->waiting_for_io) {
 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     int pnum;
 168     int64_t ret;
 169
 170     s->sector_num = hbitmap_iter_next(&s->hbi);
 171     if (s->sector_num < 0) {
 172         bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
 173         s->sector_num = hbitmap_iter_next(&s->hbi);
 174         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
 175         assert(s->sector_num >= 0);
 176     }
 177
 178     hbitmap_next_sector = s->sector_num;
 179     sector_num = s->sector_num;
 180     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 181     end = s->bdev_length / BDRV_SECTOR_SIZE;
 182
 183     /* Extend the QEMUIOVector to include all adjacent blocks that will
 184      * be copied in this operation.
 185      *
 186      * We have to do this if we have no backing file yet in the destination,
 187      * and the cluster size is very large.  Then we need to do COW ourselves.
 188      * The first time a cluster is copied, copy it entirely.  Note that,
 189      * because both the granularity and the cluster size are powers of two,
 190      * the number of sectors to copy cannot exceed one cluster.
 191      *
 192      * We also want to extend the QEMUIOVector to include more adjacent
 193      * dirty blocks if possible, to limit the number of I/O operations and
 194      * run efficiently even with a small granularity.
 195      */
 196     nb_chunks = 0;
 197     nb_sectors = 0;
 198     next_sector = sector_num;
 199     next_chunk = sector_num / sectors_per_chunk;
 200
 201     /* Wait for I/O to this cluster (from a previous iteration) to be done.  */
 202     while (test_bit(next_chunk, s->in_flight_bitmap)) {
 203         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 204         s->waiting_for_io = true;
 205         qemu_coroutine_yield();
 206         s->waiting_for_io = false;
 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             s->waiting_for_io = true;
 243             qemu_coroutine_yield();
 244             s->waiting_for_io = false;
 245         }
 246         if (s->buf_free_count < nb_chunks + added_chunks) {
 247             trace_mirror_break_buf_busy(s, nb_chunks, s->in_flight);
 248             break;
 249         }
 250         if (IOV_MAX < nb_chunks + added_chunks) {
 251             trace_mirror_break_iov_max(s, nb_chunks, added_chunks);
 252             break;
 253         }
 254
 255         /* We have enough free space to copy these sectors.  */
 256         bitmap_set(s->in_flight_bitmap, next_chunk, added_chunks);
 257
 258         nb_sectors += added_sectors;
 259         nb_chunks += added_chunks;
 260         next_sector += added_sectors;
 261         next_chunk += added_chunks;
 262         if (!s->synced && s->common.speed) {
 263             delay_ns = ratelimit_calculate_delay(&s->limit, added_sectors);
 264         }
 265     } while (delay_ns == 0 && next_sector < end);
 266
 267     /* Allocate a MirrorOp that is used as an AIO callback.  */
 268     op = g_new(MirrorOp, 1);
 269     op->s = s;
 270     op->sector_num = sector_num;
 271     op->nb_sectors = nb_sectors;
 272
 273     /* Now make a QEMUIOVector taking enough granularity-sized chunks
 274      * from s->buf_free.
 275      */
 276     qemu_iovec_init(&op->qiov, nb_chunks);
 277     next_sector = sector_num;
 278     while (nb_chunks-- > 0) {
 279         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
 280         size_t remaining = (nb_sectors * BDRV_SECTOR_SIZE) - op->qiov.size;
 281
 282         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
 283         s->buf_free_count--;
 284         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
 285
 286         /* Advance the HBitmapIter in parallel, so that we do not examine
 287          * the same sector twice.
