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, 2014 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 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RWTUN,
  74     &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
  75 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RWTUN,
  76     &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
  77 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RWTUN,
  78     &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
  79 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RWTUN,
  80     &zfs_scan_idle, 0, "Idle scan window in clock ticks");
  81 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RWTUN,
  82     &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
  83 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RWTUN,
  84     &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
  85 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RWTUN,
  86     &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
  87 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RWTUN,
  88     &zfs_no_scrub_io, 0, "Disable scrub I/O");
  89 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RWTUN,
  90     &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
  91
  92 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
  93 /* max number of blocks to free in a single TXG */
  94 uint64_t zfs_free_max_blocks = UINT64_MAX;
  95 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
  96     &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
  97
  98
  99 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
 100         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
 101         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
 102
 103 extern int zfs_txg_timeout;
 104
 105 /* the order has to match pool_scan_type */
 106 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
 107         NULL,
 108         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
 109         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
 110 };
 111
 112 int
 113 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
 114 {
 115         int err;
 116         dsl_scan_t *scn;
 117         spa_t *spa = dp->dp_spa;
 118         uint64_t f;
 119
 120         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
 121         scn->scn_dp = dp;
 122
 123         /*
 124          * It's possible that we're resuming a scan after a reboot so
 125          * make sure that the scan_async_destroying flag is initialized
 126          * appropriately.
 127          */
 128         ASSERT(!scn->scn_async_destroying);
 129         scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
 130             SPA_FEATURE_ASYNC_DESTROY);
 131
 132         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 133             "scrub_func", sizeof (uint64_t), 1, &f);
 134         if (err == 0) {
 135                 /*
 136                  * There was an old-style scrub in progress.  Restart a
 137                  * new-style scrub from the beginning.
 138                  */
 139                 scn->scn_restart_txg = txg;
 140                 zfs_dbgmsg("old-style scrub was in progress; "
 141                     "restarting new-style scrub in txg %llu",
 142                     scn->scn_restart_txg);
 143
 144                 /*
 145                  * Load the queue obj from the old location so that it
 146                  * can be freed by dsl_scan_done().
 147                  */
 148                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 149                     "scrub_queue", sizeof (uint64_t), 1,
 150                     &scn->scn_phys.scn_queue_obj);
 151         } else {
 152                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 153                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 154                     &scn->scn_phys);
 155                 if (err == ENOENT)
 156                         return (0);
 157                 else if (err)
 158                         return (err);
 159
 160                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 161                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 162                         /*
 163                          * A new-type scrub was in progress on an old
 164                          * pool, and the pool was accessed by old
 165                          * software.  Restart from the beginning, since
 166                          * the old software may have changed the pool in
 167                          * the meantime.
 168                          */
 169                         scn->scn_restart_txg = txg;
 170                         zfs_dbgmsg("new-style scrub was modified "
 171                             "by old software; restarting in txg %llu",
 172                             scn->scn_restart_txg);
 173                 }
 174         }
 175
 176         spa_scan_stat_init(spa);
 177         return (0);
 178 }
 179
 180 void
 181 dsl_scan_fini(dsl_pool_t *dp)
 182 {
 183         if (dp->dp_scan) {
 184                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 185                 dp->dp_scan = NULL;
 186         }
 187 }
 188
 189 /* ARGSUSED */
 190 static int
 191 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 192 {
 193         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 194
 195         if (scn->scn_phys.scn_state == DSS_SCANNING)
 196                 return (SET_ERROR(EBUSY));
 197
 198         return (0);
 199 }
 200
 201 static void
 202 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 203 {
 204         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 205         pool_scan_func_t *funcp = arg;
 206         dmu_object_type_t ot = 0;
 207         dsl_pool_t *dp = scn->scn_dp;
 208         spa_t *spa = dp->dp_spa;
 209
 210         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 211         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 212         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 213         scn->scn_phys.scn_func = *funcp;
 214         scn->scn_phys.scn_state = DSS_SCANNING;
 215         scn->scn_phys.scn_min_txg = 0;
 216         scn->scn_phys.scn_max_txg = tx->tx_txg;
 217         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 218         scn->scn_phys.scn_start_time = gethrestime_sec();
 219         scn->scn_phys.scn_errors = 0;
 220         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 221         scn->scn_restart_txg = 0;
 222         scn->scn_done_txg = 0;
 223         spa_scan_stat_init(spa);
 224
 225         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 226                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 227
 228                 /* rewrite all disk labels */
 229                 vdev_config_dirty(spa->spa_root_vdev);
 230
 231                 if (vdev_resilver_needed(spa->spa_root_vdev,
 232                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 233                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 234                 } else {
 235                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 236                 }
 237
 238                 spa->spa_scrub_started = B_TRUE;
 239                 /*
 240                  * If this is an incremental scrub, limit the DDT scrub phase
 241                  * to just the auto-ditto class (for correctness); the rest
 242                  * of the scrub should go faster using top-down pruning.
 243                  */
 244                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 245                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 246
 247         }
 248
 249         /* back to the generic stuff */
 250
 251         if (dp->dp_blkstats == NULL) {
 252                 dp->dp_blkstats =
 253                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 254         }
 255         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 256
 257         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 258                 ot = DMU_OT_ZAP_OTHER;
 259
 260         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 261             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 262
 263         dsl_scan_sync_state(scn, tx);
 264
 265         spa_history_log_internal(spa, "scan setup", tx,
 266             "func=%u mintxg=%llu maxtxg=%llu",
 267             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 268 }
 269
 270 /* ARGSUSED */
 271 static void
 272 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 273 {
 274         static const char *old_names[] = {
 275                 "scrub_bookmark",
 276                 "scrub_ddt_bookmark",
 277                 "scrub_ddt_class_max",
 278                 "scrub_queue",
 279                 "scrub_min_txg",
 280                 "scrub_max_txg",
 281                 "scrub_func",
 282                 "scrub_errors",
 283                 NULL
 284         };
 285
 286         dsl_pool_t *dp = scn->scn_dp;
 287         spa_t *spa = dp->dp_spa;
 288         int i;
 289
 290         /* Remove any remnants of an old-style scrub. */
 291         for (i = 0; old_names[i]; i++) {
 292                 (void) zap_remove(dp->dp_meta_objset,
 293                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 294         }
 295
 296         if (scn->scn_phys.scn_queue_obj != 0) {
 297                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 298                     scn->scn_phys.scn_queue_obj, tx));
 299                 scn->scn_phys.scn_queue_obj = 0;
 300         }
 301
 302         /*
 303          * If we were "restarted" from a stopped state, don't bother
 304          * with anything else.
