1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2012 by Delphix. All rights reserved.
  24  */
  25 
  26 #include <sys/dsl_scan.h>
  27 #include <sys/dsl_pool.h>
  28 #include <sys/dsl_dataset.h>
  29 #include <sys/dsl_prop.h>
  30 #include <sys/dsl_dir.h>
  31 #include <sys/dsl_synctask.h>
  32 #include <sys/dnode.h>
  33 #include <sys/dmu_tx.h>
  34 #include <sys/dmu_objset.h>
  35 #include <sys/arc.h>
  36 #include <sys/zap.h>
  37 #include <sys/zio.h>
  38 #include <sys/zfs_context.h>
  39 #include <sys/fs/zfs.h>
  40 #include <sys/zfs_znode.h>
  41 #include <sys/spa_impl.h>
  42 #include <sys/vdev_impl.h>
  43 #include <sys/zil_impl.h>
  44 #include <sys/zio_checksum.h>
  45 #include <sys/ddt.h>
  46 #include <sys/sa.h>
  47 #include <sys/sa_impl.h>
  48 #include <sys/zfeature.h>
  49 #ifdef _KERNEL
  50 #include <sys/zfs_vfsops.h>
  51 #endif
  52 
  53 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
  54 
  55 static scan_cb_t dsl_scan_defrag_cb;
  56 static scan_cb_t dsl_scan_scrub_cb;
  57 static scan_cb_t dsl_scan_remove_cb;
  58 static dsl_syncfunc_t dsl_scan_cancel_sync;
  59 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
  60 
  61 int zfs_top_maxinflight = 32;           /* maximum I/Os per top-level */
  62 int zfs_resilver_delay = 2;             /* number of ticks to delay resilver */
  63 int zfs_scrub_delay = 4;                /* number of ticks to delay scrub */
  64 int zfs_scan_idle = 50;                 /* idle window in clock ticks */
  65 
  66 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
  67 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
  68 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
  69 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  70 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable srub prefetching */
  71 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
  72 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
  73 
  74 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
  75         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
  76         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
  77 
  78 extern int zfs_txg_timeout;
  79 
  80 /* the order has to match pool_scan_type */
  81 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
  82         NULL,
  83         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
  84         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
  85 };
  86 
  87 int
  88 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
  89 {
  90         int err;
  91         dsl_scan_t *scn;
  92         spa_t *spa = dp->dp_spa;
  93         uint64_t f;
  94 
  95         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
  96         scn->scn_dp = dp;
  97 
  98         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
  99             "scrub_func", sizeof (uint64_t), 1, &f);
 100         if (err == 0) {
 101                 /*
 102                  * There was an old-style scrub in progress.  Restart a
 103                  * new-style scrub from the beginning.
 104                  */
 105                 scn->scn_restart_txg = txg;
 106                 zfs_dbgmsg("old-style scrub was in progress; "
 107                     "restarting new-style scrub in txg %llu",
 108                     scn->scn_restart_txg);
 109 
 110                 /*
 111                  * Load the queue obj from the old location so that it
 112                  * can be freed by dsl_scan_done().
 113                  */
 114                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 115                     "scrub_queue", sizeof (uint64_t), 1,
 116                     &scn->scn_phys.scn_queue_obj);
 117         } else {
 118                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 119                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 120                     &scn->scn_phys);
 121                 if (err == ENOENT)
 122                         return (0);
 123                 else if (err)
 124                         return (err);
 125 
 126                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 127                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 128                         /*
 129                          * A new-type scrub was in progress on an old
 130                          * pool, and the pool was accessed by old
 131                          * software.  Restart from the beginning, since
 132                          * the old software may have changed the pool in
 133                          * the meantime.
 134                          */
 135                         scn->scn_restart_txg = txg;
 136                         zfs_dbgmsg("new-style scrub was modified "
 137                             "by old software; restarting in txg %llu",
 138                             scn->scn_restart_txg);
 139                 }
 140         }
 141 
 142         spa_scan_stat_init(spa);
 143         return (0);
 144 }
 145 
 146 void
 147 dsl_scan_fini(dsl_pool_t *dp)
 148 {
 149         if (dp->dp_scan) {
 150                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 151                 dp->dp_scan = NULL;
 152         }
 153 }
 154 
 155 /* ARGSUSED */
 156 static int
 157 dsl_scan_setup_check(void *arg1, void *arg2, dmu_tx_t *tx)
 158 {
 159         dsl_scan_t *scn = arg1;
 160 
 161         if (scn->scn_phys.scn_state == DSS_SCANNING)
 162                 return (EBUSY);
 163 
 164         return (0);
 165 }
 166 
 167 /* ARGSUSED */
 168 static void
 169 dsl_scan_setup_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 170 {
 171         dsl_scan_t *scn = arg1;
 172         pool_scan_func_t *funcp = arg2;
 173         dmu_object_type_t ot = 0;
 174         dsl_pool_t *dp = scn->scn_dp;
 175         spa_t *spa = dp->dp_spa;
 176 
 177         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 178         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 179         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 180         scn->scn_phys.scn_func = *funcp;
 181         scn->scn_phys.scn_state = DSS_SCANNING;
 182         scn->scn_phys.scn_min_txg = 0;
 183         scn->scn_phys.scn_max_txg = tx->tx_txg;
 184         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 185         scn->scn_phys.scn_start_time = gethrestime_sec();
 186         scn->scn_phys.scn_errors = 0;
 187         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 188         scn->scn_restart_txg = 0;
 189         spa_scan_stat_init(spa);
 190 
 191         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 192                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 193 
 194                 /* rewrite all disk labels */
 195                 vdev_config_dirty(spa->spa_root_vdev);
 196 
 197                 if (vdev_resilver_needed(spa->spa_root_vdev,
 198                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 199                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 200                 } else {
 201                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 202                 }
 203 
 204                 spa->spa_scrub_started = B_TRUE;
 205                 /*
 206                  * If this is an incremental scrub, limit the DDT scrub phase
 207                  * to just the auto-ditto class (for correctness); the rest
 208                  * of the scrub should go faster using top-down pruning.
