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, 2021 by Delphix. All rights reserved.
24 * Copyright 2016 Gary Mills
25 * Copyright (c) 2017, 2019, Datto Inc. All rights reserved.
26 * Copyright (c) 2015, Nexenta Systems, Inc. All rights reserved.
27 * Copyright 2019 Joyent, Inc.
30 #include <sys/dsl_scan.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_prop.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_synctask.h>
36 #include <sys/dnode.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/dmu_objset.h>
42 #include <sys/zfs_context.h>
43 #include <sys/fs/zfs.h>
44 #include <sys/zfs_znode.h>
45 #include <sys/spa_impl.h>
46 #include <sys/vdev_impl.h>
47 #include <sys/zil_impl.h>
48 #include <sys/zio_checksum.h>
51 #include <sys/sa_impl.h>
52 #include <sys/zfeature.h>
54 #include <sys/range_tree.h>
56 #include <sys/zfs_vfsops.h>
60 * Grand theory statement on scan queue sorting
62 * Scanning is implemented by recursively traversing all indirection levels
63 * in an object and reading all blocks referenced from said objects. This
64 * results in us approximately traversing the object from lowest logical
65 * offset to the highest. For best performance, we would want the logical
66 * blocks to be physically contiguous. However, this is frequently not the
67 * case with pools given the allocation patterns of copy-on-write filesystems.
68 * So instead, we put the I/Os into a reordering queue and issue them in a
69 * way that will most benefit physical disks (LBA-order).
73 * Ideally, we would want to scan all metadata and queue up all block I/O
74 * prior to starting to issue it, because that allows us to do an optimal
75 * sorting job. This can however consume large amounts of memory. Therefore
76 * we continuously monitor the size of the queues and constrain them to 5%
77 * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this
78 * limit, we clear out a few of the largest extents at the head of the queues
79 * to make room for more scanning. Hopefully, these extents will be fairly
80 * large and contiguous, allowing us to approach sequential I/O throughput
81 * even without a fully sorted tree.
83 * Metadata scanning takes place in dsl_scan_visit(), which is called from
84 * dsl_scan_sync() every spa_sync(). If we have either fully scanned all
85 * metadata on the pool, or we need to make room in memory because our
86 * queues are too large, dsl_scan_visit() is postponed and
87 * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies
88 * that metadata scanning and queued I/O issuing are mutually exclusive. This
89 * allows us to provide maximum sequential I/O throughput for the majority of
90 * I/O's issued since sequential I/O performance is significantly negatively
91 * impacted if it is interleaved with random I/O.
93 * Implementation Notes
95 * One side effect of the queued scanning algorithm is that the scanning code
96 * needs to be notified whenever a block is freed. This is needed to allow
97 * the scanning code to remove these I/Os from the issuing queue. Additionally,
98 * we do not attempt to queue gang blocks to be issued sequentially since this
99 * is very hard to do and would have an extremely limited performance benefit.
100 * Instead, we simply issue gang I/Os as soon as we find them using the legacy
103 * Backwards compatibility
105 * This new algorithm is backwards compatible with the legacy on-disk data
106 * structures (and therefore does not require a new feature flag).
107 * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan
108 * will stop scanning metadata (in logical order) and wait for all outstanding
109 * sorted I/O to complete. Once this is done, we write out a checkpoint
110 * bookmark, indicating that we have scanned everything logically before it.
111 * If the pool is imported on a machine without the new sorting algorithm,
112 * the scan simply resumes from the last checkpoint using the legacy algorithm.
115 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
116 const zbookmark_phys_t *);
118 static scan_cb_t dsl_scan_scrub_cb;
120 static int scan_ds_queue_compare(const void *a, const void *b);
121 static int scan_prefetch_queue_compare(const void *a, const void *b);
122 static void scan_ds_queue_clear(dsl_scan_t *scn);
123 static void scan_ds_prefetch_queue_clear(dsl_scan_t *scn);
124 static boolean_t scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj,
126 static void scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg);
127 static void scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj);
128 static void scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx);
129 static uint64_t dsl_scan_count_data_disks(vdev_t *vd);
131 extern int zfs_vdev_async_write_active_min_dirty_percent;
134 * By default zfs will check to ensure it is not over the hard memory
135 * limit before each txg. If finer-grained control of this is needed
136 * this value can be set to 1 to enable checking before scanning each
139 static int zfs_scan_strict_mem_lim = B_FALSE;
142 * Maximum number of parallelly executed bytes per leaf vdev. We attempt
143 * to strike a balance here between keeping the vdev queues full of I/Os
144 * at all times and not overflowing the queues to cause long latency,
145 * which would cause long txg sync times. No matter what, we will not
146 * overload the drives with I/O, since that is protected by
147 * zfs_vdev_scrub_max_active.
149 static unsigned long zfs_scan_vdev_limit = 4 << 20;
151 static int zfs_scan_issue_strategy = 0;
152 static int zfs_scan_legacy = B_FALSE; /* don't queue & sort zios, go direct */
153 static unsigned long zfs_scan_max_ext_gap = 2 << 20; /* in bytes */
156 * fill_weight is non-tunable at runtime, so we copy it at module init from
157 * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
158 * break queue sorting.
160 static int zfs_scan_fill_weight = 3;
161 static uint64_t fill_weight;
163 /* See dsl_scan_should_clear() for details on the memory limit tunables */
164 static const uint64_t zfs_scan_mem_lim_min = 16 << 20; /* bytes */
165 static const uint64_t zfs_scan_mem_lim_soft_max = 128 << 20; /* bytes */
166 static int zfs_scan_mem_lim_fact = 20; /* fraction of physmem */
167 static int zfs_scan_mem_lim_soft_fact = 20; /* fraction of mem lim above */
169 static int zfs_scrub_min_time_ms = 1000; /* min millis to scrub per txg */
170 static int zfs_obsolete_min_time_ms = 500; /* min millis to obsolete per txg */
171 static int zfs_free_min_time_ms = 1000; /* min millis to free per txg */
172 static int zfs_resilver_min_time_ms = 3000; /* min millis to resilver per txg */
173 static int zfs_scan_checkpoint_intval = 7200; /* in seconds */
174 int zfs_scan_suspend_progress = 0; /* set to prevent scans from progressing */
175 static int zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
176 static int zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
177 static const enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
178 /* max number of blocks to free in a single TXG */
179 static unsigned long zfs_async_block_max_blocks = ULONG_MAX;
180 /* max number of dedup blocks to free in a single TXG */
181 static unsigned long zfs_max_async_dedup_frees = 100000;
183 /* set to disable resilver deferring */
184 static int zfs_resilver_disable_defer = B_FALSE;
187 * We wait a few txgs after importing a pool to begin scanning so that
188 * the import / mounting code isn't held up by scrub / resilver IO.
189 * Unfortunately, it is a bit difficult to determine exactly how long
190 * this will take since userspace will trigger fs mounts asynchronously
191 * and the kernel will create zvol minors asynchronously. As a result,
192 * the value provided here is a bit arbitrary, but represents a
193 * reasonable estimate of how many txgs it will take to finish fully
196 #define SCAN_IMPORT_WAIT_TXGS 5
198 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
199 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
200 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
203 * Enable/disable the processing of the free_bpobj object.
205 static int zfs_free_bpobj_enabled = 1;
207 /* the order has to match pool_scan_type */
208 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
210 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
211 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
214 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */
222 * This controls what conditions are placed on dsl_scan_sync_state():
223 * SYNC_OPTIONAL) write out scn_phys iff scn_bytes_pending == 0
224 * SYNC_MANDATORY) write out scn_phys always. scn_bytes_pending must be 0.
225 * SYNC_CACHED) if scn_bytes_pending == 0, write out scn_phys. Otherwise
226 * write out the scn_phys_cached version.
227 * See dsl_scan_sync_state for details.
236 * This struct represents the minimum information needed to reconstruct a
237 * zio for sequential scanning. This is useful because many of these will
238 * accumulate in the sequential IO queues before being issued, so saving
239 * memory matters here.
241 typedef struct scan_io {
242 /* fields from blkptr_t */
243 uint64_t sio_blk_prop;
244 uint64_t sio_phys_birth;
246 zio_cksum_t sio_cksum;
247 uint32_t sio_nr_dvas;
249 /* fields from zio_t */
251 zbookmark_phys_t sio_zb;
253 /* members for queue sorting */
255 avl_node_t sio_addr_node; /* link into issuing queue */
256 list_node_t sio_list_node; /* link for issuing to disk */
260 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp,
261 * depending on how many were in the original bp. Only the
262 * first DVA is really used for sorting and issuing purposes.
263 * The other DVAs (if provided) simply exist so that the zio
264 * layer can find additional copies to repair from in the
265 * event of an error. This array must go at the end of the
266 * struct to allow this for the variable number of elements.
271 #define SIO_SET_OFFSET(sio, x) DVA_SET_OFFSET(&(sio)->sio_dva[0], x)
272 #define SIO_SET_ASIZE(sio, x) DVA_SET_ASIZE(&(sio)->sio_dva[0], x)
273 #define SIO_GET_OFFSET(sio) DVA_GET_OFFSET(&(sio)->sio_dva[0])
274 #define SIO_GET_ASIZE(sio) DVA_GET_ASIZE(&(sio)->sio_dva[0])
275 #define SIO_GET_END_OFFSET(sio) \
276 (SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio))
277 #define SIO_GET_MUSED(sio) \
278 (sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t)))
280 struct dsl_scan_io_queue {
281 dsl_scan_t *q_scn; /* associated dsl_scan_t */
282 vdev_t *q_vd; /* top-level vdev that this queue represents */
284 /* trees used for sorting I/Os and extents of I/Os */
285 range_tree_t *q_exts_by_addr;
286 zfs_btree_t q_exts_by_size;
287 avl_tree_t q_sios_by_addr;
288 uint64_t q_sio_memused;
290 /* members for zio rate limiting */
291 uint64_t q_maxinflight_bytes;
292 uint64_t q_inflight_bytes;
293 kcondvar_t q_zio_cv; /* used under vd->vdev_scan_io_queue_lock */
295 /* per txg statistics */
296 uint64_t q_total_seg_size_this_txg;
297 uint64_t q_segs_this_txg;
298 uint64_t q_total_zio_size_this_txg;
299 uint64_t q_zios_this_txg;
302 /* private data for dsl_scan_prefetch_cb() */
303 typedef struct scan_prefetch_ctx {
304 zfs_refcount_t spc_refcnt; /* refcount for memory management */
305 dsl_scan_t *spc_scn; /* dsl_scan_t for the pool */
306 boolean_t spc_root; /* is this prefetch for an objset? */
307 uint8_t spc_indblkshift; /* dn_indblkshift of current dnode */
308 uint16_t spc_datablkszsec; /* dn_idatablkszsec of current dnode */
309 } scan_prefetch_ctx_t;
311 /* private data for dsl_scan_prefetch() */
312 typedef struct scan_prefetch_issue_ctx {
313 avl_node_t spic_avl_node; /* link into scn->scn_prefetch_queue */
314 scan_prefetch_ctx_t *spic_spc; /* spc for the callback */
315 blkptr_t spic_bp; /* bp to prefetch */
316 zbookmark_phys_t spic_zb; /* bookmark to prefetch */
317 } scan_prefetch_issue_ctx_t;
319 static void scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
320 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue);
321 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue,
324 static dsl_scan_io_queue_t *scan_io_queue_create(vdev_t *vd);
325 static void scan_io_queues_destroy(dsl_scan_t *scn);
327 static kmem_cache_t *sio_cache[SPA_DVAS_PER_BP];
329 /* sio->sio_nr_dvas must be set so we know which cache to free from */
331 sio_free(scan_io_t *sio)
333 ASSERT3U(sio->sio_nr_dvas, >, 0);
334 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
336 kmem_cache_free(sio_cache[sio->sio_nr_dvas - 1], sio);
339 /* It is up to the caller to set sio->sio_nr_dvas for freeing */
341 sio_alloc(unsigned short nr_dvas)
343 ASSERT3U(nr_dvas, >, 0);
344 ASSERT3U(nr_dvas, <=, SPA_DVAS_PER_BP);
346 return (kmem_cache_alloc(sio_cache[nr_dvas - 1], KM_SLEEP));
353 * This is used in ext_size_compare() to weight segments
354 * based on how sparse they are. This cannot be changed
355 * mid-scan and the tree comparison functions don't currently
356 * have a mechanism for passing additional context to the
357 * compare functions. Thus we store this value globally and
358 * we only allow it to be set at module initialization time
360 fill_weight = zfs_scan_fill_weight;
362 for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
365 (void) snprintf(name, sizeof (name), "sio_cache_%d", i);
366 sio_cache[i] = kmem_cache_create(name,
367 (sizeof (scan_io_t) + ((i + 1) * sizeof (dva_t))),
368 0, NULL, NULL, NULL, NULL, NULL, 0);
375 for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
376 kmem_cache_destroy(sio_cache[i]);
380 static inline boolean_t
381 dsl_scan_is_running(const dsl_scan_t *scn)
383 return (scn->scn_phys.scn_state == DSS_SCANNING);
387 dsl_scan_resilvering(dsl_pool_t *dp)
389 return (dsl_scan_is_running(dp->dp_scan) &&
390 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
394 sio2bp(const scan_io_t *sio, blkptr_t *bp)
396 memset(bp, 0, sizeof (*bp));
397 bp->blk_prop = sio->sio_blk_prop;
398 bp->blk_phys_birth = sio->sio_phys_birth;
399 bp->blk_birth = sio->sio_birth;
400 bp->blk_fill = 1; /* we always only work with data pointers */
401 bp->blk_cksum = sio->sio_cksum;
403 ASSERT3U(sio->sio_nr_dvas, >, 0);
404 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
406 memcpy(bp->blk_dva, sio->sio_dva, sio->sio_nr_dvas * sizeof (dva_t));
410 bp2sio(const blkptr_t *bp, scan_io_t *sio, int dva_i)
412 sio->sio_blk_prop = bp->blk_prop;
413 sio->sio_phys_birth = bp->blk_phys_birth;
414 sio->sio_birth = bp->blk_birth;
415 sio->sio_cksum = bp->blk_cksum;
416 sio->sio_nr_dvas = BP_GET_NDVAS(bp);
419 * Copy the DVAs to the sio. We need all copies of the block so
420 * that the self healing code can use the alternate copies if the
421 * first is corrupted. We want the DVA at index dva_i to be first
422 * in the sio since this is the primary one that we want to issue.
424 for (int i = 0, j = dva_i; i < sio->sio_nr_dvas; i++, j++) {
425 sio->sio_dva[i] = bp->blk_dva[j % sio->sio_nr_dvas];
430 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
434 spa_t *spa = dp->dp_spa;
437 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
441 * It's possible that we're resuming a scan after a reboot so
442 * make sure that the scan_async_destroying flag is initialized
445 ASSERT(!scn->scn_async_destroying);
446 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
447 SPA_FEATURE_ASYNC_DESTROY);
450 * Calculate the max number of in-flight bytes for pool-wide
451 * scanning operations (minimum 1MB). Limits for the issuing
452 * phase are done per top-level vdev and are handled separately.
454 scn->scn_maxinflight_bytes = MAX(zfs_scan_vdev_limit *
455 dsl_scan_count_data_disks(spa->spa_root_vdev), 1ULL << 20);
457 avl_create(&scn->scn_queue, scan_ds_queue_compare, sizeof (scan_ds_t),
458 offsetof(scan_ds_t, sds_node));
459 avl_create(&scn->scn_prefetch_queue, scan_prefetch_queue_compare,
460 sizeof (scan_prefetch_issue_ctx_t),
461 offsetof(scan_prefetch_issue_ctx_t, spic_avl_node));
463 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
464 "scrub_func", sizeof (uint64_t), 1, &f);
467 * There was an old-style scrub in progress. Restart a
468 * new-style scrub from the beginning.
470 scn->scn_restart_txg = txg;
471 zfs_dbgmsg("old-style scrub was in progress for %s; "
472 "restarting new-style scrub in txg %llu",
474 (longlong_t)scn->scn_restart_txg);
477 * Load the queue obj from the old location so that it
478 * can be freed by dsl_scan_done().
480 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
481 "scrub_queue", sizeof (uint64_t), 1,
482 &scn->scn_phys.scn_queue_obj);
484 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
485 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
488 * Detect if the pool contains the signature of #2094. If it
489 * does properly update the scn->scn_phys structure and notify
490 * the administrator by setting an errata for the pool.
492 if (err == EOVERFLOW) {
493 uint64_t zaptmp[SCAN_PHYS_NUMINTS + 1];
494 VERIFY3S(SCAN_PHYS_NUMINTS, ==, 24);
495 VERIFY3S(offsetof(dsl_scan_phys_t, scn_flags), ==,
496 (23 * sizeof (uint64_t)));
498 err = zap_lookup(dp->dp_meta_objset,
499 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SCAN,
500 sizeof (uint64_t), SCAN_PHYS_NUMINTS + 1, &zaptmp);
502 uint64_t overflow = zaptmp[SCAN_PHYS_NUMINTS];
504 if (overflow & ~DSL_SCAN_FLAGS_MASK ||
505 scn->scn_async_destroying) {
507 ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY;
511 memcpy(&scn->scn_phys, zaptmp,
512 SCAN_PHYS_NUMINTS * sizeof (uint64_t));
513 scn->scn_phys.scn_flags = overflow;
515 /* Required scrub already in progress. */
516 if (scn->scn_phys.scn_state == DSS_FINISHED ||
517 scn->scn_phys.scn_state == DSS_CANCELED)
519 ZPOOL_ERRATA_ZOL_2094_SCRUB;
529 * We might be restarting after a reboot, so jump the issued
530 * counter to how far we've scanned. We know we're consistent
533 scn->scn_issued_before_pass = scn->scn_phys.scn_examined;
535 if (dsl_scan_is_running(scn) &&
536 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
538 * A new-type scrub was in progress on an old
539 * pool, and the pool was accessed by old
540 * software. Restart from the beginning, since
541 * the old software may have changed the pool in
544 scn->scn_restart_txg = txg;
545 zfs_dbgmsg("new-style scrub for %s was modified "
546 "by old software; restarting in txg %llu",
548 (longlong_t)scn->scn_restart_txg);
549 } else if (dsl_scan_resilvering(dp)) {
551 * If a resilver is in progress and there are already
552 * errors, restart it instead of finishing this scan and
553 * then restarting it. If there haven't been any errors
554 * then remember that the incore DTL is valid.