 305          */
 306         if (scn->scn_phys.scn_state != DSS_SCANNING)
 307                 return;
 308
 309         if (complete)
 310                 scn->scn_phys.scn_state = DSS_FINISHED;
 311         else
 312                 scn->scn_phys.scn_state = DSS_CANCELED;
 313
 314         spa_history_log_internal(spa, "scan done", tx,
 315             "complete=%u", complete);
 316
 317         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 318                 mutex_enter(&spa->spa_scrub_lock);
 319                 while (spa->spa_scrub_inflight > 0) {
 320                         cv_wait(&spa->spa_scrub_io_cv,
 321                             &spa->spa_scrub_lock);
 322                 }
 323                 mutex_exit(&spa->spa_scrub_lock);
 324                 spa->spa_scrub_started = B_FALSE;
 325                 spa->spa_scrub_active = B_FALSE;
 326
 327                 /*
 328                  * If the scrub/resilver completed, update all DTLs to
 329                  * reflect this.  Whether it succeeded or not, vacate
 330                  * all temporary scrub DTLs.
 331                  */
 332                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 333                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 334                 if (complete) {
 335                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 336                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 337                 }
 338                 spa_errlog_rotate(spa);
 339
 340                 /*
 341                  * We may have finished replacing a device.
 342                  * Let the async thread assess this and handle the detach.
 343                  */
 344                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 345         }
 346
 347         scn->scn_phys.scn_end_time = gethrestime_sec();
 348 }
 349
 350 /* ARGSUSED */
 351 static int
 352 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 353 {
 354         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 355
 356         if (scn->scn_phys.scn_state != DSS_SCANNING)
 357                 return (SET_ERROR(ENOENT));
 358         return (0);
 359 }
 360
 361 /* ARGSUSED */
 362 static void
 363 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 364 {
 365         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 366
 367         dsl_scan_done(scn, B_FALSE, tx);
 368         dsl_scan_sync_state(scn, tx);
 369 }
 370
 371 int
 372 dsl_scan_cancel(dsl_pool_t *dp)
 373 {
 374         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 375             dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
 376 }
 377
 378 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 379     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 380     dmu_objset_type_t ostype, dmu_tx_t *tx);
 381 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 382     dmu_objset_type_t ostype,
 383     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
 384
 385 void
 386 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 387 {
 388         zio_free(dp->dp_spa, txg, bp);
 389 }
 390
 391 void
 392 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 393 {
 394         ASSERT(dsl_pool_sync_context(dp));
 395         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 396             pio->io_flags));
 397 }
 398
 399 static uint64_t
 400 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 401 {
 402         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 403         if (ds->ds_is_snapshot)
 404                 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
 405         return (smt);
 406 }
 407
 408 static void
 409 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 410 {
 411         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 412             DMU_POOL_DIRECTORY_OBJECT,
 413             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 414             &scn->scn_phys, tx));
 415 }
 416
 417 extern int zfs_vdev_async_write_active_min_dirty_percent;
 418
 419 static boolean_t
 420 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
 421 {
 422         /* we never skip user/group accounting objects */
 423         if (zb && (int64_t)zb->zb_object < 0)
 424                 return (B_FALSE);
 425
 426         if (scn->scn_pausing)
 427                 return (B_TRUE); /* we're already pausing */
 428
 429         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 430                 return (B_FALSE); /* we're resuming */
 431
 432         /* We only know how to resume from level-0 blocks. */
 433         if (zb && zb->zb_level != 0)
 434                 return (B_FALSE);
 435
 436         /*
 437          * We pause if:
 438          *  - we have scanned for the maximum time: an entire txg
 439          *    timeout (default 5 sec)
 440          *  or
 441          *  - we have scanned for at least the minimum time (default 1 sec
 442          *    for scrub, 3 sec for resilver), and either we have sufficient
 443          *    dirty data that we are starting to write more quickly
 444          *    (default 30%), or someone is explicitly waiting for this txg
 445          *    to complete.
 446          *  or
 447          *  - the spa is shutting down because this pool is being exported
 448          *    or the machine is rebooting.
 449          */
 450         int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 451             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 452         uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 453         int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
 454         if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
 455             (NSEC2MSEC(elapsed_nanosecs) > mintime &&
 456             (txg_sync_waiting(scn->scn_dp) ||
 457             dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
 458             spa_shutting_down(scn->scn_dp->dp_spa)) {
 459                 if (zb) {
 460                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 461                             (longlong_t)zb->zb_objset,
 462                             (longlong_t)zb->zb_object,
 463                             (longlong_t)zb->zb_level,
 464                             (longlong_t)zb->zb_blkid);
 465                         scn->scn_phys.scn_bookmark = *zb;
 466                 }
 467                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 468                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 469                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 470                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 471                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 472                 scn->scn_pausing = B_TRUE;
 473                 return (B_TRUE);
 474         }
 475         return (B_FALSE);
 476 }
 477
 478 typedef struct zil_scan_arg {
 479         dsl_pool_t      *zsa_dp;
 480         zil_header_t    *zsa_zh;
 481 } zil_scan_arg_t;
 482
 483 /* ARGSUSED */
 484 static int
 485 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 486 {
 487         zil_scan_arg_t *zsa = arg;
 488         dsl_pool_t *dp = zsa->zsa_dp;
 489         dsl_scan_t *scn = dp->dp_scan;
 490         zil_header_t *zh = zsa->zsa_zh;
 491         zbookmark_phys_t zb;
 492
 493         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 494                 return (0);
 495
 496         /*
 497          * One block ("stubby") can be allocated a long time ago; we
 498          * want to visit that one because it has been allocated
 499          * (on-disk) even if it hasn't been claimed (even though for
 500          * scrub there's nothing to do to it).
