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.
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.
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]
22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
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>
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>
46 #include <sys/sa_impl.h>
48 #include <sys/zfs_vfsops.h>
51 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
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);
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 */
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 */
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)
76 extern int zfs_txg_timeout;
78 /* the order has to match pool_scan_type */
79 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
81 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
82 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
86 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
90 spa_t *spa = dp->dp_spa;
93 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
96 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
97 "scrub_func", sizeof (uint64_t), 1, &f);
100 * There was an old-style scrub in progress. Restart a
101 * new-style scrub from the beginning.
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);
109 * Load the queue obj from the old location so that it
110 * can be freed by dsl_scan_done().
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);
116 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
117 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
124 if (scn->scn_phys.scn_state == DSS_SCANNING &&
125 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
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
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);
140 spa_scan_stat_init(spa);
145 dsl_scan_fini(dsl_pool_t *dp)
148 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
155 dsl_scan_setup_check(void *arg1, void *arg2, dmu_tx_t *tx)
157 dsl_scan_t *scn = arg1;
159 if (scn->scn_phys.scn_state == DSS_SCANNING)
167 dsl_scan_setup_sync(void *arg1, void *arg2, dmu_tx_t *tx)
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;
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);
189 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
190 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
192 /* rewrite all disk labels */
193 vdev_config_dirty(spa->spa_root_vdev);
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);
199 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
202 spa->spa_scrub_started = B_TRUE;
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.
208 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
209 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
213 /* back to the generic stuff */
215 if (dp->dp_blkstats == NULL) {
217 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
219 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
221 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
222 ot = DMU_OT_ZAP_OTHER;
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);
227 dsl_scan_sync_state(scn, tx);
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);
236 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
238 static const char *old_names[] = {
240 "scrub_ddt_bookmark",
241 "scrub_ddt_class_max",
250 dsl_pool_t *dp = scn->scn_dp;
251 spa_t *spa = dp->dp_spa;
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);
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;
267 * If we were "restarted" from a stopped state, don't bother
268 * with anything else.
270 if (scn->scn_phys.scn_state != DSS_SCANNING)
274 scn->scn_phys.scn_state = DSS_FINISHED;
276 scn->scn_phys.scn_state = DSS_CANCELED;
278 spa_history_log_internal(LOG_POOL_SCAN_DONE, spa, tx,
279 "complete=%u", complete);
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);
287 mutex_exit(&spa->spa_scrub_lock);
288 spa->spa_scrub_started = B_FALSE;
289 spa->spa_scrub_active = B_FALSE;
292 * If the scrub/resilver completed, update all DTLs to
293 * reflect this. Whether it succeeded or not, vacate
294 * all temporary scrub DTLs.
296 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
297 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
299 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
300 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
302 spa_errlog_rotate(spa);
305 * We may have finished replacing a device.
306 * Let the async thread assess this and handle the detach.
308 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
311 scn->scn_phys.scn_end_time = gethrestime_sec();
316 dsl_scan_cancel_check(void *arg1, void *arg2, dmu_tx_t *tx)
318 dsl_scan_t *scn = arg1;
320 if (scn->scn_phys.scn_state != DSS_SCANNING)
327 dsl_scan_cancel_sync(void *arg1, void *arg2, dmu_tx_t *tx)
329 dsl_scan_t *scn = arg1;
331 dsl_scan_done(scn, B_FALSE, tx);
332 dsl_scan_sync_state(scn, tx);
336 dsl_scan_cancel(dsl_pool_t *dp)
338 boolean_t complete = B_FALSE;
341 err = dsl_sync_task_do(dp, dsl_scan_cancel_check,
342 dsl_scan_cancel_sync, dp->dp_scan, &complete, 3);
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,
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);
355 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
357 zio_free(dp->dp_spa, txg, bp);
361 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
363 ASSERT(dsl_pool_sync_context(dp));
364 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
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)
372 return (arc_read(pio, spa, bpp, pbuf, done, private,
373 priority, zio_flags, arc_flags, zb));
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)
381 return (arc_read_nolock(pio, spa, bpp, done, private,
382 priority, zio_flags, arc_flags, zb));
386 bookmark_is_zero(const zbookmark_t *zb)
388 return (zb->zb_objset == 0 && zb->zb_object == 0 &&
