granicus.if.org Git - zfs/blob - module/zfs/dsl_scan.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2012 by Delphix. All rights reserved.
  24  */
  25
  26 #include <sys/dsl_scan.h>
  27 #include <sys/dsl_pool.h>
  28 #include <sys/dsl_dataset.h>
  29 #include <sys/dsl_prop.h>
  30 #include <sys/dsl_dir.h>
  31 #include <sys/dsl_synctask.h>
  32 #include <sys/dnode.h>
  33 #include <sys/dmu_tx.h>
  34 #include <sys/dmu_objset.h>
  35 #include <sys/arc.h>
  36 #include <sys/zap.h>
  37 #include <sys/zio.h>
  38 #include <sys/zfs_context.h>
  39 #include <sys/fs/zfs.h>
  40 #include <sys/zfs_znode.h>
  41 #include <sys/spa_impl.h>
  42 #include <sys/vdev_impl.h>
  43 #include <sys/zil_impl.h>
  44 #include <sys/zio_checksum.h>
  45 #include <sys/ddt.h>
  46 #include <sys/sa.h>
  47 #include <sys/sa_impl.h>
  48 #include <sys/zfeature.h>
  49 #ifdef _KERNEL
  50 #include <sys/zfs_vfsops.h>
  51 #endif
  52
  53 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
  54
  55 static scan_cb_t dsl_scan_scrub_cb;
  56 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
  57 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
  58
  59 int zfs_top_maxinflight = 32;           /* maximum I/Os per top-level */
  60 int zfs_resilver_delay = 2;             /* number of ticks to delay resilver */
  61 int zfs_scrub_delay = 4;                /* number of ticks to delay scrub */
  62 int zfs_scan_idle = 50;                 /* idle window in clock ticks */
  63
  64 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
  65 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
  66 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
  67 int zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  68 int zfs_no_scrub_prefetch = B_FALSE; /* set to disable srub prefetching */
  69 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
  70 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
  71
  72 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
  73         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
  74         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
  75
  76 /* the order has to match pool_scan_type */
  77 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
  78         NULL,
  79         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
  80         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
  81 };
  82
  83 int
  84 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
  85 {
  86         int err;
  87         dsl_scan_t *scn;
  88         spa_t *spa = dp->dp_spa;
  89         uint64_t f;
  90
  91         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
  92         scn->scn_dp = dp;
  93
  94         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
  95             "scrub_func", sizeof (uint64_t), 1, &f);
  96         if (err == 0) {
  97                 /*
  98                  * There was an old-style scrub in progress.  Restart a
  99                  * new-style scrub from the beginning.
 100                  */
 101                 scn->scn_restart_txg = txg;
 102                 zfs_dbgmsg("old-style scrub was in progress; "
 103                     "restarting new-style scrub in txg %llu",
 104                     scn->scn_restart_txg);
 105
 106                 /*
 107                  * Load the queue obj from the old location so that it
 108                  * can be freed by dsl_scan_done().
 109                  */
 110                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 111                     "scrub_queue", sizeof (uint64_t), 1,
 112                     &scn->scn_phys.scn_queue_obj);
 113         } else {
 114                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 115                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 116                     &scn->scn_phys);
 117                 if (err == ENOENT)
 118                         return (0);
 119                 else if (err)
 120                         return (err);
 121
 122                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 123                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 124                         /*
 125                          * A new-type scrub was in progress on an old
 126                          * pool, and the pool was accessed by old
 127                          * software.  Restart from the beginning, since
 128                          * the old software may have changed the pool in
 129                          * the meantime.
 130                          */
 131                         scn->scn_restart_txg = txg;
 132                         zfs_dbgmsg("new-style scrub was modified "
 133                             "by old software; restarting in txg %llu",
 134                             scn->scn_restart_txg);
 135                 }
 136         }
 137
 138         spa_scan_stat_init(spa);
 139         return (0);
 140 }
 141
 142 void
 143 dsl_scan_fini(dsl_pool_t *dp)
 144 {
 145         if (dp->dp_scan) {
 146                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 147                 dp->dp_scan = NULL;
 148         }
 149 }
 150
 151 /* ARGSUSED */
 152 static int
 153 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 154 {
 155         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 156
 157         if (scn->scn_phys.scn_state == DSS_SCANNING)
 158                 return (EBUSY);
 159
 160         return (0);
 161 }
 162
 163 static void
 164 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 165 {
 166         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 167         pool_scan_func_t *funcp = arg;
 168         dmu_object_type_t ot = 0;
 169         dsl_pool_t *dp = scn->scn_dp;
 170         spa_t *spa = dp->dp_spa;
 171
 172         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 173         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 174         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 175         scn->scn_phys.scn_func = *funcp;
 176         scn->scn_phys.scn_state = DSS_SCANNING;
 177         scn->scn_phys.scn_min_txg = 0;
 178         scn->scn_phys.scn_max_txg = tx->tx_txg;
 179         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 180         scn->scn_phys.scn_start_time = gethrestime_sec();
 181         scn->scn_phys.scn_errors = 0;
 182         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 183         scn->scn_restart_txg = 0;
 184         spa_scan_stat_init(spa);
 185
 186         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 187                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 188
 189                 /* rewrite all disk labels */
 190                 vdev_config_dirty(spa->spa_root_vdev);
 191
 192                 if (vdev_resilver_needed(spa->spa_root_vdev,
 193                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 194                         spa_event_notify(spa, NULL, FM_EREPORT_ZFS_RESILVER_START);
 195                 } else {
 196                         spa_event_notify(spa, NULL, FM_EREPORT_ZFS_SCRUB_START);
 197                 }
 198
 199                 spa->spa_scrub_started = B_TRUE;
 200                 /*
 201                  * If this is an incremental scrub, limit the DDT scrub phase
 202                  * to just the auto-ditto class (for correctness); the rest
 203                  * of the scrub should go faster using top-down pruning.