 501          */
 502         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 503                 return (0);
 504
 505         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 506             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 507
 508         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 509         return (0);
 510 }
 511
 512 /* ARGSUSED */
 513 static int
 514 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 515 {
 516         if (lrc->lrc_txtype == TX_WRITE) {
 517                 zil_scan_arg_t *zsa = arg;
 518                 dsl_pool_t *dp = zsa->zsa_dp;
 519                 dsl_scan_t *scn = dp->dp_scan;
 520                 zil_header_t *zh = zsa->zsa_zh;
 521                 lr_write_t *lr = (lr_write_t *)lrc;
 522                 blkptr_t *bp = &lr->lr_blkptr;
 523                 zbookmark_phys_t zb;
 524
 525                 if (BP_IS_HOLE(bp) ||
 526                     bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 527                         return (0);
 528
 529                 /*
 530                  * birth can be < claim_txg if this record's txg is
 531                  * already txg sync'ed (but this log block contains
 532                  * other records that are not synced)
 533                  */
 534                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 535                         return (0);
 536
 537                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 538                     lr->lr_foid, ZB_ZIL_LEVEL,
 539                     lr->lr_offset / BP_GET_LSIZE(bp));
 540
 541                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 542         }
 543         return (0);
 544 }
 545
 546 static void
 547 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 548 {
 549         uint64_t claim_txg = zh->zh_claim_txg;
 550         zil_scan_arg_t zsa = { dp, zh };
 551         zilog_t *zilog;
 552
 553         /*
 554          * We only want to visit blocks that have been claimed but not yet
 555          * replayed (or, in read-only mode, blocks that *would* be claimed).
 556          */
 557         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 558                 return;
 559
 560         zilog = zil_alloc(dp->dp_meta_objset, zh);
 561
 562         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 563             claim_txg);
 564
 565         zil_free(zilog);
 566 }
 567
 568 /* ARGSUSED */
 569 static void
 570 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 571     uint64_t objset, uint64_t object, uint64_t blkid)
 572 {
 573         zbookmark_phys_t czb;
 574         arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
 575
 576         if (zfs_no_scrub_prefetch)
 577                 return;
 578
 579         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 580             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 581                 return;
 582
 583         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 584
 585         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 586             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 587             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 588 }
 589
 590 static boolean_t
 591 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 592     const zbookmark_phys_t *zb)
 593 {
 594         /*
 595          * We never skip over user/group accounting objects (obj<0)
 596          */
 597         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 598             (int64_t)zb->zb_object >= 0) {
 599                 /*
 600                  * If we already visited this bp & everything below (in
 601                  * a prior txg sync), don't bother doing it again.
 602                  */
 603                 if (zbookmark_subtree_completed(dnp, zb,
 604                     &scn->scn_phys.scn_bookmark))
 605                         return (B_TRUE);
 606
 607                 /*
 608                  * If we found the block we're trying to resume from, or
 609                  * we went past it to a different object, zero it out to
 610                  * indicate that it's OK to start checking for pausing
 611                  * again.
 612                  */
 613                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 614                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 615                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 616                             (longlong_t)zb->zb_objset,
 617                             (longlong_t)zb->zb_object,
 618                             (longlong_t)zb->zb_level,
 619                             (longlong_t)zb->zb_blkid);
 620                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 621                 }
 622         }
 623         return (B_FALSE);
 624 }
 625
 626 /*
 627  * Return nonzero on i/o error.
 628  * Return new buf to write out in *bufp.
 629  */
 630 static int
 631 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 632     dnode_phys_t *dnp, const blkptr_t *bp,
 633     const zbookmark_phys_t *zb, dmu_tx_t *tx)
 634 {
 635         dsl_pool_t *dp = scn->scn_dp;
 636         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 637         int err;
 638
 639         if (BP_GET_LEVEL(bp) > 0) {
 640                 arc_flags_t flags = ARC_FLAG_WAIT;
 641                 int i;
 642                 blkptr_t *cbp;
 643                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 644                 arc_buf_t *buf;
 645
 646                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 647                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 648                 if (err) {
 649                         scn->scn_phys.scn_errors++;
 650                         return (err);
 651                 }
 652                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 653                         dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
 654                             zb->zb_object, zb->zb_blkid * epb + i);
 655                 }
 656                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 657                         zbookmark_phys_t czb;
 658
 659                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 660                             zb->zb_level - 1,
 661                             zb->zb_blkid * epb + i);
 662                         dsl_scan_visitbp(cbp, &czb, dnp,
 663                             ds, scn, ostype, tx);
 664                 }
 665                 (void) arc_buf_remove_ref(buf, &buf);
 666         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 667                 arc_flags_t flags = ARC_FLAG_WAIT;
 668                 dnode_phys_t *cdnp;
 669                 int i, j;
 670                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 671                 arc_buf_t *buf;
 672
 673                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 674                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 675                 if (err) {
 676                         scn->scn_phys.scn_errors++;
 677                         return (err);
 678                 }
 679                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 680                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 681                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 682                                 dsl_scan_prefetch(scn, buf, cbp,
 683                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 684                         }
 685                 }
 686                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 687                         dsl_scan_visitdnode(scn, ds, ostype,
 688                             cdnp, zb->zb_blkid * epb + i, tx);
 689                 }
 690
 691                 (void) arc_buf_remove_ref(buf, &buf);
 692         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 693                 arc_flags_t flags = ARC_FLAG_WAIT;
 694                 objset_phys_t *osp;
 695                 arc_buf_t *buf;
 696
 697                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 698                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 699                 if (err) {
 700                         scn->scn_phys.scn_errors++;
 701                         return (err);
 702                 }
 703
 704                 osp = buf->b_data;
 705
 706                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 707                     &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
 708
 709                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
 710                         /*
 711                          * We also always visit user/group accounting
 712                          * objects, and never skip them, even if we are
 713                          * pausing.  This is necessary so that the space
 714                          * deltas from this txg get integrated.
 715                          */
 716                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 717                             &osp->os_groupused_dnode,
 718                             DMU_GROUPUSED_OBJECT, tx);
 719                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 720                             &osp->os_userused_dnode,
 721                             DMU_USERUSED_OBJECT, tx);
 722                 }
 723                 (void) arc_buf_remove_ref(buf, &buf);
 724         }
 725
 726         return (0);
 727 }
 728
 729 static void
 730 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 731     dmu_objset_type_t ostype, dnode_phys_t *dnp,
 732     uint64_t object, dmu_tx_t *tx)
 733 {
 734         int j;
 735
 736         for (j = 0; j < dnp->dn_nblkptr; j++) {
 737                 zbookmark_phys_t czb;
 738
 739                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 740                     dnp->dn_nlevels - 1, j);
 741                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 742                     &czb, dnp, ds, scn, ostype, tx);
 743         }
 744
 745         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 746                 zbookmark_phys_t czb;
 747                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 748                     0, DMU_SPILL_BLKID);
 749                 dsl_scan_visitbp(&dnp->dn_spill,
 750                     &czb, dnp, ds, scn, ostype, tx);
 751         }
 752 }
 753
 754 /*
 755  * The arguments are in this order because mdb can only print the
 756  * first 5; we want them to be useful.
