sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_scan.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2012 by Delphix. All rights reserved.
  24  */
  25
  26 #include <sys/dsl_scan.h>
  27 #include <sys/dsl_pool.h>
  28 #include <sys/dsl_dataset.h>
  29 #include <sys/dsl_prop.h>
  30 #include <sys/dsl_dir.h>
  31 #include <sys/dsl_synctask.h>
  32 #include <sys/dnode.h>
  33 #include <sys/dmu_tx.h>
  34 #include <sys/dmu_objset.h>
  35 #include <sys/arc.h>
  36 #include <sys/zap.h>
  37 #include <sys/zio.h>
  38 #include <sys/zfs_context.h>
  39 #include <sys/fs/zfs.h>
  40 #include <sys/zfs_znode.h>
  41 #include <sys/spa_impl.h>
  42 #include <sys/vdev_impl.h>
  43 #include <sys/zil_impl.h>
  44 #include <sys/zio_checksum.h>
  45 #include <sys/ddt.h>
  46 #include <sys/sa.h>
  47 #include <sys/sa_impl.h>
  48 #include <sys/zfeature.h>
  49 #ifdef _KERNEL
  50 #include <sys/zfs_vfsops.h>
  51 #endif
  52
  53 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
  54
  55 static scan_cb_t dsl_scan_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
  73 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
  74         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
  75         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
  76
  77 extern int zfs_txg_timeout;
  78
  79 /* the order has to match pool_scan_type */
  80 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
  81         NULL,
  82         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
  83         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
  84 };
  85
  86 int
  87 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
  88 {
  89         int err;
  90         dsl_scan_t *scn;
  91         spa_t *spa = dp->dp_spa;
  92         uint64_t f;
  93
  94         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
  95         scn->scn_dp = dp;
  96
  97         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
  98             "scrub_func", sizeof (uint64_t), 1, &f);
  99         if (err == 0) {
 100                 /*
 101                  * There was an old-style scrub in progress.  Restart a
 102                  * new-style scrub from the beginning.
 103                  */
 104                 scn->scn_restart_txg = txg;
 105                 zfs_dbgmsg("old-style scrub was in progress; "
 106                     "restarting new-style scrub in txg %llu",
 107                     scn->scn_restart_txg);
 108
 109                 /*
 110                  * Load the queue obj from the old location so that it
 111                  * can be freed by dsl_scan_done().
 112                  */
 113                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 114                     "scrub_queue", sizeof (uint64_t), 1,
 115                     &scn->scn_phys.scn_queue_obj);
 116         } else {
 117                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 118                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 119                     &scn->scn_phys);
 120                 if (err == ENOENT)
 121                         return (0);
 122                 else if (err)
 123                         return (err);
 124
 125                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 126                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 127                         /*
 128                          * A new-type scrub was in progress on an old
 129                          * pool, and the pool was accessed by old
 130                          * software.  Restart from the beginning, since
 131                          * the old software may have changed the pool in
 132                          * the meantime.
 133                          */
 134                         scn->scn_restart_txg = txg;
 135                         zfs_dbgmsg("new-style scrub was modified "
 136                             "by old software; restarting in txg %llu",
 137                             scn->scn_restart_txg);
 138                 }
 139         }
 140
 141         spa_scan_stat_init(spa);
 142         return (0);
 143 }
 144
 145 void
 146 dsl_scan_fini(dsl_pool_t *dp)
 147 {
 148         if (dp->dp_scan) {
 149                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 150                 dp->dp_scan = NULL;
 151         }
 152 }
 153
 154 /* ARGSUSED */
 155 static int
 156 dsl_scan_setup_check(void *arg1, void *arg2, dmu_tx_t *tx)
 157 {
 158         dsl_scan_t *scn = arg1;
 159
 160         if (scn->scn_phys.scn_state == DSS_SCANNING)
 161                 return (EBUSY);
 162
 163         return (0);
 164 }
 165
 166 /* ARGSUSED */
 167 static void
 168 dsl_scan_setup_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 169 {
 170         dsl_scan_t *scn = arg1;
 171         pool_scan_func_t *funcp = arg2;
 172         dmu_object_type_t ot = 0;
 173         dsl_pool_t *dp = scn->scn_dp;
 174         spa_t *spa = dp->dp_spa;
 175
 176         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 177         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 178         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 179         scn->scn_phys.scn_func = *funcp;
 180         scn->scn_phys.scn_state = DSS_SCANNING;
 181         scn->scn_phys.scn_min_txg = 0;
 182         scn->scn_phys.scn_max_txg = tx->tx_txg;
 183         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 184         scn->scn_phys.scn_start_time = gethrestime_sec();
 185         scn->scn_phys.scn_errors = 0;
 186         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 187         scn->scn_restart_txg = 0;
 188         spa_scan_stat_init(spa);
 189
 190         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 191                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 192
 193                 /* rewrite all disk labels */
 194                 vdev_config_dirty(spa->spa_root_vdev);
 195
 196                 if (vdev_resilver_needed(spa->spa_root_vdev,
 197                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 198                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 199                 } else {
 200                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 201                 }
 202
 203                 spa->spa_scrub_started = B_TRUE;
 204                 /*
 205                  * If this is an incremental scrub, limit the DDT scrub phase
 206                  * to just the auto-ditto class (for correctness); the rest
 207                  * of the scrub should go faster using top-down pruning.
 208                  */
 209                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 210                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 211
 212         }
 213
 214         /* back to the generic stuff */
 215
 216         if (dp->dp_blkstats == NULL) {
 217                 dp->dp_blkstats =
 218                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 219         }
 220         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 221
 222         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 223                 ot = DMU_OT_ZAP_OTHER;
 224
 225         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 226             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 227
 228         dsl_scan_sync_state(scn, tx);
 229
 230         spa_history_log_internal(LOG_POOL_SCAN, spa, tx,
 231             "func=%u mintxg=%llu maxtxg=%llu",
 232             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 233 }
 234
 235 /* ARGSUSED */
 236 static void
 237 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 238 {
 239         static const char *old_names[] = {
 240                 "scrub_bookmark",
 241                 "scrub_ddt_bookmark",
 242                 "scrub_ddt_class_max",
 243                 "scrub_queue",
 244                 "scrub_min_txg",
 245                 "scrub_max_txg",
 246                 "scrub_func",
 247                 "scrub_errors",
 248                 NULL
 249         };
 250
 251         dsl_pool_t *dp = scn->scn_dp;
 252         spa_t *spa = dp->dp_spa;
 253         int i;
 254
 255         /* Remove any remnants of an old-style scrub. */
 256         for (i = 0; old_names[i]; i++) {
 257                 (void) zap_remove(dp->dp_meta_objset,
 258                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 259         }
 260
 261         if (scn->scn_phys.scn_queue_obj != 0) {
 262                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 263                     scn->scn_phys.scn_queue_obj, tx));
 264                 scn->scn_phys.scn_queue_obj = 0;
 265         }
 266
 267         /*
 268          * If we were "restarted" from a stopped state, don't bother
 269          * with anything else.
