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