module/zfs/dsl_scan.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  23  */
  24
  25 #include <sys/dsl_scan.h>
  26 #include <sys/dsl_pool.h>
  27 #include <sys/dsl_dataset.h>
  28 #include <sys/dsl_prop.h>
  29 #include <sys/dsl_dir.h>
  30 #include <sys/dsl_synctask.h>
  31 #include <sys/dnode.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/dmu_objset.h>
  34 #include <sys/arc.h>
  35 #include <sys/zap.h>
  36 #include <sys/zio.h>
  37 #include <sys/zfs_context.h>
  38 #include <sys/fs/zfs.h>
  39 #include <sys/zfs_znode.h>
  40 #include <sys/spa_impl.h>
  41 #include <sys/vdev_impl.h>
  42 #include <sys/zil_impl.h>
  43 #include <sys/zio_checksum.h>
  44 #include <sys/ddt.h>
  45 #include <sys/sa.h>
  46 #include <sys/sa_impl.h>
  47 #ifdef _KERNEL
  48 #include <sys/zfs_vfsops.h>
  49 #endif
  50
  51 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
  52
  53 static scan_cb_t dsl_scan_defrag_cb;
  54 static scan_cb_t dsl_scan_scrub_cb;
  55 static scan_cb_t dsl_scan_remove_cb;
  56 static dsl_syncfunc_t dsl_scan_cancel_sync;
  57 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
  58
  59 int zfs_top_maxinflight = 32;           /* maximum I/Os per top-level */
  60 int zfs_resilver_delay = 2;             /* number of ticks to delay resilver */
  61 int zfs_scrub_delay = 4;                /* number of ticks to delay scrub */
  62 int zfs_scan_idle = 50;                 /* idle window in clock ticks */
  63
  64 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
  65 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
  66 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
  67 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  68 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable srub prefetching */
  69 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
  70 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
  71
  72 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
  73         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
  74         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
  75
  76 extern int zfs_txg_timeout;
  77
  78 /* the order has to match pool_scan_type */
  79 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
  80         NULL,
  81         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
  82         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
  83 };
  84
  85 int
  86 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
  87 {
  88         int err;
  89         dsl_scan_t *scn;
  90         spa_t *spa = dp->dp_spa;
  91         uint64_t f;
  92
  93         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
  94         scn->scn_dp = dp;
  95
  96         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
  97             "scrub_func", sizeof (uint64_t), 1, &f);
  98         if (err == 0) {
  99                 /*
 100                  * There was an old-style scrub in progress.  Restart a
 101                  * new-style scrub from the beginning.
 102                  */
 103                 scn->scn_restart_txg = txg;
 104                 zfs_dbgmsg("old-style scrub was in progress; "
 105                     "restarting new-style scrub in txg %llu",
 106                     scn->scn_restart_txg);
 107
 108                 /*
 109                  * Load the queue obj from the old location so that it
 110                  * can be freed by dsl_scan_done().
 111                  */
 112                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 113                     "scrub_queue", sizeof (uint64_t), 1,
 114                     &scn->scn_phys.scn_queue_obj);
 115         } else {
 116                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 117                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 118                     &scn->scn_phys);
 119                 if (err == ENOENT)
 120                         return (0);
 121                 else if (err)
 122                         return (err);
 123
 124                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 125                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 126                         /*
 127                          * A new-type scrub was in progress on an old
 128                          * pool, and the pool was accessed by old
 129                          * software.  Restart from the beginning, since
 130                          * the old software may have changed the pool in
 131                          * the meantime.
 132                          */
 133                         scn->scn_restart_txg = txg;
 134                         zfs_dbgmsg("new-style scrub was modified "
 135                             "by old software; restarting in txg %llu",
 136                             scn->scn_restart_txg);
 137                 }
 138         }
 139
 140         spa_scan_stat_init(spa);
 141         return (0);
 142 }
 143
 144 void
 145 dsl_scan_fini(dsl_pool_t *dp)
 146 {
 147         if (dp->dp_scan) {
 148                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 149                 dp->dp_scan = NULL;
 150         }
 151 }
 152
 153 /* ARGSUSED */
 154 static int
 155 dsl_scan_setup_check(void *arg1, void *arg2, dmu_tx_t *tx)
 156 {
 157         dsl_scan_t *scn = arg1;
 158
 159         if (scn->scn_phys.scn_state == DSS_SCANNING)
 160                 return (EBUSY);
 161
 162         return (0);
 163 }
 164
 165 /* ARGSUSED */
 166 static void
 167 dsl_scan_setup_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 168 {
 169         dsl_scan_t *scn = arg1;
 170         pool_scan_func_t *funcp = arg2;
 171         dmu_object_type_t ot = 0;
 172         dsl_pool_t *dp = scn->scn_dp;
 173         spa_t *spa = dp->dp_spa;
 174
 175         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 176         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 177         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 178         scn->scn_phys.scn_func = *funcp;
 179         scn->scn_phys.scn_state = DSS_SCANNING;
 180         scn->scn_phys.scn_min_txg = 0;
 181         scn->scn_phys.scn_max_txg = tx->tx_txg;
 182         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 183         scn->scn_phys.scn_start_time = gethrestime_sec();
 184         scn->scn_phys.scn_errors = 0;
 185         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 186         scn->scn_restart_txg = 0;
 187         spa_scan_stat_init(spa);
 188
 189         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 190                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 191
 192                 /* rewrite all disk labels */
 193                 vdev_config_dirty(spa->spa_root_vdev);
 194
 195                 if (vdev_resilver_needed(spa->spa_root_vdev,
 196                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 197                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 198                 } else {
 199                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 200                 }
 201
 202                 spa->spa_scrub_started = B_TRUE;
 203                 /*
 204                  * If this is an incremental scrub, limit the DDT scrub phase
 205                  * to just the auto-ditto class (for correctness); the rest
 206                  * of the scrub should go faster using top-down pruning.