556 if (scn->scn_phys.scn_errors > 0) {
557 scn->scn_restart_txg = txg;
558 zfs_dbgmsg("resilver can't excise DTL_MISSING "
559 "when finished; restarting on %s in txg "
562 (u_longlong_t)scn->scn_restart_txg);
564 /* it's safe to excise DTL when finished */
565 spa->spa_scrub_started = B_TRUE;
570 memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys));
572 /* reload the queue into the in-core state */
573 if (scn->scn_phys.scn_queue_obj != 0) {
577 for (zap_cursor_init(&zc, dp->dp_meta_objset,
578 scn->scn_phys.scn_queue_obj);
579 zap_cursor_retrieve(&zc, &za) == 0;
580 (void) zap_cursor_advance(&zc)) {
581 scan_ds_queue_insert(scn,
582 zfs_strtonum(za.za_name, NULL),
583 za.za_first_integer);
585 zap_cursor_fini(&zc);
588 spa_scan_stat_init(spa);
593 dsl_scan_fini(dsl_pool_t *dp)
595 if (dp->dp_scan != NULL) {
596 dsl_scan_t *scn = dp->dp_scan;
598 if (scn->scn_taskq != NULL)
599 taskq_destroy(scn->scn_taskq);
601 scan_ds_queue_clear(scn);
602 avl_destroy(&scn->scn_queue);
603 scan_ds_prefetch_queue_clear(scn);
604 avl_destroy(&scn->scn_prefetch_queue);
606 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
612 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
614 return (scn->scn_restart_txg != 0 &&
615 scn->scn_restart_txg <= tx->tx_txg);
619 dsl_scan_resilver_scheduled(dsl_pool_t *dp)
621 return ((dp->dp_scan && dp->dp_scan->scn_restart_txg != 0) ||
622 (spa_async_tasks(dp->dp_spa) & SPA_ASYNC_RESILVER));
626 dsl_scan_scrubbing(const dsl_pool_t *dp)
628 dsl_scan_phys_t *scn_phys = &dp->dp_scan->scn_phys;
630 return (scn_phys->scn_state == DSS_SCANNING &&
631 scn_phys->scn_func == POOL_SCAN_SCRUB);
635 dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
637 return (dsl_scan_scrubbing(scn->scn_dp) &&
638 scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED);
642 * Writes out a persistent dsl_scan_phys_t record to the pool directory.
643 * Because we can be running in the block sorting algorithm, we do not always
644 * want to write out the record, only when it is "safe" to do so. This safety
645 * condition is achieved by making sure that the sorting queues are empty
646 * (scn_bytes_pending == 0). When this condition is not true, the sync'd state
647 * is inconsistent with how much actual scanning progress has been made. The
648 * kind of sync to be performed is specified by the sync_type argument. If the
649 * sync is optional, we only sync if the queues are empty. If the sync is
650 * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
651 * third possible state is a "cached" sync. This is done in response to:
652 * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
653 * destroyed, so we wouldn't be able to restart scanning from it.
654 * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
655 * superseded by a newer snapshot.
656 * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
657 * swapped with its clone.
658 * In all cases, a cached sync simply rewrites the last record we've written,
659 * just slightly modified. For the modifications that are performed to the
660 * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
661 * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
664 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx, state_sync_type_t sync_type)
667 spa_t *spa = scn->scn_dp->dp_spa;
669 ASSERT(sync_type != SYNC_MANDATORY || scn->scn_bytes_pending == 0);
670 if (scn->scn_bytes_pending == 0) {
671 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
672 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
673 dsl_scan_io_queue_t *q = vd->vdev_scan_io_queue;
678 mutex_enter(&vd->vdev_scan_io_queue_lock);
679 ASSERT3P(avl_first(&q->q_sios_by_addr), ==, NULL);
680 ASSERT3P(zfs_btree_first(&q->q_exts_by_size, NULL), ==,
682 ASSERT3P(range_tree_first(q->q_exts_by_addr), ==, NULL);
683 mutex_exit(&vd->vdev_scan_io_queue_lock);
686 if (scn->scn_phys.scn_queue_obj != 0)
687 scan_ds_queue_sync(scn, tx);
688 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
689 DMU_POOL_DIRECTORY_OBJECT,
690 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
691 &scn->scn_phys, tx));
692 memcpy(&scn->scn_phys_cached, &scn->scn_phys,
693 sizeof (scn->scn_phys));
695 if (scn->scn_checkpointing)
696 zfs_dbgmsg("finish scan checkpoint for %s",
699 scn->scn_checkpointing = B_FALSE;
700 scn->scn_last_checkpoint = ddi_get_lbolt();
701 } else if (sync_type == SYNC_CACHED) {
702 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
703 DMU_POOL_DIRECTORY_OBJECT,
704 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
705 &scn->scn_phys_cached, tx));
710 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
713 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
714 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
716 if (dsl_scan_is_running(scn) || vdev_rebuild_active(rvd))
717 return (SET_ERROR(EBUSY));
723 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
725 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
726 pool_scan_func_t *funcp = arg;
727 dmu_object_type_t ot = 0;
728 dsl_pool_t *dp = scn->scn_dp;
729 spa_t *spa = dp->dp_spa;
731 ASSERT(!dsl_scan_is_running(scn));
732 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
733 memset(&scn->scn_phys, 0, sizeof (scn->scn_phys));
734 scn->scn_phys.scn_func = *funcp;
735 scn->scn_phys.scn_state = DSS_SCANNING;
736 scn->scn_phys.scn_min_txg = 0;
737 scn->scn_phys.scn_max_txg = tx->tx_txg;
738 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
739 scn->scn_phys.scn_start_time = gethrestime_sec();
740 scn->scn_phys.scn_errors = 0;
741 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
742 scn->scn_issued_before_pass = 0;
743 scn->scn_restart_txg = 0;
744 scn->scn_done_txg = 0;
745 scn->scn_last_checkpoint = 0;
746 scn->scn_checkpointing = B_FALSE;
747 spa_scan_stat_init(spa);
749 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
750 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
752 /* rewrite all disk labels */
753 vdev_config_dirty(spa->spa_root_vdev);
755 if (vdev_resilver_needed(spa->spa_root_vdev,
756 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
757 nvlist_t *aux = fnvlist_alloc();
758 fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE,
760 spa_event_notify(spa, NULL, aux,
761 ESC_ZFS_RESILVER_START);
764 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
767 spa->spa_scrub_started = B_TRUE;
769 * If this is an incremental scrub, limit the DDT scrub phase
770 * to just the auto-ditto class (for correctness); the rest
771 * of the scrub should go faster using top-down pruning.
773 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
774 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
777 * When starting a resilver clear any existing rebuild state.
778 * This is required to prevent stale rebuild status from
779 * being reported when a rebuild is run, then a resilver and
780 * finally a scrub. In which case only the scrub status
781 * should be reported by 'zpool status'.
783 if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
784 vdev_t *rvd = spa->spa_root_vdev;
785 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
786 vdev_t *vd = rvd->vdev_child[i];
787 vdev_rebuild_clear_sync(
788 (void *)(uintptr_t)vd->vdev_id, tx);
793 /* back to the generic stuff */
795 if (dp->dp_blkstats == NULL) {
797 vmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
798 mutex_init(&dp->dp_blkstats->zab_lock, NULL,
799 MUTEX_DEFAULT, NULL);
801 memset(&dp->dp_blkstats->zab_type, 0,
802 sizeof (dp->dp_blkstats->zab_type));
804 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
805 ot = DMU_OT_ZAP_OTHER;
807 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
808 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
810 memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys));
812 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
814 spa_history_log_internal(spa, "scan setup", tx,
815 "func=%u mintxg=%llu maxtxg=%llu",
816 *funcp, (u_longlong_t)scn->scn_phys.scn_min_txg,
817 (u_longlong_t)scn->scn_phys.scn_max_txg);
821 * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
822 * Can also be called to resume a paused scrub.
825 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
827 spa_t *spa = dp->dp_spa;
828 dsl_scan_t *scn = dp->dp_scan;
831 * Purge all vdev caches and probe all devices. We do this here
832 * rather than in sync context because this requires a writer lock
833 * on the spa_config lock, which we can't do from sync context. The
834 * spa_scrub_reopen flag indicates that vdev_open() should not
835 * attempt to start another scrub.
837 spa_vdev_state_enter(spa, SCL_NONE);
838 spa->spa_scrub_reopen = B_TRUE;
839 vdev_reopen(spa->spa_root_vdev);
840 spa->spa_scrub_reopen = B_FALSE;
841 (void) spa_vdev_state_exit(spa, NULL, 0);
843 if (func == POOL_SCAN_RESILVER) {
844 dsl_scan_restart_resilver(spa->spa_dsl_pool, 0);
848 if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
849 /* got scrub start cmd, resume paused scrub */
850 int err = dsl_scrub_set_pause_resume(scn->scn_dp,
853 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
854 return (SET_ERROR(ECANCELED));
857 return (SET_ERROR(err));
860 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
861 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED));
865 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
867 static const char *old_names[] = {
869 "scrub_ddt_bookmark",
870 "scrub_ddt_class_max",
879 dsl_pool_t *dp = scn->scn_dp;
880 spa_t *spa = dp->dp_spa;
883 /* Remove any remnants of an old-style scrub. */
884 for (i = 0; old_names[i]; i++) {
885 (void) zap_remove(dp->dp_meta_objset,
886 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
889 if (scn->scn_phys.scn_queue_obj != 0) {
890 VERIFY0(dmu_object_free(dp->dp_meta_objset,
891 scn->scn_phys.scn_queue_obj, tx));
892 scn->scn_phys.scn_queue_obj = 0;
894 scan_ds_queue_clear(scn);
895 scan_ds_prefetch_queue_clear(scn);
897 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
900 * If we were "restarted" from a stopped state, don't bother
901 * with anything else.
903 if (!dsl_scan_is_running(scn)) {
904 ASSERT(!scn->scn_is_sorted);
908 if (scn->scn_is_sorted) {
909 scan_io_queues_destroy(scn);
910 scn->scn_is_sorted = B_FALSE;
912 if (scn->scn_taskq != NULL) {
913 taskq_destroy(scn->scn_taskq);
914 scn->scn_taskq = NULL;
918 scn->scn_phys.scn_state = complete ? DSS_FINISHED : DSS_CANCELED;
920 spa_notify_waiters(spa);
922 if (dsl_scan_restarting(scn, tx))
923 spa_history_log_internal(spa, "scan aborted, restarting", tx,
924 "errors=%llu", (u_longlong_t)spa_get_errlog_size(spa));
926 spa_history_log_internal(spa, "scan cancelled", tx,
927 "errors=%llu", (u_longlong_t)spa_get_errlog_size(spa));
929 spa_history_log_internal(spa, "scan done", tx,
930 "errors=%llu", (u_longlong_t)spa_get_errlog_size(spa));
932 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
933 spa->spa_scrub_active = B_FALSE;
936 * If the scrub/resilver completed, update all DTLs to
937 * reflect this. Whether it succeeded or not, vacate
938 * all temporary scrub DTLs.
940 * As the scrub does not currently support traversing
941 * data that have been freed but are part of a checkpoint,
942 * we don't mark the scrub as done in the DTLs as faults
943 * may still exist in those vdevs.
946 !spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
947 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
948 scn->scn_phys.scn_max_txg, B_TRUE, B_FALSE);
950 if (scn->scn_phys.scn_min_txg) {
951 nvlist_t *aux = fnvlist_alloc();
952 fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE,
954 spa_event_notify(spa, NULL, aux,
955 ESC_ZFS_RESILVER_FINISH);
958 spa_event_notify(spa, NULL, NULL,
959 ESC_ZFS_SCRUB_FINISH);
962 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
965 spa_errlog_rotate(spa);
968 * Don't clear flag until after vdev_dtl_reassess to ensure that
969 * DTL_MISSING will get updated when possible.
971 spa->spa_scrub_started = B_FALSE;
974 * We may have finished replacing a device.
975 * Let the async thread assess this and handle the detach.
977 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
980 * Clear any resilver_deferred flags in the config.
981 * If there are drives that need resilvering, kick
982 * off an asynchronous request to start resilver.
983 * vdev_clear_resilver_deferred() may update the config
984 * before the resilver can restart. In the event of
985 * a crash during this period, the spa loading code
986 * will find the drives that need to be resilvered
987 * and start the resilver then.
989 if (spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER) &&
990 vdev_clear_resilver_deferred(spa->spa_root_vdev, tx)) {
991 spa_history_log_internal(spa,
992 "starting deferred resilver", tx, "errors=%llu",
993 (u_longlong_t)spa_get_errlog_size(spa));
994 spa_async_request(spa, SPA_ASYNC_RESILVER);
997 /* Clear recent error events (i.e. duplicate events tracking) */
999 zfs_ereport_clear(spa, NULL);
1002 scn->scn_phys.scn_end_time = gethrestime_sec();
1004 if (spa->spa_errata == ZPOOL_ERRATA_ZOL_2094_SCRUB)
1005 spa->spa_errata = 0;
1007 ASSERT(!dsl_scan_is_running(scn));
1011 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
1014 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1016 if (!dsl_scan_is_running(scn))
1017 return (SET_ERROR(ENOENT));
1022 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
1025 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1027 dsl_scan_done(scn, B_FALSE, tx);
1028 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
1029 spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
1033 dsl_scan_cancel(dsl_pool_t *dp)
1035 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
1036 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
1040 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
1042 pool_scrub_cmd_t *cmd = arg;
1043 dsl_pool_t *dp = dmu_tx_pool(tx);
1044 dsl_scan_t *scn = dp->dp_scan;
1046 if (*cmd == POOL_SCRUB_PAUSE) {
1047 /* can't pause a scrub when there is no in-progress scrub */
1048 if (!dsl_scan_scrubbing(dp))
1049 return (SET_ERROR(ENOENT));
1051 /* can't pause a paused scrub */
1052 if (dsl_scan_is_paused_scrub(scn))
1053 return (SET_ERROR(EBUSY));
1054 } else if (*cmd != POOL_SCRUB_NORMAL) {
1055 return (SET_ERROR(ENOTSUP));
1062 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
1064 pool_scrub_cmd_t *cmd = arg;
1065 dsl_pool_t *dp = dmu_tx_pool(tx);
1066 spa_t *spa = dp->dp_spa;
1067 dsl_scan_t *scn = dp->dp_scan;
1069 if (*cmd == POOL_SCRUB_PAUSE) {
1070 /* can't pause a scrub when there is no in-progress scrub */
1071 spa->spa_scan_pass_scrub_pause = gethrestime_sec();
1072 scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
1073 scn->scn_phys_cached.scn_flags |= DSF_SCRUB_PAUSED;
1074 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1075 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
1076 spa_notify_waiters(spa);
1078 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
1079 if (dsl_scan_is_paused_scrub(scn)) {
1081 * We need to keep track of how much time we spend
1082 * paused per pass so that we can adjust the scrub rate
1083 * shown in the output of 'zpool status'
1085 spa->spa_scan_pass_scrub_spent_paused +=
1086 gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
1087 spa->spa_scan_pass_scrub_pause = 0;
1088 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
1089 scn->scn_phys_cached.scn_flags &= ~DSF_SCRUB_PAUSED;
1090 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1096 * Set scrub pause/resume state if it makes sense to do so
1099 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
1101 return (dsl_sync_task(spa_name(dp->dp_spa),
1102 dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
1103 ZFS_SPACE_CHECK_RESERVED));
1107 /* start a new scan, or restart an existing one. */
1109 dsl_scan_restart_resilver(dsl_pool_t *dp, uint64_t txg)
1113 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1114 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1116 txg = dmu_tx_get_txg(tx);
1117 dp->dp_scan->scn_restart_txg = txg;
1120 dp->dp_scan->scn_restart_txg = txg;
1122 zfs_dbgmsg("restarting resilver for %s at txg=%llu",
1123 dp->dp_spa->spa_name, (longlong_t)txg);
1127 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
1129 zio_free(dp->dp_spa, txg, bp);
1133 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
1135 ASSERT(dsl_pool_sync_context(dp));
1136 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
1140 scan_ds_queue_compare(const void *a, const void *b)
1142 const scan_ds_t *sds_a = a, *sds_b = b;
1144 if (sds_a->sds_dsobj < sds_b->sds_dsobj)
1146 if (sds_a->sds_dsobj == sds_b->sds_dsobj)
1152 scan_ds_queue_clear(dsl_scan_t *scn)
1154 void *cookie = NULL;
1156 while ((sds = avl_destroy_nodes(&scn->scn_queue, &cookie)) != NULL) {
1157 kmem_free(sds, sizeof (*sds));
1162 scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, uint64_t *txg)
1164 scan_ds_t srch, *sds;
1166 srch.sds_dsobj = dsobj;
1167 sds = avl_find(&scn->scn_queue, &srch, NULL);
1168 if (sds != NULL && txg != NULL)
1169 *txg = sds->sds_txg;
1170 return (sds != NULL);
1174 scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg)
1179 sds = kmem_zalloc(sizeof (*sds), KM_SLEEP);
1180 sds->sds_dsobj = dsobj;
1183 VERIFY3P(avl_find(&scn->scn_queue, sds, &where), ==, NULL);
1184 avl_insert(&scn->scn_queue, sds, where);
1188 scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj)
1190 scan_ds_t srch, *sds;
1192 srch.sds_dsobj = dsobj;
1194 sds = avl_find(&scn->scn_queue, &srch, NULL);
1195 VERIFY(sds != NULL);
1196 avl_remove(&scn->scn_queue, sds);
1197 kmem_free(sds, sizeof (*sds));
1201 scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx)
1203 dsl_pool_t *dp = scn->scn_dp;
1204 spa_t *spa = dp->dp_spa;
1205 dmu_object_type_t ot = (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) ?
