block/backup.c

   1 /*
   2  * QEMU backup
   3  *
   4  * Copyright (C) 2013 Proxmox Server Solutions
   5  *
   6  * Authors:
   7  *  Dietmar Maurer ([email protected])
   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 "block/block_backup.h"
  21 #include "qapi/error.h"
  22 #include "qapi/qmp/qerror.h"
  23 #include "qemu/ratelimit.h"
  24 #include "qemu/cutils.h"
  25 #include "sysemu/block-backend.h"
  26 #include "qemu/bitmap.h"
  27
  28 #define BACKUP_CLUSTER_SIZE_DEFAULT (1 << 16)
  29 #define SLICE_TIME 100000000ULL /* ns */
  30
  31 typedef struct BackupBlockJob {
  32     BlockJob common;
  33     BlockBackend *target;
  34     /* bitmap for sync=incremental */
  35     BdrvDirtyBitmap *sync_bitmap;
  36     MirrorSyncMode sync_mode;
  37     RateLimit limit;
  38     BlockdevOnError on_source_error;
  39     BlockdevOnError on_target_error;
  40     CoRwlock flush_rwlock;
  41     uint64_t sectors_read;
  42     unsigned long *done_bitmap;
  43     int64_t cluster_size;
  44     bool compress;
  45     NotifierWithReturn before_write;
  46     QLIST_HEAD(, CowRequest) inflight_reqs;
  47 } BackupBlockJob;
  48
  49 /* Size of a cluster in sectors, instead of bytes. */
  50 static inline int64_t cluster_size_sectors(BackupBlockJob *job)
  51 {
  52   return job->cluster_size / BDRV_SECTOR_SIZE;
  53 }
  54
  55 /* See if in-flight requests overlap and wait for them to complete */
  56 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
  57                                                        int64_t start,
  58                                                        int64_t end)
  59 {
  60     CowRequest *req;
  61     bool retry;
  62
  63     do {
  64         retry = false;
  65         QLIST_FOREACH(req, &job->inflight_reqs, list) {
  66             if (end > req->start && start < req->end) {
  67                 qemu_co_queue_wait(&req->wait_queue);
  68                 retry = true;
  69                 break;
  70             }
  71         }
  72     } while (retry);
  73 }
  74
  75 /* Keep track of an in-flight request */
  76 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
  77                                      int64_t start, int64_t end)
  78 {
  79     req->start = start;
  80     req->end = end;
  81     qemu_co_queue_init(&req->wait_queue);
  82     QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
  83 }
  84
  85 /* Forget about a completed request */
  86 static void cow_request_end(CowRequest *req)
  87 {
  88     QLIST_REMOVE(req, list);
  89     qemu_co_queue_restart_all(&req->wait_queue);
  90 }
  91
  92 static int coroutine_fn backup_do_cow(BackupBlockJob *job,
  93                                       int64_t sector_num, int nb_sectors,
  94                                       bool *error_is_read,
  95                                       bool is_write_notifier)
  96 {
  97     BlockBackend *blk = job->common.blk;
  98     CowRequest cow_request;
  99     struct iovec iov;
 100     QEMUIOVector bounce_qiov;
 101     void *bounce_buffer = NULL;
 102     int ret = 0;
 103     int64_t sectors_per_cluster = cluster_size_sectors(job);
 104     int64_t start, end;
 105     int n;
 106
 107     qemu_co_rwlock_rdlock(&job->flush_rwlock);
 108
 109     start = sector_num / sectors_per_cluster;
 110     end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
 111
 112     trace_backup_do_cow_enter(job, start, sector_num, nb_sectors);
 113
 114     wait_for_overlapping_requests(job, start, end);
 115     cow_request_begin(&cow_request, job, start, end);
 116
 117     for (; start < end; start++) {
 118         if (test_bit(start, job->done_bitmap)) {
 119             trace_backup_do_cow_skip(job, start);
 120             continue; /* already copied */
 121         }
 122
 123         trace_backup_do_cow_process(job, start);
 124
 125         n = MIN(sectors_per_cluster,
 126                 job->common.len / BDRV_SECTOR_SIZE -
 127                 start * sectors_per_cluster);
 128
 129         if (!bounce_buffer) {
 130             bounce_buffer = blk_blockalign(blk, job->cluster_size);
 131         }
 132         iov.iov_base = bounce_buffer;
 133         iov.iov_len = n * BDRV_SECTOR_SIZE;
 134         qemu_iovec_init_external(&bounce_qiov, &iov, 1);
 135
 136         ret = blk_co_preadv(blk, start * job->cluster_size,
 137                             bounce_qiov.size, &bounce_qiov,
 138                             is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0);
 139         if (ret < 0) {
 140             trace_backup_do_cow_read_fail(job, start, ret);
 141             if (error_is_read) {
 142                 *error_is_read = true;
