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 https://opensource.org/licenses/CDDL-1.0.
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 uint_t zfs_vdev_async_write_active_min_dirty_percent;
132 static int zfs_scan_blkstats = 0;
135 * By default zfs will check to ensure it is not over the hard memory
136 * limit before each txg. If finer-grained control of this is needed
137 * this value can be set to 1 to enable checking before scanning each
140 static int zfs_scan_strict_mem_lim = B_FALSE;
143 * Maximum number of parallelly executed bytes per leaf vdev. We attempt
144 * to strike a balance here between keeping the vdev queues full of I/Os
145 * at all times and not overflowing the queues to cause long latency,
146 * which would cause long txg sync times. No matter what, we will not
147 * overload the drives with I/O, since that is protected by
148 * zfs_vdev_scrub_max_active.
150 static uint64_t zfs_scan_vdev_limit = 4 << 20;
152 static uint_t zfs_scan_issue_strategy = 0;
154 /* don't queue & sort zios, go direct */
155 static int zfs_scan_legacy = B_FALSE;
156 static uint64_t zfs_scan_max_ext_gap = 2 << 20; /* in bytes */
159 * fill_weight is non-tunable at runtime, so we copy it at module init from
160 * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
161 * break queue sorting.
163 static uint_t zfs_scan_fill_weight = 3;
164 static uint64_t fill_weight;
166 /* See dsl_scan_should_clear() for details on the memory limit tunables */
167 static const uint64_t zfs_scan_mem_lim_min = 16 << 20; /* bytes */
168 static const uint64_t zfs_scan_mem_lim_soft_max = 128 << 20; /* bytes */
171 /* fraction of physmem */
172 static uint_t zfs_scan_mem_lim_fact = 20;
174 /* fraction of mem lim above */
175 static uint_t zfs_scan_mem_lim_soft_fact = 20;
177 /* minimum milliseconds to scrub per txg */
178 static uint_t zfs_scrub_min_time_ms = 1000;
180 /* minimum milliseconds to obsolete per txg */
181 static uint_t zfs_obsolete_min_time_ms = 500;
183 /* minimum milliseconds to free per txg */
184 static uint_t zfs_free_min_time_ms = 1000;
186 /* minimum milliseconds to resilver per txg */
187 static uint_t zfs_resilver_min_time_ms = 3000;
189 static uint_t zfs_scan_checkpoint_intval = 7200; /* in seconds */
190 int zfs_scan_suspend_progress = 0; /* set to prevent scans from progressing */
191 static int zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
192 static int zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
193 static const enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
194 /* max number of blocks to free in a single TXG */
195 static uint64_t zfs_async_block_max_blocks = UINT64_MAX;
196 /* max number of dedup blocks to free in a single TXG */
197 static uint64_t zfs_max_async_dedup_frees = 100000;
199 /* set to disable resilver deferring */
200 static int zfs_resilver_disable_defer = B_FALSE;
203 * We wait a few txgs after importing a pool to begin scanning so that
204 * the import / mounting code isn't held up by scrub / resilver IO.
205 * Unfortunately, it is a bit difficult to determine exactly how long
206 * this will take since userspace will trigger fs mounts asynchronously
207 * and the kernel will create zvol minors asynchronously. As a result,
208 * the value provided here is a bit arbitrary, but represents a
209 * reasonable estimate of how many txgs it will take to finish fully
212 #define SCAN_IMPORT_WAIT_TXGS 5
214 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
215 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
216 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
219 * Enable/disable the processing of the free_bpobj object.
221 static int zfs_free_bpobj_enabled = 1;
223 /* the order has to match pool_scan_type */
224 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
226 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
227 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
230 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */
238 * This controls what conditions are placed on dsl_scan_sync_state():
239 * SYNC_OPTIONAL) write out scn_phys iff scn_queues_pending == 0
240 * SYNC_MANDATORY) write out scn_phys always. scn_queues_pending must be 0.
241 * SYNC_CACHED) if scn_queues_pending == 0, write out scn_phys. Otherwise
242 * write out the scn_phys_cached version.
243 * See dsl_scan_sync_state for details.
252 * This struct represents the minimum information needed to reconstruct a
253 * zio for sequential scanning. This is useful because many of these will
254 * accumulate in the sequential IO queues before being issued, so saving
255 * memory matters here.
257 typedef struct scan_io {
258 /* fields from blkptr_t */
259 uint64_t sio_blk_prop;
260 uint64_t sio_phys_birth;
262 zio_cksum_t sio_cksum;
263 uint32_t sio_nr_dvas;
265 /* fields from zio_t */
267 zbookmark_phys_t sio_zb;
269 /* members for queue sorting */
271 avl_node_t sio_addr_node; /* link into issuing queue */
272 list_node_t sio_list_node; /* link for issuing to disk */
276 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp,
277 * depending on how many were in the original bp. Only the
278 * first DVA is really used for sorting and issuing purposes.
279 * The other DVAs (if provided) simply exist so that the zio
280 * layer can find additional copies to repair from in the
281 * event of an error. This array must go at the end of the
282 * struct to allow this for the variable number of elements.
287 #define SIO_SET_OFFSET(sio, x) DVA_SET_OFFSET(&(sio)->sio_dva[0], x)
288 #define SIO_SET_ASIZE(sio, x) DVA_SET_ASIZE(&(sio)->sio_dva[0], x)
289 #define SIO_GET_OFFSET(sio) DVA_GET_OFFSET(&(sio)->sio_dva[0])
290 #define SIO_GET_ASIZE(sio) DVA_GET_ASIZE(&(sio)->sio_dva[0])
291 #define SIO_GET_END_OFFSET(sio) \
292 (SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio))
293 #define SIO_GET_MUSED(sio) \
294 (sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t)))
296 struct dsl_scan_io_queue {
297 dsl_scan_t *q_scn; /* associated dsl_scan_t */
298 vdev_t *q_vd; /* top-level vdev that this queue represents */
299 zio_t *q_zio; /* scn_zio_root child for waiting on IO */
301 /* trees used for sorting I/Os and extents of I/Os */
302 range_tree_t *q_exts_by_addr;
303 zfs_btree_t q_exts_by_size;
304 avl_tree_t q_sios_by_addr;
305 uint64_t q_sio_memused;
306 uint64_t q_last_ext_addr;
308 /* members for zio rate limiting */
309 uint64_t q_maxinflight_bytes;
310 uint64_t q_inflight_bytes;
311 kcondvar_t q_zio_cv; /* used under vd->vdev_scan_io_queue_lock */
313 /* per txg statistics */
314 uint64_t q_total_seg_size_this_txg;
315 uint64_t q_segs_this_txg;
316 uint64_t q_total_zio_size_this_txg;
317 uint64_t q_zios_this_txg;
320 /* private data for dsl_scan_prefetch_cb() */
321 typedef struct scan_prefetch_ctx {
322 zfs_refcount_t spc_refcnt; /* refcount for memory management */
323 dsl_scan_t *spc_scn; /* dsl_scan_t for the pool */
324 boolean_t spc_root; /* is this prefetch for an objset? */
325 uint8_t spc_indblkshift; /* dn_indblkshift of current dnode */
326 uint16_t spc_datablkszsec; /* dn_idatablkszsec of current dnode */
327 } scan_prefetch_ctx_t;
329 /* private data for dsl_scan_prefetch() */
330 typedef struct scan_prefetch_issue_ctx {
331 avl_node_t spic_avl_node; /* link into scn->scn_prefetch_queue */
332 scan_prefetch_ctx_t *spic_spc; /* spc for the callback */
333 blkptr_t spic_bp; /* bp to prefetch */
334 zbookmark_phys_t spic_zb; /* bookmark to prefetch */
335 } scan_prefetch_issue_ctx_t;
337 static void scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
338 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue);
339 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue,
342 static dsl_scan_io_queue_t *scan_io_queue_create(vdev_t *vd);
343 static void scan_io_queues_destroy(dsl_scan_t *scn);
345 static kmem_cache_t *sio_cache[SPA_DVAS_PER_BP];
347 /* sio->sio_nr_dvas must be set so we know which cache to free from */
349 sio_free(scan_io_t *sio)
351 ASSERT3U(sio->sio_nr_dvas, >, 0);
352 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
354 kmem_cache_free(sio_cache[sio->sio_nr_dvas - 1], sio);
357 /* It is up to the caller to set sio->sio_nr_dvas for freeing */
359 sio_alloc(unsigned short nr_dvas)
361 ASSERT3U(nr_dvas, >, 0);
362 ASSERT3U(nr_dvas, <=, SPA_DVAS_PER_BP);
364 return (kmem_cache_alloc(sio_cache[nr_dvas - 1], KM_SLEEP));
371 * This is used in ext_size_compare() to weight segments
372 * based on how sparse they are. This cannot be changed
373 * mid-scan and the tree comparison functions don't currently
374 * have a mechanism for passing additional context to the
375 * compare functions. Thus we store this value globally and
376 * we only allow it to be set at module initialization time
378 fill_weight = zfs_scan_fill_weight;
380 for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
383 (void) snprintf(name, sizeof (name), "sio_cache_%d", i);
384 sio_cache[i] = kmem_cache_create(name,
385 (sizeof (scan_io_t) + ((i + 1) * sizeof (dva_t))),
386 0, NULL, NULL, NULL, NULL, NULL, 0);
393 for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
394 kmem_cache_destroy(sio_cache[i]);
398 static inline boolean_t
399 dsl_scan_is_running(const dsl_scan_t *scn)
401 return (scn->scn_phys.scn_state == DSS_SCANNING);
405 dsl_scan_resilvering(dsl_pool_t *dp)
407 return (dsl_scan_is_running(dp->dp_scan) &&
408 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
412 sio2bp(const scan_io_t *sio, blkptr_t *bp)
414 memset(bp, 0, sizeof (*bp));
415 bp->blk_prop = sio->sio_blk_prop;
416 bp->blk_phys_birth = sio->sio_phys_birth;
417 bp->blk_birth = sio->sio_birth;
418 bp->blk_fill = 1; /* we always only work with data pointers */
419 bp->blk_cksum = sio->sio_cksum;
421 ASSERT3U(sio->sio_nr_dvas, >, 0);
422 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
424 memcpy(bp->blk_dva, sio->sio_dva, sio->sio_nr_dvas * sizeof (dva_t));
428 bp2sio(const blkptr_t *bp, scan_io_t *sio, int dva_i)
430 sio->sio_blk_prop = bp->blk_prop;
431 sio->sio_phys_birth = bp->blk_phys_birth;
432 sio->sio_birth = bp->blk_birth;
433 sio->sio_cksum = bp->blk_cksum;
434 sio->sio_nr_dvas = BP_GET_NDVAS(bp);
437 * Copy the DVAs to the sio. We need all copies of the block so
438 * that the self healing code can use the alternate copies if the
439 * first is corrupted. We want the DVA at index dva_i to be first
440 * in the sio since this is the primary one that we want to issue.
442 for (int i = 0, j = dva_i; i < sio->sio_nr_dvas; i++, j++) {
443 sio->sio_dva[i] = bp->blk_dva[j % sio->sio_nr_dvas];
448 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
452 spa_t *spa = dp->dp_spa;
455 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
459 * It's possible that we're resuming a scan after a reboot so
460 * make sure that the scan_async_destroying flag is initialized
463 ASSERT(!scn->scn_async_destroying);
464 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
465 SPA_FEATURE_ASYNC_DESTROY);
468 * Calculate the max number of in-flight bytes for pool-wide
469 * scanning operations (minimum 1MB). Limits for the issuing
470 * phase are done per top-level vdev and are handled separately.
472 scn->scn_maxinflight_bytes = MAX(zfs_scan_vdev_limit *
473 dsl_scan_count_data_disks(spa->spa_root_vdev), 1ULL << 20);
475 avl_create(&scn->scn_queue, scan_ds_queue_compare, sizeof (scan_ds_t),
476 offsetof(scan_ds_t, sds_node));
477 avl_create(&scn->scn_prefetch_queue, scan_prefetch_queue_compare,
478 sizeof (scan_prefetch_issue_ctx_t),
479 offsetof(scan_prefetch_issue_ctx_t, spic_avl_node));
481 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
482 "scrub_func", sizeof (uint64_t), 1, &f);
485 * There was an old-style scrub in progress. Restart a
486 * new-style scrub from the beginning.
488 scn->scn_restart_txg = txg;
489 zfs_dbgmsg("old-style scrub was in progress for %s; "
490 "restarting new-style scrub in txg %llu",
492 (longlong_t)scn->scn_restart_txg);
495 * Load the queue obj from the old location so that it
496 * can be freed by dsl_scan_done().
498 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
499 "scrub_queue", sizeof (uint64_t), 1,
500 &scn->scn_phys.scn_queue_obj);
502 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
503 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
506 * Detect if the pool contains the signature of #2094. If it
507 * does properly update the scn->scn_phys structure and notify
508 * the administrator by setting an errata for the pool.
510 if (err == EOVERFLOW) {
511 uint64_t zaptmp[SCAN_PHYS_NUMINTS + 1];
512 VERIFY3S(SCAN_PHYS_NUMINTS, ==, 24);
513 VERIFY3S(offsetof(dsl_scan_phys_t, scn_flags), ==,
514 (23 * sizeof (uint64_t)));
516 err = zap_lookup(dp->dp_meta_objset,
517 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SCAN,
518 sizeof (uint64_t), SCAN_PHYS_NUMINTS + 1, &zaptmp);
520 uint64_t overflow = zaptmp[SCAN_PHYS_NUMINTS];
522 if (overflow & ~DSL_SCAN_FLAGS_MASK ||
523 scn->scn_async_destroying) {
525 ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY;
529 memcpy(&scn->scn_phys, zaptmp,
530 SCAN_PHYS_NUMINTS * sizeof (uint64_t));
531 scn->scn_phys.scn_flags = overflow;
533 /* Required scrub already in progress. */
534 if (scn->scn_phys.scn_state == DSS_FINISHED ||
535 scn->scn_phys.scn_state == DSS_CANCELED)
537 ZPOOL_ERRATA_ZOL_2094_SCRUB;
547 * We might be restarting after a reboot, so jump the issued
548 * counter to how far we've scanned. We know we're consistent
551 scn->scn_issued_before_pass = scn->scn_phys.scn_examined;
553 if (dsl_scan_is_running(scn) &&
554 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
556 * A new-type scrub was in progress on an old
557 * pool, and the pool was accessed by old
558 * software. Restart from the beginning, since
559 * the old software may have changed the pool in
562 scn->scn_restart_txg = txg;
563 zfs_dbgmsg("new-style scrub for %s was modified "
564 "by old software; restarting in txg %llu",
566 (longlong_t)scn->scn_restart_txg);
567 } else if (dsl_scan_resilvering(dp)) {
569 * If a resilver is in progress and there are already
570 * errors, restart it instead of finishing this scan and
571 * then restarting it. If there haven't been any errors
572 * then remember that the incore DTL is valid.
574 if (scn->scn_phys.scn_errors > 0) {
575 scn->scn_restart_txg = txg;
576 zfs_dbgmsg("resilver can't excise DTL_MISSING "
577 "when finished; restarting on %s in txg "
580 (u_longlong_t)scn->scn_restart_txg);
582 /* it's safe to excise DTL when finished */
583 spa->spa_scrub_started = B_TRUE;
588 memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys));
590 /* reload the queue into the in-core state */
591 if (scn->scn_phys.scn_queue_obj != 0) {
595 for (zap_cursor_init(&zc, dp->dp_meta_objset,
596 scn->scn_phys.scn_queue_obj);
597 zap_cursor_retrieve(&zc, &za) == 0;
598 (void) zap_cursor_advance(&zc)) {
599 scan_ds_queue_insert(scn,
600 zfs_strtonum(za.za_name, NULL),
601 za.za_first_integer);
603 zap_cursor_fini(&zc);
606 spa_scan_stat_init(spa);
611 dsl_scan_fini(dsl_pool_t *dp)
613 if (dp->dp_scan != NULL) {
614 dsl_scan_t *scn = dp->dp_scan;
616 if (scn->scn_taskq != NULL)
617 taskq_destroy(scn->scn_taskq);
619 scan_ds_queue_clear(scn);
620 avl_destroy(&scn->scn_queue);
621 scan_ds_prefetch_queue_clear(scn);
622 avl_destroy(&scn->scn_prefetch_queue);
624 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
630 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
632 return (scn->scn_restart_txg != 0 &&
633 scn->scn_restart_txg <= tx->tx_txg);
637 dsl_scan_resilver_scheduled(dsl_pool_t *dp)
639 return ((dp->dp_scan && dp->dp_scan->scn_restart_txg != 0) ||
640 (spa_async_tasks(dp->dp_spa) & SPA_ASYNC_RESILVER));
644 dsl_scan_scrubbing(const dsl_pool_t *dp)
646 dsl_scan_phys_t *scn_phys = &dp->dp_scan->scn_phys;
648 return (scn_phys->scn_state == DSS_SCANNING &&
649 scn_phys->scn_func == POOL_SCAN_SCRUB);
653 dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
655 return (dsl_scan_scrubbing(scn->scn_dp) &&
656 scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED);
660 * Writes out a persistent dsl_scan_phys_t record to the pool directory.
