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.
23 * Copyright (c) 2013 by Delphix. All rights reserved.
26 #include <sys/dsl_scan.h>
27 #include <sys/dsl_pool.h>
28 #include <sys/dsl_dataset.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_synctask.h>
32 #include <sys/dnode.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
38 #include <sys/zfs_context.h>
39 #include <sys/fs/zfs.h>
40 #include <sys/zfs_znode.h>
41 #include <sys/spa_impl.h>
42 #include <sys/vdev_impl.h>
43 #include <sys/zil_impl.h>
44 #include <sys/zio_checksum.h>
47 #include <sys/sa_impl.h>
48 #include <sys/zfeature.h>
50 #include <sys/zfs_vfsops.h>
53 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
55 static scan_cb_t dsl_scan_defrag_cb;
56 static scan_cb_t dsl_scan_scrub_cb;
57 static scan_cb_t dsl_scan_remove_cb;
58 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
59 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
61 unsigned int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */
62 unsigned int zfs_resilver_delay = 2; /* number of ticks to delay resilver */
63 unsigned int zfs_scrub_delay = 4; /* number of ticks to delay scrub */
64 unsigned int zfs_scan_idle = 50; /* idle window in clock ticks */
66 unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
67 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
68 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
70 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
71 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable srub prefetching */
73 SYSCTL_DECL(_vfs_zfs);
74 TUNABLE_INT("vfs.zfs.top_maxinflight", &zfs_top_maxinflight);
75 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RW,
76 &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
77 TUNABLE_INT("vfs.zfs.resilver_delay", &zfs_resilver_delay);
78 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RW,
79 &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
80 TUNABLE_INT("vfs.zfs.scrub_delay", &zfs_scrub_delay);
81 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RW,
82 &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
83 TUNABLE_INT("vfs.zfs.scan_idle", &zfs_scan_idle);
84 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RW,
85 &zfs_scan_idle, 0, "Idle scan window in clock ticks");
86 TUNABLE_INT("vfs.zfs.scan_min_time_ms", &zfs_scan_min_time_ms);
87 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RW,
88 &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
89 TUNABLE_INT("vfs.zfs.free_min_time_ms", &zfs_free_min_time_ms);
90 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RW,
91 &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
92 TUNABLE_INT("vfs.zfs.resilver_min_time_ms", &zfs_resilver_min_time_ms);
93 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RW,
94 &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
95 TUNABLE_INT("vfs.zfs.no_scrub_io", &zfs_no_scrub_io);
96 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RW,
97 &zfs_no_scrub_io, 0, "Disable scrub I/O");
98 TUNABLE_INT("vfs.zfs.no_scrub_prefetch", &zfs_no_scrub_prefetch);
99 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RW,
100 &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
102 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
104 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
105 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
106 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
108 extern int zfs_txg_timeout;
110 /* the order has to match pool_scan_type */
111 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
113 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
114 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
118 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
122 spa_t *spa = dp->dp_spa;
125 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
129 * It's possible that we're resuming a scan after a reboot so
130 * make sure that the scan_async_destroying flag is initialized
133 ASSERT(!scn->scn_async_destroying);
134 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
135 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY]);
137 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
138 "scrub_func", sizeof (uint64_t), 1, &f);
141 * There was an old-style scrub in progress. Restart a
142 * new-style scrub from the beginning.
144 scn->scn_restart_txg = txg;
145 zfs_dbgmsg("old-style scrub was in progress; "
146 "restarting new-style scrub in txg %llu",
147 scn->scn_restart_txg);
150 * Load the queue obj from the old location so that it
151 * can be freed by dsl_scan_done().
153 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
154 "scrub_queue", sizeof (uint64_t), 1,
155 &scn->scn_phys.scn_queue_obj);
157 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
158 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
165 if (scn->scn_phys.scn_state == DSS_SCANNING &&
166 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
168 * A new-type scrub was in progress on an old
169 * pool, and the pool was accessed by old
170 * software. Restart from the beginning, since
171 * the old software may have changed the pool in
174 scn->scn_restart_txg = txg;
175 zfs_dbgmsg("new-style scrub was modified "
176 "by old software; restarting in txg %llu",
177 scn->scn_restart_txg);
181 spa_scan_stat_init(spa);
186 dsl_scan_fini(dsl_pool_t *dp)
189 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
196 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
198 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
200 if (scn->scn_phys.scn_state == DSS_SCANNING)
201 return (SET_ERROR(EBUSY));
207 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
209 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
210 pool_scan_func_t *funcp = arg;
211 dmu_object_type_t ot = 0;
212 dsl_pool_t *dp = scn->scn_dp;
213 spa_t *spa = dp->dp_spa;
215 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
216 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
217 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
218 scn->scn_phys.scn_func = *funcp;
219 scn->scn_phys.scn_state = DSS_SCANNING;
220 scn->scn_phys.scn_min_txg = 0;
221 scn->scn_phys.scn_max_txg = tx->tx_txg;
222 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
223 scn->scn_phys.scn_start_time = gethrestime_sec();
224 scn->scn_phys.scn_errors = 0;
225 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
226 scn->scn_restart_txg = 0;
227 spa_scan_stat_init(spa);
229 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
230 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
232 /* rewrite all disk labels */
233 vdev_config_dirty(spa->spa_root_vdev);
235 if (vdev_resilver_needed(spa->spa_root_vdev,
236 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
237 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
239 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
242 spa->spa_scrub_started = B_TRUE;
244 * If this is an incremental scrub, limit the DDT scrub phase
245 * to just the auto-ditto class (for correctness); the rest
246 * of the scrub should go faster using top-down pruning.