 143             }
 144             goto out;
 145         }
 146
 147         if (buffer_is_zero(iov.iov_base, iov.iov_len)) {
 148             ret = blk_co_pwrite_zeroes(job->target, start * job->cluster_size,
 149                                        bounce_qiov.size, BDRV_REQ_MAY_UNMAP);
 150         } else {
 151             ret = blk_co_pwritev(job->target, start * job->cluster_size,
 152                                  bounce_qiov.size, &bounce_qiov,
 153                                  job->compress ? BDRV_REQ_WRITE_COMPRESSED : 0);
 154         }
 155         if (ret < 0) {
 156             trace_backup_do_cow_write_fail(job, start, ret);
 157             if (error_is_read) {
 158                 *error_is_read = false;
 159             }
 160             goto out;
 161         }
 162
 163         set_bit(start, job->done_bitmap);
 164
 165         /* Publish progress, guest I/O counts as progress too.  Note that the
 166          * offset field is an opaque progress value, it is not a disk offset.
 167          */
 168         job->sectors_read += n;
 169         job->common.offset += n * BDRV_SECTOR_SIZE;
 170     }
 171
 172 out:
 173     if (bounce_buffer) {
 174         qemu_vfree(bounce_buffer);
 175     }
 176
 177     cow_request_end(&cow_request);
 178
 179     trace_backup_do_cow_return(job, sector_num, nb_sectors, ret);
 180
 181     qemu_co_rwlock_unlock(&job->flush_rwlock);
 182
 183     return ret;
 184 }
 185
 186 static int coroutine_fn backup_before_write_notify(
 187         NotifierWithReturn *notifier,
 188         void *opaque)
 189 {
 190     BackupBlockJob *job = container_of(notifier, BackupBlockJob, before_write);
 191     BdrvTrackedRequest *req = opaque;
 192     int64_t sector_num = req->offset >> BDRV_SECTOR_BITS;
 193     int nb_sectors = req->bytes >> BDRV_SECTOR_BITS;
 194
 195     assert(req->bs == blk_bs(job->common.blk));
 196     assert((req->offset & (BDRV_SECTOR_SIZE - 1)) == 0);
 197     assert((req->bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
 198
 199     return backup_do_cow(job, sector_num, nb_sectors, NULL, true);
 200 }
 201
 202 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp)
 203 {
 204     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 205
 206     if (speed < 0) {
 207         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 208         return;
 209     }
 210     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 211 }
 212
 213 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
 214 {
 215     BdrvDirtyBitmap *bm;
 216     BlockDriverState *bs = blk_bs(job->common.blk);
 217
 218     if (ret < 0 || block_job_is_cancelled(&job->common)) {
 219         /* Merge the successor back into the parent, delete nothing. */
 220         bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL);
 221         assert(bm);
 222     } else {
 223         /* Everything is fine, delete this bitmap and install the backup. */
 224         bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL);
 225         assert(bm);
 226     }
 227 }
 228
 229 static void backup_commit(BlockJob *job)
 230 {
 231     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 232     if (s->sync_bitmap) {
 233         backup_cleanup_sync_bitmap(s, 0);
 234     }
 235 }
 236
 237 static void backup_abort(BlockJob *job)
 238 {
 239     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 240     if (s->sync_bitmap) {
 241         backup_cleanup_sync_bitmap(s, -1);
 242     }
 243 }
 244
 245 static void backup_attached_aio_context(BlockJob *job, AioContext *aio_context)
 246 {
 247     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 248
 249     blk_set_aio_context(s->target, aio_context);
 250 }
 251
 252 void backup_do_checkpoint(BlockJob *job, Error **errp)
 253 {
 254     BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
 255     int64_t len;
 256
 257     assert(job->driver->job_type == BLOCK_JOB_TYPE_BACKUP);
 258
 259     if (backup_job->sync_mode != MIRROR_SYNC_MODE_NONE) {
 260         error_setg(errp, "The backup job only supports block checkpoint in"
 261                    " sync=none mode");
 262         return;
 263     }
 264
 265     len = DIV_ROUND_UP(backup_job->common.len, backup_job->cluster_size);
 266     bitmap_zero(backup_job->done_bitmap, len);
 267 }
 268
 269 void backup_wait_for_overlapping_requests(BlockJob *job, int64_t sector_num,
 270                                           int nb_sectors)
 271 {
 272     BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
 273     int64_t sectors_per_cluster = cluster_size_sectors(backup_job);
 274     int64_t start, end;
 275
 276     assert(job->driver->job_type == BLOCK_JOB_TYPE_BACKUP);
