sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_scan.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
  24  * Copyright 2016 Gary Mills
  25  */
  26
  27 #include <sys/dsl_scan.h>
  28 #include <sys/dsl_pool.h>
  29 #include <sys/dsl_dataset.h>
  30 #include <sys/dsl_prop.h>
  31 #include <sys/dsl_dir.h>
  32 #include <sys/dsl_synctask.h>
  33 #include <sys/dnode.h>
  34 #include <sys/dmu_tx.h>
  35 #include <sys/dmu_objset.h>
  36 #include <sys/arc.h>
  37 #include <sys/zap.h>
  38 #include <sys/zio.h>
  39 #include <sys/zfs_context.h>
  40 #include <sys/fs/zfs.h>
  41 #include <sys/zfs_znode.h>
  42 #include <sys/spa_impl.h>
  43 #include <sys/vdev_impl.h>
  44 #include <sys/zil_impl.h>
  45 #include <sys/zio_checksum.h>
  46 #include <sys/ddt.h>
  47 #include <sys/sa.h>
  48 #include <sys/sa_impl.h>
  49 #include <sys/zfeature.h>
  50 #ifdef _KERNEL
  51 #include <sys/zfs_vfsops.h>
  52 #endif
  53
  54 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
  55     const zbookmark_phys_t *);
  56
  57 static scan_cb_t dsl_scan_scrub_cb;
  58 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
  59 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
  60
  61 unsigned int zfs_top_maxinflight = 32;  /* maximum I/Os per top-level */
  62 unsigned int zfs_resilver_delay = 2;    /* number of ticks to delay resilver */
  63 unsigned int zfs_scrub_delay = 4;       /* number of ticks to delay scrub */
  64 unsigned int zfs_scan_idle = 50;        /* idle window in clock ticks */
  65
  66 unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
  67 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
  68 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
  69                                                  per txg */
  70 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  71 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
  72
  73 SYSCTL_DECL(_vfs_zfs);
  74 TUNABLE_INT("vfs.zfs.top_maxinflight", &zfs_top_maxinflight);
  75 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RW,
  76     &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
  77 TUNABLE_INT("vfs.zfs.resilver_delay", &zfs_resilver_delay);
  78 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RW,
  79     &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
  80 TUNABLE_INT("vfs.zfs.scrub_delay", &zfs_scrub_delay);
  81 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RW,
  82     &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
  83 TUNABLE_INT("vfs.zfs.scan_idle", &zfs_scan_idle);
  84 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RW,
  85     &zfs_scan_idle, 0, "Idle scan window in clock ticks");
  86 TUNABLE_INT("vfs.zfs.scan_min_time_ms", &zfs_scan_min_time_ms);
  87 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RW,
  88     &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
  89 TUNABLE_INT("vfs.zfs.free_min_time_ms", &zfs_free_min_time_ms);
  90 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RW,
  91     &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
  92 TUNABLE_INT("vfs.zfs.resilver_min_time_ms", &zfs_resilver_min_time_ms);
  93 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RW,
  94     &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
  95 TUNABLE_INT("vfs.zfs.no_scrub_io", &zfs_no_scrub_io);
  96 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RW,
  97     &zfs_no_scrub_io, 0, "Disable scrub I/O");
  98 TUNABLE_INT("vfs.zfs.no_scrub_prefetch", &zfs_no_scrub_prefetch);
  99 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RW,
 100     &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
 101
 102 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
 103 /* max number of blocks to free in a single TXG */
 104 uint64_t zfs_free_max_blocks = UINT64_MAX;
 105 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
 106     &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
 107
 108
 109 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
 110         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
 111         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
 112
 113 extern int zfs_txg_timeout;
 114
 115 /*
 116  * Enable/disable the processing of the free_bpobj object.
 117  */
 118 boolean_t zfs_free_bpobj_enabled = B_TRUE;
 119
 120 SYSCTL_INT(_vfs_zfs, OID_AUTO, free_bpobj_enabled, CTLFLAG_RWTUN,
 121     &zfs_free_bpobj_enabled, 0, "Enable free_bpobj processing");
 122
 123 /* the order has to match pool_scan_type */
 124 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
 125         NULL,
 126         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
 127         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
 128 };
 129
 130 int
 131 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
 132 {
 133         int err;
 134         dsl_scan_t *scn;
 135         spa_t *spa = dp->dp_spa;
 136         uint64_t f;
 137
 138         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
 139         scn->scn_dp = dp;
 140
 141         /*
 142          * It's possible that we're resuming a scan after a reboot so
 143          * make sure that the scan_async_destroying flag is initialized
 144          * appropriately.
 145          */
 146         ASSERT(!scn->scn_async_destroying);
 147         scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
 148             SPA_FEATURE_ASYNC_DESTROY);
 149
 150         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 151             "scrub_func", sizeof (uint64_t), 1, &f);
 152         if (err == 0) {
 153                 /*
 154                  * There was an old-style scrub in progress.  Restart a
 155                  * new-style scrub from the beginning.
 156                  */
 157                 scn->scn_restart_txg = txg;
 158                 zfs_dbgmsg("old-style scrub was in progress; "
 159                     "restarting new-style scrub in txg %llu",
 160                     scn->scn_restart_txg);
 161
 162                 /*
 163                  * Load the queue obj from the old location so that it
 164                  * can be freed by dsl_scan_done().
 165                  */
 166                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 167                     "scrub_queue", sizeof (uint64_t), 1,
 168                     &scn->scn_phys.scn_queue_obj);
 169         } else {
 170                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 171                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 172                     &scn->scn_phys);
 173                 if (err == ENOENT)
 174                         return (0);
 175                 else if (err)
 176                         return (err);
 177
 178                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 179                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 180                         /*
 181                          * A new-type scrub was in progress on an old
 182                          * pool, and the pool was accessed by old
 183                          * software.  Restart from the beginning, since
 184                          * the old software may have changed the pool in
 185                          * the meantime.
 186                          */
 187                         scn->scn_restart_txg = txg;
 188                         zfs_dbgmsg("new-style scrub was modified "
 189                             "by old software; restarting in txg %llu",
 190                             scn->scn_restart_txg);
 191                 }
 192         }
 193
 194         spa_scan_stat_init(spa);
 195         return (0);
 196 }
 197
 198 void
 199 dsl_scan_fini(dsl_pool_t *dp)
 200 {
 201         if (dp->dp_scan) {
 202                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 203                 dp->dp_scan = NULL;
 204         }
 205 }
 206
 207 /* ARGSUSED */
 208 static int
 209 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 210 {
 211         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 212
 213         if (scn->scn_phys.scn_state == DSS_SCANNING)
 214                 return (SET_ERROR(EBUSY));
 215
 216         return (0);
 217 }
 218
 219 static void
 220 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 221 {
 222         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 223         pool_scan_func_t *funcp = arg;
 224         dmu_object_type_t ot = 0;
 225         dsl_pool_t *dp = scn->scn_dp;
 226         spa_t *spa = dp->dp_spa;
 227
 228         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 229         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 230         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 231         scn->scn_phys.scn_func = *funcp;
 232         scn->scn_phys.scn_state = DSS_SCANNING;
 233         scn->scn_phys.scn_min_txg = 0;
 234         scn->scn_phys.scn_max_txg = tx->tx_txg;
 235         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 236         scn->scn_phys.scn_start_time = gethrestime_sec();
 237         scn->scn_phys.scn_errors = 0;
 238         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 239         scn->scn_restart_txg = 0;
 240         scn->scn_done_txg = 0;
 241         spa_scan_stat_init(spa);
 242
 243         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 244                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 245
 246                 /* rewrite all disk labels */
 247                 vdev_config_dirty(spa->spa_root_vdev);
 248
 249                 if (vdev_resilver_needed(spa->spa_root_vdev,
 250                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 251                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 252                 } else {
 253                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 254                 }
 255
 256                 spa->spa_scrub_started = B_TRUE;
 257                 /*
 258                  * If this is an incremental scrub, limit the DDT scrub phase
 259                  * to just the auto-ditto class (for correctness); the rest
 260                  * of the scrub should go faster using top-down pruning.
