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) 2012 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);
128 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
129 "scrub_func", sizeof (uint64_t), 1, &f);
132 * There was an old-style scrub in progress. Restart a
133 * new-style scrub from the beginning.
135 scn->scn_restart_txg = txg;
136 zfs_dbgmsg("old-style scrub was in progress; "
137 "restarting new-style scrub in txg %llu",
138 scn->scn_restart_txg);
141 * Load the queue obj from the old location so that it
142 * can be freed by dsl_scan_done().
144 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
145 "scrub_queue", sizeof (uint64_t), 1,
146 &scn->scn_phys.scn_queue_obj);
148 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
149 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
156 if (scn->scn_phys.scn_state == DSS_SCANNING &&
157 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
159 * A new-type scrub was in progress on an old
160 * pool, and the pool was accessed by old
161 * software. Restart from the beginning, since
162 * the old software may have changed the pool in
165 scn->scn_restart_txg = txg;
166 zfs_dbgmsg("new-style scrub was modified "
167 "by old software; restarting in txg %llu",
168 scn->scn_restart_txg);
172 spa_scan_stat_init(spa);
177 dsl_scan_fini(dsl_pool_t *dp)
180 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
187 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
189 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
191 if (scn->scn_phys.scn_state == DSS_SCANNING)
198 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
200 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
201 pool_scan_func_t *funcp = arg;
202 dmu_object_type_t ot = 0;
203 dsl_pool_t *dp = scn->scn_dp;
204 spa_t *spa = dp->dp_spa;
206 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
207 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
208 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
209 scn->scn_phys.scn_func = *funcp;
210 scn->scn_phys.scn_state = DSS_SCANNING;
211 scn->scn_phys.scn_min_txg = 0;
212 scn->scn_phys.scn_max_txg = tx->tx_txg;
213 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
214 scn->scn_phys.scn_start_time = gethrestime_sec();
215 scn->scn_phys.scn_errors = 0;
216 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
217 scn->scn_restart_txg = 0;
218 spa_scan_stat_init(spa);
220 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
221 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
223 /* rewrite all disk labels */
224 vdev_config_dirty(spa->spa_root_vdev);
226 if (vdev_resilver_needed(spa->spa_root_vdev,
227 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
228 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
230 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
233 spa->spa_scrub_started = B_TRUE;
235 * If this is an incremental scrub, limit the DDT scrub phase
236 * to just the auto-ditto class (for correctness); the rest
237 * of the scrub should go faster using top-down pruning.
239 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
240 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
244 /* back to the generic stuff */
246 if (dp->dp_blkstats == NULL) {
248 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
250 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
252 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
253 ot = DMU_OT_ZAP_OTHER;
255 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
256 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
258 dsl_scan_sync_state(scn, tx);
260 spa_history_log_internal(spa, "scan setup", tx,
261 "func=%u mintxg=%llu maxtxg=%llu",
262 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
267 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
269 static const char *old_names[] = {
271 "scrub_ddt_bookmark",
272 "scrub_ddt_class_max",
281 dsl_pool_t *dp = scn->scn_dp;
282 spa_t *spa = dp->dp_spa;
285 /* Remove any remnants of an old-style scrub. */
286 for (i = 0; old_names[i]; i++) {
287 (void) zap_remove(dp->dp_meta_objset,
288 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
291 if (scn->scn_phys.scn_queue_obj != 0) {
292 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
293 scn->scn_phys.scn_queue_obj, tx));
294 scn->scn_phys.scn_queue_obj = 0;
298 * If we were "restarted" from a stopped state, don't bother
299 * with anything else.
301 if (scn->scn_phys.scn_state != DSS_SCANNING)
305 scn->scn_phys.scn_state = DSS_FINISHED;
307 scn->scn_phys.scn_state = DSS_CANCELED;
309 spa_history_log_internal(spa, "scan done", tx,
310 "complete=%u", complete);
312 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
313 mutex_enter(&spa->spa_scrub_lock);
314 while (spa->spa_scrub_inflight > 0) {
315 cv_wait(&spa->spa_scrub_io_cv,
316 &spa->spa_scrub_lock);
318 mutex_exit(&spa->spa_scrub_lock);
319 spa->spa_scrub_started = B_FALSE;
320 spa->spa_scrub_active = B_FALSE;
323 * If the scrub/resilver completed, update all DTLs to
324 * reflect this. Whether it succeeded or not, vacate
325 * all temporary scrub DTLs.
327 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
328 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
330 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
331 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
333 spa_errlog_rotate(spa);
336 * We may have finished replacing a device.
337 * Let the async thread assess this and handle the detach.
