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 CowRequest {
  32     int64_t start_byte;
  33     int64_t end_byte;
  34     QLIST_ENTRY(CowRequest) list;
  35     CoQueue wait_queue; /* coroutines blocked on this request */
  36 } CowRequest;
  37
  38 typedef void (*ProgressBytesCallbackFunc)(int64_t bytes, void *opaque);
  39 typedef void (*ProgressResetCallbackFunc)(void *opaque);
  40 typedef struct BlockCopyState {
  41     BlockBackend *source;
  42     BlockBackend *target;
  43     BdrvDirtyBitmap *copy_bitmap;
  44     int64_t cluster_size;
  45     bool use_copy_range;
  46     int64_t copy_range_size;
  47     uint64_t len;
  48
  49     BdrvRequestFlags write_flags;
  50
  51     /*
  52      * skip_unallocated:
  53      *
  54      * Used by sync=top jobs, which first scan the source node for unallocated
  55      * areas and clear them in the copy_bitmap.  During this process, the bitmap
  56      * is thus not fully initialized: It may still have bits set for areas that
  57      * are unallocated and should actually not be copied.
  58      *
  59      * This is indicated by skip_unallocated.
  60      *
  61      * In this case, block_copy() will query the source’s allocation status,
  62      * skip unallocated regions, clear them in the copy_bitmap, and invoke
  63      * block_copy_reset_unallocated() every time it does.
  64      */
  65     bool skip_unallocated;
  66
  67     /* progress_bytes_callback: called when some copying progress is done. */
  68     ProgressBytesCallbackFunc progress_bytes_callback;
  69
  70     /*
  71      * progress_reset_callback: called when some bytes reset from copy_bitmap
  72      * (see @skip_unallocated above). The callee is assumed to recalculate how
  73      * many bytes remain based on the dirty bit count of copy_bitmap.
  74      */
  75     ProgressResetCallbackFunc progress_reset_callback;
  76     void *progress_opaque;
  77 } BlockCopyState;
  78
  79 typedef struct BackupBlockJob {
  80     BlockJob common;
  81     BlockDriverState *source_bs;
  82
  83     BdrvDirtyBitmap *sync_bitmap;
  84
  85     MirrorSyncMode sync_mode;
  86     BitmapSyncMode bitmap_mode;
  87     BlockdevOnError on_source_error;
  88     BlockdevOnError on_target_error;
  89     CoRwlock flush_rwlock;
  90     uint64_t len;
  91     uint64_t bytes_read;
  92     int64_t cluster_size;
  93     NotifierWithReturn before_write;
  94     QLIST_HEAD(, CowRequest) inflight_reqs;
  95
  96     BlockCopyState *bcs;
  97 } BackupBlockJob;
  98
  99 static const BlockJobDriver backup_job_driver;
 100
 101 /* See if in-flight requests overlap and wait for them to complete */
 102 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
 103                                                        int64_t start,
 104                                                        int64_t end)
 105 {
 106     CowRequest *req;
 107     bool retry;
 108
 109     do {
 110         retry = false;
 111         QLIST_FOREACH(req, &job->inflight_reqs, list) {
 112             if (end > req->start_byte && start < req->end_byte) {
 113                 qemu_co_queue_wait(&req->wait_queue, NULL);
 114                 retry = true;
 115                 break;
 116             }
 117         }
 118     } while (retry);
 119 }
 120
 121 /* Keep track of an in-flight request */
 122 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
 123                               int64_t start, int64_t end)
 124 {
 125     req->start_byte = start;
 126     req->end_byte = end;
 127     qemu_co_queue_init(&req->wait_queue);
 128     QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
 129 }
 130
 131 /* Forget about a completed request */
 132 static void cow_request_end(CowRequest *req)
 133 {
 134     QLIST_REMOVE(req, list);
 135     qemu_co_queue_restart_all(&req->wait_queue);
 136 }
 137
 138 static void block_copy_state_free(BlockCopyState *s)
 139 {
 140     if (!s) {
 141         return;
 142     }
 143
 144     bdrv_release_dirty_bitmap(blk_bs(s->source), s->copy_bitmap);
