4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright 2016 Gary Mills
25 * Copyright (c) 2017, 2019, Datto Inc. All rights reserved.
26 * Copyright 2019 Joyent, Inc.
29 #include <sys/dsl_scan.h>
30 #include <sys/dsl_pool.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_prop.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_synctask.h>
35 #include <sys/dnode.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/dmu_objset.h>
41 #include <sys/zfs_context.h>
42 #include <sys/fs/zfs.h>
43 #include <sys/zfs_znode.h>
44 #include <sys/spa_impl.h>
45 #include <sys/vdev_impl.h>
46 #include <sys/zil_impl.h>
47 #include <sys/zio_checksum.h>
50 #include <sys/sa_impl.h>
51 #include <sys/zfeature.h>
53 #include <sys/range_tree.h>
55 #include <sys/zfs_vfsops.h>
59 * Grand theory statement on scan queue sorting
61 * Scanning is implemented by recursively traversing all indirection levels
62 * in an object and reading all blocks referenced from said objects. This
63 * results in us approximately traversing the object from lowest logical
64 * offset to the highest. For best performance, we would want the logical
65 * blocks to be physically contiguous. However, this is frequently not the
66 * case with pools given the allocation patterns of copy-on-write filesystems.
67 * So instead, we put the I/Os into a reordering queue and issue them in a
68 * way that will most benefit physical disks (LBA-order).
72 * Ideally, we would want to scan all metadata and queue up all block I/O
73 * prior to starting to issue it, because that allows us to do an optimal
74 * sorting job. This can however consume large amounts of memory. Therefore
75 * we continuously monitor the size of the queues and constrain them to 5%
76 * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this
77 * limit, we clear out a few of the largest extents at the head of the queues
78 * to make room for more scanning. Hopefully, these extents will be fairly
79 * large and contiguous, allowing us to approach sequential I/O throughput
80 * even without a fully sorted tree.
82 * Metadata scanning takes place in dsl_scan_visit(), which is called from
83 * dsl_scan_sync() every spa_sync(). If we have either fully scanned all
84 * metadata on the pool, or we need to make room in memory because our
85 * queues are too large, dsl_scan_visit() is postponed and
86 * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies
87 * that metadata scanning and queued I/O issuing are mutually exclusive. This
88 * allows us to provide maximum sequential I/O throughput for the majority of
89 * I/O's issued since sequential I/O performance is significantly negatively
90 * impacted if it is interleaved with random I/O.
92 * Implementation Notes
94 * One side effect of the queued scanning algorithm is that the scanning code
95 * needs to be notified whenever a block is freed. This is needed to allow
96 * the scanning code to remove these I/Os from the issuing queue. Additionally,
97 * we do not attempt to queue gang blocks to be issued sequentially since this
98 * is very hard to do and would have an extremely limited performance benefit.
99 * Instead, we simply issue gang I/Os as soon as we find them using the legacy
102 * Backwards compatibility
104 * This new algorithm is backwards compatible with the legacy on-disk data
105 * structures (and therefore does not require a new feature flag).
106 * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan
107 * will stop scanning metadata (in logical order) and wait for all outstanding
108 * sorted I/O to complete. Once this is done, we write out a checkpoint
109 * bookmark, indicating that we have scanned everything logically before it.
110 * If the pool is imported on a machine without the new sorting algorithm,
111 * the scan simply resumes from the last checkpoint using the legacy algorithm.
114 typedef int (scan_cb_t
)(dsl_pool_t
*, const blkptr_t
*,
115 const zbookmark_phys_t
*);
117 static scan_cb_t dsl_scan_scrub_cb
;
119 static int scan_ds_queue_compare(const void *a
, const void *b
);
120 static int scan_prefetch_queue_compare(const void *a
, const void *b
);
121 static void scan_ds_queue_clear(dsl_scan_t
*scn
);
122 static void scan_ds_prefetch_queue_clear(dsl_scan_t
*scn
);
123 static boolean_t
scan_ds_queue_contains(dsl_scan_t
*scn
, uint64_t dsobj
,
125 static void scan_ds_queue_insert(dsl_scan_t
*scn
, uint64_t dsobj
, uint64_t txg
);
126 static void scan_ds_queue_remove(dsl_scan_t
*scn
, uint64_t dsobj
);
127 static void scan_ds_queue_sync(dsl_scan_t
*scn
, dmu_tx_t
*tx
);
128 static uint64_t dsl_scan_count_leaves(vdev_t
*vd
);
130 extern int zfs_vdev_async_write_active_min_dirty_percent
;
133 * By default zfs will check to ensure it is not over the hard memory
134 * limit before each txg. If finer-grained control of this is needed
135 * this value can be set to 1 to enable checking before scanning each
138 int zfs_scan_strict_mem_lim
= B_FALSE
;
141 * Maximum number of parallelly executed bytes per leaf vdev. We attempt
142 * to strike a balance here between keeping the vdev queues full of I/Os
143 * at all times and not overflowing the queues to cause long latency,
144 * which would cause long txg sync times. No matter what, we will not
145 * overload the drives with I/O, since that is protected by
146 * zfs_vdev_scrub_max_active.
148 unsigned long zfs_scan_vdev_limit
= 4 << 20;
150 int zfs_scan_issue_strategy
= 0;
151 int zfs_scan_legacy
= B_FALSE
; /* don't queue & sort zios, go direct */
152 unsigned long zfs_scan_max_ext_gap
= 2 << 20; /* in bytes */
155 * fill_weight is non-tunable at runtime, so we copy it at module init from
156 * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
157 * break queue sorting.
159 int zfs_scan_fill_weight
= 3;
160 static uint64_t fill_weight
;
162 /* See dsl_scan_should_clear() for details on the memory limit tunables */
163 uint64_t zfs_scan_mem_lim_min
= 16 << 20; /* bytes */
164 uint64_t zfs_scan_mem_lim_soft_max
= 128 << 20; /* bytes */
165 int zfs_scan_mem_lim_fact
= 20; /* fraction of physmem */
166 int zfs_scan_mem_lim_soft_fact
= 20; /* fraction of mem lim above */
168 int zfs_scrub_min_time_ms
= 1000; /* min millisecs to scrub per txg */
169 int zfs_obsolete_min_time_ms
= 500; /* min millisecs to obsolete per txg */
170 int zfs_free_min_time_ms
= 1000; /* min millisecs to free per txg */
171 int zfs_resilver_min_time_ms
= 3000; /* min millisecs to resilver per txg */
172 int zfs_scan_checkpoint_intval
= 7200; /* in seconds */
173 int zfs_scan_suspend_progress
= 0; /* set to prevent scans from progressing */
174 int zfs_no_scrub_io
= B_FALSE
; /* set to disable scrub i/o */
175 int zfs_no_scrub_prefetch
= B_FALSE
; /* set to disable scrub prefetch */
176 enum ddt_class zfs_scrub_ddt_class_max
= DDT_CLASS_DUPLICATE
;
177 /* max number of blocks to free in a single TXG */
178 unsigned long zfs_async_block_max_blocks
= ULONG_MAX
;
179 /* max number of dedup blocks to free in a single TXG */
180 unsigned long zfs_max_async_dedup_frees
= 100000;
182 int zfs_resilver_disable_defer
= 0; /* set to disable resilver deferring */
185 * We wait a few txgs after importing a pool to begin scanning so that
186 * the import / mounting code isn't held up by scrub / resilver IO.
187 * Unfortunately, it is a bit difficult to determine exactly how long
188 * this will take since userspace will trigger fs mounts asynchronously
189 * and the kernel will create zvol minors asynchronously. As a result,
190 * the value provided here is a bit arbitrary, but represents a
191 * reasonable estimate of how many txgs it will take to finish fully
194 #define SCAN_IMPORT_WAIT_TXGS 5
196 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
197 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
198 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
201 * Enable/disable the processing of the free_bpobj object.
203 int zfs_free_bpobj_enabled
= 1;
205 /* the order has to match pool_scan_type */
206 static scan_cb_t
*scan_funcs
[POOL_SCAN_FUNCS
] = {
208 dsl_scan_scrub_cb
, /* POOL_SCAN_SCRUB */
209 dsl_scan_scrub_cb
, /* POOL_SCAN_RESILVER */
212 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */
220 * This controls what conditions are placed on dsl_scan_sync_state():
221 * SYNC_OPTIONAL) write out scn_phys iff scn_bytes_pending == 0
222 * SYNC_MANDATORY) write out scn_phys always. scn_bytes_pending must be 0.
223 * SYNC_CACHED) if scn_bytes_pending == 0, write out scn_phys. Otherwise
224 * write out the scn_phys_cached version.
225 * See dsl_scan_sync_state for details.
234 * This struct represents the minimum information needed to reconstruct a
235 * zio for sequential scanning. This is useful because many of these will
236 * accumulate in the sequential IO queues before being issued, so saving
237 * memory matters here.
239 typedef struct scan_io
{
240 /* fields from blkptr_t */
241 uint64_t sio_blk_prop
;
242 uint64_t sio_phys_birth
;
244 zio_cksum_t sio_cksum
;
245 uint32_t sio_nr_dvas
;
247 /* fields from zio_t */
249 zbookmark_phys_t sio_zb
;
251 /* members for queue sorting */
253 avl_node_t sio_addr_node
; /* link into issuing queue */
254 list_node_t sio_list_node
; /* link for issuing to disk */
258 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp,
259 * depending on how many were in the original bp. Only the
260 * first DVA is really used for sorting and issuing purposes.
261 * The other DVAs (if provided) simply exist so that the zio
262 * layer can find additional copies to repair from in the
263 * event of an error. This array must go at the end of the
264 * struct to allow this for the variable number of elements.
