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
 103 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
 104         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
 105         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
 106
 107 extern int zfs_txg_timeout;
 108
 109 /* the order has to match pool_scan_type */
 110 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
 111         NULL,
 112         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
 113         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
 114 };
 115
 116 int
 117 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
 118 {
 119         int err;
 120         dsl_scan_t *scn;
 121         spa_t *spa = dp->dp_spa;
 122         uint64_t f;
 123
 124         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
 125         scn->scn_dp = dp;
 126
 127         /*
 128          * It's possible that we're resuming a scan after a reboot so
 129          * make sure that the scan_async_destroying flag is initialized
 130          * appropriately.
 131          */
 132         ASSERT(!scn->scn_async_destroying);
 133         scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
 134             SPA_FEATURE_ASYNC_DESTROY);
 135
 136         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 137             "scrub_func", sizeof (uint64_t), 1, &f);
 138         if (err == 0) {
 139                 /*
 140                  * There was an old-style scrub in progress.  Restart a
 141                  * new-style scrub from the beginning.
 142                  */
 143                 scn->scn_restart_txg = txg;
 144                 zfs_dbgmsg("old-style scrub was in progress; "
 145                     "restarting new-style scrub in txg %llu",
 146                     scn->scn_restart_txg);
 147
 148                 /*
 149                  * Load the queue obj from the old location so that it
 150                  * can be freed by dsl_scan_done().
 151                  */
 152                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 153                     "scrub_queue", sizeof (uint64_t), 1,
 154                     &scn->scn_phys.scn_queue_obj);
 155         } else {
 156                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 157                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 158                     &scn->scn_phys);
 159                 if (err == ENOENT)
 160                         return (0);
 161                 else if (err)
 162                         return (err);
 163
 164                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 165                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 166                         /*
 167                          * A new-type scrub was in progress on an old
 168                          * pool, and the pool was accessed by old
 169                          * software.  Restart from the beginning, since
 170                          * the old software may have changed the pool in
 171                          * the meantime.
 172                          */
 173                         scn->scn_restart_txg = txg;
 174                         zfs_dbgmsg("new-style scrub was modified "
 175                             "by old software; restarting in txg %llu",
 176                             scn->scn_restart_txg);
 177                 }
 178         }
 179
 180         spa_scan_stat_init(spa);
 181         return (0);
 182 }
 183
 184 void
 185 dsl_scan_fini(dsl_pool_t *dp)
 186 {
 187         if (dp->dp_scan) {
 188                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 189                 dp->dp_scan = NULL;
 190         }
 191 }
 192
 193 /* ARGSUSED */
 194 static int
 195 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 196 {
 197         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 198
 199         if (scn->scn_phys.scn_state == DSS_SCANNING)
 200                 return (SET_ERROR(EBUSY));
 201
 202         return (0);
 203 }
 204
 205 static void
 206 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 207 {
 208         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 209         pool_scan_func_t *funcp = arg;
 210         dmu_object_type_t ot = 0;
 211         dsl_pool_t *dp = scn->scn_dp;
 212         spa_t *spa = dp->dp_spa;
 213
 214         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 215         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 216         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 217         scn->scn_phys.scn_func = *funcp;
 218         scn->scn_phys.scn_state = DSS_SCANNING;
 219         scn->scn_phys.scn_min_txg = 0;
 220         scn->scn_phys.scn_max_txg = tx->tx_txg;
 221         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 222         scn->scn_phys.scn_start_time = gethrestime_sec();
 223         scn->scn_phys.scn_errors = 0;
 224         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 225         scn->scn_restart_txg = 0;
 226         scn->scn_done_txg = 0;
 227         spa_scan_stat_init(spa);
 228
 229         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 230                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 231
 232                 /* rewrite all disk labels */
 233                 vdev_config_dirty(spa->spa_root_vdev);
 234
 235                 if (vdev_resilver_needed(spa->spa_root_vdev,
 236                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 237                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 238                 } else {
 239                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 240                 }
 241
 242                 spa->spa_scrub_started = B_TRUE;
 243                 /*
 244                  * If this is an incremental scrub, limit the DDT scrub phase
 245                  * to just the auto-ditto class (for correctness); the rest
 246                  * of the scrub should go faster using top-down pruning.
 247                  */
 248                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 249                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 250
 251         }
 252
 253         /* back to the generic stuff */
 254
 255         if (dp->dp_blkstats == NULL) {
 256                 dp->dp_blkstats =
 257                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 258         }
 259         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 260
 261         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 262                 ot = DMU_OT_ZAP_OTHER;
 263
 264         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 265             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 266
 267         dsl_scan_sync_state(scn, tx);
 268
 269         spa_history_log_internal(spa, "scan setup", tx,
 270             "func=%u mintxg=%llu maxtxg=%llu",
 271             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 272 }
 273
 274 /* ARGSUSED */
 275 static void
 276 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 277 {
 278         static const char *old_names[] = {
 279                 "scrub_bookmark",
 280                 "scrub_ddt_bookmark",
 281                 "scrub_ddt_class_max",
 282                 "scrub_queue",
 283                 "scrub_min_txg",
 284                 "scrub_max_txg",
 285                 "scrub_func",
 286                 "scrub_errors",
 287                 NULL
 288         };
 289
 290         dsl_pool_t *dp = scn->scn_dp;
 291         spa_t *spa = dp->dp_spa;
 292         int i;
 293
 294         /* Remove any remnants of an old-style scrub. */
 295         for (i = 0; old_names[i]; i++) {
 296                 (void) zap_remove(dp->dp_meta_objset,
 297                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 298         }
 299
 300         if (scn->scn_phys.scn_queue_obj != 0) {
 301                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 302                     scn->scn_phys.scn_queue_obj, tx));
 303                 scn->scn_phys.scn_queue_obj = 0;
 304         }
 305
 306         /*
 307          * If we were "restarted" from a stopped state, don't bother
 308          * with anything else.
