module/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  */
  24
  25 #include <sys/dsl_scan.h>
  26 #include <sys/dsl_pool.h>
  27 #include <sys/dsl_dataset.h>
  28 #include <sys/dsl_prop.h>
  29 #include <sys/dsl_dir.h>
  30 #include <sys/dsl_synctask.h>
  31 #include <sys/dnode.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/dmu_objset.h>
  34 #include <sys/arc.h>
  35 #include <sys/zap.h>
  36 #include <sys/zio.h>
  37 #include <sys/zfs_context.h>
  38 #include <sys/fs/zfs.h>
  39 #include <sys/zfs_znode.h>
  40 #include <sys/spa_impl.h>
  41 #include <sys/vdev_impl.h>
  42 #include <sys/zil_impl.h>
  43 #include <sys/zio_checksum.h>
  44 #include <sys/ddt.h>
  45 #include <sys/sa.h>
  46 #include <sys/sa_impl.h>
  47 #ifdef _KERNEL
  48 #include <sys/zfs_vfsops.h>
  49 #endif
  50
  51 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
  52
  53 static scan_cb_t dsl_scan_defrag_cb;
  54 static scan_cb_t dsl_scan_scrub_cb;
  55 static scan_cb_t dsl_scan_remove_cb;
  56 static dsl_syncfunc_t dsl_scan_cancel_sync;
  57 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
  58
  59 int zfs_top_maxinflight = 32;           /* maximum I/Os per top-level */
  60 int zfs_resilver_delay = 2;             /* number of ticks to delay resilver */
  61 int zfs_scrub_delay = 4;                /* number of ticks to delay scrub */
  62 int zfs_scan_idle = 50;                 /* idle window in clock ticks */
  63
  64 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
  65 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
  66 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
  67 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  68 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable srub prefetching */
  69 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
  70 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
  71
  72 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
  73         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
  74         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
  75
  76 extern int zfs_txg_timeout;
  77
  78 /* the order has to match pool_scan_type */
  79 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
  80         NULL,
  81         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
  82         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
  83 };
  84
  85 int
  86 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
  87 {
  88         int err;
  89         dsl_scan_t *scn;
  90         spa_t *spa = dp->dp_spa;
  91         uint64_t f;
  92
  93         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
  94         scn->scn_dp = dp;
  95
  96         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
  97             "scrub_func", sizeof (uint64_t), 1, &f);
  98         if (err == 0) {
  99                 /*
 100                  * There was an old-style scrub in progress.  Restart a
 101                  * new-style scrub from the beginning.
 102                  */
 103                 scn->scn_restart_txg = txg;
 104                 zfs_dbgmsg("old-style scrub was in progress; "
 105                     "restarting new-style scrub in txg %llu",
 106                     scn->scn_restart_txg);
 107
 108                 /*
 109                  * Load the queue obj from the old location so that it
 110                  * can be freed by dsl_scan_done().
 111                  */
 112                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 113                     "scrub_queue", sizeof (uint64_t), 1,
 114                     &scn->scn_phys.scn_queue_obj);
 115         } else {
 116                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 117                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 118                     &scn->scn_phys);
 119                 if (err == ENOENT)
 120                         return (0);
 121                 else if (err)
 122                         return (err);
 123
 124                 if (scn->scn_phys.scn_state == DSS_SCANNING &&
 125                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 126                         /*
 127                          * A new-type scrub was in progress on an old
 128                          * pool, and the pool was accessed by old
 129                          * software.  Restart from the beginning, since
 130                          * the old software may have changed the pool in
 131                          * the meantime.
 132                          */
 133                         scn->scn_restart_txg = txg;
 134                         zfs_dbgmsg("new-style scrub was modified "
 135                             "by old software; restarting in txg %llu",
 136                             scn->scn_restart_txg);
 137                 }
 138         }
 139
 140         spa_scan_stat_init(spa);
 141         return (0);
 142 }
 143
 144 void
 145 dsl_scan_fini(dsl_pool_t *dp)
 146 {
 147         if (dp->dp_scan) {
 148                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 149                 dp->dp_scan = NULL;
 150         }
 151 }
 152
 153 /* ARGSUSED */
 154 static int
 155 dsl_scan_setup_check(void *arg1, void *arg2, dmu_tx_t *tx)
 156 {
 157         dsl_scan_t *scn = arg1;
 158
 159         if (scn->scn_phys.scn_state == DSS_SCANNING)
 160                 return (EBUSY);
 161
 162         return (0);
 163 }
 164
 165 /* ARGSUSED */
 166 static void
 167 dsl_scan_setup_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 168 {
 169         dsl_scan_t *scn = arg1;
 170         pool_scan_func_t *funcp = arg2;
 171         dmu_object_type_t ot = 0;
 172         dsl_pool_t *dp = scn->scn_dp;
 173         spa_t *spa = dp->dp_spa;
 174
 175         ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
 176         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 177         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 178         scn->scn_phys.scn_func = *funcp;
 179         scn->scn_phys.scn_state = DSS_SCANNING;
 180         scn->scn_phys.scn_min_txg = 0;
 181         scn->scn_phys.scn_max_txg = tx->tx_txg;
 182         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 183         scn->scn_phys.scn_start_time = gethrestime_sec();
 184         scn->scn_phys.scn_errors = 0;
 185         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 186         scn->scn_restart_txg = 0;
 187         spa_scan_stat_init(spa);
 188
 189         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 190                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 191
 192                 /* rewrite all disk labels */
 193                 vdev_config_dirty(spa->spa_root_vdev);
 194
 195                 if (vdev_resilver_needed(spa->spa_root_vdev,
 196                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 197                         spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
 198                 } else {
 199                         spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
 200                 }
 201
 202                 spa->spa_scrub_started = B_TRUE;
 203                 /*
 204                  * If this is an incremental scrub, limit the DDT scrub phase
 205                  * to just the auto-ditto class (for correctness); the rest
 206                  * of the scrub should go faster using top-down pruning.
 207                  */
 208                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 209                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 210
 211         }
 212
 213         /* back to the generic stuff */
 214
 215         if (dp->dp_blkstats == NULL) {
 216                 dp->dp_blkstats =
 217                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 218         }
 219         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 220
 221         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 222                 ot = DMU_OT_ZAP_OTHER;
 223
 224         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 225             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 226
 227         dsl_scan_sync_state(scn, tx);
 228
 229         spa_history_log_internal(LOG_POOL_SCAN, spa, tx,
 230             "func=%u mintxg=%llu maxtxg=%llu",
 231             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 232 }
 233
 234 /* ARGSUSED */
 235 static void
 236 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 237 {
 238         static const char *old_names[] = {
 239                 "scrub_bookmark",
 240                 "scrub_ddt_bookmark",
 241                 "scrub_ddt_class_max",
 242                 "scrub_queue",
 243                 "scrub_min_txg",
 244                 "scrub_max_txg",
 245                 "scrub_func",
 246                 "scrub_errors",
 247                 NULL
 248         };
 249
 250         dsl_pool_t *dp = scn->scn_dp;
 251         spa_t *spa = dp->dp_spa;
 252         int i;
 253
 254         /* Remove any remnants of an old-style scrub. */
 255         for (i = 0; old_names[i]; i++) {
 256                 (void) zap_remove(dp->dp_meta_objset,
 257                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 258         }
 259
 260         if (scn->scn_phys.scn_queue_obj != 0) {
 261                 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 262                     scn->scn_phys.scn_queue_obj, tx));
 263                 scn->scn_phys.scn_queue_obj = 0;
 264         }
 265
 266         /*
 267          * If we were "restarted" from a stopped state, don't bother
 268          * with anything else.