 145     blk_unref(s->source);
 146     blk_unref(s->target);
 147     g_free(s);
 148 }
 149
 150 static BlockCopyState *block_copy_state_new(
 151         BlockDriverState *source, BlockDriverState *target,
 152         int64_t cluster_size, BdrvRequestFlags write_flags,
 153         ProgressBytesCallbackFunc progress_bytes_callback,
 154         ProgressResetCallbackFunc progress_reset_callback,
 155         void *progress_opaque, Error **errp)
 156 {
 157     BlockCopyState *s;
 158     int ret;
 159     uint64_t no_resize = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE |
 160                          BLK_PERM_WRITE_UNCHANGED | BLK_PERM_GRAPH_MOD;
 161     BdrvDirtyBitmap *copy_bitmap;
 162
 163     copy_bitmap = bdrv_create_dirty_bitmap(source, cluster_size, NULL, errp);
 164     if (!copy_bitmap) {
 165         return NULL;
 166     }
 167     bdrv_disable_dirty_bitmap(copy_bitmap);
 168
 169     s = g_new(BlockCopyState, 1);
 170     *s = (BlockCopyState) {
 171         .source = blk_new(bdrv_get_aio_context(source),
 172                           BLK_PERM_CONSISTENT_READ, no_resize),
 173         .target = blk_new(bdrv_get_aio_context(target),
 174                           BLK_PERM_WRITE, no_resize),
 175         .copy_bitmap = copy_bitmap,
 176         .cluster_size = cluster_size,
 177         .len = bdrv_dirty_bitmap_size(copy_bitmap),
 178         .write_flags = write_flags,
 179         .progress_bytes_callback = progress_bytes_callback,
 180         .progress_reset_callback = progress_reset_callback,
 181         .progress_opaque = progress_opaque,
 182     };
 183
 184     s->copy_range_size = QEMU_ALIGN_DOWN(MIN(blk_get_max_transfer(s->source),
 185                                              blk_get_max_transfer(s->target)),
 186                                          s->cluster_size);
 187     /*
 188      * Set use_copy_range, consider the following:
 189      * 1. Compression is not supported for copy_range.
 190      * 2. copy_range does not respect max_transfer (it's a TODO), so we factor
 191      *    that in here. If max_transfer is smaller than the job->cluster_size,
 192      *    we do not use copy_range (in that case it's zero after aligning down
 193      *    above).
 194      */
 195     s->use_copy_range =
 196         !(write_flags & BDRV_REQ_WRITE_COMPRESSED) && s->copy_range_size > 0;
 197
 198     /*
 199      * We just allow aio context change on our block backends. block_copy() user
 200      * (now it's only backup) is responsible for source and target being in same
 201      * aio context.
 202      */
 203     blk_set_disable_request_queuing(s->source, true);
 204     blk_set_allow_aio_context_change(s->source, true);
 205     blk_set_disable_request_queuing(s->target, true);
 206     blk_set_allow_aio_context_change(s->target, true);
 207
 208     ret = blk_insert_bs(s->source, source, errp);
 209     if (ret < 0) {
 210         goto fail;
 211     }
 212
 213     ret = blk_insert_bs(s->target, target, errp);
 214     if (ret < 0) {
 215         goto fail;
 216     }
 217
 218     return s;
 219
 220 fail:
 221     block_copy_state_free(s);
 222
 223     return NULL;
 224 }
 225
 226 /*
 227  * Copy range to target with a bounce buffer and return the bytes copied. If
 228  * error occurred, return a negative error number
 229  */
 230 static int coroutine_fn block_copy_with_bounce_buffer(BlockCopyState *s,
 231                                                       int64_t start,
 232                                                       int64_t end,
 233                                                       bool is_write_notifier,
 234                                                       bool *error_is_read,
 235                                                       void **bounce_buffer)
 236 {
 237     int ret;
 238     int nbytes;
 239     int read_flags = is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0;
 240
 241     assert(QEMU_IS_ALIGNED(start, s->cluster_size));