 309          */
 310         if (scn->scn_phys.scn_state != DSS_SCANNING)
 311                 return;
 312
 313         if (complete)
 314                 scn->scn_phys.scn_state = DSS_FINISHED;
 315         else
 316                 scn->scn_phys.scn_state = DSS_CANCELED;
 317
 318         spa_history_log_internal(spa, "scan done", tx,
 319             "complete=%u", complete);
 320
 321         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 322                 mutex_enter(&spa->spa_scrub_lock);
 323                 while (spa->spa_scrub_inflight > 0) {
 324                         cv_wait(&spa->spa_scrub_io_cv,
 325                             &spa->spa_scrub_lock);
 326                 }
 327                 mutex_exit(&spa->spa_scrub_lock);
 328                 spa->spa_scrub_started = B_FALSE;
 329                 spa->spa_scrub_active = B_FALSE;
 330
 331                 /*
 332                  * If the scrub/resilver completed, update all DTLs to
 333                  * reflect this.  Whether it succeeded or not, vacate
 334                  * all temporary scrub DTLs.
 335                  */
 336                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 337                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 338                 if (complete) {
 339                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 340                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 341                 }
 342                 spa_errlog_rotate(spa);
 343
 344                 /*
 345                  * We may have finished replacing a device.
 346                  * Let the async thread assess this and handle the detach.
 347                  */
 348                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 349         }
 350
 351         scn->scn_phys.scn_end_time = gethrestime_sec();
 352 }
 353
 354 /* ARGSUSED */
 355 static int
 356 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 357 {
 358         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 359
 360         if (scn->scn_phys.scn_state != DSS_SCANNING)
 361                 return (SET_ERROR(ENOENT));
 362         return (0);
 363 }
 364
 365 /* ARGSUSED */
 366 static void
 367 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 368 {
 369         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 370
 371         dsl_scan_done(scn, B_FALSE, tx);
 372         dsl_scan_sync_state(scn, tx);
 373 }
 374
 375 int
 376 dsl_scan_cancel(dsl_pool_t *dp)
 377 {
 378         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 379             dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
 380 }
 381
 382 static void dsl_scan_visitbp(blkptr_t *bp,
 383     const zbookmark_phys_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 384     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 385     dmu_tx_t *tx);
 386 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 387     dmu_objset_type_t ostype,
 388     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 389
 390 void
 391 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 392 {
 393         zio_free(dp->dp_spa, txg, bp);
 394 }
 395
 396 void
 397 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 398 {
 399         ASSERT(dsl_pool_sync_context(dp));
 400         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 401             pio->io_flags));
 402 }
 403
 404 static uint64_t
 405 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 406 {
 407         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 408         if (dsl_dataset_is_snapshot(ds))
 409                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 410         return (smt);
 411 }
 412
 413 static void
 414 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 415 {
 416         VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 417             DMU_POOL_DIRECTORY_OBJECT,
 418             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 419             &scn->scn_phys, tx));
 420 }
 421
 422 static boolean_t
 423 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
 424 {
 425         uint64_t elapsed_nanosecs;
 426         unsigned int mintime;
 427
 428         /* we never skip user/group accounting objects */
 429         if (zb && (int64_t)zb->zb_object < 0)
 430                 return (B_FALSE);
 431
 432         if (scn->scn_pausing)
 433                 return (B_TRUE); /* we're already pausing */
 434
 435         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
 436                 return (B_FALSE); /* we're resuming */
 437
 438         /* We only know how to resume from level-0 blocks. */
 439         if (zb && zb->zb_level != 0)
 440                 return (B_FALSE);
 441
 442         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 443             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 444         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 445         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 446             (NSEC2MSEC(elapsed_nanosecs) > mintime &&
 447             txg_sync_waiting(scn->scn_dp)) ||
 448             spa_shutting_down(scn->scn_dp->dp_spa)) {
 449                 if (zb) {
 450                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 451                             (longlong_t)zb->zb_objset,
 452                             (longlong_t)zb->zb_object,
 453                             (longlong_t)zb->zb_level,
 454                             (longlong_t)zb->zb_blkid);
 455                         scn->scn_phys.scn_bookmark = *zb;
 456                 }
 457                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 458                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 459                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 460                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 461                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 462                 scn->scn_pausing = B_TRUE;
 463                 return (B_TRUE);
 464         }
 465         return (B_FALSE);
 466 }
 467
 468 typedef struct zil_scan_arg {
 469         dsl_pool_t      *zsa_dp;
 470         zil_header_t    *zsa_zh;
 471 } zil_scan_arg_t;
 472
 473 /* ARGSUSED */
 474 static int
 475 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 476 {
 477         zil_scan_arg_t *zsa = arg;
 478         dsl_pool_t *dp = zsa->zsa_dp;
 479         dsl_scan_t *scn = dp->dp_scan;
 480         zil_header_t *zh = zsa->zsa_zh;
 481         zbookmark_phys_t zb;
 482
 483         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 484                 return (0);
 485
 486         /*
 487          * One block ("stubby") can be allocated a long time ago; we
 488          * want to visit that one because it has been allocated
 489          * (on-disk) even if it hasn't been claimed (even though for
 490          * scrub there's nothing to do to it).
 491          */
 492         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 493                 return (0);
 494
 495         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 496             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 497
 498         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 499         return (0);
 500 }
 501
 502 /* ARGSUSED */
 503 static int
 504 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 505 {
 506         if (lrc->lrc_txtype == TX_WRITE) {
 507                 zil_scan_arg_t *zsa = arg;
 508                 dsl_pool_t *dp = zsa->zsa_dp;
 509                 dsl_scan_t *scn = dp->dp_scan;
 510                 zil_header_t *zh = zsa->zsa_zh;
 511                 lr_write_t *lr = (lr_write_t *)lrc;
 512                 blkptr_t *bp = &lr->lr_blkptr;
 513                 zbookmark_phys_t zb;
 514
 515                 if (BP_IS_HOLE(bp) ||
 516                     bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 517                         return (0);
 518
 519                 /*
 520                  * birth can be < claim_txg if this record's txg is
 521                  * already txg sync'ed (but this log block contains
 522                  * other records that are not synced)
 523                  */
 524                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 525                         return (0);
 526
 527                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 528                     lr->lr_foid, ZB_ZIL_LEVEL,
 529                     lr->lr_offset / BP_GET_LSIZE(bp));
 530
 531                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 532         }
 533         return (0);
 534 }
 535
 536 static void
 537 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 538 {
 539         uint64_t claim_txg = zh->zh_claim_txg;
 540         zil_scan_arg_t zsa = { dp, zh };
 541         zilog_t *zilog;
 542
 543         /*
 544          * We only want to visit blocks that have been claimed but not yet
 545          * replayed (or, in read-only mode, blocks that *would* be claimed).
