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) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright 2016 Gary Mills
25 * Copyright (c) 2017 Datto Inc.
26 * Copyright 2019 Joyent, Inc.
29 #include <sys/dsl_scan.h>
30 #include <sys/dsl_pool.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_prop.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_synctask.h>
35 #include <sys/dnode.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/dmu_objset.h>
41 #include <sys/zfs_context.h>
42 #include <sys/fs/zfs.h>
43 #include <sys/zfs_znode.h>
44 #include <sys/spa_impl.h>
45 #include <sys/vdev_impl.h>
46 #include <sys/zil_impl.h>
47 #include <sys/zio_checksum.h>
50 #include <sys/sa_impl.h>
51 #include <sys/zfeature.h>
53 #include <sys/range_tree.h>
55 #include <sys/zfs_vfsops.h>
59 * Grand theory statement on scan queue sorting
61 * Scanning is implemented by recursively traversing all indirection levels
62 * in an object and reading all blocks referenced from said objects. This
63 * results in us approximately traversing the object from lowest logical
64 * offset to the highest. For best performance, we would want the logical
65 * blocks to be physically contiguous. However, this is frequently not the
66 * case with pools given the allocation patterns of copy-on-write filesystems.
67 * So instead, we put the I/Os into a reordering queue and issue them in a
68 * way that will most benefit physical disks (LBA-order).
72 * Ideally, we would want to scan all metadata and queue up all block I/O
73 * prior to starting to issue it, because that allows us to do an optimal
74 * sorting job. This can however consume large amounts of memory. Therefore
75 * we continuously monitor the size of the queues and constrain them to 5%
76 * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this
77 * limit, we clear out a few of the largest extents at the head of the queues
78 * to make room for more scanning. Hopefully, these extents will be fairly
79 * large and contiguous, allowing us to approach sequential I/O throughput
80 * even without a fully sorted tree.
82 * Metadata scanning takes place in dsl_scan_visit(), which is called from
83 * dsl_scan_sync() every spa_sync(). If we have either fully scanned all
84 * metadata on the pool, or we need to make room in memory because our
85 * queues are too large, dsl_scan_visit() is postponed and
86 * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies
87 * that metadata scanning and queued I/O issuing are mutually exclusive. This
88 * allows us to provide maximum sequential I/O throughput for the majority of
89 * I/O's issued since sequential I/O performance is significantly negatively
90 * impacted if it is interleaved with random I/O.
92 * Implementation Notes
94 * One side effect of the queued scanning algorithm is that the scanning code
95 * needs to be notified whenever a block is freed. This is needed to allow
96 * the scanning code to remove these I/Os from the issuing queue. Additionally,
97 * we do not attempt to queue gang blocks to be issued sequentially since this
98 * is very hard to do and would have an extremely limited performance benefit.
99 * Instead, we simply issue gang I/Os as soon as we find them using the legacy
102 * Backwards compatibility
104 * This new algorithm is backwards compatible with the legacy on-disk data
105 * structures (and therefore does not require a new feature flag).
106 * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan
107 * will stop scanning metadata (in logical order) and wait for all outstanding
108 * sorted I/O to complete. Once this is done, we write out a checkpoint
109 * bookmark, indicating that we have scanned everything logically before it.
110 * If the pool is imported on a machine without the new sorting algorithm,
111 * the scan simply resumes from the last checkpoint using the legacy algorithm.
114 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
115 const zbookmark_phys_t *);
117 static scan_cb_t dsl_scan_scrub_cb;
119 static int scan_ds_queue_compare(const void *a, const void *b);
120 static int scan_prefetch_queue_compare(const void *a, const void *b);
121 static void scan_ds_queue_clear(dsl_scan_t *scn);
122 static void scan_ds_prefetch_queue_clear(dsl_scan_t *scn);
123 static boolean_t scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj,
125 static void scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg);
126 static void scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj);
127 static void scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx);
128 static uint64_t dsl_scan_count_leaves(vdev_t *vd);
130 extern int zfs_vdev_async_write_active_min_dirty_percent;
133 * By default zfs will check to ensure it is not over the hard memory
134 * limit before each txg. If finer-grained control of this is needed
135 * this value can be set to 1 to enable checking before scanning each
138 int zfs_scan_strict_mem_lim = B_FALSE;
141 * Maximum number of parallelly executed bytes per leaf vdev. We attempt
142 * to strike a balance here between keeping the vdev queues full of I/Os
143 * at all times and not overflowing the queues to cause long latency,
144 * which would cause long txg sync times. No matter what, we will not
145 * overload the drives with I/O, since that is protected by
146 * zfs_vdev_scrub_max_active.
148 unsigned long zfs_scan_vdev_limit = 4 << 20;
150 int zfs_scan_issue_strategy = 0;
151 int zfs_scan_legacy = B_FALSE; /* don't queue & sort zios, go direct */
152 unsigned long zfs_scan_max_ext_gap = 2 << 20; /* in bytes */
155 * fill_weight is non-tunable at runtime, so we copy it at module init from
156 * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
157 * break queue sorting.
159 int zfs_scan_fill_weight = 3;
160 static uint64_t fill_weight;
162 /* See dsl_scan_should_clear() for details on the memory limit tunables */
163 uint64_t zfs_scan_mem_lim_min = 16 << 20; /* bytes */
164 uint64_t zfs_scan_mem_lim_soft_max = 128 << 20; /* bytes */
165 int zfs_scan_mem_lim_fact = 20; /* fraction of physmem */
166 int zfs_scan_mem_lim_soft_fact = 20; /* fraction of mem lim above */
168 int zfs_scrub_min_time_ms = 1000; /* min millisecs to scrub per txg */
169 int zfs_obsolete_min_time_ms = 500; /* min millisecs to obsolete per txg */
170 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
171 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
172 int zfs_scan_checkpoint_intval = 7200; /* in seconds */
173 int zfs_scan_suspend_progress = 0; /* set to prevent scans from progressing */
174 int zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
175 int zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
176 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
177 /* max number of blocks to free in a single TXG */
178 unsigned long zfs_async_block_max_blocks = 100000;
180 int zfs_resilver_disable_defer = 0; /* set to disable resilver deferring */
183 * We wait a few txgs after importing a pool to begin scanning so that
184 * the import / mounting code isn't held up by scrub / resilver IO.
185 * Unfortunately, it is a bit difficult to determine exactly how long
186 * this will take since userspace will trigger fs mounts asynchronously
187 * and the kernel will create zvol minors asynchronously. As a result,
188 * the value provided here is a bit arbitrary, but represents a
189 * reasonable estimate of how many txgs it will take to finish fully
192 #define SCAN_IMPORT_WAIT_TXGS 5
194 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
195 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
196 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
199 * Enable/disable the processing of the free_bpobj object.
201 int zfs_free_bpobj_enabled = 1;
203 /* the order has to match pool_scan_type */
204 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
206 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
207 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
210 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */
218 * This controls what conditions are placed on dsl_scan_sync_state():
219 * SYNC_OPTIONAL) write out scn_phys iff scn_bytes_pending == 0
220 * SYNC_MANDATORY) write out scn_phys always. scn_bytes_pending must be 0.
221 * SYNC_CACHED) if scn_bytes_pending == 0, write out scn_phys. Otherwise
222 * write out the scn_phys_cached version.
223 * See dsl_scan_sync_state for details.
232 * This struct represents the minimum information needed to reconstruct a
233 * zio for sequential scanning. This is useful because many of these will
234 * accumulate in the sequential IO queues before being issued, so saving
235 * memory matters here.
237 typedef struct scan_io {
238 /* fields from blkptr_t */
239 uint64_t sio_blk_prop;
240 uint64_t sio_phys_birth;
242 zio_cksum_t sio_cksum;
243 uint32_t sio_nr_dvas;
245 /* fields from zio_t */
247 zbookmark_phys_t sio_zb;
249 /* members for queue sorting */
251 avl_node_t sio_addr_node; /* link into issuing queue */
252 list_node_t sio_list_node; /* link for issuing to disk */
256 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp,
257 * depending on how many were in the original bp. Only the
258 * first DVA is really used for sorting and issuing purposes.
259 * The other DVAs (if provided) simply exist so that the zio
260 * layer can find additional copies to repair from in the
261 * event of an error. This array must go at the end of the
262 * struct to allow this for the variable number of elements.
267 #define SIO_SET_OFFSET(sio, x) DVA_SET_OFFSET(&(sio)->sio_dva[0], x)
268 #define SIO_SET_ASIZE(sio, x) DVA_SET_ASIZE(&(sio)->sio_dva[0], x)
269 #define SIO_GET_OFFSET(sio) DVA_GET_OFFSET(&(sio)->sio_dva[0])
270 #define SIO_GET_ASIZE(sio) DVA_GET_ASIZE(&(sio)->sio_dva[0])
271 #define SIO_GET_END_OFFSET(sio) \
272 (SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio))
273 #define SIO_GET_MUSED(sio) \
274 (sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t)))
276 struct dsl_scan_io_queue {
277 dsl_scan_t *q_scn; /* associated dsl_scan_t */
278 vdev_t *q_vd; /* top-level vdev that this queue represents */
280 /* trees used for sorting I/Os and extents of I/Os */
281 range_tree_t *q_exts_by_addr;
282 zfs_btree_t q_exts_by_size;
283 avl_tree_t q_sios_by_addr;
284 uint64_t q_sio_memused;
286 /* members for zio rate limiting */
287 uint64_t q_maxinflight_bytes;
288 uint64_t q_inflight_bytes;
289 kcondvar_t q_zio_cv; /* used under vd->vdev_scan_io_queue_lock */
291 /* per txg statistics */
292 uint64_t q_total_seg_size_this_txg;
293 uint64_t q_segs_this_txg;
294 uint64_t q_total_zio_size_this_txg;
295 uint64_t q_zios_this_txg;
298 /* private data for dsl_scan_prefetch_cb() */
299 typedef struct scan_prefetch_ctx {
300 zfs_refcount_t spc_refcnt; /* refcount for memory management */
301 dsl_scan_t *spc_scn; /* dsl_scan_t for the pool */
302 boolean_t spc_root; /* is this prefetch for an objset? */
303 uint8_t spc_indblkshift; /* dn_indblkshift of current dnode */
304 uint16_t spc_datablkszsec; /* dn_idatablkszsec of current dnode */
305 } scan_prefetch_ctx_t;
307 /* private data for dsl_scan_prefetch() */
308 typedef struct scan_prefetch_issue_ctx {
309 avl_node_t spic_avl_node; /* link into scn->scn_prefetch_queue */
310 scan_prefetch_ctx_t *spic_spc; /* spc for the callback */
311 blkptr_t spic_bp; /* bp to prefetch */
312 zbookmark_phys_t spic_zb; /* bookmark to prefetch */
313 } scan_prefetch_issue_ctx_t;
315 static void scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
316 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue);
317 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue,
320 static dsl_scan_io_queue_t *scan_io_queue_create(vdev_t *vd);
321 static void scan_io_queues_destroy(dsl_scan_t *scn);
323 static kmem_cache_t *sio_cache[SPA_DVAS_PER_BP];
325 /* sio->sio_nr_dvas must be set so we know which cache to free from */
327 sio_free(scan_io_t *sio)
329 ASSERT3U(sio->sio_nr_dvas, >, 0);
330 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
332 kmem_cache_free(sio_cache[sio->sio_nr_dvas - 1], sio);
335 /* It is up to the caller to set sio->sio_nr_dvas for freeing */
337 sio_alloc(unsigned short nr_dvas)
339 ASSERT3U(nr_dvas, >, 0);
340 ASSERT3U(nr_dvas, <=, SPA_DVAS_PER_BP);
342 return (kmem_cache_alloc(sio_cache[nr_dvas - 1], KM_SLEEP));
349 * This is used in ext_size_compare() to weight segments
350 * based on how sparse they are. This cannot be changed
351 * mid-scan and the tree comparison functions don't currently
352 * have a mechanism for passing additional context to the
353 * compare functions. Thus we store this value globally and
354 * we only allow it to be set at module initialization time
356 fill_weight = zfs_scan_fill_weight;
358 for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
361 (void) sprintf(name, "sio_cache_%d", i);
362 sio_cache[i] = kmem_cache_create(name,
363 (sizeof (scan_io_t) + ((i + 1) * sizeof (dva_t))),
364 0, NULL, NULL, NULL, NULL, NULL, 0);
371 for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
372 kmem_cache_destroy(sio_cache[i]);
376 static inline boolean_t
377 dsl_scan_is_running(const dsl_scan_t *scn)
379 return (scn->scn_phys.scn_state == DSS_SCANNING);
383 dsl_scan_resilvering(dsl_pool_t *dp)
385 return (dsl_scan_is_running(dp->dp_scan) &&
386 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
390 sio2bp(const scan_io_t *sio, blkptr_t *bp)
392 bzero(bp, sizeof (*bp));
393 bp->blk_prop = sio->sio_blk_prop;
394 bp->blk_phys_birth = sio->sio_phys_birth;
395 bp->blk_birth = sio->sio_birth;
396 bp->blk_fill = 1; /* we always only work with data pointers */
397 bp->blk_cksum = sio->sio_cksum;
399 ASSERT3U(sio->sio_nr_dvas, >, 0);
400 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
402 bcopy(sio->sio_dva, bp->blk_dva, sio->sio_nr_dvas * sizeof (dva_t));
406 bp2sio(const blkptr_t *bp, scan_io_t *sio, int dva_i)
408 sio->sio_blk_prop = bp->blk_prop;
409 sio->sio_phys_birth = bp->blk_phys_birth;
410 sio->sio_birth = bp->blk_birth;
411 sio->sio_cksum = bp->blk_cksum;
412 sio->sio_nr_dvas = BP_GET_NDVAS(bp);
415 * Copy the DVAs to the sio. We need all copies of the block so
416 * that the self healing code can use the alternate copies if the
417 * first is corrupted. We want the DVA at index dva_i to be first
418 * in the sio since this is the primary one that we want to issue.
420 for (int i = 0, j = dva_i; i < sio->sio_nr_dvas; i++, j++) {
421 sio->sio_dva[i] = bp->blk_dva[j % sio->sio_nr_dvas];
426 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
430 spa_t *spa = dp->dp_spa;
433 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
437 * It's possible that we're resuming a scan after a reboot so
438 * make sure that the scan_async_destroying flag is initialized
441 ASSERT(!scn->scn_async_destroying);
442 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
443 SPA_FEATURE_ASYNC_DESTROY);
446 * Calculate the max number of in-flight bytes for pool-wide
447 * scanning operations (minimum 1MB). Limits for the issuing
448 * phase are done per top-level vdev and are handled separately.
450 scn->scn_maxinflight_bytes = MAX(zfs_scan_vdev_limit *
451 dsl_scan_count_leaves(spa->spa_root_vdev), 1ULL << 20);
453 avl_create(&scn->scn_queue, scan_ds_queue_compare, sizeof (scan_ds_t),
454 offsetof(scan_ds_t, sds_node));
455 avl_create(&scn->scn_prefetch_queue, scan_prefetch_queue_compare,
456 sizeof (scan_prefetch_issue_ctx_t),
457 offsetof(scan_prefetch_issue_ctx_t, spic_avl_node));
459 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
460 "scrub_func", sizeof (uint64_t), 1, &f);
463 * There was an old-style scrub in progress. Restart a
464 * new-style scrub from the beginning.
466 scn->scn_restart_txg = txg;
467 zfs_dbgmsg("old-style scrub was in progress; "
468 "restarting new-style scrub in txg %llu",
469 (longlong_t)scn->scn_restart_txg);
472 * Load the queue obj from the old location so that it
473 * can be freed by dsl_scan_done().
475 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
476 "scrub_queue", sizeof (uint64_t), 1,
477 &scn->scn_phys.scn_queue_obj);
479 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
480 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
483 * Detect if the pool contains the signature of #2094. If it
484 * does properly update the scn->scn_phys structure and notify
485 * the administrator by setting an errata for the pool.
