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         scn->scn_done_txg = 0;
 194         spa_scan_stat_init(spa);
 195
 196         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 197                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 198
 199                 /* rewrite all disk labels */
 200                 vdev_config_dirty(spa->spa_root_vdev);
 201
 202                 if (vdev_resilver_needed(spa->spa_root_vdev,
 203                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 204                         spa_event_notify(spa, NULL, FM_EREPORT_ZFS_RESILVER_START);
 205                 } else {
 206                         spa_event_notify(spa, NULL, FM_EREPORT_ZFS_SCRUB_START);
 207                 }
 208
 209                 spa->spa_scrub_started = B_TRUE;
 210                 /*
 211                  * If this is an incremental scrub, limit the DDT scrub phase
 212                  * to just the auto-ditto class (for correctness); the rest
 213                  * of the scrub should go faster using top-down pruning.
 214                  */
 215                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 216                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 217
 218         }
 219
 220         /* back to the generic stuff */
 221
 222         if (dp->dp_blkstats == NULL) {
 223                 dp->dp_blkstats = kmem_alloc(sizeof (zfs_all_blkstats_t),
 224                     KM_PUSHPAGE | KM_NODEBUG);
 225         }
 226         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 227
 228         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 229                 ot = DMU_OT_ZAP_OTHER;
 230
 231         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 232             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 233
 234         dsl_scan_sync_state(scn, tx);
 235
 236         spa_history_log_internal(spa, "scan setup", tx,
 237             "func=%u mintxg=%llu maxtxg=%llu",
 238             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 239 }
 240
 241 /* ARGSUSED */
 242 static void
 243 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 244 {
 245         static const char *old_names[] = {
 246                 "scrub_bookmark",
 247                 "scrub_ddt_bookmark",
 248                 "scrub_ddt_class_max",
 249                 "scrub_queue",
 250                 "scrub_min_txg",
 251                 "scrub_max_txg",
 252                 "scrub_func",
 253                 "scrub_errors",
 254                 NULL
 255         };
 256
 257         dsl_pool_t *dp = scn->scn_dp;
 258         spa_t *spa = dp->dp_spa;
 259         int i;
 260
 261         /* Remove any remnants of an old-style scrub. */
 262         for (i = 0; old_names[i]; i++) {
 263                 (void) zap_remove(dp->dp_meta_objset,
 264                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 265         }
 266
 267         if (scn->scn_phys.scn_queue_obj != 0) {
 268                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 269                     scn->scn_phys.scn_queue_obj, tx));
 270                 scn->scn_phys.scn_queue_obj = 0;
 271         }
 272
 273         /*
 274          * If we were "restarted" from a stopped state, don't bother
 275          * with anything else.
 276          */
 277         if (scn->scn_phys.scn_state != DSS_SCANNING)
 278                 return;
 279
 280         if (complete)
 281                 scn->scn_phys.scn_state = DSS_FINISHED;
 282         else
 283                 scn->scn_phys.scn_state = DSS_CANCELED;
 284
 285         spa_history_log_internal(spa, "scan done", tx,
 286             "complete=%u", complete);
 287
 288         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 289                 mutex_enter(&spa->spa_scrub_lock);
 290                 while (spa->spa_scrub_inflight > 0) {
 291                         cv_wait(&spa->spa_scrub_io_cv,
 292                             &spa->spa_scrub_lock);
 293                 }
 294                 mutex_exit(&spa->spa_scrub_lock);
 295                 spa->spa_scrub_started = B_FALSE;
 296                 spa->spa_scrub_active = B_FALSE;
 297
 298                 /*
 299                  * If the scrub/resilver completed, update all DTLs to
 300                  * reflect this.  Whether it succeeded or not, vacate
 301                  * all temporary scrub DTLs.
 302                  */
 303                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 304                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 305                 if (complete) {
 306                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 307                             FM_EREPORT_ZFS_RESILVER_FINISH :
 308                             FM_EREPORT_ZFS_SCRUB_FINISH);
 309                 }
 310                 spa_errlog_rotate(spa);
 311
 312                 /*
 313                  * We may have finished replacing a device.
 314                  * Let the async thread assess this and handle the detach.
 315                  */
 316                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 317         }
 318
 319         scn->scn_phys.scn_end_time = gethrestime_sec();
 320 }
 321
 322 /* ARGSUSED */
 323 static int
 324 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 325 {
 326         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 327
 328         if (scn->scn_phys.scn_state != DSS_SCANNING)
 329                 return (SET_ERROR(ENOENT));
 330         return (0);
 331 }
 332
 333 /* ARGSUSED */
 334 static void
 335 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 336 {
 337         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 338
 339         dsl_scan_done(scn, B_FALSE, tx);
 340         dsl_scan_sync_state(scn, tx);
 341 }
 342
 343 int
 344 dsl_scan_cancel(dsl_pool_t *dp)
 345 {
 346         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 347             dsl_scan_cancel_sync, NULL, 3));
 348 }
 349
 350 static void dsl_scan_visitbp(blkptr_t *bp,
 351     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 352     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 353     dmu_tx_t *tx);
 354 inline __attribute__((always_inline)) static void dsl_scan_visitdnode(
 355     dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 356     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 357
 358 void
 359 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 360 {
 361         zio_free(dp->dp_spa, txg, bp);
 362 }
 363
 364 void
 365 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 366 {
 367         ASSERT(dsl_pool_sync_context(dp));
 368         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
 369 }
 370
 371 static uint64_t
 372 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 373 {
 374         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 375         if (dsl_dataset_is_snapshot(ds))
 376                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 377         return (smt);
 378 }
 379
 380 static void
 381 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 382 {
 383         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 384             DMU_POOL_DIRECTORY_OBJECT,
 385             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 386             &scn->scn_phys, tx));
 387 }
 388
 389 static boolean_t
 390 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 391 {
 392         uint64_t elapsed_nanosecs;
 393         int mintime;
 394
