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/error.h"
  21 #include "qapi/qmp/qerror.h"
  22 #include "qemu/ratelimit.h"
  23 #include "qemu/cutils.h"
  24 #include "sysemu/block-backend.h"
  25 #include "qemu/bitmap.h"
  26
  27 #define BACKUP_CLUSTER_SIZE_DEFAULT (1 << 16)
  28 #define SLICE_TIME 100000000ULL /* ns */
  29
  30 typedef struct CowRequest {
  31     int64_t start;
  32     int64_t end;
  33     QLIST_ENTRY(CowRequest) list;
  34     CoQueue wait_queue; /* coroutines blocked on this request */
  35 } CowRequest;
  36
  37 typedef struct BackupBlockJob {
  38     BlockJob common;
  39     BlockDriverState *target;
  40     /* bitmap for sync=incremental */
  41     BdrvDirtyBitmap *sync_bitmap;
  42     MirrorSyncMode sync_mode;
  43     RateLimit limit;
  44     BlockdevOnError on_source_error;
  45     BlockdevOnError on_target_error;
  46     CoRwlock flush_rwlock;
  47     uint64_t sectors_read;
  48     unsigned long *done_bitmap;
  49     int64_t cluster_size;
  50     QLIST_HEAD(, CowRequest) inflight_reqs;
  51 } BackupBlockJob;
  52
  53 /* Size of a cluster in sectors, instead of bytes. */
  54 static inline int64_t cluster_size_sectors(BackupBlockJob *job)
  55 {
  56   return job->cluster_size / BDRV_SECTOR_SIZE;
  57 }
  58
  59 /* See if in-flight requests overlap and wait for them to complete */
  60 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
  61                                                        int64_t start,
  62                                                        int64_t end)
  63 {
  64     CowRequest *req;
  65     bool retry;
  66
  67     do {
  68         retry = false;
  69         QLIST_FOREACH(req, &job->inflight_reqs, list) {
  70             if (end > req->start && start < req->end) {
  71                 qemu_co_queue_wait(&req->wait_queue);
  72                 retry = true;
  73                 break;
  74             }
  75         }
  76     } while (retry);
  77 }
  78
  79 /* Keep track of an in-flight request */
  80 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
  81                                      int64_t start, int64_t end)
  82 {
  83     req->start = start;
  84     req->end = end;
  85     qemu_co_queue_init(&req->wait_queue);
  86     QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
  87 }
  88
  89 /* Forget about a completed request */
  90 static void cow_request_end(CowRequest *req)
  91 {
  92     QLIST_REMOVE(req, list);
  93     qemu_co_queue_restart_all(&req->wait_queue);
  94 }
  95
  96 static int coroutine_fn backup_do_cow(BlockDriverState *bs,
  97                                       int64_t sector_num, int nb_sectors,
  98                                       bool *error_is_read,
  99                                       bool is_write_notifier)
 100 {
 101     BackupBlockJob *job = (BackupBlockJob *)bs->job;
 102     CowRequest cow_request;
 103     struct iovec iov;
 104     QEMUIOVector bounce_qiov;
 105     void *bounce_buffer = NULL;
 106     int ret = 0;
 107     int64_t sectors_per_cluster = cluster_size_sectors(job);
 108     int64_t start, end;
 109     int n;
 110
 111     qemu_co_rwlock_rdlock(&job->flush_rwlock);
 112
 113     start = sector_num / sectors_per_cluster;
 114     end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
 115
 116     trace_backup_do_cow_enter(job, start, sector_num, nb_sectors);
 117
 118     wait_for_overlapping_requests(job, start, end);
 119     cow_request_begin(&cow_request, job, start, end);
 120
 121     for (; start < end; start++) {
 122         if (test_bit(start, job->done_bitmap)) {
 123             trace_backup_do_cow_skip(job, start);
 124             continue; /* already copied */
 125         }
 126
 127         trace_backup_do_cow_process(job, start);
 128
 129         n = MIN(sectors_per_cluster,
 130                 job->common.len / BDRV_SECTOR_SIZE -
 131                 start * sectors_per_cluster);
 132
 133         if (!bounce_buffer) {
 134             bounce_buffer = qemu_blockalign(bs, job->cluster_size);
 135         }
 136         iov.iov_base = bounce_buffer;
 137         iov.iov_len = n * BDRV_SECTOR_SIZE;
 138         qemu_iovec_init_external(&bounce_qiov, &iov, 1);
 139
 140         if (is_write_notifier) {
 141             ret = bdrv_co_readv_no_serialising(bs,