389 zb->zb_level == 0 && zb->zb_blkid == 0);
392 /* dnp is the dnode for zb1->zb_object */
394 bookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
395 const zbookmark_t *zb2)
397 uint64_t zb1nextL0, zb2thisobj;
399 ASSERT(zb1->zb_objset == zb2->zb_objset);
400 ASSERT(zb2->zb_level == 0);
403 * A bookmark in the deadlist is considered to be after
406 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
409 /* The objset_phys_t isn't before anything. */
413 zb1nextL0 = (zb1->zb_blkid + 1) <<
414 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
416 zb2thisobj = zb2->zb_object ? zb2->zb_object :
417 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
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);
425 if (zb1->zb_object < zb2thisobj)
427 if (zb1->zb_object > zb2thisobj)
429 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
431 return (zb1nextL0 <= zb2->zb_blkid);
435 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
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));
444 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
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));
453 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
455 uint64_t elapsed_nanosecs;
458 /* we never skip user/group accounting objects */
459 if (zb && (int64_t)zb->zb_object < 0)
462 if (scn->scn_pausing)
463 return (B_TRUE); /* we're already pausing */
465 if (!bookmark_is_zero(&scn->scn_phys.scn_bookmark))
466 return (B_FALSE); /* we're resuming */
468 /* We only know how to resume from level-0 blocks. */
469 if (zb && zb->zb_level != 0)
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)) {
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;
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;
498 typedef struct zil_scan_arg {
500 zil_header_t *zsa_zh;
505 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
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;
513 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
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).
522 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
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]);
528 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
534 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
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;
545 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
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)
553 if (claim_txg == 0 || bp->blk_birth < claim_txg)
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));
560 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
566 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
568 uint64_t claim_txg = zh->zh_claim_txg;
569 zil_scan_arg_t zsa = { dp, zh };
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).
576 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
579 zilog = zil_alloc(dp->dp_meta_objset, zh);
581 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
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)
593 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
595 if (zfs_no_scrub_prefetch)
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))
602 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
605 * XXX need to make sure all of these arc_read() prefetches are
606 * done before setting xlateall (similar to dsl_read())
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);
614 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
615 const zbookmark_t *zb)
618 * We never skip over user/group accounting objects (obj<0)
620 if (!bookmark_is_zero(&scn->scn_phys.scn_bookmark) &&
621 (int64_t)zb->zb_object >= 0) {
623 * If we already visited this bp & everything below (in
624 * a prior txg sync), don't bother doing it again.
626 if (bookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
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
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));
649 * Return nonzero on i/o error.
650 * Return new buf to write out in *bufp.
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)
657 dsl_pool_t *dp = scn->scn_dp;
658 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
661 if (BP_GET_LEVEL(bp) > 0) {
662 uint32_t flags = ARC_WAIT;
665 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
667 err = arc_read_nolock(NULL, dp->dp_spa, bp,
668 arc_getbuf_func, bufp,
669 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
671 scn->scn_phys.scn_errors++;
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);
678 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
681 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
683 zb->zb_blkid * epb + i);
684 dsl_scan_visitbp(cbp, &czb, dnp,
685 *bufp, ds, scn, ostype, tx);
687 } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
688 uint32_t flags = ARC_WAIT;
690 err = arc_read_nolock(NULL, dp->dp_spa, bp,
691 arc_getbuf_func, bufp,
692 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
694 scn->scn_phys.scn_errors++;
697 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
698 uint32_t flags = ARC_WAIT;
701 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
703 err = arc_read_nolock(NULL, dp->dp_spa, bp,
704 arc_getbuf_func, bufp,
705 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
707 scn->scn_phys.scn_errors++;
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);
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);
722 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
723 uint32_t flags = ARC_WAIT;
726 err = arc_read_nolock(NULL, dp->dp_spa, bp,
727 arc_getbuf_func, bufp,
728 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
730 scn->scn_phys.scn_errors++;
734 osp = (*bufp)->b_data;
736 dsl_scan_visitdnode(scn, ds, osp->os_type,
737 &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
739 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
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.
