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 dsl_syncfunc_t dsl_scan_cancel_sync;
  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 *arg1, void *arg2, dmu_tx_t *tx)
 154 {
 155         dsl_scan_t *scn = arg1;
 156
 157         if (scn->scn_phys.scn_state == DSS_SCANNING)
 158                 return (EBUSY);
 159
 160         return (0);
 161 }
 162
 163 /* ARGSUSED */
 164 static void
 165 dsl_scan_setup_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 166 {
 167         dsl_scan_t *scn = arg1;
 168         pool_scan_func_t *funcp = arg2;
 169         dmu_object_type_t ot = 0;
 170         dsl_pool_t *dp = scn->scn_dp;
 171         spa_t *spa = dp->dp_spa;
 172
 173         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 174         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 175         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 176         scn->scn_phys.scn_func = *funcp;
 177         scn->scn_phys.scn_state = DSS_SCANNING;
 178         scn->scn_phys.scn_min_txg = 0;
 179         scn->scn_phys.scn_max_txg = tx->tx_txg;
 180         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 181         scn->scn_phys.scn_start_time = gethrestime_sec();
 182         scn->scn_phys.scn_errors = 0;
 183         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 184         scn->scn_restart_txg = 0;
 185         spa_scan_stat_init(spa);
 186
 187         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 188                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 189
 190                 /* rewrite all disk labels */
 191                 vdev_config_dirty(spa->spa_root_vdev);
 192
 193                 if (vdev_resilver_needed(spa->spa_root_vdev,
 194                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 195                         spa_event_notify(spa, NULL, FM_EREPORT_ZFS_RESILVER_START);
 196                 } else {
 197                         spa_event_notify(spa, NULL, FM_EREPORT_ZFS_SCRUB_START);
 198                 }
 199
 200                 spa->spa_scrub_started = B_TRUE;
 201                 /*
 202                  * If this is an incremental scrub, limit the DDT scrub phase
 203                  * to just the auto-ditto class (for correctness); the rest
 204                  * of the scrub should go faster using top-down pruning.
 205                  */
 206                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 207                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 208
 209         }
 210
 211         /* back to the generic stuff */
 212
 213         if (dp->dp_blkstats == NULL) {
 214                 dp->dp_blkstats = kmem_alloc(sizeof (zfs_all_blkstats_t),
 215                     KM_PUSHPAGE | KM_NODEBUG);
 216         }
 217         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 218
 219         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 220                 ot = DMU_OT_ZAP_OTHER;
 221
 222         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 223             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 224
 225         dsl_scan_sync_state(scn, tx);
 226
 227         spa_history_log_internal(LOG_POOL_SCAN, spa, tx,
 228             "func=%u mintxg=%llu maxtxg=%llu",
 229             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 230 }
 231
 232 /* ARGSUSED */
 233 static void
 234 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 235 {
 236         static const char *old_names[] = {
 237                 "scrub_bookmark",
 238                 "scrub_ddt_bookmark",
 239                 "scrub_ddt_class_max",
 240                 "scrub_queue",
 241                 "scrub_min_txg",
 242                 "scrub_max_txg",
 243                 "scrub_func",
 244                 "scrub_errors",
 245                 NULL
 246         };
 247
 248         dsl_pool_t *dp = scn->scn_dp;
 249         spa_t *spa = dp->dp_spa;
 250         int i;
 251
 252         /* Remove any remnants of an old-style scrub. */
 253         for (i = 0; old_names[i]; i++) {
 254                 (void) zap_remove(dp->dp_meta_objset,
 255                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 256         }
 257
 258         if (scn->scn_phys.scn_queue_obj != 0) {
 259                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 260                     scn->scn_phys.scn_queue_obj, tx));
 261                 scn->scn_phys.scn_queue_obj = 0;
 262         }
 263
 264         /*
 265          * If we were "restarted" from a stopped state, don't bother
 266          * with anything else.
 267          */
 268         if (scn->scn_phys.scn_state != DSS_SCANNING)
 269                 return;
 270
 271         if (complete)
 272                 scn->scn_phys.scn_state = DSS_FINISHED;
 273         else
 274                 scn->scn_phys.scn_state = DSS_CANCELED;
 275
 276         spa_history_log_internal(LOG_POOL_SCAN_DONE, spa, tx,
 277             "complete=%u", complete);
 278
 279         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 280                 mutex_enter(&spa->spa_scrub_lock);
 281                 while (spa->spa_scrub_inflight > 0) {
 282                         cv_wait(&spa->spa_scrub_io_cv,
 283                             &spa->spa_scrub_lock);
 284                 }
 285                 mutex_exit(&spa->spa_scrub_lock);
 286                 spa->spa_scrub_started = B_FALSE;
 287                 spa->spa_scrub_active = B_FALSE;
 288
 289                 /*
 290                  * If the scrub/resilver completed, update all DTLs to
 291                  * reflect this.  Whether it succeeded or not, vacate
 292                  * all temporary scrub DTLs.
 293                  */
 294                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 295                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 296                 if (complete) {
 297                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 298                             FM_EREPORT_ZFS_RESILVER_FINISH :
 299                             FM_EREPORT_ZFS_SCRUB_FINISH);
 300                 }
 301                 spa_errlog_rotate(spa);
 302
 303                 /*
 304                  * We may have finished replacing a device.
 305                  * Let the async thread assess this and handle the detach.
