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 (dsl_dataset_is_snapshot(ds))
 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_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 604                         return (B_TRUE);
 605
 606                 /*
 607                  * If we found the block we're trying to resume from, or
 608                  * we went past it to a different object, zero it out to
 609                  * indicate that it's OK to start checking for pausing
 610                  * again.
 611                  */
 612                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 613                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 614                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 615                             (longlong_t)zb->zb_objset,
 616                             (longlong_t)zb->zb_object,
 617                             (longlong_t)zb->zb_level,
 618                             (longlong_t)zb->zb_blkid);
 619                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 620                 }
 621         }
 622         return (B_FALSE);
 623 }
 624
 625 /*
 626  * Return nonzero on i/o error.
 627  * Return new buf to write out in *bufp.
 628  */
 629 static int
 630 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 631     dnode_phys_t *dnp, const blkptr_t *bp,
 632     const zbookmark_phys_t *zb, dmu_tx_t *tx)
 633 {
 634         dsl_pool_t *dp = scn->scn_dp;
 635         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 636         int err;
 637
 638         if (BP_GET_LEVEL(bp) > 0) {
 639                 arc_flags_t flags = ARC_FLAG_WAIT;
 640                 int i;
 641                 blkptr_t *cbp;
 642                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 643                 arc_buf_t *buf;
 644
 645                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 646                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 647                 if (err) {
 648                         scn->scn_phys.scn_errors++;
 649                         return (err);
 650                 }
 651                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 652                         dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
 653                             zb->zb_object, zb->zb_blkid * epb + i);
 654                 }
 655                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 656                         zbookmark_phys_t czb;
 657
 658                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 659                             zb->zb_level - 1,
 660                             zb->zb_blkid * epb + i);
 661                         dsl_scan_visitbp(cbp, &czb, dnp,
 662                             ds, scn, ostype, tx);
 663                 }
 664                 (void) arc_buf_remove_ref(buf, &buf);
 665         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 666                 arc_flags_t flags = ARC_FLAG_WAIT;
 667                 dnode_phys_t *cdnp;
 668                 int i, j;
 669                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 670                 arc_buf_t *buf;
 671
 672                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 673                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 674                 if (err) {
 675                         scn->scn_phys.scn_errors++;
 676                         return (err);
 677                 }
 678                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 679                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 680                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 681                                 dsl_scan_prefetch(scn, buf, cbp,
 682                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 683                         }
 684                 }
 685                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 686                         dsl_scan_visitdnode(scn, ds, ostype,
 687                             cdnp, zb->zb_blkid * epb + i, tx);
 688                 }
 689
 690                 (void) arc_buf_remove_ref(buf, &buf);
 691         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 692                 arc_flags_t flags = ARC_FLAG_WAIT;
 693                 objset_phys_t *osp;
 694                 arc_buf_t *buf;
 695
 696                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 697                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 698                 if (err) {
 699                         scn->scn_phys.scn_errors++;
 700                         return (err);
 701                 }
 702
 703                 osp = buf->b_data;
 704
 705                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 706                     &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
 707
 708                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
 709                         /*
 710                          * We also always visit user/group accounting
 711                          * objects, and never skip them, even if we are
 712                          * pausing.  This is necessary so that the space
 713                          * deltas from this txg get integrated.
 714                          */
 715                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 716                             &osp->os_groupused_dnode,
 717                             DMU_GROUPUSED_OBJECT, tx);
 718                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 719                             &osp->os_userused_dnode,
 720                             DMU_USERUSED_OBJECT, tx);
 721                 }
 722                 (void) arc_buf_remove_ref(buf, &buf);
 723         }
 724
 725         return (0);
 726 }
 727
 728 static void
 729 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 730     dmu_objset_type_t ostype, dnode_phys_t *dnp,
 731     uint64_t object, dmu_tx_t *tx)
 732 {
 733         int j;
 734
 735         for (j = 0; j < dnp->dn_nblkptr; j++) {
 736                 zbookmark_phys_t czb;
 737
 738                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 739                     dnp->dn_nlevels - 1, j);
 740                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 741                     &czb, dnp, ds, scn, ostype, tx);
 742         }
 743
 744         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 745                 zbookmark_phys_t czb;
 746                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 747                     0, DMU_SPILL_BLKID);
 748                 dsl_scan_visitbp(&dnp->dn_spill,
 749                     &czb, dnp, ds, scn, ostype, tx);
 750         }
 751 }
 752
 753 /*
 754  * The arguments are in this order because mdb can only print the
 755  * first 5; we want them to be useful.
