qemu/block/mirror.c
<<
>>
Prefs
   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 "qemu/osdep.h"
  15#include "qemu/cutils.h"
  16#include "trace.h"
  17#include "block/blockjob_int.h"
  18#include "block/block_int.h"
  19#include "sysemu/block-backend.h"
  20#include "qapi/error.h"
  21#include "qapi/qmp/qerror.h"
  22#include "qemu/ratelimit.h"
  23#include "qemu/bitmap.h"
  24
  25#define SLICE_TIME    100000000ULL /* ns */
  26#define MAX_IN_FLIGHT 16
  27#define MAX_IO_BYTES (1 << 20) /* 1 Mb */
  28#define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
  29
  30/* The mirroring buffer is a list of granularity-sized chunks.
  31 * Free chunks are organized in a list.
  32 */
  33typedef struct MirrorBuffer {
  34    QSIMPLEQ_ENTRY(MirrorBuffer) next;
  35} MirrorBuffer;
  36
  37typedef struct MirrorBlockJob {
  38    BlockJob common;
  39    RateLimit limit;
  40    BlockBackend *target;
  41    BlockDriverState *mirror_top_bs;
  42    BlockDriverState *source;
  43    BlockDriverState *base;
  44
  45    /* The name of the graph node to replace */
  46    char *replaces;
  47    /* The BDS to replace */
  48    BlockDriverState *to_replace;
  49    /* Used to block operations on the drive-mirror-replace target */
  50    Error *replace_blocker;
  51    bool is_none_mode;
  52    BlockMirrorBackingMode backing_mode;
  53    BlockdevOnError on_source_error, on_target_error;
  54    bool synced;
  55    bool should_complete;
  56    int64_t granularity;
  57    size_t buf_size;
  58    int64_t bdev_length;
  59    unsigned long *cow_bitmap;
  60    BdrvDirtyBitmap *dirty_bitmap;
  61    BdrvDirtyBitmapIter *dbi;
  62    uint8_t *buf;
  63    QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
  64    int buf_free_count;
  65
  66    uint64_t last_pause_ns;
  67    unsigned long *in_flight_bitmap;
  68    int in_flight;
  69    int64_t bytes_in_flight;
  70    int ret;
  71    bool unmap;
  72    bool waiting_for_io;
  73    int target_cluster_size;
  74    int max_iov;
  75    bool initial_zeroing_ongoing;
  76} MirrorBlockJob;
  77
  78typedef struct MirrorOp {
  79    MirrorBlockJob *s;
  80    QEMUIOVector qiov;
  81    int64_t offset;
  82    uint64_t bytes;
  83} MirrorOp;
  84
  85static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  86                                            int error)
  87{
  88    s->synced = false;
  89    if (read) {
  90        return block_job_error_action(&s->common, s->on_source_error,
  91                                      true, error);
  92    } else {
  93        return block_job_error_action(&s->common, s->on_target_error,
  94                                      false, error);
  95    }
  96}
  97
  98static void mirror_iteration_done(MirrorOp *op, int ret)
  99{
 100    MirrorBlockJob *s = op->s;
 101    struct iovec *iov;
 102    int64_t chunk_num;
 103    int i, nb_chunks;
 104
 105    trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
 106
 107    s->in_flight--;
 108    s->bytes_in_flight -= op->bytes;
 109    iov = op->qiov.iov;
 110    for (i = 0; i < op->qiov.niov; i++) {
 111        MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
 112        QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
 113        s->buf_free_count++;
 114    }
 115
 116    chunk_num = op->offset / s->granularity;
 117    nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
 118    bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
 119    if (ret >= 0) {
 120        if (s->cow_bitmap) {
 121            bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
 122        }
 123        if (!s->initial_zeroing_ongoing) {
 124            s->common.offset += op->bytes;
 125        }
 126    }
 127    qemu_iovec_destroy(&op->qiov);
 128    g_free(op);
 129
 130    if (s->waiting_for_io) {
 131        qemu_coroutine_enter(s->common.co);
 132    }
 133}
 134
 135static void mirror_write_complete(void *opaque, int ret)
 136{
 137    MirrorOp *op = opaque;
 138    MirrorBlockJob *s = op->s;
 139
 140    aio_context_acquire(blk_get_aio_context(s->common.blk));
 141    if (ret < 0) {
 142        BlockErrorAction action;
 143
 144        bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
 145        action = mirror_error_action(s, false, -ret);
 146        if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 147            s->ret = ret;
 148        }
 149    }
 150    mirror_iteration_done(op, ret);
 151    aio_context_release(blk_get_aio_context(s->common.blk));
 152}
 153
 154static void mirror_read_complete(void *opaque, int ret)
 155{
 156    MirrorOp *op = opaque;
 157    MirrorBlockJob *s = op->s;
 158
 159    aio_context_acquire(blk_get_aio_context(s->common.blk));
 160    if (ret < 0) {
 161        BlockErrorAction action;
 162
 163        bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
 164        action = mirror_error_action(s, true, -ret);
 165        if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 166            s->ret = ret;
 167        }
 168
 169        mirror_iteration_done(op, ret);
 170    } else {
 171        blk_aio_pwritev(s->target, op->offset, &op->qiov,
 172                        0, mirror_write_complete, op);
 173    }
 174    aio_context_release(blk_get_aio_context(s->common.blk));
 175}
 176
 177/* Clip bytes relative to offset to not exceed end-of-file */
 178static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
 179                                        int64_t offset,
 180                                        int64_t bytes)
 181{
 182    return MIN(bytes, s->bdev_length - offset);
 183}
 184
 185/* Round offset and/or bytes to target cluster if COW is needed, and
 186 * return the offset of the adjusted tail against original. */
 187static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
 188                            uint64_t *bytes)
 189{
 190    bool need_cow;
 191    int ret = 0;
 192    int64_t align_offset = *offset;
 193    int64_t align_bytes = *bytes;
 194    int max_bytes = s->granularity * s->max_iov;
 195
 196    need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
 197    need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
 198                          s->cow_bitmap);
 199    if (need_cow) {
 200        bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
 201                               &align_offset, &align_bytes);
 202    }
 203
 204    if (align_bytes > max_bytes) {
 205        align_bytes = max_bytes;
 206        if (need_cow) {
 207            align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
 208        }
 209    }
 210    /* Clipping may result in align_bytes unaligned to chunk boundary, but
 211     * that doesn't matter because it's already the end of source image. */
 212    align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
 213
 214    ret = align_offset + align_bytes - (*offset + *bytes);
 215    *offset = align_offset;
 216    *bytes = align_bytes;
 217    assert(ret >= 0);
 218    return ret;
 219}
 220
 221static inline void mirror_wait_for_io(MirrorBlockJob *s)
 222{
 223    assert(!s->waiting_for_io);
 224    s->waiting_for_io = true;
 225    qemu_coroutine_yield();
 226    s->waiting_for_io = false;
 227}
 228
 229/* Submit async read while handling COW.