 207                  */
 208                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 209                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 210
 211         }
 212
 213         /* back to the generic stuff */
 214
 215         if (dp->dp_blkstats == NULL) {
 216                 dp->dp_blkstats =
 217                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 218         }
 219         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 220
 221         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 222                 ot = DMU_OT_ZAP_OTHER;
 223
 224         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 225             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 226
 227         dsl_scan_sync_state(scn, tx);
 228
 229         spa_history_log_internal(LOG_POOL_SCAN, spa, tx,
 230             "func=%u mintxg=%llu maxtxg=%llu",
 231             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 232 }
 233
 234 /* ARGSUSED */
 235 static void
 236 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 237 {
 238         static const char *old_names[] = {
 239                 "scrub_bookmark",
 240                 "scrub_ddt_bookmark",
 241                 "scrub_ddt_class_max",
 242                 "scrub_queue",
 243                 "scrub_min_txg",
 244                 "scrub_max_txg",
 245                 "scrub_func",
 246                 "scrub_errors",
 247                 NULL
 248         };
 249
 250         dsl_pool_t *dp = scn->scn_dp;
 251         spa_t *spa = dp->dp_spa;
 252         int i;
 253
 254         /* Remove any remnants of an old-style scrub. */
 255         for (i = 0; old_names[i]; i++) {
 256                 (void) zap_remove(dp->dp_meta_objset,
 257                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 258         }
 259
 260         if (scn->scn_phys.scn_queue_obj != 0) {
 261                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 262                     scn->scn_phys.scn_queue_obj, tx));
 263                 scn->scn_phys.scn_queue_obj = 0;
 264         }
 265
 266         /*
 267          * If we were "restarted" from a stopped state, don't bother
 268          * with anything else.
 269          */
 270         if (scn->scn_phys.scn_state != DSS_SCANNING)
 271                 return;
 272
 273         if (complete)
 274                 scn->scn_phys.scn_state = DSS_FINISHED;
 275         else
 276                 scn->scn_phys.scn_state = DSS_CANCELED;
 277
 278         spa_history_log_internal(LOG_POOL_SCAN_DONE, spa, tx,
 279             "complete=%u", complete);
 280
 281         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 282                 mutex_enter(&spa->spa_scrub_lock);
 283                 while (spa->spa_scrub_inflight > 0) {
 284                         cv_wait(&spa->spa_scrub_io_cv,
 285                             &spa->spa_scrub_lock);
 286                 }
 287                 mutex_exit(&spa->spa_scrub_lock);
 288                 spa->spa_scrub_started = B_FALSE;
 289                 spa->spa_scrub_active = B_FALSE;
 290
 291                 /*
 292                  * If the scrub/resilver completed, update all DTLs to
 293                  * reflect this.  Whether it succeeded or not, vacate
 294                  * all temporary scrub DTLs.
 295                  */
 296                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 297                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 298                 if (complete) {
 299                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 300                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 301                 }
 302                 spa_errlog_rotate(spa);
 303
 304                 /*
 305                  * We may have finished replacing a device.
 306                  * Let the async thread assess this and handle the detach.
 307                  */
 308                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 309         }
 310
 311         scn->scn_phys.scn_end_time = gethrestime_sec();
 312 }
 313
 314 /* ARGSUSED */
 315 static int
 316 dsl_scan_cancel_check(void *arg1, void *arg2, dmu_tx_t *tx)
 317 {
 318         dsl_scan_t *scn = arg1;
 319
 320         if (scn->scn_phys.scn_state != DSS_SCANNING)
 321                 return (ENOENT);
 322         return (0);
 323 }
 324
 325 /* ARGSUSED */
 326 static void
 327 dsl_scan_cancel_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 328 {
 329         dsl_scan_t *scn = arg1;
 330
 331         dsl_scan_done(scn, B_FALSE, tx);
 332         dsl_scan_sync_state(scn, tx);
 333 }
 334
 335 int
 336 dsl_scan_cancel(dsl_pool_t *dp)
 337 {
 338         boolean_t complete = B_FALSE;
 339         int err;
 340
 341         err = dsl_sync_task_do(dp, dsl_scan_cancel_check,
 342             dsl_scan_cancel_sync, dp->dp_scan, &complete, 3);
 343         return (err);
 344 }
 345
 346 static void dsl_scan_visitbp(blkptr_t *bp,
 347     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 348     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 349     dmu_tx_t *tx);
 350 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 351     dmu_objset_type_t ostype,
 352     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 353
 354 void
 355 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 356 {
 357         zio_free(dp->dp_spa, txg, bp);
 358 }
 359
 360 void
 361 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 362 {
 363         ASSERT(dsl_pool_sync_context(dp));
 364         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
 365 }
 366
 367 int
 368 dsl_read(zio_t *pio, spa_t *spa, const blkptr_t *bpp, arc_buf_t *pbuf,
 369     arc_done_func_t *done, void *private, int priority, int zio_flags,
 370     uint32_t *arc_flags, const zbookmark_t *zb)
 371 {
 372         return (arc_read(pio, spa, bpp, pbuf, done, private,
 373             priority, zio_flags, arc_flags, zb));
 374 }
 375
 376 int
 377 dsl_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bpp,
 378     arc_done_func_t *done, void *private, int priority, int zio_flags,
 379     uint32_t *arc_flags, const zbookmark_t *zb)
 380 {
 381         return (arc_read_nolock(pio, spa, bpp, done, private,
 382             priority, zio_flags, arc_flags, zb));
 383 }
 384
 385 static boolean_t
 386 bookmark_is_zero(const zbookmark_t *zb)
 387 {
 388         return (zb->zb_objset == 0 && zb->zb_object == 0 &&
 389             zb->zb_level == 0 && zb->zb_blkid == 0);
 390 }
 391
 392 /* dnp is the dnode for zb1->zb_object */
 393 static boolean_t
 394 bookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
 395     const zbookmark_t *zb2)
 396 {
 397         uint64_t zb1nextL0, zb2thisobj;
 398
 399         ASSERT(zb1->zb_objset == zb2->zb_objset);
 400         ASSERT(zb2->zb_level == 0);
 401
 402         /*
 403          * A bookmark in the deadlist is considered to be after
 404          * everything else.
