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,
 388     const zbookmark_phys_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 389     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 390     dmu_tx_t *tx);
 391 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 392     dmu_objset_type_t ostype,
 393     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 394
 395 void
 396 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 397 {
 398         zio_free(dp->dp_spa, txg, bp);
 399 }
 400
 401 void
 402 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 403 {
 404         ASSERT(dsl_pool_sync_context(dp));
 405         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 406             pio->io_flags));
 407 }
 408
 409 static uint64_t
 410 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 411 {
 412         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 413         if (dsl_dataset_is_snapshot(ds))
 414                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 415         return (smt);
 416 }
 417
 418 static void
 419 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 420 {
 421         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 422             DMU_POOL_DIRECTORY_OBJECT,
 423             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 424             &scn->scn_phys, tx));
 425 }
 426
 427 static boolean_t
 428 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
 429 {
 430         uint64_t elapsed_nanosecs;
 431         unsigned int mintime;
 432
 433         /* we never skip user/group accounting objects */
 434         if (zb && (int64_t)zb->zb_object < 0)
 435                 return (B_FALSE);
 436
 437         if (scn->scn_pausing)
 438                 return (B_TRUE); /* we're already pausing */
 439
 440         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 441                 return (B_FALSE); /* we're resuming */
 442
 443         /* We only know how to resume from level-0 blocks. */
 444         if (zb && zb->zb_level != 0)
 445                 return (B_FALSE);
 446
 447         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 448             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 449         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 450         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 451             (NSEC2MSEC(elapsed_nanosecs) > mintime &&
 452             txg_sync_waiting(scn->scn_dp)) ||
 453             spa_shutting_down(scn->scn_dp->dp_spa)) {
 454                 if (zb) {
 455                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 456                             (longlong_t)zb->zb_objset,
 457                             (longlong_t)zb->zb_object,
 458                             (longlong_t)zb->zb_level,
 459                             (longlong_t)zb->zb_blkid);
 460                         scn->scn_phys.scn_bookmark = *zb;
 461                 }
 462                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 463                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 464                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 465                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 466                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 467                 scn->scn_pausing = B_TRUE;
 468                 return (B_TRUE);
 469         }
 470         return (B_FALSE);
 471 }
 472
 473 typedef struct zil_scan_arg {
 474         dsl_pool_t      *zsa_dp;
 475         zil_header_t    *zsa_zh;
 476 } zil_scan_arg_t;
 477
 478 /* ARGSUSED */
 479 static int
 480 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 481 {
 482         zil_scan_arg_t *zsa = arg;
 483         dsl_pool_t *dp = zsa->zsa_dp;
 484         dsl_scan_t *scn = dp->dp_scan;
 485         zil_header_t *zh = zsa->zsa_zh;
 486         zbookmark_phys_t zb;
 487
 488         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 489                 return (0);
 490
 491         /*
 492          * One block ("stubby") can be allocated a long time ago; we
 493          * want to visit that one because it has been allocated
 494          * (on-disk) even if it hasn't been claimed (even though for
 495          * scrub there's nothing to do to it).
 496          */
 497         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 498                 return (0);
 499
 500         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 501             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 502
 503         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 504         return (0);
 505 }
 506
 507 /* ARGSUSED */
 508 static int
 509 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 510 {
 511         if (lrc->lrc_txtype == TX_WRITE) {
 512                 zil_scan_arg_t *zsa = arg;
 513                 dsl_pool_t *dp = zsa->zsa_dp;
 514                 dsl_scan_t *scn = dp->dp_scan;
 515                 zil_header_t *zh = zsa->zsa_zh;
 516                 lr_write_t *lr = (lr_write_t *)lrc;
 517                 blkptr_t *bp = &lr->lr_blkptr;
 518                 zbookmark_phys_t zb;
 519
 520                 if (BP_IS_HOLE(bp) ||
 521                     bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 522                         return (0);
 523
 524                 /*
 525                  * birth can be < claim_txg if this record's txg is
 526                  * already txg sync'ed (but this log block contains
 527                  * other records that are not synced)
 528                  */
 529                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 530                         return (0);
 531
 532                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 533                     lr->lr_foid, ZB_ZIL_LEVEL,
 534                     lr->lr_offset / BP_GET_LSIZE(bp));
 535
 536                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 537         }
 538         return (0);
 539 }
 540
 541 static void
 542 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 543 {
 544         uint64_t claim_txg = zh->zh_claim_txg;
 545         zil_scan_arg_t zsa = { dp, zh };
 546         zilog_t *zilog;
 547
 548         /*
 549          * We only want to visit blocks that have been claimed but not yet
 550          * replayed (or, in read-only mode, blocks that *would* be claimed).
 551          */
 552         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 553                 return;
 554
 555         zilog = zil_alloc(dp->dp_meta_objset, zh);
 556
 557         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 558             claim_txg);
 559
 560         zil_free(zilog);
 561 }
 562
 563 /* ARGSUSED */
 564 static void
 565 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 566     uint64_t objset, uint64_t object, uint64_t blkid)
 567 {
 568         zbookmark_phys_t czb;
 569         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 570
 571         if (zfs_no_scrub_prefetch)
 572                 return;
 573
 574         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 575             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 576                 return;
 577
 578         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 579
 580         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 581             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 582             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 583 }
 584
 585 static boolean_t
 586 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 587     const zbookmark_phys_t *zb)
 588 {
 589         /*
 590          * We never skip over user/group accounting objects (obj<0)
 591          */
 592         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 593             (int64_t)zb->zb_object >= 0) {
 594                 /*
 595                  * If we already visited this bp & everything below (in
 596                  * a prior txg sync), don't bother doing it again.