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);
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)
765 for (j = 0; j < dnp->dn_nblkptr; j++) {
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);
774 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
776 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
778 dsl_scan_visitbp(&dnp->dn_spill,
779 &czb, dnp, buf, ds, scn, ostype, tx);
784 * The arguments are in this order because mdb can only print the
785 * first 5; we want them to be useful.
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,
793 dsl_pool_t *dp = scn->scn_dp;
794 arc_buf_t *buf = NULL;
795 blkptr_t bp_toread = *bp;
797 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
799 if (dsl_scan_check_pause(scn, zb))
802 if (dsl_scan_check_resume(scn, dnp, zb))
805 if (bp->blk_birth == 0)
808 scn->scn_visited_this_txg++;
811 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
812 ds, ds ? ds->ds_object : 0,
813 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
816 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
819 if (BP_GET_TYPE(bp) != DMU_OT_USERGROUP_USED) {
821 * For non-user-accounting blocks, we need to read the
822 * new bp (from a deleted snapshot, found in
823 * check_existing_xlation). If we used the old bp,
824 * pointers inside this block from before we resumed
825 * would be untranslated.
827 * For user-accounting blocks, we need to read the old
828 * bp, because we will apply the entire space delta to
829 * it (original untranslated -> translations from
830 * deleted snap -> now).
835 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
840 * If dsl_scan_ddt() has aready visited this block, it will have
841 * already done any translations or scrubbing, so don't call the
844 if (ddt_class_contains(dp->dp_spa,
845 scn->scn_phys.scn_ddt_class_max, bp)) {
851 * If this block is from the future (after cur_max_txg), then we
852 * are doing this on behalf of a deleted snapshot, and we will
853 * revisit the future block on the next pass of this dataset.
854 * Don't scan it now unless we need to because something
855 * under it was modified.
857 if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
858 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
861 (void) arc_buf_remove_ref(buf, &buf);
865 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
870 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
871 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
872 dsl_scan_visitbp(bp, &zb, NULL, NULL,
873 ds, scn, DMU_OST_NONE, tx);
875 dprintf_ds(ds, "finished scan%s", "");
879 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
881 dsl_pool_t *dp = ds->ds_dir->dd_pool;
882 dsl_scan_t *scn = dp->dp_scan;
885 if (scn->scn_phys.scn_state != DSS_SCANNING)
888 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
889 if (dsl_dataset_is_snapshot(ds)) {
890 /* Note, scn_cur_{min,max}_txg stays the same. */
891 scn->scn_phys.scn_bookmark.zb_objset =
892 ds->ds_phys->ds_next_snap_obj;
893 zfs_dbgmsg("destroying ds %llu; currently traversing; "
894 "reset zb_objset to %llu",
895 (u_longlong_t)ds->ds_object,
896 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
897 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
899 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
900 ZB_DESTROYED_OBJSET, 0, 0, 0);
901 zfs_dbgmsg("destroying ds %llu; currently traversing; "
902 "reset bookmark to -1,0,0,0",
903 (u_longlong_t)ds->ds_object);
905 } else if (zap_lookup_int_key(dp->dp_meta_objset,
906 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
907 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
908 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
909 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
910 if (dsl_dataset_is_snapshot(ds)) {
912 * We keep the same mintxg; it could be >
913 * ds_creation_txg if the previous snapshot was
916 VERIFY(zap_add_int_key(dp->dp_meta_objset,
917 scn->scn_phys.scn_queue_obj,
918 ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
919 zfs_dbgmsg("destroying ds %llu; in queue; "
920 "replacing with %llu",
921 (u_longlong_t)ds->ds_object,
922 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
924 zfs_dbgmsg("destroying ds %llu; in queue; removing",
925 (u_longlong_t)ds->ds_object);
928 zfs_dbgmsg("destroying ds %llu; ignoring",
929 (u_longlong_t)ds->ds_object);
933 * dsl_scan_sync() should be called after this, and should sync
934 * out our changed state, but just to be safe, do it here.
