block/backup.c

   1 /*
   2  * QEMU backup
   3  *
   4  * Copyright (C) 2013 Proxmox Server Solutions
   5  * Copyright (c) 2019 Virtuozzo International GmbH.
   6  *
   7  * Authors:
   8  *  Dietmar Maurer ([email protected])
   9  *
  10  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  11  * See the COPYING file in the top-level directory.
  12  *
  13  */
  14
  15 #include "qemu/osdep.h"
  16
  17 #include "trace.h"
  18 #include "block/block.h"
  19 #include "block/block_int.h"
  20 #include "block/blockjob_int.h"
  21 #include "block/block_backup.h"
  22 #include "block/block-copy.h"
  23 #include "qapi/error.h"
  24 #include "qapi/qmp/qerror.h"
  25 #include "qemu/cutils.h"
  26 #include "sysemu/block-backend.h"
  27 #include "qemu/bitmap.h"
  28 #include "qemu/error-report.h"
  29
  30 #include "block/backup-top.h"
  31
  32 #define BACKUP_CLUSTER_SIZE_DEFAULT (1 << 16)
  33
  34 typedef struct BackupBlockJob {
  35     BlockJob common;
  36     BlockDriverState *backup_top;
  37     BlockDriverState *source_bs;
  38
  39     BdrvDirtyBitmap *sync_bitmap;
  40
  41     MirrorSyncMode sync_mode;
  42     BitmapSyncMode bitmap_mode;
  43     BlockdevOnError on_source_error;
  44     BlockdevOnError on_target_error;
  45     uint64_t len;
  46     int64_t cluster_size;
  47     BackupPerf perf;
  48
  49     BlockCopyState *bcs;
  50
  51     bool wait;
  52     BlockCopyCallState *bg_bcs_call;
  53 } BackupBlockJob;
  54
  55 static const BlockJobDriver backup_job_driver;
  56
  57 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
  58 {
  59     BdrvDirtyBitmap *bm;
  60     bool sync = (((ret == 0) || (job->bitmap_mode == BITMAP_SYNC_MODE_ALWAYS)) \
  61                  && (job->bitmap_mode != BITMAP_SYNC_MODE_NEVER));
  62
  63     if (sync) {
  64         /*
  65          * We succeeded, or we always intended to sync the bitmap.
  66          * Delete this bitmap and install the child.
  67          */
  68         bm = bdrv_dirty_bitmap_abdicate(job->sync_bitmap, NULL);
  69     } else {
  70         /*
  71          * We failed, or we never intended to sync the bitmap anyway.
  72          * Merge the successor back into the parent, keeping all data.
  73          */
  74         bm = bdrv_reclaim_dirty_bitmap(job->sync_bitmap, NULL);
  75     }
  76
  77     assert(bm);
  78
  79     if (ret < 0 && job->bitmap_mode == BITMAP_SYNC_MODE_ALWAYS) {
  80         /* If we failed and synced, merge in the bits we didn't copy: */
  81         bdrv_dirty_bitmap_merge_internal(bm, block_copy_dirty_bitmap(job->bcs),
  82                                          NULL, true);
  83     }
  84 }
  85
  86 static void backup_commit(Job *job)
  87 {
  88     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
  89     if (s->sync_bitmap) {
  90         backup_cleanup_sync_bitmap(s, 0);
  91     }
  92 }
  93
  94 static void backup_abort(Job *job)
  95 {
  96     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
  97     if (s->sync_bitmap) {
  98         backup_cleanup_sync_bitmap(s, -1);
  99     }
 100 }
 101
 102 static void backup_clean(Job *job)
 103 {
 104     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 105     bdrv_backup_top_drop(s->backup_top);
 106 }
 107
 108 void backup_do_checkpoint(BlockJob *job, Error **errp)
 109 {
 110     BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
 111
 112     assert(block_job_driver(job) == &backup_job_driver);
 113
 114     if (backup_job->sync_mode != MIRROR_SYNC_MODE_NONE) {
 115         error_setg(errp, "The backup job only supports block checkpoint in"
 116                    " sync=none mode");
 117         return;
 118     }
 119
 120     bdrv_set_dirty_bitmap(block_copy_dirty_bitmap(backup_job->bcs), 0,
 121                           backup_job->len);
 122 }
 123
 124 static BlockErrorAction backup_error_action(BackupBlockJob *job,
 125                                             bool read, int error)
 126 {
 127     if (read) {