487 if (err == EOVERFLOW) {
488 uint64_t zaptmp[SCAN_PHYS_NUMINTS + 1];
489 VERIFY3S(SCAN_PHYS_NUMINTS, ==, 24);
490 VERIFY3S(offsetof(dsl_scan_phys_t, scn_flags), ==,
491 (23 * sizeof (uint64_t)));
493 err = zap_lookup(dp->dp_meta_objset,
494 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SCAN,
495 sizeof (uint64_t), SCAN_PHYS_NUMINTS + 1, &zaptmp);
497 uint64_t overflow = zaptmp[SCAN_PHYS_NUMINTS];
499 if (overflow & ~DSL_SCAN_FLAGS_MASK ||
500 scn->scn_async_destroying) {
502 ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY;
506 bcopy(zaptmp, &scn->scn_phys,
507 SCAN_PHYS_NUMINTS * sizeof (uint64_t));
508 scn->scn_phys.scn_flags = overflow;
510 /* Required scrub already in progress. */
511 if (scn->scn_phys.scn_state == DSS_FINISHED ||
512 scn->scn_phys.scn_state == DSS_CANCELED)
514 ZPOOL_ERRATA_ZOL_2094_SCRUB;
524 * We might be restarting after a reboot, so jump the issued
525 * counter to how far we've scanned. We know we're consistent
528 scn->scn_issued_before_pass = scn->scn_phys.scn_examined;
530 if (dsl_scan_is_running(scn) &&
531 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
533 * A new-type scrub was in progress on an old
534 * pool, and the pool was accessed by old
535 * software. Restart from the beginning, since
536 * the old software may have changed the pool in
539 scn->scn_restart_txg = txg;
540 zfs_dbgmsg("new-style scrub was modified "
541 "by old software; restarting in txg %llu",
542 (longlong_t)scn->scn_restart_txg);
546 bcopy(&scn->scn_phys, &scn->scn_phys_cached, sizeof (scn->scn_phys));
548 /* reload the queue into the in-core state */
549 if (scn->scn_phys.scn_queue_obj != 0) {
553 for (zap_cursor_init(&zc, dp->dp_meta_objset,
554 scn->scn_phys.scn_queue_obj);
555 zap_cursor_retrieve(&zc, &za) == 0;
556 (void) zap_cursor_advance(&zc)) {
557 scan_ds_queue_insert(scn,
558 zfs_strtonum(za.za_name, NULL),
559 za.za_first_integer);
561 zap_cursor_fini(&zc);
564 spa_scan_stat_init(spa);
569 dsl_scan_fini(dsl_pool_t *dp)
571 if (dp->dp_scan != NULL) {
572 dsl_scan_t *scn = dp->dp_scan;
574 if (scn->scn_taskq != NULL)
575 taskq_destroy(scn->scn_taskq);
577 scan_ds_queue_clear(scn);
578 avl_destroy(&scn->scn_queue);
579 scan_ds_prefetch_queue_clear(scn);
580 avl_destroy(&scn->scn_prefetch_queue);
582 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
588 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
590 return (scn->scn_restart_txg != 0 &&
591 scn->scn_restart_txg <= tx->tx_txg);
595 dsl_scan_scrubbing(const dsl_pool_t *dp)
597 dsl_scan_phys_t *scn_phys = &dp->dp_scan->scn_phys;
599 return (scn_phys->scn_state == DSS_SCANNING &&
600 scn_phys->scn_func == POOL_SCAN_SCRUB);
604 dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
606 return (dsl_scan_scrubbing(scn->scn_dp) &&
607 scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED);
611 * Writes out a persistent dsl_scan_phys_t record to the pool directory.
612 * Because we can be running in the block sorting algorithm, we do not always
613 * want to write out the record, only when it is "safe" to do so. This safety
614 * condition is achieved by making sure that the sorting queues are empty
615 * (scn_bytes_pending == 0). When this condition is not true, the sync'd state
616 * is inconsistent with how much actual scanning progress has been made. The
617 * kind of sync to be performed is specified by the sync_type argument. If the
618 * sync is optional, we only sync if the queues are empty. If the sync is
619 * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
620 * third possible state is a "cached" sync. This is done in response to:
621 * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
622 * destroyed, so we wouldn't be able to restart scanning from it.
623 * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
624 * superseded by a newer snapshot.
625 * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
626 * swapped with its clone.
627 * In all cases, a cached sync simply rewrites the last record we've written,
628 * just slightly modified. For the modifications that are performed to the
629 * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
630 * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
633 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx, state_sync_type_t sync_type)
636 spa_t *spa = scn->scn_dp->dp_spa;
638 ASSERT(sync_type != SYNC_MANDATORY || scn->scn_bytes_pending == 0);
639 if (scn->scn_bytes_pending == 0) {
640 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
641 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
642 dsl_scan_io_queue_t *q = vd->vdev_scan_io_queue;
647 mutex_enter(&vd->vdev_scan_io_queue_lock);
648 ASSERT3P(avl_first(&q->q_sios_by_addr), ==, NULL);
649 ASSERT3P(zfs_btree_first(&q->q_exts_by_size, NULL), ==,
651 ASSERT3P(range_tree_first(q->q_exts_by_addr), ==, NULL);
652 mutex_exit(&vd->vdev_scan_io_queue_lock);
655 if (scn->scn_phys.scn_queue_obj != 0)
656 scan_ds_queue_sync(scn, tx);
657 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
658 DMU_POOL_DIRECTORY_OBJECT,
659 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
660 &scn->scn_phys, tx));
661 bcopy(&scn->scn_phys, &scn->scn_phys_cached,
662 sizeof (scn->scn_phys));
664 if (scn->scn_checkpointing)
665 zfs_dbgmsg("finish scan checkpoint");
667 scn->scn_checkpointing = B_FALSE;
668 scn->scn_last_checkpoint = ddi_get_lbolt();
669 } else if (sync_type == SYNC_CACHED) {
670 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
671 DMU_POOL_DIRECTORY_OBJECT,
672 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
673 &scn->scn_phys_cached, tx));
679 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
681 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
683 if (dsl_scan_is_running(scn))
684 return (SET_ERROR(EBUSY));
690 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
692 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
693 pool_scan_func_t *funcp = arg;
694 dmu_object_type_t ot = 0;
695 dsl_pool_t *dp = scn->scn_dp;
696 spa_t *spa = dp->dp_spa;
698 ASSERT(!dsl_scan_is_running(scn));
699 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
700 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
701 scn->scn_phys.scn_func = *funcp;
702 scn->scn_phys.scn_state = DSS_SCANNING;
703 scn->scn_phys.scn_min_txg = 0;
704 scn->scn_phys.scn_max_txg = tx->tx_txg;
705 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
706 scn->scn_phys.scn_start_time = gethrestime_sec();
707 scn->scn_phys.scn_errors = 0;
708 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
709 scn->scn_issued_before_pass = 0;
710 scn->scn_restart_txg = 0;
711 scn->scn_done_txg = 0;
712 scn->scn_last_checkpoint = 0;
713 scn->scn_checkpointing = B_FALSE;
714 spa_scan_stat_init(spa);
716 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
717 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
719 /* rewrite all disk labels */
720 vdev_config_dirty(spa->spa_root_vdev);
722 if (vdev_resilver_needed(spa->spa_root_vdev,
723 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
724 spa_event_notify(spa, NULL, NULL,
725 ESC_ZFS_RESILVER_START);
727 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
730 spa->spa_scrub_started = B_TRUE;
732 * If this is an incremental scrub, limit the DDT scrub phase
733 * to just the auto-ditto class (for correctness); the rest
734 * of the scrub should go faster using top-down pruning.
736 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
737 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
741 /* back to the generic stuff */
743 if (dp->dp_blkstats == NULL) {
745 vmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
746 mutex_init(&dp->dp_blkstats->zab_lock, NULL,
747 MUTEX_DEFAULT, NULL);
749 bzero(&dp->dp_blkstats->zab_type, sizeof (dp->dp_blkstats->zab_type));
751 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
752 ot = DMU_OT_ZAP_OTHER;
754 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
755 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
757 bcopy(&scn->scn_phys, &scn->scn_phys_cached, sizeof (scn->scn_phys));
759 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
761 spa_history_log_internal(spa, "scan setup", tx,
762 "func=%u mintxg=%llu maxtxg=%llu",
763 *funcp, (u_longlong_t)scn->scn_phys.scn_min_txg,
764 (u_longlong_t)scn->scn_phys.scn_max_txg);
768 * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
769 * Can also be called to resume a paused scrub.
772 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
774 spa_t *spa = dp->dp_spa;
775 dsl_scan_t *scn = dp->dp_scan;
778 * Purge all vdev caches and probe all devices. We do this here
779 * rather than in sync context because this requires a writer lock
780 * on the spa_config lock, which we can't do from sync context. The
781 * spa_scrub_reopen flag indicates that vdev_open() should not
782 * attempt to start another scrub.
784 spa_vdev_state_enter(spa, SCL_NONE);
785 spa->spa_scrub_reopen = B_TRUE;
786 vdev_reopen(spa->spa_root_vdev);
787 spa->spa_scrub_reopen = B_FALSE;
788 (void) spa_vdev_state_exit(spa, NULL, 0);
790 if (func == POOL_SCAN_RESILVER) {
791 dsl_resilver_restart(spa->spa_dsl_pool, 0);
795 if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
796 /* got scrub start cmd, resume paused scrub */
797 int err = dsl_scrub_set_pause_resume(scn->scn_dp,
800 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
804 return (SET_ERROR(err));
807 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
808 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED));
812 * Sets the resilver defer flag to B_FALSE on all leaf devs under vd. Returns
813 * B_TRUE if we have devices that need to be resilvered and are available to
814 * accept resilver I/Os.
817 dsl_scan_clear_deferred(vdev_t *vd, dmu_tx_t *tx)
819 boolean_t resilver_needed = B_FALSE;
820 spa_t *spa = vd->vdev_spa;
822 for (int c = 0; c < vd->vdev_children; c++) {
824 dsl_scan_clear_deferred(vd->vdev_child[c], tx);
827 if (vd == spa->spa_root_vdev &&
828 spa_feature_is_active(spa, SPA_FEATURE_RESILVER_DEFER)) {
829 spa_feature_decr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
830 vdev_config_dirty(vd);
831 spa->spa_resilver_deferred = B_FALSE;
832 return (resilver_needed);
835 if (!vdev_is_concrete(vd) || vd->vdev_aux ||
836 !vd->vdev_ops->vdev_op_leaf)
837 return (resilver_needed);
839 if (vd->vdev_resilver_deferred)
840 vd->vdev_resilver_deferred = B_FALSE;
842 return (!vdev_is_dead(vd) && !vd->vdev_offline &&
843 vdev_resilver_needed(vd, NULL, NULL));
848 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
850 static const char *old_names[] = {
852 "scrub_ddt_bookmark",
853 "scrub_ddt_class_max",
862 dsl_pool_t *dp = scn->scn_dp;
863 spa_t *spa = dp->dp_spa;
866 /* Remove any remnants of an old-style scrub. */
867 for (i = 0; old_names[i]; i++) {
868 (void) zap_remove(dp->dp_meta_objset,
869 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
872 if (scn->scn_phys.scn_queue_obj != 0) {
873 VERIFY0(dmu_object_free(dp->dp_meta_objset,
874 scn->scn_phys.scn_queue_obj, tx));
875 scn->scn_phys.scn_queue_obj = 0;
877 scan_ds_queue_clear(scn);
878 scan_ds_prefetch_queue_clear(scn);
880 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
883 * If we were "restarted" from a stopped state, don't bother
884 * with anything else.
886 if (!dsl_scan_is_running(scn)) {
887 ASSERT(!scn->scn_is_sorted);
891 if (scn->scn_is_sorted) {
892 scan_io_queues_destroy(scn);
893 scn->scn_is_sorted = B_FALSE;
895 if (scn->scn_taskq != NULL) {
896 taskq_destroy(scn->scn_taskq);
897 scn->scn_taskq = NULL;
901 scn->scn_phys.scn_state = complete ? DSS_FINISHED : DSS_CANCELED;
903 spa_notify_waiters(spa);
905 if (dsl_scan_restarting(scn, tx))
906 spa_history_log_internal(spa, "scan aborted, restarting", tx,
907 "errors=%llu", (u_longlong_t)spa_get_errlog_size(spa));
909 spa_history_log_internal(spa, "scan cancelled", tx,
910 "errors=%llu", (u_longlong_t)spa_get_errlog_size(spa));
912 spa_history_log_internal(spa, "scan done", tx,
913 "errors=%llu", (u_longlong_t)spa_get_errlog_size(spa));
915 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
916 spa->spa_scrub_started = B_FALSE;
917 spa->spa_scrub_active = B_FALSE;
920 * If the scrub/resilver completed, update all DTLs to
921 * reflect this. Whether it succeeded or not, vacate
922 * all temporary scrub DTLs.
924 * As the scrub does not currently support traversing
925 * data that have been freed but are part of a checkpoint,
926 * we don't mark the scrub as done in the DTLs as faults
927 * may still exist in those vdevs.
930 !spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
931 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
932 scn->scn_phys.scn_max_txg, B_TRUE);
934 spa_event_notify(spa, NULL, NULL,
935 scn->scn_phys.scn_min_txg ?
936 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
938 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
941 spa_errlog_rotate(spa);
944 * We may have finished replacing a device.
945 * Let the async thread assess this and handle the detach.
947 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
950 * Clear any deferred_resilver flags in the config.
951 * If there are drives that need resilvering, kick
952 * off an asynchronous request to start resilver.
953 * dsl_scan_clear_deferred() may update the config
954 * before the resilver can restart. In the event of
955 * a crash during this period, the spa loading code
956 * will find the drives that need to be resilvered
957 * when the machine reboots and start the resilver then.
959 if (spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) {
960 boolean_t resilver_needed =
961 dsl_scan_clear_deferred(spa->spa_root_vdev, tx);
962 if (resilver_needed) {
963 spa_history_log_internal(spa,
964 "starting deferred resilver", tx,
966 (u_longlong_t)spa_get_errlog_size(spa));
967 spa_async_request(spa, SPA_ASYNC_RESILVER);
972 scn->scn_phys.scn_end_time = gethrestime_sec();
974 if (spa->spa_errata == ZPOOL_ERRATA_ZOL_2094_SCRUB)
977 ASSERT(!dsl_scan_is_running(scn));
982 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
984 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
986 if (!dsl_scan_is_running(scn))
987 return (SET_ERROR(ENOENT));
993 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
995 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
997 dsl_scan_done(scn, B_FALSE, tx);
998 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
999 spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
1003 dsl_scan_cancel(dsl_pool_t *dp)
1005 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
1006 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
1010 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
1012 pool_scrub_cmd_t *cmd = arg;
1013 dsl_pool_t *dp = dmu_tx_pool(tx);
1014 dsl_scan_t *scn = dp->dp_scan;
1016 if (*cmd == POOL_SCRUB_PAUSE) {
1017 /* can't pause a scrub when there is no in-progress scrub */
1018 if (!dsl_scan_scrubbing(dp))
1019 return (SET_ERROR(ENOENT));
1021 /* can't pause a paused scrub */
1022 if (dsl_scan_is_paused_scrub(scn))
1023 return (SET_ERROR(EBUSY));
1024 } else if (*cmd != POOL_SCRUB_NORMAL) {
1025 return (SET_ERROR(ENOTSUP));
1032 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
1034 pool_scrub_cmd_t *cmd = arg;
1035 dsl_pool_t *dp = dmu_tx_pool(tx);
1036 spa_t *spa = dp->dp_spa;
1037 dsl_scan_t *scn = dp->dp_scan;
1039 if (*cmd == POOL_SCRUB_PAUSE) {
1040 /* can't pause a scrub when there is no in-progress scrub */
1041 spa->spa_scan_pass_scrub_pause = gethrestime_sec();
1042 scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
1043 scn->scn_phys_cached.scn_flags |= DSF_SCRUB_PAUSED;
1044 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1045 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
1046 spa_notify_waiters(spa);
1048 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
1049 if (dsl_scan_is_paused_scrub(scn)) {
1051 * We need to keep track of how much time we spend
1052 * paused per pass so that we can adjust the scrub rate
1053 * shown in the output of 'zpool status'
1055 spa->spa_scan_pass_scrub_spent_paused +=
1056 gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
1057 spa->spa_scan_pass_scrub_pause = 0;
1058 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
1059 scn->scn_phys_cached.scn_flags &= ~DSF_SCRUB_PAUSED;
1060 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1066 * Set scrub pause/resume state if it makes sense to do so
1069 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
1071 return (dsl_sync_task(spa_name(dp->dp_spa),
1072 dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
1073 ZFS_SPACE_CHECK_RESERVED));
1077 /* start a new scan, or restart an existing one. */
1079 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1083 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1084 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1086 txg = dmu_tx_get_txg(tx);
1087 dp->dp_scan->scn_restart_txg = txg;
1090 dp->dp_scan->scn_restart_txg = txg;
1092 zfs_dbgmsg("restarting resilver txg=%llu", (longlong_t)txg);
1096 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
1098 zio_free(dp->dp_spa, txg, bp);
1102 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
1104 ASSERT(dsl_pool_sync_context(dp));
1105 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
1109 scan_ds_queue_compare(const void *a, const void *b)
1111 const scan_ds_t *sds_a = a, *sds_b = b;
1113 if (sds_a->sds_dsobj < sds_b->sds_dsobj)
1115 if (sds_a->sds_dsobj == sds_b->sds_dsobj)
1121 scan_ds_queue_clear(dsl_scan_t *scn)
1123 void *cookie = NULL;
1125 while ((sds = avl_destroy_nodes(&scn->scn_queue, &cookie)) != NULL) {
1126 kmem_free(sds, sizeof (*sds));
1131 scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, uint64_t *txg)
1133 scan_ds_t srch, *sds;
1135 srch.sds_dsobj = dsobj;
1136 sds = avl_find(&scn->scn_queue, &srch, NULL);
1137 if (sds != NULL && txg != NULL)
1138 *txg = sds->sds_txg;
1139 return (sds != NULL);
1143 scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg)
1148 sds = kmem_zalloc(sizeof (*sds), KM_SLEEP);
1149 sds->sds_dsobj = dsobj;
1152 VERIFY3P(avl_find(&scn->scn_queue, sds, &where), ==, NULL);
1153 avl_insert(&scn->scn_queue, sds, where);
1157 scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj)
1159 scan_ds_t srch, *sds;
1161 srch.sds_dsobj = dsobj;
1163 sds = avl_find(&scn->scn_queue, &srch, NULL);
1164 VERIFY(sds != NULL);
1165 avl_remove(&scn->scn_queue, sds);
1166 kmem_free(sds, sizeof (*sds));
1170 scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx)
1172 dsl_pool_t *dp = scn->scn_dp;
1173 spa_t *spa = dp->dp_spa;
1174 dmu_object_type_t ot = (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) ?