936 dsl_scan_sync_state(scn, tx);
940 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
942 dsl_pool_t *dp = ds->ds_dir->dd_pool;
943 dsl_scan_t *scn = dp->dp_scan;
946 if (scn->scn_phys.scn_state != DSS_SCANNING)
949 ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
951 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
952 scn->scn_phys.scn_bookmark.zb_objset =
953 ds->ds_phys->ds_prev_snap_obj;
954 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
955 "reset zb_objset to %llu",
956 (u_longlong_t)ds->ds_object,
957 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
958 } else if (zap_lookup_int_key(dp->dp_meta_objset,
959 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
960 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
961 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
962 VERIFY(zap_add_int_key(dp->dp_meta_objset,
963 scn->scn_phys.scn_queue_obj,
964 ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
965 zfs_dbgmsg("snapshotting ds %llu; in queue; "
966 "replacing with %llu",
967 (u_longlong_t)ds->ds_object,
968 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
970 dsl_scan_sync_state(scn, tx);
974 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
976 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
977 dsl_scan_t *scn = dp->dp_scan;
980 if (scn->scn_phys.scn_state != DSS_SCANNING)
983 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
984 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
985 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
986 "reset zb_objset to %llu",
987 (u_longlong_t)ds1->ds_object,
988 (u_longlong_t)ds2->ds_object);
989 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
990 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
991 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
992 "reset zb_objset to %llu",
993 (u_longlong_t)ds2->ds_object,
994 (u_longlong_t)ds1->ds_object);
997 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
998 ds1->ds_object, &mintxg) == 0) {
1001 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
1002 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
1003 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1004 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
1005 err = zap_add_int_key(dp->dp_meta_objset,
1006 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
1007 VERIFY(err == 0 || err == EEXIST);
1008 if (err == EEXIST) {
1009 /* Both were there to begin with */
1010 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1011 scn->scn_phys.scn_queue_obj,
1012 ds1->ds_object, mintxg, tx));
1014 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1015 "replacing with %llu",
1016 (u_longlong_t)ds1->ds_object,
1017 (u_longlong_t)ds2->ds_object);
1018 } else if (zap_lookup_int_key(dp->dp_meta_objset,
1019 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
1020 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
1021 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
1022 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1023 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
1024 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
1025 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
1026 zfs_dbgmsg("clone_swap ds %llu; in queue; "
1027 "replacing with %llu",
1028 (u_longlong_t)ds2->ds_object,
1029 (u_longlong_t)ds1->ds_object);
1032 dsl_scan_sync_state(scn, tx);
1035 struct enqueue_clones_arg {
1042 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1044 struct enqueue_clones_arg *eca = arg;
1047 dsl_pool_t *dp = spa->spa_dsl_pool;
1048 dsl_scan_t *scn = dp->dp_scan;
1050 err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1054 if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
1055 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1056 dsl_dataset_t *prev;
1057 err = dsl_dataset_hold_obj(dp,
1058 ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1060 dsl_dataset_rele(ds, FTAG);
1065 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1066 scn->scn_phys.scn_queue_obj, ds->ds_object,
1067 ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1069 dsl_dataset_rele(ds, FTAG);
1074 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1076 dsl_pool_t *dp = scn->scn_dp;
1081 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1083 if (dmu_objset_from_ds(ds, &os))
1087 * Only the ZIL in the head (non-snapshot) is valid. Even though
1088 * snapshots can have ZIL block pointers (which may be the same
1089 * BP as in the head), they must be ignored. So we traverse the
1090 * ZIL here, rather than in scan_recurse(), because the regular
1091 * snapshot block-sharing rules don't apply to it.
1093 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1094 dsl_scan_zil(dp, &os->os_zil_header);
1097 * Iterate over the bps in this ds.
1099 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1100 dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1102 dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1103 dsl_dataset_name(ds, dsname);
1104 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1106 (longlong_t)dsobj, dsname,
1107 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1108 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1109 (int)scn->scn_pausing);
1110 kmem_free(dsname, ZFS_MAXNAMELEN);
1112 if (scn->scn_pausing)
1116 * We've finished this pass over this dataset.