 306                  */
 307                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 308         }
 309
 310         scn->scn_phys.scn_end_time = gethrestime_sec();
 311 }
 312
 313 /* ARGSUSED */
 314 static int
 315 dsl_scan_cancel_check(void *arg1, void *arg2, dmu_tx_t *tx)
 316 {
 317         dsl_scan_t *scn = arg1;
 318
 319         if (scn->scn_phys.scn_state != DSS_SCANNING)
 320                 return (ENOENT);
 321         return (0);
 322 }
 323
 324 /* ARGSUSED */
 325 static void
 326 dsl_scan_cancel_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 327 {
 328         dsl_scan_t *scn = arg1;
 329
 330         dsl_scan_done(scn, B_FALSE, tx);
 331         dsl_scan_sync_state(scn, tx);
 332 }
 333
 334 int
 335 dsl_scan_cancel(dsl_pool_t *dp)
 336 {
 337         boolean_t complete = B_FALSE;
 338         int err;
 339
 340         err = dsl_sync_task_do(dp, dsl_scan_cancel_check,
 341             dsl_scan_cancel_sync, dp->dp_scan, &complete, 3);
 342         return (err);
 343 }
 344
 345 static void dsl_scan_visitbp(blkptr_t *bp,
 346     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 347     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 348     dmu_tx_t *tx);
 349 inline __attribute__((always_inline)) static void dsl_scan_visitdnode(
 350     dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 351     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 352
 353 void
 354 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 355 {
 356         zio_free(dp->dp_spa, txg, bp);
 357 }
 358
 359 void
 360 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 361 {
 362         ASSERT(dsl_pool_sync_context(dp));
 363         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
 364 }
 365
 366 static uint64_t
 367 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 368 {
 369         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 370         if (dsl_dataset_is_snapshot(ds))
 371                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 372         return (smt);
 373 }
 374
 375 static void
 376 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 377 {
 378         VERIFY(0 == zap_update(scn->scn_dp->dp_meta_objset,
 379             DMU_POOL_DIRECTORY_OBJECT,
 380             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 381             &scn->scn_phys, tx));
 382 }
 383
 384 static boolean_t
 385 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 386 {
 387         uint64_t elapsed_nanosecs;
 388         int mintime;
 389
 390         /* we never skip user/group accounting objects */
 391         if (zb && (int64_t)zb->zb_object < 0)
 392                 return (B_FALSE);
 393
 394         if (scn->scn_pausing)
 395                 return (B_TRUE); /* we're already pausing */
 396
 397         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 398                 return (B_FALSE); /* we're resuming */
 399
 400         /* We only know how to resume from level-0 blocks. */
 401         if (zb && zb->zb_level != 0)
 402                 return (B_FALSE);
 403
 404         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 405             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 406         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 407         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 408             (elapsed_nanosecs / MICROSEC > mintime &&
 409             txg_sync_waiting(scn->scn_dp)) ||
 410             spa_shutting_down(scn->scn_dp->dp_spa)) {
 411                 if (zb) {
 412                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 413                             (longlong_t)zb->zb_objset,
 414                             (longlong_t)zb->zb_object,
 415                             (longlong_t)zb->zb_level,
 416                             (longlong_t)zb->zb_blkid);
 417                         scn->scn_phys.scn_bookmark = *zb;
 418                 }
 419                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 420                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 421                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 422                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 423                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 424                 scn->scn_pausing = B_TRUE;
 425                 return (B_TRUE);
 426         }
 427         return (B_FALSE);
 428 }
 429
 430 typedef struct zil_scan_arg {
 431         dsl_pool_t      *zsa_dp;
 432         zil_header_t    *zsa_zh;
 433 } zil_scan_arg_t;
 434
 435 /* ARGSUSED */
 436 static int
 437 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 438 {
 439         zil_scan_arg_t *zsa = arg;
 440         dsl_pool_t *dp = zsa->zsa_dp;
 441         dsl_scan_t *scn = dp->dp_scan;
 442         zil_header_t *zh = zsa->zsa_zh;
 443         zbookmark_t zb;
 444
 445         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 446                 return (0);
 447
 448         /*
 449          * One block ("stubby") can be allocated a long time ago; we
 450          * want to visit that one because it has been allocated
 451          * (on-disk) even if it hasn't been claimed (even though for
 452          * scrub there's nothing to do to it).
 453          */
 454         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 455                 return (0);
 456
 457         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 458             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 459
 460         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 461         return (0);
 462 }
 463
 464 /* ARGSUSED */
 465 static int
 466 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 467 {
 468         if (lrc->lrc_txtype == TX_WRITE) {
 469                 zil_scan_arg_t *zsa = arg;
 470                 dsl_pool_t *dp = zsa->zsa_dp;
 471                 dsl_scan_t *scn = dp->dp_scan;
 472                 zil_header_t *zh = zsa->zsa_zh;
 473                 lr_write_t *lr = (lr_write_t *)lrc;
 474                 blkptr_t *bp = &lr->lr_blkptr;
 475                 zbookmark_t zb;
 476
 477                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 478                         return (0);
 479
 480                 /*
 481                  * birth can be < claim_txg if this record's txg is
 482                  * already txg sync'ed (but this log block contains
 483                  * other records that are not synced)
 484                  */
 485                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 486                         return (0);
 487
 488                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 489                     lr->lr_foid, ZB_ZIL_LEVEL,
 490                     lr->lr_offset / BP_GET_LSIZE(bp));
 491
 492                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 493         }
 494         return (0);
 495 }
 496
 497 static void
 498 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 499 {
 500         uint64_t claim_txg = zh->zh_claim_txg;
 501         zil_scan_arg_t zsa = { dp, zh };
 502         zilog_t *zilog;
 503
 504         /*
 505          * We only want to visit blocks that have been claimed but not yet
 506          * replayed (or, in read-only mode, blocks that *would* be claimed).
 507          */
 508         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 509                 return;
 510
 511         zilog = zil_alloc(dp->dp_meta_objset, zh);
 512
 513         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 514             claim_txg);
 515
 516         zil_free(zilog);
 517 }
 518
 519 /* ARGSUSED */
 520 static void
 521 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 522     uint64_t objset, uint64_t object, uint64_t blkid)
 523 {
 524         zbookmark_t czb;
 525         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 526
 527         if (zfs_no_scrub_prefetch)
 528                 return;
 529
 530         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 531             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 532                 return;
 533
 534         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 535
 536         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 537             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 538             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 539 }
 540
 541 static boolean_t
 542 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 543     const zbookmark_t *zb)
 544 {
 545         /*
 546          * We never skip over user/group accounting objects (obj<0)
 547          */
 548         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 549             (int64_t)zb->zb_object >= 0) {
 550                 /*
 551                  * If we already visited this bp & everything below (in
 552                  * a prior txg sync), don't bother doing it again.
 553                  */
 554                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 555                         return (B_TRUE);
 556
 557                 /*
 558                  * If we found the block we're trying to resume from, or
 559                  * we went past it to a different object, zero it out to
 560                  * indicate that it's OK to start checking for pausing
 561                  * again.