661 * Because we can be running in the block sorting algorithm, we do not always
662 * want to write out the record, only when it is "safe" to do so. This safety
663 * condition is achieved by making sure that the sorting queues are empty
664 * (scn_queues_pending == 0). When this condition is not true, the sync'd state
665 * is inconsistent with how much actual scanning progress has been made. The
666 * kind of sync to be performed is specified by the sync_type argument. If the
667 * sync is optional, we only sync if the queues are empty. If the sync is
668 * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
669 * third possible state is a "cached" sync. This is done in response to:
670 * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
671 * destroyed, so we wouldn't be able to restart scanning from it.
672 * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
673 * superseded by a newer snapshot.
674 * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
675 * swapped with its clone.
676 * In all cases, a cached sync simply rewrites the last record we've written,
677 * just slightly modified. For the modifications that are performed to the
678 * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
679 * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
682 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx, state_sync_type_t sync_type)
685 spa_t *spa = scn->scn_dp->dp_spa;
687 ASSERT(sync_type != SYNC_MANDATORY || scn->scn_queues_pending == 0);
688 if (scn->scn_queues_pending == 0) {
689 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
690 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
691 dsl_scan_io_queue_t *q = vd->vdev_scan_io_queue;
696 mutex_enter(&vd->vdev_scan_io_queue_lock);
697 ASSERT3P(avl_first(&q->q_sios_by_addr), ==, NULL);
698 ASSERT3P(zfs_btree_first(&q->q_exts_by_size, NULL), ==,
700 ASSERT3P(range_tree_first(q->q_exts_by_addr), ==, NULL);
701 mutex_exit(&vd->vdev_scan_io_queue_lock);
704 if (scn->scn_phys.scn_queue_obj != 0)
705 scan_ds_queue_sync(scn, tx);
706 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
707 DMU_POOL_DIRECTORY_OBJECT,
708 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
709 &scn->scn_phys, tx));
710 memcpy(&scn->scn_phys_cached, &scn->scn_phys,
711 sizeof (scn->scn_phys));
713 if (scn->scn_checkpointing)
714 zfs_dbgmsg("finish scan checkpoint for %s",
717 scn->scn_checkpointing = B_FALSE;
718 scn->scn_last_checkpoint = ddi_get_lbolt();
719 } else if (sync_type == SYNC_CACHED) {
720 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
721 DMU_POOL_DIRECTORY_OBJECT,
722 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
723 &scn->scn_phys_cached, tx));
728 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
731 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
732 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
734 if (dsl_scan_is_running(scn) || vdev_rebuild_active(rvd))
735 return (SET_ERROR(EBUSY));
741 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
743 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
744 pool_scan_func_t *funcp = arg;
745 dmu_object_type_t ot = 0;
746 dsl_pool_t *dp = scn->scn_dp;
747 spa_t *spa = dp->dp_spa;
749 ASSERT(!dsl_scan_is_running(scn));
750 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
751 memset(&scn->scn_phys, 0, sizeof (scn->scn_phys));
752 scn->scn_phys.scn_func = *funcp;
753 scn->scn_phys.scn_state = DSS_SCANNING;
754 scn->scn_phys.scn_min_txg = 0;
755 scn->scn_phys.scn_max_txg = tx->tx_txg;
756 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
757 scn->scn_phys.scn_start_time = gethrestime_sec();
758 scn->scn_phys.scn_errors = 0;
759 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
760 scn->scn_issued_before_pass = 0;
761 scn->scn_restart_txg = 0;
762 scn->scn_done_txg = 0;
763 scn->scn_last_checkpoint = 0;
764 scn->scn_checkpointing = B_FALSE;
765 spa_scan_stat_init(spa);
767 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
768 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
770 /* rewrite all disk labels */
771 vdev_config_dirty(spa->spa_root_vdev);
773 if (vdev_resilver_needed(spa->spa_root_vdev,
774 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
775 nvlist_t *aux = fnvlist_alloc();
776 fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE,
778 spa_event_notify(spa, NULL, aux,
779 ESC_ZFS_RESILVER_START);
782 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
785 spa->spa_scrub_started = B_TRUE;
787 * If this is an incremental scrub, limit the DDT scrub phase
788 * to just the auto-ditto class (for correctness); the rest
789 * of the scrub should go faster using top-down pruning.
791 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
792 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
795 * When starting a resilver clear any existing rebuild state.
796 * This is required to prevent stale rebuild status from
797 * being reported when a rebuild is run, then a resilver and
798 * finally a scrub. In which case only the scrub status
799 * should be reported by 'zpool status'.
801 if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
802 vdev_t *rvd = spa->spa_root_vdev;
803 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
804 vdev_t *vd = rvd->vdev_child[i];
805 vdev_rebuild_clear_sync(
806 (void *)(uintptr_t)vd->vdev_id, tx);
811 /* back to the generic stuff */
813 if (zfs_scan_blkstats) {
814 if (dp->dp_blkstats == NULL) {
816 vmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
818 memset(&dp->dp_blkstats->zab_type, 0,
819 sizeof (dp->dp_blkstats->zab_type));
821 if (dp->dp_blkstats) {
822 vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
823 dp->dp_blkstats = NULL;
827 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
828 ot = DMU_OT_ZAP_OTHER;
830 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
831 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
833 memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys));
835 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
837 spa_history_log_internal(spa, "scan setup", tx,
838 "func=%u mintxg=%llu maxtxg=%llu",
839 *funcp, (u_longlong_t)scn->scn_phys.scn_min_txg,
840 (u_longlong_t)scn->scn_phys.scn_max_txg);
844 * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
845 * Can also be called to resume a paused scrub.
848 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
850 spa_t *spa = dp->dp_spa;
851 dsl_scan_t *scn = dp->dp_scan;
854 * Purge all vdev caches and probe all devices. We do this here
855 * rather than in sync context because this requires a writer lock
856 * on the spa_config lock, which we can't do from sync context. The
857 * spa_scrub_reopen flag indicates that vdev_open() should not
858 * attempt to start another scrub.
860 spa_vdev_state_enter(spa, SCL_NONE);
861 spa->spa_scrub_reopen = B_TRUE;
862 vdev_reopen(spa->spa_root_vdev);
863 spa->spa_scrub_reopen = B_FALSE;
864 (void) spa_vdev_state_exit(spa, NULL, 0);
866 if (func == POOL_SCAN_RESILVER) {
867 dsl_scan_restart_resilver(spa->spa_dsl_pool, 0);
871 if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
872 /* got scrub start cmd, resume paused scrub */
873 int err = dsl_scrub_set_pause_resume(scn->scn_dp,
876 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
877 return (SET_ERROR(ECANCELED));
880 return (SET_ERROR(err));
883 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
884 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED));
888 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
890 static const char *old_names[] = {
892 "scrub_ddt_bookmark",
893 "scrub_ddt_class_max",
902 dsl_pool_t *dp = scn->scn_dp;
903 spa_t *spa = dp->dp_spa;
906 /* Remove any remnants of an old-style scrub. */
907 for (i = 0; old_names[i]; i++) {
908 (void) zap_remove(dp->dp_meta_objset,
909 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
912 if (scn->scn_phys.scn_queue_obj != 0) {
913 VERIFY0(dmu_object_free(dp->dp_meta_objset,
914 scn->scn_phys.scn_queue_obj, tx));
915 scn->scn_phys.scn_queue_obj = 0;
917 scan_ds_queue_clear(scn);
918 scan_ds_prefetch_queue_clear(scn);
920 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
923 * If we were "restarted" from a stopped state, don't bother
924 * with anything else.
926 if (!dsl_scan_is_running(scn)) {
927 ASSERT(!scn->scn_is_sorted);
931 if (scn->scn_is_sorted) {
932 scan_io_queues_destroy(scn);
933 scn->scn_is_sorted = B_FALSE;
935 if (scn->scn_taskq != NULL) {
936 taskq_destroy(scn->scn_taskq);
937 scn->scn_taskq = NULL;
941 scn->scn_phys.scn_state = complete ? DSS_FINISHED : DSS_CANCELED;
943 spa_notify_waiters(spa);
945 if (dsl_scan_restarting(scn, tx))
946 spa_history_log_internal(spa, "scan aborted, restarting", tx,
947 "errors=%llu", (u_longlong_t)spa_get_errlog_size(spa));
949 spa_history_log_internal(spa, "scan cancelled", tx,
950 "errors=%llu", (u_longlong_t)spa_get_errlog_size(spa));
952 spa_history_log_internal(spa, "scan done", tx,
953 "errors=%llu", (u_longlong_t)spa_get_errlog_size(spa));
955 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
956 spa->spa_scrub_active = B_FALSE;
959 * If the scrub/resilver completed, update all DTLs to
960 * reflect this. Whether it succeeded or not, vacate
961 * all temporary scrub DTLs.
963 * As the scrub does not currently support traversing
964 * data that have been freed but are part of a checkpoint,
965 * we don't mark the scrub as done in the DTLs as faults
966 * may still exist in those vdevs.
969 !spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
970 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
971 scn->scn_phys.scn_max_txg, B_TRUE, B_FALSE);
973 if (scn->scn_phys.scn_min_txg) {
974 nvlist_t *aux = fnvlist_alloc();
975 fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE,
977 spa_event_notify(spa, NULL, aux,
978 ESC_ZFS_RESILVER_FINISH);
981 spa_event_notify(spa, NULL, NULL,
982 ESC_ZFS_SCRUB_FINISH);
985 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
988 spa_errlog_rotate(spa);
991 * Don't clear flag until after vdev_dtl_reassess to ensure that
992 * DTL_MISSING will get updated when possible.
994 spa->spa_scrub_started = B_FALSE;
997 * We may have finished replacing a device.
998 * Let the async thread assess this and handle the detach.
1000 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
1003 * Clear any resilver_deferred flags in the config.
1004 * If there are drives that need resilvering, kick
1005 * off an asynchronous request to start resilver.
1006 * vdev_clear_resilver_deferred() may update the config
1007 * before the resilver can restart. In the event of
1008 * a crash during this period, the spa loading code
1009 * will find the drives that need to be resilvered
1010 * and start the resilver then.
1012 if (spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER) &&
1013 vdev_clear_resilver_deferred(spa->spa_root_vdev, tx)) {
1014 spa_history_log_internal(spa,
1015 "starting deferred resilver", tx, "errors=%llu",
1016 (u_longlong_t)spa_get_errlog_size(spa));
1017 spa_async_request(spa, SPA_ASYNC_RESILVER);
1020 /* Clear recent error events (i.e. duplicate events tracking) */
1022 zfs_ereport_clear(spa, NULL);
1025 scn->scn_phys.scn_end_time = gethrestime_sec();
1027 if (spa->spa_errata == ZPOOL_ERRATA_ZOL_2094_SCRUB)
1028 spa->spa_errata = 0;
1030 ASSERT(!dsl_scan_is_running(scn));
1034 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
1037 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1039 if (!dsl_scan_is_running(scn))
1040 return (SET_ERROR(ENOENT));
1045 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
1048 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1050 dsl_scan_done(scn, B_FALSE, tx);
1051 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
1052 spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
1056 dsl_scan_cancel(dsl_pool_t *dp)
1058 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
1059 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
1063 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
1065 pool_scrub_cmd_t *cmd = arg;
1066 dsl_pool_t *dp = dmu_tx_pool(tx);
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 if (!dsl_scan_scrubbing(dp))
1072 return (SET_ERROR(ENOENT));
1074 /* can't pause a paused scrub */
1075 if (dsl_scan_is_paused_scrub(scn))
1076 return (SET_ERROR(EBUSY));
1077 } else if (*cmd != POOL_SCRUB_NORMAL) {
1078 return (SET_ERROR(ENOTSUP));
1085 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
1087 pool_scrub_cmd_t *cmd = arg;
1088 dsl_pool_t *dp = dmu_tx_pool(tx);
1089 spa_t *spa = dp->dp_spa;
1090 dsl_scan_t *scn = dp->dp_scan;
1092 if (*cmd == POOL_SCRUB_PAUSE) {
1093 /* can't pause a scrub when there is no in-progress scrub */
1094 spa->spa_scan_pass_scrub_pause = gethrestime_sec();
1095 scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
1096 scn->scn_phys_cached.scn_flags |= DSF_SCRUB_PAUSED;
1097 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1098 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
1099 spa_notify_waiters(spa);
1101 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
1102 if (dsl_scan_is_paused_scrub(scn)) {
1104 * We need to keep track of how much time we spend
1105 * paused per pass so that we can adjust the scrub rate
1106 * shown in the output of 'zpool status'
1108 spa->spa_scan_pass_scrub_spent_paused +=
1109 gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
1110 spa->spa_scan_pass_scrub_pause = 0;
1111 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
1112 scn->scn_phys_cached.scn_flags &= ~DSF_SCRUB_PAUSED;
1113 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1119 * Set scrub pause/resume state if it makes sense to do so
1122 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
1124 return (dsl_sync_task(spa_name(dp->dp_spa),
1125 dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
1126 ZFS_SPACE_CHECK_RESERVED));
1130 /* start a new scan, or restart an existing one. */
1132 dsl_scan_restart_resilver(dsl_pool_t *dp, uint64_t txg)
1136 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1137 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1139 txg = dmu_tx_get_txg(tx);
1140 dp->dp_scan->scn_restart_txg = txg;
1143 dp->dp_scan->scn_restart_txg = txg;
1145 zfs_dbgmsg("restarting resilver for %s at txg=%llu",
1146 dp->dp_spa->spa_name, (longlong_t)txg);
1150 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
1152 zio_free(dp->dp_spa, txg, bp);
1156 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
1158 ASSERT(dsl_pool_sync_context(dp));
1159 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
1163 scan_ds_queue_compare(const void *a, const void *b)
1165 const scan_ds_t *sds_a = a, *sds_b = b;
1167 if (sds_a->sds_dsobj < sds_b->sds_dsobj)
1169 if (sds_a->sds_dsobj == sds_b->sds_dsobj)
1175 scan_ds_queue_clear(dsl_scan_t *scn)
1177 void *cookie = NULL;
1179 while ((sds = avl_destroy_nodes(&scn->scn_queue, &cookie)) != NULL) {
1180 kmem_free(sds, sizeof (*sds));
1185 scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, uint64_t *txg)
1187 scan_ds_t srch, *sds;
1189 srch.sds_dsobj = dsobj;
1190 sds = avl_find(&scn->scn_queue, &srch, NULL);
1191 if (sds != NULL && txg != NULL)
1192 *txg = sds->sds_txg;
1193 return (sds != NULL);
1197 scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg)
1202 sds = kmem_zalloc(sizeof (*sds), KM_SLEEP);
1203 sds->sds_dsobj = dsobj;
1206 VERIFY3P(avl_find(&scn->scn_queue, sds, &where), ==, NULL);
1207 avl_insert(&scn->scn_queue, sds, where);
1211 scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj)
1213 scan_ds_t srch, *sds;
1215 srch.sds_dsobj = dsobj;
1217 sds = avl_find(&scn->scn_queue, &srch, NULL);
1218 VERIFY(sds != NULL);
1219 avl_remove(&scn->scn_queue, sds);
1220 kmem_free(sds, sizeof (*sds));
1224 scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx)
1226 dsl_pool_t *dp = scn->scn_dp;
1227 spa_t *spa = dp->dp_spa;
1228 dmu_object_type_t ot = (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) ?
1229 DMU_OT_SCAN_QUEUE : DMU_OT_ZAP_OTHER;
1231 ASSERT0(scn->scn_queues_pending);
1232 ASSERT(scn->scn_phys.scn_queue_obj != 0);
1234 VERIFY0(dmu_object_free(dp->dp_meta_objset,
1235 scn->scn_phys.scn_queue_obj, tx));
1236 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, ot,
1237 DMU_OT_NONE, 0, tx);
1238 for (scan_ds_t *sds = avl_first(&scn->scn_queue);
1239 sds != NULL; sds = AVL_NEXT(&scn->scn_queue, sds)) {
1240 VERIFY0(zap_add_int_key(dp->dp_meta_objset,
1241 scn->scn_phys.scn_queue_obj, sds->sds_dsobj,
1247 * Computes the memory limit state that we're currently in. A sorted scan
1248 * needs quite a bit of memory to hold the sorting queue, so we need to
1249 * reasonably constrain the size so it doesn't impact overall system
1250 * performance. We compute two limits:
1251 * 1) Hard memory limit: if the amount of memory used by the sorting
1252 * queues on a pool gets above this value, we stop the metadata
1253 * scanning portion and start issuing the queued up and sorted
1254 * I/Os to reduce memory usage.
1255 * This limit is calculated as a fraction of physmem (by default 5%).
1256 * We constrain the lower bound of the hard limit to an absolute
1257 * minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1258 * the upper bound to 5% of the total pool size - no chance we'll
1259 * ever need that much memory, but just to keep the value in check.
1260 * 2) Soft memory limit: once we hit the hard memory limit, we start
1261 * issuing I/O to reduce queue memory usage, but we don't want to
1262 * completely empty out the queues, since we might be able to find I/Os
1263 * that will fill in the gaps of our non-sequential IOs at some point
1264 * in the future. So we stop the issuing of I/Os once the amount of
1265 * memory used drops below the soft limit (at which point we stop issuing
1266 * I/O and start scanning metadata again).
1268 * This limit is calculated by subtracting a fraction of the hard
1269 * limit from the hard limit. By default this fraction is 5%, so
1270 * the soft limit is 95% of the hard limit. We cap the size of the
1271 * difference between the hard and soft limits at an absolute
1272 * maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1273 * sufficient to not cause too frequent switching between the
1274 * metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1275 * worth of queues is about 1.2 GiB of on-pool data, so scanning
1276 * that should take at least a decent fraction of a second).