 277
 278     start = sector_num / sectors_per_cluster;
 279     end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
 280     wait_for_overlapping_requests(backup_job, start, end);
 281 }
 282
 283 void backup_cow_request_begin(CowRequest *req, BlockJob *job,
 284                               int64_t sector_num,
 285                               int nb_sectors)
 286 {
 287     BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
 288     int64_t sectors_per_cluster = cluster_size_sectors(backup_job);
 289     int64_t start, end;
 290
 291     assert(job->driver->job_type == BLOCK_JOB_TYPE_BACKUP);
 292
 293     start = sector_num / sectors_per_cluster;
 294     end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
 295     cow_request_begin(req, backup_job, start, end);
 296 }
 297
 298 void backup_cow_request_end(CowRequest *req)
 299 {
 300     cow_request_end(req);
 301 }
 302
 303 static const BlockJobDriver backup_job_driver = {
 304     .instance_size          = sizeof(BackupBlockJob),
 305     .job_type               = BLOCK_JOB_TYPE_BACKUP,
 306     .set_speed              = backup_set_speed,
 307     .commit                 = backup_commit,
 308     .abort                  = backup_abort,
 309     .attached_aio_context   = backup_attached_aio_context,
 310 };
 311
 312 static BlockErrorAction backup_error_action(BackupBlockJob *job,
 313                                             bool read, int error)
 314 {
 315     if (read) {
 316         return block_job_error_action(&job->common, job->on_source_error,
 317                                       true, error);
 318     } else {
 319         return block_job_error_action(&job->common, job->on_target_error,
 320                                       false, error);
 321     }
 322 }
 323
 324 typedef struct {
 325     int ret;
 326 } BackupCompleteData;
 327
 328 static void backup_complete(BlockJob *job, void *opaque)
 329 {
 330     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 331     BackupCompleteData *data = opaque;
 332
 333     blk_unref(s->target);
 334
 335     block_job_completed(job, data->ret);
 336     g_free(data);
 337 }
 338
 339 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
 340 {
 341     if (block_job_is_cancelled(&job->common)) {
 342         return true;
 343     }
 344
 345     /* we need to yield so that bdrv_drain_all() returns.
 346      * (without, VM does not reboot)
 347      */
 348     if (job->common.speed) {
 349         uint64_t delay_ns = ratelimit_calculate_delay(&job->limit,
 350                                                       job->sectors_read);
 351         job->sectors_read = 0;
 352         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, delay_ns);
 353     } else {
 354         block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, 0);
 355     }
 356
 357     if (block_job_is_cancelled(&job->common)) {
 358         return true;
 359     }
 360
 361     return false;
 362 }
 363
 364 static int coroutine_fn backup_run_incremental(BackupBlockJob *job)
 365 {
 366     bool error_is_read;
 367     int ret = 0;
 368     int clusters_per_iter;
 369     uint32_t granularity;
 370     int64_t sector;
 371     int64_t cluster;
 372     int64_t end;
 373     int64_t last_cluster = -1;
 374     int64_t sectors_per_cluster = cluster_size_sectors(job);
 375     HBitmapIter hbi;
 376
 377     granularity = bdrv_dirty_bitmap_granularity(job->sync_bitmap);
 378     clusters_per_iter = MAX((granularity / job->cluster_size), 1);
 379     bdrv_dirty_iter_init(job->sync_bitmap, &hbi);
 380
 381     /* Find the next dirty sector(s) */
 382     while ((sector = hbitmap_iter_next(&hbi)) != -1) {
 383         cluster = sector / sectors_per_cluster;
 384
 385         /* Fake progress updates for any clusters we skipped */
 386         if (cluster != last_cluster + 1) {
 387             job->common.offset += ((cluster - last_cluster - 1) *
 388                                    job->cluster_size);
 389         }
 390
 391         for (end = cluster + clusters_per_iter; cluster < end; cluster++) {
 392             do {
 393                 if (yield_and_check(job)) {
 394                     return ret;
 395                 }
 396                 ret = backup_do_cow(job, cluster * sectors_per_cluster,
 397                                     sectors_per_cluster, &error_is_read,
 398                                     false);
 399                 if ((ret < 0) &&
 400                     backup_error_action(job, error_is_read, -ret) ==
 401                     BLOCK_ERROR_ACTION_REPORT) {
 402                     return ret;
 403                 }
 404             } while (ret < 0);