1175 DMU_OT_SCAN_QUEUE : DMU_OT_ZAP_OTHER;
1177 ASSERT0(scn->scn_bytes_pending);
1178 ASSERT(scn->scn_phys.scn_queue_obj != 0);
1180 VERIFY0(dmu_object_free(dp->dp_meta_objset,
1181 scn->scn_phys.scn_queue_obj, tx));
1182 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, ot,
1183 DMU_OT_NONE, 0, tx);
1184 for (scan_ds_t *sds = avl_first(&scn->scn_queue);
1185 sds != NULL; sds = AVL_NEXT(&scn->scn_queue, sds)) {
1186 VERIFY0(zap_add_int_key(dp->dp_meta_objset,
1187 scn->scn_phys.scn_queue_obj, sds->sds_dsobj,
1193 * Computes the memory limit state that we're currently in. A sorted scan
1194 * needs quite a bit of memory to hold the sorting queue, so we need to
1195 * reasonably constrain the size so it doesn't impact overall system
1196 * performance. We compute two limits:
1197 * 1) Hard memory limit: if the amount of memory used by the sorting
1198 * queues on a pool gets above this value, we stop the metadata
1199 * scanning portion and start issuing the queued up and sorted
1200 * I/Os to reduce memory usage.
1201 * This limit is calculated as a fraction of physmem (by default 5%).
1202 * We constrain the lower bound of the hard limit to an absolute
1203 * minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1204 * the upper bound to 5% of the total pool size - no chance we'll
1205 * ever need that much memory, but just to keep the value in check.
1206 * 2) Soft memory limit: once we hit the hard memory limit, we start
1207 * issuing I/O to reduce queue memory usage, but we don't want to
1208 * completely empty out the queues, since we might be able to find I/Os
1209 * that will fill in the gaps of our non-sequential IOs at some point
1210 * in the future. So we stop the issuing of I/Os once the amount of
1211 * memory used drops below the soft limit (at which point we stop issuing
1212 * I/O and start scanning metadata again).
1214 * This limit is calculated by subtracting a fraction of the hard
1215 * limit from the hard limit. By default this fraction is 5%, so
1216 * the soft limit is 95% of the hard limit. We cap the size of the
1217 * difference between the hard and soft limits at an absolute
1218 * maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1219 * sufficient to not cause too frequent switching between the
1220 * metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1221 * worth of queues is about 1.2 GiB of on-pool data, so scanning
1222 * that should take at least a decent fraction of a second).
1225 dsl_scan_should_clear(dsl_scan_t *scn)
1227 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
1228 uint64_t mlim_hard, mlim_soft, mused;
1229 uint64_t alloc = metaslab_class_get_alloc(spa_normal_class(
1230 scn->scn_dp->dp_spa));
1232 mlim_hard = MAX((physmem / zfs_scan_mem_lim_fact) * PAGESIZE,
1233 zfs_scan_mem_lim_min);
1234 mlim_hard = MIN(mlim_hard, alloc / 20);
1235 mlim_soft = mlim_hard - MIN(mlim_hard / zfs_scan_mem_lim_soft_fact,
1236 zfs_scan_mem_lim_soft_max);
1238 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1239 vdev_t *tvd = rvd->vdev_child[i];
1240 dsl_scan_io_queue_t *queue;
1242 mutex_enter(&tvd->vdev_scan_io_queue_lock);
1243 queue = tvd->vdev_scan_io_queue;
1244 if (queue != NULL) {
1245 /* # extents in exts_by_size = # in exts_by_addr */
1246 mused += zfs_btree_numnodes(&queue->q_exts_by_size) *
1247 sizeof (range_seg_gap_t) + queue->q_sio_memused;
1249 mutex_exit(&tvd->vdev_scan_io_queue_lock);
1252 dprintf("current scan memory usage: %llu bytes\n", (longlong_t)mused);
1255 ASSERT0(scn->scn_bytes_pending);
1258 * If we are above our hard limit, we need to clear out memory.
1259 * If we are below our soft limit, we need to accumulate sequential IOs.
1260 * Otherwise, we should keep doing whatever we are currently doing.
1262 if (mused >= mlim_hard)
1264 else if (mused < mlim_soft)
1267 return (scn->scn_clearing);
1271 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1273 /* we never skip user/group accounting objects */
1274 if (zb && (int64_t)zb->zb_object < 0)
1277 if (scn->scn_suspending)
1278 return (B_TRUE); /* we're already suspending */
1280 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
1281 return (B_FALSE); /* we're resuming */
1283 /* We only know how to resume from level-0 and objset blocks. */
1284 if (zb && (zb->zb_level != 0 && zb->zb_level != ZB_ROOT_LEVEL))
1289 * - we have scanned for at least the minimum time (default 1 sec
1290 * for scrub, 3 sec for resilver), and either we have sufficient
1291 * dirty data that we are starting to write more quickly
1292 * (default 30%), someone is explicitly waiting for this txg
1293 * to complete, or we have used up all of the time in the txg
1294 * timeout (default 5 sec).
1296 * - the spa is shutting down because this pool is being exported
1297 * or the machine is rebooting.
1299 * - the scan queue has reached its memory use limit
1301 uint64_t curr_time_ns = gethrtime();
1302 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
1303 uint64_t sync_time_ns = curr_time_ns -
1304 scn->scn_dp->dp_spa->spa_sync_starttime;
1305 int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
1306 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
1307 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
1309 if ((NSEC2MSEC(scan_time_ns) > mintime &&
1310 (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
1311 txg_sync_waiting(scn->scn_dp) ||
1312 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1313 spa_shutting_down(scn->scn_dp->dp_spa) ||
1314 (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn))) {
1315 if (zb && zb->zb_level == ZB_ROOT_LEVEL) {
1316 dprintf("suspending at first available bookmark "
1317 "%llx/%llx/%llx/%llx\n",
1318 (longlong_t)zb->zb_objset,
1319 (longlong_t)zb->zb_object,
1320 (longlong_t)zb->zb_level,
1321 (longlong_t)zb->zb_blkid);
1322 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
1323 zb->zb_objset, 0, 0, 0);
1324 } else if (zb != NULL) {
1325 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1326 (longlong_t)zb->zb_objset,
1327 (longlong_t)zb->zb_object,
1328 (longlong_t)zb->zb_level,
1329 (longlong_t)zb->zb_blkid);
1330 scn->scn_phys.scn_bookmark = *zb;
1333 dsl_scan_phys_t *scnp = &scn->scn_phys;
1334 dprintf("suspending at at DDT bookmark "
1335 "%llx/%llx/%llx/%llx\n",
1336 (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
1337 (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
1338 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
1339 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
1342 scn->scn_suspending = B_TRUE;
1348 typedef struct zil_scan_arg {
1350 zil_header_t *zsa_zh;
1355 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1357 zil_scan_arg_t *zsa = arg;
1358 dsl_pool_t *dp = zsa->zsa_dp;
1359 dsl_scan_t *scn = dp->dp_scan;
1360 zil_header_t *zh = zsa->zsa_zh;
1361 zbookmark_phys_t zb;
1363 ASSERT(!BP_IS_REDACTED(bp));
1364 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1368 * One block ("stubby") can be allocated a long time ago; we
1369 * want to visit that one because it has been allocated
1370 * (on-disk) even if it hasn't been claimed (even though for
1371 * scrub there's nothing to do to it).
1373 if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(dp->dp_spa))
1376 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1377 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
1379 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1385 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
1387 if (lrc->lrc_txtype == TX_WRITE) {
1388 zil_scan_arg_t *zsa = arg;
1389 dsl_pool_t *dp = zsa->zsa_dp;
1390 dsl_scan_t *scn = dp->dp_scan;
1391 zil_header_t *zh = zsa->zsa_zh;
1392 lr_write_t *lr = (lr_write_t *)lrc;
1393 blkptr_t *bp = &lr->lr_blkptr;
1394 zbookmark_phys_t zb;
1396 ASSERT(!BP_IS_REDACTED(bp));
1397 if (BP_IS_HOLE(bp) ||
1398 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1402 * birth can be < claim_txg if this record's txg is
1403 * already txg sync'ed (but this log block contains
1404 * other records that are not synced)
1406 if (claim_txg == 0 || bp->blk_birth < claim_txg)
1409 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1410 lr->lr_foid, ZB_ZIL_LEVEL,
1411 lr->lr_offset / BP_GET_LSIZE(bp));
1413 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1419 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
1421 uint64_t claim_txg = zh->zh_claim_txg;
1422 zil_scan_arg_t zsa = { dp, zh };
1425 ASSERT(spa_writeable(dp->dp_spa));
1428 * We only want to visit blocks that have been claimed but not yet
1429 * replayed (or, in read-only mode, blocks that *would* be claimed).
1434 zilog = zil_alloc(dp->dp_meta_objset, zh);
1436 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
1437 claim_txg, B_FALSE);
1443 * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1444 * here is to sort the AVL tree by the order each block will be needed.
1447 scan_prefetch_queue_compare(const void *a, const void *b)
1449 const scan_prefetch_issue_ctx_t *spic_a = a, *spic_b = b;
1450 const scan_prefetch_ctx_t *spc_a = spic_a->spic_spc;
1451 const scan_prefetch_ctx_t *spc_b = spic_b->spic_spc;
1453 return (zbookmark_compare(spc_a->spc_datablkszsec,
1454 spc_a->spc_indblkshift, spc_b->spc_datablkszsec,
1455 spc_b->spc_indblkshift, &spic_a->spic_zb, &spic_b->spic_zb));
1459 scan_prefetch_ctx_rele(scan_prefetch_ctx_t *spc, void *tag)
1461 if (zfs_refcount_remove(&spc->spc_refcnt, tag) == 0) {
1462 zfs_refcount_destroy(&spc->spc_refcnt);
1463 kmem_free(spc, sizeof (scan_prefetch_ctx_t));
1467 static scan_prefetch_ctx_t *
1468 scan_prefetch_ctx_create(dsl_scan_t *scn, dnode_phys_t *dnp, void *tag)
1470 scan_prefetch_ctx_t *spc;
1472 spc = kmem_alloc(sizeof (scan_prefetch_ctx_t), KM_SLEEP);
1473 zfs_refcount_create(&spc->spc_refcnt);
1474 zfs_refcount_add(&spc->spc_refcnt, tag);
1477 spc->spc_datablkszsec = dnp->dn_datablkszsec;
1478 spc->spc_indblkshift = dnp->dn_indblkshift;
1479 spc->spc_root = B_FALSE;
1481 spc->spc_datablkszsec = 0;
1482 spc->spc_indblkshift = 0;
1483 spc->spc_root = B_TRUE;
1490 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t *spc, void *tag)
1492 zfs_refcount_add(&spc->spc_refcnt, tag);
1496 scan_ds_prefetch_queue_clear(dsl_scan_t *scn)
1498 spa_t *spa = scn->scn_dp->dp_spa;
1499 void *cookie = NULL;
1500 scan_prefetch_issue_ctx_t *spic = NULL;
1502 mutex_enter(&spa->spa_scrub_lock);
1503 while ((spic = avl_destroy_nodes(&scn->scn_prefetch_queue,
1504 &cookie)) != NULL) {
1505 scan_prefetch_ctx_rele(spic->spic_spc, scn);
1506 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1508 mutex_exit(&spa->spa_scrub_lock);
1512 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t *spc,
1513 const zbookmark_phys_t *zb)
1515 zbookmark_phys_t *last_zb = &spc->spc_scn->scn_prefetch_bookmark;
1516 dnode_phys_t tmp_dnp;
1517 dnode_phys_t *dnp = (spc->spc_root) ? NULL : &tmp_dnp;
1519 if (zb->zb_objset != last_zb->zb_objset)
1521 if ((int64_t)zb->zb_object < 0)
1524 tmp_dnp.dn_datablkszsec = spc->spc_datablkszsec;
1525 tmp_dnp.dn_indblkshift = spc->spc_indblkshift;
1527 if (zbookmark_subtree_completed(dnp, zb, last_zb))
1534 dsl_scan_prefetch(scan_prefetch_ctx_t *spc, blkptr_t *bp, zbookmark_phys_t *zb)
1537 dsl_scan_t *scn = spc->spc_scn;
1538 spa_t *spa = scn->scn_dp->dp_spa;
1539 scan_prefetch_issue_ctx_t *spic;
1541 if (zfs_no_scrub_prefetch || BP_IS_REDACTED(bp))
1544 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg ||
1545 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE &&
1546 BP_GET_TYPE(bp) != DMU_OT_OBJSET))
1549 if (dsl_scan_check_prefetch_resume(spc, zb))
1552 scan_prefetch_ctx_add_ref(spc, scn);
1553 spic = kmem_alloc(sizeof (scan_prefetch_issue_ctx_t), KM_SLEEP);
1554 spic->spic_spc = spc;
1555 spic->spic_bp = *bp;
1556 spic->spic_zb = *zb;
1559 * Add the IO to the queue of blocks to prefetch. This allows us to
1560 * prioritize blocks that we will need first for the main traversal
1563 mutex_enter(&spa->spa_scrub_lock);
1564 if (avl_find(&scn->scn_prefetch_queue, spic, &idx) != NULL) {
1565 /* this block is already queued for prefetch */
1566 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1567 scan_prefetch_ctx_rele(spc, scn);
1568 mutex_exit(&spa->spa_scrub_lock);
1572 avl_insert(&scn->scn_prefetch_queue, spic, idx);
1573 cv_broadcast(&spa->spa_scrub_io_cv);
1574 mutex_exit(&spa->spa_scrub_lock);
1578 dsl_scan_prefetch_dnode(dsl_scan_t *scn, dnode_phys_t *dnp,
1579 uint64_t objset, uint64_t object)
1582 zbookmark_phys_t zb;
1583 scan_prefetch_ctx_t *spc;
1585 if (dnp->dn_nblkptr == 0 && !(dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1588 SET_BOOKMARK(&zb, objset, object, 0, 0);
1590 spc = scan_prefetch_ctx_create(scn, dnp, FTAG);
1592 for (i = 0; i < dnp->dn_nblkptr; i++) {
1593 zb.zb_level = BP_GET_LEVEL(&dnp->dn_blkptr[i]);
1595 dsl_scan_prefetch(spc, &dnp->dn_blkptr[i], &zb);
1598 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1600 zb.zb_blkid = DMU_SPILL_BLKID;
1601 dsl_scan_prefetch(spc, DN_SPILL_BLKPTR(dnp), &zb);
1604 scan_prefetch_ctx_rele(spc, FTAG);
1608 dsl_scan_prefetch_cb(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1609 arc_buf_t *buf, void *private)
1611 scan_prefetch_ctx_t *spc = private;
1612 dsl_scan_t *scn = spc->spc_scn;
1613 spa_t *spa = scn->scn_dp->dp_spa;
1615 /* broadcast that the IO has completed for rate limiting purposes */
1616 mutex_enter(&spa->spa_scrub_lock);
1617 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
1618 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
1619 cv_broadcast(&spa->spa_scrub_io_cv);
1620 mutex_exit(&spa->spa_scrub_lock);
1622 /* if there was an error or we are done prefetching, just cleanup */
1623 if (buf == NULL || scn->scn_prefetch_stop)
1626 if (BP_GET_LEVEL(bp) > 0) {
1629 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1630 zbookmark_phys_t czb;
1632 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1633 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1634 zb->zb_level - 1, zb->zb_blkid * epb + i);
1635 dsl_scan_prefetch(spc, cbp, &czb);
1637 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1640 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1642 for (i = 0, cdnp = buf->b_data; i < epb;
1643 i += cdnp->dn_extra_slots + 1,
1644 cdnp += cdnp->dn_extra_slots + 1) {
1645 dsl_scan_prefetch_dnode(scn, cdnp,
1646 zb->zb_objset, zb->zb_blkid * epb + i);
1648 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1649 objset_phys_t *osp = buf->b_data;
1651 dsl_scan_prefetch_dnode(scn, &osp->os_meta_dnode,
1652 zb->zb_objset, DMU_META_DNODE_OBJECT);
1654 if (OBJSET_BUF_HAS_USERUSED(buf)) {
1655 dsl_scan_prefetch_dnode(scn,
1656 &osp->os_groupused_dnode, zb->zb_objset,
1657 DMU_GROUPUSED_OBJECT);
1658 dsl_scan_prefetch_dnode(scn,
1659 &osp->os_userused_dnode, zb->zb_objset,
1660 DMU_USERUSED_OBJECT);
1666 arc_buf_destroy(buf, private);
1667 scan_prefetch_ctx_rele(spc, scn);
1672 dsl_scan_prefetch_thread(void *arg)
1674 dsl_scan_t *scn = arg;
1675 spa_t *spa = scn->scn_dp->dp_spa;
1676 scan_prefetch_issue_ctx_t *spic;
1678 /* loop until we are told to stop */
1679 while (!scn->scn_prefetch_stop) {
1680 arc_flags_t flags = ARC_FLAG_NOWAIT |
1681 ARC_FLAG_PRESCIENT_PREFETCH | ARC_FLAG_PREFETCH;
1682 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1684 mutex_enter(&spa->spa_scrub_lock);
1687 * Wait until we have an IO to issue and are not above our
1688 * maximum in flight limit.
