block/mirror.c

   1 /*
   2  * Image mirroring
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Paolo Bonzini  <[email protected]>
   8  *
   9  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
  10  * See the COPYING.LIB file in the top-level directory.
  11  *
  12  */
  13
  14 #include "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_SECTORS ((1 << 20) >> BDRV_SECTOR_BITS) /* 1 Mb */
  28 #define DEFAULT_MIRROR_BUF_SIZE \
  29     (MAX_IN_FLIGHT * MAX_IO_SECTORS * BDRV_SECTOR_SIZE)
  30
  31 /* The mirroring buffer is a list of granularity-sized chunks.
  32  * Free chunks are organized in a list.
  33  */
  34 typedef struct MirrorBuffer {
  35     QSIMPLEQ_ENTRY(MirrorBuffer) next;
  36 } MirrorBuffer;
  37
  38 typedef struct MirrorBlockJob {
  39     BlockJob common;
  40     RateLimit limit;
  41     BlockBackend *target;
  42     BlockDriverState *mirror_top_bs;
  43     BlockDriverState *source;
  44     BlockDriverState *base;
  45
  46     /* The name of the graph node to replace */
  47     char *replaces;
  48     /* The BDS to replace */
  49     BlockDriverState *to_replace;
  50     /* Used to block operations on the drive-mirror-replace target */
  51     Error *replace_blocker;
  52     bool is_none_mode;
  53     BlockMirrorBackingMode backing_mode;
  54     BlockdevOnError on_source_error, on_target_error;
  55     bool synced;
  56     bool should_complete;
  57     int64_t granularity;
  58     size_t buf_size;
  59     int64_t bdev_length;
  60     unsigned long *cow_bitmap;
  61     BdrvDirtyBitmap *dirty_bitmap;
  62     BdrvDirtyBitmapIter *dbi;
  63     uint8_t *buf;
  64     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
  65     int buf_free_count;
  66
  67     uint64_t last_pause_ns;
  68     unsigned long *in_flight_bitmap;
  69     int in_flight;
  70     int64_t sectors_in_flight;
  71     int ret;
  72     bool unmap;
  73     bool waiting_for_io;
  74     int target_cluster_sectors;
  75     int max_iov;
  76     bool initial_zeroing_ongoing;
  77 } MirrorBlockJob;
  78
  79 typedef struct MirrorOp {
  80     MirrorBlockJob *s;
  81     QEMUIOVector qiov;
  82     int64_t sector_num;
  83     int nb_sectors;
  84 } MirrorOp;
  85
  86 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  87                                             int error)
  88 {
  89     s->synced = false;
  90     if (read) {
  91         return block_job_error_action(&s->common, s->on_source_error,
  92                                       true, error);
  93     } else {
  94         return block_job_error_action(&s->common, s->on_target_error,
  95                                       false, error);
  96     }
  97 }
  98
  99 static void mirror_iteration_done(MirrorOp *op, int ret)
 100 {
 101     MirrorBlockJob *s = op->s;
 102     struct iovec *iov;
 103     int64_t chunk_num;
 104     int i, nb_chunks, sectors_per_chunk;
 105
 106     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
 107
 108     s->in_flight--;
 109     s->sectors_in_flight -= op->nb_sectors;
 110     iov = op->qiov.iov;
 111     for (i = 0; i < op->qiov.niov; i++) {
 112         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
 113         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
 114         s->buf_free_count++;
 115     }
 116
 117     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 118     chunk_num = op->sector_num / sectors_per_chunk;
 119     nb_chunks = DIV_ROUND_UP(op->nb_sectors, sectors_per_chunk);
 120     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
 121     if (ret >= 0) {
 122         if (s->cow_bitmap) {
 123             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
 124         }
 125         if (!s->initial_zeroing_ongoing) {
 126             s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
 127         }
 128     }
 129     qemu_iovec_destroy(&op->qiov);
 130     g_free(op);
 131
 132     if (s->waiting_for_io) {
 133         qemu_coroutine_enter(s->common.co);
 134     }
 135 }
 136
 137 static void mirror_write_complete(void *opaque, int ret)
 138 {
 139     MirrorOp *op = opaque;
 140     MirrorBlockJob *s = op->s;
 141
 142     aio_context_acquire(blk_get_aio_context(s->common.blk));
 143     if (ret < 0) {
 144         BlockErrorAction action;
 145
 146         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
 147         action = mirror_error_action(s, false, -ret);
 148         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 149             s->ret = ret;
 150         }
 151     }
 152     mirror_iteration_done(op, ret);
 153     aio_context_release(blk_get_aio_context(s->common.blk));
 154 }
 155
 156 static void mirror_read_complete(void *opaque, int ret)
 157 {
 158     MirrorOp *op = opaque;
 159     MirrorBlockJob *s = op->s;
 160
 161     aio_context_acquire(blk_get_aio_context(s->common.blk));
 162     if (ret < 0) {