1206 DMU_OT_SCAN_QUEUE : DMU_OT_ZAP_OTHER;
1208 ASSERT0(scn->scn_bytes_pending);
1209 ASSERT(scn->scn_phys.scn_queue_obj != 0);
1211 VERIFY0(dmu_object_free(dp->dp_meta_objset,
1212 scn->scn_phys.scn_queue_obj, tx));
1213 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, ot,
1214 DMU_OT_NONE, 0, tx);
1215 for (scan_ds_t *sds = avl_first(&scn->scn_queue);
1216 sds != NULL; sds = AVL_NEXT(&scn->scn_queue, sds)) {
1217 VERIFY0(zap_add_int_key(dp->dp_meta_objset,
1218 scn->scn_phys.scn_queue_obj, sds->sds_dsobj,
1224 * Computes the memory limit state that we're currently in. A sorted scan
1225 * needs quite a bit of memory to hold the sorting queue, so we need to
1226 * reasonably constrain the size so it doesn't impact overall system
1227 * performance. We compute two limits:
1228 * 1) Hard memory limit: if the amount of memory used by the sorting
1229 * queues on a pool gets above this value, we stop the metadata
1230 * scanning portion and start issuing the queued up and sorted
1231 * I/Os to reduce memory usage.
1232 * This limit is calculated as a fraction of physmem (by default 5%).
1233 * We constrain the lower bound of the hard limit to an absolute
1234 * minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1235 * the upper bound to 5% of the total pool size - no chance we'll
1236 * ever need that much memory, but just to keep the value in check.
1237 * 2) Soft memory limit: once we hit the hard memory limit, we start
1238 * issuing I/O to reduce queue memory usage, but we don't want to
1239 * completely empty out the queues, since we might be able to find I/Os
1240 * that will fill in the gaps of our non-sequential IOs at some point
1241 * in the future. So we stop the issuing of I/Os once the amount of
1242 * memory used drops below the soft limit (at which point we stop issuing
1243 * I/O and start scanning metadata again).
1245 * This limit is calculated by subtracting a fraction of the hard
1246 * limit from the hard limit. By default this fraction is 5%, so
1247 * the soft limit is 95% of the hard limit. We cap the size of the
1248 * difference between the hard and soft limits at an absolute
1249 * maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1250 * sufficient to not cause too frequent switching between the
1251 * metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1252 * worth of queues is about 1.2 GiB of on-pool data, so scanning
1253 * that should take at least a decent fraction of a second).
1256 dsl_scan_should_clear(dsl_scan_t *scn)
1258 spa_t *spa = scn->scn_dp->dp_spa;
1259 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
1260 uint64_t alloc, mlim_hard, mlim_soft, mused;
1262 alloc = metaslab_class_get_alloc(spa_normal_class(spa));
1263 alloc += metaslab_class_get_alloc(spa_special_class(spa));
1264 alloc += metaslab_class_get_alloc(spa_dedup_class(spa));
1266 mlim_hard = MAX((physmem / zfs_scan_mem_lim_fact) * PAGESIZE,
1267 zfs_scan_mem_lim_min);
1268 mlim_hard = MIN(mlim_hard, alloc / 20);
1269 mlim_soft = mlim_hard - MIN(mlim_hard / zfs_scan_mem_lim_soft_fact,
1270 zfs_scan_mem_lim_soft_max);
1272 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1273 vdev_t *tvd = rvd->vdev_child[i];
1274 dsl_scan_io_queue_t *queue;
1276 mutex_enter(&tvd->vdev_scan_io_queue_lock);
1277 queue = tvd->vdev_scan_io_queue;
1278 if (queue != NULL) {
1279 /* # extents in exts_by_size = # in exts_by_addr */
1280 mused += zfs_btree_numnodes(&queue->q_exts_by_size) *
1281 sizeof (range_seg_gap_t) + queue->q_sio_memused;
1283 mutex_exit(&tvd->vdev_scan_io_queue_lock);
1286 dprintf("current scan memory usage: %llu bytes\n", (longlong_t)mused);
1289 ASSERT0(scn->scn_bytes_pending);
1292 * If we are above our hard limit, we need to clear out memory.
1293 * If we are below our soft limit, we need to accumulate sequential IOs.
1294 * Otherwise, we should keep doing whatever we are currently doing.
1296 if (mused >= mlim_hard)
1298 else if (mused < mlim_soft)
1301 return (scn->scn_clearing);
1305 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1307 /* we never skip user/group accounting objects */
1308 if (zb && (int64_t)zb->zb_object < 0)
1311 if (scn->scn_suspending)
1312 return (B_TRUE); /* we're already suspending */
1314 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
1315 return (B_FALSE); /* we're resuming */
1317 /* We only know how to resume from level-0 and objset blocks. */
1318 if (zb && (zb->zb_level != 0 && zb->zb_level != ZB_ROOT_LEVEL))
1323 * - we have scanned for at least the minimum time (default 1 sec
1324 * for scrub, 3 sec for resilver), and either we have sufficient
1325 * dirty data that we are starting to write more quickly
1326 * (default 30%), someone is explicitly waiting for this txg
1327 * to complete, or we have used up all of the time in the txg
1328 * timeout (default 5 sec).
1330 * - the spa is shutting down because this pool is being exported
1331 * or the machine is rebooting.
1333 * - the scan queue has reached its memory use limit
1335 uint64_t curr_time_ns = gethrtime();
1336 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
1337 uint64_t sync_time_ns = curr_time_ns -
1338 scn->scn_dp->dp_spa->spa_sync_starttime;
1339 int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
1340 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
1341 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
1343 if ((NSEC2MSEC(scan_time_ns) > mintime &&
1344 (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
1345 txg_sync_waiting(scn->scn_dp) ||
1346 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1347 spa_shutting_down(scn->scn_dp->dp_spa) ||
1348 (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn))) {
1349 if (zb && zb->zb_level == ZB_ROOT_LEVEL) {
1350 dprintf("suspending at first available bookmark "
1351 "%llx/%llx/%llx/%llx\n",
1352 (longlong_t)zb->zb_objset,
1353 (longlong_t)zb->zb_object,
1354 (longlong_t)zb->zb_level,
1355 (longlong_t)zb->zb_blkid);
1356 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
1357 zb->zb_objset, 0, 0, 0);
1358 } else if (zb != NULL) {
1359 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1360 (longlong_t)zb->zb_objset,
1361 (longlong_t)zb->zb_object,
1362 (longlong_t)zb->zb_level,
1363 (longlong_t)zb->zb_blkid);
1364 scn->scn_phys.scn_bookmark = *zb;
1367 dsl_scan_phys_t *scnp = &scn->scn_phys;
1368 dprintf("suspending at at DDT bookmark "
1369 "%llx/%llx/%llx/%llx\n",
1370 (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
1371 (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
1372 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
1373 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
1376 scn->scn_suspending = B_TRUE;
1382 typedef struct zil_scan_arg {
1384 zil_header_t *zsa_zh;
1388 dsl_scan_zil_block(zilog_t *zilog, const blkptr_t *bp, void *arg,
1392 zil_scan_arg_t *zsa = arg;
1393 dsl_pool_t *dp = zsa->zsa_dp;
1394 dsl_scan_t *scn = dp->dp_scan;
1395 zil_header_t *zh = zsa->zsa_zh;
1396 zbookmark_phys_t zb;
1398 ASSERT(!BP_IS_REDACTED(bp));
1399 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1403 * One block ("stubby") can be allocated a long time ago; we
1404 * want to visit that one because it has been allocated
1405 * (on-disk) even if it hasn't been claimed (even though for
1406 * scrub there's nothing to do to it).
1408 if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(dp->dp_spa))
1411 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1412 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
1414 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1419 dsl_scan_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg,
1423 if (lrc->lrc_txtype == TX_WRITE) {
1424 zil_scan_arg_t *zsa = arg;
1425 dsl_pool_t *dp = zsa->zsa_dp;
1426 dsl_scan_t *scn = dp->dp_scan;
1427 zil_header_t *zh = zsa->zsa_zh;
1428 const lr_write_t *lr = (const lr_write_t *)lrc;
1429 const blkptr_t *bp = &lr->lr_blkptr;
1430 zbookmark_phys_t zb;
1432 ASSERT(!BP_IS_REDACTED(bp));
1433 if (BP_IS_HOLE(bp) ||
1434 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1438 * birth can be < claim_txg if this record's txg is
1439 * already txg sync'ed (but this log block contains
1440 * other records that are not synced)
1442 if (claim_txg == 0 || bp->blk_birth < claim_txg)
1445 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1446 lr->lr_foid, ZB_ZIL_LEVEL,
1447 lr->lr_offset / BP_GET_LSIZE(bp));
1449 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1455 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
1457 uint64_t claim_txg = zh->zh_claim_txg;
1458 zil_scan_arg_t zsa = { dp, zh };
1461 ASSERT(spa_writeable(dp->dp_spa));
1464 * We only want to visit blocks that have been claimed but not yet
1465 * replayed (or, in read-only mode, blocks that *would* be claimed).
1470 zilog = zil_alloc(dp->dp_meta_objset, zh);
1472 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
1473 claim_txg, B_FALSE);
1479 * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1480 * here is to sort the AVL tree by the order each block will be needed.
1483 scan_prefetch_queue_compare(const void *a, const void *b)
1485 const scan_prefetch_issue_ctx_t *spic_a = a, *spic_b = b;
1486 const scan_prefetch_ctx_t *spc_a = spic_a->spic_spc;
1487 const scan_prefetch_ctx_t *spc_b = spic_b->spic_spc;
1489 return (zbookmark_compare(spc_a->spc_datablkszsec,
1490 spc_a->spc_indblkshift, spc_b->spc_datablkszsec,
1491 spc_b->spc_indblkshift, &spic_a->spic_zb, &spic_b->spic_zb));
1495 scan_prefetch_ctx_rele(scan_prefetch_ctx_t *spc, void *tag)
1497 if (zfs_refcount_remove(&spc->spc_refcnt, tag) == 0) {
1498 zfs_refcount_destroy(&spc->spc_refcnt);
1499 kmem_free(spc, sizeof (scan_prefetch_ctx_t));
1503 static scan_prefetch_ctx_t *
1504 scan_prefetch_ctx_create(dsl_scan_t *scn, dnode_phys_t *dnp, void *tag)
1506 scan_prefetch_ctx_t *spc;
1508 spc = kmem_alloc(sizeof (scan_prefetch_ctx_t), KM_SLEEP);
1509 zfs_refcount_create(&spc->spc_refcnt);
1510 zfs_refcount_add(&spc->spc_refcnt, tag);
1513 spc->spc_datablkszsec = dnp->dn_datablkszsec;
1514 spc->spc_indblkshift = dnp->dn_indblkshift;
1515 spc->spc_root = B_FALSE;
1517 spc->spc_datablkszsec = 0;
1518 spc->spc_indblkshift = 0;
1519 spc->spc_root = B_TRUE;
1526 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t *spc, void *tag)
1528 zfs_refcount_add(&spc->spc_refcnt, tag);
1532 scan_ds_prefetch_queue_clear(dsl_scan_t *scn)
1534 spa_t *spa = scn->scn_dp->dp_spa;
1535 void *cookie = NULL;
1536 scan_prefetch_issue_ctx_t *spic = NULL;
1538 mutex_enter(&spa->spa_scrub_lock);
1539 while ((spic = avl_destroy_nodes(&scn->scn_prefetch_queue,
1540 &cookie)) != NULL) {
1541 scan_prefetch_ctx_rele(spic->spic_spc, scn);
1542 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1544 mutex_exit(&spa->spa_scrub_lock);
1548 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t *spc,
1549 const zbookmark_phys_t *zb)
1551 zbookmark_phys_t *last_zb = &spc->spc_scn->scn_prefetch_bookmark;
1552 dnode_phys_t tmp_dnp;
1553 dnode_phys_t *dnp = (spc->spc_root) ? NULL : &tmp_dnp;
1555 if (zb->zb_objset != last_zb->zb_objset)
1557 if ((int64_t)zb->zb_object < 0)
1560 tmp_dnp.dn_datablkszsec = spc->spc_datablkszsec;
1561 tmp_dnp.dn_indblkshift = spc->spc_indblkshift;
1563 if (zbookmark_subtree_completed(dnp, zb, last_zb))
1570 dsl_scan_prefetch(scan_prefetch_ctx_t *spc, blkptr_t *bp, zbookmark_phys_t *zb)
1573 dsl_scan_t *scn = spc->spc_scn;
1574 spa_t *spa = scn->scn_dp->dp_spa;
1575 scan_prefetch_issue_ctx_t *spic;
1577 if (zfs_no_scrub_prefetch || BP_IS_REDACTED(bp))
1580 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg ||
1581 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE &&
1582 BP_GET_TYPE(bp) != DMU_OT_OBJSET))
1585 if (dsl_scan_check_prefetch_resume(spc, zb))
1588 scan_prefetch_ctx_add_ref(spc, scn);
1589 spic = kmem_alloc(sizeof (scan_prefetch_issue_ctx_t), KM_SLEEP);
1590 spic->spic_spc = spc;
1591 spic->spic_bp = *bp;
1592 spic->spic_zb = *zb;
1595 * Add the IO to the queue of blocks to prefetch. This allows us to
1596 * prioritize blocks that we will need first for the main traversal
1599 mutex_enter(&spa->spa_scrub_lock);
1600 if (avl_find(&scn->scn_prefetch_queue, spic, &idx) != NULL) {
1601 /* this block is already queued for prefetch */
1602 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1603 scan_prefetch_ctx_rele(spc, scn);
1604 mutex_exit(&spa->spa_scrub_lock);
1608 avl_insert(&scn->scn_prefetch_queue, spic, idx);
1609 cv_broadcast(&spa->spa_scrub_io_cv);
1610 mutex_exit(&spa->spa_scrub_lock);
1614 dsl_scan_prefetch_dnode(dsl_scan_t *scn, dnode_phys_t *dnp,
1615 uint64_t objset, uint64_t object)
1618 zbookmark_phys_t zb;
1619 scan_prefetch_ctx_t *spc;
1621 if (dnp->dn_nblkptr == 0 && !(dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1624 SET_BOOKMARK(&zb, objset, object, 0, 0);
1626 spc = scan_prefetch_ctx_create(scn, dnp, FTAG);
1628 for (i = 0; i < dnp->dn_nblkptr; i++) {
1629 zb.zb_level = BP_GET_LEVEL(&dnp->dn_blkptr[i]);
1631 dsl_scan_prefetch(spc, &dnp->dn_blkptr[i], &zb);
1634 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1636 zb.zb_blkid = DMU_SPILL_BLKID;
1637 dsl_scan_prefetch(spc, DN_SPILL_BLKPTR(dnp), &zb);
1640 scan_prefetch_ctx_rele(spc, FTAG);
1644 dsl_scan_prefetch_cb(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1645 arc_buf_t *buf, void *private)
1648 scan_prefetch_ctx_t *spc = private;
1649 dsl_scan_t *scn = spc->spc_scn;
1650 spa_t *spa = scn->scn_dp->dp_spa;
1652 /* broadcast that the IO has completed for rate limiting purposes */
1653 mutex_enter(&spa->spa_scrub_lock);
1654 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
1655 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
1656 cv_broadcast(&spa->spa_scrub_io_cv);
1657 mutex_exit(&spa->spa_scrub_lock);
1659 /* if there was an error or we are done prefetching, just cleanup */
1660 if (buf == NULL || scn->scn_prefetch_stop)
1663 if (BP_GET_LEVEL(bp) > 0) {
1666 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1667 zbookmark_phys_t czb;
1669 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1670 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1671 zb->zb_level - 1, zb->zb_blkid * epb + i);
1672 dsl_scan_prefetch(spc, cbp, &czb);
1674 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1677 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1679 for (i = 0, cdnp = buf->b_data; i < epb;
1680 i += cdnp->dn_extra_slots + 1,
1681 cdnp += cdnp->dn_extra_slots + 1) {
1682 dsl_scan_prefetch_dnode(scn, cdnp,
1683 zb->zb_objset, zb->zb_blkid * epb + i);
1685 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1686 objset_phys_t *osp = buf->b_data;
1688 dsl_scan_prefetch_dnode(scn, &osp->os_meta_dnode,
1689 zb->zb_objset, DMU_META_DNODE_OBJECT);
1691 if (OBJSET_BUF_HAS_USERUSED(buf)) {
1692 dsl_scan_prefetch_dnode(scn,
1693 &osp->os_groupused_dnode, zb->zb_objset,
1694 DMU_GROUPUSED_OBJECT);
1695 dsl_scan_prefetch_dnode(scn,
1696 &osp->os_userused_dnode, zb->zb_objset,
1697 DMU_USERUSED_OBJECT);
1703 arc_buf_destroy(buf, private);
1704 scan_prefetch_ctx_rele(spc, scn);
1708 dsl_scan_prefetch_thread(void *arg)
1710 dsl_scan_t *scn = arg;
1711 spa_t *spa = scn->scn_dp->dp_spa;
1712 scan_prefetch_issue_ctx_t *spic;
1714 /* loop until we are told to stop */
1715 while (!scn->scn_prefetch_stop) {
1716 arc_flags_t flags = ARC_FLAG_NOWAIT |
1717 ARC_FLAG_PRESCIENT_PREFETCH | ARC_FLAG_PREFETCH;
1718 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1720 mutex_enter(&spa->spa_scrub_lock);
1723 * Wait until we have an IO to issue and are not above our
1724 * maximum in flight limit.