 128         return block_job_error_action(&job->common, job->on_source_error,
 129                                       true, error);
 130     } else {
 131         return block_job_error_action(&job->common, job->on_target_error,
 132                                       false, error);
 133     }
 134 }
 135
 136 static void coroutine_fn backup_block_copy_callback(void *opaque)
 137 {
 138     BackupBlockJob *s = opaque;
 139
 140     if (s->wait) {
 141         s->wait = false;
 142         aio_co_wake(s->common.job.co);
 143     } else {
 144         job_enter(&s->common.job);
 145     }
 146 }
 147
 148 static int coroutine_fn backup_loop(BackupBlockJob *job)
 149 {
 150     BlockCopyCallState *s = NULL;
 151     int ret = 0;
 152     bool error_is_read;
 153     BlockErrorAction act;
 154
 155     while (true) { /* retry loop */
 156         job->bg_bcs_call = s = block_copy_async(job->bcs, 0,
 157                 QEMU_ALIGN_UP(job->len, job->cluster_size),
 158                 job->perf.max_workers, job->perf.max_chunk,
 159                 backup_block_copy_callback, job);
 160
 161         while (!block_copy_call_finished(s) &&
 162                !job_is_cancelled(&job->common.job))
 163         {
 164             job_yield(&job->common.job);
 165         }
 166
 167         if (!block_copy_call_finished(s)) {
 168             assert(job_is_cancelled(&job->common.job));
 169             /*
 170              * Note that we can't use job_yield() here, as it doesn't work for
 171              * cancelled job.
 172              */
 173             block_copy_call_cancel(s);
 174             job->wait = true;
 175             qemu_coroutine_yield();
 176             assert(block_copy_call_finished(s));
 177             ret = 0;
 178             goto out;
 179         }
 180
 181         if (job_is_cancelled(&job->common.job) ||
 182             block_copy_call_succeeded(s))
 183         {
 184             ret = 0;
 185             goto out;
 186         }
 187
 188         if (block_copy_call_cancelled(s)) {
 189             /*
 190              * Job is not cancelled but only block-copy call. This is possible
 191              * after job pause. Now the pause is finished, start new block-copy
 192              * iteration.
 193              */
 194             block_copy_call_free(s);
 195             continue;
 196         }
 197
 198         /* The only remaining case is failed block-copy call. */
 199         assert(block_copy_call_failed(s));
 200
 201         ret = block_copy_call_status(s, &error_is_read);
 202         act = backup_error_action(job, error_is_read, -ret);
 203         switch (act) {
 204         case BLOCK_ERROR_ACTION_REPORT:
 205             goto out;
 206         case BLOCK_ERROR_ACTION_STOP:
 207             /*
 208              * Go to pause prior to starting new block-copy call on the next
 209              * iteration.
 210              */
 211             job_pause_point(&job->common.job);
 212             break;
 213         case BLOCK_ERROR_ACTION_IGNORE:
 214             /* Proceed to new block-copy call to retry. */
 215             break;
 216         default:
 217             abort();
 218         }
 219
 220         block_copy_call_free(s);
 221     }
 222
 223 out:
 224     block_copy_call_free(s);
 225     job->bg_bcs_call = NULL;
 226     return ret;
 227 }
 228
 229 static void backup_init_bcs_bitmap(BackupBlockJob *job)
 230 {
 231     bool ret;
 232     uint64_t estimate;
 233     BdrvDirtyBitmap *bcs_bitmap = block_copy_dirty_bitmap(job->bcs);
 234
 235     if (job->sync_mode == MIRROR_SYNC_MODE_BITMAP) {
 236         ret = bdrv_dirty_bitmap_merge_internal(bcs_bitmap, job->sync_bitmap,
 237                                                NULL, true);
 238         assert(ret);
 239     } else {
 240         if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
 241             /*
 242              * We can't hog the coroutine to initialize this thoroughly.