 163         BlockErrorAction action;
 164
 165         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
 166         action = mirror_error_action(s, true, -ret);
 167         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 168             s->ret = ret;
 169         }
 170
 171         mirror_iteration_done(op, ret);
 172     } else {
 173         blk_aio_pwritev(s->target, op->sector_num * BDRV_SECTOR_SIZE, &op->qiov,
 174                         0, mirror_write_complete, op);
 175     }
 176     aio_context_release(blk_get_aio_context(s->common.blk));
 177 }
 178
 179 static inline void mirror_clip_sectors(MirrorBlockJob *s,
 180                                        int64_t sector_num,
 181                                        int *nb_sectors)
 182 {
 183     *nb_sectors = MIN(*nb_sectors,
 184                       s->bdev_length / BDRV_SECTOR_SIZE - sector_num);
 185 }
 186
 187 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
 188  * return the offset of the adjusted tail sector against original. */
 189 static int mirror_cow_align(MirrorBlockJob *s,
 190                             int64_t *sector_num,
 191                             int *nb_sectors)
 192 {
 193     bool need_cow;
 194     int ret = 0;
 195     int chunk_sectors = s->granularity >> BDRV_SECTOR_BITS;
 196     int64_t align_sector_num = *sector_num;
 197     int align_nb_sectors = *nb_sectors;
 198     int max_sectors = chunk_sectors * s->max_iov;
 199
 200     need_cow = !test_bit(*sector_num / chunk_sectors, s->cow_bitmap);
 201     need_cow |= !test_bit((*sector_num + *nb_sectors - 1) / chunk_sectors,
 202                           s->cow_bitmap);
 203     if (need_cow) {
 204         bdrv_round_sectors_to_clusters(blk_bs(s->target), *sector_num,
 205                                        *nb_sectors, &align_sector_num,
 206                                        &align_nb_sectors);
 207     }
 208
 209     if (align_nb_sectors > max_sectors) {
 210         align_nb_sectors = max_sectors;
 211         if (need_cow) {
 212             align_nb_sectors = QEMU_ALIGN_DOWN(align_nb_sectors,
 213                                                s->target_cluster_sectors);
 214         }
 215     }
 216     /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but
 217      * that doesn't matter because it's already the end of source image. */
 218     mirror_clip_sectors(s, align_sector_num, &align_nb_sectors);
 219
 220     ret = align_sector_num + align_nb_sectors - (*sector_num + *nb_sectors);
 221     *sector_num = align_sector_num;
 222     *nb_sectors = align_nb_sectors;
 223     assert(ret >= 0);
 224     return ret;
 225 }
 226
 227 static inline void mirror_wait_for_io(MirrorBlockJob *s)
 228 {
 229     assert(!s->waiting_for_io);
 230     s->waiting_for_io = true;
 231     qemu_coroutine_yield();
 232     s->waiting_for_io = false;
 233 }
 234
 235 /* Submit async read while handling COW.
 236  * Returns: The number of sectors copied after and including sector_num,
 237  *          excluding any sectors copied prior to sector_num due to alignment.
 238  *          This will be nb_sectors if no alignment is necessary, or
 239  *          (new_end - sector_num) if tail is rounded up or down due to
 240  *          alignment or buffer limit.
 241  */
 242 static int mirror_do_read(MirrorBlockJob *s, int64_t sector_num,
 243                           int nb_sectors)
 244 {
 245     BlockBackend *source = s->common.blk;
 246     int sectors_per_chunk, nb_chunks;
 247     int ret;
 248     MirrorOp *op;
 249     int max_sectors;
 250
 251     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 252     max_sectors = sectors_per_chunk * s->max_iov;
 253
 254     /* We can only handle as much as buf_size at a time. */
 255     nb_sectors = MIN(s->buf_size >> BDRV_SECTOR_BITS, nb_sectors);
 256     nb_sectors = MIN(max_sectors, nb_sectors);
 257     assert(nb_sectors);
 258     ret = nb_sectors;
 259
 260     if (s->cow_bitmap) {
 261         ret += mirror_cow_align(s, &sector_num, &nb_sectors);
 262     }
 263     assert(nb_sectors << BDRV_SECTOR_BITS <= s->buf_size);
 264     /* The sector range must meet granularity because:
 265      * 1) Caller passes in aligned values;
 266      * 2) mirror_cow_align is used only when target cluster is larger. */
 267     assert(!(sector_num % sectors_per_chunk));
 268     nb_chunks = DIV_ROUND_UP(nb_sectors, sectors_per_chunk);
 269
 270     while (s->buf_free_count < nb_chunks) {
 271         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 272         mirror_wait_for_io(s);
 273     }
 274
 275     /* Allocate a MirrorOp that is used as an AIO callback.  */
 276     op = g_new(MirrorOp, 1);
 277     op->s = s;
 278     op->sector_num = sector_num;
 279     op->nb_sectors = nb_sectors;
 280
 281     /* Now make a QEMUIOVector taking enough granularity-sized chunks
 282      * from s->buf_free.