1120 * If we did not completely visit this dataset, do another pass.
1122 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1123 zfs_dbgmsg("incomplete pass; visiting again");
1124 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1125 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1126 scn->scn_phys.scn_queue_obj, ds->ds_object,
1127 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1132 * Add descendent datasets to work queue.
1134 if (ds->ds_phys->ds_next_snap_obj != 0) {
1135 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1136 scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1137 ds->ds_phys->ds_creation_txg, tx) == 0);
1139 if (ds->ds_phys->ds_num_children > 1) {
1140 boolean_t usenext = B_FALSE;
1141 if (ds->ds_phys->ds_next_clones_obj != 0) {
1144 * A bug in a previous version of the code could
1145 * cause upgrade_clones_cb() to not set
1146 * ds_next_snap_obj when it should, leading to a
1147 * missing entry. Therefore we can only use the
1148 * next_clones_obj when its count is correct.
1150 int err = zap_count(dp->dp_meta_objset,
1151 ds->ds_phys->ds_next_clones_obj, &count);
1153 count == ds->ds_phys->ds_num_children - 1)
1158 VERIFY(zap_join_key(dp->dp_meta_objset,
1159 ds->ds_phys->ds_next_clones_obj,
1160 scn->scn_phys.scn_queue_obj,
1161 ds->ds_phys->ds_creation_txg, tx) == 0);
1163 struct enqueue_clones_arg eca;
1165 eca.originobj = ds->ds_object;
1167 (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
1168 NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
1173 dsl_dataset_rele(ds, FTAG);
1178 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1183 dsl_pool_t *dp = spa->spa_dsl_pool;
1184 dsl_scan_t *scn = dp->dp_scan;
1186 err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
1190 while (ds->ds_phys->ds_prev_snap_obj != 0) {
1191 dsl_dataset_t *prev;
1192 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1195 dsl_dataset_rele(ds, FTAG);
1200 * If this is a clone, we don't need to worry about it for now.
1202 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1203 dsl_dataset_rele(ds, FTAG);
1204 dsl_dataset_rele(prev, FTAG);
1207 dsl_dataset_rele(ds, FTAG);
1211 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1212 ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1213 dsl_dataset_rele(ds, FTAG);
1218 * Scrub/dedup interaction.
1220 * If there are N references to a deduped block, we don't want to scrub it
1221 * N times -- ideally, we should scrub it exactly once.
1223 * We leverage the fact that the dde's replication class (enum ddt_class)
1224 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1225 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1227 * To prevent excess scrubbing, the scrub begins by walking the DDT
1228 * to find all blocks with refcnt > 1, and scrubs each of these once.
1229 * Since there are two replication classes which contain blocks with
1230 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1231 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1233 * There would be nothing more to say if a block's refcnt couldn't change
1234 * during a scrub, but of course it can so we must account for changes
1235 * in a block's replication class.
1237 * Here's an example of what can occur:
1239 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1240 * when visited during the top-down scrub phase, it will be scrubbed twice.
1241 * This negates our scrub optimization, but is otherwise harmless.
1243 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1244 * on each visit during the top-down scrub phase, it will never be scrubbed.
1245 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1246 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1247 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1248 * while a scrub is in progress, it scrubs the block right then.