 757  */
 758 static void
 759 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 760     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 761     dmu_objset_type_t ostype, dmu_tx_t *tx)
 762 {
 763         dsl_pool_t *dp = scn->scn_dp;
 764         arc_buf_t *buf = NULL;
 765         blkptr_t bp_toread = *bp;
 766
 767         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 768
 769         if (dsl_scan_check_pause(scn, zb))
 770                 return;
 771
 772         if (dsl_scan_check_resume(scn, dnp, zb))
 773                 return;
 774
 775         if (BP_IS_HOLE(bp))
 776                 return;
 777
 778         scn->scn_visited_this_txg++;
 779
 780         dprintf_bp(bp,
 781             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
 782             ds, ds ? ds->ds_object : 0,
 783             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 784             bp);
 785
 786         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 787                 return;
 788
 789         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
 790                 return;
 791
 792         /*
 793          * If dsl_scan_ddt() has aready visited this block, it will have
 794          * already done any translations or scrubbing, so don't call the
 795          * callback again.
 796          */
 797         if (ddt_class_contains(dp->dp_spa,
 798             scn->scn_phys.scn_ddt_class_max, bp)) {
 799                 ASSERT(buf == NULL);
 800                 return;
 801         }
 802
 803         /*
 804          * If this block is from the future (after cur_max_txg), then we
 805          * are doing this on behalf of a deleted snapshot, and we will
 806          * revisit the future block on the next pass of this dataset.
 807          * Don't scan it now unless we need to because something
 808          * under it was modified.
 809          */
 810         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 811                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 812         }
 813 }
 814
 815 static void
 816 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 817     dmu_tx_t *tx)
 818 {
 819         zbookmark_phys_t zb;
 820
 821         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 822             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 823         dsl_scan_visitbp(bp, &zb, NULL,
 824             ds, scn, DMU_OST_NONE, tx);
 825
 826         dprintf_ds(ds, "finished scan%s", "");
 827 }
 828
 829 void
 830 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 831 {
 832         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 833         dsl_scan_t *scn = dp->dp_scan;
 834         uint64_t mintxg;
 835
 836         if (scn->scn_phys.scn_state != DSS_SCANNING)
 837                 return;
 838
 839         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 840                 if (ds->ds_is_snapshot) {
 841                         /* Note, scn_cur_{min,max}_txg stays the same. */
 842                         scn->scn_phys.scn_bookmark.zb_objset =
 843                             dsl_dataset_phys(ds)->ds_next_snap_obj;
 844                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 845                             "reset zb_objset to %llu",
 846                             (u_longlong_t)ds->ds_object,
 847                             (u_longlong_t)dsl_dataset_phys(ds)->
 848                             ds_next_snap_obj);
 849                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 850                 } else {
 851                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 852                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 853                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 854                             "reset bookmark to -1,0,0,0",
 855                             (u_longlong_t)ds->ds_object);
 856                 }
 857         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 858             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 859                 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
 860                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 861                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 862                 if (ds->ds_is_snapshot) {
 863                         /*
 864                          * We keep the same mintxg; it could be >
 865                          * ds_creation_txg if the previous snapshot was
 866                          * deleted too.
 867                          */
 868                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 869                             scn->scn_phys.scn_queue_obj,
 870                             dsl_dataset_phys(ds)->ds_next_snap_obj,
 871                             mintxg, tx) == 0);
 872                         zfs_dbgmsg("destroying ds %llu; in queue; "
 873                             "replacing with %llu",
 874                             (u_longlong_t)ds->ds_object,
 875                             (u_longlong_t)dsl_dataset_phys(ds)->
 876                             ds_next_snap_obj);
 877                 } else {
 878                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 879                             (u_longlong_t)ds->ds_object);
 880                 }
 881         } else {
 882                 zfs_dbgmsg("destroying ds %llu; ignoring",
 883                     (u_longlong_t)ds->ds_object);
 884         }
 885
 886         /*
 887          * dsl_scan_sync() should be called after this, and should sync
 888          * out our changed state, but just to be safe, do it here.
 889          */
 890         dsl_scan_sync_state(scn, tx);
 891 }
 892
 893 void
 894 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 895 {
 896         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 897         dsl_scan_t *scn = dp->dp_scan;
 898         uint64_t mintxg;
 899
 900         if (scn->scn_phys.scn_state != DSS_SCANNING)
 901                 return;
 902
 903         ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
 904
 905         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 906                 scn->scn_phys.scn_bookmark.zb_objset =
 907                     dsl_dataset_phys(ds)->ds_prev_snap_obj;
 908                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 909                     "reset zb_objset to %llu",
 910                     (u_longlong_t)ds->ds_object,
 911                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 912         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 913             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 914                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 915                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 916                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 917                     scn->scn_phys.scn_queue_obj,
 918                     dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
 919                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 920                     "replacing with %llu",
 921                     (u_longlong_t)ds->ds_object,
 922                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 923         }
 924         dsl_scan_sync_state(scn, tx);
 925 }
 926
 927 void
 928 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 929 {
 930         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 931         dsl_scan_t *scn = dp->dp_scan;
 932         uint64_t mintxg;
 933
 934         if (scn->scn_phys.scn_state != DSS_SCANNING)
 935                 return;
 936
 937         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 938                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 939                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 940                     "reset zb_objset to %llu",
 941                     (u_longlong_t)ds1->ds_object,
 942                     (u_longlong_t)ds2->ds_object);
 943         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 944                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 945                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 946                     "reset zb_objset to %llu",
 947                     (u_longlong_t)ds2->ds_object,
 948                     (u_longlong_t)ds1->ds_object);
 949         }
 950
 951         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 952             ds1->ds_object, &mintxg) == 0) {
 953                 int err;
 954
 955                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 956                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 957                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 958                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 959                 err = zap_add_int_key(dp->dp_meta_objset,
 960                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 961                 VERIFY(err == 0 || err == EEXIST);
 962                 if (err == EEXIST) {
 963                         /* Both were there to begin with */
 964                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 965                             scn->scn_phys.scn_queue_obj,
 966                             ds1->ds_object, mintxg, tx));
 967                 }
 968                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 969                     "replacing with %llu",
 970                     (u_longlong_t)ds1->ds_object,
 971                     (u_longlong_t)ds2->ds_object);
 972         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 973             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 974                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 975                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 976                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 977                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 978                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 979                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 980                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 981                     "replacing with %llu",
 982                     (u_longlong_t)ds2->ds_object,
 983                     (u_longlong_t)ds1->ds_object);
 984         }
 985
 986         dsl_scan_sync_state(scn, tx);
 987 }
 988
 989 struct enqueue_clones_arg {
 990         dmu_tx_t *tx;
 991         uint64_t originobj;
 992 };
 993
 994 /* ARGSUSED */
 995 static int
 996 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 997 {
 998         struct enqueue_clones_arg *eca = arg;
 999         dsl_dataset_t *ds;
1000         int err;
1001         dsl_scan_t *scn = dp->dp_scan;
1002
1003         if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1004                 return (0);
1005
1006         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1007         if (err)
1008                 return (err);
1009
1010         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1011                 dsl_dataset_t *prev;
1012                 err = dsl_dataset_hold_obj(dp,
1013                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1014
1015                 dsl_dataset_rele(ds, FTAG);
1016                 if (err)
1017                         return (err);
1018                 ds = prev;
1019         }
1020         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1021             scn->scn_phys.scn_queue_obj, ds->ds_object,
1022             dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1023         dsl_dataset_rele(ds, FTAG);
1024         return (0);
1025 }
1026
1027 static void
1028 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1029 {
1030         dsl_pool_t *dp = scn->scn_dp;
1031         dsl_dataset_t *ds;
1032         objset_t *os;
1033
1034         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1035
1036         if (dmu_objset_from_ds(ds, &os))
1037                 goto out;
1038
1039         /*
1040          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1041          * snapshots can have ZIL block pointers (which may be the same
1042          * BP as in the head), they must be ignored.  So we traverse the
1043          * ZIL here, rather than in scan_recurse(), because the regular
1044          * snapshot block-sharing rules don't apply to it.
