block/backup.c

   1 /*
   2  * QEMU backup
   3  *
   4  * Copyright (C) 2013 Proxmox Server Solutions
   5  *
   6  * Authors:
   7  *  Dietmar Maurer ([email protected])
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  10  * See the COPYING file in the top-level directory.
  11  *
  12  */
  13
  14 #include "qemu/osdep.h"
  15
  16 #include "trace.h"
  17 #include "block/block.h"
  18 #include "block/block_int.h"
  19 #include "block/blockjob_int.h"
  20 #include "block/block_backup.h"
  21 #include "qapi/error.h"
  22 #include "qapi/qmp/qerror.h"
  23 #include "qemu/ratelimit.h"
  24 #include "qemu/cutils.h"
  25 #include "sysemu/block-backend.h"
  26 #include "qemu/bitmap.h"
  27 #include "qemu/error-report.h"
  28
  29 #define BACKUP_CLUSTER_SIZE_DEFAULT (1 << 16)
  30
  31 typedef struct BackupBlockJob {
  32     BlockJob common;
  33     BlockBackend *target;
  34     /* bitmap for sync=incremental */
  35     BdrvDirtyBitmap *sync_bitmap;
  36     MirrorSyncMode sync_mode;
  37     BlockdevOnError on_source_error;
  38     BlockdevOnError on_target_error;
  39     CoRwlock flush_rwlock;
  40     uint64_t len;
  41     uint64_t bytes_read;
  42     int64_t cluster_size;
  43     bool compress;
  44     NotifierWithReturn before_write;
  45     QLIST_HEAD(, CowRequest) inflight_reqs;
  46
  47     HBitmap *copy_bitmap;
  48     bool use_copy_range;
  49     int64_t copy_range_size;
  50 } BackupBlockJob;
  51
  52 static const BlockJobDriver backup_job_driver;
  53
  54 /* See if in-flight requests overlap and wait for them to complete */
  55 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
  56                                                        int64_t start,
  57                                                        int64_t end)
  58 {
  59     CowRequest *req;
  60     bool retry;
  61
  62     do {
  63         retry = false;
  64         QLIST_FOREACH(req, &job->inflight_reqs, list) {
  65             if (end > req->start_byte && start < req->end_byte) {
  66                 qemu_co_queue_wait(&req->wait_queue, NULL);
  67                 retry = true;
  68                 break;
  69             }
  70         }
  71     } while (retry);
  72 }
  73
  74 /* Keep track of an in-flight request */
  75 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
  76                               int64_t start, int64_t end)
  77 {
  78     req->start_byte = start;
  79     req->end_byte = end;
  80     qemu_co_queue_init(&req->wait_queue);
  81     QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
  82 }
  83
  84 /* Forget about a completed request */
  85 static void cow_request_end(CowRequest *req)
  86 {
  87     QLIST_REMOVE(req, list);
  88     qemu_co_queue_restart_all(&req->wait_queue);
  89 }
  90
  91 /* Copy range to target with a bounce buffer and return the bytes copied. If
  92  * error occured, return a negative error number */
  93 static int coroutine_fn backup_cow_with_bounce_buffer(BackupBlockJob *job,
  94                                                       int64_t start,
  95                                                       int64_t end,
  96                                                       bool is_write_notifier,
  97                                                       bool *error_is_read,
  98                                                       void **bounce_buffer)
  99 {
 100     int ret;
 101     struct iovec iov;
 102     QEMUIOVector qiov;
 103     BlockBackend *blk = job->common.blk;
 104     int nbytes;
 105
 106     hbitmap_reset(job->copy_bitmap, start / job->cluster_size, 1);
 107     nbytes = MIN(job->cluster_size, job->len - start);
 108     if (!*bounce_buffer) {
 109         *bounce_buffer = blk_blockalign(blk, job->cluster_size);
 110     }
 111     iov.iov_base = *bounce_buffer;
 112     iov.iov_len = nbytes;
 113     qemu_iovec_init_external(&qiov, &iov, 1);
 114
 115     ret = blk_co_preadv(blk, start, qiov.size, &qiov,
 116                         is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0);