 242     bdrv_reset_dirty_bitmap(s->copy_bitmap, start, s->cluster_size);
 243     nbytes = MIN(s->cluster_size, s->len - start);
 244     if (!*bounce_buffer) {
 245         *bounce_buffer = blk_blockalign(s->source, s->cluster_size);
 246     }
 247
 248     ret = blk_co_pread(s->source, start, nbytes, *bounce_buffer, read_flags);
 249     if (ret < 0) {
 250         trace_block_copy_with_bounce_buffer_read_fail(s, start, ret);
 251         if (error_is_read) {
 252             *error_is_read = true;
 253         }
 254         goto fail;
 255     }
 256
 257     ret = blk_co_pwrite(s->target, start, nbytes, *bounce_buffer,
 258                         s->write_flags);
 259     if (ret < 0) {
 260         trace_block_copy_with_bounce_buffer_write_fail(s, start, ret);
 261         if (error_is_read) {
 262             *error_is_read = false;
 263         }
 264         goto fail;
 265     }
 266
 267     return nbytes;
 268 fail:
 269     bdrv_set_dirty_bitmap(s->copy_bitmap, start, s->cluster_size);
 270     return ret;
 271
 272 }
 273
 274 /*
 275  * Copy range to target and return the bytes copied. If error occurred, return a
 276  * negative error number.
 277  */
 278 static int coroutine_fn block_copy_with_offload(BlockCopyState *s,
 279                                                 int64_t start,
 280                                                 int64_t end,
 281                                                 bool is_write_notifier)
 282 {
 283     int ret;
 284     int nr_clusters;
 285     int nbytes;
 286     int read_flags = is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0;
 287
 288     assert(QEMU_IS_ALIGNED(s->copy_range_size, s->cluster_size));
 289     assert(QEMU_IS_ALIGNED(start, s->cluster_size));
 290     nbytes = MIN(s->copy_range_size, MIN(end, s->len) - start);
 291     nr_clusters = DIV_ROUND_UP(nbytes, s->cluster_size);
 292     bdrv_reset_dirty_bitmap(s->copy_bitmap, start,
 293                             s->cluster_size * nr_clusters);
 294     ret = blk_co_copy_range(s->source, start, s->target, start, nbytes,
 295                             read_flags, s->write_flags);
 296     if (ret < 0) {
 297         trace_block_copy_with_offload_fail(s, start, ret);
 298         bdrv_set_dirty_bitmap(s->copy_bitmap, start,
 299                               s->cluster_size * nr_clusters);
 300         return ret;
 301     }
 302
 303     return nbytes;
 304 }
 305
 306 /*
 307  * Check if the cluster starting at offset is allocated or not.
 308  * return via pnum the number of contiguous clusters sharing this allocation.
 309  */
 310 static int block_copy_is_cluster_allocated(BlockCopyState *s, int64_t offset,
 311                                            int64_t *pnum)
 312 {
 313     BlockDriverState *bs = blk_bs(s->source);
 314     int64_t count, total_count = 0;
 315     int64_t bytes = s->len - offset;
 316     int ret;
 317
 318     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
 319
 320     while (true) {
 321         ret = bdrv_is_allocated(bs, offset, bytes, &count);
 322         if (ret < 0) {
 323             return ret;
 324         }
 325
 326         total_count += count;
 327
 328         if (ret || count == 0) {
 329             /*
 330              * ret: partial segment(s) are considered allocated.
 331              * otherwise: unallocated tail is treated as an entire segment.
 332              */
 333             *pnum = DIV_ROUND_UP(total_count, s->cluster_size);
 334             return ret;
 335         }
 336
 337         /* Unallocated segment(s) with uncertain following segment(s) */
 338         if (total_count >= s->cluster_size) {
 339             *pnum = total_count / s->cluster_size;
 340             return 0;
 341         }
 342
 343         offset += count;
 344         bytes -= count;
 345     }
 346 }
 347
 348 /*
 349  * Reset bits in copy_bitmap starting at offset if they represent unallocated
 350  * data in the image. May reset subsequent contiguous bits.