 562                  */
 563                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 564                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 565                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 566                             (longlong_t)zb->zb_objset,
 567                             (longlong_t)zb->zb_object,
 568                             (longlong_t)zb->zb_level,
 569                             (longlong_t)zb->zb_blkid);
 570                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 571                 }
 572         }
 573         return (B_FALSE);
 574 }
 575
 576 /*
 577  * Return nonzero on i/o error.
 578  * Return new buf to write out in *bufp.
 579  */
 580 inline __attribute__((always_inline)) static int
 581 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 582     dnode_phys_t *dnp, const blkptr_t *bp,
 583     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 584 {
 585         dsl_pool_t *dp = scn->scn_dp;
 586         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 587         int err;
 588
 589         if (BP_GET_LEVEL(bp) > 0) {
 590                 uint32_t flags = ARC_WAIT;
 591                 int i;
 592                 blkptr_t *cbp;
 593                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 594
 595                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 596                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 597                 if (err) {
 598                         scn->scn_phys.scn_errors++;
 599                         return (err);
 600                 }
 601                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 602                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 603                             zb->zb_object, zb->zb_blkid * epb + i);
 604                 }
 605                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 606                         zbookmark_t czb;
 607
 608                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 609                             zb->zb_level - 1,
 610                             zb->zb_blkid * epb + i);
 611                         dsl_scan_visitbp(cbp, &czb, dnp,
 612                             *bufp, ds, scn, ostype, tx);
 613                 }
 614         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 615                 uint32_t flags = ARC_WAIT;
 616
 617                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 618                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 619                 if (err) {
 620                         scn->scn_phys.scn_errors++;
 621                         return (err);
 622                 }
 623         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 624                 uint32_t flags = ARC_WAIT;
 625                 dnode_phys_t *cdnp;
 626                 int i, j;
 627                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 628
 629                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 630                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 631                 if (err) {
 632                         scn->scn_phys.scn_errors++;
 633                         return (err);
 634                 }
 635                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 636                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 637                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 638                                 dsl_scan_prefetch(scn, *bufp, cbp,
 639                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 640                         }
 641                 }
 642                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 643                         dsl_scan_visitdnode(scn, ds, ostype,
 644                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 645                 }
 646
 647         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 648                 uint32_t flags = ARC_WAIT;
 649                 objset_phys_t *osp;
 650
 651                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 652                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 653                 if (err) {
 654                         scn->scn_phys.scn_errors++;
 655                         return (err);
 656                 }
 657
 658                 osp = (*bufp)->b_data;
 659
 660                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 661                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 662
 663                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 664                         /*
 665                          * We also always visit user/group accounting
 666                          * objects, and never skip them, even if we are
 667                          * pausing.  This is necessary so that the space
 668                          * deltas from this txg get integrated.
 669                          */
 670                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 671                             &osp->os_groupused_dnode, *bufp,
 672                             DMU_GROUPUSED_OBJECT, tx);
 673                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 674                             &osp->os_userused_dnode, *bufp,
 675                             DMU_USERUSED_OBJECT, tx);
 676                 }
 677         }
 678
 679         return (0);
 680 }
 681
 682 inline __attribute__((always_inline)) static void
 683 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 684     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 685     uint64_t object, dmu_tx_t *tx)
 686 {
 687         int j;
 688
 689         for (j = 0; j < dnp->dn_nblkptr; j++) {
 690                 zbookmark_t czb;
 691
 692                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 693                     dnp->dn_nlevels - 1, j);
 694                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 695                     &czb, dnp, buf, ds, scn, ostype, tx);
 696         }
 697
 698         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 699                 zbookmark_t czb;
 700                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 701                     0, DMU_SPILL_BLKID);
 702                 dsl_scan_visitbp(&dnp->dn_spill,
 703                     &czb, dnp, buf, ds, scn, ostype, tx);
 704         }
 705 }
 706
 707 /*
 708  * The arguments are in this order because mdb can only print the
 709  * first 5; we want them to be useful.
 710  */
 711 static void
 712 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 713     dnode_phys_t *dnp, arc_buf_t *pbuf,
 714     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 715     dmu_tx_t *tx)
 716 {
 717         dsl_pool_t *dp = scn->scn_dp;
 718         arc_buf_t *buf = NULL;
 719         blkptr_t *bp_toread;
 720
 721         bp_toread = kmem_alloc(sizeof (blkptr_t), KM_PUSHPAGE);
 722         *bp_toread = *bp;
 723
 724         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 725
 726         if (dsl_scan_check_pause(scn, zb))
 727                 goto out;
 728
 729         if (dsl_scan_check_resume(scn, dnp, zb))
 730                 goto out;
 731
 732         if (bp->blk_birth == 0)
 733                 goto out;
 734
 735         scn->scn_visited_this_txg++;
 736
 737         /*
 738          * This debugging is commented out to conserve stack space.  This
 739          * function is called recursively and the debugging addes several
 740          * bytes to the stack for each call.  It can be commented back in
 741          * if required to debug an issue in dsl_scan_visitbp().
 742          *
 743          * dprintf_bp(bp,
 744          *    "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 745          *    ds, ds ? ds->ds_object : 0,
 746          *    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 747          *    pbuf, bp);
 748          */
 749
 750         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 751                 goto out;
 752
 753         if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx,
 754             &buf) != 0)
 755                 goto out;
 756
 757         /*
 758          * If dsl_scan_ddt() has aready visited this block, it will have
 759          * already done any translations or scrubbing, so don't call the
 760          * callback again.
 761          */
 762         if (ddt_class_contains(dp->dp_spa,
 763             scn->scn_phys.scn_ddt_class_max, bp)) {
 764                 ASSERT(buf == NULL);
 765                 goto out;
 766         }
 767
 768         /*
 769          * If this block is from the future (after cur_max_txg), then we
 770          * are doing this on behalf of a deleted snapshot, and we will
 771          * revisit the future block on the next pass of this dataset.
 772          * Don't scan it now unless we need to because something
 773          * under it was modified.