 243              * Set a flag and resume work when we are able to yield safely.
 244              */
 245             block_copy_set_skip_unallocated(job->bcs, true);
 246         }
 247         bdrv_set_dirty_bitmap(bcs_bitmap, 0, job->len);
 248     }
 249
 250     estimate = bdrv_get_dirty_count(bcs_bitmap);
 251     job_progress_set_remaining(&job->common.job, estimate);
 252 }
 253
 254 static int coroutine_fn backup_run(Job *job, Error **errp)
 255 {
 256     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 257     int ret;
 258
 259     backup_init_bcs_bitmap(s);
 260
 261     if (s->sync_mode == MIRROR_SYNC_MODE_TOP) {
 262         int64_t offset = 0;
 263         int64_t count;
 264
 265         for (offset = 0; offset < s->len; ) {
 266             if (job_is_cancelled(job)) {
 267                 return -ECANCELED;
 268             }
 269
 270             job_pause_point(job);
 271
 272             if (job_is_cancelled(job)) {
 273                 return -ECANCELED;
 274             }
 275
 276             ret = block_copy_reset_unallocated(s->bcs, offset, &count);
 277             if (ret < 0) {
 278                 return ret;
 279             }
 280
 281             offset += count;
 282         }
 283         block_copy_set_skip_unallocated(s->bcs, false);
 284     }
 285
 286     if (s->sync_mode == MIRROR_SYNC_MODE_NONE) {
 287         /*
 288          * All bits are set in bcs bitmap to allow any cluster to be copied.
 289          * This does not actually require them to be copied.
 290          */
 291         while (!job_is_cancelled(job)) {
 292             /*
 293              * Yield until the job is cancelled.  We just let our before_write
 294              * notify callback service CoW requests.
 295              */
 296             job_yield(job);
 297         }
 298     } else {
 299         return backup_loop(s);
 300     }
 301
 302     return 0;
 303 }
 304
 305 static void coroutine_fn backup_pause(Job *job)
 306 {
 307     BackupBlockJob *s = container_of(job, BackupBlockJob, common.job);
 308
 309     if (s->bg_bcs_call && !block_copy_call_finished(s->bg_bcs_call)) {
 310         block_copy_call_cancel(s->bg_bcs_call);
 311         s->wait = true;
 312         qemu_coroutine_yield();
 313     }
 314 }
 315
 316 static void coroutine_fn backup_set_speed(BlockJob *job, int64_t speed)
 317 {
 318     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 319
 320     /*
 321      * block_job_set_speed() is called first from block_job_create(), when we
 322      * don't yet have s->bcs.
 323      */
 324     if (s->bcs) {
 325         block_copy_set_speed(s->bcs, speed);
 326         if (s->bg_bcs_call) {
 327             block_copy_kick(s->bg_bcs_call);
 328         }
 329     }
 330 }
 331
 332 static const BlockJobDriver backup_job_driver = {
 333     .job_driver = {
 334         .instance_size          = sizeof(BackupBlockJob),
 335         .job_type               = JOB_TYPE_BACKUP,
 336         .free                   = block_job_free,
 337         .user_resume            = block_job_user_resume,
 338         .run                    = backup_run,
 339         .commit                 = backup_commit,
 340         .abort                  = backup_abort,
 341         .clean                  = backup_clean,
 342         .pause                  = backup_pause,
 343     },
 344     .set_speed = backup_set_speed,
 345 };
 346
 347 static int64_t backup_calculate_cluster_size(BlockDriverState *target,
 348                                              Error **errp)
 349 {
 350     int ret;
 351     BlockDriverInfo bdi;
 352     bool target_does_cow = bdrv_backing_chain_next(target);
 353
 354     /*
 355      * If there is no backing file on the target, we cannot rely on COW if our
 356      * backup cluster size is smaller than the target cluster size. Even for
 357      * targets with a backing file, try to avoid COW if possible.