 230 * Returns: The number of bytes copied after and including offset,
 231 *          excluding any bytes copied prior to offset due to alignment.
 232 *          This will be @bytes if no alignment is necessary, or
 233 *          (new_end - offset) if tail is rounded up or down due to
 234 *          alignment or buffer limit.
 235 */
 236static uint64_t mirror_do_read(MirrorBlockJob *s, int64_t offset,
 237                               uint64_t bytes)
 238{
 239    BlockBackend *source = s->common.blk;
 240    int nb_chunks;
 241    uint64_t ret;
 242    MirrorOp *op;
 243    uint64_t max_bytes;
 244
 245    max_bytes = s->granularity * s->max_iov;
 246
 247    /* We can only handle as much as buf_size at a time. */
 248    bytes = MIN(s->buf_size, MIN(max_bytes, bytes));
 249    assert(bytes);
 250    assert(bytes < BDRV_REQUEST_MAX_BYTES);
 251    ret = bytes;
 252
 253    if (s->cow_bitmap) {
 254        ret += mirror_cow_align(s, &offset, &bytes);
 255    }
 256    assert(bytes <= s->buf_size);
 257    /* The offset is granularity-aligned because:
 258     * 1) Caller passes in aligned values;
 259     * 2) mirror_cow_align is used only when target cluster is larger. */
 260    assert(QEMU_IS_ALIGNED(offset, s->granularity));
 261    /* The range is sector-aligned, since bdrv_getlength() rounds up. */
 262    assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE));
 263    nb_chunks = DIV_ROUND_UP(bytes, s->granularity);
 264
 265    while (s->buf_free_count < nb_chunks) {
 266        trace_mirror_yield_in_flight(s, offset, s->in_flight);
 267        mirror_wait_for_io(s);
 268    }
 269
 270    /* Allocate a MirrorOp that is used as an AIO callback.  */
 271    op = g_new(MirrorOp, 1);
 272    op->s = s;
 273    op->offset = offset;
 274    op->bytes = bytes;
 275
 276    /* Now make a QEMUIOVector taking enough granularity-sized chunks
 277     * from s->buf_free.
 278     */
 279    qemu_iovec_init(&op->qiov, nb_chunks);
 280    while (nb_chunks-- > 0) {
 281        MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
 282        size_t remaining = bytes - op->qiov.size;
 283
 284        QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
 285        s->buf_free_count--;
 286        qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
 287    }
 288
 289    /* Copy the dirty cluster.  */
 290    s->in_flight++;
 291    s->bytes_in_flight += bytes;
 292    trace_mirror_one_iteration(s, offset, bytes);
 293
 294    blk_aio_preadv(source, offset, &op->qiov, 0, mirror_read_complete, op);
 295    return ret;
 296}
 297
 298static void mirror_do_zero_or_discard(MirrorBlockJob *s,
 299                                      int64_t offset,
 300                                      uint64_t bytes,
 301                                      bool is_discard)
 302{
 303    MirrorOp *op;
 304
 305    /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
 306     * so the freeing in mirror_iteration_done is nop. */
 307    op = g_new0(MirrorOp, 1);
 308    op->s = s;
 309    op->offset = offset;
 310    op->bytes = bytes;
 311
 312    s->in_flight++;
 313    s->bytes_in_flight += bytes;
 314    if (is_discard) {
 315        blk_aio_pdiscard(s->target, offset,
 316                         op->bytes, mirror_write_complete, op);
 317    } else {
 318        blk_aio_pwrite_zeroes(s->target, offset,
 319                              op->bytes, s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
 320                              mirror_write_complete, op);
 321    }
 322}
 323
 324static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
 325{
 326    BlockDriverState *source = s->source;
 327    int64_t offset, first_chunk;
 328    uint64_t delay_ns = 0;
 329    /* At least the first dirty chunk is mirrored in one iteration. */
 330    int nb_chunks = 1;
 331    bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
 332    int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
 333
 334    bdrv_dirty_bitmap_lock(s->dirty_bitmap);
 335    offset = bdrv_dirty_iter_next(s->dbi);
 336    if (offset < 0) {
 337        bdrv_set_dirty_iter(s->dbi, 0);
 338        offset = bdrv_dirty_iter_next(s->dbi);
 339        trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
 340        assert(offset >= 0);
 341    }
 342    bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
 343
 344    first_chunk = offset / s->granularity;
 345    while (test_bit(first_chunk, s->in_flight_bitmap)) {
 346        trace_mirror_yield_in_flight(s, offset, s->in_flight);
 347        mirror_wait_for_io(s);
 348    }
 349
 350    block_job_pause_point(&s->common);
 351
 352    /* Find the number of consective dirty chunks following the first dirty
 353     * one, and wait for in flight requests in them. */
 354    bdrv_dirty_bitmap_lock(s->dirty_bitmap);
 355    while (nb_chunks * s->granularity < s->buf_size) {
 356        int64_t next_dirty;
 357        int64_t next_offset = offset + nb_chunks * s->granularity;
 358        int64_t next_chunk = next_offset / s->granularity;
 359        if (next_offset >= s->bdev_length ||
 360            !bdrv_get_dirty_locked(source, s->dirty_bitmap, next_offset)) {
 361            break;
 362        }
 363        if (test_bit(next_chunk, s->in_flight_bitmap)) {
 364            break;
 365        }
 366
 367        next_dirty = bdrv_dirty_iter_next(s->dbi);
 368        if (next_dirty > next_offset || next_dirty < 0) {
 369            /* The bitmap iterator's cache is stale, refresh it */
 370            bdrv_set_dirty_iter(s->dbi, next_offset);
 371            next_dirty = bdrv_dirty_iter_next(s->dbi);
 372        }
 373        assert(next_dirty == next_offset);
 374        nb_chunks++;
 375    }
 376
 377    /* Clear dirty bits before querying the block status, because
 378     * calling bdrv_block_status_above could yield - if some blocks are
 379     * marked dirty in this window, we need to know.