 117     if (ret < 0) {
 118         trace_backup_do_cow_read_fail(job, start, ret);
 119         if (error_is_read) {
 120             *error_is_read = true;
 121         }
 122         goto fail;
 123     }
 124
 125     if (qemu_iovec_is_zero(&qiov)) {
 126         ret = blk_co_pwrite_zeroes(job->target, start,
 127                                    qiov.size, BDRV_REQ_MAY_UNMAP);
 128     } else {
 129         ret = blk_co_pwritev(job->target, start,
 130                              qiov.size, &qiov,
 131                              job->compress ? BDRV_REQ_WRITE_COMPRESSED : 0);
 132     }
 133     if (ret < 0) {
 134         trace_backup_do_cow_write_fail(job, start, ret);
 135         if (error_is_read) {
 136             *error_is_read = false;
 137         }
 138         goto fail;
 139     }
 140
 141     return nbytes;
 142 fail:
 143     hbitmap_set(job->copy_bitmap, start / job->cluster_size, 1);
 144     return ret;
 145
 146 }
 147
 148 /* Copy range to target and return the bytes copied. If error occured, return a
 149  * negative error number. */
 150 static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job,
 151                                                 int64_t start,
 152                                                 int64_t end,
 153                                                 bool is_write_notifier)
 154 {
 155     int ret;
 156     int nr_clusters;
 157     BlockBackend *blk = job->common.blk;
 158     int nbytes;
 159
 160     assert(QEMU_IS_ALIGNED(job->copy_range_size, job->cluster_size));
 161     nbytes = MIN(job->copy_range_size, end - start);
 162     nr_clusters = DIV_ROUND_UP(nbytes, job->cluster_size);
 163     hbitmap_reset(job->copy_bitmap, start / job->cluster_size,
 164                   nr_clusters);
 165     ret = blk_co_copy_range(blk, start, job->target, start, nbytes,
 166                             is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0);
 167     if (ret < 0) {
 168         trace_backup_do_cow_copy_range_fail(job, start, ret);
 169         hbitmap_set(job->copy_bitmap, start / job->cluster_size,
 170                     nr_clusters);
 171         return ret;
 172     }
 173
 174     return nbytes;
 175 }
 176
 177 static int coroutine_fn backup_do_cow(BackupBlockJob *job,
 178                                       int64_t offset, uint64_t bytes,
 179                                       bool *error_is_read,
 180                                       bool is_write_notifier)
 181 {
 182     CowRequest cow_request;
 183     int ret = 0;
 184     int64_t start, end; /* bytes */
 185     void *bounce_buffer = NULL;
 186
 187     qemu_co_rwlock_rdlock(&job->flush_rwlock);
 188
 189     start = QEMU_ALIGN_DOWN(offset, job->cluster_size);
 190     end = QEMU_ALIGN_UP(bytes + offset, job->cluster_size);
 191
 192     trace_backup_do_cow_enter(job, start, offset, bytes);
 193
 194     wait_for_overlapping_requests(job, start, end);
 195     cow_request_begin(&cow_request, job, start, end);
 196
 197     while (start < end) {
 198         if (!hbitmap_get(job->copy_bitmap, start / job->cluster_size)) {
 199             trace_backup_do_cow_skip(job, start);
 200             start += job->cluster_size;
 201             continue; /* already copied */
 202         }
 203
 204         trace_backup_do_cow_process(job, start);
 205
 206         if (job->use_copy_range) {
 207             ret = backup_cow_with_offload(job, start, end, is_write_notifier);
 208             if (ret < 0) {
 209                 job->use_copy_range = false;
 210             }
 211         }
 212         if (!job->use_copy_range) {
 213             ret = backup_cow_with_bounce_buffer(job, start, end, is_write_notifier,
 214                                                 error_is_read, &bounce_buffer);
 215         }
 216         if (ret < 0) {
 217             break;
 218         }
 219
 220         /* Publish progress, guest I/O counts as progress too.  Note that the
 221          * offset field is an opaque progress value, it is not a disk offset.