339 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
342 scn->scn_phys.scn_end_time = gethrestime_sec();
347 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
349 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
351 if (scn->scn_phys.scn_state != DSS_SCANNING)
358 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
360 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
362 dsl_scan_done(scn, B_FALSE, tx);
363 dsl_scan_sync_state(scn, tx);
367 dsl_scan_cancel(dsl_pool_t *dp)
369 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
370 dsl_scan_cancel_sync, NULL, 3));
373 static void dsl_scan_visitbp(blkptr_t *bp,
374 const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
375 dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
377 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
378 dmu_objset_type_t ostype,
379 dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
382 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
384 zio_free(dp->dp_spa, txg, bp);
388 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
390 ASSERT(dsl_pool_sync_context(dp));
391 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
396 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
398 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
399 if (dsl_dataset_is_snapshot(ds))
400 return (MIN(smt, ds->ds_phys->ds_creation_txg));
405 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
407 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
408 DMU_POOL_DIRECTORY_OBJECT,
409 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
410 &scn->scn_phys, tx));
414 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
416 uint64_t elapsed_nanosecs;
417 unsigned int mintime;
419 /* we never skip user/group accounting objects */
420 if (zb && (int64_t)zb->zb_object < 0)
423 if (scn->scn_pausing)
424 return (B_TRUE); /* we're already pausing */
426 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
427 return (B_FALSE); /* we're resuming */
429 /* We only know how to resume from level-0 blocks. */
430 if (zb && zb->zb_level != 0)
433 mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
434 zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
435 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
436 if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
437 (elapsed_nanosecs / MICROSEC > mintime &&
438 txg_sync_waiting(scn->scn_dp)) ||
439 spa_shutting_down(scn->scn_dp->dp_spa)) {
441 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
442 (longlong_t)zb->zb_objset,
443 (longlong_t)zb->zb_object,
444 (longlong_t)zb->zb_level,
445 (longlong_t)zb->zb_blkid);
446 scn->scn_phys.scn_bookmark = *zb;
448 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
449 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
450 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
451 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
452 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
453 scn->scn_pausing = B_TRUE;
459 typedef struct zil_scan_arg {
461 zil_header_t *zsa_zh;
466 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
468 zil_scan_arg_t *zsa = arg;
469 dsl_pool_t *dp = zsa->zsa_dp;
470 dsl_scan_t *scn = dp->dp_scan;
471 zil_header_t *zh = zsa->zsa_zh;
474 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
478 * One block ("stubby") can be allocated a long time ago; we
479 * want to visit that one because it has been allocated
480 * (on-disk) even if it hasn't been claimed (even though for
481 * scrub there's nothing to do to it).
483 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
486 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
487 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
489 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
495 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
497 if (lrc->lrc_txtype == TX_WRITE) {
498 zil_scan_arg_t *zsa = arg;
499 dsl_pool_t *dp = zsa->zsa_dp;
500 dsl_scan_t *scn = dp->dp_scan;
501 zil_header_t *zh = zsa->zsa_zh;
502 lr_write_t *lr = (lr_write_t *)lrc;
503 blkptr_t *bp = &lr->lr_blkptr;
506 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
510 * birth can be < claim_txg if this record's txg is
511 * already txg sync'ed (but this log block contains
512 * other records that are not synced)
514 if (claim_txg == 0 || bp->blk_birth < claim_txg)
517 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
518 lr->lr_foid, ZB_ZIL_LEVEL,
519 lr->lr_offset / BP_GET_LSIZE(bp));
521 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
527 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
529 uint64_t claim_txg = zh->zh_claim_txg;
530 zil_scan_arg_t zsa = { dp, zh };
534 * We only want to visit blocks that have been claimed but not yet
535 * replayed (or, in read-only mode, blocks that *would* be claimed).
537 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
540 zilog = zil_alloc(dp->dp_meta_objset, zh);
542 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
550 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
551 uint64_t objset, uint64_t object, uint64_t blkid)
554 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
556 if (zfs_no_scrub_prefetch)
559 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
560 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
563 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
565 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
566 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
567 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
571 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
572 const zbookmark_t *zb)
575 * We never skip over user/group accounting objects (obj<0)
577 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
578 (int64_t)zb->zb_object >= 0) {
580 * If we already visited this bp & everything below (in
581 * a prior txg sync), don't bother doing it again.
583 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
587 * If we found the block we're trying to resume from, or
588 * we went past it to a different object, zero it out to
589 * indicate that it's OK to start checking for pausing
592 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
593 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
594 dprintf("resuming at %llx/%llx/%llx/%llx\n",
595 (longlong_t)zb->zb_objset,
596 (longlong_t)zb->zb_object,
597 (longlong_t)zb->zb_level,
598 (longlong_t)zb->zb_blkid);
599 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
606 * Return nonzero on i/o error.