 261                  */
 262                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 263                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 264
 265         }
 266
 267         /* back to the generic stuff */
 268
 269         if (dp->dp_blkstats == NULL) {
 270                 dp->dp_blkstats =
 271                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 272         }
 273         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 274
 275         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 276                 ot = DMU_OT_ZAP_OTHER;
 277
 278         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 279             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 280
 281         dsl_scan_sync_state(scn, tx);
 282
 283         spa_history_log_internal(spa, "scan setup", tx,
 284             "func=%u mintxg=%llu maxtxg=%llu",
 285             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 286 }
 287
 288 /* ARGSUSED */
 289 static void
 290 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 291 {
 292         static const char *old_names[] = {
 293                 "scrub_bookmark",
 294                 "scrub_ddt_bookmark",
 295                 "scrub_ddt_class_max",
 296                 "scrub_queue",
 297                 "scrub_min_txg",
 298                 "scrub_max_txg",
 299                 "scrub_func",
 300                 "scrub_errors",
 301                 NULL
 302         };
 303
 304         dsl_pool_t *dp = scn->scn_dp;
 305         spa_t *spa = dp->dp_spa;
 306         int i;
 307
 308         /* Remove any remnants of an old-style scrub. */
 309         for (i = 0; old_names[i]; i++) {
 310                 (void) zap_remove(dp->dp_meta_objset,
 311                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 312         }
 313
 314         if (scn->scn_phys.scn_queue_obj != 0) {
 315                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 316                     scn->scn_phys.scn_queue_obj, tx));
 317                 scn->scn_phys.scn_queue_obj = 0;
 318         }
 319
 320         /*
 321          * If we were "restarted" from a stopped state, don't bother
 322          * with anything else.
 323          */
 324         if (scn->scn_phys.scn_state != DSS_SCANNING)
 325                 return;
 326
 327         if (complete)
 328                 scn->scn_phys.scn_state = DSS_FINISHED;
 329         else
 330                 scn->scn_phys.scn_state = DSS_CANCELED;
 331
 332         spa_history_log_internal(spa, "scan done", tx,
 333             "complete=%u", complete);
 334
 335         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 336                 mutex_enter(&spa->spa_scrub_lock);
 337                 while (spa->spa_scrub_inflight > 0) {
 338                         cv_wait(&spa->spa_scrub_io_cv,
 339                             &spa->spa_scrub_lock);
 340                 }
 341                 mutex_exit(&spa->spa_scrub_lock);
 342                 spa->spa_scrub_started = B_FALSE;
 343                 spa->spa_scrub_active = B_FALSE;
 344
 345                 /*
 346                  * If the scrub/resilver completed, update all DTLs to
 347                  * reflect this.  Whether it succeeded or not, vacate
 348                  * all temporary scrub DTLs.
 349                  */
 350                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 351                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 352                 if (complete) {
 353                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 354                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 355                 }
 356                 spa_errlog_rotate(spa);
 357
 358                 /*
 359                  * We may have finished replacing a device.
 360                  * Let the async thread assess this and handle the detach.
 361                  */
 362                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 363         }
 364
 365         scn->scn_phys.scn_end_time = gethrestime_sec();
 366 }
 367
 368 /* ARGSUSED */
 369 static int
 370 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 371 {
 372         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 373
 374         if (scn->scn_phys.scn_state != DSS_SCANNING)
 375                 return (SET_ERROR(ENOENT));
 376         return (0);
 377 }
 378
 379 /* ARGSUSED */
 380 static void
 381 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 382 {
 383         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 384
 385         dsl_scan_done(scn, B_FALSE, tx);
 386         dsl_scan_sync_state(scn, tx);
 387 }
 388
 389 int
 390 dsl_scan_cancel(dsl_pool_t *dp)
 391 {
 392         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 393             dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
 394 }
 395
 396 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 397     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 398     dmu_objset_type_t ostype, dmu_tx_t *tx);
 399 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 400     dmu_objset_type_t ostype,
 401     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
 402
 403 void
 404 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 405 {
 406         zio_free(dp->dp_spa, txg, bp);
 407 }
 408
 409 void
 410 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 411 {
 412         ASSERT(dsl_pool_sync_context(dp));
 413         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 414             pio->io_flags));
 415 }
 416
 417 static uint64_t
 418 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 419 {
 420         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 421         if (ds->ds_is_snapshot)
 422                 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
 423         return (smt);
 424 }
 425
 426 static void
 427 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 428 {
 429         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 430             DMU_POOL_DIRECTORY_OBJECT,
 431             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 432             &scn->scn_phys, tx));
 433 }
 434
 435 extern int zfs_vdev_async_write_active_min_dirty_percent;
 436
 437 static boolean_t
 438 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
 439 {
 440         /* we never skip user/group accounting objects */
 441         if (zb && (int64_t)zb->zb_object < 0)
 442                 return (B_FALSE);
 443
 444         if (scn->scn_pausing)
 445                 return (B_TRUE); /* we're already pausing */
 446
 447         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 448                 return (B_FALSE); /* we're resuming */
 449
 450         /* We only know how to resume from level-0 blocks. */
 451         if (zb && zb->zb_level != 0)
 452                 return (B_FALSE);
 453
 454         /*
 455          * We pause if:
 456          *  - we have scanned for the maximum time: an entire txg
 457          *    timeout (default 5 sec)
 458          *  or
 459          *  - we have scanned for at least the minimum time (default 1 sec
 460          *    for scrub, 3 sec for resilver), and either we have sufficient
 461          *    dirty data that we are starting to write more quickly
 462          *    (default 30%), or someone is explicitly waiting for this txg
 463          *    to complete.
 464          *  or
 465          *  - the spa is shutting down because this pool is being exported
 466          *    or the machine is rebooting.