 222          */
 223         start += ret;
 224         job->bytes_read += ret;
 225         job_progress_update(&job->common.job, ret);
 226         ret = 0;
 227     }
 228
 229     if (bounce_buffer) {
 230         qemu_vfree(bounce_buffer);
 231     }
 232
 233     cow_request_end(&cow_request);
 234
 235     trace_backup_do_cow_return(job, offset, bytes, ret);
 236
 237     qemu_co_rwlock_unlock(&job->flush_rwlock);
 238
 239     return ret;
 240 }
 241
 242 static int coroutine_fn backup_before_write_notify(
 243         NotifierWithReturn *notifier,
 244         void *opaque)
 245 {
 246     BackupBlockJob *job = container_of(notifier, BackupBlockJob, before_write);
 247     BdrvTrackedRequest *req = opaque;
 248
 249     assert(req->bs == blk_bs(job->common.blk));
 250     assert(QEMU_IS_ALIGNED(req->offset, BDRV_SECTOR_SIZE));
 251     assert(QEMU_IS_ALIGNED(req->bytes, BDRV_SECTOR_SIZE));
 252
 253     return backup_do_cow(job, req->offset, req->bytes, NULL, true);
 254 }
 255
 256 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
 257 {
 258     BdrvDirtyBitmap *bm;
 259     BlockDriverState *bs = blk_bs(job->common.blk);
 260
 261     if (ret < 0) {
 262         /* Merge the successor back into the parent, delete nothing. */
 263         bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL);
 264         assert(bm);
 265     } else {
 266         /* Everything is fine, delete this bitmap and install the backup. */
 267         bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL);
 268         assert(bm);
 269     }
 270 }
 271
 272 static void backup_commit(Job *job)
 273 {
 274     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 275     if (s->sync_bitmap) {
 276         backup_cleanup_sync_bitmap(s, 0);
 277     }
 278 }
 279
 280 static void backup_abort(Job *job)
 281 {
 282     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 283     if (s->sync_bitmap) {
 284         backup_cleanup_sync_bitmap(s, -1);
 285     }
 286 }
 287
 288 static void backup_clean(Job *job)
 289 {
 290     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 291     assert(s->target);
 292     blk_unref(s->target);
 293     s->target = NULL;
 294 }
 295
 296 static void backup_attached_aio_context(BlockJob *job, AioContext *aio_context)
 297 {
 298     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 299
 300     blk_set_aio_context(s->target, aio_context);
 301 }
 302
 303 void backup_do_checkpoint(BlockJob *job, Error **errp)
 304 {
 305     BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
 306     int64_t len;
 307
 308     assert(block_job_driver(job) == &backup_job_driver);
 309
 310     if (backup_job->sync_mode != MIRROR_SYNC_MODE_NONE) {
 311         error_setg(errp, "The backup job only supports block checkpoint in"
 312                    " sync=none mode");
 313         return;
 314     }
 315
 316     len = DIV_ROUND_UP(backup_job->len, backup_job->cluster_size);
 317     hbitmap_set(backup_job->copy_bitmap, 0, len);
 318 }
 319
 320 void backup_wait_for_overlapping_requests(BlockJob *job, int64_t offset,
 321                                           uint64_t bytes)
 322 {
 323     BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
 324     int64_t start, end;
 325
 326     assert(block_job_driver(job) == &backup_job_driver);
 327
 328     start = QEMU_ALIGN_DOWN(offset, backup_job->cluster_size);
 329     end = QEMU_ALIGN_UP(offset + bytes, backup_job->cluster_size);
 330     wait_for_overlapping_requests(backup_job, start, end);
 331 }
 332
 333 void backup_cow_request_begin(CowRequest *req, BlockJob *job,
 334                               int64_t offset, uint64_t bytes)
 335 {
 336     BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
 337     int64_t start, end;
 338
 339     assert(block_job_driver(job) == &backup_job_driver);
 340
 341     start = QEMU_ALIGN_DOWN(offset, backup_job->cluster_size);
 342     end = QEMU_ALIGN_UP(offset + bytes, backup_job->cluster_size);
 343     cow_request_begin(req, backup_job, start, end);
 344 }
 345
 346 void backup_cow_request_end(CowRequest *req)
 347 {
 348     cow_request_end(req);
 349 }
 350
 351 static void backup_drain(BlockJob *job)
 352 {
 353     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 354
 355     /* Need to keep a reference in case blk_drain triggers execution
 356      * of backup_complete...
