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) 2013 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         /*
  95          * It's possible that we're resuming a scan after a reboot so
  96          * make sure that the scan_async_destroying flag is initialized
  97          * appropriately.
  98          */
  99         ASSERT(!scn->scn_async_destroying);
 100         scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
 101             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]);
 102
 103         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 104             "scrub_func", sizeof (uint64_t), 1, &f);
 105         if (err == 0) {
 106                 /*
 107                  * There was an old-style scrub in progress.  Restart a
 108                  * new-style scrub from the beginning.
 109                  */
 110                 scn->scn_restart_txg = txg;
 111                 zfs_dbgmsg("old-style scrub was in progress; "
 112                     "restarting new-style scrub in txg %llu",
 113                     scn->scn_restart_txg);
 114
 115                 /*
 116                  * Load the queue obj from the old location so that it
 117                  * can be freed by dsl_scan_done().
 118                  */
 119                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 120                     "scrub_queue", sizeof (uint64_t), 1,
 121                     &scn->scn_phys.scn_queue_obj);
 122         } else {
 123                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 124                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 125                     &scn->scn_phys);
 126                 if (err == ENOENT)
 127                         return (0);
 128                 else if (err)
 129                         return (err);
 130
 131                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 132                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 133                         /*
 134                          * A new-type scrub was in progress on an old
 135                          * pool, and the pool was accessed by old
 136                          * software.  Restart from the beginning, since
 137                          * the old software may have changed the pool in
 138                          * the meantime.
 139                          */
 140                         scn->scn_restart_txg = txg;
 141                         zfs_dbgmsg("new-style scrub was modified "
 142                             "by old software; restarting in txg %llu",
 143                             scn->scn_restart_txg);
 144                 }
 145         }
 146
 147         spa_scan_stat_init(spa);
 148         return (0);
 149 }
 150
 151 void
 152 dsl_scan_fini(dsl_pool_t *dp)
 153 {
 154         if (dp->dp_scan) {
 155                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 156                 dp->dp_scan = NULL;
 157         }
 158 }
 159
 160 /* ARGSUSED */
 161 static int
 162 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 163 {
 164         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 165
 166         if (scn->scn_phys.scn_state == DSS_SCANNING)
 167                 return (SET_ERROR(EBUSY));
 168
 169         return (0);
 170 }
 171
 172 static void
 173 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 174 {
 175         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 176         pool_scan_func_t *funcp = arg;
 177         dmu_object_type_t ot = 0;
 178         dsl_pool_t *dp = scn->scn_dp;
 179         spa_t *spa = dp->dp_spa;
 180
 181         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 182         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 183         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 184         scn->scn_phys.scn_func = *funcp;
 185         scn->scn_phys.scn_state = DSS_SCANNING;
 186         scn->scn_phys.scn_min_txg = 0;
 187         scn->scn_phys.scn_max_txg = tx->tx_txg;
 188         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 189         scn->scn_phys.scn_start_time = gethrestime_sec();
 190         scn->scn_phys.scn_errors = 0;
 191         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 192         scn->scn_restart_txg = 0;
 193         spa_scan_stat_init(spa);
 194
 195         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 196                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 197
 198                 /* rewrite all disk labels */
 199                 vdev_config_dirty(spa->spa_root_vdev);
 200
 201                 if (vdev_resilver_needed(spa->spa_root_vdev,
 202                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 203                         spa_event_notify(spa, NULL, FM_EREPORT_ZFS_RESILVER_START);
 204                 } else {
 205                         spa_event_notify(spa, NULL, FM_EREPORT_ZFS_SCRUB_START);
 206                 }
 207
 208                 spa->spa_scrub_started = B_TRUE;
 209                 /*
 210                  * If this is an incremental scrub, limit the DDT scrub phase
 211                  * to just the auto-ditto class (for correctness); the rest
 212                  * of the scrub should go faster using top-down pruning.
 213                  */
 214                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 215                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 216
 217         }
 218
 219         /* back to the generic stuff */
 220
 221         if (dp->dp_blkstats == NULL) {
 222                 dp->dp_blkstats = kmem_alloc(sizeof (zfs_all_blkstats_t),
 223                     KM_PUSHPAGE | KM_NODEBUG);
 224         }
 225         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 226
 227         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 228                 ot = DMU_OT_ZAP_OTHER;
 229
 230         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 231             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 232
 233         dsl_scan_sync_state(scn, tx);
 234
 235         spa_history_log_internal(spa, "scan setup", tx,
 236             "func=%u mintxg=%llu maxtxg=%llu",
 237             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 238 }
 239
 240 /* ARGSUSED */
 241 static void
 242 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 243 {
 244         static const char *old_names[] = {
 245                 "scrub_bookmark",
 246                 "scrub_ddt_bookmark",
 247                 "scrub_ddt_class_max",
 248                 "scrub_queue",
 249                 "scrub_min_txg",
 250                 "scrub_max_txg",
 251                 "scrub_func",
 252                 "scrub_errors",
 253                 NULL
 254         };
 255
 256         dsl_pool_t *dp = scn->scn_dp;
 257         spa_t *spa = dp->dp_spa;
 258         int i;
 259
 260         /* Remove any remnants of an old-style scrub. */
 261         for (i = 0; old_names[i]; i++) {
 262                 (void) zap_remove(dp->dp_meta_objset,
 263                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 264         }
 265
 266         if (scn->scn_phys.scn_queue_obj != 0) {
 267                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 268                     scn->scn_phys.scn_queue_obj, tx));
 269                 scn->scn_phys.scn_queue_obj = 0;
 270         }
 271
 272         /*
 273          * If we were "restarted" from a stopped state, don't bother
 274          * with anything else.