 774          */
 775         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 776                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 777         }
 778         if (buf)
 779                 (void) arc_buf_remove_ref(buf, &buf);
 780 out:
 781         kmem_free(bp_toread, sizeof(blkptr_t));
 782 }
 783
 784 static void
 785 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 786     dmu_tx_t *tx)
 787 {
 788         zbookmark_t zb;
 789
 790         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 791             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 792         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 793             ds, scn, DMU_OST_NONE, tx);
 794
 795         dprintf_ds(ds, "finished scan%s", "");
 796 }
 797
 798 void
 799 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 800 {
 801         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 802         dsl_scan_t *scn = dp->dp_scan;
 803         uint64_t mintxg;
 804
 805         if (scn->scn_phys.scn_state != DSS_SCANNING)
 806                 return;
 807
 808         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 809                 if (dsl_dataset_is_snapshot(ds)) {
 810                         /* Note, scn_cur_{min,max}_txg stays the same. */
 811                         scn->scn_phys.scn_bookmark.zb_objset =
 812                             ds->ds_phys->ds_next_snap_obj;
 813                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 814                             "reset zb_objset to %llu",
 815                             (u_longlong_t)ds->ds_object,
 816                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 817                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 818                 } else {
 819                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 820                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 821                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 822                             "reset bookmark to -1,0,0,0",
 823                             (u_longlong_t)ds->ds_object);
 824                 }
 825         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 826             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 827                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 828                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 829                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 830                 if (dsl_dataset_is_snapshot(ds)) {
 831                         /*
 832                          * We keep the same mintxg; it could be >
 833                          * ds_creation_txg if the previous snapshot was
 834                          * deleted too.
 835                          */
 836                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 837                             scn->scn_phys.scn_queue_obj,
 838                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 839                         zfs_dbgmsg("destroying ds %llu; in queue; "
 840                             "replacing with %llu",
 841                             (u_longlong_t)ds->ds_object,
 842                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 843                 } else {
 844                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 845                             (u_longlong_t)ds->ds_object);
 846                 }
 847         } else {
 848                 zfs_dbgmsg("destroying ds %llu; ignoring",
 849                     (u_longlong_t)ds->ds_object);
 850         }
 851
 852         /*
 853          * dsl_scan_sync() should be called after this, and should sync
 854          * out our changed state, but just to be safe, do it here.
 855          */
 856         dsl_scan_sync_state(scn, tx);
 857 }
 858
 859 void
 860 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 861 {
 862         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 863         dsl_scan_t *scn = dp->dp_scan;
 864         uint64_t mintxg;
 865
 866         if (scn->scn_phys.scn_state != DSS_SCANNING)
 867                 return;
 868
 869         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 870
 871         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 872                 scn->scn_phys.scn_bookmark.zb_objset =
 873                     ds->ds_phys->ds_prev_snap_obj;
 874                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 875                     "reset zb_objset to %llu",
 876                     (u_longlong_t)ds->ds_object,
 877                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 878         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 879             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 880                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 881                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 882                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 883                     scn->scn_phys.scn_queue_obj,
 884                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 885                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 886                     "replacing with %llu",
 887                     (u_longlong_t)ds->ds_object,
 888                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 889         }
 890         dsl_scan_sync_state(scn, tx);
 891 }
 892
 893 void
 894 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 895 {
 896         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 897         dsl_scan_t *scn = dp->dp_scan;
 898         uint64_t mintxg;
 899
 900         if (scn->scn_phys.scn_state != DSS_SCANNING)
 901                 return;
 902
 903         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 904                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 905                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 906                     "reset zb_objset to %llu",
 907                     (u_longlong_t)ds1->ds_object,
 908                     (u_longlong_t)ds2->ds_object);
 909         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 910                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 911                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 912                     "reset zb_objset to %llu",
 913                     (u_longlong_t)ds2->ds_object,
 914                     (u_longlong_t)ds1->ds_object);
 915         }
 916
 917         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 918             ds1->ds_object, &mintxg) == 0) {
 919                 int err;
 920
 921                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 922                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 923                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 924                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 925                 err = zap_add_int_key(dp->dp_meta_objset,
 926                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 927                 VERIFY(err == 0 || err == EEXIST);
 928                 if (err == EEXIST) {
 929                         /* Both were there to begin with */
 930                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 931                             scn->scn_phys.scn_queue_obj,
 932                             ds1->ds_object, mintxg, tx));
 933                 }
 934                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 935                     "replacing with %llu",
 936                     (u_longlong_t)ds1->ds_object,
 937                     (u_longlong_t)ds2->ds_object);
 938         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 939             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 940                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 941                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 942                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 943                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 944                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 945                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 946                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 947                     "replacing with %llu",
 948                     (u_longlong_t)ds2->ds_object,
 949                     (u_longlong_t)ds1->ds_object);
 950         }
 951
 952         dsl_scan_sync_state(scn, tx);
 953 }
 954
 955 struct enqueue_clones_arg {
 956         dmu_tx_t *tx;
 957         uint64_t originobj;
 958 };
 959
 960 /* ARGSUSED */
 961 static int
 962 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 963 {
 964         struct enqueue_clones_arg *eca = arg;
 965         dsl_dataset_t *ds;
 966         int err;
 967         dsl_pool_t *dp = spa->spa_dsl_pool;
 968         dsl_scan_t *scn = dp->dp_scan;
 969
 970         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
 971         if (err)
 972                 return (err);
 973
 974         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
 975                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 976                         dsl_dataset_t *prev;
 977                         err = dsl_dataset_hold_obj(dp,
 978                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 979
 980                         dsl_dataset_rele(ds, FTAG);
 981                         if (err)
 982                                 return (err);
 983                         ds = prev;
 984                 }
 985                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 986                     scn->scn_phys.scn_queue_obj, ds->ds_object,
 987                     ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
 988         }
 989         dsl_dataset_rele(ds, FTAG);
 990         return (0);
 991 }
 992
 993 static void
 994 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
 995 {
 996         dsl_pool_t *dp = scn->scn_dp;
 997         dsl_dataset_t *ds;
 998         objset_t *os;
 999         char *dsname;
1000
1001         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1002
1003         if (dmu_objset_from_ds(ds, &os))
1004                 goto out;
1005
1006         /*
1007          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1008          * snapshots can have ZIL block pointers (which may be the same
1009          * BP as in the head), they must be ignored.  So we traverse the
1010          * ZIL here, rather than in scan_recurse(), because the regular
1011          * snapshot block-sharing rules don't apply to it.