 357      */
 358     if (s->target) {
 359         BlockBackend *target = s->target;
 360         blk_ref(target);
 361         blk_drain(target);
 362         blk_unref(target);
 363     }
 364 }
 365
 366 static BlockErrorAction backup_error_action(BackupBlockJob *job,
 367                                             bool read, int error)
 368 {
 369     if (read) {
 370         return block_job_error_action(&job->common, job->on_source_error,
 371                                       true, error);
 372     } else {
 373         return block_job_error_action(&job->common, job->on_target_error,
 374                                       false, error);
 375     }
 376 }
 377
 378 typedef struct {
 379     int ret;
 380 } BackupCompleteData;
 381
 382 static void backup_complete(Job *job, void *opaque)
 383 {
 384     BackupCompleteData *data = opaque;
 385
 386     job_completed(job, data->ret, NULL);
 387     g_free(data);
 388 }
 389
 390 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
 391 {
 392     uint64_t delay_ns;
 393
 394     if (job_is_cancelled(&job->common.job)) {
 395         return true;
 396     }
 397
 398     /* We need to yield even for delay_ns = 0 so that bdrv_drain_all() can
 399      * return. Without a yield, the VM would not reboot. */
 400     delay_ns = block_job_ratelimit_get_delay(&job->common, job->bytes_read);
 401     job->bytes_read = 0;
 402     job_sleep_ns(&job->common.job, delay_ns);
 403
 404     if (job_is_cancelled(&job->common.job)) {
 405         return true;
 406     }
 407
 408     return false;
 409 }
 410
 411 static int coroutine_fn backup_run_incremental(BackupBlockJob *job)
 412 {
 413     int ret;
 414     bool error_is_read;
 415     int64_t cluster;
 416     HBitmapIter hbi;
 417
 418     hbitmap_iter_init(&hbi, job->copy_bitmap, 0);
 419     while ((cluster = hbitmap_iter_next(&hbi, true)) != -1) {
 420         do {
 421             if (yield_and_check(job)) {
 422                 return 0;
 423             }
 424             ret = backup_do_cow(job, cluster * job->cluster_size,
 425                                 job->cluster_size, &error_is_read, false);
 426             if (ret < 0 && backup_error_action(job, error_is_read, -ret) ==
 427                            BLOCK_ERROR_ACTION_REPORT)
 428             {
 429                 return ret;
 430             }
 431         } while (ret < 0);
 432     }
 433
 434     return 0;
 435 }
 436
 437 /* init copy_bitmap from sync_bitmap */
 438 static void backup_incremental_init_copy_bitmap(BackupBlockJob *job)
 439 {
 440     BdrvDirtyBitmapIter *dbi;
 441     int64_t offset;
 442     int64_t end = DIV_ROUND_UP(bdrv_dirty_bitmap_size(job->sync_bitmap),
 443                                job->cluster_size);
 444
 445     dbi = bdrv_dirty_iter_new(job->sync_bitmap);
 446     while ((offset = bdrv_dirty_iter_next(dbi)) != -1) {
 447         int64_t cluster = offset / job->cluster_size;
 448         int64_t next_cluster;
 449
 450         offset += bdrv_dirty_bitmap_granularity(job->sync_bitmap);
 451         if (offset >= bdrv_dirty_bitmap_size(job->sync_bitmap)) {
 452             hbitmap_set(job->copy_bitmap, cluster, end - cluster);
 453             break;