 209                  */
 210                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 211                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 212 
 213         }
 214 
 215         /* back to the generic stuff */
 216 
 217         if (dp->dp_blkstats == NULL) {
 218                 dp->dp_blkstats =
 219                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 220         }
 221         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 222 
 223         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 224                 ot = DMU_OT_ZAP_OTHER;
 225 
 226         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 227             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 228 
 229         dsl_scan_sync_state(scn, tx);
 230 
 231         spa_history_log_internal(spa, "scan setup", tx,
 232             "func=%u mintxg=%llu maxtxg=%llu",
 233             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 234 }
 235 
 236 /* ARGSUSED */
 237 static void
 238 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 239 {
 240         static const char *old_names[] = {
 241                 "scrub_bookmark",
 242                 "scrub_ddt_bookmark",
 243                 "scrub_ddt_class_max",
 244                 "scrub_queue",
 245                 "scrub_min_txg",
 246                 "scrub_max_txg",
 247                 "scrub_func",
 248                 "scrub_errors",
 249                 NULL
 250         };
 251 
 252         dsl_pool_t *dp = scn->scn_dp;
 253         spa_t *spa = dp->dp_spa;
 254         int i;
 255 
 256         /* Remove any remnants of an old-style scrub. */
 257         for (i = 0; old_names[i]; i++) {
 258                 (void) zap_remove(dp->dp_meta_objset,
 259                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 260         }
 261 
 262         if (scn->scn_phys.scn_queue_obj != 0) {
 263                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 264                     scn->scn_phys.scn_queue_obj, tx));
 265                 scn->scn_phys.scn_queue_obj = 0;
 266         }
 267 
 268         /*
 269          * If we were "restarted" from a stopped state, don't bother
 270          * with anything else.
 271          */
 272         if (scn->scn_phys.scn_state != DSS_SCANNING)
 273                 return;
 274 
 275         if (complete)
 276                 scn->scn_phys.scn_state = DSS_FINISHED;
 277         else
 278                 scn->scn_phys.scn_state = DSS_CANCELED;
 279 
 280         spa_history_log_internal(spa, "scan done", tx,
 281             "complete=%u", complete);
 282 
 283         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 284                 mutex_enter(&spa->spa_scrub_lock);
 285                 while (spa->spa_scrub_inflight > 0) {
 286                         cv_wait(&spa->spa_scrub_io_cv,
 287                             &spa->spa_scrub_lock);
 288                 }
 289                 mutex_exit(&spa->spa_scrub_lock);
 290                 spa->spa_scrub_started = B_FALSE;
 291                 spa->spa_scrub_active = B_FALSE;
 292 
 293                 /*
 294                  * If the scrub/resilver completed, update all DTLs to
 295                  * reflect this.  Whether it succeeded or not, vacate
 296                  * all temporary scrub DTLs.
 297                  */
 298                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 299                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 300                 if (complete) {
 301                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 302                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 303                 }
 304                 spa_errlog_rotate(spa);
 305 
 306                 /*
 307                  * We may have finished replacing a device.
 308                  * Let the async thread assess this and handle the detach.
 309                  */
 310                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 311         }
 312 
 313         scn->scn_phys.scn_end_time = gethrestime_sec();
 314 }
 315 
 316 /* ARGSUSED */
 317 static int
 318 dsl_scan_cancel_check(void *arg1, void *arg2, dmu_tx_t *tx)
 319 {
 320         dsl_scan_t *scn = arg1;
 321 
 322         if (scn->scn_phys.scn_state != DSS_SCANNING)
 323                 return (ENOENT);
 324         return (0);
 325 }
 326 
 327 /* ARGSUSED */
 328 static void
 329 dsl_scan_cancel_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 330 {
 331         dsl_scan_t *scn = arg1;
 332 
 333         dsl_scan_done(scn, B_FALSE, tx);
 334         dsl_scan_sync_state(scn, tx);
 335 }
 336 
 337 int
 338 dsl_scan_cancel(dsl_pool_t *dp)
 339 {
 340         boolean_t complete = B_FALSE;
 341         int err;
 342 
 343         err = dsl_sync_task_do(dp, dsl_scan_cancel_check,
 344             dsl_scan_cancel_sync, dp->dp_scan, &complete, 3);
 345         return (err);
 346 }
 347 
 348 static void dsl_scan_visitbp(blkptr_t *bp,
 349     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 350     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 351     dmu_tx_t *tx);
 352 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 353     dmu_objset_type_t ostype,
 354     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 355 
 356 void
 357 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 358 {
 359         zio_free(dp->dp_spa, txg, bp);
 360 }
 361 
 362 void
 363 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 364 {
 365         ASSERT(dsl_pool_sync_context(dp));
 366         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
 367 }
 368 
 369 int
 370 dsl_read(zio_t *pio, spa_t *spa, const blkptr_t *bpp, arc_buf_t *pbuf,
 371     arc_done_func_t *done, void *private, int priority, int zio_flags,
 372     uint32_t *arc_flags, const zbookmark_t *zb)
 373 {
 374         return (arc_read(pio, spa, bpp, pbuf, done, private,
 375             priority, zio_flags, arc_flags, zb));
 376 }
 377 
 378 int
 379 dsl_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bpp,
 380     arc_done_func_t *done, void *private, int priority, int zio_flags,
 381     uint32_t *arc_flags, const zbookmark_t *zb)
 382 {
 383         return (arc_read_nolock(pio, spa, bpp, done, private,
 384             priority, zio_flags, arc_flags, zb));
 385 }
 386 
 387 static uint64_t
 388 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 389 {
 390         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 391         if (dsl_dataset_is_snapshot(ds))
 392                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 393         return (smt);
 394 }
 395 
 396 static void
 397 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 398 {
 399         VERIFY(0 == zap_update(scn->scn_dp->dp_meta_objset,
 400             DMU_POOL_DIRECTORY_OBJECT,
 401             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 402             &scn->scn_phys, tx));
 403 }
 404 
 405 static boolean_t
 406 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 407 {
 408         uint64_t elapsed_nanosecs;
 409         int mintime;
 410 
 411         /* we never skip user/group accounting objects */
 412         if (zb && (int64_t)zb->zb_object < 0)
 413                 return (B_FALSE);
 414 
 415         if (scn->scn_pausing)
 416                 return (B_TRUE); /* we're already pausing */
 417 