 351  * @return 0 when the cluster at @offset was unallocated,
 352  *         1 otherwise, and -ret on error.
 353  */
 354 static int64_t block_copy_reset_unallocated(BlockCopyState *s,
 355                                             int64_t offset, int64_t *count)
 356 {
 357     int ret;
 358     int64_t clusters, bytes;
 359
 360     ret = block_copy_is_cluster_allocated(s, offset, &clusters);
 361     if (ret < 0) {
 362         return ret;
 363     }
 364
 365     bytes = clusters * s->cluster_size;
 366
 367     if (!ret) {
 368         bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
 369         s->progress_reset_callback(s->progress_opaque);
 370     }
 371
 372     *count = bytes;
 373     return ret;
 374 }
 375
 376 static int coroutine_fn block_copy(BlockCopyState *s,
 377                                    int64_t start, uint64_t bytes,
 378                                    bool *error_is_read,
 379                                    bool is_write_notifier)
 380 {
 381     int ret = 0;
 382     int64_t end = bytes + start; /* bytes */
 383     void *bounce_buffer = NULL;
 384     int64_t status_bytes;
 385
 386     /*
 387      * block_copy() user is responsible for keeping source and target in same
 388      * aio context
 389      */
 390     assert(blk_get_aio_context(s->source) == blk_get_aio_context(s->target));
 391
 392     assert(QEMU_IS_ALIGNED(start, s->cluster_size));
 393     assert(QEMU_IS_ALIGNED(end, s->cluster_size));
 394
 395     while (start < end) {
 396         int64_t dirty_end;
 397
 398         if (!bdrv_dirty_bitmap_get(s->copy_bitmap, start)) {
 399             trace_block_copy_skip(s, start);
 400             start += s->cluster_size;
 401             continue; /* already copied */
 402         }
 403
 404         dirty_end = bdrv_dirty_bitmap_next_zero(s->copy_bitmap, start,
 405                                                 (end - start));
 406         if (dirty_end < 0) {
 407             dirty_end = end;
 408         }
 409
 410         if (s->skip_unallocated) {
 411             ret = block_copy_reset_unallocated(s, start, &status_bytes);
 412             if (ret == 0) {
 413                 trace_block_copy_skip_range(s, start, status_bytes);
 414                 start += status_bytes;
 415                 continue;
 416             }
 417             /* Clamp to known allocated region */
 418             dirty_end = MIN(dirty_end, start + status_bytes);
 419         }
 420
 421         trace_block_copy_process(s, start);
 422
 423         if (s->use_copy_range) {
 424             ret = block_copy_with_offload(s, start, dirty_end,
 425                                           is_write_notifier);
 426             if (ret < 0) {
 427                 s->use_copy_range = false;
 428             }
 429         }
 430         if (!s->use_copy_range) {
 431             ret = block_copy_with_bounce_buffer(s, start, dirty_end,
 432                                                 is_write_notifier,
 433                                                 error_is_read, &bounce_buffer);
 434         }
 435         if (ret < 0) {
 436             break;
 437         }
 438
 439         start += ret;
 440         s->progress_bytes_callback(ret, s->progress_opaque);
 441         ret = 0;
 442     }
 443
 444     if (bounce_buffer) {
 445         qemu_vfree(bounce_buffer);
 446     }
 447
 448     return ret;
 449 }
 450
 451 static void backup_progress_bytes_callback(int64_t bytes, void *opaque)
 452 {
 453     BackupBlockJob *s = opaque;
 454
 455     s->bytes_read += bytes;
 456     job_progress_update(&s->common.job, bytes);
 457 }
 458
 459 static void backup_progress_reset_callback(void *opaque)
 460 {
 461     BackupBlockJob *s = opaque;
 462     uint64_t estimate = bdrv_get_dirty_count(s->bcs->copy_bitmap);
 463
 464     job_progress_set_remaining(&s->common.job, estimate);
 465 }
 466
 467 static int coroutine_fn backup_do_cow(BackupBlockJob *job,
 468                                       int64_t offset, uint64_t bytes,
 469                                       bool *error_is_read,
 470                                       bool is_write_notifier)
 471 {
 472     CowRequest cow_request;
 473     int ret = 0;
 474     int64_t start, end; /* bytes */
 475