 454         }
 455
 456         offset = bdrv_dirty_bitmap_next_zero(job->sync_bitmap, offset);
 457         if (offset == -1) {
 458             hbitmap_set(job->copy_bitmap, cluster, end - cluster);
 459             break;
 460         }
 461
 462         next_cluster = DIV_ROUND_UP(offset, job->cluster_size);
 463         hbitmap_set(job->copy_bitmap, cluster, next_cluster - cluster);
 464         if (next_cluster >= end) {
 465             break;
 466         }
 467
 468         bdrv_set_dirty_iter(dbi, next_cluster * job->cluster_size);
 469     }
 470
 471     /* TODO job_progress_set_remaining() would make more sense */
 472     job_progress_update(&job->common.job,
 473         job->len - hbitmap_count(job->copy_bitmap) * job->cluster_size);
 474
 475     bdrv_dirty_iter_free(dbi);
 476 }
 477
 478 static void coroutine_fn backup_run(void *opaque)
 479 {
 480     BackupBlockJob *job = opaque;
 481     BackupCompleteData *data;
 482     BlockDriverState *bs = blk_bs(job->common.blk);
 483     int64_t offset, nb_clusters;
 484     int ret = 0;
 485
 486     QLIST_INIT(&job->inflight_reqs);
 487     qemu_co_rwlock_init(&job->flush_rwlock);
 488
 489     nb_clusters = DIV_ROUND_UP(job->len, job->cluster_size);
 490     job_progress_set_remaining(&job->common.job, job->len);
 491
 492     job->copy_bitmap = hbitmap_alloc(nb_clusters, 0);
 493     if (job->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 494         backup_incremental_init_copy_bitmap(job);
 495     } else {
 496         hbitmap_set(job->copy_bitmap, 0, nb_clusters);
 497     }
 498
 499
 500     job->before_write.notify = backup_before_write_notify;
 501     bdrv_add_before_write_notifier(bs, &job->before_write);
 502
 503     if (job->sync_mode == MIRROR_SYNC_MODE_NONE) {
 504         /* All bits are set in copy_bitmap to allow any cluster to be copied.
 505          * This does not actually require them to be copied. */
 506         while (!job_is_cancelled(&job->common.job)) {
 507             /* Yield until the job is cancelled.  We just let our before_write
 508              * notify callback service CoW requests. */
 509             job_yield(&job->common.job);
 510         }
 511     } else if (job->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 512         ret = backup_run_incremental(job);
 513     } else {
 514         /* Both FULL and TOP SYNC_MODE's require copying.. */
 515         for (offset = 0; offset < job->len;
 516              offset += job->cluster_size) {
 517             bool error_is_read;
 518             int alloced = 0;
 519
 520             if (yield_and_check(job)) {
 521                 break;
 522             }
 523
 524             if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
 525                 int i;
 526                 int64_t n;
 527
 528                 /* Check to see if these blocks are already in the
 529                  * backing file. */
 530
 531                 for (i = 0; i < job->cluster_size;) {
 532                     /* bdrv_is_allocated() only returns true/false based
 533                      * on the first set of sectors it comes across that
 534                      * are are all in the same state.