 597                  */
 598                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 599                         return (B_TRUE);
 600
 601                 /*
 602                  * If we found the block we're trying to resume from, or
 603                  * we went past it to a different object, zero it out to
 604                  * indicate that it's OK to start checking for pausing
 605                  * again.
 606                  */
 607                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 608                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 609                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 610                             (longlong_t)zb->zb_objset,
 611                             (longlong_t)zb->zb_object,
 612                             (longlong_t)zb->zb_level,
 613                             (longlong_t)zb->zb_blkid);
 614                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 615                 }
 616         }
 617         return (B_FALSE);
 618 }
 619
 620 /*
 621  * Return nonzero on i/o error.
 622  * Return new buf to write out in *bufp.
 623  */
 624 static int
 625 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 626     dnode_phys_t *dnp, const blkptr_t *bp,
 627     const zbookmark_phys_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 628 {
 629         dsl_pool_t *dp = scn->scn_dp;
 630         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 631         int err;
 632
 633         if (BP_GET_LEVEL(bp) > 0) {
 634                 uint32_t flags = ARC_WAIT;
 635                 int i;
 636                 blkptr_t *cbp;
 637                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 638
 639                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 640                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 641                 if (err) {
 642                         scn->scn_phys.scn_errors++;
 643                         return (err);
 644                 }
 645                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 646                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 647                             zb->zb_object, zb->zb_blkid * epb + i);
 648                 }
 649                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 650                         zbookmark_phys_t czb;
 651
 652                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 653                             zb->zb_level - 1,
 654                             zb->zb_blkid * epb + i);
 655                         dsl_scan_visitbp(cbp, &czb, dnp,
 656                             *bufp, ds, scn, ostype, tx);
 657                 }
 658         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 659                 uint32_t flags = ARC_WAIT;
 660
 661                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 662                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 663                 if (err) {
 664                         scn->scn_phys.scn_errors++;
 665                         return (err);
 666                 }
 667         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 668                 uint32_t flags = ARC_WAIT;
 669                 dnode_phys_t *cdnp;
 670                 int i, j;
 671                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 672
 673                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 674                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 675                 if (err) {
 676                         scn->scn_phys.scn_errors++;
 677                         return (err);
 678                 }
 679                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 680                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 681                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 682                                 dsl_scan_prefetch(scn, *bufp, cbp,
 683                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 684                         }
 685                 }
 686                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 687                         dsl_scan_visitdnode(scn, ds, ostype,
 688                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 689                 }
 690
 691         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 692                 uint32_t flags = ARC_WAIT;
 693                 objset_phys_t *osp;
 694
 695                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 696                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 697                 if (err) {
 698                         scn->scn_phys.scn_errors++;
 699                         return (err);
 700                 }
 701
 702                 osp = (*bufp)->b_data;
 703
 704                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 705                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 706
 707                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 708                         /*
 709                          * We also always visit user/group accounting
 710                          * objects, and never skip them, even if we are
 711                          * pausing.  This is necessary so that the space
 712                          * deltas from this txg get integrated.
 713                          */
 714                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 715                             &osp->os_groupused_dnode, *bufp,
 716                             DMU_GROUPUSED_OBJECT, tx);
 717                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 718                             &osp->os_userused_dnode, *bufp,
 719                             DMU_USERUSED_OBJECT, tx);
 720                 }
 721         }
 722
 723         return (0);
 724 }
 725
 726 static void
 727 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 728     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 729     uint64_t object, dmu_tx_t *tx)
 730 {
 731         int j;
 732
 733         for (j = 0; j < dnp->dn_nblkptr; j++) {
 734                 zbookmark_phys_t czb;
 735
 736                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 737                     dnp->dn_nlevels - 1, j);
 738                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 739                     &czb, dnp, buf, ds, scn, ostype, tx);
 740         }
 741
 742         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 743                 zbookmark_phys_t czb;
 744                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 745                     0, DMU_SPILL_BLKID);
 746                 dsl_scan_visitbp(&dnp->dn_spill,
 747                     &czb, dnp, buf, ds, scn, ostype, tx);
 748         }
 749 }
 750
 751 /*
 752  * The arguments are in this order because mdb can only print the
 753  * first 5; we want them to be useful.
 754  */
 755 static void
 756 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 757     dnode_phys_t *dnp, arc_buf_t *pbuf,
 758     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 759     dmu_tx_t *tx)
 760 {
 761         dsl_pool_t *dp = scn->scn_dp;
 762         arc_buf_t *buf = NULL;
 763         blkptr_t bp_toread = *bp;
 764
 765         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 766
 767         if (dsl_scan_check_pause(scn, zb))
 768                 return;
 769
 770         if (dsl_scan_check_resume(scn, dnp, zb))
 771                 return;
 772
 773         if (BP_IS_HOLE(bp))
 774                 return;
 775
 776         scn->scn_visited_this_txg++;
 777
 778         dprintf_bp(bp,
 779             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 780             ds, ds ? ds->ds_object : 0,
 781             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 782             pbuf, bp);
 783
 784         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 785                 return;
 786
 787         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
 788             &buf) != 0)
 789                 return;
 790
 791         /*
 792          * If dsl_scan_ddt() has aready visited this block, it will have
 793          * already done any translations or scrubbing, so don't call the
 794          * callback again.