 756  */
 757 static void
 758 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 759     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 760     dmu_objset_type_t ostype, dmu_tx_t *tx)
 761 {
 762         dsl_pool_t *dp = scn->scn_dp;
 763         arc_buf_t *buf = NULL;
 764         blkptr_t bp_toread = *bp;
 765
 766         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 767
 768         if (dsl_scan_check_pause(scn, zb))
 769                 return;
 770
 771         if (dsl_scan_check_resume(scn, dnp, zb))
 772                 return;
 773
 774         if (BP_IS_HOLE(bp))
 775                 return;
 776
 777         scn->scn_visited_this_txg++;
 778
 779         dprintf_bp(bp,
 780             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
 781             ds, ds ? ds->ds_object : 0,
 782             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 783             bp);
 784
 785         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 786                 return;
 787
 788         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
 789                 return;
 790
 791         /*
 792          * If dsl_scan_ddt() has aready visited this block, it will have
 793          * already done any translations or scrubbing, so don't call the
 794          * callback again.
 795          */
 796         if (ddt_class_contains(dp->dp_spa,
 797             scn->scn_phys.scn_ddt_class_max, bp)) {
 798                 ASSERT(buf == NULL);
 799                 return;
 800         }
 801
 802         /*
 803          * If this block is from the future (after cur_max_txg), then we
 804          * are doing this on behalf of a deleted snapshot, and we will
 805          * revisit the future block on the next pass of this dataset.
 806          * Don't scan it now unless we need to because something
 807          * under it was modified.
 808          */
 809         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 810                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 811         }
 812 }
 813
 814 static void
 815 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 816     dmu_tx_t *tx)
 817 {
 818         zbookmark_phys_t zb;
 819
 820         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 821             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 822         dsl_scan_visitbp(bp, &zb, NULL,
 823             ds, scn, DMU_OST_NONE, tx);
 824
 825         dprintf_ds(ds, "finished scan%s", "");
 826 }
 827
 828 void
 829 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 830 {
 831         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 832         dsl_scan_t *scn = dp->dp_scan;
 833         uint64_t mintxg;
 834
 835         if (scn->scn_phys.scn_state != DSS_SCANNING)
 836                 return;
 837
 838         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 839                 if (dsl_dataset_is_snapshot(ds)) {
 840                         /* Note, scn_cur_{min,max}_txg stays the same. */
 841                         scn->scn_phys.scn_bookmark.zb_objset =
 842                             dsl_dataset_phys(ds)->ds_next_snap_obj;
 843                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 844                             "reset zb_objset to %llu",
 845                             (u_longlong_t)ds->ds_object,
 846                             (u_longlong_t)dsl_dataset_phys(ds)->
 847                             ds_next_snap_obj);
 848                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 849                 } else {
 850                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 851                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 852                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 853                             "reset bookmark to -1,0,0,0",
 854                             (u_longlong_t)ds->ds_object);
 855                 }
 856         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 857             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 858                 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
 859                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 860                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 861                 if (dsl_dataset_is_snapshot(ds)) {
 862                         /*
 863                          * We keep the same mintxg; it could be >
 864                          * ds_creation_txg if the previous snapshot was
 865                          * deleted too.
 866                          */
 867                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 868                             scn->scn_phys.scn_queue_obj,
 869                             dsl_dataset_phys(ds)->ds_next_snap_obj,
 870                             mintxg, tx) == 0);
 871                         zfs_dbgmsg("destroying ds %llu; in queue; "
 872                             "replacing with %llu",
 873                             (u_longlong_t)ds->ds_object,
 874                             (u_longlong_t)dsl_dataset_phys(ds)->
 875                             ds_next_snap_obj);
 876                 } else {
 877                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 878                             (u_longlong_t)ds->ds_object);
 879                 }
 880         } else {
 881                 zfs_dbgmsg("destroying ds %llu; ignoring",
 882                     (u_longlong_t)ds->ds_object);
 883         }
 884
 885         /*
 886          * dsl_scan_sync() should be called after this, and should sync
 887          * out our changed state, but just to be safe, do it here.
