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