1045          */
1046         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1047                 dsl_scan_zil(dp, &os->os_zil_header);
1048
1049         /*
1050          * Iterate over the bps in this ds.
1051          */
1052         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1053         dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1054
1055         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1056         dsl_dataset_name(ds, dsname);
1057         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1058             "pausing=%u",
1059             (longlong_t)dsobj, dsname,
1060             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1061             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1062             (int)scn->scn_pausing);
1063         kmem_free(dsname, ZFS_MAXNAMELEN);
1064
1065         if (scn->scn_pausing)
1066                 goto out;
1067
1068         /*
1069          * We've finished this pass over this dataset.
1070          */
1071
1072         /*
1073          * If we did not completely visit this dataset, do another pass.
1074          */
1075         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1076                 zfs_dbgmsg("incomplete pass; visiting again");
1077                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1078                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1079                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1080                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1081                 goto out;
1082         }
1083
1084         /*
1085          * Add descendent datasets to work queue.
1086          */
1087         if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1088                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1089                     scn->scn_phys.scn_queue_obj,
1090                     dsl_dataset_phys(ds)->ds_next_snap_obj,
1091                     dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1092         }
1093         if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1094                 boolean_t usenext = B_FALSE;
1095                 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1096                         uint64_t count;
1097                         /*
1098                          * A bug in a previous version of the code could
1099                          * cause upgrade_clones_cb() to not set
1100                          * ds_next_snap_obj when it should, leading to a
1101                          * missing entry.  Therefore we can only use the
1102                          * next_clones_obj when its count is correct.
1103                          */
1104                         int err = zap_count(dp->dp_meta_objset,
1105                             dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1106                         if (err == 0 &&
1107                             count == dsl_dataset_phys(ds)->ds_num_children - 1)
1108                                 usenext = B_TRUE;
1109                 }
1110
1111                 if (usenext) {
1112                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1113                             dsl_dataset_phys(ds)->ds_next_clones_obj,
1114                             scn->scn_phys.scn_queue_obj,
1115                             dsl_dataset_phys(ds)->ds_creation_txg, tx));
1116                 } else {
1117                         struct enqueue_clones_arg eca;
1118                         eca.tx = tx;
1119                         eca.originobj = ds->ds_object;
1120
1121                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1122                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1123                 }
1124         }
1125
1126 out:
1127         dsl_dataset_rele(ds, FTAG);
1128 }
1129
1130 /* ARGSUSED */
1131 static int
1132 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1133 {
1134         dmu_tx_t *tx = arg;
1135         dsl_dataset_t *ds;
1136         int err;
1137         dsl_scan_t *scn = dp->dp_scan;
1138
1139         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1140         if (err)
1141                 return (err);
1142
1143         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1144                 dsl_dataset_t *prev;
1145                 err = dsl_dataset_hold_obj(dp,
1146                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1147                 if (err) {
1148                         dsl_dataset_rele(ds, FTAG);
1149                         return (err);
1150                 }
1151
1152                 /*
1153                  * If this is a clone, we don't need to worry about it for now.
1154                  */
1155                 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1156                         dsl_dataset_rele(ds, FTAG);
1157                         dsl_dataset_rele(prev, FTAG);
1158                         return (0);
1159                 }
1160                 dsl_dataset_rele(ds, FTAG);
1161                 ds = prev;
1162         }
1163
1164         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1165             ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1166         dsl_dataset_rele(ds, FTAG);
1167         return (0);
1168 }
1169
1170 /*
1171  * Scrub/dedup interaction.
1172  *
1173  * If there are N references to a deduped block, we don't want to scrub it
1174  * N times -- ideally, we should scrub it exactly once.
1175  *
1176  * We leverage the fact that the dde's replication class (enum ddt_class)
1177  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1178  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1179  *
1180  * To prevent excess scrubbing, the scrub begins by walking the DDT
1181  * to find all blocks with refcnt > 1, and scrubs each of these once.
1182  * Since there are two replication classes which contain blocks with
1183  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1184  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1185  *
1186  * There would be nothing more to say if a block's refcnt couldn't change
1187  * during a scrub, but of course it can so we must account for changes
1188  * in a block's replication class.
1189  *
1190  * Here's an example of what can occur:
1191  *
1192  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1193  * when visited during the top-down scrub phase, it will be scrubbed twice.
1194  * This negates our scrub optimization, but is otherwise harmless.
1195  *
1196  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1197  * on each visit during the top-down scrub phase, it will never be scrubbed.
1198  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1199  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1200  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1201  * while a scrub is in progress, it scrubs the block right then.