 888          */
 889         dsl_scan_sync_state(scn, tx);
 890 }
 891
 892 void
 893 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 894 {
 895         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 896         dsl_scan_t *scn = dp->dp_scan;
 897         uint64_t mintxg;
 898
 899         if (scn->scn_phys.scn_state != DSS_SCANNING)
 900                 return;
 901
 902         ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
 903
 904         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 905                 scn->scn_phys.scn_bookmark.zb_objset =
 906                     dsl_dataset_phys(ds)->ds_prev_snap_obj;
 907                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 908                     "reset zb_objset to %llu",
 909                     (u_longlong_t)ds->ds_object,
 910                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 911         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 912             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 913                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 914                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 915                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 916                     scn->scn_phys.scn_queue_obj,
 917                     dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
 918                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 919                     "replacing with %llu",
 920                     (u_longlong_t)ds->ds_object,
 921                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 922         }
 923         dsl_scan_sync_state(scn, tx);
 924 }
 925
 926 void
 927 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 928 {
 929         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 930         dsl_scan_t *scn = dp->dp_scan;
 931         uint64_t mintxg;
 932
 933         if (scn->scn_phys.scn_state != DSS_SCANNING)
 934                 return;
 935
 936         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 937                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 938                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 939                     "reset zb_objset to %llu",
 940                     (u_longlong_t)ds1->ds_object,
 941                     (u_longlong_t)ds2->ds_object);
 942         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 943                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 944                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 945                     "reset zb_objset to %llu",
 946                     (u_longlong_t)ds2->ds_object,
 947                     (u_longlong_t)ds1->ds_object);
 948         }
 949
 950         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 951             ds1->ds_object, &mintxg) == 0) {
 952                 int err;
 953
 954                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 955                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 956                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 957                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 958                 err = zap_add_int_key(dp->dp_meta_objset,
 959                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 960                 VERIFY(err == 0 || err == EEXIST);
 961                 if (err == EEXIST) {
 962                         /* Both were there to begin with */
 963                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 964                             scn->scn_phys.scn_queue_obj,
 965                             ds1->ds_object, mintxg, tx));
 966                 }
 967                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 968                     "replacing with %llu",
 969                     (u_longlong_t)ds1->ds_object,
 970                     (u_longlong_t)ds2->ds_object);
 971         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 972             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 973                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 974                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 975                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 976                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 977                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 978                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 979                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 980                     "replacing with %llu",
 981                     (u_longlong_t)ds2->ds_object,
 982                     (u_longlong_t)ds1->ds_object);
 983         }
 984
 985         dsl_scan_sync_state(scn, tx);
 986 }
 987
 988 struct enqueue_clones_arg {
 989         dmu_tx_t *tx;
 990         uint64_t originobj;
 991 };
 992
 993 /* ARGSUSED */
 994 static int
 995 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 996 {
 997         struct enqueue_clones_arg *eca = arg;
 998         dsl_dataset_t *ds;
 999         int err;
1000         dsl_scan_t *scn = dp->dp_scan;
1001
1002         if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1003                 return (0);
1004
1005         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1006         if (err)
1007                 return (err);
1008
1009         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1010                 dsl_dataset_t *prev;
1011                 err = dsl_dataset_hold_obj(dp,
1012                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1013
1014                 dsl_dataset_rele(ds, FTAG);
1015                 if (err)
1016                         return (err);
1017                 ds = prev;
1018         }
1019         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1020             scn->scn_phys.scn_queue_obj, ds->ds_object,
1021             dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1022         dsl_dataset_rele(ds, FTAG);
1023         return (0);
1024 }
1025
1026 static void
1027 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1028 {
1029         dsl_pool_t *dp = scn->scn_dp;
1030         dsl_dataset_t *ds;
1031         objset_t *os;
1032
1033         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1034
1035         if (dmu_objset_from_ds(ds, &os))
1036                 goto out;
1037
1038         /*
1039          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1040          * snapshots can have ZIL block pointers (which may be the same
1041          * BP as in the head), they must be ignored.  So we traverse the
1042          * ZIL here, rather than in scan_recurse(), because the regular
1043          * snapshot block-sharing rules don't apply to it.
1044          */
1045         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1046                 dsl_scan_zil(dp, &os->os_zil_header);
1047
1048         /*
1049          * Iterate over the bps in this ds.
1050          */
1051         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1052         dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1053
1054         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1055         dsl_dataset_name(ds, dsname);
1056         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1057             "pausing=%u",
1058             (longlong_t)dsobj, dsname,
1059             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1060             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1061             (int)scn->scn_pausing);
1062         kmem_free(dsname, ZFS_MAXNAMELEN);
1063
1064         if (scn->scn_pausing)
1065                 goto out;
1066
1067         /*
1068          * We've finished this pass over this dataset.
1069          */
1070
1071         /*
1072          * If we did not completely visit this dataset, do another pass.
1073          */
1074         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1075                 zfs_dbgmsg("incomplete pass; visiting again");
1076                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1077                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1078                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1079                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1080                 goto out;
1081         }
1082
1083         /*
1084          * Add descendent datasets to work queue.
1085          */
1086         if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1087                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1088                     scn->scn_phys.scn_queue_obj,
1089                     dsl_dataset_phys(ds)->ds_next_snap_obj,
1090                     dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1091         }
1092         if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1093                 boolean_t usenext = B_FALSE;
1094                 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1095                         uint64_t count;
1096                         /*
1097                          * A bug in a previous version of the code could
1098                          * cause upgrade_clones_cb() to not set
1099                          * ds_next_snap_obj when it should, leading to a
1100                          * missing entry.  Therefore we can only use the
1101                          * next_clones_obj when its count is correct.
