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) 2011, 2015 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 *,
  54     const zbookmark_phys_t *);
  55
  56 static scan_cb_t dsl_scan_scrub_cb;
  57 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
  58 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
  59
  60 unsigned int zfs_top_maxinflight = 32;  /* maximum I/Os per top-level */
  61 unsigned int zfs_resilver_delay = 2;    /* number of ticks to delay resilver */
  62 unsigned int zfs_scrub_delay = 4;       /* number of ticks to delay scrub */
  63 unsigned int zfs_scan_idle = 50;        /* idle window in clock ticks */
  64
  65 unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
  66 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
  67 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
  68                                                  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 scrub prefetch */
  71
  72 SYSCTL_DECL(_vfs_zfs);
  73 TUNABLE_INT("vfs.zfs.top_maxinflight", &zfs_top_maxinflight);
  74 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RW,
  75     &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
  76 TUNABLE_INT("vfs.zfs.resilver_delay", &zfs_resilver_delay);
  77 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RW,
  78     &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
  79 TUNABLE_INT("vfs.zfs.scrub_delay", &zfs_scrub_delay);
  80 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RW,
  81     &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
  82 TUNABLE_INT("vfs.zfs.scan_idle", &zfs_scan_idle);
  83 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RW,
  84     &zfs_scan_idle, 0, "Idle scan window in clock ticks");
  85 TUNABLE_INT("vfs.zfs.scan_min_time_ms", &zfs_scan_min_time_ms);
  86 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RW,
  87     &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
  88 TUNABLE_INT("vfs.zfs.free_min_time_ms", &zfs_free_min_time_ms);
  89 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RW,
  90     &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
  91 TUNABLE_INT("vfs.zfs.resilver_min_time_ms", &zfs_resilver_min_time_ms);
  92 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RW,
  93     &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
  94 TUNABLE_INT("vfs.zfs.no_scrub_io", &zfs_no_scrub_io);
  95 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RW,
  96     &zfs_no_scrub_io, 0, "Disable scrub I/O");
  97 TUNABLE_INT("vfs.zfs.no_scrub_prefetch", &zfs_no_scrub_prefetch);
  98 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RW,
  99     &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
 100
 101 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
 102 /* max number of blocks to free in a single TXG */
 103 uint64_t zfs_free_max_blocks = UINT64_MAX;
 104 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
 105     &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
 106
 107
 108 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
 109         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
 110         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
 111
 112 extern int zfs_txg_timeout;
 113
 114 /*
 115  * Enable/disable the processing of the free_bpobj object.
 116  */
 117 boolean_t zfs_free_bpobj_enabled = B_TRUE;
 118
 119 SYSCTL_INT(_vfs_zfs, OID_AUTO, free_bpobj_enabled, CTLFLAG_RWTUN,
 120     &zfs_free_bpobj_enabled, 0, "Enable free_bpobj processing");
 121
 122 /* the order has to match pool_scan_type */
 123 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
 124         NULL,
 125         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
 126         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
 127 };
 128
 129 int
 130 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
 131 {
 132         int err;
 133         dsl_scan_t *scn;
 134         spa_t *spa = dp->dp_spa;
 135         uint64_t f;
 136
 137         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
 138         scn->scn_dp = dp;
 139
 140         /*
 141          * It's possible that we're resuming a scan after a reboot so
 142          * make sure that the scan_async_destroying flag is initialized
 143          * appropriately.
 144          */
 145         ASSERT(!scn->scn_async_destroying);
 146         scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
 147             SPA_FEATURE_ASYNC_DESTROY);
 148
 149         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 150             "scrub_func", sizeof (uint64_t), 1, &f);
 151         if (err == 0) {
 152                 /*
 153                  * There was an old-style scrub in progress.  Restart a
 154                  * new-style scrub from the beginning.
 155                  */
 156                 scn->scn_restart_txg = txg;
 157                 zfs_dbgmsg("old-style scrub was in progress; "
 158                     "restarting new-style scrub in txg %llu",
 159                     scn->scn_restart_txg);
 160
 161                 /*
 162                  * Load the queue obj from the old location so that it
 163                  * can be freed by dsl_scan_done().
 164                  */
 165                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 166                     "scrub_queue", sizeof (uint64_t), 1,
 167                     &scn->scn_phys.scn_queue_obj);
 168         } else {
 169                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 170                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 171                     &scn->scn_phys);
 172                 if (err == ENOENT)
 173                         return (0);
 174                 else if (err)
 175                         return (err);
 176
 177                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 178                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 179                         /*
 180                          * A new-type scrub was in progress on an old
 181                          * pool, and the pool was accessed by old
 182                          * software.  Restart from the beginning, since
 183                          * the old software may have changed the pool in
 184                          * the meantime.
 185                          */
 186                         scn->scn_restart_txg = txg;
 187                         zfs_dbgmsg("new-style scrub was modified "
 188                             "by old software; restarting in txg %llu",
 189                             scn->scn_restart_txg);
 190                 }
 191         }
 192
 193         spa_scan_stat_init(spa);
 194         return (0);
 195 }
 196
 197 void
 198 dsl_scan_fini(dsl_pool_t *dp)
 199 {
 200         if (dp->dp_scan) {
 201                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 202                 dp->dp_scan = NULL;
 203         }
 204 }
 205
 206 /* ARGSUSED */
 207 static int
 208 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 209 {
 210         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 211
 212         if (scn->scn_phys.scn_state == DSS_SCANNING)
 213                 return (SET_ERROR(EBUSY));
 214
 215         return (0);
 216 }
 217
 218 static void
 219 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 220 {
 221         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 222         pool_scan_func_t *funcp = arg;
 223         dmu_object_type_t ot = 0;
 224         dsl_pool_t *dp = scn->scn_dp;
 225         spa_t *spa = dp->dp_spa;
 226
 227         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 228         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 229         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 230         scn->scn_phys.scn_func = *funcp;
 231         scn->scn_phys.scn_state = DSS_SCANNING;
 232         scn->scn_phys.scn_min_txg = 0;
 233         scn->scn_phys.scn_max_txg = tx->tx_txg;
 234         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 235         scn->scn_phys.scn_start_time = gethrestime_sec();
 236         scn->scn_phys.scn_errors = 0;
 237         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 238         scn->scn_restart_txg = 0;
 239         scn->scn_done_txg = 0;
 240         spa_scan_stat_init(spa);
 241
 242         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 243                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 244
 245                 /* rewrite all disk labels */
 246                 vdev_config_dirty(spa->spa_root_vdev);
 247
 248                 if (vdev_resilver_needed(spa->spa_root_vdev,
 249                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 250                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 251                 } else {
 252                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 253                 }
 254
 255                 spa->spa_scrub_started = B_TRUE;
 256                 /*
 257                  * If this is an incremental scrub, limit the DDT scrub phase
 258                  * to just the auto-ditto class (for correctness); the rest
 259                  * of the scrub should go faster using top-down pruning.