 418         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 419                 return (B_FALSE); /* we're resuming */
 420 
 421         /* We only know how to resume from level-0 blocks. */
 422         if (zb && zb->zb_level != 0)
 423                 return (B_FALSE);
 424 
 425         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 426             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 427         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 428         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 429             (elapsed_nanosecs / MICROSEC > mintime &&
 430             txg_sync_waiting(scn->scn_dp)) ||
 431             spa_shutting_down(scn->scn_dp->dp_spa)) {
 432                 if (zb) {
 433                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 434                             (longlong_t)zb->zb_objset,
 435                             (longlong_t)zb->zb_object,
 436                             (longlong_t)zb->zb_level,
 437                             (longlong_t)zb->zb_blkid);
 438                         scn->scn_phys.scn_bookmark = *zb;
 439                 }
 440                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 441                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 442                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 443                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 444                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 445                 scn->scn_pausing = B_TRUE;
 446                 return (B_TRUE);
 447         }
 448         return (B_FALSE);
 449 }
 450 
 451 typedef struct zil_scan_arg {
 452         dsl_pool_t      *zsa_dp;
 453         zil_header_t    *zsa_zh;
 454 } zil_scan_arg_t;
 455 
 456 /* ARGSUSED */
 457 static int
 458 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 459 {
 460         zil_scan_arg_t *zsa = arg;
 461         dsl_pool_t *dp = zsa->zsa_dp;
 462         dsl_scan_t *scn = dp->dp_scan;
 463         zil_header_t *zh = zsa->zsa_zh;
 464         zbookmark_t zb;
 465 
 466         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 467                 return (0);
 468 
 469         /*
 470          * One block ("stubby") can be allocated a long time ago; we
 471          * want to visit that one because it has been allocated
 472          * (on-disk) even if it hasn't been claimed (even though for
 473          * scrub there's nothing to do to it).
 474          */
 475         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 476                 return (0);
 477 
 478         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 479             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 480 
 481         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 482         return (0);
 483 }
 484 
 485 /* ARGSUSED */
 486 static int
 487 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 488 {
 489         if (lrc->lrc_txtype == TX_WRITE) {
 490                 zil_scan_arg_t *zsa = arg;
 491                 dsl_pool_t *dp = zsa->zsa_dp;
 492                 dsl_scan_t *scn = dp->dp_scan;
 493                 zil_header_t *zh = zsa->zsa_zh;
 494                 lr_write_t *lr = (lr_write_t *)lrc;
 495                 blkptr_t *bp = &lr->lr_blkptr;
 496                 zbookmark_t zb;
 497 
 498                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 499                         return (0);
 500 
 501                 /*
 502                  * birth can be < claim_txg if this record's txg is
 503                  * already txg sync'ed (but this log block contains
 504                  * other records that are not synced)
 505                  */
 506                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 507                         return (0);
 508 
 509                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 510                     lr->lr_foid, ZB_ZIL_LEVEL,
 511                     lr->lr_offset / BP_GET_LSIZE(bp));
 512 
 513                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 514         }
 515         return (0);
 516 }
 517 
 518 static void
 519 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 520 {
 521         uint64_t claim_txg = zh->zh_claim_txg;
 522         zil_scan_arg_t zsa = { dp, zh };
 523         zilog_t *zilog;
 524 
 525         /*
 526          * We only want to visit blocks that have been claimed but not yet
 527          * replayed (or, in read-only mode, blocks that *would* be claimed).
 528          */
 529         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 530                 return;
 531 
 532         zilog = zil_alloc(dp->dp_meta_objset, zh);
 533 
 534         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 535             claim_txg);
 536 
 537         zil_free(zilog);
 538 }
 539 
 540 /* ARGSUSED */
 541 static void
 542 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 543     uint64_t objset, uint64_t object, uint64_t blkid)
 544 {
 545         zbookmark_t czb;
 546         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 547 
 548         if (zfs_no_scrub_prefetch)
 549                 return;
 550 
 551         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 552             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 553                 return;
 554 
 555         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 556 
 557         /*
 558          * XXX need to make sure all of these arc_read() prefetches are
 559          * done before setting xlateall (similar to dsl_read())
 560          */
 561         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 562             buf, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 563             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 564 }
 565 
 566 static boolean_t
 567 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 568     const zbookmark_t *zb)
 569 {
 570         /*
 571          * We never skip over user/group accounting objects (obj<0)
 572          */
 573         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 574             (int64_t)zb->zb_object >= 0) {
 575                 /*
 576                  * If we already visited this bp & everything below (in
 577                  * a prior txg sync), don't bother doing it again.
 578                  */
 579                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 580                         return (B_TRUE);
 581 
 582                 /*
 583                  * If we found the block we're trying to resume from, or
 584                  * we went past it to a different object, zero it out to
 585                  * indicate that it's OK to start checking for pausing
 586                  * again.
 587                  */
 588                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 589                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 590                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 591                             (longlong_t)zb->zb_objset,
 592                             (longlong_t)zb->zb_object,
 593                             (longlong_t)zb->zb_level,
 594                             (longlong_t)zb->zb_blkid);
 595                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 596                 }
 597         }
 598         return (B_FALSE);
 599 }
 600 
 601 /*
 602  * Return nonzero on i/o error.
 603  * Return new buf to write out in *bufp.