 395         /* we never skip user/group accounting objects */
 396         if (zb && (int64_t)zb->zb_object < 0)
 397                 return (B_FALSE);
 398
 399         if (scn->scn_pausing)
 400                 return (B_TRUE); /* we're already pausing */
 401
 402         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 403                 return (B_FALSE); /* we're resuming */
 404
 405         /* We only know how to resume from level-0 blocks. */
 406         if (zb && zb->zb_level != 0)
 407                 return (B_FALSE);
 408
 409         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 410             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 411         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 412         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 413             (NSEC2MSEC(elapsed_nanosecs) > mintime &&
 414             txg_sync_waiting(scn->scn_dp)) ||
 415             spa_shutting_down(scn->scn_dp->dp_spa)) {
 416                 if (zb) {
 417                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 418                             (longlong_t)zb->zb_objset,
 419                             (longlong_t)zb->zb_object,
 420                             (longlong_t)zb->zb_level,
 421                             (longlong_t)zb->zb_blkid);
 422                         scn->scn_phys.scn_bookmark = *zb;
 423                 }
 424                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 425                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 426                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 427                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 428                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 429                 scn->scn_pausing = B_TRUE;
 430                 return (B_TRUE);
 431         }
 432         return (B_FALSE);
 433 }
 434
 435 typedef struct zil_scan_arg {
 436         dsl_pool_t      *zsa_dp;
 437         zil_header_t    *zsa_zh;
 438 } zil_scan_arg_t;
 439
 440 /* ARGSUSED */
 441 static int
 442 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 443 {
 444         zil_scan_arg_t *zsa = arg;
 445         dsl_pool_t *dp = zsa->zsa_dp;
 446         dsl_scan_t *scn = dp->dp_scan;
 447         zil_header_t *zh = zsa->zsa_zh;
 448         zbookmark_t zb;
 449
 450         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 451                 return (0);
 452
 453         /*
 454          * One block ("stubby") can be allocated a long time ago; we
 455          * want to visit that one because it has been allocated
 456          * (on-disk) even if it hasn't been claimed (even though for
 457          * scrub there's nothing to do to it).
 458          */
 459         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 460                 return (0);
 461
 462         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 463             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 464
 465         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 466         return (0);
 467 }
 468
 469 /* ARGSUSED */
 470 static int
 471 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 472 {
 473         if (lrc->lrc_txtype == TX_WRITE) {
 474                 zil_scan_arg_t *zsa = arg;
 475                 dsl_pool_t *dp = zsa->zsa_dp;
 476                 dsl_scan_t *scn = dp->dp_scan;
 477                 zil_header_t *zh = zsa->zsa_zh;
 478                 lr_write_t *lr = (lr_write_t *)lrc;
 479                 blkptr_t *bp = &lr->lr_blkptr;
 480                 zbookmark_t zb;
 481
 482                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 483                         return (0);
 484
 485                 /*
 486                  * birth can be < claim_txg if this record's txg is
 487                  * already txg sync'ed (but this log block contains
 488                  * other records that are not synced)
 489                  */
 490                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 491                         return (0);
 492
 493                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 494                     lr->lr_foid, ZB_ZIL_LEVEL,
 495                     lr->lr_offset / BP_GET_LSIZE(bp));
 496
 497                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 498         }
 499         return (0);
 500 }
 501
 502 static void
 503 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 504 {
 505         uint64_t claim_txg = zh->zh_claim_txg;
 506         zil_scan_arg_t zsa = { dp, zh };
 507         zilog_t *zilog;
 508
 509         /*
 510          * We only want to visit blocks that have been claimed but not yet
 511          * replayed (or, in read-only mode, blocks that *would* be claimed).
 512          */
 513         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 514                 return;
 515
 516         zilog = zil_alloc(dp->dp_meta_objset, zh);
 517
 518         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 519             claim_txg);
 520
 521         zil_free(zilog);
 522 }
 523
 524 /* ARGSUSED */
 525 static void
 526 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 527     uint64_t objset, uint64_t object, uint64_t blkid)
 528 {
 529         zbookmark_t czb;
 530         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 531
 532         if (zfs_no_scrub_prefetch)
 533                 return;
 534
 535         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 536             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 537                 return;
 538
 539         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 540
 541         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 542             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 543             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 544 }
 545
 546 static boolean_t
 547 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 548     const zbookmark_t *zb)
 549 {
 550         /*
 551          * We never skip over user/group accounting objects (obj<0)
 552          */
 553         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 554             (int64_t)zb->zb_object >= 0) {
 555                 /*
 556                  * If we already visited this bp & everything below (in
 557                  * a prior txg sync), don't bother doing it again.
 558                  */
 559                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 560                         return (B_TRUE);
 561
 562                 /*
 563                  * If we found the block we're trying to resume from, or
 564                  * we went past it to a different object, zero it out to
 565                  * indicate that it's OK to start checking for pausing
 566                  * again.
 567                  */
 568                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 569                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 570                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 571                             (longlong_t)zb->zb_objset,
 572                             (longlong_t)zb->zb_object,
 573                             (longlong_t)zb->zb_level,
 574                             (longlong_t)zb->zb_blkid);
 575                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 576                 }
 577         }
 578         return (B_FALSE);
 579 }
 580
 581 /*
 582  * Return nonzero on i/o error.
 583  * Return new buf to write out in *bufp.