 380     */
 381    bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
 382                                   nb_chunks * s->granularity);
 383    bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
 384
 385    bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
 386    while (nb_chunks > 0 && offset < s->bdev_length) {
 387        int ret;
 388        int64_t io_bytes;
 389        int64_t io_bytes_acct;
 390        enum MirrorMethod {
 391            MIRROR_METHOD_COPY,
 392            MIRROR_METHOD_ZERO,
 393            MIRROR_METHOD_DISCARD
 394        } mirror_method = MIRROR_METHOD_COPY;
 395
 396        assert(!(offset % s->granularity));
 397        ret = bdrv_block_status_above(source, NULL, offset,
 398                                      nb_chunks * s->granularity,
 399                                      &io_bytes, NULL, NULL);
 400        if (ret < 0) {
 401            io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
 402        } else if (ret & BDRV_BLOCK_DATA) {
 403            io_bytes = MIN(io_bytes, max_io_bytes);
 404        }
 405
 406        io_bytes -= io_bytes % s->granularity;
 407        if (io_bytes < s->granularity) {
 408            io_bytes = s->granularity;
 409        } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
 410            int64_t target_offset;
 411            int64_t target_bytes;
 412            bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
 413                                   &target_offset, &target_bytes);
 414            if (target_offset == offset &&
 415                target_bytes == io_bytes) {
 416                mirror_method = ret & BDRV_BLOCK_ZERO ?
 417                                    MIRROR_METHOD_ZERO :
 418                                    MIRROR_METHOD_DISCARD;
 419            }
 420        }
 421
 422        while (s->in_flight >= MAX_IN_FLIGHT) {
 423            trace_mirror_yield_in_flight(s, offset, s->in_flight);
 424            mirror_wait_for_io(s);
 425        }
 426
 427        if (s->ret < 0) {
 428            return 0;
 429        }
 430
 431        io_bytes = mirror_clip_bytes(s, offset, io_bytes);
 432        switch (mirror_method) {
 433        case MIRROR_METHOD_COPY:
 434            io_bytes = io_bytes_acct = mirror_do_read(s, offset, io_bytes);
 435            break;
 436        case MIRROR_METHOD_ZERO:
 437        case MIRROR_METHOD_DISCARD:
 438            mirror_do_zero_or_discard(s, offset, io_bytes,
 439                                      mirror_method == MIRROR_METHOD_DISCARD);
 440            if (write_zeroes_ok) {
 441                io_bytes_acct = 0;
 442            } else {
 443                io_bytes_acct = io_bytes;
 444            }
 445            break;
 446        default:
 447            abort();
 448        }
 449        assert(io_bytes);
 450        offset += io_bytes;
 451        nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
 452        if (s->common.speed) {
 453            delay_ns = ratelimit_calculate_delay(&s->limit, io_bytes_acct);
 454        }
 455    }
 456    return delay_ns;
 457}
 458
 459static void mirror_free_init(MirrorBlockJob *s)
 460{
 461    int granularity = s->granularity;
 462    size_t buf_size = s->buf_size;
 463    uint8_t *buf = s->buf;
 464
 465    assert(s->buf_free_count == 0);
 466    QSIMPLEQ_INIT(&s->buf_free);
 467    while (buf_size != 0) {
 468        MirrorBuffer *cur = (MirrorBuffer *)buf;
 469        QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 470        s->buf_free_count++;
 471        buf_size -= granularity;
 472        buf += granularity;
 473    }
 474}
 475
 476/* This is also used for the .pause callback. There is no matching
 477 * mirror_resume() because mirror_run() will begin iterating again
 478 * when the job is resumed.