 270          */
 271         if (scn->scn_phys.scn_state != DSS_SCANNING)
 272                 return;
 273
 274         if (complete)
 275                 scn->scn_phys.scn_state = DSS_FINISHED;
 276         else
 277                 scn->scn_phys.scn_state = DSS_CANCELED;
 278
 279         spa_history_log_internal(LOG_POOL_SCAN_DONE, spa, tx,
 280             "complete=%u", complete);
 281
 282         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 283                 mutex_enter(&spa->spa_scrub_lock);
 284                 while (spa->spa_scrub_inflight > 0) {
 285                         cv_wait(&spa->spa_scrub_io_cv,
 286                             &spa->spa_scrub_lock);
 287                 }
 288                 mutex_exit(&spa->spa_scrub_lock);
 289                 spa->spa_scrub_started = B_FALSE;
 290                 spa->spa_scrub_active = B_FALSE;
 291
 292                 /*
 293                  * If the scrub/resilver completed, update all DTLs to
 294                  * reflect this.  Whether it succeeded or not, vacate
 295                  * all temporary scrub DTLs.
 296                  */
 297                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 298                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 299                 if (complete) {
 300                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 301                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 302                 }
 303                 spa_errlog_rotate(spa);
 304
 305                 /*
 306                  * We may have finished replacing a device.
 307                  * Let the async thread assess this and handle the detach.
 308                  */
 309                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 310         }
 311
 312         scn->scn_phys.scn_end_time = gethrestime_sec();
 313 }
 314
 315 /* ARGSUSED */
 316 static int
 317 dsl_scan_cancel_check(void *arg1, void *arg2, dmu_tx_t *tx)
 318 {
 319         dsl_scan_t *scn = arg1;
 320
 321         if (scn->scn_phys.scn_state != DSS_SCANNING)
 322                 return (ENOENT);
 323         return (0);
 324 }
 325
 326 /* ARGSUSED */
 327 static void
 328 dsl_scan_cancel_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 329 {
 330         dsl_scan_t *scn = arg1;
 331
 332         dsl_scan_done(scn, B_FALSE, tx);
 333         dsl_scan_sync_state(scn, tx);
 334 }
 335
 336 int
 337 dsl_scan_cancel(dsl_pool_t *dp)
 338 {
 339         boolean_t complete = B_FALSE;
 340         int err;
 341
 342         err = dsl_sync_task_do(dp, dsl_scan_cancel_check,
 343             dsl_scan_cancel_sync, dp->dp_scan, &complete, 3);
 344         return (err);
 345 }
 346
 347 static void dsl_scan_visitbp(blkptr_t *bp,
 348     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 349     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 350     dmu_tx_t *tx);
 351 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 352     dmu_objset_type_t ostype,
 353     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 354
 355 void
 356 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 357 {
 358         zio_free(dp->dp_spa, txg, bp);
 359 }
 360
 361 void
 362 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 363 {
 364         ASSERT(dsl_pool_sync_context(dp));
 365         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
 366 }
 367
 368 int
 369 dsl_read(zio_t *pio, spa_t *spa, const blkptr_t *bpp, arc_buf_t *pbuf,
 370     arc_done_func_t *done, void *private, int priority, int zio_flags,
 371     uint32_t *arc_flags, const zbookmark_t *zb)
 372 {
 373         return (arc_read(pio, spa, bpp, pbuf, done, private,
 374             priority, zio_flags, arc_flags, zb));
 375 }
 376
 377 int
 378 dsl_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bpp,
 379     arc_done_func_t *done, void *private, int priority, int zio_flags,
 380     uint32_t *arc_flags, const zbookmark_t *zb)
 381 {
 382         return (arc_read_nolock(pio, spa, bpp, done, private,
 383             priority, zio_flags, arc_flags, zb));
 384 }
 385
 386 static uint64_t
 387 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 388 {
 389         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 390         if (dsl_dataset_is_snapshot(ds))
 391                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 392         return (smt);
 393 }
 394
 395 static void
 396 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 397 {
 398         VERIFY(0 == zap_update(scn->scn_dp->dp_meta_objset,
 399             DMU_POOL_DIRECTORY_OBJECT,
 400             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 401             &scn->scn_phys, tx));
 402 }
 403
 404 static boolean_t
 405 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 406 {
 407         uint64_t elapsed_nanosecs;
 408         int mintime;
 409
 410         /* we never skip user/group accounting objects */
 411         if (zb && (int64_t)zb->zb_object < 0)
 412                 return (B_FALSE);
 413
 414         if (scn->scn_pausing)
 415                 return (B_TRUE); /* we're already pausing */
 416
 417         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 418                 return (B_FALSE); /* we're resuming */
 419
 420         /* We only know how to resume from level-0 blocks. */
 421         if (zb && zb->zb_level != 0)
 422                 return (B_FALSE);
 423
 424         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 425             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 426         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 427         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 428             (elapsed_nanosecs / MICROSEC > mintime &&
 429             txg_sync_waiting(scn->scn_dp)) ||
 430             spa_shutting_down(scn->scn_dp->dp_spa)) {
 431                 if (zb) {
 432                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 433                             (longlong_t)zb->zb_objset,
 434                             (longlong_t)zb->zb_object,
 435                             (longlong_t)zb->zb_level,
 436                             (longlong_t)zb->zb_blkid);
 437                         scn->scn_phys.scn_bookmark = *zb;
 438                 }
 439                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 440                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 441                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 442                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 443                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 444                 scn->scn_pausing = B_TRUE;
 445                 return (B_TRUE);
 446         }
 447         return (B_FALSE);
 448 }
 449
 450 typedef struct zil_scan_arg {
 451         dsl_pool_t      *zsa_dp;
 452         zil_header_t    *zsa_zh;
 453 } zil_scan_arg_t;
 454
 455 /* ARGSUSED */
 456 static int
 457 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 458 {
 459         zil_scan_arg_t *zsa = arg;
 460         dsl_pool_t *dp = zsa->zsa_dp;
 461         dsl_scan_t *scn = dp->dp_scan;
 462         zil_header_t *zh = zsa->zsa_zh;
 463         zbookmark_t zb;
 464
 465         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 466                 return (0);
 467
 468         /*
 469          * One block ("stubby") can be allocated a long time ago; we
 470          * want to visit that one because it has been allocated
 471          * (on-disk) even if it hasn't been claimed (even though for
 472          * scrub there's nothing to do to it).