607 * Return new buf to write out in *bufp.
610 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
611 dnode_phys_t *dnp, const blkptr_t *bp,
612 const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
614 dsl_pool_t *dp = scn->scn_dp;
615 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
618 if (BP_GET_LEVEL(bp) > 0) {
619 uint32_t flags = ARC_WAIT;
622 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
624 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
625 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
627 scn->scn_phys.scn_errors++;
630 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
631 dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
632 zb->zb_object, zb->zb_blkid * epb + i);
634 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
637 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
639 zb->zb_blkid * epb + i);
640 dsl_scan_visitbp(cbp, &czb, dnp,
641 *bufp, ds, scn, ostype, tx);
643 } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
644 uint32_t flags = ARC_WAIT;
646 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
647 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
649 scn->scn_phys.scn_errors++;
652 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
653 uint32_t flags = ARC_WAIT;
656 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
658 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
659 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
661 scn->scn_phys.scn_errors++;
664 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
665 for (j = 0; j < cdnp->dn_nblkptr; j++) {
666 blkptr_t *cbp = &cdnp->dn_blkptr[j];
667 dsl_scan_prefetch(scn, *bufp, cbp,
668 zb->zb_objset, zb->zb_blkid * epb + i, j);
671 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
672 dsl_scan_visitdnode(scn, ds, ostype,
673 cdnp, *bufp, zb->zb_blkid * epb + i, tx);
676 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
677 uint32_t flags = ARC_WAIT;
680 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
681 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
683 scn->scn_phys.scn_errors++;
687 osp = (*bufp)->b_data;
689 dsl_scan_visitdnode(scn, ds, osp->os_type,
690 &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
692 if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
694 * We also always visit user/group accounting
695 * objects, and never skip them, even if we are
696 * pausing. This is necessary so that the space
697 * deltas from this txg get integrated.
699 dsl_scan_visitdnode(scn, ds, osp->os_type,
700 &osp->os_groupused_dnode, *bufp,
701 DMU_GROUPUSED_OBJECT, tx);
702 dsl_scan_visitdnode(scn, ds, osp->os_type,
703 &osp->os_userused_dnode, *bufp,
704 DMU_USERUSED_OBJECT, tx);
712 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
713 dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
714 uint64_t object, dmu_tx_t *tx)
718 for (j = 0; j < dnp->dn_nblkptr; j++) {
721 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
722 dnp->dn_nlevels - 1, j);
723 dsl_scan_visitbp(&dnp->dn_blkptr[j],
724 &czb, dnp, buf, ds, scn, ostype, tx);
727 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
729 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
731 dsl_scan_visitbp(&dnp->dn_spill,
732 &czb, dnp, buf, ds, scn, ostype, tx);
737 * The arguments are in this order because mdb can only print the
738 * first 5; we want them to be useful.
741 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
742 dnode_phys_t *dnp, arc_buf_t *pbuf,
743 dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
746 dsl_pool_t *dp = scn->scn_dp;
747 arc_buf_t *buf = NULL;
748 blkptr_t bp_toread = *bp;
750 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
752 if (dsl_scan_check_pause(scn, zb))
755 if (dsl_scan_check_resume(scn, dnp, zb))
758 if (bp->blk_birth == 0)
761 scn->scn_visited_this_txg++;
764 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
765 ds, ds ? ds->ds_object : 0,
766 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
769 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
772 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
777 * If dsl_scan_ddt() has aready visited this block, it will have
778 * already done any translations or scrubbing, so don't call the
781 if (ddt_class_contains(dp->dp_spa,
782 scn->scn_phys.scn_ddt_class_max, bp)) {
788 * If this block is from the future (after cur_max_txg), then we
789 * are doing this on behalf of a deleted snapshot, and we will
790 * revisit the future block on the next pass of this dataset.
791 * Don't scan it now unless we need to because something
792 * under it was modified.