1202  */
1203 static void
1204 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1205 {
1206         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1207         ddt_entry_t dde = { 0 };
1208         int error;
1209         uint64_t n = 0;
1210
1211         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1212                 ddt_t *ddt;
1213
1214                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1215                         break;
1216                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1217                     (longlong_t)ddb->ddb_class,
1218                     (longlong_t)ddb->ddb_type,
1219                     (longlong_t)ddb->ddb_checksum,
1220                     (longlong_t)ddb->ddb_cursor);
1221
1222                 /* There should be no pending changes to the dedup table */
1223                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1224                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1225
1226                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1227                 n++;
1228
1229                 if (dsl_scan_check_pause(scn, NULL))
1230                         break;
1231         }
1232
1233         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1234             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1235             (int)scn->scn_pausing);
1236
1237         ASSERT(error == 0 || error == ENOENT);
1238         ASSERT(error != ENOENT ||
1239             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1240 }
1241
1242 /* ARGSUSED */
1243 void
1244 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1245     ddt_entry_t *dde, dmu_tx_t *tx)
1246 {
1247         const ddt_key_t *ddk = &dde->dde_key;
1248         ddt_phys_t *ddp = dde->dde_phys;
1249         blkptr_t bp;
1250         zbookmark_phys_t zb = { 0 };
1251
1252         if (scn->scn_phys.scn_state != DSS_SCANNING)
1253                 return;
1254
1255         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1256                 if (ddp->ddp_phys_birth == 0 ||
1257                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1258                         continue;
1259                 ddt_bp_create(checksum, ddk, ddp, &bp);
1260
1261                 scn->scn_visited_this_txg++;
1262                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1263         }
1264 }
1265
1266 static void
1267 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1268 {
1269         dsl_pool_t *dp = scn->scn_dp;
1270         zap_cursor_t zc;
1271         zap_attribute_t za;
1272
1273         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1274             scn->scn_phys.scn_ddt_class_max) {
1275                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1276                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1277                 dsl_scan_ddt(scn, tx);
1278                 if (scn->scn_pausing)
1279                         return;
1280         }
1281
1282         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1283                 /* First do the MOS & ORIGIN */
1284
1285                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1286                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1287                 dsl_scan_visit_rootbp(scn, NULL,
1288                     &dp->dp_meta_rootbp, tx);
1289                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1290                 if (scn->scn_pausing)
1291                         return;
1292
1293                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1294                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1295                             enqueue_cb, tx, DS_FIND_CHILDREN));
1296                 } else {
1297                         dsl_scan_visitds(scn,
1298                             dp->dp_origin_snap->ds_object, tx);
1299                 }
1300                 ASSERT(!scn->scn_pausing);
1301         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1302             ZB_DESTROYED_OBJSET) {
1303                 /*
1304                  * If we were paused, continue from here.  Note if the
1305                  * ds we were paused on was deleted, the zb_objset may
1306                  * be -1, so we will skip this and find a new objset
1307                  * below.
1308                  */
1309                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1310                 if (scn->scn_pausing)
1311                         return;
1312         }
1313
1314         /*
1315          * In case we were paused right at the end of the ds, zero the
1316          * bookmark so we don't think that we're still trying to resume.
1317          */
1318         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1319
1320         /* keep pulling things out of the zap-object-as-queue */
1321         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1322             scn->scn_phys.scn_queue_obj),
1323             zap_cursor_retrieve(&zc, &za) == 0) {
1324                 dsl_dataset_t *ds;
1325                 uint64_t dsobj;
1326
1327                 dsobj = strtonum(za.za_name, NULL);
1328                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1329                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1330
1331                 /* Set up min/max txg */
1332                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1333                 if (za.za_first_integer != 0) {
1334                         scn->scn_phys.scn_cur_min_txg =
1335                             MAX(scn->scn_phys.scn_min_txg,
1336                             za.za_first_integer);
1337                 } else {
1338                         scn->scn_phys.scn_cur_min_txg =
1339                             MAX(scn->scn_phys.scn_min_txg,
1340                             dsl_dataset_phys(ds)->ds_prev_snap_txg);
1341                 }
1342                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1343                 dsl_dataset_rele(ds, FTAG);
1344
1345                 dsl_scan_visitds(scn, dsobj, tx);
1346                 zap_cursor_fini(&zc);
1347                 if (scn->scn_pausing)
1348                         return;
1349         }
1350         zap_cursor_fini(&zc);
1351 }
1352
1353 static boolean_t
1354 dsl_scan_free_should_pause(dsl_scan_t *scn)
1355 {
1356         uint64_t elapsed_nanosecs;
1357
1358         if (zfs_recover)
1359                 return (B_FALSE);
1360
1361         if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1362                 return (B_TRUE);
1363
1364         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1365         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1366             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1367             txg_sync_waiting(scn->scn_dp)) ||
1368             spa_shutting_down(scn->scn_dp->dp_spa));
1369 }
1370
1371 static int
1372 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1373 {
1374         dsl_scan_t *scn = arg;
1375
1376         if (!scn->scn_is_bptree ||
1377             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1378                 if (dsl_scan_free_should_pause(scn))
1379                         return (SET_ERROR(ERESTART));
1380         }
1381
1382         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1383             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1384         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1385             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1386             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1387         scn->scn_visited_this_txg++;
1388         return (0);
1389 }
1390
1391 boolean_t
1392 dsl_scan_active(dsl_scan_t *scn)
1393 {
1394         spa_t *spa = scn->scn_dp->dp_spa;
1395         uint64_t used = 0, comp, uncomp;
1396
1397         if (spa->spa_load_state != SPA_LOAD_NONE)
1398                 return (B_FALSE);
1399         if (spa_shutting_down(spa))
1400                 return (B_FALSE);
1401         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1402             (scn->scn_async_destroying && !scn->scn_async_stalled))
1403                 return (B_TRUE);
1404
1405         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1406                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1407                     &used, &comp, &uncomp);
1408         }
1409         return (used != 0);
1410 }
1411
1412 void
1413 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1414 {
1415         dsl_scan_t *scn = dp->dp_scan;
1416         spa_t *spa = dp->dp_spa;
1417         int err = 0;
1418
1419         /*
1420          * Check for scn_restart_txg before checking spa_load_state, so
1421          * that we can restart an old-style scan while the pool is being
1422          * imported (see dsl_scan_init).
1423          */
1424         if (scn->scn_restart_txg != 0 &&
1425             scn->scn_restart_txg <= tx->tx_txg) {
1426                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1427                 dsl_scan_done(scn, B_FALSE, tx);
1428                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1429                         func = POOL_SCAN_RESILVER;
1430                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1431                     func, tx->tx_txg);
1432                 dsl_scan_setup_sync(&func, tx);
1433         }
1434
1435         /*
1436          * If the scan is inactive due to a stalled async destroy, try again.
