block/backup.c

   1 /*
   2  * QEMU backup
   3  *
   4  * Copyright (C) 2013 Proxmox Server Solutions
   5  *
   6  * Authors:
   7  *  Dietmar Maurer (dietmar@proxmox.com)
   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.h"
  20 #include "qapi/qmp/qerror.h"
  21 #include "qemu/ratelimit.h"
  22 #include "sysemu/block-backend.h"
  23 #include "qemu/bitmap.h"
  24
  25 #define BACKUP_CLUSTER_SIZE_DEFAULT (1 << 16)
  26 #define SLICE_TIME 100000000ULL /* ns */
  27
  28 typedef struct CowRequest {
  29     int64_t start;
  30     int64_t end;
  31     QLIST_ENTRY(CowRequest) list;
  32     CoQueue wait_queue; /* coroutines blocked on this request */
  33 } CowRequest;
  34
  35 typedef struct BackupBlockJob {
  36     BlockJob common;
  37     BlockDriverState *target;
  38     /* bitmap for sync=incremental */
  39     BdrvDirtyBitmap *sync_bitmap;
  40     MirrorSyncMode sync_mode;
  41     RateLimit limit;
  42     BlockdevOnError on_source_error;
  43     BlockdevOnError on_target_error;
  44     CoRwlock flush_rwlock;
  45     uint64_t sectors_read;
  46     unsigned long *done_bitmap;
  47     int64_t cluster_size;
  48     QLIST_HEAD(, CowRequest) inflight_reqs;
  49 } BackupBlockJob;
  50
  51 /* Size of a cluster in sectors, instead of bytes. */
  52 static inline int64_t cluster_size_sectors(BackupBlockJob *job)
  53 {
  54   return job->cluster_size / BDRV_SECTOR_SIZE;
  55 }
  56
  57 /* See if in-flight requests overlap and wait for them to complete */
  58 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
  59                                                        int64_t start,
  60                                                        int64_t end)
  61 {
  62     CowRequest *req;
  63     bool retry;
  64
  65     do {
  66         retry = false;
  67         QLIST_FOREACH(req, &job->inflight_reqs, list) {
  68             if (end > req->start && start < req->end) {
  69                 qemu_co_queue_wait(&req->wait_queue);
  70                 retry = true;
  71                 break;
  72             }
  73         }
  74     } while (retry);
  75 }
  76
  77 /* Keep track of an in-flight request */
  78 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
  79                                      int64_t start, int64_t end)
  80 {
  81     req->start = start;
  82     req->end = end;
  83     qemu_co_queue_init(&req->wait_queue);
  84     QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
  85 }
  86
  87 /* Forget about a completed request */
  88 static void cow_request_end(CowRequest *req)
  89 {
  90     QLIST_REMOVE(req, list);
  91     qemu_co_queue_restart_all(&req->wait_queue);
  92 }
  93
  94 static int coroutine_fn backup_do_cow(BlockDriverState *bs,
  95                                       int64_t sector_num, int nb_sectors,
  96                                       bool *error_is_read,
  97                                       bool is_write_notifier)
  98 {
  99     BackupBlockJob *job = (BackupBlockJob *)bs->job;
 100     CowRequest cow_request;
 101     struct iovec iov;
 102     QEMUIOVector bounce_qiov;
 103     void *bounce_buffer = NULL;
 104     int ret = 0;
 105     int64_t sectors_per_cluster = cluster_size_sectors(job);
 106     int64_t start, end;
 107     int n;
 108
 109     qemu_co_rwlock_rdlock(&job->flush_rwlock);
 110
 111     start = sector_num / sectors_per_cluster;
 112     end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
 113
 114     trace_backup_do_cow_enter(job, start, sector_num, nb_sectors);
 115
 116     wait_for_overlapping_requests(job, start, end);
 117     cow_request_begin(&cow_request, job, start, end);
 118
 119     for (; start < end; start++) {
 120         if (test_bit(start, job->done_bitmap)) {
 121             trace_backup_do_cow_skip(job, start);
 122             continue; /* already copied */
 123         }
 124
 125         trace_backup_do_cow_process(job, start);
 126
 127         n = MIN(sectors_per_cluster,
 128                 job->common.len / BDRV_SECTOR_SIZE -
 129                 start * sectors_per_cluster);
 130
 131         if (!bounce_buffer) {
 132             bounce_buffer = qemu_blockalign(bs, job->cluster_size);
 133         }
 134         iov.iov_base = bounce_buffer;
 135         iov.iov_len = n * BDRV_SECTOR_SIZE;
 136         qemu_iovec_init_external(&bounce_qiov, &iov, 1);
 137
 138         if (is_write_notifier) {