 584  */
 585 inline __attribute__((always_inline)) static int
 586 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 587     dnode_phys_t *dnp, const blkptr_t *bp,
 588     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 589 {
 590         dsl_pool_t *dp = scn->scn_dp;
 591         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 592         int err;
 593
 594         if (BP_GET_LEVEL(bp) > 0) {
 595                 uint32_t flags = ARC_WAIT;
 596                 int i;
 597                 blkptr_t *cbp;
 598                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 599
 600                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 601                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 602                 if (err) {
 603                         scn->scn_phys.scn_errors++;
 604                         return (err);
 605                 }
 606                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 607                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 608                             zb->zb_object, zb->zb_blkid * epb + i);
 609                 }
 610                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 611                         zbookmark_t czb;
 612
 613                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 614                             zb->zb_level - 1,
 615                             zb->zb_blkid * epb + i);
 616                         dsl_scan_visitbp(cbp, &czb, dnp,
 617                             *bufp, ds, scn, ostype, tx);
 618                 }
 619         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 620                 uint32_t flags = ARC_WAIT;
 621
 622                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 623                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 624                 if (err) {
 625                         scn->scn_phys.scn_errors++;
 626                         return (err);
 627                 }
 628         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 629                 uint32_t flags = ARC_WAIT;
 630                 dnode_phys_t *cdnp;
 631                 int i, j;
 632                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 633
 634                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 635                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 636                 if (err) {
 637                         scn->scn_phys.scn_errors++;
 638                         return (err);
 639                 }
 640                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 641                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 642                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 643                                 dsl_scan_prefetch(scn, *bufp, cbp,
 644                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 645                         }
 646                 }
 647                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 648                         dsl_scan_visitdnode(scn, ds, ostype,
 649                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 650                 }
 651
 652         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 653                 uint32_t flags = ARC_WAIT;
 654                 objset_phys_t *osp;
 655
 656                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 657                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 658                 if (err) {
 659                         scn->scn_phys.scn_errors++;
 660                         return (err);
 661                 }
 662
 663                 osp = (*bufp)->b_data;
 664
 665                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 666                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 667
 668                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 669                         /*
 670                          * We also always visit user/group accounting
 671                          * objects, and never skip them, even if we are
 672                          * pausing.  This is necessary so that the space
 673                          * deltas from this txg get integrated.
 674                          */
 675                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 676                             &osp->os_groupused_dnode, *bufp,
 677                             DMU_GROUPUSED_OBJECT, tx);
 678                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 679                             &osp->os_userused_dnode, *bufp,
 680                             DMU_USERUSED_OBJECT, tx);
 681                 }
 682         }
 683
 684         return (0);
 685 }
 686
 687 inline __attribute__((always_inline)) static void
 688 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 689     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 690     uint64_t object, dmu_tx_t *tx)
 691 {
 692         int j;
 693
 694         for (j = 0; j < dnp->dn_nblkptr; j++) {
 695                 zbookmark_t czb;
 696
 697                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 698                     dnp->dn_nlevels - 1, j);
 699                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 700                     &czb, dnp, buf, ds, scn, ostype, tx);
 701         }
 702
 703         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 704                 zbookmark_t czb;
 705                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 706                     0, DMU_SPILL_BLKID);
 707                 dsl_scan_visitbp(&dnp->dn_spill,
 708                     &czb, dnp, buf, ds, scn, ostype, tx);
 709         }
 710 }
 711
 712 /*
 713  * The arguments are in this order because mdb can only print the
 714  * first 5; we want them to be useful.
 715  */
 716 static void
 717 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 718     dnode_phys_t *dnp, arc_buf_t *pbuf,
 719     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 720     dmu_tx_t *tx)
 721 {
 722         dsl_pool_t *dp = scn->scn_dp;
 723         arc_buf_t *buf = NULL;
 724         blkptr_t *bp_toread;
 725
 726         bp_toread = kmem_alloc(sizeof (blkptr_t), KM_PUSHPAGE);
 727         *bp_toread = *bp;
 728
 729         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 730
 731         if (dsl_scan_check_pause(scn, zb))
 732                 goto out;
 733
 734         if (dsl_scan_check_resume(scn, dnp, zb))
 735                 goto out;
 736
 737         if (bp->blk_birth == 0)
 738                 goto out;
 739
 740         scn->scn_visited_this_txg++;
 741
 742         /*
 743          * This debugging is commented out to conserve stack space.  This
 744          * function is called recursively and the debugging addes several
 745          * bytes to the stack for each call.  It can be commented back in
 746          * if required to debug an issue in dsl_scan_visitbp().
 747          *
 748          * dprintf_bp(bp,
 749          *    "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 750          *    ds, ds ? ds->ds_object : 0,
 751          *    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 752          *    pbuf, bp);
 753          */
 754
 755         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 756                 goto out;
 757
 758         if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx,
 759             &buf) != 0)
 760                 goto out;
 761
 762         /*
 763          * If dsl_scan_ddt() has aready visited this block, it will have
 764          * already done any translations or scrubbing, so don't call the
 765          * callback again.
 766          */
 767         if (ddt_class_contains(dp->dp_spa,
 768             scn->scn_phys.scn_ddt_class_max, bp)) {
 769                 ASSERT(buf == NULL);
 770                 goto out;
 771         }
 772
 773         /*
 774          * If this block is from the future (after cur_max_txg), then we
 775          * are doing this on behalf of a deleted snapshot, and we will
 776          * revisit the future block on the next pass of this dataset.
 777          * Don't scan it now unless we need to because something
 778          * under it was modified.
 779          */
 780         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 781                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 782         }
 783         if (buf)
 784                 (void) arc_buf_remove_ref(buf, &buf);
 785 out:
 786         kmem_free(bp_toread, sizeof(blkptr_t));
 787 }
 788
 789 static void
 790 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 791     dmu_tx_t *tx)
 792 {
 793         zbookmark_t zb;
 794
 795         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 796             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 797         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 798             ds, scn, DMU_OST_NONE, tx);
 799
 800         dprintf_ds(ds, "finished scan%s", "");
 801 }
 802
 803 void
 804 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 805 {
 806         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 807         dsl_scan_t *scn = dp->dp_scan;
 808         uint64_t mintxg;
 809
 810         if (scn->scn_phys.scn_state != DSS_SCANNING)
 811                 return;
 812
 813         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 814                 if (dsl_dataset_is_snapshot(ds)) {
 815                         /* Note, scn_cur_{min,max}_txg stays the same. */
 816                         scn->scn_phys.scn_bookmark.zb_objset =
 817                             ds->ds_phys->ds_next_snap_obj;
 818                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 819                             "reset zb_objset to %llu",
 820                             (u_longlong_t)ds->ds_object,
 821                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 822                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 823                 } else {
 824                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 825                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 826                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 827                             "reset bookmark to -1,0,0,0",
 828                             (u_longlong_t)ds->ds_object);
 829                 }
 830         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 831             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 832                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 833                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 834                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 835                 if (dsl_dataset_is_snapshot(ds)) {
 836                         /*
 837                          * We keep the same mintxg; it could be >
 838                          * ds_creation_txg if the previous snapshot was
 839                          * deleted too.