 405         }
 406
 407         /* If the bitmap granularity is smaller than the backup granularity,
 408          * we need to advance the iterator pointer to the next cluster. */
 409         if (granularity < job->cluster_size) {
 410             bdrv_set_dirty_iter(&hbi, cluster * sectors_per_cluster);
 411         }
 412
 413         last_cluster = cluster - 1;
 414     }
 415
 416     /* Play some final catchup with the progress meter */
 417     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 418     if (last_cluster + 1 < end) {
 419         job->common.offset += ((end - last_cluster - 1) * job->cluster_size);
 420     }
 421
 422     return ret;
 423 }
 424
 425 static void coroutine_fn backup_run(void *opaque)
 426 {
 427     BackupBlockJob *job = opaque;
 428     BackupCompleteData *data;
 429     BlockDriverState *bs = blk_bs(job->common.blk);
 430     BlockBackend *target = job->target;
 431     int64_t start, end;
 432     int64_t sectors_per_cluster = cluster_size_sectors(job);
 433     int ret = 0;
 434
 435     QLIST_INIT(&job->inflight_reqs);
 436     qemu_co_rwlock_init(&job->flush_rwlock);
 437
 438     start = 0;
 439     end = DIV_ROUND_UP(job->common.len, job->cluster_size);
 440
 441     job->done_bitmap = bitmap_new(end);
 442
 443     job->before_write.notify = backup_before_write_notify;
 444     bdrv_add_before_write_notifier(bs, &job->before_write);
 445
 446     if (job->sync_mode == MIRROR_SYNC_MODE_NONE) {
 447         while (!block_job_is_cancelled(&job->common)) {
 448             /* Yield until the job is cancelled.  We just let our before_write
 449              * notify callback service CoW requests. */
 450             block_job_yield(&job->common);
 451         }
 452     } else if (job->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 453         ret = backup_run_incremental(job);
 454     } else {
 455         /* Both FULL and TOP SYNC_MODE's require copying.. */
 456         for (; start < end; start++) {
 457             bool error_is_read;
 458             if (yield_and_check(job)) {
 459                 break;
 460             }
 461
 462             if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
 463                 int i, n;
 464                 int alloced = 0;
 465
 466                 /* Check to see if these blocks are already in the
 467                  * backing file. */
 468
 469                 for (i = 0; i < sectors_per_cluster;) {
 470                     /* bdrv_is_allocated() only returns true/false based
 471                      * on the first set of sectors it comes across that
 472                      * are are all in the same state.
 473                      * For that reason we must verify each sector in the
 474                      * backup cluster length.  We end up copying more than
 475                      * needed but at some point that is always the case. */
 476                     alloced =
 477                         bdrv_is_allocated(bs,
 478                                 start * sectors_per_cluster + i,
 479                                 sectors_per_cluster - i, &n);
 480                     i += n;
 481
 482                     if (alloced == 1 || n == 0) {
 483                         break;
 484                     }
 485                 }
 486
 487                 /* If the above loop never found any sectors that are in
 488                  * the topmost image, skip this backup. */
 489                 if (alloced == 0) {
 490                     continue;
 491                 }
 492             }
 493             /* FULL sync mode we copy the whole drive. */
 494             ret = backup_do_cow(job, start * sectors_per_cluster,
 495                                 sectors_per_cluster, &error_is_read, false);
 496             if (ret < 0) {
 497                 /* Depending on error action, fail now or retry cluster */
 498                 BlockErrorAction action =
 499                     backup_error_action(job, error_is_read, -ret);
 500                 if (action == BLOCK_ERROR_ACTION_REPORT) {
 501                     break;
 502                 } else {
 503                     start--;
 504                     continue;
 505                 }
 506             }
 507         }
 508     }
 509
 510     notifier_with_return_remove(&job->before_write);
 511
 512     /* wait until pending backup_do_cow() calls have completed */
 513     qemu_co_rwlock_wrlock(&job->flush_rwlock);
 514     qemu_co_rwlock_unlock(&job->flush_rwlock);
 515     g_free(job->done_bitmap);
 516
 517     bdrv_op_unblock_all(blk_bs(target), job->common.blocker);
 518
 519     data = g_malloc(sizeof(*data));
 520     data->ret = ret;
