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