 275          */
 276         if (scn->scn_phys.scn_state != DSS_SCANNING)
 277                 return;
 278
 279         if (complete)
 280                 scn->scn_phys.scn_state = DSS_FINISHED;
 281         else
 282                 scn->scn_phys.scn_state = DSS_CANCELED;
 283
 284         spa_history_log_internal(spa, "scan done", tx,
 285             "complete=%u", complete);
 286
 287         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 288                 mutex_enter(&spa->spa_scrub_lock);
 289                 while (spa->spa_scrub_inflight > 0) {
 290                         cv_wait(&spa->spa_scrub_io_cv,
 291                             &spa->spa_scrub_lock);
 292                 }
 293                 mutex_exit(&spa->spa_scrub_lock);
 294                 spa->spa_scrub_started = B_FALSE;
 295                 spa->spa_scrub_active = B_FALSE;
 296
 297                 /*
 298                  * If the scrub/resilver completed, update all DTLs to
 299                  * reflect this.  Whether it succeeded or not, vacate
 300                  * all temporary scrub DTLs.
 301                  */
 302                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 303                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 304                 if (complete) {
 305                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 306                             FM_EREPORT_ZFS_RESILVER_FINISH :
 307                             FM_EREPORT_ZFS_SCRUB_FINISH);
 308                 }
 309                 spa_errlog_rotate(spa);
 310
 311                 /*
 312                  * We may have finished replacing a device.
 313                  * Let the async thread assess this and handle the detach.
 314                  */
 315                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 316         }
 317
 318         scn->scn_phys.scn_end_time = gethrestime_sec();
 319 }
 320
 321 /* ARGSUSED */
 322 static int
 323 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 324 {
 325         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 326
 327         if (scn->scn_phys.scn_state != DSS_SCANNING)
 328                 return (SET_ERROR(ENOENT));
 329         return (0);
 330 }
 331
 332 /* ARGSUSED */
 333 static void
 334 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 335 {
 336         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 337
 338         dsl_scan_done(scn, B_FALSE, tx);
 339         dsl_scan_sync_state(scn, tx);
 340 }
 341
 342 int
 343 dsl_scan_cancel(dsl_pool_t *dp)
 344 {
 345         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 346             dsl_scan_cancel_sync, NULL, 3));
 347 }
 348
 349 static void dsl_scan_visitbp(blkptr_t *bp,
 350     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 351     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 352     dmu_tx_t *tx);
 353 inline __attribute__((always_inline)) static void dsl_scan_visitdnode(
 354     dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 355     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 356
 357 void
 358 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 359 {
 360         zio_free(dp->dp_spa, txg, bp);
 361 }
 362
 363 void
 364 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 365 {
 366         ASSERT(dsl_pool_sync_context(dp));
 367         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
 368 }
 369
 370 static uint64_t
 371 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 372 {
 373         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 374         if (dsl_dataset_is_snapshot(ds))
 375                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 376         return (smt);
 377 }
 378
 379 static void
 380 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 381 {
 382         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 383             DMU_POOL_DIRECTORY_OBJECT,
 384             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 385             &scn->scn_phys, tx));
 386 }
 387
 388 static boolean_t
 389 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 390 {
 391         uint64_t elapsed_nanosecs;
 392         int mintime;
 393
 394         /* we never skip user/group accounting objects */
 395         if (zb && (int64_t)zb->zb_object < 0)
 396                 return (B_FALSE);
 397
 398         if (scn->scn_pausing)
 399                 return (B_TRUE); /* we're already pausing */
 400
 401         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 402                 return (B_FALSE); /* we're resuming */
 403
 404         /* We only know how to resume from level-0 blocks. */
 405         if (zb && zb->zb_level != 0)
 406                 return (B_FALSE);
 407
 408         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 409             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 410         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 411         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 412             (elapsed_nanosecs / MICROSEC > mintime &&
 413             txg_sync_waiting(scn->scn_dp)) ||
 414             spa_shutting_down(scn->scn_dp->dp_spa)) {
 415                 if (zb) {
 416                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 417                             (longlong_t)zb->zb_objset,
 418                             (longlong_t)zb->zb_object,
 419                             (longlong_t)zb->zb_level,
 420                             (longlong_t)zb->zb_blkid);
 421                         scn->scn_phys.scn_bookmark = *zb;
 422                 }
 423                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 424                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 425                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 426                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 427                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 428                 scn->scn_pausing = B_TRUE;
 429                 return (B_TRUE);
 430         }
 431         return (B_FALSE);
 432 }
 433
 434 typedef struct zil_scan_arg {
 435         dsl_pool_t      *zsa_dp;
 436         zil_header_t    *zsa_zh;
 437 } zil_scan_arg_t;
 438
 439 /* ARGSUSED */
 440 static int
 441 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 442 {
 443         zil_scan_arg_t *zsa = arg;
 444         dsl_pool_t *dp = zsa->zsa_dp;
 445         dsl_scan_t *scn = dp->dp_scan;
 446         zil_header_t *zh = zsa->zsa_zh;
 447         zbookmark_t zb;
 448
 449         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 450                 return (0);
 451
 452         /*
 453          * One block ("stubby") can be allocated a long time ago; we
 454          * want to visit that one because it has been allocated
 455          * (on-disk) even if it hasn't been claimed (even though for
 456          * scrub there's nothing to do to it).
 457          */
 458         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 459                 return (0);
 460
 461         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 462             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 463
 464         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 465         return (0);
 466 }
 467
 468 /* ARGSUSED */
 469 static int
 470 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 471 {
 472         if (lrc->lrc_txtype == TX_WRITE) {
 473                 zil_scan_arg_t *zsa = arg;
 474                 dsl_pool_t *dp = zsa->zsa_dp;
 475                 dsl_scan_t *scn = dp->dp_scan;
 476                 zil_header_t *zh = zsa->zsa_zh;
 477                 lr_write_t *lr = (lr_write_t *)lrc;
 478                 blkptr_t *bp = &lr->lr_blkptr;
 479                 zbookmark_t zb;
 480
 481                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 482                         return (0);
 483
 484                 /*
 485                  * birth can be < claim_txg if this record's txg is
 486                  * already txg sync'ed (but this log block contains
 487                  * other records that are not synced)
 488                  */
 489                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 490                         return (0);
 491
 492                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 493                     lr->lr_foid, ZB_ZIL_LEVEL,
 494                     lr->lr_offset / BP_GET_LSIZE(bp));
 495
 496                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 497         }
 498         return (0);
 499 }
 500
 501 static void
 502 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 503 {
 504         uint64_t claim_txg = zh->zh_claim_txg;
 505         zil_scan_arg_t zsa = { dp, zh };
 506         zilog_t *zilog;
 507
 508         /*
 509          * We only want to visit blocks that have been claimed but not yet
 510          * replayed (or, in read-only mode, blocks that *would* be claimed).