269 #define SIO_SET_OFFSET(sio, x) DVA_SET_OFFSET(&(sio)->sio_dva[0], x)
270 #define SIO_SET_ASIZE(sio, x) DVA_SET_ASIZE(&(sio)->sio_dva[0], x)
271 #define SIO_GET_OFFSET(sio) DVA_GET_OFFSET(&(sio)->sio_dva[0])
272 #define SIO_GET_ASIZE(sio) DVA_GET_ASIZE(&(sio)->sio_dva[0])
273 #define SIO_GET_END_OFFSET(sio) \
274 (SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio))
275 #define SIO_GET_MUSED(sio) \
276 (sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t)))
278 struct dsl_scan_io_queue
{
279 dsl_scan_t
*q_scn
; /* associated dsl_scan_t */
280 vdev_t
*q_vd
; /* top-level vdev that this queue represents */
282 /* trees used for sorting I/Os and extents of I/Os */
283 range_tree_t
*q_exts_by_addr
;
284 zfs_btree_t q_exts_by_size
;
285 avl_tree_t q_sios_by_addr
;
286 uint64_t q_sio_memused
;
288 /* members for zio rate limiting */
289 uint64_t q_maxinflight_bytes
;
290 uint64_t q_inflight_bytes
;
291 kcondvar_t q_zio_cv
; /* used under vd->vdev_scan_io_queue_lock */
293 /* per txg statistics */
294 uint64_t q_total_seg_size_this_txg
;
295 uint64_t q_segs_this_txg
;
296 uint64_t q_total_zio_size_this_txg
;
297 uint64_t q_zios_this_txg
;
300 /* private data for dsl_scan_prefetch_cb() */
301 typedef struct scan_prefetch_ctx
{
302 zfs_refcount_t spc_refcnt
; /* refcount for memory management */
303 dsl_scan_t
*spc_scn
; /* dsl_scan_t for the pool */
304 boolean_t spc_root
; /* is this prefetch for an objset? */
305 uint8_t spc_indblkshift
; /* dn_indblkshift of current dnode */
306 uint16_t spc_datablkszsec
; /* dn_idatablkszsec of current dnode */
307 } scan_prefetch_ctx_t
;
309 /* private data for dsl_scan_prefetch() */
310 typedef struct scan_prefetch_issue_ctx
{
311 avl_node_t spic_avl_node
; /* link into scn->scn_prefetch_queue */
312 scan_prefetch_ctx_t
*spic_spc
; /* spc for the callback */
313 blkptr_t spic_bp
; /* bp to prefetch */
314 zbookmark_phys_t spic_zb
; /* bookmark to prefetch */
315 } scan_prefetch_issue_ctx_t
;
317 static void scan_exec_io(dsl_pool_t
*dp
, const blkptr_t
*bp
, int zio_flags
,
318 const zbookmark_phys_t
*zb
, dsl_scan_io_queue_t
*queue
);
319 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t
*queue
,
322 static dsl_scan_io_queue_t
*scan_io_queue_create(vdev_t
*vd
);
323 static void scan_io_queues_destroy(dsl_scan_t
*scn
);
325 static kmem_cache_t
*sio_cache
[SPA_DVAS_PER_BP
];
327 /* sio->sio_nr_dvas must be set so we know which cache to free from */
329 sio_free(scan_io_t
*sio
)
331 ASSERT3U(sio
->sio_nr_dvas
, >, 0);
332 ASSERT3U(sio
->sio_nr_dvas
, <=, SPA_DVAS_PER_BP
);
334 kmem_cache_free(sio_cache
[sio
->sio_nr_dvas
- 1], sio
);
337 /* It is up to the caller to set sio->sio_nr_dvas for freeing */
339 sio_alloc(unsigned short nr_dvas
)
341 ASSERT3U(nr_dvas
, >, 0);
342 ASSERT3U(nr_dvas
, <=, SPA_DVAS_PER_BP
);
344 return (kmem_cache_alloc(sio_cache
[nr_dvas
- 1], KM_SLEEP
));
351 * This is used in ext_size_compare() to weight segments
352 * based on how sparse they are. This cannot be changed
353 * mid-scan and the tree comparison functions don't currently
354 * have a mechanism for passing additional context to the
355 * compare functions. Thus we store this value globally and
356 * we only allow it to be set at module initialization time
358 fill_weight
= zfs_scan_fill_weight
;
360 for (int i
= 0; i
< SPA_DVAS_PER_BP
; i
++) {
363 (void) snprintf(name
, sizeof (name
), "sio_cache_%d", i
);
364 sio_cache
[i
] = kmem_cache_create(name
,
365 (sizeof (scan_io_t
) + ((i
+ 1) * sizeof (dva_t
))),
366 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
373 for (int i
= 0; i
< SPA_DVAS_PER_BP
; i
++) {
374 kmem_cache_destroy(sio_cache
[i
]);
378 static inline boolean_t
379 dsl_scan_is_running(const dsl_scan_t
*scn
)
381 return (scn
->scn_phys
.scn_state
== DSS_SCANNING
);
385 dsl_scan_resilvering(dsl_pool_t
*dp
)
387 return (dsl_scan_is_running(dp
->dp_scan
) &&
388 dp
->dp_scan
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
);
392 sio2bp(const scan_io_t
*sio
, blkptr_t
*bp
)
394 bzero(bp
, sizeof (*bp
));
395 bp
->blk_prop
= sio
->sio_blk_prop
;
396 bp
->blk_phys_birth
= sio
->sio_phys_birth
;
397 bp
->blk_birth
= sio
->sio_birth
;
398 bp
->blk_fill
= 1; /* we always only work with data pointers */
399 bp
->blk_cksum
= sio
->sio_cksum
;
401 ASSERT3U(sio
->sio_nr_dvas
, >, 0);
402 ASSERT3U(sio
->sio_nr_dvas
, <=, SPA_DVAS_PER_BP
);
404 bcopy(sio
->sio_dva
, bp
->blk_dva
, sio
->sio_nr_dvas
* sizeof (dva_t
));
408 bp2sio(const blkptr_t
*bp
, scan_io_t
*sio
, int dva_i
)
410 sio
->sio_blk_prop
= bp
->blk_prop
;
411 sio
->sio_phys_birth
= bp
->blk_phys_birth
;
412 sio
->sio_birth
= bp
->blk_birth
;
413 sio
->sio_cksum
= bp
->blk_cksum
;
414 sio
->sio_nr_dvas
= BP_GET_NDVAS(bp
);
417 * Copy the DVAs to the sio. We need all copies of the block so
418 * that the self healing code can use the alternate copies if the
419 * first is corrupted. We want the DVA at index dva_i to be first
420 * in the sio since this is the primary one that we want to issue.
422 for (int i
= 0, j
= dva_i
; i
< sio
->sio_nr_dvas
; i
++, j
++) {
423 sio
->sio_dva
[i
] = bp
->blk_dva
[j
% sio
->sio_nr_dvas
];
428 dsl_scan_init(dsl_pool_t
*dp
, uint64_t txg
)
432 spa_t
*spa
= dp
->dp_spa
;
435 scn
= dp
->dp_scan
= kmem_zalloc(sizeof (dsl_scan_t
), KM_SLEEP
);
439 * It's possible that we're resuming a scan after a reboot so
440 * make sure that the scan_async_destroying flag is initialized
443 ASSERT(!scn
->scn_async_destroying
);
444 scn
->scn_async_destroying
= spa_feature_is_active(dp
->dp_spa
,
445 SPA_FEATURE_ASYNC_DESTROY
);
448 * Calculate the max number of in-flight bytes for pool-wide
449 * scanning operations (minimum 1MB). Limits for the issuing
450 * phase are done per top-level vdev and are handled separately.
452 scn
->scn_maxinflight_bytes
= MAX(zfs_scan_vdev_limit
*
453 dsl_scan_count_leaves(spa
->spa_root_vdev
), 1ULL << 20);
455 avl_create(&scn
->scn_queue
, scan_ds_queue_compare
, sizeof (scan_ds_t
),
456 offsetof(scan_ds_t
, sds_node
));
457 avl_create(&scn
->scn_prefetch_queue
, scan_prefetch_queue_compare
,
458 sizeof (scan_prefetch_issue_ctx_t
),
459 offsetof(scan_prefetch_issue_ctx_t
, spic_avl_node
));
461 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
462 "scrub_func", sizeof (uint64_t), 1, &f
);
465 * There was an old-style scrub in progress. Restart a
466 * new-style scrub from the beginning.
468 scn
->scn_restart_txg
= txg
;
469 zfs_dbgmsg("old-style scrub was in progress; "
470 "restarting new-style scrub in txg %llu",
471 (longlong_t
)scn
->scn_restart_txg
);
474 * Load the queue obj from the old location so that it
475 * can be freed by dsl_scan_done().
477 (void) zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
478 "scrub_queue", sizeof (uint64_t), 1,
479 &scn
->scn_phys
.scn_queue_obj
);
481 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
482 DMU_POOL_SCAN
, sizeof (uint64_t), SCAN_PHYS_NUMINTS
,
485 * Detect if the pool contains the signature of #2094. If it
486 * does properly update the scn->scn_phys structure and notify
487 * the administrator by setting an errata for the pool.
489 if (err
== EOVERFLOW
) {
490 uint64_t zaptmp
[SCAN_PHYS_NUMINTS
+ 1];
491 VERIFY3S(SCAN_PHYS_NUMINTS
, ==, 24);
492 VERIFY3S(offsetof(dsl_scan_phys_t
, scn_flags
), ==,
493 (23 * sizeof (uint64_t)));
495 err
= zap_lookup(dp
->dp_meta_objset
,
496 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_SCAN
,
497 sizeof (uint64_t), SCAN_PHYS_NUMINTS
+ 1, &zaptmp
);
499 uint64_t overflow
= zaptmp
[SCAN_PHYS_NUMINTS
];
501 if (overflow
& ~DSL_SCAN_FLAGS_MASK
||
502 scn
->scn_async_destroying
) {
504 ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY
;
508 bcopy(zaptmp
, &scn
->scn_phys
,
509 SCAN_PHYS_NUMINTS
* sizeof (uint64_t));
510 scn
->scn_phys
.scn_flags
= overflow
;
512 /* Required scrub already in progress. */
513 if (scn
->scn_phys
.scn_state
== DSS_FINISHED
||
514 scn
->scn_phys
.scn_state
== DSS_CANCELED
)
516 ZPOOL_ERRATA_ZOL_2094_SCRUB
;
526 * We might be restarting after a reboot, so jump the issued
527 * counter to how far we've scanned. We know we're consistent
530 scn
->scn_issued_before_pass
= scn
->scn_phys
.scn_examined
;
532 if (dsl_scan_is_running(scn
) &&
533 spa_prev_software_version(dp
->dp_spa
) < SPA_VERSION_SCAN
) {
535 * A new-type scrub was in progress on an old
536 * pool, and the pool was accessed by old
537 * software. Restart from the beginning, since
538 * the old software may have changed the pool in
541 scn
->scn_restart_txg
= txg
;
542 zfs_dbgmsg("new-style scrub was modified "
543 "by old software; restarting in txg %llu",
544 (longlong_t
)scn
->scn_restart_txg
);
545 } else if (dsl_scan_resilvering(dp
)) {
547 * If a resilver is in progress and there are already
548 * errors, restart it instead of finishing this scan and
549 * then restarting it. If there haven't been any errors
550 * then remember that the incore DTL is valid.