 204                  */
 205                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 206                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 207
 208         }
 209
 210         /* back to the generic stuff */
 211
 212         if (dp->dp_blkstats == NULL) {
 213                 dp->dp_blkstats = kmem_alloc(sizeof (zfs_all_blkstats_t),
 214                     KM_PUSHPAGE | KM_NODEBUG);
 215         }
 216         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 217
 218         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 219                 ot = DMU_OT_ZAP_OTHER;
 220
 221         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 222             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 223
 224         dsl_scan_sync_state(scn, tx);
 225
 226         spa_history_log_internal(spa, "scan setup", tx,
 227             "func=%u mintxg=%llu maxtxg=%llu",
 228             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 229 }
 230
 231 /* ARGSUSED */
 232 static void
 233 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 234 {
 235         static const char *old_names[] = {
 236                 "scrub_bookmark",
 237                 "scrub_ddt_bookmark",
 238                 "scrub_ddt_class_max",
 239                 "scrub_queue",
 240                 "scrub_min_txg",
 241                 "scrub_max_txg",
 242                 "scrub_func",
 243                 "scrub_errors",
 244                 NULL
 245         };
 246
 247         dsl_pool_t *dp = scn->scn_dp;
 248         spa_t *spa = dp->dp_spa;
 249         int i;
 250
 251         /* Remove any remnants of an old-style scrub. */
 252         for (i = 0; old_names[i]; i++) {
 253                 (void) zap_remove(dp->dp_meta_objset,
 254                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 255         }
 256
 257         if (scn->scn_phys.scn_queue_obj != 0) {
 258                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 259                     scn->scn_phys.scn_queue_obj, tx));
 260                 scn->scn_phys.scn_queue_obj = 0;
 261         }
 262
 263         /*
 264          * If we were "restarted" from a stopped state, don't bother
 265          * with anything else.
 266          */
 267         if (scn->scn_phys.scn_state != DSS_SCANNING)
 268                 return;
 269
 270         if (complete)
 271                 scn->scn_phys.scn_state = DSS_FINISHED;
 272         else
 273                 scn->scn_phys.scn_state = DSS_CANCELED;
 274
 275         spa_history_log_internal(spa, "scan done", tx,
 276             "complete=%u", complete);
 277
 278         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 279                 mutex_enter(&spa->spa_scrub_lock);
 280                 while (spa->spa_scrub_inflight > 0) {
 281                         cv_wait(&spa->spa_scrub_io_cv,
 282                             &spa->spa_scrub_lock);
 283                 }
 284                 mutex_exit(&spa->spa_scrub_lock);
 285                 spa->spa_scrub_started = B_FALSE;
 286                 spa->spa_scrub_active = B_FALSE;
 287
 288                 /*
 289                  * If the scrub/resilver completed, update all DTLs to
 290                  * reflect this.  Whether it succeeded or not, vacate
 291                  * all temporary scrub DTLs.
 292                  */
 293                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 294                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 295                 if (complete) {
 296                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 297                             FM_EREPORT_ZFS_RESILVER_FINISH :
 298                             FM_EREPORT_ZFS_SCRUB_FINISH);
 299                 }
 300                 spa_errlog_rotate(spa);
 301
 302                 /*
 303                  * We may have finished replacing a device.
 304                  * Let the async thread assess this and handle the detach.
 305                  */
 306                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 307         }
 308
 309         scn->scn_phys.scn_end_time = gethrestime_sec();
 310 }
 311
 312 /* ARGSUSED */
 313 static int
 314 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 315 {
 316         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 317
 318         if (scn->scn_phys.scn_state != DSS_SCANNING)
 319                 return (ENOENT);
 320         return (0);
 321 }
 322
 323 /* ARGSUSED */
 324 static void
 325 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 326 {
 327         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 328
 329         dsl_scan_done(scn, B_FALSE, tx);
 330         dsl_scan_sync_state(scn, tx);
 331 }
 332
 333 int
 334 dsl_scan_cancel(dsl_pool_t *dp)
 335 {
 336         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 337             dsl_scan_cancel_sync, NULL, 3));
 338 }
 339
 340 static void dsl_scan_visitbp(blkptr_t *bp,
 341     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 342     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 343     dmu_tx_t *tx);
 344 inline __attribute__((always_inline)) static void dsl_scan_visitdnode(
 345     dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 346     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 347
 348 void
 349 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 350 {
 351         zio_free(dp->dp_spa, txg, bp);
 352 }
 353
 354 void
 355 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 356 {
 357         ASSERT(dsl_pool_sync_context(dp));
 358         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
 359 }
 360
 361 static uint64_t
 362 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 363 {
 364         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 365         if (dsl_dataset_is_snapshot(ds))
 366                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 367         return (smt);
 368 }
 369
 370 static void
 371 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 372 {
 373         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 374             DMU_POOL_DIRECTORY_OBJECT,
 375             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 376             &scn->scn_phys, tx));
 377 }
 378
 379 static boolean_t
 380 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 381 {
 382         uint64_t elapsed_nanosecs;
 383         int mintime;
 384
 385         /* we never skip user/group accounting objects */
 386         if (zb && (int64_t)zb->zb_object < 0)
 387                 return (B_FALSE);
 388
 389         if (scn->scn_pausing)
 390                 return (B_TRUE); /* we're already pausing */
 391
 392         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 393                 return (B_FALSE); /* we're resuming */
 394
 395         /* We only know how to resume from level-0 blocks. */
 396         if (zb && zb->zb_level != 0)
 397                 return (B_FALSE);
 398
 399         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 400             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 401         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 402         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 403             (elapsed_nanosecs / MICROSEC > mintime &&
 404             txg_sync_waiting(scn->scn_dp)) ||
 405             spa_shutting_down(scn->scn_dp->dp_spa)) {
 406                 if (zb) {
 407                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 408                             (longlong_t)zb->zb_objset,
 409                             (longlong_t)zb->zb_object,
 410                             (longlong_t)zb->zb_level,
 411                             (longlong_t)zb->zb_blkid);
 412                         scn->scn_phys.scn_bookmark = *zb;
 413                 }
 414                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 415                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 416                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 417                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 418                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 419                 scn->scn_pausing = B_TRUE;
 420                 return (B_TRUE);
 421         }
 422         return (B_FALSE);
 423 }
 424
 425 typedef struct zil_scan_arg {
 426         dsl_pool_t      *zsa_dp;
 427         zil_header_t    *zsa_zh;
 428 } zil_scan_arg_t;
 429
 430 /* ARGSUSED */
 431 static int
 432 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 433 {
 434         zil_scan_arg_t *zsa = arg;
 435         dsl_pool_t *dp = zsa->zsa_dp;
 436         dsl_scan_t *scn = dp->dp_scan;
 437         zil_header_t *zh = zsa->zsa_zh;
 438         zbookmark_t zb;
 439
 440         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 441                 return (0);
 442
 443         /*
 444          * One block ("stubby") can be allocated a long time ago; we
 445          * want to visit that one because it has been allocated
 446          * (on-disk) even if it hasn't been claimed (even though for
 447          * scrub there's nothing to do to it).