 139             ret = bdrv_co_readv_no_serialising(bs,
 140                                            start * sectors_per_cluster,
 141                                            n, &bounce_qiov);
 142         } else {
 143             ret = bdrv_co_readv(bs, start * sectors_per_cluster, n,
 144                                 &bounce_qiov);
 145         }
 146         if (ret < 0) {
 147             trace_backup_do_cow_read_fail(job, start, ret);
 148             if (error_is_read) {
 149                 *error_is_read = true;
 150             }
 151             goto out;
 152         }
 153
 154         if (buffer_is_zero(iov.iov_base, iov.iov_len)) {
 155             ret = bdrv_co_write_zeroes(job->target,
 156                                        start * sectors_per_cluster,
 157                                        n, BDRV_REQ_MAY_UNMAP);
 158         } else {
 159             ret = bdrv_co_writev(job->target,
 160                                  start * sectors_per_cluster, n,
 161                                  &bounce_qiov);
 162         }
 163         if (ret < 0) {
 164             trace_backup_do_cow_write_fail(job, start, ret);
 165             if (error_is_read) {
 166                 *error_is_read = false;
 167             }
 168             goto out;
 169         }
 170
 171         set_bit(start, job->done_bitmap);
 172
 173         /* Publish progress, guest I/O counts as progress too.  Note that the
 174          * offset field is an opaque progress value, it is not a disk offset.
 175          */
 176         job->sectors_read += n;
 177         job->common.offset += n * BDRV_SECTOR_SIZE;
 178     }
 179
 180 out:
 181     if (bounce_buffer) {
 182         qemu_vfree(bounce_buffer);
 183     }
 184
 185     cow_request_end(&cow_request);
 186
 187     trace_backup_do_cow_return(job, sector_num, nb_sectors, ret);
 188
 189     qemu_co_rwlock_unlock(&job->flush_rwlock);
 190
 191     return ret;
 192 }
 193
 194 static int coroutine_fn backup_before_write_notify(
 195         NotifierWithReturn *notifier,
 196         void *opaque)
 197 {
 198     BdrvTrackedRequest *req = opaque;
 199     int64_t sector_num = req->offset >> BDRV_SECTOR_BITS;
 200     int nb_sectors = req->bytes >> BDRV_SECTOR_BITS;
 201
 202     assert((req->offset & (BDRV_SECTOR_SIZE - 1)) == 0);
 203     assert((req->bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
 204
 205     return backup_do_cow(req->bs, sector_num, nb_sectors, NULL, true);
 206 }
 207
 208 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp)
 209 {
 210     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 211
 212     if (speed < 0) {
 213         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 214         return;
 215     }
 216     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 217 }
 218
 219 static void backup_iostatus_reset(BlockJob *job)
 220 {
 221     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 222
 223     if (s->target->blk) {
 224         blk_iostatus_reset(s->target->blk);
 225     }
 226 }
 227
 228 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
 229 {
 230     BdrvDirtyBitmap *bm;
 231     BlockDriverState *bs = job->common.bs;
 232
 233     if (ret < 0 || block_job_is_cancelled(&job->common)) {
 234         /* Merge the successor back into the parent, delete nothing. */
 235         bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL);
 236         assert(bm);
 237     } else {
 238         /* Everything is fine, delete this bitmap and install the backup. */
 239         bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL);
 240         assert(bm);
 241     }
 242 }
 243
 244 static void backup_commit(BlockJob *job)
 245 {
 246     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 247     if (s->sync_bitmap) {
 248         backup_cleanup_sync_bitmap(s, 0);
 249     }
 250 }
 251
 252 static void backup_abort(BlockJob *job)
 253 {
 254     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 255     if (s->sync_bitmap) {
 256         backup_cleanup_sync_bitmap(s, -1);
 257     }
 258 }
 259
 260 static const BlockJobDriver backup_job_driver = {
 261     .instance_size  = sizeof(BackupBlockJob),
 262     .job_type       = BLOCK_JOB_TYPE_BACKUP,
 263     .set_speed      = backup_set_speed,
 264     .iostatus_reset = backup_iostatus_reset,
 265     .commit         = backup_commit,