794 if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
795 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
798 (void) arc_buf_remove_ref(buf, &buf);
802 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
807 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
808 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
809 dsl_scan_visitbp(bp, &zb, NULL, NULL,
810 ds, scn, DMU_OST_NONE, tx);
812 dprintf_ds(ds, "finished scan%s", "");
816 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
818 dsl_pool_t *dp = ds->ds_dir->dd_pool;
819 dsl_scan_t *scn = dp->dp_scan;
822 if (scn->scn_phys.scn_state != DSS_SCANNING)
825 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
826 if (dsl_dataset_is_snapshot(ds)) {
827 /* Note, scn_cur_{min,max}_txg stays the same. */
828 scn->scn_phys.scn_bookmark.zb_objset =
829 ds->ds_phys->ds_next_snap_obj;
830 zfs_dbgmsg("destroying ds %llu; currently traversing; "
831 "reset zb_objset to %llu",
832 (u_longlong_t)ds->ds_object,
833 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
834 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
836 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
837 ZB_DESTROYED_OBJSET, 0, 0, 0);
838 zfs_dbgmsg("destroying ds %llu; currently traversing; "
839 "reset bookmark to -1,0,0,0",
840 (u_longlong_t)ds->ds_object);
842 } else if (zap_lookup_int_key(dp->dp_meta_objset,
843 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
844 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
845 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
846 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
847 if (dsl_dataset_is_snapshot(ds)) {
849 * We keep the same mintxg; it could be >
850 * ds_creation_txg if the previous snapshot was
853 VERIFY(zap_add_int_key(dp->dp_meta_objset,
854 scn->scn_phys.scn_queue_obj,
855 ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
856 zfs_dbgmsg("destroying ds %llu; in queue; "
857 "replacing with %llu",
858 (u_longlong_t)ds->ds_object,
859 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
861 zfs_dbgmsg("destroying ds %llu; in queue; removing",
862 (u_longlong_t)ds->ds_object);
865 zfs_dbgmsg("destroying ds %llu; ignoring",
866 (u_longlong_t)ds->ds_object);
870 * dsl_scan_sync() should be called after this, and should sync
871 * out our changed state, but just to be safe, do it here.
873 dsl_scan_sync_state(scn, tx);
877 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
879 dsl_pool_t *dp = ds->ds_dir->dd_pool;
880 dsl_scan_t *scn = dp->dp_scan;
883 if (scn->scn_phys.scn_state != DSS_SCANNING)
886 ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
888 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
889 scn->scn_phys.scn_bookmark.zb_objset =
890 ds->ds_phys->ds_prev_snap_obj;
891 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
892 "reset zb_objset to %llu",
893 (u_longlong_t)ds->ds_object,
894 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
895 } else if (zap_lookup_int_key(dp->dp_meta_objset,
896 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
897 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
898 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
899 VERIFY(zap_add_int_key(dp->dp_meta_objset,
900 scn->scn_phys.scn_queue_obj,
901 ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
902 zfs_dbgmsg("snapshotting ds %llu; in queue; "
903 "replacing with %llu",
904 (u_longlong_t)ds->ds_object,
905 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
907 dsl_scan_sync_state(scn, tx);
911 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
913 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
914 dsl_scan_t *scn = dp->dp_scan;
917 if (scn->scn_phys.scn_state != DSS_SCANNING)
920 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
921 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
922 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
923 "reset zb_objset to %llu",
924 (u_longlong_t)ds1->ds_object,
925 (u_longlong_t)ds2->ds_object);
926 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
927 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
928 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
929 "reset zb_objset to %llu",
930 (u_longlong_t)ds2->ds_object,
931 (u_longlong_t)ds1->ds_object);
934 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
935 ds1->ds_object, &mintxg) == 0) {
938 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
939 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
940 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
941 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
942 err = zap_add_int_key(dp->dp_meta_objset,
943 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
944 VERIFY(err == 0 || err == EEXIST);
946 /* Both were there to begin with */
947 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
948 scn->scn_phys.scn_queue_obj,
949 ds1->ds_object, mintxg, tx));
951 zfs_dbgmsg("clone_swap ds %llu; in queue; "
952 "replacing with %llu",
953 (u_longlong_t)ds1->ds_object,
954 (u_longlong_t)ds2->ds_object);
955 } else if (zap_lookup_int_key(dp->dp_meta_objset,
956 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
957 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
958 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
959 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
960 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
961 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
962 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
963 zfs_dbgmsg("clone_swap ds %llu; in queue; "
964 "replacing with %llu",
965 (u_longlong_t)ds2->ds_object,
966 (u_longlong_t)ds1->ds_object);
969 dsl_scan_sync_state(scn, tx);
972 struct enqueue_clones_arg {
979 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
981 struct enqueue_clones_arg *eca = arg;
984 dsl_scan_t *scn = dp->dp_scan;
986 if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
989 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
993 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
995 err = dsl_dataset_hold_obj(dp,
996 ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
998 dsl_dataset_rele(ds, FTAG);
1003 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1004 scn->scn_phys.scn_queue_obj, ds->ds_object,
1005 ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1006 dsl_dataset_rele(ds, FTAG);
1011 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1013 dsl_pool_t *dp = scn->scn_dp;
1017 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1019 if (dmu_objset_from_ds(ds, &os))
1023 * Only the ZIL in the head (non-snapshot) is valid. Even though
1024 * snapshots can have ZIL block pointers (which may be the same
1025 * BP as in the head), they must be ignored. So we traverse the
1026 * ZIL here, rather than in scan_recurse(), because the regular
1027 * snapshot block-sharing rules don't apply to it.