 479 */
 480static void mirror_wait_for_all_io(MirrorBlockJob *s)
 481{
 482    while (s->in_flight > 0) {
 483        mirror_wait_for_io(s);
 484    }
 485}
 486
 487typedef struct {
 488    int ret;
 489} MirrorExitData;
 490
 491static void mirror_exit(BlockJob *job, void *opaque)
 492{
 493    MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 494    MirrorExitData *data = opaque;
 495    AioContext *replace_aio_context = NULL;
 496    BlockDriverState *src = s->source;
 497    BlockDriverState *target_bs = blk_bs(s->target);
 498    BlockDriverState *mirror_top_bs = s->mirror_top_bs;
 499    Error *local_err = NULL;
 500
 501    bdrv_release_dirty_bitmap(src, s->dirty_bitmap);
 502
 503    /* Make sure that the source BDS doesn't go away before we called
 504     * block_job_completed(). */
 505    bdrv_ref(src);
 506    bdrv_ref(mirror_top_bs);
 507    bdrv_ref(target_bs);
 508
 509    /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
 510     * inserting target_bs at s->to_replace, where we might not be able to get
 511     * these permissions.
 512     *
 513     * Note that blk_unref() alone doesn't necessarily drop permissions because
 514     * we might be running nested inside mirror_drain(), which takes an extra
 515     * reference, so use an explicit blk_set_perm() first. */
 516    blk_set_perm(s->target, 0, BLK_PERM_ALL, &error_abort);
 517    blk_unref(s->target);
 518    s->target = NULL;
 519
 520    /* We don't access the source any more. Dropping any WRITE/RESIZE is
 521     * required before it could become a backing file of target_bs. */
 522    bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
 523                            &error_abort);
 524    if (s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
 525        BlockDriverState *backing = s->is_none_mode ? src : s->base;
 526        if (backing_bs(target_bs) != backing) {
 527            bdrv_set_backing_hd(target_bs, backing, &local_err);
 528            if (local_err) {
 529                error_report_err(local_err);
 530                data->ret = -EPERM;
 531            }
 532        }
 533    }
 534
 535    if (s->to_replace) {
 536        replace_aio_context = bdrv_get_aio_context(s->to_replace);
 537        aio_context_acquire(replace_aio_context);
 538    }
 539
 540    if (s->should_complete && data->ret == 0) {
 541        BlockDriverState *to_replace = src;
 542        if (s->to_replace) {
 543            to_replace = s->to_replace;
 544        }
 545
 546        if (bdrv_get_flags(target_bs) != bdrv_get_flags(to_replace)) {
 547            bdrv_reopen(target_bs, bdrv_get_flags(to_replace), NULL);
 548        }
 549
 550        /* The mirror job has no requests in flight any more, but we need to
 551         * drain potential other users of the BDS before changing the graph. */
 552        bdrv_drained_begin(target_bs);
 553        bdrv_replace_node(to_replace, target_bs, &local_err);
 554        bdrv_drained_end(target_bs);
 555        if (local_err) {
 556            error_report_err(local_err);
 557            data->ret = -EPERM;
 558        }
 559    }
 560    if (s->to_replace) {
 561        bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
 562        error_free(s->replace_blocker);
 563        bdrv_unref(s->to_replace);
 564    }
 565    if (replace_aio_context) {
 566        aio_context_release(replace_aio_context);
 567    }
 568    g_free(s->replaces);
 569    bdrv_unref(target_bs);
 570
 571    /* Remove the mirror filter driver from the graph. Before this, get rid of
 572     * the blockers on the intermediate nodes so that the resulting state is
 573     * valid. Also give up permissions on mirror_top_bs->backing, which might
 574     * block the removal. */
 575    block_job_remove_all_bdrv(job);
 576    bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
 577                            &error_abort);
 578    bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
 579
 580    /* We just changed the BDS the job BB refers to (with either or both of the
 581     * bdrv_replace_node() calls), so switch the BB back so the cleanup does
 582     * the right thing. We don't need any permissions any more now. */
 583    blk_remove_bs(job->blk);
 584    blk_set_perm(job->blk, 0, BLK_PERM_ALL, &error_abort);
 585    blk_insert_bs(job->blk, mirror_top_bs, &error_abort);
 586
 587    block_job_completed(&s->common, data->ret);
 588
 589    g_free(data);
 590    bdrv_drained_end(src);
 591    bdrv_unref(mirror_top_bs);
 592    bdrv_unref(src);
 593}
 594
 595static void mirror_throttle(MirrorBlockJob *s)
 596{
 597    int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 598
 599    if (now - s->last_pause_ns > SLICE_TIME) {
 600        s->last_pause_ns = now;
 601        block_job_sleep_ns(&s->common, 0);
 602    } else {
 603        block_job_pause_point(&s->common);
 604    }
 605}
 606
 607static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
 608{
 609    int64_t offset;
 610    BlockDriverState *base = s->base;
 611    BlockDriverState *bs = s->source;
 612    BlockDriverState *target_bs = blk_bs(s->target);
 613    int ret;
 614    int64_t count;
 615
 616    if (base == NULL && !bdrv_has_zero_init(target_bs)) {
 617        if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
 618            bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
 619            return 0;
 620        }
 621
 622        s->initial_zeroing_ongoing = true;
 623        for (offset = 0; offset < s->bdev_length; ) {
 624            int bytes = MIN(s->bdev_length - offset,
 625                            QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
 626
 627            mirror_throttle(s);
 628
 629            if (block_job_is_cancelled(&s->common)) {
 630                s->initial_zeroing_ongoing = false;
 631                return 0;
 632            }
 633
 634            if (s->in_flight >= MAX_IN_FLIGHT) {
 635                trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
 636                                   s->in_flight);
 637                mirror_wait_for_io(s);
 638                continue;
 639            }
 640
 641            mirror_do_zero_or_discard(s, offset, bytes, false);
 642            offset += bytes;
 643        }
 644
 645        mirror_wait_for_all_io(s);
 646        s->initial_zeroing_ongoing = false;
 647    }
 648
 649    /* First part, loop on the sectors and initialize the dirty bitmap.  */
 650    for (offset = 0; offset < s->bdev_length; ) {
 651        /* Just to make sure we are not exceeding int limit. */
 652        int bytes = MIN(s->bdev_length - offset,
 653                        QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
 654
 655        mirror_throttle(s);
 656
 657        if (block_job_is_cancelled(&s->common)) {
 658            return 0;
 659        }
 660
 661        ret = bdrv_is_allocated_above(bs, base, offset, bytes, &count);
 662        if (ret < 0) {
 663            return ret;
 664        }
 665
 666        assert(count);
 667        if (ret == 1) {
 668            bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
 669        }
 670        offset += count;
 671    }
 672    return 0;
 673}
 674
 675/* Called when going out of the streaming phase to flush the bulk of the
 676 * data to the medium, or just before completing.