248 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
249 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
253 /* back to the generic stuff */
255 if (dp->dp_blkstats == NULL) {
257 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
259 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
261 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
262 ot = DMU_OT_ZAP_OTHER;
264 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
265 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
267 dsl_scan_sync_state(scn, tx);
269 spa_history_log_internal(spa, "scan setup", tx,
270 "func=%u mintxg=%llu maxtxg=%llu",
271 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
276 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
278 static const char *old_names[] = {
280 "scrub_ddt_bookmark",
281 "scrub_ddt_class_max",
290 dsl_pool_t *dp = scn->scn_dp;
291 spa_t *spa = dp->dp_spa;
294 /* Remove any remnants of an old-style scrub. */
295 for (i = 0; old_names[i]; i++) {
296 (void) zap_remove(dp->dp_meta_objset,
297 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
300 if (scn->scn_phys.scn_queue_obj != 0) {
301 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
302 scn->scn_phys.scn_queue_obj, tx));
303 scn->scn_phys.scn_queue_obj = 0;
307 * If we were "restarted" from a stopped state, don't bother
308 * with anything else.
310 if (scn->scn_phys.scn_state != DSS_SCANNING)
314 scn->scn_phys.scn_state = DSS_FINISHED;
316 scn->scn_phys.scn_state = DSS_CANCELED;
318 spa_history_log_internal(spa, "scan done", tx,
319 "complete=%u", complete);
321 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
322 mutex_enter(&spa->spa_scrub_lock);
323 while (spa->spa_scrub_inflight > 0) {
324 cv_wait(&spa->spa_scrub_io_cv,
325 &spa->spa_scrub_lock);
327 mutex_exit(&spa->spa_scrub_lock);
328 spa->spa_scrub_started = B_FALSE;
329 spa->spa_scrub_active = B_FALSE;
332 * If the scrub/resilver completed, update all DTLs to
333 * reflect this. Whether it succeeded or not, vacate
334 * all temporary scrub DTLs.
336 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
337 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
339 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
340 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
342 spa_errlog_rotate(spa);
345 * We may have finished replacing a device.
346 * Let the async thread assess this and handle the detach.
348 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
351 scn->scn_phys.scn_end_time = gethrestime_sec();
356 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
358 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
360 if (scn->scn_phys.scn_state != DSS_SCANNING)
361 return (SET_ERROR(ENOENT));
367 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
369 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
371 dsl_scan_done(scn, B_FALSE, tx);
372 dsl_scan_sync_state(scn, tx);
376 dsl_scan_cancel(dsl_pool_t *dp)
378 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
379 dsl_scan_cancel_sync, NULL, 3));
382 static void dsl_scan_visitbp(blkptr_t *bp,
383 const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
384 dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
386 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
387 dmu_objset_type_t ostype,
388 dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
391 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
393 zio_free(dp->dp_spa, txg, bp);
397 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
399 ASSERT(dsl_pool_sync_context(dp));
400 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
404 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
406 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
407 if (dsl_dataset_is_snapshot(ds))
408 return (MIN(smt, ds->ds_phys->ds_creation_txg));
413 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
415 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
416 DMU_POOL_DIRECTORY_OBJECT,
417 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
418 &scn->scn_phys, tx));
422 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
424 uint64_t elapsed_nanosecs;
425 unsigned int mintime;
427 /* we never skip user/group accounting objects */
428 if (zb && (int64_t)zb->zb_object < 0)
431 if (scn->scn_pausing)
432 return (B_TRUE); /* we're already pausing */
434 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
435 return (B_FALSE); /* we're resuming */
437 /* We only know how to resume from level-0 blocks. */
438 if (zb && zb->zb_level != 0)
441 mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
442 zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
443 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
444 if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
445 (elapsed_nanosecs / MICROSEC > mintime &&
446 txg_sync_waiting(scn->scn_dp)) ||
447 spa_shutting_down(scn->scn_dp->dp_spa)) {
449 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
450 (longlong_t)zb->zb_objset,
451 (longlong_t)zb->zb_object,
452 (longlong_t)zb->zb_level,
453 (longlong_t)zb->zb_blkid);
454 scn->scn_phys.scn_bookmark = *zb;
456 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
457 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
458 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
459 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
460 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
461 scn->scn_pausing = B_TRUE;
467 typedef struct zil_scan_arg {
469 zil_header_t *zsa_zh;
474 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
476 zil_scan_arg_t *zsa = arg;
477 dsl_pool_t *dp = zsa->zsa_dp;
478 dsl_scan_t *scn = dp->dp_scan;
479 zil_header_t *zh = zsa->zsa_zh;
482 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
486 * One block ("stubby") can be allocated a long time ago; we
487 * want to visit that one because it has been allocated
488 * (on-disk) even if it hasn't been claimed (even though for
489 * scrub there's nothing to do to it).