 511          */
 512         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 513                 return;
 514
 515         zilog = zil_alloc(dp->dp_meta_objset, zh);
 516
 517         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 518             claim_txg);
 519
 520         zil_free(zilog);
 521 }
 522
 523 /* ARGSUSED */
 524 static void
 525 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 526     uint64_t objset, uint64_t object, uint64_t blkid)
 527 {
 528         zbookmark_t czb;
 529         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 530
 531         if (zfs_no_scrub_prefetch)
 532                 return;
 533
 534         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 535             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 536                 return;
 537
 538         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 539
 540         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 541             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 542             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 543 }
 544
 545 static boolean_t
 546 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 547     const zbookmark_t *zb)
 548 {
 549         /*
 550          * We never skip over user/group accounting objects (obj<0)
 551          */
 552         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 553             (int64_t)zb->zb_object >= 0) {
 554                 /*
 555                  * If we already visited this bp & everything below (in
 556                  * a prior txg sync), don't bother doing it again.
 557                  */
 558                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 559                         return (B_TRUE);
 560
 561                 /*
 562                  * If we found the block we're trying to resume from, or
 563                  * we went past it to a different object, zero it out to
 564                  * indicate that it's OK to start checking for pausing
 565                  * again.
 566                  */
 567                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 568                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 569                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 570                             (longlong_t)zb->zb_objset,
 571                             (longlong_t)zb->zb_object,
 572                             (longlong_t)zb->zb_level,
 573                             (longlong_t)zb->zb_blkid);
 574                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 575                 }
 576         }
 577         return (B_FALSE);
 578 }
 579
 580 /*
 581  * Return nonzero on i/o error.
 582  * Return new buf to write out in *bufp.
 583  */
 584 inline __attribute__((always_inline)) static int
 585 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 586     dnode_phys_t *dnp, const blkptr_t *bp,
 587     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 588 {
 589         dsl_pool_t *dp = scn->scn_dp;
 590         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 591         int err;
 592
 593         if (BP_GET_LEVEL(bp) > 0) {
 594                 uint32_t flags = ARC_WAIT;
 595                 int i;
 596                 blkptr_t *cbp;
 597                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 598
 599                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 600                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 601                 if (err) {
 602                         scn->scn_phys.scn_errors++;
 603                         return (err);
 604                 }
 605                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 606                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 607                             zb->zb_object, zb->zb_blkid * epb + i);
 608                 }
 609                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 610                         zbookmark_t czb;
 611
 612                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 613                             zb->zb_level - 1,
 614                             zb->zb_blkid * epb + i);
 615                         dsl_scan_visitbp(cbp, &czb, dnp,
 616                             *bufp, ds, scn, ostype, tx);
 617                 }
 618         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 619                 uint32_t flags = ARC_WAIT;
 620
 621                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 622                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 623                 if (err) {
 624                         scn->scn_phys.scn_errors++;
 625                         return (err);
 626                 }
 627         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 628                 uint32_t flags = ARC_WAIT;
 629                 dnode_phys_t *cdnp;
 630                 int i, j;
 631                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 632
 633                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 634                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 635                 if (err) {
 636                         scn->scn_phys.scn_errors++;
 637                         return (err);
 638                 }
 639                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 640                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 641                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 642                                 dsl_scan_prefetch(scn, *bufp, cbp,
 643                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 644                         }
 645                 }
 646                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 647                         dsl_scan_visitdnode(scn, ds, ostype,
 648                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 649                 }
 650
 651         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 652                 uint32_t flags = ARC_WAIT;
 653                 objset_phys_t *osp;
 654
 655                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 656                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 657                 if (err) {
 658                         scn->scn_phys.scn_errors++;
 659                         return (err);
 660                 }
 661
 662                 osp = (*bufp)->b_data;
 663
 664                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 665                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 666
 667                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 668                         /*
 669                          * We also always visit user/group accounting
 670                          * objects, and never skip them, even if we are
 671                          * pausing.  This is necessary so that the space
 672                          * deltas from this txg get integrated.
 673                          */
 674                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 675                             &osp->os_groupused_dnode, *bufp,
 676                             DMU_GROUPUSED_OBJECT, tx);
 677                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 678                             &osp->os_userused_dnode, *bufp,
 679                             DMU_USERUSED_OBJECT, tx);
 680                 }
 681         }
 682
 683         return (0);
 684 }
 685
 686 inline __attribute__((always_inline)) static void
 687 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 688     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 689     uint64_t object, dmu_tx_t *tx)
 690 {
 691         int j;
 692
 693         for (j = 0; j < dnp->dn_nblkptr; j++) {
 694                 zbookmark_t czb;
 695
 696                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 697                     dnp->dn_nlevels - 1, j);
 698                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 699                     &czb, dnp, buf, ds, scn, ostype, tx);
 700         }
 701
 702         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 703                 zbookmark_t czb;
 704                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 705                     0, DMU_SPILL_BLKID);
 706                 dsl_scan_visitbp(&dnp->dn_spill,
 707                     &czb, dnp, buf, ds, scn, ostype, tx);
 708         }
 709 }
 710
 711 /*
 712  * The arguments are in this order because mdb can only print the
 713  * first 5; we want them to be useful.
 714  */
 715 static void
 716 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 717     dnode_phys_t *dnp, arc_buf_t *pbuf,
 718     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 719     dmu_tx_t *tx)
 720 {
 721         dsl_pool_t *dp = scn->scn_dp;
 722         arc_buf_t *buf = NULL;
 723         blkptr_t *bp_toread;
 724
 725         bp_toread = kmem_alloc(sizeof (blkptr_t), KM_PUSHPAGE);
 726         *bp_toread = *bp;
 727
 728         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 729
 730         if (dsl_scan_check_pause(scn, zb))
 731                 goto out;
 732
 733         if (dsl_scan_check_resume(scn, dnp, zb))
 734                 goto out;
 735
 736         if (bp->blk_birth == 0)
 737                 goto out;
 738
 739         scn->scn_visited_this_txg++;
 740
 741         /*
 742          * This debugging is commented out to conserve stack space.  This
 743          * function is called recursively and the debugging addes several
 744          * bytes to the stack for each call.  It can be commented back in
 745          * if required to debug an issue in dsl_scan_visitbp().