1029 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1030 dsl_scan_zil(dp, &os->os_zil_header);
1033 * Iterate over the bps in this ds.
1035 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1036 dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1038 char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1039 dsl_dataset_name(ds, dsname);
1040 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1042 (longlong_t)dsobj, dsname,
1043 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1044 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1045 (int)scn->scn_pausing);
1046 kmem_free(dsname, ZFS_MAXNAMELEN);
1048 if (scn->scn_pausing)
1052 * We've finished this pass over this dataset.
1056 * If we did not completely visit this dataset, do another pass.
1058 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1059 zfs_dbgmsg("incomplete pass; visiting again");
1060 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1061 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1062 scn->scn_phys.scn_queue_obj, ds->ds_object,
1063 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1068 * Add descendent datasets to work queue.
1070 if (ds->ds_phys->ds_next_snap_obj != 0) {
1071 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1072 scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1073 ds->ds_phys->ds_creation_txg, tx) == 0);
1075 if (ds->ds_phys->ds_num_children > 1) {
1076 boolean_t usenext = B_FALSE;
1077 if (ds->ds_phys->ds_next_clones_obj != 0) {
1080 * A bug in a previous version of the code could
1081 * cause upgrade_clones_cb() to not set
1082 * ds_next_snap_obj when it should, leading to a
1083 * missing entry. Therefore we can only use the
1084 * next_clones_obj when its count is correct.
1086 int err = zap_count(dp->dp_meta_objset,
1087 ds->ds_phys->ds_next_clones_obj, &count);
1089 count == ds->ds_phys->ds_num_children - 1)
1094 VERIFY0(zap_join_key(dp->dp_meta_objset,
1095 ds->ds_phys->ds_next_clones_obj,
1096 scn->scn_phys.scn_queue_obj,
1097 ds->ds_phys->ds_creation_txg, tx));
1099 struct enqueue_clones_arg eca;
1101 eca.originobj = ds->ds_object;
1103 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1104 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1109 dsl_dataset_rele(ds, FTAG);
1114 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1119 dsl_scan_t *scn = dp->dp_scan;
1121 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1125 while (ds->ds_phys->ds_prev_snap_obj != 0) {
1126 dsl_dataset_t *prev;
1127 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1130 dsl_dataset_rele(ds, FTAG);
1135 * If this is a clone, we don't need to worry about it for now.
1137 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1138 dsl_dataset_rele(ds, FTAG);
1139 dsl_dataset_rele(prev, FTAG);
1142 dsl_dataset_rele(ds, FTAG);
1146 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1147 ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1148 dsl_dataset_rele(ds, FTAG);
1153 * Scrub/dedup interaction.
1155 * If there are N references to a deduped block, we don't want to scrub it
1156 * N times -- ideally, we should scrub it exactly once.
1158 * We leverage the fact that the dde's replication class (enum ddt_class)
1159 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1160 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1162 * To prevent excess scrubbing, the scrub begins by walking the DDT
1163 * to find all blocks with refcnt > 1, and scrubs each of these once.
1164 * Since there are two replication classes which contain blocks with
1165 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1166 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1168 * There would be nothing more to say if a block's refcnt couldn't change
1169 * during a scrub, but of course it can so we must account for changes
1170 * in a block's replication class.
1172 * Here's an example of what can occur:
1174 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1175 * when visited during the top-down scrub phase, it will be scrubbed twice.
1176 * This negates our scrub optimization, but is otherwise harmless.
1178 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1179 * on each visit during the top-down scrub phase, it will never be scrubbed.
1180 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1181 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1182 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1183 * while a scrub is in progress, it scrubs the block right then.