552 if (scn
->scn_phys
.scn_errors
> 0) {
553 scn
->scn_restart_txg
= txg
;
554 zfs_dbgmsg("resilver can't excise DTL_MISSING "
555 "when finished; restarting in txg %llu",
556 (u_longlong_t
)scn
->scn_restart_txg
);
558 /* it's safe to excise DTL when finished */
559 spa
->spa_scrub_started
= B_TRUE
;
564 bcopy(&scn
->scn_phys
, &scn
->scn_phys_cached
, sizeof (scn
->scn_phys
));
566 /* reload the queue into the in-core state */
567 if (scn
->scn_phys
.scn_queue_obj
!= 0) {
571 for (zap_cursor_init(&zc
, dp
->dp_meta_objset
,
572 scn
->scn_phys
.scn_queue_obj
);
573 zap_cursor_retrieve(&zc
, &za
) == 0;
574 (void) zap_cursor_advance(&zc
)) {
575 scan_ds_queue_insert(scn
,
576 zfs_strtonum(za
.za_name
, NULL
),
577 za
.za_first_integer
);
579 zap_cursor_fini(&zc
);
582 spa_scan_stat_init(spa
);
587 dsl_scan_fini(dsl_pool_t
*dp
)
589 if (dp
->dp_scan
!= NULL
) {
590 dsl_scan_t
*scn
= dp
->dp_scan
;
592 if (scn
->scn_taskq
!= NULL
)
593 taskq_destroy(scn
->scn_taskq
);
595 scan_ds_queue_clear(scn
);
596 avl_destroy(&scn
->scn_queue
);
597 scan_ds_prefetch_queue_clear(scn
);
598 avl_destroy(&scn
->scn_prefetch_queue
);
600 kmem_free(dp
->dp_scan
, sizeof (dsl_scan_t
));
606 dsl_scan_restarting(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
608 return (scn
->scn_restart_txg
!= 0 &&
609 scn
->scn_restart_txg
<= tx
->tx_txg
);
613 dsl_scan_resilver_scheduled(dsl_pool_t
*dp
)
615 return ((dp
->dp_scan
&& dp
->dp_scan
->scn_restart_txg
!= 0) ||
616 (spa_async_tasks(dp
->dp_spa
) & SPA_ASYNC_RESILVER
));
620 dsl_scan_scrubbing(const dsl_pool_t
*dp
)
622 dsl_scan_phys_t
*scn_phys
= &dp
->dp_scan
->scn_phys
;
624 return (scn_phys
->scn_state
== DSS_SCANNING
&&
625 scn_phys
->scn_func
== POOL_SCAN_SCRUB
);
629 dsl_scan_is_paused_scrub(const dsl_scan_t
*scn
)
631 return (dsl_scan_scrubbing(scn
->scn_dp
) &&
632 scn
->scn_phys
.scn_flags
& DSF_SCRUB_PAUSED
);
636 * Writes out a persistent dsl_scan_phys_t record to the pool directory.
637 * Because we can be running in the block sorting algorithm, we do not always
638 * want to write out the record, only when it is "safe" to do so. This safety
639 * condition is achieved by making sure that the sorting queues are empty
640 * (scn_bytes_pending == 0). When this condition is not true, the sync'd state
641 * is inconsistent with how much actual scanning progress has been made. The
642 * kind of sync to be performed is specified by the sync_type argument. If the
643 * sync is optional, we only sync if the queues are empty. If the sync is
644 * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
645 * third possible state is a "cached" sync. This is done in response to:
646 * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
647 * destroyed, so we wouldn't be able to restart scanning from it.
648 * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
649 * superseded by a newer snapshot.
650 * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
651 * swapped with its clone.
652 * In all cases, a cached sync simply rewrites the last record we've written,
653 * just slightly modified. For the modifications that are performed to the
654 * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
655 * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
658 dsl_scan_sync_state(dsl_scan_t
*scn
, dmu_tx_t
*tx
, state_sync_type_t sync_type
)
661 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
663 ASSERT(sync_type
!= SYNC_MANDATORY
|| scn
->scn_bytes_pending
== 0);
664 if (scn
->scn_bytes_pending
== 0) {
665 for (i
= 0; i
< spa
->spa_root_vdev
->vdev_children
; i
++) {
666 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[i
];
667 dsl_scan_io_queue_t
*q
= vd
->vdev_scan_io_queue
;
672 mutex_enter(&vd
->vdev_scan_io_queue_lock
);
673 ASSERT3P(avl_first(&q
->q_sios_by_addr
), ==, NULL
);
674 ASSERT3P(zfs_btree_first(&q
->q_exts_by_size
, NULL
), ==,
676 ASSERT3P(range_tree_first(q
->q_exts_by_addr
), ==, NULL
);
677 mutex_exit(&vd
->vdev_scan_io_queue_lock
);
680 if (scn
->scn_phys
.scn_queue_obj
!= 0)
681 scan_ds_queue_sync(scn
, tx
);
682 VERIFY0(zap_update(scn
->scn_dp
->dp_meta_objset
,
683 DMU_POOL_DIRECTORY_OBJECT
,
684 DMU_POOL_SCAN
, sizeof (uint64_t), SCAN_PHYS_NUMINTS
,
685 &scn
->scn_phys
, tx
));
686 bcopy(&scn
->scn_phys
, &scn
->scn_phys_cached
,
687 sizeof (scn
->scn_phys
));
689 if (scn
->scn_checkpointing
)
690 zfs_dbgmsg("finish scan checkpoint");
692 scn
->scn_checkpointing
= B_FALSE
;
693 scn
->scn_last_checkpoint
= ddi_get_lbolt();
694 } else if (sync_type
== SYNC_CACHED
) {
695 VERIFY0(zap_update(scn
->scn_dp
->dp_meta_objset
,
696 DMU_POOL_DIRECTORY_OBJECT
,
697 DMU_POOL_SCAN
, sizeof (uint64_t), SCAN_PHYS_NUMINTS
,
698 &scn
->scn_phys_cached
, tx
));
704 dsl_scan_setup_check(void *arg
, dmu_tx_t
*tx
)
706 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
707 vdev_t
*rvd
= scn
->scn_dp
->dp_spa
->spa_root_vdev
;
709 if (dsl_scan_is_running(scn
) || vdev_rebuild_active(rvd
))
710 return (SET_ERROR(EBUSY
));
716 dsl_scan_setup_sync(void *arg
, dmu_tx_t
*tx
)
718 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
719 pool_scan_func_t
*funcp
= arg
;
720 dmu_object_type_t ot
= 0;
721 dsl_pool_t
*dp
= scn
->scn_dp
;
722 spa_t
*spa
= dp
->dp_spa
;
724 ASSERT(!dsl_scan_is_running(scn
));
725 ASSERT(*funcp
> POOL_SCAN_NONE
&& *funcp
< POOL_SCAN_FUNCS
);
726 bzero(&scn
->scn_phys
, sizeof (scn
->scn_phys
));
727 scn
->scn_phys
.scn_func
= *funcp
;
728 scn
->scn_phys
.scn_state
= DSS_SCANNING
;
729 scn
->scn_phys
.scn_min_txg
= 0;
730 scn
->scn_phys
.scn_max_txg
= tx
->tx_txg
;
731 scn
->scn_phys
.scn_ddt_class_max
= DDT_CLASSES
- 1; /* the entire DDT */
732 scn
->scn_phys
.scn_start_time
= gethrestime_sec();
733 scn
->scn_phys
.scn_errors
= 0;
734 scn
->scn_phys
.scn_to_examine
= spa
->spa_root_vdev
->vdev_stat
.vs_alloc
;
735 scn
->scn_issued_before_pass
= 0;
736 scn
->scn_restart_txg
= 0;
737 scn
->scn_done_txg
= 0;
738 scn
->scn_last_checkpoint
= 0;
739 scn
->scn_checkpointing
= B_FALSE
;
740 spa_scan_stat_init(spa
);
742 if (DSL_SCAN_IS_SCRUB_RESILVER(scn
)) {
743 scn
->scn_phys
.scn_ddt_class_max
= zfs_scrub_ddt_class_max
;
745 /* rewrite all disk labels */
746 vdev_config_dirty(spa
->spa_root_vdev
);
748 if (vdev_resilver_needed(spa
->spa_root_vdev
,
749 &scn
->scn_phys
.scn_min_txg
, &scn
->scn_phys
.scn_max_txg
)) {
750 nvlist_t
*aux
= fnvlist_alloc();
751 fnvlist_add_string(aux
, ZFS_EV_RESILVER_TYPE
,
753 spa_event_notify(spa
, NULL
, aux
,
754 ESC_ZFS_RESILVER_START
);
757 spa_event_notify(spa
, NULL
, NULL
, ESC_ZFS_SCRUB_START
);
760 spa
->spa_scrub_started
= B_TRUE
;
762 * If this is an incremental scrub, limit the DDT scrub phase
763 * to just the auto-ditto class (for correctness); the rest
764 * of the scrub should go faster using top-down pruning.
766 if (scn
->scn_phys
.scn_min_txg
> TXG_INITIAL
)
767 scn
->scn_phys
.scn_ddt_class_max
= DDT_CLASS_DITTO
;
770 * When starting a resilver clear any existing rebuild state.
771 * This is required to prevent stale rebuild status from
772 * being reported when a rebuild is run, then a resilver and
773 * finally a scrub. In which case only the scrub status
774 * should be reported by 'zpool status'.
776 if (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) {
777 vdev_t
*rvd
= spa
->spa_root_vdev
;
778 for (uint64_t i
= 0; i
< rvd
->vdev_children
; i
++) {
779 vdev_t
*vd
= rvd
->vdev_child
[i
];
780 vdev_rebuild_clear_sync(
781 (void *)(uintptr_t)vd
->vdev_id
, tx
);
786 /* back to the generic stuff */
788 if (dp
->dp_blkstats
== NULL
) {
790 vmem_alloc(sizeof (zfs_all_blkstats_t
), KM_SLEEP
);
791 mutex_init(&dp
->dp_blkstats
->zab_lock
, NULL
,
792 MUTEX_DEFAULT
, NULL
);
794 bzero(&dp
->dp_blkstats
->zab_type
, sizeof (dp
->dp_blkstats
->zab_type
));
796 if (spa_version(spa
) < SPA_VERSION_DSL_SCRUB
)
797 ot
= DMU_OT_ZAP_OTHER
;
799 scn
->scn_phys
.scn_queue_obj
= zap_create(dp
->dp_meta_objset
,
800 ot
? ot
: DMU_OT_SCAN_QUEUE
, DMU_OT_NONE
, 0, tx
);
802 bcopy(&scn
->scn_phys
, &scn
->scn_phys_cached
, sizeof (scn
->scn_phys
));
804 dsl_scan_sync_state(scn
, tx
, SYNC_MANDATORY
);
806 spa_history_log_internal(spa
, "scan setup", tx
,
807 "func=%u mintxg=%llu maxtxg=%llu",
808 *funcp
, (u_longlong_t
)scn
->scn_phys
.scn_min_txg
,
809 (u_longlong_t
)scn
->scn_phys
.scn_max_txg
);
813 * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
814 * Can also be called to resume a paused scrub.