 746          *
 747          * dprintf_bp(bp,
 748          *    "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 749          *    ds, ds ? ds->ds_object : 0,
 750          *    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 751          *    pbuf, bp);
 752          */
 753
 754         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 755                 goto out;
 756
 757         if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx,
 758             &buf) != 0)
 759                 goto out;
 760
 761         /*
 762          * If dsl_scan_ddt() has aready visited this block, it will have
 763          * already done any translations or scrubbing, so don't call the
 764          * callback again.
 765          */
 766         if (ddt_class_contains(dp->dp_spa,
 767             scn->scn_phys.scn_ddt_class_max, bp)) {
 768                 ASSERT(buf == NULL);
 769                 goto out;
 770         }
 771
 772         /*
 773          * If this block is from the future (after cur_max_txg), then we
 774          * are doing this on behalf of a deleted snapshot, and we will
 775          * revisit the future block on the next pass of this dataset.
 776          * Don't scan it now unless we need to because something
 777          * under it was modified.
 778          */
 779         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 780                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 781         }
 782         if (buf)
 783                 (void) arc_buf_remove_ref(buf, &buf);
 784 out:
 785         kmem_free(bp_toread, sizeof(blkptr_t));
 786 }
 787
 788 static void
 789 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 790     dmu_tx_t *tx)
 791 {
 792         zbookmark_t zb;
 793
 794         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 795             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 796         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 797             ds, scn, DMU_OST_NONE, tx);
 798
 799         dprintf_ds(ds, "finished scan%s", "");
 800 }
 801
 802 void
 803 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 804 {
 805         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 806         dsl_scan_t *scn = dp->dp_scan;
 807         uint64_t mintxg;
 808
 809         if (scn->scn_phys.scn_state != DSS_SCANNING)
 810                 return;
 811
 812         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 813                 if (dsl_dataset_is_snapshot(ds)) {
 814                         /* Note, scn_cur_{min,max}_txg stays the same. */
 815                         scn->scn_phys.scn_bookmark.zb_objset =
 816                             ds->ds_phys->ds_next_snap_obj;
 817                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 818                             "reset zb_objset to %llu",
 819                             (u_longlong_t)ds->ds_object,
 820                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 821                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 822                 } else {
 823                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 824                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 825                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 826                             "reset bookmark to -1,0,0,0",
 827                             (u_longlong_t)ds->ds_object);
 828                 }
 829         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 830             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 831                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 832                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 833                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 834                 if (dsl_dataset_is_snapshot(ds)) {
 835                         /*
 836                          * We keep the same mintxg; it could be >
 837                          * ds_creation_txg if the previous snapshot was
 838                          * deleted too.
 839                          */
 840                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 841                             scn->scn_phys.scn_queue_obj,
 842                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 843                         zfs_dbgmsg("destroying ds %llu; in queue; "
 844                             "replacing with %llu",
 845                             (u_longlong_t)ds->ds_object,
 846                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 847                 } else {
 848                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 849                             (u_longlong_t)ds->ds_object);
 850                 }
 851         } else {
 852                 zfs_dbgmsg("destroying ds %llu; ignoring",
 853                     (u_longlong_t)ds->ds_object);
 854         }
 855
 856         /*
 857          * dsl_scan_sync() should be called after this, and should sync
 858          * out our changed state, but just to be safe, do it here.
 859          */
 860         dsl_scan_sync_state(scn, tx);
 861 }
 862
 863 void
 864 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 865 {
 866         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 867         dsl_scan_t *scn = dp->dp_scan;
 868         uint64_t mintxg;
 869
 870         if (scn->scn_phys.scn_state != DSS_SCANNING)
 871                 return;
 872
 873         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 874
 875         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 876                 scn->scn_phys.scn_bookmark.zb_objset =
 877                     ds->ds_phys->ds_prev_snap_obj;
 878                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 879                     "reset zb_objset to %llu",
 880                     (u_longlong_t)ds->ds_object,
 881                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 882         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 883             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 884                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 885                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 886                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 887                     scn->scn_phys.scn_queue_obj,
 888                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 889                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 890                     "replacing with %llu",
 891                     (u_longlong_t)ds->ds_object,
 892                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 893         }
 894         dsl_scan_sync_state(scn, tx);
 895 }
 896
 897 void
 898 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 899 {
 900         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 901         dsl_scan_t *scn = dp->dp_scan;
 902         uint64_t mintxg;
 903
 904         if (scn->scn_phys.scn_state != DSS_SCANNING)
 905                 return;
 906
 907         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 908                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 909                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 910                     "reset zb_objset to %llu",
 911                     (u_longlong_t)ds1->ds_object,
 912                     (u_longlong_t)ds2->ds_object);
 913         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 914                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 915                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 916                     "reset zb_objset to %llu",
 917                     (u_longlong_t)ds2->ds_object,
 918                     (u_longlong_t)ds1->ds_object);
 919         }
 920
 921         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 922             ds1->ds_object, &mintxg) == 0) {
 923                 int err;
 924
 925                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 926                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 927                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 928                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 929                 err = zap_add_int_key(dp->dp_meta_objset,
 930                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 931                 VERIFY(err == 0 || err == EEXIST);
 932                 if (err == EEXIST) {
 933                         /* Both were there to begin with */
 934                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 935                             scn->scn_phys.scn_queue_obj,
 936                             ds1->ds_object, mintxg, tx));
 937                 }
 938                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 939                     "replacing with %llu",
 940                     (u_longlong_t)ds1->ds_object,
 941                     (u_longlong_t)ds2->ds_object);
 942         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 943             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 944                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 945                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 946                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 947                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 948                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 949                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 950                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 951                     "replacing with %llu",
 952                     (u_longlong_t)ds2->ds_object,
 953                     (u_longlong_t)ds1->ds_object);
 954         }
 955
 956         dsl_scan_sync_state(scn, tx);
 957 }
 958
 959 struct enqueue_clones_arg {
 960         dmu_tx_t *tx;
 961         uint64_t originobj;
 962 };
 963
 964 /* ARGSUSED */
 965 static int
 966 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 967 {
 968         struct enqueue_clones_arg *eca = arg;
 969         dsl_dataset_t *ds;
 970         int err;
 971         dsl_scan_t *scn = dp->dp_scan;
 972
 973         if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
 974                 return (0);
 975
 976         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
 977         if (err)
 978                 return (err);
 979
 980         while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 981                 dsl_dataset_t *prev;
 982                 err = dsl_dataset_hold_obj(dp,
 983                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 984
 985                 dsl_dataset_rele(ds, FTAG);
 986                 if (err)
 987                         return (err);
 988                 ds = prev;
 989         }
 990         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 991             scn->scn_phys.scn_queue_obj, ds->ds_object,
 992             ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
 993         dsl_dataset_rele(ds, FTAG);
 994         return (0);
 995 }
 996
 997 static void
 998 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
 999 {
1000         dsl_pool_t *dp = scn->scn_dp;
1001         dsl_dataset_t *ds;
1002         objset_t *os;
1003         char *dsname;
1004
1005         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1006
1007         if (dmu_objset_from_ds(ds, &os))
1008                 goto out;
1009
1010         /*
1011          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1012          * snapshots can have ZIL block pointers (which may be the same
1013          * BP as in the head), they must be ignored.  So we traverse the
1014          * ZIL here, rather than in scan_recurse(), because the regular
1015          * snapshot block-sharing rules don't apply to it.