1012          */
1013         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1014                 dsl_scan_zil(dp, &os->os_zil_header);
1015
1016         /*
1017          * Iterate over the bps in this ds.
1018          */
1019         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1020         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1021
1022         dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_PUSHPAGE);
1023         dsl_dataset_name(ds, dsname);
1024         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1025             "pausing=%u",
1026             (longlong_t)dsobj, dsname,
1027             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1028             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1029             (int)scn->scn_pausing);
1030         kmem_free(dsname, ZFS_MAXNAMELEN);
1031
1032         if (scn->scn_pausing)
1033                 goto out;
1034
1035         /*
1036          * We've finished this pass over this dataset.
1037          */
1038
1039         /*
1040          * If we did not completely visit this dataset, do another pass.
1041          */
1042         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1043                 zfs_dbgmsg("incomplete pass; visiting again");
1044                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1045                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1046                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1047                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1048                 goto out;
1049         }
1050
1051         /*
1052          * Add descendent datasets to work queue.
1053          */
1054         if (ds->ds_phys->ds_next_snap_obj != 0) {
1055                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1056                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1057                     ds->ds_phys->ds_creation_txg, tx) == 0);
1058         }
1059         if (ds->ds_phys->ds_num_children > 1) {
1060                 boolean_t usenext = B_FALSE;
1061                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1062                         uint64_t count;
1063                         /*
1064                          * A bug in a previous version of the code could
1065                          * cause upgrade_clones_cb() to not set
1066                          * ds_next_snap_obj when it should, leading to a
1067                          * missing entry.  Therefore we can only use the
1068                          * next_clones_obj when its count is correct.
1069                          */
1070                         int err = zap_count(dp->dp_meta_objset,
1071                             ds->ds_phys->ds_next_clones_obj, &count);
1072                         if (err == 0 &&
1073                             count == ds->ds_phys->ds_num_children - 1)
1074                                 usenext = B_TRUE;
1075                 }
1076
1077                 if (usenext) {
1078                         VERIFY(zap_join_key(dp->dp_meta_objset,
1079                             ds->ds_phys->ds_next_clones_obj,
1080                             scn->scn_phys.scn_queue_obj,
1081                             ds->ds_phys->ds_creation_txg, tx) == 0);
1082                 } else {
1083                         struct enqueue_clones_arg eca;
1084                         eca.tx = tx;
1085                         eca.originobj = ds->ds_object;
1086
1087                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
1088                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
1089                 }
1090         }
1091
1092 out:
1093         dsl_dataset_rele(ds, FTAG);
1094 }
1095
1096 /* ARGSUSED */
1097 static int
1098 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1099 {
1100         dmu_tx_t *tx = arg;
1101         dsl_dataset_t *ds;
1102         int err;
1103         dsl_pool_t *dp = spa->spa_dsl_pool;
1104         dsl_scan_t *scn = dp->dp_scan;
1105
1106         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1107         if (err)
1108                 return (err);
1109
1110         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1111                 dsl_dataset_t *prev;
1112                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1113                     FTAG, &prev);
1114                 if (err) {
1115                         dsl_dataset_rele(ds, FTAG);
1116                         return (err);
1117                 }
1118
1119                 /*
1120                  * If this is a clone, we don't need to worry about it for now.
1121                  */
1122                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1123                         dsl_dataset_rele(ds, FTAG);
1124                         dsl_dataset_rele(prev, FTAG);
1125                         return (0);
1126                 }
1127                 dsl_dataset_rele(ds, FTAG);
1128                 ds = prev;
1129         }
1130
1131         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1132             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1133         dsl_dataset_rele(ds, FTAG);
1134         return (0);
1135 }
1136
1137 /*
1138  * Scrub/dedup interaction.
1139  *
1140  * If there are N references to a deduped block, we don't want to scrub it
1141  * N times -- ideally, we should scrub it exactly once.
1142  *
1143  * We leverage the fact that the dde's replication class (enum ddt_class)
1144  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1145  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1146  *
1147  * To prevent excess scrubbing, the scrub begins by walking the DDT
1148  * to find all blocks with refcnt > 1, and scrubs each of these once.
1149  * Since there are two replication classes which contain blocks with
1150  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1151  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1152  *
1153  * There would be nothing more to say if a block's refcnt couldn't change
1154  * during a scrub, but of course it can so we must account for changes
1155  * in a block's replication class.
1156  *
1157  * Here's an example of what can occur:
1158  *
1159  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1160  * when visited during the top-down scrub phase, it will be scrubbed twice.
1161  * This negates our scrub optimization, but is otherwise harmless.
1162  *
1163  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1164  * on each visit during the top-down scrub phase, it will never be scrubbed.
1165  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1166  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1167  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1168  * while a scrub is in progress, it scrubs the block right then.