 405          */
 406         if (zb2->zb_object == DMU_DEADLIST_OBJECT)
 407                 return (B_TRUE);
 408
 409         /* The objset_phys_t isn't before anything. */
 410         if (dnp == NULL)
 411                 return (B_FALSE);
 412
 413         zb1nextL0 = (zb1->zb_blkid + 1) <<
 414             ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
 415
 416         zb2thisobj = zb2->zb_object ? zb2->zb_object :
 417             zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
 418
 419         if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
 420                 uint64_t nextobj = zb1nextL0 *
 421                     (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
 422                 return (nextobj <= zb2thisobj);
 423         }
 424
 425         if (zb1->zb_object < zb2thisobj)
 426                 return (B_TRUE);
 427         if (zb1->zb_object > zb2thisobj)
 428                 return (B_FALSE);
 429         if (zb2->zb_object == DMU_META_DNODE_OBJECT)
 430                 return (B_FALSE);
 431         return (zb1nextL0 <= zb2->zb_blkid);
 432 }
 433
 434 static uint64_t
 435 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 436 {
 437         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 438         if (dsl_dataset_is_snapshot(ds))
 439                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 440         return (smt);
 441 }
 442
 443 static void
 444 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 445 {
 446         VERIFY(0 == zap_update(scn->scn_dp->dp_meta_objset,
 447             DMU_POOL_DIRECTORY_OBJECT,
 448             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 449             &scn->scn_phys, tx));
 450 }
 451
 452 static boolean_t
 453 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 454 {
 455         uint64_t elapsed_nanosecs;
 456         int mintime;
 457
 458         /* we never skip user/group accounting objects */
 459         if (zb && (int64_t)zb->zb_object < 0)
 460                 return (B_FALSE);
 461
 462         if (scn->scn_pausing)
 463                 return (B_TRUE); /* we're already pausing */
 464
 465         if (!bookmark_is_zero(&scn->scn_phys.scn_bookmark))
 466                 return (B_FALSE); /* we're resuming */
 467
 468         /* We only know how to resume from level-0 blocks. */
 469         if (zb && zb->zb_level != 0)
 470                 return (B_FALSE);
 471
 472         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 473             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 474         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 475         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 476             (elapsed_nanosecs / MICROSEC > mintime &&
 477             txg_sync_waiting(scn->scn_dp)) ||
 478             spa_shutting_down(scn->scn_dp->dp_spa)) {
 479                 if (zb) {
 480                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 481                             (longlong_t)zb->zb_objset,
 482                             (longlong_t)zb->zb_object,
 483                             (longlong_t)zb->zb_level,
 484                             (longlong_t)zb->zb_blkid);
 485                         scn->scn_phys.scn_bookmark = *zb;
 486                 }
 487                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 488                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 489                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 490                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 491                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 492                 scn->scn_pausing = B_TRUE;
 493                 return (B_TRUE);
 494         }
 495         return (B_FALSE);
 496 }
 497
 498 typedef struct zil_scan_arg {
 499         dsl_pool_t      *zsa_dp;
 500         zil_header_t    *zsa_zh;
 501 } zil_scan_arg_t;
 502
 503 /* ARGSUSED */
 504 static int
 505 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 506 {
 507         zil_scan_arg_t *zsa = arg;
 508         dsl_pool_t *dp = zsa->zsa_dp;
 509         dsl_scan_t *scn = dp->dp_scan;
 510         zil_header_t *zh = zsa->zsa_zh;
 511         zbookmark_t zb;
 512
 513         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 514                 return (0);
 515
 516         /*
 517          * One block ("stubby") can be allocated a long time ago; we
 518          * want to visit that one because it has been allocated
 519          * (on-disk) even if it hasn't been claimed (even though for
 520          * scrub there's nothing to do to it).
 521          */
 522         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 523                 return (0);
 524
 525         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 526             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 527
 528         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 529         return (0);
 530 }
 531
 532 /* ARGSUSED */
 533 static int
 534 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 535 {
 536         if (lrc->lrc_txtype == TX_WRITE) {
 537                 zil_scan_arg_t *zsa = arg;
 538                 dsl_pool_t *dp = zsa->zsa_dp;
 539                 dsl_scan_t *scn = dp->dp_scan;
 540                 zil_header_t *zh = zsa->zsa_zh;
 541                 lr_write_t *lr = (lr_write_t *)lrc;
 542                 blkptr_t *bp = &lr->lr_blkptr;
 543                 zbookmark_t zb;
 544
 545                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 546                         return (0);
 547
 548                 /*
 549                  * birth can be < claim_txg if this record's txg is
 550                  * already txg sync'ed (but this log block contains
 551                  * other records that are not synced)
 552                  */
 553                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 554                         return (0);
 555
 556                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 557                     lr->lr_foid, ZB_ZIL_LEVEL,
 558                     lr->lr_offset / BP_GET_LSIZE(bp));
 559
 560                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 561         }
 562         return (0);
 563 }
 564
 565 static void
 566 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 567 {
 568         uint64_t claim_txg = zh->zh_claim_txg;
 569         zil_scan_arg_t zsa = { dp, zh };
 570         zilog_t *zilog;
 571
 572         /*
 573          * We only want to visit blocks that have been claimed but not yet
 574          * replayed (or, in read-only mode, blocks that *would* be claimed).
 575          */
 576         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 577                 return;
 578
 579         zilog = zil_alloc(dp->dp_meta_objset, zh);
 580
 581         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 582             claim_txg);
 583
 584         zil_free(zilog);
 585 }
 586
 587 /* ARGSUSED */
 588 static void
 589 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 590     uint64_t objset, uint64_t object, uint64_t blkid)
 591 {
 592         zbookmark_t czb;
 593         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 594
 595         if (zfs_no_scrub_prefetch)
 596                 return;
 597
 598         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 599             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 600                 return;
 601
 602         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 603
 604         /*
 605          * XXX need to make sure all of these arc_read() prefetches are
 606          * done before setting xlateall (similar to dsl_read())
 607          */
 608         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 609             buf, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 610             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 611 }
 612
 613 static boolean_t
 614 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 615     const zbookmark_t *zb)
 616 {
 617         /*
 618          * We never skip over user/group accounting objects (obj<0)
 619          */
 620         if (!bookmark_is_zero(&scn->scn_phys.scn_bookmark) &&
 621             (int64_t)zb->zb_object >= 0) {
 622                 /*
 623                  * If we already visited this bp & everything below (in
 624                  * a prior txg sync), don't bother doing it again.
 625                  */
 626                 if (bookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 627                         return (B_TRUE);
 628
 629                 /*
 630                  * If we found the block we're trying to resume from, or
 631                  * we went past it to a different object, zero it out to
 632                  * indicate that it's OK to start checking for pausing
 633                  * again.
 634                  */
 635                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 636                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 637                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 638                             (longlong_t)zb->zb_objset,
 639                             (longlong_t)zb->zb_object,
 640                             (longlong_t)zb->zb_level,
 641                             (longlong_t)zb->zb_blkid);
 642                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 643                 }
 644         }
 645         return (B_FALSE);
 646 }
 647
 648 /*
 649  * Return nonzero on i/o error.
 650  * Return new buf to write out in *bufp.