817 dsl_scan(dsl_pool_t
*dp
, pool_scan_func_t func
)
819 spa_t
*spa
= dp
->dp_spa
;
820 dsl_scan_t
*scn
= dp
->dp_scan
;
823 * Purge all vdev caches and probe all devices. We do this here
824 * rather than in sync context because this requires a writer lock
825 * on the spa_config lock, which we can't do from sync context. The
826 * spa_scrub_reopen flag indicates that vdev_open() should not
827 * attempt to start another scrub.
829 spa_vdev_state_enter(spa
, SCL_NONE
);
830 spa
->spa_scrub_reopen
= B_TRUE
;
831 vdev_reopen(spa
->spa_root_vdev
);
832 spa
->spa_scrub_reopen
= B_FALSE
;
833 (void) spa_vdev_state_exit(spa
, NULL
, 0);
835 if (func
== POOL_SCAN_RESILVER
) {
836 dsl_scan_restart_resilver(spa
->spa_dsl_pool
, 0);
840 if (func
== POOL_SCAN_SCRUB
&& dsl_scan_is_paused_scrub(scn
)) {
841 /* got scrub start cmd, resume paused scrub */
842 int err
= dsl_scrub_set_pause_resume(scn
->scn_dp
,
845 spa_event_notify(spa
, NULL
, NULL
, ESC_ZFS_SCRUB_RESUME
);
846 return (SET_ERROR(ECANCELED
));
849 return (SET_ERROR(err
));
852 return (dsl_sync_task(spa_name(spa
), dsl_scan_setup_check
,
853 dsl_scan_setup_sync
, &func
, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED
));
858 dsl_scan_done(dsl_scan_t
*scn
, boolean_t complete
, dmu_tx_t
*tx
)
860 static const char *old_names
[] = {
862 "scrub_ddt_bookmark",
863 "scrub_ddt_class_max",
872 dsl_pool_t
*dp
= scn
->scn_dp
;
873 spa_t
*spa
= dp
->dp_spa
;
876 /* Remove any remnants of an old-style scrub. */
877 for (i
= 0; old_names
[i
]; i
++) {
878 (void) zap_remove(dp
->dp_meta_objset
,
879 DMU_POOL_DIRECTORY_OBJECT
, old_names
[i
], tx
);
882 if (scn
->scn_phys
.scn_queue_obj
!= 0) {
883 VERIFY0(dmu_object_free(dp
->dp_meta_objset
,
884 scn
->scn_phys
.scn_queue_obj
, tx
));
885 scn
->scn_phys
.scn_queue_obj
= 0;
887 scan_ds_queue_clear(scn
);
888 scan_ds_prefetch_queue_clear(scn
);
890 scn
->scn_phys
.scn_flags
&= ~DSF_SCRUB_PAUSED
;
893 * If we were "restarted" from a stopped state, don't bother
894 * with anything else.
896 if (!dsl_scan_is_running(scn
)) {
897 ASSERT(!scn
->scn_is_sorted
);
901 if (scn
->scn_is_sorted
) {
902 scan_io_queues_destroy(scn
);
903 scn
->scn_is_sorted
= B_FALSE
;
905 if (scn
->scn_taskq
!= NULL
) {
906 taskq_destroy(scn
->scn_taskq
);
907 scn
->scn_taskq
= NULL
;
911 scn
->scn_phys
.scn_state
= complete
? DSS_FINISHED
: DSS_CANCELED
;
913 spa_notify_waiters(spa
);
915 if (dsl_scan_restarting(scn
, tx
))
916 spa_history_log_internal(spa
, "scan aborted, restarting", tx
,
917 "errors=%llu", (u_longlong_t
)spa_get_errlog_size(spa
));
919 spa_history_log_internal(spa
, "scan cancelled", tx
,
920 "errors=%llu", (u_longlong_t
)spa_get_errlog_size(spa
));
922 spa_history_log_internal(spa
, "scan done", tx
,
923 "errors=%llu", (u_longlong_t
)spa_get_errlog_size(spa
));
925 if (DSL_SCAN_IS_SCRUB_RESILVER(scn
)) {
926 spa
->spa_scrub_active
= B_FALSE
;
929 * If the scrub/resilver completed, update all DTLs to
930 * reflect this. Whether it succeeded or not, vacate
931 * all temporary scrub DTLs.
933 * As the scrub does not currently support traversing
934 * data that have been freed but are part of a checkpoint,
935 * we don't mark the scrub as done in the DTLs as faults
936 * may still exist in those vdevs.
939 !spa_feature_is_active(spa
, SPA_FEATURE_POOL_CHECKPOINT
)) {
940 vdev_dtl_reassess(spa
->spa_root_vdev
, tx
->tx_txg
,
941 scn
->scn_phys
.scn_max_txg
, B_TRUE
, B_FALSE
);
943 if (scn
->scn_phys
.scn_min_txg
) {
944 nvlist_t
*aux
= fnvlist_alloc();
945 fnvlist_add_string(aux
, ZFS_EV_RESILVER_TYPE
,
947 spa_event_notify(spa
, NULL
, aux
,
948 ESC_ZFS_RESILVER_FINISH
);
951 spa_event_notify(spa
, NULL
, NULL
,
952 ESC_ZFS_SCRUB_FINISH
);
955 vdev_dtl_reassess(spa
->spa_root_vdev
, tx
->tx_txg
,
958 spa_errlog_rotate(spa
);
961 * Don't clear flag until after vdev_dtl_reassess to ensure that
962 * DTL_MISSING will get updated when possible.
964 spa
->spa_scrub_started
= B_FALSE
;
967 * We may have finished replacing a device.
968 * Let the async thread assess this and handle the detach.
970 spa_async_request(spa
, SPA_ASYNC_RESILVER_DONE
);
973 * Clear any resilver_deferred flags in the config.
974 * If there are drives that need resilvering, kick
975 * off an asynchronous request to start resilver.
976 * vdev_clear_resilver_deferred() may update the config
977 * before the resilver can restart. In the event of
978 * a crash during this period, the spa loading code
979 * will find the drives that need to be resilvered
980 * and start the resilver then.
982 if (spa_feature_is_enabled(spa
, SPA_FEATURE_RESILVER_DEFER
) &&
983 vdev_clear_resilver_deferred(spa
->spa_root_vdev
, tx
)) {
984 spa_history_log_internal(spa
,
985 "starting deferred resilver", tx
, "errors=%llu",
986 (u_longlong_t
)spa_get_errlog_size(spa
));
987 spa_async_request(spa
, SPA_ASYNC_RESILVER
);
991 scn
->scn_phys
.scn_end_time
= gethrestime_sec();
993 if (spa
->spa_errata
== ZPOOL_ERRATA_ZOL_2094_SCRUB
)
996 ASSERT(!dsl_scan_is_running(scn
));
1001 dsl_scan_cancel_check(void *arg
, dmu_tx_t
*tx
)
1003 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
1005 if (!dsl_scan_is_running(scn
))
1006 return (SET_ERROR(ENOENT
));
1012 dsl_scan_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1014 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
1016 dsl_scan_done(scn
, B_FALSE
, tx
);
1017 dsl_scan_sync_state(scn
, tx
, SYNC_MANDATORY
);
1018 spa_event_notify(scn
->scn_dp
->dp_spa
, NULL
, NULL
, ESC_ZFS_SCRUB_ABORT
);
1022 dsl_scan_cancel(dsl_pool_t
*dp
)
1024 return (dsl_sync_task(spa_name(dp
->dp_spa
), dsl_scan_cancel_check
,
1025 dsl_scan_cancel_sync
, NULL
, 3, ZFS_SPACE_CHECK_RESERVED
));
1029 dsl_scrub_pause_resume_check(void *arg
, dmu_tx_t
*tx
)
1031 pool_scrub_cmd_t
*cmd
= arg
;
1032 dsl_pool_t
*dp
= dmu_tx_pool(tx
);
1033 dsl_scan_t
*scn
= dp
->dp_scan
;
1035 if (*cmd
== POOL_SCRUB_PAUSE
) {
1036 /* can't pause a scrub when there is no in-progress scrub */
1037 if (!dsl_scan_scrubbing(dp
))
1038 return (SET_ERROR(ENOENT
));
1040 /* can't pause a paused scrub */
1041 if (dsl_scan_is_paused_scrub(scn
))
1042 return (SET_ERROR(EBUSY
));
1043 } else if (*cmd
!= POOL_SCRUB_NORMAL
) {
1044 return (SET_ERROR(ENOTSUP
));
1051 dsl_scrub_pause_resume_sync(void *arg
, dmu_tx_t
*tx
)
1053 pool_scrub_cmd_t
*cmd
= arg
;
1054 dsl_pool_t
*dp
= dmu_tx_pool(tx
);
1055 spa_t
*spa
= dp
->dp_spa
;
1056 dsl_scan_t
*scn
= dp
->dp_scan
;
1058 if (*cmd
== POOL_SCRUB_PAUSE
) {
1059 /* can't pause a scrub when there is no in-progress scrub */
1060 spa
->spa_scan_pass_scrub_pause
= gethrestime_sec();
1061 scn
->scn_phys
.scn_flags
|= DSF_SCRUB_PAUSED
;
1062 scn
->scn_phys_cached
.scn_flags
|= DSF_SCRUB_PAUSED
;
1063 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
1064 spa_event_notify(spa
, NULL
, NULL
, ESC_ZFS_SCRUB_PAUSED
);
1065 spa_notify_waiters(spa
);
1067 ASSERT3U(*cmd
, ==, POOL_SCRUB_NORMAL
);
1068 if (dsl_scan_is_paused_scrub(scn
)) {
1070 * We need to keep track of how much time we spend
1071 * paused per pass so that we can adjust the scrub rate
1072 * shown in the output of 'zpool status'
1074 spa
->spa_scan_pass_scrub_spent_paused
+=
1075 gethrestime_sec() - spa
->spa_scan_pass_scrub_pause
;
1076 spa
->spa_scan_pass_scrub_pause
= 0;
1077 scn
->scn_phys
.scn_flags
&= ~DSF_SCRUB_PAUSED
;
1078 scn
->scn_phys_cached
.scn_flags
&= ~DSF_SCRUB_PAUSED
;
1079 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
1085 * Set scrub pause/resume state if it makes sense to do so
1088 dsl_scrub_set_pause_resume(const dsl_pool_t
*dp
, pool_scrub_cmd_t cmd
)
1090 return (dsl_sync_task(spa_name(dp
->dp_spa
),
1091 dsl_scrub_pause_resume_check
, dsl_scrub_pause_resume_sync
, &cmd
, 3,