 677 */
 678static int mirror_flush(MirrorBlockJob *s)
 679{
 680    int ret = blk_flush(s->target);
 681    if (ret < 0) {
 682        if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
 683            s->ret = ret;
 684        }
 685    }
 686    return ret;
 687}
 688
 689static void coroutine_fn mirror_run(void *opaque)
 690{
 691    MirrorBlockJob *s = opaque;
 692    MirrorExitData *data;
 693    BlockDriverState *bs = s->source;
 694    BlockDriverState *target_bs = blk_bs(s->target);
 695    bool need_drain = true;
 696    int64_t length;
 697    BlockDriverInfo bdi;
 698    char backing_filename[2]; /* we only need 2 characters because we are only
 699                                 checking for a NULL string */
 700    int ret = 0;
 701
 702    if (block_job_is_cancelled(&s->common)) {
 703        goto immediate_exit;
 704    }
 705
 706    s->bdev_length = bdrv_getlength(bs);
 707    if (s->bdev_length < 0) {
 708        ret = s->bdev_length;
 709        goto immediate_exit;
 710    }
 711
 712    /* Active commit must resize the base image if its size differs from the
 713     * active layer. */
 714    if (s->base == blk_bs(s->target)) {
 715        int64_t base_length;
 716
 717        base_length = blk_getlength(s->target);
 718        if (base_length < 0) {
 719            ret = base_length;
 720            goto immediate_exit;
 721        }
 722
 723        if (s->bdev_length > base_length) {
 724            ret = blk_truncate(s->target, s->bdev_length, PREALLOC_MODE_OFF,
 725                               NULL);
 726            if (ret < 0) {
 727                goto immediate_exit;
 728            }
 729        }
 730    }
 731
 732    if (s->bdev_length == 0) {
 733        /* Report BLOCK_JOB_READY and wait for complete. */
 734        block_job_event_ready(&s->common);
 735        s->synced = true;
 736        while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
 737            block_job_yield(&s->common);
 738        }
 739        s->common.cancelled = false;
 740        goto immediate_exit;
 741    }
 742
 743    length = DIV_ROUND_UP(s->bdev_length, s->granularity);
 744    s->in_flight_bitmap = bitmap_new(length);
 745
 746    /* If we have no backing file yet in the destination, we cannot let
 747     * the destination do COW.  Instead, we copy sectors around the
 748     * dirty data if needed.  We need a bitmap to do that.
 749     */
 750    bdrv_get_backing_filename(target_bs, backing_filename,
 751                              sizeof(backing_filename));
 752    if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
 753        s->target_cluster_size = bdi.cluster_size;
 754    } else {
 755        s->target_cluster_size = BDRV_SECTOR_SIZE;
 756    }
 757    if (backing_filename[0] && !target_bs->backing &&
 758        s->granularity < s->target_cluster_size) {
 759        s->buf_size = MAX(s->buf_size, s->target_cluster_size);
 760        s->cow_bitmap = bitmap_new(length);
 761    }
 762    s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
 763
 764    s->buf = qemu_try_blockalign(bs, s->buf_size);
 765    if (s->buf == NULL) {
 766        ret = -ENOMEM;
 767        goto immediate_exit;
 768    }
 769
 770    mirror_free_init(s);
 771
 772    s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 773    if (!s->is_none_mode) {
 774        ret = mirror_dirty_init(s);
 775        if (ret < 0 || block_job_is_cancelled(&s->common)) {
 776            goto immediate_exit;
 777        }
 778    }
 779
 780    assert(!s->dbi);
 781    s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
 782    for (;;) {
 783        uint64_t delay_ns = 0;
 784        int64_t cnt, delta;
 785        bool should_complete;
 786
 787        if (s->ret < 0) {
 788            ret = s->ret;
 789            goto immediate_exit;
 790        }
 791
 792        block_job_pause_point(&s->common);
 793
 794        cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 795        /* s->common.offset contains the number of bytes already processed so
 796         * far, cnt is the number of dirty bytes remaining and
 797         * s->bytes_in_flight is the number of bytes currently being
 798         * processed; together those are the current total operation length */
 799        s->common.len = s->common.offset + s->bytes_in_flight + cnt;
 800
 801        /* Note that even when no rate limit is applied we need to yield
 802         * periodically with no pending I/O so that bdrv_drain_all() returns.
 803         * We do so every SLICE_TIME nanoseconds, or when there is an error,
 804         * or when the source is clean, whichever comes first.
 805         */
 806        delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
 807        if (delta < SLICE_TIME &&
 808            s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 809            if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 810                (cnt == 0 && s->in_flight > 0)) {
 811                trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
 812                mirror_wait_for_io(s);
 813                continue;
 814            } else if (cnt != 0) {
 815                delay_ns = mirror_iteration(s);
 816            }
 817        }
 818
 819        should_complete = false;
 820        if (s->in_flight == 0 && cnt == 0) {
 821            trace_mirror_before_flush(s);
 822            if (!s->synced) {
 823                if (mirror_flush(s) < 0) {
 824                    /* Go check s->ret.  */
 825                    continue;
 826                }
 827                /* We're out of the streaming phase.  From now on, if the job
 828                 * is cancelled we will actually complete all pending I/O and
 829                 * report completion.  This way, block-job-cancel will leave
 830                 * the target in a consistent state.