 448          */
 449         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 450                 return (0);
 451
 452         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 453             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 454
 455         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 456         return (0);
 457 }
 458
 459 /* ARGSUSED */
 460 static int
 461 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 462 {
 463         if (lrc->lrc_txtype == TX_WRITE) {
 464                 zil_scan_arg_t *zsa = arg;
 465                 dsl_pool_t *dp = zsa->zsa_dp;
 466                 dsl_scan_t *scn = dp->dp_scan;
 467                 zil_header_t *zh = zsa->zsa_zh;
 468                 lr_write_t *lr = (lr_write_t *)lrc;
 469                 blkptr_t *bp = &lr->lr_blkptr;
 470                 zbookmark_t zb;
 471
 472                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 473                         return (0);
 474
 475                 /*
 476                  * birth can be < claim_txg if this record's txg is
 477                  * already txg sync'ed (but this log block contains
 478                  * other records that are not synced)
 479                  */
 480                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 481                         return (0);
 482
 483                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 484                     lr->lr_foid, ZB_ZIL_LEVEL,
 485                     lr->lr_offset / BP_GET_LSIZE(bp));
 486
 487                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 488         }
 489         return (0);
 490 }
 491
 492 static void
 493 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 494 {
 495         uint64_t claim_txg = zh->zh_claim_txg;
 496         zil_scan_arg_t zsa = { dp, zh };
 497         zilog_t *zilog;
 498
 499         /*
 500          * We only want to visit blocks that have been claimed but not yet
 501          * replayed (or, in read-only mode, blocks that *would* be claimed).
 502          */
 503         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 504                 return;
 505
 506         zilog = zil_alloc(dp->dp_meta_objset, zh);
 507
 508         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 509             claim_txg);
 510
 511         zil_free(zilog);
 512 }
 513
 514 /* ARGSUSED */
 515 static void
 516 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 517     uint64_t objset, uint64_t object, uint64_t blkid)
 518 {
 519         zbookmark_t czb;
 520         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 521
 522         if (zfs_no_scrub_prefetch)
 523                 return;
 524
 525         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 526             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 527                 return;
 528
 529         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 530
 531         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 532             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 533             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 534 }
 535
 536 static boolean_t
 537 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 538     const zbookmark_t *zb)
 539 {
 540         /*
 541          * We never skip over user/group accounting objects (obj<0)
 542          */
 543         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 544             (int64_t)zb->zb_object >= 0) {
 545                 /*
 546                  * If we already visited this bp & everything below (in
 547                  * a prior txg sync), don't bother doing it again.
 548                  */
 549                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 550                         return (B_TRUE);
 551
 552                 /*
 553                  * If we found the block we're trying to resume from, or
 554                  * we went past it to a different object, zero it out to
 555                  * indicate that it's OK to start checking for pausing
 556                  * again.
 557                  */
 558                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 559                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 560                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 561                             (longlong_t)zb->zb_objset,
 562                             (longlong_t)zb->zb_object,
 563                             (longlong_t)zb->zb_level,
 564                             (longlong_t)zb->zb_blkid);
 565                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 566                 }
 567         }
 568         return (B_FALSE);
 569 }
 570
 571 /*
 572  * Return nonzero on i/o error.
 573  * Return new buf to write out in *bufp.
 574  */
 575 inline __attribute__((always_inline)) static int
 576 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 577     dnode_phys_t *dnp, const blkptr_t *bp,
 578     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 579 {
 580         dsl_pool_t *dp = scn->scn_dp;
 581         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 582         int err;
 583
 584         if (BP_GET_LEVEL(bp) > 0) {
 585                 uint32_t flags = ARC_WAIT;
 586                 int i;
 587                 blkptr_t *cbp;
 588                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 589
 590                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 591                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 592                 if (err) {
 593                         scn->scn_phys.scn_errors++;
 594                         return (err);
 595                 }
 596                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 597                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 598                             zb->zb_object, zb->zb_blkid * epb + i);
 599                 }
 600                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 601                         zbookmark_t czb;
 602
 603                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 604                             zb->zb_level - 1,
 605                             zb->zb_blkid * epb + i);
 606                         dsl_scan_visitbp(cbp, &czb, dnp,
 607                             *bufp, ds, scn, ostype, tx);
 608                 }
 609         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 610                 uint32_t flags = ARC_WAIT;
 611
 612                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 613                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 614                 if (err) {
 615                         scn->scn_phys.scn_errors++;
 616                         return (err);
 617                 }
 618         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 619                 uint32_t flags = ARC_WAIT;
 620                 dnode_phys_t *cdnp;
 621                 int i, j;
 622                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 623
 624                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 625                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 626                 if (err) {
 627                         scn->scn_phys.scn_errors++;
 628                         return (err);
 629                 }
 630                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 631                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 632                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 633                                 dsl_scan_prefetch(scn, *bufp, cbp,
 634                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 635                         }
 636                 }
 637                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 638                         dsl_scan_visitdnode(scn, ds, ostype,
 639                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 640                 }
 641
 642         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 643                 uint32_t flags = ARC_WAIT;
 644                 objset_phys_t *osp;
 645
 646                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 647                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 648                 if (err) {
 649                         scn->scn_phys.scn_errors++;
 650                         return (err);
 651                 }
 652
 653                 osp = (*bufp)->b_data;
 654
 655                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 656                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 657
 658                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 659                         /*
 660                          * We also always visit user/group accounting
 661                          * objects, and never skip them, even if we are
 662                          * pausing.  This is necessary so that the space
 663                          * deltas from this txg get integrated.
 664                          */
 665                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 666                             &osp->os_groupused_dnode, *bufp,
 667                             DMU_GROUPUSED_OBJECT, tx);
 668                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 669                             &osp->os_userused_dnode, *bufp,
 670                             DMU_USERUSED_OBJECT, tx);
 671                 }
 672         }
 673
 674         return (0);
 675 }
 676
 677 inline __attribute__((always_inline)) static void
 678 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 679     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 680     uint64_t object, dmu_tx_t *tx)
 681 {
 682         int j;
 683
 684         for (j = 0; j < dnp->dn_nblkptr; j++) {
 685                 zbookmark_t czb;
 686
 687                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 688                     dnp->dn_nlevels - 1, j);
 689                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 690                     &czb, dnp, buf, ds, scn, ostype, tx);
 691         }
 692
 693         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 694                 zbookmark_t czb;
 695                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 696                     0, DMU_SPILL_BLKID);
 697                 dsl_scan_visitbp(&dnp->dn_spill,
 698                     &czb, dnp, buf, ds, scn, ostype, tx);
 699         }
 700 }
 701
 702 /*
 703  * The arguments are in this order because mdb can only print the
 704  * first 5; we want them to be useful.