 840                          */
 841                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 842                             scn->scn_phys.scn_queue_obj,
 843                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 844                         zfs_dbgmsg("destroying ds %llu; in queue; "
 845                             "replacing with %llu",
 846                             (u_longlong_t)ds->ds_object,
 847                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 848                 } else {
 849                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 850                             (u_longlong_t)ds->ds_object);
 851                 }
 852         } else {
 853                 zfs_dbgmsg("destroying ds %llu; ignoring",
 854                     (u_longlong_t)ds->ds_object);
 855         }
 856
 857         /*
 858          * dsl_scan_sync() should be called after this, and should sync
 859          * out our changed state, but just to be safe, do it here.
 860          */
 861         dsl_scan_sync_state(scn, tx);
 862 }
 863
 864 void
 865 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 866 {
 867         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 868         dsl_scan_t *scn = dp->dp_scan;
 869         uint64_t mintxg;
 870
 871         if (scn->scn_phys.scn_state != DSS_SCANNING)
 872                 return;
 873
 874         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 875
 876         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 877                 scn->scn_phys.scn_bookmark.zb_objset =
 878                     ds->ds_phys->ds_prev_snap_obj;
 879                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 880                     "reset zb_objset to %llu",
 881                     (u_longlong_t)ds->ds_object,
 882                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 883         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 884             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 885                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 886                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 887                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 888                     scn->scn_phys.scn_queue_obj,
 889                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 890                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 891                     "replacing with %llu",
 892                     (u_longlong_t)ds->ds_object,
 893                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 894         }
 895         dsl_scan_sync_state(scn, tx);
 896 }
 897
 898 void
 899 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 900 {
 901         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 902         dsl_scan_t *scn = dp->dp_scan;
 903         uint64_t mintxg;
 904
 905         if (scn->scn_phys.scn_state != DSS_SCANNING)
 906                 return;
 907
 908         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 909                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 910                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 911                     "reset zb_objset to %llu",
 912                     (u_longlong_t)ds1->ds_object,
 913                     (u_longlong_t)ds2->ds_object);
 914         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 915                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 916                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 917                     "reset zb_objset to %llu",
 918                     (u_longlong_t)ds2->ds_object,
 919                     (u_longlong_t)ds1->ds_object);
 920         }
 921
 922         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 923             ds1->ds_object, &mintxg) == 0) {
 924                 int err;
 925
 926                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 927                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 928                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 929                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 930                 err = zap_add_int_key(dp->dp_meta_objset,
 931                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 932                 VERIFY(err == 0 || err == EEXIST);
 933                 if (err == EEXIST) {
 934                         /* Both were there to begin with */
 935                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 936                             scn->scn_phys.scn_queue_obj,
 937                             ds1->ds_object, mintxg, tx));
 938                 }
 939                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 940                     "replacing with %llu",
 941                     (u_longlong_t)ds1->ds_object,
 942                     (u_longlong_t)ds2->ds_object);
 943         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 944             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 945                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 946                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 947                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 948                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 949                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 950                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 951                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 952                     "replacing with %llu",
 953                     (u_longlong_t)ds2->ds_object,
 954                     (u_longlong_t)ds1->ds_object);
 955         }
 956
 957         dsl_scan_sync_state(scn, tx);
 958 }
 959
 960 struct enqueue_clones_arg {
 961         dmu_tx_t *tx;
 962         uint64_t originobj;
 963 };
 964
 965 /* ARGSUSED */
 966 static int
 967 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 968 {
 969         struct enqueue_clones_arg *eca = arg;
 970         dsl_dataset_t *ds;
 971         int err;
 972         dsl_scan_t *scn = dp->dp_scan;
 973
 974         if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
 975                 return (0);
 976
 977         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
 978         if (err)
 979                 return (err);
 980
 981         while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 982                 dsl_dataset_t *prev;
 983                 err = dsl_dataset_hold_obj(dp,
 984                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 985
 986                 dsl_dataset_rele(ds, FTAG);
 987                 if (err)
 988                         return (err);
 989                 ds = prev;
 990         }
 991         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 992             scn->scn_phys.scn_queue_obj, ds->ds_object,
 993             ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
 994         dsl_dataset_rele(ds, FTAG);
 995         return (0);
 996 }
 997
 998 static void
 999 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1000 {
1001         dsl_pool_t *dp = scn->scn_dp;
1002         dsl_dataset_t *ds;
1003         objset_t *os;
1004         char *dsname;
1005
1006         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1007
1008         if (dmu_objset_from_ds(ds, &os))
1009                 goto out;
1010
1011         /*
1012          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1013          * snapshots can have ZIL block pointers (which may be the same
1014          * BP as in the head), they must be ignored.  So we traverse the
1015          * ZIL here, rather than in scan_recurse(), because the regular
1016          * snapshot block-sharing rules don't apply to it.
1017          */
1018         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1019                 dsl_scan_zil(dp, &os->os_zil_header);
1020
1021         /*
1022          * Iterate over the bps in this ds.
1023          */
1024         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1025         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1026
1027         dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_PUSHPAGE);
1028         dsl_dataset_name(ds, dsname);
1029         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1030             "pausing=%u",
1031             (longlong_t)dsobj, dsname,
1032             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1033             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1034             (int)scn->scn_pausing);
1035         kmem_free(dsname, ZFS_MAXNAMELEN);
1036
1037         if (scn->scn_pausing)
1038                 goto out;
1039
1040         /*
1041          * We've finished this pass over this dataset.
1042          */
1043
1044         /*
1045          * If we did not completely visit this dataset, do another pass.
1046          */
1047         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1048                 zfs_dbgmsg("incomplete pass; visiting again");
1049                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1050                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1051                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1052                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1053                 goto out;
1054         }
1055
1056         /*
1057          * Add descendent datasets to work queue.