 795          */
 796         if (ddt_class_contains(dp->dp_spa,
 797             scn->scn_phys.scn_ddt_class_max, bp)) {
 798                 ASSERT(buf == NULL);
 799                 return;
 800         }
 801
 802         /*
 803          * If this block is from the future (after cur_max_txg), then we
 804          * are doing this on behalf of a deleted snapshot, and we will
 805          * revisit the future block on the next pass of this dataset.
 806          * Don't scan it now unless we need to because something
 807          * under it was modified.
 808          */
 809         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 810                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 811         }
 812         if (buf)
 813                 (void) arc_buf_remove_ref(buf, &buf);
 814 }
 815
 816 static void
 817 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 818     dmu_tx_t *tx)
 819 {
 820         zbookmark_phys_t zb;
 821
 822         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 823             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 824         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 825             ds, scn, DMU_OST_NONE, tx);
 826
 827         dprintf_ds(ds, "finished scan%s", "");
 828 }
 829
 830 void
 831 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 832 {
 833         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 834         dsl_scan_t *scn = dp->dp_scan;
 835         uint64_t mintxg;
 836
 837         if (scn->scn_phys.scn_state != DSS_SCANNING)
 838                 return;
 839
 840         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 841                 if (dsl_dataset_is_snapshot(ds)) {
 842                         /* Note, scn_cur_{min,max}_txg stays the same. */
 843                         scn->scn_phys.scn_bookmark.zb_objset =
 844                             ds->ds_phys->ds_next_snap_obj;
 845                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 846                             "reset zb_objset to %llu",
 847                             (u_longlong_t)ds->ds_object,
 848                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 849                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 850                 } else {
 851                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 852                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 853                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 854                             "reset bookmark to -1,0,0,0",
 855                             (u_longlong_t)ds->ds_object);
 856                 }
 857         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 858             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 859                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 860                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 861                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 862                 if (dsl_dataset_is_snapshot(ds)) {
 863                         /*
 864                          * We keep the same mintxg; it could be >
 865                          * ds_creation_txg if the previous snapshot was
 866                          * deleted too.
 867                          */
 868                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 869                             scn->scn_phys.scn_queue_obj,
 870                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 871                         zfs_dbgmsg("destroying ds %llu; in queue; "
 872                             "replacing with %llu",
 873                             (u_longlong_t)ds->ds_object,
 874                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 875                 } else {
 876                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 877                             (u_longlong_t)ds->ds_object);
 878                 }
 879         } else {
 880                 zfs_dbgmsg("destroying ds %llu; ignoring",
 881                     (u_longlong_t)ds->ds_object);
 882         }
 883
 884         /*
 885          * dsl_scan_sync() should be called after this, and should sync
 886          * out our changed state, but just to be safe, do it here.
 887          */
 888         dsl_scan_sync_state(scn, tx);
 889 }
 890
 891 void
 892 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 893 {
 894         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 895         dsl_scan_t *scn = dp->dp_scan;
 896         uint64_t mintxg;
 897
 898         if (scn->scn_phys.scn_state != DSS_SCANNING)
 899                 return;
 900
 901         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 902
 903         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 904                 scn->scn_phys.scn_bookmark.zb_objset =
 905                     ds->ds_phys->ds_prev_snap_obj;
 906                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 907                     "reset zb_objset to %llu",
 908                     (u_longlong_t)ds->ds_object,
 909                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 910         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 911             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 912                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 913                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 914                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 915                     scn->scn_phys.scn_queue_obj,
 916                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 917                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 918                     "replacing with %llu",
 919                     (u_longlong_t)ds->ds_object,
 920                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 921         }
 922         dsl_scan_sync_state(scn, tx);
 923 }
 924
 925 void
 926 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 927 {
 928         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 929         dsl_scan_t *scn = dp->dp_scan;
 930         uint64_t mintxg;
 931
 932         if (scn->scn_phys.scn_state != DSS_SCANNING)
 933                 return;
 934
 935         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 936                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 937                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 938                     "reset zb_objset to %llu",
 939                     (u_longlong_t)ds1->ds_object,
 940                     (u_longlong_t)ds2->ds_object);
 941         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 942                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 943                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 944                     "reset zb_objset to %llu",
 945                     (u_longlong_t)ds2->ds_object,
 946                     (u_longlong_t)ds1->ds_object);
 947         }
 948
 949         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 950             ds1->ds_object, &mintxg) == 0) {
 951                 int err;
 952
 953                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 954                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 955                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 956                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 957                 err = zap_add_int_key(dp->dp_meta_objset,
 958                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 959                 VERIFY(err == 0 || err == EEXIST);
 960                 if (err == EEXIST) {
 961                         /* Both were there to begin with */
 962                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 963                             scn->scn_phys.scn_queue_obj,
 964                             ds1->ds_object, mintxg, tx));
 965                 }
 966                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 967                     "replacing with %llu",
 968                     (u_longlong_t)ds1->ds_object,
 969                     (u_longlong_t)ds2->ds_object);