 705  */
 706 static void
 707 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 708     dnode_phys_t *dnp, arc_buf_t *pbuf,
 709     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 710     dmu_tx_t *tx)
 711 {
 712         dsl_pool_t *dp = scn->scn_dp;
 713         arc_buf_t *buf = NULL;
 714         blkptr_t *bp_toread;
 715
 716         bp_toread = kmem_alloc(sizeof (blkptr_t), KM_PUSHPAGE);
 717         *bp_toread = *bp;
 718
 719         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 720
 721         if (dsl_scan_check_pause(scn, zb))
 722                 goto out;
 723
 724         if (dsl_scan_check_resume(scn, dnp, zb))
 725                 goto out;
 726
 727         if (bp->blk_birth == 0)
 728                 goto out;
 729
 730         scn->scn_visited_this_txg++;
 731
 732         /*
 733          * This debugging is commented out to conserve stack space.  This
 734          * function is called recursively and the debugging addes several
 735          * bytes to the stack for each call.  It can be commented back in
 736          * if required to debug an issue in dsl_scan_visitbp().
 737          *
 738          * dprintf_bp(bp,
 739          *    "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 740          *    ds, ds ? ds->ds_object : 0,
 741          *    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 742          *    pbuf, bp);
 743          */
 744
 745         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 746                 goto out;
 747
 748         if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx,
 749             &buf) != 0)
 750                 goto out;
 751
 752         /*
 753          * If dsl_scan_ddt() has aready visited this block, it will have
 754          * already done any translations or scrubbing, so don't call the
 755          * callback again.
 756          */
 757         if (ddt_class_contains(dp->dp_spa,
 758             scn->scn_phys.scn_ddt_class_max, bp)) {
 759                 ASSERT(buf == NULL);
 760                 goto out;
 761         }
 762
 763         /*
 764          * If this block is from the future (after cur_max_txg), then we
 765          * are doing this on behalf of a deleted snapshot, and we will
 766          * revisit the future block on the next pass of this dataset.
 767          * Don't scan it now unless we need to because something
 768          * under it was modified.
 769          */
 770         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 771                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 772         }
 773         if (buf)
 774                 (void) arc_buf_remove_ref(buf, &buf);
 775 out:
 776         kmem_free(bp_toread, sizeof(blkptr_t));
 777 }
 778
 779 static void
 780 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 781     dmu_tx_t *tx)
 782 {
 783         zbookmark_t zb;
 784
 785         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 786             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 787         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 788             ds, scn, DMU_OST_NONE, tx);
 789
 790         dprintf_ds(ds, "finished scan%s", "");
 791 }
 792
 793 void
 794 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 795 {
 796         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 797         dsl_scan_t *scn = dp->dp_scan;
 798         uint64_t mintxg;
 799
 800         if (scn->scn_phys.scn_state != DSS_SCANNING)
 801                 return;
 802
 803         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 804                 if (dsl_dataset_is_snapshot(ds)) {
 805                         /* Note, scn_cur_{min,max}_txg stays the same. */
 806                         scn->scn_phys.scn_bookmark.zb_objset =
 807                             ds->ds_phys->ds_next_snap_obj;
 808                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 809                             "reset zb_objset to %llu",
 810                             (u_longlong_t)ds->ds_object,
 811                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 812                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 813                 } else {
 814                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 815                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 816                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 817                             "reset bookmark to -1,0,0,0",
 818                             (u_longlong_t)ds->ds_object);
 819                 }
 820         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 821             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 822                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 823                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 824                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 825                 if (dsl_dataset_is_snapshot(ds)) {
 826                         /*
 827                          * We keep the same mintxg; it could be >
 828                          * ds_creation_txg if the previous snapshot was
 829                          * deleted too.
 830                          */
 831                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 832                             scn->scn_phys.scn_queue_obj,
 833                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 834                         zfs_dbgmsg("destroying ds %llu; in queue; "
 835                             "replacing with %llu",
 836                             (u_longlong_t)ds->ds_object,
 837                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 838                 } else {
 839                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 840                             (u_longlong_t)ds->ds_object);
 841                 }
 842         } else {
 843                 zfs_dbgmsg("destroying ds %llu; ignoring",
 844                     (u_longlong_t)ds->ds_object);
 845         }
 846
 847         /*
 848          * dsl_scan_sync() should be called after this, and should sync
 849          * out our changed state, but just to be safe, do it here.
 850          */
 851         dsl_scan_sync_state(scn, tx);
 852 }
 853
 854 void
 855 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 856 {
 857         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 858         dsl_scan_t *scn = dp->dp_scan;
 859         uint64_t mintxg;
 860
 861         if (scn->scn_phys.scn_state != DSS_SCANNING)
 862                 return;
 863
 864         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 865
 866         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 867                 scn->scn_phys.scn_bookmark.zb_objset =
 868                     ds->ds_phys->ds_prev_snap_obj;
 869                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 870                     "reset zb_objset to %llu",
 871                     (u_longlong_t)ds->ds_object,
 872                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 873         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 874             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 875                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 876                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 877                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 878                     scn->scn_phys.scn_queue_obj,
 879                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 880                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 881                     "replacing with %llu",
 882                     (u_longlong_t)ds->ds_object,
 883                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 884         }
 885         dsl_scan_sync_state(scn, tx);
 886 }
 887
 888 void
 889 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 890 {
 891         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 892         dsl_scan_t *scn = dp->dp_scan;
 893         uint64_t mintxg;
 894
 895         if (scn->scn_phys.scn_state != DSS_SCANNING)
 896                 return;
 897
 898         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 899                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 900                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 901                     "reset zb_objset to %llu",
 902                     (u_longlong_t)ds1->ds_object,
 903                     (u_longlong_t)ds2->ds_object);
 904         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 905                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 906                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 907                     "reset zb_objset to %llu",
 908                     (u_longlong_t)ds2->ds_object,
 909                     (u_longlong_t)ds1->ds_object);
 910         }
 911
 912         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 913             ds1->ds_object, &mintxg) == 0) {
 914                 int err;
 915
 916                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 917                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 918                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 919                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 920                 err = zap_add_int_key(dp->dp_meta_objset,
 921                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 922                 VERIFY(err == 0 || err == EEXIST);
 923                 if (err == EEXIST) {
 924                         /* Both were there to begin with */
 925                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 926                             scn->scn_phys.scn_queue_obj,
 927                             ds1->ds_object, mintxg, tx));
 928                 }
 929                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 930                     "replacing with %llu",
 931                     (u_longlong_t)ds1->ds_object,
 932                     (u_longlong_t)ds2->ds_object);
 933         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 934             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 935                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 936                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 937                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 938                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 939                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 940                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 941                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 942                     "replacing with %llu",
 943                     (u_longlong_t)ds2->ds_object,
 944                     (u_longlong_t)ds1->ds_object);
 945         }
 946
 947         dsl_scan_sync_state(scn, tx);
 948 }
 949
 950 struct enqueue_clones_arg {
 951         dmu_tx_t *tx;
 952         uint64_t originobj;
 953 };
 954
 955 /* ARGSUSED */
 956 static int
 957 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 958 {
 959         struct enqueue_clones_arg *eca = arg;
 960         dsl_dataset_t *ds;
 961         int err;
 962         dsl_scan_t *scn = dp->dp_scan;
 963
 964         if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
 965                 return (0);
 966
 967         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
 968         if (err)
 969                 return (err);
 970
 971         while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 972                 dsl_dataset_t *prev;
 973                 err = dsl_dataset_hold_obj(dp,
 974                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 975
 976                 dsl_dataset_rele(ds, FTAG);
 977                 if (err)
 978                         return (err);
 979                 ds = prev;
 980         }
 981         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 982             scn->scn_phys.scn_queue_obj, ds->ds_object,
 983             ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
 984         dsl_dataset_rele(ds, FTAG);
 985         return (0);
 986 }
 987
 988 static void
 989 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
 990 {
 991         dsl_pool_t *dp = scn->scn_dp;
 992         dsl_dataset_t *ds;
 993         objset_t *os;
 994         char *dsname;
 995
 996         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 997
 998         if (dmu_objset_from_ds(ds, &os))
 999                 goto out;
1000
1001         /*
1002          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1003          * snapshots can have ZIL block pointers (which may be the same
1004          * BP as in the head), they must be ignored.  So we traverse the
1005          * ZIL here, rather than in scan_recurse(), because the regular
1006          * snapshot block-sharing rules don't apply to it.