1690 while (!scn->scn_prefetch_stop &&
1691 (avl_numnodes(&scn->scn_prefetch_queue) == 0 ||
1692 spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)) {
1693 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1696 /* recheck if we should stop since we waited for the cv */
1697 if (scn->scn_prefetch_stop) {
1698 mutex_exit(&spa->spa_scrub_lock);
1702 /* remove the prefetch IO from the tree */
1703 spic = avl_first(&scn->scn_prefetch_queue);
1704 spa->spa_scrub_inflight += BP_GET_PSIZE(&spic->spic_bp);
1705 avl_remove(&scn->scn_prefetch_queue, spic);
1707 mutex_exit(&spa->spa_scrub_lock);
1709 if (BP_IS_PROTECTED(&spic->spic_bp)) {
1710 ASSERT(BP_GET_TYPE(&spic->spic_bp) == DMU_OT_DNODE ||
1711 BP_GET_TYPE(&spic->spic_bp) == DMU_OT_OBJSET);
1712 ASSERT3U(BP_GET_LEVEL(&spic->spic_bp), ==, 0);
1713 zio_flags |= ZIO_FLAG_RAW;
1716 /* issue the prefetch asynchronously */
1717 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa,
1718 &spic->spic_bp, dsl_scan_prefetch_cb, spic->spic_spc,
1719 ZIO_PRIORITY_SCRUB, zio_flags, &flags, &spic->spic_zb);
1721 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1724 ASSERT(scn->scn_prefetch_stop);
1726 /* free any prefetches we didn't get to complete */
1727 mutex_enter(&spa->spa_scrub_lock);
1728 while ((spic = avl_first(&scn->scn_prefetch_queue)) != NULL) {
1729 avl_remove(&scn->scn_prefetch_queue, spic);
1730 scan_prefetch_ctx_rele(spic->spic_spc, scn);
1731 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1733 ASSERT0(avl_numnodes(&scn->scn_prefetch_queue));
1734 mutex_exit(&spa->spa_scrub_lock);
1738 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
1739 const zbookmark_phys_t *zb)
1742 * We never skip over user/group accounting objects (obj<0)
1744 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
1745 (int64_t)zb->zb_object >= 0) {
1747 * If we already visited this bp & everything below (in
1748 * a prior txg sync), don't bother doing it again.
1750 if (zbookmark_subtree_completed(dnp, zb,
1751 &scn->scn_phys.scn_bookmark))
1755 * If we found the block we're trying to resume from, or
1756 * we went past it to a different object, zero it out to
1757 * indicate that it's OK to start checking for suspending
1760 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
1761 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
1762 dprintf("resuming at %llx/%llx/%llx/%llx\n",
1763 (longlong_t)zb->zb_objset,
1764 (longlong_t)zb->zb_object,
1765 (longlong_t)zb->zb_level,
1766 (longlong_t)zb->zb_blkid);
1767 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
1773 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1774 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1775 dmu_objset_type_t ostype, dmu_tx_t *tx);
1776 inline __attribute__((always_inline)) static void dsl_scan_visitdnode(
1777 dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1778 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
1781 * Return nonzero on i/o error.
1782 * Return new buf to write out in *bufp.
1784 inline __attribute__((always_inline)) static int
1785 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1786 dnode_phys_t *dnp, const blkptr_t *bp,
1787 const zbookmark_phys_t *zb, dmu_tx_t *tx)
1789 dsl_pool_t *dp = scn->scn_dp;
1790 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1793 ASSERT(!BP_IS_REDACTED(bp));
1795 if (BP_GET_LEVEL(bp) > 0) {
1796 arc_flags_t flags = ARC_FLAG_WAIT;
1799 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1802 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1803 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1805 scn->scn_phys.scn_errors++;
1808 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1809 zbookmark_phys_t czb;
1811 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1813 zb->zb_blkid * epb + i);
1814 dsl_scan_visitbp(cbp, &czb, dnp,
1815 ds, scn, ostype, tx);
1817 arc_buf_destroy(buf, &buf);
1818 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1819 arc_flags_t flags = ARC_FLAG_WAIT;
1822 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1825 if (BP_IS_PROTECTED(bp)) {
1826 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
1827 zio_flags |= ZIO_FLAG_RAW;
1830 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1831 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1833 scn->scn_phys.scn_errors++;
1836 for (i = 0, cdnp = buf->b_data; i < epb;
1837 i += cdnp->dn_extra_slots + 1,
1838 cdnp += cdnp->dn_extra_slots + 1) {
1839 dsl_scan_visitdnode(scn, ds, ostype,
1840 cdnp, zb->zb_blkid * epb + i, tx);
1843 arc_buf_destroy(buf, &buf);
1844 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1845 arc_flags_t flags = ARC_FLAG_WAIT;
1849 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1850 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1852 scn->scn_phys.scn_errors++;
1858 dsl_scan_visitdnode(scn, ds, osp->os_type,
1859 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
1861 if (OBJSET_BUF_HAS_USERUSED(buf)) {
1863 * We also always visit user/group/project accounting
1864 * objects, and never skip them, even if we are
1865 * suspending. This is necessary so that the
1866 * space deltas from this txg get integrated.
1868 if (OBJSET_BUF_HAS_PROJECTUSED(buf))
1869 dsl_scan_visitdnode(scn, ds, osp->os_type,
1870 &osp->os_projectused_dnode,
1871 DMU_PROJECTUSED_OBJECT, tx);
1872 dsl_scan_visitdnode(scn, ds, osp->os_type,
1873 &osp->os_groupused_dnode,
1874 DMU_GROUPUSED_OBJECT, tx);
1875 dsl_scan_visitdnode(scn, ds, osp->os_type,
1876 &osp->os_userused_dnode,
1877 DMU_USERUSED_OBJECT, tx);
1879 arc_buf_destroy(buf, &buf);
1885 inline __attribute__((always_inline)) static void
1886 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
1887 dmu_objset_type_t ostype, dnode_phys_t *dnp,
1888 uint64_t object, dmu_tx_t *tx)
1892 for (j = 0; j < dnp->dn_nblkptr; j++) {
1893 zbookmark_phys_t czb;
1895 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1896 dnp->dn_nlevels - 1, j);
1897 dsl_scan_visitbp(&dnp->dn_blkptr[j],
1898 &czb, dnp, ds, scn, ostype, tx);
1901 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1902 zbookmark_phys_t czb;
1903 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1904 0, DMU_SPILL_BLKID);
1905 dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp),
1906 &czb, dnp, ds, scn, ostype, tx);
1911 * The arguments are in this order because mdb can only print the
1912 * first 5; we want them to be useful.
1915 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1916 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1917 dmu_objset_type_t ostype, dmu_tx_t *tx)
1919 dsl_pool_t *dp = scn->scn_dp;
1920 blkptr_t *bp_toread = NULL;
1922 if (dsl_scan_check_suspend(scn, zb))
1925 if (dsl_scan_check_resume(scn, dnp, zb))
1928 scn->scn_visited_this_txg++;
1931 * This debugging is commented out to conserve stack space. This
1932 * function is called recursively and the debugging adds several
1933 * bytes to the stack for each call. It can be commented back in
1934 * if required to debug an issue in dsl_scan_visitbp().
1937 * "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
1938 * ds, ds ? ds->ds_object : 0,
1939 * zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1943 if (BP_IS_HOLE(bp)) {
1944 scn->scn_holes_this_txg++;
1948 if (BP_IS_REDACTED(bp)) {
1949 ASSERT(dsl_dataset_feature_is_active(ds,
1950 SPA_FEATURE_REDACTED_DATASETS));
1954 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) {
1955 scn->scn_lt_min_this_txg++;
1959 bp_toread = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
1962 if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx) != 0)
1966 * If dsl_scan_ddt() has already visited this block, it will have
1967 * already done any translations or scrubbing, so don't call the
1970 if (ddt_class_contains(dp->dp_spa,
1971 scn->scn_phys.scn_ddt_class_max, bp)) {
1972 scn->scn_ddt_contained_this_txg++;
1977 * If this block is from the future (after cur_max_txg), then we
1978 * are doing this on behalf of a deleted snapshot, and we will
1979 * revisit the future block on the next pass of this dataset.
1980 * Don't scan it now unless we need to because something
1981 * under it was modified.
1983 if (BP_PHYSICAL_BIRTH(bp) > scn->scn_phys.scn_cur_max_txg) {
1984 scn->scn_gt_max_this_txg++;
1988 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
1991 kmem_free(bp_toread, sizeof (blkptr_t));
1995 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
1998 zbookmark_phys_t zb;
1999 scan_prefetch_ctx_t *spc;
2001 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
2002 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
2004 if (ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) {
2005 SET_BOOKMARK(&scn->scn_prefetch_bookmark,
2006 zb.zb_objset, 0, 0, 0);
2008 scn->scn_prefetch_bookmark = scn->scn_phys.scn_bookmark;
2011 scn->scn_objsets_visited_this_txg++;
2013 spc = scan_prefetch_ctx_create(scn, NULL, FTAG);
2014 dsl_scan_prefetch(spc, bp, &zb);
2015 scan_prefetch_ctx_rele(spc, FTAG);
2017 dsl_scan_visitbp(bp, &zb, NULL, ds, scn, DMU_OST_NONE, tx);
2019 dprintf_ds(ds, "finished scan%s", "");
2023 ds_destroyed_scn_phys(dsl_dataset_t *ds, dsl_scan_phys_t *scn_phys)
2025 if (scn_phys->scn_bookmark.zb_objset == ds->ds_object) {
2026 if (ds->ds_is_snapshot) {
2029 * - scn_cur_{min,max}_txg stays the same.
2030 * - Setting the flag is not really necessary if
2031 * scn_cur_max_txg == scn_max_txg, because there
2032 * is nothing after this snapshot that we care
2033 * about. However, we set it anyway and then
2034 * ignore it when we retraverse it in
2035 * dsl_scan_visitds().
2037 scn_phys->scn_bookmark.zb_objset =
2038 dsl_dataset_phys(ds)->ds_next_snap_obj;
2039 zfs_dbgmsg("destroying ds %llu; currently traversing; "
2040 "reset zb_objset to %llu",
2041 (u_longlong_t)ds->ds_object,
2042 (u_longlong_t)dsl_dataset_phys(ds)->
2044 scn_phys->scn_flags |= DSF_VISIT_DS_AGAIN;
2046 SET_BOOKMARK(&scn_phys->scn_bookmark,
2047 ZB_DESTROYED_OBJSET, 0, 0, 0);
2048 zfs_dbgmsg("destroying ds %llu; currently traversing; "
2049 "reset bookmark to -1,0,0,0",
2050 (u_longlong_t)ds->ds_object);
2056 * Invoked when a dataset is destroyed. We need to make sure that:
2058 * 1) If it is the dataset that was currently being scanned, we write
2059 * a new dsl_scan_phys_t and marking the objset reference in it
2061 * 2) Remove it from the work queue, if it was present.
2063 * If the dataset was actually a snapshot, instead of marking the dataset
2064 * as destroyed, we instead substitute the next snapshot in line.
2067 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
2069 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2070 dsl_scan_t *scn = dp->dp_scan;
2073 if (!dsl_scan_is_running(scn))
2076 ds_destroyed_scn_phys(ds, &scn->scn_phys);
2077 ds_destroyed_scn_phys(ds, &scn->scn_phys_cached);
2079 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2080 scan_ds_queue_remove(scn, ds->ds_object);
2081 if (ds->ds_is_snapshot)
2082 scan_ds_queue_insert(scn,
2083 dsl_dataset_phys(ds)->ds_next_snap_obj, mintxg);
2086 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2087 ds->ds_object, &mintxg) == 0) {
2088 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
2089 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2090 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2091 if (ds->ds_is_snapshot) {
2093 * We keep the same mintxg; it could be >
2094 * ds_creation_txg if the previous snapshot was
2097 VERIFY(zap_add_int_key(dp->dp_meta_objset,
2098 scn->scn_phys.scn_queue_obj,
2099 dsl_dataset_phys(ds)->ds_next_snap_obj,
2101 zfs_dbgmsg("destroying ds %llu; in queue; "
2102 "replacing with %llu",
2103 (u_longlong_t)ds->ds_object,
2104 (u_longlong_t)dsl_dataset_phys(ds)->
2107 zfs_dbgmsg("destroying ds %llu; in queue; removing",
2108 (u_longlong_t)ds->ds_object);
2113 * dsl_scan_sync() should be called after this, and should sync
2114 * out our changed state, but just to be safe, do it here.
2116 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2120 ds_snapshotted_bookmark(dsl_dataset_t *ds, zbookmark_phys_t *scn_bookmark)
2122 if (scn_bookmark->zb_objset == ds->ds_object) {
2123 scn_bookmark->zb_objset =
2124 dsl_dataset_phys(ds)->ds_prev_snap_obj;
2125 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
2126 "reset zb_objset to %llu",
2127 (u_longlong_t)ds->ds_object,
2128 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2133 * Called when a dataset is snapshotted. If we were currently traversing
2134 * this snapshot, we reset our bookmark to point at the newly created
2135 * snapshot. We also modify our work queue to remove the old snapshot and
2136 * replace with the new one.
2139 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
2141 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2142 dsl_scan_t *scn = dp->dp_scan;
2145 if (!dsl_scan_is_running(scn))
2148 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
2150 ds_snapshotted_bookmark(ds, &scn->scn_phys.scn_bookmark);
2151 ds_snapshotted_bookmark(ds, &scn->scn_phys_cached.scn_bookmark);
2153 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2154 scan_ds_queue_remove(scn, ds->ds_object);
2155 scan_ds_queue_insert(scn,
2156 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg);
2159 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2160 ds->ds_object, &mintxg) == 0) {
2161 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2162 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2163 VERIFY(zap_add_int_key(dp->dp_meta_objset,
2164 scn->scn_phys.scn_queue_obj,
2165 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
2166 zfs_dbgmsg("snapshotting ds %llu; in queue; "
2167 "replacing with %llu",
2168 (u_longlong_t)ds->ds_object,
2169 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2172 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2176 ds_clone_swapped_bookmark(dsl_dataset_t *ds1, dsl_dataset_t *ds2,
2177 zbookmark_phys_t *scn_bookmark)
2179 if (scn_bookmark->zb_objset == ds1->ds_object) {
2180 scn_bookmark->zb_objset = ds2->ds_object;
2181 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2182 "reset zb_objset to %llu",
2183 (u_longlong_t)ds1->ds_object,
2184 (u_longlong_t)ds2->ds_object);
2185 } else if (scn_bookmark->zb_objset == ds2->ds_object) {
2186 scn_bookmark->zb_objset = ds1->ds_object;
2187 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2188 "reset zb_objset to %llu",
2189 (u_longlong_t)ds2->ds_object,
2190 (u_longlong_t)ds1->ds_object);
2195 * Called when an origin dataset and its clone are swapped. If we were
2196 * currently traversing the dataset, we need to switch to traversing the
2197 * newly promoted clone.
2200 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
2202 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
2203 dsl_scan_t *scn = dp->dp_scan;
2204 uint64_t mintxg1, mintxg2;
2205 boolean_t ds1_queued, ds2_queued;
2207 if (!dsl_scan_is_running(scn))
2210 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys.scn_bookmark);
2211 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys_cached.scn_bookmark);
2214 * Handle the in-memory scan queue.
2216 ds1_queued = scan_ds_queue_contains(scn, ds1->ds_object, &mintxg1);
2217 ds2_queued = scan_ds_queue_contains(scn, ds2->ds_object, &mintxg2);
2219 /* Sanity checking. */
2221 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2222 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2225 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2226 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2229 if (ds1_queued && ds2_queued) {
2231 * If both are queued, we don't need to do anything.
2232 * The swapping code below would not handle this case correctly,
2233 * since we can't insert ds2 if it is already there. That's
2234 * because scan_ds_queue_insert() prohibits a duplicate insert
2237 } else if (ds1_queued) {
2238 scan_ds_queue_remove(scn, ds1->ds_object);
2239 scan_ds_queue_insert(scn, ds2->ds_object, mintxg1);
2240 } else if (ds2_queued) {
2241 scan_ds_queue_remove(scn, ds2->ds_object);
2242 scan_ds_queue_insert(scn, ds1->ds_object, mintxg2);
2246 * Handle the on-disk scan queue.