 269          */
 270         if (scn->scn_phys.scn_state != DSS_SCANNING)
 271                 return;
 272
 273         if (complete)
 274                 scn->scn_phys.scn_state = DSS_FINISHED;
 275         else
 276                 scn->scn_phys.scn_state = DSS_CANCELED;
 277
 278         spa_history_log_internal(LOG_POOL_SCAN_DONE, spa, tx,
 279             "complete=%u", complete);
 280
 281         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 282                 mutex_enter(&spa->spa_scrub_lock);
 283                 while (spa->spa_scrub_inflight > 0) {
 284                         cv_wait(&spa->spa_scrub_io_cv,
 285                             &spa->spa_scrub_lock);
 286                 }
 287                 mutex_exit(&spa->spa_scrub_lock);
 288                 spa->spa_scrub_started = B_FALSE;
 289                 spa->spa_scrub_active = B_FALSE;
 290
 291                 /*
 292                  * If the scrub/resilver completed, update all DTLs to
 293                  * reflect this.  Whether it succeeded or not, vacate
 294                  * all temporary scrub DTLs.
 295                  */
 296                 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 297                     complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
 298                 if (complete) {
 299                         spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
 300                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 301                 }
 302                 spa_errlog_rotate(spa);
 303
 304                 /*
 305                  * We may have finished replacing a device.
 306                  * Let the async thread assess this and handle the detach.
 307                  */
 308                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 309         }
 310
 311         scn->scn_phys.scn_end_time = gethrestime_sec();
 312 }
 313
 314 /* ARGSUSED */
 315 static int
 316 dsl_scan_cancel_check(void *arg1, void *arg2, dmu_tx_t *tx)
 317 {
 318         dsl_scan_t *scn = arg1;
 319
 320         if (scn->scn_phys.scn_state != DSS_SCANNING)
 321                 return (ENOENT);
 322         return (0);
 323 }
 324
 325 /* ARGSUSED */
 326 static void
 327 dsl_scan_cancel_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 328 {
 329         dsl_scan_t *scn = arg1;
 330
 331         dsl_scan_done(scn, B_FALSE, tx);
 332         dsl_scan_sync_state(scn, tx);
 333 }
 334
 335 int
 336 dsl_scan_cancel(dsl_pool_t *dp)
 337 {
 338         boolean_t complete = B_FALSE;
 339         int err;
 340
 341         err = dsl_sync_task_do(dp, dsl_scan_cancel_check,
 342             dsl_scan_cancel_sync, dp->dp_scan, &complete, 3);
 343         return (err);
 344 }
 345
 346 static void dsl_scan_visitbp(blkptr_t *bp,
 347     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
 348     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 349     dmu_tx_t *tx);
 350 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
 351     dmu_objset_type_t ostype,
 352     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
 353
 354 void
 355 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 356 {
 357         zio_free(dp->dp_spa, txg, bp);
 358 }
 359
 360 void
 361 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 362 {
 363         ASSERT(dsl_pool_sync_context(dp));
 364         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
 365 }
 366
 367 int
 368 dsl_read(zio_t *pio, spa_t *spa, const blkptr_t *bpp, arc_buf_t *pbuf,
 369     arc_done_func_t *done, void *private, int priority, int zio_flags,
 370     uint32_t *arc_flags, const zbookmark_t *zb)
 371 {
 372         return (arc_read(pio, spa, bpp, pbuf, done, private,
 373             priority, zio_flags, arc_flags, zb));
 374 }
 375
 376 int
 377 dsl_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bpp,
 378     arc_done_func_t *done, void *private, int priority, int zio_flags,
 379     uint32_t *arc_flags, const zbookmark_t *zb)
 380 {
 381         return (arc_read_nolock(pio, spa, bpp, done, private,
 382             priority, zio_flags, arc_flags, zb));
 383 }
 384
 385 static boolean_t
 386 bookmark_is_zero(const zbookmark_t *zb)
 387 {
 388         return (zb->zb_objset == 0 && zb->zb_object == 0 &&
 389             zb->zb_level == 0 && zb->zb_blkid == 0);
 390 }
 391
 392 /* dnp is the dnode for zb1->zb_object */
 393 static boolean_t
 394 bookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
 395     const zbookmark_t *zb2)
 396 {
 397         uint64_t zb1nextL0, zb2thisobj;
 398
 399         ASSERT(zb1->zb_objset == zb2->zb_objset);
 400         ASSERT(zb2->zb_level == 0);
 401
 402         /*
 403          * A bookmark in the deadlist is considered to be after
 404          * everything else.