 535                      * For that reason we must verify each sector in the
 536                      * backup cluster length.  We end up copying more than
 537                      * needed but at some point that is always the case. */
 538                     alloced =
 539                         bdrv_is_allocated(bs, offset + i,
 540                                           job->cluster_size - i, &n);
 541                     i += n;
 542
 543                     if (alloced || n == 0) {
 544                         break;
 545                     }
 546                 }
 547
 548                 /* If the above loop never found any sectors that are in
 549                  * the topmost image, skip this backup. */
 550                 if (alloced == 0) {
 551                     continue;
 552                 }
 553             }
 554             /* FULL sync mode we copy the whole drive. */
 555             if (alloced < 0) {
 556                 ret = alloced;
 557             } else {
 558                 ret = backup_do_cow(job, offset, job->cluster_size,
 559                                     &error_is_read, false);
 560             }
 561             if (ret < 0) {
 562                 /* Depending on error action, fail now or retry cluster */
 563                 BlockErrorAction action =
 564                     backup_error_action(job, error_is_read, -ret);
 565                 if (action == BLOCK_ERROR_ACTION_REPORT) {
 566                     break;
 567                 } else {
 568                     offset -= job->cluster_size;
 569                     continue;
 570                 }
 571             }
 572         }
 573     }
 574
 575     notifier_with_return_remove(&job->before_write);
 576
 577     /* wait until pending backup_do_cow() calls have completed */
 578     qemu_co_rwlock_wrlock(&job->flush_rwlock);
 579     qemu_co_rwlock_unlock(&job->flush_rwlock);
 580     hbitmap_free(job->copy_bitmap);
 581
 582     data = g_malloc(sizeof(*data));
 583     data->ret = ret;
 584     job_defer_to_main_loop(&job->common.job, backup_complete, data);
 585 }
 586
 587 static const BlockJobDriver backup_job_driver = {
 588     .job_driver = {
 589         .instance_size          = sizeof(BackupBlockJob),
 590         .job_type               = JOB_TYPE_BACKUP,
 591         .free                   = block_job_free,
 592         .user_resume            = block_job_user_resume,
 593         .drain                  = block_job_drain,
 594         .start                  = backup_run,
 595         .commit                 = backup_commit,
 596         .abort                  = backup_abort,
 597         .clean                  = backup_clean,
 598     },
 599     .attached_aio_context   = backup_attached_aio_context,
 600     .drain                  = backup_drain,
 601 };
 602
 603 BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
 604                   BlockDriverState *target, int64_t speed,
 605                   MirrorSyncMode sync_mode, BdrvDirtyBitmap *sync_bitmap,
 606                   bool compress,
 607                   BlockdevOnError on_source_error,
 608                   BlockdevOnError on_target_error,
 609                   int creation_flags,
 610                   BlockCompletionFunc *cb, void *opaque,
 611                   JobTxn *txn, Error **errp)
 612 {
 613     int64_t len;
 614     BlockDriverInfo bdi;
 615     BackupBlockJob *job = NULL;
 616     int ret;
 617
 618     assert(bs);
 619     assert(target);
 620
 621     if (bs == target) {
 622         error_setg(errp, "Source and target cannot be the same");
 623         return NULL;
 624     }
 625
 626     if (!bdrv_is_inserted(bs)) {
 627         error_setg(errp, "Device is not inserted: %s",
 628                    bdrv_get_device_name(bs));
 629         return NULL;
 630     }
 631
 632     if (!bdrv_is_inserted(target)) {
 633         error_setg(errp, "Device is not inserted: %s",
 634                    bdrv_get_device_name(target));
 635         return NULL;
 636     }
 637
 638     if (compress && target->drv->bdrv_co_pwritev_compressed == NULL) {
 639         error_setg(errp, "Compression is not supported for this drive %s",
 640                    bdrv_get_device_name(target));
 641         return NULL;
 642     }
 643
 644     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
 645         return NULL;
 646     }
 647
 648     if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
 649         return NULL;
 650     }
 651
 652     if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 653         if (!sync_bitmap) {
