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