 283      */
 284     qemu_iovec_init(&op->qiov, nb_chunks);
 285     while (nb_chunks-- > 0) {
 286         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
 287         size_t remaining = nb_sectors * BDRV_SECTOR_SIZE - op->qiov.size;
 288
 289         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
 290         s->buf_free_count--;
 291         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
 292     }
 293
 294     /* Copy the dirty cluster.  */
 295     s->in_flight++;
 296     s->sectors_in_flight += nb_sectors;
 297     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 298
 299     blk_aio_preadv(source, sector_num * BDRV_SECTOR_SIZE, &op->qiov, 0,
 300                    mirror_read_complete, op);
 301     return ret;
 302 }
 303
 304 static void mirror_do_zero_or_discard(MirrorBlockJob *s,
 305                                       int64_t sector_num,
 306                                       int nb_sectors,
 307                                       bool is_discard)
 308 {
 309     MirrorOp *op;
 310
 311     /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
 312      * so the freeing in mirror_iteration_done is nop. */
 313     op = g_new0(MirrorOp, 1);
 314     op->s = s;
 315     op->sector_num = sector_num;
 316     op->nb_sectors = nb_sectors;
 317
 318     s->in_flight++;
 319     s->sectors_in_flight += nb_sectors;
 320     if (is_discard) {
 321         blk_aio_pdiscard(s->target, sector_num << BDRV_SECTOR_BITS,
 322                          op->nb_sectors << BDRV_SECTOR_BITS,
 323                          mirror_write_complete, op);
 324     } else {
 325         blk_aio_pwrite_zeroes(s->target, sector_num * BDRV_SECTOR_SIZE,
 326                               op->nb_sectors * BDRV_SECTOR_SIZE,
 327                               s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
 328                               mirror_write_complete, op);
 329     }
 330 }
 331
 332 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
 333 {
 334     BlockDriverState *source = s->source;
 335     int64_t sector_num, first_chunk;
 336     uint64_t delay_ns = 0;
 337     /* At least the first dirty chunk is mirrored in one iteration. */
 338     int nb_chunks = 1;
 339     int64_t end = s->bdev_length / BDRV_SECTOR_SIZE;
 340     int sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 341     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
 342     int max_io_sectors = MAX((s->buf_size >> BDRV_SECTOR_BITS) / MAX_IN_FLIGHT,
 343                              MAX_IO_SECTORS);
 344
 345     sector_num = bdrv_dirty_iter_next(s->dbi);
 346     if (sector_num < 0) {
 347         bdrv_set_dirty_iter(s->dbi, 0);
 348         sector_num = bdrv_dirty_iter_next(s->dbi);
 349         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
 350         assert(sector_num >= 0);
 351     }
 352
 353     first_chunk = sector_num / sectors_per_chunk;
 354     while (test_bit(first_chunk, s->in_flight_bitmap)) {
 355         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 356         mirror_wait_for_io(s);
 357     }
 358
 359     block_job_pause_point(&s->common);
 360
 361     /* Find the number of consective dirty chunks following the first dirty
 362      * one, and wait for in flight requests in them. */
 363     while (nb_chunks * sectors_per_chunk < (s->buf_size >> BDRV_SECTOR_BITS)) {
 364         int64_t next_dirty;
 365         int64_t next_sector = sector_num + nb_chunks * sectors_per_chunk;
 366         int64_t next_chunk = next_sector / sectors_per_chunk;
 367         if (next_sector >= end ||
 368             !bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
 369             break;
 370         }
 371         if (test_bit(next_chunk, s->in_flight_bitmap)) {
 372             break;
 373         }
 374
 375         next_dirty = bdrv_dirty_iter_next(s->dbi);
 376         if (next_dirty > next_sector || next_dirty < 0) {
 377             /* The bitmap iterator's cache is stale, refresh it */
 378             bdrv_set_dirty_iter(s->dbi, next_sector);
 379             next_dirty = bdrv_dirty_iter_next(s->dbi);
 380         }
 381         assert(next_dirty == next_sector);
 382         nb_chunks++;
 383     }
 384
 385     /* Clear dirty bits before querying the block status, because
 386      * calling bdrv_get_block_status_above could yield - if some blocks are
 387      * marked dirty in this window, we need to know.
 388      */
 389     bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num,
 390                             nb_chunks * sectors_per_chunk);
 391     bitmap_set(s->in_flight_bitmap, sector_num / sectors_per_chunk, nb_chunks);
 392     while (nb_chunks > 0 && sector_num < end) {
 393         int64_t ret;
 394         int io_sectors, io_sectors_acct;
 395         BlockDriverState *file;
 396         enum MirrorMethod {
 397             MIRROR_METHOD_COPY,
 398             MIRROR_METHOD_ZERO,
 399             MIRROR_METHOD_DISCARD
 400         } mirror_method = MIRROR_METHOD_COPY;
 401
 402         assert(!(sector_num % sectors_per_chunk));
 403         ret = bdrv_get_block_status_above(source, NULL, sector_num,
 404                                           nb_chunks * sectors_per_chunk,
 405                                           &io_sectors, &file);
 406         if (ret < 0) {
 407             io_sectors = MIN(nb_chunks * sectors_per_chunk, max_io_sectors);
 408         } else if (ret & BDRV_BLOCK_DATA) {
 409             io_sectors = MIN(io_sectors, max_io_sectors);
 410         }
 411
 412         io_sectors -= io_sectors % sectors_per_chunk;
 413         if (io_sectors < sectors_per_chunk) {
 414             io_sectors = sectors_per_chunk;
 415         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
 416             int64_t target_sector_num;
 417             int target_nb_sectors;
 418             bdrv_round_sectors_to_clusters(blk_bs(s->target), sector_num,
 419                                            io_sectors,  &target_sector_num,
 420                                            &target_nb_sectors);
 421             if (target_sector_num == sector_num &&
 422                 target_nb_sectors == io_sectors) {
 423                 mirror_method = ret & BDRV_BLOCK_ZERO ?
 424                                     MIRROR_METHOD_ZERO :
 425                                     MIRROR_METHOD_DISCARD;
 426             }
 427         }
 428
 429         while (s->in_flight >= MAX_IN_FLIGHT) {
 430             trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 431             mirror_wait_for_io(s);
 432         }
 433
 434         if (s->ret < 0) {
 435             return 0;
 436         }
 437
 438         mirror_clip_sectors(s, sector_num, &io_sectors);
 439         switch (mirror_method) {
 440         case MIRROR_METHOD_COPY:
 441             io_sectors = mirror_do_read(s, sector_num, io_sectors);
 442             io_sectors_acct = io_sectors;
 443             break;
 444         case MIRROR_METHOD_ZERO:
 445         case MIRROR_METHOD_DISCARD:
 446             mirror_do_zero_or_discard(s, sector_num, io_sectors,
 447                                       mirror_method == MIRROR_METHOD_DISCARD);
 448             if (write_zeroes_ok) {
 449                 io_sectors_acct = 0;
 450             } else {
 451                 io_sectors_acct = io_sectors;
 452             }
 453             break;
 454         default:
 455             abort();
 456         }
 457         assert(io_sectors);
 458         sector_num += io_sectors;
 459         nb_chunks -= DIV_ROUND_UP(io_sectors, sectors_per_chunk);
 460         if (s->common.speed) {
 461             delay_ns = ratelimit_calculate_delay(&s->limit, io_sectors_acct);
 462         }
 463     }
 464     return delay_ns;
 465 }
 466
 467 static void mirror_free_init(MirrorBlockJob *s)
 468 {
 469     int granularity = s->granularity;
 470     size_t buf_size = s->buf_size;
 471     uint8_t *buf = s->buf;
 472
 473     assert(s->buf_free_count == 0);
 474     QSIMPLEQ_INIT(&s->buf_free);
 475     while (buf_size != 0) {
 476         MirrorBuffer *cur = (MirrorBuffer *)buf;
 477         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 478         s->buf_free_count++;
 479         buf_size -= granularity;
 480         buf += granularity;
 481     }
 482 }
 483
 484 /* This is also used for the .pause callback. There is no matching
 485  * mirror_resume() because mirror_run() will begin iterating again
 486  * when the job is resumed.