 142                                            start * sectors_per_cluster,
 143                                            n, &bounce_qiov);
 144         } else {
 145             ret = bdrv_co_readv(bs, start * sectors_per_cluster, n,
 146                                 &bounce_qiov);
 147         }
 148         if (ret < 0) {
 149             trace_backup_do_cow_read_fail(job, start, ret);
 150             if (error_is_read) {
 151                 *error_is_read = true;
 152             }
 153             goto out;
 154         }
 155
 156         if (buffer_is_zero(iov.iov_base, iov.iov_len)) {
 157             ret = bdrv_co_write_zeroes(job->target,
 158                                        start * sectors_per_cluster,
 159                                        n, BDRV_REQ_MAY_UNMAP);
 160         } else {
 161             ret = bdrv_co_writev(job->target,
 162                                  start * sectors_per_cluster, n,
 163                                  &bounce_qiov);
 164         }
 165         if (ret < 0) {
 166             trace_backup_do_cow_write_fail(job, start, ret);
 167             if (error_is_read) {
 168                 *error_is_read = false;
 169             }
 170             goto out;
 171         }
 172
 173         set_bit(start, job->done_bitmap);
 174
 175         /* Publish progress, guest I/O counts as progress too.  Note that the
 176          * offset field is an opaque progress value, it is not a disk offset.
 177          */
 178         job->sectors_read += n;
 179         job->common.offset += n * BDRV_SECTOR_SIZE;
 180     }
 181
 182 out:
 183     if (bounce_buffer) {
 184         qemu_vfree(bounce_buffer);
 185     }
 186
 187     cow_request_end(&cow_request);
 188
 189     trace_backup_do_cow_return(job, sector_num, nb_sectors, ret);
 190
 191     qemu_co_rwlock_unlock(&job->flush_rwlock);
 192
 193     return ret;
 194 }
 195
 196 static int coroutine_fn backup_before_write_notify(
 197         NotifierWithReturn *notifier,
 198         void *opaque)
 199 {
 200     BdrvTrackedRequest *req = opaque;
 201     int64_t sector_num = req->offset >> BDRV_SECTOR_BITS;
 202     int nb_sectors = req->bytes >> BDRV_SECTOR_BITS;
 203
 204     assert((req->offset & (BDRV_SECTOR_SIZE - 1)) == 0);
 205     assert((req->bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
 206
 207     return backup_do_cow(req->bs, sector_num, nb_sectors, NULL, true);
 208 }
 209
 210 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp)
 211 {
 212     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 213
 214     if (speed < 0) {
 215         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 216         return;
 217     }
 218     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 219 }
 220
 221 static void backup_iostatus_reset(BlockJob *job)
 222 {
 223     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 224
 225     if (s->target->blk) {
 226         blk_iostatus_reset(s->target->blk);
 227     }
 228 }
 229
 230 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
 231 {
 232     BdrvDirtyBitmap *bm;
 233     BlockDriverState *bs = job->common.bs;
 234
 235     if (ret < 0 || block_job_is_cancelled(&job->common)) {
 236         /* Merge the successor back into the parent, delete nothing. */
 237         bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL);
 238         assert(bm);
 239     } else {
 240         /* Everything is fine, delete this bitmap and install the backup. */
 241         bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL);
 242         assert(bm);
 243     }
 244 }
 245
 246 static void backup_commit(BlockJob *job)
 247 {
 248     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 249     if (s->sync_bitmap) {
 250         backup_cleanup_sync_bitmap(s, 0);
 251     }
 252 }
 253
 254 static void backup_abort(BlockJob *job)
 255 {
 256     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 257     if (s->sync_bitmap) {
 258         backup_cleanup_sync_bitmap(s, -1);
 259     }
 260 }
 261
 262 static const BlockJobDriver backup_job_driver = {
 263     .instance_size  = sizeof(BackupBlockJob),
 264     .job_type       = BLOCK_JOB_TYPE_BACKUP,
 265     .set_speed      = backup_set_speed,
 266     .iostatus_reset = backup_iostatus_reset,
 267     .commit         = backup_commit,
 268     .abort          = backup_abort,
 269 };
 270