1007          */
1008         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1009                 dsl_scan_zil(dp, &os->os_zil_header);
1010
1011         /*
1012          * Iterate over the bps in this ds.
1013          */
1014         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1015         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1016
1017         dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_PUSHPAGE);
1018         dsl_dataset_name(ds, dsname);
1019         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1020             "pausing=%u",
1021             (longlong_t)dsobj, dsname,
1022             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1023             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1024             (int)scn->scn_pausing);
1025         kmem_free(dsname, ZFS_MAXNAMELEN);
1026
1027         if (scn->scn_pausing)
1028                 goto out;
1029
1030         /*
1031          * We've finished this pass over this dataset.
1032          */
1033
1034         /*
1035          * If we did not completely visit this dataset, do another pass.
1036          */
1037         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1038                 zfs_dbgmsg("incomplete pass; visiting again");
1039                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1040                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1041                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1042                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1043                 goto out;
1044         }
1045
1046         /*
1047          * Add descendent datasets to work queue.
1048          */
1049         if (ds->ds_phys->ds_next_snap_obj != 0) {
1050                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1051                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1052                     ds->ds_phys->ds_creation_txg, tx) == 0);
1053         }
1054         if (ds->ds_phys->ds_num_children > 1) {
1055                 boolean_t usenext = B_FALSE;
1056                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1057                         uint64_t count;
1058                         /*
1059                          * A bug in a previous version of the code could
1060                          * cause upgrade_clones_cb() to not set
1061                          * ds_next_snap_obj when it should, leading to a
1062                          * missing entry.  Therefore we can only use the
1063                          * next_clones_obj when its count is correct.
1064                          */
1065                         int err = zap_count(dp->dp_meta_objset,
1066                             ds->ds_phys->ds_next_clones_obj, &count);
1067                         if (err == 0 &&
1068                             count == ds->ds_phys->ds_num_children - 1)
1069                                 usenext = B_TRUE;
1070                 }
1071
1072                 if (usenext) {
1073                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1074                             ds->ds_phys->ds_next_clones_obj,
1075                             scn->scn_phys.scn_queue_obj,
1076                             ds->ds_phys->ds_creation_txg, tx));
1077                 } else {
1078                         struct enqueue_clones_arg eca;
1079                         eca.tx = tx;
1080                         eca.originobj = ds->ds_object;
1081
1082                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1083                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1084                 }
1085         }
1086
1087 out:
1088         dsl_dataset_rele(ds, FTAG);
1089 }
1090
1091 /* ARGSUSED */
1092 static int
1093 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1094 {
1095         dmu_tx_t *tx = arg;
1096         dsl_dataset_t *ds;
1097         int err;
1098         dsl_scan_t *scn = dp->dp_scan;
1099
1100         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1101         if (err)
1102                 return (err);
1103
1104         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1105                 dsl_dataset_t *prev;
1106                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1107                     FTAG, &prev);
1108                 if (err) {
1109                         dsl_dataset_rele(ds, FTAG);
1110                         return (err);
1111                 }
1112
1113                 /*
1114                  * If this is a clone, we don't need to worry about it for now.
1115                  */
1116                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1117                         dsl_dataset_rele(ds, FTAG);
1118                         dsl_dataset_rele(prev, FTAG);
1119                         return (0);
1120                 }
1121                 dsl_dataset_rele(ds, FTAG);
1122                 ds = prev;
1123         }
1124
1125         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1126             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1127         dsl_dataset_rele(ds, FTAG);
1128         return (0);
1129 }
1130
1131 /*
1132  * Scrub/dedup interaction.
1133  *
1134  * If there are N references to a deduped block, we don't want to scrub it
1135  * N times -- ideally, we should scrub it exactly once.
1136  *
1137  * We leverage the fact that the dde's replication class (enum ddt_class)
1138  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1139  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1140  *
1141  * To prevent excess scrubbing, the scrub begins by walking the DDT
1142  * to find all blocks with refcnt > 1, and scrubs each of these once.
1143  * Since there are two replication classes which contain blocks with
1144  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1145  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1146  *
1147  * There would be nothing more to say if a block's refcnt couldn't change
1148  * during a scrub, but of course it can so we must account for changes
1149  * in a block's replication class.
1150  *
1151  * Here's an example of what can occur:
1152  *
1153  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1154  * when visited during the top-down scrub phase, it will be scrubbed twice.
1155  * This negates our scrub optimization, but is otherwise harmless.
1156  *
1157  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1158  * on each visit during the top-down scrub phase, it will never be scrubbed.
1159  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1160  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1161  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1162  * while a scrub is in progress, it scrubs the block right then.