 405          */
 406         if (zb2->zb_object == DMU_DEADLIST_OBJECT)
 407                 return (B_TRUE);
 408
 409         /* The objset_phys_t isn't before anything. */
 410         if (dnp == NULL)
 411                 return (B_FALSE);
 412
 413         zb1nextL0 = (zb1->zb_blkid + 1) <<
 414             ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
 415
 416         zb2thisobj = zb2->zb_object ? zb2->zb_object :
 417             zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
 418
 419         if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
 420                 uint64_t nextobj = zb1nextL0 *
 421                     (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
 422                 return (nextobj <= zb2thisobj);
 423         }
 424
 425         if (zb1->zb_object < zb2thisobj)
 426                 return (B_TRUE);
 427         if (zb1->zb_object > zb2thisobj)
 428                 return (B_FALSE);
 429         if (zb2->zb_object == DMU_META_DNODE_OBJECT)
 430                 return (B_FALSE);
 431         return (zb1nextL0 <= zb2->zb_blkid);
 432 }
 433
 434 static uint64_t
 435 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
 436 {
 437         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
 438         if (dsl_dataset_is_snapshot(ds))
 439                 return (MIN(smt, ds->ds_phys->ds_creation_txg));
 440         return (smt);
 441 }
 442
 443 static void
 444 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
 445 {
 446         VERIFY(0 == zap_update(scn->scn_dp->dp_meta_objset,
 447             DMU_POOL_DIRECTORY_OBJECT,
 448             DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 449             &scn->scn_phys, tx));
 450 }
 451
 452 static boolean_t
 453 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
 454 {
 455         uint64_t elapsed_nanosecs;
 456         int mintime;
 457
 458         /* we never skip user/group accounting objects */
 459         if (zb && (int64_t)zb->zb_object < 0)
 460                 return (B_FALSE);
 461
 462         if (scn->scn_pausing)
 463                 return (B_TRUE); /* we're already pausing */
 464
 465         if (!bookmark_is_zero(&scn->scn_phys.scn_bookmark))
 466                 return (B_FALSE); /* we're resuming */
 467
 468         /* We only know how to resume from level-0 blocks. */
 469         if (zb && zb->zb_level != 0)
 470                 return (B_FALSE);
 471
 472         mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
 473             zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
 474         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 475         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 476             (elapsed_nanosecs / MICROSEC > mintime &&
 477             txg_sync_waiting(scn->scn_dp)) ||
 478             spa_shutting_down(scn->scn_dp->dp_spa)) {
 479                 if (zb) {
 480                         dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
 481                             (longlong_t)zb->zb_objset,
 482                             (longlong_t)zb->zb_object,
 483                             (longlong_t)zb->zb_level,
 484                             (longlong_t)zb->zb_blkid);
 485                         scn->scn_phys.scn_bookmark = *zb;
 486                 }
 487                 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
 488                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
 489                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
 490                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
 491                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
 492                 scn->scn_pausing = B_TRUE;
 493                 return (B_TRUE);
 494         }
 495         return (B_FALSE);
 496 }
 497
 498 typedef struct zil_scan_arg {
 499         dsl_pool_t      *zsa_dp;
 500         zil_header_t    *zsa_zh;
 501 } zil_scan_arg_t;
 502
 503 /* ARGSUSED */
 504 static int
 505 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 506 {
 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         zbookmark_t zb;
 512
 513         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 514                 return (0);
 515
 516         /*
 517          * One block ("stubby") can be allocated a long time ago; we
 518          * want to visit that one because it has been allocated
 519          * (on-disk) even if it hasn't been claimed (even though for
 520          * scrub there's nothing to do to it).
 521          */
 522         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 523                 return (0);
 524
 525         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 526             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
 527
 528         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 529         return (0);
 530 }
 531
 532 /* ARGSUSED */
 533 static int
 534 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 535 {
 536         if (lrc->lrc_txtype == TX_WRITE) {
 537                 zil_scan_arg_t *zsa = arg;
 538                 dsl_pool_t *dp = zsa->zsa_dp;
 539                 dsl_scan_t *scn = dp->dp_scan;
 540                 zil_header_t *zh = zsa->zsa_zh;
 541                 lr_write_t *lr = (lr_write_t *)lrc;
 542                 blkptr_t *bp = &lr->lr_blkptr;
 543                 zbookmark_t zb;
 544
 545                 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 546                         return (0);
 547
 548                 /*
 549                  * birth can be < claim_txg if this record's txg is
 550                  * already txg sync'ed (but this log block contains
 551                  * other records that are not synced)
 552                  */
 553                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 554                         return (0);
 555
 556                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 557                     lr->lr_foid, ZB_ZIL_LEVEL,
 558                     lr->lr_offset / BP_GET_LSIZE(bp));
 559
 560                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
 561         }
 562         return (0);
 563 }
 564
 565 static void
 566 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
 567 {
 568         uint64_t claim_txg = zh->zh_claim_txg;
 569         zil_scan_arg_t zsa = { dp, zh };
 570         zilog_t *zilog;
 571
 572         /*
 573          * We only want to visit blocks that have been claimed but not yet
 574          * replayed (or, in read-only mode, blocks that *would* be claimed).
 575          */
 576         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 577                 return;
 578
 579         zilog = zil_alloc(dp->dp_meta_objset, zh);
 580
 581         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
 582             claim_txg);
 583
 584         zil_free(zilog);
 585 }
 586
 587 /* ARGSUSED */
 588 static void
 589 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
 590     uint64_t objset, uint64_t object, uint64_t blkid)
 591 {
 592         zbookmark_t czb;
 593         uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
 594
 595         if (zfs_no_scrub_prefetch)
 596                 return;
 597
 598         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
 599             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
 600                 return;
 601
 602         SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
 603
 604         /*
 605          * XXX need to make sure all of these arc_read() prefetches are
 606          * done before setting xlateall (similar to dsl_read())
 607          */
 608         (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
 609             buf, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
 610             ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
 611 }
 612
 613 static boolean_t
 614 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 615     const zbookmark_t *zb)
 616 {
 617         /*
 618          * We never skip over user/group accounting objects (obj<0)
 619          */
 620         if (!bookmark_is_zero(&scn->scn_phys.scn_bookmark) &&
 621             (int64_t)zb->zb_object >= 0) {
 622                 /*
 623                  * If we already visited this bp & everything below (in
 624                  * a prior txg sync), don't bother doing it again.
 625                  */
 626                 if (bookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
 627                         return (B_TRUE);
 628
 629                 /*
 630                  * If we found the block we're trying to resume from, or
 631                  * we went past it to a different object, zero it out to
 632                  * indicate that it's OK to start checking for pausing
 633                  * again.
 634                  */
 635                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
 636                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
 637                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 638                             (longlong_t)zb->zb_objset,
 639                             (longlong_t)zb->zb_object,
 640                             (longlong_t)zb->zb_level,
 641                             (longlong_t)zb->zb_blkid);
 642                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
 643                 }
 644         }
 645         return (B_FALSE);
 646 }
 647
 648 /*
 649  * Return nonzero on i/o error.
 650  * Return new buf to write out in *bufp.