 473          */
 474         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 475                 return (0);
 476
 477         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 478             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 479
 480         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 481         return (0);
 482 }
 483
 484 /* ARGSUSED */
 485 static int
 486 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 487 {
 488         if (lrc->lrc_txtype == TX_WRITE) {
 489                 zil_scan_arg_t *zsa = arg;
 490                 dsl_pool_t *dp = zsa->zsa_dp;
 491                 dsl_scan_t *scn = dp->dp_scan;
 492                 zil_header_t *zh = zsa->zsa_zh;
 493                 lr_write_t *lr = (lr_write_t *)lrc;
 494                 blkptr_t *bp = &lr->lr_blkptr;
 495                 zbookmark_t zb;
 496
 497                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 498                         return (0);
 499
 500                 /*
 501                  * birth can be < claim_txg if this record's txg is
 502                  * already txg sync'ed (but this log block contains
 503                  * other records that are not synced)
 504                  */
 505                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 506                         return (0);
 507
 508                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 509                     lr->lr_foid, ZB_ZIL_LEVEL,
 510                     lr->lr_offset / BP_GET_LSIZE(bp));
 511
 512                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 513         }
 514         return (0);
 515 }
 516
 517 static void
 518 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 519 {
 520         uint64_t claim_txg = zh->zh_claim_txg;
 521         zil_scan_arg_t zsa = { dp, zh };
 522         zilog_t *zilog;
 523
 524         /*
 525          * We only want to visit blocks that have been claimed but not yet
 526          * replayed (or, in read-only mode, blocks that *would* be claimed).
 527          */
 528         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 529                 return;
 530
 531         zilog = zil_alloc(dp->dp_meta_objset, zh);
 532
 533         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 534             claim_txg);
 535
 536         zil_free(zilog);
 537 }
 538
 539 /* ARGSUSED */
 540 static void
 541 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 542     uint64_t objset, uint64_t object, uint64_t blkid)
 543 {
 544         zbookmark_t czb;
 545         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 546
 547         if (zfs_no_scrub_prefetch)
 548                 return;
 549
 550         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 551             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 552                 return;
 553
 554         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 555
 556         /*
 557          * XXX need to make sure all of these arc_read() prefetches are
 558          * done before setting xlateall (similar to dsl_read())
 559          */
 560         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 561             buf, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 562             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 563 }
 564
 565 static boolean_t
 566 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 567     const zbookmark_t *zb)
 568 {
 569         /*
 570          * We never skip over user/group accounting objects (obj<0)
 571          */
 572         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 573             (int64_t)zb->zb_object >= 0) {
 574                 /*
 575                  * If we already visited this bp & everything below (in
 576                  * a prior txg sync), don't bother doing it again.
 577                  */
 578                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 579                         return (B_TRUE);
 580
 581                 /*
 582                  * If we found the block we're trying to resume from, or
 583                  * we went past it to a different object, zero it out to
 584                  * indicate that it's OK to start checking for pausing
 585                  * again.
 586                  */
 587                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 588                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 589                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 590                             (longlong_t)zb->zb_objset,
 591                             (longlong_t)zb->zb_object,
 592                             (longlong_t)zb->zb_level,
 593                             (longlong_t)zb->zb_blkid);
 594                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 595                 }
 596         }
 597         return (B_FALSE);
 598 }
 599
 600 /*
 601  * Return nonzero on i/o error.
 602  * Return new buf to write out in *bufp.
 603  */
 604 static int
 605 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 606     dnode_phys_t *dnp, const blkptr_t *bp,
 607     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 608 {
 609         dsl_pool_t *dp = scn->scn_dp;
 610         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 611         int err;
 612
 613         if (BP_GET_LEVEL(bp) > 0) {
 614                 uint32_t flags = ARC_WAIT;
 615                 int i;
 616                 blkptr_t *cbp;
 617                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 618
 619                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 620                     arc_getbuf_func, bufp,
 621                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 622                 if (err) {
 623                         scn->scn_phys.scn_errors++;
 624                         return (err);
 625                 }
 626                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 627                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 628                             zb->zb_object, zb->zb_blkid * epb + i);
 629                 }
 630                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 631                         zbookmark_t czb;
 632
 633                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 634                             zb->zb_level - 1,
 635                             zb->zb_blkid * epb + i);
 636                         dsl_scan_visitbp(cbp, &czb, dnp,
 637                             *bufp, ds, scn, ostype, tx);
 638                 }
 639         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 640                 uint32_t flags = ARC_WAIT;
 641
 642                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 643                     arc_getbuf_func, bufp,
 644                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 645                 if (err) {
 646                         scn->scn_phys.scn_errors++;
 647                         return (err);
 648                 }
 649         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 650                 uint32_t flags = ARC_WAIT;
 651                 dnode_phys_t *cdnp;
 652                 int i, j;
 653                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 654
 655                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 656                     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                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 663                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 664                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 665                                 dsl_scan_prefetch(scn, *bufp, cbp,
 666                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 667                         }
 668                 }
 669                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 670                         dsl_scan_visitdnode(scn, ds, ostype,
 671                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 672                 }
 673
 674         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 675                 uint32_t flags = ARC_WAIT;
 676                 objset_phys_t *osp;
 677
 678                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 679                     arc_getbuf_func, bufp,
 680                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 681                 if (err) {
 682                         scn->scn_phys.scn_errors++;
 683                         return (err);
 684                 }
 685
 686                 osp = (*bufp)->b_data;
 687
 688                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 689                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 690
 691                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 692                         /*
 693                          * We also always visit user/group accounting
 694                          * objects, and never skip them, even if we are
 695                          * pausing.  This is necessary so that the space
 696                          * deltas from this txg get integrated.