 651  */
 652 static int
 653 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 654     dnode_phys_t *dnp, const blkptr_t *bp,
 655     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 656 {
 657         dsl_pool_t *dp = scn->scn_dp;
 658         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 659         int err;
 660
 661         if (BP_GET_LEVEL(bp) > 0) {
 662                 uint32_t flags = ARC_WAIT;
 663                 int i;
 664                 blkptr_t *cbp;
 665                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 666
 667                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 668                     arc_getbuf_func, bufp,
 669                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 670                 if (err) {
 671                         scn->scn_phys.scn_errors++;
 672                         return (err);
 673                 }
 674                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 675                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 676                             zb->zb_object, zb->zb_blkid * epb + i);
 677                 }
 678                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 679                         zbookmark_t czb;
 680
 681                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 682                             zb->zb_level - 1,
 683                             zb->zb_blkid * epb + i);
 684                         dsl_scan_visitbp(cbp, &czb, dnp,
 685                             *bufp, ds, scn, ostype, tx);
 686                 }
 687         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 688                 uint32_t flags = ARC_WAIT;
 689
 690                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 691                     arc_getbuf_func, bufp,
 692                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 693                 if (err) {
 694                         scn->scn_phys.scn_errors++;
 695                         return (err);
 696                 }
 697         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 698                 uint32_t flags = ARC_WAIT;
 699                 dnode_phys_t *cdnp;
 700                 int i, j;
 701                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 702
 703                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 704                     arc_getbuf_func, bufp,
 705                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 706                 if (err) {
 707                         scn->scn_phys.scn_errors++;
 708                         return (err);
 709                 }
 710                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 711                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 712                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 713                                 dsl_scan_prefetch(scn, *bufp, cbp,
 714                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 715                         }
 716                 }
 717                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 718                         dsl_scan_visitdnode(scn, ds, ostype,
 719                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 720                 }
 721
 722         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 723                 uint32_t flags = ARC_WAIT;
 724                 objset_phys_t *osp;
 725
 726                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 727                     arc_getbuf_func, bufp,
 728                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 729                 if (err) {
 730                         scn->scn_phys.scn_errors++;
 731                         return (err);
 732                 }
 733
 734                 osp = (*bufp)->b_data;
 735
 736                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 737                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 738
 739                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 740                         /*
 741                          * We also always visit user/group accounting
 742                          * objects, and never skip them, even if we are
 743                          * pausing.  This is necessary so that the space
 744                          * deltas from this txg get integrated.
 745                          */
 746                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 747                             &osp->os_groupused_dnode, *bufp,
 748                             DMU_GROUPUSED_OBJECT, tx);
 749                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 750                             &osp->os_userused_dnode, *bufp,
 751                             DMU_USERUSED_OBJECT, tx);
 752                 }
 753         }
 754
 755         return (0);
 756 }
 757
 758 static void
 759 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 760     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 761     uint64_t object, dmu_tx_t *tx)
 762 {
 763         int j;
 764
 765         for (j = 0; j < dnp->dn_nblkptr; j++) {
 766                 zbookmark_t czb;
 767
 768                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 769                     dnp->dn_nlevels - 1, j);
 770                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 771                     &czb, dnp, buf, ds, scn, ostype, tx);
 772         }
 773
 774         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 775                 zbookmark_t czb;
 776                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 777                     0, DMU_SPILL_BLKID);
 778                 dsl_scan_visitbp(&dnp->dn_spill,
 779                     &czb, dnp, buf, ds, scn, ostype, tx);
 780         }
 781 }
 782
 783 /*
 784  * The arguments are in this order because mdb can only print the
 785  * first 5; we want them to be useful.
 786  */
 787 static void
 788 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 789     dnode_phys_t *dnp, arc_buf_t *pbuf,
 790     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 791     dmu_tx_t *tx)
 792 {
 793         dsl_pool_t *dp = scn->scn_dp;
 794         arc_buf_t *buf = NULL;
 795         blkptr_t bp_toread = *bp;
 796
 797         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 798
 799         if (dsl_scan_check_pause(scn, zb))
 800                 return;
 801
 802         if (dsl_scan_check_resume(scn, dnp, zb))
 803                 return;
 804
 805         if (bp->blk_birth == 0)
 806                 return;
 807
 808         scn->scn_visited_this_txg++;
 809
 810         dprintf_bp(bp,
 811             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 812             ds, ds ? ds->ds_object : 0,
 813             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 814             pbuf, bp);
 815
 816         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 817                 return;
 818
 819         if (BP_GET_TYPE(bp) != DMU_OT_USERGROUP_USED) {
 820                 /*
 821                  * For non-user-accounting blocks, we need to read the
 822                  * new bp (from a deleted snapshot, found in
 823                  * check_existing_xlation).  If we used the old bp,
 824                  * pointers inside this block from before we resumed
 825                  * would be untranslated.
 826                  *
 827                  * For user-accounting blocks, we need to read the old
 828                  * bp, because we will apply the entire space delta to
 829                  * it (original untranslated -> translations from
 830                  * deleted snap -> now).
 831                  */
 832                 bp_toread = *bp;
 833         }
 834
 835         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
 836             &buf) != 0)
 837                 return;
 838
 839         /*
 840          * If dsl_scan_ddt() has aready visited this block, it will have
 841          * already done any translations or scrubbing, so don't call the
 842          * callback again.
 843          */
 844         if (ddt_class_contains(dp->dp_spa,
 845             scn->scn_phys.scn_ddt_class_max, bp)) {
 846                 ASSERT(buf == NULL);
 847                 return;
 848         }
 849
 850         /*
 851          * If this block is from the future (after cur_max_txg), then we
 852          * are doing this on behalf of a deleted snapshot, and we will
 853          * revisit the future block on the next pass of this dataset.
 854          * Don't scan it now unless we need to because something
 855          * under it was modified.
 856          */
 857         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 858                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 859         }
 860         if (buf)
 861                 (void) arc_buf_remove_ref(buf, &buf);
 862 }
 863
 864 static void
 865 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 866     dmu_tx_t *tx)
 867 {
 868         zbookmark_t zb;
 869
 870         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 871             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 872         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 873             ds, scn, DMU_OST_NONE, tx);
 874
 875         dprintf_ds(ds, "finished scan%s", "");
 876 }
 877
 878 void
 879 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 880 {
 881         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 882         dsl_scan_t *scn = dp->dp_scan;
 883         uint64_t mintxg;
 884
 885         if (scn->scn_phys.scn_state != DSS_SCANNING)
 886                 return;
 887
 888         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 889                 if (dsl_dataset_is_snapshot(ds)) {
 890                         /* Note, scn_cur_{min,max}_txg stays the same. */
 891                         scn->scn_phys.scn_bookmark.zb_objset =
 892                             ds->ds_phys->ds_next_snap_obj;
 893                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 894                             "reset zb_objset to %llu",
 895                             (u_longlong_t)ds->ds_object,
 896                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 897                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 898                 } else {
 899                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 900                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 901                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 902                             "reset bookmark to -1,0,0,0",
 903                             (u_longlong_t)ds->ds_object);
 904                 }
 905         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 906             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 907                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 908                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 909                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 910                 if (dsl_dataset_is_snapshot(ds)) {
 911                         /*
 912                          * We keep the same mintxg; it could be >
 913                          * ds_creation_txg if the previous snapshot was
 914                          * deleted too.