 476     qemu_co_rwlock_rdlock(&job->flush_rwlock);
 477
 478     start = QEMU_ALIGN_DOWN(offset, job->cluster_size);
 479     end = QEMU_ALIGN_UP(bytes + offset, job->cluster_size);
 480
 481     trace_backup_do_cow_enter(job, start, offset, bytes);
 482
 483     wait_for_overlapping_requests(job, start, end);
 484     cow_request_begin(&cow_request, job, start, end);
 485
 486     ret = block_copy(job->bcs, start, end - start, error_is_read,
 487                      is_write_notifier);
 488
 489     cow_request_end(&cow_request);
 490
 491     trace_backup_do_cow_return(job, offset, bytes, ret);
 492
 493     qemu_co_rwlock_unlock(&job->flush_rwlock);
 494
 495     return ret;
 496 }
 497
 498 static int coroutine_fn backup_before_write_notify(
 499         NotifierWithReturn *notifier,
 500         void *opaque)
 501 {
 502     BackupBlockJob *job = container_of(notifier, BackupBlockJob, before_write);
 503     BdrvTrackedRequest *req = opaque;
 504
 505     assert(req->bs == job->source_bs);
 506     assert(QEMU_IS_ALIGNED(req->offset, BDRV_SECTOR_SIZE));
 507     assert(QEMU_IS_ALIGNED(req->bytes, BDRV_SECTOR_SIZE));
 508
 509     return backup_do_cow(job, req->offset, req->bytes, NULL, true);
 510 }
 511
 512 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
 513 {
 514     BdrvDirtyBitmap *bm;
 515     bool sync = (((ret == 0) || (job->bitmap_mode == BITMAP_SYNC_MODE_ALWAYS)) \
 516                  && (job->bitmap_mode != BITMAP_SYNC_MODE_NEVER));
 517
 518     if (sync) {
 519         /*
 520          * We succeeded, or we always intended to sync the bitmap.
 521          * Delete this bitmap and install the child.
 522          */
 523         bm = bdrv_dirty_bitmap_abdicate(job->source_bs, job->sync_bitmap, NULL);
 524     } else {
 525         /*
 526          * We failed, or we never intended to sync the bitmap anyway.
 527          * Merge the successor back into the parent, keeping all data.
 528          */
 529         bm = bdrv_reclaim_dirty_bitmap(job->source_bs, job->sync_bitmap, NULL);
 530     }
 531
 532     assert(bm);
 533
 534     if (ret < 0 && job->bitmap_mode == BITMAP_SYNC_MODE_ALWAYS) {
 535         /* If we failed and synced, merge in the bits we didn't copy: */
 536         bdrv_dirty_bitmap_merge_internal(bm, job->bcs->copy_bitmap,
 537                                          NULL, true);
 538     }
 539 }
 540
 541 static void backup_commit(Job *job)
 542 {
 543     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 544     if (s->sync_bitmap) {
 545         backup_cleanup_sync_bitmap(s, 0);
 546     }
 547 }
 548
 549 static void backup_abort(Job *job)
 550 {
 551     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 552     if (s->sync_bitmap) {
 553         backup_cleanup_sync_bitmap(s, -1);
 554     }
 555 }
 556
 557 static void backup_clean(Job *job)
 558 {
 559     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 560
 561     block_copy_state_free(s->bcs);
 562 }
 563
 564 void backup_do_checkpoint(BlockJob *job, Error **errp)
 565 {
 566     BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
 567
 568     assert(block_job_driver(job) == &backup_job_driver);
 569
 570     if (backup_job->sync_mode != MIRROR_SYNC_MODE_NONE) {
 571         error_setg(errp, "The backup job only supports block checkpoint in"
 572                    " sync=none mode");
 573         return;
 574     }
 575
 576     bdrv_set_dirty_bitmap(backup_job->bcs->copy_bitmap, 0, backup_job->len);
 577 }
 578
 579 static BlockErrorAction backup_error_action(BackupBlockJob *job,
 580                                             bool read, int error)
 581 {
 582     if (read) {
 583         return block_job_error_action(&job->common, job->on_source_error,
 584                                       true, error);
 585     } else {
 586         return block_job_error_action(&job->common, job->on_target_error,
 587                                       false, error);
 588     }
 589 }
 590
 591 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
 592 {
 593     uint64_t delay_ns;
 594
 595     if (job_is_cancelled(&job->common.job)) {
 596         return true;
 597     }
 598
 599     /*
 600      * We need to yield even for delay_ns = 0 so that bdrv_drain_all() can
 601      * return. Without a yield, the VM would not reboot.