 651  */
 652 static int
 653 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
 654     dnode_phys_t *dnp, const blkptr_t *bp,
 655     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
 656 {
 657         dsl_pool_t *dp = scn->scn_dp;
 658         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
 659         int err;
 660
 661         if (BP_GET_LEVEL(bp) > 0) {
 662                 uint32_t flags = ARC_WAIT;
 663                 int i;
 664                 blkptr_t *cbp;
 665                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 666
 667                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 668                     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, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 675                         dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
 676                             zb->zb_object, zb->zb_blkid * epb + i);
 677                 }
 678                 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
 679                         zbookmark_t czb;
 680
 681                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 682                             zb->zb_level - 1,
 683                             zb->zb_blkid * epb + i);
 684                         dsl_scan_visitbp(cbp, &czb, dnp,
 685                             *bufp, ds, scn, ostype, tx);
 686                 }
 687         } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
 688                 uint32_t flags = ARC_WAIT;
 689
 690                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 691                     arc_getbuf_func, bufp,
 692                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 693                 if (err) {
 694                         scn->scn_phys.scn_errors++;
 695                         return (err);
 696                 }
 697         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 698                 uint32_t flags = ARC_WAIT;
 699                 dnode_phys_t *cdnp;
 700                 int i, j;
 701                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 702
 703                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 704                     arc_getbuf_func, bufp,
 705                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 706                 if (err) {
 707                         scn->scn_phys.scn_errors++;
 708                         return (err);
 709                 }
 710                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 711                         for (j = 0; j < cdnp->dn_nblkptr; j++) {
 712                                 blkptr_t *cbp = &cdnp->dn_blkptr[j];
 713                                 dsl_scan_prefetch(scn, *bufp, cbp,
 714                                     zb->zb_objset, zb->zb_blkid * epb + i, j);
 715                         }
 716                 }
 717                 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
 718                         dsl_scan_visitdnode(scn, ds, ostype,
 719                             cdnp, *bufp, zb->zb_blkid * epb + i, tx);
 720                 }
 721
 722         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 723                 uint32_t flags = ARC_WAIT;
 724                 objset_phys_t *osp;
 725
 726                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 727                     arc_getbuf_func, bufp,
 728                     ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
 729                 if (err) {
 730                         scn->scn_phys.scn_errors++;
 731                         return (err);
 732                 }
 733
 734                 osp = (*bufp)->b_data;
 735
 736                 dsl_scan_visitdnode(scn, ds, osp->os_type,
 737                     &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
 738
 739                 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
 740                         /*
 741                          * We also always visit user/group accounting
 742                          * objects, and never skip them, even if we are
 743                          * pausing.  This is necessary so that the space
 744                          * deltas from this txg get integrated.
 745                          */
 746                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 747                             &osp->os_groupused_dnode, *bufp,
 748                             DMU_GROUPUSED_OBJECT, tx);
 749                         dsl_scan_visitdnode(scn, ds, osp->os_type,
 750                             &osp->os_userused_dnode, *bufp,
 751                             DMU_USERUSED_OBJECT, tx);
 752                 }
 753         }
 754
 755         return (0);
 756 }
 757
 758 static void
 759 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
 760     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
 761     uint64_t object, dmu_tx_t *tx)
 762 {
 763         int j;
 764
 765         for (j = 0; j < dnp->dn_nblkptr; j++) {
 766                 zbookmark_t czb;
 767
 768                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 769                     dnp->dn_nlevels - 1, j);
 770                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
 771                     &czb, dnp, buf, ds, scn, ostype, tx);
 772         }
 773
 774         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
 775                 zbookmark_t czb;
 776                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
 777                     0, DMU_SPILL_BLKID);
 778                 dsl_scan_visitbp(&dnp->dn_spill,
 779                     &czb, dnp, buf, ds, scn, ostype, tx);
 780         }
 781 }
 782
 783 /*
 784  * The arguments are in this order because mdb can only print the
 785  * first 5; we want them to be useful.
 786  */
 787 static void
 788 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
 789     dnode_phys_t *dnp, arc_buf_t *pbuf,
 790     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
 791     dmu_tx_t *tx)
 792 {
 793         dsl_pool_t *dp = scn->scn_dp;
 794         arc_buf_t *buf = NULL;
 795         blkptr_t *bp_toread;
 796
 797         bp_toread = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
 798         *bp_toread = *bp;
 799
 800         /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
 801
 802         if (dsl_scan_check_pause(scn, zb))
 803                 goto out;
 804
 805         if (dsl_scan_check_resume(scn, dnp, zb))
 806                 goto out;
 807
 808         if (bp->blk_birth == 0)
 809                 goto out;
 810
 811         scn->scn_visited_this_txg++;
 812
 813         dprintf_bp(bp,
 814             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
 815             ds, ds ? ds->ds_object : 0,
 816             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
 817             pbuf, bp);
 818
 819         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
 820                 goto out;
 821
 822         if (BP_GET_TYPE(bp) != DMU_OT_USERGROUP_USED) {
 823                 /*
 824                  * For non-user-accounting blocks, we need to read the
 825                  * new bp (from a deleted snapshot, found in
 826                  * check_existing_xlation).  If we used the old bp,
 827                  * pointers inside this block from before we resumed
 828                  * would be untranslated.
 829                  *
 830                  * For user-accounting blocks, we need to read the old
 831                  * bp, because we will apply the entire space delta to
 832                  * it (original untranslated -> translations from
 833                  * deleted snap -> now).
 834                  */
 835                 *bp_toread = *bp;
 836         }
 837
 838         if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx,
 839             &buf) != 0)
 840                 goto out;
 841
 842         /*
 843          * If dsl_scan_ddt() has aready visited this block, it will have
 844          * already done any translations or scrubbing, so don't call the
 845          * callback again.
 846          */
 847         if (ddt_class_contains(dp->dp_spa,
 848             scn->scn_phys.scn_ddt_class_max, bp)) {
 849                 ASSERT(buf == NULL);
 850                 goto out;
 851         }
 852
 853         /*
 854          * If this block is from the future (after cur_max_txg), then we
 855          * are doing this on behalf of a deleted snapshot, and we will
 856          * revisit the future block on the next pass of this dataset.
 857          * Don't scan it now unless we need to because something
 858          * under it was modified.
 859          */
 860         if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
 861                 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
 862         }
 863         if (buf)
 864                 (void) arc_buf_remove_ref(buf, &buf);
 865 out:
 866         kmem_free(bp_toread, sizeof(blkptr_t));
 867 }
 868
 869 static void
 870 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
 871     dmu_tx_t *tx)
 872 {
 873         zbookmark_t zb;
 874
 875         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 876             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 877         dsl_scan_visitbp(bp, &zb, NULL, NULL,
 878             ds, scn, DMU_OST_NONE, tx);
 879
 880         dprintf_ds(ds, "finished scan%s", "");
 881 }
 882
 883 void
 884 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 885 {
 886         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 887         dsl_scan_t *scn = dp->dp_scan;
 888         uint64_t mintxg;
 889
 890         if (scn->scn_phys.scn_state != DSS_SCANNING)
 891                 return;
 892
 893         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 894                 if (dsl_dataset_is_snapshot(ds)) {
 895                         /* Note, scn_cur_{min,max}_txg stays the same. */
 896                         scn->scn_phys.scn_bookmark.zb_objset =
 897                             ds->ds_phys->ds_next_snap_obj;
 898                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 899                             "reset zb_objset to %llu",
 900                             (u_longlong_t)ds->ds_object,
 901                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 902                         scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
 903                 } else {
 904                         SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
 905                             ZB_DESTROYED_OBJSET, 0, 0, 0);
 906                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
 907                             "reset bookmark to -1,0,0,0",
 908                             (u_longlong_t)ds->ds_object);
 909                 }
 910         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 911             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 912                 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 913                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 914                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 915                 if (dsl_dataset_is_snapshot(ds)) {
 916                         /*
 917                          * We keep the same mintxg; it could be >
 918                          * ds_creation_txg if the previous snapshot was
 919                          * deleted too.