 546          */
 547         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 548                 return;
 549
 550         zilog = zil_alloc(dp->dp_meta_objset, zh);
 551
 552         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 553             claim_txg);
 554
 555         zil_free(zilog);
 556 }
 557
 558 /* ARGSUSED */
 559 static void
 560 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 561     uint64_t objset, uint64_t object, uint64_t blkid)
 562 {
 563         zbookmark_phys_t czb;
 564         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 565
 566         if (zfs_no_scrub_prefetch)
 567                 return;
 568
 569         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 570             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 571                 return;
 572
 573         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 574
 575         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 576             NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 577             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 578 }
 579
 580 static boolean_t
 581 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 582     const zbookmark_phys_t *zb)
 583 {
 584         /*
 585          * We never skip over user/group accounting objects (obj<0)
 586          */
 587         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
 588             (int64_t)zb->zb_object >= 0) {
 589                 /*
 590                  * If we already visited this bp & everything below (in
 591                  * a prior txg sync), don't bother doing it again.
 592                  */
 593                 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 594                         return (B_TRUE);
 595
 596                 /*
 597                  * If we found the block we're trying to resume from, or
 598                  * we went past it to a different object, zero it out to
 599                  * indicate that it's OK to start checking for pausing
 600                  * again.
 601                  */
 602                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 603                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 604                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 605                             (longlong_t)zb->zb_objset,
 606                             (longlong_t)zb->zb_object,
 607                             (longlong_t)zb->zb_level,
 608                             (longlong_t)zb->zb_blkid);
 609                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 610                 }
 611         }
 612         return (B_FALSE);
 613 }
 614
 615 /*
 616  * Return nonzero on i/o error.
 617  * Return new buf to write out in *bufp.
 618  */
 619 static int
 620 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 621     dnode_phys_t *dnp, const blkptr_t *bp,
 622     const zbookmark_phys_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 623 {
 624         dsl_pool_t *dp = scn->scn_dp;
 625         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 626         int err;
 627
 628         if (BP_GET_LEVEL(bp) > 0) {
 629                 uint32_t flags = ARC_WAIT;
 630                 int i;
 631                 blkptr_t *cbp;
 632                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 633
 634                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 635                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 636                 if (err) {
 637                         scn->scn_phys.scn_errors++;
 638                         return (err);
 639                 }
 640                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 641                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 642                             zb->zb_object, zb->zb_blkid * epb + i);
 643                 }
 644                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 645                         zbookmark_phys_t czb;
 646
 647                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 648                             zb->zb_level - 1,
 649                             zb->zb_blkid * epb + i);
 650                         dsl_scan_visitbp(cbp, &czb, dnp,
 651                             *bufp, ds, scn, ostype, tx);
 652                 }
 653         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 654                 uint32_t flags = ARC_WAIT;
 655
 656                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 657                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 658                 if (err) {
 659                         scn->scn_phys.scn_errors++;
 660                         return (err);
 661                 }
 662         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 663                 uint32_t flags = ARC_WAIT;
 664                 dnode_phys_t *cdnp;
 665                 int i, j;
 666                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 667
 668                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 669                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 670                 if (err) {
 671                         scn->scn_phys.scn_errors++;
 672                         return (err);
 673                 }
 674                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 675                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 676                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 677                                 dsl_scan_prefetch(scn, *bufp, cbp,
 678                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 679                         }
 680                 }
 681                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 682                         dsl_scan_visitdnode(scn, ds, ostype,
 683                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 684                 }
 685
 686         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 687                 uint32_t flags = ARC_WAIT;
 688                 objset_phys_t *osp;
 689
 690                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
 691                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 692                 if (err) {
 693                         scn->scn_phys.scn_errors++;
 694                         return (err);
 695                 }
 696
 697                 osp = (*bufp)->b_data;
 698
 699                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 700                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 701
 702                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 703                         /*
 704                          * We also always visit user/group accounting
 705                          * objects, and never skip them, even if we are
 706                          * pausing.  This is necessary so that the space
 707                          * deltas from this txg get integrated.
 708                          */
 709                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 710                             &osp->os_groupused_dnode, *bufp,
 711                             DMU_GROUPUSED_OBJECT, tx);
 712                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 713                             &osp->os_userused_dnode, *bufp,
 714                             DMU_USERUSED_OBJECT, tx);
 715                 }
 716         }
 717
 718         return (0);
 719 }
 720
 721 static void
 722 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 723     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 724     uint64_t object, dmu_tx_t *tx)
 725 {
 726         int j;
 727
 728         for (j = 0; j < dnp->dn_nblkptr; j++) {
 729                 zbookmark_phys_t czb;
 730
 731                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 732                     dnp->dn_nlevels - 1, j);
 733                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 734                     &czb, dnp, buf, ds, scn, ostype, tx);
 735         }
 736
 737         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 738                 zbookmark_phys_t czb;
 739                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 740                     0, DMU_SPILL_BLKID);
 741                 dsl_scan_visitbp(&dnp->dn_spill,
 742                     &czb, dnp, buf, ds, scn, ostype, tx);
 743         }
 744 }
 745
 746 /*
 747  * The arguments are in this order because mdb can only print the
 748  * first 5; we want them to be useful.
 749  */
 750 static void
 751 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
 752     dnode_phys_t *dnp, arc_buf_t *pbuf,
 753     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 754     dmu_tx_t *tx)
 755 {
 756         dsl_pool_t *dp = scn->scn_dp;
 757         arc_buf_t *buf = NULL;
 758         blkptr_t bp_toread = *bp;
 759
 760         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 761
 762         if (dsl_scan_check_pause(scn, zb))
 763                 return;
 764
 765         if (dsl_scan_check_resume(scn, dnp, zb))
 766                 return;
 767
 768         if (BP_IS_HOLE(bp))
 769                 return;
 770
 771         scn->scn_visited_this_txg++;
 772
 773         dprintf_bp(bp,
 774             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 775             ds, ds ? ds->ds_object : 0,
 776             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 777             pbuf, bp);
 778
 779         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 780                 return;
 781
 782         if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
 783             &buf) != 0)
 784                 return;
 785
 786         /*
 787          * If dsl_scan_ddt() has aready visited this block, it will have
 788          * already done any translations or scrubbing, so don't call the
 789          * callback again.