1279 dsl_scan_should_clear(dsl_scan_t *scn)
1281 spa_t *spa = scn->scn_dp->dp_spa;
1282 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
1283 uint64_t alloc, mlim_hard, mlim_soft, mused;
1285 alloc = metaslab_class_get_alloc(spa_normal_class(spa));
1286 alloc += metaslab_class_get_alloc(spa_special_class(spa));
1287 alloc += metaslab_class_get_alloc(spa_dedup_class(spa));
1289 mlim_hard = MAX((physmem / zfs_scan_mem_lim_fact) * PAGESIZE,
1290 zfs_scan_mem_lim_min);
1291 mlim_hard = MIN(mlim_hard, alloc / 20);
1292 mlim_soft = mlim_hard - MIN(mlim_hard / zfs_scan_mem_lim_soft_fact,
1293 zfs_scan_mem_lim_soft_max);
1295 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1296 vdev_t *tvd = rvd->vdev_child[i];
1297 dsl_scan_io_queue_t *queue;
1299 mutex_enter(&tvd->vdev_scan_io_queue_lock);
1300 queue = tvd->vdev_scan_io_queue;
1301 if (queue != NULL) {
1303 * # of extents in exts_by_addr = # in exts_by_size.
1304 * B-tree efficiency is ~75%, but can be as low as 50%.
1306 mused += zfs_btree_numnodes(&queue->q_exts_by_size) *
1307 ((sizeof (range_seg_gap_t) + sizeof (uint64_t)) *
1308 3 / 2) + queue->q_sio_memused;
1310 mutex_exit(&tvd->vdev_scan_io_queue_lock);
1313 dprintf("current scan memory usage: %llu bytes\n", (longlong_t)mused);
1316 ASSERT0(scn->scn_queues_pending);
1319 * If we are above our hard limit, we need to clear out memory.
1320 * If we are below our soft limit, we need to accumulate sequential IOs.
1321 * Otherwise, we should keep doing whatever we are currently doing.
1323 if (mused >= mlim_hard)
1325 else if (mused < mlim_soft)
1328 return (scn->scn_clearing);
1332 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1334 /* we never skip user/group accounting objects */
1335 if (zb && (int64_t)zb->zb_object < 0)
1338 if (scn->scn_suspending)
1339 return (B_TRUE); /* we're already suspending */
1341 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
1342 return (B_FALSE); /* we're resuming */
1344 /* We only know how to resume from level-0 and objset blocks. */
1345 if (zb && (zb->zb_level != 0 && zb->zb_level != ZB_ROOT_LEVEL))
1350 * - we have scanned for at least the minimum time (default 1 sec
1351 * for scrub, 3 sec for resilver), and either we have sufficient
1352 * dirty data that we are starting to write more quickly
1353 * (default 30%), someone is explicitly waiting for this txg
1354 * to complete, or we have used up all of the time in the txg
1355 * timeout (default 5 sec).
1357 * - the spa is shutting down because this pool is being exported
1358 * or the machine is rebooting.
1360 * - the scan queue has reached its memory use limit
1362 uint64_t curr_time_ns = gethrtime();
1363 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
1364 uint64_t sync_time_ns = curr_time_ns -
1365 scn->scn_dp->dp_spa->spa_sync_starttime;
1366 uint64_t dirty_min_bytes = zfs_dirty_data_max *
1367 zfs_vdev_async_write_active_min_dirty_percent / 100;
1368 uint_t mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
1369 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
1371 if ((NSEC2MSEC(scan_time_ns) > mintime &&
1372 (scn->scn_dp->dp_dirty_total >= dirty_min_bytes ||
1373 txg_sync_waiting(scn->scn_dp) ||
1374 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1375 spa_shutting_down(scn->scn_dp->dp_spa) ||
1376 (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn))) {
1377 if (zb && zb->zb_level == ZB_ROOT_LEVEL) {
1378 dprintf("suspending at first available bookmark "
1379 "%llx/%llx/%llx/%llx\n",
1380 (longlong_t)zb->zb_objset,
1381 (longlong_t)zb->zb_object,
1382 (longlong_t)zb->zb_level,
1383 (longlong_t)zb->zb_blkid);
1384 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
1385 zb->zb_objset, 0, 0, 0);
1386 } else if (zb != NULL) {
1387 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1388 (longlong_t)zb->zb_objset,
1389 (longlong_t)zb->zb_object,
1390 (longlong_t)zb->zb_level,
1391 (longlong_t)zb->zb_blkid);
1392 scn->scn_phys.scn_bookmark = *zb;
1395 dsl_scan_phys_t *scnp = &scn->scn_phys;
1396 dprintf("suspending at at DDT bookmark "
1397 "%llx/%llx/%llx/%llx\n",
1398 (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
1399 (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
1400 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
1401 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
1404 scn->scn_suspending = B_TRUE;
1410 typedef struct zil_scan_arg {
1412 zil_header_t *zsa_zh;
1416 dsl_scan_zil_block(zilog_t *zilog, const blkptr_t *bp, void *arg,
1420 zil_scan_arg_t *zsa = arg;
1421 dsl_pool_t *dp = zsa->zsa_dp;
1422 dsl_scan_t *scn = dp->dp_scan;
1423 zil_header_t *zh = zsa->zsa_zh;
1424 zbookmark_phys_t zb;
1426 ASSERT(!BP_IS_REDACTED(bp));
1427 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1431 * One block ("stubby") can be allocated a long time ago; we
1432 * want to visit that one because it has been allocated
1433 * (on-disk) even if it hasn't been claimed (even though for
1434 * scrub there's nothing to do to it).
1436 if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(dp->dp_spa))
1439 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1440 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
1442 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1447 dsl_scan_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg,
1451 if (lrc->lrc_txtype == TX_WRITE) {
1452 zil_scan_arg_t *zsa = arg;
1453 dsl_pool_t *dp = zsa->zsa_dp;
1454 dsl_scan_t *scn = dp->dp_scan;
1455 zil_header_t *zh = zsa->zsa_zh;
1456 const lr_write_t *lr = (const lr_write_t *)lrc;
1457 const blkptr_t *bp = &lr->lr_blkptr;
1458 zbookmark_phys_t zb;
1460 ASSERT(!BP_IS_REDACTED(bp));
1461 if (BP_IS_HOLE(bp) ||
1462 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1466 * birth can be < claim_txg if this record's txg is
1467 * already txg sync'ed (but this log block contains
1468 * other records that are not synced)
1470 if (claim_txg == 0 || bp->blk_birth < claim_txg)
1473 ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
1474 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1475 lr->lr_foid, ZB_ZIL_LEVEL,
1476 lr->lr_offset / BP_GET_LSIZE(bp));
1478 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1484 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
1486 uint64_t claim_txg = zh->zh_claim_txg;
1487 zil_scan_arg_t zsa = { dp, zh };
1490 ASSERT(spa_writeable(dp->dp_spa));
1493 * We only want to visit blocks that have been claimed but not yet
1494 * replayed (or, in read-only mode, blocks that *would* be claimed).
1499 zilog = zil_alloc(dp->dp_meta_objset, zh);
1501 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
1502 claim_txg, B_FALSE);
1508 * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1509 * here is to sort the AVL tree by the order each block will be needed.
1512 scan_prefetch_queue_compare(const void *a, const void *b)
1514 const scan_prefetch_issue_ctx_t *spic_a = a, *spic_b = b;
1515 const scan_prefetch_ctx_t *spc_a = spic_a->spic_spc;
1516 const scan_prefetch_ctx_t *spc_b = spic_b->spic_spc;
1518 return (zbookmark_compare(spc_a->spc_datablkszsec,
1519 spc_a->spc_indblkshift, spc_b->spc_datablkszsec,
1520 spc_b->spc_indblkshift, &spic_a->spic_zb, &spic_b->spic_zb));
1524 scan_prefetch_ctx_rele(scan_prefetch_ctx_t *spc, const void *tag)
1526 if (zfs_refcount_remove(&spc->spc_refcnt, tag) == 0) {
1527 zfs_refcount_destroy(&spc->spc_refcnt);
1528 kmem_free(spc, sizeof (scan_prefetch_ctx_t));
1532 static scan_prefetch_ctx_t *
1533 scan_prefetch_ctx_create(dsl_scan_t *scn, dnode_phys_t *dnp, const void *tag)
1535 scan_prefetch_ctx_t *spc;
1537 spc = kmem_alloc(sizeof (scan_prefetch_ctx_t), KM_SLEEP);
1538 zfs_refcount_create(&spc->spc_refcnt);
1539 zfs_refcount_add(&spc->spc_refcnt, tag);
1542 spc->spc_datablkszsec = dnp->dn_datablkszsec;
1543 spc->spc_indblkshift = dnp->dn_indblkshift;
1544 spc->spc_root = B_FALSE;
1546 spc->spc_datablkszsec = 0;
1547 spc->spc_indblkshift = 0;
1548 spc->spc_root = B_TRUE;
1555 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t *spc, const void *tag)
1557 zfs_refcount_add(&spc->spc_refcnt, tag);
1561 scan_ds_prefetch_queue_clear(dsl_scan_t *scn)
1563 spa_t *spa = scn->scn_dp->dp_spa;
1564 void *cookie = NULL;
1565 scan_prefetch_issue_ctx_t *spic = NULL;
1567 mutex_enter(&spa->spa_scrub_lock);
1568 while ((spic = avl_destroy_nodes(&scn->scn_prefetch_queue,
1569 &cookie)) != NULL) {
1570 scan_prefetch_ctx_rele(spic->spic_spc, scn);
1571 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1573 mutex_exit(&spa->spa_scrub_lock);
1577 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t *spc,
1578 const zbookmark_phys_t *zb)
1580 zbookmark_phys_t *last_zb = &spc->spc_scn->scn_prefetch_bookmark;
1581 dnode_phys_t tmp_dnp;
1582 dnode_phys_t *dnp = (spc->spc_root) ? NULL : &tmp_dnp;
1584 if (zb->zb_objset != last_zb->zb_objset)
1586 if ((int64_t)zb->zb_object < 0)
1589 tmp_dnp.dn_datablkszsec = spc->spc_datablkszsec;
1590 tmp_dnp.dn_indblkshift = spc->spc_indblkshift;
1592 if (zbookmark_subtree_completed(dnp, zb, last_zb))
1599 dsl_scan_prefetch(scan_prefetch_ctx_t *spc, blkptr_t *bp, zbookmark_phys_t *zb)
1602 dsl_scan_t *scn = spc->spc_scn;
1603 spa_t *spa = scn->scn_dp->dp_spa;
1604 scan_prefetch_issue_ctx_t *spic;
1606 if (zfs_no_scrub_prefetch || BP_IS_REDACTED(bp))
1609 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg ||
1610 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE &&
1611 BP_GET_TYPE(bp) != DMU_OT_OBJSET))
1614 if (dsl_scan_check_prefetch_resume(spc, zb))
1617 scan_prefetch_ctx_add_ref(spc, scn);
1618 spic = kmem_alloc(sizeof (scan_prefetch_issue_ctx_t), KM_SLEEP);
1619 spic->spic_spc = spc;
1620 spic->spic_bp = *bp;
1621 spic->spic_zb = *zb;
1624 * Add the IO to the queue of blocks to prefetch. This allows us to
1625 * prioritize blocks that we will need first for the main traversal
1628 mutex_enter(&spa->spa_scrub_lock);
1629 if (avl_find(&scn->scn_prefetch_queue, spic, &idx) != NULL) {
1630 /* this block is already queued for prefetch */
1631 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1632 scan_prefetch_ctx_rele(spc, scn);
1633 mutex_exit(&spa->spa_scrub_lock);
1637 avl_insert(&scn->scn_prefetch_queue, spic, idx);
1638 cv_broadcast(&spa->spa_scrub_io_cv);
1639 mutex_exit(&spa->spa_scrub_lock);
1643 dsl_scan_prefetch_dnode(dsl_scan_t *scn, dnode_phys_t *dnp,
1644 uint64_t objset, uint64_t object)
1647 zbookmark_phys_t zb;
1648 scan_prefetch_ctx_t *spc;
1650 if (dnp->dn_nblkptr == 0 && !(dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1653 SET_BOOKMARK(&zb, objset, object, 0, 0);
1655 spc = scan_prefetch_ctx_create(scn, dnp, FTAG);
1657 for (i = 0; i < dnp->dn_nblkptr; i++) {
1658 zb.zb_level = BP_GET_LEVEL(&dnp->dn_blkptr[i]);
1660 dsl_scan_prefetch(spc, &dnp->dn_blkptr[i], &zb);
1663 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1665 zb.zb_blkid = DMU_SPILL_BLKID;
1666 dsl_scan_prefetch(spc, DN_SPILL_BLKPTR(dnp), &zb);
1669 scan_prefetch_ctx_rele(spc, FTAG);
1673 dsl_scan_prefetch_cb(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1674 arc_buf_t *buf, void *private)
1677 scan_prefetch_ctx_t *spc = private;
1678 dsl_scan_t *scn = spc->spc_scn;
1679 spa_t *spa = scn->scn_dp->dp_spa;
1681 /* broadcast that the IO has completed for rate limiting purposes */
1682 mutex_enter(&spa->spa_scrub_lock);
1683 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
1684 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
1685 cv_broadcast(&spa->spa_scrub_io_cv);
1686 mutex_exit(&spa->spa_scrub_lock);
1688 /* if there was an error or we are done prefetching, just cleanup */
1689 if (buf == NULL || scn->scn_prefetch_stop)
1692 if (BP_GET_LEVEL(bp) > 0) {
1695 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1696 zbookmark_phys_t czb;
1698 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1699 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1700 zb->zb_level - 1, zb->zb_blkid * epb + i);
1701 dsl_scan_prefetch(spc, cbp, &czb);
1703 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1706 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1708 for (i = 0, cdnp = buf->b_data; i < epb;
1709 i += cdnp->dn_extra_slots + 1,
1710 cdnp += cdnp->dn_extra_slots + 1) {
1711 dsl_scan_prefetch_dnode(scn, cdnp,
1712 zb->zb_objset, zb->zb_blkid * epb + i);
1714 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1715 objset_phys_t *osp = buf->b_data;
1717 dsl_scan_prefetch_dnode(scn, &osp->os_meta_dnode,
1718 zb->zb_objset, DMU_META_DNODE_OBJECT);
1720 if (OBJSET_BUF_HAS_USERUSED(buf)) {
1721 dsl_scan_prefetch_dnode(scn,
1722 &osp->os_groupused_dnode, zb->zb_objset,
1723 DMU_GROUPUSED_OBJECT);
1724 dsl_scan_prefetch_dnode(scn,
1725 &osp->os_userused_dnode, zb->zb_objset,
1726 DMU_USERUSED_OBJECT);
1732 arc_buf_destroy(buf, private);
1733 scan_prefetch_ctx_rele(spc, scn);
1737 dsl_scan_prefetch_thread(void *arg)
1739 dsl_scan_t *scn = arg;
1740 spa_t *spa = scn->scn_dp->dp_spa;
1741 scan_prefetch_issue_ctx_t *spic;
1743 /* loop until we are told to stop */
1744 while (!scn->scn_prefetch_stop) {
1745 arc_flags_t flags = ARC_FLAG_NOWAIT |
1746 ARC_FLAG_PRESCIENT_PREFETCH | ARC_FLAG_PREFETCH;
1747 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1749 mutex_enter(&spa->spa_scrub_lock);
1752 * Wait until we have an IO to issue and are not above our
1753 * maximum in flight limit.
1755 while (!scn->scn_prefetch_stop &&
1756 (avl_numnodes(&scn->scn_prefetch_queue) == 0 ||
1757 spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)) {
1758 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1761 /* recheck if we should stop since we waited for the cv */
1762 if (scn->scn_prefetch_stop) {
1763 mutex_exit(&spa->spa_scrub_lock);
1767 /* remove the prefetch IO from the tree */
1768 spic = avl_first(&scn->scn_prefetch_queue);
1769 spa->spa_scrub_inflight += BP_GET_PSIZE(&spic->spic_bp);
1770 avl_remove(&scn->scn_prefetch_queue, spic);
1772 mutex_exit(&spa->spa_scrub_lock);
1774 if (BP_IS_PROTECTED(&spic->spic_bp)) {
1775 ASSERT(BP_GET_TYPE(&spic->spic_bp) == DMU_OT_DNODE ||
1776 BP_GET_TYPE(&spic->spic_bp) == DMU_OT_OBJSET);
1777 ASSERT3U(BP_GET_LEVEL(&spic->spic_bp), ==, 0);
1778 zio_flags |= ZIO_FLAG_RAW;
1781 /* issue the prefetch asynchronously */
1782 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa,
1783 &spic->spic_bp, dsl_scan_prefetch_cb, spic->spic_spc,
1784 ZIO_PRIORITY_SCRUB, zio_flags, &flags, &spic->spic_zb);
1786 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1789 ASSERT(scn->scn_prefetch_stop);
1791 /* free any prefetches we didn't get to complete */
1792 mutex_enter(&spa->spa_scrub_lock);
1793 while ((spic = avl_first(&scn->scn_prefetch_queue)) != NULL) {
1794 avl_remove(&scn->scn_prefetch_queue, spic);
1795 scan_prefetch_ctx_rele(spic->spic_spc, scn);
1796 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1798 ASSERT0(avl_numnodes(&scn->scn_prefetch_queue));
1799 mutex_exit(&spa->spa_scrub_lock);
1803 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
1804 const zbookmark_phys_t *zb)
1807 * We never skip over user/group accounting objects (obj<0)
1809 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
1810 (int64_t)zb->zb_object >= 0) {
1812 * If we already visited this bp & everything below (in
1813 * a prior txg sync), don't bother doing it again.