 288          */
 289         if (next_sector > hbitmap_next_sector
 290             && bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
 291             hbitmap_next_sector = hbitmap_iter_next(&s->hbi);
 292         }
 293
 294         next_sector += sectors_per_chunk;
 295     }
 296
 297     bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num, nb_sectors);
 298
 299     /* Copy the dirty cluster.  */
 300     s->in_flight++;
 301     s->sectors_in_flight += nb_sectors;
 302     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 303
 304     ret = bdrv_get_block_status_above(source, NULL, sector_num,
 305                                       nb_sectors, &pnum);
 306     if (ret < 0 || pnum < nb_sectors ||
 307             (ret & BDRV_BLOCK_DATA && !(ret & BDRV_BLOCK_ZERO))) {
 308         bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
 309                        mirror_read_complete, op);
 310     } else if (ret & BDRV_BLOCK_ZERO) {
 311         bdrv_aio_write_zeroes(s->target, sector_num, op->nb_sectors,
 312                               s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
 313                               mirror_write_complete, op);
 314     } else {
 315         assert(!(ret & BDRV_BLOCK_DATA));
 316         bdrv_aio_discard(s->target, sector_num, op->nb_sectors,
 317                          mirror_write_complete, op);
 318     }
 319     return delay_ns;
 320 }
 321
 322 static void mirror_free_init(MirrorBlockJob *s)
 323 {
 324     int granularity = s->granularity;
 325     size_t buf_size = s->buf_size;
 326     uint8_t *buf = s->buf;
 327
 328     assert(s->buf_free_count == 0);
 329     QSIMPLEQ_INIT(&s->buf_free);
 330     while (buf_size != 0) {
 331         MirrorBuffer *cur = (MirrorBuffer *)buf;
 332         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 333         s->buf_free_count++;
 334         buf_size -= granularity;
 335         buf += granularity;
 336     }
 337 }
 338
 339 static void mirror_drain(MirrorBlockJob *s)
 340 {
 341     while (s->in_flight > 0) {
 342         s->waiting_for_io = true;
 343         qemu_coroutine_yield();
 344         s->waiting_for_io = false;
 345     }
 346 }
 347
 348 typedef struct {
 349     int ret;
 350 } MirrorExitData;
 351
 352 static void mirror_exit(BlockJob *job, void *opaque)
 353 {
 354     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 355     MirrorExitData *data = opaque;
 356     AioContext *replace_aio_context = NULL;
 357     BlockDriverState *src = s->common.bs;
 358
 359     /* Make sure that the source BDS doesn't go away before we called
 360      * block_job_completed(). */
 361     bdrv_ref(src);
 362
 363     if (s->to_replace) {
 364         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 365         aio_context_acquire(replace_aio_context);
 366     }
 367
 368     if (s->should_complete && data->ret == 0) {
 369         BlockDriverState *to_replace = s->common.bs;
 370         if (s->to_replace) {
 371             to_replace = s->to_replace;
 372         }
 373         if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
 374             bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
 375         }
 376         bdrv_replace_in_backing_chain(to_replace, s->target);
 377     }
 378     if (s->to_replace) {
 379         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
 380         error_free(s->replace_blocker);
 381         bdrv_unref(s->to_replace);
 382     }
 383     if (replace_aio_context) {
 384         aio_context_release(replace_aio_context);
 385     }
 386     g_free(s->replaces);
 387     bdrv_op_unblock_all(s->target, s->common.blocker);
 388     bdrv_unref(s->target);
 389     block_job_completed(&s->common, data->ret);
 390     g_free(data);
 391     bdrv_drained_end(src);
 392     bdrv_unref(src);
 393 }
 394
 395 static void coroutine_fn mirror_run(void *opaque)
 396 {
 397     MirrorBlockJob *s = opaque;
 398     MirrorExitData *data;
 399     BlockDriverState *bs = s->common.bs;
 400     int64_t sector_num, end, length;
 401     uint64_t last_pause_ns;
 402     BlockDriverInfo bdi;
 403     char backing_filename[2]; /* we only need 2 characters because we are only
 404                                  checking for a NULL string */
 405     int ret = 0;
 406     int n;
 407
 408     if (block_job_is_cancelled(&s->common)) {
 409         goto immediate_exit;
 410     }
 411
 412     s->bdev_length = bdrv_getlength(bs);
 413     if (s->bdev_length < 0) {
 414         ret = s->bdev_length;
 415         goto immediate_exit;
 416     } else if (s->bdev_length == 0) {
 417         /* Report BLOCK_JOB_READY and wait for complete. */
 418         block_job_event_ready(&s->common);
 419         s->synced = true;
 420         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
 421             block_job_yield(&s->common);
 422         }
 423         s->common.cancelled = false;
 424         goto immediate_exit;
 425     }
 426
 427     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
 428     s->in_flight_bitmap = bitmap_new(length);
 429
 430     /* If we have no backing file yet in the destination, we cannot let
 431      * the destination do COW.  Instead, we copy sectors around the
 432      * dirty data if needed.  We need a bitmap to do that.