 920                          */
 921                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
 922                             scn->scn_phys.scn_queue_obj,
 923                             ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
 924                         zfs_dbgmsg("destroying ds %llu; in queue; "
 925                             "replacing with %llu",
 926                             (u_longlong_t)ds->ds_object,
 927                             (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
 928                 } else {
 929                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
 930                             (u_longlong_t)ds->ds_object);
 931                 }
 932         } else {
 933                 zfs_dbgmsg("destroying ds %llu; ignoring",
 934                     (u_longlong_t)ds->ds_object);
 935         }
 936
 937         /*
 938          * dsl_scan_sync() should be called after this, and should sync
 939          * out our changed state, but just to be safe, do it here.
 940          */
 941         dsl_scan_sync_state(scn, tx);
 942 }
 943
 944 void
 945 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 946 {
 947         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 948         dsl_scan_t *scn = dp->dp_scan;
 949         uint64_t mintxg;
 950
 951         if (scn->scn_phys.scn_state != DSS_SCANNING)
 952                 return;
 953
 954         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 955
 956         if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
 957                 scn->scn_phys.scn_bookmark.zb_objset =
 958                     ds->ds_phys->ds_prev_snap_obj;
 959                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
 960                     "reset zb_objset to %llu",
 961                     (u_longlong_t)ds->ds_object,
 962                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 963         } else if (zap_lookup_int_key(dp->dp_meta_objset,
 964             scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
 965                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
 966                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
 967                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
 968                     scn->scn_phys.scn_queue_obj,
 969                     ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
 970                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
 971                     "replacing with %llu",
 972                     (u_longlong_t)ds->ds_object,
 973                     (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
 974         }
 975         dsl_scan_sync_state(scn, tx);
 976 }
 977
 978 void
 979 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 980 {
 981         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 982         dsl_scan_t *scn = dp->dp_scan;
 983         uint64_t mintxg;
 984
 985         if (scn->scn_phys.scn_state != DSS_SCANNING)
 986                 return;
 987
 988         if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
 989                 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
 990                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 991                     "reset zb_objset to %llu",
 992                     (u_longlong_t)ds1->ds_object,
 993                     (u_longlong_t)ds2->ds_object);
 994         } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
 995                 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
 996                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
 997                     "reset zb_objset to %llu",
 998                     (u_longlong_t)ds2->ds_object,
 999                     (u_longlong_t)ds1->ds_object);
1000         }
1001
1002         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1003             ds1->ds_object, &mintxg) == 0) {
1004                 int err;
1005
1006                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
1007                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
1008                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1009                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
1010                 err = zap_add_int_key(dp->dp_meta_objset,
1011                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
1012                 VERIFY(err == 0 || err == EEXIST);
1013                 if (err == EEXIST) {
1014                         /* Both were there to begin with */
1015                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1016                             scn->scn_phys.scn_queue_obj,
1017                             ds1->ds_object, mintxg, tx));
1018                 }
1019                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1020                     "replacing with %llu",
1021                     (u_longlong_t)ds1->ds_object,
1022                     (u_longlong_t)ds2->ds_object);
1023         } else if (zap_lookup_int_key(dp->dp_meta_objset,
1024             scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
1025                 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
1026                 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
1027                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1028                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
1029                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1030                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
1031                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1032                     "replacing with %llu",
1033                     (u_longlong_t)ds2->ds_object,
1034                     (u_longlong_t)ds1->ds_object);
1035         }
1036
1037         dsl_scan_sync_state(scn, tx);
1038 }
1039
1040 struct enqueue_clones_arg {
1041         dmu_tx_t *tx;
1042         uint64_t originobj;
1043 };
1044
1045 /* ARGSUSED */
1046 static int
1047 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1048 {
1049         struct enqueue_clones_arg *eca = arg;
1050         dsl_dataset_t *ds;
1051         int err;
1052         dsl_pool_t *dp = spa->spa_dsl_pool;
1053         dsl_scan_t *scn = dp->dp_scan;
1054
1055         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1056         if (err)
1057                 return (err);
1058
1059         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
1060                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1061                         dsl_dataset_t *prev;
1062                         err = dsl_dataset_hold_obj(dp,
1063                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1064
1065                         dsl_dataset_rele(ds, FTAG);
1066                         if (err)
1067                                 return (err);
1068                         ds = prev;
1069                 }
1070                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1071                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1072                     ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1073         }
1074         dsl_dataset_rele(ds, FTAG);
1075         return (0);
1076 }
1077
1078 static void
1079 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1080 {
1081         dsl_pool_t *dp = scn->scn_dp;
1082         dsl_dataset_t *ds;
1083         objset_t *os;
1084         char *dsname;
1085
1086         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1087
1088         if (dmu_objset_from_ds(ds, &os))
1089                 goto out;
1090
1091         /*
1092          * Only the ZIL in the head (non-snapshot) is valid.  Even though
1093          * snapshots can have ZIL block pointers (which may be the same
1094          * BP as in the head), they must be ignored.  So we traverse the
1095          * ZIL here, rather than in scan_recurse(), because the regular
1096          * snapshot block-sharing rules don't apply to it.
1097          */
1098         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1099                 dsl_scan_zil(dp, &os->os_zil_header);
1100
1101         /*
1102          * Iterate over the bps in this ds.
1103          */
1104         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1105         dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1106
1107         dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1108         dsl_dataset_name(ds, dsname);
1109         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1110             "pausing=%u",
1111             (longlong_t)dsobj, dsname,
1112             (longlong_t)scn->scn_phys.scn_cur_min_txg,
1113             (longlong_t)scn->scn_phys.scn_cur_max_txg,
1114             (int)scn->scn_pausing);
1115         kmem_free(dsname, ZFS_MAXNAMELEN);
1116
1117         if (scn->scn_pausing)
1118                 goto out;
1119
1120         /*
1121          * We've finished this pass over this dataset.
1122          */
1123
1124         /*
1125          * If we did not completely visit this dataset, do another pass.
1126          */
1127         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1128                 zfs_dbgmsg("incomplete pass; visiting again");
1129                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1130                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1131                     scn->scn_phys.scn_queue_obj, ds->ds_object,
1132                     scn->scn_phys.scn_cur_max_txg, tx) == 0);
1133                 goto out;
1134         }
1135
1136         /*
1137          * Add descendent datasets to work queue.
1138          */
1139         if (ds->ds_phys->ds_next_snap_obj != 0) {
1140                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1141                     scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1142                     ds->ds_phys->ds_creation_txg, tx) == 0);
1143         }
1144         if (ds->ds_phys->ds_num_children > 1) {
1145                 boolean_t usenext = B_FALSE;
1146                 if (ds->ds_phys->ds_next_clones_obj != 0) {
1147                         uint64_t count;
1148                         /*
1149                          * A bug in a previous version of the code could
1150                          * cause upgrade_clones_cb() to not set
1151                          * ds_next_snap_obj when it should, leading to a
1152                          * missing entry.  Therefore we can only use the
1153                          * next_clones_obj when its count is correct.