 790          */
 791         if (ddt_class_contains(dp->dp_spa,
 792             scn->scn_phys.scn_ddt_class_max, bp)) {
 793                 ASSERT(buf == NULL);
 794                 return;
 795         }
 796
 797         /*
 798          * If this block is from the future (after cur_max_txg), then we
 799          * are doing this on behalf of a deleted snapshot, and we will
 800          * revisit the future block on the next pass of this dataset.
 801          * Don't scan it now unless we need to because something
 802          * under it was modified.
 803          */
 804         if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
 805                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 806         }
 807         if (buf)
 808                 (void) arc_buf_remove_ref(buf, &buf);
 809 }
 810
 811 static void
 812 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 813     dmu_tx_t *tx)
 814 {
 815         zbookmark_phys_t zb;
 816
 817         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 818             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 819         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 820             ds, scn, DMU_OST_NONE, tx);
 821
 822         dprintf_ds(ds, "finished scan%s", "");
 823 }
 824
 825 void
 826 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 827 {
 828         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 829         dsl_scan_t *scn = dp->dp_scan;
 830         uint64_t mintxg;
 831
 832         if (scn->scn_phys.scn_state != DSS_SCANNING)
 833                 return;
 834
 835         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 836                 if (dsl_dataset_is_snapshot(ds)) {
 837                         /* Note, scn_cur_{min,max}_txg stays the same. */
 838                         scn->scn_phys.scn_bookmark.zb_objset =
 839                             ds->ds_phys->ds_next_snap_obj;
 840                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 841                             "reset zb_objset to %llu",
 842                             (u_longlong_t)ds->ds_object,
 843                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 844                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 845                 } else {
 846                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 847                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 848                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 849                             "reset bookmark to -1,0,0,0",
 850                             (u_longlong_t)ds->ds_object);
 851                 }
 852         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 853             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 854                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 855                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 856                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 857                 if (dsl_dataset_is_snapshot(ds)) {
 858                         /*
 859                          * We keep the same mintxg; it could be >
 860                          * ds_creation_txg if the previous snapshot was
 861                          * deleted too.
 862                          */
 863                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 864                             scn->scn_phys.scn_queue_obj,
 865                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 866                         zfs_dbgmsg("destroying ds %llu; in queue; "
 867                             "replacing with %llu",
 868                             (u_longlong_t)ds->ds_object,
 869                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 870                 } else {
 871                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 872                             (u_longlong_t)ds->ds_object);
 873                 }
 874         } else {
 875                 zfs_dbgmsg("destroying ds %llu; ignoring",
 876                     (u_longlong_t)ds->ds_object);
 877         }
 878
 879         /*
 880          * dsl_scan_sync() should be called after this, and should sync
 881          * out our changed state, but just to be safe, do it here.
 882          */
 883         dsl_scan_sync_state(scn, tx);
 884 }
 885
 886 void
 887 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 888 {
 889         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 890         dsl_scan_t *scn = dp->dp_scan;
 891         uint64_t mintxg;
 892
 893         if (scn->scn_phys.scn_state != DSS_SCANNING)
 894                 return;
 895
 896         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 897
 898         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 899                 scn->scn_phys.scn_bookmark.zb_objset =
 900                     ds->ds_phys->ds_prev_snap_obj;
 901                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 902                     "reset zb_objset to %llu",
 903                     (u_longlong_t)ds->ds_object,
 904                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 905         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 906             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 907                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 908                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 909                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 910                     scn->scn_phys.scn_queue_obj,
 911                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 912                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 913                     "replacing with %llu",
 914                     (u_longlong_t)ds->ds_object,
 915                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 916         }
 917         dsl_scan_sync_state(scn, tx);
 918 }
 919
 920 void
 921 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 922 {
 923         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 924         dsl_scan_t *scn = dp->dp_scan;
 925         uint64_t mintxg;
 926
 927         if (scn->scn_phys.scn_state != DSS_SCANNING)
 928                 return;
 929
 930         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 931                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 932                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 933                     "reset zb_objset to %llu",
 934                     (u_longlong_t)ds1->ds_object,
 935                     (u_longlong_t)ds2->ds_object);
 936         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 937                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 938                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 939                     "reset zb_objset to %llu",
 940                     (u_longlong_t)ds2->ds_object,
 941                     (u_longlong_t)ds1->ds_object);
 942         }
 943
 944         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
 945             ds1->ds_object, &mintxg) == 0) {
 946                 int err;
 947
 948                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 949                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 950                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 951                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
 952                 err = zap_add_int_key(dp->dp_meta_objset,
 953                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
 954                 VERIFY(err == 0 || err == EEXIST);
 955                 if (err == EEXIST) {
 956                         /* Both were there to begin with */
 957                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 958                             scn->scn_phys.scn_queue_obj,
 959                             ds1->ds_object, mintxg, tx));
 960                 }
 961                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 962                     "replacing with %llu",
 963                     (u_longlong_t)ds1->ds_object,
 964                     (u_longlong_t)ds2->ds_object);
 965         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 966             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
 967                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
 968                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
 969                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 970                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
 971                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
 972                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
 973                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
 974                     "replacing with %llu",
 975                     (u_longlong_t)ds2->ds_object,
 976                     (u_longlong_t)ds1->ds_object);
 977         }
 978
 979         dsl_scan_sync_state(scn, tx);
 980 }
 981
 982 struct enqueue_clones_arg {
 983         dmu_tx_t *tx;
 984         uint64_t originobj;
 985 };
 986
 987 /* ARGSUSED */
 988 static int
 989 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
 990 {
 991         struct enqueue_clones_arg *eca = arg;
 992         dsl_dataset_t *ds;
 993         int err;
 994         dsl_scan_t *scn = dp->dp_scan;
 995
 996         if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
 997                 return (0);
 998
 999         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1000         if (err)
1001                 return (err);
1002
1003         while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1004                 dsl_dataset_t *prev;
1005                 err = dsl_dataset_hold_obj(dp,
1006                     ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1007
1008                 dsl_dataset_rele(ds, FTAG);
1009                 if (err)
1010                         return (err);
1011                 ds = prev;
1012         }
1013         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1014             scn->scn_phys.scn_queue_obj, ds->ds_object,
1015             ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1016         dsl_dataset_rele(ds, FTAG);
1017         return (0);
1018 }
1019
1020 static void
1021 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1022 {
1023         dsl_pool_t *dp = scn->scn_dp;
1024         dsl_dataset_t *ds;
1025         objset_t *os;
1026
1027         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1028
1029         if (dmu_objset_from_ds(ds, &os))
1030                 goto out;
1031
1032         /*
1033          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1034          * snapshots can have ZIL block pointers (which may be the same
1035          * BP as in the head), they must be ignored.  So we traverse the
1036          * ZIL here, rather than in scan_recurse(), because the regular
1037          * snapshot block-sharing rules don't apply to it.