1092 ZFS_SPACE_CHECK_RESERVED
));
1096 /* start a new scan, or restart an existing one. */
1098 dsl_scan_restart_resilver(dsl_pool_t
*dp
, uint64_t txg
)
1102 tx
= dmu_tx_create_dd(dp
->dp_mos_dir
);
1103 VERIFY(0 == dmu_tx_assign(tx
, TXG_WAIT
));
1105 txg
= dmu_tx_get_txg(tx
);
1106 dp
->dp_scan
->scn_restart_txg
= txg
;
1109 dp
->dp_scan
->scn_restart_txg
= txg
;
1111 zfs_dbgmsg("restarting resilver txg=%llu", (longlong_t
)txg
);
1115 dsl_free(dsl_pool_t
*dp
, uint64_t txg
, const blkptr_t
*bp
)
1117 zio_free(dp
->dp_spa
, txg
, bp
);
1121 dsl_free_sync(zio_t
*pio
, dsl_pool_t
*dp
, uint64_t txg
, const blkptr_t
*bpp
)
1123 ASSERT(dsl_pool_sync_context(dp
));
1124 zio_nowait(zio_free_sync(pio
, dp
->dp_spa
, txg
, bpp
, pio
->io_flags
));
1128 scan_ds_queue_compare(const void *a
, const void *b
)
1130 const scan_ds_t
*sds_a
= a
, *sds_b
= b
;
1132 if (sds_a
->sds_dsobj
< sds_b
->sds_dsobj
)
1134 if (sds_a
->sds_dsobj
== sds_b
->sds_dsobj
)
1140 scan_ds_queue_clear(dsl_scan_t
*scn
)
1142 void *cookie
= NULL
;
1144 while ((sds
= avl_destroy_nodes(&scn
->scn_queue
, &cookie
)) != NULL
) {
1145 kmem_free(sds
, sizeof (*sds
));
1150 scan_ds_queue_contains(dsl_scan_t
*scn
, uint64_t dsobj
, uint64_t *txg
)
1152 scan_ds_t srch
, *sds
;
1154 srch
.sds_dsobj
= dsobj
;
1155 sds
= avl_find(&scn
->scn_queue
, &srch
, NULL
);
1156 if (sds
!= NULL
&& txg
!= NULL
)
1157 *txg
= sds
->sds_txg
;
1158 return (sds
!= NULL
);
1162 scan_ds_queue_insert(dsl_scan_t
*scn
, uint64_t dsobj
, uint64_t txg
)
1167 sds
= kmem_zalloc(sizeof (*sds
), KM_SLEEP
);
1168 sds
->sds_dsobj
= dsobj
;
1171 VERIFY3P(avl_find(&scn
->scn_queue
, sds
, &where
), ==, NULL
);
1172 avl_insert(&scn
->scn_queue
, sds
, where
);
1176 scan_ds_queue_remove(dsl_scan_t
*scn
, uint64_t dsobj
)
1178 scan_ds_t srch
, *sds
;
1180 srch
.sds_dsobj
= dsobj
;
1182 sds
= avl_find(&scn
->scn_queue
, &srch
, NULL
);
1183 VERIFY(sds
!= NULL
);
1184 avl_remove(&scn
->scn_queue
, sds
);
1185 kmem_free(sds
, sizeof (*sds
));
1189 scan_ds_queue_sync(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
1191 dsl_pool_t
*dp
= scn
->scn_dp
;
1192 spa_t
*spa
= dp
->dp_spa
;
1193 dmu_object_type_t ot
= (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
) ?
1194 DMU_OT_SCAN_QUEUE
: DMU_OT_ZAP_OTHER
;
1196 ASSERT0(scn
->scn_bytes_pending
);
1197 ASSERT(scn
->scn_phys
.scn_queue_obj
!= 0);
1199 VERIFY0(dmu_object_free(dp
->dp_meta_objset
,
1200 scn
->scn_phys
.scn_queue_obj
, tx
));
1201 scn
->scn_phys
.scn_queue_obj
= zap_create(dp
->dp_meta_objset
, ot
,
1202 DMU_OT_NONE
, 0, tx
);
1203 for (scan_ds_t
*sds
= avl_first(&scn
->scn_queue
);
1204 sds
!= NULL
; sds
= AVL_NEXT(&scn
->scn_queue
, sds
)) {
1205 VERIFY0(zap_add_int_key(dp
->dp_meta_objset
,
1206 scn
->scn_phys
.scn_queue_obj
, sds
->sds_dsobj
,
1212 * Computes the memory limit state that we're currently in. A sorted scan
1213 * needs quite a bit of memory to hold the sorting queue, so we need to
1214 * reasonably constrain the size so it doesn't impact overall system
1215 * performance. We compute two limits:
1216 * 1) Hard memory limit: if the amount of memory used by the sorting
1217 * queues on a pool gets above this value, we stop the metadata
1218 * scanning portion and start issuing the queued up and sorted
1219 * I/Os to reduce memory usage.
1220 * This limit is calculated as a fraction of physmem (by default 5%).
1221 * We constrain the lower bound of the hard limit to an absolute
1222 * minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1223 * the upper bound to 5% of the total pool size - no chance we'll
1224 * ever need that much memory, but just to keep the value in check.
1225 * 2) Soft memory limit: once we hit the hard memory limit, we start
1226 * issuing I/O to reduce queue memory usage, but we don't want to
1227 * completely empty out the queues, since we might be able to find I/Os
1228 * that will fill in the gaps of our non-sequential IOs at some point
1229 * in the future. So we stop the issuing of I/Os once the amount of
1230 * memory used drops below the soft limit (at which point we stop issuing
1231 * I/O and start scanning metadata again).
1233 * This limit is calculated by subtracting a fraction of the hard
1234 * limit from the hard limit. By default this fraction is 5%, so
1235 * the soft limit is 95% of the hard limit. We cap the size of the
1236 * difference between the hard and soft limits at an absolute
1237 * maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1238 * sufficient to not cause too frequent switching between the
1239 * metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1240 * worth of queues is about 1.2 GiB of on-pool data, so scanning
1241 * that should take at least a decent fraction of a second).
1244 dsl_scan_should_clear(dsl_scan_t
*scn
)
1246 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1247 vdev_t
*rvd
= scn
->scn_dp
->dp_spa
->spa_root_vdev
;
1248 uint64_t alloc
, mlim_hard
, mlim_soft
, mused
;
1250 alloc
= metaslab_class_get_alloc(spa_normal_class(spa
));
1251 alloc
+= metaslab_class_get_alloc(spa_special_class(spa
));
1252 alloc
+= metaslab_class_get_alloc(spa_dedup_class(spa
));
1254 mlim_hard
= MAX((physmem
/ zfs_scan_mem_lim_fact
) * PAGESIZE
,
1255 zfs_scan_mem_lim_min
);
1256 mlim_hard
= MIN(mlim_hard
, alloc
/ 20);
1257 mlim_soft
= mlim_hard
- MIN(mlim_hard
/ zfs_scan_mem_lim_soft_fact
,
1258 zfs_scan_mem_lim_soft_max
);
1260 for (uint64_t i
= 0; i
< rvd
->vdev_children
; i
++) {
1261 vdev_t
*tvd
= rvd
->vdev_child
[i
];
1262 dsl_scan_io_queue_t
*queue
;
1264 mutex_enter(&tvd
->vdev_scan_io_queue_lock
);
1265 queue
= tvd
->vdev_scan_io_queue
;
1266 if (queue
!= NULL
) {
1267 /* # extents in exts_by_size = # in exts_by_addr */
1268 mused
+= zfs_btree_numnodes(&queue
->q_exts_by_size
) *
1269 sizeof (range_seg_gap_t
) + queue
->q_sio_memused
;
1271 mutex_exit(&tvd
->vdev_scan_io_queue_lock
);
1274 dprintf("current scan memory usage: %llu bytes\n", (longlong_t
)mused
);
1277 ASSERT0(scn
->scn_bytes_pending
);
1280 * If we are above our hard limit, we need to clear out memory.
1281 * If we are below our soft limit, we need to accumulate sequential IOs.
1282 * Otherwise, we should keep doing whatever we are currently doing.
1284 if (mused
>= mlim_hard
)
1286 else if (mused
< mlim_soft
)
1289 return (scn
->scn_clearing
);
1293 dsl_scan_check_suspend(dsl_scan_t
*scn
, const zbookmark_phys_t
*zb
)
1295 /* we never skip user/group accounting objects */
1296 if (zb
&& (int64_t)zb
->zb_object
< 0)
1299 if (scn
->scn_suspending
)
1300 return (B_TRUE
); /* we're already suspending */
1302 if (!ZB_IS_ZERO(&scn
->scn_phys
.scn_bookmark
))
1303 return (B_FALSE
); /* we're resuming */
1305 /* We only know how to resume from level-0 and objset blocks. */
1306 if (zb
&& (zb
->zb_level
!= 0 && zb
->zb_level
!= ZB_ROOT_LEVEL
))
1311 * - we have scanned for at least the minimum time (default 1 sec
1312 * for scrub, 3 sec for resilver), and either we have sufficient
1313 * dirty data that we are starting to write more quickly
1314 * (default 30%), someone is explicitly waiting for this txg
1315 * to complete, or we have used up all of the time in the txg
1316 * timeout (default 5 sec).
1318 * - the spa is shutting down because this pool is being exported
1319 * or the machine is rebooting.
1321 * - the scan queue has reached its memory use limit
1323 uint64_t curr_time_ns
= gethrtime();
1324 uint64_t scan_time_ns
= curr_time_ns
- scn
->scn_sync_start_time
;
1325 uint64_t sync_time_ns
= curr_time_ns
-
1326 scn
->scn_dp
->dp_spa
->spa_sync_starttime
;
1327 int dirty_pct
= scn
->scn_dp
->dp_dirty_total
* 100 / zfs_dirty_data_max
;
1328 int mintime
= (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) ?