 260                  */
 261                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 262                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 263
 264         }
 265
 266         /* back to the generic stuff */
 267
 268         if (dp->dp_blkstats == NULL) {
 269                 dp->dp_blkstats =
 270                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 271         }
 272         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 273
 274         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 275                 ot = DMU_OT_ZAP_OTHER;
 276
 277         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 278             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 279
 280         dsl_scan_sync_state(scn, tx);
 281
 282         spa_history_log_internal(spa, "scan setup", tx,
 283             "func=%u mintxg=%llu maxtxg=%llu",
 284             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 285 }
 286
 287 /* ARGSUSED */
 288 static void
 289 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 290 {
 291         static const char *old_names[] = {
 292                 "scrub_bookmark",
 293                 "scrub_ddt_bookmark",
 294                 "scrub_ddt_class_max",
 295                 "scrub_queue",
 296                 "scrub_min_txg",
 297                 "scrub_max_txg",
 298                 "scrub_func",
 299                 "scrub_errors",
 300                 NULL
 301         };
 302
 303         dsl_pool_t *dp = scn->scn_dp;
 304         spa_t *spa = dp->dp_spa;
 305         int i;
 306
 307         /* Remove any remnants of an old-style scrub. */
 308         for (i = 0; old_names[i]; i++) {
 309                 (void) zap_remove(dp->dp_meta_objset,
 310                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 311         }
 312
 313         if (scn->scn_phys.scn_queue_obj != 0) {
 314                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 315                     scn->scn_phys.scn_queue_obj, tx));
 316                 scn->scn_phys.scn_queue_obj = 0;
 317         }
 318
 319         /*
 320          * If we were "restarted" from a stopped state, don't bother
 321          * with anything else.
 322          */
 323         if (scn->scn_phys.scn_state != DSS_SCANNING)
 324                 return;
 325
 326         if (complete)
 327                 scn->scn_phys.scn_state = DSS_FINISHED;
 328         else
 329                 scn->scn_phys.scn_state = DSS_CANCELED;
 330
 331         spa_history_log_internal(spa, "scan done", tx,
 332             "complete=%u", complete);
 333
 334         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 335                 mutex_enter(&spa->spa_scrub_lock);
 336                 while (spa->spa_scrub_inflight > 0) {
 337                         cv_wait(&spa->spa_scrub_io_cv,
 338                             &spa->spa_scrub_lock);
 339                 }
 340                 mutex_exit(&spa->spa_scrub_lock);
 341                 spa->spa_scrub_started = B_FALSE;
 342                 spa->spa_scrub_active = B_FALSE;
 343
 344                 /*
 345                  * If the scrub/resilver completed, update all DTLs to
 346                  * reflect this.  Whether it succeeded or not, vacate
 347                  * all temporary scrub DTLs.
 348                  */
 349                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 350                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 351                 if (complete) {
 352                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 353                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 354                 }
 355                 spa_errlog_rotate(spa);
 356
 357                 /*
 358                  * We may have finished replacing a device.
 359                  * Let the async thread assess this and handle the detach.
 360                  */
 361                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 362         }
 363
 364         scn->scn_phys.scn_end_time = gethrestime_sec();
 365 }
 366
 367 /* ARGSUSED */
 368 static int
 369 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 370 {
 371         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 372
 373         if (scn->scn_phys.scn_state != DSS_SCANNING)
 374                 return (SET_ERROR(ENOENT));
 375         return (0);
 376 }
 377
 378 /* ARGSUSED */
 379 static void
 380 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 381 {
 382         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 383
 384         dsl_scan_done(scn, B_FALSE, tx);
 385         dsl_scan_sync_state(scn, tx);
 386 }
 387
 388 int
 389 dsl_scan_cancel(dsl_pool_t *dp)
 390 {
 391         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 392             dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
 393 }
 394
 395 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 396     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 397     dmu_objset_type_t ostype, dmu_tx_t *tx);
 398 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 399     dmu_objset_type_t ostype,
 400     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
 401
 402 void
 403 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 404 {
 405         zio_free(dp->dp_spa, txg, bp);
 406 }
 407
 408 void
 409 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 410 {
 411         ASSERT(dsl_pool_sync_context(dp));
 412         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 413             pio->io_flags));
 414 }
 415
 416 static uint64_t
 417 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 418 {
 419         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 420         if (ds->ds_is_snapshot)
 421                 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
 422         return (smt);
 423 }
 424
 425 static void
 426 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 427 {
 428         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 429             DMU_POOL_DIRECTORY_OBJECT,
 430             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 431             &scn->scn_phys, tx));
 432 }
 433
 434 extern int zfs_vdev_async_write_active_min_dirty_percent;
 435
 436 static boolean_t
 437 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
 438 {
 439         /* we never skip user/group accounting objects */
 440         if (zb && (int64_t)zb->zb_object < 0)
 441                 return (B_FALSE);
 442
 443         if (scn->scn_pausing)
 444                 return (B_TRUE); /* we're already pausing */
 445
 446         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 447                 return (B_FALSE); /* we're resuming */
 448
 449         /* We only know how to resume from level-0 blocks. */
 450         if (zb && zb->zb_level != 0)
 451                 return (B_FALSE);
 452
 453         /*
 454          * We pause if:
 455          *  - we have scanned for the maximum time: an entire txg
 456          *    timeout (default 5 sec)
 457          *  or
 458          *  - we have scanned for at least the minimum time (default 1 sec
 459          *    for scrub, 3 sec for resilver), and either we have sufficient
 460          *    dirty data that we are starting to write more quickly
 461          *    (default 30%), or someone is explicitly waiting for this txg
 462          *    to complete.
 463          *  or
 464          *  - the spa is shutting down because this pool is being exported
 465          *    or the machine is rebooting.
 466          */
 467         int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 468             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 469         uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 470         int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
 471         if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
 472             (NSEC2MSEC(elapsed_nanosecs) > mintime &&
 473             (txg_sync_waiting(scn->scn_dp) ||
 474             dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
 475             spa_shutting_down(scn->scn_dp->dp_spa)) {
 476                 if (zb) {
 477                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 478                             (longlong_t)zb->zb_objset,
 479                             (longlong_t)zb->zb_object,
 480                             (longlong_t)zb->zb_level,
 481                             (longlong_t)zb->zb_blkid);
 482                         scn->scn_phys.scn_bookmark = *zb;
 483                 }
 484                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 485                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 486                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 487                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 488                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 489                 scn->scn_pausing = B_TRUE;
 490                 return (B_TRUE);
 491         }
 492         return (B_FALSE);
 493 }
 494
 495 typedef struct zil_scan_arg {
 496         dsl_pool_t      *zsa_dp;
 497         zil_header_t    *zsa_zh;
 498 } zil_scan_arg_t;
 499
 500 /* ARGSUSED */
 501 static int
 502 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 503 {
 504         zil_scan_arg_t *zsa = arg;
 505         dsl_pool_t *dp = zsa->zsa_dp;
 506         dsl_scan_t *scn = dp->dp_scan;
 507         zil_header_t *zh = zsa->zsa_zh;
 508         zbookmark_phys_t zb;
 509
 510         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 511                 return (0);
 512
 513         /*
 514          * One block ("stubby") can be allocated a long time ago; we
 515          * want to visit that one because it has been allocated
 516          * (on-disk) even if it hasn't been claimed (even though for
 517          * scrub there's nothing to do to it).