 970         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 971             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 972                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 973                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 974                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 975                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 976                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 977                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 978                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 979                     "replacing with %llu",
 980                     (u_longlong_t)ds2->ds_object,
 981                     (u_longlong_t)ds1->ds_object);
 982         }
 983
 984         dsl_scan_sync_state(scn, tx);
 985 }
 986
 987 struct enqueue_clones_arg {
 988         dmu_tx_t *tx;
 989         uint64_t originobj;
 990 };
 991
 992 /* ARGSUSED */
 993 static int
 994 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 995 {
 996         struct enqueue_clones_arg *eca = arg;
 997         dsl_dataset_t *ds;
 998         int err;
 999         dsl_scan_t *scn = dp->dp_scan;
1000
1001         if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
1002                 return (0);
1003
1004         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1005         if (err)
1006                 return (err);
1007
1008         while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1009                 dsl_dataset_t *prev;
1010                 err = dsl_dataset_hold_obj(dp,
1011                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1012
1013                 dsl_dataset_rele(ds, FTAG);
1014                 if (err)
1015                         return (err);
1016                 ds = prev;
1017         }
1018         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1019             scn->scn_phys.scn_queue_obj, ds->ds_object,
1020             ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1021         dsl_dataset_rele(ds, FTAG);
1022         return (0);
1023 }
1024
1025 static void
1026 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1027 {
1028         dsl_pool_t *dp = scn->scn_dp;
1029         dsl_dataset_t *ds;
1030         objset_t *os;
1031
1032         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1033
1034         if (dmu_objset_from_ds(ds, &os))
1035                 goto out;
1036
1037         /*
1038          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1039          * snapshots can have ZIL block pointers (which may be the same
1040          * BP as in the head), they must be ignored.  So we traverse the
1041          * ZIL here, rather than in scan_recurse(), because the regular
1042          * snapshot block-sharing rules don't apply to it.
1043          */
1044         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1045                 dsl_scan_zil(dp, &os->os_zil_header);
1046
1047         /*
1048          * Iterate over the bps in this ds.
1049          */
1050         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1051         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1052
1053         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1054         dsl_dataset_name(ds, dsname);
1055         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1056             "pausing=%u",
1057             (longlong_t)dsobj, dsname,
1058             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1059             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1060             (int)scn->scn_pausing);
1061         kmem_free(dsname, ZFS_MAXNAMELEN);
1062
1063         if (scn->scn_pausing)
1064                 goto out;
1065
1066         /*
1067          * We've finished this pass over this dataset.
1068          */
1069
1070         /*
1071          * If we did not completely visit this dataset, do another pass.
1072          */
1073         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1074                 zfs_dbgmsg("incomplete pass; visiting again");
1075                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1076                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1077                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1078                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1079                 goto out;
1080         }
1081
1082         /*
1083          * Add descendent datasets to work queue.
1084          */
1085         if (ds->ds_phys->ds_next_snap_obj != 0) {
1086                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1087                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1088                     ds->ds_phys->ds_creation_txg, tx) == 0);
1089         }
1090         if (ds->ds_phys->ds_num_children > 1) {
1091                 boolean_t usenext = B_FALSE;
1092                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1093                         uint64_t count;
1094                         /*
1095                          * A bug in a previous version of the code could
1096                          * cause upgrade_clones_cb() to not set
1097                          * ds_next_snap_obj when it should, leading to a
1098                          * missing entry.  Therefore we can only use the
1099                          * next_clones_obj when its count is correct.
1100                          */
1101                         int err = zap_count(dp->dp_meta_objset,
1102                             ds->ds_phys->ds_next_clones_obj, &count);
1103                         if (err == 0 &&
1104                             count == ds->ds_phys->ds_num_children - 1)
1105                                 usenext = B_TRUE;
1106                 }
1107
1108                 if (usenext) {
1109                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1110                             ds->ds_phys->ds_next_clones_obj,
1111                             scn->scn_phys.scn_queue_obj,
1112                             ds->ds_phys->ds_creation_txg, tx));
1113                 } else {
1114                         struct enqueue_clones_arg eca;
1115                         eca.tx = tx;
1116                         eca.originobj = ds->ds_object;
1117
1118                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1119                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1120                 }
1121         }
1122
1123 out:
1124         dsl_dataset_rele(ds, FTAG);
1125 }
1126
1127 /* ARGSUSED */
1128 static int
1129 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1130 {
1131         dmu_tx_t *tx = arg;
1132         dsl_dataset_t *ds;
1133         int err;
1134         dsl_scan_t *scn = dp->dp_scan;
1135
1136         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1137         if (err)
1138                 return (err);
1139
1140         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1141                 dsl_dataset_t *prev;
1142                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1143                     FTAG, &prev);
1144                 if (err) {
1145                         dsl_dataset_rele(ds, FTAG);
1146                         return (err);
1147                 }
1148
1149                 /*
1150                  * If this is a clone, we don't need to worry about it for now.
1151                  */
1152                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1153                         dsl_dataset_rele(ds, FTAG);
1154                         dsl_dataset_rele(prev, FTAG);
1155                         return (0);
1156                 }
1157                 dsl_dataset_rele(ds, FTAG);
1158                 ds = prev;
1159         }
1160
1161         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1162             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1163         dsl_dataset_rele(ds, FTAG);
1164         return (0);
1165 }
1166
1167 /*
1168  * Scrub/dedup interaction.
1169  *
1170  * If there are N references to a deduped block, we don't want to scrub it
1171  * N times -- ideally, we should scrub it exactly once.