 433      */
 434     bdrv_get_backing_filename(s->target, backing_filename,
 435                               sizeof(backing_filename));
 436     if (backing_filename[0] && !s->target->backing) {
 437         ret = bdrv_get_info(s->target, &bdi);
 438         if (ret < 0) {
 439             goto immediate_exit;
 440         }
 441         if (s->granularity < bdi.cluster_size) {
 442             s->buf_size = MAX(s->buf_size, bdi.cluster_size);
 443             s->cow_bitmap = bitmap_new(length);
 444         }
 445     }
 446
 447     end = s->bdev_length / BDRV_SECTOR_SIZE;
 448     s->buf = qemu_try_blockalign(bs, s->buf_size);
 449     if (s->buf == NULL) {
 450         ret = -ENOMEM;
 451         goto immediate_exit;
 452     }
 453
 454     mirror_free_init(s);
 455
 456     last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 457     if (!s->is_none_mode) {
 458         /* First part, loop on the sectors and initialize the dirty bitmap.  */
 459         BlockDriverState *base = s->base;
 460         bool mark_all_dirty = s->base == NULL && !bdrv_has_zero_init(s->target);
 461
 462         for (sector_num = 0; sector_num < end; ) {
 463             /* Just to make sure we are not exceeding int limit. */
 464             int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS,
 465                                  end - sector_num);
 466             int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 467
 468             if (now - last_pause_ns > SLICE_TIME) {
 469                 last_pause_ns = now;
 470                 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
 471             }
 472
 473             if (block_job_is_cancelled(&s->common)) {
 474                 goto immediate_exit;
 475             }
 476
 477             ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n);
 478
 479             if (ret < 0) {
 480                 goto immediate_exit;
 481             }
 482
 483             assert(n > 0);
 484             if (ret == 1 || mark_all_dirty) {
 485                 bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
 486             }
 487             sector_num += n;
 488         }
 489     }
 490
 491     bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
 492     for (;;) {
 493         uint64_t delay_ns = 0;
 494         int64_t cnt;
 495         bool should_complete;
 496
 497         if (s->ret < 0) {
 498             ret = s->ret;
 499             goto immediate_exit;
 500         }
 501
 502         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 503         /* s->common.offset contains the number of bytes already processed so
 504          * far, cnt is the number of dirty sectors remaining and
 505          * s->sectors_in_flight is the number of sectors currently being
 506          * processed; together those are the current total operation length */
 507         s->common.len = s->common.offset +
 508                         (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
 509
 510         /* Note that even when no rate limit is applied we need to yield
 511          * periodically with no pending I/O so that bdrv_drain_all() returns.
 512          * We do so every SLICE_TIME nanoseconds, or when there is an error,
 513          * or when the source is clean, whichever comes first.
 514          */
 515         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
 516             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 517             if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 518                 (cnt == 0 && s->in_flight > 0)) {
 519                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
 520                 s->waiting_for_io = true;
 521                 qemu_coroutine_yield();
 522                 s->waiting_for_io = false;
 523                 continue;
 524             } else if (cnt != 0) {
 525                 delay_ns = mirror_iteration(s);
 526             }
 527         }
 528
 529         should_complete = false;
 530         if (s->in_flight == 0 && cnt == 0) {
 531             trace_mirror_before_flush(s);
 532             ret = bdrv_flush(s->target);
 533             if (ret < 0) {
 534                 if (mirror_error_action(s, false, -ret) ==
 535                     BLOCK_ERROR_ACTION_REPORT) {
 536                     goto immediate_exit;
 537                 }
 538             } else {
 539                 /* We're out of the streaming phase.  From now on, if the job
 540                  * is cancelled we will actually complete all pending I/O and
 541                  * report completion.  This way, block-job-cancel will leave
 542                  * the target in a consistent state.
 543                  */
 544                 if (!s->synced) {
 545                     block_job_event_ready(&s->common);
 546                     s->synced = true;
 547                 }
 548
 549                 should_complete = s->should_complete ||
 550                     block_job_is_cancelled(&s->common);
 551                 cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 552             }
 553         }
 554
 555         if (cnt == 0 && should_complete) {
 556             /* The dirty bitmap is not updated while operations are pending.
 557              * If we're about to exit, wait for pending operations before
 558              * calling bdrv_get_dirty_count(bs), or we may exit while the
 559              * source has dirty data to copy!
 560              *
 561              * Note that I/O can be submitted by the guest while
 562              * mirror_populate runs.
 563              */
 564             trace_mirror_before_drain(s, cnt);
 565             bdrv_drain(bs);
 566             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 567         }
 568
 569         ret = 0;
 570         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
 571         if (!s->synced) {
 572             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 573             if (block_job_is_cancelled(&s->common)) {
 574                 break;
 575             }
 576         } else if (!should_complete) {
 577             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 578             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 579         } else if (cnt == 0) {
 580             /* The two disks are in sync.  Exit and report successful
 581              * completion.