 266     .abort          = backup_abort,
 267 };
 268
 269 static BlockErrorAction backup_error_action(BackupBlockJob *job,
 270                                             bool read, int error)
 271 {
 272     if (read) {
 273         return block_job_error_action(&job->common, job->common.bs,
 274                                       job->on_source_error, true, error);
 275     } else {
 276         return block_job_error_action(&job->common, job->target,
 277                                       job->on_target_error, false, error);
 278     }
 279 }
 280
 281 typedef struct {
 282     int ret;
 283 } BackupCompleteData;
 284
 285 static void backup_complete(BlockJob *job, void *opaque)
 286 {
 287     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 288     BackupCompleteData *data = opaque;
 289
 290     bdrv_unref(s->target);
 291
 292     block_job_completed(job, data->ret);
 293     g_free(data);
 294 }
 295
 296 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
 297 {
 298     if (block_job_is_cancelled(&job->common)) {
 299         return true;
 300     }
 301
 302     /* we need to yield so that bdrv_drain_all() returns.
 303      * (without, VM does not reboot)
 304      */
 305     if (job->common.speed) {
 306         uint64_t delay_ns = ratelimit_calculate_delay(&job->limit,
 307                                                       job->sectors_read);
 308         job->sectors_read = 0;
 309         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, delay_ns);
 310     } else {
 311         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, 0);
 312     }
 313
 314     if (block_job_is_cancelled(&job->common)) {
 315         return true;
 316     }
 317
 318     return false;
 319 }
 320
 321 static int coroutine_fn backup_run_incremental(BackupBlockJob *job)
 322 {
 323     bool error_is_read;
 324     int ret = 0;
 325     int clusters_per_iter;
 326     uint32_t granularity;
 327     int64_t sector;
 328     int64_t cluster;
 329     int64_t end;
 330     int64_t last_cluster = -1;
 331     int64_t sectors_per_cluster = cluster_size_sectors(job);
 332     BlockDriverState *bs = job->common.bs;
 333     HBitmapIter hbi;
 334
 335     granularity = bdrv_dirty_bitmap_granularity(job->sync_bitmap);
 336     clusters_per_iter = MAX((granularity / job->cluster_size), 1);
 337     bdrv_dirty_iter_init(job->sync_bitmap, &hbi);
 338
 339     /* Find the next dirty sector(s) */
 340     while ((sector = hbitmap_iter_next(&hbi)) != -1) {
 341         cluster = sector / sectors_per_cluster;
 342
 343         /* Fake progress updates for any clusters we skipped */
 344         if (cluster != last_cluster + 1) {
 345             job->common.offset += ((cluster - last_cluster - 1) *
 346                                    job->cluster_size);
 347         }
 348
 349         for (end = cluster + clusters_per_iter; cluster < end; cluster++) {
 350             do {
 351                 if (yield_and_check(job)) {
 352                     return ret;
 353                 }
 354                 ret = backup_do_cow(bs, cluster * sectors_per_cluster,
 355                                     sectors_per_cluster, &error_is_read,
 356                                     false);
 357                 if ((ret < 0) &&
 358                     backup_error_action(job, error_is_read, -ret) ==
 359                     BLOCK_ERROR_ACTION_REPORT) {
 360                     return ret;
 361                 }
 362             } while (ret < 0);
 363         }
 364
 365         /* If the bitmap granularity is smaller than the backup granularity,
 366          * we need to advance the iterator pointer to the next cluster. */
 367         if (granularity < job->cluster_size) {
 368             bdrv_set_dirty_iter(&hbi, cluster * sectors_per_cluster);
 369         }
 370
 371         last_cluster = cluster - 1;
 372     }
 373
 374     /* Play some final catchup with the progress meter */
 375     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 376     if (last_cluster + 1 < end) {
 377         job->common.offset += ((end - last_cluster - 1) * job->cluster_size);
 378     }
 379
 380     return ret;
 381 }
 382
 383 static void coroutine_fn backup_run(void *opaque)
 384 {
 385     BackupBlockJob *job = opaque;
 386     BackupCompleteData *data;
 387     BlockDriverState *bs = job->common.bs;