1102                          */
1103                         int err = zap_count(dp->dp_meta_objset,
1104                             dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1105                         if (err == 0 &&
1106                             count == dsl_dataset_phys(ds)->ds_num_children - 1)
1107                                 usenext = B_TRUE;
1108                 }
1109
1110                 if (usenext) {
1111                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1112                             dsl_dataset_phys(ds)->ds_next_clones_obj,
1113                             scn->scn_phys.scn_queue_obj,
1114                             dsl_dataset_phys(ds)->ds_creation_txg, tx));
1115                 } else {
1116                         struct enqueue_clones_arg eca;
1117                         eca.tx = tx;
1118                         eca.originobj = ds->ds_object;
1119
1120                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1121                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1122                 }
1123         }
1124
1125 out:
1126         dsl_dataset_rele(ds, FTAG);
1127 }
1128
1129 /* ARGSUSED */
1130 static int
1131 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1132 {
1133         dmu_tx_t *tx = arg;
1134         dsl_dataset_t *ds;
1135         int err;
1136         dsl_scan_t *scn = dp->dp_scan;
1137
1138         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1139         if (err)
1140                 return (err);
1141
1142         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1143                 dsl_dataset_t *prev;
1144                 err = dsl_dataset_hold_obj(dp,
1145                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1146                 if (err) {
1147                         dsl_dataset_rele(ds, FTAG);
1148                         return (err);
1149                 }
1150
1151                 /*
1152                  * If this is a clone, we don't need to worry about it for now.
1153                  */
1154                 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1155                         dsl_dataset_rele(ds, FTAG);
1156                         dsl_dataset_rele(prev, FTAG);
1157                         return (0);
1158                 }
1159                 dsl_dataset_rele(ds, FTAG);
1160                 ds = prev;
1161         }
1162
1163         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1164             ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1165         dsl_dataset_rele(ds, FTAG);
1166         return (0);
1167 }
1168
1169 /*
1170  * Scrub/dedup interaction.
1171  *
1172  * If there are N references to a deduped block, we don't want to scrub it
1173  * N times -- ideally, we should scrub it exactly once.
1174  *
1175  * We leverage the fact that the dde's replication class (enum ddt_class)
1176  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1177  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1178  *
1179  * To prevent excess scrubbing, the scrub begins by walking the DDT
1180  * to find all blocks with refcnt > 1, and scrubs each of these once.
1181  * Since there are two replication classes which contain blocks with
1182  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1183  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1184  *
1185  * There would be nothing more to say if a block's refcnt couldn't change
1186  * during a scrub, but of course it can so we must account for changes
1187  * in a block's replication class.
1188  *
1189  * Here's an example of what can occur:
1190  *
1191  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1192  * when visited during the top-down scrub phase, it will be scrubbed twice.
1193  * This negates our scrub optimization, but is otherwise harmless.
1194  *
1195  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1196  * on each visit during the top-down scrub phase, it will never be scrubbed.
1197  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1198  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1199  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1200  * while a scrub is in progress, it scrubs the block right then.
1201  */
1202 static void
1203 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1204 {
1205         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1206         ddt_entry_t dde = { 0 };
1207         int error;
1208         uint64_t n = 0;
1209
1210         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1211                 ddt_t *ddt;
1212
1213                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1214                         break;
1215                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1216                     (longlong_t)ddb->ddb_class,
1217                     (longlong_t)ddb->ddb_type,
1218                     (longlong_t)ddb->ddb_checksum,
1219                     (longlong_t)ddb->ddb_cursor);
1220
1221                 /* There should be no pending changes to the dedup table */
1222                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1223                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1224
1225                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1226                 n++;
1227
1228                 if (dsl_scan_check_pause(scn, NULL))
1229                         break;
1230         }
1231
1232         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1233             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1234             (int)scn->scn_pausing);
1235
1236         ASSERT(error == 0 || error == ENOENT);
1237         ASSERT(error != ENOENT ||
1238             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1239 }
1240
1241 /* ARGSUSED */
1242 void
1243 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1244     ddt_entry_t *dde, dmu_tx_t *tx)
1245 {
1246         const ddt_key_t *ddk = &dde->dde_key;
1247         ddt_phys_t *ddp = dde->dde_phys;
1248         blkptr_t bp;
1249         zbookmark_phys_t zb = { 0 };
1250
1251         if (scn->scn_phys.scn_state != DSS_SCANNING)
1252                 return;
1253
1254         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1255                 if (ddp->ddp_phys_birth == 0 ||
1256                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1257                         continue;
1258                 ddt_bp_create(checksum, ddk, ddp, &bp);
1259
1260                 scn->scn_visited_this_txg++;