1815 if (zbookmark_subtree_completed(dnp, zb,
1816 &scn->scn_phys.scn_bookmark))
1820 * If we found the block we're trying to resume from, or
1821 * we went past it, zero it out to indicate that it's OK
1822 * to start checking for suspending again.
1824 if (zbookmark_subtree_tbd(dnp, zb,
1825 &scn->scn_phys.scn_bookmark)) {
1826 dprintf("resuming at %llx/%llx/%llx/%llx\n",
1827 (longlong_t)zb->zb_objset,
1828 (longlong_t)zb->zb_object,
1829 (longlong_t)zb->zb_level,
1830 (longlong_t)zb->zb_blkid);
1831 memset(&scn->scn_phys.scn_bookmark, 0, sizeof (*zb));
1837 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1838 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1839 dmu_objset_type_t ostype, dmu_tx_t *tx);
1840 inline __attribute__((always_inline)) static void dsl_scan_visitdnode(
1841 dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1842 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
1845 * Return nonzero on i/o error.
1846 * Return new buf to write out in *bufp.
1848 inline __attribute__((always_inline)) static int
1849 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1850 dnode_phys_t *dnp, const blkptr_t *bp,
1851 const zbookmark_phys_t *zb, dmu_tx_t *tx)
1853 dsl_pool_t *dp = scn->scn_dp;
1854 spa_t *spa = dp->dp_spa;
1855 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1858 ASSERT(!BP_IS_REDACTED(bp));
1861 * There is an unlikely case of encountering dnodes with contradicting
1862 * dn_bonuslen and DNODE_FLAG_SPILL_BLKPTR flag before in files created
1863 * or modified before commit 4254acb was merged. As it is not possible
1864 * to know which of the two is correct, report an error.
1867 dnp->dn_bonuslen > DN_MAX_BONUS_LEN(dnp)) {
1868 scn->scn_phys.scn_errors++;
1869 spa_log_error(spa, zb);
1870 return (SET_ERROR(EINVAL));
1873 if (BP_GET_LEVEL(bp) > 0) {
1874 arc_flags_t flags = ARC_FLAG_WAIT;
1877 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1880 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
1881 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1883 scn->scn_phys.scn_errors++;
1886 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1887 zbookmark_phys_t czb;
1889 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1891 zb->zb_blkid * epb + i);
1892 dsl_scan_visitbp(cbp, &czb, dnp,
1893 ds, scn, ostype, tx);
1895 arc_buf_destroy(buf, &buf);
1896 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1897 arc_flags_t flags = ARC_FLAG_WAIT;
1900 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1903 if (BP_IS_PROTECTED(bp)) {
1904 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
1905 zio_flags |= ZIO_FLAG_RAW;
1908 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
1909 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1911 scn->scn_phys.scn_errors++;
1914 for (i = 0, cdnp = buf->b_data; i < epb;
1915 i += cdnp->dn_extra_slots + 1,
1916 cdnp += cdnp->dn_extra_slots + 1) {
1917 dsl_scan_visitdnode(scn, ds, ostype,
1918 cdnp, zb->zb_blkid * epb + i, tx);
1921 arc_buf_destroy(buf, &buf);
1922 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1923 arc_flags_t flags = ARC_FLAG_WAIT;
1927 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
1928 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1930 scn->scn_phys.scn_errors++;
1936 dsl_scan_visitdnode(scn, ds, osp->os_type,
1937 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
1939 if (OBJSET_BUF_HAS_USERUSED(buf)) {
1941 * We also always visit user/group/project accounting
1942 * objects, and never skip them, even if we are
1943 * suspending. This is necessary so that the
1944 * space deltas from this txg get integrated.
1946 if (OBJSET_BUF_HAS_PROJECTUSED(buf))
1947 dsl_scan_visitdnode(scn, ds, osp->os_type,
1948 &osp->os_projectused_dnode,
1949 DMU_PROJECTUSED_OBJECT, tx);
1950 dsl_scan_visitdnode(scn, ds, osp->os_type,
1951 &osp->os_groupused_dnode,
1952 DMU_GROUPUSED_OBJECT, tx);
1953 dsl_scan_visitdnode(scn, ds, osp->os_type,
1954 &osp->os_userused_dnode,
1955 DMU_USERUSED_OBJECT, tx);
1957 arc_buf_destroy(buf, &buf);
1958 } else if (!zfs_blkptr_verify(spa, bp, B_FALSE, BLK_VERIFY_LOG)) {
1960 * Sanity check the block pointer contents, this is handled
1961 * by arc_read() for the cases above.
1963 scn->scn_phys.scn_errors++;
1964 spa_log_error(spa, zb);
1965 return (SET_ERROR(EINVAL));
1971 inline __attribute__((always_inline)) static void
1972 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
1973 dmu_objset_type_t ostype, dnode_phys_t *dnp,
1974 uint64_t object, dmu_tx_t *tx)
1978 for (j = 0; j < dnp->dn_nblkptr; j++) {
1979 zbookmark_phys_t czb;
1981 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1982 dnp->dn_nlevels - 1, j);
1983 dsl_scan_visitbp(&dnp->dn_blkptr[j],
1984 &czb, dnp, ds, scn, ostype, tx);
1987 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1988 zbookmark_phys_t czb;
1989 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1990 0, DMU_SPILL_BLKID);
1991 dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp),
1992 &czb, dnp, ds, scn, ostype, tx);
1997 * The arguments are in this order because mdb can only print the
1998 * first 5; we want them to be useful.
2001 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
2002 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
2003 dmu_objset_type_t ostype, dmu_tx_t *tx)
2005 dsl_pool_t *dp = scn->scn_dp;
2006 blkptr_t *bp_toread = NULL;
2008 if (dsl_scan_check_suspend(scn, zb))
2011 if (dsl_scan_check_resume(scn, dnp, zb))
2014 scn->scn_visited_this_txg++;
2016 if (BP_IS_HOLE(bp)) {
2017 scn->scn_holes_this_txg++;
2021 if (BP_IS_REDACTED(bp)) {
2022 ASSERT(dsl_dataset_feature_is_active(ds,
2023 SPA_FEATURE_REDACTED_DATASETS));
2027 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) {
2028 scn->scn_lt_min_this_txg++;
2032 bp_toread = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
2035 if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx) != 0)
2039 * If dsl_scan_ddt() has already visited this block, it will have
2040 * already done any translations or scrubbing, so don't call the
2043 if (ddt_class_contains(dp->dp_spa,
2044 scn->scn_phys.scn_ddt_class_max, bp)) {
2045 scn->scn_ddt_contained_this_txg++;
2050 * If this block is from the future (after cur_max_txg), then we
2051 * are doing this on behalf of a deleted snapshot, and we will
2052 * revisit the future block on the next pass of this dataset.
2053 * Don't scan it now unless we need to because something
2054 * under it was modified.
2056 if (BP_PHYSICAL_BIRTH(bp) > scn->scn_phys.scn_cur_max_txg) {
2057 scn->scn_gt_max_this_txg++;
2061 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
2064 kmem_free(bp_toread, sizeof (blkptr_t));
2068 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
2071 zbookmark_phys_t zb;
2072 scan_prefetch_ctx_t *spc;
2074 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
2075 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
2077 if (ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) {
2078 SET_BOOKMARK(&scn->scn_prefetch_bookmark,
2079 zb.zb_objset, 0, 0, 0);
2081 scn->scn_prefetch_bookmark = scn->scn_phys.scn_bookmark;
2084 scn->scn_objsets_visited_this_txg++;
2086 spc = scan_prefetch_ctx_create(scn, NULL, FTAG);
2087 dsl_scan_prefetch(spc, bp, &zb);
2088 scan_prefetch_ctx_rele(spc, FTAG);
2090 dsl_scan_visitbp(bp, &zb, NULL, ds, scn, DMU_OST_NONE, tx);
2092 dprintf_ds(ds, "finished scan%s", "");
2096 ds_destroyed_scn_phys(dsl_dataset_t *ds, dsl_scan_phys_t *scn_phys)
2098 if (scn_phys->scn_bookmark.zb_objset == ds->ds_object) {
2099 if (ds->ds_is_snapshot) {
2102 * - scn_cur_{min,max}_txg stays the same.
2103 * - Setting the flag is not really necessary if
2104 * scn_cur_max_txg == scn_max_txg, because there
2105 * is nothing after this snapshot that we care
2106 * about. However, we set it anyway and then
2107 * ignore it when we retraverse it in
2108 * dsl_scan_visitds().
2110 scn_phys->scn_bookmark.zb_objset =
2111 dsl_dataset_phys(ds)->ds_next_snap_obj;
2112 zfs_dbgmsg("destroying ds %llu on %s; currently "
2113 "traversing; reset zb_objset to %llu",
2114 (u_longlong_t)ds->ds_object,
2115 ds->ds_dir->dd_pool->dp_spa->spa_name,
2116 (u_longlong_t)dsl_dataset_phys(ds)->
2118 scn_phys->scn_flags |= DSF_VISIT_DS_AGAIN;
2120 SET_BOOKMARK(&scn_phys->scn_bookmark,
2121 ZB_DESTROYED_OBJSET, 0, 0, 0);
2122 zfs_dbgmsg("destroying ds %llu on %s; currently "
2123 "traversing; reset bookmark to -1,0,0,0",
2124 (u_longlong_t)ds->ds_object,
2125 ds->ds_dir->dd_pool->dp_spa->spa_name);
2131 * Invoked when a dataset is destroyed. We need to make sure that:
2133 * 1) If it is the dataset that was currently being scanned, we write
2134 * a new dsl_scan_phys_t and marking the objset reference in it
2136 * 2) Remove it from the work queue, if it was present.
2138 * If the dataset was actually a snapshot, instead of marking the dataset
2139 * as destroyed, we instead substitute the next snapshot in line.
2142 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
2144 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2145 dsl_scan_t *scn = dp->dp_scan;
2148 if (!dsl_scan_is_running(scn))
2151 ds_destroyed_scn_phys(ds, &scn->scn_phys);
2152 ds_destroyed_scn_phys(ds, &scn->scn_phys_cached);
2154 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2155 scan_ds_queue_remove(scn, ds->ds_object);
2156 if (ds->ds_is_snapshot)
2157 scan_ds_queue_insert(scn,
2158 dsl_dataset_phys(ds)->ds_next_snap_obj, mintxg);
2161 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2162 ds->ds_object, &mintxg) == 0) {
2163 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
2164 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2165 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2166 if (ds->ds_is_snapshot) {
2168 * We keep the same mintxg; it could be >
2169 * ds_creation_txg if the previous snapshot was
2172 VERIFY(zap_add_int_key(dp->dp_meta_objset,
2173 scn->scn_phys.scn_queue_obj,
2174 dsl_dataset_phys(ds)->ds_next_snap_obj,
2176 zfs_dbgmsg("destroying ds %llu on %s; in queue; "
2177 "replacing with %llu",
2178 (u_longlong_t)ds->ds_object,
2179 dp->dp_spa->spa_name,
2180 (u_longlong_t)dsl_dataset_phys(ds)->
2183 zfs_dbgmsg("destroying ds %llu on %s; in queue; "
2185 (u_longlong_t)ds->ds_object,
2186 dp->dp_spa->spa_name);
2191 * dsl_scan_sync() should be called after this, and should sync
2192 * out our changed state, but just to be safe, do it here.
2194 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2198 ds_snapshotted_bookmark(dsl_dataset_t *ds, zbookmark_phys_t *scn_bookmark)
2200 if (scn_bookmark->zb_objset == ds->ds_object) {
2201 scn_bookmark->zb_objset =
2202 dsl_dataset_phys(ds)->ds_prev_snap_obj;
2203 zfs_dbgmsg("snapshotting ds %llu on %s; currently traversing; "
2204 "reset zb_objset to %llu",
2205 (u_longlong_t)ds->ds_object,
2206 ds->ds_dir->dd_pool->dp_spa->spa_name,
2207 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2212 * Called when a dataset is snapshotted. If we were currently traversing
2213 * this snapshot, we reset our bookmark to point at the newly created
2214 * snapshot. We also modify our work queue to remove the old snapshot and
2215 * replace with the new one.
2218 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
2220 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2221 dsl_scan_t *scn = dp->dp_scan;
2224 if (!dsl_scan_is_running(scn))
2227 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
2229 ds_snapshotted_bookmark(ds, &scn->scn_phys.scn_bookmark);
2230 ds_snapshotted_bookmark(ds, &scn->scn_phys_cached.scn_bookmark);
2232 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2233 scan_ds_queue_remove(scn, ds->ds_object);
2234 scan_ds_queue_insert(scn,
2235 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg);
2238 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2239 ds->ds_object, &mintxg) == 0) {
2240 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2241 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2242 VERIFY(zap_add_int_key(dp->dp_meta_objset,
2243 scn->scn_phys.scn_queue_obj,
2244 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
2245 zfs_dbgmsg("snapshotting ds %llu on %s; in queue; "
2246 "replacing with %llu",
2247 (u_longlong_t)ds->ds_object,
2248 dp->dp_spa->spa_name,
2249 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2252 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2256 ds_clone_swapped_bookmark(dsl_dataset_t *ds1, dsl_dataset_t *ds2,
2257 zbookmark_phys_t *scn_bookmark)
2259 if (scn_bookmark->zb_objset == ds1->ds_object) {
2260 scn_bookmark->zb_objset = ds2->ds_object;
2261 zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; "
2262 "reset zb_objset to %llu",
2263 (u_longlong_t)ds1->ds_object,
2264 ds1->ds_dir->dd_pool->dp_spa->spa_name,
2265 (u_longlong_t)ds2->ds_object);
2266 } else if (scn_bookmark->zb_objset == ds2->ds_object) {
2267 scn_bookmark->zb_objset = ds1->ds_object;
2268 zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; "
2269 "reset zb_objset to %llu",
2270 (u_longlong_t)ds2->ds_object,
2271 ds2->ds_dir->dd_pool->dp_spa->spa_name,
2272 (u_longlong_t)ds1->ds_object);
2277 * Called when an origin dataset and its clone are swapped. If we were
2278 * currently traversing the dataset, we need to switch to traversing the
2279 * newly promoted clone.
2282 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
2284 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
2285 dsl_scan_t *scn = dp->dp_scan;
2286 uint64_t mintxg1, mintxg2;
2287 boolean_t ds1_queued, ds2_queued;
2289 if (!dsl_scan_is_running(scn))
2292 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys.scn_bookmark);
2293 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys_cached.scn_bookmark);
2296 * Handle the in-memory scan queue.
2298 ds1_queued = scan_ds_queue_contains(scn, ds1->ds_object, &mintxg1);
2299 ds2_queued = scan_ds_queue_contains(scn, ds2->ds_object, &mintxg2);
2301 /* Sanity checking. */
2303 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2304 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2307 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2308 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2311 if (ds1_queued && ds2_queued) {
2313 * If both are queued, we don't need to do anything.
2314 * The swapping code below would not handle this case correctly,
2315 * since we can't insert ds2 if it is already there. That's
2316 * because scan_ds_queue_insert() prohibits a duplicate insert
2319 } else if (ds1_queued) {
2320 scan_ds_queue_remove(scn, ds1->ds_object);
2321 scan_ds_queue_insert(scn, ds2->ds_object, mintxg1);
2322 } else if (ds2_queued) {
2323 scan_ds_queue_remove(scn, ds2->ds_object);
2324 scan_ds_queue_insert(scn, ds1->ds_object, mintxg2);
2328 * Handle the on-disk scan queue.
2329 * The on-disk state is an out-of-date version of the in-memory state,
2330 * so the in-memory and on-disk values for ds1_queued and ds2_queued may
2331 * be different. Therefore we need to apply the swap logic to the
2332 * on-disk state independently of the in-memory state.
2334 ds1_queued = zap_lookup_int_key(dp->dp_meta_objset,
2335 scn->scn_phys.scn_queue_obj, ds1->ds_object, &mintxg1) == 0;
2336 ds2_queued = zap_lookup_int_key(dp->dp_meta_objset,
2337 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg2) == 0;
2339 /* Sanity checking. */
2341 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2342 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2345 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2346 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2349 if (ds1_queued && ds2_queued) {
2351 * If both are queued, we don't need to do anything.
2352 * Alternatively, we could check for EEXIST from
2353 * zap_add_int_key() and back out to the original state, but
2354 * that would be more work than checking for this case upfront.
2356 } else if (ds1_queued) {
2357 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2358 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
2359 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2360 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg1, tx));
2361 zfs_dbgmsg("clone_swap ds %llu on %s; in queue; "
2362 "replacing with %llu",
2363 (u_longlong_t)ds1->ds_object,
2364 dp->dp_spa->spa_name,
2365 (u_longlong_t)ds2->ds_object);
2366 } else if (ds2_queued) {
2367 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2368 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
2369 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2370 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg2, tx));
2371 zfs_dbgmsg("clone_swap ds %llu on %s; in queue; "
2372 "replacing with %llu",
2373 (u_longlong_t)ds2->ds_object,
2374 dp->dp_spa->spa_name,
2375 (u_longlong_t)ds1->ds_object);
2378 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2382 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2384 uint64_t originobj = *(uint64_t *)arg;
2387 dsl_scan_t *scn = dp->dp_scan;
2389 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != originobj)
2392 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2396 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != originobj) {
2397 dsl_dataset_t *prev;
2398 err = dsl_dataset_hold_obj(dp,
2399 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2401 dsl_dataset_rele(ds, FTAG);
2406 scan_ds_queue_insert(scn, ds->ds_object,
2407 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2408 dsl_dataset_rele(ds, FTAG);
2413 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
2415 dsl_pool_t *dp = scn->scn_dp;
2418 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2420 if (scn->scn_phys.scn_cur_min_txg >=
2421 scn->scn_phys.scn_max_txg) {
2423 * This can happen if this snapshot was created after the
2424 * scan started, and we already completed a previous snapshot
2425 * that was created after the scan started. This snapshot
2426 * only references blocks with:
2428 * birth < our ds_creation_txg
2429 * cur_min_txg is no less than ds_creation_txg.