 654             error_setg(errp, "must provide a valid bitmap name for "
 655                              "\"incremental\" sync mode");
 656             return NULL;
 657         }
 658
 659         /* Create a new bitmap, and freeze/disable this one. */
 660         if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
 661             return NULL;
 662         }
 663     } else if (sync_bitmap) {
 664         error_setg(errp,
 665                    "a sync_bitmap was provided to backup_run, "
 666                    "but received an incompatible sync_mode (%s)",
 667                    MirrorSyncMode_str(sync_mode));
 668         return NULL;
 669     }
 670
 671     len = bdrv_getlength(bs);
 672     if (len < 0) {
 673         error_setg_errno(errp, -len, "unable to get length for '%s'",
 674                          bdrv_get_device_name(bs));
 675         goto error;
 676     }
 677
 678     /* job->len is fixed, so we can't allow resize */
 679     job = block_job_create(job_id, &backup_job_driver, txn, bs,
 680                            BLK_PERM_CONSISTENT_READ,
 681                            BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE |
 682                            BLK_PERM_WRITE_UNCHANGED | BLK_PERM_GRAPH_MOD,
 683                            speed, creation_flags, cb, opaque, errp);
 684     if (!job) {
 685         goto error;
 686     }
 687
 688     /* The target must match the source in size, so no resize here either */
 689     job->target = blk_new(BLK_PERM_WRITE,
 690                           BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE |
 691                           BLK_PERM_WRITE_UNCHANGED | BLK_PERM_GRAPH_MOD);
 692     ret = blk_insert_bs(job->target, target, errp);
 693     if (ret < 0) {
 694         goto error;
 695     }
 696
 697     job->on_source_error = on_source_error;
 698     job->on_target_error = on_target_error;
 699     job->sync_mode = sync_mode;
 700     job->sync_bitmap = sync_mode == MIRROR_SYNC_MODE_INCREMENTAL ?
 701                        sync_bitmap : NULL;
 702     job->compress = compress;
 703
 704     /* If there is no backing file on the target, we cannot rely on COW if our
 705      * backup cluster size is smaller than the target cluster size. Even for
 706      * targets with a backing file, try to avoid COW if possible. */
 707     ret = bdrv_get_info(target, &bdi);
 708     if (ret == -ENOTSUP && !target->backing) {
 709         /* Cluster size is not defined */
 710         warn_report("The target block device doesn't provide "
 711                     "information about the block size and it doesn't have a "
 712                     "backing file. The default block size of %u bytes is "
 713                     "used. If the actual block size of the target exceeds "
 714                     "this default, the backup may be unusable",
 715                     BACKUP_CLUSTER_SIZE_DEFAULT);
 716         job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
 717     } else if (ret < 0 && !target->backing) {
 718         error_setg_errno(errp, -ret,
 719             "Couldn't determine the cluster size of the target image, "
 720             "which has no backing file");
 721         error_append_hint(errp,
 722             "Aborting, since this may create an unusable destination image\n");
 723         goto error;
 724     } else if (ret < 0 && target->backing) {
 725         /* Not fatal; just trudge on ahead. */
 726         job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
 727     } else {
 728         job->cluster_size = MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
 729     }
 730     job->use_copy_range = true;
 731     job->copy_range_size = MIN_NON_ZERO(blk_get_max_transfer(job->common.blk),
 732                                         blk_get_max_transfer(job->target));
 733     job->copy_range_size = MAX(job->cluster_size,
 734                                QEMU_ALIGN_UP(job->copy_range_size,
 735                                              job->cluster_size));
 736
 737     /* Required permissions are already taken with target's blk_new() */
 738     block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,
 739                        &error_abort);
 740     job->len = len;
 741
 742     return &job->common;
 743
 744  error:
 745     if (sync_bitmap) {
 746         bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL);
 747     }
 748     if (job) {
 749         backup_clean(&job->common.job);
 750         job_early_fail(&job->common.job);
 751     }
 752
 753     return NULL;
 754 }