1186 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1188 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1189 ddt_entry_t dde = { 0 };
1193 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1196 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1198 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1199 (longlong_t)ddb->ddb_class,
1200 (longlong_t)ddb->ddb_type,
1201 (longlong_t)ddb->ddb_checksum,
1202 (longlong_t)ddb->ddb_cursor);
1204 /* There should be no pending changes to the dedup table */
1205 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1206 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1208 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1211 if (dsl_scan_check_pause(scn, NULL))
1215 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1216 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1217 (int)scn->scn_pausing);
1219 ASSERT(error == 0 || error == ENOENT);
1220 ASSERT(error != ENOENT ||
1221 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1226 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1227 ddt_entry_t *dde, dmu_tx_t *tx)
1229 const ddt_key_t *ddk = &dde->dde_key;
1230 ddt_phys_t *ddp = dde->dde_phys;
1232 zbookmark_t zb = { 0 };
1234 if (scn->scn_phys.scn_state != DSS_SCANNING)
1237 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1238 if (ddp->ddp_phys_birth == 0 ||
1239 ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1241 ddt_bp_create(checksum, ddk, ddp, &bp);
1243 scn->scn_visited_this_txg++;
1244 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1249 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1251 dsl_pool_t *dp = scn->scn_dp;
1255 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1256 scn->scn_phys.scn_ddt_class_max) {
1257 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1258 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1259 dsl_scan_ddt(scn, tx);
1260 if (scn->scn_pausing)
1264 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1265 /* First do the MOS & ORIGIN */
1267 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1268 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1269 dsl_scan_visit_rootbp(scn, NULL,
1270 &dp->dp_meta_rootbp, tx);
1271 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1272 if (scn->scn_pausing)
1275 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1276 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1277 enqueue_cb, tx, DS_FIND_CHILDREN));
1279 dsl_scan_visitds(scn,
1280 dp->dp_origin_snap->ds_object, tx);
1282 ASSERT(!scn->scn_pausing);
1283 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1284 ZB_DESTROYED_OBJSET) {
1286 * If we were paused, continue from here. Note if the
1287 * ds we were paused on was deleted, the zb_objset may
1288 * be -1, so we will skip this and find a new objset
1291 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1292 if (scn->scn_pausing)
1297 * In case we were paused right at the end of the ds, zero the
1298 * bookmark so we don't think that we're still trying to resume.
1300 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1302 /* keep pulling things out of the zap-object-as-queue */
1303 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1304 scn->scn_phys.scn_queue_obj),
1305 zap_cursor_retrieve(&zc, &za) == 0) {
1309 dsobj = strtonum(za.za_name, NULL);
1310 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1311 scn->scn_phys.scn_queue_obj, dsobj, tx));
1313 /* Set up min/max txg */
1314 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1315 if (za.za_first_integer != 0) {
1316 scn->scn_phys.scn_cur_min_txg =
1317 MAX(scn->scn_phys.scn_min_txg,
1318 za.za_first_integer);
1320 scn->scn_phys.scn_cur_min_txg =
1321 MAX(scn->scn_phys.scn_min_txg,
1322 ds->ds_phys->ds_prev_snap_txg);
1324 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1325 dsl_dataset_rele(ds, FTAG);
1327 dsl_scan_visitds(scn, dsobj, tx);
1328 zap_cursor_fini(&zc);
1329 if (scn->scn_pausing)
1332 zap_cursor_fini(&zc);
1336 dsl_scan_free_should_pause(dsl_scan_t *scn)
1338 uint64_t elapsed_nanosecs;
1340 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1341 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1342 (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1343 txg_sync_waiting(scn->scn_dp)) ||
1344 spa_shutting_down(scn->scn_dp->dp_spa));
1348 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1350 dsl_scan_t *scn = arg;
1352 if (!scn->scn_is_bptree ||
1353 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1354 if (dsl_scan_free_should_pause(scn))
1358 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1359 dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1360 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1361 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1362 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1363 scn->scn_visited_this_txg++;
1368 dsl_scan_active(dsl_scan_t *scn)
1370 spa_t *spa = scn->scn_dp->dp_spa;
1371 uint64_t used = 0, comp, uncomp;
1373 if (spa->spa_load_state != SPA_LOAD_NONE)
1375 if (spa_shutting_down(spa))
1378 if (scn->scn_phys.scn_state == DSS_SCANNING)
1381 if (spa_feature_is_active(spa,
1382 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1385 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1386 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1387 &used, &comp, &uncomp);
1393 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1395 dsl_scan_t *scn = dp->dp_scan;
1396 spa_t *spa = dp->dp_spa;
1400 * Check for scn_restart_txg before checking spa_load_state, so
1401 * that we can restart an old-style scan while the pool is being
1402 * imported (see dsl_scan_init).