 487  */
 488 static void mirror_wait_for_all_io(MirrorBlockJob *s)
 489 {
 490     while (s->in_flight > 0) {
 491         mirror_wait_for_io(s);
 492     }
 493 }
 494
 495 typedef struct {
 496     int ret;
 497 } MirrorExitData;
 498
 499 static void mirror_exit(BlockJob *job, void *opaque)
 500 {
 501     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 502     MirrorExitData *data = opaque;
 503     AioContext *replace_aio_context = NULL;
 504     BlockDriverState *src = s->source;
 505     BlockDriverState *target_bs = blk_bs(s->target);
 506     BlockDriverState *mirror_top_bs = s->mirror_top_bs;
 507     Error *local_err = NULL;
 508
 509     /* Make sure that the source BDS doesn't go away before we called
 510      * block_job_completed(). */
 511     bdrv_ref(src);
 512     bdrv_ref(mirror_top_bs);
 513     bdrv_ref(target_bs);
 514
 515     /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
 516      * inserting target_bs at s->to_replace, where we might not be able to get
 517      * these permissions.
 518      *
 519      * Note that blk_unref() alone doesn't necessarily drop permissions because
 520      * we might be running nested inside mirror_drain(), which takes an extra
 521      * reference, so use an explicit blk_set_perm() first. */
 522     blk_set_perm(s->target, 0, BLK_PERM_ALL, &error_abort);
 523     blk_unref(s->target);
 524     s->target = NULL;
 525
 526     /* We don't access the source any more. Dropping any WRITE/RESIZE is
 527      * required before it could become a backing file of target_bs. */
 528     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
 529                             &error_abort);
 530     if (s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
 531         BlockDriverState *backing = s->is_none_mode ? src : s->base;
 532         if (backing_bs(target_bs) != backing) {
 533             bdrv_set_backing_hd(target_bs, backing, &local_err);
 534             if (local_err) {
 535                 error_report_err(local_err);
 536                 data->ret = -EPERM;
 537             }
 538         }
 539     }
 540
 541     if (s->to_replace) {
 542         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 543         aio_context_acquire(replace_aio_context);
 544     }
 545
 546     if (s->should_complete && data->ret == 0) {
 547         BlockDriverState *to_replace = src;
 548         if (s->to_replace) {
 549             to_replace = s->to_replace;
 550         }
 551
 552         if (bdrv_get_flags(target_bs) != bdrv_get_flags(to_replace)) {
 553             bdrv_reopen(target_bs, bdrv_get_flags(to_replace), NULL);
 554         }
 555
 556         /* The mirror job has no requests in flight any more, but we need to
 557          * drain potential other users of the BDS before changing the graph. */
 558         bdrv_drained_begin(target_bs);
 559         bdrv_replace_node(to_replace, target_bs, &local_err);
 560         bdrv_drained_end(target_bs);
 561         if (local_err) {
 562             error_report_err(local_err);
 563             data->ret = -EPERM;
 564         }
 565     }
 566     if (s->to_replace) {
 567         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
 568         error_free(s->replace_blocker);
 569         bdrv_unref(s->to_replace);
 570     }
 571     if (replace_aio_context) {
 572         aio_context_release(replace_aio_context);
 573     }
 574     g_free(s->replaces);
 575     bdrv_unref(target_bs);
 576
 577     /* Remove the mirror filter driver from the graph. Before this, get rid of
 578      * the blockers on the intermediate nodes so that the resulting state is
 579      * valid. Also give up permissions on mirror_top_bs->backing, which might
 580      * block the removal. */
 581     block_job_remove_all_bdrv(job);
 582     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
 583                             &error_abort);
 584     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
 585
 586     /* We just changed the BDS the job BB refers to (with either or both of the
 587      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
 588      * the right thing. We don't need any permissions any more now. */
 589     blk_remove_bs(job->blk);
 590     blk_set_perm(job->blk, 0, BLK_PERM_ALL, &error_abort);
 591     blk_insert_bs(job->blk, mirror_top_bs, &error_abort);
 592
 593     block_job_completed(&s->common, data->ret);
 594
 595     g_free(data);
 596     bdrv_drained_end(src);
 597     bdrv_unref(mirror_top_bs);
 598     bdrv_unref(src);
 599 }
 600
 601 static void mirror_throttle(MirrorBlockJob *s)
 602 {
 603     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 604
 605     if (now - s->last_pause_ns > SLICE_TIME) {
 606         s->last_pause_ns = now;
 607         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
 608     } else {
 609         block_job_pause_point(&s->common);
 610     }
 611 }
 612
 613 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
 614 {
 615     int64_t sector_num, end;
 616     BlockDriverState *base = s->base;
 617     BlockDriverState *bs = s->source;
 618     BlockDriverState *target_bs = blk_bs(s->target);
 619     int ret, n;
 620
 621     end = s->bdev_length / BDRV_SECTOR_SIZE;
 622
 623     if (base == NULL && !bdrv_has_zero_init(target_bs)) {
 624         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
 625             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, end);
 626             return 0;
 627         }
 628
 629         s->initial_zeroing_ongoing = true;
 630         for (sector_num = 0; sector_num < end; ) {
 631             int nb_sectors = MIN(end - sector_num,
 632                 QEMU_ALIGN_DOWN(INT_MAX, s->granularity) >> BDRV_SECTOR_BITS);
 633
 634             mirror_throttle(s);
 635
 636             if (block_job_is_cancelled(&s->common)) {
 637                 s->initial_zeroing_ongoing = false;
 638                 return 0;
 639             }
 640
 641             if (s->in_flight >= MAX_IN_FLIGHT) {