 271 static BlockErrorAction backup_error_action(BackupBlockJob *job,
 272                                             bool read, int error)
 273 {
 274     if (read) {
 275         return block_job_error_action(&job->common, job->common.bs,
 276                                       job->on_source_error, true, error);
 277     } else {
 278         return block_job_error_action(&job->common, job->target,
 279                                       job->on_target_error, false, error);
 280     }
 281 }
 282
 283 typedef struct {
 284     int ret;
 285 } BackupCompleteData;
 286
 287 static void backup_complete(BlockJob *job, void *opaque)
 288 {
 289     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 290     BackupCompleteData *data = opaque;
 291
 292     bdrv_unref(s->target);
 293
 294     block_job_completed(job, data->ret);
 295     g_free(data);
 296 }
 297
 298 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
 299 {
 300     if (block_job_is_cancelled(&job->common)) {
 301         return true;
 302     }
 303
 304     /* we need to yield so that bdrv_drain_all() returns.
 305      * (without, VM does not reboot)
 306      */
 307     if (job->common.speed) {
 308         uint64_t delay_ns = ratelimit_calculate_delay(&job->limit,
 309                                                       job->sectors_read);
 310         job->sectors_read = 0;
 311         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, delay_ns);
 312     } else {
 313         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, 0);
 314     }
 315
 316     if (block_job_is_cancelled(&job->common)) {
 317         return true;
 318     }
 319
 320     return false;
 321 }
 322
 323 static int coroutine_fn backup_run_incremental(BackupBlockJob *job)
 324 {
 325     bool error_is_read;
 326     int ret = 0;
 327     int clusters_per_iter;
 328     uint32_t granularity;
 329     int64_t sector;
 330     int64_t cluster;
 331     int64_t end;
 332     int64_t last_cluster = -1;
 333     int64_t sectors_per_cluster = cluster_size_sectors(job);
 334     BlockDriverState *bs = job->common.bs;
 335     HBitmapIter hbi;
 336
 337     granularity = bdrv_dirty_bitmap_granularity(job->sync_bitmap);
 338     clusters_per_iter = MAX((granularity / job->cluster_size), 1);
 339     bdrv_dirty_iter_init(job->sync_bitmap, &hbi);
 340
 341     /* Find the next dirty sector(s) */
 342     while ((sector = hbitmap_iter_next(&hbi)) != -1) {
 343         cluster = sector / sectors_per_cluster;
 344
 345         /* Fake progress updates for any clusters we skipped */
 346         if (cluster != last_cluster + 1) {
 347             job->common.offset += ((cluster - last_cluster - 1) *
 348                                    job->cluster_size);
 349         }
 350
 351         for (end = cluster + clusters_per_iter; cluster < end; cluster++) {
 352             do {
 353                 if (yield_and_check(job)) {
 354                     return ret;
 355                 }
 356                 ret = backup_do_cow(bs, cluster * sectors_per_cluster,
 357                                     sectors_per_cluster, &error_is_read,
 358                                     false);
 359                 if ((ret < 0) &&
 360                     backup_error_action(job, error_is_read, -ret) ==
 361                     BLOCK_ERROR_ACTION_REPORT) {
 362                     return ret;
 363                 }
 364             } while (ret < 0);
 365         }
 366
 367         /* If the bitmap granularity is smaller than the backup granularity,
 368          * we need to advance the iterator pointer to the next cluster. */
 369         if (granularity < job->cluster_size) {
 370             bdrv_set_dirty_iter(&hbi, cluster * sectors_per_cluster);
 371         }
 372
 373         last_cluster = cluster - 1;
 374     }
 375
 376     /* Play some final catchup with the progress meter */
 377     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 378     if (last_cluster + 1 < end) {
 379         job->common.offset += ((end - last_cluster - 1) * job->cluster_size);
 380     }
 381
 382     return ret;
 383 }
 384
 385 static void coroutine_fn backup_run(void *opaque)
 386 {
 387     BackupBlockJob *job = opaque;
 388     BackupCompleteData *data;
 389     BlockDriverState *bs = job->common.bs;
 390     BlockDriverState *target = job->target;
