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