 604  */
 605 static int
 606 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 607     dnode_phys_t *dnp, const blkptr_t *bp,
 608     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 609 {
 610         dsl_pool_t *dp = scn->scn_dp;
 611         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 612         int err;
 613 
 614         if (BP_GET_LEVEL(bp) > 0) {
 615                 uint32_t flags = ARC_WAIT;
 616                 int i;
 617                 blkptr_t *cbp;
 618                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 619 
 620                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 621                     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                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 628                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 629                             zb->zb_object, zb->zb_blkid * epb + i);
 630                 }
 631                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 632                         zbookmark_t czb;
 633 
 634                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 635                             zb->zb_level - 1,
 636                             zb->zb_blkid * epb + i);
 637                         dsl_scan_visitbp(cbp, &czb, dnp,
 638                             *bufp, ds, scn, ostype, tx);
 639                 }
 640         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 641                 uint32_t flags = ARC_WAIT;
 642 
 643                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 644                     arc_getbuf_func, bufp,
 645                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 646                 if (err) {
 647                         scn->scn_phys.scn_errors++;
 648                         return (err);
 649                 }
 650         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 651                 uint32_t flags = ARC_WAIT;
 652                 dnode_phys_t *cdnp;
 653                 int i, j;
 654                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 655 
 656                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 657                     arc_getbuf_func, bufp,
 658                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 659                 if (err) {
 660                         scn->scn_phys.scn_errors++;
 661                         return (err);
 662                 }
 663                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 664                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 665                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 666                                 dsl_scan_prefetch(scn, *bufp, cbp,
 667                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 668                         }
 669                 }
 670                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 671                         dsl_scan_visitdnode(scn, ds, ostype,
 672                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 673                 }
 674 
 675         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 676                 uint32_t flags = ARC_WAIT;
 677                 objset_phys_t *osp;
 678 
 679                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 680                     arc_getbuf_func, bufp,
 681                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 682                 if (err) {
 683                         scn->scn_phys.scn_errors++;
 684                         return (err);
 685                 }
 686 
 687                 osp = (*bufp)->b_data;
 688 
 689                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 690                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 691 
 692                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 693                         /*
 694                          * We also always visit user/group accounting
 695                          * objects, and never skip them, even if we are
 696                          * pausing.  This is necessary so that the space
 697                          * deltas from this txg get integrated.
 698                          */
 699                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 700                             &osp->os_groupused_dnode, *bufp,
 701                             DMU_GROUPUSED_OBJECT, tx);
 702                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 703                             &osp->os_userused_dnode, *bufp,
 704                             DMU_USERUSED_OBJECT, tx);
 705                 }
 706         }
 707 
 708         return (0);
 709 }
 710 
 711 static void
 712 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 713     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 714     uint64_t object, dmu_tx_t *tx)
 715 {
 716         int j;
 717 
 718         for (j = 0; j < dnp->dn_nblkptr; j++) {
 719                 zbookmark_t czb;
 720 
 721                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 722                     dnp->dn_nlevels - 1, j);
 723                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 724                     &czb, dnp, buf, ds, scn, ostype, tx);
 725         }
 726 
 727         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 728                 zbookmark_t czb;
 729                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 730                     0, DMU_SPILL_BLKID);
 731                 dsl_scan_visitbp(&dnp->dn_spill,
 732                     &czb, dnp, buf, ds, scn, ostype, tx);
 733         }
 734 }
 735 
 736 /*
 737  * The arguments are in this order because mdb can only print the
 738  * first 5; we want them to be useful.
 739  */
 740 static void
 741 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 742     dnode_phys_t *dnp, arc_buf_t *pbuf,
 743     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 744     dmu_tx_t *tx)
 745 {
 746         dsl_pool_t *dp = scn->scn_dp;
 747         arc_buf_t *buf = NULL;
 748         blkptr_t bp_toread = *bp;
 749 
 750         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 751 
 752         if (dsl_scan_check_pause(scn, zb))
 753                 return;
 754 
 755         if (dsl_scan_check_resume(scn, dnp, zb))
 756                 return;
 757 
 758         if (bp->blk_birth == 0)
 759                 return;
 760 
 761         scn->scn_visited_this_txg++;
 762 
 763         dprintf_bp(bp,
 764             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 765             ds, ds ? ds->ds_object : 0,
 766             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 767             pbuf, bp);
 768 
 769         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 770                 return;
 771 
 772         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
 773             &buf) != 0)
 774                 return;
 775 
 776         /*
 777          * If dsl_scan_ddt() has aready visited this block, it will have
 778          * already done any translations or scrubbing, so don't call the
 779          * callback again.
 780          */
 781         if (ddt_class_contains(dp->dp_spa,
 782             scn->scn_phys.scn_ddt_class_max, bp)) {
 783                 ASSERT(buf == NULL);
 784                 return;
 785         }
 786 
 787         /*
 788          * If this block is from the future (after cur_max_txg), then we
 789          * are doing this on behalf of a deleted snapshot, and we will
 790          * revisit the future block on the next pass of this dataset.
 791          * Don't scan it now unless we need to because something
 792          * under it was modified.
 793          */
 794         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 795                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 796         }
 797         if (buf)
 798                 (void) arc_buf_remove_ref(buf, &buf);
 799 }
 800 
 801 static void
 802 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 803     dmu_tx_t *tx)
 804 {
 805         zbookmark_t zb;
 806 
 807         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 808             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 809         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 810             ds, scn, DMU_OST_NONE, tx);
 811 
 812         dprintf_ds(ds, "finished scan%s", "");
 813 }
 814 
 815 void
 816 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 817 {
 818         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 819         dsl_scan_t *scn = dp->dp_scan;
 820         uint64_t mintxg;
 821 
 822         if (scn->scn_phys.scn_state != DSS_SCANNING)
 823                 return;
 824 
 825         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 826                 if (dsl_dataset_is_snapshot(ds)) {
 827                         /* Note, scn_cur_{min,max}_txg stays the same. */
 828                         scn->scn_phys.scn_bookmark.zb_objset =
 829                             ds->ds_phys->ds_next_snap_obj;
 830                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 831                             "reset zb_objset to %llu",
 832                             (u_longlong_t)ds->ds_object,
 833                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 834                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 835                 } else {
 836                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 837                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 838                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 839                             "reset bookmark to -1,0,0,0",
 840                             (u_longlong_t)ds->ds_object);
 841                 }
 842         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 843             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 844                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 845                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 846                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 847                 if (dsl_dataset_is_snapshot(ds)) {
 848                         /*
 849                          * We keep the same mintxg; it could be >
 850                          * ds_creation_txg if the previous snapshot was
 851                          * deleted too.