1058          */
1059         if (ds->ds_phys->ds_next_snap_obj != 0) {
1060                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1061                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1062                     ds->ds_phys->ds_creation_txg, tx) == 0);
1063         }
1064         if (ds->ds_phys->ds_num_children > 1) {
1065                 boolean_t usenext = B_FALSE;
1066                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1067                         uint64_t count;
1068                         /*
1069                          * A bug in a previous version of the code could
1070                          * cause upgrade_clones_cb() to not set
1071                          * ds_next_snap_obj when it should, leading to a
1072                          * missing entry.  Therefore we can only use the
1073                          * next_clones_obj when its count is correct.
1074                          */
1075                         int err = zap_count(dp->dp_meta_objset,
1076                             ds->ds_phys->ds_next_clones_obj, &count);
1077                         if (err == 0 &&
1078                             count == ds->ds_phys->ds_num_children - 1)
1079                                 usenext = B_TRUE;
1080                 }
1081
1082                 if (usenext) {
1083                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1084                             ds->ds_phys->ds_next_clones_obj,
1085                             scn->scn_phys.scn_queue_obj,
1086                             ds->ds_phys->ds_creation_txg, tx));
1087                 } else {
1088                         struct enqueue_clones_arg eca;
1089                         eca.tx = tx;
1090                         eca.originobj = ds->ds_object;
1091
1092                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1093                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1094                 }
1095         }
1096
1097 out:
1098         dsl_dataset_rele(ds, FTAG);
1099 }
1100
1101 /* ARGSUSED */
1102 static int
1103 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1104 {
1105         dmu_tx_t *tx = arg;
1106         dsl_dataset_t *ds;
1107         int err;
1108         dsl_scan_t *scn = dp->dp_scan;
1109
1110         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1111         if (err)
1112                 return (err);
1113
1114         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1115                 dsl_dataset_t *prev;
1116                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1117                     FTAG, &prev);
1118                 if (err) {
1119                         dsl_dataset_rele(ds, FTAG);
1120                         return (err);
1121                 }
1122
1123                 /*
1124                  * If this is a clone, we don't need to worry about it for now.
1125                  */
1126                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1127                         dsl_dataset_rele(ds, FTAG);
1128                         dsl_dataset_rele(prev, FTAG);
1129                         return (0);
1130                 }
1131                 dsl_dataset_rele(ds, FTAG);
1132                 ds = prev;
1133         }
1134
1135         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1136             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1137         dsl_dataset_rele(ds, FTAG);
1138         return (0);
1139 }
1140
1141 /*
1142  * Scrub/dedup interaction.
1143  *
1144  * If there are N references to a deduped block, we don't want to scrub it
1145  * N times -- ideally, we should scrub it exactly once.
1146  *
1147  * We leverage the fact that the dde's replication class (enum ddt_class)
1148  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1149  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1150  *
1151  * To prevent excess scrubbing, the scrub begins by walking the DDT
1152  * to find all blocks with refcnt > 1, and scrubs each of these once.
1153  * Since there are two replication classes which contain blocks with
1154  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1155  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1156  *
1157  * There would be nothing more to say if a block's refcnt couldn't change
1158  * during a scrub, but of course it can so we must account for changes
1159  * in a block's replication class.
1160  *
1161  * Here's an example of what can occur:
1162  *
1163  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1164  * when visited during the top-down scrub phase, it will be scrubbed twice.
1165  * This negates our scrub optimization, but is otherwise harmless.
1166  *
1167  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1168  * on each visit during the top-down scrub phase, it will never be scrubbed.
1169  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1170  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1171  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1172  * while a scrub is in progress, it scrubs the block right then.
1173  */
1174 static void
1175 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1176 {
1177         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1178         ddt_entry_t dde;
1179         int error;
1180         uint64_t n = 0;
1181
1182         bzero(&dde, sizeof (ddt_entry_t));
1183
1184         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1185                 ddt_t *ddt;
1186
1187                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1188                         break;
1189                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1190                     (longlong_t)ddb->ddb_class,
1191                     (longlong_t)ddb->ddb_type,
1192                     (longlong_t)ddb->ddb_checksum,
1193                     (longlong_t)ddb->ddb_cursor);
1194
1195                 /* There should be no pending changes to the dedup table */
1196                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1197                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1198
1199                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1200                 n++;
1201
1202                 if (dsl_scan_check_pause(scn, NULL))
1203                         break;
1204         }
1205
1206         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1207             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1208             (int)scn->scn_pausing);
1209
1210         ASSERT(error == 0 || error == ENOENT);
1211         ASSERT(error != ENOENT ||
1212             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1213 }
1214
1215 /* ARGSUSED */
1216 void
1217 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1218     ddt_entry_t *dde, dmu_tx_t *tx)
1219 {
1220         const ddt_key_t *ddk = &dde->dde_key;
1221         ddt_phys_t *ddp = dde->dde_phys;
1222         blkptr_t bp;
1223         zbookmark_t zb = { 0 };
1224         int p;
1225
1226         if (scn->scn_phys.scn_state != DSS_SCANNING)
1227                 return;
1228
1229         for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1230                 if (ddp->ddp_phys_birth == 0 ||
1231                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1232                         continue;
1233                 ddt_bp_create(checksum, ddk, ddp, &bp);
1234
1235                 scn->scn_visited_this_txg++;
1236                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1237         }
1238 }
1239
1240 static void
1241 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1242 {
1243         dsl_pool_t *dp = scn->scn_dp;
1244         zap_cursor_t *zc;
1245         zap_attribute_t *za;
1246
1247         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1248             scn->scn_phys.scn_ddt_class_max) {
1249                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1250                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1251                 dsl_scan_ddt(scn, tx);
1252                 if (scn->scn_pausing)
1253                         return;
1254         }
1255
1256         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1257                 /* First do the MOS & ORIGIN */
1258
1259                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1260                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1261                 dsl_scan_visit_rootbp(scn, NULL,
1262                     &dp->dp_meta_rootbp, tx);
1263                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1264                 if (scn->scn_pausing)
1265                         return;
1266
1267                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1268                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1269                             enqueue_cb, tx, DS_FIND_CHILDREN));
1270                 } else {
1271                         dsl_scan_visitds(scn,
1272                             dp->dp_origin_snap->ds_object, tx);
1273                 }
1274                 ASSERT(!scn->scn_pausing);
1275         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1276             ZB_DESTROYED_OBJSET) {
1277                 /*
1278                  * If we were paused, continue from here.  Note if the
1279                  * ds we were paused on was deleted, the zb_objset may
1280                  * be -1, so we will skip this and find a new objset
1281                  * below.