 518          */
 519         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 520                 return (0);
 521
 522         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 523             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 524
 525         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 526         return (0);
 527 }
 528
 529 /* ARGSUSED */
 530 static int
 531 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 532 {
 533         if (lrc->lrc_txtype == TX_WRITE) {
 534                 zil_scan_arg_t *zsa = arg;
 535                 dsl_pool_t *dp = zsa->zsa_dp;
 536                 dsl_scan_t *scn = dp->dp_scan;
 537                 zil_header_t *zh = zsa->zsa_zh;
 538                 lr_write_t *lr = (lr_write_t *)lrc;
 539                 blkptr_t *bp = &lr->lr_blkptr;
 540                 zbookmark_phys_t zb;
 541
 542                 if (BP_IS_HOLE(bp) ||
 543                     bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 544                         return (0);
 545
 546                 /*
 547                  * birth can be < claim_txg if this record's txg is
 548                  * already txg sync'ed (but this log block contains
 549                  * other records that are not synced)
 550                  */
 551                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 552                         return (0);
 553
 554                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 555                     lr->lr_foid, ZB_ZIL_LEVEL,
 556                     lr->lr_offset / BP_GET_LSIZE(bp));
 557
 558                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 559         }
 560         return (0);
 561 }
 562
 563 static void
 564 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 565 {
 566         uint64_t claim_txg = zh->zh_claim_txg;
 567         zil_scan_arg_t zsa = { dp, zh };
 568         zilog_t *zilog;
 569
 570         /*
 571          * We only want to visit blocks that have been claimed but not yet
 572          * replayed (or, in read-only mode, blocks that *would* be claimed).
 573          */
 574         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 575                 return;
 576
 577         zilog = zil_alloc(dp->dp_meta_objset, zh);
 578
 579         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 580             claim_txg);
 581
 582         zil_free(zilog);
 583 }
 584
 585 /* ARGSUSED */
 586 static void
 587 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 588     uint64_t objset, uint64_t object, uint64_t blkid)
 589 {
 590         zbookmark_phys_t czb;
 591         arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
 592
 593         if (zfs_no_scrub_prefetch)
 594                 return;
 595
 596         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 597             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 598                 return;
 599
 600         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 601
 602         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 603             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 604             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 605 }
 606
 607 static boolean_t
 608 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 609     const zbookmark_phys_t *zb)
 610 {
 611         /*
 612          * We never skip over user/group accounting objects (obj<0)
 613          */
 614         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 615             (int64_t)zb->zb_object >= 0) {
 616                 /*
 617                  * If we already visited this bp & everything below (in
 618                  * a prior txg sync), don't bother doing it again.
 619                  */
 620                 if (zbookmark_subtree_completed(dnp, zb,
 621                     &scn->scn_phys.scn_bookmark))
 622                         return (B_TRUE);
 623
 624                 /*
 625                  * If we found the block we're trying to resume from, or
 626                  * we went past it to a different object, zero it out to
 627                  * indicate that it's OK to start checking for pausing
 628                  * again.
 629                  */
 630                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 631                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 632                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 633                             (longlong_t)zb->zb_objset,
 634                             (longlong_t)zb->zb_object,
 635                             (longlong_t)zb->zb_level,
 636                             (longlong_t)zb->zb_blkid);
 637                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 638                 }
 639         }
 640         return (B_FALSE);
 641 }
 642
 643 /*
 644  * Return nonzero on i/o error.
 645  * Return new buf to write out in *bufp.
 646  */
 647 static int
 648 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 649     dnode_phys_t *dnp, const blkptr_t *bp,
 650     const zbookmark_phys_t *zb, dmu_tx_t *tx)
 651 {
 652         dsl_pool_t *dp = scn->scn_dp;
 653         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 654         int err;
 655
 656         if (BP_GET_LEVEL(bp) > 0) {
 657                 arc_flags_t flags = ARC_FLAG_WAIT;
 658                 int i;
 659                 blkptr_t *cbp;
 660                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 661                 arc_buf_t *buf;
 662
 663                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 664                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 665                 if (err) {
 666                         scn->scn_phys.scn_errors++;
 667                         return (err);
 668                 }
 669                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 670                         dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
 671                             zb->zb_object, zb->zb_blkid * epb + i);
 672                 }
 673                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 674                         zbookmark_phys_t czb;
 675
 676                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 677                             zb->zb_level - 1,
 678                             zb->zb_blkid * epb + i);
 679                         dsl_scan_visitbp(cbp, &czb, dnp,
 680                             ds, scn, ostype, tx);
 681                 }
 682                 (void) arc_buf_remove_ref(buf, &buf);
 683         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 684                 arc_flags_t flags = ARC_FLAG_WAIT;
 685                 dnode_phys_t *cdnp;
 686                 int i, j;
 687                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 688                 arc_buf_t *buf;
 689
 690                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 691                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 692                 if (err) {
 693                         scn->scn_phys.scn_errors++;
 694                         return (err);
 695                 }
 696                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 697                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 698                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 699                                 dsl_scan_prefetch(scn, buf, cbp,
 700                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 701                         }
 702                 }
 703                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 704                         dsl_scan_visitdnode(scn, ds, ostype,
 705                             cdnp, zb->zb_blkid * epb + i, tx);
 706                 }
 707
 708                 (void) arc_buf_remove_ref(buf, &buf);
 709         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 710                 arc_flags_t flags = ARC_FLAG_WAIT;
 711                 objset_phys_t *osp;
 712                 arc_buf_t *buf;
 713
 714                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 715                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 716                 if (err) {
 717                         scn->scn_phys.scn_errors++;
 718                         return (err);
 719                 }
 720
 721                 osp = buf->b_data;
 722
 723                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 724                     &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
 725
 726                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
 727                         /*
 728                          * We also always visit user/group accounting
 729                          * objects, and never skip them, even if we are
 730                          * pausing.  This is necessary so that the space
 731                          * deltas from this txg get integrated.