 467          */
 468         int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 469             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 470         uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 471         int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
 472         if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
 473             (NSEC2MSEC(elapsed_nanosecs) > mintime &&
 474             (txg_sync_waiting(scn->scn_dp) ||
 475             dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
 476             spa_shutting_down(scn->scn_dp->dp_spa)) {
 477                 if (zb) {
 478                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 479                             (longlong_t)zb->zb_objset,
 480                             (longlong_t)zb->zb_object,
 481                             (longlong_t)zb->zb_level,
 482                             (longlong_t)zb->zb_blkid);
 483                         scn->scn_phys.scn_bookmark = *zb;
 484                 }
 485                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 486                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 487                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 488                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 489                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 490                 scn->scn_pausing = B_TRUE;
 491                 return (B_TRUE);
 492         }
 493         return (B_FALSE);
 494 }
 495
 496 typedef struct zil_scan_arg {
 497         dsl_pool_t      *zsa_dp;
 498         zil_header_t    *zsa_zh;
 499 } zil_scan_arg_t;
 500
 501 /* ARGSUSED */
 502 static int
 503 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 504 {
 505         zil_scan_arg_t *zsa = arg;
 506         dsl_pool_t *dp = zsa->zsa_dp;
 507         dsl_scan_t *scn = dp->dp_scan;
 508         zil_header_t *zh = zsa->zsa_zh;
 509         zbookmark_phys_t zb;
 510
 511         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 512                 return (0);
 513
 514         /*
 515          * One block ("stubby") can be allocated a long time ago; we
 516          * want to visit that one because it has been allocated
 517          * (on-disk) even if it hasn't been claimed (even though for
 518          * scrub there's nothing to do to it).
 519          */
 520         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 521                 return (0);
 522
 523         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 524             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 525
 526         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 527         return (0);
 528 }
 529
 530 /* ARGSUSED */
 531 static int
 532 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 533 {
 534         if (lrc->lrc_txtype == TX_WRITE) {
 535                 zil_scan_arg_t *zsa = arg;
 536                 dsl_pool_t *dp = zsa->zsa_dp;
 537                 dsl_scan_t *scn = dp->dp_scan;
 538                 zil_header_t *zh = zsa->zsa_zh;
 539                 lr_write_t *lr = (lr_write_t *)lrc;
 540                 blkptr_t *bp = &lr->lr_blkptr;
 541                 zbookmark_phys_t zb;
 542
 543                 if (BP_IS_HOLE(bp) ||
 544                     bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 545                         return (0);
 546
 547                 /*
 548                  * birth can be < claim_txg if this record's txg is
 549                  * already txg sync'ed (but this log block contains
 550                  * other records that are not synced)
 551                  */
 552                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 553                         return (0);
 554
 555                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 556                     lr->lr_foid, ZB_ZIL_LEVEL,
 557                     lr->lr_offset / BP_GET_LSIZE(bp));
 558
 559                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 560         }
 561         return (0);
 562 }
 563
 564 static void
 565 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 566 {
 567         uint64_t claim_txg = zh->zh_claim_txg;
 568         zil_scan_arg_t zsa = { dp, zh };
 569         zilog_t *zilog;
 570
 571         /*
 572          * We only want to visit blocks that have been claimed but not yet
 573          * replayed (or, in read-only mode, blocks that *would* be claimed).
 574          */
 575         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 576                 return;
 577
 578         zilog = zil_alloc(dp->dp_meta_objset, zh);
 579
 580         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 581             claim_txg);
 582
 583         zil_free(zilog);
 584 }
 585
 586 /* ARGSUSED */
 587 static void
 588 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 589     uint64_t objset, uint64_t object, uint64_t blkid)
 590 {
 591         zbookmark_phys_t czb;
 592         arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
 593
 594         if (zfs_no_scrub_prefetch)
 595                 return;
 596
 597         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 598             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 599                 return;
 600
 601         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 602
 603         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 604             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 605             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 606 }
 607
 608 static boolean_t
 609 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 610     const zbookmark_phys_t *zb)
 611 {
 612         /*
 613          * We never skip over user/group accounting objects (obj<0)
 614          */
 615         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 616             (int64_t)zb->zb_object >= 0) {
 617                 /*
 618                  * If we already visited this bp & everything below (in
 619                  * a prior txg sync), don't bother doing it again.
 620                  */
 621                 if (zbookmark_subtree_completed(dnp, zb,
 622                     &scn->scn_phys.scn_bookmark))
 623                         return (B_TRUE);
 624
 625                 /*
 626                  * If we found the block we're trying to resume from, or
 627                  * we went past it to a different object, zero it out to
 628                  * indicate that it's OK to start checking for pausing
 629                  * again.
 630                  */
 631                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 632                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 633                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 634                             (longlong_t)zb->zb_objset,
 635                             (longlong_t)zb->zb_object,
 636                             (longlong_t)zb->zb_level,
 637                             (longlong_t)zb->zb_blkid);
 638                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 639                 }
 640         }
 641         return (B_FALSE);
 642 }
 643
 644 /*
 645  * Return nonzero on i/o error.
 646  * Return new buf to write out in *bufp.
 647  */
 648 static int
 649 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 650     dnode_phys_t *dnp, const blkptr_t *bp,
 651     const zbookmark_phys_t *zb, dmu_tx_t *tx)
 652 {
 653         dsl_pool_t *dp = scn->scn_dp;
 654         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 655         int err;
 656
 657         if (BP_GET_LEVEL(bp) > 0) {
 658                 arc_flags_t flags = ARC_FLAG_WAIT;
 659                 int i;
 660                 blkptr_t *cbp;
 661                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 662                 arc_buf_t *buf;
 663
 664                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 665                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 666                 if (err) {
 667                         scn->scn_phys.scn_errors++;
 668                         return (err);
 669                 }
 670                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 671                         dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
 672                             zb->zb_object, zb->zb_blkid * epb + i);
 673                 }
 674                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
 675                         zbookmark_phys_t czb;
 676
 677                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 678                             zb->zb_level - 1,
 679                             zb->zb_blkid * epb + i);
 680                         dsl_scan_visitbp(cbp, &czb, dnp,
 681                             ds, scn, ostype, tx);
 682                 }
 683                 (void) arc_buf_remove_ref(buf, &buf);
 684         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 685                 arc_flags_t flags = ARC_FLAG_WAIT;
 686                 dnode_phys_t *cdnp;
 687                 int i, j;
 688                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 689                 arc_buf_t *buf;
 690
 691                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 692                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 693                 if (err) {
 694                         scn->scn_phys.scn_errors++;
 695                         return (err);
 696                 }
 697                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 698                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 699                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 700                                 dsl_scan_prefetch(scn, buf, cbp,
 701                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 702                         }
 703                 }
 704                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
 705                         dsl_scan_visitdnode(scn, ds, ostype,
 706                             cdnp, zb->zb_blkid * epb + i, tx);
 707                 }
 708
 709                 (void) arc_buf_remove_ref(buf, &buf);
 710         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 711                 arc_flags_t flags = ARC_FLAG_WAIT;
 712                 objset_phys_t *osp;
 713                 arc_buf_t *buf;
 714
 715                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
 716                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 717                 if (err) {
 718                         scn->scn_phys.scn_errors++;
 719                         return (err);
 720                 }
 721
 722                 osp = buf->b_data;
 723
 724                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 725                     &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
 726
 727                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
 728                         /*
 729                          * We also always visit user/group accounting
 730                          * objects, and never skip them, even if we are
 731                          * pausing.  This is necessary so that the space
 732                          * deltas from this txg get integrated.