 915                          */
 916                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 917                             scn->scn_phys.scn_queue_obj,
 918                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 919                         zfs_dbgmsg("destroying ds %llu; in queue; "
 920                             "replacing with %llu",
 921                             (u_longlong_t)ds->ds_object,
 922                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 923                 } else {
 924                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 925                             (u_longlong_t)ds->ds_object);
 926                 }
 927         } else {
 928                 zfs_dbgmsg("destroying ds %llu; ignoring",
 929                     (u_longlong_t)ds->ds_object);
 930         }
 931
 932         /*
 933          * dsl_scan_sync() should be called after this, and should sync
 934          * out our changed state, but just to be safe, do it here.
 935          */
 936         dsl_scan_sync_state(scn, tx);
 937 }
 938
 939 void
 940 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 941 {
 942         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 943         dsl_scan_t *scn = dp->dp_scan;
 944         uint64_t mintxg;
 945
 946         if (scn->scn_phys.scn_state != DSS_SCANNING)
 947                 return;
 948
 949         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 950
 951         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 952                 scn->scn_phys.scn_bookmark.zb_objset =
 953                     ds->ds_phys->ds_prev_snap_obj;
 954                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 955                     "reset zb_objset to %llu",
 956                     (u_longlong_t)ds->ds_object,
 957                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 958         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 959             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 960                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 961                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 962                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 963                     scn->scn_phys.scn_queue_obj,
 964                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 965                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 966                     "replacing with %llu",
 967                     (u_longlong_t)ds->ds_object,
 968                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 969         }
 970         dsl_scan_sync_state(scn, tx);
 971 }
 972
 973 void
 974 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 975 {
 976         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 977         dsl_scan_t *scn = dp->dp_scan;
 978         uint64_t mintxg;
 979
 980         if (scn->scn_phys.scn_state != DSS_SCANNING)
 981                 return;
 982
 983         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 984                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 985                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 986                     "reset zb_objset to %llu",
 987                     (u_longlong_t)ds1->ds_object,
 988                     (u_longlong_t)ds2->ds_object);
 989         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 990                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 991                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 992                     "reset zb_objset to %llu",
 993                     (u_longlong_t)ds2->ds_object,
 994                     (u_longlong_t)ds1->ds_object);
 995         }
 996
 997         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 998             ds1->ds_object, &mintxg) == 0) {
 999                 int err;
1000
1001                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
1002                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
1003                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1004                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
1005                 err = zap_add_int_key(dp->dp_meta_objset,
1006                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
1007                 VERIFY(err == 0 || err == EEXIST);
1008                 if (err == EEXIST) {
1009                         /* Both were there to begin with */
1010                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1011                             scn->scn_phys.scn_queue_obj,
1012                             ds1->ds_object, mintxg, tx));
1013                 }
1014                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1015                     "replacing with %llu",
1016                     (u_longlong_t)ds1->ds_object,
1017                     (u_longlong_t)ds2->ds_object);
1018         } else if (zap_lookup_int_key(dp->dp_meta_objset,
1019             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
1020                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
1021                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
1022                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1023                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
1024                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1025                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
1026                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1027                     "replacing with %llu",
1028                     (u_longlong_t)ds2->ds_object,
1029                     (u_longlong_t)ds1->ds_object);
1030         }
1031
1032         dsl_scan_sync_state(scn, tx);
1033 }
1034
1035 struct enqueue_clones_arg {
1036         dmu_tx_t *tx;
1037         uint64_t originobj;
1038 };
1039
1040 /* ARGSUSED */
1041 static int
1042 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1043 {
1044         struct enqueue_clones_arg *eca = arg;
1045         dsl_dataset_t *ds;
1046         int err;
1047         dsl_pool_t *dp = spa->spa_dsl_pool;
1048         dsl_scan_t *scn = dp->dp_scan;
1049
1050         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1051         if (err)
1052                 return (err);
1053
1054         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
1055                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1056                         dsl_dataset_t *prev;
1057                         err = dsl_dataset_hold_obj(dp,
1058                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1059
1060                         dsl_dataset_rele(ds, FTAG);
1061                         if (err)
1062                                 return (err);
1063                         ds = prev;
1064                 }
1065                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1066                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1067                     ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1068         }
1069         dsl_dataset_rele(ds, FTAG);
1070         return (0);
1071 }
1072
1073 static void
1074 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1075 {
1076         dsl_pool_t *dp = scn->scn_dp;
1077         dsl_dataset_t *ds;
1078         objset_t *os;
1079         char *dsname;
1080
1081         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1082
1083         if (dmu_objset_from_ds(ds, &os))
1084                 goto out;
1085
1086         /*
1087          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1088          * snapshots can have ZIL block pointers (which may be the same
1089          * BP as in the head), they must be ignored.  So we traverse the
1090          * ZIL here, rather than in scan_recurse(), because the regular
1091          * snapshot block-sharing rules don't apply to it.
1092          */
1093         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1094                 dsl_scan_zil(dp, &os->os_zil_header);
1095
1096         /*
1097          * Iterate over the bps in this ds.
1098          */
1099         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1100         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1101
1102         dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1103         dsl_dataset_name(ds, dsname);
1104         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1105             "pausing=%u",
1106             (longlong_t)dsobj, dsname,
1107             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1108             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1109             (int)scn->scn_pausing);
1110         kmem_free(dsname, ZFS_MAXNAMELEN);
1111
1112         if (scn->scn_pausing)
1113                 goto out;
1114
1115         /*
1116          * We've finished this pass over this dataset.
1117          */
1118
1119         /*
1120          * If we did not completely visit this dataset, do another pass.
1121          */
1122         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1123                 zfs_dbgmsg("incomplete pass; visiting again");
1124                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1125                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1126                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1127                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1128                 goto out;
1129         }
1130
1131         /*
1132          * Add descendent datasets to work queue.