 732                          */
 733                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 734                             &osp->os_groupused_dnode,
 735                             DMU_GROUPUSED_OBJECT, tx);
 736                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 737                             &osp->os_userused_dnode,
 738                             DMU_USERUSED_OBJECT, tx);
 739                 }
 740                 (void) arc_buf_remove_ref(buf, &buf);
 741         }
 742
 743         return (0);
 744 }
 745
 746 static void
 747 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 748     dmu_objset_type_t ostype, dnode_phys_t *dnp,
 749     uint64_t object, dmu_tx_t *tx)
 750 {
 751         int j;
 752
 753         for (j = 0; j < dnp->dn_nblkptr; j++) {
 754                 zbookmark_phys_t czb;
 755
 756                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 757                     dnp->dn_nlevels - 1, j);
 758                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 759                     &czb, dnp, ds, scn, ostype, tx);
 760         }
 761
 762         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 763                 zbookmark_phys_t czb;
 764                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 765                     0, DMU_SPILL_BLKID);
 766                 dsl_scan_visitbp(&dnp->dn_spill,
 767                     &czb, dnp, ds, scn, ostype, tx);
 768         }
 769 }
 770
 771 /*
 772  * The arguments are in this order because mdb can only print the
 773  * first 5; we want them to be useful.
 774  */
 775 static void
 776 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 777     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 778     dmu_objset_type_t ostype, dmu_tx_t *tx)
 779 {
 780         dsl_pool_t *dp = scn->scn_dp;
 781         arc_buf_t *buf = NULL;
 782         blkptr_t bp_toread = *bp;
 783
 784         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 785
 786         if (dsl_scan_check_pause(scn, zb))
 787                 return;
 788
 789         if (dsl_scan_check_resume(scn, dnp, zb))
 790                 return;
 791
 792         if (BP_IS_HOLE(bp))
 793                 return;
 794
 795         scn->scn_visited_this_txg++;
 796
 797         dprintf_bp(bp,
 798             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
 799             ds, ds ? ds->ds_object : 0,
 800             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 801             bp);
 802
 803         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 804                 return;
 805
 806         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
 807                 return;
 808
 809         /*
 810          * If dsl_scan_ddt() has aready visited this block, it will have
 811          * already done any translations or scrubbing, so don't call the
 812          * callback again.
 813          */
 814         if (ddt_class_contains(dp->dp_spa,
 815             scn->scn_phys.scn_ddt_class_max, bp)) {
 816                 ASSERT(buf == NULL);
 817                 return;
 818         }
 819
 820         /*
 821          * If this block is from the future (after cur_max_txg), then we
 822          * are doing this on behalf of a deleted snapshot, and we will
 823          * revisit the future block on the next pass of this dataset.
 824          * Don't scan it now unless we need to because something
 825          * under it was modified.
 826          */
 827         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 828                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 829         }
 830 }
 831
 832 static void
 833 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 834     dmu_tx_t *tx)
 835 {
 836         zbookmark_phys_t zb;
 837
 838         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 839             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 840         dsl_scan_visitbp(bp, &zb, NULL,
 841             ds, scn, DMU_OST_NONE, tx);
 842
 843         dprintf_ds(ds, "finished scan%s", "");
 844 }
 845
 846 void
 847 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 848 {
 849         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 850         dsl_scan_t *scn = dp->dp_scan;
 851         uint64_t mintxg;
 852
 853         if (scn->scn_phys.scn_state != DSS_SCANNING)
 854                 return;
 855
 856         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 857                 if (ds->ds_is_snapshot) {
 858                         /* Note, scn_cur_{min,max}_txg stays the same. */
 859                         scn->scn_phys.scn_bookmark.zb_objset =
 860                             dsl_dataset_phys(ds)->ds_next_snap_obj;
 861                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 862                             "reset zb_objset to %llu",
 863                             (u_longlong_t)ds->ds_object,
 864                             (u_longlong_t)dsl_dataset_phys(ds)->
 865                             ds_next_snap_obj);
 866                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 867                 } else {
 868                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 869                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 870                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 871                             "reset bookmark to -1,0,0,0",
 872                             (u_longlong_t)ds->ds_object);
 873                 }
 874         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 875             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 876                 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
 877                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 878                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 879                 if (ds->ds_is_snapshot) {
 880                         /*
 881                          * We keep the same mintxg; it could be >
 882                          * ds_creation_txg if the previous snapshot was
 883                          * deleted too.
 884                          */
 885                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 886                             scn->scn_phys.scn_queue_obj,
 887                             dsl_dataset_phys(ds)->ds_next_snap_obj,
 888                             mintxg, tx) == 0);
 889                         zfs_dbgmsg("destroying ds %llu; in queue; "
 890                             "replacing with %llu",
 891                             (u_longlong_t)ds->ds_object,
 892                             (u_longlong_t)dsl_dataset_phys(ds)->
 893                             ds_next_snap_obj);
 894                 } else {
 895                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 896                             (u_longlong_t)ds->ds_object);
 897                 }
 898         } else {
 899                 zfs_dbgmsg("destroying ds %llu; ignoring",
 900                     (u_longlong_t)ds->ds_object);
 901         }
 902
 903         /*
 904          * dsl_scan_sync() should be called after this, and should sync
 905          * out our changed state, but just to be safe, do it here.