1329 zfs_resilver_min_time_ms
: zfs_scrub_min_time_ms
;
1331 if ((NSEC2MSEC(scan_time_ns
) > mintime
&&
1332 (dirty_pct
>= zfs_vdev_async_write_active_min_dirty_percent
||
1333 txg_sync_waiting(scn
->scn_dp
) ||
1334 NSEC2SEC(sync_time_ns
) >= zfs_txg_timeout
)) ||
1335 spa_shutting_down(scn
->scn_dp
->dp_spa
) ||
1336 (zfs_scan_strict_mem_lim
&& dsl_scan_should_clear(scn
))) {
1337 if (zb
&& zb
->zb_level
== ZB_ROOT_LEVEL
) {
1338 dprintf("suspending at first available bookmark "
1339 "%llx/%llx/%llx/%llx\n",
1340 (longlong_t
)zb
->zb_objset
,
1341 (longlong_t
)zb
->zb_object
,
1342 (longlong_t
)zb
->zb_level
,
1343 (longlong_t
)zb
->zb_blkid
);
1344 SET_BOOKMARK(&scn
->scn_phys
.scn_bookmark
,
1345 zb
->zb_objset
, 0, 0, 0);
1346 } else if (zb
!= NULL
) {
1347 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1348 (longlong_t
)zb
->zb_objset
,
1349 (longlong_t
)zb
->zb_object
,
1350 (longlong_t
)zb
->zb_level
,
1351 (longlong_t
)zb
->zb_blkid
);
1352 scn
->scn_phys
.scn_bookmark
= *zb
;
1355 dsl_scan_phys_t
*scnp
= &scn
->scn_phys
;
1356 dprintf("suspending at at DDT bookmark "
1357 "%llx/%llx/%llx/%llx\n",
1358 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_class
,
1359 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_type
,
1360 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_checksum
,
1361 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_cursor
);
1364 scn
->scn_suspending
= B_TRUE
;
1370 typedef struct zil_scan_arg
{
1372 zil_header_t
*zsa_zh
;
1377 dsl_scan_zil_block(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
1379 zil_scan_arg_t
*zsa
= arg
;
1380 dsl_pool_t
*dp
= zsa
->zsa_dp
;
1381 dsl_scan_t
*scn
= dp
->dp_scan
;
1382 zil_header_t
*zh
= zsa
->zsa_zh
;
1383 zbookmark_phys_t zb
;
1385 ASSERT(!BP_IS_REDACTED(bp
));
1386 if (BP_IS_HOLE(bp
) || bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
)
1390 * One block ("stubby") can be allocated a long time ago; we
1391 * want to visit that one because it has been allocated
1392 * (on-disk) even if it hasn't been claimed (even though for
1393 * scrub there's nothing to do to it).
1395 if (claim_txg
== 0 && bp
->blk_birth
>= spa_min_claim_txg(dp
->dp_spa
))
1398 SET_BOOKMARK(&zb
, zh
->zh_log
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1399 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
1401 VERIFY(0 == scan_funcs
[scn
->scn_phys
.scn_func
](dp
, bp
, &zb
));
1407 dsl_scan_zil_record(zilog_t
*zilog
, lr_t
*lrc
, void *arg
, uint64_t claim_txg
)
1409 if (lrc
->lrc_txtype
== TX_WRITE
) {
1410 zil_scan_arg_t
*zsa
= arg
;
1411 dsl_pool_t
*dp
= zsa
->zsa_dp
;
1412 dsl_scan_t
*scn
= dp
->dp_scan
;
1413 zil_header_t
*zh
= zsa
->zsa_zh
;
1414 lr_write_t
*lr
= (lr_write_t
*)lrc
;
1415 blkptr_t
*bp
= &lr
->lr_blkptr
;
1416 zbookmark_phys_t zb
;
1418 ASSERT(!BP_IS_REDACTED(bp
));
1419 if (BP_IS_HOLE(bp
) ||
1420 bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
)
1424 * birth can be < claim_txg if this record's txg is
1425 * already txg sync'ed (but this log block contains
1426 * other records that are not synced)
1428 if (claim_txg
== 0 || bp
->blk_birth
< claim_txg
)
1431 SET_BOOKMARK(&zb
, zh
->zh_log
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1432 lr
->lr_foid
, ZB_ZIL_LEVEL
,
1433 lr
->lr_offset
/ BP_GET_LSIZE(bp
));
1435 VERIFY(0 == scan_funcs
[scn
->scn_phys
.scn_func
](dp
, bp
, &zb
));
1441 dsl_scan_zil(dsl_pool_t
*dp
, zil_header_t
*zh
)
1443 uint64_t claim_txg
= zh
->zh_claim_txg
;
1444 zil_scan_arg_t zsa
= { dp
, zh
};
1447 ASSERT(spa_writeable(dp
->dp_spa
));
1450 * We only want to visit blocks that have been claimed but not yet
1451 * replayed (or, in read-only mode, blocks that *would* be claimed).
1456 zilog
= zil_alloc(dp
->dp_meta_objset
, zh
);
1458 (void) zil_parse(zilog
, dsl_scan_zil_block
, dsl_scan_zil_record
, &zsa
,
1459 claim_txg
, B_FALSE
);
1465 * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1466 * here is to sort the AVL tree by the order each block will be needed.
1469 scan_prefetch_queue_compare(const void *a
, const void *b
)
1471 const scan_prefetch_issue_ctx_t
*spic_a
= a
, *spic_b
= b
;
1472 const scan_prefetch_ctx_t
*spc_a
= spic_a
->spic_spc
;
1473 const scan_prefetch_ctx_t
*spc_b
= spic_b
->spic_spc
;
1475 return (zbookmark_compare(spc_a
->spc_datablkszsec
,
1476 spc_a
->spc_indblkshift
, spc_b
->spc_datablkszsec
,
1477 spc_b
->spc_indblkshift
, &spic_a
->spic_zb
, &spic_b
->spic_zb
));
1481 scan_prefetch_ctx_rele(scan_prefetch_ctx_t
*spc
, void *tag
)
1483 if (zfs_refcount_remove(&spc
->spc_refcnt
, tag
) == 0) {
1484 zfs_refcount_destroy(&spc
->spc_refcnt
);
1485 kmem_free(spc
, sizeof (scan_prefetch_ctx_t
));
1489 static scan_prefetch_ctx_t
*
1490 scan_prefetch_ctx_create(dsl_scan_t
*scn
, dnode_phys_t
*dnp
, void *tag
)
1492 scan_prefetch_ctx_t
*spc
;
1494 spc
= kmem_alloc(sizeof (scan_prefetch_ctx_t
), KM_SLEEP
);
1495 zfs_refcount_create(&spc
->spc_refcnt
);
1496 zfs_refcount_add(&spc
->spc_refcnt
, tag
);
1499 spc
->spc_datablkszsec
= dnp
->dn_datablkszsec
;
1500 spc
->spc_indblkshift
= dnp
->dn_indblkshift
;
1501 spc
->spc_root
= B_FALSE
;
1503 spc
->spc_datablkszsec
= 0;
1504 spc
->spc_indblkshift
= 0;
1505 spc
->spc_root
= B_TRUE
;
1512 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t
*spc
, void *tag
)
1514 zfs_refcount_add(&spc
->spc_refcnt
, tag
);
1518 scan_ds_prefetch_queue_clear(dsl_scan_t
*scn
)
1520 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1521 void *cookie
= NULL
;
1522 scan_prefetch_issue_ctx_t
*spic
= NULL
;
1524 mutex_enter(&spa
->spa_scrub_lock
);
1525 while ((spic
= avl_destroy_nodes(&scn
->scn_prefetch_queue
,
1526 &cookie
)) != NULL
) {
1527 scan_prefetch_ctx_rele(spic
->spic_spc
, scn
);
1528 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1530 mutex_exit(&spa
->spa_scrub_lock
);
1534 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t
*spc
,
1535 const zbookmark_phys_t
*zb
)
1537 zbookmark_phys_t
*last_zb
= &spc
->spc_scn
->scn_prefetch_bookmark
;
1538 dnode_phys_t tmp_dnp
;
1539 dnode_phys_t
*dnp
= (spc
->spc_root
) ? NULL
: &tmp_dnp
;
1541 if (zb
->zb_objset
!= last_zb
->zb_objset
)
1543 if ((int64_t)zb
->zb_object
< 0)
1546 tmp_dnp
.dn_datablkszsec
= spc
->spc_datablkszsec
;
1547 tmp_dnp
.dn_indblkshift
= spc
->spc_indblkshift
;
1549 if (zbookmark_subtree_completed(dnp
, zb
, last_zb
))
1556 dsl_scan_prefetch(scan_prefetch_ctx_t
*spc
, blkptr_t
*bp
, zbookmark_phys_t
*zb
)
1559 dsl_scan_t
*scn
= spc
->spc_scn
;
1560 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1561 scan_prefetch_issue_ctx_t
*spic
;
1563 if (zfs_no_scrub_prefetch
|| BP_IS_REDACTED(bp
))
1566 if (BP_IS_HOLE(bp
) || bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
||
1567 (BP_GET_LEVEL(bp
) == 0 && BP_GET_TYPE(bp
) != DMU_OT_DNODE
&&
1568 BP_GET_TYPE(bp
) != DMU_OT_OBJSET
))
1571 if (dsl_scan_check_prefetch_resume(spc
, zb
))
1574 scan_prefetch_ctx_add_ref(spc
, scn
);
1575 spic
= kmem_alloc(sizeof (scan_prefetch_issue_ctx_t
), KM_SLEEP
);
1576 spic
->spic_spc
= spc
;
1577 spic
->spic_bp