1261                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1262         }
1263 }
1264
1265 static void
1266 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1267 {
1268         dsl_pool_t *dp = scn->scn_dp;
1269         zap_cursor_t zc;
1270         zap_attribute_t za;
1271
1272         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1273             scn->scn_phys.scn_ddt_class_max) {
1274                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1275                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1276                 dsl_scan_ddt(scn, tx);
1277                 if (scn->scn_pausing)
1278                         return;
1279         }
1280
1281         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1282                 /* First do the MOS & ORIGIN */
1283
1284                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1285                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1286                 dsl_scan_visit_rootbp(scn, NULL,
1287                     &dp->dp_meta_rootbp, tx);
1288                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1289                 if (scn->scn_pausing)
1290                         return;
1291
1292                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1293                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1294                             enqueue_cb, tx, DS_FIND_CHILDREN));
1295                 } else {
1296                         dsl_scan_visitds(scn,
1297                             dp->dp_origin_snap->ds_object, tx);
1298                 }
1299                 ASSERT(!scn->scn_pausing);
1300         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1301             ZB_DESTROYED_OBJSET) {
1302                 /*
1303                  * If we were paused, continue from here.  Note if the
1304                  * ds we were paused on was deleted, the zb_objset may
1305                  * be -1, so we will skip this and find a new objset
1306                  * below.
1307                  */
1308                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1309                 if (scn->scn_pausing)
1310                         return;
1311         }
1312
1313         /*
1314          * In case we were paused right at the end of the ds, zero the
1315          * bookmark so we don't think that we're still trying to resume.
1316          */
1317         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1318
1319         /* keep pulling things out of the zap-object-as-queue */
1320         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1321             scn->scn_phys.scn_queue_obj),
1322             zap_cursor_retrieve(&zc, &za) == 0) {
1323                 dsl_dataset_t *ds;
1324                 uint64_t dsobj;
1325
1326                 dsobj = strtonum(za.za_name, NULL);
1327                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1328                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1329
1330                 /* Set up min/max txg */
1331                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1332                 if (za.za_first_integer != 0) {
1333                         scn->scn_phys.scn_cur_min_txg =
1334                             MAX(scn->scn_phys.scn_min_txg,
1335                             za.za_first_integer);
1336                 } else {
1337                         scn->scn_phys.scn_cur_min_txg =
1338                             MAX(scn->scn_phys.scn_min_txg,
1339                             dsl_dataset_phys(ds)->ds_prev_snap_txg);
1340                 }
1341                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1342                 dsl_dataset_rele(ds, FTAG);
1343
1344                 dsl_scan_visitds(scn, dsobj, tx);
1345                 zap_cursor_fini(&zc);
1346                 if (scn->scn_pausing)
1347                         return;
1348         }
1349         zap_cursor_fini(&zc);
1350 }
1351
1352 static boolean_t
1353 dsl_scan_free_should_pause(dsl_scan_t *scn)
1354 {
1355         uint64_t elapsed_nanosecs;
1356
1357         if (zfs_recover)
1358                 return (B_FALSE);
1359
1360         if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1361                 return (B_TRUE);
1362
1363         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1364         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1365             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1366             txg_sync_waiting(scn->scn_dp)) ||
1367             spa_shutting_down(scn->scn_dp->dp_spa));
1368 }
1369
1370 static int
1371 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1372 {
1373         dsl_scan_t *scn = arg;
1374
1375         if (!scn->scn_is_bptree ||
1376             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1377                 if (dsl_scan_free_should_pause(scn))
1378                         return (SET_ERROR(ERESTART));
1379         }
1380
1381         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1382             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1383         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1384             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1385             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1386         scn->scn_visited_this_txg++;
1387         return (0);
1388 }
1389
1390 boolean_t
1391 dsl_scan_active(dsl_scan_t *scn)
1392 {
1393         spa_t *spa = scn->scn_dp->dp_spa;
1394         uint64_t used = 0, comp, uncomp;
1395
1396         if (spa->spa_load_state != SPA_LOAD_NONE)
1397                 return (B_FALSE);
1398         if (spa_shutting_down(spa))
1399                 return (B_FALSE);
1400         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1401             (scn->scn_async_destroying && !scn->scn_async_stalled))
1402                 return (B_TRUE);
1403
1404         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1405                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1406                     &used, &comp, &uncomp);
1407         }
1408         return (used != 0);
1409 }
1410
1411 void
1412 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1413 {
1414         dsl_scan_t *scn = dp->dp_scan;
1415         spa_t *spa = dp->dp_spa;
1416         int err = 0;
1417
1418         /*
1419          * Check for scn_restart_txg before checking spa_load_state, so
1420          * that we can restart an old-style scan while the pool is being
1421          * imported (see dsl_scan_init).