1154                          */
1155                         int err = zap_count(dp->dp_meta_objset,
1156                             ds->ds_phys->ds_next_clones_obj, &count);
1157                         if (err == 0 &&
1158                             count == ds->ds_phys->ds_num_children - 1)
1159                                 usenext = B_TRUE;
1160                 }
1161
1162                 if (usenext) {
1163                         VERIFY(zap_join_key(dp->dp_meta_objset,
1164                             ds->ds_phys->ds_next_clones_obj,
1165                             scn->scn_phys.scn_queue_obj,
1166                             ds->ds_phys->ds_creation_txg, tx) == 0);
1167                 } else {
1168                         struct enqueue_clones_arg eca;
1169                         eca.tx = tx;
1170                         eca.originobj = ds->ds_object;
1171
1172                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
1173                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
1174                 }
1175         }
1176
1177 out:
1178         dsl_dataset_rele(ds, FTAG);
1179 }
1180
1181 /* ARGSUSED */
1182 static int
1183 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1184 {
1185         dmu_tx_t *tx = arg;
1186         dsl_dataset_t *ds;
1187         int err;
1188         dsl_pool_t *dp = spa->spa_dsl_pool;
1189         dsl_scan_t *scn = dp->dp_scan;
1190
1191         err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1192         if (err)
1193                 return (err);
1194
1195         while (ds->ds_phys->ds_prev_snap_obj != 0) {
1196                 dsl_dataset_t *prev;
1197                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1198                     FTAG, &prev);
1199                 if (err) {
1200                         dsl_dataset_rele(ds, FTAG);
1201                         return (err);
1202                 }
1203
1204                 /*
1205                  * If this is a clone, we don't need to worry about it for now.
1206                  */
1207                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1208                         dsl_dataset_rele(ds, FTAG);
1209                         dsl_dataset_rele(prev, FTAG);
1210                         return (0);
1211                 }
1212                 dsl_dataset_rele(ds, FTAG);
1213                 ds = prev;
1214         }
1215
1216         VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1217             ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1218         dsl_dataset_rele(ds, FTAG);
1219         return (0);
1220 }
1221
1222 /*
1223  * Scrub/dedup interaction.
1224  *
1225  * If there are N references to a deduped block, we don't want to scrub it
1226  * N times -- ideally, we should scrub it exactly once.
1227  *
1228  * We leverage the fact that the dde's replication class (enum ddt_class)
1229  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1230  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1231  *
1232  * To prevent excess scrubbing, the scrub begins by walking the DDT
1233  * to find all blocks with refcnt > 1, and scrubs each of these once.
1234  * Since there are two replication classes which contain blocks with
1235  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1236  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1237  *
1238  * There would be nothing more to say if a block's refcnt couldn't change
1239  * during a scrub, but of course it can so we must account for changes
1240  * in a block's replication class.
1241  *
1242  * Here's an example of what can occur:
1243  *
1244  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1245  * when visited during the top-down scrub phase, it will be scrubbed twice.
1246  * This negates our scrub optimization, but is otherwise harmless.
1247  *
1248  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1249  * on each visit during the top-down scrub phase, it will never be scrubbed.
1250  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1251  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1252  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1253  * while a scrub is in progress, it scrubs the block right then.
1254  */
1255 static void
1256 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1257 {
1258         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1259         ddt_entry_t dde;
1260         int error;
1261         uint64_t n = 0;
1262
1263         bzero(&dde, sizeof (ddt_entry_t));
1264
1265         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1266                 ddt_t *ddt;
1267
1268                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1269                         break;
1270                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1271                     (longlong_t)ddb->ddb_class,
1272                     (longlong_t)ddb->ddb_type,
1273                     (longlong_t)ddb->ddb_checksum,
1274                     (longlong_t)ddb->ddb_cursor);
1275
1276                 /* There should be no pending changes to the dedup table */
1277                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1278                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1279
1280                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1281                 n++;
1282
1283                 if (dsl_scan_check_pause(scn, NULL))
1284                         break;
1285         }
1286
1287         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1288             (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1289             (int)scn->scn_pausing);
1290
1291         ASSERT(error == 0 || error == ENOENT);
1292         ASSERT(error != ENOENT ||
1293             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1294 }
1295
1296 /* ARGSUSED */
1297 void
1298 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1299     ddt_entry_t *dde, dmu_tx_t *tx)
1300 {
1301         const ddt_key_t *ddk = &dde->dde_key;
1302         ddt_phys_t *ddp = dde->dde_phys;
1303         blkptr_t bp;
1304         zbookmark_t zb = { 0 };
1305         int p;
1306
1307         if (scn->scn_phys.scn_state != DSS_SCANNING)
1308                 return;
1309
1310         for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1311                 if (ddp->ddp_phys_birth == 0 ||
1312                     ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1313                         continue;
1314                 ddt_bp_create(checksum, ddk, ddp, &bp);
1315
1316                 scn->scn_visited_this_txg++;
1317                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1318         }
1319 }
1320
1321 static void
1322 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1323 {
1324         dsl_pool_t *dp = scn->scn_dp;
1325         zap_cursor_t zc;
1326         zap_attribute_t za;
1327
1328         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1329             scn->scn_phys.scn_ddt_class_max) {
1330                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1331                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1332                 dsl_scan_ddt(scn, tx);
1333                 if (scn->scn_pausing)
1334                         return;
1335         }
1336
1337         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1338                 /* First do the MOS & ORIGIN */
1339
1340                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1341                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1342                 dsl_scan_visit_rootbp(scn, NULL,
1343                     &dp->dp_meta_rootbp, tx);
1344                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1345                 if (scn->scn_pausing)
1346                         return;
1347
1348                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1349                         VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
1350                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
1351                 } else {
1352                         dsl_scan_visitds(scn,
1353                             dp->dp_origin_snap->ds_object, tx);
1354                 }
1355                 ASSERT(!scn->scn_pausing);
1356         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1357             ZB_DESTROYED_OBJSET) {
1358                 /*
1359                  * If we were paused, continue from here.  Note if the
1360                  * ds we were paused on was deleted, the zb_objset may
1361                  * be -1, so we will skip this and find a new objset
1362                  * below.
1363                  */
1364                 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1365                 if (scn->scn_pausing)
1366                         return;
1367         }
1368
1369         /*
1370          * In case we were paused right at the end of the ds, zero the
1371          * bookmark so we don't think that we're still trying to resume.