491 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
494 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
495 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
497 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
503 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
505 if (lrc->lrc_txtype == TX_WRITE) {
506 zil_scan_arg_t *zsa = arg;
507 dsl_pool_t *dp = zsa->zsa_dp;
508 dsl_scan_t *scn = dp->dp_scan;
509 zil_header_t *zh = zsa->zsa_zh;
510 lr_write_t *lr = (lr_write_t *)lrc;
511 blkptr_t *bp = &lr->lr_blkptr;
514 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
518 * birth can be < claim_txg if this record's txg is
519 * already txg sync'ed (but this log block contains
520 * other records that are not synced)
522 if (claim_txg == 0 || bp->blk_birth < claim_txg)
525 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
526 lr->lr_foid, ZB_ZIL_LEVEL,
527 lr->lr_offset / BP_GET_LSIZE(bp));
529 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
535 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
537 uint64_t claim_txg = zh->zh_claim_txg;
538 zil_scan_arg_t zsa = { dp, zh };
542 * We only want to visit blocks that have been claimed but not yet
543 * replayed (or, in read-only mode, blocks that *would* be claimed).
545 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
548 zilog = zil_alloc(dp->dp_meta_objset, zh);
550 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
558 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
559 uint64_t objset, uint64_t object, uint64_t blkid)
562 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
564 if (zfs_no_scrub_prefetch)
567 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
568 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
571 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
573 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
574 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
575 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
579 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
580 const zbookmark_t *zb)
583 * We never skip over user/group accounting objects (obj<0)
585 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
586 (int64_t)zb->zb_object >= 0) {
588 * If we already visited this bp & everything below (in
589 * a prior txg sync), don't bother doing it again.
591 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
595 * If we found the block we're trying to resume from, or
596 * we went past it to a different object, zero it out to
597 * indicate that it's OK to start checking for pausing
600 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
601 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
602 dprintf("resuming at %llx/%llx/%llx/%llx\n",
603 (longlong_t)zb->zb_objset,
604 (longlong_t)zb->zb_object,
605 (longlong_t)zb->zb_level,
606 (longlong_t)zb->zb_blkid);
607 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
614 * Return nonzero on i/o error.
615 * Return new buf to write out in *bufp.
618 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
619 dnode_phys_t *dnp, const blkptr_t *bp,
620 const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
622 dsl_pool_t *dp = scn->scn_dp;
623 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
626 if (BP_GET_LEVEL(bp) > 0) {
627 uint32_t flags = ARC_WAIT;
630 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
632 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
633 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
635 scn->scn_phys.scn_errors++;
638 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
639 dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
640 zb->zb_object, zb->zb_blkid * epb + i);
642 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
645 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
647 zb->zb_blkid * epb + i);
648 dsl_scan_visitbp(cbp, &czb, dnp,
649 *bufp, ds, scn, ostype, tx);
651 } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
652 uint32_t flags = ARC_WAIT;
654 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
655 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
657 scn->scn_phys.scn_errors++;
660 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
661 uint32_t flags = ARC_WAIT;
664 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
666 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
667 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
669 scn->scn_phys.scn_errors++;
672 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
673 for (j = 0; j < cdnp->dn_nblkptr; j++) {
674 blkptr_t *cbp = &cdnp->dn_blkptr[j];
675 dsl_scan_prefetch(scn, *bufp, cbp,
676 zb->zb_objset, zb->zb_blkid * epb + i, j);
679 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
680 dsl_scan_visitdnode(scn, ds, ostype,
681 cdnp, *bufp, zb->zb_blkid * epb + i, tx);
684 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
685 uint32_t flags = ARC_WAIT;
688 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
689 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
691 scn->scn_phys.scn_errors++;
695 osp = (*bufp)->b_data;
697 dsl_scan_visitdnode(scn, ds, osp->os_type,
698 &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
700 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
702 * We also always visit user/group accounting
703 * objects, and never skip them, even if we are
704 * pausing. This is necessary so that the space
705 * deltas from this txg get integrated.
707 dsl_scan_visitdnode(scn, ds, osp->os_type,
708 &osp->os_groupused_dnode, *bufp,
709 DMU_GROUPUSED_OBJECT, tx);
710 dsl_scan_visitdnode(scn, ds, osp->os_type,
711 &osp->os_userused_dnode, *bufp,
712 DMU_USERUSED_OBJECT, tx);
720 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
721 dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
722 uint64_t object, dmu_tx_t *tx)
726 for (j = 0; j < dnp->dn_nblkptr; j++) {
729 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
730 dnp->dn_nlevels - 1, j);
731 dsl_scan_visitbp(&dnp->dn_blkptr[j],
732 &czb, dnp, buf, ds, scn, ostype, tx);
735 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
737 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
739 dsl_scan_visitbp(&dnp->dn_spill,
740 &czb, dnp, buf, ds, scn, ostype, tx);
745 * The arguments are in this order because mdb can only print the
746 * first 5; we want them to be useful.