1016          */
1017         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1018                 dsl_scan_zil(dp, &os->os_zil_header);
1019
1020         /*
1021          * Iterate over the bps in this ds.
1022          */
1023         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1024         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1025
1026         dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_PUSHPAGE);
1027         dsl_dataset_name(ds, dsname);
1028         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1029             "pausing=%u",
1030             (longlong_t)dsobj, dsname,
1031             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1032             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1033             (int)scn->scn_pausing);
1034         kmem_free(dsname, ZFS_MAXNAMELEN);
1035
1036         if (scn->scn_pausing)
1037                 goto out;
1038
1039         /*
1040          * We've finished this pass over this dataset.
1041          */
1042
1043         /*
1044          * If we did not completely visit this dataset, do another pass.
1045          */
1046         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1047                 zfs_dbgmsg("incomplete pass; visiting again");
1048                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1049                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1050                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1051                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1052                 goto out;
1053         }
1054
1055         /*
1056          * Add descendent datasets to work queue.
1057          */
1058         if (ds->ds_phys->ds_next_snap_obj != 0) {
1059                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1060                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1061                     ds->ds_phys->ds_creation_txg, tx) == 0);
1062         }
1063         if (ds->ds_phys->ds_num_children > 1) {
1064                 boolean_t usenext = B_FALSE;
1065                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1066                         uint64_t count;
1067                         /*
1068                          * A bug in a previous version of the code could
1069                          * cause upgrade_clones_cb() to not set
1070                          * ds_next_snap_obj when it should, leading to a
1071                          * missing entry.  Therefore we can only use the
1072                          * next_clones_obj when its count is correct.
1073                          */
1074                         int err = zap_count(dp->dp_meta_objset,
1075                             ds->ds_phys->ds_next_clones_obj, &count);
1076                         if (err == 0 &&
1077                             count == ds->ds_phys->ds_num_children - 1)
1078                                 usenext = B_TRUE;
1079                 }
1080
1081                 if (usenext) {
1082                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1083                             ds->ds_phys->ds_next_clones_obj,
1084                             scn->scn_phys.scn_queue_obj,
1085                             ds->ds_phys->ds_creation_txg, tx));
1086                 } else {
1087                         struct enqueue_clones_arg eca;
1088                         eca.tx = tx;
1089                         eca.originobj = ds->ds_object;
1090
1091                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1092                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1093                 }
1094         }
1095
1096 out:
1097         dsl_dataset_rele(ds, FTAG);
1098 }
1099
1100 /* ARGSUSED */
1101 static int
1102 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1103 {
1104         dmu_tx_t *tx = arg;
1105         dsl_dataset_t *ds;
1106         int err;
1107         dsl_scan_t *scn = dp->dp_scan;
1108
1109         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1110         if (err)
1111                 return (err);
1112
1113         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1114                 dsl_dataset_t *prev;
1115                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1116                     FTAG, &prev);
1117                 if (err) {
1118                         dsl_dataset_rele(ds, FTAG);
1119                         return (err);
1120                 }
1121
1122                 /*
1123                  * If this is a clone, we don't need to worry about it for now.
1124                  */
1125                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1126                         dsl_dataset_rele(ds, FTAG);
1127                         dsl_dataset_rele(prev, FTAG);
1128                         return (0);
1129                 }
1130                 dsl_dataset_rele(ds, FTAG);
1131                 ds = prev;
1132         }
1133
1134         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1135             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1136         dsl_dataset_rele(ds, FTAG);
1137         return (0);
1138 }
1139
1140 /*
1141  * Scrub/dedup interaction.
1142  *
1143  * If there are N references to a deduped block, we don't want to scrub it
1144  * N times -- ideally, we should scrub it exactly once.
1145  *
1146  * We leverage the fact that the dde's replication class (enum ddt_class)
1147  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1148  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1149  *
1150  * To prevent excess scrubbing, the scrub begins by walking the DDT
1151  * to find all blocks with refcnt > 1, and scrubs each of these once.
1152  * Since there are two replication classes which contain blocks with
1153  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1154  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1155  *
1156  * There would be nothing more to say if a block's refcnt couldn't change
1157  * during a scrub, but of course it can so we must account for changes
1158  * in a block's replication class.
1159  *
1160  * Here's an example of what can occur:
1161  *
1162  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1163  * when visited during the top-down scrub phase, it will be scrubbed twice.
1164  * This negates our scrub optimization, but is otherwise harmless.
1165  *
1166  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1167  * on each visit during the top-down scrub phase, it will never be scrubbed.
1168  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1169  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1170  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1171  * while a scrub is in progress, it scrubs the block right then.