1372          */
1373         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1374
1375         /* keep pulling things out of the zap-object-as-queue */
1376         while (zap_cursor_init(&zc, dp->dp_meta_objset,
1377             scn->scn_phys.scn_queue_obj),
1378             zap_cursor_retrieve(&zc, &za) == 0) {
1379                 dsl_dataset_t *ds;
1380                 uint64_t dsobj;
1381
1382                 dsobj = strtonum(za.za_name, NULL);
1383                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1384                     scn->scn_phys.scn_queue_obj, dsobj, tx));
1385
1386                 /* Set up min/max txg */
1387                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1388                 if (za.za_first_integer != 0) {
1389                         scn->scn_phys.scn_cur_min_txg =
1390                             MAX(scn->scn_phys.scn_min_txg,
1391                             za.za_first_integer);
1392                 } else {
1393                         scn->scn_phys.scn_cur_min_txg =
1394                             MAX(scn->scn_phys.scn_min_txg,
1395                             ds->ds_phys->ds_prev_snap_txg);
1396                 }
1397                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1398                 dsl_dataset_rele(ds, FTAG);
1399
1400                 dsl_scan_visitds(scn, dsobj, tx);
1401                 zap_cursor_fini(&zc);
1402                 if (scn->scn_pausing)
1403                         return;
1404         }
1405         zap_cursor_fini(&zc);
1406 }
1407
1408 static int
1409 dsl_scan_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1410 {
1411         dsl_scan_t *scn = arg;
1412         uint64_t elapsed_nanosecs;
1413
1414         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1415
1416         if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1417             (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1418             txg_sync_waiting(scn->scn_dp)) ||
1419             spa_shutting_down(scn->scn_dp->dp_spa))
1420                 return (ERESTART);
1421
1422         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1423             dmu_tx_get_txg(tx), bp, 0));
1424         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1425             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1426             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1427         scn->scn_visited_this_txg++;
1428         return (0);
1429 }
1430
1431 boolean_t
1432 dsl_scan_active(dsl_scan_t *scn)
1433 {
1434         spa_t *spa = scn->scn_dp->dp_spa;
1435         uint64_t used = 0, comp, uncomp;
1436
1437         if (spa->spa_load_state != SPA_LOAD_NONE)
1438                 return (B_FALSE);
1439         if (spa_shutting_down(spa))
1440                 return (B_FALSE);
1441
1442         if (scn->scn_phys.scn_state == DSS_SCANNING)
1443                 return (B_TRUE);
1444
1445         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1446                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1447                     &used, &comp, &uncomp);
1448         }
1449         return (used != 0);
1450 }
1451
1452 void
1453 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1454 {
1455         dsl_scan_t *scn = dp->dp_scan;
1456         spa_t *spa = dp->dp_spa;
1457         int err;
1458
1459         /*
1460          * Check for scn_restart_txg before checking spa_load_state, so
1461          * that we can restart an old-style scan while the pool is being
1462          * imported (see dsl_scan_init).
1463          */
1464         if (scn->scn_restart_txg != 0 &&
1465             scn->scn_restart_txg <= tx->tx_txg) {
1466                 pool_scan_func_t func = POOL_SCAN_SCRUB;
1467                 dsl_scan_done(scn, B_FALSE, tx);
1468                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1469                         func = POOL_SCAN_RESILVER;
1470                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1471                     func, tx->tx_txg);
1472                 dsl_scan_setup_sync(scn, &func, tx);
1473         }
1474
1475         if (!dsl_scan_active(scn) ||
1476             spa_sync_pass(dp->dp_spa) > 1)
1477                 return;
1478
1479         scn->scn_visited_this_txg = 0;
1480         scn->scn_pausing = B_FALSE;
1481         scn->scn_sync_start_time = gethrtime();
1482         spa->spa_scrub_active = B_TRUE;
1483
1484         /*
1485          * First process the free list.  If we pause the free, don't do
1486          * any scanning.  This ensures that there is no free list when
1487          * we are scanning, so the scan code doesn't have to worry about
1488          * traversing it.
1489          */
1490         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1491                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1492                     NULL, ZIO_FLAG_MUSTSUCCEED);
1493                 err = bpobj_iterate(&dp->dp_free_bpobj,
1494                     dsl_scan_free_cb, scn, tx);
1495                 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1496                 if (scn->scn_visited_this_txg) {
1497                         zfs_dbgmsg("freed %llu blocks in %llums from "
1498                             "free_bpobj txg %llu",
1499                             (longlong_t)scn->scn_visited_this_txg,
1500                             (longlong_t)
1501                             (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1502                             (longlong_t)tx->tx_txg);
1503                         scn->scn_visited_this_txg = 0;
1504                         /*
1505                          * Re-sync the ddt so that we can further modify
1506                          * it when doing bprewrite.
1507                          */
1508                         ddt_sync(spa, tx->tx_txg);
1509                 }
1510                 if (err == ERESTART)
1511                         return;
1512         }
1513
1514         if (scn->scn_phys.scn_state != DSS_SCANNING)
1515                 return;
1516
1517         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1518             scn->scn_phys.scn_ddt_class_max) {
1519                 zfs_dbgmsg("doing scan sync txg %llu; "
1520                     "ddt bm=%llu/%llu/%llu/%llx",
1521                     (longlong_t)tx->tx_txg,
1522                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1523                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1524                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1525                     (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1526                 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1527                 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1528                 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1529                 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1530         } else {
1531                 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1532                     (longlong_t)tx->tx_txg,
1533                     (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1534                     (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1535                     (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1536                     (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1537         }
1538
1539         scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1540             NULL, ZIO_FLAG_CANFAIL);
1541         dsl_scan_visit(scn, tx);
1542         (void) zio_wait(scn->scn_zio_root);
1543         scn->scn_zio_root = NULL;
1544
1545         zfs_dbgmsg("visited %llu blocks in %llums",
1546             (longlong_t)scn->scn_visited_this_txg,
1547             (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1548
1549         if (!scn->scn_pausing) {
1550                 /* finished with scan. */
1551                 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1552                 dsl_scan_done(scn, B_TRUE, tx);
1553         }
1554
1555         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1556                 mutex_enter(&spa->spa_scrub_lock);
1557                 while (spa->spa_scrub_inflight > 0) {
1558                         cv_wait(&spa->spa_scrub_io_cv,
1559                             &spa->spa_scrub_lock);
1560                 }
1561                 mutex_exit(&spa->spa_scrub_lock);
1562         }
1563
1564         dsl_scan_sync_state(scn, tx);
1565 }
1566
1567 /*
1568  * This will start a new scan, or restart an existing one.