749 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
750 dnode_phys_t *dnp, arc_buf_t *pbuf,
751 dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
754 dsl_pool_t *dp = scn->scn_dp;
755 arc_buf_t *buf = NULL;
756 blkptr_t bp_toread = *bp;
758 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
760 if (dsl_scan_check_pause(scn, zb))
763 if (dsl_scan_check_resume(scn, dnp, zb))
766 if (bp->blk_birth == 0)
769 scn->scn_visited_this_txg++;
772 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
773 ds, ds ? ds->ds_object : 0,
774 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
777 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
780 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
785 * If dsl_scan_ddt() has aready visited this block, it will have
786 * already done any translations or scrubbing, so don't call the
789 if (ddt_class_contains(dp->dp_spa,
790 scn->scn_phys.scn_ddt_class_max, bp)) {
796 * If this block is from the future (after cur_max_txg), then we
797 * are doing this on behalf of a deleted snapshot, and we will
798 * revisit the future block on the next pass of this dataset.
799 * Don't scan it now unless we need to because something
800 * under it was modified.
802 if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
803 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
806 (void) arc_buf_remove_ref(buf, &buf);
810 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
815 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
816 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
817 dsl_scan_visitbp(bp, &zb, NULL, NULL,
818 ds, scn, DMU_OST_NONE, tx);
820 dprintf_ds(ds, "finished scan%s", "");
824 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
826 dsl_pool_t *dp = ds->ds_dir->dd_pool;
827 dsl_scan_t *scn = dp->dp_scan;
830 if (scn->scn_phys.scn_state != DSS_SCANNING)
833 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
834 if (dsl_dataset_is_snapshot(ds)) {
835 /* Note, scn_cur_{min,max}_txg stays the same. */
836 scn->scn_phys.scn_bookmark.zb_objset =
837 ds->ds_phys->ds_next_snap_obj;
838 zfs_dbgmsg("destroying ds %llu; currently traversing; "
839 "reset zb_objset to %llu",
840 (u_longlong_t)ds->ds_object,
841 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
842 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
844 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
845 ZB_DESTROYED_OBJSET, 0, 0, 0);
846 zfs_dbgmsg("destroying ds %llu; currently traversing; "
847 "reset bookmark to -1,0,0,0",
848 (u_longlong_t)ds->ds_object);
850 } else if (zap_lookup_int_key(dp->dp_meta_objset,
851 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
852 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
853 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
854 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
855 if (dsl_dataset_is_snapshot(ds)) {
857 * We keep the same mintxg; it could be >
858 * ds_creation_txg if the previous snapshot was
861 VERIFY(zap_add_int_key(dp->dp_meta_objset,
862 scn->scn_phys.scn_queue_obj,
863 ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
864 zfs_dbgmsg("destroying ds %llu; in queue; "
865 "replacing with %llu",
866 (u_longlong_t)ds->ds_object,
867 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
869 zfs_dbgmsg("destroying ds %llu; in queue; removing",
870 (u_longlong_t)ds->ds_object);
873 zfs_dbgmsg("destroying ds %llu; ignoring",
874 (u_longlong_t)ds->ds_object);
878 * dsl_scan_sync() should be called after this, and should sync
879 * out our changed state, but just to be safe, do it here.
881 dsl_scan_sync_state(scn, tx);
885 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
887 dsl_pool_t *dp = ds->ds_dir->dd_pool;
888 dsl_scan_t *scn = dp->dp_scan;
891 if (scn->scn_phys.scn_state != DSS_SCANNING)
894 ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
896 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
897 scn->scn_phys.scn_bookmark.zb_objset =
898 ds->ds_phys->ds_prev_snap_obj;
899 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
900 "reset zb_objset to %llu",
901 (u_longlong_t)ds->ds_object,
902 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
903 } else if (zap_lookup_int_key(dp->dp_meta_objset,
904 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
905 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
906 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
907 VERIFY(zap_add_int_key(dp->dp_meta_objset,
908 scn->scn_phys.scn_queue_obj,
909 ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
910 zfs_dbgmsg("snapshotting ds %llu; in queue; "
911 "replacing with %llu",
912 (u_longlong_t)ds->ds_object,
913 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
915 dsl_scan_sync_state(scn, tx);
919 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
921 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
922 dsl_scan_t *scn = dp->dp_scan;
925 if (scn->scn_phys.scn_state != DSS_SCANNING)
928 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
929 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
930 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
931 "reset zb_objset to %llu",
932 (u_longlong_t)ds1->ds_object,
933 (u_longlong_t)ds2->ds_object);
934 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
935 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
936 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
937 "reset zb_objset to %llu",
938 (u_longlong_t)ds2->ds_object,
939 (u_longlong_t)ds1->ds_object);
942 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
943 ds1->ds_object, &mintxg) == 0) {
946 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
947 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
948 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
949 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