1172  */
1173 static void
1174 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1175 {
1176         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1177         ddt_entry_t dde;
1178         int error;
1179         uint64_t n = 0;
1180
1181         bzero(&dde, sizeof (ddt_entry_t));
1182
1183         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1184                 ddt_t *ddt;
1185
1186                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1187                         break;
1188                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1189                     (longlong_t)ddb->ddb_class,
1190                     (longlong_t)ddb->ddb_type,
1191                     (longlong_t)ddb->ddb_checksum,
1192                     (longlong_t)ddb->ddb_cursor);
1193
1194                 /* There should be no pending changes to the dedup table */
1195                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1196                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1197
1198                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1199                 n++;
1200
1201                 if (dsl_scan_check_pause(scn, NULL))
1202                         break;
1203         }
1204
1205         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1206             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1207             (int)scn->scn_pausing);
1208
1209         ASSERT(error == 0 || error == ENOENT);
1210         ASSERT(error != ENOENT ||
1211             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1212 }
1213
1214 /* ARGSUSED */
1215 void
1216 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1217     ddt_entry_t *dde, dmu_tx_t *tx)
1218 {
1219         const ddt_key_t *ddk = &dde->dde_key;
1220         ddt_phys_t *ddp = dde->dde_phys;
1221         blkptr_t bp;
1222         zbookmark_t zb = { 0 };
1223         int p;
1224
1225         if (scn->scn_phys.scn_state != DSS_SCANNING)
1226                 return;
1227
1228         for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1229                 if (ddp->ddp_phys_birth == 0 ||
1230                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1231                         continue;
1232                 ddt_bp_create(checksum, ddk, ddp, &bp);
1233
1234                 scn->scn_visited_this_txg++;
1235                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1236         }
1237 }
1238
1239 static void
1240 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1241 {
1242         dsl_pool_t *dp = scn->scn_dp;
1243         zap_cursor_t *zc;
1244         zap_attribute_t *za;
1245
1246         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1247             scn->scn_phys.scn_ddt_class_max) {
1248                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1249                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1250                 dsl_scan_ddt(scn, tx);
1251                 if (scn->scn_pausing)
1252                         return;
1253         }
1254
1255         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1256                 /* First do the MOS & ORIGIN */
1257
1258                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1259                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1260                 dsl_scan_visit_rootbp(scn, NULL,
1261                     &dp->dp_meta_rootbp, tx);
1262                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1263                 if (scn->scn_pausing)
1264                         return;
1265
1266                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1267                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1268                             enqueue_cb, tx, DS_FIND_CHILDREN));
1269                 } else {
1270                         dsl_scan_visitds(scn,
1271                             dp->dp_origin_snap->ds_object, tx);
1272                 }
1273                 ASSERT(!scn->scn_pausing);
1274         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1275             ZB_DESTROYED_OBJSET) {
1276                 /*
1277                  * If we were paused, continue from here.  Note if the
1278                  * ds we were paused on was deleted, the zb_objset may
1279                  * be -1, so we will skip this and find a new objset
1280                  * below.
1281                  */
1282                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1283                 if (scn->scn_pausing)
1284                         return;
1285         }
1286
1287         /*
1288          * In case we were paused right at the end of the ds, zero the
1289          * bookmark so we don't think that we're still trying to resume.
1290          */
1291         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1292         zc = kmem_alloc(sizeof(zap_cursor_t), KM_PUSHPAGE);
1293         za = kmem_alloc(sizeof(zap_attribute_t), KM_PUSHPAGE);
1294
1295         /* keep pulling things out of the zap-object-as-queue */
1296         while (zap_cursor_init(zc, dp->dp_meta_objset,
1297             scn->scn_phys.scn_queue_obj),
1298             zap_cursor_retrieve(zc, za) == 0) {
1299                 dsl_dataset_t *ds;
1300                 uint64_t dsobj;
1301
1302                 dsobj = strtonum(za->za_name, NULL);
1303                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1304                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1305
1306                 /* Set up min/max txg */
1307                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1308                 if (za->za_first_integer != 0) {
1309                         scn->scn_phys.scn_cur_min_txg =
1310                             MAX(scn->scn_phys.scn_min_txg,
1311                             za->za_first_integer);
1312                 } else {
1313                         scn->scn_phys.scn_cur_min_txg =
1314                             MAX(scn->scn_phys.scn_min_txg,
1315                             ds->ds_phys->ds_prev_snap_txg);
1316                 }
1317                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1318                 dsl_dataset_rele(ds, FTAG);
1319
1320                 dsl_scan_visitds(scn, dsobj, tx);
1321                 zap_cursor_fini(zc);
1322                 if (scn->scn_pausing)
1323                         goto out;
1324         }
1325         zap_cursor_fini(zc);
1326 out:
1327         kmem_free(za, sizeof(zap_attribute_t));
1328         kmem_free(zc, sizeof(zap_cursor_t));
1329 }
1330
1331 static boolean_t
1332 dsl_scan_free_should_pause(dsl_scan_t *scn)
1333 {
1334         uint64_t elapsed_nanosecs;
1335
1336         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1337         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1338             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1339             txg_sync_waiting(scn->scn_dp)) ||
1340             spa_shutting_down(scn->scn_dp->dp_spa));
1341 }
1342
1343 static int
1344 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1345 {
1346         dsl_scan_t *scn = arg;
1347
1348         if (!scn->scn_is_bptree ||
1349             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1350                 if (dsl_scan_free_should_pause(scn))
1351                         return (SET_ERROR(ERESTART));
1352         }
1353
1354         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1355             dmu_tx_get_txg(tx), bp, 0));
1356         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1357             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1358             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1359         scn->scn_visited_this_txg++;
1360         return (0);
1361 }
1362
1363 boolean_t
1364 dsl_scan_active(dsl_scan_t *scn)
1365 {
1366         spa_t *spa = scn->scn_dp->dp_spa;
1367         uint64_t used = 0, comp, uncomp;
1368
1369         if (spa->spa_load_state != SPA_LOAD_NONE)
1370                 return (B_FALSE);
1371         if (spa_shutting_down(spa))
1372                 return (B_FALSE);
1373
1374         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1375             scn->scn_async_destroying)
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(&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                         ASSERT(scn->scn_async_destroying);
1434                         scn->scn_is_bptree = B_TRUE;
1435                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1436                             NULL, ZIO_FLAG_MUSTSUCCEED);
1437                         err = bptree_iterate(dp->dp_meta_objset,
1438                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1439                             scn, tx);
1440                         VERIFY0(zio_wait(scn->scn_zio_root));
1441
1442                         if (err == 0) {
1443                                 zfeature_info_t *feat = &spa_feature_table
1444                                     [SPA_FEATURE_ASYNC_DESTROY];
1445                                 /* finished; deactivate async destroy feature */
1446                                 spa_feature_decr(spa, feat, tx);
1447                                 ASSERT(!spa_feature_is_active(spa, feat));
1448                                 VERIFY0(zap_remove(dp->dp_meta_objset,
1449                                     DMU_POOL_DIRECTORY_OBJECT,
1450                                     DMU_POOL_BPTREE_OBJ, tx));
1451                                 VERIFY0(bptree_free(dp->dp_meta_objset,
1452                                     dp->dp_bptree_obj, tx));
1453                                 dp->dp_bptree_obj = 0;
1454                                 scn->scn_async_destroying = B_FALSE;
1455                         }
1456                 }
1457                 if (scn->scn_visited_this_txg) {
1458                         zfs_dbgmsg("freed %llu blocks in %llums from "
1459                             "free_bpobj/bptree txg %llu",
1460                             (longlong_t)scn->scn_visited_this_txg,
1461                             (longlong_t)
1462                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1463                             (longlong_t)tx->tx_txg);
1464                         scn->scn_visited_this_txg = 0;
1465                         /*
1466                          * Re-sync the ddt so that we can further modify
1467                          * it when doing bprewrite.