1437          */
1438         if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1439             spa_sync_pass(dp->dp_spa) > 1)
1440                 return;
1441
1442         scn->scn_visited_this_txg = 0;
1443         scn->scn_pausing = B_FALSE;
1444         scn->scn_sync_start_time = gethrtime();
1445         spa->spa_scrub_active = B_TRUE;
1446
1447         /*
1448          * First process the async destroys.  If we pause, don't do
1449          * any scrubbing or resilvering.  This ensures that there are no
1450          * async destroys while we are scanning, so the scan code doesn't
1451          * have to worry about traversing it.  It is also faster to free the
1452          * blocks than to scrub them.
1453          */
1454         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1455                 scn->scn_is_bptree = B_FALSE;
1456                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1457                     NULL, ZIO_FLAG_MUSTSUCCEED);
1458                 err = bpobj_iterate(&dp->dp_free_bpobj,
1459                     dsl_scan_free_block_cb, scn, tx);
1460                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1461
1462                 if (err != 0 && err != ERESTART)
1463                         zfs_panic_recover("error %u from bpobj_iterate()", err);
1464         }
1465
1466         if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1467                 ASSERT(scn->scn_async_destroying);
1468                 scn->scn_is_bptree = B_TRUE;
1469                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1470                     NULL, ZIO_FLAG_MUSTSUCCEED);
1471                 err = bptree_iterate(dp->dp_meta_objset,
1472                     dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1473                 VERIFY0(zio_wait(scn->scn_zio_root));
1474
1475                 if (err == EIO || err == ECKSUM) {
1476                         err = 0;
1477                 } else if (err != 0 && err != ERESTART) {
1478                         zfs_panic_recover("error %u from "
1479                             "traverse_dataset_destroyed()", err);
1480                 }
1481
1482                 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1483                         /* finished; deactivate async destroy feature */
1484                         spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1485                         ASSERT(!spa_feature_is_active(spa,
1486                             SPA_FEATURE_ASYNC_DESTROY));
1487                         VERIFY0(zap_remove(dp->dp_meta_objset,
1488                             DMU_POOL_DIRECTORY_OBJECT,
1489                             DMU_POOL_BPTREE_OBJ, tx));
1490                         VERIFY0(bptree_free(dp->dp_meta_objset,
1491                             dp->dp_bptree_obj, tx));
1492                         dp->dp_bptree_obj = 0;
1493                         scn->scn_async_destroying = B_FALSE;
1494                         scn->scn_async_stalled = B_FALSE;
1495                 } else {
1496                         /*
1497                          * If we didn't make progress, mark the async
1498                          * destroy as stalled, so that we will not initiate
1499                          * a spa_sync() on its behalf.  Note that we only
1500                          * check this if we are not finished, because if the
1501                          * bptree had no blocks for us to visit, we can
1502                          * finish without "making progress".
1503                          */
1504                         scn->scn_async_stalled =
1505                             (scn->scn_visited_this_txg == 0);
1506                 }
1507         }
1508         if (scn->scn_visited_this_txg) {
1509                 zfs_dbgmsg("freed %llu blocks in %llums from "
1510                     "free_bpobj/bptree txg %llu; err=%d",
1511                     (longlong_t)scn->scn_visited_this_txg,
1512                     (longlong_t)
1513                     NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1514                     (longlong_t)tx->tx_txg, err);
1515                 scn->scn_visited_this_txg = 0;
1516
1517                 /*
1518                  * Write out changes to the DDT that may be required as a
1519                  * result of the blocks freed.  This ensures that the DDT
1520                  * is clean when a scrub/resilver runs.
1521                  */
1522                 ddt_sync(spa, tx->tx_txg);
1523         }
1524         if (err != 0)
1525                 return;
1526         if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1527             (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1528             dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1529             dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1530                 /*
1531                  * We have finished background destroying, but there is still
1532                  * some space left in the dp_free_dir. Transfer this leaked
1533                  * space to the dp_leak_dir.
1534                  */
1535                 if (dp->dp_leak_dir == NULL) {
1536                         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1537                         (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1538                             LEAK_DIR_NAME, tx);
1539                         VERIFY0(dsl_pool_open_special_dir(dp,
1540                             LEAK_DIR_NAME, &dp->dp_leak_dir));
1541                         rrw_exit(&dp->dp_config_rwlock, FTAG);
1542                 }
1543                 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1544                     dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1545                     dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1546                     dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1547                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1548                     -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1549                     -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1550                     -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1551         }
1552         if (!scn->scn_async_destroying) {
1553                 /* finished; verify that space accounting went to zero */
1554                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1555                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1556                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1557         }
1558
1559         if (scn->scn_phys.scn_state != DSS_SCANNING)
1560                 return;
1561
1562         if (scn->scn_done_txg == tx->tx_txg) {
1563                 ASSERT(!scn->scn_pausing);
1564                 /* finished with scan. */
1565                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1566                 dsl_scan_done(scn, B_TRUE, tx);
1567                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1568                 dsl_scan_sync_state(scn, tx);
1569                 return;
1570         }
1571
1572         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1573             scn->scn_phys.scn_ddt_class_max) {
1574                 zfs_dbgmsg("doing scan sync txg %llu; "
1575                     "ddt bm=%llu/%llu/%llu/%llx",
1576                     (longlong_t)tx->tx_txg,
1577                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1578                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1579                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1580                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1581                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1582                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1583                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1584                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1585         } else {
1586                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1587                     (longlong_t)tx->tx_txg,
1588                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1589                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1590                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1591                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1592         }
1593
1594         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1595             NULL, ZIO_FLAG_CANFAIL);
1596         dsl_pool_config_enter(dp, FTAG);
1597         dsl_scan_visit(scn, tx);
1598         dsl_pool_config_exit(dp, FTAG);
1599         (void) zio_wait(scn->scn_zio_root);
1600         scn->scn_zio_root = NULL;
1601
1602         zfs_dbgmsg("visited %llu blocks in %llums",
1603             (longlong_t)scn->scn_visited_this_txg,
1604             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1605
1606         if (!scn->scn_pausing) {
1607                 scn->scn_done_txg = tx->tx_txg + 1;
1608                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1609                     tx->tx_txg, scn->scn_done_txg);
1610         }
1611
1612         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1613                 mutex_enter(&spa->spa_scrub_lock);
1614                 while (spa->spa_scrub_inflight > 0) {
1615                         cv_wait(&spa->spa_scrub_io_cv,
1616                             &spa->spa_scrub_lock);
1617                 }
1618                 mutex_exit(&spa->spa_scrub_lock);
1619         }
1620
1621         dsl_scan_sync_state(scn, tx);
1622 }
1623
1624 /*
1625  * This will start a new scan, or restart an existing one.