1726 while (!scn->scn_prefetch_stop &&
1727 (avl_numnodes(&scn->scn_prefetch_queue) == 0 ||
1728 spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)) {
1729 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1732 /* recheck if we should stop since we waited for the cv */
1733 if (scn->scn_prefetch_stop) {
1734 mutex_exit(&spa->spa_scrub_lock);
1738 /* remove the prefetch IO from the tree */
1739 spic = avl_first(&scn->scn_prefetch_queue);
1740 spa->spa_scrub_inflight += BP_GET_PSIZE(&spic->spic_bp);
1741 avl_remove(&scn->scn_prefetch_queue, spic);
1743 mutex_exit(&spa->spa_scrub_lock);
1745 if (BP_IS_PROTECTED(&spic->spic_bp)) {
1746 ASSERT(BP_GET_TYPE(&spic->spic_bp) == DMU_OT_DNODE ||
1747 BP_GET_TYPE(&spic->spic_bp) == DMU_OT_OBJSET);
1748 ASSERT3U(BP_GET_LEVEL(&spic->spic_bp), ==, 0);
1749 zio_flags |= ZIO_FLAG_RAW;
1752 /* issue the prefetch asynchronously */
1753 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa,
1754 &spic->spic_bp, dsl_scan_prefetch_cb, spic->spic_spc,
1755 ZIO_PRIORITY_SCRUB, zio_flags, &flags, &spic->spic_zb);
1757 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1760 ASSERT(scn->scn_prefetch_stop);
1762 /* free any prefetches we didn't get to complete */
1763 mutex_enter(&spa->spa_scrub_lock);
1764 while ((spic = avl_first(&scn->scn_prefetch_queue)) != NULL) {
1765 avl_remove(&scn->scn_prefetch_queue, spic);
1766 scan_prefetch_ctx_rele(spic->spic_spc, scn);
1767 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1769 ASSERT0(avl_numnodes(&scn->scn_prefetch_queue));
1770 mutex_exit(&spa->spa_scrub_lock);
1774 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
1775 const zbookmark_phys_t *zb)
1778 * We never skip over user/group accounting objects (obj<0)
1780 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
1781 (int64_t)zb->zb_object >= 0) {
1783 * If we already visited this bp & everything below (in
1784 * a prior txg sync), don't bother doing it again.
1786 if (zbookmark_subtree_completed(dnp, zb,
1787 &scn->scn_phys.scn_bookmark))
1791 * If we found the block we're trying to resume from, or
1792 * we went past it to a different object, zero it out to
1793 * indicate that it's OK to start checking for suspending
1796 if (memcmp(zb, &scn->scn_phys.scn_bookmark,
1797 sizeof (*zb)) == 0 ||
1798 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
1799 dprintf("resuming at %llx/%llx/%llx/%llx\n",
1800 (longlong_t)zb->zb_objset,
1801 (longlong_t)zb->zb_object,
1802 (longlong_t)zb->zb_level,
1803 (longlong_t)zb->zb_blkid);
1804 memset(&scn->scn_phys.scn_bookmark, 0, sizeof (*zb));
1810 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1811 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1812 dmu_objset_type_t ostype, dmu_tx_t *tx);
1813 inline __attribute__((always_inline)) static void dsl_scan_visitdnode(
1814 dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1815 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
1818 * Return nonzero on i/o error.
1819 * Return new buf to write out in *bufp.
1821 inline __attribute__((always_inline)) static int
1822 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1823 dnode_phys_t *dnp, const blkptr_t *bp,
1824 const zbookmark_phys_t *zb, dmu_tx_t *tx)
1826 dsl_pool_t *dp = scn->scn_dp;
1827 spa_t *spa = dp->dp_spa;
1828 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1831 ASSERT(!BP_IS_REDACTED(bp));
1834 * There is an unlikely case of encountering dnodes with contradicting
1835 * dn_bonuslen and DNODE_FLAG_SPILL_BLKPTR flag before in files created
1836 * or modified before commit 4254acb was merged. As it is not possible
1837 * to know which of the two is correct, report an error.
1840 dnp->dn_bonuslen > DN_MAX_BONUS_LEN(dnp)) {
1841 scn->scn_phys.scn_errors++;
1842 spa_log_error(spa, zb);
1843 return (SET_ERROR(EINVAL));
1846 if (BP_GET_LEVEL(bp) > 0) {
1847 arc_flags_t flags = ARC_FLAG_WAIT;
1850 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1853 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
1854 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1856 scn->scn_phys.scn_errors++;
1859 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1860 zbookmark_phys_t czb;
1862 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1864 zb->zb_blkid * epb + i);
1865 dsl_scan_visitbp(cbp, &czb, dnp,
1866 ds, scn, ostype, tx);
1868 arc_buf_destroy(buf, &buf);
1869 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1870 arc_flags_t flags = ARC_FLAG_WAIT;
1873 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1876 if (BP_IS_PROTECTED(bp)) {
1877 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
1878 zio_flags |= ZIO_FLAG_RAW;
1881 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
1882 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1884 scn->scn_phys.scn_errors++;
1887 for (i = 0, cdnp = buf->b_data; i < epb;
1888 i += cdnp->dn_extra_slots + 1,
1889 cdnp += cdnp->dn_extra_slots + 1) {
1890 dsl_scan_visitdnode(scn, ds, ostype,
1891 cdnp, zb->zb_blkid * epb + i, tx);
1894 arc_buf_destroy(buf, &buf);
1895 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1896 arc_flags_t flags = ARC_FLAG_WAIT;
1900 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
1901 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1903 scn->scn_phys.scn_errors++;
1909 dsl_scan_visitdnode(scn, ds, osp->os_type,
1910 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
1912 if (OBJSET_BUF_HAS_USERUSED(buf)) {
1914 * We also always visit user/group/project accounting
1915 * objects, and never skip them, even if we are
1916 * suspending. This is necessary so that the
1917 * space deltas from this txg get integrated.
1919 if (OBJSET_BUF_HAS_PROJECTUSED(buf))
1920 dsl_scan_visitdnode(scn, ds, osp->os_type,
1921 &osp->os_projectused_dnode,
1922 DMU_PROJECTUSED_OBJECT, tx);
1923 dsl_scan_visitdnode(scn, ds, osp->os_type,
1924 &osp->os_groupused_dnode,
1925 DMU_GROUPUSED_OBJECT, tx);
1926 dsl_scan_visitdnode(scn, ds, osp->os_type,
1927 &osp->os_userused_dnode,
1928 DMU_USERUSED_OBJECT, tx);
1930 arc_buf_destroy(buf, &buf);
1931 } else if (!zfs_blkptr_verify(spa, bp, B_FALSE, BLK_VERIFY_LOG)) {
1933 * Sanity check the block pointer contents, this is handled
1934 * by arc_read() for the cases above.
1936 scn->scn_phys.scn_errors++;
1937 spa_log_error(spa, zb);
1938 return (SET_ERROR(EINVAL));
1944 inline __attribute__((always_inline)) static void
1945 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
1946 dmu_objset_type_t ostype, dnode_phys_t *dnp,
1947 uint64_t object, dmu_tx_t *tx)
1951 for (j = 0; j < dnp->dn_nblkptr; j++) {
1952 zbookmark_phys_t czb;
1954 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1955 dnp->dn_nlevels - 1, j);
1956 dsl_scan_visitbp(&dnp->dn_blkptr[j],
1957 &czb, dnp, ds, scn, ostype, tx);
1960 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1961 zbookmark_phys_t czb;
1962 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1963 0, DMU_SPILL_BLKID);
1964 dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp),
1965 &czb, dnp, ds, scn, ostype, tx);
1970 * The arguments are in this order because mdb can only print the
1971 * first 5; we want them to be useful.
1974 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1975 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1976 dmu_objset_type_t ostype, dmu_tx_t *tx)
1978 dsl_pool_t *dp = scn->scn_dp;
1979 blkptr_t *bp_toread = NULL;
1981 if (dsl_scan_check_suspend(scn, zb))
1984 if (dsl_scan_check_resume(scn, dnp, zb))
1987 scn->scn_visited_this_txg++;
1989 if (BP_IS_HOLE(bp)) {
1990 scn->scn_holes_this_txg++;
1994 if (BP_IS_REDACTED(bp)) {
1995 ASSERT(dsl_dataset_feature_is_active(ds,
1996 SPA_FEATURE_REDACTED_DATASETS));
2000 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) {
2001 scn->scn_lt_min_this_txg++;
2005 bp_toread = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
2008 if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx) != 0)
2012 * If dsl_scan_ddt() has already visited this block, it will have
2013 * already done any translations or scrubbing, so don't call the
2016 if (ddt_class_contains(dp->dp_spa,
2017 scn->scn_phys.scn_ddt_class_max, bp)) {
2018 scn->scn_ddt_contained_this_txg++;
2023 * If this block is from the future (after cur_max_txg), then we
2024 * are doing this on behalf of a deleted snapshot, and we will
2025 * revisit the future block on the next pass of this dataset.
2026 * Don't scan it now unless we need to because something
2027 * under it was modified.
2029 if (BP_PHYSICAL_BIRTH(bp) > scn->scn_phys.scn_cur_max_txg) {
2030 scn->scn_gt_max_this_txg++;
2034 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
2037 kmem_free(bp_toread, sizeof (blkptr_t));
2041 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
2044 zbookmark_phys_t zb;
2045 scan_prefetch_ctx_t *spc;
2047 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
2048 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
2050 if (ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) {
2051 SET_BOOKMARK(&scn->scn_prefetch_bookmark,
2052 zb.zb_objset, 0, 0, 0);
2054 scn->scn_prefetch_bookmark = scn->scn_phys.scn_bookmark;
2057 scn->scn_objsets_visited_this_txg++;
2059 spc = scan_prefetch_ctx_create(scn, NULL, FTAG);
2060 dsl_scan_prefetch(spc, bp, &zb);
2061 scan_prefetch_ctx_rele(spc, FTAG);
2063 dsl_scan_visitbp(bp, &zb, NULL, ds, scn, DMU_OST_NONE, tx);
2065 dprintf_ds(ds, "finished scan%s", "");
2069 ds_destroyed_scn_phys(dsl_dataset_t *ds, dsl_scan_phys_t *scn_phys)
2071 if (scn_phys->scn_bookmark.zb_objset == ds->ds_object) {
2072 if (ds->ds_is_snapshot) {
2075 * - scn_cur_{min,max}_txg stays the same.
2076 * - Setting the flag is not really necessary if
2077 * scn_cur_max_txg == scn_max_txg, because there
2078 * is nothing after this snapshot that we care
2079 * about. However, we set it anyway and then
2080 * ignore it when we retraverse it in
2081 * dsl_scan_visitds().
2083 scn_phys->scn_bookmark.zb_objset =
2084 dsl_dataset_phys(ds)->ds_next_snap_obj;
2085 zfs_dbgmsg("destroying ds %llu on %s; currently "
2086 "traversing; reset zb_objset to %llu",
2087 (u_longlong_t)ds->ds_object,
2088 ds->ds_dir->dd_pool->dp_spa->spa_name,
2089 (u_longlong_t)dsl_dataset_phys(ds)->
2091 scn_phys->scn_flags |= DSF_VISIT_DS_AGAIN;
2093 SET_BOOKMARK(&scn_phys->scn_bookmark,
2094 ZB_DESTROYED_OBJSET, 0, 0, 0);
2095 zfs_dbgmsg("destroying ds %llu on %s; currently "
2096 "traversing; reset bookmark to -1,0,0,0",
2097 (u_longlong_t)ds->ds_object,
2098 ds->ds_dir->dd_pool->dp_spa->spa_name);
2104 * Invoked when a dataset is destroyed. We need to make sure that:
2106 * 1) If it is the dataset that was currently being scanned, we write
2107 * a new dsl_scan_phys_t and marking the objset reference in it
2109 * 2) Remove it from the work queue, if it was present.
2111 * If the dataset was actually a snapshot, instead of marking the dataset
2112 * as destroyed, we instead substitute the next snapshot in line.
2115 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
2117 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2118 dsl_scan_t *scn = dp->dp_scan;
2121 if (!dsl_scan_is_running(scn))
2124 ds_destroyed_scn_phys(ds, &scn->scn_phys);
2125 ds_destroyed_scn_phys(ds, &scn->scn_phys_cached);
2127 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2128 scan_ds_queue_remove(scn, ds->ds_object);
2129 if (ds->ds_is_snapshot)
2130 scan_ds_queue_insert(scn,
2131 dsl_dataset_phys(ds)->ds_next_snap_obj, mintxg);
2134 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2135 ds->ds_object, &mintxg) == 0) {
2136 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
2137 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2138 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2139 if (ds->ds_is_snapshot) {
2141 * We keep the same mintxg; it could be >
2142 * ds_creation_txg if the previous snapshot was
2145 VERIFY(zap_add_int_key(dp->dp_meta_objset,
2146 scn->scn_phys.scn_queue_obj,
2147 dsl_dataset_phys(ds)->ds_next_snap_obj,
2149 zfs_dbgmsg("destroying ds %llu on %s; in queue; "
2150 "replacing with %llu",
2151 (u_longlong_t)ds->ds_object,
2152 dp->dp_spa->spa_name,
2153 (u_longlong_t)dsl_dataset_phys(ds)->
2156 zfs_dbgmsg("destroying ds %llu on %s; in queue; "
2158 (u_longlong_t)ds->ds_object,
2159 dp->dp_spa->spa_name);
2164 * dsl_scan_sync() should be called after this, and should sync
2165 * out our changed state, but just to be safe, do it here.
2167 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2171 ds_snapshotted_bookmark(dsl_dataset_t *ds, zbookmark_phys_t *scn_bookmark)
2173 if (scn_bookmark->zb_objset == ds->ds_object) {
2174 scn_bookmark->zb_objset =
2175 dsl_dataset_phys(ds)->ds_prev_snap_obj;
2176 zfs_dbgmsg("snapshotting ds %llu on %s; currently traversing; "
2177 "reset zb_objset to %llu",
2178 (u_longlong_t)ds->ds_object,
2179 ds->ds_dir->dd_pool->dp_spa->spa_name,
2180 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2185 * Called when a dataset is snapshotted. If we were currently traversing
2186 * this snapshot, we reset our bookmark to point at the newly created
2187 * snapshot. We also modify our work queue to remove the old snapshot and
2188 * replace with the new one.
2191 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
2193 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2194 dsl_scan_t *scn = dp->dp_scan;
2197 if (!dsl_scan_is_running(scn))
2200 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
2202 ds_snapshotted_bookmark(ds, &scn->scn_phys.scn_bookmark);
2203 ds_snapshotted_bookmark(ds, &scn->scn_phys_cached.scn_bookmark);
2205 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2206 scan_ds_queue_remove(scn, ds->ds_object);
2207 scan_ds_queue_insert(scn,
2208 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg);
2211 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2212 ds->ds_object, &mintxg) == 0) {
2213 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2214 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2215 VERIFY(zap_add_int_key(dp->dp_meta_objset,
2216 scn->scn_phys.scn_queue_obj,
2217 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
2218 zfs_dbgmsg("snapshotting ds %llu on %s; in queue; "
2219 "replacing with %llu",
2220 (u_longlong_t)ds->ds_object,
2221 dp->dp_spa->spa_name,
2222 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2225 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2229 ds_clone_swapped_bookmark(dsl_dataset_t *ds1, dsl_dataset_t *ds2,
2230 zbookmark_phys_t *scn_bookmark)
2232 if (scn_bookmark->zb_objset == ds1->ds_object) {
2233 scn_bookmark->zb_objset = ds2->ds_object;
2234 zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; "
2235 "reset zb_objset to %llu",
2236 (u_longlong_t)ds1->ds_object,
2237 ds1->ds_dir->dd_pool->dp_spa->spa_name,
2238 (u_longlong_t)ds2->ds_object);
2239 } else if (scn_bookmark->zb_objset == ds2->ds_object) {
2240 scn_bookmark->zb_objset = ds1->ds_object;
2241 zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; "
2242 "reset zb_objset to %llu",
2243 (u_longlong_t)ds2->ds_object,
2244 ds2->ds_dir->dd_pool->dp_spa->spa_name,
2245 (u_longlong_t)ds1->ds_object);
2250 * Called when an origin dataset and its clone are swapped. If we were
2251 * currently traversing the dataset, we need to switch to traversing the
2252 * newly promoted clone.
2255 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
2257 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
2258 dsl_scan_t *scn = dp->dp_scan;
2259 uint64_t mintxg1, mintxg2;
2260 boolean_t ds1_queued, ds2_queued;
2262 if (!dsl_scan_is_running(scn))
2265 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys.scn_bookmark);
2266 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys_cached.scn_bookmark);
2269 * Handle the in-memory scan queue.
2271 ds1_queued = scan_ds_queue_contains(scn, ds1->ds_object, &mintxg1);
2272 ds2_queued = scan_ds_queue_contains(scn, ds2->ds_object, &mintxg2);
2274 /* Sanity checking. */
2276 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2277 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2280 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2281 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2284 if (ds1_queued && ds2_queued) {
2286 * If both are queued, we don't need to do anything.