 906          */
 907         dsl_scan_sync_state(scn, tx);
 908 }
 909
 910 void
 911 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 912 {
 913         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 914         dsl_scan_t *scn = dp->dp_scan;
 915         uint64_t mintxg;
 916
 917         if (scn->scn_phys.scn_state != DSS_SCANNING)
 918                 return;
 919
 920         ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
 921
 922         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 923                 scn->scn_phys.scn_bookmark.zb_objset =
 924                     dsl_dataset_phys(ds)->ds_prev_snap_obj;
 925                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 926                     "reset zb_objset to %llu",
 927                     (u_longlong_t)ds->ds_object,
 928                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 929         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 930             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 931                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 932                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 933                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 934                     scn->scn_phys.scn_queue_obj,
 935                     dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
 936                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 937                     "replacing with %llu",
 938                     (u_longlong_t)ds->ds_object,
 939                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 940         }
 941         dsl_scan_sync_state(scn, tx);
 942 }
 943
 944 void
 945 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 946 {
 947         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 948         dsl_scan_t *scn = dp->dp_scan;
 949         uint64_t mintxg;
 950
 951         if (scn->scn_phys.scn_state != DSS_SCANNING)
 952                 return;
 953
 954         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 955                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 956                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 957                     "reset zb_objset to %llu",
 958                     (u_longlong_t)ds1->ds_object,
 959                     (u_longlong_t)ds2->ds_object);
 960         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 961                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 962                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 963                     "reset zb_objset to %llu",
 964                     (u_longlong_t)ds2->ds_object,
 965                     (u_longlong_t)ds1->ds_object);
 966         }
 967
 968         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 969             ds1->ds_object, &mintxg) == 0) {
 970                 int err;
 971
 972                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 973                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 974                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 975                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 976                 err = zap_add_int_key(dp->dp_meta_objset,
 977                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 978                 VERIFY(err == 0 || err == EEXIST);
 979                 if (err == EEXIST) {
 980                         /* Both were there to begin with */
 981                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 982                             scn->scn_phys.scn_queue_obj,
 983                             ds1->ds_object, mintxg, tx));
 984                 }
 985                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 986                     "replacing with %llu",
 987                     (u_longlong_t)ds1->ds_object,
 988                     (u_longlong_t)ds2->ds_object);
 989         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 990             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 991                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 992                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 993                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 994                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 995                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 996                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 997                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 998                     "replacing with %llu",
 999                     (u_longlong_t)ds2->ds_object,
1000                     (u_longlong_t)ds1->ds_object);
1001         }
1002
1003         dsl_scan_sync_state(scn, tx);
1004 }
1005
1006 struct enqueue_clones_arg {
1007         dmu_tx_t *tx;
1008         uint64_t originobj;
1009 };
1010
1011 /* ARGSUSED */
1012 static int
1013 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1014 {
1015         struct enqueue_clones_arg *eca = arg;
1016         dsl_dataset_t *ds;
1017         int err;
1018         dsl_scan_t *scn = dp->dp_scan;
1019
1020         if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1021                 return (0);
1022
1023         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1024         if (err)
1025                 return (err);
1026
1027         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1028                 dsl_dataset_t *prev;
1029                 err = dsl_dataset_hold_obj(dp,
1030                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1031
1032                 dsl_dataset_rele(ds, FTAG);
1033                 if (err)
1034                         return (err);
1035                 ds = prev;
1036         }
1037         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1038             scn->scn_phys.scn_queue_obj, ds->ds_object,
1039             dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1040         dsl_dataset_rele(ds, FTAG);
1041         return (0);
1042 }
1043
1044 static void
1045 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1046 {
1047         dsl_pool_t *dp = scn->scn_dp;
1048         dsl_dataset_t *ds;
1049         objset_t *os;
1050
1051         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1052
1053         if (dmu_objset_from_ds(ds, &os))
1054                 goto out;
1055
1056         /*
1057          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1058          * snapshots can have ZIL block pointers (which may be the same
1059          * BP as in the head), they must be ignored.  So we traverse the
1060          * ZIL here, rather than in scan_recurse(), because the regular
1061          * snapshot block-sharing rules don't apply to it.
1062          */
1063         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1064                 dsl_scan_zil(dp, &os->os_zil_header);
1065
1066         /*
1067          * Iterate over the bps in this ds.
1068          */
1069         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1070         dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1071
1072         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1073         dsl_dataset_name(ds, dsname);
1074         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1075             "pausing=%u",
1076             (longlong_t)dsobj, dsname,
1077             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1078             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1079             (int)scn->scn_pausing);
1080         kmem_free(dsname, ZFS_MAXNAMELEN);
1081
1082         if (scn->scn_pausing)
1083                 goto out;
1084
1085         /*
1086          * We've finished this pass over this dataset.
1087          */
1088
1089         /*
1090          * If we did not completely visit this dataset, do another pass.
1091          */
1092         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1093                 zfs_dbgmsg("incomplete pass; visiting again");
1094                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1095                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1096                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1097                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1098                 goto out;
1099         }
1100
1101         /*
1102          * Add descendent datasets to work queue.
1103          */
1104         if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1105                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1106                     scn->scn_phys.scn_queue_obj,
1107                     dsl_dataset_phys(ds)->ds_next_snap_obj,
1108                     dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1109         }
1110         if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1111                 boolean_t usenext = B_FALSE;
1112                 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1113                         uint64_t count;
1114                         /*
1115                          * A bug in a previous version of the code could
1116                          * cause upgrade_clones_cb() to not set
1117                          * ds_next_snap_obj when it should, leading to a
1118                          * missing entry.  Therefore we can only use the
1119                          * next_clones_obj when its count is correct.
1120                          */
1121                         int err = zap_count(dp->dp_meta_objset,
1122                             dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1123                         if (err == 0 &&
1124                             count == dsl_dataset_phys(ds)->ds_num_children - 1)
1125                                 usenext = B_TRUE;
1126                 }
1127
1128                 if (usenext) {
1129                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1130                             dsl_dataset_phys(ds)->ds_next_clones_obj,
1131                             scn->scn_phys.scn_queue_obj,
1132                             dsl_dataset_phys(ds)->ds_creation_txg, tx));
1133                 } else {
1134                         struct enqueue_clones_arg eca;
1135                         eca.tx = tx;
1136                         eca.originobj = ds->ds_object;
1137
1138                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1139                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1140                 }
1141         }
1142
1143 out:
1144         dsl_dataset_rele(ds, FTAG);
1145 }
1146
1147 /* ARGSUSED */
1148 static int
1149 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1150 {
1151         dmu_tx_t *tx = arg;
1152         dsl_dataset_t *ds;
1153         int err;
1154         dsl_scan_t *scn = dp->dp_scan;
1155
1156         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1157         if (err)
1158                 return (err);
1159
1160         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1161                 dsl_dataset_t *prev;
1162                 err = dsl_dataset_hold_obj(dp,
1163                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1164                 if (err) {
1165                         dsl_dataset_rele(ds, FTAG);
1166                         return (err);
1167                 }
1168
1169                 /*
1170                  * If this is a clone, we don't need to worry about it for now.
1171                  */
1172                 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1173                         dsl_dataset_rele(ds, FTAG);
1174                         dsl_dataset_rele(prev, FTAG);
1175                         return (0);
1176                 }
1177                 dsl_dataset_rele(ds, FTAG);
1178                 ds = prev;
1179         }
1180
1181         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1182             ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1183         dsl_dataset_rele(ds, FTAG);
1184         return (0);
1185 }
1186
1187 /*
1188  * Scrub/dedup interaction.
1189  *
1190  * If there are N references to a deduped block, we don't want to scrub it
1191  * N times -- ideally, we should scrub it exactly once.
1192  *
1193  * We leverage the fact that the dde's replication class (enum ddt_class)
1194  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1195  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1196  *
1197  * To prevent excess scrubbing, the scrub begins by walking the DDT
1198  * to find all blocks with refcnt > 1, and scrubs each of these once.
1199  * Since there are two replication classes which contain blocks with
1200  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1201  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1202  *
1203  * There would be nothing more to say if a block's refcnt couldn't change
1204  * during a scrub, but of course it can so we must account for changes
1205  * in a block's replication class.
1206  *
1207  * Here's an example of what can occur:
1208  *
1209  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1210  * when visited during the top-down scrub phase, it will be scrubbed twice.
1211  * This negates our scrub optimization, but is otherwise harmless.