 733                          */
 734                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 735                             &osp->os_groupused_dnode,
 736                             DMU_GROUPUSED_OBJECT, tx);
 737                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 738                             &osp->os_userused_dnode,
 739                             DMU_USERUSED_OBJECT, tx);
 740                 }
 741                 (void) arc_buf_remove_ref(buf, &buf);
 742         }
 743
 744         return (0);
 745 }
 746
 747 static void
 748 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 749     dmu_objset_type_t ostype, dnode_phys_t *dnp,
 750     uint64_t object, dmu_tx_t *tx)
 751 {
 752         int j;
 753
 754         for (j = 0; j < dnp->dn_nblkptr; j++) {
 755                 zbookmark_phys_t czb;
 756
 757                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 758                     dnp->dn_nlevels - 1, j);
 759                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 760                     &czb, dnp, ds, scn, ostype, tx);
 761         }
 762
 763         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 764                 zbookmark_phys_t czb;
 765                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 766                     0, DMU_SPILL_BLKID);
 767                 dsl_scan_visitbp(&dnp->dn_spill,
 768                     &czb, dnp, ds, scn, ostype, tx);
 769         }
 770 }
 771
 772 /*
 773  * The arguments are in this order because mdb can only print the
 774  * first 5; we want them to be useful.
 775  */
 776 static void
 777 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 778     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
 779     dmu_objset_type_t ostype, dmu_tx_t *tx)
 780 {
 781         dsl_pool_t *dp = scn->scn_dp;
 782         arc_buf_t *buf = NULL;
 783         blkptr_t bp_toread = *bp;
 784
 785         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 786
 787         if (dsl_scan_check_pause(scn, zb))
 788                 return;
 789
 790         if (dsl_scan_check_resume(scn, dnp, zb))
 791                 return;
 792
 793         if (BP_IS_HOLE(bp))
 794                 return;
 795
 796         scn->scn_visited_this_txg++;
 797
 798         dprintf_bp(bp,
 799             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
 800             ds, ds ? ds->ds_object : 0,
 801             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 802             bp);
 803
 804         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 805                 return;
 806
 807         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
 808                 return;
 809
 810         /*
 811          * If dsl_scan_ddt() has aready visited this block, it will have
 812          * already done any translations or scrubbing, so don't call the
 813          * callback again.
 814          */
 815         if (ddt_class_contains(dp->dp_spa,
 816             scn->scn_phys.scn_ddt_class_max, bp)) {
 817                 ASSERT(buf == NULL);
 818                 return;
 819         }
 820
 821         /*
 822          * If this block is from the future (after cur_max_txg), then we
 823          * are doing this on behalf of a deleted snapshot, and we will
 824          * revisit the future block on the next pass of this dataset.
 825          * Don't scan it now unless we need to because something
 826          * under it was modified.
 827          */
 828         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 829                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 830         }
 831 }
 832
 833 static void
 834 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 835     dmu_tx_t *tx)
 836 {
 837         zbookmark_phys_t zb;
 838
 839         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 840             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 841         dsl_scan_visitbp(bp, &zb, NULL,
 842             ds, scn, DMU_OST_NONE, tx);
 843
 844         dprintf_ds(ds, "finished scan%s", "");
 845 }
 846
 847 void
 848 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 849 {
 850         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 851         dsl_scan_t *scn = dp->dp_scan;
 852         uint64_t mintxg;
 853
 854         if (scn->scn_phys.scn_state != DSS_SCANNING)
 855                 return;
 856
 857         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 858                 if (ds->ds_is_snapshot) {
 859                         /*
 860                          * Note:
 861                          *  - scn_cur_{min,max}_txg stays the same.
 862                          *  - Setting the flag is not really necessary if
 863                          *    scn_cur_max_txg == scn_max_txg, because there
 864                          *    is nothing after this snapshot that we care
 865                          *    about.  However, we set it anyway and then
 866                          *    ignore it when we retraverse it in
 867                          *    dsl_scan_visitds().
 868                          */
 869                         scn->scn_phys.scn_bookmark.zb_objset =
 870                             dsl_dataset_phys(ds)->ds_next_snap_obj;
 871                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 872                             "reset zb_objset to %llu",
 873                             (u_longlong_t)ds->ds_object,
 874                             (u_longlong_t)dsl_dataset_phys(ds)->
 875                             ds_next_snap_obj);
 876                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 877                 } else {
 878                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 879                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 880                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 881                             "reset bookmark to -1,0,0,0",
 882                             (u_longlong_t)ds->ds_object);
 883                 }
 884         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 885             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 886                 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
 887                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 888                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 889                 if (ds->ds_is_snapshot) {
 890                         /*
 891                          * We keep the same mintxg; it could be >
 892                          * ds_creation_txg if the previous snapshot was
 893                          * deleted too.
 894                          */
 895                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 896                             scn->scn_phys.scn_queue_obj,
 897                             dsl_dataset_phys(ds)->ds_next_snap_obj,
 898                             mintxg, tx) == 0);
 899                         zfs_dbgmsg("destroying ds %llu; in queue; "
 900                             "replacing with %llu",
 901                             (u_longlong_t)ds->ds_object,
 902                             (u_longlong_t)dsl_dataset_phys(ds)->
 903                             ds_next_snap_obj);
 904                 } else {
 905                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 906                             (u_longlong_t)ds->ds_object);
 907                 }
 908         }
 909
 910         /*
 911          * dsl_scan_sync() should be called after this, and should sync
 912          * out our changed state, but just to be safe, do it here.