2287 * The swapping code below would not handle this case correctly,
2288 * since we can't insert ds2 if it is already there. That's
2289 * because scan_ds_queue_insert() prohibits a duplicate insert
2292 } else if (ds1_queued) {
2293 scan_ds_queue_remove(scn, ds1->ds_object);
2294 scan_ds_queue_insert(scn, ds2->ds_object, mintxg1);
2295 } else if (ds2_queued) {
2296 scan_ds_queue_remove(scn, ds2->ds_object);
2297 scan_ds_queue_insert(scn, ds1->ds_object, mintxg2);
2301 * Handle the on-disk scan queue.
2302 * The on-disk state is an out-of-date version of the in-memory state,
2303 * so the in-memory and on-disk values for ds1_queued and ds2_queued may
2304 * be different. Therefore we need to apply the swap logic to the
2305 * on-disk state independently of the in-memory state.
2307 ds1_queued = zap_lookup_int_key(dp->dp_meta_objset,
2308 scn->scn_phys.scn_queue_obj, ds1->ds_object, &mintxg1) == 0;
2309 ds2_queued = zap_lookup_int_key(dp->dp_meta_objset,
2310 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg2) == 0;
2312 /* Sanity checking. */
2314 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2315 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2318 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2319 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2322 if (ds1_queued && ds2_queued) {
2324 * If both are queued, we don't need to do anything.
2325 * Alternatively, we could check for EEXIST from
2326 * zap_add_int_key() and back out to the original state, but
2327 * that would be more work than checking for this case upfront.
2329 } else if (ds1_queued) {
2330 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2331 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
2332 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2333 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg1, tx));
2334 zfs_dbgmsg("clone_swap ds %llu on %s; in queue; "
2335 "replacing with %llu",
2336 (u_longlong_t)ds1->ds_object,
2337 dp->dp_spa->spa_name,
2338 (u_longlong_t)ds2->ds_object);
2339 } else if (ds2_queued) {
2340 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2341 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
2342 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2343 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg2, tx));
2344 zfs_dbgmsg("clone_swap ds %llu on %s; in queue; "
2345 "replacing with %llu",
2346 (u_longlong_t)ds2->ds_object,
2347 dp->dp_spa->spa_name,
2348 (u_longlong_t)ds1->ds_object);
2351 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2355 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2357 uint64_t originobj = *(uint64_t *)arg;
2360 dsl_scan_t *scn = dp->dp_scan;
2362 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != originobj)
2365 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2369 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != originobj) {
2370 dsl_dataset_t *prev;
2371 err = dsl_dataset_hold_obj(dp,
2372 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2374 dsl_dataset_rele(ds, FTAG);
2379 scan_ds_queue_insert(scn, ds->ds_object,
2380 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2381 dsl_dataset_rele(ds, FTAG);
2386 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
2388 dsl_pool_t *dp = scn->scn_dp;
2391 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2393 if (scn->scn_phys.scn_cur_min_txg >=
2394 scn->scn_phys.scn_max_txg) {
2396 * This can happen if this snapshot was created after the
2397 * scan started, and we already completed a previous snapshot
2398 * that was created after the scan started. This snapshot
2399 * only references blocks with:
2401 * birth < our ds_creation_txg
2402 * cur_min_txg is no less than ds_creation_txg.
2403 * We have already visited these blocks.
2405 * birth > scn_max_txg
2406 * The scan requested not to visit these blocks.
2408 * Subsequent snapshots (and clones) can reference our
2409 * blocks, or blocks with even higher birth times.
2410 * Therefore we do not need to visit them either,
2411 * so we do not add them to the work queue.
2413 * Note that checking for cur_min_txg >= cur_max_txg
2414 * is not sufficient, because in that case we may need to
2415 * visit subsequent snapshots. This happens when min_txg > 0,
2416 * which raises cur_min_txg. In this case we will visit
2417 * this dataset but skip all of its blocks, because the
2418 * rootbp's birth time is < cur_min_txg. Then we will
2419 * add the next snapshots/clones to the work queue.
2421 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2422 dsl_dataset_name(ds, dsname);
2423 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2424 "cur_min_txg (%llu) >= max_txg (%llu)",
2425 (longlong_t)dsobj, dsname,
2426 (longlong_t)scn->scn_phys.scn_cur_min_txg,
2427 (longlong_t)scn->scn_phys.scn_max_txg);
2428 kmem_free(dsname, MAXNAMELEN);
2434 * Only the ZIL in the head (non-snapshot) is valid. Even though
2435 * snapshots can have ZIL block pointers (which may be the same
2436 * BP as in the head), they must be ignored. In addition, $ORIGIN
2437 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2438 * need to look for a ZIL in it either. So we traverse the ZIL here,
2439 * rather than in scan_recurse(), because the regular snapshot
2440 * block-sharing rules don't apply to it.
2442 if (!dsl_dataset_is_snapshot(ds) &&
2443 (dp->dp_origin_snap == NULL ||
2444 ds->ds_dir != dp->dp_origin_snap->ds_dir)) {
2446 if (dmu_objset_from_ds(ds, &os) != 0) {
2449 dsl_scan_zil(dp, &os->os_zil_header);
2453 * Iterate over the bps in this ds.
2455 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2456 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
2457 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
2458 rrw_exit(&ds->ds_bp_rwlock, FTAG);
2460 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2461 dsl_dataset_name(ds, dsname);
2462 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2464 (longlong_t)dsobj, dsname,
2465 (longlong_t)scn->scn_phys.scn_cur_min_txg,
2466 (longlong_t)scn->scn_phys.scn_cur_max_txg,
2467 (int)scn->scn_suspending);
2468 kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
2470 if (scn->scn_suspending)
2474 * We've finished this pass over this dataset.
2478 * If we did not completely visit this dataset, do another pass.
2480 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
2481 zfs_dbgmsg("incomplete pass on %s; visiting again",
2482 dp->dp_spa->spa_name);
2483 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
2484 scan_ds_queue_insert(scn, ds->ds_object,
2485 scn->scn_phys.scn_cur_max_txg);
2490 * Add descendant datasets to work queue.
2492 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
2493 scan_ds_queue_insert(scn,
2494 dsl_dataset_phys(ds)->ds_next_snap_obj,
2495 dsl_dataset_phys(ds)->ds_creation_txg);
2497 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
2498 boolean_t usenext = B_FALSE;
2499 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
2502 * A bug in a previous version of the code could
2503 * cause upgrade_clones_cb() to not set
2504 * ds_next_snap_obj when it should, leading to a
2505 * missing entry. Therefore we can only use the
2506 * next_clones_obj when its count is correct.
2508 int err = zap_count(dp->dp_meta_objset,
2509 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
2511 count == dsl_dataset_phys(ds)->ds_num_children - 1)
2518 for (zap_cursor_init(&zc, dp->dp_meta_objset,
2519 dsl_dataset_phys(ds)->ds_next_clones_obj);
2520 zap_cursor_retrieve(&zc, &za) == 0;
2521 (void) zap_cursor_advance(&zc)) {
2522 scan_ds_queue_insert(scn,
2523 zfs_strtonum(za.za_name, NULL),
2524 dsl_dataset_phys(ds)->ds_creation_txg);
2526 zap_cursor_fini(&zc);
2528 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2529 enqueue_clones_cb, &ds->ds_object,
2535 dsl_dataset_rele(ds, FTAG);
2539 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2544 dsl_scan_t *scn = dp->dp_scan;
2546 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2550 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
2551 dsl_dataset_t *prev;
2552 err = dsl_dataset_hold_obj(dp,
2553 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2555 dsl_dataset_rele(ds, FTAG);
2560 * If this is a clone, we don't need to worry about it for now.
2562 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
2563 dsl_dataset_rele(ds, FTAG);
2564 dsl_dataset_rele(prev, FTAG);
2567 dsl_dataset_rele(ds, FTAG);
2571 scan_ds_queue_insert(scn, ds->ds_object,
2572 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2573 dsl_dataset_rele(ds, FTAG);
2578 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
2579 ddt_entry_t *dde, dmu_tx_t *tx)
2582 const ddt_key_t *ddk = &dde->dde_key;
2583 ddt_phys_t *ddp = dde->dde_phys;
2585 zbookmark_phys_t zb = { 0 };
2587 if (!dsl_scan_is_running(scn))
2591 * This function is special because it is the only thing
2592 * that can add scan_io_t's to the vdev scan queues from
2593 * outside dsl_scan_sync(). For the most part this is ok
2594 * as long as it is called from within syncing context.
2595 * However, dsl_scan_sync() expects that no new sio's will
2596 * be added between when all the work for a scan is done
2597 * and the next txg when the scan is actually marked as
2598 * completed. This check ensures we do not issue new sio's
2599 * during this period.
2601 if (scn->scn_done_txg != 0)
2604 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2605 if (ddp->ddp_phys_birth == 0 ||
2606 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
2608 ddt_bp_create(checksum, ddk, ddp, &bp);
2610 scn->scn_visited_this_txg++;
2611 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
2616 * Scrub/dedup interaction.
2618 * If there are N references to a deduped block, we don't want to scrub it
2619 * N times -- ideally, we should scrub it exactly once.
2621 * We leverage the fact that the dde's replication class (enum ddt_class)
2622 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2623 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2625 * To prevent excess scrubbing, the scrub begins by walking the DDT
2626 * to find all blocks with refcnt > 1, and scrubs each of these once.
2627 * Since there are two replication classes which contain blocks with
2628 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2629 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2631 * There would be nothing more to say if a block's refcnt couldn't change
2632 * during a scrub, but of course it can so we must account for changes
2633 * in a block's replication class.
2635 * Here's an example of what can occur:
2637 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2638 * when visited during the top-down scrub phase, it will be scrubbed twice.
2639 * This negates our scrub optimization, but is otherwise harmless.
2641 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2642 * on each visit during the top-down scrub phase, it will never be scrubbed.
2643 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2644 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2645 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2646 * while a scrub is in progress, it scrubs the block right then.
2649 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
2651 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
2652 ddt_entry_t dde = {{{{0}}}};
2656 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
2659 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
2661 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
2662 (longlong_t)ddb->ddb_class,
2663 (longlong_t)ddb->ddb_type,
2664 (longlong_t)ddb->ddb_checksum,
2665 (longlong_t)ddb->ddb_cursor);
2667 /* There should be no pending changes to the dedup table */
2668 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
2669 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
2671 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
2674 if (dsl_scan_check_suspend(scn, NULL))
2678 zfs_dbgmsg("scanned %llu ddt entries on %s with class_max = %u; "
2679 "suspending=%u", (longlong_t)n, scn->scn_dp->dp_spa->spa_name,
2680 (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
2682 ASSERT(error == 0 || error == ENOENT);
2683 ASSERT(error != ENOENT ||
2684 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
2688 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
2690 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
2691 if (ds->ds_is_snapshot)
2692 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
2697 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
2700 dsl_pool_t *dp = scn->scn_dp;
2702 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
2703 scn->scn_phys.scn_ddt_class_max) {
2704 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2705 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2706 dsl_scan_ddt(scn, tx);
2707 if (scn->scn_suspending)
2711 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
2712 /* First do the MOS & ORIGIN */
2714 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2715 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2716 dsl_scan_visit_rootbp(scn, NULL,
2717 &dp->dp_meta_rootbp, tx);
2718 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
2719 if (scn->scn_suspending)
2722 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
2723 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2724 enqueue_cb, NULL, DS_FIND_CHILDREN));
2726 dsl_scan_visitds(scn,
2727 dp->dp_origin_snap->ds_object, tx);
2729 ASSERT(!scn->scn_suspending);
2730 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
2731 ZB_DESTROYED_OBJSET) {
2732 uint64_t dsobj = scn->scn_phys.scn_bookmark.zb_objset;
2734 * If we were suspended, continue from here. Note if the
2735 * ds we were suspended on was deleted, the zb_objset may
2736 * be -1, so we will skip this and find a new objset
2739 dsl_scan_visitds(scn, dsobj, tx);
2740 if (scn->scn_suspending)
2745 * In case we suspended right at the end of the ds, zero the
2746 * bookmark so we don't think that we're still trying to resume.
2748 memset(&scn->scn_phys.scn_bookmark, 0, sizeof (zbookmark_phys_t));
2751 * Keep pulling things out of the dataset avl queue. Updates to the
2752 * persistent zap-object-as-queue happen only at checkpoints.
2754 while ((sds = avl_first(&scn->scn_queue)) != NULL) {
2756 uint64_t dsobj = sds->sds_dsobj;
2757 uint64_t txg = sds->sds_txg;
2759 /* dequeue and free the ds from the queue */
2760 scan_ds_queue_remove(scn, dsobj);
2763 /* set up min / max txg */
2764 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2766 scn->scn_phys.scn_cur_min_txg =
2767 MAX(scn->scn_phys.scn_min_txg, txg);
2769 scn->scn_phys.scn_cur_min_txg =
2770 MAX(scn->scn_phys.scn_min_txg,
2771 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2773 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
2774 dsl_dataset_rele(ds, FTAG);
2776 dsl_scan_visitds(scn, dsobj, tx);
2777 if (scn->scn_suspending)
2781 /* No more objsets to fetch, we're done */
2782 scn->scn_phys.scn_bookmark.zb_objset = ZB_DESTROYED_OBJSET;
2783 ASSERT0(scn->scn_suspending);
2787 dsl_scan_count_data_disks(vdev_t *rvd)
2789 uint64_t i, leaves = 0;
2791 for (i = 0; i < rvd->vdev_children; i++) {
2792 vdev_t *vd = rvd->vdev_child[i];
2793 if (vd->vdev_islog || vd->vdev_isspare || vd->vdev_isl2cache)
2795 leaves += vdev_get_ndisks(vd) - vdev_get_nparity(vd);
2801 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t *q, const blkptr_t *bp)
2804 uint64_t cur_size = 0;
2806 for (i = 0; i < BP_GET_NDVAS(bp); i++) {
2807 cur_size += DVA_GET_ASIZE(&bp->blk_dva[i]);
2810 q->q_total_zio_size_this_txg += cur_size;
2811 q->q_zios_this_txg++;
2815 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t *q, uint64_t start,
2818 q->q_total_seg_size_this_txg += end - start;
2819 q->q_segs_this_txg++;
2823 scan_io_queue_check_suspend(dsl_scan_t *scn)
2825 /* See comment in dsl_scan_check_suspend() */
2826 uint64_t curr_time_ns = gethrtime();
2827 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
2828 uint64_t sync_time_ns = curr_time_ns -
2829 scn->scn_dp->dp_spa->spa_sync_starttime;
2830 int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
2831 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
2832 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
2834 return ((NSEC2MSEC(scan_time_ns) > mintime &&
2835 (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
2836 txg_sync_waiting(scn->scn_dp) ||
2837 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
2838 spa_shutting_down(scn->scn_dp->dp_spa));
2842 * Given a list of scan_io_t's in io_list, this issues the I/Os out to
2843 * disk. This consumes the io_list and frees the scan_io_t's. This is
2844 * called when emptying queues, either when we're up against the memory
2845 * limit or when we have finished scanning. Returns B_TRUE if we stopped
2846 * processing the list before we finished. Any sios that were not issued
2847 * will remain in the io_list.
2850 scan_io_queue_issue(dsl_scan_io_queue_t *queue, list_t *io_list)
2852 dsl_scan_t *scn = queue->q_scn;
2854 int64_t bytes_issued = 0;
2855 boolean_t suspended = B_FALSE;
2857 while ((sio = list_head(io_list)) != NULL) {
2860 if (scan_io_queue_check_suspend(scn)) {
2866 bytes_issued += SIO_GET_ASIZE(sio);
2867 scan_exec_io(scn->scn_dp, &bp, sio->sio_flags,
2868 &sio->sio_zb, queue);
2869 (void) list_remove_head(io_list);
2870 scan_io_queues_update_zio_stats(queue, &bp);
2874 atomic_add_64(&scn->scn_bytes_pending, -bytes_issued);
2880 * This function removes sios from an IO queue which reside within a given
2881 * range_seg_t and inserts them (in offset order) into a list. Note that
2882 * we only ever return a maximum of 32 sios at once. If there are more sios
2883 * to process within this segment that did not make it onto the list we
2884 * return B_TRUE and otherwise B_FALSE.
2887 scan_io_queue_gather(dsl_scan_io_queue_t *queue, range_seg_t *rs, list_t *list)
2889 scan_io_t *srch_sio, *sio, *next_sio;
2891 uint_t num_sios = 0;
2892 int64_t bytes_issued = 0;
2895 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2897 srch_sio = sio_alloc(1);
2898 srch_sio->sio_nr_dvas = 1;
2899 SIO_SET_OFFSET(srch_sio, rs_get_start(rs, queue->q_exts_by_addr));
2902 * The exact start of the extent might not contain any matching zios,
2903 * so if that's the case, examine the next one in the tree.
2905 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
2909 sio = avl_nearest(&queue->q_sios_by_addr, idx, AVL_AFTER);
2911 while (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2912 queue->q_exts_by_addr) && num_sios <= 32) {
2913 ASSERT3U(SIO_GET_OFFSET(sio), >=, rs_get_start(rs,
2914 queue->q_exts_by_addr));
2915 ASSERT3U(SIO_GET_END_OFFSET(sio), <=, rs_get_end(rs,
2916 queue->q_exts_by_addr));
2918 next_sio = AVL_NEXT(&queue->q_sios_by_addr, sio);
2919 avl_remove(&queue->q_sios_by_addr, sio);
2920 queue->q_sio_memused -= SIO_GET_MUSED(sio);
2922 bytes_issued += SIO_GET_ASIZE(sio);
2924 list_insert_tail(list, sio);
2929 * We limit the number of sios we process at once to 32 to avoid
2930 * biting off more than we can chew. If we didn't take everything
2931 * in the segment we update it to reflect the work we were able to
2932 * complete. Otherwise, we remove it from the range tree entirely.