 697                          */
 698                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 699                             &osp->os_groupused_dnode, *bufp,
 700                             DMU_GROUPUSED_OBJECT, tx);
 701                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 702                             &osp->os_userused_dnode, *bufp,
 703                             DMU_USERUSED_OBJECT, tx);
 704                 }
 705         }
 706
 707         return (0);
 708 }
 709
 710 static void
 711 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 712     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 713     uint64_t object, dmu_tx_t *tx)
 714 {
 715         int j;
 716
 717         for (j = 0; j < dnp->dn_nblkptr; j++) {
 718                 zbookmark_t czb;
 719
 720                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 721                     dnp->dn_nlevels - 1, j);
 722                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 723                     &czb, dnp, buf, ds, scn, ostype, tx);
 724         }
 725
 726         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 727                 zbookmark_t czb;
 728                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 729                     0, DMU_SPILL_BLKID);
 730                 dsl_scan_visitbp(&dnp->dn_spill,
 731                     &czb, dnp, buf, ds, scn, ostype, tx);
 732         }
 733 }
 734
 735 /*
 736  * The arguments are in this order because mdb can only print the
 737  * first 5; we want them to be useful.
 738  */
 739 static void
 740 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 741     dnode_phys_t *dnp, arc_buf_t *pbuf,
 742     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 743     dmu_tx_t *tx)
 744 {
 745         dsl_pool_t *dp = scn->scn_dp;
 746         arc_buf_t *buf = NULL;
 747         blkptr_t bp_toread = *bp;
 748
 749         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 750
 751         if (dsl_scan_check_pause(scn, zb))
 752                 return;
 753
 754         if (dsl_scan_check_resume(scn, dnp, zb))
 755                 return;
 756
 757         if (bp->blk_birth == 0)
 758                 return;
 759
 760         scn->scn_visited_this_txg++;
 761
 762         dprintf_bp(bp,
 763             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 764             ds, ds ? ds->ds_object : 0,
 765             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 766             pbuf, bp);
 767
 768         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 769                 return;
 770
 771         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
 772             &buf) != 0)
 773                 return;
 774
 775         /*
 776          * If dsl_scan_ddt() has aready visited this block, it will have
 777          * already done any translations or scrubbing, so don't call the
 778          * callback again.
 779          */
 780         if (ddt_class_contains(dp->dp_spa,
 781             scn->scn_phys.scn_ddt_class_max, bp)) {
 782                 ASSERT(buf == NULL);
 783                 return;
 784         }
 785
 786         /*
 787          * If this block is from the future (after cur_max_txg), then we
 788          * are doing this on behalf of a deleted snapshot, and we will
 789          * revisit the future block on the next pass of this dataset.
 790          * Don't scan it now unless we need to because something
 791          * under it was modified.
 792          */
 793         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 794                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 795         }
 796         if (buf)
 797                 (void) arc_buf_remove_ref(buf, &buf);
 798 }
 799
 800 static void
 801 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 802     dmu_tx_t *tx)
 803 {
 804         zbookmark_t zb;
 805
 806         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 807             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 808         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 809             ds, scn, DMU_OST_NONE, tx);
 810
 811         dprintf_ds(ds, "finished scan%s", "");
 812 }
 813
 814 void
 815 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 816 {
 817         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 818         dsl_scan_t *scn = dp->dp_scan;
 819         uint64_t mintxg;
 820
 821         if (scn->scn_phys.scn_state != DSS_SCANNING)
 822                 return;
 823
 824         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 825                 if (dsl_dataset_is_snapshot(ds)) {
 826                         /* Note, scn_cur_{min,max}_txg stays the same. */
 827                         scn->scn_phys.scn_bookmark.zb_objset =
 828                             ds->ds_phys->ds_next_snap_obj;
 829                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 830                             "reset zb_objset to %llu",
 831                             (u_longlong_t)ds->ds_object,
 832                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 833                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 834                 } else {
 835                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 836                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 837                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 838                             "reset bookmark to -1,0,0,0",
 839                             (u_longlong_t)ds->ds_object);
 840                 }
 841         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 842             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 843                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 844                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 845                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 846                 if (dsl_dataset_is_snapshot(ds)) {
 847                         /*
 848                          * We keep the same mintxg; it could be >
 849                          * ds_creation_txg if the previous snapshot was
 850                          * deleted too.
 851                          */
 852                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 853                             scn->scn_phys.scn_queue_obj,
 854                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 855                         zfs_dbgmsg("destroying ds %llu; in queue; "
 856                             "replacing with %llu",
 857                             (u_longlong_t)ds->ds_object,
 858                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 859                 } else {
 860                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 861                             (u_longlong_t)ds->ds_object);
 862                 }
 863         } else {
 864                 zfs_dbgmsg("destroying ds %llu; ignoring",
 865                     (u_longlong_t)ds->ds_object);
 866         }
 867
 868         /*
 869          * dsl_scan_sync() should be called after this, and should sync
 870          * out our changed state, but just to be safe, do it here.