1212  *
1213  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1214  * on each visit during the top-down scrub phase, it will never be scrubbed.
1215  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1216  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1217  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1218  * while a scrub is in progress, it scrubs the block right then.
1219  */
1220 static void
1221 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1222 {
1223         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1224         ddt_entry_t dde = { 0 };
1225         int error;
1226         uint64_t n = 0;
1227
1228         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1229                 ddt_t *ddt;
1230
1231                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1232                         break;
1233                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1234                     (longlong_t)ddb->ddb_class,
1235                     (longlong_t)ddb->ddb_type,
1236                     (longlong_t)ddb->ddb_checksum,
1237                     (longlong_t)ddb->ddb_cursor);
1238
1239                 /* There should be no pending changes to the dedup table */
1240                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1241                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1242
1243                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1244                 n++;
1245
1246                 if (dsl_scan_check_pause(scn, NULL))
1247                         break;
1248         }
1249
1250         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1251             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1252             (int)scn->scn_pausing);
1253
1254         ASSERT(error == 0 || error == ENOENT);
1255         ASSERT(error != ENOENT ||
1256             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1257 }
1258
1259 /* ARGSUSED */
1260 void
1261 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1262     ddt_entry_t *dde, dmu_tx_t *tx)
1263 {
1264         const ddt_key_t *ddk = &dde->dde_key;
1265         ddt_phys_t *ddp = dde->dde_phys;
1266         blkptr_t bp;
1267         zbookmark_phys_t zb = { 0 };
1268
1269         if (scn->scn_phys.scn_state != DSS_SCANNING)
1270                 return;
1271
1272         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1273                 if (ddp->ddp_phys_birth == 0 ||
1274                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1275                         continue;
1276                 ddt_bp_create(checksum, ddk, ddp, &bp);
1277
1278                 scn->scn_visited_this_txg++;
1279                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1280         }
1281 }
1282
1283 static void
1284 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1285 {
1286         dsl_pool_t *dp = scn->scn_dp;
1287         zap_cursor_t zc;
1288         zap_attribute_t za;
1289
1290         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1291             scn->scn_phys.scn_ddt_class_max) {
1292                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1293                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1294                 dsl_scan_ddt(scn, tx);
1295                 if (scn->scn_pausing)
1296                         return;
1297         }
1298
1299         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1300                 /* First do the MOS & ORIGIN */
1301
1302                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1303                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1304                 dsl_scan_visit_rootbp(scn, NULL,
1305                     &dp->dp_meta_rootbp, tx);
1306                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1307                 if (scn->scn_pausing)
1308                         return;
1309
1310                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1311                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1312                             enqueue_cb, tx, DS_FIND_CHILDREN));
1313                 } else {
1314                         dsl_scan_visitds(scn,
1315                             dp->dp_origin_snap->ds_object, tx);
1316                 }
1317                 ASSERT(!scn->scn_pausing);
1318         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1319             ZB_DESTROYED_OBJSET) {
1320                 /*
1321                  * If we were paused, continue from here.  Note if the
1322                  * ds we were paused on was deleted, the zb_objset may
1323                  * be -1, so we will skip this and find a new objset
1324                  * below.
1325                  */
1326                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1327                 if (scn->scn_pausing)
1328                         return;
1329         }
1330
1331         /*
1332          * In case we were paused right at the end of the ds, zero the
1333          * bookmark so we don't think that we're still trying to resume.
1334          */
1335         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1336
1337         /* keep pulling things out of the zap-object-as-queue */
1338         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1339             scn->scn_phys.scn_queue_obj),
1340             zap_cursor_retrieve(&zc, &za) == 0) {
1341                 dsl_dataset_t *ds;
1342                 uint64_t dsobj;
1343
1344                 dsobj = strtonum(za.za_name, NULL);
1345                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1346                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1347
1348                 /* Set up min/max txg */
1349                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1350                 if (za.za_first_integer != 0) {
1351                         scn->scn_phys.scn_cur_min_txg =
1352                             MAX(scn->scn_phys.scn_min_txg,
1353                             za.za_first_integer);
1354                 } else {
1355                         scn->scn_phys.scn_cur_min_txg =
1356                             MAX(scn->scn_phys.scn_min_txg,
1357                             dsl_dataset_phys(ds)->ds_prev_snap_txg);
1358                 }
1359                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1360                 dsl_dataset_rele(ds, FTAG);
1361
1362                 dsl_scan_visitds(scn, dsobj, tx);
1363                 zap_cursor_fini(&zc);
1364                 if (scn->scn_pausing)
1365                         return;
1366         }
1367         zap_cursor_fini(&zc);
1368 }
1369
1370 static boolean_t
1371 dsl_scan_free_should_pause(dsl_scan_t *scn)
1372 {
1373         uint64_t elapsed_nanosecs;
1374
1375         if (zfs_recover)
1376                 return (B_FALSE);
1377
1378         if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1379                 return (B_TRUE);
1380
1381         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1382         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1383             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1384             txg_sync_waiting(scn->scn_dp)) ||
1385             spa_shutting_down(scn->scn_dp->dp_spa));
1386 }
1387
1388 static int
1389 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1390 {
1391         dsl_scan_t *scn = arg;
1392
1393         if (!scn->scn_is_bptree ||
1394             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1395                 if (dsl_scan_free_should_pause(scn))
1396                         return (SET_ERROR(ERESTART));
1397         }
1398
1399         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1400             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1401         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1402             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1403             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1404         scn->scn_visited_this_txg++;
1405         return (0);
1406 }
1407
1408 boolean_t
1409 dsl_scan_active(dsl_scan_t *scn)
1410 {
1411         spa_t *spa = scn->scn_dp->dp_spa;
1412         uint64_t used = 0, comp, uncomp;
1413
1414         if (spa->spa_load_state != SPA_LOAD_NONE)
1415                 return (B_FALSE);
1416         if (spa_shutting_down(spa))
1417                 return (B_FALSE);
1418         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1419             (scn->scn_async_destroying && !scn->scn_async_stalled))
1420                 return (B_TRUE);
1421
1422         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1423                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1424                     &used, &comp, &uncomp);
1425         }
1426         return (used != 0);
1427 }
1428
1429 void
1430 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1431 {
1432         dsl_scan_t *scn = dp->dp_scan;
1433         spa_t *spa = dp->dp_spa;
1434         int err = 0;
1435
1436         /*
1437          * Check for scn_restart_txg before checking spa_load_state, so
1438          * that we can restart an old-style scan while the pool is being
1439          * imported (see dsl_scan_init).
1440          */
1441         if (scn->scn_restart_txg != 0 &&
1442             scn->scn_restart_txg <= tx->tx_txg) {
1443                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1444                 dsl_scan_done(scn, B_FALSE, tx);
1445                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1446                         func = POOL_SCAN_RESILVER;
1447                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1448                     func, tx->tx_txg);
1449                 dsl_scan_setup_sync(&func, tx);
1450         }
1451
1452         /*
1453          * If the scan is inactive due to a stalled async destroy, try again.