1133          */
1134         if (ds->ds_phys->ds_next_snap_obj != 0) {
1135                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1136                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1137                     ds->ds_phys->ds_creation_txg, tx) == 0);
1138         }
1139         if (ds->ds_phys->ds_num_children > 1) {
1140                 boolean_t usenext = B_FALSE;
1141                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1142                         uint64_t count;
1143                         /*
1144                          * A bug in a previous version of the code could
1145                          * cause upgrade_clones_cb() to not set
1146                          * ds_next_snap_obj when it should, leading to a
1147                          * missing entry.  Therefore we can only use the
1148                          * next_clones_obj when its count is correct.
1149                          */
1150                         int err = zap_count(dp->dp_meta_objset,
1151                             ds->ds_phys->ds_next_clones_obj, &count);
1152                         if (err == 0 &&
1153                             count == ds->ds_phys->ds_num_children - 1)
1154                                 usenext = B_TRUE;
1155                 }
1156
1157                 if (usenext) {
1158                         VERIFY(zap_join_key(dp->dp_meta_objset,
1159                             ds->ds_phys->ds_next_clones_obj,
1160                             scn->scn_phys.scn_queue_obj,
1161                             ds->ds_phys->ds_creation_txg, tx) == 0);
1162                 } else {
1163                         struct enqueue_clones_arg eca;
1164                         eca.tx = tx;
1165                         eca.originobj = ds->ds_object;
1166
1167                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
1168                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
1169                 }
1170         }
1171
1172 out:
1173         dsl_dataset_rele(ds, FTAG);
1174 }
1175
1176 /* ARGSUSED */
1177 static int
1178 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1179 {
1180         dmu_tx_t *tx = arg;
1181         dsl_dataset_t *ds;
1182         int err;
1183         dsl_pool_t *dp = spa->spa_dsl_pool;
1184         dsl_scan_t *scn = dp->dp_scan;
1185
1186         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1187         if (err)
1188                 return (err);
1189
1190         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1191                 dsl_dataset_t *prev;
1192                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1193                     FTAG, &prev);
1194                 if (err) {
1195                         dsl_dataset_rele(ds, FTAG);
1196                         return (err);
1197                 }
1198
1199                 /*
1200                  * If this is a clone, we don't need to worry about it for now.
1201                  */
1202                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1203                         dsl_dataset_rele(ds, FTAG);
1204                         dsl_dataset_rele(prev, FTAG);
1205                         return (0);
1206                 }
1207                 dsl_dataset_rele(ds, FTAG);
1208                 ds = prev;
1209         }
1210
1211         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1212             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1213         dsl_dataset_rele(ds, FTAG);
1214         return (0);
1215 }
1216
1217 /*
1218  * Scrub/dedup interaction.
1219  *
1220  * If there are N references to a deduped block, we don't want to scrub it
1221  * N times -- ideally, we should scrub it exactly once.
1222  *
1223  * We leverage the fact that the dde's replication class (enum ddt_class)
1224  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1225  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1226  *
1227  * To prevent excess scrubbing, the scrub begins by walking the DDT
1228  * to find all blocks with refcnt > 1, and scrubs each of these once.
1229  * Since there are two replication classes which contain blocks with
1230  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1231  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1232  *
1233  * There would be nothing more to say if a block's refcnt couldn't change
1234  * during a scrub, but of course it can so we must account for changes
1235  * in a block's replication class.
1236  *
1237  * Here's an example of what can occur:
1238  *
1239  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1240  * when visited during the top-down scrub phase, it will be scrubbed twice.
1241  * This negates our scrub optimization, but is otherwise harmless.
1242  *
1243  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1244  * on each visit during the top-down scrub phase, it will never be scrubbed.
1245  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1246  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1247  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1248  * while a scrub is in progress, it scrubs the block right then.
1249  */
1250 static void
1251 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1252 {
1253         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1254         ddt_entry_t dde = { 0 };
1255         int error;
1256         uint64_t n = 0;
1257
1258         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1259                 ddt_t *ddt;
1260
1261                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1262                         break;
1263                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1264                     (longlong_t)ddb->ddb_class,
1265                     (longlong_t)ddb->ddb_type,
1266                     (longlong_t)ddb->ddb_checksum,
1267                     (longlong_t)ddb->ddb_cursor);
1268
1269                 /* There should be no pending changes to the dedup table */
1270                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1271                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1272
1273                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1274                 n++;
1275
1276                 if (dsl_scan_check_pause(scn, NULL))
1277                         break;
1278         }
1279
1280         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1281             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1282             (int)scn->scn_pausing);
1283
1284         ASSERT(error == 0 || error == ENOENT);
1285         ASSERT(error != ENOENT ||
1286             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1287 }
1288
1289 /* ARGSUSED */
1290 void
1291 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1292     ddt_entry_t *dde, dmu_tx_t *tx)
1293 {
1294         const ddt_key_t *ddk = &dde->dde_key;
1295         ddt_phys_t *ddp = dde->dde_phys;
1296         blkptr_t bp;
1297         zbookmark_t zb = { 0 };
1298         int p;
1299
1300         if (scn->scn_phys.scn_state != DSS_SCANNING)
1301                 return;
1302
1303         for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1304                 if (ddp->ddp_phys_birth == 0 ||
1305                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1306                         continue;
1307                 ddt_bp_create(checksum, ddk, ddp, &bp);
1308
1309                 scn->scn_visited_this_txg++;
1310                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1311         }
1312 }
1313
1314 static void
1315 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1316 {
1317         dsl_pool_t *dp = scn->scn_dp;
1318         zap_cursor_t zc;
1319         zap_attribute_t za;
1320
1321         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1322             scn->scn_phys.scn_ddt_class_max) {
1323                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1324                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1325                 dsl_scan_ddt(scn, tx);
1326                 if (scn->scn_pausing)
1327                         return;
1328         }
1329
1330         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1331                 /* First do the MOS & ORIGIN */
1332
1333                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1334                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1335                 dsl_scan_visit_rootbp(scn, NULL,
1336                     &dp->dp_meta_rootbp, tx);
1337                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1338                 if (scn->scn_pausing)
1339                         return;
1340
1341                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1342                         VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
1343                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
1344                 } else {
1345                         dsl_scan_visitds(scn,
1346                             dp->dp_origin_snap->ds_object, tx);
1347                 }
1348                 ASSERT(!scn->scn_pausing);
1349         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1350             ZB_DESTROYED_OBJSET) {
1351                 /*
1352                  * If we were paused, continue from here.  Note if the
1353                  * ds we were paused on was deleted, the zb_objset may
1354                  * be -1, so we will skip this and find a new objset
1355                  * below.
1356                  */
1357                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1358                 if (scn->scn_pausing)
1359                         return;
1360         }
1361
1362         /*
1363          * In case we were paused right at the end of the ds, zero the
1364          * bookmark so we don't think that we're still trying to resume.