 871          */
 872         dsl_scan_sync_state(scn, tx);
 873 }
 874
 875 void
 876 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 877 {
 878         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 879         dsl_scan_t *scn = dp->dp_scan;
 880         uint64_t mintxg;
 881
 882         if (scn->scn_phys.scn_state != DSS_SCANNING)
 883                 return;
 884
 885         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 886
 887         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 888                 scn->scn_phys.scn_bookmark.zb_objset =
 889                     ds->ds_phys->ds_prev_snap_obj;
 890                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 891                     "reset zb_objset to %llu",
 892                     (u_longlong_t)ds->ds_object,
 893                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 894         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 895             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 896                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 897                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 898                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 899                     scn->scn_phys.scn_queue_obj,
 900                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 901                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 902                     "replacing with %llu",
 903                     (u_longlong_t)ds->ds_object,
 904                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 905         }
 906         dsl_scan_sync_state(scn, tx);
 907 }
 908
 909 void
 910 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 911 {
 912         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 913         dsl_scan_t *scn = dp->dp_scan;
 914         uint64_t mintxg;
 915
 916         if (scn->scn_phys.scn_state != DSS_SCANNING)
 917                 return;
 918
 919         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 920                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 921                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 922                     "reset zb_objset to %llu",
 923                     (u_longlong_t)ds1->ds_object,
 924                     (u_longlong_t)ds2->ds_object);
 925         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 926                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 927                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 928                     "reset zb_objset to %llu",
 929                     (u_longlong_t)ds2->ds_object,
 930                     (u_longlong_t)ds1->ds_object);
 931         }
 932
 933         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 934             ds1->ds_object, &mintxg) == 0) {
 935                 int err;
 936
 937                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 938                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 939                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 940                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 941                 err = zap_add_int_key(dp->dp_meta_objset,
 942                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 943                 VERIFY(err == 0 || err == EEXIST);
 944                 if (err == EEXIST) {
 945                         /* Both were there to begin with */
 946                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 947                             scn->scn_phys.scn_queue_obj,
 948                             ds1->ds_object, mintxg, tx));
 949                 }
 950                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 951                     "replacing with %llu",
 952                     (u_longlong_t)ds1->ds_object,
 953                     (u_longlong_t)ds2->ds_object);
 954         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 955             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 956                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 957                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 958                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 959                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 960                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 961                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 962                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 963                     "replacing with %llu",
 964                     (u_longlong_t)ds2->ds_object,
 965                     (u_longlong_t)ds1->ds_object);
 966         }
 967
 968         dsl_scan_sync_state(scn, tx);
 969 }
 970
 971 struct enqueue_clones_arg {
 972         dmu_tx_t *tx;
 973         uint64_t originobj;
 974 };
 975
 976 /* ARGSUSED */
 977 static int
 978 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 979 {
 980         struct enqueue_clones_arg *eca = arg;
 981         dsl_dataset_t *ds;
 982         int err;
 983         dsl_pool_t *dp = spa->spa_dsl_pool;
 984         dsl_scan_t *scn = dp->dp_scan;
 985
 986         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
 987         if (err)
 988                 return (err);
 989
 990         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
 991                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 992                         dsl_dataset_t *prev;
 993                         err = dsl_dataset_hold_obj(dp,
 994                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 995
 996                         dsl_dataset_rele(ds, FTAG);
 997                         if (err)
 998                                 return (err);
 999                         ds = prev;
1000                 }
1001                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1002                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1003                     ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1004         }
1005         dsl_dataset_rele(ds, FTAG);
1006         return (0);
1007 }
1008
1009 static void
1010 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1011 {
1012         dsl_pool_t *dp = scn->scn_dp;
1013         dsl_dataset_t *ds;
1014         objset_t *os;
1015
1016         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1017
1018         if (dmu_objset_from_ds(ds, &os))
1019                 goto out;
1020
1021         /*
1022          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1023          * snapshots can have ZIL block pointers (which may be the same
1024          * BP as in the head), they must be ignored.  So we traverse the
1025          * ZIL here, rather than in scan_recurse(), because the regular
1026          * snapshot block-sharing rules don't apply to it.
1027          */
1028         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1029                 dsl_scan_zil(dp, &os->os_zil_header);
1030
1031         /*
1032          * Iterate over the bps in this ds.
1033          */
1034         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1035         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1036
1037         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1038         dsl_dataset_name(ds, dsname);
1039         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1040             "pausing=%u",
1041             (longlong_t)dsobj, dsname,
1042             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1043             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1044             (int)scn->scn_pausing);
1045         kmem_free(dsname, ZFS_MAXNAMELEN);
1046
1047         if (scn->scn_pausing)
1048                 goto out;
1049
1050         /*
1051          * We've finished this pass over this dataset.
1052          */
1053
1054         /*
1055          * If we did not completely visit this dataset, do another pass.
1056          */
1057         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1058                 zfs_dbgmsg("incomplete pass; visiting again");
1059                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1060                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1061                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1062                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1063                 goto out;
1064         }
1065
1066         /*
1067          * Add descendent datasets to work queue.
1068          */
1069         if (ds->ds_phys->ds_next_snap_obj != 0) {
1070                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1071                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1072                     ds->ds_phys->ds_creation_txg, tx) == 0);
1073         }
1074         if (ds->ds_phys->ds_num_children > 1) {
1075                 boolean_t usenext = B_FALSE;
1076                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1077                         uint64_t count;
1078                         /*
1079                          * A bug in a previous version of the code could
1080                          * cause upgrade_clones_cb() to not set
1081                          * ds_next_snap_obj when it should, leading to a
1082                          * missing entry.  Therefore we can only use the
1083                          * next_clones_obj when its count is correct.
1084                          */
1085                         int err = zap_count(dp->dp_meta_objset,
1086                             ds->ds_phys->ds_next_clones_obj, &count);
1087                         if (err == 0 &&
1088                             count == ds->ds_phys->ds_num_children - 1)
1089                                 usenext = B_TRUE;
1090                 }
1091
1092                 if (usenext) {
1093                         VERIFY(zap_join_key(dp->dp_meta_objset,
1094                             ds->ds_phys->ds_next_clones_obj,
1095                             scn->scn_phys.scn_queue_obj,
1096                             ds->ds_phys->ds_creation_txg, tx) == 0);
1097                 } else {
1098                         struct enqueue_clones_arg eca;
1099                         eca.tx = tx;
1100                         eca.originobj = ds->ds_object;
1101
1102                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
1103                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
1104                 }
1105         }
1106
1107 out:
1108         dsl_dataset_rele(ds, FTAG);
1109 }
1110
1111 /* ARGSUSED */
1112 static int
1113 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1114 {
1115         dmu_tx_t *tx = arg;
1116         dsl_dataset_t *ds;
1117         int err;
1118         dsl_pool_t *dp = spa->spa_dsl_pool;
1119         dsl_scan_t *scn = dp->dp_scan;
1120
1121         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1122         if (err)
1123                 return (err);
1124
1125         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1126                 dsl_dataset_t *prev;
1127                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1128                     FTAG, &prev);
1129                 if (err) {
1130                         dsl_dataset_rele(ds, FTAG);
1131                         return (err);
1132                 }
1133
1134                 /*
1135                  * If this is a clone, we don't need to worry about it for now.