1169  */
1170 static void
1171 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1172 {
1173         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1174         ddt_entry_t dde;
1175         int error;
1176         uint64_t n = 0;
1177
1178         bzero(&dde, sizeof (ddt_entry_t));
1179
1180         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1181                 ddt_t *ddt;
1182
1183                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1184                         break;
1185                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1186                     (longlong_t)ddb->ddb_class,
1187                     (longlong_t)ddb->ddb_type,
1188                     (longlong_t)ddb->ddb_checksum,
1189                     (longlong_t)ddb->ddb_cursor);
1190
1191                 /* There should be no pending changes to the dedup table */
1192                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1193                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1194
1195                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1196                 n++;
1197
1198                 if (dsl_scan_check_pause(scn, NULL))
1199                         break;
1200         }
1201
1202         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1203             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1204             (int)scn->scn_pausing);
1205
1206         ASSERT(error == 0 || error == ENOENT);
1207         ASSERT(error != ENOENT ||
1208             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1209 }
1210
1211 /* ARGSUSED */
1212 void
1213 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1214     ddt_entry_t *dde, dmu_tx_t *tx)
1215 {
1216         const ddt_key_t *ddk = &dde->dde_key;
1217         ddt_phys_t *ddp = dde->dde_phys;
1218         blkptr_t bp;
1219         zbookmark_t zb = { 0 };
1220         int p;
1221
1222         if (scn->scn_phys.scn_state != DSS_SCANNING)
1223                 return;
1224
1225         for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1226                 if (ddp->ddp_phys_birth == 0 ||
1227                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1228                         continue;
1229                 ddt_bp_create(checksum, ddk, ddp, &bp);
1230
1231                 scn->scn_visited_this_txg++;
1232                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1233         }
1234 }
1235
1236 static void
1237 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1238 {
1239         dsl_pool_t *dp = scn->scn_dp;
1240         zap_cursor_t *zc;
1241         zap_attribute_t *za;
1242
1243         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1244             scn->scn_phys.scn_ddt_class_max) {
1245                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1246                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1247                 dsl_scan_ddt(scn, tx);
1248                 if (scn->scn_pausing)
1249                         return;
1250         }
1251
1252         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1253                 /* First do the MOS & ORIGIN */
1254
1255                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1256                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1257                 dsl_scan_visit_rootbp(scn, NULL,
1258                     &dp->dp_meta_rootbp, tx);
1259                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1260                 if (scn->scn_pausing)
1261                         return;
1262
1263                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1264                         VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
1265                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
1266                 } else {
1267                         dsl_scan_visitds(scn,
1268                             dp->dp_origin_snap->ds_object, tx);
1269                 }
1270                 ASSERT(!scn->scn_pausing);
1271         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1272             ZB_DESTROYED_OBJSET) {
1273                 /*
1274                  * If we were paused, continue from here.  Note if the
1275                  * ds we were paused on was deleted, the zb_objset may
1276                  * be -1, so we will skip this and find a new objset
1277                  * below.
1278                  */
1279                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1280                 if (scn->scn_pausing)
1281                         return;
1282         }
1283
1284         /*
1285          * In case we were paused right at the end of the ds, zero the
1286          * bookmark so we don't think that we're still trying to resume.
1287          */
1288         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1289         zc = kmem_alloc(sizeof(zap_cursor_t), KM_PUSHPAGE);
1290         za = kmem_alloc(sizeof(zap_attribute_t), KM_PUSHPAGE);
1291
1292         /* keep pulling things out of the zap-object-as-queue */
1293         while (zap_cursor_init(zc, dp->dp_meta_objset,
1294             scn->scn_phys.scn_queue_obj),
1295             zap_cursor_retrieve(zc, za) == 0) {
1296                 dsl_dataset_t *ds;
1297                 uint64_t dsobj;
1298
1299                 dsobj = strtonum(za->za_name, NULL);
1300                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1301                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1302
1303                 /* Set up min/max txg */
1304                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1305                 if (za->za_first_integer != 0) {
1306                         scn->scn_phys.scn_cur_min_txg =
1307                             MAX(scn->scn_phys.scn_min_txg,
1308                             za->za_first_integer);
1309                 } else {
1310                         scn->scn_phys.scn_cur_min_txg =
1311                             MAX(scn->scn_phys.scn_min_txg,
1312                             ds->ds_phys->ds_prev_snap_txg);
1313                 }
1314                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1315                 dsl_dataset_rele(ds, FTAG);
1316
1317                 dsl_scan_visitds(scn, dsobj, tx);
1318                 zap_cursor_fini(zc);
1319                 if (scn->scn_pausing)
1320                         goto out;
1321         }
1322         zap_cursor_fini(zc);
1323 out:
1324         kmem_free(za, sizeof(zap_attribute_t));
1325         kmem_free(zc, sizeof(zap_cursor_t));
1326 }
1327
1328 static boolean_t
1329 dsl_scan_free_should_pause(dsl_scan_t *scn)
1330 {
1331         uint64_t elapsed_nanosecs;
1332
1333         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1334         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1335             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1336             txg_sync_waiting(scn->scn_dp)) ||
1337             spa_shutting_down(scn->scn_dp->dp_spa));
1338 }
1339
1340 static int
1341 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1342 {
1343         dsl_scan_t *scn = arg;
1344
1345         if (!scn->scn_is_bptree ||
1346             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1347                 if (dsl_scan_free_should_pause(scn))
1348                         return (ERESTART);
1349         }
1350
1351         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1352             dmu_tx_get_txg(tx), bp, 0));
1353         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1354             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1355             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1356         scn->scn_visited_this_txg++;
1357         return (0);
1358 }
1359
1360 boolean_t
1361 dsl_scan_active(dsl_scan_t *scn)
1362 {
1363         spa_t *spa = scn->scn_dp->dp_spa;
1364         uint64_t used = 0, comp, uncomp;
1365
1366         if (spa->spa_load_state != SPA_LOAD_NONE)
1367                 return (B_FALSE);
1368         if (spa_shutting_down(spa))
1369                 return (B_FALSE);
1370
1371         if (scn->scn_phys.scn_state == DSS_SCANNING)
1372                 return (B_TRUE);
1373
1374         if (spa_feature_is_active(spa,
1375             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1376                 return (B_TRUE);
1377         }
1378         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1379                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1380                     &used, &comp, &uncomp);
1381         }
1382         return (used != 0);
1383 }
1384
1385 void
1386 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1387 {
1388         dsl_scan_t *scn = dp->dp_scan;
1389         spa_t *spa = dp->dp_spa;
1390         int err;
1391
1392         /*
1393          * Check for scn_restart_txg before checking spa_load_state, so
1394          * that we can restart an old-style scan while the pool is being
1395          * imported (see dsl_scan_init).
1396          */
1397         if (scn->scn_restart_txg != 0 &&
1398             scn->scn_restart_txg <= tx->tx_txg) {
1399                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1400                 dsl_scan_done(scn, B_FALSE, tx);
1401                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1402                         func = POOL_SCAN_RESILVER;
1403                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1404                     func, tx->tx_txg);
1405                 dsl_scan_setup_sync(scn, &func, tx);
1406         }
1407
1408         if (!dsl_scan_active(scn) ||
1409             spa_sync_pass(dp->dp_spa) > 1)
1410                 return;
1411
1412         scn->scn_visited_this_txg = 0;
1413         scn->scn_pausing = B_FALSE;
1414         scn->scn_sync_start_time = gethrtime();
1415         spa->spa_scrub_active = B_TRUE;
1416
1417         /*
1418          * First process the free list.  If we pause the free, don't do
1419          * any scanning.  This ensures that there is no free list when
1420          * we are scanning, so the scan code doesn't have to worry about
1421          * traversing it.