1422          */
1423         if (scn->scn_restart_txg != 0 &&
1424             scn->scn_restart_txg <= tx->tx_txg) {
1425                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1426                 dsl_scan_done(scn, B_FALSE, tx);
1427                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1428                         func = POOL_SCAN_RESILVER;
1429                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1430                     func, tx->tx_txg);
1431                 dsl_scan_setup_sync(&func, tx);
1432         }
1433
1434         /*
1435          * If the scan is inactive due to a stalled async destroy, try again.
1436          */
1437         if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1438             spa_sync_pass(dp->dp_spa) > 1)
1439                 return;
1440
1441         scn->scn_visited_this_txg = 0;
1442         scn->scn_pausing = B_FALSE;
1443         scn->scn_sync_start_time = gethrtime();
1444         spa->spa_scrub_active = B_TRUE;
1445
1446         /*
1447          * First process the async destroys.  If we pause, don't do
1448          * any scrubbing or resilvering.  This ensures that there are no
1449          * async destroys while we are scanning, so the scan code doesn't
1450          * have to worry about traversing it.  It is also faster to free the
1451          * blocks than to scrub them.
1452          */
1453         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1454                 scn->scn_is_bptree = B_FALSE;
1455                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1456                     NULL, ZIO_FLAG_MUSTSUCCEED);
1457                 err = bpobj_iterate(&dp->dp_free_bpobj,
1458                     dsl_scan_free_block_cb, scn, tx);
1459                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1460
1461                 if (err != 0 && err != ERESTART)
1462                         zfs_panic_recover("error %u from bpobj_iterate()", err);
1463         }
1464
1465         if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1466                 ASSERT(scn->scn_async_destroying);
1467                 scn->scn_is_bptree = B_TRUE;
1468                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1469                     NULL, ZIO_FLAG_MUSTSUCCEED);
1470                 err = bptree_iterate(dp->dp_meta_objset,
1471                     dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1472                 VERIFY0(zio_wait(scn->scn_zio_root));
1473
1474                 if (err == EIO || err == ECKSUM) {
1475                         err = 0;
1476                 } else if (err != 0 && err != ERESTART) {
1477                         zfs_panic_recover("error %u from "
1478                             "traverse_dataset_destroyed()", err);
1479                 }
1480
1481                 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1482                         /* finished; deactivate async destroy feature */
1483                         spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1484                         ASSERT(!spa_feature_is_active(spa,
1485                             SPA_FEATURE_ASYNC_DESTROY));
1486                         VERIFY0(zap_remove(dp->dp_meta_objset,
1487                             DMU_POOL_DIRECTORY_OBJECT,
1488                             DMU_POOL_BPTREE_OBJ, tx));
1489                         VERIFY0(bptree_free(dp->dp_meta_objset,
1490                             dp->dp_bptree_obj, tx));
1491                         dp->dp_bptree_obj = 0;
1492                         scn->scn_async_destroying = B_FALSE;
1493                         scn->scn_async_stalled = B_FALSE;
1494                 } else {
1495                         /*
1496                          * If we didn't make progress, mark the async
1497                          * destroy as stalled, so that we will not initiate
1498                          * a spa_sync() on its behalf.  Note that we only
1499                          * check this if we are not finished, because if the
1500                          * bptree had no blocks for us to visit, we can
1501                          * finish without "making progress".
1502                          */
1503                         scn->scn_async_stalled =
1504                             (scn->scn_visited_this_txg == 0);
1505                 }
1506         }
1507         if (scn->scn_visited_this_txg) {
1508                 zfs_dbgmsg("freed %llu blocks in %llums from "
1509                     "free_bpobj/bptree txg %llu; err=%d",
1510                     (longlong_t)scn->scn_visited_this_txg,
1511                     (longlong_t)
1512                     NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1513                     (longlong_t)tx->tx_txg, err);
1514                 scn->scn_visited_this_txg = 0;
1515
1516                 /*
1517                  * Write out changes to the DDT that may be required as a
1518                  * result of the blocks freed.  This ensures that the DDT
1519                  * is clean when a scrub/resilver runs.
1520                  */
1521                 ddt_sync(spa, tx->tx_txg);
1522         }
1523         if (err != 0)
1524                 return;
1525         if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1526             (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1527             dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1528             dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1529                 /*
1530                  * We have finished background destroying, but there is still
1531                  * some space left in the dp_free_dir. Transfer this leaked
1532                  * space to the dp_leak_dir.