 391     BlockdevOnError on_target_error = job->on_target_error;
 392     NotifierWithReturn before_write = {
 393         .notify = backup_before_write_notify,
 394     };
 395     int64_t start, end;
 396     int64_t sectors_per_cluster = cluster_size_sectors(job);
 397     int ret = 0;
 398
 399     QLIST_INIT(&job->inflight_reqs);
 400     qemu_co_rwlock_init(&job->flush_rwlock);
 401
 402     start = 0;
 403     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 404
 405     job->done_bitmap = bitmap_new(end);
 406
 407     if (target->blk) {
 408         blk_set_on_error(target->blk, on_target_error, on_target_error);
 409         blk_iostatus_enable(target->blk);
 410     }
 411
 412     bdrv_add_before_write_notifier(bs, &before_write);
 413
 414     if (job->sync_mode == MIRROR_SYNC_MODE_NONE) {
 415         while (!block_job_is_cancelled(&job->common)) {
 416             /* Yield until the job is cancelled.  We just let our before_write
 417              * notify callback service CoW requests. */
 418             job->common.busy = false;
 419             qemu_coroutine_yield();
 420             job->common.busy = true;
 421         }
 422     } else if (job->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 423         ret = backup_run_incremental(job);
 424     } else {
 425         /* Both FULL and TOP SYNC_MODE's require copying.. */
 426         for (; start < end; start++) {
 427             bool error_is_read;
 428             if (yield_and_check(job)) {
 429                 break;
 430             }
 431
 432             if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
 433                 int i, n;
 434                 int alloced = 0;
 435
 436                 /* Check to see if these blocks are already in the
 437                  * backing file. */
 438
 439                 for (i = 0; i < sectors_per_cluster;) {
 440                     /* bdrv_is_allocated() only returns true/false based
 441                      * on the first set of sectors it comes across that
 442                      * are are all in the same state.
 443                      * For that reason we must verify each sector in the
 444                      * backup cluster length.  We end up copying more than
 445                      * needed but at some point that is always the case. */
 446                     alloced =
 447                         bdrv_is_allocated(bs,
 448                                 start * sectors_per_cluster + i,
 449                                 sectors_per_cluster - i, &n);
 450                     i += n;
 451
 452                     if (alloced == 1 || n == 0) {
 453                         break;
 454                     }
 455                 }
 456
 457                 /* If the above loop never found any sectors that are in
 458                  * the topmost image, skip this backup. */
 459                 if (alloced == 0) {
 460                     continue;
 461                 }
 462             }
 463             /* FULL sync mode we copy the whole drive. */
 464             ret = backup_do_cow(bs, start * sectors_per_cluster,
 465                                 sectors_per_cluster, &error_is_read, false);
 466             if (ret < 0) {
 467                 /* Depending on error action, fail now or retry cluster */
 468                 BlockErrorAction action =
 469                     backup_error_action(job, error_is_read, -ret);
 470                 if (action == BLOCK_ERROR_ACTION_REPORT) {
 471                     break;
 472                 } else {
 473                     start--;
 474                     continue;
 475                 }
 476             }
 477         }
 478     }
 479
 480     notifier_with_return_remove(&before_write);
 481
 482     /* wait until pending backup_do_cow() calls have completed */
 483     qemu_co_rwlock_wrlock(&job->flush_rwlock);
 484     qemu_co_rwlock_unlock(&job->flush_rwlock);
 485     g_free(job->done_bitmap);
 486
 487     if (target->blk) {
 488         blk_iostatus_disable(target->blk);
 489     }
 490     bdrv_op_unblock_all(target, job->common.blocker);
 491
 492     data = g_malloc(sizeof(*data));
 493     data->ret = ret;
 494     block_job_defer_to_main_loop(&job->common, backup_complete, data);
 495 }
 496
 497 void backup_start(BlockDriverState *bs, BlockDriverState *target,