2247 * The on-disk state is an out-of-date version of the in-memory state,
2248 * so the in-memory and on-disk values for ds1_queued and ds2_queued may
2249 * be different. Therefore we need to apply the swap logic to the
2250 * on-disk state independently of the in-memory state.
2252 ds1_queued = zap_lookup_int_key(dp->dp_meta_objset,
2253 scn->scn_phys.scn_queue_obj, ds1->ds_object, &mintxg1) == 0;
2254 ds2_queued = zap_lookup_int_key(dp->dp_meta_objset,
2255 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg2) == 0;
2257 /* Sanity checking. */
2259 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2260 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2263 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2264 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2267 if (ds1_queued && ds2_queued) {
2269 * If both are queued, we don't need to do anything.
2270 * Alternatively, we could check for EEXIST from
2271 * zap_add_int_key() and back out to the original state, but
2272 * that would be more work than checking for this case upfront.
2274 } else if (ds1_queued) {
2275 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2276 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
2277 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2278 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg1, tx));
2279 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2280 "replacing with %llu",
2281 (u_longlong_t)ds1->ds_object,
2282 (u_longlong_t)ds2->ds_object);
2283 } else if (ds2_queued) {
2284 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2285 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
2286 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2287 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg2, tx));
2288 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2289 "replacing with %llu",
2290 (u_longlong_t)ds2->ds_object,
2291 (u_longlong_t)ds1->ds_object);
2294 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2299 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2301 uint64_t originobj = *(uint64_t *)arg;
2304 dsl_scan_t *scn = dp->dp_scan;
2306 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != originobj)
2309 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2313 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != originobj) {
2314 dsl_dataset_t *prev;
2315 err = dsl_dataset_hold_obj(dp,
2316 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2318 dsl_dataset_rele(ds, FTAG);
2323 scan_ds_queue_insert(scn, ds->ds_object,
2324 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2325 dsl_dataset_rele(ds, FTAG);
2330 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
2332 dsl_pool_t *dp = scn->scn_dp;
2335 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2337 if (scn->scn_phys.scn_cur_min_txg >=
2338 scn->scn_phys.scn_max_txg) {
2340 * This can happen if this snapshot was created after the
2341 * scan started, and we already completed a previous snapshot
2342 * that was created after the scan started. This snapshot
2343 * only references blocks with:
2345 * birth < our ds_creation_txg
2346 * cur_min_txg is no less than ds_creation_txg.
2347 * We have already visited these blocks.
2349 * birth > scn_max_txg
2350 * The scan requested not to visit these blocks.
2352 * Subsequent snapshots (and clones) can reference our
2353 * blocks, or blocks with even higher birth times.
2354 * Therefore we do not need to visit them either,
2355 * so we do not add them to the work queue.
2357 * Note that checking for cur_min_txg >= cur_max_txg
2358 * is not sufficient, because in that case we may need to
2359 * visit subsequent snapshots. This happens when min_txg > 0,
2360 * which raises cur_min_txg. In this case we will visit
2361 * this dataset but skip all of its blocks, because the
2362 * rootbp's birth time is < cur_min_txg. Then we will
2363 * add the next snapshots/clones to the work queue.
2365 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2366 dsl_dataset_name(ds, dsname);
2367 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2368 "cur_min_txg (%llu) >= max_txg (%llu)",
2369 (longlong_t)dsobj, dsname,
2370 (longlong_t)scn->scn_phys.scn_cur_min_txg,
2371 (longlong_t)scn->scn_phys.scn_max_txg);
2372 kmem_free(dsname, MAXNAMELEN);
2378 * Only the ZIL in the head (non-snapshot) is valid. Even though
2379 * snapshots can have ZIL block pointers (which may be the same
2380 * BP as in the head), they must be ignored. In addition, $ORIGIN
2381 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2382 * need to look for a ZIL in it either. So we traverse the ZIL here,
2383 * rather than in scan_recurse(), because the regular snapshot
2384 * block-sharing rules don't apply to it.
2386 if (!dsl_dataset_is_snapshot(ds) &&
2387 (dp->dp_origin_snap == NULL ||
2388 ds->ds_dir != dp->dp_origin_snap->ds_dir)) {
2390 if (dmu_objset_from_ds(ds, &os) != 0) {
2393 dsl_scan_zil(dp, &os->os_zil_header);
2397 * Iterate over the bps in this ds.
2399 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2400 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
2401 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
2402 rrw_exit(&ds->ds_bp_rwlock, FTAG);
2404 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2405 dsl_dataset_name(ds, dsname);
2406 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2408 (longlong_t)dsobj, dsname,
2409 (longlong_t)scn->scn_phys.scn_cur_min_txg,
2410 (longlong_t)scn->scn_phys.scn_cur_max_txg,
2411 (int)scn->scn_suspending);
2412 kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
2414 if (scn->scn_suspending)
2418 * We've finished this pass over this dataset.
2422 * If we did not completely visit this dataset, do another pass.
2424 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
2425 zfs_dbgmsg("incomplete pass; visiting again");
2426 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
2427 scan_ds_queue_insert(scn, ds->ds_object,
2428 scn->scn_phys.scn_cur_max_txg);
2433 * Add descendant datasets to work queue.
2435 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
2436 scan_ds_queue_insert(scn,
2437 dsl_dataset_phys(ds)->ds_next_snap_obj,
2438 dsl_dataset_phys(ds)->ds_creation_txg);
2440 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
2441 boolean_t usenext = B_FALSE;
2442 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
2445 * A bug in a previous version of the code could
2446 * cause upgrade_clones_cb() to not set
2447 * ds_next_snap_obj when it should, leading to a
2448 * missing entry. Therefore we can only use the
2449 * next_clones_obj when its count is correct.
2451 int err = zap_count(dp->dp_meta_objset,
2452 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
2454 count == dsl_dataset_phys(ds)->ds_num_children - 1)
2461 for (zap_cursor_init(&zc, dp->dp_meta_objset,
2462 dsl_dataset_phys(ds)->ds_next_clones_obj);
2463 zap_cursor_retrieve(&zc, &za) == 0;
2464 (void) zap_cursor_advance(&zc)) {
2465 scan_ds_queue_insert(scn,
2466 zfs_strtonum(za.za_name, NULL),
2467 dsl_dataset_phys(ds)->ds_creation_txg);
2469 zap_cursor_fini(&zc);
2471 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2472 enqueue_clones_cb, &ds->ds_object,
2478 dsl_dataset_rele(ds, FTAG);
2483 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2487 dsl_scan_t *scn = dp->dp_scan;
2489 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2493 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
2494 dsl_dataset_t *prev;
2495 err = dsl_dataset_hold_obj(dp,
2496 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2498 dsl_dataset_rele(ds, FTAG);
2503 * If this is a clone, we don't need to worry about it for now.
2505 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
2506 dsl_dataset_rele(ds, FTAG);
2507 dsl_dataset_rele(prev, FTAG);
2510 dsl_dataset_rele(ds, FTAG);
2514 scan_ds_queue_insert(scn, ds->ds_object,
2515 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2516 dsl_dataset_rele(ds, FTAG);
2522 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
2523 ddt_entry_t *dde, dmu_tx_t *tx)
2525 const ddt_key_t *ddk = &dde->dde_key;
2526 ddt_phys_t *ddp = dde->dde_phys;
2528 zbookmark_phys_t zb = { 0 };
2531 if (!dsl_scan_is_running(scn))
2535 * This function is special because it is the only thing
2536 * that can add scan_io_t's to the vdev scan queues from
2537 * outside dsl_scan_sync(). For the most part this is ok
2538 * as long as it is called from within syncing context.
2539 * However, dsl_scan_sync() expects that no new sio's will
2540 * be added between when all the work for a scan is done
2541 * and the next txg when the scan is actually marked as
2542 * completed. This check ensures we do not issue new sio's
2543 * during this period.
2545 if (scn->scn_done_txg != 0)
2548 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2549 if (ddp->ddp_phys_birth == 0 ||
2550 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
2552 ddt_bp_create(checksum, ddk, ddp, &bp);
2554 scn->scn_visited_this_txg++;
2555 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
2560 * Scrub/dedup interaction.
2562 * If there are N references to a deduped block, we don't want to scrub it
2563 * N times -- ideally, we should scrub it exactly once.
2565 * We leverage the fact that the dde's replication class (enum ddt_class)
2566 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2567 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2569 * To prevent excess scrubbing, the scrub begins by walking the DDT
2570 * to find all blocks with refcnt > 1, and scrubs each of these once.
2571 * Since there are two replication classes which contain blocks with
2572 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2573 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2575 * There would be nothing more to say if a block's refcnt couldn't change
2576 * during a scrub, but of course it can so we must account for changes
2577 * in a block's replication class.
2579 * Here's an example of what can occur:
2581 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2582 * when visited during the top-down scrub phase, it will be scrubbed twice.
2583 * This negates our scrub optimization, but is otherwise harmless.
2585 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2586 * on each visit during the top-down scrub phase, it will never be scrubbed.
2587 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2588 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2589 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2590 * while a scrub is in progress, it scrubs the block right then.
2593 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
2595 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
2600 bzero(&dde, sizeof (ddt_entry_t));
2602 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
2605 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
2607 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
2608 (longlong_t)ddb->ddb_class,
2609 (longlong_t)ddb->ddb_type,
2610 (longlong_t)ddb->ddb_checksum,
2611 (longlong_t)ddb->ddb_cursor);
2613 /* There should be no pending changes to the dedup table */
2614 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
2615 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
2617 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
2620 if (dsl_scan_check_suspend(scn, NULL))
2624 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; "
2625 "suspending=%u", (longlong_t)n,
2626 (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
2628 ASSERT(error == 0 || error == ENOENT);
2629 ASSERT(error != ENOENT ||
2630 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
2634 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
2636 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
2637 if (ds->ds_is_snapshot)
2638 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
2643 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
2646 dsl_pool_t *dp = scn->scn_dp;
2648 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
2649 scn->scn_phys.scn_ddt_class_max) {
2650 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2651 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2652 dsl_scan_ddt(scn, tx);
2653 if (scn->scn_suspending)
2657 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
2658 /* First do the MOS & ORIGIN */
2660 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2661 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2662 dsl_scan_visit_rootbp(scn, NULL,
2663 &dp->dp_meta_rootbp, tx);
2664 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
2665 if (scn->scn_suspending)
2668 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
2669 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2670 enqueue_cb, NULL, DS_FIND_CHILDREN));
2672 dsl_scan_visitds(scn,
2673 dp->dp_origin_snap->ds_object, tx);
2675 ASSERT(!scn->scn_suspending);
2676 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
2677 ZB_DESTROYED_OBJSET) {
2678 uint64_t dsobj = scn->scn_phys.scn_bookmark.zb_objset;
2680 * If we were suspended, continue from here. Note if the
2681 * ds we were suspended on was deleted, the zb_objset may
2682 * be -1, so we will skip this and find a new objset
2685 dsl_scan_visitds(scn, dsobj, tx);
2686 if (scn->scn_suspending)
2691 * In case we suspended right at the end of the ds, zero the
2692 * bookmark so we don't think that we're still trying to resume.
2694 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
2697 * Keep pulling things out of the dataset avl queue. Updates to the
2698 * persistent zap-object-as-queue happen only at checkpoints.
2700 while ((sds = avl_first(&scn->scn_queue)) != NULL) {
2702 uint64_t dsobj = sds->sds_dsobj;
2703 uint64_t txg = sds->sds_txg;
2705 /* dequeue and free the ds from the queue */
2706 scan_ds_queue_remove(scn, dsobj);
2709 /* set up min / max txg */
2710 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2712 scn->scn_phys.scn_cur_min_txg =
2713 MAX(scn->scn_phys.scn_min_txg, txg);
2715 scn->scn_phys.scn_cur_min_txg =
2716 MAX(scn->scn_phys.scn_min_txg,
2717 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2719 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
2720 dsl_dataset_rele(ds, FTAG);
2722 dsl_scan_visitds(scn, dsobj, tx);
2723 if (scn->scn_suspending)
2727 /* No more objsets to fetch, we're done */
2728 scn->scn_phys.scn_bookmark.zb_objset = ZB_DESTROYED_OBJSET;
2729 ASSERT0(scn->scn_suspending);
2733 dsl_scan_count_leaves(vdev_t *vd)
2735 uint64_t i, leaves = 0;
2737 /* we only count leaves that belong to the main pool and are readable */
2738 if (vd->vdev_islog || vd->vdev_isspare ||
2739 vd->vdev_isl2cache || !vdev_readable(vd))
2742 if (vd->vdev_ops->vdev_op_leaf)
2745 for (i = 0; i < vd->vdev_children; i++) {
2746 leaves += dsl_scan_count_leaves(vd->vdev_child[i]);
2753 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t *q, const blkptr_t *bp)
2756 uint64_t cur_size = 0;
2758 for (i = 0; i < BP_GET_NDVAS(bp); i++) {
2759 cur_size += DVA_GET_ASIZE(&bp->blk_dva[i]);
2762 q->q_total_zio_size_this_txg += cur_size;
2763 q->q_zios_this_txg++;
2767 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t *q, uint64_t start,
2770 q->q_total_seg_size_this_txg += end - start;
2771 q->q_segs_this_txg++;
2775 scan_io_queue_check_suspend(dsl_scan_t *scn)
2777 /* See comment in dsl_scan_check_suspend() */
2778 uint64_t curr_time_ns = gethrtime();
2779 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
2780 uint64_t sync_time_ns = curr_time_ns -
2781 scn->scn_dp->dp_spa->spa_sync_starttime;
2782 int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
2783 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
2784 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
2786 return ((NSEC2MSEC(scan_time_ns) > mintime &&
2787 (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
2788 txg_sync_waiting(scn->scn_dp) ||
2789 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
2790 spa_shutting_down(scn->scn_dp->dp_spa));
2794 * Given a list of scan_io_t's in io_list, this issues the I/Os out to
2795 * disk. This consumes the io_list and frees the scan_io_t's. This is
2796 * called when emptying queues, either when we're up against the memory
2797 * limit or when we have finished scanning. Returns B_TRUE if we stopped
2798 * processing the list before we finished. Any sios that were not issued
2799 * will remain in the io_list.
2802 scan_io_queue_issue(dsl_scan_io_queue_t *queue, list_t *io_list)
2804 dsl_scan_t *scn = queue->q_scn;
2806 int64_t bytes_issued = 0;
2807 boolean_t suspended = B_FALSE;
2809 while ((sio = list_head(io_list)) != NULL) {
2812 if (scan_io_queue_check_suspend(scn)) {
2818 bytes_issued += SIO_GET_ASIZE(sio);
2819 scan_exec_io(scn->scn_dp, &bp, sio->sio_flags,
2820 &sio->sio_zb, queue);
2821 (void) list_remove_head(io_list);
2822 scan_io_queues_update_zio_stats(queue, &bp);
2826 atomic_add_64(&scn->scn_bytes_pending, -bytes_issued);
2832 * This function removes sios from an IO queue which reside within a given
2833 * range_seg_t and inserts them (in offset order) into a list. Note that
2834 * we only ever return a maximum of 32 sios at once. If there are more sios
2835 * to process within this segment that did not make it onto the list we
2836 * return B_TRUE and otherwise B_FALSE.
2839 scan_io_queue_gather(dsl_scan_io_queue_t *queue, range_seg_t *rs, list_t *list)
2841 scan_io_t *srch_sio, *sio, *next_sio;
2843 uint_t num_sios = 0;
2844 int64_t bytes_issued = 0;
2847 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2849 srch_sio = sio_alloc(1);
2850 srch_sio->sio_nr_dvas = 1;
2851 SIO_SET_OFFSET(srch_sio, rs_get_start(rs, queue->q_exts_by_addr));
2854 * The exact start of the extent might not contain any matching zios,
2855 * so if that's the case, examine the next one in the tree.
2857 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
2861 sio = avl_nearest(&queue->q_sios_by_addr, idx, AVL_AFTER);
2863 while (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2864 queue->q_exts_by_addr) && num_sios <= 32) {
2865 ASSERT3U(SIO_GET_OFFSET(sio), >=, rs_get_start(rs,
2866 queue->q_exts_by_addr));
2867 ASSERT3U(SIO_GET_END_OFFSET(sio), <=, rs_get_end(rs,
2868 queue->q_exts_by_addr));
2870 next_sio = AVL_NEXT(&queue->q_sios_by_addr, sio);
2871 avl_remove(&queue->q_sios_by_addr, sio);
2872 queue->q_sio_memused -= SIO_GET_MUSED(sio);
2874 bytes_issued += SIO_GET_ASIZE(sio);
2876 list_insert_tail(list, sio);
2881 * We limit the number of sios we process at once to 32 to avoid
2882 * biting off more than we can chew. If we didn't take everything
2883 * in the segment we update it to reflect the work we were able to
2884 * complete. Otherwise, we remove it from the range tree entirely.