1163  */
1164 static void
1165 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1166 {
1167         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1168         ddt_entry_t dde;
1169         int error;
1170         uint64_t n = 0;
1171
1172         bzero(&dde, sizeof (ddt_entry_t));
1173
1174         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1175                 ddt_t *ddt;
1176
1177                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1178                         break;
1179                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1180                     (longlong_t)ddb->ddb_class,
1181                     (longlong_t)ddb->ddb_type,
1182                     (longlong_t)ddb->ddb_checksum,
1183                     (longlong_t)ddb->ddb_cursor);
1184
1185                 /* There should be no pending changes to the dedup table */
1186                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1187                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1188
1189                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1190                 n++;
1191
1192                 if (dsl_scan_check_pause(scn, NULL))
1193                         break;
1194         }
1195
1196         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1197             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1198             (int)scn->scn_pausing);
1199
1200         ASSERT(error == 0 || error == ENOENT);
1201         ASSERT(error != ENOENT ||
1202             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1203 }
1204
1205 /* ARGSUSED */
1206 void
1207 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1208     ddt_entry_t *dde, dmu_tx_t *tx)
1209 {
1210         const ddt_key_t *ddk = &dde->dde_key;
1211         ddt_phys_t *ddp = dde->dde_phys;
1212         blkptr_t bp;
1213         zbookmark_t zb = { 0 };
1214         int p;
1215
1216         if (scn->scn_phys.scn_state != DSS_SCANNING)
1217                 return;
1218
1219         for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1220                 if (ddp->ddp_phys_birth == 0 ||
1221                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1222                         continue;
1223                 ddt_bp_create(checksum, ddk, ddp, &bp);
1224
1225                 scn->scn_visited_this_txg++;
1226                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1227         }
1228 }
1229
1230 static void
1231 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1232 {
1233         dsl_pool_t *dp = scn->scn_dp;
1234         zap_cursor_t *zc;
1235         zap_attribute_t *za;
1236
1237         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1238             scn->scn_phys.scn_ddt_class_max) {
1239                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1240                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1241                 dsl_scan_ddt(scn, tx);
1242                 if (scn->scn_pausing)
1243                         return;
1244         }
1245
1246         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1247                 /* First do the MOS & ORIGIN */
1248
1249                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1250                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1251                 dsl_scan_visit_rootbp(scn, NULL,
1252                     &dp->dp_meta_rootbp, tx);
1253                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1254                 if (scn->scn_pausing)
1255                         return;
1256
1257                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1258                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1259                             enqueue_cb, tx, DS_FIND_CHILDREN));
1260                 } else {
1261                         dsl_scan_visitds(scn,
1262                             dp->dp_origin_snap->ds_object, tx);
1263                 }
1264                 ASSERT(!scn->scn_pausing);
1265         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1266             ZB_DESTROYED_OBJSET) {
1267                 /*
1268                  * If we were paused, continue from here.  Note if the
1269                  * ds we were paused on was deleted, the zb_objset may
1270                  * be -1, so we will skip this and find a new objset
1271                  * below.
1272                  */
1273                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1274                 if (scn->scn_pausing)
1275                         return;
1276         }
1277
1278         /*
1279          * In case we were paused right at the end of the ds, zero the
1280          * bookmark so we don't think that we're still trying to resume.
1281          */
1282         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1283         zc = kmem_alloc(sizeof(zap_cursor_t), KM_PUSHPAGE);
1284         za = kmem_alloc(sizeof(zap_attribute_t), KM_PUSHPAGE);
1285
1286         /* keep pulling things out of the zap-object-as-queue */
1287         while (zap_cursor_init(zc, dp->dp_meta_objset,
1288             scn->scn_phys.scn_queue_obj),
1289             zap_cursor_retrieve(zc, za) == 0) {
1290                 dsl_dataset_t *ds;
1291                 uint64_t dsobj;
1292
1293                 dsobj = strtonum(za->za_name, NULL);
1294                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1295                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1296
1297                 /* Set up min/max txg */
1298                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1299                 if (za->za_first_integer != 0) {
1300                         scn->scn_phys.scn_cur_min_txg =
1301                             MAX(scn->scn_phys.scn_min_txg,
1302                             za->za_first_integer);
1303                 } else {
1304                         scn->scn_phys.scn_cur_min_txg =
1305                             MAX(scn->scn_phys.scn_min_txg,
1306                             ds->ds_phys->ds_prev_snap_txg);
1307                 }
1308                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1309                 dsl_dataset_rele(ds, FTAG);
1310
1311                 dsl_scan_visitds(scn, dsobj, tx);
1312                 zap_cursor_fini(zc);
1313                 if (scn->scn_pausing)
1314                         goto out;
1315         }
1316         zap_cursor_fini(zc);
1317 out:
1318         kmem_free(za, sizeof(zap_attribute_t));
1319         kmem_free(zc, sizeof(zap_cursor_t));
1320 }
1321
1322 static boolean_t
1323 dsl_scan_free_should_pause(dsl_scan_t *scn)
1324 {
1325         uint64_t elapsed_nanosecs;
1326
1327         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1328         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1329             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1330             txg_sync_waiting(scn->scn_dp)) ||
1331             spa_shutting_down(scn->scn_dp->dp_spa));
1332 }
1333
1334 static int
1335 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1336 {
1337         dsl_scan_t *scn = arg;
1338
1339         if (!scn->scn_is_bptree ||
1340             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1341                 if (dsl_scan_free_should_pause(scn))
1342                         return (ERESTART);
1343         }
1344
1345         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1346             dmu_tx_get_txg(tx), bp, 0));
1347         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1348             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1349             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1350         scn->scn_visited_this_txg++;
1351         return (0);
1352 }
1353
1354 boolean_t
1355 dsl_scan_active(dsl_scan_t *scn)
1356 {
1357         spa_t *spa = scn->scn_dp->dp_spa;
1358         uint64_t used = 0, comp, uncomp;
1359
1360         if (spa->spa_load_state != SPA_LOAD_NONE)
1361                 return (B_FALSE);
1362         if (spa_shutting_down(spa))
1363                 return (B_FALSE);
1364
1365         if (scn->scn_phys.scn_state == DSS_SCANNING)
1366                 return (B_TRUE);
1367
1368         if (spa_feature_is_active(spa,
1369             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1370                 return (B_TRUE);
1371         }
1372         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1373                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1374                     &used, &comp, &uncomp);
1375         }
1376         return (used != 0);
1377 }
1378
1379 void
1380 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1381 {
1382         dsl_scan_t *scn = dp->dp_scan;
1383         spa_t *spa = dp->dp_spa;
1384         int err;
1385
1386         /*
1387          * Check for scn_restart_txg before checking spa_load_state, so
1388          * that we can restart an old-style scan while the pool is being
1389          * imported (see dsl_scan_init).