1454          */
1455         if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1456             spa_sync_pass(dp->dp_spa) > 1)
1457                 return;
1458
1459         scn->scn_visited_this_txg = 0;
1460         scn->scn_pausing = B_FALSE;
1461         scn->scn_sync_start_time = gethrtime();
1462         spa->spa_scrub_active = B_TRUE;
1463
1464         /*
1465          * First process the async destroys.  If we pause, don't do
1466          * any scrubbing or resilvering.  This ensures that there are no
1467          * async destroys while we are scanning, so the scan code doesn't
1468          * have to worry about traversing it.  It is also faster to free the
1469          * blocks than to scrub them.
1470          */
1471         if (zfs_free_bpobj_enabled &&
1472             spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1473                 scn->scn_is_bptree = B_FALSE;
1474                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1475                     NULL, ZIO_FLAG_MUSTSUCCEED);
1476                 err = bpobj_iterate(&dp->dp_free_bpobj,
1477                     dsl_scan_free_block_cb, scn, tx);
1478                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1479
1480                 if (err != 0 && err != ERESTART)
1481                         zfs_panic_recover("error %u from bpobj_iterate()", err);
1482         }
1483
1484         if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1485                 ASSERT(scn->scn_async_destroying);
1486                 scn->scn_is_bptree = B_TRUE;
1487                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1488                     NULL, ZIO_FLAG_MUSTSUCCEED);
1489                 err = bptree_iterate(dp->dp_meta_objset,
1490                     dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1491                 VERIFY0(zio_wait(scn->scn_zio_root));
1492
1493                 if (err == EIO || err == ECKSUM) {
1494                         err = 0;
1495                 } else if (err != 0 && err != ERESTART) {
1496                         zfs_panic_recover("error %u from "
1497                             "traverse_dataset_destroyed()", err);
1498                 }
1499
1500                 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1501                         /* finished; deactivate async destroy feature */
1502                         spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1503                         ASSERT(!spa_feature_is_active(spa,
1504                             SPA_FEATURE_ASYNC_DESTROY));
1505                         VERIFY0(zap_remove(dp->dp_meta_objset,
1506                             DMU_POOL_DIRECTORY_OBJECT,
1507                             DMU_POOL_BPTREE_OBJ, tx));
1508                         VERIFY0(bptree_free(dp->dp_meta_objset,
1509                             dp->dp_bptree_obj, tx));
1510                         dp->dp_bptree_obj = 0;
1511                         scn->scn_async_destroying = B_FALSE;
1512                         scn->scn_async_stalled = B_FALSE;
1513                 } else {
1514                         /*
1515                          * If we didn't make progress, mark the async
1516                          * destroy as stalled, so that we will not initiate
1517                          * a spa_sync() on its behalf.  Note that we only
1518                          * check this if we are not finished, because if the
1519                          * bptree had no blocks for us to visit, we can
1520                          * finish without "making progress".
1521                          */
1522                         scn->scn_async_stalled =
1523                             (scn->scn_visited_this_txg == 0);
1524                 }
1525         }
1526         if (scn->scn_visited_this_txg) {
1527                 zfs_dbgmsg("freed %llu blocks in %llums from "
1528                     "free_bpobj/bptree txg %llu; err=%d",
1529                     (longlong_t)scn->scn_visited_this_txg,
1530                     (longlong_t)
1531                     NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1532                     (longlong_t)tx->tx_txg, err);
1533                 scn->scn_visited_this_txg = 0;
1534
1535                 /*
1536                  * Write out changes to the DDT that may be required as a
1537                  * result of the blocks freed.  This ensures that the DDT
1538                  * is clean when a scrub/resilver runs.
1539                  */
1540                 ddt_sync(spa, tx->tx_txg);
1541         }
1542         if (err != 0)
1543                 return;
1544         if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1545             (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1546             dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1547             dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1548                 /*
1549                  * We have finished background destroying, but there is still
1550                  * some space left in the dp_free_dir. Transfer this leaked
1551                  * space to the dp_leak_dir.
1552                  */
1553                 if (dp->dp_leak_dir == NULL) {
1554                         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1555                         (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1556                             LEAK_DIR_NAME, tx);
1557                         VERIFY0(dsl_pool_open_special_dir(dp,
1558                             LEAK_DIR_NAME, &dp->dp_leak_dir));
1559                         rrw_exit(&dp->dp_config_rwlock, FTAG);
1560                 }
1561                 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1562                     dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1563                     dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1564                     dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1565                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1566                     -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1567                     -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1568                     -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1569         }
1570         if (!scn->scn_async_destroying) {
1571                 /* finished; verify that space accounting went to zero */
1572                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1573                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1574                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1575         }
1576
1577         if (scn->scn_phys.scn_state != DSS_SCANNING)
1578                 return;
1579
1580         if (scn->scn_done_txg == tx->tx_txg) {
1581                 ASSERT(!scn->scn_pausing);
1582                 /* finished with scan. */
1583                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1584                 dsl_scan_done(scn, B_TRUE, tx);
1585                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1586                 dsl_scan_sync_state(scn, tx);
1587                 return;
1588         }
1589
1590         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1591             scn->scn_phys.scn_ddt_class_max) {
1592                 zfs_dbgmsg("doing scan sync txg %llu; "
1593                     "ddt bm=%llu/%llu/%llu/%llx",
1594                     (longlong_t)tx->tx_txg,
1595                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1596                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1597                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1598                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1599                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1600                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1601                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1602                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1603         } else {
1604                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1605                     (longlong_t)tx->tx_txg,
1606                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1607                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1608                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1609                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1610         }
1611
1612         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1613             NULL, ZIO_FLAG_CANFAIL);
1614         dsl_pool_config_enter(dp, FTAG);
1615         dsl_scan_visit(scn, tx);
1616         dsl_pool_config_exit(dp, FTAG);
1617         (void) zio_wait(scn->scn_zio_root);
1618         scn->scn_zio_root = NULL;
1619
1620         zfs_dbgmsg("visited %llu blocks in %llums",
1621             (longlong_t)scn->scn_visited_this_txg,
1622             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1623
1624         if (!scn->scn_pausing) {
1625                 scn->scn_done_txg = tx->tx_txg + 1;
1626                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1627                     tx->tx_txg, scn->scn_done_txg);
1628         }
1629
1630         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1631                 mutex_enter(&spa->spa_scrub_lock);
1632                 while (spa->spa_scrub_inflight > 0) {
1633                         cv_wait(&spa->spa_scrub_io_cv,
1634                             &spa->spa_scrub_lock);
1635                 }
1636                 mutex_exit(&spa->spa_scrub_lock);
1637         }
1638
1639         dsl_scan_sync_state(scn, tx);
1640 }
1641
1642 /*
1643  * This will start a new scan, or restart an existing one.