1251 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1253 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1258 bzero(&dde, sizeof (ddt_entry_t));
1260 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1263 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1265 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1266 (longlong_t)ddb->ddb_class,
1267 (longlong_t)ddb->ddb_type,
1268 (longlong_t)ddb->ddb_checksum,
1269 (longlong_t)ddb->ddb_cursor);
1271 /* There should be no pending changes to the dedup table */
1272 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1273 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1275 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1278 if (dsl_scan_check_pause(scn, NULL))
1282 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1283 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1284 (int)scn->scn_pausing);
1286 ASSERT(error == 0 || error == ENOENT);
1287 ASSERT(error != ENOENT ||
1288 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1293 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1294 ddt_entry_t *dde, dmu_tx_t *tx)
1296 const ddt_key_t *ddk = &dde->dde_key;
1297 ddt_phys_t *ddp = dde->dde_phys;
1299 zbookmark_t zb = { 0 };
1302 if (scn->scn_phys.scn_state != DSS_SCANNING)
1305 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1306 if (ddp->ddp_phys_birth == 0 ||
1307 ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1309 ddt_bp_create(checksum, ddk, ddp, &bp);
1311 scn->scn_visited_this_txg++;
1312 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1317 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1319 dsl_pool_t *dp = scn->scn_dp;
1323 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1324 scn->scn_phys.scn_ddt_class_max) {
1325 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1326 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1327 dsl_scan_ddt(scn, tx);
1328 if (scn->scn_pausing)
1332 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1333 /* First do the MOS & ORIGIN */
1335 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1336 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1337 dsl_scan_visit_rootbp(scn, NULL,
1338 &dp->dp_meta_rootbp, tx);
1339 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1340 if (scn->scn_pausing)
1343 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1344 VERIFY(0 == dmu_objset_find_spa(dp->dp_spa,
1345 NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
1347 dsl_scan_visitds(scn,
1348 dp->dp_origin_snap->ds_object, tx);
1350 ASSERT(!scn->scn_pausing);
1351 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1352 ZB_DESTROYED_OBJSET) {
1354 * If we were paused, continue from here. Note if the
1355 * ds we were paused on was deleted, the zb_objset may
1356 * be -1, so we will skip this and find a new objset
1359 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1360 if (scn->scn_pausing)
1365 * In case we were paused right at the end of the ds, zero the
1366 * bookmark so we don't think that we're still trying to resume.
1368 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1370 /* keep pulling things out of the zap-object-as-queue */
1371 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1372 scn->scn_phys.scn_queue_obj),
1373 zap_cursor_retrieve(&zc, &za) == 0) {
1377 dsobj = strtonum(za.za_name, NULL);
1378 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1379 scn->scn_phys.scn_queue_obj, dsobj, tx));
1381 /* Set up min/max txg */
1382 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1383 if (za.za_first_integer != 0) {
1384 scn->scn_phys.scn_cur_min_txg =
1385 MAX(scn->scn_phys.scn_min_txg,
1386 za.za_first_integer);
1388 scn->scn_phys.scn_cur_min_txg =
1389 MAX(scn->scn_phys.scn_min_txg,
1390 ds->ds_phys->ds_prev_snap_txg);
1392 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1393 dsl_dataset_rele(ds, FTAG);
1395 dsl_scan_visitds(scn, dsobj, tx);
1396 zap_cursor_fini(&zc);
1397 if (scn->scn_pausing)
1400 zap_cursor_fini(&zc);
1404 dsl_scan_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1406 dsl_scan_t *scn = arg;
1407 uint64_t elapsed_nanosecs;
1409 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1411 if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1412 (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1413 txg_sync_waiting(scn->scn_dp)) ||
1414 spa_shutting_down(scn->scn_dp->dp_spa))
1417 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1418 dmu_tx_get_txg(tx), bp, 0));
1419 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1420 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1421 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1422 scn->scn_visited_this_txg++;
1427 dsl_scan_active(dsl_scan_t *scn)
1429 spa_t *spa = scn->scn_dp->dp_spa;
1430 uint64_t used = 0, comp, uncomp;
1432 if (spa->spa_load_state != SPA_LOAD_NONE)
1434 if (spa_shutting_down(spa))
1437 if (scn->scn_phys.scn_state == DSS_SCANNING)
1440 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1441 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1442 &used, &comp, &uncomp);
1448 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1450 dsl_scan_t *scn = dp->dp_scan;
1451 spa_t *spa = dp->dp_spa;
1455 * Check for scn_restart_txg before checking spa_load_state, so
1456 * that we can restart an old-style scan while the pool is being
1457 * imported (see dsl_scan_init).