 602      */
 603     delay_ns = block_job_ratelimit_get_delay(&job->common, job->bytes_read);
 604     job->bytes_read = 0;
 605     job_sleep_ns(&job->common.job, delay_ns);
 606
 607     if (job_is_cancelled(&job->common.job)) {
 608         return true;
 609     }
 610
 611     return false;
 612 }
 613
 614 static int coroutine_fn backup_loop(BackupBlockJob *job)
 615 {
 616     bool error_is_read;
 617     int64_t offset;
 618     BdrvDirtyBitmapIter *bdbi;
 619     int ret = 0;
 620
 621     bdbi = bdrv_dirty_iter_new(job->bcs->copy_bitmap);
 622     while ((offset = bdrv_dirty_iter_next(bdbi)) != -1) {
 623         do {
 624             if (yield_and_check(job)) {
 625                 goto out;
 626             }
 627             ret = backup_do_cow(job, offset,
 628                                 job->cluster_size, &error_is_read, false);
 629             if (ret < 0 && backup_error_action(job, error_is_read, -ret) ==
 630                            BLOCK_ERROR_ACTION_REPORT)
 631             {
 632                 goto out;
 633             }
 634         } while (ret < 0);
 635     }
 636
 637  out:
 638     bdrv_dirty_iter_free(bdbi);
 639     return ret;
 640 }
 641
 642 static void backup_init_copy_bitmap(BackupBlockJob *job)
 643 {
 644     bool ret;
 645     uint64_t estimate;
 646
 647     if (job->sync_mode == MIRROR_SYNC_MODE_BITMAP) {
 648         ret = bdrv_dirty_bitmap_merge_internal(job->bcs->copy_bitmap,
 649                                                job->sync_bitmap,
 650                                                NULL, true);
 651         assert(ret);
 652     } else {
 653         if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
 654             /*
 655              * We can't hog the coroutine to initialize this thoroughly.
 656              * Set a flag and resume work when we are able to yield safely.
 657              */
 658             job->bcs->skip_unallocated = true;
 659         }
 660         bdrv_set_dirty_bitmap(job->bcs->copy_bitmap, 0, job->len);
 661     }
 662
 663     estimate = bdrv_get_dirty_count(job->bcs->copy_bitmap);
 664     job_progress_set_remaining(&job->common.job, estimate);
 665 }
 666
 667 static int coroutine_fn backup_run(Job *job, Error **errp)
 668 {
 669     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 670     int ret = 0;
 671
 672     QLIST_INIT(&s->inflight_reqs);
 673     qemu_co_rwlock_init(&s->flush_rwlock);
 674
 675     backup_init_copy_bitmap(s);
 676
 677     s->before_write.notify = backup_before_write_notify;
 678     bdrv_add_before_write_notifier(s->source_bs, &s->before_write);
 679
 680     if (s->sync_mode == MIRROR_SYNC_MODE_TOP) {
 681         int64_t offset = 0;
 682         int64_t count;
 683
 684         for (offset = 0; offset < s->len; ) {
 685             if (yield_and_check(s)) {
 686                 ret = -ECANCELED;
 687                 goto out;
 688             }
 689
 690             ret = block_copy_reset_unallocated(s->bcs, offset, &count);
 691             if (ret < 0) {
 692                 goto out;
 693             }
 694
 695             offset += count;
 696         }
 697         s->bcs->skip_unallocated = false;
 698     }
 699
 700     if (s->sync_mode == MIRROR_SYNC_MODE_NONE) {
 701         /*
 702          * All bits are set in copy_bitmap to allow any cluster to be copied.