 358      */
 359     ret = bdrv_get_info(target, &bdi);
 360     if (ret == -ENOTSUP && !target_does_cow) {
 361         /* Cluster size is not defined */
 362         warn_report("The target block device doesn't provide "
 363                     "information about the block size and it doesn't have a "
 364                     "backing file. The default block size of %u bytes is "
 365                     "used. If the actual block size of the target exceeds "
 366                     "this default, the backup may be unusable",
 367                     BACKUP_CLUSTER_SIZE_DEFAULT);
 368         return BACKUP_CLUSTER_SIZE_DEFAULT;
 369     } else if (ret < 0 && !target_does_cow) {
 370         error_setg_errno(errp, -ret,
 371             "Couldn't determine the cluster size of the target image, "
 372             "which has no backing file");
 373         error_append_hint(errp,
 374             "Aborting, since this may create an unusable destination image\n");
 375         return ret;
 376     } else if (ret < 0 && target_does_cow) {
 377         /* Not fatal; just trudge on ahead. */
 378         return BACKUP_CLUSTER_SIZE_DEFAULT;
 379     }
 380
 381     return MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
 382 }
 383
 384 BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
 385                   BlockDriverState *target, int64_t speed,
 386                   MirrorSyncMode sync_mode, BdrvDirtyBitmap *sync_bitmap,
 387                   BitmapSyncMode bitmap_mode,
 388                   bool compress,
 389                   const char *filter_node_name,
 390                   BackupPerf *perf,
 391                   BlockdevOnError on_source_error,
 392                   BlockdevOnError on_target_error,
 393                   int creation_flags,
 394                   BlockCompletionFunc *cb, void *opaque,
 395                   JobTxn *txn, Error **errp)
 396 {
 397     int64_t len, target_len;
 398     BackupBlockJob *job = NULL;
 399     int64_t cluster_size;
 400     BdrvRequestFlags write_flags;
 401     BlockDriverState *backup_top = NULL;
 402     BlockCopyState *bcs = NULL;
 403
 404     assert(bs);
 405     assert(target);
 406
 407     /* QMP interface protects us from these cases */
 408     assert(sync_mode != MIRROR_SYNC_MODE_INCREMENTAL);
 409     assert(sync_bitmap || sync_mode != MIRROR_SYNC_MODE_BITMAP);
 410
 411     if (bs == target) {
 412         error_setg(errp, "Source and target cannot be the same");
 413         return NULL;
 414     }
 415
 416     if (!bdrv_is_inserted(bs)) {
 417         error_setg(errp, "Device is not inserted: %s",
 418                    bdrv_get_device_name(bs));
 419         return NULL;
 420     }
 421
 422     if (!bdrv_is_inserted(target)) {
 423         error_setg(errp, "Device is not inserted: %s",
 424                    bdrv_get_device_name(target));
 425         return NULL;
 426     }
 427
 428     if (compress && !bdrv_supports_compressed_writes(target)) {
 429         error_setg(errp, "Compression is not supported for this drive %s",
 430                    bdrv_get_device_name(target));
 431         return NULL;
 432     }
 433
 434     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
 435         return NULL;
 436     }
 437
 438     if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
 439         return NULL;
 440     }
 441
 442     cluster_size = backup_calculate_cluster_size(target, errp);
 443     if (cluster_size < 0) {
 444         goto error;
 445     }
 446
 447     if (perf->max_workers < 1) {
 448         error_setg(errp, "max-workers must be greater than zero");
 449         return NULL;
 450     }
 451
 452     if (perf->max_chunk < 0) {
 453         error_setg(errp, "max-chunk must be zero (which means no limit) or "
 454                    "positive");
 455         return NULL;
 456     }
 457
 458     if (perf->max_chunk && perf->max_chunk < cluster_size) {