2430 * We have already visited these blocks.
2432 * birth > scn_max_txg
2433 * The scan requested not to visit these blocks.
2435 * Subsequent snapshots (and clones) can reference our
2436 * blocks, or blocks with even higher birth times.
2437 * Therefore we do not need to visit them either,
2438 * so we do not add them to the work queue.
2440 * Note that checking for cur_min_txg >= cur_max_txg
2441 * is not sufficient, because in that case we may need to
2442 * visit subsequent snapshots. This happens when min_txg > 0,
2443 * which raises cur_min_txg. In this case we will visit
2444 * this dataset but skip all of its blocks, because the
2445 * rootbp's birth time is < cur_min_txg. Then we will
2446 * add the next snapshots/clones to the work queue.
2448 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2449 dsl_dataset_name(ds, dsname);
2450 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2451 "cur_min_txg (%llu) >= max_txg (%llu)",
2452 (longlong_t)dsobj, dsname,
2453 (longlong_t)scn->scn_phys.scn_cur_min_txg,
2454 (longlong_t)scn->scn_phys.scn_max_txg);
2455 kmem_free(dsname, MAXNAMELEN);
2461 * Only the ZIL in the head (non-snapshot) is valid. Even though
2462 * snapshots can have ZIL block pointers (which may be the same
2463 * BP as in the head), they must be ignored. In addition, $ORIGIN
2464 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2465 * need to look for a ZIL in it either. So we traverse the ZIL here,
2466 * rather than in scan_recurse(), because the regular snapshot
2467 * block-sharing rules don't apply to it.
2469 if (!dsl_dataset_is_snapshot(ds) &&
2470 (dp->dp_origin_snap == NULL ||
2471 ds->ds_dir != dp->dp_origin_snap->ds_dir)) {
2473 if (dmu_objset_from_ds(ds, &os) != 0) {
2476 dsl_scan_zil(dp, &os->os_zil_header);
2480 * Iterate over the bps in this ds.
2482 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2483 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
2484 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
2485 rrw_exit(&ds->ds_bp_rwlock, FTAG);
2487 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2488 dsl_dataset_name(ds, dsname);
2489 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2491 (longlong_t)dsobj, dsname,
2492 (longlong_t)scn->scn_phys.scn_cur_min_txg,
2493 (longlong_t)scn->scn_phys.scn_cur_max_txg,
2494 (int)scn->scn_suspending);
2495 kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
2497 if (scn->scn_suspending)
2501 * We've finished this pass over this dataset.
2505 * If we did not completely visit this dataset, do another pass.
2507 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
2508 zfs_dbgmsg("incomplete pass on %s; visiting again",
2509 dp->dp_spa->spa_name);
2510 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
2511 scan_ds_queue_insert(scn, ds->ds_object,
2512 scn->scn_phys.scn_cur_max_txg);
2517 * Add descendant datasets to work queue.
2519 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
2520 scan_ds_queue_insert(scn,
2521 dsl_dataset_phys(ds)->ds_next_snap_obj,
2522 dsl_dataset_phys(ds)->ds_creation_txg);
2524 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
2525 boolean_t usenext = B_FALSE;
2526 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
2529 * A bug in a previous version of the code could
2530 * cause upgrade_clones_cb() to not set
2531 * ds_next_snap_obj when it should, leading to a
2532 * missing entry. Therefore we can only use the
2533 * next_clones_obj when its count is correct.
2535 int err = zap_count(dp->dp_meta_objset,
2536 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
2538 count == dsl_dataset_phys(ds)->ds_num_children - 1)
2545 for (zap_cursor_init(&zc, dp->dp_meta_objset,
2546 dsl_dataset_phys(ds)->ds_next_clones_obj);
2547 zap_cursor_retrieve(&zc, &za) == 0;
2548 (void) zap_cursor_advance(&zc)) {
2549 scan_ds_queue_insert(scn,
2550 zfs_strtonum(za.za_name, NULL),
2551 dsl_dataset_phys(ds)->ds_creation_txg);
2553 zap_cursor_fini(&zc);
2555 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2556 enqueue_clones_cb, &ds->ds_object,
2562 dsl_dataset_rele(ds, FTAG);
2566 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2571 dsl_scan_t *scn = dp->dp_scan;
2573 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2577 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
2578 dsl_dataset_t *prev;
2579 err = dsl_dataset_hold_obj(dp,
2580 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2582 dsl_dataset_rele(ds, FTAG);
2587 * If this is a clone, we don't need to worry about it for now.
2589 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
2590 dsl_dataset_rele(ds, FTAG);
2591 dsl_dataset_rele(prev, FTAG);
2594 dsl_dataset_rele(ds, FTAG);
2598 scan_ds_queue_insert(scn, ds->ds_object,
2599 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2600 dsl_dataset_rele(ds, FTAG);
2605 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
2606 ddt_entry_t *dde, dmu_tx_t *tx)
2609 const ddt_key_t *ddk = &dde->dde_key;
2610 ddt_phys_t *ddp = dde->dde_phys;
2612 zbookmark_phys_t zb = { 0 };
2614 if (!dsl_scan_is_running(scn))
2618 * This function is special because it is the only thing
2619 * that can add scan_io_t's to the vdev scan queues from
2620 * outside dsl_scan_sync(). For the most part this is ok
2621 * as long as it is called from within syncing context.
2622 * However, dsl_scan_sync() expects that no new sio's will
2623 * be added between when all the work for a scan is done
2624 * and the next txg when the scan is actually marked as
2625 * completed. This check ensures we do not issue new sio's
2626 * during this period.
2628 if (scn->scn_done_txg != 0)
2631 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2632 if (ddp->ddp_phys_birth == 0 ||
2633 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
2635 ddt_bp_create(checksum, ddk, ddp, &bp);
2637 scn->scn_visited_this_txg++;
2638 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
2643 * Scrub/dedup interaction.
2645 * If there are N references to a deduped block, we don't want to scrub it
2646 * N times -- ideally, we should scrub it exactly once.
2648 * We leverage the fact that the dde's replication class (enum ddt_class)
2649 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2650 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2652 * To prevent excess scrubbing, the scrub begins by walking the DDT
2653 * to find all blocks with refcnt > 1, and scrubs each of these once.
2654 * Since there are two replication classes which contain blocks with
2655 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2656 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2658 * There would be nothing more to say if a block's refcnt couldn't change
2659 * during a scrub, but of course it can so we must account for changes
2660 * in a block's replication class.
2662 * Here's an example of what can occur:
2664 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2665 * when visited during the top-down scrub phase, it will be scrubbed twice.
2666 * This negates our scrub optimization, but is otherwise harmless.
2668 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2669 * on each visit during the top-down scrub phase, it will never be scrubbed.
2670 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2671 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2672 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2673 * while a scrub is in progress, it scrubs the block right then.
2676 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
2678 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
2679 ddt_entry_t dde = {{{{0}}}};
2683 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
2686 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
2688 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
2689 (longlong_t)ddb->ddb_class,
2690 (longlong_t)ddb->ddb_type,
2691 (longlong_t)ddb->ddb_checksum,
2692 (longlong_t)ddb->ddb_cursor);
2694 /* There should be no pending changes to the dedup table */
2695 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
2696 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
2698 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
2701 if (dsl_scan_check_suspend(scn, NULL))
2705 zfs_dbgmsg("scanned %llu ddt entries on %s with class_max = %u; "
2706 "suspending=%u", (longlong_t)n, scn->scn_dp->dp_spa->spa_name,
2707 (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
2709 ASSERT(error == 0 || error == ENOENT);
2710 ASSERT(error != ENOENT ||
2711 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
2715 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
2717 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
2718 if (ds->ds_is_snapshot)
2719 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
2724 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
2727 dsl_pool_t *dp = scn->scn_dp;
2729 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
2730 scn->scn_phys.scn_ddt_class_max) {
2731 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2732 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2733 dsl_scan_ddt(scn, tx);
2734 if (scn->scn_suspending)
2738 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
2739 /* First do the MOS & ORIGIN */
2741 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2742 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2743 dsl_scan_visit_rootbp(scn, NULL,
2744 &dp->dp_meta_rootbp, tx);
2745 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
2746 if (scn->scn_suspending)
2749 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
2750 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2751 enqueue_cb, NULL, DS_FIND_CHILDREN));
2753 dsl_scan_visitds(scn,
2754 dp->dp_origin_snap->ds_object, tx);
2756 ASSERT(!scn->scn_suspending);
2757 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
2758 ZB_DESTROYED_OBJSET) {
2759 uint64_t dsobj = scn->scn_phys.scn_bookmark.zb_objset;
2761 * If we were suspended, continue from here. Note if the
2762 * ds we were suspended on was deleted, the zb_objset may
2763 * be -1, so we will skip this and find a new objset
2766 dsl_scan_visitds(scn, dsobj, tx);
2767 if (scn->scn_suspending)
2772 * In case we suspended right at the end of the ds, zero the
2773 * bookmark so we don't think that we're still trying to resume.
2775 memset(&scn->scn_phys.scn_bookmark, 0, sizeof (zbookmark_phys_t));
2778 * Keep pulling things out of the dataset avl queue. Updates to the
2779 * persistent zap-object-as-queue happen only at checkpoints.
2781 while ((sds = avl_first(&scn->scn_queue)) != NULL) {
2783 uint64_t dsobj = sds->sds_dsobj;
2784 uint64_t txg = sds->sds_txg;
2786 /* dequeue and free the ds from the queue */
2787 scan_ds_queue_remove(scn, dsobj);
2790 /* set up min / max txg */
2791 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2793 scn->scn_phys.scn_cur_min_txg =
2794 MAX(scn->scn_phys.scn_min_txg, txg);
2796 scn->scn_phys.scn_cur_min_txg =
2797 MAX(scn->scn_phys.scn_min_txg,
2798 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2800 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
2801 dsl_dataset_rele(ds, FTAG);
2803 dsl_scan_visitds(scn, dsobj, tx);
2804 if (scn->scn_suspending)
2808 /* No more objsets to fetch, we're done */
2809 scn->scn_phys.scn_bookmark.zb_objset = ZB_DESTROYED_OBJSET;
2810 ASSERT0(scn->scn_suspending);
2814 dsl_scan_count_data_disks(vdev_t *rvd)
2816 uint64_t i, leaves = 0;
2818 for (i = 0; i < rvd->vdev_children; i++) {
2819 vdev_t *vd = rvd->vdev_child[i];
2820 if (vd->vdev_islog || vd->vdev_isspare || vd->vdev_isl2cache)
2822 leaves += vdev_get_ndisks(vd) - vdev_get_nparity(vd);
2828 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t *q, const blkptr_t *bp)
2831 uint64_t cur_size = 0;
2833 for (i = 0; i < BP_GET_NDVAS(bp); i++) {
2834 cur_size += DVA_GET_ASIZE(&bp->blk_dva[i]);
2837 q->q_total_zio_size_this_txg += cur_size;
2838 q->q_zios_this_txg++;
2842 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t *q, uint64_t start,
2845 q->q_total_seg_size_this_txg += end - start;
2846 q->q_segs_this_txg++;
2850 scan_io_queue_check_suspend(dsl_scan_t *scn)
2852 /* See comment in dsl_scan_check_suspend() */
2853 uint64_t curr_time_ns = gethrtime();
2854 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
2855 uint64_t sync_time_ns = curr_time_ns -
2856 scn->scn_dp->dp_spa->spa_sync_starttime;
2857 uint64_t dirty_min_bytes = zfs_dirty_data_max *
2858 zfs_vdev_async_write_active_min_dirty_percent / 100;
2859 uint_t mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
2860 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
2862 return ((NSEC2MSEC(scan_time_ns) > mintime &&
2863 (scn->scn_dp->dp_dirty_total >= dirty_min_bytes ||
2864 txg_sync_waiting(scn->scn_dp) ||
2865 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
2866 spa_shutting_down(scn->scn_dp->dp_spa));
2870 * Given a list of scan_io_t's in io_list, this issues the I/Os out to
2871 * disk. This consumes the io_list and frees the scan_io_t's. This is
2872 * called when emptying queues, either when we're up against the memory
2873 * limit or when we have finished scanning. Returns B_TRUE if we stopped
2874 * processing the list before we finished. Any sios that were not issued
2875 * will remain in the io_list.
2878 scan_io_queue_issue(dsl_scan_io_queue_t *queue, list_t *io_list)
2880 dsl_scan_t *scn = queue->q_scn;
2882 boolean_t suspended = B_FALSE;
2884 while ((sio = list_head(io_list)) != NULL) {
2887 if (scan_io_queue_check_suspend(scn)) {
2893 scan_exec_io(scn->scn_dp, &bp, sio->sio_flags,
2894 &sio->sio_zb, queue);
2895 (void) list_remove_head(io_list);
2896 scan_io_queues_update_zio_stats(queue, &bp);
2903 * This function removes sios from an IO queue which reside within a given
2904 * range_seg_t and inserts them (in offset order) into a list. Note that
2905 * we only ever return a maximum of 32 sios at once. If there are more sios
2906 * to process within this segment that did not make it onto the list we
2907 * return B_TRUE and otherwise B_FALSE.
2910 scan_io_queue_gather(dsl_scan_io_queue_t *queue, range_seg_t *rs, list_t *list)
2912 scan_io_t *srch_sio, *sio, *next_sio;
2914 uint_t num_sios = 0;
2915 int64_t bytes_issued = 0;
2918 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2920 srch_sio = sio_alloc(1);
2921 srch_sio->sio_nr_dvas = 1;
2922 SIO_SET_OFFSET(srch_sio, rs_get_start(rs, queue->q_exts_by_addr));
2925 * The exact start of the extent might not contain any matching zios,
2926 * so if that's the case, examine the next one in the tree.
2928 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
2932 sio = avl_nearest(&queue->q_sios_by_addr, idx, AVL_AFTER);
2934 while (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2935 queue->q_exts_by_addr) && num_sios <= 32) {
2936 ASSERT3U(SIO_GET_OFFSET(sio), >=, rs_get_start(rs,
2937 queue->q_exts_by_addr));
2938 ASSERT3U(SIO_GET_END_OFFSET(sio), <=, rs_get_end(rs,
2939 queue->q_exts_by_addr));
2941 next_sio = AVL_NEXT(&queue->q_sios_by_addr, sio);
2942 avl_remove(&queue->q_sios_by_addr, sio);
2943 if (avl_is_empty(&queue->q_sios_by_addr))
2944 atomic_add_64(&queue->q_scn->scn_queues_pending, -1);
2945 queue->q_sio_memused -= SIO_GET_MUSED(sio);
2947 bytes_issued += SIO_GET_ASIZE(sio);
2949 list_insert_tail(list, sio);
2954 * We limit the number of sios we process at once to 32 to avoid
2955 * biting off more than we can chew. If we didn't take everything
2956 * in the segment we update it to reflect the work we were able to
2957 * complete. Otherwise, we remove it from the range tree entirely.
2959 if (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2960 queue->q_exts_by_addr)) {
2961 range_tree_adjust_fill(queue->q_exts_by_addr, rs,
2963 range_tree_resize_segment(queue->q_exts_by_addr, rs,
2964 SIO_GET_OFFSET(sio), rs_get_end(rs,
2965 queue->q_exts_by_addr) - SIO_GET_OFFSET(sio));
2966 queue->q_last_ext_addr = SIO_GET_OFFSET(sio);
2969 uint64_t rstart = rs_get_start(rs, queue->q_exts_by_addr);
2970 uint64_t rend = rs_get_end(rs, queue->q_exts_by_addr);
2971 range_tree_remove(queue->q_exts_by_addr, rstart, rend - rstart);
2972 queue->q_last_ext_addr = -1;
2978 * This is called from the queue emptying thread and selects the next
2979 * extent from which we are to issue I/Os. The behavior of this function
2980 * depends on the state of the scan, the current memory consumption and
2981 * whether or not we are performing a scan shutdown.
2982 * 1) We select extents in an elevator algorithm (LBA-order) if the scan
2983 * needs to perform a checkpoint
2984 * 2) We select the largest available extent if we are up against the
2986 * 3) Otherwise we don't select any extents.
2988 static range_seg_t *
2989 scan_io_queue_fetch_ext(dsl_scan_io_queue_t *queue)
2991 dsl_scan_t *scn = queue->q_scn;
2992 range_tree_t *rt = queue->q_exts_by_addr;
2994 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2995 ASSERT(scn->scn_is_sorted);
2997 if (!scn->scn_checkpointing && !scn->scn_clearing)
3001 * During normal clearing, we want to issue our largest segments
3002 * first, keeping IO as sequential as possible, and leaving the
3003 * smaller extents for later with the hope that they might eventually
3004 * grow to larger sequential segments. However, when the scan is
3005 * checkpointing, no new extents will be added to the sorting queue,
3006 * so the way we are sorted now is as good as it will ever get.