 913          */
 914         dsl_scan_sync_state(scn, tx);
 915 }
 916
 917 void
 918 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 919 {
 920         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 921         dsl_scan_t *scn = dp->dp_scan;
 922         uint64_t mintxg;
 923
 924         if (scn->scn_phys.scn_state != DSS_SCANNING)
 925                 return;
 926
 927         ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
 928
 929         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 930                 scn->scn_phys.scn_bookmark.zb_objset =
 931                     dsl_dataset_phys(ds)->ds_prev_snap_obj;
 932                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 933                     "reset zb_objset to %llu",
 934                     (u_longlong_t)ds->ds_object,
 935                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 936         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 937             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 938                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 939                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 940                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 941                     scn->scn_phys.scn_queue_obj,
 942                     dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
 943                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 944                     "replacing with %llu",
 945                     (u_longlong_t)ds->ds_object,
 946                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
 947         }
 948         dsl_scan_sync_state(scn, tx);
 949 }
 950
 951 void
 952 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 953 {
 954         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 955         dsl_scan_t *scn = dp->dp_scan;
 956         uint64_t mintxg;
 957
 958         if (scn->scn_phys.scn_state != DSS_SCANNING)
 959                 return;
 960
 961         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 962                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 963                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 964                     "reset zb_objset to %llu",
 965                     (u_longlong_t)ds1->ds_object,
 966                     (u_longlong_t)ds2->ds_object);
 967         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 968                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 969                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 970                     "reset zb_objset to %llu",
 971                     (u_longlong_t)ds2->ds_object,
 972                     (u_longlong_t)ds1->ds_object);
 973         }
 974
 975         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 976             ds1->ds_object, &mintxg) == 0) {
 977                 int err;
 978
 979                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 980                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
 981                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 982                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 983                 err = zap_add_int_key(dp->dp_meta_objset,
 984                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 985                 VERIFY(err == 0 || err == EEXIST);
 986                 if (err == EEXIST) {
 987                         /* Both were there to begin with */
 988                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 989                             scn->scn_phys.scn_queue_obj,
 990                             ds1->ds_object, mintxg, tx));
 991                 }
 992                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 993                     "replacing with %llu",
 994                     (u_longlong_t)ds1->ds_object,
 995                     (u_longlong_t)ds2->ds_object);
 996         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 997             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 998                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
 999                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
1000                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1001                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
1002                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1003                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
1004                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1005                     "replacing with %llu",
1006                     (u_longlong_t)ds2->ds_object,
1007                     (u_longlong_t)ds1->ds_object);
1008         }
1009
1010         dsl_scan_sync_state(scn, tx);
1011 }
1012
1013 struct enqueue_clones_arg {
1014         dmu_tx_t *tx;
1015         uint64_t originobj;
1016 };
1017
1018 /* ARGSUSED */
1019 static int
1020 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1021 {
1022         struct enqueue_clones_arg *eca = arg;
1023         dsl_dataset_t *ds;
1024         int err;
1025         dsl_scan_t *scn = dp->dp_scan;
1026
1027         if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1028                 return (0);
1029
1030         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1031         if (err)
1032                 return (err);
1033
1034         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1035                 dsl_dataset_t *prev;
1036                 err = dsl_dataset_hold_obj(dp,
1037                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1038
1039                 dsl_dataset_rele(ds, FTAG);
1040                 if (err)
1041                         return (err);
1042                 ds = prev;
1043         }
1044         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1045             scn->scn_phys.scn_queue_obj, ds->ds_object,
1046             dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1047         dsl_dataset_rele(ds, FTAG);
1048         return (0);
1049 }
1050
1051 static void
1052 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1053 {
1054         dsl_pool_t *dp = scn->scn_dp;
1055         dsl_dataset_t *ds;
1056         objset_t *os;
1057
1058         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1059
1060         if (scn->scn_phys.scn_cur_min_txg >=
1061             scn->scn_phys.scn_max_txg) {
1062                 /*
1063                  * This can happen if this snapshot was created after the
1064                  * scan started, and we already completed a previous snapshot
1065                  * that was created after the scan started.  This snapshot
1066                  * only references blocks with:
1067                  *
1068                  *      birth < our ds_creation_txg
1069                  *      cur_min_txg is no less than ds_creation_txg.
1070                  *      We have already visited these blocks.
1071                  * or
1072                  *      birth > scn_max_txg
1073                  *      The scan requested not to visit these blocks.
1074                  *
1075                  * Subsequent snapshots (and clones) can reference our
1076                  * blocks, or blocks with even higher birth times.
1077                  * Therefore we do not need to visit them either,
1078                  * so we do not add them to the work queue.
1079                  *
1080                  * Note that checking for cur_min_txg >= cur_max_txg
1081                  * is not sufficient, because in that case we may need to
1082                  * visit subsequent snapshots.  This happens when min_txg > 0,
1083                  * which raises cur_min_txg.  In this case we will visit
1084                  * this dataset but skip all of its blocks, because the
1085                  * rootbp's birth time is < cur_min_txg.  Then we will
1086                  * add the next snapshots/clones to the work queue.
1087                  */
1088                 char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1089                 dsl_dataset_name(ds, dsname);
1090                 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
1091                     "cur_min_txg (%llu) >= max_txg (%llu)",
1092                     dsobj, dsname,
1093                     scn->scn_phys.scn_cur_min_txg,
1094                     scn->scn_phys.scn_max_txg);
1095                 kmem_free(dsname, MAXNAMELEN);
1096
1097                 goto out;
1098         }
1099
1100         if (dmu_objset_from_ds(ds, &os))
1101                 goto out;
1102
1103         /*
1104          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1105          * snapshots can have ZIL block pointers (which may be the same
1106          * BP as in the head), they must be ignored.  So we traverse the
1107          * ZIL here, rather than in scan_recurse(), because the regular
1108          * snapshot block-sharing rules don't apply to it.
1109          */
1110         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1111                 dsl_scan_zil(dp, &os->os_zil_header);
1112
1113         /*
1114          * Iterate over the bps in this ds.
1115          */
1116         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1117         dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1118
1119         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1120         dsl_dataset_name(ds, dsname);
1121         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1122             "pausing=%u",
1123             (longlong_t)dsobj, dsname,
1124             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1125             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1126             (int)scn->scn_pausing);
1127         kmem_free(dsname, ZFS_MAXNAMELEN);
1128
1129         if (scn->scn_pausing)
1130                 goto out;
1131
1132         /*
1133          * We've finished this pass over this dataset.
1134          */
1135
1136         /*
1137          * If we did not completely visit this dataset, do another pass.
1138          */
1139         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1140                 zfs_dbgmsg("incomplete pass; visiting again");
1141                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1142                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1143                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1144                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1145                 goto out;
1146         }
1147
1148         /*
1149          * Add descendent datasets to work queue.
1150          */
1151         if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1152                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1153                     scn->scn_phys.scn_queue_obj,
1154                     dsl_dataset_phys(ds)->ds_next_snap_obj,
1155                     dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1156         }
1157         if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1158                 boolean_t usenext = B_FALSE;
1159                 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1160                         uint64_t count;
1161                         /*
1162                          * A bug in a previous version of the code could
1163                          * cause upgrade_clones_cb() to not set
1164                          * ds_next_snap_obj when it should, leading to a
1165                          * missing entry.  Therefore we can only use the
1166                          * next_clones_obj when its count is correct.
1167                          */
1168                         int err = zap_count(dp->dp_meta_objset,
1169                             dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1170                         if (err == 0 &&
1171                             count == dsl_dataset_phys(ds)->ds_num_children - 1)
1172                                 usenext = B_TRUE;
1173                 }
1174
1175                 if (usenext) {
1176                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1177                             dsl_dataset_phys(ds)->ds_next_clones_obj,
1178                             scn->scn_phys.scn_queue_obj,
1179                             dsl_dataset_phys(ds)->ds_creation_txg, tx));
1180                 } else {
1181                         struct enqueue_clones_arg eca;
1182                         eca.tx = tx;
1183                         eca.originobj = ds->ds_object;
1184
1185                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1186                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1187                 }
1188         }
1189
1190 out:
1191         dsl_dataset_rele(ds, FTAG);
1192 }
1193
1194 /* ARGSUSED */
1195 static int
1196 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1197 {
1198         dmu_tx_t *tx = arg;
1199         dsl_dataset_t *ds;
1200         int err;
1201         dsl_scan_t *scn = dp->dp_scan;
1202
1203         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1204         if (err)
1205                 return (err);
1206
1207         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1208                 dsl_dataset_t *prev;
1209                 err = dsl_dataset_hold_obj(dp,
1210                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1211                 if (err) {
1212                         dsl_dataset_rele(ds, FTAG);
1213                         return (err);
1214                 }
1215
1216                 /*
1217                  * If this is a clone, we don't need to worry about it for now.
1218                  */
1219                 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1220                         dsl_dataset_rele(ds, FTAG);
1221                         dsl_dataset_rele(prev, FTAG);
1222                         return (0);
1223                 }
1224                 dsl_dataset_rele(ds, FTAG);
1225                 ds = prev;
1226         }
1227
1228         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1229             ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1230         dsl_dataset_rele(ds, FTAG);
1231         return (0);
1232 }
1233
1234 /*
1235  * Scrub/dedup interaction.