 459         error_setg(errp, "Required max-chunk (%" PRIi64 ") is less than backup "
 460                    "cluster size (%" PRIi64 ")", perf->max_chunk, cluster_size);
 461         return NULL;
 462     }
 463
 464
 465     if (sync_bitmap) {
 466         /* If we need to write to this bitmap, check that we can: */
 467         if (bitmap_mode != BITMAP_SYNC_MODE_NEVER &&
 468             bdrv_dirty_bitmap_check(sync_bitmap, BDRV_BITMAP_DEFAULT, errp)) {
 469             return NULL;
 470         }
 471
 472         /* Create a new bitmap, and freeze/disable this one. */
 473         if (bdrv_dirty_bitmap_create_successor(sync_bitmap, errp) < 0) {
 474             return NULL;
 475         }
 476     }
 477
 478     len = bdrv_getlength(bs);
 479     if (len < 0) {
 480         error_setg_errno(errp, -len, "Unable to get length for '%s'",
 481                          bdrv_get_device_or_node_name(bs));
 482         goto error;
 483     }
 484
 485     target_len = bdrv_getlength(target);
 486     if (target_len < 0) {
 487         error_setg_errno(errp, -target_len, "Unable to get length for '%s'",
 488                          bdrv_get_device_or_node_name(bs));
 489         goto error;
 490     }
 491
 492     if (target_len != len) {
 493         error_setg(errp, "Source and target image have different sizes");
 494         goto error;
 495     }
 496
 497     /*
 498      * If source is in backing chain of target assume that target is going to be
 499      * used for "image fleecing", i.e. it should represent a kind of snapshot of
 500      * source at backup-start point in time. And target is going to be read by
 501      * somebody (for example, used as NBD export) during backup job.
 502      *
 503      * In this case, we need to add BDRV_REQ_SERIALISING write flag to avoid
 504      * intersection of backup writes and third party reads from target,
 505      * otherwise reading from target we may occasionally read already updated by
 506      * guest data.
 507      *
 508      * For more information see commit f8d59dfb40bb and test
 509      * tests/qemu-iotests/222
 510      */
 511     write_flags = (bdrv_chain_contains(target, bs) ? BDRV_REQ_SERIALISING : 0) |
 512                   (compress ? BDRV_REQ_WRITE_COMPRESSED : 0),
 513
 514     backup_top = bdrv_backup_top_append(bs, target, filter_node_name,
 515                                         cluster_size, perf,
 516                                         write_flags, &bcs, errp);
 517     if (!backup_top) {
 518         goto error;
 519     }
 520
 521     /* job->len is fixed, so we can't allow resize */
 522     job = block_job_create(job_id, &backup_job_driver, txn, backup_top,
 523                            0, BLK_PERM_ALL,
 524                            speed, creation_flags, cb, opaque, errp);
 525     if (!job) {
 526         goto error;
 527     }
 528
 529     job->backup_top = backup_top;
 530     job->source_bs = bs;
 531     job->on_source_error = on_source_error;
 532     job->on_target_error = on_target_error;
 533     job->sync_mode = sync_mode;
 534     job->sync_bitmap = sync_bitmap;
 535     job->bitmap_mode = bitmap_mode;
 536     job->bcs = bcs;
 537     job->cluster_size = cluster_size;
 538     job->len = len;
 539     job->perf = *perf;
 540
 541     block_copy_set_progress_meter(bcs, &job->common.job.progress);
 542     block_copy_set_speed(bcs, speed);
 543
 544     /* Required permissions are already taken by backup-top target */
 545     block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,
 546                        &error_abort);
 547
 548     return &job->common;
 549
 550  error:
 551     if (sync_bitmap) {
 552         bdrv_reclaim_dirty_bitmap(sync_bitmap, NULL);
 553     }
 554     if (backup_top) {
 555         bdrv_backup_top_drop(backup_top);
 556     }
 557
 558     return NULL;
 559 }