 852                          */
 853                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 854                             scn->scn_phys.scn_queue_obj,
 855                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 856                         zfs_dbgmsg("destroying ds %llu; in queue; "
 857                             "replacing with %llu",
 858                             (u_longlong_t)ds->ds_object,
 859                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 860                 } else {
 861                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 862                             (u_longlong_t)ds->ds_object);
 863                 }
 864         } else {
 865                 zfs_dbgmsg("destroying ds %llu; ignoring",
 866                     (u_longlong_t)ds->ds_object);
 867         }
 868 
 869         /*
 870          * dsl_scan_sync() should be called after this, and should sync
 871          * out our changed state, but just to be safe, do it here.
 872          */
 873         dsl_scan_sync_state(scn, tx);
 874 }
 875 
 876 void
 877 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 878 {
 879         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 880         dsl_scan_t *scn = dp->dp_scan;
 881         uint64_t mintxg;
 882 
 883         if (scn->scn_phys.scn_state != DSS_SCANNING)
 884                 return;
 885 
 886         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 887 
 888         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 889                 scn->scn_phys.scn_bookmark.zb_objset =
 890                     ds->ds_phys->ds_prev_snap_obj;
 891                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 892                     "reset zb_objset to %llu",
 893                     (u_longlong_t)ds->ds_object,
 894                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 895         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 896             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 897                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 898                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 899                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 900                     scn->scn_phys.scn_queue_obj,
 901                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 902                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 903                     "replacing with %llu",
 904                     (u_longlong_t)ds->ds_object,
 905                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 906         }
 907         dsl_scan_sync_state(scn, tx);
 908 }
 909 
 910 void
 911 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 912 {
 913         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 914         dsl_scan_t *scn = dp->dp_scan;
 915         uint64_t mintxg;
 916 
 917         if (scn->scn_phys.scn_state != DSS_SCANNING)
 918                 return;
 919 
 920         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 921                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 922                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 923                     "reset zb_objset to %llu",
 924                     (u_longlong_t)ds1->ds_object,
 925                     (u_longlong_t)ds2->ds_object);
 926         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 927                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 928                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 929                     "reset zb_objset to %llu",
 930                     (u_longlong_t)ds2->ds_object,
 931                     (u_longlong_t)ds1->ds_object);
 932         }
 933 
 934         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 935             ds1->ds_object, &mintxg) == 0) {
 936                 int err;
 937 
 938                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 939                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 940                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 941                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 942                 err = zap_add_int_key(dp->dp_meta_objset,
 943                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 944                 VERIFY(err == 0 || err == EEXIST);
 945                 if (err == EEXIST) {
 946                         /* Both were there to begin with */
 947                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 948                             scn->scn_phys.scn_queue_obj,
 949                             ds1->ds_object, mintxg, tx));
 950                 }
 951                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 952                     "replacing with %llu",
 953                     (u_longlong_t)ds1->ds_object,
 954                     (u_longlong_t)ds2->ds_object);
 955         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 956             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 957                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 958                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 959                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 960                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 961                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 962                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 963                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 964                     "replacing with %llu",
 965                     (u_longlong_t)ds2->ds_object,
 966                     (u_longlong_t)ds1->ds_object);
 967         }
 968 
 969         dsl_scan_sync_state(scn, tx);
 970 }
 971 
 972 struct enqueue_clones_arg {
 973         dmu_tx_t *tx;
 974         uint64_t originobj;
 975 };
 976 
 977 /* ARGSUSED */
 978 static int
 979 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 980 {
 981         struct enqueue_clones_arg *eca = arg;
 982         dsl_dataset_t *ds;
 983         int err;
 984         dsl_pool_t *dp = spa->spa_dsl_pool;
 985         dsl_scan_t *scn = dp->dp_scan;
 986 
 987         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
 988         if (err)
 989                 return (err);
 990 
 991         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
 992                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 993                         dsl_dataset_t *prev;
 994                         err = dsl_dataset_hold_obj(dp,
 995                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 996 
 997                         dsl_dataset_rele(ds, FTAG);
 998                         if (err)
 999                                 return (err);
1000                         ds = prev;
1001                 }
1002                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1003                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1004                     ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1005         }
1006         dsl_dataset_rele(ds, FTAG);
1007         return (0);
1008 }
1009 
1010 static void
1011 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1012 {
1013         dsl_pool_t *dp = scn->scn_dp;
1014         dsl_dataset_t *ds;
1015         objset_t *os;
1016 
1017         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1018 
1019         if (dmu_objset_from_ds(ds, &os))
1020                 goto out;
1021 
1022         /*
1023          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1024          * snapshots can have ZIL block pointers (which may be the same
1025          * BP as in the head), they must be ignored.  So we traverse the
1026          * ZIL here, rather than in scan_recurse(), because the regular
1027          * snapshot block-sharing rules don't apply to it.
1028          */
1029         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1030                 dsl_scan_zil(dp, &os->os_zil_header);
1031 
1032         /*
1033          * Iterate over the bps in this ds.
1034          */
1035         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1036         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1037 
1038         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1039         dsl_dataset_name(ds, dsname);
1040         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1041             "pausing=%u",
1042             (longlong_t)dsobj, dsname,
1043             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1044             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1045             (int)scn->scn_pausing);
1046         kmem_free(dsname, ZFS_MAXNAMELEN);
1047 
1048         if (scn->scn_pausing)
1049                 goto out;
1050 
1051         /*
1052          * We've finished this pass over this dataset.
1053          */
1054 
1055         /*
1056          * If we did not completely visit this dataset, do another pass.
1057          */
1058         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1059                 zfs_dbgmsg("incomplete pass; visiting again");
1060                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1061                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1062                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1063                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1064                 goto out;
1065         }
1066 
1067         /*
1068          * Add descendent datasets to work queue.
1069          */
1070         if (ds->ds_phys->ds_next_snap_obj != 0) {
1071                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1072                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1073                     ds->ds_phys->ds_creation_txg, tx) == 0);
1074         }
1075         if (ds->ds_phys->ds_num_children > 1) {
1076                 boolean_t usenext = B_FALSE;
1077                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1078                         uint64_t count;
1079                         /*
1080                          * A bug in a previous version of the code could
1081                          * cause upgrade_clones_cb() to not set
1082                          * ds_next_snap_obj when it should, leading to a
1083                          * missing entry.  Therefore we can only use the
1084                          * next_clones_obj when its count is correct.