1136                  */
1137                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1138                         dsl_dataset_rele(ds, FTAG);
1139                         dsl_dataset_rele(prev, FTAG);
1140                         return (0);
1141                 }
1142                 dsl_dataset_rele(ds, FTAG);
1143                 ds = prev;
1144         }
1145
1146         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1147             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1148         dsl_dataset_rele(ds, FTAG);
1149         return (0);
1150 }
1151
1152 /*
1153  * Scrub/dedup interaction.
1154  *
1155  * If there are N references to a deduped block, we don't want to scrub it
1156  * N times -- ideally, we should scrub it exactly once.
1157  *
1158  * We leverage the fact that the dde's replication class (enum ddt_class)
1159  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1160  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1161  *
1162  * To prevent excess scrubbing, the scrub begins by walking the DDT
1163  * to find all blocks with refcnt > 1, and scrubs each of these once.
1164  * Since there are two replication classes which contain blocks with
1165  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1166  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1167  *
1168  * There would be nothing more to say if a block's refcnt couldn't change
1169  * during a scrub, but of course it can so we must account for changes
1170  * in a block's replication class.
1171  *
1172  * Here's an example of what can occur:
1173  *
1174  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1175  * when visited during the top-down scrub phase, it will be scrubbed twice.
1176  * This negates our scrub optimization, but is otherwise harmless.
1177  *
1178  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1179  * on each visit during the top-down scrub phase, it will never be scrubbed.
1180  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1181  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1182  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1183  * while a scrub is in progress, it scrubs the block right then.
1184  */
1185 static void
1186 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1187 {
1188         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1189         ddt_entry_t dde = { 0 };
1190         int error;
1191         uint64_t n = 0;
1192
1193         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1194                 ddt_t *ddt;
1195
1196                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1197                         break;
1198                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1199                     (longlong_t)ddb->ddb_class,
1200                     (longlong_t)ddb->ddb_type,
1201                     (longlong_t)ddb->ddb_checksum,
1202                     (longlong_t)ddb->ddb_cursor);
1203
1204                 /* There should be no pending changes to the dedup table */
1205                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1206                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1207
1208                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1209                 n++;
1210
1211                 if (dsl_scan_check_pause(scn, NULL))
1212                         break;
1213         }
1214
1215         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1216             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1217             (int)scn->scn_pausing);
1218
1219         ASSERT(error == 0 || error == ENOENT);
1220         ASSERT(error != ENOENT ||
1221             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1222 }
1223
1224 /* ARGSUSED */
1225 void
1226 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1227     ddt_entry_t *dde, dmu_tx_t *tx)
1228 {
1229         const ddt_key_t *ddk = &dde->dde_key;
1230         ddt_phys_t *ddp = dde->dde_phys;
1231         blkptr_t bp;
1232         zbookmark_t zb = { 0 };
1233
1234         if (scn->scn_phys.scn_state != DSS_SCANNING)
1235                 return;
1236
1237         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1238                 if (ddp->ddp_phys_birth == 0 ||
1239                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1240                         continue;
1241                 ddt_bp_create(checksum, ddk, ddp, &bp);
1242
1243                 scn->scn_visited_this_txg++;
1244                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1245         }
1246 }
1247
1248 static void
1249 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1250 {
1251         dsl_pool_t *dp = scn->scn_dp;
1252         zap_cursor_t zc;
1253         zap_attribute_t za;
1254
1255         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1256             scn->scn_phys.scn_ddt_class_max) {
1257                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1258                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1259                 dsl_scan_ddt(scn, tx);
1260                 if (scn->scn_pausing)
1261                         return;
1262         }
1263
1264         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1265                 /* First do the MOS & ORIGIN */
1266
1267                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1268                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1269                 dsl_scan_visit_rootbp(scn, NULL,
1270                     &dp->dp_meta_rootbp, tx);
1271                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1272                 if (scn->scn_pausing)
1273                         return;
1274
1275                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1276                         VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
1277                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
1278                 } else {
1279                         dsl_scan_visitds(scn,
1280                             dp->dp_origin_snap->ds_object, tx);
1281                 }
1282                 ASSERT(!scn->scn_pausing);
1283         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1284             ZB_DESTROYED_OBJSET) {
1285                 /*
1286                  * If we were paused, continue from here.  Note if the
1287                  * ds we were paused on was deleted, the zb_objset may
1288                  * be -1, so we will skip this and find a new objset
1289                  * below.
1290                  */
1291                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1292                 if (scn->scn_pausing)
1293                         return;
1294         }
1295
1296         /*
1297          * In case we were paused right at the end of the ds, zero the
1298          * bookmark so we don't think that we're still trying to resume.
1299          */
1300         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1301
1302         /* keep pulling things out of the zap-object-as-queue */
1303         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1304             scn->scn_phys.scn_queue_obj),
1305             zap_cursor_retrieve(&zc, &za) == 0) {
1306                 dsl_dataset_t *ds;
1307                 uint64_t dsobj;
1308
1309                 dsobj = strtonum(za.za_name, NULL);
1310                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1311                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1312
1313                 /* Set up min/max txg */
1314                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1315                 if (za.za_first_integer != 0) {
1316                         scn->scn_phys.scn_cur_min_txg =
1317                             MAX(scn->scn_phys.scn_min_txg,
1318                             za.za_first_integer);
1319                 } else {
1320                         scn->scn_phys.scn_cur_min_txg =
1321                             MAX(scn->scn_phys.scn_min_txg,
1322                             ds->ds_phys->ds_prev_snap_txg);
1323                 }
1324                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1325                 dsl_dataset_rele(ds, FTAG);
1326
1327                 dsl_scan_visitds(scn, dsobj, tx);
1328                 zap_cursor_fini(&zc);
1329                 if (scn->scn_pausing)
1330                         return;
1331         }
1332         zap_cursor_fini(&zc);
1333 }
1334
1335 static boolean_t
1336 dsl_scan_free_should_pause(dsl_scan_t *scn)
1337 {
1338         uint64_t elapsed_nanosecs;
1339
1340         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1341         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1342             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1343             txg_sync_waiting(scn->scn_dp)) ||
1344             spa_shutting_down(scn->scn_dp->dp_spa));
1345 }
1346
1347 static int
1348 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1349 {
1350         dsl_scan_t *scn = arg;
1351
1352         if (!scn->scn_is_bptree ||
1353             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1354                 if (dsl_scan_free_should_pause(scn))
1355                         return (ERESTART);
1356         }
1357
1358         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1359             dmu_tx_get_txg(tx), bp, 0));
1360         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1361             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1362             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1363         scn->scn_visited_this_txg++;
1364         return (0);
1365 }
1366
1367 boolean_t
1368 dsl_scan_active(dsl_scan_t *scn)
1369 {
1370         spa_t *spa = scn->scn_dp->dp_spa;
1371         uint64_t used = 0, comp, uncomp;
1372
1373         if (spa->spa_load_state != SPA_LOAD_NONE)
1374                 return (B_FALSE);
1375         if (spa_shutting_down(spa))
1376                 return (B_FALSE);
1377
1378         if (scn->scn_phys.scn_state == DSS_SCANNING)
1379                 return (B_TRUE);
1380
1381         if (spa_feature_is_active(spa,
1382             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1383                 return (B_TRUE);
1384         }
1385         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1386                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1387                     &used, &comp, &uncomp);
1388         }
1389         return (used != 0);
1390 }
1391
1392 void
1393 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1394 {
1395         dsl_scan_t *scn = dp->dp_scan;
1396         spa_t *spa = dp->dp_spa;
1397         int err;
1398
1399         /*
1400          * Check for scn_restart_txg before checking spa_load_state, so
1401          * that we can restart an old-style scan while the pool is being
1402          * imported (see dsl_scan_init).