1404 if (scn->scn_restart_txg != 0 &&
1405 scn->scn_restart_txg <= tx->tx_txg) {
1406 pool_scan_func_t func = POOL_SCAN_SCRUB;
1407 dsl_scan_done(scn, B_FALSE, tx);
1408 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1409 func = POOL_SCAN_RESILVER;
1410 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1412 dsl_scan_setup_sync(&func, tx);
1415 if (!dsl_scan_active(scn) ||
1416 spa_sync_pass(dp->dp_spa) > 1)
1419 scn->scn_visited_this_txg = 0;
1420 scn->scn_pausing = B_FALSE;
1421 scn->scn_sync_start_time = gethrtime();
1422 spa->spa_scrub_active = B_TRUE;
1425 * First process the free list. If we pause the free, don't do
1426 * any scanning. This ensures that there is no free list when
1427 * we are scanning, so the scan code doesn't have to worry about
1430 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1431 scn->scn_is_bptree = B_FALSE;
1432 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1433 NULL, ZIO_FLAG_MUSTSUCCEED);
1434 err = bpobj_iterate(&dp->dp_free_bpobj,
1435 dsl_scan_free_block_cb, scn, tx);
1436 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1438 if (err == 0 && spa_feature_is_active(spa,
1439 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1440 scn->scn_is_bptree = B_TRUE;
1441 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1442 NULL, ZIO_FLAG_MUSTSUCCEED);
1443 err = bptree_iterate(dp->dp_meta_objset,
1444 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1446 VERIFY0(zio_wait(scn->scn_zio_root));
1449 zfeature_info_t *feat = &spa_feature_table
1450 [SPA_FEATURE_ASYNC_DESTROY];
1451 /* finished; deactivate async destroy feature */
1452 spa_feature_decr(spa, feat, tx);
1453 ASSERT(!spa_feature_is_active(spa, feat));
1454 VERIFY0(zap_remove(dp->dp_meta_objset,
1455 DMU_POOL_DIRECTORY_OBJECT,
1456 DMU_POOL_BPTREE_OBJ, tx));
1457 VERIFY0(bptree_free(dp->dp_meta_objset,
1458 dp->dp_bptree_obj, tx));
1459 dp->dp_bptree_obj = 0;
1462 if (scn->scn_visited_this_txg) {
1463 zfs_dbgmsg("freed %llu blocks in %llums from "
1464 "free_bpobj/bptree txg %llu",
1465 (longlong_t)scn->scn_visited_this_txg,
1467 (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1468 (longlong_t)tx->tx_txg);
1469 scn->scn_visited_this_txg = 0;
1471 * Re-sync the ddt so that we can further modify
1472 * it when doing bprewrite.
1474 ddt_sync(spa, tx->tx_txg);
1476 if (err == ERESTART)
1480 if (scn->scn_phys.scn_state != DSS_SCANNING)
1483 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1484 scn->scn_phys.scn_ddt_class_max) {
1485 zfs_dbgmsg("doing scan sync txg %llu; "
1486 "ddt bm=%llu/%llu/%llu/%llx",
1487 (longlong_t)tx->tx_txg,
1488 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1489 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1490 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1491 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1492 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1493 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1494 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1495 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1497 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1498 (longlong_t)tx->tx_txg,
1499 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1500 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1501 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1502 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1505 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1506 NULL, ZIO_FLAG_CANFAIL);
1507 dsl_pool_config_enter(dp, FTAG);
1508 dsl_scan_visit(scn, tx);
1509 dsl_pool_config_exit(dp, FTAG);
1510 (void) zio_wait(scn->scn_zio_root);
1511 scn->scn_zio_root = NULL;
1513 zfs_dbgmsg("visited %llu blocks in %llums",
1514 (longlong_t)scn->scn_visited_this_txg,
1515 (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1517 if (!scn->scn_pausing) {
1518 /* finished with scan. */
1519 zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1520 dsl_scan_done(scn, B_TRUE, tx);
1523 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1524 mutex_enter(&spa->spa_scrub_lock);
1525 while (spa->spa_scrub_inflight > 0) {
1526 cv_wait(&spa->spa_scrub_io_cv,
1527 &spa->spa_scrub_lock);
1529 mutex_exit(&spa->spa_scrub_lock);
1532 dsl_scan_sync_state(scn, tx);
1536 * This will start a new scan, or restart an existing one.
1539 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1543 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1544 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1546 txg = dmu_tx_get_txg(tx);
1547 dp->dp_scan->scn_restart_txg = txg;
1550 dp->dp_scan->scn_restart_txg = txg;
1552 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1556 dsl_scan_resilvering(dsl_pool_t *dp)
1558 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1559 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1567 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1572 * If we resume after a reboot, zab will be NULL; don't record
1573 * incomplete stats in that case.