1085                          */
1086                         int err = zap_count(dp->dp_meta_objset,
1087                             ds->ds_phys->ds_next_clones_obj, &count);
1088                         if (err == 0 &&
1089                             count == ds->ds_phys->ds_num_children - 1)
1090                                 usenext = B_TRUE;
1091                 }
1092 
1093                 if (usenext) {
1094                         VERIFY(zap_join_key(dp->dp_meta_objset,
1095                             ds->ds_phys->ds_next_clones_obj,
1096                             scn->scn_phys.scn_queue_obj,
1097                             ds->ds_phys->ds_creation_txg, tx) == 0);
1098                 } else {
1099                         struct enqueue_clones_arg eca;
1100                         eca.tx = tx;
1101                         eca.originobj = ds->ds_object;
1102 
1103                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
1104                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
1105                 }
1106         }
1107 
1108 out:
1109         dsl_dataset_rele(ds, FTAG);
1110 }
1111 
1112 /* ARGSUSED */
1113 static int
1114 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1115 {
1116         dmu_tx_t *tx = arg;
1117         dsl_dataset_t *ds;
1118         int err;
1119         dsl_pool_t *dp = spa->spa_dsl_pool;
1120         dsl_scan_t *scn = dp->dp_scan;
1121 
1122         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1123         if (err)
1124                 return (err);
1125 
1126         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1127                 dsl_dataset_t *prev;
1128                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1129                     FTAG, &prev);
1130                 if (err) {
1131                         dsl_dataset_rele(ds, FTAG);
1132                         return (err);
1133                 }
1134 
1135                 /*
1136                  * If this is a clone, we don't need to worry about it for now.
1137                  */
1138                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1139                         dsl_dataset_rele(ds, FTAG);
1140                         dsl_dataset_rele(prev, FTAG);
1141                         return (0);
1142                 }
1143                 dsl_dataset_rele(ds, FTAG);
1144                 ds = prev;
1145         }
1146 
1147         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1148             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1149         dsl_dataset_rele(ds, FTAG);
1150         return (0);
1151 }
1152 
1153 /*
1154  * Scrub/dedup interaction.
1155  *
1156  * If there are N references to a deduped block, we don't want to scrub it
1157  * N times -- ideally, we should scrub it exactly once.
1158  *
1159  * We leverage the fact that the dde's replication class (enum ddt_class)
1160  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1161  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1162  *
1163  * To prevent excess scrubbing, the scrub begins by walking the DDT
1164  * to find all blocks with refcnt > 1, and scrubs each of these once.
1165  * Since there are two replication classes which contain blocks with
1166  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1167  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1168  *
1169  * There would be nothing more to say if a block's refcnt couldn't change
1170  * during a scrub, but of course it can so we must account for changes
1171  * in a block's replication class.
1172  *
1173  * Here's an example of what can occur:
1174  *
1175  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1176  * when visited during the top-down scrub phase, it will be scrubbed twice.
1177  * This negates our scrub optimization, but is otherwise harmless.
1178  *
1179  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1180  * on each visit during the top-down scrub phase, it will never be scrubbed.
1181  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1182  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1183  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1184  * while a scrub is in progress, it scrubs the block right then.
1185  */
1186 static void
1187 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1188 {
1189         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1190         ddt_entry_t dde = { 0 };
1191         int error;
1192         uint64_t n = 0;
1193 
1194         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1195                 ddt_t *ddt;
1196 
1197                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1198                         break;
1199                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1200                     (longlong_t)ddb->ddb_class,
1201                     (longlong_t)ddb->ddb_type,
1202                     (longlong_t)ddb->ddb_checksum,
1203                     (longlong_t)ddb->ddb_cursor);
1204 
1205                 /* There should be no pending changes to the dedup table */
1206                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1207                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1208 
1209                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1210                 n++;
1211 
1212                 if (dsl_scan_check_pause(scn, NULL))
1213                         break;
1214         }
1215 
1216         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1217             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1218             (int)scn->scn_pausing);
1219 
1220         ASSERT(error == 0 || error == ENOENT);
1221         ASSERT(error != ENOENT ||
1222             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1223 }
1224 
1225 /* ARGSUSED */
1226 void
1227 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1228     ddt_entry_t *dde, dmu_tx_t *tx)
1229 {
1230         const ddt_key_t *ddk = &dde->dde_key;
1231         ddt_phys_t *ddp = dde->dde_phys;
1232         blkptr_t bp;
1233         zbookmark_t zb = { 0 };
1234 
1235         if (scn->scn_phys.scn_state != DSS_SCANNING)
1236                 return;
1237 
1238         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1239                 if (ddp->ddp_phys_birth == 0 ||
1240                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1241                         continue;
1242                 ddt_bp_create(checksum, ddk, ddp, &bp);
1243 
1244                 scn->scn_visited_this_txg++;
1245                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1246         }
1247 }
1248 
1249 static void
1250 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1251 {
1252         dsl_pool_t *dp = scn->scn_dp;
1253         zap_cursor_t zc;
1254         zap_attribute_t za;
1255 
1256         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1257             scn->scn_phys.scn_ddt_class_max) {
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_ddt(scn, tx);
1261                 if (scn->scn_pausing)
1262                         return;
1263         }
1264 
1265         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1266                 /* First do the MOS & ORIGIN */
1267 
1268                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1269                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1270                 dsl_scan_visit_rootbp(scn, NULL,
1271                     &dp->dp_meta_rootbp, tx);
1272                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1273                 if (scn->scn_pausing)
1274                         return;
1275 
1276                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1277                         VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
1278                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
1279                 } else {
1280                         dsl_scan_visitds(scn,
1281                             dp->dp_origin_snap->ds_object, tx);
1282                 }
1283                 ASSERT(!scn->scn_pausing);
1284         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1285             ZB_DESTROYED_OBJSET) {
1286                 /*
1287                  * If we were paused, continue from here.  Note if the
1288                  * ds we were paused on was deleted, the zb_objset may
1289                  * be -1, so we will skip this and find a new objset
1290                  * below.