1533                  */
1534                 if (dp->dp_leak_dir == NULL) {
1535                         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1536                         (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1537                             LEAK_DIR_NAME, tx);
1538                         VERIFY0(dsl_pool_open_special_dir(dp,
1539                             LEAK_DIR_NAME, &dp->dp_leak_dir));
1540                         rrw_exit(&dp->dp_config_rwlock, FTAG);
1541                 }
1542                 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1543                     dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1544                     dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1545                     dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1546                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1547                     -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1548                     -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1549                     -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1550         }
1551         if (!scn->scn_async_destroying) {
1552                 /* finished; verify that space accounting went to zero */
1553                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1554                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1555                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1556         }
1557
1558         if (scn->scn_phys.scn_state != DSS_SCANNING)
1559                 return;
1560
1561         if (scn->scn_done_txg == tx->tx_txg) {
1562                 ASSERT(!scn->scn_pausing);
1563                 /* finished with scan. */
1564                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1565                 dsl_scan_done(scn, B_TRUE, tx);
1566                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1567                 dsl_scan_sync_state(scn, tx);
1568                 return;
1569         }
1570
1571         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1572             scn->scn_phys.scn_ddt_class_max) {
1573                 zfs_dbgmsg("doing scan sync txg %llu; "
1574                     "ddt bm=%llu/%llu/%llu/%llx",
1575                     (longlong_t)tx->tx_txg,
1576                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1577                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1578                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1579                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1580                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1581                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1582                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1583                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1584         } else {
1585                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1586                     (longlong_t)tx->tx_txg,
1587                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1588                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1589                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1590                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1591         }
1592
1593         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1594             NULL, ZIO_FLAG_CANFAIL);
1595         dsl_pool_config_enter(dp, FTAG);
1596         dsl_scan_visit(scn, tx);
1597         dsl_pool_config_exit(dp, FTAG);
1598         (void) zio_wait(scn->scn_zio_root);
1599         scn->scn_zio_root = NULL;
1600
1601         zfs_dbgmsg("visited %llu blocks in %llums",
1602             (longlong_t)scn->scn_visited_this_txg,
1603             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1604
1605         if (!scn->scn_pausing) {
1606                 scn->scn_done_txg = tx->tx_txg + 1;
1607                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1608                     tx->tx_txg, scn->scn_done_txg);
1609         }
1610
1611         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1612                 mutex_enter(&spa->spa_scrub_lock);
1613                 while (spa->spa_scrub_inflight > 0) {
1614                         cv_wait(&spa->spa_scrub_io_cv,
1615                             &spa->spa_scrub_lock);
1616                 }
1617                 mutex_exit(&spa->spa_scrub_lock);
1618         }
1619
1620         dsl_scan_sync_state(scn, tx);
1621 }
1622
1623 /*
1624  * This will start a new scan, or restart an existing one.
1625  */
1626 void
1627 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1628 {
1629         if (txg == 0) {
1630                 dmu_tx_t *tx;
1631                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1632                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1633
1634                 txg = dmu_tx_get_txg(tx);
1635                 dp->dp_scan->scn_restart_txg = txg;
1636                 dmu_tx_commit(tx);
1637         } else {
1638                 dp->dp_scan->scn_restart_txg = txg;
1639         }
1640         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1641 }
1642
1643 boolean_t
1644 dsl_scan_resilvering(dsl_pool_t *dp)
1645 {
1646         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1647             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1648 }
1649
1650 /*
1651  * scrub consumers
1652  */
1653
1654 static void
1655 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1656 {
1657         int i;
1658
1659         /*
1660          * If we resume after a reboot, zab will be NULL; don't record
1661          * incomplete stats in that case.