 642                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
 643                                    s->in_flight);
 644                 mirror_wait_for_io(s);
 645                 continue;
 646             }
 647
 648             mirror_do_zero_or_discard(s, sector_num, nb_sectors, false);
 649             sector_num += nb_sectors;
 650         }
 651
 652         mirror_wait_for_all_io(s);
 653         s->initial_zeroing_ongoing = false;
 654     }
 655
 656     /* First part, loop on the sectors and initialize the dirty bitmap.  */
 657     for (sector_num = 0; sector_num < end; ) {
 658         /* Just to make sure we are not exceeding int limit. */
 659         int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS,
 660                              end - sector_num);
 661
 662         mirror_throttle(s);
 663
 664         if (block_job_is_cancelled(&s->common)) {
 665             return 0;
 666         }
 667
 668         ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n);
 669         if (ret < 0) {
 670             return ret;
 671         }
 672
 673         assert(n > 0);
 674         if (ret == 1) {
 675             bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
 676         }
 677         sector_num += n;
 678     }
 679     return 0;
 680 }
 681
 682 /* Called when going out of the streaming phase to flush the bulk of the
 683  * data to the medium, or just before completing.
 684  */
 685 static int mirror_flush(MirrorBlockJob *s)
 686 {
 687     int ret = blk_flush(s->target);
 688     if (ret < 0) {
 689         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
 690             s->ret = ret;
 691         }
 692     }
 693     return ret;
 694 }
 695
 696 static void coroutine_fn mirror_run(void *opaque)
 697 {
 698     MirrorBlockJob *s = opaque;
 699     MirrorExitData *data;
 700     BlockDriverState *bs = s->source;
 701     BlockDriverState *target_bs = blk_bs(s->target);
 702     bool need_drain = true;
 703     int64_t length;
 704     BlockDriverInfo bdi;
 705     char backing_filename[2]; /* we only need 2 characters because we are only
 706                                  checking for a NULL string */
 707     int ret = 0;
 708     int target_cluster_size = BDRV_SECTOR_SIZE;
 709
 710     if (block_job_is_cancelled(&s->common)) {
 711         goto immediate_exit;
 712     }
 713
 714     s->bdev_length = bdrv_getlength(bs);
 715     if (s->bdev_length < 0) {
 716         ret = s->bdev_length;
 717         goto immediate_exit;
 718     }
 719
 720     /* Active commit must resize the base image if its size differs from the
 721      * active layer. */
 722     if (s->base == blk_bs(s->target)) {
 723         int64_t base_length;
 724
 725         base_length = blk_getlength(s->target);
 726         if (base_length < 0) {
 727             ret = base_length;
 728             goto immediate_exit;
 729         }
 730
 731         if (s->bdev_length > base_length) {
 732             ret = blk_truncate(s->target, s->bdev_length, NULL);
 733             if (ret < 0) {
 734                 goto immediate_exit;
 735             }
 736         }
 737     }
 738
 739     if (s->bdev_length == 0) {
 740         /* Report BLOCK_JOB_READY and wait for complete. */
 741         block_job_event_ready(&s->common);
 742         s->synced = true;
 743         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
 744             block_job_yield(&s->common);
 745         }
 746         s->common.cancelled = false;
 747         goto immediate_exit;
 748     }
 749
 750     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
 751     s->in_flight_bitmap = bitmap_new(length);
 752
 753     /* If we have no backing file yet in the destination, we cannot let
 754      * the destination do COW.  Instead, we copy sectors around the
 755      * dirty data if needed.  We need a bitmap to do that.
 756      */
 757     bdrv_get_backing_filename(target_bs, backing_filename,
 758                               sizeof(backing_filename));
 759     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
 760         target_cluster_size = bdi.cluster_size;
 761     }
 762     if (backing_filename[0] && !target_bs->backing
 763         && s->granularity < target_cluster_size) {
 764         s->buf_size = MAX(s->buf_size, target_cluster_size);
 765         s->cow_bitmap = bitmap_new(length);
 766     }
 767     s->target_cluster_sectors = target_cluster_size >> BDRV_SECTOR_BITS;
 768     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
 769
 770     s->buf = qemu_try_blockalign(bs, s->buf_size);
 771     if (s->buf == NULL) {
 772         ret = -ENOMEM;
 773         goto immediate_exit;
 774     }
 775
 776     mirror_free_init(s);
 777
 778     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 779     if (!s->is_none_mode) {
 780         ret = mirror_dirty_init(s);
 781         if (ret < 0 || block_job_is_cancelled(&s->common)) {
 782             goto immediate_exit;
 783         }
 784     }
 785
 786     assert(!s->dbi);
 787     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap, 0);
 788     for (;;) {
 789         uint64_t delay_ns = 0;
 790         int64_t cnt, delta;
 791         bool should_complete;
 792
 793         if (s->ret < 0) {
 794             ret = s->ret;
 795             goto immediate_exit;
 796         }
 797
 798         block_job_pause_point(&s->common);
 799
 800         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 801         /* s->common.offset contains the number of bytes already processed so
 802          * far, cnt is the number of dirty sectors remaining and
 803          * s->sectors_in_flight is the number of sectors currently being
 804          * processed; together those are the current total operation length */
 805         s->common.len = s->common.offset +
 806                         (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
 807
 808         /* Note that even when no rate limit is applied we need to yield
 809          * periodically with no pending I/O so that bdrv_drain_all() returns.