2886 if (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2887 queue->q_exts_by_addr)) {
2888 range_tree_adjust_fill(queue->q_exts_by_addr, rs,
2890 range_tree_resize_segment(queue->q_exts_by_addr, rs,
2891 SIO_GET_OFFSET(sio), rs_get_end(rs,
2892 queue->q_exts_by_addr) - SIO_GET_OFFSET(sio));
2896 uint64_t rstart = rs_get_start(rs, queue->q_exts_by_addr);
2897 uint64_t rend = rs_get_end(rs, queue->q_exts_by_addr);
2898 range_tree_remove(queue->q_exts_by_addr, rstart, rend - rstart);
2904 * This is called from the queue emptying thread and selects the next
2905 * extent from which we are to issue I/Os. The behavior of this function
2906 * depends on the state of the scan, the current memory consumption and
2907 * whether or not we are performing a scan shutdown.
2908 * 1) We select extents in an elevator algorithm (LBA-order) if the scan
2909 * needs to perform a checkpoint
2910 * 2) We select the largest available extent if we are up against the
2912 * 3) Otherwise we don't select any extents.
2914 static range_seg_t *
2915 scan_io_queue_fetch_ext(dsl_scan_io_queue_t *queue)
2917 dsl_scan_t *scn = queue->q_scn;
2918 range_tree_t *rt = queue->q_exts_by_addr;
2920 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2921 ASSERT(scn->scn_is_sorted);
2923 /* handle tunable overrides */
2924 if (scn->scn_checkpointing || scn->scn_clearing) {
2925 if (zfs_scan_issue_strategy == 1) {
2926 return (range_tree_first(rt));
2927 } else if (zfs_scan_issue_strategy == 2) {
2929 * We need to get the original entry in the by_addr
2930 * tree so we can modify it.
2932 range_seg_t *size_rs =
2933 zfs_btree_first(&queue->q_exts_by_size, NULL);
2934 if (size_rs == NULL)
2936 uint64_t start = rs_get_start(size_rs, rt);
2937 uint64_t size = rs_get_end(size_rs, rt) - start;
2938 range_seg_t *addr_rs = range_tree_find(rt, start,
2940 ASSERT3P(addr_rs, !=, NULL);
2941 ASSERT3U(rs_get_start(size_rs, rt), ==,
2942 rs_get_start(addr_rs, rt));
2943 ASSERT3U(rs_get_end(size_rs, rt), ==,
2944 rs_get_end(addr_rs, rt));
2950 * During normal clearing, we want to issue our largest segments
2951 * first, keeping IO as sequential as possible, and leaving the
2952 * smaller extents for later with the hope that they might eventually
2953 * grow to larger sequential segments. However, when the scan is
2954 * checkpointing, no new extents will be added to the sorting queue,
2955 * so the way we are sorted now is as good as it will ever get.
2956 * In this case, we instead switch to issuing extents in LBA order.
2958 if (scn->scn_checkpointing) {
2959 return (range_tree_first(rt));
2960 } else if (scn->scn_clearing) {
2962 * We need to get the original entry in the by_addr
2963 * tree so we can modify it.
2965 range_seg_t *size_rs = zfs_btree_first(&queue->q_exts_by_size,
2967 if (size_rs == NULL)
2969 uint64_t start = rs_get_start(size_rs, rt);
2970 uint64_t size = rs_get_end(size_rs, rt) - start;
2971 range_seg_t *addr_rs = range_tree_find(rt, start, size);
2972 ASSERT3P(addr_rs, !=, NULL);
2973 ASSERT3U(rs_get_start(size_rs, rt), ==, rs_get_start(addr_rs,
2975 ASSERT3U(rs_get_end(size_rs, rt), ==, rs_get_end(addr_rs, rt));
2983 scan_io_queues_run_one(void *arg)
2985 dsl_scan_io_queue_t *queue = arg;
2986 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
2987 boolean_t suspended = B_FALSE;
2988 range_seg_t *rs = NULL;
2989 scan_io_t *sio = NULL;
2991 uint64_t bytes_per_leaf = zfs_scan_vdev_limit;
2992 uint64_t nr_leaves = dsl_scan_count_leaves(queue->q_vd);
2994 ASSERT(queue->q_scn->scn_is_sorted);
2996 list_create(&sio_list, sizeof (scan_io_t),
2997 offsetof(scan_io_t, sio_nodes.sio_list_node));
2998 mutex_enter(q_lock);
3000 /* calculate maximum in-flight bytes for this txg (min 1MB) */
3001 queue->q_maxinflight_bytes =
3002 MAX(nr_leaves * bytes_per_leaf, 1ULL << 20);
3004 /* reset per-queue scan statistics for this txg */
3005 queue->q_total_seg_size_this_txg = 0;
3006 queue->q_segs_this_txg = 0;
3007 queue->q_total_zio_size_this_txg = 0;
3008 queue->q_zios_this_txg = 0;
3010 /* loop until we run out of time or sios */
3011 while ((rs = scan_io_queue_fetch_ext(queue)) != NULL) {
3012 uint64_t seg_start = 0, seg_end = 0;
3013 boolean_t more_left = B_TRUE;
3015 ASSERT(list_is_empty(&sio_list));
3017 /* loop while we still have sios left to process in this rs */
3019 scan_io_t *first_sio, *last_sio;
3022 * We have selected which extent needs to be
3023 * processed next. Gather up the corresponding sios.
3025 more_left = scan_io_queue_gather(queue, rs, &sio_list);
3026 ASSERT(!list_is_empty(&sio_list));
3027 first_sio = list_head(&sio_list);
3028 last_sio = list_tail(&sio_list);
3030 seg_end = SIO_GET_END_OFFSET(last_sio);
3032 seg_start = SIO_GET_OFFSET(first_sio);
3035 * Issuing sios can take a long time so drop the
3036 * queue lock. The sio queue won't be updated by
3037 * other threads since we're in syncing context so
3038 * we can be sure that our trees will remain exactly
3042 suspended = scan_io_queue_issue(queue, &sio_list);
3043 mutex_enter(q_lock);
3049 /* update statistics for debugging purposes */
3050 scan_io_queues_update_seg_stats(queue, seg_start, seg_end);
3057 * If we were suspended in the middle of processing,
3058 * requeue any unfinished sios and exit.
3060 while ((sio = list_head(&sio_list)) != NULL) {
3061 list_remove(&sio_list, sio);
3062 scan_io_queue_insert_impl(queue, sio);
3066 list_destroy(&sio_list);
3070 * Performs an emptying run on all scan queues in the pool. This just
3071 * punches out one thread per top-level vdev, each of which processes
3072 * only that vdev's scan queue. We can parallelize the I/O here because
3073 * we know that each queue's I/Os only affect its own top-level vdev.
3075 * This function waits for the queue runs to complete, and must be
3076 * called from dsl_scan_sync (or in general, syncing context).
3079 scan_io_queues_run(dsl_scan_t *scn)
3081 spa_t *spa = scn->scn_dp->dp_spa;
3083 ASSERT(scn->scn_is_sorted);
3084 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
3086 if (scn->scn_bytes_pending == 0)
3089 if (scn->scn_taskq == NULL) {
3090 int nthreads = spa->spa_root_vdev->vdev_children;
3093 * We need to make this taskq *always* execute as many
3094 * threads in parallel as we have top-level vdevs and no
3095 * less, otherwise strange serialization of the calls to
3096 * scan_io_queues_run_one can occur during spa_sync runs
3097 * and that significantly impacts performance.
3099 scn->scn_taskq = taskq_create("dsl_scan_iss", nthreads,
3100 minclsyspri, nthreads, nthreads, TASKQ_PREPOPULATE);
3103 for (uint64_t i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3104 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3106 mutex_enter(&vd->vdev_scan_io_queue_lock);
3107 if (vd->vdev_scan_io_queue != NULL) {
3108 VERIFY(taskq_dispatch(scn->scn_taskq,
3109 scan_io_queues_run_one, vd->vdev_scan_io_queue,
3110 TQ_SLEEP) != TASKQID_INVALID);
3112 mutex_exit(&vd->vdev_scan_io_queue_lock);
3116 * Wait for the queues to finish issuing their IOs for this run
3117 * before we return. There may still be IOs in flight at this
3120 taskq_wait(scn->scn_taskq);
3124 dsl_scan_async_block_should_pause(dsl_scan_t *scn)
3126 uint64_t elapsed_nanosecs;
3131 if (zfs_async_block_max_blocks != 0 &&
3132 scn->scn_visited_this_txg >= zfs_async_block_max_blocks) {
3136 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
3137 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
3138 (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms &&
3139 txg_sync_waiting(scn->scn_dp)) ||
3140 spa_shutting_down(scn->scn_dp->dp_spa));
3144 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
3146 dsl_scan_t *scn = arg;
3148 if (!scn->scn_is_bptree ||
3149 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
3150 if (dsl_scan_async_block_should_pause(scn))
3151 return (SET_ERROR(ERESTART));
3154 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
3155 dmu_tx_get_txg(tx), bp, 0));
3156 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
3157 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
3158 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
3159 scn->scn_visited_this_txg++;
3164 dsl_scan_update_stats(dsl_scan_t *scn)
3166 spa_t *spa = scn->scn_dp->dp_spa;
3168 uint64_t seg_size_total = 0, zio_size_total = 0;
3169 uint64_t seg_count_total = 0, zio_count_total = 0;
3171 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3172 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3173 dsl_scan_io_queue_t *queue = vd->vdev_scan_io_queue;
3178 seg_size_total += queue->q_total_seg_size_this_txg;
3179 zio_size_total += queue->q_total_zio_size_this_txg;
3180 seg_count_total += queue->q_segs_this_txg;
3181 zio_count_total += queue->q_zios_this_txg;
3184 if (seg_count_total == 0 || zio_count_total == 0) {
3185 scn->scn_avg_seg_size_this_txg = 0;
3186 scn->scn_avg_zio_size_this_txg = 0;
3187 scn->scn_segs_this_txg = 0;
3188 scn->scn_zios_this_txg = 0;
3192 scn->scn_avg_seg_size_this_txg = seg_size_total / seg_count_total;
3193 scn->scn_avg_zio_size_this_txg = zio_size_total / zio_count_total;
3194 scn->scn_segs_this_txg = seg_count_total;
3195 scn->scn_zios_this_txg = zio_count_total;
3199 bpobj_dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3203 return (dsl_scan_free_block_cb(arg, bp, tx));
3207 dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3211 dsl_scan_t *scn = arg;
3212 const dva_t *dva = &bp->blk_dva[0];
3214 if (dsl_scan_async_block_should_pause(scn))
3215 return (SET_ERROR(ERESTART));
3217 spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa,
3218 DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva),
3219 DVA_GET_ASIZE(dva), tx);
3220 scn->scn_visited_this_txg++;
3225 dsl_scan_active(dsl_scan_t *scn)
3227 spa_t *spa = scn->scn_dp->dp_spa;
3228 uint64_t used = 0, comp, uncomp;
3229 boolean_t clones_left;
3231 if (spa->spa_load_state != SPA_LOAD_NONE)
3233 if (spa_shutting_down(spa))
3235 if ((dsl_scan_is_running(scn) && !dsl_scan_is_paused_scrub(scn)) ||
3236 (scn->scn_async_destroying && !scn->scn_async_stalled))
3239 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
3240 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
3241 &used, &comp, &uncomp);
3243 clones_left = spa_livelist_delete_check(spa);
3244 return ((used != 0) || (clones_left));
3248 dsl_scan_check_deferred(vdev_t *vd)
3250 boolean_t need_resilver = B_FALSE;
3252 for (int c = 0; c < vd->vdev_children; c++) {
3254 dsl_scan_check_deferred(vd->vdev_child[c]);
3257 if (!vdev_is_concrete(vd) || vd->vdev_aux ||
3258 !vd->vdev_ops->vdev_op_leaf)
3259 return (need_resilver);
3261 if (!vd->vdev_resilver_deferred)
3262 need_resilver = B_TRUE;
3264 return (need_resilver);
3268 dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize,
3269 uint64_t phys_birth)
3273 vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
3275 if (vd->vdev_ops == &vdev_indirect_ops) {
3277 * The indirect vdev can point to multiple
3278 * vdevs. For simplicity, always create
3279 * the resilver zio_t. zio_vdev_io_start()
3280 * will bypass the child resilver i/o's if
3281 * they are on vdevs that don't have DTL's.
3286 if (DVA_GET_GANG(dva)) {
3288 * Gang members may be spread across multiple
3289 * vdevs, so the best estimate we have is the
3290 * scrub range, which has already been checked.
3291 * XXX -- it would be better to change our
3292 * allocation policy to ensure that all
3293 * gang members reside on the same vdev.
3299 * Check if the txg falls within the range which must be
3300 * resilvered. DVAs outside this range can always be skipped.
3302 if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1))
3306 * Check if the top-level vdev must resilver this offset.
3307 * When the offset does not intersect with a dirty leaf DTL
3308 * then it may be possible to skip the resilver IO. The psize
3309 * is provided instead of asize to simplify the check for RAIDZ.
3311 if (!vdev_dtl_need_resilver(vd, DVA_GET_OFFSET(dva), psize))
3315 * Check that this top-level vdev has a device under it which
3316 * is resilvering and is not deferred.
3318 if (!dsl_scan_check_deferred(vd))
3325 dsl_process_async_destroys(dsl_pool_t *dp, dmu_tx_t *tx)
3327 dsl_scan_t *scn = dp->dp_scan;
3328 spa_t *spa = dp->dp_spa;
3331 if (spa_suspend_async_destroy(spa))
3334 if (zfs_free_bpobj_enabled &&
3335 spa_version(spa) >= SPA_VERSION_DEADLISTS) {
3336 scn->scn_is_bptree = B_FALSE;
3337 scn->scn_async_block_min_time_ms = zfs_free_min_time_ms;
3338 scn->scn_zio_root = zio_root(spa, NULL,
3339 NULL, ZIO_FLAG_MUSTSUCCEED);
3340 err = bpobj_iterate(&dp->dp_free_bpobj,
3341 bpobj_dsl_scan_free_block_cb, scn, tx);
3342 VERIFY0(zio_wait(scn->scn_zio_root));
3343 scn->scn_zio_root = NULL;
3345 if (err != 0 && err != ERESTART)
3346 zfs_panic_recover("error %u from bpobj_iterate()", err);
3349 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
3350 ASSERT(scn->scn_async_destroying);
3351 scn->scn_is_bptree = B_TRUE;
3352 scn->scn_zio_root = zio_root(spa, NULL,
3353 NULL, ZIO_FLAG_MUSTSUCCEED);
3354 err = bptree_iterate(dp->dp_meta_objset,
3355 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
3356 VERIFY0(zio_wait(scn->scn_zio_root));
3357 scn->scn_zio_root = NULL;
3359 if (err == EIO || err == ECKSUM) {
3361 } else if (err != 0 && err != ERESTART) {
3362 zfs_panic_recover("error %u from "
3363 "traverse_dataset_destroyed()", err);
3366 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
3367 /* finished; deactivate async destroy feature */
3368 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
3369 ASSERT(!spa_feature_is_active(spa,
3370 SPA_FEATURE_ASYNC_DESTROY));
3371 VERIFY0(zap_remove(dp->dp_meta_objset,
3372 DMU_POOL_DIRECTORY_OBJECT,
3373 DMU_POOL_BPTREE_OBJ, tx));
3374 VERIFY0(bptree_free(dp->dp_meta_objset,
3375 dp->dp_bptree_obj, tx));
3376 dp->dp_bptree_obj = 0;
3377 scn->scn_async_destroying = B_FALSE;
3378 scn->scn_async_stalled = B_FALSE;
3381 * If we didn't make progress, mark the async
3382 * destroy as stalled, so that we will not initiate
3383 * a spa_sync() on its behalf. Note that we only
3384 * check this if we are not finished, because if the
3385 * bptree had no blocks for us to visit, we can
3386 * finish without "making progress".
3388 scn->scn_async_stalled =
3389 (scn->scn_visited_this_txg == 0);
3392 if (scn->scn_visited_this_txg) {
3393 zfs_dbgmsg("freed %llu blocks in %llums from "
3394 "free_bpobj/bptree txg %llu; err=%u",
3395 (longlong_t)scn->scn_visited_this_txg,
3397 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
3398 (longlong_t)tx->tx_txg, err);
3399 scn->scn_visited_this_txg = 0;
3402 * Write out changes to the DDT that may be required as a
3403 * result of the blocks freed. This ensures that the DDT
3404 * is clean when a scrub/resilver runs.
3406 ddt_sync(spa, tx->tx_txg);
3410 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3411 zfs_free_leak_on_eio &&
3412 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
3413 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
3414 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
3416 * We have finished background destroying, but there is still
3417 * some space left in the dp_free_dir. Transfer this leaked
3418 * space to the dp_leak_dir.