1468                          */
1469                         ddt_sync(spa, tx->tx_txg);
1470                 }
1471                 if (err == ERESTART)
1472                         return;
1473         }
1474
1475         if (scn->scn_phys.scn_state != DSS_SCANNING)
1476                 return;
1477
1478         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1479             scn->scn_phys.scn_ddt_class_max) {
1480                 zfs_dbgmsg("doing scan sync txg %llu; "
1481                     "ddt bm=%llu/%llu/%llu/%llx",
1482                     (longlong_t)tx->tx_txg,
1483                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1484                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1485                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1486                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1487                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1488                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1489                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1490                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1491         } else {
1492                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1493                     (longlong_t)tx->tx_txg,
1494                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1495                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1496                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1497                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1498         }
1499
1500         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1501             NULL, ZIO_FLAG_CANFAIL);
1502         dsl_pool_config_enter(dp, FTAG);
1503         dsl_scan_visit(scn, tx);
1504         dsl_pool_config_exit(dp, FTAG);
1505         (void) zio_wait(scn->scn_zio_root);
1506         scn->scn_zio_root = NULL;
1507
1508         zfs_dbgmsg("visited %llu blocks in %llums",
1509             (longlong_t)scn->scn_visited_this_txg,
1510             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1511
1512         if (!scn->scn_pausing) {
1513                 /* finished with scan. */
1514                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1515                 dsl_scan_done(scn, B_TRUE, tx);
1516         }
1517
1518         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1519                 mutex_enter(&spa->spa_scrub_lock);
1520                 while (spa->spa_scrub_inflight > 0) {
1521                         cv_wait(&spa->spa_scrub_io_cv,
1522                             &spa->spa_scrub_lock);
1523                 }
1524                 mutex_exit(&spa->spa_scrub_lock);
1525         }
1526
1527         dsl_scan_sync_state(scn, tx);
1528 }
1529
1530 /*
1531  * This will start a new scan, or restart an existing one.
1532  */
1533 void
1534 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1535 {
1536         if (txg == 0) {
1537                 dmu_tx_t *tx;
1538                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1539                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1540
1541                 txg = dmu_tx_get_txg(tx);
1542                 dp->dp_scan->scn_restart_txg = txg;
1543                 dmu_tx_commit(tx);
1544         } else {
1545                 dp->dp_scan->scn_restart_txg = txg;
1546         }
1547         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1548 }
1549
1550 boolean_t
1551 dsl_scan_resilvering(dsl_pool_t *dp)
1552 {
1553         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1554             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1555 }
1556
1557 /*
1558  * scrub consumers
1559  */
1560
1561 static void
1562 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1563 {
1564         int i;
1565
1566         /*
1567          * If we resume after a reboot, zab will be NULL; don't record
1568          * incomplete stats in that case.