 388     BlockDriverState *target = job->target;
 389     BlockdevOnError on_target_error = job->on_target_error;
 390     NotifierWithReturn before_write = {
 391         .notify = backup_before_write_notify,
 392     };
 393     int64_t start, end;
 394     int64_t sectors_per_cluster = cluster_size_sectors(job);
 395     int ret = 0;
 396
 397     QLIST_INIT(&job->inflight_reqs);
 398     qemu_co_rwlock_init(&job->flush_rwlock);
 399
 400     start = 0;
 401     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 402
 403     job->done_bitmap = bitmap_new(end);
 404
 405     bdrv_set_enable_write_cache(target, true);
 406     if (target->blk) {
 407         blk_set_on_error(target->blk, on_target_error, on_target_error);
 408         blk_iostatus_enable(target->blk);
 409     }
 410
 411     bdrv_add_before_write_notifier(bs, &before_write);
 412
 413     if (job->sync_mode == MIRROR_SYNC_MODE_NONE) {
 414         while (!block_job_is_cancelled(&job->common)) {
 415             /* Yield until the job is cancelled.  We just let our before_write
 416              * notify callback service CoW requests. */
 417             job->common.busy = false;
 418             qemu_coroutine_yield();
 419             job->common.busy = true;
 420         }
 421     } else if (job->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 422         ret = backup_run_incremental(job);
 423     } else {
 424         /* Both FULL and TOP SYNC_MODE's require copying.. */
 425         for (; start < end; start++) {
 426             bool error_is_read;
 427             if (yield_and_check(job)) {
 428                 break;
 429             }
 430
 431             if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
 432                 int i, n;
 433                 int alloced = 0;
 434
 435                 /* Check to see if these blocks are already in the
 436                  * backing file. */
 437
 438                 for (i = 0; i < sectors_per_cluster;) {
 439                     /* bdrv_is_allocated() only returns true/false based
 440                      * on the first set of sectors it comes across that
 441                      * are are all in the same state.
 442                      * For that reason we must verify each sector in the
 443                      * backup cluster length.  We end up copying more than
 444                      * needed but at some point that is always the case. */
 445                     alloced =
 446                         bdrv_is_allocated(bs,
 447                                 start * sectors_per_cluster + i,
 448                                 sectors_per_cluster - i, &n);
 449                     i += n;
 450
 451                     if (alloced == 1 || n == 0) {
 452                         break;
 453                     }
 454                 }
 455
 456                 /* If the above loop never found any sectors that are in
 457                  * the topmost image, skip this backup. */
 458                 if (alloced == 0) {
 459                     continue;
 460                 }
 461             }
 462             /* FULL sync mode we copy the whole drive. */
 463             ret = backup_do_cow(bs, start * sectors_per_cluster,
 464                                 sectors_per_cluster, &error_is_read, false);
 465             if (ret < 0) {
 466                 /* Depending on error action, fail now or retry cluster */
 467                 BlockErrorAction action =
 468                     backup_error_action(job, error_is_read, -ret);
 469                 if (action == BLOCK_ERROR_ACTION_REPORT) {
 470                     break;
 471                 } else {
 472                     start--;
 473                     continue;
 474                 }
 475             }
 476         }
 477     }
 478
 479     notifier_with_return_remove(&before_write);
 480
 481     /* wait until pending backup_do_cow() calls have completed */
 482     qemu_co_rwlock_wrlock(&job->flush_rwlock);
 483     qemu_co_rwlock_unlock(&job->flush_rwlock);
 484     g_free(job->done_bitmap);
 485
 486     if (target->blk) {
 487         blk_iostatus_disable(target->blk);
 488     }
 489     bdrv_op_unblock_all(target, job->common.blocker);
 490
 491     data = g_malloc(sizeof(*data));
 492     data->ret = ret;
 493     block_job_defer_to_main_loop(&job->common, backup_complete, data);
 494 }
 495