3007 * In this case, we instead switch to issuing extents in LBA order.
3009 if ((zfs_scan_issue_strategy < 1 && scn->scn_checkpointing) ||
3010 zfs_scan_issue_strategy == 1)
3011 return (range_tree_first(rt));
3014 * Try to continue previous extent if it is not completed yet. After
3015 * shrink in scan_io_queue_gather() it may no longer be the best, but
3016 * otherwise we leave shorter remnant every txg.
3019 uint64_t size = 1ULL << rt->rt_shift;
3020 range_seg_t *addr_rs;
3021 if (queue->q_last_ext_addr != -1) {
3022 start = queue->q_last_ext_addr;
3023 addr_rs = range_tree_find(rt, start, size);
3024 if (addr_rs != NULL)
3029 * Nothing to continue, so find new best extent.
3031 uint64_t *v = zfs_btree_first(&queue->q_exts_by_size, NULL);
3034 queue->q_last_ext_addr = start = *v << rt->rt_shift;
3037 * We need to get the original entry in the by_addr tree so we can
3040 addr_rs = range_tree_find(rt, start, size);
3041 ASSERT3P(addr_rs, !=, NULL);
3042 ASSERT3U(rs_get_start(addr_rs, rt), ==, start);
3043 ASSERT3U(rs_get_end(addr_rs, rt), >, start);
3048 scan_io_queues_run_one(void *arg)
3050 dsl_scan_io_queue_t *queue = arg;
3051 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
3052 boolean_t suspended = B_FALSE;
3058 ASSERT(queue->q_scn->scn_is_sorted);
3060 list_create(&sio_list, sizeof (scan_io_t),
3061 offsetof(scan_io_t, sio_nodes.sio_list_node));
3062 zio = zio_null(queue->q_scn->scn_zio_root, queue->q_scn->scn_dp->dp_spa,
3063 NULL, NULL, NULL, ZIO_FLAG_CANFAIL);
3064 mutex_enter(q_lock);
3067 /* Calculate maximum in-flight bytes for this vdev. */
3068 queue->q_maxinflight_bytes = MAX(1, zfs_scan_vdev_limit *
3069 (vdev_get_ndisks(queue->q_vd) - vdev_get_nparity(queue->q_vd)));
3071 /* reset per-queue scan statistics for this txg */
3072 queue->q_total_seg_size_this_txg = 0;
3073 queue->q_segs_this_txg = 0;
3074 queue->q_total_zio_size_this_txg = 0;
3075 queue->q_zios_this_txg = 0;
3077 /* loop until we run out of time or sios */
3078 while ((rs = scan_io_queue_fetch_ext(queue)) != NULL) {
3079 uint64_t seg_start = 0, seg_end = 0;
3080 boolean_t more_left;
3082 ASSERT(list_is_empty(&sio_list));
3084 /* loop while we still have sios left to process in this rs */
3086 scan_io_t *first_sio, *last_sio;
3089 * We have selected which extent needs to be
3090 * processed next. Gather up the corresponding sios.
3092 more_left = scan_io_queue_gather(queue, rs, &sio_list);
3093 ASSERT(!list_is_empty(&sio_list));
3094 first_sio = list_head(&sio_list);
3095 last_sio = list_tail(&sio_list);
3097 seg_end = SIO_GET_END_OFFSET(last_sio);
3099 seg_start = SIO_GET_OFFSET(first_sio);
3102 * Issuing sios can take a long time so drop the
3103 * queue lock. The sio queue won't be updated by
3104 * other threads since we're in syncing context so
3105 * we can be sure that our trees will remain exactly
3109 suspended = scan_io_queue_issue(queue, &sio_list);
3110 mutex_enter(q_lock);
3114 } while (more_left);
3116 /* update statistics for debugging purposes */
3117 scan_io_queues_update_seg_stats(queue, seg_start, seg_end);
3124 * If we were suspended in the middle of processing,
3125 * requeue any unfinished sios and exit.
3127 while ((sio = list_head(&sio_list)) != NULL) {
3128 list_remove(&sio_list, sio);
3129 scan_io_queue_insert_impl(queue, sio);
3132 queue->q_zio = NULL;
3135 list_destroy(&sio_list);
3139 * Performs an emptying run on all scan queues in the pool. This just
3140 * punches out one thread per top-level vdev, each of which processes
3141 * only that vdev's scan queue. We can parallelize the I/O here because
3142 * we know that each queue's I/Os only affect its own top-level vdev.
3144 * This function waits for the queue runs to complete, and must be
3145 * called from dsl_scan_sync (or in general, syncing context).
3148 scan_io_queues_run(dsl_scan_t *scn)
3150 spa_t *spa = scn->scn_dp->dp_spa;
3152 ASSERT(scn->scn_is_sorted);
3153 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
3155 if (scn->scn_queues_pending == 0)
3158 if (scn->scn_taskq == NULL) {
3159 int nthreads = spa->spa_root_vdev->vdev_children;
3162 * We need to make this taskq *always* execute as many
3163 * threads in parallel as we have top-level vdevs and no
3164 * less, otherwise strange serialization of the calls to
3165 * scan_io_queues_run_one can occur during spa_sync runs
3166 * and that significantly impacts performance.
3168 scn->scn_taskq = taskq_create("dsl_scan_iss", nthreads,
3169 minclsyspri, nthreads, nthreads, TASKQ_PREPOPULATE);
3172 for (uint64_t i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3173 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3175 mutex_enter(&vd->vdev_scan_io_queue_lock);
3176 if (vd->vdev_scan_io_queue != NULL) {
3177 VERIFY(taskq_dispatch(scn->scn_taskq,
3178 scan_io_queues_run_one, vd->vdev_scan_io_queue,
3179 TQ_SLEEP) != TASKQID_INVALID);
3181 mutex_exit(&vd->vdev_scan_io_queue_lock);
3185 * Wait for the queues to finish issuing their IOs for this run
3186 * before we return. There may still be IOs in flight at this
3189 taskq_wait(scn->scn_taskq);
3193 dsl_scan_async_block_should_pause(dsl_scan_t *scn)
3195 uint64_t elapsed_nanosecs;
3200 if (zfs_async_block_max_blocks != 0 &&
3201 scn->scn_visited_this_txg >= zfs_async_block_max_blocks) {
3205 if (zfs_max_async_dedup_frees != 0 &&
3206 scn->scn_dedup_frees_this_txg >= zfs_max_async_dedup_frees) {
3210 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
3211 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
3212 (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms &&
3213 txg_sync_waiting(scn->scn_dp)) ||
3214 spa_shutting_down(scn->scn_dp->dp_spa));
3218 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
3220 dsl_scan_t *scn = arg;
3222 if (!scn->scn_is_bptree ||
3223 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
3224 if (dsl_scan_async_block_should_pause(scn))
3225 return (SET_ERROR(ERESTART));
3228 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
3229 dmu_tx_get_txg(tx), bp, 0));
3230 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
3231 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
3232 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
3233 scn->scn_visited_this_txg++;
3234 if (BP_GET_DEDUP(bp))
3235 scn->scn_dedup_frees_this_txg++;
3240 dsl_scan_update_stats(dsl_scan_t *scn)
3242 spa_t *spa = scn->scn_dp->dp_spa;
3244 uint64_t seg_size_total = 0, zio_size_total = 0;
3245 uint64_t seg_count_total = 0, zio_count_total = 0;
3247 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3248 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3249 dsl_scan_io_queue_t *queue = vd->vdev_scan_io_queue;
3254 seg_size_total += queue->q_total_seg_size_this_txg;
3255 zio_size_total += queue->q_total_zio_size_this_txg;
3256 seg_count_total += queue->q_segs_this_txg;
3257 zio_count_total += queue->q_zios_this_txg;
3260 if (seg_count_total == 0 || zio_count_total == 0) {
3261 scn->scn_avg_seg_size_this_txg = 0;
3262 scn->scn_avg_zio_size_this_txg = 0;
3263 scn->scn_segs_this_txg = 0;
3264 scn->scn_zios_this_txg = 0;
3268 scn->scn_avg_seg_size_this_txg = seg_size_total / seg_count_total;
3269 scn->scn_avg_zio_size_this_txg = zio_size_total / zio_count_total;
3270 scn->scn_segs_this_txg = seg_count_total;
3271 scn->scn_zios_this_txg = zio_count_total;
3275 bpobj_dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3279 return (dsl_scan_free_block_cb(arg, bp, tx));
3283 dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3287 dsl_scan_t *scn = arg;
3288 const dva_t *dva = &bp->blk_dva[0];
3290 if (dsl_scan_async_block_should_pause(scn))
3291 return (SET_ERROR(ERESTART));
3293 spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa,
3294 DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva),
3295 DVA_GET_ASIZE(dva), tx);
3296 scn->scn_visited_this_txg++;
3301 dsl_scan_active(dsl_scan_t *scn)
3303 spa_t *spa = scn->scn_dp->dp_spa;
3304 uint64_t used = 0, comp, uncomp;
3305 boolean_t clones_left;
3307 if (spa->spa_load_state != SPA_LOAD_NONE)
3309 if (spa_shutting_down(spa))
3311 if ((dsl_scan_is_running(scn) && !dsl_scan_is_paused_scrub(scn)) ||
3312 (scn->scn_async_destroying && !scn->scn_async_stalled))
3315 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
3316 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
3317 &used, &comp, &uncomp);
3319 clones_left = spa_livelist_delete_check(spa);
3320 return ((used != 0) || (clones_left));
3324 dsl_scan_check_deferred(vdev_t *vd)
3326 boolean_t need_resilver = B_FALSE;
3328 for (int c = 0; c < vd->vdev_children; c++) {
3330 dsl_scan_check_deferred(vd->vdev_child[c]);
3333 if (!vdev_is_concrete(vd) || vd->vdev_aux ||
3334 !vd->vdev_ops->vdev_op_leaf)
3335 return (need_resilver);
3337 if (!vd->vdev_resilver_deferred)
3338 need_resilver = B_TRUE;
3340 return (need_resilver);
3344 dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize,
3345 uint64_t phys_birth)
3349 vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
3351 if (vd->vdev_ops == &vdev_indirect_ops) {
3353 * The indirect vdev can point to multiple
3354 * vdevs. For simplicity, always create
3355 * the resilver zio_t. zio_vdev_io_start()
3356 * will bypass the child resilver i/o's if
3357 * they are on vdevs that don't have DTL's.
3362 if (DVA_GET_GANG(dva)) {
3364 * Gang members may be spread across multiple
3365 * vdevs, so the best estimate we have is the
3366 * scrub range, which has already been checked.
3367 * XXX -- it would be better to change our
3368 * allocation policy to ensure that all
3369 * gang members reside on the same vdev.
3375 * Check if the top-level vdev must resilver this offset.
3376 * When the offset does not intersect with a dirty leaf DTL
3377 * then it may be possible to skip the resilver IO. The psize
3378 * is provided instead of asize to simplify the check for RAIDZ.
3380 if (!vdev_dtl_need_resilver(vd, dva, psize, phys_birth))
3384 * Check that this top-level vdev has a device under it which
3385 * is resilvering and is not deferred.
3387 if (!dsl_scan_check_deferred(vd))
3394 dsl_process_async_destroys(dsl_pool_t *dp, dmu_tx_t *tx)
3396 dsl_scan_t *scn = dp->dp_scan;
3397 spa_t *spa = dp->dp_spa;
3400 if (spa_suspend_async_destroy(spa))
3403 if (zfs_free_bpobj_enabled &&
3404 spa_version(spa) >= SPA_VERSION_DEADLISTS) {
3405 scn->scn_is_bptree = B_FALSE;
3406 scn->scn_async_block_min_time_ms = zfs_free_min_time_ms;
3407 scn->scn_zio_root = zio_root(spa, NULL,
3408 NULL, ZIO_FLAG_MUSTSUCCEED);
3409 err = bpobj_iterate(&dp->dp_free_bpobj,
3410 bpobj_dsl_scan_free_block_cb, scn, tx);
3411 VERIFY0(zio_wait(scn->scn_zio_root));
3412 scn->scn_zio_root = NULL;
3414 if (err != 0 && err != ERESTART)
3415 zfs_panic_recover("error %u from bpobj_iterate()", err);
3418 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
3419 ASSERT(scn->scn_async_destroying);
3420 scn->scn_is_bptree = B_TRUE;
3421 scn->scn_zio_root = zio_root(spa, NULL,
3422 NULL, ZIO_FLAG_MUSTSUCCEED);
3423 err = bptree_iterate(dp->dp_meta_objset,
3424 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
3425 VERIFY0(zio_wait(scn->scn_zio_root));
3426 scn->scn_zio_root = NULL;
3428 if (err == EIO || err == ECKSUM) {
3430 } else if (err != 0 && err != ERESTART) {
3431 zfs_panic_recover("error %u from "
3432 "traverse_dataset_destroyed()", err);
3435 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
3436 /* finished; deactivate async destroy feature */
3437 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
3438 ASSERT(!spa_feature_is_active(spa,
3439 SPA_FEATURE_ASYNC_DESTROY));
3440 VERIFY0(zap_remove(dp->dp_meta_objset,
3441 DMU_POOL_DIRECTORY_OBJECT,
3442 DMU_POOL_BPTREE_OBJ, tx));
3443 VERIFY0(bptree_free(dp->dp_meta_objset,
3444 dp->dp_bptree_obj, tx));
3445 dp->dp_bptree_obj = 0;
3446 scn->scn_async_destroying = B_FALSE;
3447 scn->scn_async_stalled = B_FALSE;
3450 * If we didn't make progress, mark the async
3451 * destroy as stalled, so that we will not initiate
3452 * a spa_sync() on its behalf. Note that we only
3453 * check this if we are not finished, because if the
3454 * bptree had no blocks for us to visit, we can
3455 * finish without "making progress".
3457 scn->scn_async_stalled =
3458 (scn->scn_visited_this_txg == 0);
3461 if (scn->scn_visited_this_txg) {
3462 zfs_dbgmsg("freed %llu blocks in %llums from "
3463 "free_bpobj/bptree on %s in txg %llu; err=%u",
3464 (longlong_t)scn->scn_visited_this_txg,
3466 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
3467 spa->spa_name, (longlong_t)tx->tx_txg, err);
3468 scn->scn_visited_this_txg = 0;
3469 scn->scn_dedup_frees_this_txg = 0;
3472 * Write out changes to the DDT that may be required as a
3473 * result of the blocks freed. This ensures that the DDT
3474 * is clean when a scrub/resilver runs.
3476 ddt_sync(spa, tx->tx_txg);
3480 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3481 zfs_free_leak_on_eio &&
3482 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
3483 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
3484 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
3486 * We have finished background destroying, but there is still
3487 * some space left in the dp_free_dir. Transfer this leaked
3488 * space to the dp_leak_dir.
3490 if (dp->dp_leak_dir == NULL) {
3491 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
3492 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
3494 VERIFY0(dsl_pool_open_special_dir(dp,
3495 LEAK_DIR_NAME, &dp->dp_leak_dir));
3496 rrw_exit(&dp->dp_config_rwlock, FTAG);
3498 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
3499 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3500 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3501 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3502 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
3503 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3504 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3505 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3508 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3509 !spa_livelist_delete_check(spa)) {
3510 /* finished; verify that space accounting went to zero */
3511 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
3512 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
3513 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
3516 spa_notify_waiters(spa);
3518 EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj),
3519 0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3520 DMU_POOL_OBSOLETE_BPOBJ));
3521 if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) {
3522 ASSERT(spa_feature_is_active(dp->dp_spa,
3523 SPA_FEATURE_OBSOLETE_COUNTS));
3525 scn->scn_is_bptree = B_FALSE;
3526 scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms;
3527 err = bpobj_iterate(&dp->dp_obsolete_bpobj,
3528 dsl_scan_obsolete_block_cb, scn, tx);
3529 if (err != 0 && err != ERESTART)
3530 zfs_panic_recover("error %u from bpobj_iterate()", err);
3532 if (bpobj_is_empty(&dp->dp_obsolete_bpobj))
3533 dsl_pool_destroy_obsolete_bpobj(dp, tx);
3539 * This is the primary entry point for scans that is called from syncing
3540 * context. Scans must happen entirely during syncing context so that we
3541 * can guarantee that blocks we are currently scanning will not change out
3542 * from under us. While a scan is active, this function controls how quickly
3543 * transaction groups proceed, instead of the normal handling provided by
3544 * txg_sync_thread().
3547 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
3550 dsl_scan_t *scn = dp->dp_scan;
3551 spa_t *spa = dp->dp_spa;
3552 state_sync_type_t sync_type = SYNC_OPTIONAL;
3554 if (spa->spa_resilver_deferred &&
3555 !spa_feature_is_active(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
3556 spa_feature_incr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
3559 * Check for scn_restart_txg before checking spa_load_state, so
3560 * that we can restart an old-style scan while the pool is being
3561 * imported (see dsl_scan_init). We also restart scans if there
3562 * is a deferred resilver and the user has manually disabled
3563 * deferred resilvers via the tunable.