1038          */
1039         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1040                 dsl_scan_zil(dp, &os->os_zil_header);
1041
1042         /*
1043          * Iterate over the bps in this ds.
1044          */
1045         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1046         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1047
1048         char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1049         dsl_dataset_name(ds, dsname);
1050         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1051             "pausing=%u",
1052             (longlong_t)dsobj, dsname,
1053             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1054             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1055             (int)scn->scn_pausing);
1056         kmem_free(dsname, ZFS_MAXNAMELEN);
1057
1058         if (scn->scn_pausing)
1059                 goto out;
1060
1061         /*
1062          * We've finished this pass over this dataset.
1063          */
1064
1065         /*
1066          * If we did not completely visit this dataset, do another pass.
1067          */
1068         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1069                 zfs_dbgmsg("incomplete pass; visiting again");
1070                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1071                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1072                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1073                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1074                 goto out;
1075         }
1076
1077         /*
1078          * Add descendent datasets to work queue.
1079          */
1080         if (ds->ds_phys->ds_next_snap_obj != 0) {
1081                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1082                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1083                     ds->ds_phys->ds_creation_txg, tx) == 0);
1084         }
1085         if (ds->ds_phys->ds_num_children > 1) {
1086                 boolean_t usenext = B_FALSE;
1087                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1088                         uint64_t count;
1089                         /*
1090                          * A bug in a previous version of the code could
1091                          * cause upgrade_clones_cb() to not set
1092                          * ds_next_snap_obj when it should, leading to a
1093                          * missing entry.  Therefore we can only use the
1094                          * next_clones_obj when its count is correct.
1095                          */
1096                         int err = zap_count(dp->dp_meta_objset,
1097                             ds->ds_phys->ds_next_clones_obj, &count);
1098                         if (err == 0 &&
1099                             count == ds->ds_phys->ds_num_children - 1)
1100                                 usenext = B_TRUE;
1101                 }
1102
1103                 if (usenext) {
1104                         VERIFY0(zap_join_key(dp->dp_meta_objset,
1105                             ds->ds_phys->ds_next_clones_obj,
1106                             scn->scn_phys.scn_queue_obj,
1107                             ds->ds_phys->ds_creation_txg, tx));
1108                 } else {
1109                         struct enqueue_clones_arg eca;
1110                         eca.tx = tx;
1111                         eca.originobj = ds->ds_object;
1112
1113                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1114                             enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1115                 }
1116         }
1117
1118 out:
1119         dsl_dataset_rele(ds, FTAG);
1120 }
1121
1122 /* ARGSUSED */
1123 static int
1124 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1125 {
1126         dmu_tx_t *tx = arg;
1127         dsl_dataset_t *ds;
1128         int err;
1129         dsl_scan_t *scn = dp->dp_scan;
1130
1131         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1132         if (err)
1133                 return (err);
1134
1135         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1136                 dsl_dataset_t *prev;
1137                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1138                     FTAG, &prev);
1139                 if (err) {
1140                         dsl_dataset_rele(ds, FTAG);
1141                         return (err);
1142                 }
1143
1144                 /*
1145                  * If this is a clone, we don't need to worry about it for now.
1146                  */
1147                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1148                         dsl_dataset_rele(ds, FTAG);
1149                         dsl_dataset_rele(prev, FTAG);
1150                         return (0);
1151                 }
1152                 dsl_dataset_rele(ds, FTAG);
1153                 ds = prev;
1154         }
1155
1156         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1157             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1158         dsl_dataset_rele(ds, FTAG);
1159         return (0);
1160 }
1161
1162 /*
1163  * Scrub/dedup interaction.
1164  *
1165  * If there are N references to a deduped block, we don't want to scrub it
1166  * N times -- ideally, we should scrub it exactly once.
1167  *
1168  * We leverage the fact that the dde's replication class (enum ddt_class)
1169  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1170  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1171  *
1172  * To prevent excess scrubbing, the scrub begins by walking the DDT
1173  * to find all blocks with refcnt > 1, and scrubs each of these once.
1174  * Since there are two replication classes which contain blocks with
1175  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1176  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1177  *
1178  * There would be nothing more to say if a block's refcnt couldn't change
1179  * during a scrub, but of course it can so we must account for changes
1180  * in a block's replication class.
1181  *
1182  * Here's an example of what can occur:
1183  *
1184  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1185  * when visited during the top-down scrub phase, it will be scrubbed twice.
1186  * This negates our scrub optimization, but is otherwise harmless.
1187  *
1188  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1189  * on each visit during the top-down scrub phase, it will never be scrubbed.
1190  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1191  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1192  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1193  * while a scrub is in progress, it scrubs the block right then.