950 err = zap_add_int_key(dp->dp_meta_objset,
951 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
952 VERIFY(err == 0 || err == EEXIST);
954 /* Both were there to begin with */
955 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
956 scn->scn_phys.scn_queue_obj,
957 ds1->ds_object, mintxg, tx));
959 zfs_dbgmsg("clone_swap ds %llu; in queue; "
960 "replacing with %llu",
961 (u_longlong_t)ds1->ds_object,
962 (u_longlong_t)ds2->ds_object);
963 } else if (zap_lookup_int_key(dp->dp_meta_objset,
964 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
965 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
966 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
967 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
968 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
969 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
970 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
971 zfs_dbgmsg("clone_swap ds %llu; in queue; "
972 "replacing with %llu",
973 (u_longlong_t)ds2->ds_object,
974 (u_longlong_t)ds1->ds_object);
977 dsl_scan_sync_state(scn, tx);
980 struct enqueue_clones_arg {
987 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
989 struct enqueue_clones_arg *eca = arg;
992 dsl_scan_t *scn = dp->dp_scan;
994 if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
997 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1001 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1002 dsl_dataset_t *prev;
1003 err = dsl_dataset_hold_obj(dp,
1004 ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1006 dsl_dataset_rele(ds, FTAG);
1011 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1012 scn->scn_phys.scn_queue_obj, ds->ds_object,
1013 ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1014 dsl_dataset_rele(ds, FTAG);
1019 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1021 dsl_pool_t *dp = scn->scn_dp;
1025 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1027 if (dmu_objset_from_ds(ds, &os))
1031 * Only the ZIL in the head (non-snapshot) is valid. Even though
1032 * snapshots can have ZIL block pointers (which may be the same
1033 * BP as in the head), they must be ignored. So we traverse the
1034 * ZIL here, rather than in scan_recurse(), because the regular
1035 * snapshot block-sharing rules don't apply to it.
1037 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1038 dsl_scan_zil(dp, &os->os_zil_header);
1041 * Iterate over the bps in this ds.
1043 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1044 dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1046 char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1047 dsl_dataset_name(ds, dsname);
1048 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1050 (longlong_t)dsobj, dsname,
1051 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1052 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1053 (int)scn->scn_pausing);
1054 kmem_free(dsname, ZFS_MAXNAMELEN);
1056 if (scn->scn_pausing)
1060 * We've finished this pass over this dataset.
1064 * If we did not completely visit this dataset, do another pass.
1066 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1067 zfs_dbgmsg("incomplete pass; visiting again");
1068 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1069 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1070 scn->scn_phys.scn_queue_obj, ds->ds_object,
1071 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1076 * Add descendent datasets to work queue.
1078 if (ds->ds_phys->ds_next_snap_obj != 0) {
1079 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1080 scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1081 ds->ds_phys->ds_creation_txg, tx) == 0);
1083 if (ds->ds_phys->ds_num_children > 1) {
1084 boolean_t usenext = B_FALSE;
1085 if (ds->ds_phys->ds_next_clones_obj != 0) {
1088 * A bug in a previous version of the code could
1089 * cause upgrade_clones_cb() to not set
1090 * ds_next_snap_obj when it should, leading to a
1091 * missing entry. Therefore we can only use the
1092 * next_clones_obj when its count is correct.
1094 int err = zap_count(dp->dp_meta_objset,
1095 ds->ds_phys->ds_next_clones_obj, &count);
1097 count == ds->ds_phys->ds_num_children - 1)
1102 VERIFY0(zap_join_key(dp->dp_meta_objset,
1103 ds->ds_phys->ds_next_clones_obj,
1104 scn->scn_phys.scn_queue_obj,
1105 ds->ds_phys->ds_creation_txg, tx));
1107 struct enqueue_clones_arg eca;
1109 eca.originobj = ds->ds_object;
1111 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1112 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1117 dsl_dataset_rele(ds, FTAG);
1122 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1127 dsl_scan_t *scn = dp->dp_scan;
1129 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1133 while (ds->ds_phys->ds_prev_snap_obj != 0) {
1134 dsl_dataset_t *prev;
1135 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1138 dsl_dataset_rele(ds, FTAG);
1143 * If this is a clone, we don't need to worry about it for now.
1145 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1146 dsl_dataset_rele(ds, FTAG);
1147 dsl_dataset_rele(prev, FTAG);
1150 dsl_dataset_rele(ds, FTAG);
1154 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1155 ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1156 dsl_dataset_rele(ds, FTAG);
1161 * Scrub/dedup interaction.
1163 * If there are N references to a deduped block, we don't want to scrub it
1164 * N times -- ideally, we should scrub it exactly once.
1166 * We leverage the fact that the dde's replication class (enum ddt_class)
1167 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1168 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1170 * To prevent excess scrubbing, the scrub begins by walking the DDT
1171 * to find all blocks with refcnt > 1, and scrubs each of these once.