 703          * This does not actually require them to be copied.
 704          */
 705         while (!job_is_cancelled(job)) {
 706             /*
 707              * Yield until the job is cancelled.  We just let our before_write
 708              * notify callback service CoW requests.
 709              */
 710             job_yield(job);
 711         }
 712     } else {
 713         ret = backup_loop(s);
 714     }
 715
 716  out:
 717     notifier_with_return_remove(&s->before_write);
 718
 719     /* wait until pending backup_do_cow() calls have completed */
 720     qemu_co_rwlock_wrlock(&s->flush_rwlock);
 721     qemu_co_rwlock_unlock(&s->flush_rwlock);
 722
 723     return ret;
 724 }
 725
 726 static const BlockJobDriver backup_job_driver = {
 727     .job_driver = {
 728         .instance_size          = sizeof(BackupBlockJob),
 729         .job_type               = JOB_TYPE_BACKUP,
 730         .free                   = block_job_free,
 731         .user_resume            = block_job_user_resume,
 732         .run                    = backup_run,
 733         .commit                 = backup_commit,
 734         .abort                  = backup_abort,
 735         .clean                  = backup_clean,
 736     }
 737 };
 738
 739 static int64_t backup_calculate_cluster_size(BlockDriverState *target,
 740                                              Error **errp)
 741 {
 742     int ret;
 743     BlockDriverInfo bdi;
 744
 745     /*
 746      * If there is no backing file on the target, we cannot rely on COW if our
 747      * backup cluster size is smaller than the target cluster size. Even for
 748      * targets with a backing file, try to avoid COW if possible.
 749      */
 750     ret = bdrv_get_info(target, &bdi);
 751     if (ret == -ENOTSUP && !target->backing) {
 752         /* Cluster size is not defined */
 753         warn_report("The target block device doesn't provide "
 754                     "information about the block size and it doesn't have a "
 755                     "backing file. The default block size of %u bytes is "
 756                     "used. If the actual block size of the target exceeds "
 757                     "this default, the backup may be unusable",
 758                     BACKUP_CLUSTER_SIZE_DEFAULT);
 759         return BACKUP_CLUSTER_SIZE_DEFAULT;
 760     } else if (ret < 0 && !target->backing) {
 761         error_setg_errno(errp, -ret,
 762             "Couldn't determine the cluster size of the target image, "
 763             "which has no backing file");
 764         error_append_hint(errp,
 765             "Aborting, since this may create an unusable destination image\n");
 766         return ret;
 767     } else if (ret < 0 && target->backing) {
 768         /* Not fatal; just trudge on ahead. */
 769         return BACKUP_CLUSTER_SIZE_DEFAULT;
 770     }
 771
 772     return MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
 773 }
 774
 775 BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
 776                   BlockDriverState *target, int64_t speed,
 777                   MirrorSyncMode sync_mode, BdrvDirtyBitmap *sync_bitmap,
 778                   BitmapSyncMode bitmap_mode,
 779                   bool compress,
 780                   BlockdevOnError on_source_error,
 781                   BlockdevOnError on_target_error,
 782                   int creation_flags,
 783                   BlockCompletionFunc *cb, void *opaque,
 784                   JobTxn *txn, Error **errp)
 785 {
 786     int64_t len;
 787     BackupBlockJob *job = NULL;
 788     int64_t cluster_size;
 789     BdrvRequestFlags write_flags;
 790
 791     assert(bs);
 792     assert(target);
 793
 794     /* QMP interface protects us from these cases */
 795     assert(sync_mode != MIRROR_SYNC_MODE_INCREMENTAL);
 796     assert(sync_bitmap || sync_mode != MIRROR_SYNC_MODE_BITMAP);
 797
 798     if (bs == target) {
 799         error_setg(errp, "Source and target cannot be the same");
 800         return NULL;
 801     }