1194  */
1195 static void
1196 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1197 {
1198         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1199         ddt_entry_t dde = { 0 };
1200         int error;
1201         uint64_t n = 0;
1202
1203         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1204                 ddt_t *ddt;
1205
1206                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1207                         break;
1208                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1209                     (longlong_t)ddb->ddb_class,
1210                     (longlong_t)ddb->ddb_type,
1211                     (longlong_t)ddb->ddb_checksum,
1212                     (longlong_t)ddb->ddb_cursor);
1213
1214                 /* There should be no pending changes to the dedup table */
1215                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1216                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1217
1218                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1219                 n++;
1220
1221                 if (dsl_scan_check_pause(scn, NULL))
1222                         break;
1223         }
1224
1225         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1226             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1227             (int)scn->scn_pausing);
1228
1229         ASSERT(error == 0 || error == ENOENT);
1230         ASSERT(error != ENOENT ||
1231             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1232 }
1233
1234 /* ARGSUSED */
1235 void
1236 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1237     ddt_entry_t *dde, dmu_tx_t *tx)
1238 {
1239         const ddt_key_t *ddk = &dde->dde_key;
1240         ddt_phys_t *ddp = dde->dde_phys;
1241         blkptr_t bp;
1242         zbookmark_phys_t zb = { 0 };
1243
1244         if (scn->scn_phys.scn_state != DSS_SCANNING)
1245                 return;
1246
1247         for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1248                 if (ddp->ddp_phys_birth == 0 ||
1249                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1250                         continue;
1251                 ddt_bp_create(checksum, ddk, ddp, &bp);
1252
1253                 scn->scn_visited_this_txg++;
1254                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1255         }
1256 }
1257
1258 static void
1259 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1260 {
1261         dsl_pool_t *dp = scn->scn_dp;
1262         zap_cursor_t zc;
1263         zap_attribute_t za;
1264
1265         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1266             scn->scn_phys.scn_ddt_class_max) {
1267                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1268                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1269                 dsl_scan_ddt(scn, tx);
1270                 if (scn->scn_pausing)
1271                         return;
1272         }
1273
1274         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1275                 /* First do the MOS & ORIGIN */
1276
1277                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1278                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1279                 dsl_scan_visit_rootbp(scn, NULL,
1280                     &dp->dp_meta_rootbp, tx);
1281                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1282                 if (scn->scn_pausing)
1283                         return;
1284
1285                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1286                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1287                             enqueue_cb, tx, DS_FIND_CHILDREN));
1288                 } else {
1289                         dsl_scan_visitds(scn,
1290                             dp->dp_origin_snap->ds_object, tx);
1291                 }
1292                 ASSERT(!scn->scn_pausing);
1293         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1294             ZB_DESTROYED_OBJSET) {
1295                 /*
1296                  * If we were paused, continue from here.  Note if the
1297                  * ds we were paused on was deleted, the zb_objset may
1298                  * be -1, so we will skip this and find a new objset
1299                  * below.
1300                  */
1301                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1302                 if (scn->scn_pausing)
1303                         return;
1304         }
1305
1306         /*
1307          * In case we were paused right at the end of the ds, zero the
1308          * bookmark so we don't think that we're still trying to resume.
1309          */
1310         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1311
1312         /* keep pulling things out of the zap-object-as-queue */
1313         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1314             scn->scn_phys.scn_queue_obj),
1315             zap_cursor_retrieve(&zc, &za) == 0) {
1316                 dsl_dataset_t *ds;
1317                 uint64_t dsobj;
1318
1319                 dsobj = strtonum(za.za_name, NULL);
1320                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1321                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1322
1323                 /* Set up min/max txg */
1324                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1325                 if (za.za_first_integer != 0) {
1326                         scn->scn_phys.scn_cur_min_txg =
1327                             MAX(scn->scn_phys.scn_min_txg,
1328                             za.za_first_integer);
1329                 } else {
1330                         scn->scn_phys.scn_cur_min_txg =
1331                             MAX(scn->scn_phys.scn_min_txg,
1332                             ds->ds_phys->ds_prev_snap_txg);
1333                 }
1334                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1335                 dsl_dataset_rele(ds, FTAG);
1336
1337                 dsl_scan_visitds(scn, dsobj, tx);
1338                 zap_cursor_fini(&zc);
1339                 if (scn->scn_pausing)
1340                         return;
1341         }
1342         zap_cursor_fini(&zc);
1343 }
1344
1345 static boolean_t
1346 dsl_scan_free_should_pause(dsl_scan_t *scn)
1347 {
1348         uint64_t elapsed_nanosecs;
1349
1350         if (zfs_recover)
1351                 return (B_FALSE);
1352
1353         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1354         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1355             (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1356             txg_sync_waiting(scn->scn_dp)) ||
1357             spa_shutting_down(scn->scn_dp->dp_spa));
1358 }
1359
1360 static int
1361 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1362 {
1363         dsl_scan_t *scn = arg;
1364
1365         if (!scn->scn_is_bptree ||
1366             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1367                 if (dsl_scan_free_should_pause(scn))
1368                         return (SET_ERROR(ERESTART));
1369         }
1370
1371         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1372             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1373         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1374             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1375             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1376         scn->scn_visited_this_txg++;
1377         return (0);
1378 }
1379
1380 boolean_t
1381 dsl_scan_active(dsl_scan_t *scn)
1382 {
1383         spa_t *spa = scn->scn_dp->dp_spa;
1384         uint64_t used = 0, comp, uncomp;
1385
1386         if (spa->spa_load_state != SPA_LOAD_NONE)
1387                 return (B_FALSE);
1388         if (spa_shutting_down(spa))
1389                 return (B_FALSE);
1390         if (scn->scn_phys.scn_state == DSS_SCANNING ||
1391             (scn->scn_async_destroying && !scn->scn_async_stalled))
1392                 return (B_TRUE);
1393
1394         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1395                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1396                     &used, &comp, &uncomp);
1397         }
1398         return (used != 0);
1399 }
1400
1401 void
1402 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1403 {
1404         dsl_scan_t *scn = dp->dp_scan;
1405         spa_t *spa = dp->dp_spa;
1406         int err = 0;
1407
1408         /*
1409          * Check for scn_restart_txg before checking spa_load_state, so
1410          * that we can restart an old-style scan while the pool is being
1411          * imported (see dsl_scan_init).
1412          */
1413         if (scn->scn_restart_txg != 0 &&
1414             scn->scn_restart_txg <= tx->tx_txg) {
1415                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1416                 dsl_scan_done(scn, B_FALSE, tx);
1417                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1418                         func = POOL_SCAN_RESILVER;
1419                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1420                     func, tx->tx_txg);
1421                 dsl_scan_setup_sync(&func, tx);
1422         }
1423
1424         /*
1425          * If the scan is inactive due to a stalled async destroy, try again.
1426          */
1427         if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1428             spa_sync_pass(dp->dp_spa) > 1)
1429                 return;
1430
1431         scn->scn_visited_this_txg = 0;
1432         scn->scn_pausing = B_FALSE;
1433         scn->scn_sync_start_time = gethrtime();
1434         spa->spa_scrub_active = B_TRUE;
1435
1436         /*
1437          * First process the async destroys.  If we pause, don't do
1438          * any scrubbing or resilvering.  This ensures that there are no
1439          * async destroys while we are scanning, so the scan code doesn't
1440          * have to worry about traversing it.  It is also faster to free the
1441          * blocks than to scrub them.