1172 * Since there are two replication classes which contain blocks with
1173 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1174 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1176 * There would be nothing more to say if a block's refcnt couldn't change
1177 * during a scrub, but of course it can so we must account for changes
1178 * in a block's replication class.
1180 * Here's an example of what can occur:
1182 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1183 * when visited during the top-down scrub phase, it will be scrubbed twice.
1184 * This negates our scrub optimization, but is otherwise harmless.
1186 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1187 * on each visit during the top-down scrub phase, it will never be scrubbed.
1188 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1189 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1190 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1191 * while a scrub is in progress, it scrubs the block right then.
1194 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1196 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1197 ddt_entry_t dde = { 0 };
1201 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1204 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1206 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1207 (longlong_t)ddb->ddb_class,
1208 (longlong_t)ddb->ddb_type,
1209 (longlong_t)ddb->ddb_checksum,
1210 (longlong_t)ddb->ddb_cursor);
1212 /* There should be no pending changes to the dedup table */
1213 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1214 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1216 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1219 if (dsl_scan_check_pause(scn, NULL))
1223 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1224 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1225 (int)scn->scn_pausing);
1227 ASSERT(error == 0 || error == ENOENT);
1228 ASSERT(error != ENOENT ||
1229 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1234 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1235 ddt_entry_t *dde, dmu_tx_t *tx)
1237 const ddt_key_t *ddk = &dde->dde_key;
1238 ddt_phys_t *ddp = dde->dde_phys;
1240 zbookmark_t zb = { 0 };
1242 if (scn->scn_phys.scn_state != DSS_SCANNING)
1245 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1246 if (ddp->ddp_phys_birth == 0 ||
1247 ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1249 ddt_bp_create(checksum, ddk, ddp, &bp);
1251 scn->scn_visited_this_txg++;
1252 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1257 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1259 dsl_pool_t *dp = scn->scn_dp;
1263 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1264 scn->scn_phys.scn_ddt_class_max) {
1265 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1266 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1267 dsl_scan_ddt(scn, tx);
1268 if (scn->scn_pausing)
1272 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1273 /* First do the MOS & ORIGIN */
1275 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1276 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1277 dsl_scan_visit_rootbp(scn, NULL,
1278 &dp->dp_meta_rootbp, tx);
1279 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1280 if (scn->scn_pausing)
1283 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1284 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1285 enqueue_cb, tx, DS_FIND_CHILDREN));
1287 dsl_scan_visitds(scn,
1288 dp->dp_origin_snap->ds_object, tx);
1290 ASSERT(!scn->scn_pausing);
1291 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1292 ZB_DESTROYED_OBJSET) {
1294 * If we were paused, continue from here. Note if the
1295 * ds we were paused on was deleted, the zb_objset may
1296 * be -1, so we will skip this and find a new objset
1299 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1300 if (scn->scn_pausing)
1305 * In case we were paused right at the end of the ds, zero the
1306 * bookmark so we don't think that we're still trying to resume.
1308 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1310 /* keep pulling things out of the zap-object-as-queue */
1311 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1312 scn->scn_phys.scn_queue_obj),
1313 zap_cursor_retrieve(&zc, &za) == 0) {
1317 dsobj = strtonum(za.za_name, NULL);
1318 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1319 scn->scn_phys.scn_queue_obj, dsobj, tx));
1321 /* Set up min/max txg */
1322 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1323 if (za.za_first_integer != 0) {
1324 scn->scn_phys.scn_cur_min_txg =
1325 MAX(scn->scn_phys.scn_min_txg,
1326 za.za_first_integer);
1328 scn->scn_phys.scn_cur_min_txg =
1329 MAX(scn->scn_phys.scn_min_txg,
1330 ds->ds_phys->ds_prev_snap_txg);
1332 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1333 dsl_dataset_rele(ds, FTAG);
1335 dsl_scan_visitds(scn, dsobj, tx);
1336 zap_cursor_fini(&zc);
1337 if (scn->scn_pausing)
1340 zap_cursor_fini(&zc);
1344 dsl_scan_free_should_pause(dsl_scan_t *scn)
1346 uint64_t elapsed_nanosecs;
1348 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1349 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1350 (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1351 txg_sync_waiting(scn->scn_dp)) ||
1352 spa_shutting_down(scn->scn_dp->dp_spa));
1356 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1358 dsl_scan_t *scn = arg;
1360 if (!scn->scn_is_bptree ||
1361 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1362 if (dsl_scan_free_should_pause(scn))
1363 return (SET_ERROR(ERESTART));
1366 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1367 dmu_tx_get_txg(tx), bp, 0));
1368 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1369 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1370 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1371 scn->scn_visited_this_txg++;
1376 dsl_scan_active(dsl_scan_t *scn)
1378 spa_t *spa = scn->scn_dp->dp_spa;
1379 uint64_t used = 0, comp, uncomp;
1381 if (spa->spa_load_state != SPA_LOAD_NONE)
1383 if (spa_shutting_down(spa))
1386 if (scn->scn_phys.scn_state == DSS_SCANNING ||
1387 scn->scn_async_destroying)
1390 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1391 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1392 &used, &comp, &uncomp);
1398 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1400 dsl_scan_t *scn = dp->dp_scan;
1401 spa_t *spa = dp->dp_spa;
1405 * Check for scn_restart_txg before checking spa_load_state, so
1406 * that we can restart an old-style scan while the pool is being
1407 * imported (see dsl_scan_init).