1365          */
1366         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1367
1368         /* keep pulling things out of the zap-object-as-queue */
1369         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1370             scn->scn_phys.scn_queue_obj),
1371             zap_cursor_retrieve(&zc, &za) == 0) {
1372                 dsl_dataset_t *ds;
1373                 uint64_t dsobj;
1374
1375                 dsobj = strtonum(za.za_name, NULL);
1376                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1377                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1378
1379                 /* Set up min/max txg */
1380                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1381                 if (za.za_first_integer != 0) {
1382                         scn->scn_phys.scn_cur_min_txg =
1383                             MAX(scn->scn_phys.scn_min_txg,
1384                             za.za_first_integer);
1385                 } else {
1386                         scn->scn_phys.scn_cur_min_txg =
1387                             MAX(scn->scn_phys.scn_min_txg,
1388                             ds->ds_phys->ds_prev_snap_txg);
1389                 }
1390                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1391                 dsl_dataset_rele(ds, FTAG);
1392
1393                 dsl_scan_visitds(scn, dsobj, tx);
1394                 zap_cursor_fini(&zc);
1395                 if (scn->scn_pausing)
1396                         return;
1397         }
1398         zap_cursor_fini(&zc);
1399 }
1400
1401 static int
1402 dsl_scan_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1403 {
1404         dsl_scan_t *scn = arg;
1405         uint64_t elapsed_nanosecs;
1406
1407         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1408
1409         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1410             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1411             txg_sync_waiting(scn->scn_dp)) ||
1412             spa_shutting_down(scn->scn_dp->dp_spa))
1413                 return (ERESTART);
1414
1415         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1416             dmu_tx_get_txg(tx), bp, 0));
1417         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1418             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1419             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1420         scn->scn_visited_this_txg++;
1421         return (0);
1422 }
1423
1424 boolean_t
1425 dsl_scan_active(dsl_scan_t *scn)
1426 {
1427         spa_t *spa = scn->scn_dp->dp_spa;
1428         uint64_t used = 0, comp, uncomp;
1429
1430         if (spa->spa_load_state != SPA_LOAD_NONE)
1431                 return (B_FALSE);
1432         if (spa_shutting_down(spa))
1433                 return (B_FALSE);
1434
1435         if (scn->scn_phys.scn_state == DSS_SCANNING)
1436                 return (B_TRUE);
1437
1438         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1439                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1440                     &used, &comp, &uncomp);
1441         }
1442         return (used != 0);
1443 }
1444
1445 void
1446 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1447 {
1448         dsl_scan_t *scn = dp->dp_scan;
1449         spa_t *spa = dp->dp_spa;
1450         int err;
1451
1452         /*
1453          * Check for scn_restart_txg before checking spa_load_state, so
1454          * that we can restart an old-style scan while the pool is being
1455          * imported (see dsl_scan_init).
1456          */
1457         if (scn->scn_restart_txg != 0 &&
1458             scn->scn_restart_txg <= tx->tx_txg) {
1459                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1460                 dsl_scan_done(scn, B_FALSE, tx);
1461                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1462                         func = POOL_SCAN_RESILVER;
1463                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1464                     func, tx->tx_txg);
1465                 dsl_scan_setup_sync(scn, &func, tx);
1466         }
1467
1468         if (!dsl_scan_active(scn) ||
1469             spa_sync_pass(dp->dp_spa) > 1)
1470                 return;
1471
1472         scn->scn_visited_this_txg = 0;
1473         scn->scn_pausing = B_FALSE;
1474         scn->scn_sync_start_time = gethrtime();
1475         spa->spa_scrub_active = B_TRUE;
1476
1477         /*
1478          * First process the free list.  If we pause the free, don't do
1479          * any scanning.  This ensures that there is no free list when
1480          * we are scanning, so the scan code doesn't have to worry about
1481          * traversing it.
1482          */
1483         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1484                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1485                     NULL, ZIO_FLAG_MUSTSUCCEED);
1486                 err = bpobj_iterate(&dp->dp_free_bpobj,
1487                     dsl_scan_free_cb, scn, tx);
1488                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1489                 if (scn->scn_visited_this_txg) {
1490                         zfs_dbgmsg("freed %llu blocks in %llums from "
1491                             "free_bpobj txg %llu",
1492                             (longlong_t)scn->scn_visited_this_txg,
1493                             (longlong_t)
1494                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1495                             (longlong_t)tx->tx_txg);
1496                         scn->scn_visited_this_txg = 0;
1497                         /*
1498                          * Re-sync the ddt so that we can further modify
1499                          * it when doing bprewrite.
1500                          */
1501                         ddt_sync(spa, tx->tx_txg);
1502                 }
1503                 if (err == ERESTART)
1504                         return;
1505         }
1506
1507         if (scn->scn_phys.scn_state != DSS_SCANNING)
1508                 return;
1509
1510         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1511             scn->scn_phys.scn_ddt_class_max) {
1512                 zfs_dbgmsg("doing scan sync txg %llu; "
1513                     "ddt bm=%llu/%llu/%llu/%llx",
1514                     (longlong_t)tx->tx_txg,
1515                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1516                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1517                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1518                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1519                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1520                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1521                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1522                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1523         } else {
1524                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1525                     (longlong_t)tx->tx_txg,
1526                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1527                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1528                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1529                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1530         }
1531
1532         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1533             NULL, ZIO_FLAG_CANFAIL);
1534         dsl_scan_visit(scn, tx);
1535         (void) zio_wait(scn->scn_zio_root);
1536         scn->scn_zio_root = NULL;
1537
1538         zfs_dbgmsg("visited %llu blocks in %llums",
1539             (longlong_t)scn->scn_visited_this_txg,
1540             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1541
1542         if (!scn->scn_pausing) {
1543                 /* finished with scan. */
1544                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1545                 dsl_scan_done(scn, B_TRUE, tx);
1546         }
1547
1548         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1549                 mutex_enter(&spa->spa_scrub_lock);
1550                 while (spa->spa_scrub_inflight > 0) {
1551                         cv_wait(&spa->spa_scrub_io_cv,
1552                             &spa->spa_scrub_lock);
1553                 }
1554                 mutex_exit(&spa->spa_scrub_lock);
1555         }
1556
1557         dsl_scan_sync_state(scn, tx);
1558 }
1559
1560 /*
1561  * This will start a new scan, or restart an existing one.