 810          * We do so every SLICE_TIME nanoseconds, or when there is an error,
 811          * or when the source is clean, whichever comes first.
 812          */
 813         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
 814         if (delta < SLICE_TIME &&
 815             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 816             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 817                 (cnt == 0 && s->in_flight > 0)) {
 818                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
 819                 mirror_wait_for_io(s);
 820                 continue;
 821             } else if (cnt != 0) {
 822                 delay_ns = mirror_iteration(s);
 823             }
 824         }
 825
 826         should_complete = false;
 827         if (s->in_flight == 0 && cnt == 0) {
 828             trace_mirror_before_flush(s);
 829             if (!s->synced) {
 830                 if (mirror_flush(s) < 0) {
 831                     /* Go check s->ret.  */
 832                     continue;
 833                 }
 834                 /* We're out of the streaming phase.  From now on, if the job
 835                  * is cancelled we will actually complete all pending I/O and
 836                  * report completion.  This way, block-job-cancel will leave
 837                  * the target in a consistent state.
 838                  */
 839                 block_job_event_ready(&s->common);
 840                 s->synced = true;
 841             }
 842
 843             should_complete = s->should_complete ||
 844                 block_job_is_cancelled(&s->common);
 845             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 846         }
 847
 848         if (cnt == 0 && should_complete) {
 849             /* The dirty bitmap is not updated while operations are pending.
 850              * If we're about to exit, wait for pending operations before
 851              * calling bdrv_get_dirty_count(bs), or we may exit while the
 852              * source has dirty data to copy!
 853              *
 854              * Note that I/O can be submitted by the guest while
 855              * mirror_populate runs, so pause it now.  Before deciding
 856              * whether to switch to target check one last time if I/O has
 857              * come in the meanwhile, and if not flush the data to disk.
 858              */
 859             trace_mirror_before_drain(s, cnt);
 860
 861             bdrv_drained_begin(bs);
 862             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 863             if (cnt > 0 || mirror_flush(s) < 0) {
 864                 bdrv_drained_end(bs);
 865                 continue;
 866             }
 867
 868             /* The two disks are in sync.  Exit and report successful
 869              * completion.
 870              */
 871             assert(QLIST_EMPTY(&bs->tracked_requests));
 872             s->common.cancelled = false;
 873             need_drain = false;
 874             break;
 875         }
 876
 877         ret = 0;
 878         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
 879         if (!s->synced) {
 880             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 881             if (block_job_is_cancelled(&s->common)) {
 882                 break;
 883             }
 884         } else if (!should_complete) {
 885             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 886             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 887         }
 888         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 889     }
 890
 891 immediate_exit:
 892     if (s->in_flight > 0) {
 893         /* We get here only if something went wrong.  Either the job failed,
 894          * or it was cancelled prematurely so that we do not guarantee that
 895          * the target is a copy of the source.
 896          */
 897         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 898         assert(need_drain);
 899         mirror_wait_for_all_io(s);
 900     }
 901
 902     assert(s->in_flight == 0);
 903     qemu_vfree(s->buf);
 904     g_free(s->cow_bitmap);
 905     g_free(s->in_flight_bitmap);
 906     bdrv_dirty_iter_free(s->dbi);
 907     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
 908
 909     data = g_malloc(sizeof(*data));
 910     data->ret = ret;
 911
 912     if (need_drain) {
 913         bdrv_drained_begin(bs);
 914     }
 915     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
 916 }
 917
 918 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 919 {
 920     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 921
 922     if (speed < 0) {
 923         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 924         return;
 925     }
 926     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 927 }
 928
 929 static void mirror_complete(BlockJob *job, Error **errp)
 930 {
 931     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 932     BlockDriverState *target;
 933
 934     target = blk_bs(s->target);
 935
 936     if (!s->synced) {
 937         error_setg(errp, "The active block job '%s' cannot be completed",
 938                    job->id);
 939         return;
 940     }
 941
 942     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
 943         int ret;
 944
 945         assert(!target->backing);
 946         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
 947         if (ret < 0) {
 948             return;
 949         }
 950     }
 951
 952     /* block all operations on to_replace bs */
 953     if (s->replaces) {
 954         AioContext *replace_aio_context;
 955
 956         s->to_replace = bdrv_find_node(s->replaces);
 957         if (!s->to_replace) {
 958             error_setg(errp, "Node name '%s' not found", s->replaces);
 959             return;
 960         }
 961
 962         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 963         aio_context_acquire(replace_aio_context);
 964
 965         /* TODO Translate this into permission system. Current definition of
 966          * GRAPH_MOD would require to request it for the parents; they might
 967          * not even be BlockDriverStates, however, so a BdrvChild can't address
 968          * them. May need redefinition of GRAPH_MOD. */
 969         error_setg(&s->replace_blocker,
 970                    "block device is in use by block-job-complete");
 971         bdrv_op_block_all(s->to_replace, s->replace_blocker);
 972         bdrv_ref(s->to_replace);
 973
 974         aio_context_release(replace_aio_context);
 975     }
 976
 977     s->should_complete = true;
 978     block_job_enter(&s->common);
 979 }
 980
 981 static void mirror_pause(BlockJob *job)
 982 {
 983     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 984
 985     mirror_wait_for_all_io(s);
 986 }
 987
 988 static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context)
 989 {
 990     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 991
 992     blk_set_aio_context(s->target, new_context);
 993 }
 994
 995 static void mirror_drain(BlockJob *job)
 996 {
 997     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 998
 999     /* Need to keep a reference in case blk_drain triggers execution
1000      * of mirror_complete...