2934 if (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2935 queue->q_exts_by_addr)) {
2936 range_tree_adjust_fill(queue->q_exts_by_addr, rs,
2938 range_tree_resize_segment(queue->q_exts_by_addr, rs,
2939 SIO_GET_OFFSET(sio), rs_get_end(rs,
2940 queue->q_exts_by_addr) - SIO_GET_OFFSET(sio));
2944 uint64_t rstart = rs_get_start(rs, queue->q_exts_by_addr);
2945 uint64_t rend = rs_get_end(rs, queue->q_exts_by_addr);
2946 range_tree_remove(queue->q_exts_by_addr, rstart, rend - rstart);
2952 * This is called from the queue emptying thread and selects the next
2953 * extent from which we are to issue I/Os. The behavior of this function
2954 * depends on the state of the scan, the current memory consumption and
2955 * whether or not we are performing a scan shutdown.
2956 * 1) We select extents in an elevator algorithm (LBA-order) if the scan
2957 * needs to perform a checkpoint
2958 * 2) We select the largest available extent if we are up against the
2960 * 3) Otherwise we don't select any extents.
2962 static range_seg_t *
2963 scan_io_queue_fetch_ext(dsl_scan_io_queue_t *queue)
2965 dsl_scan_t *scn = queue->q_scn;
2966 range_tree_t *rt = queue->q_exts_by_addr;
2968 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2969 ASSERT(scn->scn_is_sorted);
2971 /* handle tunable overrides */
2972 if (scn->scn_checkpointing || scn->scn_clearing) {
2973 if (zfs_scan_issue_strategy == 1) {
2974 return (range_tree_first(rt));
2975 } else if (zfs_scan_issue_strategy == 2) {
2977 * We need to get the original entry in the by_addr
2978 * tree so we can modify it.
2980 range_seg_t *size_rs =
2981 zfs_btree_first(&queue->q_exts_by_size, NULL);
2982 if (size_rs == NULL)
2984 uint64_t start = rs_get_start(size_rs, rt);
2985 uint64_t size = rs_get_end(size_rs, rt) - start;
2986 range_seg_t *addr_rs = range_tree_find(rt, start,
2988 ASSERT3P(addr_rs, !=, NULL);
2989 ASSERT3U(rs_get_start(size_rs, rt), ==,
2990 rs_get_start(addr_rs, rt));
2991 ASSERT3U(rs_get_end(size_rs, rt), ==,
2992 rs_get_end(addr_rs, rt));
2998 * During normal clearing, we want to issue our largest segments
2999 * first, keeping IO as sequential as possible, and leaving the
3000 * smaller extents for later with the hope that they might eventually
3001 * grow to larger sequential segments. However, when the scan is
3002 * checkpointing, no new extents will be added to the sorting queue,
3003 * so the way we are sorted now is as good as it will ever get.
3004 * In this case, we instead switch to issuing extents in LBA order.
3006 if (scn->scn_checkpointing) {
3007 return (range_tree_first(rt));
3008 } else if (scn->scn_clearing) {
3010 * We need to get the original entry in the by_addr
3011 * tree so we can modify it.
3013 range_seg_t *size_rs = zfs_btree_first(&queue->q_exts_by_size,
3015 if (size_rs == NULL)
3017 uint64_t start = rs_get_start(size_rs, rt);
3018 uint64_t size = rs_get_end(size_rs, rt) - start;
3019 range_seg_t *addr_rs = range_tree_find(rt, start, size);
3020 ASSERT3P(addr_rs, !=, NULL);
3021 ASSERT3U(rs_get_start(size_rs, rt), ==, rs_get_start(addr_rs,
3023 ASSERT3U(rs_get_end(size_rs, rt), ==, rs_get_end(addr_rs, rt));
3031 scan_io_queues_run_one(void *arg)
3033 dsl_scan_io_queue_t *queue = arg;
3034 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
3035 boolean_t suspended = B_FALSE;
3036 range_seg_t *rs = NULL;
3037 scan_io_t *sio = NULL;
3040 ASSERT(queue->q_scn->scn_is_sorted);
3042 list_create(&sio_list, sizeof (scan_io_t),
3043 offsetof(scan_io_t, sio_nodes.sio_list_node));
3044 mutex_enter(q_lock);
3046 /* Calculate maximum in-flight bytes for this vdev. */
3047 queue->q_maxinflight_bytes = MAX(1, zfs_scan_vdev_limit *
3048 (vdev_get_ndisks(queue->q_vd) - vdev_get_nparity(queue->q_vd)));
3050 /* reset per-queue scan statistics for this txg */
3051 queue->q_total_seg_size_this_txg = 0;
3052 queue->q_segs_this_txg = 0;
3053 queue->q_total_zio_size_this_txg = 0;
3054 queue->q_zios_this_txg = 0;
3056 /* loop until we run out of time or sios */
3057 while ((rs = scan_io_queue_fetch_ext(queue)) != NULL) {
3058 uint64_t seg_start = 0, seg_end = 0;
3059 boolean_t more_left = B_TRUE;
3061 ASSERT(list_is_empty(&sio_list));
3063 /* loop while we still have sios left to process in this rs */
3065 scan_io_t *first_sio, *last_sio;
3068 * We have selected which extent needs to be
3069 * processed next. Gather up the corresponding sios.
3071 more_left = scan_io_queue_gather(queue, rs, &sio_list);
3072 ASSERT(!list_is_empty(&sio_list));
3073 first_sio = list_head(&sio_list);
3074 last_sio = list_tail(&sio_list);
3076 seg_end = SIO_GET_END_OFFSET(last_sio);
3078 seg_start = SIO_GET_OFFSET(first_sio);
3081 * Issuing sios can take a long time so drop the
3082 * queue lock. The sio queue won't be updated by
3083 * other threads since we're in syncing context so
3084 * we can be sure that our trees will remain exactly
3088 suspended = scan_io_queue_issue(queue, &sio_list);
3089 mutex_enter(q_lock);
3095 /* update statistics for debugging purposes */
3096 scan_io_queues_update_seg_stats(queue, seg_start, seg_end);
3103 * If we were suspended in the middle of processing,
3104 * requeue any unfinished sios and exit.
3106 while ((sio = list_head(&sio_list)) != NULL) {
3107 list_remove(&sio_list, sio);
3108 scan_io_queue_insert_impl(queue, sio);
3112 list_destroy(&sio_list);
3116 * Performs an emptying run on all scan queues in the pool. This just
3117 * punches out one thread per top-level vdev, each of which processes
3118 * only that vdev's scan queue. We can parallelize the I/O here because
3119 * we know that each queue's I/Os only affect its own top-level vdev.
3121 * This function waits for the queue runs to complete, and must be
3122 * called from dsl_scan_sync (or in general, syncing context).
3125 scan_io_queues_run(dsl_scan_t *scn)
3127 spa_t *spa = scn->scn_dp->dp_spa;
3129 ASSERT(scn->scn_is_sorted);
3130 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
3132 if (scn->scn_bytes_pending == 0)
3135 if (scn->scn_taskq == NULL) {
3136 int nthreads = spa->spa_root_vdev->vdev_children;
3139 * We need to make this taskq *always* execute as many
3140 * threads in parallel as we have top-level vdevs and no
3141 * less, otherwise strange serialization of the calls to
3142 * scan_io_queues_run_one can occur during spa_sync runs
3143 * and that significantly impacts performance.
3145 scn->scn_taskq = taskq_create("dsl_scan_iss", nthreads,
3146 minclsyspri, nthreads, nthreads, TASKQ_PREPOPULATE);
3149 for (uint64_t i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3150 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3152 mutex_enter(&vd->vdev_scan_io_queue_lock);
3153 if (vd->vdev_scan_io_queue != NULL) {
3154 VERIFY(taskq_dispatch(scn->scn_taskq,
3155 scan_io_queues_run_one, vd->vdev_scan_io_queue,
3156 TQ_SLEEP) != TASKQID_INVALID);
3158 mutex_exit(&vd->vdev_scan_io_queue_lock);
3162 * Wait for the queues to finish issuing their IOs for this run
3163 * before we return. There may still be IOs in flight at this
3166 taskq_wait(scn->scn_taskq);
3170 dsl_scan_async_block_should_pause(dsl_scan_t *scn)
3172 uint64_t elapsed_nanosecs;
3177 if (zfs_async_block_max_blocks != 0 &&
3178 scn->scn_visited_this_txg >= zfs_async_block_max_blocks) {
3182 if (zfs_max_async_dedup_frees != 0 &&
3183 scn->scn_dedup_frees_this_txg >= zfs_max_async_dedup_frees) {
3187 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
3188 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
3189 (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms &&
3190 txg_sync_waiting(scn->scn_dp)) ||
3191 spa_shutting_down(scn->scn_dp->dp_spa));
3195 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
3197 dsl_scan_t *scn = arg;
3199 if (!scn->scn_is_bptree ||
3200 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
3201 if (dsl_scan_async_block_should_pause(scn))
3202 return (SET_ERROR(ERESTART));
3205 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
3206 dmu_tx_get_txg(tx), bp, 0));
3207 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
3208 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
3209 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
3210 scn->scn_visited_this_txg++;
3211 if (BP_GET_DEDUP(bp))
3212 scn->scn_dedup_frees_this_txg++;
3217 dsl_scan_update_stats(dsl_scan_t *scn)
3219 spa_t *spa = scn->scn_dp->dp_spa;
3221 uint64_t seg_size_total = 0, zio_size_total = 0;
3222 uint64_t seg_count_total = 0, zio_count_total = 0;
3224 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3225 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3226 dsl_scan_io_queue_t *queue = vd->vdev_scan_io_queue;
3231 seg_size_total += queue->q_total_seg_size_this_txg;
3232 zio_size_total += queue->q_total_zio_size_this_txg;
3233 seg_count_total += queue->q_segs_this_txg;
3234 zio_count_total += queue->q_zios_this_txg;
3237 if (seg_count_total == 0 || zio_count_total == 0) {
3238 scn->scn_avg_seg_size_this_txg = 0;
3239 scn->scn_avg_zio_size_this_txg = 0;
3240 scn->scn_segs_this_txg = 0;
3241 scn->scn_zios_this_txg = 0;
3245 scn->scn_avg_seg_size_this_txg = seg_size_total / seg_count_total;
3246 scn->scn_avg_zio_size_this_txg = zio_size_total / zio_count_total;
3247 scn->scn_segs_this_txg = seg_count_total;
3248 scn->scn_zios_this_txg = zio_count_total;
3252 bpobj_dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3256 return (dsl_scan_free_block_cb(arg, bp, tx));
3260 dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3264 dsl_scan_t *scn = arg;
3265 const dva_t *dva = &bp->blk_dva[0];
3267 if (dsl_scan_async_block_should_pause(scn))
3268 return (SET_ERROR(ERESTART));
3270 spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa,
3271 DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva),
3272 DVA_GET_ASIZE(dva), tx);
3273 scn->scn_visited_this_txg++;
3278 dsl_scan_active(dsl_scan_t *scn)
3280 spa_t *spa = scn->scn_dp->dp_spa;
3281 uint64_t used = 0, comp, uncomp;
3282 boolean_t clones_left;
3284 if (spa->spa_load_state != SPA_LOAD_NONE)
3286 if (spa_shutting_down(spa))
3288 if ((dsl_scan_is_running(scn) && !dsl_scan_is_paused_scrub(scn)) ||
3289 (scn->scn_async_destroying && !scn->scn_async_stalled))
3292 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
3293 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
3294 &used, &comp, &uncomp);
3296 clones_left = spa_livelist_delete_check(spa);
3297 return ((used != 0) || (clones_left));
3301 dsl_scan_check_deferred(vdev_t *vd)
3303 boolean_t need_resilver = B_FALSE;
3305 for (int c = 0; c < vd->vdev_children; c++) {
3307 dsl_scan_check_deferred(vd->vdev_child[c]);
3310 if (!vdev_is_concrete(vd) || vd->vdev_aux ||
3311 !vd->vdev_ops->vdev_op_leaf)
3312 return (need_resilver);
3314 if (!vd->vdev_resilver_deferred)
3315 need_resilver = B_TRUE;
3317 return (need_resilver);
3321 dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize,
3322 uint64_t phys_birth)
3326 vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
3328 if (vd->vdev_ops == &vdev_indirect_ops) {
3330 * The indirect vdev can point to multiple
3331 * vdevs. For simplicity, always create
3332 * the resilver zio_t. zio_vdev_io_start()
3333 * will bypass the child resilver i/o's if
3334 * they are on vdevs that don't have DTL's.
3339 if (DVA_GET_GANG(dva)) {
3341 * Gang members may be spread across multiple
3342 * vdevs, so the best estimate we have is the
3343 * scrub range, which has already been checked.
3344 * XXX -- it would be better to change our
3345 * allocation policy to ensure that all
3346 * gang members reside on the same vdev.
3352 * Check if the top-level vdev must resilver this offset.
3353 * When the offset does not intersect with a dirty leaf DTL
3354 * then it may be possible to skip the resilver IO. The psize
3355 * is provided instead of asize to simplify the check for RAIDZ.
3357 if (!vdev_dtl_need_resilver(vd, dva, psize, phys_birth))
3361 * Check that this top-level vdev has a device under it which
3362 * is resilvering and is not deferred.
3364 if (!dsl_scan_check_deferred(vd))
3371 dsl_process_async_destroys(dsl_pool_t *dp, dmu_tx_t *tx)
3373 dsl_scan_t *scn = dp->dp_scan;
3374 spa_t *spa = dp->dp_spa;
3377 if (spa_suspend_async_destroy(spa))
3380 if (zfs_free_bpobj_enabled &&
3381 spa_version(spa) >= SPA_VERSION_DEADLISTS) {
3382 scn->scn_is_bptree = B_FALSE;
3383 scn->scn_async_block_min_time_ms = zfs_free_min_time_ms;
3384 scn->scn_zio_root = zio_root(spa, NULL,
3385 NULL, ZIO_FLAG_MUSTSUCCEED);
3386 err = bpobj_iterate(&dp->dp_free_bpobj,
3387 bpobj_dsl_scan_free_block_cb, scn, tx);
3388 VERIFY0(zio_wait(scn->scn_zio_root));
3389 scn->scn_zio_root = NULL;
3391 if (err != 0 && err != ERESTART)
3392 zfs_panic_recover("error %u from bpobj_iterate()", err);
3395 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
3396 ASSERT(scn->scn_async_destroying);
3397 scn->scn_is_bptree = B_TRUE;
3398 scn->scn_zio_root = zio_root(spa, NULL,
3399 NULL, ZIO_FLAG_MUSTSUCCEED);
3400 err = bptree_iterate(dp->dp_meta_objset,
3401 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
3402 VERIFY0(zio_wait(scn->scn_zio_root));
3403 scn->scn_zio_root = NULL;
3405 if (err == EIO || err == ECKSUM) {
3407 } else if (err != 0 && err != ERESTART) {
3408 zfs_panic_recover("error %u from "
3409 "traverse_dataset_destroyed()", err);
3412 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
3413 /* finished; deactivate async destroy feature */
3414 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
3415 ASSERT(!spa_feature_is_active(spa,
3416 SPA_FEATURE_ASYNC_DESTROY));
3417 VERIFY0(zap_remove(dp->dp_meta_objset,
3418 DMU_POOL_DIRECTORY_OBJECT,
3419 DMU_POOL_BPTREE_OBJ, tx));
3420 VERIFY0(bptree_free(dp->dp_meta_objset,
3421 dp->dp_bptree_obj, tx));
3422 dp->dp_bptree_obj = 0;
3423 scn->scn_async_destroying = B_FALSE;
3424 scn->scn_async_stalled = B_FALSE;
3427 * If we didn't make progress, mark the async
3428 * destroy as stalled, so that we will not initiate
3429 * a spa_sync() on its behalf. Note that we only
3430 * check this if we are not finished, because if the
3431 * bptree had no blocks for us to visit, we can
3432 * finish without "making progress".
3434 scn->scn_async_stalled =
3435 (scn->scn_visited_this_txg == 0);
3438 if (scn->scn_visited_this_txg) {
3439 zfs_dbgmsg("freed %llu blocks in %llums from "
3440 "free_bpobj/bptree on %s in txg %llu; err=%u",
3441 (longlong_t)scn->scn_visited_this_txg,
3443 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
3444 spa->spa_name, (longlong_t)tx->tx_txg, err);
3445 scn->scn_visited_this_txg = 0;
3446 scn->scn_dedup_frees_this_txg = 0;
3449 * Write out changes to the DDT that may be required as a
3450 * result of the blocks freed. This ensures that the DDT
3451 * is clean when a scrub/resilver runs.
3453 ddt_sync(spa, tx->tx_txg);
3457 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3458 zfs_free_leak_on_eio &&
3459 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
3460 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
3461 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
3463 * We have finished background destroying, but there is still
3464 * some space left in the dp_free_dir. Transfer this leaked
3465 * space to the dp_leak_dir.