1172  *
1173  * We leverage the fact that the dde's replication class (enum ddt_class)
1174  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1175  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1176  *
1177  * To prevent excess scrubbing, the scrub begins by walking the DDT
1178  * to find all blocks with refcnt > 1, and scrubs each of these once.
1179  * Since there are two replication classes which contain blocks with
1180  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1181  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1182  *
1183  * There would be nothing more to say if a block's refcnt couldn't change
1184  * during a scrub, but of course it can so we must account for changes
1185  * in a block's replication class.
1186  *
1187  * Here's an example of what can occur:
1188  *
1189  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1190  * when visited during the top-down scrub phase, it will be scrubbed twice.
1191  * This negates our scrub optimization, but is otherwise harmless.
1192  *
1193  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1194  * on each visit during the top-down scrub phase, it will never be scrubbed.
1195  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1196  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1197  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1198  * while a scrub is in progress, it scrubs the block right then.
1199  */
1200 static void
1201 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1202 {
1203         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1204         ddt_entry_t dde = { 0 };
1205         int error;
1206         uint64_t n = 0;
1207
1208         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1209                 ddt_t *ddt;
1210
1211                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1212                         break;
1213                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1214                     (longlong_t)ddb->ddb_class,
1215                     (longlong_t)ddb->ddb_type,
1216                     (longlong_t)ddb->ddb_checksum,
1217                     (longlong_t)ddb->ddb_cursor);
1218
1219                 /* There should be no pending changes to the dedup table */
1220                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1221                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1222
1223                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1224                 n++;
1225
1226                 if (dsl_scan_check_pause(scn, NULL))
1227                         break;
1228         }
1229
1230         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1231             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1232             (int)scn->scn_pausing);
1233
1234         ASSERT(error == 0 || error == ENOENT);
1235         ASSERT(error != ENOENT ||
1236             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1237 }
1238
1239 /* ARGSUSED */
1240 void
1241 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1242     ddt_entry_t *dde, dmu_tx_t *tx)
1243 {
1244         const ddt_key_t *ddk = &dde->dde_key;
1245         ddt_phys_t *ddp = dde->dde_phys;
1246         blkptr_t bp;
1247         zbookmark_phys_t zb = { 0 };
1248
1249         if (scn->scn_phys.scn_state != DSS_SCANNING)
1250                 return;
1251
1252         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1253                 if (ddp->ddp_phys_birth == 0 ||
1254                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1255                         continue;
1256                 ddt_bp_create(checksum, ddk, ddp, &bp);
1257
1258                 scn->scn_visited_this_txg++;
1259                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1260         }
1261 }
1262
1263 static void
1264 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1265 {
1266         dsl_pool_t *dp = scn->scn_dp;
1267         zap_cursor_t zc;
1268         zap_attribute_t za;
1269
1270         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1271             scn->scn_phys.scn_ddt_class_max) {
1272                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1273                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1274                 dsl_scan_ddt(scn, tx);
1275                 if (scn->scn_pausing)
1276                         return;
1277         }
1278
1279         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1280                 /* First do the MOS & ORIGIN */
1281
1282                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1283                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1284                 dsl_scan_visit_rootbp(scn, NULL,
1285                     &dp->dp_meta_rootbp, tx);
1286                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1287                 if (scn->scn_pausing)
1288                         return;
1289
1290                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1291                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1292                             enqueue_cb, tx, DS_FIND_CHILDREN));
1293                 } else {
1294                         dsl_scan_visitds(scn,
1295                             dp->dp_origin_snap->ds_object, tx);
1296                 }
1297                 ASSERT(!scn->scn_pausing);
1298         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1299             ZB_DESTROYED_OBJSET) {
1300                 /*
1301                  * If we were paused, continue from here.  Note if the
1302                  * ds we were paused on was deleted, the zb_objset may
1303                  * be -1, so we will skip this and find a new objset
1304                  * below.
1305                  */
1306                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1307                 if (scn->scn_pausing)
1308                         return;
1309         }
1310
1311         /*
1312          * In case we were paused right at the end of the ds, zero the
1313          * bookmark so we don't think that we're still trying to resume.