 831                 */
 832                block_job_event_ready(&s->common);
 833                s->synced = true;
 834            }
 835
 836            should_complete = s->should_complete ||
 837                block_job_is_cancelled(&s->common);
 838            cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 839        }
 840
 841        if (cnt == 0 && should_complete) {
 842            /* The dirty bitmap is not updated while operations are pending.
 843             * If we're about to exit, wait for pending operations before
 844             * calling bdrv_get_dirty_count(bs), or we may exit while the
 845             * source has dirty data to copy!
 846             *
 847             * Note that I/O can be submitted by the guest while
 848             * mirror_populate runs, so pause it now.  Before deciding
 849             * whether to switch to target check one last time if I/O has
 850             * come in the meanwhile, and if not flush the data to disk.
 851             */
 852            trace_mirror_before_drain(s, cnt);
 853
 854            bdrv_drained_begin(bs);
 855            cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 856            if (cnt > 0 || mirror_flush(s) < 0) {
 857                bdrv_drained_end(bs);
 858                continue;
 859            }
 860
 861            /* The two disks are in sync.  Exit and report successful
 862             * completion.
 863             */
 864            assert(QLIST_EMPTY(&bs->tracked_requests));
 865            s->common.cancelled = false;
 866            need_drain = false;
 867            break;
 868        }
 869
 870        ret = 0;
 871        trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
 872        if (!s->synced) {
 873            block_job_sleep_ns(&s->common, delay_ns);
 874            if (block_job_is_cancelled(&s->common)) {
 875                break;
 876            }
 877        } else if (!should_complete) {
 878            delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 879            block_job_sleep_ns(&s->common, delay_ns);
 880        }
 881        s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 882    }
 883
 884immediate_exit:
 885    if (s->in_flight > 0) {
 886        /* We get here only if something went wrong.  Either the job failed,
 887         * or it was cancelled prematurely so that we do not guarantee that
 888         * the target is a copy of the source.
 889         */
 890        assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 891        assert(need_drain);
 892        mirror_wait_for_all_io(s);
 893    }
 894
 895    assert(s->in_flight == 0);
 896    qemu_vfree(s->buf);
 897    g_free(s->cow_bitmap);
 898    g_free(s->in_flight_bitmap);
 899    bdrv_dirty_iter_free(s->dbi);
 900
 901    data = g_malloc(sizeof(*data));
 902    data->ret = ret;
 903
 904    if (need_drain) {
 905        bdrv_drained_begin(bs);
 906    }
 907    block_job_defer_to_main_loop(&s->common, mirror_exit, data);
 908}
 909
 910static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 911{
 912    MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 913
 914    if (speed < 0) {
 915        error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 916        return;
 917    }
 918    ratelimit_set_speed(&s->limit, speed, SLICE_TIME);
 919}
 920
 921static void mirror_complete(BlockJob *job, Error **errp)
 922{
 923    MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 924    BlockDriverState *target;
 925
 926    target = blk_bs(s->target);
 927
 928    if (!s->synced) {
 929        error_setg(errp, "The active block job '%s' cannot be completed",
 930                   job->id);
 931        return;
 932    }
 933
 934    if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
 935        int ret;
 936
 937        assert(!target->backing);
 938        ret = bdrv_open_backing_file(target, NULL, "backing", errp);
 939        if (ret < 0) {
 940            return;
 941        }
 942    }
 943
 944    /* block all operations on to_replace bs */
 945    if (s->replaces) {
 946        AioContext *replace_aio_context;
 947
 948        s->to_replace = bdrv_find_node(s->replaces);
 949        if (!s->to_replace) {
 950            error_setg(errp, "Node name '%s' not found", s->replaces);
 951            return;
 952        }
 953
 954        replace_aio_context = bdrv_get_aio_context(s->to_replace);
 955        aio_context_acquire(replace_aio_context);
 956
 957        /* TODO Translate this into permission system. Current definition of
 958         * GRAPH_MOD would require to request it for the parents; they might
 959         * not even be BlockDriverStates, however, so a BdrvChild can't address
 960         * them. May need redefinition of GRAPH_MOD. */
 961        error_setg(&s->replace_blocker,
 962                   "block device is in use by block-job-complete");
 963        bdrv_op_block_all(s->to_replace, s->replace_blocker);
 964        bdrv_ref(s->to_replace);
 965
 966        aio_context_release(replace_aio_context);
 967    }
 968
 969    s->should_complete = true;
 970    block_job_enter(&s->common);
 971}
 972
 973static void mirror_pause(BlockJob *job)
 974{
 975    MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 976
 977    mirror_wait_for_all_io(s);
 978}
 979
 980static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context)
 981{
 982    MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 983
 984    blk_set_aio_context(s->target, new_context);
 985}
 986
 987static void mirror_drain(BlockJob *job)
 988{
 989    MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 990
 991    /* Need to keep a reference in case blk_drain triggers execution
 992     * of mirror_complete...