1390          */
1391         if (scn->scn_restart_txg != 0 &&
1392             scn->scn_restart_txg <= tx->tx_txg) {
1393                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1394                 dsl_scan_done(scn, B_FALSE, tx);
1395                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1396                         func = POOL_SCAN_RESILVER;
1397                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1398                     func, tx->tx_txg);
1399                 dsl_scan_setup_sync(&func, tx);
1400         }
1401
1402         if (!dsl_scan_active(scn) ||
1403             spa_sync_pass(dp->dp_spa) > 1)
1404                 return;
1405
1406         scn->scn_visited_this_txg = 0;
1407         scn->scn_pausing = B_FALSE;
1408         scn->scn_sync_start_time = gethrtime();
1409         spa->spa_scrub_active = B_TRUE;
1410
1411         /*
1412          * First process the free list.  If we pause the free, don't do
1413          * any scanning.  This ensures that there is no free list when
1414          * we are scanning, so the scan code doesn't have to worry about
1415          * traversing it.
1416          */
1417         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1418                 scn->scn_is_bptree = B_FALSE;
1419                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1420                     NULL, ZIO_FLAG_MUSTSUCCEED);
1421                 err = bpobj_iterate(&dp->dp_free_bpobj,
1422                     dsl_scan_free_block_cb, scn, tx);
1423                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1424
1425                 if (err == 0 && spa_feature_is_active(spa,
1426                     &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1427                         scn->scn_is_bptree = B_TRUE;
1428                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1429                             NULL, ZIO_FLAG_MUSTSUCCEED);
1430                         err = bptree_iterate(dp->dp_meta_objset,
1431                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1432                             scn, tx);
1433                         VERIFY0(zio_wait(scn->scn_zio_root));
1434
1435                         if (err == 0) {
1436                                 zfeature_info_t *feat = &spa_feature_table
1437                                     [SPA_FEATURE_ASYNC_DESTROY];
1438                                 /* finished; deactivate async destroy feature */
1439                                 spa_feature_decr(spa, feat, tx);
1440                                 ASSERT(!spa_feature_is_active(spa, feat));
1441                                 VERIFY0(zap_remove(dp->dp_meta_objset,
1442                                     DMU_POOL_DIRECTORY_OBJECT,
1443                                     DMU_POOL_BPTREE_OBJ, tx));
1444                                 VERIFY0(bptree_free(dp->dp_meta_objset,
1445                                     dp->dp_bptree_obj, tx));
1446                                 dp->dp_bptree_obj = 0;
1447                         }
1448                 }
1449                 if (scn->scn_visited_this_txg) {
1450                         zfs_dbgmsg("freed %llu blocks in %llums from "
1451                             "free_bpobj/bptree txg %llu",
1452                             (longlong_t)scn->scn_visited_this_txg,
1453                             (longlong_t)
1454                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1455                             (longlong_t)tx->tx_txg);
1456                         scn->scn_visited_this_txg = 0;
1457                         /*
1458                          * Re-sync the ddt so that we can further modify
1459                          * it when doing bprewrite.
1460                          */
1461                         ddt_sync(spa, tx->tx_txg);
1462                 }
1463                 if (err == ERESTART)
1464                         return;
1465         }
1466
1467         if (scn->scn_phys.scn_state != DSS_SCANNING)
1468                 return;
1469
1470         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1471             scn->scn_phys.scn_ddt_class_max) {
1472                 zfs_dbgmsg("doing scan sync txg %llu; "
1473                     "ddt bm=%llu/%llu/%llu/%llx",
1474                     (longlong_t)tx->tx_txg,
1475                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1476                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1477                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1478                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1479                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1480                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1481                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1482                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1483         } else {
1484                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1485                     (longlong_t)tx->tx_txg,
1486                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1487                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1488                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1489                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1490         }
1491
1492         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1493             NULL, ZIO_FLAG_CANFAIL);
1494         dsl_pool_config_enter(dp, FTAG);
1495         dsl_scan_visit(scn, tx);
1496         dsl_pool_config_exit(dp, FTAG);
1497         (void) zio_wait(scn->scn_zio_root);
1498         scn->scn_zio_root = NULL;
1499
1500         zfs_dbgmsg("visited %llu blocks in %llums",
1501             (longlong_t)scn->scn_visited_this_txg,
1502             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1503
1504         if (!scn->scn_pausing) {
1505                 /* finished with scan. */
1506                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1507                 dsl_scan_done(scn, B_TRUE, tx);
1508         }
1509
1510         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1511                 mutex_enter(&spa->spa_scrub_lock);
1512                 while (spa->spa_scrub_inflight > 0) {
1513                         cv_wait(&spa->spa_scrub_io_cv,
1514                             &spa->spa_scrub_lock);
1515                 }
1516                 mutex_exit(&spa->spa_scrub_lock);
1517         }
1518
1519         dsl_scan_sync_state(scn, tx);
1520 }
1521
1522 /*
1523  * This will start a new scan, or restart an existing one.
1524  */
1525 void
1526 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1527 {
1528         if (txg == 0) {
1529                 dmu_tx_t *tx;
1530                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1531                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1532
1533                 txg = dmu_tx_get_txg(tx);
1534                 dp->dp_scan->scn_restart_txg = txg;
1535                 dmu_tx_commit(tx);
1536         } else {
1537                 dp->dp_scan->scn_restart_txg = txg;
1538         }
1539         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1540 }
1541
1542 boolean_t
1543 dsl_scan_resilvering(dsl_pool_t *dp)
1544 {
1545         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1546             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1547 }
1548
1549 /*
1550  * scrub consumers
1551  */
1552
1553 static void
1554 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1555 {
1556         int i;
1557
1558         /*
1559          * If we resume after a reboot, zab will be NULL; don't record
1560          * incomplete stats in that case.