1314          */
1315         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1316
1317         /* keep pulling things out of the zap-object-as-queue */
1318         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1319             scn->scn_phys.scn_queue_obj),
1320             zap_cursor_retrieve(&zc, &za) == 0) {
1321                 dsl_dataset_t *ds;
1322                 uint64_t dsobj;
1323
1324                 dsobj = strtonum(za.za_name, NULL);
1325                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1326                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1327
1328                 /* Set up min/max txg */
1329                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1330                 if (za.za_first_integer != 0) {
1331                         scn->scn_phys.scn_cur_min_txg =
1332                             MAX(scn->scn_phys.scn_min_txg,
1333                             za.za_first_integer);
1334                 } else {
1335                         scn->scn_phys.scn_cur_min_txg =
1336                             MAX(scn->scn_phys.scn_min_txg,
1337                             ds->ds_phys->ds_prev_snap_txg);
1338                 }
1339                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1340                 dsl_dataset_rele(ds, FTAG);
1341
1342                 dsl_scan_visitds(scn, dsobj, tx);
1343                 zap_cursor_fini(&zc);
1344                 if (scn->scn_pausing)
1345                         return;
1346         }
1347         zap_cursor_fini(&zc);
1348 }
1349
1350 static boolean_t
1351 dsl_scan_free_should_pause(dsl_scan_t *scn)
1352 {
1353         uint64_t elapsed_nanosecs;
1354
1355         if (zfs_recover)
1356                 return (B_FALSE);
1357
1358         if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1359                 return (B_TRUE);
1360
1361         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1362         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1363             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1364             txg_sync_waiting(scn->scn_dp)) ||
1365             spa_shutting_down(scn->scn_dp->dp_spa));
1366 }
1367
1368 static int
1369 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1370 {
1371         dsl_scan_t *scn = arg;
1372
1373         if (!scn->scn_is_bptree ||
1374             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1375                 if (dsl_scan_free_should_pause(scn))
1376                         return (SET_ERROR(ERESTART));
1377         }
1378
1379         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1380             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1381         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1382             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1383             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1384         scn->scn_visited_this_txg++;
1385         return (0);
1386 }
1387
1388 boolean_t
1389 dsl_scan_active(dsl_scan_t *scn)
1390 {
1391         spa_t *spa = scn->scn_dp->dp_spa;
1392         uint64_t used = 0, comp, uncomp;
1393
1394         if (spa->spa_load_state != SPA_LOAD_NONE)
1395                 return (B_FALSE);
1396         if (spa_shutting_down(spa))
1397                 return (B_FALSE);
1398         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1399             (scn->scn_async_destroying && !scn->scn_async_stalled))
1400                 return (B_TRUE);
1401
1402         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1403                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1404                     &used, &comp, &uncomp);
1405         }
1406         return (used != 0);
1407 }
1408
1409 void
1410 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1411 {
1412         dsl_scan_t *scn = dp->dp_scan;
1413         spa_t *spa = dp->dp_spa;
1414         int err = 0;
1415
1416         /*
1417          * Check for scn_restart_txg before checking spa_load_state, so
1418          * that we can restart an old-style scan while the pool is being
1419          * imported (see dsl_scan_init).
1420          */
1421         if (scn->scn_restart_txg != 0 &&
1422             scn->scn_restart_txg <= tx->tx_txg) {
1423                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1424                 dsl_scan_done(scn, B_FALSE, tx);
1425                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1426                         func = POOL_SCAN_RESILVER;
1427                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1428                     func, tx->tx_txg);
1429                 dsl_scan_setup_sync(&func, tx);
1430         }
1431
1432         /*
1433          * If the scan is inactive due to a stalled async destroy, try again.
1434          */
1435         if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1436             spa_sync_pass(dp->dp_spa) > 1)
1437                 return;
1438
1439         scn->scn_visited_this_txg = 0;
1440         scn->scn_pausing = B_FALSE;
1441         scn->scn_sync_start_time = gethrtime();
1442         spa->spa_scrub_active = B_TRUE;
1443
1444         /*
1445          * First process the async destroys.  If we pause, don't do
1446          * any scrubbing or resilvering.  This ensures that there are no
1447          * async destroys while we are scanning, so the scan code doesn't
1448          * have to worry about traversing it.  It is also faster to free the
1449          * blocks than to scrub them.
1450          */
1451         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1452                 scn->scn_is_bptree = B_FALSE;
1453                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1454                     NULL, ZIO_FLAG_MUSTSUCCEED);
1455                 err = bpobj_iterate(&dp->dp_free_bpobj,
1456                     dsl_scan_free_block_cb, scn, tx);
1457                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1458
1459                 if (err != 0 && err != ERESTART)
1460                         zfs_panic_recover("error %u from bpobj_iterate()", err);
1461         }
1462
1463         if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1464                 ASSERT(scn->scn_async_destroying);
1465                 scn->scn_is_bptree = B_TRUE;
1466                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1467                     NULL, ZIO_FLAG_MUSTSUCCEED);
1468                 err = bptree_iterate(dp->dp_meta_objset,
1469                     dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1470                 VERIFY0(zio_wait(scn->scn_zio_root));
1471
1472                 if (err == EIO || err == ECKSUM) {
1473                         err = 0;
1474                 } else if (err != 0 && err != ERESTART) {
1475                         zfs_panic_recover("error %u from "
1476                             "traverse_dataset_destroyed()", err);
1477                 }
1478
1479                 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1480                         /* finished; deactivate async destroy feature */
1481                         spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1482                         ASSERT(!spa_feature_is_active(spa,
1483                             SPA_FEATURE_ASYNC_DESTROY));
1484                         VERIFY0(zap_remove(dp->dp_meta_objset,
1485                             DMU_POOL_DIRECTORY_OBJECT,
1486                             DMU_POOL_BPTREE_OBJ, tx));
1487                         VERIFY0(bptree_free(dp->dp_meta_objset,
1488                             dp->dp_bptree_obj, tx));
1489                         dp->dp_bptree_obj = 0;
1490                         scn->scn_async_destroying = B_FALSE;
1491                 } else {
1492                         /*
1493                          * If we didn't make progress, mark the async destroy as
1494                          * stalled, so that we will not initiate a spa_sync() on
1495                          * its behalf.