1001      */
1002     if (s->target) {
1003         BlockBackend *target = s->target;
1004         blk_ref(target);
1005         blk_drain(target);
1006         blk_unref(target);
1007     }
1008 }
1009
1010 static const BlockJobDriver mirror_job_driver = {
1011     .instance_size          = sizeof(MirrorBlockJob),
1012     .job_type               = BLOCK_JOB_TYPE_MIRROR,
1013     .set_speed              = mirror_set_speed,
1014     .start                  = mirror_run,
1015     .complete               = mirror_complete,
1016     .pause                  = mirror_pause,
1017     .attached_aio_context   = mirror_attached_aio_context,
1018     .drain                  = mirror_drain,
1019 };
1020
1021 static const BlockJobDriver commit_active_job_driver = {
1022     .instance_size          = sizeof(MirrorBlockJob),
1023     .job_type               = BLOCK_JOB_TYPE_COMMIT,
1024     .set_speed              = mirror_set_speed,
1025     .start                  = mirror_run,
1026     .complete               = mirror_complete,
1027     .pause                  = mirror_pause,
1028     .attached_aio_context   = mirror_attached_aio_context,
1029     .drain                  = mirror_drain,
1030 };
1031
1032 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1033     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1034 {
1035     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1036 }
1037
1038 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1039     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1040 {
1041     return bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1042 }
1043
1044 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1045 {
1046     return bdrv_co_flush(bs->backing->bs);
1047 }
1048
1049 static int64_t coroutine_fn bdrv_mirror_top_get_block_status(
1050     BlockDriverState *bs, int64_t sector_num, int nb_sectors, int *pnum,
1051     BlockDriverState **file)
1052 {
1053     *pnum = nb_sectors;
1054     *file = bs->backing->bs;
1055     return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID | BDRV_BLOCK_DATA |
1056            (sector_num << BDRV_SECTOR_BITS);
1057 }
1058
1059 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1060     int64_t offset, int count, BdrvRequestFlags flags)
1061 {
1062     return bdrv_co_pwrite_zeroes(bs->backing, offset, count, flags);
1063 }
1064
1065 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1066     int64_t offset, int count)
1067 {
1068     return bdrv_co_pdiscard(bs->backing->bs, offset, count);
1069 }
1070
1071 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs, QDict *opts)
1072 {
1073     bdrv_refresh_filename(bs->backing->bs);
1074     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1075             bs->backing->bs->filename);
1076 }
1077
1078 static void bdrv_mirror_top_close(BlockDriverState *bs)
1079 {
1080 }
1081
1082 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1083                                        const BdrvChildRole *role,
1084                                        uint64_t perm, uint64_t shared,
1085                                        uint64_t *nperm, uint64_t *nshared)
1086 {
1087     /* Must be able to forward guest writes to the real image */
1088     *nperm = 0;
1089     if (perm & BLK_PERM_WRITE) {
1090         *nperm |= BLK_PERM_WRITE;
1091     }
1092
1093     *nshared = BLK_PERM_ALL;
1094 }
1095
1096 /* Dummy node that provides consistent read to its users without requiring it
1097  * from its backing file and that allows writes on the backing file chain. */
1098 static BlockDriver bdrv_mirror_top = {
1099     .format_name                = "mirror_top",
1100     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1101     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1102     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1103     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1104     .bdrv_co_flush              = bdrv_mirror_top_flush,
1105     .bdrv_co_get_block_status   = bdrv_mirror_top_get_block_status,
1106     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1107     .bdrv_close                 = bdrv_mirror_top_close,
1108     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1109 };
1110
1111 static void mirror_start_job(const char *job_id, BlockDriverState *bs,
1112                              int creation_flags, BlockDriverState *target,
1113                              const char *replaces, int64_t speed,
1114                              uint32_t granularity, int64_t buf_size,
1115                              BlockMirrorBackingMode backing_mode,
1116                              BlockdevOnError on_source_error,
1117                              BlockdevOnError on_target_error,
1118                              bool unmap,
1119                              BlockCompletionFunc *cb,
1120                              void *opaque,
1121                              const BlockJobDriver *driver,
1122                              bool is_none_mode, BlockDriverState *base,
1123                              bool auto_complete, const char *filter_node_name,
1124                              Error **errp)
1125 {
1126     MirrorBlockJob *s;
1127     BlockDriverState *mirror_top_bs;
1128     bool target_graph_mod;
1129     bool target_is_backing;
1130     Error *local_err = NULL;
1131     int ret;
1132
1133     if (granularity == 0) {
1134         granularity = bdrv_get_default_bitmap_granularity(target);
1135     }
1136
1137     assert ((granularity & (granularity - 1)) == 0);
1138
1139     if (buf_size < 0) {
1140         error_setg(errp, "Invalid parameter 'buf-size'");
1141         return;
1142     }
1143
1144     if (buf_size == 0) {
1145         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1146     }
1147
1148     /* In the case of active commit, add dummy driver to provide consistent
1149      * reads on the top, while disabling it in the intermediate nodes, and make
1150      * the backing chain writable. */
1151     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1152                                          BDRV_O_RDWR, errp);
1153     if (mirror_top_bs == NULL) {
1154         return;
1155     }
1156     mirror_top_bs->total_sectors = bs->total_sectors;
1157     bdrv_set_aio_context(mirror_top_bs, bdrv_get_aio_context(bs));
1158
1159     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1160      * it alive until block_job_create() succeeds even if bs has no parent. */
1161     bdrv_ref(mirror_top_bs);
1162     bdrv_drained_begin(bs);
1163     bdrv_append(mirror_top_bs, bs, &local_err);
1164     bdrv_drained_end(bs);
1165
1166     if (local_err) {
1167         bdrv_unref(mirror_top_bs);
1168         error_propagate(errp, local_err);
1169         return;
1170     }
1171
1172     /* Make sure that the source is not resized while the job is running */
1173     s = block_job_create(job_id, driver, mirror_top_bs,
1174                          BLK_PERM_CONSISTENT_READ,
1175                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1176                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1177                          creation_flags, cb, opaque, errp);
1178     if (!s) {
1179         goto fail;
1180     }
1181     /* The block job now has a reference to this node */
1182     bdrv_unref(mirror_top_bs);
1183
1184     s->source = bs;
1185     s->mirror_top_bs = mirror_top_bs;
1186
1187     /* No resize for the target either; while the mirror is still running, a
1188      * consistent read isn't necessarily possible. We could possibly allow
1189      * writes and graph modifications, though it would likely defeat the
1190      * purpose of a mirror, so leave them blocked for now.