 582              */
 583             assert(QLIST_EMPTY(&bs->tracked_requests));
 584             s->common.cancelled = false;
 585             break;
 586         }
 587         last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 588     }
 589
 590 immediate_exit:
 591     if (s->in_flight > 0) {
 592         /* We get here only if something went wrong.  Either the job failed,
 593          * or it was cancelled prematurely so that we do not guarantee that
 594          * the target is a copy of the source.
 595          */
 596         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 597         mirror_drain(s);
 598     }
 599
 600     assert(s->in_flight == 0);
 601     qemu_vfree(s->buf);
 602     g_free(s->cow_bitmap);
 603     g_free(s->in_flight_bitmap);
 604     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
 605     if (s->target->blk) {
 606         blk_iostatus_disable(s->target->blk);
 607     }
 608
 609     data = g_malloc(sizeof(*data));
 610     data->ret = ret;
 611     /* Before we switch to target in mirror_exit, make sure data doesn't
 612      * change. */
 613     bdrv_drained_begin(s->common.bs);
 614     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
 615 }
 616
 617 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 618 {
 619     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 620
 621     if (speed < 0) {
 622         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 623         return;
 624     }
 625     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 626 }
 627
 628 static void mirror_iostatus_reset(BlockJob *job)
 629 {
 630     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 631
 632     if (s->target->blk) {
 633         blk_iostatus_reset(s->target->blk);
 634     }
 635 }
 636
 637 static void mirror_complete(BlockJob *job, Error **errp)
 638 {
 639     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 640     Error *local_err = NULL;
 641     int ret;
 642
 643     ret = bdrv_open_backing_file(s->target, NULL, &local_err);
 644     if (ret < 0) {
 645         error_propagate(errp, local_err);
 646         return;
 647     }
 648     if (!s->synced) {
 649         error_setg(errp, QERR_BLOCK_JOB_NOT_READY, job->id);
 650         return;
 651     }
 652
 653     /* check the target bs is not blocked and block all operations on it */
 654     if (s->replaces) {
 655         AioContext *replace_aio_context;
 656
 657         s->to_replace = bdrv_find_node(s->replaces);
 658         if (!s->to_replace) {
 659             error_setg(errp, "Node name '%s' not found", s->replaces);
 660             return;
 661         }
 662
 663         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 664         aio_context_acquire(replace_aio_context);
 665
 666         error_setg(&s->replace_blocker,
 667                    "block device is in use by block-job-complete");
 668         bdrv_op_block_all(s->to_replace, s->replace_blocker);
 669         bdrv_ref(s->to_replace);
 670
 671         aio_context_release(replace_aio_context);
 672     }
 673
 674     s->should_complete = true;
 675     block_job_enter(&s->common);
 676 }
 677
 678 static const BlockJobDriver mirror_job_driver = {
 679     .instance_size = sizeof(MirrorBlockJob),
 680     .job_type      = BLOCK_JOB_TYPE_MIRROR,
 681     .set_speed     = mirror_set_speed,
 682     .iostatus_reset= mirror_iostatus_reset,
 683     .complete      = mirror_complete,
 684 };
 685
 686 static const BlockJobDriver commit_active_job_driver = {
 687     .instance_size = sizeof(MirrorBlockJob),
 688     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 689     .set_speed     = mirror_set_speed,
 690     .iostatus_reset
 691                    = mirror_iostatus_reset,
 692     .complete      = mirror_complete,
 693 };
 694
 695 static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
 696                              const char *replaces,
 697                              int64_t speed, uint32_t granularity,
 698                              int64_t buf_size,
 699                              BlockdevOnError on_source_error,
 700                              BlockdevOnError on_target_error,
 701                              bool unmap,
 702                              BlockCompletionFunc *cb,
 703                              void *opaque, Error **errp,
 704                              const BlockJobDriver *driver,
 705                              bool is_none_mode, BlockDriverState *base)
 706 {
 707     MirrorBlockJob *s;
 708
 709     if (granularity == 0) {
 710         granularity = bdrv_get_default_bitmap_granularity(target);
 711     }
 712
 713     assert ((granularity & (granularity - 1)) == 0);
 714