1442          */
1443         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1444                 scn->scn_is_bptree = B_FALSE;
1445                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1446                     NULL, ZIO_FLAG_MUSTSUCCEED);
1447                 err = bpobj_iterate(&dp->dp_free_bpobj,
1448                     dsl_scan_free_block_cb, scn, tx);
1449                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1450
1451                 if (err != 0 && err != ERESTART)
1452                         zfs_panic_recover("error %u from bpobj_iterate()", err);
1453         }
1454
1455         if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1456                 ASSERT(scn->scn_async_destroying);
1457                 scn->scn_is_bptree = B_TRUE;
1458                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1459                     NULL, ZIO_FLAG_MUSTSUCCEED);
1460                 err = bptree_iterate(dp->dp_meta_objset,
1461                     dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1462                 VERIFY0(zio_wait(scn->scn_zio_root));
1463
1464                 if (err == EIO || err == ECKSUM) {
1465                         err = 0;
1466                 } else if (err != 0 && err != ERESTART) {
1467                         zfs_panic_recover("error %u from "
1468                             "traverse_dataset_destroyed()", err);
1469                 }
1470
1471                 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1472                         /* finished; deactivate async destroy feature */
1473                         spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1474                         ASSERT(!spa_feature_is_active(spa,
1475                             SPA_FEATURE_ASYNC_DESTROY));
1476                         VERIFY0(zap_remove(dp->dp_meta_objset,
1477                             DMU_POOL_DIRECTORY_OBJECT,
1478                             DMU_POOL_BPTREE_OBJ, tx));
1479                         VERIFY0(bptree_free(dp->dp_meta_objset,
1480                             dp->dp_bptree_obj, tx));
1481                         dp->dp_bptree_obj = 0;
1482                         scn->scn_async_destroying = B_FALSE;
1483                 } else {
1484                         /*
1485                          * If we didn't make progress, mark the async destroy as
1486                          * stalled, so that we will not initiate a spa_sync() on
1487                          * its behalf.
1488                          */
1489                         scn->scn_async_stalled =
1490                             (scn->scn_visited_this_txg == 0);
1491                 }
1492         }
1493         if (scn->scn_visited_this_txg) {
1494                 zfs_dbgmsg("freed %llu blocks in %llums from "
1495                     "free_bpobj/bptree txg %llu; err=%u",
1496                     (longlong_t)scn->scn_visited_this_txg,
1497                     (longlong_t)
1498                     NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1499                     (longlong_t)tx->tx_txg, err);
1500                 scn->scn_visited_this_txg = 0;
1501
1502                 /*
1503                  * Write out changes to the DDT that may be required as a
1504                  * result of the blocks freed.  This ensures that the DDT
1505                  * is clean when a scrub/resilver runs.
1506                  */
1507                 ddt_sync(spa, tx->tx_txg);
1508         }
1509         if (err != 0)
1510                 return;
1511         if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1512             (dp->dp_free_dir->dd_phys->dd_used_bytes != 0 ||
1513             dp->dp_free_dir->dd_phys->dd_compressed_bytes != 0 ||
1514             dp->dp_free_dir->dd_phys->dd_uncompressed_bytes != 0)) {
1515                 /*
1516                  * We have finished background destroying, but there is still
1517                  * some space left in the dp_free_dir. Transfer this leaked
1518                  * space to the dp_leak_dir.
1519                  */
1520                 if (dp->dp_leak_dir == NULL) {
1521                         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1522                         (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1523                             LEAK_DIR_NAME, tx);
1524                         VERIFY0(dsl_pool_open_special_dir(dp,
1525                             LEAK_DIR_NAME, &dp->dp_leak_dir));
1526                         rrw_exit(&dp->dp_config_rwlock, FTAG);
1527                 }
1528                 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1529                     dp->dp_free_dir->dd_phys->dd_used_bytes,
1530                     dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1531                     dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1532                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1533                     -dp->dp_free_dir->dd_phys->dd_used_bytes,
1534                     -dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1535                     -dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1536         }
1537         if (!scn->scn_async_destroying) {
1538                 /* finished; verify that space accounting went to zero */
1539                 ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes);
1540                 ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes);
1541                 ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes);
1542         }
1543
1544         if (scn->scn_phys.scn_state != DSS_SCANNING)
1545                 return;
1546
1547         if (scn->scn_done_txg == tx->tx_txg) {
1548                 ASSERT(!scn->scn_pausing);
1549                 /* finished with scan. */
1550                 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1551                 dsl_scan_done(scn, B_TRUE, tx);
1552                 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1553                 dsl_scan_sync_state(scn, tx);
1554                 return;
1555         }
1556
1557         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1558             scn->scn_phys.scn_ddt_class_max) {
1559                 zfs_dbgmsg("doing scan sync txg %llu; "
1560                     "ddt bm=%llu/%llu/%llu/%llx",
1561                     (longlong_t)tx->tx_txg,
1562                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1563                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1564                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1565                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1566                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1567                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1568                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1569                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1570         } else {
1571                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1572                     (longlong_t)tx->tx_txg,
1573                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1574                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1575                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1576                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1577         }
1578
1579         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1580             NULL, ZIO_FLAG_CANFAIL);
1581         dsl_pool_config_enter(dp, FTAG);
1582         dsl_scan_visit(scn, tx);
1583         dsl_pool_config_exit(dp, FTAG);
1584         (void) zio_wait(scn->scn_zio_root);
1585         scn->scn_zio_root = NULL;
1586
1587         zfs_dbgmsg("visited %llu blocks in %llums",
1588             (longlong_t)scn->scn_visited_this_txg,
1589             (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1590
1591         if (!scn->scn_pausing) {
1592                 scn->scn_done_txg = tx->tx_txg + 1;
1593                 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1594                     tx->tx_txg, scn->scn_done_txg);
1595         }
1596
1597         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1598                 mutex_enter(&spa->spa_scrub_lock);
1599                 while (spa->spa_scrub_inflight > 0) {
1600                         cv_wait(&spa->spa_scrub_io_cv,
1601                             &spa->spa_scrub_lock);
1602                 }
1603                 mutex_exit(&spa->spa_scrub_lock);
1604         }
1605
1606         dsl_scan_sync_state(scn, tx);
1607 }
1608
1609 /*
1610  * This will start a new scan, or restart an existing one.