1236  *
1237  * If there are N references to a deduped block, we don't want to scrub it
1238  * N times -- ideally, we should scrub it exactly once.
1239  *
1240  * We leverage the fact that the dde's replication class (enum ddt_class)
1241  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1242  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1243  *
1244  * To prevent excess scrubbing, the scrub begins by walking the DDT
1245  * to find all blocks with refcnt > 1, and scrubs each of these once.
1246  * Since there are two replication classes which contain blocks with
1247  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1248  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1249  *
1250  * There would be nothing more to say if a block's refcnt couldn't change
1251  * during a scrub, but of course it can so we must account for changes
1252  * in a block's replication class.
1253  *
1254  * Here's an example of what can occur:
1255  *
1256  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1257  * when visited during the top-down scrub phase, it will be scrubbed twice.
1258  * This negates our scrub optimization, but is otherwise harmless.
1259  *
1260  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1261  * on each visit during the top-down scrub phase, it will never be scrubbed.
1262  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1263  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1264  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1265  * while a scrub is in progress, it scrubs the block right then.
1266  */
1267 static void
1268 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1269 {
1270         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1271         ddt_entry_t dde = { 0 };
1272         int error;
1273         uint64_t n = 0;
1274
1275         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1276                 ddt_t *ddt;
1277
1278                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1279                         break;
1280                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1281                     (longlong_t)ddb->ddb_class,
1282                     (longlong_t)ddb->ddb_type,
1283                     (longlong_t)ddb->ddb_checksum,
1284                     (longlong_t)ddb->ddb_cursor);
1285
1286                 /* There should be no pending changes to the dedup table */
1287                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1288                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1289
1290                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1291                 n++;
1292
1293                 if (dsl_scan_check_pause(scn, NULL))
1294                         break;
1295         }
1296
1297         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1298             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1299             (int)scn->scn_pausing);
1300
1301         ASSERT(error == 0 || error == ENOENT);
1302         ASSERT(error != ENOENT ||
1303             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1304 }
1305
1306 /* ARGSUSED */
1307 void
1308 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1309     ddt_entry_t *dde, dmu_tx_t *tx)
1310 {
1311         const ddt_key_t *ddk = &dde->dde_key;
1312         ddt_phys_t *ddp = dde->dde_phys;
1313         blkptr_t bp;
1314         zbookmark_phys_t zb = { 0 };
1315
1316         if (scn->scn_phys.scn_state != DSS_SCANNING)
1317                 return;
1318
1319         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1320                 if (ddp->ddp_phys_birth == 0 ||
1321                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1322                         continue;
1323                 ddt_bp_create(checksum, ddk, ddp, &bp);
1324
1325                 scn->scn_visited_this_txg++;
1326                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1327         }
1328 }
1329
1330 static void
1331 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1332 {
1333         dsl_pool_t *dp = scn->scn_dp;
1334         zap_cursor_t zc;
1335         zap_attribute_t za;
1336
1337         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1338             scn->scn_phys.scn_ddt_class_max) {
1339                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1340                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1341                 dsl_scan_ddt(scn, tx);
1342                 if (scn->scn_pausing)
1343                         return;
1344         }
1345
1346         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1347                 /* First do the MOS & ORIGIN */
1348
1349                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1350                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1351                 dsl_scan_visit_rootbp(scn, NULL,
1352                     &dp->dp_meta_rootbp, tx);
1353                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1354                 if (scn->scn_pausing)
1355                         return;
1356
1357                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1358                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1359                             enqueue_cb, tx, DS_FIND_CHILDREN));
1360                 } else {
1361                         dsl_scan_visitds(scn,
1362                             dp->dp_origin_snap->ds_object, tx);
1363                 }
1364                 ASSERT(!scn->scn_pausing);
1365         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1366             ZB_DESTROYED_OBJSET) {
1367                 /*
1368                  * If we were paused, continue from here.  Note if the
1369                  * ds we were paused on was deleted, the zb_objset may
1370                  * be -1, so we will skip this and find a new objset
1371                  * below.
1372                  */
1373                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1374                 if (scn->scn_pausing)
1375                         return;
1376         }
1377
1378         /*
1379          * In case we were paused right at the end of the ds, zero the
1380          * bookmark so we don't think that we're still trying to resume.
1381          */
1382         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1383
1384         /* keep pulling things out of the zap-object-as-queue */
1385         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1386             scn->scn_phys.scn_queue_obj),
1387             zap_cursor_retrieve(&zc, &za) == 0) {
1388                 dsl_dataset_t *ds;
1389                 uint64_t dsobj;
1390
1391                 dsobj = strtonum(za.za_name, NULL);
1392                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1393                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1394
1395                 /* Set up min/max txg */
1396                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1397                 if (za.za_first_integer != 0) {
1398                         scn->scn_phys.scn_cur_min_txg =
1399                             MAX(scn->scn_phys.scn_min_txg,
1400                             za.za_first_integer);
1401                 } else {
1402                         scn->scn_phys.scn_cur_min_txg =
1403                             MAX(scn->scn_phys.scn_min_txg,
1404                             dsl_dataset_phys(ds)->ds_prev_snap_txg);
1405                 }
1406                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1407                 dsl_dataset_rele(ds, FTAG);
1408
1409                 dsl_scan_visitds(scn, dsobj, tx);
1410                 zap_cursor_fini(&zc);
1411                 if (scn->scn_pausing)
1412                         return;
1413         }
1414         zap_cursor_fini(&zc);
1415 }
1416
1417 static boolean_t
1418 dsl_scan_free_should_pause(dsl_scan_t *scn)
1419 {
1420         uint64_t elapsed_nanosecs;
1421
1422         if (zfs_recover)
1423                 return (B_FALSE);
1424
1425         if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1426                 return (B_TRUE);
1427
1428         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1429         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1430             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1431             txg_sync_waiting(scn->scn_dp)) ||
1432             spa_shutting_down(scn->scn_dp->dp_spa));
1433 }
1434
1435 static int
1436 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1437 {
1438         dsl_scan_t *scn = arg;
1439
1440         if (!scn->scn_is_bptree ||
1441             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1442                 if (dsl_scan_free_should_pause(scn))
1443                         return (SET_ERROR(ERESTART));
1444         }
1445
1446         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1447             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1448         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1449             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1450             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1451         scn->scn_visited_this_txg++;
1452         return (0);
1453 }
1454
1455 boolean_t
1456 dsl_scan_active(dsl_scan_t *scn)
1457 {
1458         spa_t *spa = scn->scn_dp->dp_spa;
1459         uint64_t used = 0, comp, uncomp;
1460
1461         if (spa->spa_load_state != SPA_LOAD_NONE)
1462                 return (B_FALSE);
1463         if (spa_shutting_down(spa))
1464                 return (B_FALSE);
1465         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1466             (scn->scn_async_destroying && !scn->scn_async_stalled))
1467                 return (B_TRUE);
1468
1469         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1470                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1471                     &used, &comp, &uncomp);
1472         }
1473         return (used != 0);
1474 }
1475
1476 void
1477 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1478 {
1479         dsl_scan_t *scn = dp->dp_scan;
1480         spa_t *spa = dp->dp_spa;
1481         int err = 0;
1482
1483         /*
1484          * Check for scn_restart_txg before checking spa_load_state, so
1485          * that we can restart an old-style scan while the pool is being
1486          * imported (see dsl_scan_init).