1459 if (scn->scn_restart_txg != 0 &&
1460 scn->scn_restart_txg <= tx->tx_txg) {
1461 pool_scan_func_t func = POOL_SCAN_SCRUB;
1462 dsl_scan_done(scn, B_FALSE, tx);
1463 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1464 func = POOL_SCAN_RESILVER;
1465 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1467 dsl_scan_setup_sync(scn, &func, tx);
1470 if (!dsl_scan_active(scn) ||
1471 spa_sync_pass(dp->dp_spa) > 1)
1474 scn->scn_visited_this_txg = 0;
1475 scn->scn_pausing = B_FALSE;
1476 scn->scn_sync_start_time = gethrtime();
1477 spa->spa_scrub_active = B_TRUE;
1480 * First process the free list. If we pause the free, don't do
1481 * any scanning. This ensures that there is no free list when
1482 * we are scanning, so the scan code doesn't have to worry about
1485 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1486 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1487 NULL, ZIO_FLAG_MUSTSUCCEED);
1488 err = bpobj_iterate(&dp->dp_free_bpobj,
1489 dsl_scan_free_cb, scn, tx);
1490 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1491 if (scn->scn_visited_this_txg) {
1492 zfs_dbgmsg("freed %llu blocks in %llums from "
1493 "free_bpobj txg %llu",
1494 (longlong_t)scn->scn_visited_this_txg,
1496 (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1497 (longlong_t)tx->tx_txg);
1498 scn->scn_visited_this_txg = 0;
1500 * Re-sync the ddt so that we can further modify
1501 * it when doing bprewrite.
1503 ddt_sync(spa, tx->tx_txg);
1505 if (err == ERESTART)
1509 if (scn->scn_phys.scn_state != DSS_SCANNING)
1512 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1513 scn->scn_phys.scn_ddt_class_max) {
1514 zfs_dbgmsg("doing scan sync txg %llu; "
1515 "ddt bm=%llu/%llu/%llu/%llx",
1516 (longlong_t)tx->tx_txg,
1517 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1518 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1519 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1520 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1521 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1522 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1523 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1524 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1526 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1527 (longlong_t)tx->tx_txg,
1528 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1529 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1530 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1531 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1534 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1535 NULL, ZIO_FLAG_CANFAIL);
1536 dsl_scan_visit(scn, tx);
1537 (void) zio_wait(scn->scn_zio_root);
1538 scn->scn_zio_root = NULL;
1540 zfs_dbgmsg("visited %llu blocks in %llums",
1541 (longlong_t)scn->scn_visited_this_txg,
1542 (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1544 if (!scn->scn_pausing) {
1545 /* finished with scan. */
1546 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1547 dsl_scan_done(scn, B_TRUE, tx);
1550 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1551 mutex_enter(&spa->spa_scrub_lock);
1552 while (spa->spa_scrub_inflight > 0) {
1553 cv_wait(&spa->spa_scrub_io_cv,
1554 &spa->spa_scrub_lock);
1556 mutex_exit(&spa->spa_scrub_lock);
1559 dsl_scan_sync_state(scn, tx);
1563 * This will start a new scan, or restart an existing one.
1566 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1570 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1571 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1573 txg = dmu_tx_get_txg(tx);
1574 dp->dp_scan->scn_restart_txg = txg;
1577 dp->dp_scan->scn_restart_txg = txg;
1579 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1583 dsl_scan_resilvering(dsl_pool_t *dp)
1585 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1586 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1594 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1599 * If we resume after a reboot, zab will be NULL; don't record
1600 * incomplete stats in that case.