 521     block_job_defer_to_main_loop(&job->common, backup_complete, data);
 522 }
 523
 524 void backup_start(const char *job_id, BlockDriverState *bs,
 525                   BlockDriverState *target, int64_t speed,
 526                   MirrorSyncMode sync_mode, BdrvDirtyBitmap *sync_bitmap,
 527                   bool compress,
 528                   BlockdevOnError on_source_error,
 529                   BlockdevOnError on_target_error,
 530                   BlockCompletionFunc *cb, void *opaque,
 531                   BlockJobTxn *txn, Error **errp)
 532 {
 533     int64_t len;
 534     BlockDriverInfo bdi;
 535     BackupBlockJob *job = NULL;
 536     int ret;
 537
 538     assert(bs);
 539     assert(target);
 540
 541     if (bs == target) {
 542         error_setg(errp, "Source and target cannot be the same");
 543         return;
 544     }
 545
 546     if (!bdrv_is_inserted(bs)) {
 547         error_setg(errp, "Device is not inserted: %s",
 548                    bdrv_get_device_name(bs));
 549         return;
 550     }
 551
 552     if (!bdrv_is_inserted(target)) {
 553         error_setg(errp, "Device is not inserted: %s",
 554                    bdrv_get_device_name(target));
 555         return;
 556     }
 557
 558     if (compress && target->drv->bdrv_co_pwritev_compressed == NULL) {
 559         error_setg(errp, "Compression is not supported for this drive %s",
 560                    bdrv_get_device_name(target));
 561         return;
 562     }
 563
 564     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
 565         return;
 566     }
 567
 568     if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
 569         return;
 570     }
 571
 572     if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
 573         if (!sync_bitmap) {
 574             error_setg(errp, "must provide a valid bitmap name for "
 575                              "\"incremental\" sync mode");
 576             return;
 577         }
 578
 579         /* Create a new bitmap, and freeze/disable this one. */
 580         if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
 581             return;
 582         }
 583     } else if (sync_bitmap) {
 584         error_setg(errp,
 585                    "a sync_bitmap was provided to backup_run, "
 586                    "but received an incompatible sync_mode (%s)",
 587                    MirrorSyncMode_lookup[sync_mode]);
 588         return;
 589     }
 590
 591     len = bdrv_getlength(bs);
 592     if (len < 0) {
 593         error_setg_errno(errp, -len, "unable to get length for '%s'",
 594                          bdrv_get_device_name(bs));
 595         goto error;
 596     }
 597
 598     job = block_job_create(job_id, &backup_job_driver, bs, speed,
 599                            cb, opaque, errp);
 600     if (!job) {
 601         goto error;
 602     }
 603
 604     job->target = blk_new();
 605     blk_insert_bs(job->target, target);
 606
 607     job->on_source_error = on_source_error;
 608     job->on_target_error = on_target_error;
 609     job->sync_mode = sync_mode;
 610     job->sync_bitmap = sync_mode == MIRROR_SYNC_MODE_INCREMENTAL ?
 611                        sync_bitmap : NULL;
 612     job->compress = compress;
 613
 614     /* If there is no backing file on the target, we cannot rely on COW if our
 615      * backup cluster size is smaller than the target cluster size. Even for
 616      * targets with a backing file, try to avoid COW if possible. */
 617     ret = bdrv_get_info(target, &bdi);
 618     if (ret < 0 && !target->backing) {
 619         error_setg_errno(errp, -ret,
 620             "Couldn't determine the cluster size of the target image, "
 621             "which has no backing file");
 622         error_append_hint(errp,
 623             "Aborting, since this may create an unusable destination image\n");
 624         goto error;
 625     } else if (ret < 0 && target->backing) {
 626         /* Not fatal; just trudge on ahead. */
 627         job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
 628     } else {
 629         job->cluster_size = MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
 630     }
 631
 632     bdrv_op_block_all(target, job->common.blocker);
 633     job->common.len = len;
 634     job->common.co = qemu_coroutine_create(backup_run, job);
 635     block_job_txn_add_job(txn, &job->common);
 636     qemu_coroutine_enter(job->common.co);
 637     return;
 638
 639  error:
 640     if (sync_bitmap) {
 641         bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL);
 642     }
 643     if (job) {
 644         blk_unref(job->target);
 645         block_job_unref(&job->common);
 646     }
 647 }