1291                  */
1292                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1293                 if (scn->scn_pausing)
1294                         return;
1295         }
1296 
1297         /*
1298          * In case we were paused right at the end of the ds, zero the
1299          * bookmark so we don't think that we're still trying to resume.
1300          */
1301         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1302 
1303         /* keep pulling things out of the zap-object-as-queue */
1304         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1305             scn->scn_phys.scn_queue_obj),
1306             zap_cursor_retrieve(&zc, &za) == 0) {
1307                 dsl_dataset_t *ds;
1308                 uint64_t dsobj;
1309 
1310                 dsobj = strtonum(za.za_name, NULL);
1311                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1312                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1313 
1314                 /* Set up min/max txg */
1315                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1316                 if (za.za_first_integer != 0) {
1317                         scn->scn_phys.scn_cur_min_txg =
1318                             MAX(scn->scn_phys.scn_min_txg,
1319                             za.za_first_integer);
1320                 } else {
1321                         scn->scn_phys.scn_cur_min_txg =
1322                             MAX(scn->scn_phys.scn_min_txg,
1323                             ds->ds_phys->ds_prev_snap_txg);
1324                 }
1325                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1326                 dsl_dataset_rele(ds, FTAG);
1327 
1328                 dsl_scan_visitds(scn, dsobj, tx);
1329                 zap_cursor_fini(&zc);
1330                 if (scn->scn_pausing)
1331                         return;
1332         }
1333         zap_cursor_fini(&zc);
1334 }
1335 
1336 static boolean_t
1337 dsl_scan_free_should_pause(dsl_scan_t *scn)
1338 {
1339         uint64_t elapsed_nanosecs;
1340 
1341         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1342         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1343             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1344             txg_sync_waiting(scn->scn_dp)) ||
1345             spa_shutting_down(scn->scn_dp->dp_spa));
1346 }
1347 
1348 static int
1349 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1350 {
1351         dsl_scan_t *scn = arg;
1352 
1353         if (!scn->scn_is_bptree ||
1354             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1355                 if (dsl_scan_free_should_pause(scn))
1356                         return (ERESTART);
1357         }
1358 
1359         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1360             dmu_tx_get_txg(tx), bp, 0));
1361         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1362             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1363             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1364         scn->scn_visited_this_txg++;
1365         return (0);
1366 }
1367 
1368 boolean_t
1369 dsl_scan_active(dsl_scan_t *scn)
1370 {
1371         spa_t *spa = scn->scn_dp->dp_spa;
1372         uint64_t used = 0, comp, uncomp;
1373 
1374         if (spa->spa_load_state != SPA_LOAD_NONE)
1375                 return (B_FALSE);
1376         if (spa_shutting_down(spa))
1377                 return (B_FALSE);
1378 
1379         if (scn->scn_phys.scn_state == DSS_SCANNING)
1380                 return (B_TRUE);
1381 
1382         if (spa_feature_is_active(spa,
1383             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1384                 return (B_TRUE);
1385         }
1386         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1387                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1388                     &used, &comp, &uncomp);
1389         }
1390         return (used != 0);
1391 }
1392 
1393 void
1394 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1395 {
1396         dsl_scan_t *scn = dp->dp_scan;
1397         spa_t *spa = dp->dp_spa;
1398         int err;
1399 
1400         /*
1401          * Check for scn_restart_txg before checking spa_load_state, so
1402          * that we can restart an old-style scan while the pool is being
1403          * imported (see dsl_scan_init).
1404          */
1405         if (scn->scn_restart_txg != 0 &&
1406             scn->scn_restart_txg <= tx->tx_txg) {
1407                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1408                 dsl_scan_done(scn, B_FALSE, tx);
1409                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1410                         func = POOL_SCAN_RESILVER;
1411                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1412                     func, tx->tx_txg);
1413                 dsl_scan_setup_sync(scn, &func, tx);
1414         }
1415 
1416         if (!dsl_scan_active(scn) ||
1417             spa_sync_pass(dp->dp_spa) > 1)
1418                 return;
1419 
1420         scn->scn_visited_this_txg = 0;
1421         scn->scn_pausing = B_FALSE;
1422         scn->scn_sync_start_time = gethrtime();
1423         spa->spa_scrub_active = B_TRUE;
1424 
1425         /*
1426          * First process the free list.  If we pause the free, don't do
1427          * any scanning.  This ensures that there is no free list when
1428          * we are scanning, so the scan code doesn't have to worry about
1429          * traversing it.
1430          */
1431         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1432                 scn->scn_is_bptree = B_FALSE;
1433                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1434                     NULL, ZIO_FLAG_MUSTSUCCEED);
1435                 err = bpobj_iterate(&dp->dp_free_bpobj,
1436                     dsl_scan_free_block_cb, scn, tx);
1437                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1438 
1439                 if (err == 0 && spa_feature_is_active(spa,
1440                     &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1441                         scn->scn_is_bptree = B_TRUE;
1442                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1443                             NULL, ZIO_FLAG_MUSTSUCCEED);
1444                         err = bptree_iterate(dp->dp_meta_objset,
1445                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1446                             scn, tx);
1447                         VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1448                         if (err != 0)
1449                                 return;
1450 
1451                         /* disable async destroy feature */
1452                         spa_feature_decr(spa,
1453                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY], tx);
1454                         ASSERT(!spa_feature_is_active(spa,
1455                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]));
1456                         VERIFY3U(0, ==, zap_remove(dp->dp_meta_objset,
1457                             DMU_POOL_DIRECTORY_OBJECT,
1458                             DMU_POOL_BPTREE_OBJ, tx));
1459                         VERIFY3U(0, ==, bptree_free(dp->dp_meta_objset,
1460                             dp->dp_bptree_obj, tx));
1461                         dp->dp_bptree_obj = 0;
1462                 }
1463                 if (scn->scn_visited_this_txg) {
1464                         zfs_dbgmsg("freed %llu blocks in %llums from "
1465                             "free_bpobj/bptree txg %llu",
1466                             (longlong_t)scn->scn_visited_this_txg,
1467                             (longlong_t)
1468                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1469                             (longlong_t)tx->tx_txg);
1470                         scn->scn_visited_this_txg = 0;
1471                         /*
1472                          * Re-sync the ddt so that we can further modify
1473                          * it when doing bprewrite.