3467 if (dp->dp_leak_dir == NULL) {
3468 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
3469 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
3471 VERIFY0(dsl_pool_open_special_dir(dp,
3472 LEAK_DIR_NAME, &dp->dp_leak_dir));
3473 rrw_exit(&dp->dp_config_rwlock, FTAG);
3475 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
3476 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3477 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3478 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3479 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
3480 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3481 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3482 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3485 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3486 !spa_livelist_delete_check(spa)) {
3487 /* finished; verify that space accounting went to zero */
3488 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
3489 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
3490 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
3493 spa_notify_waiters(spa);
3495 EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj),
3496 0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3497 DMU_POOL_OBSOLETE_BPOBJ));
3498 if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) {
3499 ASSERT(spa_feature_is_active(dp->dp_spa,
3500 SPA_FEATURE_OBSOLETE_COUNTS));
3502 scn->scn_is_bptree = B_FALSE;
3503 scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms;
3504 err = bpobj_iterate(&dp->dp_obsolete_bpobj,
3505 dsl_scan_obsolete_block_cb, scn, tx);
3506 if (err != 0 && err != ERESTART)
3507 zfs_panic_recover("error %u from bpobj_iterate()", err);
3509 if (bpobj_is_empty(&dp->dp_obsolete_bpobj))
3510 dsl_pool_destroy_obsolete_bpobj(dp, tx);
3516 * This is the primary entry point for scans that is called from syncing
3517 * context. Scans must happen entirely during syncing context so that we
3518 * can guarantee that blocks we are currently scanning will not change out
3519 * from under us. While a scan is active, this function controls how quickly
3520 * transaction groups proceed, instead of the normal handling provided by
3521 * txg_sync_thread().
3524 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
3527 dsl_scan_t *scn = dp->dp_scan;
3528 spa_t *spa = dp->dp_spa;
3529 state_sync_type_t sync_type = SYNC_OPTIONAL;
3531 if (spa->spa_resilver_deferred &&
3532 !spa_feature_is_active(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
3533 spa_feature_incr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
3536 * Check for scn_restart_txg before checking spa_load_state, so
3537 * that we can restart an old-style scan while the pool is being
3538 * imported (see dsl_scan_init). We also restart scans if there
3539 * is a deferred resilver and the user has manually disabled
3540 * deferred resilvers via the tunable.
3542 if (dsl_scan_restarting(scn, tx) ||
3543 (spa->spa_resilver_deferred && zfs_resilver_disable_defer)) {
3544 pool_scan_func_t func = POOL_SCAN_SCRUB;
3545 dsl_scan_done(scn, B_FALSE, tx);
3546 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
3547 func = POOL_SCAN_RESILVER;
3548 zfs_dbgmsg("restarting scan func=%u on %s txg=%llu",
3549 func, dp->dp_spa->spa_name, (longlong_t)tx->tx_txg);
3550 dsl_scan_setup_sync(&func, tx);
3554 * Only process scans in sync pass 1.
3556 if (spa_sync_pass(spa) > 1)
3560 * If the spa is shutting down, then stop scanning. This will
3561 * ensure that the scan does not dirty any new data during the
3564 if (spa_shutting_down(spa))
3568 * If the scan is inactive due to a stalled async destroy, try again.
3570 if (!scn->scn_async_stalled && !dsl_scan_active(scn))
3573 /* reset scan statistics */
3574 scn->scn_visited_this_txg = 0;
3575 scn->scn_dedup_frees_this_txg = 0;
3576 scn->scn_holes_this_txg = 0;
3577 scn->scn_lt_min_this_txg = 0;
3578 scn->scn_gt_max_this_txg = 0;
3579 scn->scn_ddt_contained_this_txg = 0;
3580 scn->scn_objsets_visited_this_txg = 0;
3581 scn->scn_avg_seg_size_this_txg = 0;
3582 scn->scn_segs_this_txg = 0;
3583 scn->scn_avg_zio_size_this_txg = 0;
3584 scn->scn_zios_this_txg = 0;
3585 scn->scn_suspending = B_FALSE;
3586 scn->scn_sync_start_time = gethrtime();
3587 spa->spa_scrub_active = B_TRUE;
3590 * First process the async destroys. If we suspend, don't do
3591 * any scrubbing or resilvering. This ensures that there are no
3592 * async destroys while we are scanning, so the scan code doesn't
3593 * have to worry about traversing it. It is also faster to free the
3594 * blocks than to scrub them.
3596 err = dsl_process_async_destroys(dp, tx);
3600 if (!dsl_scan_is_running(scn) || dsl_scan_is_paused_scrub(scn))
3604 * Wait a few txgs after importing to begin scanning so that
3605 * we can get the pool imported quickly.
3607 if (spa->spa_syncing_txg < spa->spa_first_txg + SCAN_IMPORT_WAIT_TXGS)
3611 * zfs_scan_suspend_progress can be set to disable scan progress.
3612 * We don't want to spin the txg_sync thread, so we add a delay
3613 * here to simulate the time spent doing a scan. This is mostly
3614 * useful for testing and debugging.
3616 if (zfs_scan_suspend_progress) {
3617 uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time;
3618 int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
3619 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
3621 while (zfs_scan_suspend_progress &&
3622 !txg_sync_waiting(scn->scn_dp) &&
3623 !spa_shutting_down(scn->scn_dp->dp_spa) &&
3624 NSEC2MSEC(scan_time_ns) < mintime) {
3626 scan_time_ns = gethrtime() - scn->scn_sync_start_time;
3632 * It is possible to switch from unsorted to sorted at any time,
3633 * but afterwards the scan will remain sorted unless reloaded from
3634 * a checkpoint after a reboot.
3636 if (!zfs_scan_legacy) {
3637 scn->scn_is_sorted = B_TRUE;
3638 if (scn->scn_last_checkpoint == 0)
3639 scn->scn_last_checkpoint = ddi_get_lbolt();
3643 * For sorted scans, determine what kind of work we will be doing
3644 * this txg based on our memory limitations and whether or not we
3645 * need to perform a checkpoint.
3647 if (scn->scn_is_sorted) {
3649 * If we are over our checkpoint interval, set scn_clearing
3650 * so that we can begin checkpointing immediately. The
3651 * checkpoint allows us to save a consistent bookmark
3652 * representing how much data we have scrubbed so far.
3653 * Otherwise, use the memory limit to determine if we should
3654 * scan for metadata or start issue scrub IOs. We accumulate
3655 * metadata until we hit our hard memory limit at which point
3656 * we issue scrub IOs until we are at our soft memory limit.
3658 if (scn->scn_checkpointing ||
3659 ddi_get_lbolt() - scn->scn_last_checkpoint >
3660 SEC_TO_TICK(zfs_scan_checkpoint_intval)) {
3661 if (!scn->scn_checkpointing)
3662 zfs_dbgmsg("begin scan checkpoint for %s",
3665 scn->scn_checkpointing = B_TRUE;
3666 scn->scn_clearing = B_TRUE;
3668 boolean_t should_clear = dsl_scan_should_clear(scn);
3669 if (should_clear && !scn->scn_clearing) {
3670 zfs_dbgmsg("begin scan clearing for %s",
3672 scn->scn_clearing = B_TRUE;
3673 } else if (!should_clear && scn->scn_clearing) {
3674 zfs_dbgmsg("finish scan clearing for %s",
3676 scn->scn_clearing = B_FALSE;
3680 ASSERT0(scn->scn_checkpointing);
3681 ASSERT0(scn->scn_clearing);
3684 if (!scn->scn_clearing && scn->scn_done_txg == 0) {
3685 /* Need to scan metadata for more blocks to scrub */
3686 dsl_scan_phys_t *scnp = &scn->scn_phys;
3687 taskqid_t prefetch_tqid;
3690 * Recalculate the max number of in-flight bytes for pool-wide
3691 * scanning operations (minimum 1MB). Limits for the issuing
3692 * phase are done per top-level vdev and are handled separately.
3694 scn->scn_maxinflight_bytes = MAX(zfs_scan_vdev_limit *
3695 dsl_scan_count_data_disks(spa->spa_root_vdev), 1ULL << 20);
3697 if (scnp->scn_ddt_bookmark.ddb_class <=
3698 scnp->scn_ddt_class_max) {
3699 ASSERT(ZB_IS_ZERO(&scnp->scn_bookmark));
3700 zfs_dbgmsg("doing scan sync for %s txg %llu; "
3701 "ddt bm=%llu/%llu/%llu/%llx",
3703 (longlong_t)tx->tx_txg,
3704 (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
3705 (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
3706 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
3707 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
3709 zfs_dbgmsg("doing scan sync for %s txg %llu; "
3710 "bm=%llu/%llu/%llu/%llu",
3712 (longlong_t)tx->tx_txg,
3713 (longlong_t)scnp->scn_bookmark.zb_objset,
3714 (longlong_t)scnp->scn_bookmark.zb_object,
3715 (longlong_t)scnp->scn_bookmark.zb_level,
3716 (longlong_t)scnp->scn_bookmark.zb_blkid);
3719 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3720 NULL, ZIO_FLAG_CANFAIL);
3722 scn->scn_prefetch_stop = B_FALSE;
3723 prefetch_tqid = taskq_dispatch(dp->dp_sync_taskq,
3724 dsl_scan_prefetch_thread, scn, TQ_SLEEP);
3725 ASSERT(prefetch_tqid != TASKQID_INVALID);
3727 dsl_pool_config_enter(dp, FTAG);
3728 dsl_scan_visit(scn, tx);
3729 dsl_pool_config_exit(dp, FTAG);
3731 mutex_enter(&dp->dp_spa->spa_scrub_lock);
3732 scn->scn_prefetch_stop = B_TRUE;
3733 cv_broadcast(&spa->spa_scrub_io_cv);
3734 mutex_exit(&dp->dp_spa->spa_scrub_lock);
3736 taskq_wait_id(dp->dp_sync_taskq, prefetch_tqid);
3737 (void) zio_wait(scn->scn_zio_root);
3738 scn->scn_zio_root = NULL;
3740 zfs_dbgmsg("scan visited %llu blocks of %s in %llums "
3741 "(%llu os's, %llu holes, %llu < mintxg, "
3742 "%llu in ddt, %llu > maxtxg)",
3743 (longlong_t)scn->scn_visited_this_txg,
3745 (longlong_t)NSEC2MSEC(gethrtime() -
3746 scn->scn_sync_start_time),
3747 (longlong_t)scn->scn_objsets_visited_this_txg,
3748 (longlong_t)scn->scn_holes_this_txg,
3749 (longlong_t)scn->scn_lt_min_this_txg,
3750 (longlong_t)scn->scn_ddt_contained_this_txg,
3751 (longlong_t)scn->scn_gt_max_this_txg);
3753 if (!scn->scn_suspending) {
3754 ASSERT0(avl_numnodes(&scn->scn_queue));
3755 scn->scn_done_txg = tx->tx_txg + 1;
3756 if (scn->scn_is_sorted) {
3757 scn->scn_checkpointing = B_TRUE;
3758 scn->scn_clearing = B_TRUE;
3760 zfs_dbgmsg("scan complete for %s txg %llu",
3762 (longlong_t)tx->tx_txg);
3764 } else if (scn->scn_is_sorted && scn->scn_bytes_pending != 0) {
3765 ASSERT(scn->scn_clearing);
3767 /* need to issue scrubbing IOs from per-vdev queues */
3768 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3769 NULL, ZIO_FLAG_CANFAIL);
3770 scan_io_queues_run(scn);
3771 (void) zio_wait(scn->scn_zio_root);
3772 scn->scn_zio_root = NULL;
3774 /* calculate and dprintf the current memory usage */
3775 (void) dsl_scan_should_clear(scn);
3776 dsl_scan_update_stats(scn);
3778 zfs_dbgmsg("scan issued %llu blocks for %s (%llu segs) "
3779 "in %llums (avg_block_size = %llu, avg_seg_size = %llu)",
3780 (longlong_t)scn->scn_zios_this_txg,
3782 (longlong_t)scn->scn_segs_this_txg,
3783 (longlong_t)NSEC2MSEC(gethrtime() -
3784 scn->scn_sync_start_time),
3785 (longlong_t)scn->scn_avg_zio_size_this_txg,
3786 (longlong_t)scn->scn_avg_seg_size_this_txg);
3787 } else if (scn->scn_done_txg != 0 && scn->scn_done_txg <= tx->tx_txg) {
3788 /* Finished with everything. Mark the scrub as complete */
3789 zfs_dbgmsg("scan issuing complete txg %llu for %s",
3790 (longlong_t)tx->tx_txg,
3792 ASSERT3U(scn->scn_done_txg, !=, 0);
3793 ASSERT0(spa->spa_scrub_inflight);
3794 ASSERT0(scn->scn_bytes_pending);
3795 dsl_scan_done(scn, B_TRUE, tx);
3796 sync_type = SYNC_MANDATORY;
3799 dsl_scan_sync_state(scn, tx, sync_type);
3803 count_block(dsl_scan_t *scn, zfs_all_blkstats_t *zab, const blkptr_t *bp)
3808 * Don't count embedded bp's, since we already did the work of
3809 * scanning these when we scanned the containing block.
3811 if (BP_IS_EMBEDDED(bp))
3815 * Update the spa's stats on how many bytes we have issued.
3816 * Sequential scrubs create a zio for each DVA of the bp. Each
3817 * of these will include all DVAs for repair purposes, but the
3818 * zio code will only try the first one unless there is an issue.
3819 * Therefore, we should only count the first DVA for these IOs.
3821 if (scn->scn_is_sorted) {
3822 atomic_add_64(&scn->scn_dp->dp_spa->spa_scan_pass_issued,
3823 DVA_GET_ASIZE(&bp->blk_dva[0]));
3825 spa_t *spa = scn->scn_dp->dp_spa;
3827 for (i = 0; i < BP_GET_NDVAS(bp); i++) {
3828 atomic_add_64(&spa->spa_scan_pass_issued,
3829 DVA_GET_ASIZE(&bp->blk_dva[i]));
3834 * If we resume after a reboot, zab will be NULL; don't record
3835 * incomplete stats in that case.
3840 mutex_enter(&zab->zab_lock);
3842 for (i = 0; i < 4; i++) {
3843 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
3844 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
3846 if (t & DMU_OT_NEWTYPE)
3848 zfs_blkstat_t *zb = &zab->zab_type[l][t];
3852 zb->zb_asize += BP_GET_ASIZE(bp);
3853 zb->zb_lsize += BP_GET_LSIZE(bp);
3854 zb->zb_psize += BP_GET_PSIZE(bp);
3855 zb->zb_gangs += BP_COUNT_GANG(bp);
3857 switch (BP_GET_NDVAS(bp)) {
3859 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3860 DVA_GET_VDEV(&bp->blk_dva[1]))
3861 zb->zb_ditto_2_of_2_samevdev++;
3864 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3865 DVA_GET_VDEV(&bp->blk_dva[1])) +
3866 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3867 DVA_GET_VDEV(&bp->blk_dva[2])) +
3868 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
3869 DVA_GET_VDEV(&bp->blk_dva[2]));
3871 zb->zb_ditto_2_of_3_samevdev++;
3872 else if (equal == 3)
3873 zb->zb_ditto_3_of_3_samevdev++;
3878 mutex_exit(&zab->zab_lock);
3882 scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue, scan_io_t *sio)
3885 int64_t asize = SIO_GET_ASIZE(sio);
3886 dsl_scan_t *scn = queue->q_scn;
3888 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3890 if (avl_find(&queue->q_sios_by_addr, sio, &idx) != NULL) {
3891 /* block is already scheduled for reading */
3892 atomic_add_64(&scn->scn_bytes_pending, -asize);
3896 avl_insert(&queue->q_sios_by_addr, sio, idx);
3897 queue->q_sio_memused += SIO_GET_MUSED(sio);
3898 range_tree_add(queue->q_exts_by_addr, SIO_GET_OFFSET(sio), asize);
3902 * Given all the info we got from our metadata scanning process, we
3903 * construct a scan_io_t and insert it into the scan sorting queue. The
3904 * I/O must already be suitable for us to process. This is controlled
3905 * by dsl_scan_enqueue().
3908 scan_io_queue_insert(dsl_scan_io_queue_t *queue, const blkptr_t *bp, int dva_i,
3909 int zio_flags, const zbookmark_phys_t *zb)
3911 dsl_scan_t *scn = queue->q_scn;
3912 scan_io_t *sio = sio_alloc(BP_GET_NDVAS(bp));
3914 ASSERT0(BP_IS_GANG(bp));
3915 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3917 bp2sio(bp, sio, dva_i);
3918 sio->sio_flags = zio_flags;
3922 * Increment the bytes pending counter now so that we can't
3923 * get an integer underflow in case the worker processes the
3924 * zio before we get to incrementing this counter.
3926 atomic_add_64(&scn->scn_bytes_pending, SIO_GET_ASIZE(sio));
3928 scan_io_queue_insert_impl(queue, sio);
3932 * Given a set of I/O parameters as discovered by the metadata traversal
3933 * process, attempts to place the I/O into the sorted queues (if allowed),
3934 * or immediately executes the I/O.
3937 dsl_scan_enqueue(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
3938 const zbookmark_phys_t *zb)
3940 spa_t *spa = dp->dp_spa;
3942 ASSERT(!BP_IS_EMBEDDED(bp));
3945 * Gang blocks are hard to issue sequentially, so we just issue them
3946 * here immediately instead of queuing them.