1409 if (scn->scn_restart_txg != 0 &&
1410 scn->scn_restart_txg <= tx->tx_txg) {
1411 pool_scan_func_t func = POOL_SCAN_SCRUB;
1412 dsl_scan_done(scn, B_FALSE, tx);
1413 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1414 func = POOL_SCAN_RESILVER;
1415 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1417 dsl_scan_setup_sync(&func, tx);
1420 if (!dsl_scan_active(scn) ||
1421 spa_sync_pass(dp->dp_spa) > 1)
1424 scn->scn_visited_this_txg = 0;
1425 scn->scn_pausing = B_FALSE;
1426 scn->scn_sync_start_time = gethrtime();
1427 spa->spa_scrub_active = B_TRUE;
1430 * First process the free list. If we pause the free, don't do
1431 * any scanning. This ensures that there is no free list when
1432 * we are scanning, so the scan code doesn't have to worry about
1435 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1436 scn->scn_is_bptree = B_FALSE;
1437 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1438 NULL, ZIO_FLAG_MUSTSUCCEED);
1439 err = bpobj_iterate(&dp->dp_free_bpobj,
1440 dsl_scan_free_block_cb, scn, tx);
1441 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1443 if (err == 0 && spa_feature_is_active(spa,
1444 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1445 ASSERT(scn->scn_async_destroying);
1446 scn->scn_is_bptree = B_TRUE;
1447 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1448 NULL, ZIO_FLAG_MUSTSUCCEED);
1449 err = bptree_iterate(dp->dp_meta_objset,
1450 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1452 VERIFY0(zio_wait(scn->scn_zio_root));
1455 zfeature_info_t *feat = &spa_feature_table
1456 [SPA_FEATURE_ASYNC_DESTROY];
1457 /* finished; deactivate async destroy feature */
1458 spa_feature_decr(spa, feat, tx);
1459 ASSERT(!spa_feature_is_active(spa, feat));
1460 VERIFY0(zap_remove(dp->dp_meta_objset,
1461 DMU_POOL_DIRECTORY_OBJECT,
1462 DMU_POOL_BPTREE_OBJ, tx));
1463 VERIFY0(bptree_free(dp->dp_meta_objset,
1464 dp->dp_bptree_obj, tx));
1465 dp->dp_bptree_obj = 0;
1466 scn->scn_async_destroying = B_FALSE;
1469 if (scn->scn_visited_this_txg) {
1470 zfs_dbgmsg("freed %llu blocks in %llums from "
1471 "free_bpobj/bptree txg %llu",
1472 (longlong_t)scn->scn_visited_this_txg,
1474 (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1475 (longlong_t)tx->tx_txg);
1476 scn->scn_visited_this_txg = 0;
1478 * Re-sync the ddt so that we can further modify
1479 * it when doing bprewrite.
1481 ddt_sync(spa, tx->tx_txg);
1483 if (err == ERESTART)
1487 if (scn->scn_phys.scn_state != DSS_SCANNING)
1490 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1491 scn->scn_phys.scn_ddt_class_max) {
1492 zfs_dbgmsg("doing scan sync txg %llu; "
1493 "ddt bm=%llu/%llu/%llu/%llx",
1494 (longlong_t)tx->tx_txg,
1495 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1496 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1497 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1498 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1499 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1500 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1501 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1502 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1504 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1505 (longlong_t)tx->tx_txg,
1506 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1507 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1508 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1509 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1512 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1513 NULL, ZIO_FLAG_CANFAIL);
1514 dsl_pool_config_enter(dp, FTAG);
1515 dsl_scan_visit(scn, tx);
1516 dsl_pool_config_exit(dp, FTAG);
1517 (void) zio_wait(scn->scn_zio_root);
1518 scn->scn_zio_root = NULL;
1520 zfs_dbgmsg("visited %llu blocks in %llums",
1521 (longlong_t)scn->scn_visited_this_txg,
1522 (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1524 if (!scn->scn_pausing) {
1525 /* finished with scan. */
1526 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1527 dsl_scan_done(scn, B_TRUE, tx);
1530 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1531 mutex_enter(&spa->spa_scrub_lock);
1532 while (spa->spa_scrub_inflight > 0) {
1533 cv_wait(&spa->spa_scrub_io_cv,
1534 &spa->spa_scrub_lock);
1536 mutex_exit(&spa->spa_scrub_lock);
1539 dsl_scan_sync_state(scn, tx);
1543 * This will start a new scan, or restart an existing one.
1546 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1550 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1551 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1553 txg = dmu_tx_get_txg(tx);
1554 dp->dp_scan->scn_restart_txg = txg;
1557 dp->dp_scan->scn_restart_txg = txg;
1559 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1563 dsl_scan_resilvering(dsl_pool_t *dp)
1565 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1566 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1574 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1579 * If we resume after a reboot, zab will be NULL; don't record
1580 * incomplete stats in that case.