1569          */
1570         if (zab == NULL)
1571                 return;
1572
1573         for (i = 0; i < 4; i++) {
1574                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1575                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1576                 int equal;
1577                 zfs_blkstat_t *zb;
1578
1579                 if (t & DMU_OT_NEWTYPE)
1580                         t = DMU_OT_OTHER;
1581
1582                 zb = &zab->zab_type[l][t];
1583                 zb->zb_count++;
1584                 zb->zb_asize += BP_GET_ASIZE(bp);
1585                 zb->zb_lsize += BP_GET_LSIZE(bp);
1586                 zb->zb_psize += BP_GET_PSIZE(bp);
1587                 zb->zb_gangs += BP_COUNT_GANG(bp);
1588
1589                 switch (BP_GET_NDVAS(bp)) {
1590                 case 2:
1591                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1592                             DVA_GET_VDEV(&bp->blk_dva[1]))
1593                                 zb->zb_ditto_2_of_2_samevdev++;
1594                         break;
1595                 case 3:
1596                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1597                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1598                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1599                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1600                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1601                             DVA_GET_VDEV(&bp->blk_dva[2]));
1602                         if (equal == 1)
1603                                 zb->zb_ditto_2_of_3_samevdev++;
1604                         else if (equal == 3)
1605                                 zb->zb_ditto_3_of_3_samevdev++;
1606                         break;
1607                 }
1608         }
1609 }
1610
1611 static void
1612 dsl_scan_scrub_done(zio_t *zio)
1613 {
1614         spa_t *spa = zio->io_spa;
1615
1616         zio_data_buf_free(zio->io_data, zio->io_size);
1617
1618         mutex_enter(&spa->spa_scrub_lock);
1619         spa->spa_scrub_inflight--;
1620         cv_broadcast(&spa->spa_scrub_io_cv);
1621
1622         if (zio->io_error && (zio->io_error != ECKSUM ||
1623             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1624                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1625         }
1626         mutex_exit(&spa->spa_scrub_lock);
1627 }
1628
1629 static int
1630 dsl_scan_scrub_cb(dsl_pool_t *dp,
1631     const blkptr_t *bp, const zbookmark_t *zb)
1632 {
1633         dsl_scan_t *scn = dp->dp_scan;
1634         size_t size = BP_GET_PSIZE(bp);
1635         spa_t *spa = dp->dp_spa;
1636         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1637         boolean_t needs_io = B_FALSE;
1638         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1639         int zio_priority = 0;
1640         int scan_delay = 0;
1641         int d;
1642
1643         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1644             phys_birth >= scn->scn_phys.scn_max_txg)
1645                 return (0);
1646
1647         count_block(dp->dp_blkstats, bp);
1648
1649         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1650         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1651                 zio_flags |= ZIO_FLAG_SCRUB;
1652                 zio_priority = ZIO_PRIORITY_SCRUB;
1653                 needs_io = B_TRUE;
1654                 scan_delay = zfs_scrub_delay;
1655         } else {
1656                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1657                 zio_flags |= ZIO_FLAG_RESILVER;
1658                 zio_priority = ZIO_PRIORITY_RESILVER;
1659                 needs_io = B_FALSE;
1660                 scan_delay = zfs_resilver_delay;
1661         }
1662
1663         /* If it's an intent log block, failure is expected. */
1664         if (zb->zb_level == ZB_ZIL_LEVEL)
1665                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1666
1667         for (d = 0; d < BP_GET_NDVAS(bp); d++) {
1668                 vdev_t *vd = vdev_lookup_top(spa,
1669                     DVA_GET_VDEV(&bp->blk_dva[d]));
1670
1671                 /*
1672                  * Keep track of how much data we've examined so that
1673                  * zpool(1M) status can make useful progress reports.
1674                  */
1675                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1676                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1677
1678                 /* if it's a resilver, this may not be in the target range */
1679                 if (!needs_io) {
1680                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1681                                 /*
1682                                  * Gang members may be spread across multiple
1683                                  * vdevs, so the best estimate we have is the
1684                                  * scrub range, which has already been checked.
1685                                  * XXX -- it would be better to change our
1686                                  * allocation policy to ensure that all
1687                                  * gang members reside on the same vdev.
1688                                  */
1689                                 needs_io = B_TRUE;
1690                         } else {
1691                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1692                                     phys_birth, 1);
1693                         }
1694                 }
1695         }
1696
1697         if (needs_io && !zfs_no_scrub_io) {
1698                 vdev_t *rvd = spa->spa_root_vdev;
1699                 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1700                 void *data = zio_data_buf_alloc(size);
1701
1702                 mutex_enter(&spa->spa_scrub_lock);
1703                 while (spa->spa_scrub_inflight >= maxinflight)
1704                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1705                 spa->spa_scrub_inflight++;
1706                 mutex_exit(&spa->spa_scrub_lock);
1707
1708                 /*
1709                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1710                  * then throttle our workload to limit the impact of a scan.
1711                  */
1712                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1713                         delay(scan_delay);
1714
1715                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1716                     dsl_scan_scrub_done, NULL, zio_priority,
1717                     zio_flags, zb));
1718         }
1719
1720         /* do not relocate this block */
1721         return (0);
1722 }
1723
1724 int
1725 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1726 {
1727         spa_t *spa = dp->dp_spa;
1728
1729         /*
1730          * Purge all vdev caches and probe all devices.  We do this here
1731          * rather than in sync context because this requires a writer lock
1732          * on the spa_config lock, which we can't do from sync context.  The
1733          * spa_scrub_reopen flag indicates that vdev_open() should not
1734          * attempt to start another scrub.
1735          */
1736         spa_vdev_state_enter(spa, SCL_NONE);
1737         spa->spa_scrub_reopen = B_TRUE;
1738         vdev_reopen(spa->spa_root_vdev);
1739         spa->spa_scrub_reopen = B_FALSE;
1740         (void) spa_vdev_state_exit(spa, NULL, 0);
1741
1742         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1743             dsl_scan_setup_sync, &func, 0));
1744 }
1745
1746 #if defined(_KERNEL) && defined(HAVE_SPL)
1747 module_param(zfs_top_maxinflight, int, 0644);
1748 MODULE_PARM_DESC(zfs_top_maxinflight, "Max I/Os per top-level");
1749
1750 module_param(zfs_resilver_delay, int, 0644);
1751 MODULE_PARM_DESC(zfs_resilver_delay, "Number of ticks to delay resilver");
1752
1753 module_param(zfs_scrub_delay, int, 0644);
1754 MODULE_PARM_DESC(zfs_scrub_delay, "Number of ticks to delay scrub");
1755
1756 module_param(zfs_scan_idle, int, 0644);
1757 MODULE_PARM_DESC(zfs_scan_idle, "Idle window in clock ticks");
1758
1759 module_param(zfs_scan_min_time_ms, int, 0644);
1760 MODULE_PARM_DESC(zfs_scan_min_time_ms, "Min millisecs to scrub per txg");
1761
1762 module_param(zfs_free_min_time_ms, int, 0644);
1763 MODULE_PARM_DESC(zfs_free_min_time_ms, "Min millisecs to free per txg");
1764
1765 module_param(zfs_resilver_min_time_ms, int, 0644);
1766 MODULE_PARM_DESC(zfs_resilver_min_time_ms, "Min millisecs to resilver per txg");
1767
1768 module_param(zfs_no_scrub_io, int, 0644);
1769 MODULE_PARM_DESC(zfs_no_scrub_io, "Set to disable scrub I/O");
1770
1771 module_param(zfs_no_scrub_prefetch, int, 0644);
1772 MODULE_PARM_DESC(zfs_no_scrub_prefetch, "Set to disable scrub prefetching");
1773 #endif