1191      *
1192      * In the case of active commit, things look a bit different, though,
1193      * because the target is an already populated backing file in active use.
1194      * We can allow anything except resize there.*/
1195     target_is_backing = bdrv_chain_contains(bs, target);
1196     target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1197     s->target = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE |
1198                         (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1199                         BLK_PERM_WRITE_UNCHANGED |
1200                         (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1201                                              BLK_PERM_WRITE |
1202                                              BLK_PERM_GRAPH_MOD : 0));
1203     ret = blk_insert_bs(s->target, target, errp);
1204     if (ret < 0) {
1205         goto fail;
1206     }
1207
1208     s->replaces = g_strdup(replaces);
1209     s->on_source_error = on_source_error;
1210     s->on_target_error = on_target_error;
1211     s->is_none_mode = is_none_mode;
1212     s->backing_mode = backing_mode;
1213     s->base = base;
1214     s->granularity = granularity;
1215     s->buf_size = ROUND_UP(buf_size, granularity);
1216     s->unmap = unmap;
1217     if (auto_complete) {
1218         s->should_complete = true;
1219     }
1220
1221     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1222     if (!s->dirty_bitmap) {
1223         goto fail;
1224     }
1225
1226     /* Required permissions are already taken with blk_new() */
1227     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1228                        &error_abort);
1229
1230     /* In commit_active_start() all intermediate nodes disappear, so
1231      * any jobs in them must be blocked */
1232     if (target_is_backing) {
1233         BlockDriverState *iter;
1234         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1235             /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1236              * ourselves at s->base (if writes are blocked for a node, they are
1237              * also blocked for its backing file). The other options would be a
1238              * second filter driver above s->base (== target). */
1239             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1240                                      BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1241                                      errp);
1242             if (ret < 0) {
1243                 goto fail;
1244             }
1245         }
1246     }
1247
1248     trace_mirror_start(bs, s, opaque);
1249     block_job_start(&s->common);
1250     return;
1251
1252 fail:
1253     if (s) {
1254         /* Make sure this BDS does not go away until we have completed the graph
1255          * changes below */
1256         bdrv_ref(mirror_top_bs);
1257
1258         g_free(s->replaces);
1259         blk_unref(s->target);
1260         block_job_early_fail(&s->common);
1261     }
1262
1263     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
1264                             &error_abort);
1265     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1266
1267     bdrv_unref(mirror_top_bs);
1268 }
1269
1270 void mirror_start(const char *job_id, BlockDriverState *bs,
1271                   BlockDriverState *target, const char *replaces,
1272                   int64_t speed, uint32_t granularity, int64_t buf_size,
1273                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1274                   BlockdevOnError on_source_error,
1275                   BlockdevOnError on_target_error,
1276                   bool unmap, const char *filter_node_name, Error **errp)
1277 {
1278     bool is_none_mode;
1279     BlockDriverState *base;
1280
1281     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
1282         error_setg(errp, "Sync mode 'incremental' not supported");
1283         return;
1284     }
1285     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1286     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1287     mirror_start_job(job_id, bs, BLOCK_JOB_DEFAULT, target, replaces,
1288                      speed, granularity, buf_size, backing_mode,
1289                      on_source_error, on_target_error, unmap, NULL, NULL,
1290                      &mirror_job_driver, is_none_mode, base, false,
1291                      filter_node_name, errp);
1292 }
1293
1294 void commit_active_start(const char *job_id, BlockDriverState *bs,
1295                          BlockDriverState *base, int creation_flags,
1296                          int64_t speed, BlockdevOnError on_error,
1297                          const char *filter_node_name,
1298                          BlockCompletionFunc *cb, void *opaque,
1299                          bool auto_complete, Error **errp)
1300 {
1301     int orig_base_flags;
1302     Error *local_err = NULL;
1303
1304     orig_base_flags = bdrv_get_flags(base);
1305
1306     if (bdrv_reopen(base, bs->open_flags, errp)) {
1307         return;
1308     }
1309
1310     mirror_start_job(job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1311                      MIRROR_LEAVE_BACKING_CHAIN,
1312                      on_error, on_error, true, cb, opaque,
1313                      &commit_active_job_driver, false, base, auto_complete,
1314                      filter_node_name, &local_err);
1315     if (local_err) {
1316         error_propagate(errp, local_err);
1317         goto error_restore_flags;
1318     }
1319
1320     return;
1321
1322 error_restore_flags:
1323     /* ignore error and errp for bdrv_reopen, because we want to propagate
1324      * the original error */
1325     bdrv_reopen(base, orig_base_flags, NULL);
1326     return;
1327 }