3565 if (dsl_scan_restarting(scn, tx) ||
3566 (spa->spa_resilver_deferred && zfs_resilver_disable_defer)) {
3567 pool_scan_func_t func = POOL_SCAN_SCRUB;
3568 dsl_scan_done(scn, B_FALSE, tx);
3569 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
3570 func = POOL_SCAN_RESILVER;
3571 zfs_dbgmsg("restarting scan func=%u on %s txg=%llu",
3572 func, dp->dp_spa->spa_name, (longlong_t)tx->tx_txg);
3573 dsl_scan_setup_sync(&func, tx);
3577 * Only process scans in sync pass 1.
3579 if (spa_sync_pass(spa) > 1)
3583 * If the spa is shutting down, then stop scanning. This will
3584 * ensure that the scan does not dirty any new data during the
3587 if (spa_shutting_down(spa))
3591 * If the scan is inactive due to a stalled async destroy, try again.
3593 if (!scn->scn_async_stalled && !dsl_scan_active(scn))
3596 /* reset scan statistics */
3597 scn->scn_visited_this_txg = 0;
3598 scn->scn_dedup_frees_this_txg = 0;
3599 scn->scn_holes_this_txg = 0;
3600 scn->scn_lt_min_this_txg = 0;
3601 scn->scn_gt_max_this_txg = 0;
3602 scn->scn_ddt_contained_this_txg = 0;
3603 scn->scn_objsets_visited_this_txg = 0;
3604 scn->scn_avg_seg_size_this_txg = 0;
3605 scn->scn_segs_this_txg = 0;
3606 scn->scn_avg_zio_size_this_txg = 0;
3607 scn->scn_zios_this_txg = 0;
3608 scn->scn_suspending = B_FALSE;
3609 scn->scn_sync_start_time = gethrtime();
3610 spa->spa_scrub_active = B_TRUE;
3613 * First process the async destroys. If we suspend, don't do
3614 * any scrubbing or resilvering. This ensures that there are no
3615 * async destroys while we are scanning, so the scan code doesn't
3616 * have to worry about traversing it. It is also faster to free the
3617 * blocks than to scrub them.
3619 err = dsl_process_async_destroys(dp, tx);
3623 if (!dsl_scan_is_running(scn) || dsl_scan_is_paused_scrub(scn))
3627 * Wait a few txgs after importing to begin scanning so that
3628 * we can get the pool imported quickly.
3630 if (spa->spa_syncing_txg < spa->spa_first_txg + SCAN_IMPORT_WAIT_TXGS)
3634 * zfs_scan_suspend_progress can be set to disable scan progress.
3635 * We don't want to spin the txg_sync thread, so we add a delay
3636 * here to simulate the time spent doing a scan. This is mostly
3637 * useful for testing and debugging.
3639 if (zfs_scan_suspend_progress) {
3640 uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time;
3641 uint_t mintime = (scn->scn_phys.scn_func ==
3642 POOL_SCAN_RESILVER) ? zfs_resilver_min_time_ms :
3643 zfs_scrub_min_time_ms;
3645 while (zfs_scan_suspend_progress &&
3646 !txg_sync_waiting(scn->scn_dp) &&
3647 !spa_shutting_down(scn->scn_dp->dp_spa) &&
3648 NSEC2MSEC(scan_time_ns) < mintime) {
3650 scan_time_ns = gethrtime() - scn->scn_sync_start_time;
3656 * It is possible to switch from unsorted to sorted at any time,
3657 * but afterwards the scan will remain sorted unless reloaded from
3658 * a checkpoint after a reboot.
3660 if (!zfs_scan_legacy) {
3661 scn->scn_is_sorted = B_TRUE;
3662 if (scn->scn_last_checkpoint == 0)
3663 scn->scn_last_checkpoint = ddi_get_lbolt();
3667 * For sorted scans, determine what kind of work we will be doing
3668 * this txg based on our memory limitations and whether or not we
3669 * need to perform a checkpoint.
3671 if (scn->scn_is_sorted) {
3673 * If we are over our checkpoint interval, set scn_clearing
3674 * so that we can begin checkpointing immediately. The
3675 * checkpoint allows us to save a consistent bookmark
3676 * representing how much data we have scrubbed so far.
3677 * Otherwise, use the memory limit to determine if we should
3678 * scan for metadata or start issue scrub IOs. We accumulate
3679 * metadata until we hit our hard memory limit at which point
3680 * we issue scrub IOs until we are at our soft memory limit.
3682 if (scn->scn_checkpointing ||
3683 ddi_get_lbolt() - scn->scn_last_checkpoint >
3684 SEC_TO_TICK(zfs_scan_checkpoint_intval)) {
3685 if (!scn->scn_checkpointing)
3686 zfs_dbgmsg("begin scan checkpoint for %s",
3689 scn->scn_checkpointing = B_TRUE;
3690 scn->scn_clearing = B_TRUE;
3692 boolean_t should_clear = dsl_scan_should_clear(scn);
3693 if (should_clear && !scn->scn_clearing) {
3694 zfs_dbgmsg("begin scan clearing for %s",
3696 scn->scn_clearing = B_TRUE;
3697 } else if (!should_clear && scn->scn_clearing) {
3698 zfs_dbgmsg("finish scan clearing for %s",
3700 scn->scn_clearing = B_FALSE;
3704 ASSERT0(scn->scn_checkpointing);
3705 ASSERT0(scn->scn_clearing);
3708 if (!scn->scn_clearing && scn->scn_done_txg == 0) {
3709 /* Need to scan metadata for more blocks to scrub */
3710 dsl_scan_phys_t *scnp = &scn->scn_phys;
3711 taskqid_t prefetch_tqid;
3714 * Recalculate the max number of in-flight bytes for pool-wide
3715 * scanning operations (minimum 1MB). Limits for the issuing
3716 * phase are done per top-level vdev and are handled separately.
3718 scn->scn_maxinflight_bytes = MAX(zfs_scan_vdev_limit *
3719 dsl_scan_count_data_disks(spa->spa_root_vdev), 1ULL << 20);
3721 if (scnp->scn_ddt_bookmark.ddb_class <=
3722 scnp->scn_ddt_class_max) {
3723 ASSERT(ZB_IS_ZERO(&scnp->scn_bookmark));
3724 zfs_dbgmsg("doing scan sync for %s txg %llu; "
3725 "ddt bm=%llu/%llu/%llu/%llx",
3727 (longlong_t)tx->tx_txg,
3728 (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
3729 (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
3730 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
3731 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
3733 zfs_dbgmsg("doing scan sync for %s txg %llu; "
3734 "bm=%llu/%llu/%llu/%llu",
3736 (longlong_t)tx->tx_txg,
3737 (longlong_t)scnp->scn_bookmark.zb_objset,
3738 (longlong_t)scnp->scn_bookmark.zb_object,
3739 (longlong_t)scnp->scn_bookmark.zb_level,
3740 (longlong_t)scnp->scn_bookmark.zb_blkid);
3743 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3744 NULL, ZIO_FLAG_CANFAIL);
3746 scn->scn_prefetch_stop = B_FALSE;
3747 prefetch_tqid = taskq_dispatch(dp->dp_sync_taskq,
3748 dsl_scan_prefetch_thread, scn, TQ_SLEEP);
3749 ASSERT(prefetch_tqid != TASKQID_INVALID);
3751 dsl_pool_config_enter(dp, FTAG);
3752 dsl_scan_visit(scn, tx);
3753 dsl_pool_config_exit(dp, FTAG);
3755 mutex_enter(&dp->dp_spa->spa_scrub_lock);
3756 scn->scn_prefetch_stop = B_TRUE;
3757 cv_broadcast(&spa->spa_scrub_io_cv);
3758 mutex_exit(&dp->dp_spa->spa_scrub_lock);
3760 taskq_wait_id(dp->dp_sync_taskq, prefetch_tqid);
3761 (void) zio_wait(scn->scn_zio_root);
3762 scn->scn_zio_root = NULL;
3764 zfs_dbgmsg("scan visited %llu blocks of %s in %llums "
3765 "(%llu os's, %llu holes, %llu < mintxg, "
3766 "%llu in ddt, %llu > maxtxg)",
3767 (longlong_t)scn->scn_visited_this_txg,
3769 (longlong_t)NSEC2MSEC(gethrtime() -
3770 scn->scn_sync_start_time),
3771 (longlong_t)scn->scn_objsets_visited_this_txg,
3772 (longlong_t)scn->scn_holes_this_txg,
3773 (longlong_t)scn->scn_lt_min_this_txg,
3774 (longlong_t)scn->scn_ddt_contained_this_txg,
3775 (longlong_t)scn->scn_gt_max_this_txg);
3777 if (!scn->scn_suspending) {
3778 ASSERT0(avl_numnodes(&scn->scn_queue));
3779 scn->scn_done_txg = tx->tx_txg + 1;
3780 if (scn->scn_is_sorted) {
3781 scn->scn_checkpointing = B_TRUE;
3782 scn->scn_clearing = B_TRUE;
3784 zfs_dbgmsg("scan complete for %s txg %llu",
3786 (longlong_t)tx->tx_txg);
3788 } else if (scn->scn_is_sorted && scn->scn_queues_pending != 0) {
3789 ASSERT(scn->scn_clearing);
3791 /* need to issue scrubbing IOs from per-vdev queues */
3792 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3793 NULL, ZIO_FLAG_CANFAIL);
3794 scan_io_queues_run(scn);
3795 (void) zio_wait(scn->scn_zio_root);
3796 scn->scn_zio_root = NULL;
3798 /* calculate and dprintf the current memory usage */
3799 (void) dsl_scan_should_clear(scn);
3800 dsl_scan_update_stats(scn);
3802 zfs_dbgmsg("scan issued %llu blocks for %s (%llu segs) "
3803 "in %llums (avg_block_size = %llu, avg_seg_size = %llu)",
3804 (longlong_t)scn->scn_zios_this_txg,
3806 (longlong_t)scn->scn_segs_this_txg,
3807 (longlong_t)NSEC2MSEC(gethrtime() -
3808 scn->scn_sync_start_time),
3809 (longlong_t)scn->scn_avg_zio_size_this_txg,
3810 (longlong_t)scn->scn_avg_seg_size_this_txg);
3811 } else if (scn->scn_done_txg != 0 && scn->scn_done_txg <= tx->tx_txg) {
3812 /* Finished with everything. Mark the scrub as complete */
3813 zfs_dbgmsg("scan issuing complete txg %llu for %s",
3814 (longlong_t)tx->tx_txg,
3816 ASSERT3U(scn->scn_done_txg, !=, 0);
3817 ASSERT0(spa->spa_scrub_inflight);
3818 ASSERT0(scn->scn_queues_pending);
3819 dsl_scan_done(scn, B_TRUE, tx);
3820 sync_type = SYNC_MANDATORY;
3823 dsl_scan_sync_state(scn, tx, sync_type);
3827 count_block_issued(spa_t *spa, const blkptr_t *bp, boolean_t all)
3830 * Don't count embedded bp's, since we already did the work of
3831 * scanning these when we scanned the containing block.
3833 if (BP_IS_EMBEDDED(bp))
3837 * Update the spa's stats on how many bytes we have issued.
3838 * Sequential scrubs create a zio for each DVA of the bp. Each
3839 * of these will include all DVAs for repair purposes, but the
3840 * zio code will only try the first one unless there is an issue.
3841 * Therefore, we should only count the first DVA for these IOs.
3843 atomic_add_64(&spa->spa_scan_pass_issued,
3844 all ? BP_GET_ASIZE(bp) : DVA_GET_ASIZE(&bp->blk_dva[0]));
3848 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
3851 * If we resume after a reboot, zab will be NULL; don't record
3852 * incomplete stats in that case.
3857 for (int i = 0; i < 4; i++) {
3858 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
3859 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
3861 if (t & DMU_OT_NEWTYPE)
3863 zfs_blkstat_t *zb = &zab->zab_type[l][t];
3867 zb->zb_asize += BP_GET_ASIZE(bp);
3868 zb->zb_lsize += BP_GET_LSIZE(bp);
3869 zb->zb_psize += BP_GET_PSIZE(bp);
3870 zb->zb_gangs += BP_COUNT_GANG(bp);
3872 switch (BP_GET_NDVAS(bp)) {
3874 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3875 DVA_GET_VDEV(&bp->blk_dva[1]))
3876 zb->zb_ditto_2_of_2_samevdev++;
3879 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3880 DVA_GET_VDEV(&bp->blk_dva[1])) +
3881 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3882 DVA_GET_VDEV(&bp->blk_dva[2])) +
3883 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
3884 DVA_GET_VDEV(&bp->blk_dva[2]));
3886 zb->zb_ditto_2_of_3_samevdev++;
3887 else if (equal == 3)
3888 zb->zb_ditto_3_of_3_samevdev++;
3895 scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue, scan_io_t *sio)
3898 dsl_scan_t *scn = queue->q_scn;
3900 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3902 if (unlikely(avl_is_empty(&queue->q_sios_by_addr)))
3903 atomic_add_64(&scn->scn_queues_pending, 1);
3904 if (avl_find(&queue->q_sios_by_addr, sio, &idx) != NULL) {
3905 /* block is already scheduled for reading */
3909 avl_insert(&queue->q_sios_by_addr, sio, idx);
3910 queue->q_sio_memused += SIO_GET_MUSED(sio);
3911 range_tree_add(queue->q_exts_by_addr, SIO_GET_OFFSET(sio),
3912 SIO_GET_ASIZE(sio));
3916 * Given all the info we got from our metadata scanning process, we
3917 * construct a scan_io_t and insert it into the scan sorting queue. The
3918 * I/O must already be suitable for us to process. This is controlled
3919 * by dsl_scan_enqueue().
3922 scan_io_queue_insert(dsl_scan_io_queue_t *queue, const blkptr_t *bp, int dva_i,
3923 int zio_flags, const zbookmark_phys_t *zb)
3925 scan_io_t *sio = sio_alloc(BP_GET_NDVAS(bp));
3927 ASSERT0(BP_IS_GANG(bp));
3928 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3930 bp2sio(bp, sio, dva_i);
3931 sio->sio_flags = zio_flags;
3934 queue->q_last_ext_addr = -1;
3935 scan_io_queue_insert_impl(queue, sio);
3939 * Given a set of I/O parameters as discovered by the metadata traversal
3940 * process, attempts to place the I/O into the sorted queues (if allowed),
3941 * or immediately executes the I/O.
3944 dsl_scan_enqueue(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
3945 const zbookmark_phys_t *zb)
3947 spa_t *spa = dp->dp_spa;
3949 ASSERT(!BP_IS_EMBEDDED(bp));
3952 * Gang blocks are hard to issue sequentially, so we just issue them
3953 * here immediately instead of queuing them.
3955 if (!dp->dp_scan->scn_is_sorted || BP_IS_GANG(bp)) {
3956 scan_exec_io(dp, bp, zio_flags, zb, NULL);
3960 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
3964 dva = bp->blk_dva[i];
3965 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&dva));
3966 ASSERT(vdev != NULL);
3968 mutex_enter(&vdev->vdev_scan_io_queue_lock);
3969 if (vdev->vdev_scan_io_queue == NULL)
3970 vdev->vdev_scan_io_queue = scan_io_queue_create(vdev);
3971 ASSERT(dp->dp_scan != NULL);
3972 scan_io_queue_insert(vdev->vdev_scan_io_queue, bp,
3974 mutex_exit(&vdev->vdev_scan_io_queue_lock);
3979 dsl_scan_scrub_cb(dsl_pool_t *dp,
3980 const blkptr_t *bp, const zbookmark_phys_t *zb)
3982 dsl_scan_t *scn = dp->dp_scan;
3983 spa_t *spa = dp->dp_spa;
3984 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
3985 size_t psize = BP_GET_PSIZE(bp);
3986 boolean_t needs_io = B_FALSE;
3987 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
3989 count_block(dp->dp_blkstats, bp);
3990 if (phys_birth <= scn->scn_phys.scn_min_txg ||
3991 phys_birth >= scn->scn_phys.scn_max_txg) {
3992 count_block_issued(spa, bp, B_TRUE);
3996 /* Embedded BP's have phys_birth==0, so we reject them above. */
3997 ASSERT(!BP_IS_EMBEDDED(bp));
3999 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
4000 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
4001 zio_flags |= ZIO_FLAG_SCRUB;
4004 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
4005 zio_flags |= ZIO_FLAG_RESILVER;
4009 /* If it's an intent log block, failure is expected. */
4010 if (zb->zb_level == ZB_ZIL_LEVEL)
4011 zio_flags |= ZIO_FLAG_SPECULATIVE;
4013 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
4014 const dva_t *dva = &bp->blk_dva[d];
4017 * Keep track of how much data we've examined so that
4018 * zpool(8) status can make useful progress reports.
4020 uint64_t asize = DVA_GET_ASIZE(dva);
4021 scn->scn_phys.scn_examined += asize;
4022 spa->spa_scan_pass_exam += asize;
4024 /* if it's a resilver, this may not be in the target range */
4026 needs_io = dsl_scan_need_resilver(spa, dva, psize,
4030 if (needs_io && !zfs_no_scrub_io) {
4031 dsl_scan_enqueue(dp, bp, zio_flags, zb);
4033 count_block_issued(spa, bp, B_TRUE);
4036 /* do not relocate this block */
4041 dsl_scan_scrub_done(zio_t *zio)
4043 spa_t *spa = zio->io_spa;
4044 blkptr_t *bp = zio->io_bp;
4045 dsl_scan_io_queue_t *queue = zio->io_private;
4047 abd_free(zio->io_abd);
4049 if (queue == NULL) {
4050 mutex_enter(&spa->spa_scrub_lock);
4051 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
4052 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
4053 cv_broadcast(&spa->spa_scrub_io_cv);
4054 mutex_exit(&spa->spa_scrub_lock);
4056 mutex_enter(&queue->q_vd->vdev_scan_io_queue_lock);
4057 ASSERT3U(queue->q_inflight_bytes, >=, BP_GET_PSIZE(bp));
4058 queue->q_inflight_bytes -= BP_GET_PSIZE(bp);
4059 cv_broadcast(&queue->q_zio_cv);
4060 mutex_exit(&queue->q_vd->vdev_scan_io_queue_lock);
4063 if (zio->io_error && (zio->io_error != ECKSUM ||
4064 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
4065 atomic_inc_64(&spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors);
4070 * Given a scanning zio's information, executes the zio. The zio need
4071 * not necessarily be only sortable, this function simply executes the
4072 * zio, no matter what it is. The optional queue argument allows the
4073 * caller to specify that they want per top level vdev IO rate limiting
4074 * instead of the legacy global limiting.