1282                  */
1283                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1284                 if (scn->scn_pausing)
1285                         return;
1286         }
1287
1288         /*
1289          * In case we were paused right at the end of the ds, zero the
1290          * bookmark so we don't think that we're still trying to resume.
1291          */
1292         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1293         zc = kmem_alloc(sizeof(zap_cursor_t), KM_PUSHPAGE);
1294         za = kmem_alloc(sizeof(zap_attribute_t), KM_PUSHPAGE);
1295
1296         /* keep pulling things out of the zap-object-as-queue */
1297         while (zap_cursor_init(zc, dp->dp_meta_objset,
1298             scn->scn_phys.scn_queue_obj),
1299             zap_cursor_retrieve(zc, za) == 0) {
1300                 dsl_dataset_t *ds;
1301                 uint64_t dsobj;
1302
1303                 dsobj = strtonum(za->za_name, NULL);
1304                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1305                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1306
1307                 /* Set up min/max txg */
1308                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1309                 if (za->za_first_integer != 0) {
1310                         scn->scn_phys.scn_cur_min_txg =
1311                             MAX(scn->scn_phys.scn_min_txg,
1312                             za->za_first_integer);
1313                 } else {
1314                         scn->scn_phys.scn_cur_min_txg =
1315                             MAX(scn->scn_phys.scn_min_txg,
1316                             ds->ds_phys->ds_prev_snap_txg);
1317                 }
1318                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1319                 dsl_dataset_rele(ds, FTAG);
1320
1321                 dsl_scan_visitds(scn, dsobj, tx);
1322                 zap_cursor_fini(zc);
1323                 if (scn->scn_pausing)
1324                         goto out;
1325         }
1326         zap_cursor_fini(zc);
1327 out:
1328         kmem_free(za, sizeof(zap_attribute_t));
1329         kmem_free(zc, sizeof(zap_cursor_t));
1330 }
1331
1332 static boolean_t
1333 dsl_scan_free_should_pause(dsl_scan_t *scn)
1334 {
1335         uint64_t elapsed_nanosecs;
1336
1337         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1338         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1339             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1340             txg_sync_waiting(scn->scn_dp)) ||
1341             spa_shutting_down(scn->scn_dp->dp_spa));
1342 }
1343
1344 static int
1345 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1346 {
1347         dsl_scan_t *scn = arg;
1348
1349         if (!scn->scn_is_bptree ||
1350             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1351                 if (dsl_scan_free_should_pause(scn))
1352                         return (SET_ERROR(ERESTART));
1353         }
1354
1355         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1356             dmu_tx_get_txg(tx), bp, 0));
1357         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1358             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1359             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1360         scn->scn_visited_this_txg++;
1361         return (0);
1362 }
1363
1364 boolean_t
1365 dsl_scan_active(dsl_scan_t *scn)
1366 {
1367         spa_t *spa = scn->scn_dp->dp_spa;
1368         uint64_t used = 0, comp, uncomp;
1369
1370         if (spa->spa_load_state != SPA_LOAD_NONE)
1371                 return (B_FALSE);
1372         if (spa_shutting_down(spa))
1373                 return (B_FALSE);
1374
1375         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1376             scn->scn_async_destroying)
1377                 return (B_TRUE);
1378
1379         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1380                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1381                     &used, &comp, &uncomp);
1382         }
1383         return (used != 0);
1384 }
1385
1386 void
1387 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1388 {
1389         dsl_scan_t *scn = dp->dp_scan;
1390         spa_t *spa = dp->dp_spa;
1391         int err;
1392
1393         /*
1394          * Check for scn_restart_txg before checking spa_load_state, so
1395          * that we can restart an old-style scan while the pool is being
1396          * imported (see dsl_scan_init).
1397          */
1398         if (scn->scn_restart_txg != 0 &&
1399             scn->scn_restart_txg <= tx->tx_txg) {
1400                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1401                 dsl_scan_done(scn, B_FALSE, tx);
1402                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1403                         func = POOL_SCAN_RESILVER;
1404                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1405                     func, tx->tx_txg);
1406                 dsl_scan_setup_sync(&func, tx);
1407         }
1408
1409         if (!dsl_scan_active(scn) ||
1410             spa_sync_pass(dp->dp_spa) > 1)
1411                 return;
1412
1413         scn->scn_visited_this_txg = 0;
1414         scn->scn_pausing = B_FALSE;
1415         scn->scn_sync_start_time = gethrtime();
1416         spa->spa_scrub_active = B_TRUE;
1417
1418         /*
1419          * First process the free list.  If we pause the free, don't do
1420          * any scanning.  This ensures that there is no free list when
1421          * we are scanning, so the scan code doesn't have to worry about
1422          * traversing it.