 993     */
 994    if (s->target) {
 995        BlockBackend *target = s->target;
 996        blk_ref(target);
 997        blk_drain(target);
 998        blk_unref(target);
 999    }
1000}
1001
1002static const BlockJobDriver mirror_job_driver = {
1003    .instance_size          = sizeof(MirrorBlockJob),
1004    .job_type               = BLOCK_JOB_TYPE_MIRROR,
1005    .set_speed              = mirror_set_speed,
1006    .start                  = mirror_run,
1007    .complete               = mirror_complete,
1008    .pause                  = mirror_pause,
1009    .attached_aio_context   = mirror_attached_aio_context,
1010    .drain                  = mirror_drain,
1011};
1012
1013static const BlockJobDriver commit_active_job_driver = {
1014    .instance_size          = sizeof(MirrorBlockJob),
1015    .job_type               = BLOCK_JOB_TYPE_COMMIT,
1016    .set_speed              = mirror_set_speed,
1017    .start                  = mirror_run,
1018    .complete               = mirror_complete,
1019    .pause                  = mirror_pause,
1020    .attached_aio_context   = mirror_attached_aio_context,
1021    .drain                  = mirror_drain,
1022};
1023
1024static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1025    uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1026{
1027    return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1028}
1029
1030static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1031    uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1032{
1033    return bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1034}
1035
1036static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1037{
1038    if (bs->backing == NULL) {
1039        /* we can be here after failed bdrv_append in mirror_start_job */
1040        return 0;
1041    }
1042    return bdrv_co_flush(bs->backing->bs);
1043}
1044
1045static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1046    int64_t offset, int bytes, BdrvRequestFlags flags)
1047{
1048    return bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1049}
1050
1051static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1052    int64_t offset, int bytes)
1053{
1054    return bdrv_co_pdiscard(bs->backing->bs, offset, bytes);
1055}
1056
1057static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs, QDict *opts)
1058{
1059    if (bs->backing == NULL) {
1060        /* we can be here after failed bdrv_attach_child in
1061         * bdrv_set_backing_hd */
1062        return;
1063    }
1064    bdrv_refresh_filename(bs->backing->bs);
1065    pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1066            bs->backing->bs->filename);
1067}
1068
1069static void bdrv_mirror_top_close(BlockDriverState *bs)
1070{
1071}
1072
1073static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1074                                       const BdrvChildRole *role,
1075                                       BlockReopenQueue *reopen_queue,
1076                                       uint64_t perm, uint64_t shared,
1077                                       uint64_t *nperm, uint64_t *nshared)
1078{
1079    /* Must be able to forward guest writes to the real image */
1080    *nperm = 0;
1081    if (perm & BLK_PERM_WRITE) {
1082        *nperm |= BLK_PERM_WRITE;
1083    }
1084
1085    *nshared = BLK_PERM_ALL;
1086}
1087
1088/* Dummy node that provides consistent read to its users without requiring it
1089 * from its backing file and that allows writes on the backing file chain. */
1090static BlockDriver bdrv_mirror_top = {
1091    .format_name                = "mirror_top",
1092    .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1093    .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1094    .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1095    .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1096    .bdrv_co_flush              = bdrv_mirror_top_flush,
1097    .bdrv_co_get_block_status   = bdrv_co_get_block_status_from_backing,
1098    .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1099    .bdrv_close                 = bdrv_mirror_top_close,
1100    .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1101};
1102
1103static void mirror_start_job(const char *job_id, BlockDriverState *bs,
1104                             int creation_flags, BlockDriverState *target,
1105                             const char *replaces, int64_t speed,
1106                             uint32_t granularity, int64_t buf_size,
1107                             BlockMirrorBackingMode backing_mode,
1108                             BlockdevOnError on_source_error,
1109                             BlockdevOnError on_target_error,
1110                             bool unmap,
1111                             BlockCompletionFunc *cb,
1112                             void *opaque,
1113                             const BlockJobDriver *driver,
1114                             bool is_none_mode, BlockDriverState *base,
1115                             bool auto_complete, const char *filter_node_name,
1116                             bool is_mirror,
1117                             Error **errp)
1118{
1119    MirrorBlockJob *s;
1120    BlockDriverState *mirror_top_bs;
1121    bool target_graph_mod;
1122    bool target_is_backing;
1123    Error *local_err = NULL;
1124    int ret;
1125
1126    if (granularity == 0) {
1127        granularity = bdrv_get_default_bitmap_granularity(target);
1128    }
1129
1130    assert(is_power_of_2(granularity));
1131
1132    if (buf_size < 0) {
1133        error_setg(errp, "Invalid parameter 'buf-size'");
1134        return;
1135    }
1136
1137    if (buf_size == 0) {
1138        buf_size = DEFAULT_MIRROR_BUF_SIZE;
1139    }
1140
1141    /* In the case of active commit, add dummy driver to provide consistent
1142     * reads on the top, while disabling it in the intermediate nodes, and make
1143     * the backing chain writable. */
1144    mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1145                                         BDRV_O_RDWR, errp);
1146    if (mirror_top_bs == NULL) {
1147        return;
1148    }
1149    if (!filter_node_name) {
1150        mirror_top_bs->implicit = true;
1151    }
1152    mirror_top_bs->total_sectors = bs->total_sectors;
1153    bdrv_set_aio_context(mirror_top_bs, bdrv_get_aio_context(bs));
1154
1155    /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1156     * it alive until block_job_create() succeeds even if bs has no parent. */
1157    bdrv_ref(mirror_top_bs);
1158    bdrv_drained_begin(bs);
1159    bdrv_append(mirror_top_bs, bs, &local_err);
1160    bdrv_drained_end(bs);
1161
1162    if (local_err) {
1163        bdrv_unref(mirror_top_bs);
1164        error_propagate(errp, local_err);
1165        return;
1166    }
1167
1168    /* Make sure that the source is not resized while the job is running */
1169    s = block_job_create(job_id, driver, mirror_top_bs,
1170                         BLK_PERM_CONSISTENT_READ,
1171                         BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1172                         BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1173                         creation_flags, cb, opaque, errp);
1174    if (!s) {
1175        goto fail;
1176    }
1177    /* The block job now has a reference to this node */
1178    bdrv_unref(mirror_top_bs);
1179
1180    s->source = bs;
1181    s->mirror_top_bs = mirror_top_bs;
1182
1183    /* No resize for the target either; while the mirror is still running, a
1184     * consistent read isn't necessarily possible. We could possibly allow
1185     * writes and graph modifications, though it would likely defeat the
1186     * purpose of a mirror, so leave them blocked for now.