1578 for (i = 0; i < 4; i++) {
1579 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1580 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1581 if (t & DMU_OT_NEWTYPE)
1583 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1587 zb->zb_asize += BP_GET_ASIZE(bp);
1588 zb->zb_lsize += BP_GET_LSIZE(bp);
1589 zb->zb_psize += BP_GET_PSIZE(bp);
1590 zb->zb_gangs += BP_COUNT_GANG(bp);
1592 switch (BP_GET_NDVAS(bp)) {
1594 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1595 DVA_GET_VDEV(&bp->blk_dva[1]))
1596 zb->zb_ditto_2_of_2_samevdev++;
1599 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1600 DVA_GET_VDEV(&bp->blk_dva[1])) +
1601 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1602 DVA_GET_VDEV(&bp->blk_dva[2])) +
1603 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1604 DVA_GET_VDEV(&bp->blk_dva[2]));
1606 zb->zb_ditto_2_of_3_samevdev++;
1607 else if (equal == 3)
1608 zb->zb_ditto_3_of_3_samevdev++;
1615 dsl_scan_scrub_done(zio_t *zio)
1617 spa_t *spa = zio->io_spa;
1619 zio_data_buf_free(zio->io_data, zio->io_size);
1621 mutex_enter(&spa->spa_scrub_lock);
1622 spa->spa_scrub_inflight--;
1623 cv_broadcast(&spa->spa_scrub_io_cv);
1625 if (zio->io_error && (zio->io_error != ECKSUM ||
1626 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1627 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1629 mutex_exit(&spa->spa_scrub_lock);
1633 dsl_scan_scrub_cb(dsl_pool_t *dp,
1634 const blkptr_t *bp, const zbookmark_t *zb)
1636 dsl_scan_t *scn = dp->dp_scan;
1637 size_t size = BP_GET_PSIZE(bp);
1638 spa_t *spa = dp->dp_spa;
1639 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1641 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1643 unsigned int scan_delay = 0;
1645 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1646 phys_birth >= scn->scn_phys.scn_max_txg)
1649 count_block(dp->dp_blkstats, bp);
1651 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1652 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1653 zio_flags |= ZIO_FLAG_SCRUB;
1654 zio_priority = ZIO_PRIORITY_SCRUB;
1656 scan_delay = zfs_scrub_delay;
1658 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1659 zio_flags |= ZIO_FLAG_RESILVER;
1660 zio_priority = ZIO_PRIORITY_RESILVER;
1662 scan_delay = zfs_resilver_delay;
1665 /* If it's an intent log block, failure is expected. */
1666 if (zb->zb_level == ZB_ZIL_LEVEL)
1667 zio_flags |= ZIO_FLAG_SPECULATIVE;
1669 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1670 vdev_t *vd = vdev_lookup_top(spa,
1671 DVA_GET_VDEV(&bp->blk_dva[d]));
1674 * Keep track of how much data we've examined so that
1675 * zpool(1M) status can make useful progress reports.
1677 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1678 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1680 /* if it's a resilver, this may not be in the target range */
1682 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1684 * Gang members may be spread across multiple
1685 * vdevs, so the best estimate we have is the
1686 * scrub range, which has already been checked.
1687 * XXX -- it would be better to change our
1688 * allocation policy to ensure that all
1689 * gang members reside on the same vdev.
1693 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1699 if (needs_io && !zfs_no_scrub_io) {
1700 vdev_t *rvd = spa->spa_root_vdev;
1701 uint64_t maxinflight = rvd->vdev_children *
1702 MAX(zfs_top_maxinflight, 1);
1703 void *data = zio_data_buf_alloc(size);
1705 mutex_enter(&spa->spa_scrub_lock);
1706 while (spa->spa_scrub_inflight >= maxinflight)
1707 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1708 spa->spa_scrub_inflight++;
1709 mutex_exit(&spa->spa_scrub_lock);
1712 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1713 * then throttle our workload to limit the impact of a scan.
1715 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1716 delay(MAX((int)scan_delay, 0));
1718 zio_nowait(zio_read(NULL, spa, bp, data, size,
1719 dsl_scan_scrub_done, NULL, zio_priority,
1723 /* do not relocate this block */
1728 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1730 spa_t *spa = dp->dp_spa;
1733 * Purge all vdev caches and probe all devices. We do this here
1734 * rather than in sync context because this requires a writer lock
1735 * on the spa_config lock, which we can't do from sync context. The
1736 * spa_scrub_reopen flag indicates that vdev_open() should not
1737 * attempt to start another scrub.
1739 spa_vdev_state_enter(spa, SCL_NONE);
1740 spa->spa_scrub_reopen = B_TRUE;
1741 vdev_reopen(spa->spa_root_vdev);
1742 spa->spa_scrub_reopen = B_FALSE;
1743 (void) spa_vdev_state_exit(spa, NULL, 0);
1745 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1746 dsl_scan_setup_sync, &func, 0));