 498                   int64_t speed, MirrorSyncMode sync_mode,
 499                   BdrvDirtyBitmap *sync_bitmap,
 500                   BlockdevOnError on_source_error,
 501                   BlockdevOnError on_target_error,
 502                   BlockCompletionFunc *cb, void *opaque,
 503                   BlockJobTxn *txn, Error **errp)
 504 {
 505     int64_t len;
 506     BlockDriverInfo bdi;
 507     int ret;
 508
 509     assert(bs);
 510     assert(target);
 511     assert(cb);
 512
 513     if (bs == target) {
 514         error_setg(errp, "Source and target cannot be the same");
 515         return;
 516     }
 517
 518     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 519          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 520         (!bs->blk || !blk_iostatus_is_enabled(bs->blk))) {
 521         error_setg(errp, QERR_INVALID_PARAMETER, "on-source-error");
 522         return;
 523     }
 524
 525     if (!bdrv_is_inserted(bs)) {
 526         error_setg(errp, "Device is not inserted: %s",
 527                    bdrv_get_device_name(bs));
 528         return;
 529     }
 530
 531     if (!bdrv_is_inserted(target)) {
 532         error_setg(errp, "Device is not inserted: %s",
 533                    bdrv_get_device_name(target));
 534         return;
 535     }
 536
 537     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
 538         return;
 539     }
 540
 541     if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
 542         return;
 543     }
 544
 545     if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 546         if (!sync_bitmap) {
 547             error_setg(errp, "must provide a valid bitmap name for "
 548                              "\"incremental\" sync mode");
 549             return;
 550         }
 551
 552         /* Create a new bitmap, and freeze/disable this one. */
 553         if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
 554             return;
 555         }
 556     } else if (sync_bitmap) {
 557         error_setg(errp,
 558                    "a sync_bitmap was provided to backup_run, "
 559                    "but received an incompatible sync_mode (%s)",
 560                    MirrorSyncMode_lookup[sync_mode]);
 561         return;
 562     }
 563
 564     len = bdrv_getlength(bs);
 565     if (len < 0) {
 566         error_setg_errno(errp, -len, "unable to get length for '%s'",
 567                          bdrv_get_device_name(bs));
 568         goto error;
 569     }
 570
 571     BackupBlockJob *job = block_job_create(&backup_job_driver, bs, speed,
 572                                            cb, opaque, errp);
 573     if (!job) {
 574         goto error;
 575     }
 576
 577     job->on_source_error = on_source_error;
 578     job->on_target_error = on_target_error;
 579     job->target = target;
 580     job->sync_mode = sync_mode;
 581     job->sync_bitmap = sync_mode == MIRROR_SYNC_MODE_INCREMENTAL ?
 582                        sync_bitmap : NULL;
 583
 584     /* If there is no backing file on the target, we cannot rely on COW if our
 585      * backup cluster size is smaller than the target cluster size. Even for
 586      * targets with a backing file, try to avoid COW if possible. */
 587     ret = bdrv_get_info(job->target, &bdi);
 588     if (ret < 0 && !target->backing) {
 589         error_setg_errno(errp, -ret,
 590             "Couldn't determine the cluster size of the target image, "
 591             "which has no backing file");
 592         error_append_hint(errp,
 593             "Aborting, since this may create an unusable destination image\n");
 594         goto error;
 595     } else if (ret < 0 && target->backing) {
 596         /* Not fatal; just trudge on ahead. */
 597         job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
 598     } else {
 599         job->cluster_size = MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
 600     }
 601
 602     bdrv_op_block_all(target, job->common.blocker);
 603     job->common.len = len;
 604     job->common.co = qemu_coroutine_create(backup_run);
 605     block_job_txn_add_job(txn, &job->common);
 606     qemu_coroutine_enter(job->common.co, job);
 607     return;
 608
 609  error:
 610     if (sync_bitmap) {
 611         bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL);
 612     }
 613 }