4077 scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
4078 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue)
4080 spa_t *spa = dp->dp_spa;
4081 dsl_scan_t *scn = dp->dp_scan;
4082 size_t size = BP_GET_PSIZE(bp);
4083 abd_t *data = abd_alloc_for_io(size, B_FALSE);
4086 if (queue == NULL) {
4087 ASSERT3U(scn->scn_maxinflight_bytes, >, 0);
4088 mutex_enter(&spa->spa_scrub_lock);
4089 while (spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)
4090 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
4091 spa->spa_scrub_inflight += BP_GET_PSIZE(bp);
4092 mutex_exit(&spa->spa_scrub_lock);
4093 pio = scn->scn_zio_root;
4095 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
4097 ASSERT3U(queue->q_maxinflight_bytes, >, 0);
4098 mutex_enter(q_lock);
4099 while (queue->q_inflight_bytes >= queue->q_maxinflight_bytes)
4100 cv_wait(&queue->q_zio_cv, q_lock);
4101 queue->q_inflight_bytes += BP_GET_PSIZE(bp);
4106 ASSERT(pio != NULL);
4107 count_block_issued(spa, bp, queue == NULL);
4108 zio_nowait(zio_read(pio, spa, bp, data, size, dsl_scan_scrub_done,
4109 queue, ZIO_PRIORITY_SCRUB, zio_flags, zb));
4113 * This is the primary extent sorting algorithm. We balance two parameters:
4114 * 1) how many bytes of I/O are in an extent
4115 * 2) how well the extent is filled with I/O (as a fraction of its total size)
4116 * Since we allow extents to have gaps between their constituent I/Os, it's
4117 * possible to have a fairly large extent that contains the same amount of
4118 * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
4119 * The algorithm sorts based on a score calculated from the extent's size,
4120 * the relative fill volume (in %) and a "fill weight" parameter that controls
4121 * the split between whether we prefer larger extents or more well populated
4124 * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
4127 * 1) assume extsz = 64 MiB
4128 * 2) assume fill = 32 MiB (extent is half full)
4129 * 3) assume fill_weight = 3
4130 * 4) SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
4131 * SCORE = 32M + (50 * 3 * 32M) / 100
4132 * SCORE = 32M + (4800M / 100)
4135 * | +--- final total relative fill-based score
4136 * +--------- final total fill-based score
4139 * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
4140 * extents that are more completely filled (in a 3:2 ratio) vs just larger.
4141 * Note that as an optimization, we replace multiplication and division by
4142 * 100 with bitshifting by 7 (which effectively multiplies and divides by 128).
4144 * Since we do not care if one extent is only few percent better than another,
4145 * compress the score into 6 bits via binary logarithm AKA highbit64() and
4146 * put into otherwise unused due to ashift high bits of offset. This allows
4147 * to reduce q_exts_by_size B-tree elements to only 64 bits and compare them
4148 * with single operation. Plus it makes scrubs more sequential and reduces
4149 * chances that minor extent change move it within the B-tree.
4152 ext_size_compare(const void *x, const void *y)
4154 const uint64_t *a = x, *b = y;
4156 return (TREE_CMP(*a, *b));
4160 ext_size_create(range_tree_t *rt, void *arg)
4163 zfs_btree_t *size_tree = arg;
4165 zfs_btree_create(size_tree, ext_size_compare, sizeof (uint64_t));
4169 ext_size_destroy(range_tree_t *rt, void *arg)
4172 zfs_btree_t *size_tree = arg;
4173 ASSERT0(zfs_btree_numnodes(size_tree));
4175 zfs_btree_destroy(size_tree);
4179 ext_size_value(range_tree_t *rt, range_seg_gap_t *rsg)
4182 uint64_t size = rsg->rs_end - rsg->rs_start;
4183 uint64_t score = rsg->rs_fill + ((((rsg->rs_fill << 7) / size) *
4184 fill_weight * rsg->rs_fill) >> 7);
4185 ASSERT3U(rt->rt_shift, >=, 8);
4186 return (((uint64_t)(64 - highbit64(score)) << 56) | rsg->rs_start);
4190 ext_size_add(range_tree_t *rt, range_seg_t *rs, void *arg)
4192 zfs_btree_t *size_tree = arg;
4193 ASSERT3U(rt->rt_type, ==, RANGE_SEG_GAP);
4194 uint64_t v = ext_size_value(rt, (range_seg_gap_t *)rs);
4195 zfs_btree_add(size_tree, &v);
4199 ext_size_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
4201 zfs_btree_t *size_tree = arg;
4202 ASSERT3U(rt->rt_type, ==, RANGE_SEG_GAP);
4203 uint64_t v = ext_size_value(rt, (range_seg_gap_t *)rs);
4204 zfs_btree_remove(size_tree, &v);
4208 ext_size_vacate(range_tree_t *rt, void *arg)
4210 zfs_btree_t *size_tree = arg;
4211 zfs_btree_clear(size_tree);
4212 zfs_btree_destroy(size_tree);
4214 ext_size_create(rt, arg);
4217 static const range_tree_ops_t ext_size_ops = {
4218 .rtop_create = ext_size_create,
4219 .rtop_destroy = ext_size_destroy,
4220 .rtop_add = ext_size_add,
4221 .rtop_remove = ext_size_remove,
4222 .rtop_vacate = ext_size_vacate
4226 * Comparator for the q_sios_by_addr tree. Sorting is simply performed
4227 * based on LBA-order (from lowest to highest).
4230 sio_addr_compare(const void *x, const void *y)
4232 const scan_io_t *a = x, *b = y;
4234 return (TREE_CMP(SIO_GET_OFFSET(a), SIO_GET_OFFSET(b)));
4237 /* IO queues are created on demand when they are needed. */
4238 static dsl_scan_io_queue_t *
4239 scan_io_queue_create(vdev_t *vd)
4241 dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan;
4242 dsl_scan_io_queue_t *q = kmem_zalloc(sizeof (*q), KM_SLEEP);
4246 q->q_sio_memused = 0;
4247 q->q_last_ext_addr = -1;
4248 cv_init(&q->q_zio_cv, NULL, CV_DEFAULT, NULL);
4249 q->q_exts_by_addr = range_tree_create_gap(&ext_size_ops, RANGE_SEG_GAP,
4250 &q->q_exts_by_size, 0, vd->vdev_ashift, zfs_scan_max_ext_gap);
4251 avl_create(&q->q_sios_by_addr, sio_addr_compare,
4252 sizeof (scan_io_t), offsetof(scan_io_t, sio_nodes.sio_addr_node));
4258 * Destroys a scan queue and all segments and scan_io_t's contained in it.
4259 * No further execution of I/O occurs, anything pending in the queue is
4260 * simply freed without being executed.
4263 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t *queue)
4265 dsl_scan_t *scn = queue->q_scn;
4267 void *cookie = NULL;
4269 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
4271 if (!avl_is_empty(&queue->q_sios_by_addr))
4272 atomic_add_64(&scn->scn_queues_pending, -1);
4273 while ((sio = avl_destroy_nodes(&queue->q_sios_by_addr, &cookie)) !=
4275 ASSERT(range_tree_contains(queue->q_exts_by_addr,
4276 SIO_GET_OFFSET(sio), SIO_GET_ASIZE(sio)));
4277 queue->q_sio_memused -= SIO_GET_MUSED(sio);
4281 ASSERT0(queue->q_sio_memused);
4282 range_tree_vacate(queue->q_exts_by_addr, NULL, queue);
4283 range_tree_destroy(queue->q_exts_by_addr);
4284 avl_destroy(&queue->q_sios_by_addr);
4285 cv_destroy(&queue->q_zio_cv);
4287 kmem_free(queue, sizeof (*queue));
4291 * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
4292 * called on behalf of vdev_top_transfer when creating or destroying
4293 * a mirror vdev due to zpool attach/detach.
4296 dsl_scan_io_queue_vdev_xfer(vdev_t *svd, vdev_t *tvd)
4298 mutex_enter(&svd->vdev_scan_io_queue_lock);
4299 mutex_enter(&tvd->vdev_scan_io_queue_lock);
4301 VERIFY3P(tvd->vdev_scan_io_queue, ==, NULL);
4302 tvd->vdev_scan_io_queue = svd->vdev_scan_io_queue;
4303 svd->vdev_scan_io_queue = NULL;
4304 if (tvd->vdev_scan_io_queue != NULL)
4305 tvd->vdev_scan_io_queue->q_vd = tvd;
4307 mutex_exit(&tvd->vdev_scan_io_queue_lock);
4308 mutex_exit(&svd->vdev_scan_io_queue_lock);
4312 scan_io_queues_destroy(dsl_scan_t *scn)
4314 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
4316 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
4317 vdev_t *tvd = rvd->vdev_child[i];
4319 mutex_enter(&tvd->vdev_scan_io_queue_lock);
4320 if (tvd->vdev_scan_io_queue != NULL)
4321 dsl_scan_io_queue_destroy(tvd->vdev_scan_io_queue);
4322 tvd->vdev_scan_io_queue = NULL;
4323 mutex_exit(&tvd->vdev_scan_io_queue_lock);
4328 dsl_scan_freed_dva(spa_t *spa, const blkptr_t *bp, int dva_i)
4330 dsl_pool_t *dp = spa->spa_dsl_pool;
4331 dsl_scan_t *scn = dp->dp_scan;
4334 dsl_scan_io_queue_t *queue;
4335 scan_io_t *srch_sio, *sio;
4337 uint64_t start, size;
4339 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&bp->blk_dva[dva_i]));
4340 ASSERT(vdev != NULL);
4341 q_lock = &vdev->vdev_scan_io_queue_lock;
4342 queue = vdev->vdev_scan_io_queue;
4344 mutex_enter(q_lock);
4345 if (queue == NULL) {
4350 srch_sio = sio_alloc(BP_GET_NDVAS(bp));
4351 bp2sio(bp, srch_sio, dva_i);
4352 start = SIO_GET_OFFSET(srch_sio);
4353 size = SIO_GET_ASIZE(srch_sio);
4356 * We can find the zio in two states:
4357 * 1) Cold, just sitting in the queue of zio's to be issued at
4358 * some point in the future. In this case, all we do is
4359 * remove the zio from the q_sios_by_addr tree, decrement
4360 * its data volume from the containing range_seg_t and
4361 * resort the q_exts_by_size tree to reflect that the
4362 * range_seg_t has lost some of its 'fill'. We don't shorten
4363 * the range_seg_t - this is usually rare enough not to be
4364 * worth the extra hassle of trying keep track of precise
4365 * extent boundaries.
4366 * 2) Hot, where the zio is currently in-flight in
4367 * dsl_scan_issue_ios. In this case, we can't simply
4368 * reach in and stop the in-flight zio's, so we instead
4369 * block the caller. Eventually, dsl_scan_issue_ios will
4370 * be done with issuing the zio's it gathered and will
4373 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
4379 /* Got it while it was cold in the queue */
4380 ASSERT3U(start, ==, SIO_GET_OFFSET(sio));
4381 ASSERT3U(size, ==, SIO_GET_ASIZE(sio));
4382 avl_remove(&queue->q_sios_by_addr, sio);
4383 if (avl_is_empty(&queue->q_sios_by_addr))
4384 atomic_add_64(&scn->scn_queues_pending, -1);
4385 queue->q_sio_memused -= SIO_GET_MUSED(sio);
4387 ASSERT(range_tree_contains(queue->q_exts_by_addr, start, size));
4388 range_tree_remove_fill(queue->q_exts_by_addr, start, size);
4390 /* count the block as though we issued it */
4391 sio2bp(sio, &tmpbp);
4392 count_block_issued(spa, &tmpbp, B_FALSE);
4400 * Callback invoked when a zio_free() zio is executing. This needs to be
4401 * intercepted to prevent the zio from deallocating a particular portion
4402 * of disk space and it then getting reallocated and written to, while we
4403 * still have it queued up for processing.
4406 dsl_scan_freed(spa_t *spa, const blkptr_t *bp)
4408 dsl_pool_t *dp = spa->spa_dsl_pool;
4409 dsl_scan_t *scn = dp->dp_scan;
4411 ASSERT(!BP_IS_EMBEDDED(bp));
4412 ASSERT(scn != NULL);
4413 if (!dsl_scan_is_running(scn))
4416 for (int i = 0; i < BP_GET_NDVAS(bp); i++)
4417 dsl_scan_freed_dva(spa, bp, i);
4421 * Check if a vdev needs resilvering (non-empty DTL), if so, and resilver has
4422 * not started, start it. Otherwise, only restart if max txg in DTL range is
4423 * greater than the max txg in the current scan. If the DTL max is less than
4424 * the scan max, then the vdev has not missed any new data since the resilver
4425 * started, so a restart is not needed.
4428 dsl_scan_assess_vdev(dsl_pool_t *dp, vdev_t *vd)
4432 if (!vdev_resilver_needed(vd, &min, &max))
4435 if (!dsl_scan_resilvering(dp)) {
4436 spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER);
4440 if (max <= dp->dp_scan->scn_phys.scn_max_txg)
4443 /* restart is needed, check if it can be deferred */
4444 if (spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
4445 vdev_defer_resilver(vd);
4447 spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER);
4450 ZFS_MODULE_PARAM(zfs, zfs_, scan_vdev_limit, U64, ZMOD_RW,
4451 "Max bytes in flight per leaf vdev for scrubs and resilvers");
4453 ZFS_MODULE_PARAM(zfs, zfs_, scrub_min_time_ms, UINT, ZMOD_RW,
4454 "Min millisecs to scrub per txg");
4456 ZFS_MODULE_PARAM(zfs, zfs_, obsolete_min_time_ms, UINT, ZMOD_RW,
4457 "Min millisecs to obsolete per txg");
4459 ZFS_MODULE_PARAM(zfs, zfs_, free_min_time_ms, UINT, ZMOD_RW,
4460 "Min millisecs to free per txg");
4462 ZFS_MODULE_PARAM(zfs, zfs_, resilver_min_time_ms, UINT, ZMOD_RW,
4463 "Min millisecs to resilver per txg");
4465 ZFS_MODULE_PARAM(zfs, zfs_, scan_suspend_progress, INT, ZMOD_RW,
4466 "Set to prevent scans from progressing");
4468 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_io, INT, ZMOD_RW,
4469 "Set to disable scrub I/O");
4471 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_prefetch, INT, ZMOD_RW,
4472 "Set to disable scrub prefetching");
4474 ZFS_MODULE_PARAM(zfs, zfs_, async_block_max_blocks, U64, ZMOD_RW,
4475 "Max number of blocks freed in one txg");
4477 ZFS_MODULE_PARAM(zfs, zfs_, max_async_dedup_frees, U64, ZMOD_RW,
4478 "Max number of dedup blocks freed in one txg");
4480 ZFS_MODULE_PARAM(zfs, zfs_, free_bpobj_enabled, INT, ZMOD_RW,
4481 "Enable processing of the free_bpobj");
4483 ZFS_MODULE_PARAM(zfs, zfs_, scan_blkstats, INT, ZMOD_RW,
4484 "Enable block statistics calculation during scrub");
4486 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_fact, UINT, ZMOD_RW,
4487 "Fraction of RAM for scan hard limit");
4489 ZFS_MODULE_PARAM(zfs, zfs_, scan_issue_strategy, UINT, ZMOD_RW,
4490 "IO issuing strategy during scrubbing. 0 = default, 1 = LBA, 2 = size");
4492 ZFS_MODULE_PARAM(zfs, zfs_, scan_legacy, INT, ZMOD_RW,
4493 "Scrub using legacy non-sequential method");
4495 ZFS_MODULE_PARAM(zfs, zfs_, scan_checkpoint_intval, UINT, ZMOD_RW,
4496 "Scan progress on-disk checkpointing interval");
4498 ZFS_MODULE_PARAM(zfs, zfs_, scan_max_ext_gap, U64, ZMOD_RW,
4499 "Max gap in bytes between sequential scrub / resilver I/Os");
4501 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_soft_fact, UINT, ZMOD_RW,
4502 "Fraction of hard limit used as soft limit");
4504 ZFS_MODULE_PARAM(zfs, zfs_, scan_strict_mem_lim, INT, ZMOD_RW,
4505 "Tunable to attempt to reduce lock contention");
4507 ZFS_MODULE_PARAM(zfs, zfs_, scan_fill_weight, UINT, ZMOD_RW,
4508 "Tunable to adjust bias towards more filled segments during scans");
4510 ZFS_MODULE_PARAM(zfs, zfs_, resilver_disable_defer, INT, ZMOD_RW,
4511 "Process all resilvers immediately");