1474                          */
1475                         ddt_sync(spa, tx->tx_txg);
1476                 }
1477                 if (err == ERESTART)
1478                         return;
1479         }
1480 
1481         if (scn->scn_phys.scn_state != DSS_SCANNING)
1482                 return;
1483 
1484         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1485             scn->scn_phys.scn_ddt_class_max) {
1486                 zfs_dbgmsg("doing scan sync txg %llu; "
1487                     "ddt bm=%llu/%llu/%llu/%llx",
1488                     (longlong_t)tx->tx_txg,
1489                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1490                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1491                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1492                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1493                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1494                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1495                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1496                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1497         } else {
1498                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1499                     (longlong_t)tx->tx_txg,
1500                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1501                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1502                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1503                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1504         }
1505 
1506         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1507             NULL, ZIO_FLAG_CANFAIL);
1508         dsl_scan_visit(scn, tx);
1509         (void) zio_wait(scn->scn_zio_root);
1510         scn->scn_zio_root = NULL;
1511 
1512         zfs_dbgmsg("visited %llu blocks in %llums",
1513             (longlong_t)scn->scn_visited_this_txg,
1514             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1515 
1516         if (!scn->scn_pausing) {
1517                 /* finished with scan. */
1518                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1519                 dsl_scan_done(scn, B_TRUE, tx);
1520         }
1521 
1522         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1523                 mutex_enter(&spa->spa_scrub_lock);
1524                 while (spa->spa_scrub_inflight > 0) {
1525                         cv_wait(&spa->spa_scrub_io_cv,
1526                             &spa->spa_scrub_lock);
1527                 }
1528                 mutex_exit(&spa->spa_scrub_lock);
1529         }
1530 
1531         dsl_scan_sync_state(scn, tx);
1532 }
1533 
1534 /*
1535  * This will start a new scan, or restart an existing one.
1536  */
1537 void
1538 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1539 {
1540         if (txg == 0) {
1541                 dmu_tx_t *tx;
1542                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1543                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1544 
1545                 txg = dmu_tx_get_txg(tx);
1546                 dp->dp_scan->scn_restart_txg = txg;
1547                 dmu_tx_commit(tx);
1548         } else {
1549                 dp->dp_scan->scn_restart_txg = txg;
1550         }
1551         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1552 }
1553 
1554 boolean_t
1555 dsl_scan_resilvering(dsl_pool_t *dp)
1556 {
1557         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1558             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1559 }
1560 
1561 /*
1562  * scrub consumers
1563  */
1564 
1565 static void
1566 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1567 {
1568         int i;
1569 
1570         /*
1571          * If we resume after a reboot, zab will be NULL; don't record
1572          * incomplete stats in that case.
1573          */
1574         if (zab == NULL)
1575                 return;
1576 
1577         for (i = 0; i < 4; i++) {
1578                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1579                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1580                 if (t & DMU_OT_NEWTYPE)
1581                         t = DMU_OT_OTHER;
1582                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1583                 int equal;
1584 
1585                 zb->zb_count++;
1586                 zb->zb_asize += BP_GET_ASIZE(bp);
1587                 zb->zb_lsize += BP_GET_LSIZE(bp);
1588                 zb->zb_psize += BP_GET_PSIZE(bp);
1589                 zb->zb_gangs += BP_COUNT_GANG(bp);
1590 
1591                 switch (BP_GET_NDVAS(bp)) {
1592                 case 2:
1593                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1594                             DVA_GET_VDEV(&bp->blk_dva[1]))
1595                                 zb->zb_ditto_2_of_2_samevdev++;
1596                         break;
1597                 case 3:
1598                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1599                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1600                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1601                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1602                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1603                             DVA_GET_VDEV(&bp->blk_dva[2]));
1604                         if (equal == 1)
1605                                 zb->zb_ditto_2_of_3_samevdev++;
1606                         else if (equal == 3)
1607                                 zb->zb_ditto_3_of_3_samevdev++;
1608                         break;
1609                 }
1610         }
1611 }
1612 
1613 static void
1614 dsl_scan_scrub_done(zio_t *zio)
1615 {
1616         spa_t *spa = zio->io_spa;
1617 
1618         zio_data_buf_free(zio->io_data, zio->io_size);
1619 
1620         mutex_enter(&spa->spa_scrub_lock);
1621         spa->spa_scrub_inflight--;
1622         cv_broadcast(&spa->spa_scrub_io_cv);
1623 
1624         if (zio->io_error && (zio->io_error != ECKSUM ||
1625             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1626                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1627         }
1628         mutex_exit(&spa->spa_scrub_lock);
1629 }
1630 
1631 static int
1632 dsl_scan_scrub_cb(dsl_pool_t *dp,
1633     const blkptr_t *bp, const zbookmark_t *zb)
1634 {
1635         dsl_scan_t *scn = dp->dp_scan;
1636         size_t size = BP_GET_PSIZE(bp);
1637         spa_t *spa = dp->dp_spa;
1638         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1639         boolean_t needs_io;
1640         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1641         int zio_priority;
1642         int scan_delay = 0;
1643 
1644         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1645             phys_birth >= scn->scn_phys.scn_max_txg)
1646                 return (0);
1647 
1648         count_block(dp->dp_blkstats, bp);
1649 
1650         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1651         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1652                 zio_flags |= ZIO_FLAG_SCRUB;
1653                 zio_priority = ZIO_PRIORITY_SCRUB;
1654                 needs_io = B_TRUE;
1655                 scan_delay = zfs_scrub_delay;
1656         } else if (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 (int 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_do(dp, dsl_scan_setup_check,
1743             dsl_scan_setup_sync, dp->dp_scan, &func, 0));
1744 }