1626  */
1627 void
1628 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1629 {
1630         if (txg == 0) {
1631                 dmu_tx_t *tx;
1632                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1633                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1634
1635                 txg = dmu_tx_get_txg(tx);
1636                 dp->dp_scan->scn_restart_txg = txg;
1637                 dmu_tx_commit(tx);
1638         } else {
1639                 dp->dp_scan->scn_restart_txg = txg;
1640         }
1641         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1642 }
1643
1644 boolean_t
1645 dsl_scan_resilvering(dsl_pool_t *dp)
1646 {
1647         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1648             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1649 }
1650
1651 /*
1652  * scrub consumers
1653  */
1654
1655 static void
1656 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1657 {
1658         int i;
1659
1660         /*
1661          * If we resume after a reboot, zab will be NULL; don't record
1662          * incomplete stats in that case.
1663          */
1664         if (zab == NULL)
1665                 return;
1666
1667         for (i = 0; i < 4; i++) {
1668                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1669                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1670                 if (t & DMU_OT_NEWTYPE)
1671                         t = DMU_OT_OTHER;
1672                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1673                 int equal;
1674
1675                 zb->zb_count++;
1676                 zb->zb_asize += BP_GET_ASIZE(bp);
1677                 zb->zb_lsize += BP_GET_LSIZE(bp);
1678                 zb->zb_psize += BP_GET_PSIZE(bp);
1679                 zb->zb_gangs += BP_COUNT_GANG(bp);
1680
1681                 switch (BP_GET_NDVAS(bp)) {
1682                 case 2:
1683                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1684                             DVA_GET_VDEV(&bp->blk_dva[1]))
1685                                 zb->zb_ditto_2_of_2_samevdev++;
1686                         break;
1687                 case 3:
1688                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1689                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1690                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1691                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1692                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1693                             DVA_GET_VDEV(&bp->blk_dva[2]));
1694                         if (equal == 1)
1695                                 zb->zb_ditto_2_of_3_samevdev++;
1696                         else if (equal == 3)
1697                                 zb->zb_ditto_3_of_3_samevdev++;
1698                         break;
1699                 }
1700         }
1701 }
1702
1703 static void
1704 dsl_scan_scrub_done(zio_t *zio)
1705 {
1706         spa_t *spa = zio->io_spa;
1707
1708         zio_data_buf_free(zio->io_data, zio->io_size);
1709
1710         mutex_enter(&spa->spa_scrub_lock);
1711         spa->spa_scrub_inflight--;
1712         cv_broadcast(&spa->spa_scrub_io_cv);
1713
1714         if (zio->io_error && (zio->io_error != ECKSUM ||
1715             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1716                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1717         }
1718         mutex_exit(&spa->spa_scrub_lock);
1719 }
1720
1721 static int
1722 dsl_scan_scrub_cb(dsl_pool_t *dp,
1723     const blkptr_t *bp, const zbookmark_phys_t *zb)
1724 {
1725         dsl_scan_t *scn = dp->dp_scan;
1726         size_t size = BP_GET_PSIZE(bp);
1727         spa_t *spa = dp->dp_spa;
1728         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1729         boolean_t needs_io;
1730         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1731         unsigned int scan_delay = 0;
1732
1733         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1734             phys_birth >= scn->scn_phys.scn_max_txg)
1735                 return (0);
1736
1737         count_block(dp->dp_blkstats, bp);
1738
1739         if (BP_IS_EMBEDDED(bp))
1740                 return (0);
1741
1742         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1743         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1744                 zio_flags |= ZIO_FLAG_SCRUB;
1745                 needs_io = B_TRUE;
1746                 scan_delay = zfs_scrub_delay;
1747         } else {
1748                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1749                 zio_flags |= ZIO_FLAG_RESILVER;
1750                 needs_io = B_FALSE;
1751                 scan_delay = zfs_resilver_delay;
1752         }
1753
1754         /* If it's an intent log block, failure is expected. */
1755         if (zb->zb_level == ZB_ZIL_LEVEL)
1756                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1757
1758         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1759                 vdev_t *vd = vdev_lookup_top(spa,
1760                     DVA_GET_VDEV(&bp->blk_dva[d]));
1761
1762                 /*
1763                  * Keep track of how much data we've examined so that
1764                  * zpool(1M) status can make useful progress reports.
1765                  */
1766                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1767                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1768
1769                 /* if it's a resilver, this may not be in the target range */
1770                 if (!needs_io) {
1771                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1772                                 /*
1773                                  * Gang members may be spread across multiple
1774                                  * vdevs, so the best estimate we have is the
1775                                  * scrub range, which has already been checked.
1776                                  * XXX -- it would be better to change our
1777                                  * allocation policy to ensure that all
1778                                  * gang members reside on the same vdev.
1779                                  */
1780                                 needs_io = B_TRUE;
1781                         } else {
1782                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1783                                     phys_birth, 1);
1784                         }
1785                 }
1786         }
1787
1788         if (needs_io && !zfs_no_scrub_io) {
1789                 vdev_t *rvd = spa->spa_root_vdev;
1790                 uint64_t maxinflight = rvd->vdev_children *
1791                     MAX(zfs_top_maxinflight, 1);
1792                 void *data = zio_data_buf_alloc(size);
1793
1794                 mutex_enter(&spa->spa_scrub_lock);
1795                 while (spa->spa_scrub_inflight >= maxinflight)
1796                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1797                 spa->spa_scrub_inflight++;
1798                 mutex_exit(&spa->spa_scrub_lock);
1799
1800                 /*
1801                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1802                  * then throttle our workload to limit the impact of a scan.
1803                  */
1804                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1805                         delay(MAX((int)scan_delay, 0));
1806
1807                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1808                     dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1809                     zio_flags, zb));
1810         }
1811
1812         /* do not relocate this block */
1813         return (0);
1814 }
1815
1816 int
1817 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1818 {
1819         spa_t *spa = dp->dp_spa;
1820
1821         /*
1822          * Purge all vdev caches and probe all devices.  We do this here
1823          * rather than in sync context because this requires a writer lock
1824          * on the spa_config lock, which we can't do from sync context.  The
1825          * spa_scrub_reopen flag indicates that vdev_open() should not
1826          * attempt to start another scrub.
1827          */
1828         spa_vdev_state_enter(spa, SCL_NONE);
1829         spa->spa_scrub_reopen = B_TRUE;
1830         vdev_reopen(spa->spa_root_vdev);
1831         spa->spa_scrub_reopen = B_FALSE;
1832         (void) spa_vdev_state_exit(spa, NULL, 0);
1833
1834         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1835             dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1836 }