1569  */
1570 void
1571 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1572 {
1573         if (txg == 0) {
1574                 dmu_tx_t *tx;
1575                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1576                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1577
1578                 txg = dmu_tx_get_txg(tx);
1579                 dp->dp_scan->scn_restart_txg = txg;
1580                 dmu_tx_commit(tx);
1581         } else {
1582                 dp->dp_scan->scn_restart_txg = txg;
1583         }
1584         zfs_dbgmsg("restarting resilver txg=%llu", txg);
1585 }
1586
1587 boolean_t
1588 dsl_scan_resilvering(dsl_pool_t *dp)
1589 {
1590         return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1591             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1592 }
1593
1594 /*
1595  * scrub consumers
1596  */
1597
1598 static void
1599 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1600 {
1601         int i;
1602
1603         /*
1604          * If we resume after a reboot, zab will be NULL; don't record
1605          * incomplete stats in that case.
1606          */
1607         if (zab == NULL)
1608                 return;
1609
1610         for (i = 0; i < 4; i++) {
1611                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1612                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1613                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1614                 int equal;
1615
1616                 zb->zb_count++;
1617                 zb->zb_asize += BP_GET_ASIZE(bp);
1618                 zb->zb_lsize += BP_GET_LSIZE(bp);
1619                 zb->zb_psize += BP_GET_PSIZE(bp);
1620                 zb->zb_gangs += BP_COUNT_GANG(bp);
1621
1622                 switch (BP_GET_NDVAS(bp)) {
1623                 case 2:
1624                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1625                             DVA_GET_VDEV(&bp->blk_dva[1]))
1626                                 zb->zb_ditto_2_of_2_samevdev++;
1627                         break;
1628                 case 3:
1629                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1630                             DVA_GET_VDEV(&bp->blk_dva[1])) +
1631                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1632                             DVA_GET_VDEV(&bp->blk_dva[2])) +
1633                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1634                             DVA_GET_VDEV(&bp->blk_dva[2]));
1635                         if (equal == 1)
1636                                 zb->zb_ditto_2_of_3_samevdev++;
1637                         else if (equal == 3)
1638                                 zb->zb_ditto_3_of_3_samevdev++;
1639                         break;
1640                 }
1641         }
1642 }
1643
1644 static void
1645 dsl_scan_scrub_done(zio_t *zio)
1646 {
1647         spa_t *spa = zio->io_spa;
1648
1649         zio_data_buf_free(zio->io_data, zio->io_size);
1650
1651         mutex_enter(&spa->spa_scrub_lock);
1652         spa->spa_scrub_inflight--;
1653         cv_broadcast(&spa->spa_scrub_io_cv);
1654
1655         if (zio->io_error && (zio->io_error != ECKSUM ||
1656             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1657                 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1658         }
1659         mutex_exit(&spa->spa_scrub_lock);
1660 }
1661
1662 static int
1663 dsl_scan_scrub_cb(dsl_pool_t *dp,
1664     const blkptr_t *bp, const zbookmark_t *zb)
1665 {
1666         dsl_scan_t *scn = dp->dp_scan;
1667         size_t size = BP_GET_PSIZE(bp);
1668         spa_t *spa = dp->dp_spa;
1669         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1670         boolean_t needs_io = B_FALSE;
1671         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1672         int zio_priority = 0;
1673         int scan_delay = 0;
1674         int d;
1675
1676         if (phys_birth <= scn->scn_phys.scn_min_txg ||
1677             phys_birth >= scn->scn_phys.scn_max_txg)
1678                 return (0);
1679
1680         count_block(dp->dp_blkstats, bp);
1681
1682         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1683         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1684                 zio_flags |= ZIO_FLAG_SCRUB;
1685                 zio_priority = ZIO_PRIORITY_SCRUB;
1686                 needs_io = B_TRUE;
1687                 scan_delay = zfs_scrub_delay;
1688         } else if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
1689                 zio_flags |= ZIO_FLAG_RESILVER;
1690                 zio_priority = ZIO_PRIORITY_RESILVER;
1691                 needs_io = B_FALSE;
1692                 scan_delay = zfs_resilver_delay;
1693         }
1694
1695         /* If it's an intent log block, failure is expected. */
1696         if (zb->zb_level == ZB_ZIL_LEVEL)
1697                 zio_flags |= ZIO_FLAG_SPECULATIVE;
1698
1699         for (d = 0; d < BP_GET_NDVAS(bp); d++) {
1700                 vdev_t *vd = vdev_lookup_top(spa,
1701                     DVA_GET_VDEV(&bp->blk_dva[d]));
1702
1703                 /*
1704                  * Keep track of how much data we've examined so that
1705                  * zpool(1M) status can make useful progress reports.
1706                  */
1707                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1708                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1709
1710                 /* if it's a resilver, this may not be in the target range */
1711                 if (!needs_io) {
1712                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
1713                                 /*
1714                                  * Gang members may be spread across multiple
1715                                  * vdevs, so the best estimate we have is the
1716                                  * scrub range, which has already been checked.
1717                                  * XXX -- it would be better to change our
1718                                  * allocation policy to ensure that all
1719                                  * gang members reside on the same vdev.
1720                                  */
1721                                 needs_io = B_TRUE;
1722                         } else {
1723                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1724                                     phys_birth, 1);
1725                         }
1726                 }
1727         }
1728
1729         if (needs_io && !zfs_no_scrub_io) {
1730                 vdev_t *rvd = spa->spa_root_vdev;
1731                 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1732                 void *data = zio_data_buf_alloc(size);
1733
1734                 mutex_enter(&spa->spa_scrub_lock);
1735                 while (spa->spa_scrub_inflight >= maxinflight)
1736                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1737                 spa->spa_scrub_inflight++;
1738                 mutex_exit(&spa->spa_scrub_lock);
1739
1740                 /*
1741                  * If we're seeing recent (zfs_scan_idle) "important" I/Os
1742                  * then throttle our workload to limit the impact of a scan.
1743                  */
1744                 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1745                         delay(scan_delay);
1746
1747                 zio_nowait(zio_read(NULL, spa, bp, data, size,
1748                     dsl_scan_scrub_done, NULL, zio_priority,
1749                     zio_flags, zb));
1750         }
1751
1752         /* do not relocate this block */
1753         return (0);
1754 }
1755
1756 int
1757 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1758 {
1759         spa_t *spa = dp->dp_spa;
1760
1761         /*
1762          * Purge all vdev caches and probe all devices.  We do this here
1763          * rather than in sync context because this requires a writer lock
1764          * on the spa_config lock, which we can't do from sync context.  The
1765          * spa_scrub_reopen flag indicates that vdev_open() should not
1766          * attempt to start another scrub.
1767          */
1768         spa_vdev_state_enter(spa, SCL_NONE);
1769         spa->spa_scrub_reopen = B_TRUE;
1770         vdev_reopen(spa->spa_root_vdev);
1771         spa->spa_scrub_reopen = B_FALSE;
1772         (void) spa_vdev_state_exit(spa, NULL, 0);
1773
1774         return (dsl_sync_task_do(dp, dsl_scan_setup_check,
1775             dsl_scan_setup_sync, dp->dp_scan, &func, 0));
1776 }