1423          */
1424         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1425                 scn->scn_is_bptree = B_FALSE;
1426                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1427                     NULL, ZIO_FLAG_MUSTSUCCEED);
1428                 err = bpobj_iterate(&dp->dp_free_bpobj,
1429                     dsl_scan_free_block_cb, scn, tx);
1430                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1431
1432                 if (err == 0 && spa_feature_is_active(spa,
1433                     &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1434                         ASSERT(scn->scn_async_destroying);
1435                         scn->scn_is_bptree = B_TRUE;
1436                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1437                             NULL, ZIO_FLAG_MUSTSUCCEED);
1438                         err = bptree_iterate(dp->dp_meta_objset,
1439                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1440                             scn, tx);
1441                         VERIFY0(zio_wait(scn->scn_zio_root));
1442
1443                         if (err == 0) {
1444                                 zfeature_info_t *feat = &spa_feature_table
1445                                     [SPA_FEATURE_ASYNC_DESTROY];
1446                                 /* finished; deactivate async destroy feature */
1447                                 spa_feature_decr(spa, feat, tx);
1448                                 ASSERT(!spa_feature_is_active(spa, feat));
1449                                 VERIFY0(zap_remove(dp->dp_meta_objset,
1450                                     DMU_POOL_DIRECTORY_OBJECT,
1451                                     DMU_POOL_BPTREE_OBJ, tx));
1452                                 VERIFY0(bptree_free(dp->dp_meta_objset,
1453                                     dp->dp_bptree_obj, tx));
1454                                 dp->dp_bptree_obj = 0;
1455                                 scn->scn_async_destroying = B_FALSE;
1456                         }
1457                 }
1458                 if (scn->scn_visited_this_txg) {
1459                         zfs_dbgmsg("freed %llu blocks in %llums from "
1460                             "free_bpobj/bptree txg %llu",
1461                             (longlong_t)scn->scn_visited_this_txg,
1462                             (longlong_t)
1463                             NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1464                             (longlong_t)tx->tx_txg);
1465                         scn->scn_visited_this_txg = 0;
1466                         /*
1467                          * Re-sync the ddt so that we can further modify
1468                          * it when doing bprewrite.
1469                          */
1470                         ddt_sync(spa, tx->tx_txg);
1471                 }
1472                 if (err == ERESTART)
1473                         return;
1474         }
1475
1476         if (scn->scn_phys.scn_state != DSS_SCANNING)
1477                 return;
1478
1479         if (scn->scn_done_txg == tx->tx_txg) {
1480                 ASSERT(!scn->scn_pausing);
1481                 /* finished with scan. */
1482                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1483                 dsl_scan_done(scn, B_TRUE, tx);
1484                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1485                 dsl_scan_sync_state(scn, tx);
1486                 return;
1487         }
1488
1489         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1490             scn->scn_phys.scn_ddt_class_max) {
1491                 zfs_dbgmsg("doing scan sync txg %llu; "
1492                     "ddt bm=%llu/%llu/%llu/%llx",
1493                     (longlong_t)tx->tx_txg,
1494                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1495                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1496                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1497                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1498                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1499                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1500                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1501                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1502         } else {
1503                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1504                     (longlong_t)tx->tx_txg,
1505                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1506                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1507                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1508                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1509         }
1510
1511         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1512             NULL, ZIO_FLAG_CANFAIL);
1513         dsl_pool_config_enter(dp, FTAG);
1514         dsl_scan_visit(scn, tx);
1515         dsl_pool_config_exit(dp, FTAG);
1516         (void) zio_wait(scn->scn_zio_root);
1517         scn->scn_zio_root = NULL;
1518
1519         zfs_dbgmsg("visited %llu blocks in %llums",
1520             (longlong_t)scn->scn_visited_this_txg,
1521             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1522
1523         if (!scn->scn_pausing) {
1524                 scn->scn_done_txg = tx->tx_txg + 1;
1525                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1526                     tx->tx_txg, scn->scn_done_txg);
1527         }
1528
1529         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1530                 mutex_enter(&spa->spa_scrub_lock);
1531                 while (spa->spa_scrub_inflight > 0) {
1532                         cv_wait(&spa->spa_scrub_io_cv,
1533                             &spa->spa_scrub_lock);
1534                 }
1535                 mutex_exit(&spa->spa_scrub_lock);
1536         }
1537
1538         dsl_scan_sync_state(scn, tx);
1539 }
1540
1541 /*
1542  * This will start a new scan, or restart an existing one.
1543  */
1544 void
1545 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1546 {
1547         if (txg == 0) {
1548                 dmu_tx_t *tx;
1549                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1550                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1551
1552                 txg = dmu_tx_get_txg(tx);
1553                 dp->dp_scan->scn_restart_txg = txg;
1554                 dmu_tx_commit(tx);
1555         } else {
1556                 dp->dp_scan->scn_restart_txg = txg;
1557         }
1558         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1559 }
1560
1561 boolean_t
1562 dsl_scan_resilvering(dsl_pool_t *dp)
1563 {
1564         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1565             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1566 }
1567
1568 /*
1569  * scrub consumers
1570  */
1571
1572 static void
1573 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1574 {
1575         int i;
1576
1577         /*
1578          * If we resume after a reboot, zab will be NULL; don't record
1579          * incomplete stats in that case.