1422          */
1423         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1424                 scn->scn_is_bptree = B_FALSE;
1425                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1426                     NULL, ZIO_FLAG_MUSTSUCCEED);
1427                 err = bpobj_iterate(&dp->dp_free_bpobj,
1428                     dsl_scan_free_block_cb, scn, tx);
1429                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1430
1431                 if (err == 0 && spa_feature_is_active(spa,
1432                     &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1433                         scn->scn_is_bptree = B_TRUE;
1434                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1435                             NULL, ZIO_FLAG_MUSTSUCCEED);
1436                         err = bptree_iterate(dp->dp_meta_objset,
1437                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1438                             scn, tx);
1439                         VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1440                         if (err != 0)
1441                                 return;
1442
1443                         /* disable async destroy feature */
1444                         spa_feature_decr(spa,
1445                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY], tx);
1446                         ASSERT(!spa_feature_is_active(spa,
1447                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]));
1448                         VERIFY3U(0, ==, zap_remove(dp->dp_meta_objset,
1449                             DMU_POOL_DIRECTORY_OBJECT,
1450                             DMU_POOL_BPTREE_OBJ, tx));
1451                         VERIFY3U(0, ==, bptree_free(dp->dp_meta_objset,
1452                             dp->dp_bptree_obj, tx));
1453                         dp->dp_bptree_obj = 0;
1454                 }
1455                 if (scn->scn_visited_this_txg) {
1456                         zfs_dbgmsg("freed %llu blocks in %llums from "
1457                             "free_bpobj/bptree txg %llu",
1458                             (longlong_t)scn->scn_visited_this_txg,
1459                             (longlong_t)
1460                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1461                             (longlong_t)tx->tx_txg);
1462                         scn->scn_visited_this_txg = 0;
1463                         /*
1464                          * Re-sync the ddt so that we can further modify
1465                          * it when doing bprewrite.
1466                          */
1467                         ddt_sync(spa, tx->tx_txg);
1468                 }
1469                 if (err == ERESTART)
1470                         return;
1471         }
1472
1473         if (scn->scn_phys.scn_state != DSS_SCANNING)
1474                 return;
1475
1476         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1477             scn->scn_phys.scn_ddt_class_max) {
1478                 zfs_dbgmsg("doing scan sync txg %llu; "
1479                     "ddt bm=%llu/%llu/%llu/%llx",
1480                     (longlong_t)tx->tx_txg,
1481                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1482                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1483                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1484                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1485                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1486                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1487                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1488                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1489         } else {
1490                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1491                     (longlong_t)tx->tx_txg,
1492                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1493                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1494                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1495                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1496         }
1497
1498         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1499             NULL, ZIO_FLAG_CANFAIL);
1500         dsl_scan_visit(scn, tx);
1501         (void) zio_wait(scn->scn_zio_root);
1502         scn->scn_zio_root = NULL;
1503
1504         zfs_dbgmsg("visited %llu blocks in %llums",
1505             (longlong_t)scn->scn_visited_this_txg,
1506             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1507
1508         if (!scn->scn_pausing) {
1509                 /* finished with scan. */
1510                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1511                 dsl_scan_done(scn, B_TRUE, tx);
1512         }
1513
1514         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1515                 mutex_enter(&spa->spa_scrub_lock);
1516                 while (spa->spa_scrub_inflight > 0) {
1517                         cv_wait(&spa->spa_scrub_io_cv,
1518                             &spa->spa_scrub_lock);
1519                 }
1520                 mutex_exit(&spa->spa_scrub_lock);
1521         }
1522
1523         dsl_scan_sync_state(scn, tx);
1524 }
1525
1526 /*
1527  * This will start a new scan, or restart an existing one.
1528  */
1529 void
1530 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1531 {
1532         if (txg == 0) {
1533                 dmu_tx_t *tx;
1534                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1535                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1536
1537                 txg = dmu_tx_get_txg(tx);
1538                 dp->dp_scan->scn_restart_txg = txg;
1539                 dmu_tx_commit(tx);
1540         } else {
1541                 dp->dp_scan->scn_restart_txg = txg;
1542         }
1543         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1544 }
1545
1546 boolean_t
1547 dsl_scan_resilvering(dsl_pool_t *dp)
1548 {
1549         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1550             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1551 }
1552
1553 /*
1554  * scrub consumers
1555  */
1556
1557 static void
1558 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1559 {
1560         int i;
1561
1562         /*
1563          * If we resume after a reboot, zab will be NULL; don't record
1564          * incomplete stats in that case.