 802
 803     if (!bdrv_is_inserted(bs)) {
 804         error_setg(errp, "Device is not inserted: %s",
 805                    bdrv_get_device_name(bs));
 806         return NULL;
 807     }
 808
 809     if (!bdrv_is_inserted(target)) {
 810         error_setg(errp, "Device is not inserted: %s",
 811                    bdrv_get_device_name(target));
 812         return NULL;
 813     }
 814
 815     if (compress && !block_driver_can_compress(target->drv)) {
 816         error_setg(errp, "Compression is not supported for this drive %s",
 817                    bdrv_get_device_name(target));
 818         return NULL;
 819     }
 820
 821     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
 822         return NULL;
 823     }
 824
 825     if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
 826         return NULL;
 827     }
 828
 829     if (sync_bitmap) {
 830         /* If we need to write to this bitmap, check that we can: */
 831         if (bitmap_mode != BITMAP_SYNC_MODE_NEVER &&
 832             bdrv_dirty_bitmap_check(sync_bitmap, BDRV_BITMAP_DEFAULT, errp)) {
 833             return NULL;
 834         }
 835
 836         /* Create a new bitmap, and freeze/disable this one. */
 837         if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
 838             return NULL;
 839         }
 840     }
 841
 842     len = bdrv_getlength(bs);
 843     if (len < 0) {
 844         error_setg_errno(errp, -len, "unable to get length for '%s'",
 845                          bdrv_get_device_name(bs));
 846         goto error;
 847     }
 848
 849     cluster_size = backup_calculate_cluster_size(target, errp);
 850     if (cluster_size < 0) {
 851         goto error;
 852     }
 853
 854     /* job->len is fixed, so we can't allow resize */
 855     job = block_job_create(job_id, &backup_job_driver, txn, bs, 0, BLK_PERM_ALL,
 856                            speed, creation_flags, cb, opaque, errp);
 857     if (!job) {
 858         goto error;
 859     }
 860
 861     job->source_bs = bs;
 862     job->on_source_error = on_source_error;
 863     job->on_target_error = on_target_error;
 864     job->sync_mode = sync_mode;
 865     job->sync_bitmap = sync_bitmap;
 866     job->bitmap_mode = bitmap_mode;
 867
 868     /*
 869      * If source is in backing chain of target assume that target is going to be
 870      * used for "image fleecing", i.e. it should represent a kind of snapshot of
 871      * source at backup-start point in time. And target is going to be read by
 872      * somebody (for example, used as NBD export) during backup job.
 873      *
 874      * In this case, we need to add BDRV_REQ_SERIALISING write flag to avoid
 875      * intersection of backup writes and third party reads from target,
 876      * otherwise reading from target we may occasionally read already updated by
 877      * guest data.
 878      *
 879      * For more information see commit f8d59dfb40bb and test
 880      * tests/qemu-iotests/222
 881      */
 882     write_flags = (bdrv_chain_contains(target, bs) ? BDRV_REQ_SERIALISING : 0) |
 883                   (compress ? BDRV_REQ_WRITE_COMPRESSED : 0),
 884
 885     job->bcs = block_copy_state_new(bs, target, cluster_size, write_flags,
 886                                     backup_progress_bytes_callback,
 887                                     backup_progress_reset_callback, job, errp);
 888     if (!job->bcs) {
 889         goto error;
 890     }
 891
 892     job->cluster_size = cluster_size;
 893
 894     /* Required permissions are already taken by block-copy-state target */
 895     block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,
 896                        &error_abort);
 897     job->len = len;
 898
 899     return &job->common;
 900
 901  error:
 902     if (sync_bitmap) {
 903         bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL);
 904     }
 905     if (job) {
 906         backup_clean(&job->common.job);
 907         job_early_fail(&job->common.job);
 908     }
 909
 910     return NULL;
 911 }