1403          */
1404         if (scn->scn_restart_txg != 0 &&
1405             scn->scn_restart_txg <= tx->tx_txg) {
1406                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1407                 dsl_scan_done(scn, B_FALSE, tx);
1408                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1409                         func = POOL_SCAN_RESILVER;
1410                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1411                     func, tx->tx_txg);
1412                 dsl_scan_setup_sync(scn, &func, tx);
1413         }
1414
1415         if (!dsl_scan_active(scn) ||
1416             spa_sync_pass(dp->dp_spa) > 1)
1417                 return;
1418
1419         scn->scn_visited_this_txg = 0;
1420         scn->scn_pausing = B_FALSE;
1421         scn->scn_sync_start_time = gethrtime();
1422         spa->spa_scrub_active = B_TRUE;
1423
1424         /*
1425          * First process the free list.  If we pause the free, don't do
1426          * any scanning.  This ensures that there is no free list when
1427          * we are scanning, so the scan code doesn't have to worry about
1428          * traversing it.
1429          */
1430         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1431                 scn->scn_is_bptree = B_FALSE;
1432                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1433                     NULL, ZIO_FLAG_MUSTSUCCEED);
1434                 err = bpobj_iterate(&dp->dp_free_bpobj,
1435                     dsl_scan_free_block_cb, scn, tx);
1436                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1437
1438                 if (err == 0 && spa_feature_is_active(spa,
1439                     &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1440                         scn->scn_is_bptree = B_TRUE;
1441                         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1442                             NULL, ZIO_FLAG_MUSTSUCCEED);
1443                         err = bptree_iterate(dp->dp_meta_objset,
1444                             dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1445                             scn, tx);
1446                         VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1447                         if (err != 0)
1448                                 return;
1449
1450                         /* disable async destroy feature */
1451                         spa_feature_decr(spa,
1452                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY], tx);
1453                         ASSERT(!spa_feature_is_active(spa,
1454                             &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]));
1455                         VERIFY3U(0, ==, zap_remove(dp->dp_meta_objset,
1456                             DMU_POOL_DIRECTORY_OBJECT,
1457                             DMU_POOL_BPTREE_OBJ, tx));
1458                         VERIFY3U(0, ==, bptree_free(dp->dp_meta_objset,
1459                             dp->dp_bptree_obj, tx));
1460                         dp->dp_bptree_obj = 0;
1461                 }
1462                 if (scn->scn_visited_this_txg) {
1463                         zfs_dbgmsg("freed %llu blocks in %llums from "
1464                             "free_bpobj/bptree txg %llu",
1465                             (longlong_t)scn->scn_visited_this_txg,
1466                             (longlong_t)
1467                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1468                             (longlong_t)tx->tx_txg);
1469                         scn->scn_visited_this_txg = 0;
1470                         /*
1471                          * Re-sync the ddt so that we can further modify
1472                          * it when doing bprewrite.
1473                          */
1474                         ddt_sync(spa, tx->tx_txg);
1475                 }
1476                 if (err == ERESTART)
1477                         return;
1478         }
1479
1480         if (scn->scn_phys.scn_state != DSS_SCANNING)
1481                 return;
1482
1483         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1484             scn->scn_phys.scn_ddt_class_max) {
1485                 zfs_dbgmsg("doing scan sync txg %llu; "
1486                     "ddt bm=%llu/%llu/%llu/%llx",
1487                     (longlong_t)tx->tx_txg,
1488                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1489                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1490                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1491                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1492                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1493                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1494                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1495                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1496         } else {
1497                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1498                     (longlong_t)tx->tx_txg,
1499                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1500                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1501                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1502                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1503         }
1504
1505         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1506             NULL, ZIO_FLAG_CANFAIL);
1507         dsl_scan_visit(scn, tx);
1508         (void) zio_wait(scn->scn_zio_root);
1509         scn->scn_zio_root = NULL;
1510
1511         zfs_dbgmsg("visited %llu blocks in %llums",
1512             (longlong_t)scn->scn_visited_this_txg,
1513             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1514
1515         if (!scn->scn_pausing) {
1516                 /* finished with scan. */
1517                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1518                 dsl_scan_done(scn, B_TRUE, tx);
1519         }
1520
1521         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1522                 mutex_enter(&spa->spa_scrub_lock);
1523                 while (spa->spa_scrub_inflight > 0) {
1524                         cv_wait(&spa->spa_scrub_io_cv,
1525                             &spa->spa_scrub_lock);
1526                 }
1527                 mutex_exit(&spa->spa_scrub_lock);
1528         }
1529
1530         dsl_scan_sync_state(scn, tx);
1531 }
1532
1533 /*
1534  * This will start a new scan, or restart an existing one.