1585 for (i = 0; i < 4; i++) {
1586 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1587 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1588 if (t & DMU_OT_NEWTYPE)
1590 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1594 zb->zb_asize += BP_GET_ASIZE(bp);
1595 zb->zb_lsize += BP_GET_LSIZE(bp);
1596 zb->zb_psize += BP_GET_PSIZE(bp);
1597 zb->zb_gangs += BP_COUNT_GANG(bp);
1599 switch (BP_GET_NDVAS(bp)) {
1601 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1602 DVA_GET_VDEV(&bp->blk_dva[1]))
1603 zb->zb_ditto_2_of_2_samevdev++;
1606 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1607 DVA_GET_VDEV(&bp->blk_dva[1])) +
1608 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1609 DVA_GET_VDEV(&bp->blk_dva[2])) +
1610 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1611 DVA_GET_VDEV(&bp->blk_dva[2]));
1613 zb->zb_ditto_2_of_3_samevdev++;
1614 else if (equal == 3)
1615 zb->zb_ditto_3_of_3_samevdev++;
1622 dsl_scan_scrub_done(zio_t *zio)
1624 spa_t *spa = zio->io_spa;
1626 zio_data_buf_free(zio->io_data, zio->io_size);
1628 mutex_enter(&spa->spa_scrub_lock);
1629 spa->spa_scrub_inflight--;
1630 cv_broadcast(&spa->spa_scrub_io_cv);
1632 if (zio->io_error && (zio->io_error != ECKSUM ||
1633 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1634 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1636 mutex_exit(&spa->spa_scrub_lock);
1640 dsl_scan_scrub_cb(dsl_pool_t *dp,
1641 const blkptr_t *bp, const zbookmark_t *zb)
1643 dsl_scan_t *scn = dp->dp_scan;
1644 size_t size = BP_GET_PSIZE(bp);
1645 spa_t *spa = dp->dp_spa;
1646 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1648 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1650 unsigned int scan_delay = 0;
1652 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1653 phys_birth >= scn->scn_phys.scn_max_txg)
1656 count_block(dp->dp_blkstats, bp);
1658 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1659 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1660 zio_flags |= ZIO_FLAG_SCRUB;
1661 zio_priority = ZIO_PRIORITY_SCRUB;
1663 scan_delay = zfs_scrub_delay;
1665 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1666 zio_flags |= ZIO_FLAG_RESILVER;
1667 zio_priority = ZIO_PRIORITY_RESILVER;
1669 scan_delay = zfs_resilver_delay;
1672 /* If it's an intent log block, failure is expected. */
1673 if (zb->zb_level == ZB_ZIL_LEVEL)
1674 zio_flags |= ZIO_FLAG_SPECULATIVE;
1676 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1677 vdev_t *vd = vdev_lookup_top(spa,
1678 DVA_GET_VDEV(&bp->blk_dva[d]));
1681 * Keep track of how much data we've examined so that
1682 * zpool(1M) status can make useful progress reports.
1684 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1685 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1687 /* if it's a resilver, this may not be in the target range */
1689 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1691 * Gang members may be spread across multiple
1692 * vdevs, so the best estimate we have is the
1693 * scrub range, which has already been checked.
1694 * XXX -- it would be better to change our
1695 * allocation policy to ensure that all
1696 * gang members reside on the same vdev.
1700 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1706 if (needs_io && !zfs_no_scrub_io) {
1707 vdev_t *rvd = spa->spa_root_vdev;
1708 uint64_t maxinflight = rvd->vdev_children *
1709 MAX(zfs_top_maxinflight, 1);
1710 void *data = zio_data_buf_alloc(size);
1712 mutex_enter(&spa->spa_scrub_lock);
1713 while (spa->spa_scrub_inflight >= maxinflight)
1714 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1715 spa->spa_scrub_inflight++;
1716 mutex_exit(&spa->spa_scrub_lock);
1719 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1720 * then throttle our workload to limit the impact of a scan.
1722 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1723 delay(MAX((int)scan_delay, 0));
1725 zio_nowait(zio_read(NULL, spa, bp, data, size,
1726 dsl_scan_scrub_done, NULL, zio_priority,
1730 /* do not relocate this block */
1735 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1737 spa_t *spa = dp->dp_spa;
1740 * Purge all vdev caches and probe all devices. We do this here
1741 * rather than in sync context because this requires a writer lock
1742 * on the spa_config lock, which we can't do from sync context. The
1743 * spa_scrub_reopen flag indicates that vdev_open() should not
1744 * attempt to start another scrub.
1746 spa_vdev_state_enter(spa, SCL_NONE);
1747 spa->spa_scrub_reopen = B_TRUE;
1748 vdev_reopen(spa->spa_root_vdev);
1749 spa->spa_scrub_reopen = B_FALSE;
1750 (void) spa_vdev_state_exit(spa, NULL, 0);
1752 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1753 dsl_scan_setup_sync, &func, 0));