1187     *
1188     * In the case of active commit, things look a bit different, though,
1189     * because the target is an already populated backing file in active use.
1190     * We can allow anything except resize there.*/
1191    target_is_backing = bdrv_chain_contains(bs, target);
1192    target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1193    s->target = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE |
1194                        (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1195                        BLK_PERM_WRITE_UNCHANGED |
1196                        (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1197                                             BLK_PERM_WRITE |
1198                                             BLK_PERM_GRAPH_MOD : 0));
1199    ret = blk_insert_bs(s->target, target, errp);
1200    if (ret < 0) {
1201        goto fail;
1202    }
1203    if (is_mirror) {
1204        /* XXX: Mirror target could be a NBD server of target QEMU in the case
1205         * of non-shared block migration. To allow migration completion, we
1206         * have to allow "inactivate" of the target BB.  When that happens, we
1207         * know the job is drained, and the vcpus are stopped, so no write
1208         * operation will be performed. Block layer already has assertions to
1209         * ensure that. */
1210        blk_set_force_allow_inactivate(s->target);
1211    }
1212
1213    s->replaces = g_strdup(replaces);
1214    s->on_source_error = on_source_error;
1215    s->on_target_error = on_target_error;
1216    s->is_none_mode = is_none_mode;
1217    s->backing_mode = backing_mode;
1218    s->base = base;
1219    s->granularity = granularity;
1220    s->buf_size = ROUND_UP(buf_size, granularity);
1221    s->unmap = unmap;
1222    if (auto_complete) {
1223        s->should_complete = true;
1224    }
1225
1226    s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1227    if (!s->dirty_bitmap) {
1228        goto fail;
1229    }
1230
1231    /* Required permissions are already taken with blk_new() */
1232    block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1233                       &error_abort);
1234
1235    /* In commit_active_start() all intermediate nodes disappear, so
1236     * any jobs in them must be blocked */
1237    if (target_is_backing) {
1238        BlockDriverState *iter;
1239        for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1240            /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1241             * ourselves at s->base (if writes are blocked for a node, they are
1242             * also blocked for its backing file). The other options would be a
1243             * second filter driver above s->base (== target). */
1244            ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1245                                     BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1246                                     errp);
1247            if (ret < 0) {
1248                goto fail;
1249            }
1250        }
1251    }
1252
1253    trace_mirror_start(bs, s, opaque);
1254    block_job_start(&s->common);
1255    return;
1256
1257fail:
1258    if (s) {
1259        /* Make sure this BDS does not go away until we have completed the graph
1260         * changes below */
1261        bdrv_ref(mirror_top_bs);
1262
1263        g_free(s->replaces);
1264        blk_unref(s->target);
1265        block_job_early_fail(&s->common);
1266    }
1267
1268    bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
1269                            &error_abort);
1270    bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1271
1272    bdrv_unref(mirror_top_bs);
1273}
1274
1275void mirror_start(const char *job_id, BlockDriverState *bs,
1276                  BlockDriverState *target, const char *replaces,
1277                  int64_t speed, uint32_t granularity, int64_t buf_size,
1278                  MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1279                  BlockdevOnError on_source_error,
1280                  BlockdevOnError on_target_error,
1281                  bool unmap, const char *filter_node_name, Error **errp)
1282{
1283    bool is_none_mode;
1284    BlockDriverState *base;
1285
1286    if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
1287        error_setg(errp, "Sync mode 'incremental' not supported");
1288        return;
1289    }
1290    is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1291    base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1292    mirror_start_job(job_id, bs, BLOCK_JOB_DEFAULT, target, replaces,
1293                     speed, granularity, buf_size, backing_mode,
1294                     on_source_error, on_target_error, unmap, NULL, NULL,
1295                     &mirror_job_driver, is_none_mode, base, false,
1296                     filter_node_name, true, errp);
1297}
1298
1299void commit_active_start(const char *job_id, BlockDriverState *bs,
1300                         BlockDriverState *base, int creation_flags,
1301                         int64_t speed, BlockdevOnError on_error,
1302                         const char *filter_node_name,
1303                         BlockCompletionFunc *cb, void *opaque,
1304                         bool auto_complete, Error **errp)
1305{
1306    int orig_base_flags;
1307    Error *local_err = NULL;
1308
1309    orig_base_flags = bdrv_get_flags(base);
1310
1311    if (bdrv_reopen(base, bs->open_flags, errp)) {
1312        return;
1313    }
1314
1315    mirror_start_job(job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1316                     MIRROR_LEAVE_BACKING_CHAIN,
1317                     on_error, on_error, true, cb, opaque,
1318                     &commit_active_job_driver, false, base, auto_complete,
1319                     filter_node_name, false, &local_err);
1320    if (local_err) {
1321        error_propagate(errp, local_err);
1322        goto error_restore_flags;
1323    }
1324
1325    return;
1326
1327error_restore_flags:
1328    /* ignore error and errp for bdrv_reopen, because we want to propagate
1329     * the original error */
1330    bdrv_reopen(base, orig_base_flags, NULL);
1331    return;
1332}
1333