1605 for (i = 0; i < 4; i++) {
1606 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1607 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1608 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1612 zb->zb_asize += BP_GET_ASIZE(bp);
1613 zb->zb_lsize += BP_GET_LSIZE(bp);
1614 zb->zb_psize += BP_GET_PSIZE(bp);
1615 zb->zb_gangs += BP_COUNT_GANG(bp);
1617 switch (BP_GET_NDVAS(bp)) {
1619 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1620 DVA_GET_VDEV(&bp->blk_dva[1]))
1621 zb->zb_ditto_2_of_2_samevdev++;
1624 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1625 DVA_GET_VDEV(&bp->blk_dva[1])) +
1626 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1627 DVA_GET_VDEV(&bp->blk_dva[2])) +
1628 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1629 DVA_GET_VDEV(&bp->blk_dva[2]));
1631 zb->zb_ditto_2_of_3_samevdev++;
1632 else if (equal == 3)
1633 zb->zb_ditto_3_of_3_samevdev++;
1640 dsl_scan_scrub_done(zio_t *zio)
1642 spa_t *spa = zio->io_spa;
1644 zio_data_buf_free(zio->io_data, zio->io_size);
1646 mutex_enter(&spa->spa_scrub_lock);
1647 spa->spa_scrub_inflight--;
1648 cv_broadcast(&spa->spa_scrub_io_cv);
1650 if (zio->io_error && (zio->io_error != ECKSUM ||
1651 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1652 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1654 mutex_exit(&spa->spa_scrub_lock);
1658 dsl_scan_scrub_cb(dsl_pool_t *dp,
1659 const blkptr_t *bp, const zbookmark_t *zb)
1661 dsl_scan_t *scn = dp->dp_scan;
1662 size_t size = BP_GET_PSIZE(bp);
1663 spa_t *spa = dp->dp_spa;
1664 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1665 boolean_t needs_io = B_FALSE;
1666 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1667 int zio_priority = 0;
1671 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1672 phys_birth >= scn->scn_phys.scn_max_txg)
1675 count_block(dp->dp_blkstats, bp);
1677 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1678 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1679 zio_flags |= ZIO_FLAG_SCRUB;
1680 zio_priority = ZIO_PRIORITY_SCRUB;
1682 scan_delay = zfs_scrub_delay;
1683 } else if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
1684 zio_flags |= ZIO_FLAG_RESILVER;
1685 zio_priority = ZIO_PRIORITY_RESILVER;
1687 scan_delay = zfs_resilver_delay;
1690 /* If it's an intent log block, failure is expected. */
1691 if (zb->zb_level == ZB_ZIL_LEVEL)
1692 zio_flags |= ZIO_FLAG_SPECULATIVE;
1694 for (d = 0; d < BP_GET_NDVAS(bp); d++) {
1695 vdev_t *vd = vdev_lookup_top(spa,
1696 DVA_GET_VDEV(&bp->blk_dva[d]));
1699 * Keep track of how much data we've examined so that
1700 * zpool(1M) status can make useful progress reports.
1702 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1703 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1705 /* if it's a resilver, this may not be in the target range */
1707 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1709 * Gang members may be spread across multiple
1710 * vdevs, so the best estimate we have is the
1711 * scrub range, which has already been checked.
1712 * XXX -- it would be better to change our
1713 * allocation policy to ensure that all
1714 * gang members reside on the same vdev.
1718 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1724 if (needs_io && !zfs_no_scrub_io) {
1725 vdev_t *rvd = spa->spa_root_vdev;
1726 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1727 void *data = zio_data_buf_alloc(size);
1729 mutex_enter(&spa->spa_scrub_lock);
1730 while (spa->spa_scrub_inflight >= maxinflight)
1731 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1732 spa->spa_scrub_inflight++;
1733 mutex_exit(&spa->spa_scrub_lock);
1736 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1737 * then throttle our workload to limit the impact of a scan.
1739 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1742 zio_nowait(zio_read(NULL, spa, bp, data, size,
1743 dsl_scan_scrub_done, NULL, zio_priority,
1747 /* do not relocate this block */
1752 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1754 spa_t *spa = dp->dp_spa;
1757 * Purge all vdev caches and probe all devices. We do this here
1758 * rather than in sync context because this requires a writer lock
1759 * on the spa_config lock, which we can't do from sync context. The
1760 * spa_scrub_reopen flag indicates that vdev_open() should not
1761 * attempt to start another scrub.
1763 spa_vdev_state_enter(spa, SCL_NONE);
1764 spa->spa_scrub_reopen = B_TRUE;
1765 vdev_reopen(spa->spa_root_vdev);
1766 spa->spa_scrub_reopen = B_FALSE;
1767 (void) spa_vdev_state_exit(spa, NULL, 0);
1769 return (dsl_sync_task_do(dp, dsl_scan_setup_check,
1770 dsl_scan_setup_sync, dp->dp_scan, &func, 0));