1535  */
1536 void
1537 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1538 {
1539         if (txg == 0) {
1540                 dmu_tx_t *tx;
1541                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1542                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1543
1544                 txg = dmu_tx_get_txg(tx);
1545                 dp->dp_scan->scn_restart_txg = txg;
1546                 dmu_tx_commit(tx);
1547         } else {
1548                 dp->dp_scan->scn_restart_txg = txg;
1549         }
1550         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1551 }
1552
1553 boolean_t
1554 dsl_scan_resilvering(dsl_pool_t *dp)
1555 {
1556         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1557             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1558 }
1559
1560 /*
1561  * scrub consumers
1562  */
1563
1564 static void
1565 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1566 {
1567         int i;
1568
1569         /*
1570          * If we resume after a reboot, zab will be NULL; don't record
1571          * incomplete stats in that case.
1572          */
1573         if (zab == NULL)
1574                 return;
1575
1576         for (i = 0; i < 4; i++) {
1577                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1578                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1579                 if (t & DMU_OT_NEWTYPE)
1580                         t = DMU_OT_OTHER;
1581                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1582                 int equal;
1583
1584                 zb->zb_count++;
1585                 zb->zb_asize += BP_GET_ASIZE(bp);
1586                 zb->zb_lsize += BP_GET_LSIZE(bp);
1587                 zb->zb_psize += BP_GET_PSIZE(bp);
1588                 zb->zb_gangs += BP_COUNT_GANG(bp);
1589
1590                 switch (BP_GET_NDVAS(bp)) {
1591                 case 2:
1592                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1593                             DVA_GET_VDEV(&bp->blk_dva[1]))
1594                                 zb->zb_ditto_2_of_2_samevdev++;
1595                         break;
1596                 case 3:
1597                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1598                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1599                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1600                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1601                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1602                             DVA_GET_VDEV(&bp->blk_dva[2]));
1603                         if (equal == 1)
1604                                 zb->zb_ditto_2_of_3_samevdev++;
1605                         else if (equal == 3)
1606                                 zb->zb_ditto_3_of_3_samevdev++;
1607                         break;
1608                 }
1609         }
1610 }
1611
1612 static void
1613 dsl_scan_scrub_done(zio_t *zio)
1614 {
1615         spa_t *spa = zio->io_spa;
1616
1617         zio_data_buf_free(zio->io_data, zio->io_size);
1618
1619         mutex_enter(&spa->spa_scrub_lock);
1620         spa->spa_scrub_inflight--;
1621         cv_broadcast(&spa->spa_scrub_io_cv);
1622
1623         if (zio->io_error && (zio->io_error != ECKSUM ||
1624             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1625                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1626         }
1627         mutex_exit(&spa->spa_scrub_lock);
1628 }
1629
1630 static int
1631 dsl_scan_scrub_cb(dsl_pool_t *dp,
1632     const blkptr_t *bp, const zbookmark_t *zb)
1633 {
1634         dsl_scan_t *scn = dp->dp_scan;
1635         size_t size = BP_GET_PSIZE(bp);
1636         spa_t *spa = dp->dp_spa;
1637         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1638         boolean_t needs_io;
1639         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1640         int zio_priority;
1641         int scan_delay = 0;
1642
1643         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1644             phys_birth >= scn->scn_phys.scn_max_txg)
1645                 return (0);
1646
1647         count_block(dp->dp_blkstats, bp);
1648
1649         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1650         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1651                 zio_flags |= ZIO_FLAG_SCRUB;
1652                 zio_priority = ZIO_PRIORITY_SCRUB;
1653                 needs_io = B_TRUE;
1654                 scan_delay = zfs_scrub_delay;
1655         } else if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
1656                 zio_flags |= ZIO_FLAG_RESILVER;
1657                 zio_priority = ZIO_PRIORITY_RESILVER;
1658                 needs_io = B_FALSE;
1659                 scan_delay = zfs_resilver_delay;
1660         }
1661
1662         /* If it's an intent log block, failure is expected. */
1663         if (zb->zb_level == ZB_ZIL_LEVEL)
1664                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1665
1666         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1667                 vdev_t *vd = vdev_lookup_top(spa,
1668                     DVA_GET_VDEV(&bp->blk_dva[d]));
1669
1670                 /*
1671                  * Keep track of how much data we've examined so that
1672                  * zpool(1M) status can make useful progress reports.
1673                  */
1674                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1675                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1676
1677                 /* if it's a resilver, this may not be in the target range */
1678                 if (!needs_io) {
1679                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1680                                 /*
1681                                  * Gang members may be spread across multiple
1682                                  * vdevs, so the best estimate we have is the
1683                                  * scrub range, which has already been checked.
1684                                  * XXX -- it would be better to change our
1685                                  * allocation policy to ensure that all
1686                                  * gang members reside on the same vdev.
1687                                  */
1688                                 needs_io = B_TRUE;
1689                         } else {
1690                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1691                                     phys_birth, 1);
1692                         }
1693                 }
1694         }
1695
1696         if (needs_io && !zfs_no_scrub_io) {
1697                 vdev_t *rvd = spa->spa_root_vdev;
1698                 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1699                 void *data = zio_data_buf_alloc(size);
1700
1701                 mutex_enter(&spa->spa_scrub_lock);
1702                 while (spa->spa_scrub_inflight >= maxinflight)
1703                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1704                 spa->spa_scrub_inflight++;
1705                 mutex_exit(&spa->spa_scrub_lock);
1706
1707                 /*
1708                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1709                  * then throttle our workload to limit the impact of a scan.
1710                  */
1711                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1712                         delay(scan_delay);
1713
1714                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1715                     dsl_scan_scrub_done, NULL, zio_priority,
1716                     zio_flags, zb));
1717         }
1718
1719         /* do not relocate this block */
1720         return (0);
1721 }
1722
1723 int
1724 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1725 {
1726         spa_t *spa = dp->dp_spa;
1727
1728         /*
1729          * Purge all vdev caches and probe all devices.  We do this here
1730          * rather than in sync context because this requires a writer lock
1731          * on the spa_config lock, which we can't do from sync context.  The
1732          * spa_scrub_reopen flag indicates that vdev_open() should not
1733          * attempt to start another scrub.
1734          */
1735         spa_vdev_state_enter(spa, SCL_NONE);
1736         spa->spa_scrub_reopen = B_TRUE;
1737         vdev_reopen(spa->spa_root_vdev);
1738         spa->spa_scrub_reopen = B_FALSE;
1739         (void) spa_vdev_state_exit(spa, NULL, 0);
1740
1741         return (dsl_sync_task_do(dp, dsl_scan_setup_check,
1742             dsl_scan_setup_sync, dp->dp_scan, &func, 0));
1743 }