1487          */
1488         if (scn->scn_restart_txg != 0 &&
1489             scn->scn_restart_txg <= tx->tx_txg) {
1490                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1491                 dsl_scan_done(scn, B_FALSE, tx);
1492                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1493                         func = POOL_SCAN_RESILVER;
1494                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1495                     func, tx->tx_txg);
1496                 dsl_scan_setup_sync(&func, tx);
1497         }
1498
1499         /*
1500          * Only process scans in sync pass 1.
1501          */
1502         if (spa_sync_pass(dp->dp_spa) > 1)
1503                 return;
1504
1505         /*
1506          * If the spa is shutting down, then stop scanning. This will
1507          * ensure that the scan does not dirty any new data during the
1508          * shutdown phase.
1509          */
1510         if (spa_shutting_down(spa))
1511                 return;
1512
1513         /*
1514          * If the scan is inactive due to a stalled async destroy, try again.
1515          */
1516         if (!scn->scn_async_stalled && !dsl_scan_active(scn))
1517                 return;
1518
1519         scn->scn_visited_this_txg = 0;
1520         scn->scn_pausing = B_FALSE;
1521         scn->scn_sync_start_time = gethrtime();
1522         spa->spa_scrub_active = B_TRUE;
1523
1524         /*
1525          * First process the async destroys.  If we pause, don't do
1526          * any scrubbing or resilvering.  This ensures that there are no
1527          * async destroys while we are scanning, so the scan code doesn't
1528          * have to worry about traversing it.  It is also faster to free the
1529          * blocks than to scrub them.
1530          */
1531         if (zfs_free_bpobj_enabled &&
1532             spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1533                 scn->scn_is_bptree = B_FALSE;
1534                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1535                     NULL, ZIO_FLAG_MUSTSUCCEED);
1536                 err = bpobj_iterate(&dp->dp_free_bpobj,
1537                     dsl_scan_free_block_cb, scn, tx);
1538                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1539
1540                 if (err != 0 && err != ERESTART)
1541                         zfs_panic_recover("error %u from bpobj_iterate()", err);
1542         }
1543
1544         if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1545                 ASSERT(scn->scn_async_destroying);
1546                 scn->scn_is_bptree = B_TRUE;
1547                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1548                     NULL, ZIO_FLAG_MUSTSUCCEED);
1549                 err = bptree_iterate(dp->dp_meta_objset,
1550                     dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1551                 VERIFY0(zio_wait(scn->scn_zio_root));
1552
1553                 if (err == EIO || err == ECKSUM) {
1554                         err = 0;
1555                 } else if (err != 0 && err != ERESTART) {
1556                         zfs_panic_recover("error %u from "
1557                             "traverse_dataset_destroyed()", err);
1558                 }
1559
1560                 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1561                         /* finished; deactivate async destroy feature */
1562                         spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1563                         ASSERT(!spa_feature_is_active(spa,
1564                             SPA_FEATURE_ASYNC_DESTROY));
1565                         VERIFY0(zap_remove(dp->dp_meta_objset,
1566                             DMU_POOL_DIRECTORY_OBJECT,
1567                             DMU_POOL_BPTREE_OBJ, tx));
1568                         VERIFY0(bptree_free(dp->dp_meta_objset,
1569                             dp->dp_bptree_obj, tx));
1570                         dp->dp_bptree_obj = 0;
1571                         scn->scn_async_destroying = B_FALSE;
1572                         scn->scn_async_stalled = B_FALSE;
1573                 } else {
1574                         /*
1575                          * If we didn't make progress, mark the async
1576                          * destroy as stalled, so that we will not initiate
1577                          * a spa_sync() on its behalf.  Note that we only
1578                          * check this if we are not finished, because if the
1579                          * bptree had no blocks for us to visit, we can
1580                          * finish without "making progress".
1581                          */
1582                         scn->scn_async_stalled =
1583                             (scn->scn_visited_this_txg == 0);
1584                 }
1585         }
1586         if (scn->scn_visited_this_txg) {
1587                 zfs_dbgmsg("freed %llu blocks in %llums from "
1588                     "free_bpobj/bptree txg %llu; err=%d",
1589                     (longlong_t)scn->scn_visited_this_txg,
1590                     (longlong_t)
1591                     NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1592                     (longlong_t)tx->tx_txg, err);
1593                 scn->scn_visited_this_txg = 0;
1594
1595                 /*
1596                  * Write out changes to the DDT that may be required as a
1597                  * result of the blocks freed.  This ensures that the DDT
1598                  * is clean when a scrub/resilver runs.
1599                  */
1600                 ddt_sync(spa, tx->tx_txg);
1601         }
1602         if (err != 0)
1603                 return;
1604         if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
1605             zfs_free_leak_on_eio &&
1606             (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1607             dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1608             dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1609                 /*
1610                  * We have finished background destroying, but there is still
1611                  * some space left in the dp_free_dir. Transfer this leaked
1612                  * space to the dp_leak_dir.
1613                  */
1614                 if (dp->dp_leak_dir == NULL) {
1615                         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1616                         (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1617                             LEAK_DIR_NAME, tx);
1618                         VERIFY0(dsl_pool_open_special_dir(dp,
1619                             LEAK_DIR_NAME, &dp->dp_leak_dir));
1620                         rrw_exit(&dp->dp_config_rwlock, FTAG);
1621                 }
1622                 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1623                     dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1624                     dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1625                     dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1626                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1627                     -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1628                     -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1629                     -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1630         }
1631         if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
1632                 /* finished; verify that space accounting went to zero */
1633                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1634                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1635                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1636         }
1637
1638         if (scn->scn_phys.scn_state != DSS_SCANNING)
1639                 return;
1640
1641         if (scn->scn_done_txg == tx->tx_txg) {
1642                 ASSERT(!scn->scn_pausing);
1643                 /* finished with scan. */
1644                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1645                 dsl_scan_done(scn, B_TRUE, tx);
1646                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1647                 dsl_scan_sync_state(scn, tx);
1648                 return;
1649         }
1650
1651         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1652             scn->scn_phys.scn_ddt_class_max) {
1653                 zfs_dbgmsg("doing scan sync txg %llu; "
1654                     "ddt bm=%llu/%llu/%llu/%llx",
1655                     (longlong_t)tx->tx_txg,
1656                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1657                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1658                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1659                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1660                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1661                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1662                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1663                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1664         } else {
1665                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1666                     (longlong_t)tx->tx_txg,
1667                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1668                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1669                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1670                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1671         }
1672
1673         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1674             NULL, ZIO_FLAG_CANFAIL);
1675         dsl_pool_config_enter(dp, FTAG);
1676         dsl_scan_visit(scn, tx);
1677         dsl_pool_config_exit(dp, FTAG);
1678         (void) zio_wait(scn->scn_zio_root);
1679         scn->scn_zio_root = NULL;
1680
1681         zfs_dbgmsg("visited %llu blocks in %llums",
1682             (longlong_t)scn->scn_visited_this_txg,
1683             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1684
1685         if (!scn->scn_pausing) {
1686                 scn->scn_done_txg = tx->tx_txg + 1;
1687                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1688                     tx->tx_txg, scn->scn_done_txg);
1689         }
1690
1691         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1692                 mutex_enter(&spa->spa_scrub_lock);
1693                 while (spa->spa_scrub_inflight > 0) {
1694                         cv_wait(&spa->spa_scrub_io_cv,
1695                             &spa->spa_scrub_lock);
1696                 }
1697                 mutex_exit(&spa->spa_scrub_lock);
1698         }
1699
1700         dsl_scan_sync_state(scn, tx);
1701 }
1702
1703 /*
1704  * This will start a new scan, or restart an existing one.