1562  */
1563 void
1564 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1565 {
1566         if (txg == 0) {
1567                 dmu_tx_t *tx;
1568                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1569                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1570
1571                 txg = dmu_tx_get_txg(tx);
1572                 dp->dp_scan->scn_restart_txg = txg;
1573                 dmu_tx_commit(tx);
1574         } else {
1575                 dp->dp_scan->scn_restart_txg = txg;
1576         }
1577         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1578 }
1579
1580 boolean_t
1581 dsl_scan_resilvering(dsl_pool_t *dp)
1582 {
1583         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1584             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1585 }
1586
1587 /*
1588  * scrub consumers
1589  */
1590
1591 static void
1592 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1593 {
1594         int i;
1595
1596         /*
1597          * If we resume after a reboot, zab will be NULL; don't record
1598          * incomplete stats in that case.
1599          */
1600         if (zab == NULL)
1601                 return;
1602
1603         for (i = 0; i < 4; i++) {
1604                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1605                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1606                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1607                 int equal;
1608
1609                 zb->zb_count++;
1610                 zb->zb_asize += BP_GET_ASIZE(bp);
1611                 zb->zb_lsize += BP_GET_LSIZE(bp);
1612                 zb->zb_psize += BP_GET_PSIZE(bp);
1613                 zb->zb_gangs += BP_COUNT_GANG(bp);
1614
1615                 switch (BP_GET_NDVAS(bp)) {
1616                 case 2:
1617                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1618                             DVA_GET_VDEV(&bp->blk_dva[1]))
1619                                 zb->zb_ditto_2_of_2_samevdev++;
1620                         break;
1621                 case 3:
1622                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1623                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1624                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1625                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1626                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1627                             DVA_GET_VDEV(&bp->blk_dva[2]));
1628                         if (equal == 1)
1629                                 zb->zb_ditto_2_of_3_samevdev++;
1630                         else if (equal == 3)
1631                                 zb->zb_ditto_3_of_3_samevdev++;
1632                         break;
1633                 }
1634         }
1635 }
1636
1637 static void
1638 dsl_scan_scrub_done(zio_t *zio)
1639 {
1640         spa_t *spa = zio->io_spa;
1641
1642         zio_data_buf_free(zio->io_data, zio->io_size);
1643
1644         mutex_enter(&spa->spa_scrub_lock);
1645         spa->spa_scrub_inflight--;
1646         cv_broadcast(&spa->spa_scrub_io_cv);
1647
1648         if (zio->io_error && (zio->io_error != ECKSUM ||
1649             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1650                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1651         }
1652         mutex_exit(&spa->spa_scrub_lock);
1653 }
1654
1655 static int
1656 dsl_scan_scrub_cb(dsl_pool_t *dp,
1657     const blkptr_t *bp, const zbookmark_t *zb)
1658 {
1659         dsl_scan_t *scn = dp->dp_scan;
1660         size_t size = BP_GET_PSIZE(bp);
1661         spa_t *spa = dp->dp_spa;
1662         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1663         boolean_t needs_io = B_FALSE;
1664         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1665         int zio_priority = 0;
1666         int scan_delay = 0;
1667         int d;
1668
1669         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1670             phys_birth >= scn->scn_phys.scn_max_txg)
1671                 return (0);
1672
1673         count_block(dp->dp_blkstats, bp);
1674
1675         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1676         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1677                 zio_flags |= ZIO_FLAG_SCRUB;
1678                 zio_priority = ZIO_PRIORITY_SCRUB;
1679                 needs_io = B_TRUE;
1680                 scan_delay = zfs_scrub_delay;
1681         } else if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
1682                 zio_flags |= ZIO_FLAG_RESILVER;
1683                 zio_priority = ZIO_PRIORITY_RESILVER;
1684                 needs_io = B_FALSE;
1685                 scan_delay = zfs_resilver_delay;
1686         }
1687
1688         /* If it's an intent log block, failure is expected. */
1689         if (zb->zb_level == ZB_ZIL_LEVEL)
1690                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1691
1692         for (d = 0; d < BP_GET_NDVAS(bp); d++) {
1693                 vdev_t *vd = vdev_lookup_top(spa,
1694                     DVA_GET_VDEV(&bp->blk_dva[d]));
1695
1696                 /*
1697                  * Keep track of how much data we've examined so that
1698                  * zpool(1M) status can make useful progress reports.
1699                  */
1700                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1701                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1702
1703                 /* if it's a resilver, this may not be in the target range */
1704                 if (!needs_io) {
1705                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1706                                 /*
1707                                  * Gang members may be spread across multiple
1708                                  * vdevs, so the best estimate we have is the
1709                                  * scrub range, which has already been checked.
1710                                  * XXX -- it would be better to change our
1711                                  * allocation policy to ensure that all
1712                                  * gang members reside on the same vdev.
1713                                  */
1714                                 needs_io = B_TRUE;
1715                         } else {
1716                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1717                                     phys_birth, 1);
1718                         }
1719                 }
1720         }
1721
1722         if (needs_io && !zfs_no_scrub_io) {
1723                 vdev_t *rvd = spa->spa_root_vdev;
1724                 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1725                 void *data = zio_data_buf_alloc(size);
1726
1727                 mutex_enter(&spa->spa_scrub_lock);
1728                 while (spa->spa_scrub_inflight >= maxinflight)
1729                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1730                 spa->spa_scrub_inflight++;
1731                 mutex_exit(&spa->spa_scrub_lock);
1732
1733                 /*
1734                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1735                  * then throttle our workload to limit the impact of a scan.
1736                  */
1737                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1738                         delay(scan_delay);
1739
1740                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1741                     dsl_scan_scrub_done, NULL, zio_priority,
1742                     zio_flags, zb));
1743         }
1744
1745         /* do not relocate this block */
1746         return (0);
1747 }
1748
1749 int
1750 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1751 {
1752         spa_t *spa = dp->dp_spa;
1753
1754         /*
1755          * Purge all vdev caches and probe all devices.  We do this here
1756          * rather than in sync context because this requires a writer lock
1757          * on the spa_config lock, which we can't do from sync context.  The
1758          * spa_scrub_reopen flag indicates that vdev_open() should not
1759          * attempt to start another scrub.
1760          */
1761         spa_vdev_state_enter(spa, SCL_NONE);
1762         spa->spa_scrub_reopen = B_TRUE;
1763         vdev_reopen(spa->spa_root_vdev);
1764         spa->spa_scrub_reopen = B_FALSE;
1765         (void) spa_vdev_state_exit(spa, NULL, 0);
1766
1767         return (dsl_sync_task_do(dp, dsl_scan_setup_check,
1768             dsl_scan_setup_sync, dp->dp_scan, &func, 0));
1769 }