1580          */
1581         if (zab == NULL)
1582                 return;
1583
1584         for (i = 0; i < 4; i++) {
1585                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1586                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1587                 int equal;
1588                 zfs_blkstat_t *zb;
1589
1590                 if (t & DMU_OT_NEWTYPE)
1591                         t = DMU_OT_OTHER;
1592
1593                 zb = &zab->zab_type[l][t];
1594                 zb->zb_count++;
1595                 zb->zb_asize += BP_GET_ASIZE(bp);
1596                 zb->zb_lsize += BP_GET_LSIZE(bp);
1597                 zb->zb_psize += BP_GET_PSIZE(bp);
1598                 zb->zb_gangs += BP_COUNT_GANG(bp);
1599
1600                 switch (BP_GET_NDVAS(bp)) {
1601                 case 2:
1602                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1603                             DVA_GET_VDEV(&bp->blk_dva[1]))
1604                                 zb->zb_ditto_2_of_2_samevdev++;
1605                         break;
1606                 case 3:
1607                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1608                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1609                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1610                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1611                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1612                             DVA_GET_VDEV(&bp->blk_dva[2]));
1613                         if (equal == 1)
1614                                 zb->zb_ditto_2_of_3_samevdev++;
1615                         else if (equal == 3)
1616                                 zb->zb_ditto_3_of_3_samevdev++;
1617                         break;
1618                 }
1619         }
1620 }
1621
1622 static void
1623 dsl_scan_scrub_done(zio_t *zio)
1624 {
1625         spa_t *spa = zio->io_spa;
1626
1627         zio_data_buf_free(zio->io_data, zio->io_size);
1628
1629         mutex_enter(&spa->spa_scrub_lock);
1630         spa->spa_scrub_inflight--;
1631         cv_broadcast(&spa->spa_scrub_io_cv);
1632
1633         if (zio->io_error && (zio->io_error != ECKSUM ||
1634             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1635                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1636         }
1637         mutex_exit(&spa->spa_scrub_lock);
1638 }
1639
1640 static int
1641 dsl_scan_scrub_cb(dsl_pool_t *dp,
1642     const blkptr_t *bp, const zbookmark_t *zb)
1643 {
1644         dsl_scan_t *scn = dp->dp_scan;
1645         size_t size = BP_GET_PSIZE(bp);
1646         spa_t *spa = dp->dp_spa;
1647         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1648         boolean_t needs_io = B_FALSE;
1649         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1650         int zio_priority = 0;
1651         int scan_delay = 0;
1652         int d;
1653
1654         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1655             phys_birth >= scn->scn_phys.scn_max_txg)
1656                 return (0);
1657
1658         count_block(dp->dp_blkstats, bp);
1659
1660         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1661         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1662                 zio_flags |= ZIO_FLAG_SCRUB;
1663                 zio_priority = ZIO_PRIORITY_SCRUB;
1664                 needs_io = B_TRUE;
1665                 scan_delay = zfs_scrub_delay;
1666         } else {
1667                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1668                 zio_flags |= ZIO_FLAG_RESILVER;
1669                 zio_priority = ZIO_PRIORITY_RESILVER;
1670                 needs_io = B_FALSE;
1671                 scan_delay = zfs_resilver_delay;
1672         }
1673
1674         /* If it's an intent log block, failure is expected. */
1675         if (zb->zb_level == ZB_ZIL_LEVEL)
1676                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1677
1678         for (d = 0; d < BP_GET_NDVAS(bp); d++) {
1679                 vdev_t *vd = vdev_lookup_top(spa,
1680                     DVA_GET_VDEV(&bp->blk_dva[d]));
1681
1682                 /*
1683                  * Keep track of how much data we've examined so that
1684                  * zpool(1M) status can make useful progress reports.
1685                  */
1686                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1687                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1688
1689                 /* if it's a resilver, this may not be in the target range */
1690                 if (!needs_io) {
1691                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1692                                 /*
1693                                  * Gang members may be spread across multiple
1694                                  * vdevs, so the best estimate we have is the
1695                                  * scrub range, which has already been checked.
1696                                  * XXX -- it would be better to change our
1697                                  * allocation policy to ensure that all
1698                                  * gang members reside on the same vdev.
1699                                  */
1700                                 needs_io = B_TRUE;
1701                         } else {
1702                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1703                                     phys_birth, 1);
1704                         }
1705                 }
1706         }
1707
1708         if (needs_io && !zfs_no_scrub_io) {
1709                 vdev_t *rvd = spa->spa_root_vdev;
1710                 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1711                 void *data = zio_data_buf_alloc(size);
1712
1713                 mutex_enter(&spa->spa_scrub_lock);
1714                 while (spa->spa_scrub_inflight >= maxinflight)
1715                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1716                 spa->spa_scrub_inflight++;
1717                 mutex_exit(&spa->spa_scrub_lock);
1718
1719                 /*
1720                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1721                  * then throttle our workload to limit the impact of a scan.
1722                  */
1723                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1724                         delay(scan_delay);
1725
1726                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1727                     dsl_scan_scrub_done, NULL, zio_priority,
1728                     zio_flags, zb));
1729         }
1730
1731         /* do not relocate this block */
1732         return (0);
1733 }
1734
1735 int
1736 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1737 {
1738         spa_t *spa = dp->dp_spa;
1739
1740         /*
1741          * Purge all vdev caches and probe all devices.  We do this here
1742          * rather than in sync context because this requires a writer lock
1743          * on the spa_config lock, which we can't do from sync context.  The
1744          * spa_scrub_reopen flag indicates that vdev_open() should not
1745          * attempt to start another scrub.
1746          */
1747         spa_vdev_state_enter(spa, SCL_NONE);
1748         spa->spa_scrub_reopen = B_TRUE;
1749         vdev_reopen(spa->spa_root_vdev);
1750         spa->spa_scrub_reopen = B_FALSE;
1751         (void) spa_vdev_state_exit(spa, NULL, 0);
1752
1753         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1754             dsl_scan_setup_sync, &func, 0));
1755 }
1756
1757 #if defined(_KERNEL) && defined(HAVE_SPL)
1758 module_param(zfs_top_maxinflight, int, 0644);
1759 MODULE_PARM_DESC(zfs_top_maxinflight, "Max I/Os per top-level");
1760
1761 module_param(zfs_resilver_delay, int, 0644);
1762 MODULE_PARM_DESC(zfs_resilver_delay, "Number of ticks to delay resilver");
1763
1764 module_param(zfs_scrub_delay, int, 0644);
1765 MODULE_PARM_DESC(zfs_scrub_delay, "Number of ticks to delay scrub");
1766
1767 module_param(zfs_scan_idle, int, 0644);
1768 MODULE_PARM_DESC(zfs_scan_idle, "Idle window in clock ticks");
1769
1770 module_param(zfs_scan_min_time_ms, int, 0644);
1771 MODULE_PARM_DESC(zfs_scan_min_time_ms, "Min millisecs to scrub per txg");
1772
1773 module_param(zfs_free_min_time_ms, int, 0644);
1774 MODULE_PARM_DESC(zfs_free_min_time_ms, "Min millisecs to free per txg");
1775
1776 module_param(zfs_resilver_min_time_ms, int, 0644);
1777 MODULE_PARM_DESC(zfs_resilver_min_time_ms, "Min millisecs to resilver per txg");
1778
1779 module_param(zfs_no_scrub_io, int, 0644);
1780 MODULE_PARM_DESC(zfs_no_scrub_io, "Set to disable scrub I/O");
1781
1782 module_param(zfs_no_scrub_prefetch, int, 0644);
1783 MODULE_PARM_DESC(zfs_no_scrub_prefetch, "Set to disable scrub prefetching");
1784 #endif