1644  */
1645 void
1646 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1647 {
1648         if (txg == 0) {
1649                 dmu_tx_t *tx;
1650                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1651                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1652
1653                 txg = dmu_tx_get_txg(tx);
1654                 dp->dp_scan->scn_restart_txg = txg;
1655                 dmu_tx_commit(tx);
1656         } else {
1657                 dp->dp_scan->scn_restart_txg = txg;
1658         }
1659         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1660 }
1661
1662 boolean_t
1663 dsl_scan_resilvering(dsl_pool_t *dp)
1664 {
1665         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1666             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1667 }
1668
1669 /*
1670  * scrub consumers
1671  */
1672
1673 static void
1674 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1675 {
1676         int i;
1677
1678         /*
1679          * If we resume after a reboot, zab will be NULL; don't record
1680          * incomplete stats in that case.
1681          */
1682         if (zab == NULL)
1683                 return;
1684
1685         for (i = 0; i < 4; i++) {
1686                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1687                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1688                 if (t & DMU_OT_NEWTYPE)
1689                         t = DMU_OT_OTHER;
1690                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1691                 int equal;
1692
1693                 zb->zb_count++;
1694                 zb->zb_asize += BP_GET_ASIZE(bp);
1695                 zb->zb_lsize += BP_GET_LSIZE(bp);
1696                 zb->zb_psize += BP_GET_PSIZE(bp);
1697                 zb->zb_gangs += BP_COUNT_GANG(bp);
1698
1699                 switch (BP_GET_NDVAS(bp)) {
1700                 case 2:
1701                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1702                             DVA_GET_VDEV(&bp->blk_dva[1]))
1703                                 zb->zb_ditto_2_of_2_samevdev++;
1704                         break;
1705                 case 3:
1706                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1707                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1708                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1709                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1710                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1711                             DVA_GET_VDEV(&bp->blk_dva[2]));
1712                         if (equal == 1)
1713                                 zb->zb_ditto_2_of_3_samevdev++;
1714                         else if (equal == 3)
1715                                 zb->zb_ditto_3_of_3_samevdev++;
1716                         break;
1717                 }
1718         }
1719 }
1720
1721 static void
1722 dsl_scan_scrub_done(zio_t *zio)
1723 {
1724         spa_t *spa = zio->io_spa;
1725
1726         zio_data_buf_free(zio->io_data, zio->io_size);
1727
1728         mutex_enter(&spa->spa_scrub_lock);
1729         spa->spa_scrub_inflight--;
1730         cv_broadcast(&spa->spa_scrub_io_cv);
1731
1732         if (zio->io_error && (zio->io_error != ECKSUM ||
1733             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1734                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1735         }
1736         mutex_exit(&spa->spa_scrub_lock);
1737 }
1738
1739 static int
1740 dsl_scan_scrub_cb(dsl_pool_t *dp,
1741     const blkptr_t *bp, const zbookmark_phys_t *zb)
1742 {
1743         dsl_scan_t *scn = dp->dp_scan;
1744         size_t size = BP_GET_PSIZE(bp);
1745         spa_t *spa = dp->dp_spa;
1746         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1747         boolean_t needs_io;
1748         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1749         unsigned int scan_delay = 0;
1750
1751         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1752             phys_birth >= scn->scn_phys.scn_max_txg)
1753                 return (0);
1754
1755         count_block(dp->dp_blkstats, bp);
1756
1757         if (BP_IS_EMBEDDED(bp))
1758                 return (0);
1759
1760         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1761         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1762                 zio_flags |= ZIO_FLAG_SCRUB;
1763                 needs_io = B_TRUE;
1764                 scan_delay = zfs_scrub_delay;
1765         } else {
1766                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1767                 zio_flags |= ZIO_FLAG_RESILVER;
1768                 needs_io = B_FALSE;
1769                 scan_delay = zfs_resilver_delay;
1770         }
1771
1772         /* If it's an intent log block, failure is expected. */
1773         if (zb->zb_level == ZB_ZIL_LEVEL)
1774                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1775
1776         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1777                 vdev_t *vd = vdev_lookup_top(spa,
1778                     DVA_GET_VDEV(&bp->blk_dva[d]));
1779
1780                 /*
1781                  * Keep track of how much data we've examined so that
1782                  * zpool(1M) status can make useful progress reports.
1783                  */
1784                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1785                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1786
1787                 /* if it's a resilver, this may not be in the target range */
1788                 if (!needs_io) {
1789                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1790                                 /*
1791                                  * Gang members may be spread across multiple
1792                                  * vdevs, so the best estimate we have is the
1793                                  * scrub range, which has already been checked.
1794                                  * XXX -- it would be better to change our
1795                                  * allocation policy to ensure that all
1796                                  * gang members reside on the same vdev.
1797                                  */
1798                                 needs_io = B_TRUE;
1799                         } else {
1800                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1801                                     phys_birth, 1);
1802                         }
1803                 }
1804         }
1805
1806         if (needs_io && !zfs_no_scrub_io) {
1807                 vdev_t *rvd = spa->spa_root_vdev;
1808                 uint64_t maxinflight = rvd->vdev_children *
1809                     MAX(zfs_top_maxinflight, 1);
1810                 void *data = zio_data_buf_alloc(size);
1811
1812                 mutex_enter(&spa->spa_scrub_lock);
1813                 while (spa->spa_scrub_inflight >= maxinflight)
1814                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1815                 spa->spa_scrub_inflight++;
1816                 mutex_exit(&spa->spa_scrub_lock);
1817
1818                 /*
1819                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1820                  * then throttle our workload to limit the impact of a scan.
1821                  */
1822                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1823                         delay(MAX((int)scan_delay, 0));
1824
1825                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1826                     dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1827                     zio_flags, zb));
1828         }
1829
1830         /* do not relocate this block */
1831         return (0);
1832 }
1833
1834 int
1835 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1836 {
1837         spa_t *spa = dp->dp_spa;
1838
1839         /*
1840          * Purge all vdev caches and probe all devices.  We do this here
1841          * rather than in sync context because this requires a writer lock
1842          * on the spa_config lock, which we can't do from sync context.  The
1843          * spa_scrub_reopen flag indicates that vdev_open() should not
1844          * attempt to start another scrub.
1845          */
1846         spa_vdev_state_enter(spa, SCL_NONE);
1847         spa->spa_scrub_reopen = B_TRUE;
1848         vdev_reopen(spa->spa_root_vdev);
1849         spa->spa_scrub_reopen = B_FALSE;
1850         (void) spa_vdev_state_exit(spa, NULL, 0);
1851
1852         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1853             dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1854 }