= *bp
;
1578 spic
->spic_zb
= *zb
;
1581 * Add the IO to the queue of blocks to prefetch. This allows us to
1582 * prioritize blocks that we will need first for the main traversal
1585 mutex_enter(&spa
->spa_scrub_lock
);
1586 if (avl_find(&scn
->scn_prefetch_queue
, spic
, &idx
) != NULL
) {
1587 /* this block is already queued for prefetch */
1588 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1589 scan_prefetch_ctx_rele(spc
, scn
);
1590 mutex_exit(&spa
->spa_scrub_lock
);
1594 avl_insert(&scn
->scn_prefetch_queue
, spic
, idx
);
1595 cv_broadcast(&spa
->spa_scrub_io_cv
);
1596 mutex_exit(&spa
->spa_scrub_lock
);
1600 dsl_scan_prefetch_dnode(dsl_scan_t
*scn
, dnode_phys_t
*dnp
,
1601 uint64_t objset
, uint64_t object
)
1604 zbookmark_phys_t zb
;
1605 scan_prefetch_ctx_t
*spc
;
1607 if (dnp
->dn_nblkptr
== 0 && !(dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1610 SET_BOOKMARK(&zb
, objset
, object
, 0, 0);
1612 spc
= scan_prefetch_ctx_create(scn
, dnp
, FTAG
);
1614 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++) {
1615 zb
.zb_level
= BP_GET_LEVEL(&dnp
->dn_blkptr
[i
]);
1617 dsl_scan_prefetch(spc
, &dnp
->dn_blkptr
[i
], &zb
);
1620 if (dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) {
1622 zb
.zb_blkid
= DMU_SPILL_BLKID
;
1623 dsl_scan_prefetch(spc
, DN_SPILL_BLKPTR(dnp
), &zb
);
1626 scan_prefetch_ctx_rele(spc
, FTAG
);
1630 dsl_scan_prefetch_cb(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1631 arc_buf_t
*buf
, void *private)
1633 scan_prefetch_ctx_t
*spc
= private;
1634 dsl_scan_t
*scn
= spc
->spc_scn
;
1635 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1637 /* broadcast that the IO has completed for rate limiting purposes */
1638 mutex_enter(&spa
->spa_scrub_lock
);
1639 ASSERT3U(spa
->spa_scrub_inflight
, >=, BP_GET_PSIZE(bp
));
1640 spa
->spa_scrub_inflight
-= BP_GET_PSIZE(bp
);
1641 cv_broadcast(&spa
->spa_scrub_io_cv
);
1642 mutex_exit(&spa
->spa_scrub_lock
);
1644 /* if there was an error or we are done prefetching, just cleanup */
1645 if (buf
== NULL
|| scn
->scn_prefetch_stop
)
1648 if (BP_GET_LEVEL(bp
) > 0) {
1651 int epb
= BP_GET_LSIZE(bp
) >> SPA_BLKPTRSHIFT
;
1652 zbookmark_phys_t czb
;
1654 for (i
= 0, cbp
= buf
->b_data
; i
< epb
; i
++, cbp
++) {
1655 SET_BOOKMARK(&czb
, zb
->zb_objset
, zb
->zb_object
,
1656 zb
->zb_level
- 1, zb
->zb_blkid
* epb
+ i
);
1657 dsl_scan_prefetch(spc
, cbp
, &czb
);
1659 } else if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
) {
1662 int epb
= BP_GET_LSIZE(bp
) >> DNODE_SHIFT
;
1664 for (i
= 0, cdnp
= buf
->b_data
; i
< epb
;
1665 i
+= cdnp
->dn_extra_slots
+ 1,
1666 cdnp
+= cdnp
->dn_extra_slots
+ 1) {
1667 dsl_scan_prefetch_dnode(scn
, cdnp
,
1668 zb
->zb_objset
, zb
->zb_blkid
* epb
+ i
);
1670 } else if (BP_GET_TYPE(bp
) == DMU_OT_OBJSET
) {
1671 objset_phys_t
*osp
= buf
->b_data
;
1673 dsl_scan_prefetch_dnode(scn
, &osp
->os_meta_dnode
,
1674 zb
->zb_objset
, DMU_META_DNODE_OBJECT
);
1676 if (OBJSET_BUF_HAS_USERUSED(buf
)) {
1677 dsl_scan_prefetch_dnode(scn
,
1678 &osp
->os_groupused_dnode
, zb
->zb_objset
,
1679 DMU_GROUPUSED_OBJECT
);
1680 dsl_scan_prefetch_dnode(scn
,
1681 &osp
->os_userused_dnode
, zb
->zb_objset
,
1682 DMU_USERUSED_OBJECT
);
1688 arc_buf_destroy(buf
, private);
1689 scan_prefetch_ctx_rele(spc
, scn
);
1694 dsl_scan_prefetch_thread(void *arg
)
1696 dsl_scan_t
*scn
= arg
;
1697 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1698 scan_prefetch_issue_ctx_t
*spic
;
1700 /* loop until we are told to stop */
1701 while (!scn
->scn_prefetch_stop
) {
1702 arc_flags_t flags
= ARC_FLAG_NOWAIT
|
1703 ARC_FLAG_PRESCIENT_PREFETCH
| ARC_FLAG_PREFETCH
;
1704 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SCAN_THREAD
;
1706 mutex_enter(&spa
->spa_scrub_lock
);
1709 * Wait until we have an IO to issue and are not above our
1710 * maximum in flight limit.
1712 while (!scn
->scn_prefetch_stop
&&
1713 (avl_numnodes(&scn
->scn_prefetch_queue
) == 0 ||
1714 spa
->spa_scrub_inflight
>= scn
->scn_maxinflight_bytes
)) {
1715 cv_wait(&spa
->spa_scrub_io_cv
, &spa
->spa_scrub_lock
);
1718 /* recheck if we should stop since we waited for the cv */
1719 if (scn
->scn_prefetch_stop
) {
1720 mutex_exit(&spa
->spa_scrub_lock
);
1724 /* remove the prefetch IO from the tree */
1725 spic
= avl_first(&scn
->scn_prefetch_queue
);
1726 spa
->spa_scrub_inflight
+= BP_GET_PSIZE(&spic
->spic_bp
);
1727 avl_remove(&scn
->scn_prefetch_queue
, spic
);
1729 mutex_exit(&spa
->spa_scrub_lock
);
1731 if (BP_IS_PROTECTED(&spic
->spic_bp
)) {
1732 ASSERT(BP_GET_TYPE(&spic
->spic_bp
) == DMU_OT_DNODE
||
1733 BP_GET_TYPE(&spic
->spic_bp
) == DMU_OT_OBJSET
);
1734 ASSERT3U(BP_GET_LEVEL(&spic
->spic_bp
), ==, 0);
1735 zio_flags
|= ZIO_FLAG_RAW
;
1738 /* issue the prefetch asynchronously */
1739 (void) arc_read(scn
->scn_zio_root
, scn
->scn_dp
->dp_spa
,
1740 &spic
->spic_bp
, dsl_scan_prefetch_cb
, spic
->spic_spc
,
1741 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, &spic
->spic_zb
);
1743 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1746 ASSERT(scn
->scn_prefetch_stop
);
1748 /* free any prefetches we didn't get to complete */
1749 mutex_enter(&spa
->spa_scrub_lock
);
1750 while ((spic
= avl_first(&scn
->scn_prefetch_queue
)) != NULL
) {
1751 avl_remove(&scn
->scn_prefetch_queue
, spic
);
1752 scan_prefetch_ctx_rele(spic
->spic_spc
, scn
);
1753 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1755 ASSERT0(avl_numnodes(&scn
->scn_prefetch_queue
));
1756 mutex_exit(&spa
->spa_scrub_lock
);
1760 dsl_scan_check_resume(dsl_scan_t
*scn
, const dnode_phys_t
*dnp
,
1761 const zbookmark_phys_t
*zb
)
1764 * We never skip over user/group accounting objects (obj<0)
1766 if (!ZB_IS_ZERO(&scn
->scn_phys
.scn_bookmark
) &&
1767 (int64_t)zb
->zb_object
>= 0) {
1769 * If we already visited this bp & everything below (in
1770 * a prior txg sync), don't bother doing it again.
1772 if (zbookmark_subtree_completed(dnp
, zb
,
1773 &scn
->scn_phys
.scn_bookmark
))
1777 * If we found the block we're trying to resume from, or
1778 * we went past it to a different object, zero it out to
1779 * indicate that it's OK to start checking for suspending
1782 if (bcmp(zb
, &scn
->scn_phys
.scn_bookmark
, sizeof (*zb
)) == 0 ||
1783 zb
->zb_object
> scn
->scn_phys
.scn_bookmark
.zb_object
) {
1784 dprintf("resuming at %llx/%llx/%llx/%llx\n",
1785 (longlong_t
)zb
->zb_objset
,
1786 (longlong_t
)zb
->zb_object
,
1787 (longlong_t
)zb
->zb_level
,
1788 (longlong_t
)zb
->zb_blkid
);
1789 bzero(&scn
->scn_phys
.scn_bookmark
, sizeof (*zb
));
1795 static void dsl_scan_visitbp(blkptr_t
*bp
, const zbookmark_phys_t
*zb
,
1796 dnode_phys_t
*dnp
, dsl_dataset_t
*ds
, dsl_scan_t
*scn
,
1797 dmu_objset_type_t ostype
, dmu_tx_t
*tx
);
1798 inline __attribute__((always_inline
)) static void dsl_scan_visitdnode(
1799 dsl_scan_t
*, dsl_dataset_t
*ds
, dmu_objset_type_t ostype
,
1800 dnode_phys_t
*dnp
, uint64_t object
, dmu_tx_t
*tx
);
1803 * Return nonzero on i/o error.
1804 * Return new buf to write out in *bufp.