3420 if (dp->dp_leak_dir == NULL) {
3421 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
3422 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
3424 VERIFY0(dsl_pool_open_special_dir(dp,
3425 LEAK_DIR_NAME, &dp->dp_leak_dir));
3426 rrw_exit(&dp->dp_config_rwlock, FTAG);
3428 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
3429 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3430 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3431 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3432 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
3433 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3434 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3435 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3438 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3439 !spa_livelist_delete_check(spa)) {
3440 /* finished; verify that space accounting went to zero */
3441 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
3442 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
3443 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
3446 spa_notify_waiters(spa);
3448 EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj),
3449 0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3450 DMU_POOL_OBSOLETE_BPOBJ));
3451 if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) {
3452 ASSERT(spa_feature_is_active(dp->dp_spa,
3453 SPA_FEATURE_OBSOLETE_COUNTS));
3455 scn->scn_is_bptree = B_FALSE;
3456 scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms;
3457 err = bpobj_iterate(&dp->dp_obsolete_bpobj,
3458 dsl_scan_obsolete_block_cb, scn, tx);
3459 if (err != 0 && err != ERESTART)
3460 zfs_panic_recover("error %u from bpobj_iterate()", err);
3462 if (bpobj_is_empty(&dp->dp_obsolete_bpobj))
3463 dsl_pool_destroy_obsolete_bpobj(dp, tx);
3469 * This is the primary entry point for scans that is called from syncing
3470 * context. Scans must happen entirely during syncing context so that we
3471 * can guarantee that blocks we are currently scanning will not change out
3472 * from under us. While a scan is active, this function controls how quickly
3473 * transaction groups proceed, instead of the normal handling provided by
3474 * txg_sync_thread().
3477 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
3480 dsl_scan_t *scn = dp->dp_scan;
3481 spa_t *spa = dp->dp_spa;
3482 state_sync_type_t sync_type = SYNC_OPTIONAL;
3484 if (spa->spa_resilver_deferred &&
3485 !spa_feature_is_active(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
3486 spa_feature_incr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
3489 * Check for scn_restart_txg before checking spa_load_state, so
3490 * that we can restart an old-style scan while the pool is being
3491 * imported (see dsl_scan_init). We also restart scans if there
3492 * is a deferred resilver and the user has manually disabled
3493 * deferred resilvers via the tunable.
3495 if (dsl_scan_restarting(scn, tx) ||
3496 (spa->spa_resilver_deferred && zfs_resilver_disable_defer)) {
3497 pool_scan_func_t func = POOL_SCAN_SCRUB;
3498 dsl_scan_done(scn, B_FALSE, tx);
3499 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
3500 func = POOL_SCAN_RESILVER;
3501 zfs_dbgmsg("restarting scan func=%u txg=%llu",
3502 func, (longlong_t)tx->tx_txg);
3503 dsl_scan_setup_sync(&func, tx);
3507 * Only process scans in sync pass 1.
3509 if (spa_sync_pass(spa) > 1)
3513 * If the spa is shutting down, then stop scanning. This will
3514 * ensure that the scan does not dirty any new data during the
3517 if (spa_shutting_down(spa))
3521 * If the scan is inactive due to a stalled async destroy, try again.
3523 if (!scn->scn_async_stalled && !dsl_scan_active(scn))
3526 /* reset scan statistics */
3527 scn->scn_visited_this_txg = 0;
3528 scn->scn_holes_this_txg = 0;
3529 scn->scn_lt_min_this_txg = 0;
3530 scn->scn_gt_max_this_txg = 0;
3531 scn->scn_ddt_contained_this_txg = 0;
3532 scn->scn_objsets_visited_this_txg = 0;
3533 scn->scn_avg_seg_size_this_txg = 0;
3534 scn->scn_segs_this_txg = 0;
3535 scn->scn_avg_zio_size_this_txg = 0;
3536 scn->scn_zios_this_txg = 0;
3537 scn->scn_suspending = B_FALSE;
3538 scn->scn_sync_start_time = gethrtime();
3539 spa->spa_scrub_active = B_TRUE;
3542 * First process the async destroys. If we suspend, don't do
3543 * any scrubbing or resilvering. This ensures that there are no
3544 * async destroys while we are scanning, so the scan code doesn't
3545 * have to worry about traversing it. It is also faster to free the
3546 * blocks than to scrub them.
3548 err = dsl_process_async_destroys(dp, tx);
3552 if (!dsl_scan_is_running(scn) || dsl_scan_is_paused_scrub(scn))
3556 * Wait a few txgs after importing to begin scanning so that
3557 * we can get the pool imported quickly.
3559 if (spa->spa_syncing_txg < spa->spa_first_txg + SCAN_IMPORT_WAIT_TXGS)
3563 * zfs_scan_suspend_progress can be set to disable scan progress.
3564 * We don't want to spin the txg_sync thread, so we add a delay
3565 * here to simulate the time spent doing a scan. This is mostly
3566 * useful for testing and debugging.
3568 if (zfs_scan_suspend_progress) {
3569 uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time;
3570 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
3571 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
3573 while (zfs_scan_suspend_progress &&
3574 !txg_sync_waiting(scn->scn_dp) &&
3575 !spa_shutting_down(scn->scn_dp->dp_spa) &&
3576 NSEC2MSEC(scan_time_ns) < mintime) {
3578 scan_time_ns = gethrtime() - scn->scn_sync_start_time;
3584 * It is possible to switch from unsorted to sorted at any time,
3585 * but afterwards the scan will remain sorted unless reloaded from
3586 * a checkpoint after a reboot.
3588 if (!zfs_scan_legacy) {
3589 scn->scn_is_sorted = B_TRUE;
3590 if (scn->scn_last_checkpoint == 0)
3591 scn->scn_last_checkpoint = ddi_get_lbolt();
3595 * For sorted scans, determine what kind of work we will be doing
3596 * this txg based on our memory limitations and whether or not we
3597 * need to perform a checkpoint.
3599 if (scn->scn_is_sorted) {
3601 * If we are over our checkpoint interval, set scn_clearing
3602 * so that we can begin checkpointing immediately. The
3603 * checkpoint allows us to save a consistent bookmark
3604 * representing how much data we have scrubbed so far.
3605 * Otherwise, use the memory limit to determine if we should
3606 * scan for metadata or start issue scrub IOs. We accumulate
3607 * metadata until we hit our hard memory limit at which point
3608 * we issue scrub IOs until we are at our soft memory limit.
3610 if (scn->scn_checkpointing ||
3611 ddi_get_lbolt() - scn->scn_last_checkpoint >
3612 SEC_TO_TICK(zfs_scan_checkpoint_intval)) {
3613 if (!scn->scn_checkpointing)
3614 zfs_dbgmsg("begin scan checkpoint");
3616 scn->scn_checkpointing = B_TRUE;
3617 scn->scn_clearing = B_TRUE;
3619 boolean_t should_clear = dsl_scan_should_clear(scn);
3620 if (should_clear && !scn->scn_clearing) {
3621 zfs_dbgmsg("begin scan clearing");
3622 scn->scn_clearing = B_TRUE;
3623 } else if (!should_clear && scn->scn_clearing) {
3624 zfs_dbgmsg("finish scan clearing");
3625 scn->scn_clearing = B_FALSE;
3629 ASSERT0(scn->scn_checkpointing);
3630 ASSERT0(scn->scn_clearing);
3633 if (!scn->scn_clearing && scn->scn_done_txg == 0) {
3634 /* Need to scan metadata for more blocks to scrub */
3635 dsl_scan_phys_t *scnp = &scn->scn_phys;
3636 taskqid_t prefetch_tqid;
3637 uint64_t bytes_per_leaf = zfs_scan_vdev_limit;
3638 uint64_t nr_leaves = dsl_scan_count_leaves(spa->spa_root_vdev);
3641 * Recalculate the max number of in-flight bytes for pool-wide
3642 * scanning operations (minimum 1MB). Limits for the issuing
3643 * phase are done per top-level vdev and are handled separately.
3645 scn->scn_maxinflight_bytes =
3646 MAX(nr_leaves * bytes_per_leaf, 1ULL << 20);
3648 if (scnp->scn_ddt_bookmark.ddb_class <=
3649 scnp->scn_ddt_class_max) {
3650 ASSERT(ZB_IS_ZERO(&scnp->scn_bookmark));
3651 zfs_dbgmsg("doing scan sync txg %llu; "
3652 "ddt bm=%llu/%llu/%llu/%llx",
3653 (longlong_t)tx->tx_txg,
3654 (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
3655 (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
3656 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
3657 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
3659 zfs_dbgmsg("doing scan sync txg %llu; "
3660 "bm=%llu/%llu/%llu/%llu",
3661 (longlong_t)tx->tx_txg,
3662 (longlong_t)scnp->scn_bookmark.zb_objset,
3663 (longlong_t)scnp->scn_bookmark.zb_object,
3664 (longlong_t)scnp->scn_bookmark.zb_level,
3665 (longlong_t)scnp->scn_bookmark.zb_blkid);
3668 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3669 NULL, ZIO_FLAG_CANFAIL);
3671 scn->scn_prefetch_stop = B_FALSE;
3672 prefetch_tqid = taskq_dispatch(dp->dp_sync_taskq,
3673 dsl_scan_prefetch_thread, scn, TQ_SLEEP);
3674 ASSERT(prefetch_tqid != TASKQID_INVALID);
3676 dsl_pool_config_enter(dp, FTAG);
3677 dsl_scan_visit(scn, tx);
3678 dsl_pool_config_exit(dp, FTAG);
3680 mutex_enter(&dp->dp_spa->spa_scrub_lock);
3681 scn->scn_prefetch_stop = B_TRUE;
3682 cv_broadcast(&spa->spa_scrub_io_cv);
3683 mutex_exit(&dp->dp_spa->spa_scrub_lock);
3685 taskq_wait_id(dp->dp_sync_taskq, prefetch_tqid);
3686 (void) zio_wait(scn->scn_zio_root);
3687 scn->scn_zio_root = NULL;
3689 zfs_dbgmsg("scan visited %llu blocks in %llums "
3690 "(%llu os's, %llu holes, %llu < mintxg, "
3691 "%llu in ddt, %llu > maxtxg)",
3692 (longlong_t)scn->scn_visited_this_txg,
3693 (longlong_t)NSEC2MSEC(gethrtime() -
3694 scn->scn_sync_start_time),
3695 (longlong_t)scn->scn_objsets_visited_this_txg,
3696 (longlong_t)scn->scn_holes_this_txg,
3697 (longlong_t)scn->scn_lt_min_this_txg,
3698 (longlong_t)scn->scn_ddt_contained_this_txg,
3699 (longlong_t)scn->scn_gt_max_this_txg);
3701 if (!scn->scn_suspending) {
3702 ASSERT0(avl_numnodes(&scn->scn_queue));
3703 scn->scn_done_txg = tx->tx_txg + 1;
3704 if (scn->scn_is_sorted) {
3705 scn->scn_checkpointing = B_TRUE;
3706 scn->scn_clearing = B_TRUE;
3708 zfs_dbgmsg("scan complete txg %llu",
3709 (longlong_t)tx->tx_txg);
3711 } else if (scn->scn_is_sorted && scn->scn_bytes_pending != 0) {
3712 ASSERT(scn->scn_clearing);
3714 /* need to issue scrubbing IOs from per-vdev queues */
3715 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3716 NULL, ZIO_FLAG_CANFAIL);
3717 scan_io_queues_run(scn);
3718 (void) zio_wait(scn->scn_zio_root);
3719 scn->scn_zio_root = NULL;
3721 /* calculate and dprintf the current memory usage */
3722 (void) dsl_scan_should_clear(scn);
3723 dsl_scan_update_stats(scn);
3725 zfs_dbgmsg("scan issued %llu blocks (%llu segs) in %llums "
3726 "(avg_block_size = %llu, avg_seg_size = %llu)",
3727 (longlong_t)scn->scn_zios_this_txg,
3728 (longlong_t)scn->scn_segs_this_txg,
3729 (longlong_t)NSEC2MSEC(gethrtime() -
3730 scn->scn_sync_start_time),
3731 (longlong_t)scn->scn_avg_zio_size_this_txg,
3732 (longlong_t)scn->scn_avg_seg_size_this_txg);
3733 } else if (scn->scn_done_txg != 0 && scn->scn_done_txg <= tx->tx_txg) {
3734 /* Finished with everything. Mark the scrub as complete */
3735 zfs_dbgmsg("scan issuing complete txg %llu",
3736 (longlong_t)tx->tx_txg);
3737 ASSERT3U(scn->scn_done_txg, !=, 0);
3738 ASSERT0(spa->spa_scrub_inflight);
3739 ASSERT0(scn->scn_bytes_pending);
3740 dsl_scan_done(scn, B_TRUE, tx);
3741 sync_type = SYNC_MANDATORY;
3744 dsl_scan_sync_state(scn, tx, sync_type);
3748 count_block(dsl_scan_t *scn, zfs_all_blkstats_t *zab, const blkptr_t *bp)
3753 * Don't count embedded bp's, since we already did the work of
3754 * scanning these when we scanned the containing block.
3756 if (BP_IS_EMBEDDED(bp))
3760 * Update the spa's stats on how many bytes we have issued.
3761 * Sequential scrubs create a zio for each DVA of the bp. Each
3762 * of these will include all DVAs for repair purposes, but the
3763 * zio code will only try the first one unless there is an issue.
3764 * Therefore, we should only count the first DVA for these IOs.
3766 if (scn->scn_is_sorted) {
3767 atomic_add_64(&scn->scn_dp->dp_spa->spa_scan_pass_issued,
3768 DVA_GET_ASIZE(&bp->blk_dva[0]));
3770 spa_t *spa = scn->scn_dp->dp_spa;
3772 for (i = 0; i < BP_GET_NDVAS(bp); i++) {
3773 atomic_add_64(&spa->spa_scan_pass_issued,
3774 DVA_GET_ASIZE(&bp->blk_dva[i]));
3779 * If we resume after a reboot, zab will be NULL; don't record
3780 * incomplete stats in that case.
3785 mutex_enter(&zab->zab_lock);
3787 for (i = 0; i < 4; i++) {
3788 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
3789 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
3791 if (t & DMU_OT_NEWTYPE)
3793 zfs_blkstat_t *zb = &zab->zab_type[l][t];
3797 zb->zb_asize += BP_GET_ASIZE(bp);
3798 zb->zb_lsize += BP_GET_LSIZE(bp);
3799 zb->zb_psize += BP_GET_PSIZE(bp);
3800 zb->zb_gangs += BP_COUNT_GANG(bp);
3802 switch (BP_GET_NDVAS(bp)) {
3804 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3805 DVA_GET_VDEV(&bp->blk_dva[1]))
3806 zb->zb_ditto_2_of_2_samevdev++;
3809 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3810 DVA_GET_VDEV(&bp->blk_dva[1])) +
3811 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3812 DVA_GET_VDEV(&bp->blk_dva[2])) +
3813 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
3814 DVA_GET_VDEV(&bp->blk_dva[2]));
3816 zb->zb_ditto_2_of_3_samevdev++;
3817 else if (equal == 3)
3818 zb->zb_ditto_3_of_3_samevdev++;
3823 mutex_exit(&zab->zab_lock);
3827 scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue, scan_io_t *sio)
3830 int64_t asize = SIO_GET_ASIZE(sio);
3831 dsl_scan_t *scn = queue->q_scn;
3833 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3835 if (avl_find(&queue->q_sios_by_addr, sio, &idx) != NULL) {
3836 /* block is already scheduled for reading */
3837 atomic_add_64(&scn->scn_bytes_pending, -asize);
3841 avl_insert(&queue->q_sios_by_addr, sio, idx);
3842 queue->q_sio_memused += SIO_GET_MUSED(sio);
3843 range_tree_add(queue->q_exts_by_addr, SIO_GET_OFFSET(sio), asize);
3847 * Given all the info we got from our metadata scanning process, we
3848 * construct a scan_io_t and insert it into the scan sorting queue. The
3849 * I/O must already be suitable for us to process. This is controlled
3850 * by dsl_scan_enqueue().
3853 scan_io_queue_insert(dsl_scan_io_queue_t *queue, const blkptr_t *bp, int dva_i,
3854 int zio_flags, const zbookmark_phys_t *zb)
3856 dsl_scan_t *scn = queue->q_scn;
3857 scan_io_t *sio = sio_alloc(BP_GET_NDVAS(bp));
3859 ASSERT0(BP_IS_GANG(bp));
3860 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3862 bp2sio(bp, sio, dva_i);
3863 sio->sio_flags = zio_flags;
3867 * Increment the bytes pending counter now so that we can't
3868 * get an integer underflow in case the worker processes the
3869 * zio before we get to incrementing this counter.
3871 atomic_add_64(&scn->scn_bytes_pending, SIO_GET_ASIZE(sio));
3873 scan_io_queue_insert_impl(queue, sio);
3877 * Given a set of I/O parameters as discovered by the metadata traversal
3878 * process, attempts to place the I/O into the sorted queues (if allowed),
3879 * or immediately executes the I/O.
3882 dsl_scan_enqueue(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
3883 const zbookmark_phys_t *zb)
3885 spa_t *spa = dp->dp_spa;
3887 ASSERT(!BP_IS_EMBEDDED(bp));
3890 * Gang blocks are hard to issue sequentially, so we just issue them
3891 * here immediately instead of queuing them.