1496                          */
1497                         scn->scn_async_stalled =
1498                             (scn->scn_visited_this_txg == 0);
1499                 }
1500         }
1501         if (scn->scn_visited_this_txg) {
1502                 zfs_dbgmsg("freed %llu blocks in %llums from "
1503                     "free_bpobj/bptree txg %llu; err=%u",
1504                     (longlong_t)scn->scn_visited_this_txg,
1505                     (longlong_t)
1506                     NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1507                     (longlong_t)tx->tx_txg, err);
1508                 scn->scn_visited_this_txg = 0;
1509
1510                 /*
1511                  * Write out changes to the DDT that may be required as a
1512                  * result of the blocks freed.  This ensures that the DDT
1513                  * is clean when a scrub/resilver runs.
1514                  */
1515                 ddt_sync(spa, tx->tx_txg);
1516         }
1517         if (err != 0)
1518                 return;
1519         if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1520             (dp->dp_free_dir->dd_phys->dd_used_bytes != 0 ||
1521             dp->dp_free_dir->dd_phys->dd_compressed_bytes != 0 ||
1522             dp->dp_free_dir->dd_phys->dd_uncompressed_bytes != 0)) {
1523                 /*
1524                  * We have finished background destroying, but there is still
1525                  * some space left in the dp_free_dir. Transfer this leaked
1526                  * space to the dp_leak_dir.
1527                  */
1528                 if (dp->dp_leak_dir == NULL) {
1529                         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1530                         (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1531                             LEAK_DIR_NAME, tx);
1532                         VERIFY0(dsl_pool_open_special_dir(dp,
1533                             LEAK_DIR_NAME, &dp->dp_leak_dir));
1534                         rrw_exit(&dp->dp_config_rwlock, FTAG);
1535                 }
1536                 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1537                     dp->dp_free_dir->dd_phys->dd_used_bytes,
1538                     dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1539                     dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1540                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1541                     -dp->dp_free_dir->dd_phys->dd_used_bytes,
1542                     -dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1543                     -dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1544         }
1545         if (!scn->scn_async_destroying) {
1546                 /* finished; verify that space accounting went to zero */
1547                 ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes);
1548                 ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes);
1549                 ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes);
1550         }
1551
1552         if (scn->scn_phys.scn_state != DSS_SCANNING)
1553                 return;
1554
1555         if (scn->scn_done_txg == tx->tx_txg) {
1556                 ASSERT(!scn->scn_pausing);
1557                 /* finished with scan. */
1558                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1559                 dsl_scan_done(scn, B_TRUE, tx);
1560                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1561                 dsl_scan_sync_state(scn, tx);
1562                 return;
1563         }
1564
1565         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1566             scn->scn_phys.scn_ddt_class_max) {
1567                 zfs_dbgmsg("doing scan sync txg %llu; "
1568                     "ddt bm=%llu/%llu/%llu/%llx",
1569                     (longlong_t)tx->tx_txg,
1570                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1571                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1572                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1573                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1574                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1575                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1576                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1577                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1578         } else {
1579                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1580                     (longlong_t)tx->tx_txg,
1581                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1582                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1583                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1584                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1585         }
1586
1587         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1588             NULL, ZIO_FLAG_CANFAIL);
1589         dsl_pool_config_enter(dp, FTAG);
1590         dsl_scan_visit(scn, tx);
1591         dsl_pool_config_exit(dp, FTAG);
1592         (void) zio_wait(scn->scn_zio_root);
1593         scn->scn_zio_root = NULL;
1594
1595         zfs_dbgmsg("visited %llu blocks in %llums",
1596             (longlong_t)scn->scn_visited_this_txg,
1597             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1598
1599         if (!scn->scn_pausing) {
1600                 scn->scn_done_txg = tx->tx_txg + 1;
1601                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1602                     tx->tx_txg, scn->scn_done_txg);
1603         }
1604
1605         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1606                 mutex_enter(&spa->spa_scrub_lock);
1607                 while (spa->spa_scrub_inflight > 0) {
1608                         cv_wait(&spa->spa_scrub_io_cv,
1609                             &spa->spa_scrub_lock);
1610                 }
1611                 mutex_exit(&spa->spa_scrub_lock);
1612         }
1613
1614         dsl_scan_sync_state(scn, tx);
1615 }
1616
1617 /*
1618  * This will start a new scan, or restart an existing one.
1619  */
1620 void
1621 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1622 {
1623         if (txg == 0) {
1624                 dmu_tx_t *tx;
1625                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1626                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1627
1628                 txg = dmu_tx_get_txg(tx);
1629                 dp->dp_scan->scn_restart_txg = txg;
1630                 dmu_tx_commit(tx);
1631         } else {
1632                 dp->dp_scan->scn_restart_txg = txg;
1633         }
1634         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1635 }
1636
1637 boolean_t
1638 dsl_scan_resilvering(dsl_pool_t *dp)
1639 {
1640         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1641             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1642 }
1643
1644 /*
1645  * scrub consumers
1646  */
1647
1648 static void
1649 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1650 {
1651         int i;
1652
1653         /*
1654          * If we resume after a reboot, zab will be NULL; don't record
1655          * incomplete stats in that case.