1662          */
1663         if (zab == NULL)
1664                 return;
1665
1666         for (i = 0; i < 4; i++) {
1667                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1668                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1669                 if (t & DMU_OT_NEWTYPE)
1670                         t = DMU_OT_OTHER;
1671                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1672                 int equal;
1673
1674                 zb->zb_count++;
1675                 zb->zb_asize += BP_GET_ASIZE(bp);
1676                 zb->zb_lsize += BP_GET_LSIZE(bp);
1677                 zb->zb_psize += BP_GET_PSIZE(bp);
1678                 zb->zb_gangs += BP_COUNT_GANG(bp);
1679
1680                 switch (BP_GET_NDVAS(bp)) {
1681                 case 2:
1682                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1683                             DVA_GET_VDEV(&bp->blk_dva[1]))
1684                                 zb->zb_ditto_2_of_2_samevdev++;
1685                         break;
1686                 case 3:
1687                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1688                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1689                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1690                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1691                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1692                             DVA_GET_VDEV(&bp->blk_dva[2]));
1693                         if (equal == 1)
1694                                 zb->zb_ditto_2_of_3_samevdev++;
1695                         else if (equal == 3)
1696                                 zb->zb_ditto_3_of_3_samevdev++;
1697                         break;
1698                 }
1699         }
1700 }
1701
1702 static void
1703 dsl_scan_scrub_done(zio_t *zio)
1704 {
1705         spa_t *spa = zio->io_spa;
1706
1707         zio_data_buf_free(zio->io_data, zio->io_size);
1708
1709         mutex_enter(&spa->spa_scrub_lock);
1710         spa->spa_scrub_inflight--;
1711         cv_broadcast(&spa->spa_scrub_io_cv);
1712
1713         if (zio->io_error && (zio->io_error != ECKSUM ||
1714             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1715                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1716         }
1717         mutex_exit(&spa->spa_scrub_lock);
1718 }
1719
1720 static int
1721 dsl_scan_scrub_cb(dsl_pool_t *dp,
1722     const blkptr_t *bp, const zbookmark_phys_t *zb)
1723 {
1724         dsl_scan_t *scn = dp->dp_scan;
1725         size_t size = BP_GET_PSIZE(bp);
1726         spa_t *spa = dp->dp_spa;
1727         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1728         boolean_t needs_io;
1729         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1730         unsigned int scan_delay = 0;
1731
1732         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1733             phys_birth >= scn->scn_phys.scn_max_txg)
1734                 return (0);
1735
1736         count_block(dp->dp_blkstats, bp);
1737
1738         if (BP_IS_EMBEDDED(bp))
1739                 return (0);
1740
1741         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1742         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1743                 zio_flags |= ZIO_FLAG_SCRUB;
1744                 needs_io = B_TRUE;
1745                 scan_delay = zfs_scrub_delay;
1746         } else {
1747                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1748                 zio_flags |= ZIO_FLAG_RESILVER;
1749                 needs_io = B_FALSE;
1750                 scan_delay = zfs_resilver_delay;
1751         }
1752
1753         /* If it's an intent log block, failure is expected. */
1754         if (zb->zb_level == ZB_ZIL_LEVEL)
1755                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1756
1757         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1758                 vdev_t *vd = vdev_lookup_top(spa,
1759                     DVA_GET_VDEV(&bp->blk_dva[d]));
1760
1761                 /*
1762                  * Keep track of how much data we've examined so that
1763                  * zpool(1M) status can make useful progress reports.
1764                  */
1765                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1766                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1767
1768                 /* if it's a resilver, this may not be in the target range */
1769                 if (!needs_io) {
1770                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1771                                 /*
1772                                  * Gang members may be spread across multiple
1773                                  * vdevs, so the best estimate we have is the
1774                                  * scrub range, which has already been checked.
1775                                  * XXX -- it would be better to change our
1776                                  * allocation policy to ensure that all
1777                                  * gang members reside on the same vdev.
1778                                  */
1779                                 needs_io = B_TRUE;
1780                         } else {
1781                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1782                                     phys_birth, 1);
1783                         }
1784                 }
1785         }
1786
1787         if (needs_io && !zfs_no_scrub_io) {
1788                 vdev_t *rvd = spa->spa_root_vdev;
1789                 uint64_t maxinflight = rvd->vdev_children *
1790                     MAX(zfs_top_maxinflight, 1);
1791                 void *data = zio_data_buf_alloc(size);
1792
1793                 mutex_enter(&spa->spa_scrub_lock);
1794                 while (spa->spa_scrub_inflight >= maxinflight)
1795                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1796                 spa->spa_scrub_inflight++;
1797                 mutex_exit(&spa->spa_scrub_lock);
1798
1799                 /*
1800                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1801                  * then throttle our workload to limit the impact of a scan.
1802                  */
1803                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1804                         delay(MAX((int)scan_delay, 0));
1805
1806                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1807                     dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1808                     zio_flags, zb));
1809         }
1810
1811         /* do not relocate this block */
1812         return (0);
1813 }
1814
1815 int
1816 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1817 {
1818         spa_t *spa = dp->dp_spa;
1819
1820         /*
1821          * Purge all vdev caches and probe all devices.  We do this here
1822          * rather than in sync context because this requires a writer lock
1823          * on the spa_config lock, which we can't do from sync context.  The
1824          * spa_scrub_reopen flag indicates that vdev_open() should not
1825          * attempt to start another scrub.
1826          */
1827         spa_vdev_state_enter(spa, SCL_NONE);
1828         spa->spa_scrub_reopen = B_TRUE;
1829         vdev_reopen(spa->spa_root_vdev);
1830         spa->spa_scrub_reopen = B_FALSE;
1831         (void) spa_vdev_state_exit(spa, NULL, 0);
1832
1833         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1834             dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1835 }