3893 if (!dp->dp_scan->scn_is_sorted || BP_IS_GANG(bp)) {
3894 scan_exec_io(dp, bp, zio_flags, zb, NULL);
3898 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
3902 dva = bp->blk_dva[i];
3903 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&dva));
3904 ASSERT(vdev != NULL);
3906 mutex_enter(&vdev->vdev_scan_io_queue_lock);
3907 if (vdev->vdev_scan_io_queue == NULL)
3908 vdev->vdev_scan_io_queue = scan_io_queue_create(vdev);
3909 ASSERT(dp->dp_scan != NULL);
3910 scan_io_queue_insert(vdev->vdev_scan_io_queue, bp,
3912 mutex_exit(&vdev->vdev_scan_io_queue_lock);
3917 dsl_scan_scrub_cb(dsl_pool_t *dp,
3918 const blkptr_t *bp, const zbookmark_phys_t *zb)
3920 dsl_scan_t *scn = dp->dp_scan;
3921 spa_t *spa = dp->dp_spa;
3922 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
3923 size_t psize = BP_GET_PSIZE(bp);
3924 boolean_t needs_io = B_FALSE;
3925 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
3928 if (phys_birth <= scn->scn_phys.scn_min_txg ||
3929 phys_birth >= scn->scn_phys.scn_max_txg) {
3930 count_block(scn, dp->dp_blkstats, bp);
3934 /* Embedded BP's have phys_birth==0, so we reject them above. */
3935 ASSERT(!BP_IS_EMBEDDED(bp));
3937 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
3938 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
3939 zio_flags |= ZIO_FLAG_SCRUB;
3942 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
3943 zio_flags |= ZIO_FLAG_RESILVER;
3947 /* If it's an intent log block, failure is expected. */
3948 if (zb->zb_level == ZB_ZIL_LEVEL)
3949 zio_flags |= ZIO_FLAG_SPECULATIVE;
3951 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
3952 const dva_t *dva = &bp->blk_dva[d];
3955 * Keep track of how much data we've examined so that
3956 * zpool(1M) status can make useful progress reports.
3958 scn->scn_phys.scn_examined += DVA_GET_ASIZE(dva);
3959 spa->spa_scan_pass_exam += DVA_GET_ASIZE(dva);
3961 /* if it's a resilver, this may not be in the target range */
3963 needs_io = dsl_scan_need_resilver(spa, dva, psize,
3967 if (needs_io && !zfs_no_scrub_io) {
3968 dsl_scan_enqueue(dp, bp, zio_flags, zb);
3970 count_block(scn, dp->dp_blkstats, bp);
3973 /* do not relocate this block */
3978 dsl_scan_scrub_done(zio_t *zio)
3980 spa_t *spa = zio->io_spa;
3981 blkptr_t *bp = zio->io_bp;
3982 dsl_scan_io_queue_t *queue = zio->io_private;
3984 abd_free(zio->io_abd);
3986 if (queue == NULL) {
3987 mutex_enter(&spa->spa_scrub_lock);
3988 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
3989 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
3990 cv_broadcast(&spa->spa_scrub_io_cv);
3991 mutex_exit(&spa->spa_scrub_lock);
3993 mutex_enter(&queue->q_vd->vdev_scan_io_queue_lock);
3994 ASSERT3U(queue->q_inflight_bytes, >=, BP_GET_PSIZE(bp));
3995 queue->q_inflight_bytes -= BP_GET_PSIZE(bp);
3996 cv_broadcast(&queue->q_zio_cv);
3997 mutex_exit(&queue->q_vd->vdev_scan_io_queue_lock);
4000 if (zio->io_error && (zio->io_error != ECKSUM ||
4001 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
4002 atomic_inc_64(&spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors);
4007 * Given a scanning zio's information, executes the zio. The zio need
4008 * not necessarily be only sortable, this function simply executes the
4009 * zio, no matter what it is. The optional queue argument allows the
4010 * caller to specify that they want per top level vdev IO rate limiting
4011 * instead of the legacy global limiting.
4014 scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
4015 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue)
4017 spa_t *spa = dp->dp_spa;
4018 dsl_scan_t *scn = dp->dp_scan;
4019 size_t size = BP_GET_PSIZE(bp);
4020 abd_t *data = abd_alloc_for_io(size, B_FALSE);
4022 ASSERT3U(scn->scn_maxinflight_bytes, >, 0);
4024 if (queue == NULL) {
4025 mutex_enter(&spa->spa_scrub_lock);
4026 while (spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)
4027 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
4028 spa->spa_scrub_inflight += BP_GET_PSIZE(bp);
4029 mutex_exit(&spa->spa_scrub_lock);
4031 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
4033 mutex_enter(q_lock);
4034 while (queue->q_inflight_bytes >= queue->q_maxinflight_bytes)
4035 cv_wait(&queue->q_zio_cv, q_lock);
4036 queue->q_inflight_bytes += BP_GET_PSIZE(bp);
4040 count_block(scn, dp->dp_blkstats, bp);
4041 zio_nowait(zio_read(scn->scn_zio_root, spa, bp, data, size,
4042 dsl_scan_scrub_done, queue, ZIO_PRIORITY_SCRUB, zio_flags, zb));
4046 * This is the primary extent sorting algorithm. We balance two parameters:
4047 * 1) how many bytes of I/O are in an extent
4048 * 2) how well the extent is filled with I/O (as a fraction of its total size)
4049 * Since we allow extents to have gaps between their constituent I/Os, it's
4050 * possible to have a fairly large extent that contains the same amount of
4051 * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
4052 * The algorithm sorts based on a score calculated from the extent's size,
4053 * the relative fill volume (in %) and a "fill weight" parameter that controls
4054 * the split between whether we prefer larger extents or more well populated
4057 * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
4060 * 1) assume extsz = 64 MiB
4061 * 2) assume fill = 32 MiB (extent is half full)
4062 * 3) assume fill_weight = 3
4063 * 4) SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
4064 * SCORE = 32M + (50 * 3 * 32M) / 100
4065 * SCORE = 32M + (4800M / 100)
4068 * | +--- final total relative fill-based score
4069 * +--------- final total fill-based score
4072 * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
4073 * extents that are more completely filled (in a 3:2 ratio) vs just larger.
4074 * Note that as an optimization, we replace multiplication and division by
4075 * 100 with bitshifting by 7 (which effectively multiplies and divides by 128).
4078 ext_size_compare(const void *x, const void *y)
4080 const range_seg_gap_t *rsa = x, *rsb = y;
4082 uint64_t sa = rsa->rs_end - rsa->rs_start;
4083 uint64_t sb = rsb->rs_end - rsb->rs_start;
4084 uint64_t score_a, score_b;
4086 score_a = rsa->rs_fill + ((((rsa->rs_fill << 7) / sa) *
4087 fill_weight * rsa->rs_fill) >> 7);
4088 score_b = rsb->rs_fill + ((((rsb->rs_fill << 7) / sb) *
4089 fill_weight * rsb->rs_fill) >> 7);
4091 if (score_a > score_b)
4093 if (score_a == score_b) {
4094 if (rsa->rs_start < rsb->rs_start)
4096 if (rsa->rs_start == rsb->rs_start)
4104 * Comparator for the q_sios_by_addr tree. Sorting is simply performed
4105 * based on LBA-order (from lowest to highest).
4108 sio_addr_compare(const void *x, const void *y)
4110 const scan_io_t *a = x, *b = y;
4112 return (TREE_CMP(SIO_GET_OFFSET(a), SIO_GET_OFFSET(b)));
4115 /* IO queues are created on demand when they are needed. */
4116 static dsl_scan_io_queue_t *
4117 scan_io_queue_create(vdev_t *vd)
4119 dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan;
4120 dsl_scan_io_queue_t *q = kmem_zalloc(sizeof (*q), KM_SLEEP);
4124 q->q_sio_memused = 0;
4125 cv_init(&q->q_zio_cv, NULL, CV_DEFAULT, NULL);
4126 q->q_exts_by_addr = range_tree_create_impl(&rt_btree_ops, RANGE_SEG_GAP,
4127 &q->q_exts_by_size, 0, 0, ext_size_compare, zfs_scan_max_ext_gap);
4128 avl_create(&q->q_sios_by_addr, sio_addr_compare,
4129 sizeof (scan_io_t), offsetof(scan_io_t, sio_nodes.sio_addr_node));
4135 * Destroys a scan queue and all segments and scan_io_t's contained in it.
4136 * No further execution of I/O occurs, anything pending in the queue is
4137 * simply freed without being executed.
4140 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t *queue)
4142 dsl_scan_t *scn = queue->q_scn;
4144 void *cookie = NULL;
4145 int64_t bytes_dequeued = 0;
4147 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
4149 while ((sio = avl_destroy_nodes(&queue->q_sios_by_addr, &cookie)) !=
4151 ASSERT(range_tree_contains(queue->q_exts_by_addr,
4152 SIO_GET_OFFSET(sio), SIO_GET_ASIZE(sio)));
4153 bytes_dequeued += SIO_GET_ASIZE(sio);
4154 queue->q_sio_memused -= SIO_GET_MUSED(sio);
4158 ASSERT0(queue->q_sio_memused);
4159 atomic_add_64(&scn->scn_bytes_pending, -bytes_dequeued);
4160 range_tree_vacate(queue->q_exts_by_addr, NULL, queue);
4161 range_tree_destroy(queue->q_exts_by_addr);
4162 avl_destroy(&queue->q_sios_by_addr);
4163 cv_destroy(&queue->q_zio_cv);
4165 kmem_free(queue, sizeof (*queue));
4169 * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
4170 * called on behalf of vdev_top_transfer when creating or destroying
4171 * a mirror vdev due to zpool attach/detach.
4174 dsl_scan_io_queue_vdev_xfer(vdev_t *svd, vdev_t *tvd)
4176 mutex_enter(&svd->vdev_scan_io_queue_lock);
4177 mutex_enter(&tvd->vdev_scan_io_queue_lock);
4179 VERIFY3P(tvd->vdev_scan_io_queue, ==, NULL);
4180 tvd->vdev_scan_io_queue = svd->vdev_scan_io_queue;
4181 svd->vdev_scan_io_queue = NULL;
4182 if (tvd->vdev_scan_io_queue != NULL)
4183 tvd->vdev_scan_io_queue->q_vd = tvd;
4185 mutex_exit(&tvd->vdev_scan_io_queue_lock);
4186 mutex_exit(&svd->vdev_scan_io_queue_lock);
4190 scan_io_queues_destroy(dsl_scan_t *scn)
4192 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
4194 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
4195 vdev_t *tvd = rvd->vdev_child[i];
4197 mutex_enter(&tvd->vdev_scan_io_queue_lock);
4198 if (tvd->vdev_scan_io_queue != NULL)
4199 dsl_scan_io_queue_destroy(tvd->vdev_scan_io_queue);
4200 tvd->vdev_scan_io_queue = NULL;
4201 mutex_exit(&tvd->vdev_scan_io_queue_lock);
4206 dsl_scan_freed_dva(spa_t *spa, const blkptr_t *bp, int dva_i)
4208 dsl_pool_t *dp = spa->spa_dsl_pool;
4209 dsl_scan_t *scn = dp->dp_scan;
4212 dsl_scan_io_queue_t *queue;
4213 scan_io_t *srch_sio, *sio;
4215 uint64_t start, size;
4217 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&bp->blk_dva[dva_i]));
4218 ASSERT(vdev != NULL);
4219 q_lock = &vdev->vdev_scan_io_queue_lock;
4220 queue = vdev->vdev_scan_io_queue;
4222 mutex_enter(q_lock);
4223 if (queue == NULL) {
4228 srch_sio = sio_alloc(BP_GET_NDVAS(bp));
4229 bp2sio(bp, srch_sio, dva_i);
4230 start = SIO_GET_OFFSET(srch_sio);
4231 size = SIO_GET_ASIZE(srch_sio);
4234 * We can find the zio in two states:
4235 * 1) Cold, just sitting in the queue of zio's to be issued at
4236 * some point in the future. In this case, all we do is
4237 * remove the zio from the q_sios_by_addr tree, decrement
4238 * its data volume from the containing range_seg_t and
4239 * resort the q_exts_by_size tree to reflect that the
4240 * range_seg_t has lost some of its 'fill'. We don't shorten
4241 * the range_seg_t - this is usually rare enough not to be
4242 * worth the extra hassle of trying keep track of precise
4243 * extent boundaries.
4244 * 2) Hot, where the zio is currently in-flight in
4245 * dsl_scan_issue_ios. In this case, we can't simply
4246 * reach in and stop the in-flight zio's, so we instead
4247 * block the caller. Eventually, dsl_scan_issue_ios will
4248 * be done with issuing the zio's it gathered and will
4251 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
4255 int64_t asize = SIO_GET_ASIZE(sio);
4258 /* Got it while it was cold in the queue */
4259 ASSERT3U(start, ==, SIO_GET_OFFSET(sio));
4260 ASSERT3U(size, ==, asize);
4261 avl_remove(&queue->q_sios_by_addr, sio);
4262 queue->q_sio_memused -= SIO_GET_MUSED(sio);
4264 ASSERT(range_tree_contains(queue->q_exts_by_addr, start, size));
4265 range_tree_remove_fill(queue->q_exts_by_addr, start, size);
4268 * We only update scn_bytes_pending in the cold path,
4269 * otherwise it will already have been accounted for as
4270 * part of the zio's execution.
4272 atomic_add_64(&scn->scn_bytes_pending, -asize);
4274 /* count the block as though we issued it */
4275 sio2bp(sio, &tmpbp);
4276 count_block(scn, dp->dp_blkstats, &tmpbp);
4284 * Callback invoked when a zio_free() zio is executing. This needs to be
4285 * intercepted to prevent the zio from deallocating a particular portion
4286 * of disk space and it then getting reallocated and written to, while we
4287 * still have it queued up for processing.
4290 dsl_scan_freed(spa_t *spa, const blkptr_t *bp)
4292 dsl_pool_t *dp = spa->spa_dsl_pool;
4293 dsl_scan_t *scn = dp->dp_scan;
4295 ASSERT(!BP_IS_EMBEDDED(bp));
4296 ASSERT(scn != NULL);
4297 if (!dsl_scan_is_running(scn))
4300 for (int i = 0; i < BP_GET_NDVAS(bp); i++)
4301 dsl_scan_freed_dva(spa, bp, i);
4305 ZFS_MODULE_PARAM(zfs, zfs_, scan_vdev_limit, ULONG, ZMOD_RW,
4306 "Max bytes in flight per leaf vdev for scrubs and resilvers");
4308 ZFS_MODULE_PARAM(zfs, zfs_, scrub_min_time_ms, INT, ZMOD_RW,
4309 "Min millisecs to scrub per txg");
4311 ZFS_MODULE_PARAM(zfs, zfs_, obsolete_min_time_ms, INT, ZMOD_RW,
4312 "Min millisecs to obsolete per txg");
4314 ZFS_MODULE_PARAM(zfs, zfs_, free_min_time_ms, INT, ZMOD_RW,
4315 "Min millisecs to free per txg");
4317 ZFS_MODULE_PARAM(zfs, zfs_, resilver_min_time_ms, INT, ZMOD_RW,
4318 "Min millisecs to resilver per txg");
4320 ZFS_MODULE_PARAM(zfs, zfs_, scan_suspend_progress, INT, ZMOD_RW,
4321 "Set to prevent scans from progressing");
4323 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_io, INT, ZMOD_RW,
4324 "Set to disable scrub I/O");
4326 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_prefetch, INT, ZMOD_RW,
4327 "Set to disable scrub prefetching");
4329 ZFS_MODULE_PARAM(zfs, zfs_, async_block_max_blocks, ULONG, ZMOD_RW,
4330 "Max number of blocks freed in one txg");
4332 ZFS_MODULE_PARAM(zfs, zfs_, free_bpobj_enabled, INT, ZMOD_RW,
4333 "Enable processing of the free_bpobj");
4335 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_fact, INT, ZMOD_RW,
4336 "Fraction of RAM for scan hard limit");
4338 ZFS_MODULE_PARAM(zfs, zfs_, scan_issue_strategy, INT, ZMOD_RW,
4339 "IO issuing strategy during scrubbing. "
4340 "0 = default, 1 = LBA, 2 = size");
4342 ZFS_MODULE_PARAM(zfs, zfs_, scan_legacy, INT, ZMOD_RW,
4343 "Scrub using legacy non-sequential method");
4345 ZFS_MODULE_PARAM(zfs, zfs_, scan_checkpoint_intval, INT, ZMOD_RW,
4346 "Scan progress on-disk checkpointing interval");
4348 ZFS_MODULE_PARAM(zfs, zfs_, scan_max_ext_gap, ULONG, ZMOD_RW,
4349 "Max gap in bytes between sequential scrub / resilver I/Os");
4351 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_soft_fact, INT, ZMOD_RW,
4352 "Fraction of hard limit used as soft limit");
4354 ZFS_MODULE_PARAM(zfs, zfs_, scan_strict_mem_lim, INT, ZMOD_RW,
4355 "Tunable to attempt to reduce lock contention");
4357 ZFS_MODULE_PARAM(zfs, zfs_, scan_fill_weight, INT, ZMOD_RW,
4358 "Tunable to adjust bias towards more filled segments during scans");
4360 ZFS_MODULE_PARAM(zfs, zfs_, resilver_disable_defer, INT, ZMOD_RW,
4361 "Process all resilvers immediately");