 715     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 716          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 717         (!bs->blk || !blk_iostatus_is_enabled(bs->blk))) {
 718         error_setg(errp, QERR_INVALID_PARAMETER, "on-source-error");
 719         return;
 720     }
 721
 722     if (buf_size < 0) {
 723         error_setg(errp, "Invalid parameter 'buf-size'");
 724         return;
 725     }
 726
 727     if (buf_size == 0) {
 728         buf_size = DEFAULT_MIRROR_BUF_SIZE;
 729     }
 730
 731     s = block_job_create(driver, bs, speed, cb, opaque, errp);
 732     if (!s) {
 733         return;
 734     }
 735
 736     s->replaces = g_strdup(replaces);
 737     s->on_source_error = on_source_error;
 738     s->on_target_error = on_target_error;
 739     s->target = target;
 740     s->is_none_mode = is_none_mode;
 741     s->base = base;
 742     s->granularity = granularity;
 743     s->buf_size = ROUND_UP(buf_size, granularity);
 744     s->unmap = unmap;
 745
 746     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
 747     if (!s->dirty_bitmap) {
 748         g_free(s->replaces);
 749         block_job_unref(&s->common);
 750         return;
 751     }
 752
 753     bdrv_op_block_all(s->target, s->common.blocker);
 754
 755     bdrv_set_enable_write_cache(s->target, true);
 756     if (s->target->blk) {
 757         blk_set_on_error(s->target->blk, on_target_error, on_target_error);
 758         blk_iostatus_enable(s->target->blk);
 759     }
 760     s->common.co = qemu_coroutine_create(mirror_run);
 761     trace_mirror_start(bs, s, s->common.co, opaque);
 762     qemu_coroutine_enter(s->common.co, s);
 763 }
 764
 765 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
 766                   const char *replaces,
 767                   int64_t speed, uint32_t granularity, int64_t buf_size,
 768                   MirrorSyncMode mode, BlockdevOnError on_source_error,
 769                   BlockdevOnError on_target_error,
 770                   bool unmap,
 771                   BlockCompletionFunc *cb,
 772                   void *opaque, Error **errp)
 773 {
 774     bool is_none_mode;
 775     BlockDriverState *base;
 776
 777     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 778         error_setg(errp, "Sync mode 'incremental' not supported");
 779         return;
 780     }
 781     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
 782     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
 783     mirror_start_job(bs, target, replaces,
 784                      speed, granularity, buf_size,
 785                      on_source_error, on_target_error, unmap, cb, opaque, errp,
 786                      &mirror_job_driver, is_none_mode, base);
 787 }
 788
 789 void commit_active_start(BlockDriverState *bs, BlockDriverState *base,
 790                          int64_t speed,
 791                          BlockdevOnError on_error,
 792                          BlockCompletionFunc *cb,
 793                          void *opaque, Error **errp)
 794 {
 795     int64_t length, base_length;
 796     int orig_base_flags;
 797     int ret;
 798     Error *local_err = NULL;
 799
 800     orig_base_flags = bdrv_get_flags(base);
 801
 802     if (bdrv_reopen(base, bs->open_flags, errp)) {
 803         return;
 804     }
 805
 806     length = bdrv_getlength(bs);
 807     if (length < 0) {
 808         error_setg_errno(errp, -length,
 809                          "Unable to determine length of %s", bs->filename);
 810         goto error_restore_flags;
 811     }
 812
 813     base_length = bdrv_getlength(base);
 814     if (base_length < 0) {
 815         error_setg_errno(errp, -base_length,
 816                          "Unable to determine length of %s", base->filename);
 817         goto error_restore_flags;
 818     }
 819
 820     if (length > base_length) {
 821         ret = bdrv_truncate(base, length);
 822         if (ret < 0) {
 823             error_setg_errno(errp, -ret,
 824                             "Top image %s is larger than base image %s, and "
 825                              "resize of base image failed",
 826                              bs->filename, base->filename);
 827             goto error_restore_flags;
 828         }
 829     }
 830
 831     bdrv_ref(base);
 832     mirror_start_job(bs, base, NULL, speed, 0, 0,
 833                      on_error, on_error, false, cb, opaque, &local_err,
 834                      &commit_active_job_driver, false, base);
 835     if (local_err) {
 836         error_propagate(errp, local_err);
 837         goto error_restore_flags;
 838     }
 839
 840     return;
 841
 842 error_restore_flags:
 843     /* ignore error and errp for bdrv_reopen, because we want to propagate
 844      * the original error */
 845     bdrv_reopen(base, orig_base_flags, NULL);
 846     return;
 847 }