1611  */
1612 void
1613 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1614 {
1615         if (txg == 0) {
1616                 dmu_tx_t *tx;
1617                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1618                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1619
1620                 txg = dmu_tx_get_txg(tx);
1621                 dp->dp_scan->scn_restart_txg = txg;
1622                 dmu_tx_commit(tx);
1623         } else {
1624                 dp->dp_scan->scn_restart_txg = txg;
1625         }
1626         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1627 }
1628
1629 boolean_t
1630 dsl_scan_resilvering(dsl_pool_t *dp)
1631 {
1632         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1633             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1634 }
1635
1636 /*
1637  * scrub consumers
1638  */
1639
1640 static void
1641 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1642 {
1643         int i;
1644
1645         /*
1646          * If we resume after a reboot, zab will be NULL; don't record
1647          * incomplete stats in that case.
1648          */
1649         if (zab == NULL)
1650                 return;
1651
1652         for (i = 0; i < 4; i++) {
1653                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1654                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1655                 if (t & DMU_OT_NEWTYPE)
1656                         t = DMU_OT_OTHER;
1657                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1658                 int equal;
1659
1660                 zb->zb_count++;
1661                 zb->zb_asize += BP_GET_ASIZE(bp);
1662                 zb->zb_lsize += BP_GET_LSIZE(bp);
1663                 zb->zb_psize += BP_GET_PSIZE(bp);
1664                 zb->zb_gangs += BP_COUNT_GANG(bp);
1665
1666                 switch (BP_GET_NDVAS(bp)) {
1667                 case 2:
1668                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1669                             DVA_GET_VDEV(&bp->blk_dva[1]))
1670                                 zb->zb_ditto_2_of_2_samevdev++;
1671                         break;
1672                 case 3:
1673                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1674                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1675                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1676                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1677                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1678                             DVA_GET_VDEV(&bp->blk_dva[2]));
1679                         if (equal == 1)
1680                                 zb->zb_ditto_2_of_3_samevdev++;
1681                         else if (equal == 3)
1682                                 zb->zb_ditto_3_of_3_samevdev++;
1683                         break;
1684                 }
1685         }
1686 }
1687
1688 static void
1689 dsl_scan_scrub_done(zio_t *zio)
1690 {
1691         spa_t *spa = zio->io_spa;
1692
1693         zio_data_buf_free(zio->io_data, zio->io_size);
1694
1695         mutex_enter(&spa->spa_scrub_lock);
1696         spa->spa_scrub_inflight--;
1697         cv_broadcast(&spa->spa_scrub_io_cv);
1698
1699         if (zio->io_error && (zio->io_error != ECKSUM ||
1700             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1701                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1702         }
1703         mutex_exit(&spa->spa_scrub_lock);
1704 }
1705
1706 static int
1707 dsl_scan_scrub_cb(dsl_pool_t *dp,
1708     const blkptr_t *bp, const zbookmark_phys_t *zb)
1709 {
1710         dsl_scan_t *scn = dp->dp_scan;
1711         size_t size = BP_GET_PSIZE(bp);
1712         spa_t *spa = dp->dp_spa;
1713         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1714         boolean_t needs_io;
1715         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1716         unsigned int scan_delay = 0;
1717
1718         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1719             phys_birth >= scn->scn_phys.scn_max_txg)
1720                 return (0);
1721
1722         count_block(dp->dp_blkstats, bp);
1723
1724         if (BP_IS_EMBEDDED(bp))
1725                 return (0);
1726
1727         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1728         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1729                 zio_flags |= ZIO_FLAG_SCRUB;
1730                 needs_io = B_TRUE;
1731                 scan_delay = zfs_scrub_delay;
1732         } else {
1733                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1734                 zio_flags |= ZIO_FLAG_RESILVER;
1735                 needs_io = B_FALSE;
1736                 scan_delay = zfs_resilver_delay;
1737         }
1738
1739         /* If it's an intent log block, failure is expected. */
1740         if (zb->zb_level == ZB_ZIL_LEVEL)
1741                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1742
1743         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1744                 vdev_t *vd = vdev_lookup_top(spa,
1745                     DVA_GET_VDEV(&bp->blk_dva[d]));
1746
1747                 /*
1748                  * Keep track of how much data we've examined so that
1749                  * zpool(1M) status can make useful progress reports.
1750                  */
1751                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1752                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1753
1754                 /* if it's a resilver, this may not be in the target range */
1755                 if (!needs_io) {
1756                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1757                                 /*
1758                                  * Gang members may be spread across multiple
1759                                  * vdevs, so the best estimate we have is the
1760                                  * scrub range, which has already been checked.
1761                                  * XXX -- it would be better to change our
1762                                  * allocation policy to ensure that all
1763                                  * gang members reside on the same vdev.
1764                                  */
1765                                 needs_io = B_TRUE;
1766                         } else {
1767                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1768                                     phys_birth, 1);
1769                         }
1770                 }
1771         }
1772
1773         if (needs_io && !zfs_no_scrub_io) {
1774                 vdev_t *rvd = spa->spa_root_vdev;
1775                 uint64_t maxinflight = rvd->vdev_children *
1776                     MAX(zfs_top_maxinflight, 1);
1777                 void *data = zio_data_buf_alloc(size);
1778
1779                 mutex_enter(&spa->spa_scrub_lock);
1780                 while (spa->spa_scrub_inflight >= maxinflight)
1781                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1782                 spa->spa_scrub_inflight++;
1783                 mutex_exit(&spa->spa_scrub_lock);
1784
1785                 /*
1786                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1787                  * then throttle our workload to limit the impact of a scan.
1788                  */
1789                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1790                         delay(MAX((int)scan_delay, 0));
1791
1792                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1793                     dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1794                     zio_flags, zb));
1795         }
1796
1797         /* do not relocate this block */
1798         return (0);
1799 }
1800
1801 int
1802 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1803 {
1804         spa_t *spa = dp->dp_spa;
1805
1806         /*
1807          * Purge all vdev caches and probe all devices.  We do this here
1808          * rather than in sync context because this requires a writer lock
1809          * on the spa_config lock, which we can't do from sync context.  The
1810          * spa_scrub_reopen flag indicates that vdev_open() should not
1811          * attempt to start another scrub.
1812          */
1813         spa_vdev_state_enter(spa, SCL_NONE);
1814         spa->spa_scrub_reopen = B_TRUE;
1815         vdev_reopen(spa->spa_root_vdev);
1816         spa->spa_scrub_reopen = B_FALSE;
1817         (void) spa_vdev_state_exit(spa, NULL, 0);
1818
1819         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1820             dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1821 }