1705  */
1706 void
1707 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1708 {
1709         if (txg == 0) {
1710                 dmu_tx_t *tx;
1711                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1712                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1713
1714                 txg = dmu_tx_get_txg(tx);
1715                 dp->dp_scan->scn_restart_txg = txg;
1716                 dmu_tx_commit(tx);
1717         } else {
1718                 dp->dp_scan->scn_restart_txg = txg;
1719         }
1720         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1721 }
1722
1723 boolean_t
1724 dsl_scan_resilvering(dsl_pool_t *dp)
1725 {
1726         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1727             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1728 }
1729
1730 /*
1731  * scrub consumers
1732  */
1733
1734 static void
1735 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1736 {
1737         int i;
1738
1739         /*
1740          * If we resume after a reboot, zab will be NULL; don't record
1741          * incomplete stats in that case.
1742          */
1743         if (zab == NULL)
1744                 return;
1745
1746         for (i = 0; i < 4; i++) {
1747                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1748                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1749                 if (t & DMU_OT_NEWTYPE)
1750                         t = DMU_OT_OTHER;
1751                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1752                 int equal;
1753
1754                 zb->zb_count++;
1755                 zb->zb_asize += BP_GET_ASIZE(bp);
1756                 zb->zb_lsize += BP_GET_LSIZE(bp);
1757                 zb->zb_psize += BP_GET_PSIZE(bp);
1758                 zb->zb_gangs += BP_COUNT_GANG(bp);
1759
1760                 switch (BP_GET_NDVAS(bp)) {
1761                 case 2:
1762                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1763                             DVA_GET_VDEV(&bp->blk_dva[1]))
1764                                 zb->zb_ditto_2_of_2_samevdev++;
1765                         break;
1766                 case 3:
1767                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1768                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1769                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1770                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1771                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1772                             DVA_GET_VDEV(&bp->blk_dva[2]));
1773                         if (equal == 1)
1774                                 zb->zb_ditto_2_of_3_samevdev++;
1775                         else if (equal == 3)
1776                                 zb->zb_ditto_3_of_3_samevdev++;
1777                         break;
1778                 }
1779         }
1780 }
1781
1782 static void
1783 dsl_scan_scrub_done(zio_t *zio)
1784 {
1785         spa_t *spa = zio->io_spa;
1786
1787         zio_data_buf_free(zio->io_data, zio->io_size);
1788
1789         mutex_enter(&spa->spa_scrub_lock);
1790         spa->spa_scrub_inflight--;
1791         cv_broadcast(&spa->spa_scrub_io_cv);
1792
1793         if (zio->io_error && (zio->io_error != ECKSUM ||
1794             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1795                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1796         }
1797         mutex_exit(&spa->spa_scrub_lock);
1798 }
1799
1800 static int
1801 dsl_scan_scrub_cb(dsl_pool_t *dp,
1802     const blkptr_t *bp, const zbookmark_phys_t *zb)
1803 {
1804         dsl_scan_t *scn = dp->dp_scan;
1805         size_t size = BP_GET_PSIZE(bp);
1806         spa_t *spa = dp->dp_spa;
1807         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1808         boolean_t needs_io;
1809         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1810         unsigned int scan_delay = 0;
1811
1812         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1813             phys_birth >= scn->scn_phys.scn_max_txg)
1814                 return (0);
1815
1816         count_block(dp->dp_blkstats, bp);
1817
1818         if (BP_IS_EMBEDDED(bp))
1819                 return (0);
1820
1821         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1822         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1823                 zio_flags |= ZIO_FLAG_SCRUB;
1824                 needs_io = B_TRUE;
1825                 scan_delay = zfs_scrub_delay;
1826         } else {
1827                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1828                 zio_flags |= ZIO_FLAG_RESILVER;
1829                 needs_io = B_FALSE;
1830                 scan_delay = zfs_resilver_delay;
1831         }
1832
1833         /* If it's an intent log block, failure is expected. */
1834         if (zb->zb_level == ZB_ZIL_LEVEL)
1835                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1836
1837         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1838                 vdev_t *vd = vdev_lookup_top(spa,
1839                     DVA_GET_VDEV(&bp->blk_dva[d]));
1840
1841                 /*
1842                  * Keep track of how much data we've examined so that
1843                  * zpool(1M) status can make useful progress reports.
1844                  */
1845                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1846                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1847
1848                 /* if it's a resilver, this may not be in the target range */
1849                 if (!needs_io) {
1850                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1851                                 /*
1852                                  * Gang members may be spread across multiple
1853                                  * vdevs, so the best estimate we have is the
1854                                  * scrub range, which has already been checked.
1855                                  * XXX -- it would be better to change our
1856                                  * allocation policy to ensure that all
1857                                  * gang members reside on the same vdev.
1858                                  */
1859                                 needs_io = B_TRUE;
1860                         } else {
1861                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1862                                     phys_birth, 1);
1863                         }
1864                 }
1865         }
1866
1867         if (needs_io && !zfs_no_scrub_io) {
1868                 vdev_t *rvd = spa->spa_root_vdev;
1869                 uint64_t maxinflight = rvd->vdev_children *
1870                     MAX(zfs_top_maxinflight, 1);
1871                 void *data = zio_data_buf_alloc(size);
1872
1873                 mutex_enter(&spa->spa_scrub_lock);
1874                 while (spa->spa_scrub_inflight >= maxinflight)
1875                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1876                 spa->spa_scrub_inflight++;
1877                 mutex_exit(&spa->spa_scrub_lock);
1878
1879                 /*
1880                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1881                  * then throttle our workload to limit the impact of a scan.
1882                  */
1883                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1884                         delay(MAX((int)scan_delay, 0));
1885
1886                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1887                     dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1888                     zio_flags, zb));
1889         }
1890
1891         /* do not relocate this block */
1892         return (0);
1893 }
1894
1895 int
1896 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1897 {
1898         spa_t *spa = dp->dp_spa;
1899
1900         /*
1901          * Purge all vdev caches and probe all devices.  We do this here
1902          * rather than in sync context because this requires a writer lock
1903          * on the spa_config lock, which we can't do from sync context.  The
1904          * spa_scrub_reopen flag indicates that vdev_open() should not
1905          * attempt to start another scrub.
1906          */
1907         spa_vdev_state_enter(spa, SCL_NONE);
1908         spa->spa_scrub_reopen = B_TRUE;
1909         vdev_reopen(spa->spa_root_vdev);
1910         spa->spa_scrub_reopen = B_FALSE;
1911         (void) spa_vdev_state_exit(spa, NULL, 0);
1912
1913         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1914             dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1915 }