1656          */
1657         if (zab == NULL)
1658                 return;
1659
1660         for (i = 0; i < 4; i++) {
1661                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1662                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1663                 if (t & DMU_OT_NEWTYPE)
1664                         t = DMU_OT_OTHER;
1665                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1666                 int equal;
1667
1668                 zb->zb_count++;
1669                 zb->zb_asize += BP_GET_ASIZE(bp);
1670                 zb->zb_lsize += BP_GET_LSIZE(bp);
1671                 zb->zb_psize += BP_GET_PSIZE(bp);
1672                 zb->zb_gangs += BP_COUNT_GANG(bp);
1673
1674                 switch (BP_GET_NDVAS(bp)) {
1675                 case 2:
1676                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1677                             DVA_GET_VDEV(&bp->blk_dva[1]))
1678                                 zb->zb_ditto_2_of_2_samevdev++;
1679                         break;
1680                 case 3:
1681                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1682                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1683                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1684                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1685                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1686                             DVA_GET_VDEV(&bp->blk_dva[2]));
1687                         if (equal == 1)
1688                                 zb->zb_ditto_2_of_3_samevdev++;
1689                         else if (equal == 3)
1690                                 zb->zb_ditto_3_of_3_samevdev++;
1691                         break;
1692                 }
1693         }
1694 }
1695
1696 static void
1697 dsl_scan_scrub_done(zio_t *zio)
1698 {
1699         spa_t *spa = zio->io_spa;
1700
1701         zio_data_buf_free(zio->io_data, zio->io_size);
1702
1703         mutex_enter(&spa->spa_scrub_lock);
1704         spa->spa_scrub_inflight--;
1705         cv_broadcast(&spa->spa_scrub_io_cv);
1706
1707         if (zio->io_error && (zio->io_error != ECKSUM ||
1708             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1709                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1710         }
1711         mutex_exit(&spa->spa_scrub_lock);
1712 }
1713
1714 static int
1715 dsl_scan_scrub_cb(dsl_pool_t *dp,
1716     const blkptr_t *bp, const zbookmark_phys_t *zb)
1717 {
1718         dsl_scan_t *scn = dp->dp_scan;
1719         size_t size = BP_GET_PSIZE(bp);
1720         spa_t *spa = dp->dp_spa;
1721         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1722         boolean_t needs_io;
1723         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1724         unsigned int scan_delay = 0;
1725
1726         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1727             phys_birth >= scn->scn_phys.scn_max_txg)
1728                 return (0);
1729
1730         count_block(dp->dp_blkstats, bp);
1731
1732         if (BP_IS_EMBEDDED(bp))
1733                 return (0);
1734
1735         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1736         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1737                 zio_flags |= ZIO_FLAG_SCRUB;
1738                 needs_io = B_TRUE;
1739                 scan_delay = zfs_scrub_delay;
1740         } else {
1741                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1742                 zio_flags |= ZIO_FLAG_RESILVER;
1743                 needs_io = B_FALSE;
1744                 scan_delay = zfs_resilver_delay;
1745         }
1746
1747         /* If it's an intent log block, failure is expected. */
1748         if (zb->zb_level == ZB_ZIL_LEVEL)
1749                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1750
1751         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1752                 vdev_t *vd = vdev_lookup_top(spa,
1753                     DVA_GET_VDEV(&bp->blk_dva[d]));
1754
1755                 /*
1756                  * Keep track of how much data we've examined so that
1757                  * zpool(1M) status can make useful progress reports.
1758                  */
1759                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1760                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1761
1762                 /* if it's a resilver, this may not be in the target range */
1763                 if (!needs_io) {
1764                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1765                                 /*
1766                                  * Gang members may be spread across multiple
1767                                  * vdevs, so the best estimate we have is the
1768                                  * scrub range, which has already been checked.
1769                                  * XXX -- it would be better to change our
1770                                  * allocation policy to ensure that all
1771                                  * gang members reside on the same vdev.
1772                                  */
1773                                 needs_io = B_TRUE;
1774                         } else {
1775                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1776                                     phys_birth, 1);
1777                         }
1778                 }
1779         }
1780
1781         if (needs_io && !zfs_no_scrub_io) {
1782                 vdev_t *rvd = spa->spa_root_vdev;
1783                 uint64_t maxinflight = rvd->vdev_children *
1784                     MAX(zfs_top_maxinflight, 1);
1785                 void *data = zio_data_buf_alloc(size);
1786
1787                 mutex_enter(&spa->spa_scrub_lock);
1788                 while (spa->spa_scrub_inflight >= maxinflight)
1789                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1790                 spa->spa_scrub_inflight++;
1791                 mutex_exit(&spa->spa_scrub_lock);
1792
1793                 /*
1794                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1795                  * then throttle our workload to limit the impact of a scan.
1796                  */
1797                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1798                         delay(MAX((int)scan_delay, 0));
1799
1800                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1801                     dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1802                     zio_flags, zb));
1803         }
1804
1805         /* do not relocate this block */
1806         return (0);
1807 }
1808
1809 int
1810 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1811 {
1812         spa_t *spa = dp->dp_spa;
1813
1814         /*
1815          * Purge all vdev caches and probe all devices.  We do this here
1816          * rather than in sync context because this requires a writer lock
1817          * on the spa_config lock, which we can't do from sync context.  The
1818          * spa_scrub_reopen flag indicates that vdev_open() should not
1819          * attempt to start another scrub.
1820          */
1821         spa_vdev_state_enter(spa, SCL_NONE);
1822         spa->spa_scrub_reopen = B_TRUE;
1823         vdev_reopen(spa->spa_root_vdev);
1824         spa->spa_scrub_reopen = B_FALSE;
1825         (void) spa_vdev_state_exit(spa, NULL, 0);
1826
1827         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1828             dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1829 }