1561          */
1562         if (zab == NULL)
1563                 return;
1564
1565         for (i = 0; i < 4; i++) {
1566                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1567                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1568                 int equal;
1569                 zfs_blkstat_t *zb;
1570
1571                 if (t & DMU_OT_NEWTYPE)
1572                         t = DMU_OT_OTHER;
1573
1574                 zb = &zab->zab_type[l][t];
1575                 zb->zb_count++;
1576                 zb->zb_asize += BP_GET_ASIZE(bp);
1577                 zb->zb_lsize += BP_GET_LSIZE(bp);
1578                 zb->zb_psize += BP_GET_PSIZE(bp);
1579                 zb->zb_gangs += BP_COUNT_GANG(bp);
1580
1581                 switch (BP_GET_NDVAS(bp)) {
1582                 case 2:
1583                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1584                             DVA_GET_VDEV(&bp->blk_dva[1]))
1585                                 zb->zb_ditto_2_of_2_samevdev++;
1586                         break;
1587                 case 3:
1588                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1589                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1590                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1591                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1592                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1593                             DVA_GET_VDEV(&bp->blk_dva[2]));
1594                         if (equal == 1)
1595                                 zb->zb_ditto_2_of_3_samevdev++;
1596                         else if (equal == 3)
1597                                 zb->zb_ditto_3_of_3_samevdev++;
1598                         break;
1599                 }
1600         }
1601 }
1602
1603 static void
1604 dsl_scan_scrub_done(zio_t *zio)
1605 {
1606         spa_t *spa = zio->io_spa;
1607
1608         zio_data_buf_free(zio->io_data, zio->io_size);
1609
1610         mutex_enter(&spa->spa_scrub_lock);
1611         spa->spa_scrub_inflight--;
1612         cv_broadcast(&spa->spa_scrub_io_cv);
1613
1614         if (zio->io_error && (zio->io_error != ECKSUM ||
1615             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1616                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1617         }
1618         mutex_exit(&spa->spa_scrub_lock);
1619 }
1620
1621 static int
1622 dsl_scan_scrub_cb(dsl_pool_t *dp,
1623     const blkptr_t *bp, const zbookmark_t *zb)
1624 {
1625         dsl_scan_t *scn = dp->dp_scan;
1626         size_t size = BP_GET_PSIZE(bp);
1627         spa_t *spa = dp->dp_spa;
1628         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1629         boolean_t needs_io = B_FALSE;
1630         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1631         int zio_priority = 0;
1632         int scan_delay = 0;
1633         int d;
1634
1635         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1636             phys_birth >= scn->scn_phys.scn_max_txg)
1637                 return (0);
1638
1639         count_block(dp->dp_blkstats, bp);
1640
1641         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1642         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1643                 zio_flags |= ZIO_FLAG_SCRUB;
1644                 zio_priority = ZIO_PRIORITY_SCRUB;
1645                 needs_io = B_TRUE;
1646                 scan_delay = zfs_scrub_delay;
1647         } else {
1648                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1649                 zio_flags |= ZIO_FLAG_RESILVER;
1650                 zio_priority = ZIO_PRIORITY_RESILVER;
1651                 needs_io = B_FALSE;
1652                 scan_delay = zfs_resilver_delay;
1653         }
1654
1655         /* If it's an intent log block, failure is expected. */
1656         if (zb->zb_level == ZB_ZIL_LEVEL)
1657                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1658
1659         for (d = 0; d < BP_GET_NDVAS(bp); d++) {
1660                 vdev_t *vd = vdev_lookup_top(spa,
1661                     DVA_GET_VDEV(&bp->blk_dva[d]));
1662
1663                 /*
1664                  * Keep track of how much data we've examined so that
1665                  * zpool(1M) status can make useful progress reports.
1666                  */
1667                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1668                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1669
1670                 /* if it's a resilver, this may not be in the target range */
1671                 if (!needs_io) {
1672                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1673                                 /*
1674                                  * Gang members may be spread across multiple
1675                                  * vdevs, so the best estimate we have is the
1676                                  * scrub range, which has already been checked.
1677                                  * XXX -- it would be better to change our
1678                                  * allocation policy to ensure that all
1679                                  * gang members reside on the same vdev.
1680                                  */
1681                                 needs_io = B_TRUE;
1682                         } else {
1683                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1684                                     phys_birth, 1);
1685                         }
1686                 }
1687         }
1688
1689         if (needs_io && !zfs_no_scrub_io) {
1690                 vdev_t *rvd = spa->spa_root_vdev;
1691                 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1692                 void *data = zio_data_buf_alloc(size);
1693
1694                 mutex_enter(&spa->spa_scrub_lock);
1695                 while (spa->spa_scrub_inflight >= maxinflight)
1696                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1697                 spa->spa_scrub_inflight++;
1698                 mutex_exit(&spa->spa_scrub_lock);
1699
1700                 /*
1701                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1702                  * then throttle our workload to limit the impact of a scan.
1703                  */
1704                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1705                         delay(scan_delay);
1706
1707                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1708                     dsl_scan_scrub_done, NULL, zio_priority,
1709                     zio_flags, zb));
1710         }
1711
1712         /* do not relocate this block */
1713         return (0);
1714 }
1715
1716 int
1717 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1718 {
1719         spa_t *spa = dp->dp_spa;
1720
1721         /*
1722          * Purge all vdev caches and probe all devices.  We do this here
1723          * rather than in sync context because this requires a writer lock
1724          * on the spa_config lock, which we can't do from sync context.  The
1725          * spa_scrub_reopen flag indicates that vdev_open() should not
1726          * attempt to start another scrub.
1727          */
1728         spa_vdev_state_enter(spa, SCL_NONE);
1729         spa->spa_scrub_reopen = B_TRUE;
1730         vdev_reopen(spa->spa_root_vdev);
1731         spa->spa_scrub_reopen = B_FALSE;
1732         (void) spa_vdev_state_exit(spa, NULL, 0);
1733
1734         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1735             dsl_scan_setup_sync, &func, 0));
1736 }
1737
1738 #if defined(_KERNEL) && defined(HAVE_SPL)
1739 module_param(zfs_top_maxinflight, int, 0644);
1740 MODULE_PARM_DESC(zfs_top_maxinflight, "Max I/Os per top-level");
1741
1742 module_param(zfs_resilver_delay, int, 0644);
1743 MODULE_PARM_DESC(zfs_resilver_delay, "Number of ticks to delay resilver");
1744
1745 module_param(zfs_scrub_delay, int, 0644);
1746 MODULE_PARM_DESC(zfs_scrub_delay, "Number of ticks to delay scrub");
1747
1748 module_param(zfs_scan_idle, int, 0644);
1749 MODULE_PARM_DESC(zfs_scan_idle, "Idle window in clock ticks");
1750
1751 module_param(zfs_scan_min_time_ms, int, 0644);
1752 MODULE_PARM_DESC(zfs_scan_min_time_ms, "Min millisecs to scrub per txg");
1753
1754 module_param(zfs_free_min_time_ms, int, 0644);
1755 MODULE_PARM_DESC(zfs_free_min_time_ms, "Min millisecs to free per txg");
1756
1757 module_param(zfs_resilver_min_time_ms, int, 0644);
1758 MODULE_PARM_DESC(zfs_resilver_min_time_ms, "Min millisecs to resilver per txg");
1759
1760 module_param(zfs_no_scrub_io, int, 0644);
1761 MODULE_PARM_DESC(zfs_no_scrub_io, "Set to disable scrub I/O");
1762
1763 module_param(zfs_no_scrub_prefetch, int, 0644);
1764 MODULE_PARM_DESC(zfs_no_scrub_prefetch, "Set to disable scrub prefetching");
1765 #endif