3948 if (!dp->dp_scan->scn_is_sorted || BP_IS_GANG(bp)) {
3949 scan_exec_io(dp, bp, zio_flags, zb, NULL);
3953 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
3957 dva = bp->blk_dva[i];
3958 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&dva));
3959 ASSERT(vdev != NULL);
3961 mutex_enter(&vdev->vdev_scan_io_queue_lock);
3962 if (vdev->vdev_scan_io_queue == NULL)
3963 vdev->vdev_scan_io_queue = scan_io_queue_create(vdev);
3964 ASSERT(dp->dp_scan != NULL);
3965 scan_io_queue_insert(vdev->vdev_scan_io_queue, bp,
3967 mutex_exit(&vdev->vdev_scan_io_queue_lock);
3972 dsl_scan_scrub_cb(dsl_pool_t *dp,
3973 const blkptr_t *bp, const zbookmark_phys_t *zb)
3975 dsl_scan_t *scn = dp->dp_scan;
3976 spa_t *spa = dp->dp_spa;
3977 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
3978 size_t psize = BP_GET_PSIZE(bp);
3979 boolean_t needs_io = B_FALSE;
3980 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
3983 if (phys_birth <= scn->scn_phys.scn_min_txg ||
3984 phys_birth >= scn->scn_phys.scn_max_txg) {
3985 count_block(scn, dp->dp_blkstats, bp);
3989 /* Embedded BP's have phys_birth==0, so we reject them above. */
3990 ASSERT(!BP_IS_EMBEDDED(bp));
3992 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
3993 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
3994 zio_flags |= ZIO_FLAG_SCRUB;
3997 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
3998 zio_flags |= ZIO_FLAG_RESILVER;
4002 /* If it's an intent log block, failure is expected. */
4003 if (zb->zb_level == ZB_ZIL_LEVEL)
4004 zio_flags |= ZIO_FLAG_SPECULATIVE;
4006 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
4007 const dva_t *dva = &bp->blk_dva[d];
4010 * Keep track of how much data we've examined so that
4011 * zpool(8) status can make useful progress reports.
4013 scn->scn_phys.scn_examined += DVA_GET_ASIZE(dva);
4014 spa->spa_scan_pass_exam += DVA_GET_ASIZE(dva);
4016 /* if it's a resilver, this may not be in the target range */
4018 needs_io = dsl_scan_need_resilver(spa, dva, psize,
4022 if (needs_io && !zfs_no_scrub_io) {
4023 dsl_scan_enqueue(dp, bp, zio_flags, zb);
4025 count_block(scn, dp->dp_blkstats, bp);
4028 /* do not relocate this block */
4033 dsl_scan_scrub_done(zio_t *zio)
4035 spa_t *spa = zio->io_spa;
4036 blkptr_t *bp = zio->io_bp;
4037 dsl_scan_io_queue_t *queue = zio->io_private;
4039 abd_free(zio->io_abd);
4041 if (queue == NULL) {
4042 mutex_enter(&spa->spa_scrub_lock);
4043 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
4044 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
4045 cv_broadcast(&spa->spa_scrub_io_cv);
4046 mutex_exit(&spa->spa_scrub_lock);
4048 mutex_enter(&queue->q_vd->vdev_scan_io_queue_lock);
4049 ASSERT3U(queue->q_inflight_bytes, >=, BP_GET_PSIZE(bp));
4050 queue->q_inflight_bytes -= BP_GET_PSIZE(bp);
4051 cv_broadcast(&queue->q_zio_cv);
4052 mutex_exit(&queue->q_vd->vdev_scan_io_queue_lock);
4055 if (zio->io_error && (zio->io_error != ECKSUM ||
4056 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
4057 atomic_inc_64(&spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors);
4062 * Given a scanning zio's information, executes the zio. The zio need
4063 * not necessarily be only sortable, this function simply executes the
4064 * zio, no matter what it is. The optional queue argument allows the
4065 * caller to specify that they want per top level vdev IO rate limiting
4066 * instead of the legacy global limiting.
4069 scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
4070 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue)
4072 spa_t *spa = dp->dp_spa;
4073 dsl_scan_t *scn = dp->dp_scan;
4074 size_t size = BP_GET_PSIZE(bp);
4075 abd_t *data = abd_alloc_for_io(size, B_FALSE);
4077 if (queue == NULL) {
4078 ASSERT3U(scn->scn_maxinflight_bytes, >, 0);
4079 mutex_enter(&spa->spa_scrub_lock);
4080 while (spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)
4081 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
4082 spa->spa_scrub_inflight += BP_GET_PSIZE(bp);
4083 mutex_exit(&spa->spa_scrub_lock);
4085 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
4087 ASSERT3U(queue->q_maxinflight_bytes, >, 0);
4088 mutex_enter(q_lock);
4089 while (queue->q_inflight_bytes >= queue->q_maxinflight_bytes)
4090 cv_wait(&queue->q_zio_cv, q_lock);
4091 queue->q_inflight_bytes += BP_GET_PSIZE(bp);
4095 count_block(scn, dp->dp_blkstats, bp);
4096 zio_nowait(zio_read(scn->scn_zio_root, spa, bp, data, size,
4097 dsl_scan_scrub_done, queue, ZIO_PRIORITY_SCRUB, zio_flags, zb));
4101 * This is the primary extent sorting algorithm. We balance two parameters:
4102 * 1) how many bytes of I/O are in an extent
4103 * 2) how well the extent is filled with I/O (as a fraction of its total size)
4104 * Since we allow extents to have gaps between their constituent I/Os, it's
4105 * possible to have a fairly large extent that contains the same amount of
4106 * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
4107 * The algorithm sorts based on a score calculated from the extent's size,
4108 * the relative fill volume (in %) and a "fill weight" parameter that controls
4109 * the split between whether we prefer larger extents or more well populated
4112 * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
4115 * 1) assume extsz = 64 MiB
4116 * 2) assume fill = 32 MiB (extent is half full)
4117 * 3) assume fill_weight = 3
4118 * 4) SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
4119 * SCORE = 32M + (50 * 3 * 32M) / 100
4120 * SCORE = 32M + (4800M / 100)
4123 * | +--- final total relative fill-based score
4124 * +--------- final total fill-based score
4127 * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
4128 * extents that are more completely filled (in a 3:2 ratio) vs just larger.
4129 * Note that as an optimization, we replace multiplication and division by
4130 * 100 with bitshifting by 7 (which effectively multiplies and divides by 128).
4133 ext_size_compare(const void *x, const void *y)
4135 const range_seg_gap_t *rsa = x, *rsb = y;
4137 uint64_t sa = rsa->rs_end - rsa->rs_start;
4138 uint64_t sb = rsb->rs_end - rsb->rs_start;
4139 uint64_t score_a, score_b;
4141 score_a = rsa->rs_fill + ((((rsa->rs_fill << 7) / sa) *
4142 fill_weight * rsa->rs_fill) >> 7);
4143 score_b = rsb->rs_fill + ((((rsb->rs_fill << 7) / sb) *
4144 fill_weight * rsb->rs_fill) >> 7);
4146 if (score_a > score_b)
4148 if (score_a == score_b) {
4149 if (rsa->rs_start < rsb->rs_start)
4151 if (rsa->rs_start == rsb->rs_start)
4159 * Comparator for the q_sios_by_addr tree. Sorting is simply performed
4160 * based on LBA-order (from lowest to highest).
4163 sio_addr_compare(const void *x, const void *y)
4165 const scan_io_t *a = x, *b = y;
4167 return (TREE_CMP(SIO_GET_OFFSET(a), SIO_GET_OFFSET(b)));
4170 /* IO queues are created on demand when they are needed. */
4171 static dsl_scan_io_queue_t *
4172 scan_io_queue_create(vdev_t *vd)
4174 dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan;
4175 dsl_scan_io_queue_t *q = kmem_zalloc(sizeof (*q), KM_SLEEP);
4179 q->q_sio_memused = 0;
4180 cv_init(&q->q_zio_cv, NULL, CV_DEFAULT, NULL);
4181 q->q_exts_by_addr = range_tree_create_impl(&rt_btree_ops, RANGE_SEG_GAP,
4182 &q->q_exts_by_size, 0, 0, ext_size_compare, zfs_scan_max_ext_gap);
4183 avl_create(&q->q_sios_by_addr, sio_addr_compare,
4184 sizeof (scan_io_t), offsetof(scan_io_t, sio_nodes.sio_addr_node));
4190 * Destroys a scan queue and all segments and scan_io_t's contained in it.
4191 * No further execution of I/O occurs, anything pending in the queue is
4192 * simply freed without being executed.
4195 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t *queue)
4197 dsl_scan_t *scn = queue->q_scn;
4199 void *cookie = NULL;
4200 int64_t bytes_dequeued = 0;
4202 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
4204 while ((sio = avl_destroy_nodes(&queue->q_sios_by_addr, &cookie)) !=
4206 ASSERT(range_tree_contains(queue->q_exts_by_addr,
4207 SIO_GET_OFFSET(sio), SIO_GET_ASIZE(sio)));
4208 bytes_dequeued += SIO_GET_ASIZE(sio);
4209 queue->q_sio_memused -= SIO_GET_MUSED(sio);
4213 ASSERT0(queue->q_sio_memused);
4214 atomic_add_64(&scn->scn_bytes_pending, -bytes_dequeued);
4215 range_tree_vacate(queue->q_exts_by_addr, NULL, queue);
4216 range_tree_destroy(queue->q_exts_by_addr);
4217 avl_destroy(&queue->q_sios_by_addr);
4218 cv_destroy(&queue->q_zio_cv);
4220 kmem_free(queue, sizeof (*queue));
4224 * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
4225 * called on behalf of vdev_top_transfer when creating or destroying
4226 * a mirror vdev due to zpool attach/detach.
4229 dsl_scan_io_queue_vdev_xfer(vdev_t *svd, vdev_t *tvd)
4231 mutex_enter(&svd->vdev_scan_io_queue_lock);
4232 mutex_enter(&tvd->vdev_scan_io_queue_lock);
4234 VERIFY3P(tvd->vdev_scan_io_queue, ==, NULL);
4235 tvd->vdev_scan_io_queue = svd->vdev_scan_io_queue;
4236 svd->vdev_scan_io_queue = NULL;
4237 if (tvd->vdev_scan_io_queue != NULL)
4238 tvd->vdev_scan_io_queue->q_vd = tvd;
4240 mutex_exit(&tvd->vdev_scan_io_queue_lock);
4241 mutex_exit(&svd->vdev_scan_io_queue_lock);
4245 scan_io_queues_destroy(dsl_scan_t *scn)
4247 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
4249 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
4250 vdev_t *tvd = rvd->vdev_child[i];
4252 mutex_enter(&tvd->vdev_scan_io_queue_lock);
4253 if (tvd->vdev_scan_io_queue != NULL)
4254 dsl_scan_io_queue_destroy(tvd->vdev_scan_io_queue);
4255 tvd->vdev_scan_io_queue = NULL;
4256 mutex_exit(&tvd->vdev_scan_io_queue_lock);
4261 dsl_scan_freed_dva(spa_t *spa, const blkptr_t *bp, int dva_i)
4263 dsl_pool_t *dp = spa->spa_dsl_pool;
4264 dsl_scan_t *scn = dp->dp_scan;
4267 dsl_scan_io_queue_t *queue;
4268 scan_io_t *srch_sio, *sio;
4270 uint64_t start, size;
4272 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&bp->blk_dva[dva_i]));
4273 ASSERT(vdev != NULL);
4274 q_lock = &vdev->vdev_scan_io_queue_lock;
4275 queue = vdev->vdev_scan_io_queue;
4277 mutex_enter(q_lock);
4278 if (queue == NULL) {
4283 srch_sio = sio_alloc(BP_GET_NDVAS(bp));
4284 bp2sio(bp, srch_sio, dva_i);
4285 start = SIO_GET_OFFSET(srch_sio);
4286 size = SIO_GET_ASIZE(srch_sio);
4289 * We can find the zio in two states:
4290 * 1) Cold, just sitting in the queue of zio's to be issued at
4291 * some point in the future. In this case, all we do is
4292 * remove the zio from the q_sios_by_addr tree, decrement
4293 * its data volume from the containing range_seg_t and
4294 * resort the q_exts_by_size tree to reflect that the
4295 * range_seg_t has lost some of its 'fill'. We don't shorten
4296 * the range_seg_t - this is usually rare enough not to be
4297 * worth the extra hassle of trying keep track of precise
4298 * extent boundaries.
4299 * 2) Hot, where the zio is currently in-flight in
4300 * dsl_scan_issue_ios. In this case, we can't simply
4301 * reach in and stop the in-flight zio's, so we instead
4302 * block the caller. Eventually, dsl_scan_issue_ios will
4303 * be done with issuing the zio's it gathered and will
4306 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
4310 int64_t asize = SIO_GET_ASIZE(sio);
4313 /* Got it while it was cold in the queue */
4314 ASSERT3U(start, ==, SIO_GET_OFFSET(sio));
4315 ASSERT3U(size, ==, asize);
4316 avl_remove(&queue->q_sios_by_addr, sio);
4317 queue->q_sio_memused -= SIO_GET_MUSED(sio);
4319 ASSERT(range_tree_contains(queue->q_exts_by_addr, start, size));
4320 range_tree_remove_fill(queue->q_exts_by_addr, start, size);
4323 * We only update scn_bytes_pending in the cold path,
4324 * otherwise it will already have been accounted for as
4325 * part of the zio's execution.
4327 atomic_add_64(&scn->scn_bytes_pending, -asize);
4329 /* count the block as though we issued it */
4330 sio2bp(sio, &tmpbp);
4331 count_block(scn, dp->dp_blkstats, &tmpbp);
4339 * Callback invoked when a zio_free() zio is executing. This needs to be
4340 * intercepted to prevent the zio from deallocating a particular portion
4341 * of disk space and it then getting reallocated and written to, while we
4342 * still have it queued up for processing.
4345 dsl_scan_freed(spa_t *spa, const blkptr_t *bp)
4347 dsl_pool_t *dp = spa->spa_dsl_pool;
4348 dsl_scan_t *scn = dp->dp_scan;
4350 ASSERT(!BP_IS_EMBEDDED(bp));
4351 ASSERT(scn != NULL);
4352 if (!dsl_scan_is_running(scn))
4355 for (int i = 0; i < BP_GET_NDVAS(bp); i++)
4356 dsl_scan_freed_dva(spa, bp, i);
4360 * Check if a vdev needs resilvering (non-empty DTL), if so, and resilver has
4361 * not started, start it. Otherwise, only restart if max txg in DTL range is
4362 * greater than the max txg in the current scan. If the DTL max is less than
4363 * the scan max, then the vdev has not missed any new data since the resilver
4364 * started, so a restart is not needed.
4367 dsl_scan_assess_vdev(dsl_pool_t *dp, vdev_t *vd)
4371 if (!vdev_resilver_needed(vd, &min, &max))
4374 if (!dsl_scan_resilvering(dp)) {
4375 spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER);
4379 if (max <= dp->dp_scan->scn_phys.scn_max_txg)
4382 /* restart is needed, check if it can be deferred */
4383 if (spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
4384 vdev_defer_resilver(vd);
4386 spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER);
4389 ZFS_MODULE_PARAM(zfs, zfs_, scan_vdev_limit, ULONG, ZMOD_RW,
4390 "Max bytes in flight per leaf vdev for scrubs and resilvers");
4392 ZFS_MODULE_PARAM(zfs, zfs_, scrub_min_time_ms, INT, ZMOD_RW,
4393 "Min millisecs to scrub per txg");
4395 ZFS_MODULE_PARAM(zfs, zfs_, obsolete_min_time_ms, INT, ZMOD_RW,
4396 "Min millisecs to obsolete per txg");
4398 ZFS_MODULE_PARAM(zfs, zfs_, free_min_time_ms, INT, ZMOD_RW,
4399 "Min millisecs to free per txg");
4401 ZFS_MODULE_PARAM(zfs, zfs_, resilver_min_time_ms, INT, ZMOD_RW,
4402 "Min millisecs to resilver per txg");
4404 ZFS_MODULE_PARAM(zfs, zfs_, scan_suspend_progress, INT, ZMOD_RW,
4405 "Set to prevent scans from progressing");
4407 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_io, INT, ZMOD_RW,
4408 "Set to disable scrub I/O");
4410 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_prefetch, INT, ZMOD_RW,
4411 "Set to disable scrub prefetching");
4413 ZFS_MODULE_PARAM(zfs, zfs_, async_block_max_blocks, ULONG, ZMOD_RW,
4414 "Max number of blocks freed in one txg");
4416 ZFS_MODULE_PARAM(zfs, zfs_, max_async_dedup_frees, ULONG, ZMOD_RW,
4417 "Max number of dedup blocks freed in one txg");
4419 ZFS_MODULE_PARAM(zfs, zfs_, free_bpobj_enabled, INT, ZMOD_RW,
4420 "Enable processing of the free_bpobj");
4422 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_fact, INT, ZMOD_RW,
4423 "Fraction of RAM for scan hard limit");
4425 ZFS_MODULE_PARAM(zfs, zfs_, scan_issue_strategy, INT, ZMOD_RW,
4426 "IO issuing strategy during scrubbing. 0 = default, 1 = LBA, 2 = size");
4428 ZFS_MODULE_PARAM(zfs, zfs_, scan_legacy, INT, ZMOD_RW,
4429 "Scrub using legacy non-sequential method");
4431 ZFS_MODULE_PARAM(zfs, zfs_, scan_checkpoint_intval, INT, ZMOD_RW,
4432 "Scan progress on-disk checkpointing interval");
4434 ZFS_MODULE_PARAM(zfs, zfs_, scan_max_ext_gap, ULONG, ZMOD_RW,
4435 "Max gap in bytes between sequential scrub / resilver I/Os");
4437 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_soft_fact, INT, ZMOD_RW,
4438 "Fraction of hard limit used as soft limit");
4440 ZFS_MODULE_PARAM(zfs, zfs_, scan_strict_mem_lim, INT, ZMOD_RW,
4441 "Tunable to attempt to reduce lock contention");
4443 ZFS_MODULE_PARAM(zfs, zfs_, scan_fill_weight, INT, ZMOD_RW,
4444 "Tunable to adjust bias towards more filled segments during scans");
4446 ZFS_MODULE_PARAM(zfs, zfs_, resilver_disable_defer, INT, ZMOD_RW,
4447 "Process all resilvers immediately");