 496 void backup_start(BlockDriverState *bs, BlockDriverState *target,
 497                   int64_t speed, MirrorSyncMode sync_mode,
 498                   BdrvDirtyBitmap *sync_bitmap,
 499                   BlockdevOnError on_source_error,
 500                   BlockdevOnError on_target_error,
 501                   BlockCompletionFunc *cb, void *opaque,
 502                   BlockJobTxn *txn, Error **errp)
 503 {
 504     int64_t len;
 505     BlockDriverInfo bdi;
 506     int ret;
 507
 508     assert(bs);
 509     assert(target);
 510     assert(cb);
 511
 512     if (bs == target) {
 513         error_setg(errp, "Source and target cannot be the same");
 514         return;
 515     }
 516
 517     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 518          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 519         (!bs->blk || !blk_iostatus_is_enabled(bs->blk))) {
 520         error_setg(errp, QERR_INVALID_PARAMETER, "on-source-error");
 521         return;
 522     }
 523
 524     if (!bdrv_is_inserted(bs)) {
 525         error_setg(errp, "Device is not inserted: %s",
 526                    bdrv_get_device_name(bs));
 527         return;
 528     }
 529
 530     if (!bdrv_is_inserted(target)) {
 531         error_setg(errp, "Device is not inserted: %s",
 532                    bdrv_get_device_name(target));
 533         return;
 534     }
 535
 536     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
 537         return;
 538     }
 539
 540     if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
 541         return;
 542     }
 543
 544     if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 545         if (!sync_bitmap) {
 546             error_setg(errp, "must provide a valid bitmap name for "
 547                              "\"incremental\" sync mode");
 548             return;
 549         }
 550
 551         /* Create a new bitmap, and freeze/disable this one. */
 552         if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
 553             return;
 554         }
 555     } else if (sync_bitmap) {
 556         error_setg(errp,
 557                    "a sync_bitmap was provided to backup_run, "
 558                    "but received an incompatible sync_mode (%s)",
 559                    MirrorSyncMode_lookup[sync_mode]);
 560         return;
 561     }
 562
 563     len = bdrv_getlength(bs);
 564     if (len < 0) {
 565         error_setg_errno(errp, -len, "unable to get length for '%s'",
 566                          bdrv_get_device_name(bs));
 567         goto error;
 568     }
 569
 570     BackupBlockJob *job = block_job_create(&backup_job_driver, bs, speed,
 571                                            cb, opaque, errp);
 572     if (!job) {
 573         goto error;
 574     }
 575
 576     job->on_source_error = on_source_error;
 577     job->on_target_error = on_target_error;
 578     job->target = target;
 579     job->sync_mode = sync_mode;
 580     job->sync_bitmap = sync_mode == MIRROR_SYNC_MODE_INCREMENTAL ?
 581                        sync_bitmap : NULL;
 582
 583     /* If there is no backing file on the target, we cannot rely on COW if our
 584      * backup cluster size is smaller than the target cluster size. Even for
 585      * targets with a backing file, try to avoid COW if possible. */
 586     ret = bdrv_get_info(job->target, &bdi);
 587     if (ret < 0 && !target->backing) {
 588         error_setg_errno(errp, -ret,
 589             "Couldn't determine the cluster size of the target image, "
 590             "which has no backing file");
 591         error_append_hint(errp,
 592             "Aborting, since this may create an unusable destination image\n");
 593         goto error;
 594     } else if (ret < 0 && target->backing) {
 595         /* Not fatal; just trudge on ahead. */
 596         job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
 597     } else {
 598         job->cluster_size = MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
 599     }
 600
 601     bdrv_op_block_all(target, job->common.blocker);
 602     job->common.len = len;
 603     job->common.co = qemu_coroutine_create(backup_run);
 604     block_job_txn_add_job(txn, &job->common);
 605     qemu_coroutine_enter(job->common.co, job);
 606     return;
 607
 608  error:
 609     if (sync_bitmap) {
 610         bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL);
 611     }
 612 }