1565          */
1566         if (zab == NULL)
1567                 return;
1568
1569         for (i = 0; i < 4; i++) {
1570                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1571                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1572                 int equal;
1573                 zfs_blkstat_t *zb;
1574
1575                 if (t & DMU_OT_NEWTYPE)
1576                         t = DMU_OT_OTHER;
1577
1578                 zb = &zab->zab_type[l][t];
1579                 zb->zb_count++;
1580                 zb->zb_asize += BP_GET_ASIZE(bp);
1581                 zb->zb_lsize += BP_GET_LSIZE(bp);
1582                 zb->zb_psize += BP_GET_PSIZE(bp);
1583                 zb->zb_gangs += BP_COUNT_GANG(bp);
1584
1585                 switch (BP_GET_NDVAS(bp)) {
1586                 case 2:
1587                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1588                             DVA_GET_VDEV(&bp->blk_dva[1]))
1589                                 zb->zb_ditto_2_of_2_samevdev++;
1590                         break;
1591                 case 3:
1592                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1593                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1594                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1595                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1596                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1597                             DVA_GET_VDEV(&bp->blk_dva[2]));
1598                         if (equal == 1)
1599                                 zb->zb_ditto_2_of_3_samevdev++;
1600                         else if (equal == 3)
1601                                 zb->zb_ditto_3_of_3_samevdev++;
1602                         break;
1603                 }
1604         }
1605 }
1606
1607 static void
1608 dsl_scan_scrub_done(zio_t *zio)
1609 {
1610         spa_t *spa = zio->io_spa;
1611
1612         zio_data_buf_free(zio->io_data, zio->io_size);
1613
1614         mutex_enter(&spa->spa_scrub_lock);
1615         spa->spa_scrub_inflight--;
1616         cv_broadcast(&spa->spa_scrub_io_cv);
1617
1618         if (zio->io_error && (zio->io_error != ECKSUM ||
1619             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1620                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1621         }
1622         mutex_exit(&spa->spa_scrub_lock);
1623 }
1624
1625 static int
1626 dsl_scan_scrub_cb(dsl_pool_t *dp,
1627     const blkptr_t *bp, const zbookmark_t *zb)
1628 {
1629         dsl_scan_t *scn = dp->dp_scan;
1630         size_t size = BP_GET_PSIZE(bp);
1631         spa_t *spa = dp->dp_spa;
1632         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1633         boolean_t needs_io = B_FALSE;
1634         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1635         int zio_priority = 0;
1636         int scan_delay = 0;
1637         int d;
1638
1639         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1640             phys_birth >= scn->scn_phys.scn_max_txg)
1641                 return (0);
1642
1643         count_block(dp->dp_blkstats, bp);
1644
1645         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1646         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1647                 zio_flags |= ZIO_FLAG_SCRUB;
1648                 zio_priority = ZIO_PRIORITY_SCRUB;
1649                 needs_io = B_TRUE;
1650                 scan_delay = zfs_scrub_delay;
1651         } else if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
1652                 zio_flags |= ZIO_FLAG_RESILVER;
1653                 zio_priority = ZIO_PRIORITY_RESILVER;
1654                 needs_io = B_FALSE;
1655                 scan_delay = zfs_resilver_delay;
1656         }
1657
1658         /* If it's an intent log block, failure is expected. */
1659         if (zb->zb_level == ZB_ZIL_LEVEL)
1660                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1661
1662         for (d = 0; d < BP_GET_NDVAS(bp); d++) {
1663                 vdev_t *vd = vdev_lookup_top(spa,
1664                     DVA_GET_VDEV(&bp->blk_dva[d]));
1665
1666                 /*
1667                  * Keep track of how much data we've examined so that
1668                  * zpool(1M) status can make useful progress reports.
1669                  */
1670                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1671                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1672
1673                 /* if it's a resilver, this may not be in the target range */
1674                 if (!needs_io) {
1675                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1676                                 /*
1677                                  * Gang members may be spread across multiple
1678                                  * vdevs, so the best estimate we have is the
1679                                  * scrub range, which has already been checked.
1680                                  * XXX -- it would be better to change our
1681                                  * allocation policy to ensure that all
1682                                  * gang members reside on the same vdev.
1683                                  */
1684                                 needs_io = B_TRUE;
1685                         } else {
1686                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1687                                     phys_birth, 1);
1688                         }
1689                 }
1690         }
1691
1692         if (needs_io && !zfs_no_scrub_io) {
1693                 vdev_t *rvd = spa->spa_root_vdev;
1694                 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1695                 void *data = zio_data_buf_alloc(size);
1696
1697                 mutex_enter(&spa->spa_scrub_lock);
1698                 while (spa->spa_scrub_inflight >= maxinflight)
1699                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1700                 spa->spa_scrub_inflight++;
1701                 mutex_exit(&spa->spa_scrub_lock);
1702
1703                 /*
1704                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1705                  * then throttle our workload to limit the impact of a scan.
1706                  */
1707                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1708                         delay(scan_delay);
1709
1710                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1711                     dsl_scan_scrub_done, NULL, zio_priority,
1712                     zio_flags, zb));
1713         }
1714
1715         /* do not relocate this block */
1716         return (0);
1717 }
1718
1719 int
1720 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1721 {
1722         spa_t *spa = dp->dp_spa;
1723
1724         /*
1725          * Purge all vdev caches and probe all devices.  We do this here
1726          * rather than in sync context because this requires a writer lock
1727          * on the spa_config lock, which we can't do from sync context.  The
1728          * spa_scrub_reopen flag indicates that vdev_open() should not
1729          * attempt to start another scrub.
1730          */
1731         spa_vdev_state_enter(spa, SCL_NONE);
1732         spa->spa_scrub_reopen = B_TRUE;
1733         vdev_reopen(spa->spa_root_vdev);
1734         spa->spa_scrub_reopen = B_FALSE;
1735         (void) spa_vdev_state_exit(spa, NULL, 0);
1736
1737         return (dsl_sync_task_do(dp, dsl_scan_setup_check,
1738             dsl_scan_setup_sync, dp->dp_scan, &func, 0));
1739 }
1740
1741 #if defined(_KERNEL) && defined(HAVE_SPL)
1742 module_param(zfs_top_maxinflight, int, 0644);
1743 MODULE_PARM_DESC(zfs_top_maxinflight, "Max I/Os per top-level");
1744
1745 module_param(zfs_resilver_delay, int, 0644);
1746 MODULE_PARM_DESC(zfs_resilver_delay, "Number of ticks to delay resilver");
1747
1748 module_param(zfs_scrub_delay, int, 0644);
1749 MODULE_PARM_DESC(zfs_scrub_delay, "Number of ticks to delay scrub");
1750
1751 module_param(zfs_scan_idle, int, 0644);
1752 MODULE_PARM_DESC(zfs_scan_idle, "Idle window in clock ticks");
1753
1754 module_param(zfs_scan_min_time_ms, int, 0644);
1755 MODULE_PARM_DESC(zfs_scan_min_time_ms, "Min millisecs to scrub per txg");
1756
1757 module_param(zfs_free_min_time_ms, int, 0644);
1758 MODULE_PARM_DESC(zfs_free_min_time_ms, "Min millisecs to free per txg");
1759
1760 module_param(zfs_resilver_min_time_ms, int, 0644);
1761 MODULE_PARM_DESC(zfs_resilver_min_time_ms, "Min millisecs to resilver per txg");
1762
1763 module_param(zfs_no_scrub_io, int, 0644);
1764 MODULE_PARM_DESC(zfs_no_scrub_io, "Set to disable scrub I/O");
1765
1766 module_param(zfs_no_scrub_prefetch, int, 0644);
1767 MODULE_PARM_DESC(zfs_no_scrub_prefetch, "Set to disable scrub prefetching");
1768 #endif