1806 inline __attribute__((always_inline
)) static int
1807 dsl_scan_recurse(dsl_scan_t
*scn
, dsl_dataset_t
*ds
, dmu_objset_type_t ostype
,
1808 dnode_phys_t
*dnp
, const blkptr_t
*bp
,
1809 const zbookmark_phys_t
*zb
, dmu_tx_t
*tx
)
1811 dsl_pool_t
*dp
= scn
->scn_dp
;
1812 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SCAN_THREAD
;
1815 ASSERT(!BP_IS_REDACTED(bp
));
1817 if (BP_GET_LEVEL(bp
) > 0) {
1818 arc_flags_t flags
= ARC_FLAG_WAIT
;
1821 int epb
= BP_GET_LSIZE(bp
) >> SPA_BLKPTRSHIFT
;
1824 err
= arc_read(NULL
, dp
->dp_spa
, bp
, arc_getbuf_func
, &buf
,
1825 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, zb
);
1827 scn
->scn_phys
.scn_errors
++;
1830 for (i
= 0, cbp
= buf
->b_data
; i
< epb
; i
++, cbp
++) {
1831 zbookmark_phys_t czb
;
1833 SET_BOOKMARK(&czb
, zb
->zb_objset
, zb
->zb_object
,
1835 zb
->zb_blkid
* epb
+ i
);
1836 dsl_scan_visitbp(cbp
, &czb
, dnp
,
1837 ds
, scn
, ostype
, tx
);
1839 arc_buf_destroy(buf
, &buf
);
1840 } else if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
) {
1841 arc_flags_t flags
= ARC_FLAG_WAIT
;
1844 int epb
= BP_GET_LSIZE(bp
) >> DNODE_SHIFT
;
1847 if (BP_IS_PROTECTED(bp
)) {
1848 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
1849 zio_flags
|= ZIO_FLAG_RAW
;
1852 err
= arc_read(NULL
, dp
->dp_spa
, bp
, arc_getbuf_func
, &buf
,
1853 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, zb
);
1855 scn
->scn_phys
.scn_errors
++;
1858 for (i
= 0, cdnp
= buf
->b_data
; i
< epb
;
1859 i
+= cdnp
->dn_extra_slots
+ 1,
1860 cdnp
+= cdnp
->dn_extra_slots
+ 1) {
1861 dsl_scan_visitdnode(scn
, ds
, ostype
,
1862 cdnp
, zb
->zb_blkid
* epb
+ i
, tx
);
1865 arc_buf_destroy(buf
, &buf
);
1866 } else if (BP_GET_TYPE(bp
) == DMU_OT_OBJSET
) {
1867 arc_flags_t flags
= ARC_FLAG_WAIT
;
1871 err
= arc_read(NULL
, dp
->dp_spa
, bp
, arc_getbuf_func
, &buf
,
1872 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, zb
);
1874 scn
->scn_phys
.scn_errors
++;
1880 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1881 &osp
->os_meta_dnode
, DMU_META_DNODE_OBJECT
, tx
);
1883 if (OBJSET_BUF_HAS_USERUSED(buf
)) {
1885 * We also always visit user/group/project accounting
1886 * objects, and never skip them, even if we are
1887 * suspending. This is necessary so that the
1888 * space deltas from this txg get integrated.
1890 if (OBJSET_BUF_HAS_PROJECTUSED(buf
))
1891 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1892 &osp
->os_projectused_dnode
,
1893 DMU_PROJECTUSED_OBJECT
, tx
);
1894 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1895 &osp
->os_groupused_dnode
,
1896 DMU_GROUPUSED_OBJECT
, tx
);
1897 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1898 &osp
->os_userused_dnode
,
1899 DMU_USERUSED_OBJECT
, tx
);
1901 arc_buf_destroy(buf
, &buf
);
1907 inline __attribute__((always_inline
)) static void
1908 dsl_scan_visitdnode(dsl_scan_t
*scn
, dsl_dataset_t
*ds
,
1909 dmu_objset_type_t ostype
, dnode_phys_t
*dnp
,
1910 uint64_t object
, dmu_tx_t
*tx
)
1914 for (j
= 0; j
< dnp
->dn_nblkptr
; j
++) {
1915 zbookmark_phys_t czb
;
1917 SET_BOOKMARK(&czb
, ds
? ds
->ds_object
: 0, object
,
1918 dnp
->dn_nlevels
- 1, j
);
1919 dsl_scan_visitbp(&dnp
->dn_blkptr
[j
],
1920 &czb
, dnp
, ds
, scn
, ostype
, tx
);
1923 if (dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) {
1924 zbookmark_phys_t czb
;
1925 SET_BOOKMARK(&czb
, ds
? ds
->ds_object
: 0, object
,
1926 0, DMU_SPILL_BLKID
);
1927 dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp
),
1928 &czb
, dnp
, ds
, scn
, ostype
, tx
);
1933 * The arguments are in this order because mdb can only print the
1934 * first 5; we want them to be useful.
1937 dsl_scan_visitbp(blkptr_t
*bp
, const zbookmark_phys_t
*zb
,
1938 dnode_phys_t
*dnp
, dsl_dataset_t
*ds
, dsl_scan_t
*scn
,
1939 dmu_objset_type_t ostype
, dmu_tx_t
*tx
)
1941 dsl_pool_t
*dp
= scn
->scn_dp
;
1942 blkptr_t
*bp_toread
= NULL
;
1944 if (dsl_scan_check_suspend(scn
, zb
))
1947 if (dsl_scan_check_resume(scn
, dnp
, zb
))
1950 scn
->scn_visited_this_txg
++;
1953 * This debugging is commented out to conserve stack space. This
1954 * function is called recursively and the debugging adds several
1955 * bytes to the stack for each call. It can be commented back in
1956 * if required to debug an issue in dsl_scan_visitbp().
1959 * "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
1960 * ds, ds ? ds->ds_object : 0,
1961 * zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1965 if (BP_IS_HOLE(bp
)) {
1966 scn
->scn_holes_this_txg
++;
1970 if (BP_IS_REDACTED(bp
)) {
1971 ASSERT(dsl_dataset_feature_is_active(ds
,
1972 SPA_FEATURE_REDACTED_DATASETS
));
1976 if (bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
) {
1977 scn
->scn_lt_min_this_txg
++;
1981 bp_toread
= kmem_alloc(sizeof (blkptr_t
), KM_SLEEP
);
1984 if (dsl_scan_recurse(scn
, ds
, ostype
, dnp
, bp_toread
, zb
, tx
) != 0)
1988 * If dsl_scan_ddt() has already visited this block, it will have
1989 * already done any translations or scrubbing, so don't call the
1992 if (ddt_class_contains(dp
->dp_spa
,
1993 scn
->scn_phys
.scn_ddt_class_max
, bp
)) {
1994 scn
->scn_ddt_contained_this_txg
++;
1999 * If this block is from the future (after cur_max_txg), then we
2000 * are doing this on behalf of a deleted snapshot, and we will
2001 * revisit the future block on the next pass of this dataset.
2002 * Don't scan it now unless we need to because something
2003 * under it was modified.
2005 if (BP_PHYSICAL_BIRTH(bp
) > scn
->scn_phys
.scn_cur_max_txg
) {
2006 scn
->scn_gt_max_this_txg
++;
2010 scan_funcs
[scn
->scn_phys
.scn_func
](dp
, bp
, zb
);
2013 kmem_free(bp_toread
, sizeof (blkptr_t
));
2017 dsl_scan_visit_rootbp(dsl_scan_t
*scn
, dsl_dataset_t
*ds
, blkptr_t
*bp
,
2020 zbookmark_phys_t zb
;
2021 scan_prefetch_ctx_t
*spc
;
2023 SET_BOOKMARK(&zb
, ds
? ds
->ds_object
: DMU_META_OBJSET
,
2024 ZB_ROOT_OBJECT
, ZB_ROOT_LEVEL
, ZB_ROOT_BLKID
);
2026 if (ZB_IS_ZERO(&scn
->scn_phys
.scn_bookmark
)) {
2027 SET_BOOKMARK(&scn
->scn_prefetch_bookmark
,
2028 zb
.zb_objset
, 0, 0, 0);
2030 scn
->scn_prefetch_bookmark
= scn
->scn_phys
.scn_bookmark
;
2033 scn
->scn_objsets_visited_this_txg
++;
2035 spc
= scan_prefetch_ctx_create(scn
, NULL
, FTAG
);
2036 dsl_scan_prefetch(spc
, bp
, &zb
);
2037 scan_prefetch_ctx_rele(spc
, FTAG
);
2039 dsl_scan_visitbp(bp
, &zb
, NULL
, ds
, scn
, DMU_OST_NONE
, tx
);
2041 dprintf_ds(ds
, "finished scan%s", "");
2045 ds_destroyed_scn_phys(dsl_dataset_t
*ds
, dsl_scan_phys_t
*scn_phys
)
2047 if (scn_phys
->scn_bookmark
.zb_objset
== ds
->ds_object
) {
2048 if (ds
->ds_is_snapshot
) {
2051 * - scn_cur_{min,max}_txg stays the same.
2052 * - Setting the flag is not really necessary if
2053 * scn_cur_max_txg == scn_max_txg, because there
2054 * is nothing after this snapshot that we care
2055 * about. However, we set it anyway and then
2056 * ignore it when we retraverse it in
2057 * dsl_scan_visitds().
2059 scn_phys
->scn_bookmark
.zb_objset
=
2060 dsl_dataset_phys(ds
)->ds_next_snap_obj
;
2061 zfs_dbgmsg("destroying ds %llu; currently traversing; "
2062 "reset zb_objset to %llu",
2063 (u_longlong_t
)ds
->ds_object
,
2064 (u_longlong_t
)dsl_dataset_phys(ds
)->
2066 scn_phys
->scn_flags
|= DSF_VISIT_DS_AGAIN
;
2068 SET_BOOKMARK(&scn_phys
->scn_bookmark
,
2069 ZB_DESTROYED_OBJSET
, 0, 0, 0);
2070 zfs_dbgmsg("destroying ds %llu; currently traversing; "
2071 "reset bookmark to -1,0,0,0",
2072 (u_longlong_t
)ds
->ds_object
);
2078 * Invoked when a dataset is destroyed. We need to make sure that:
2080 * 1) If it is the dataset that was currently being scanned, we write
2081 * a new dsl_scan_phys_t and marking the objset reference in it
2083 * 2) Remove it from the work queue, if it was present.
2085 * If the dataset was actually a snapshot, instead of marking the dataset
2086 * as destroyed, we instead substitute the next snapshot in line.
2089 dsl_scan_ds_destroyed(dsl_dataset_t
*ds
, dmu_tx_t
*tx
)
2091 dsl_pool_t
*dp
= ds
->ds_dir
->dd_pool
;
2092 dsl_scan_t
*scn
= dp
->dp_scan
;
2095 if (!dsl_scan_is_running(scn
))
2098 ds_destroyed_scn_phys(ds
, &scn
->scn_phys
);
2099 ds_destroyed_scn_phys(ds
, &scn
->scn_phys_cached
);
2101 if (scan_ds_queue_contains(scn
, ds
->ds_object
, &mintxg
)) {
2102 scan_ds_queue_remove(scn
, ds
->ds_object
);
2103 if (ds
->ds_is_snapshot
)
2104 scan_ds_queue_insert(scn
,
2105 dsl_dataset_phys(ds
)->ds_next_snap_obj
, mintxg
);
2108 if (zap_lookup_int_key(dp
->dp_meta_objset
, scn
->scn_phys
.scn_queue_obj
,
2109 ds
->ds_object
, &mintxg
) == 0) {
2110 ASSERT3U(dsl_dataset_phys(ds
)->ds_num_children
, <=, 1);
2111 VERIFY3U(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2112 scn
->scn_phys
.scn_queue_obj
, ds
->ds_object
, tx
));
2113 if (ds
->ds_is_snapshot
) {
2115 * We keep the same mintxg; it could be >
2116 * ds_creation_txg if the previous snapshot was
2119 VERIFY(zap_add_int_key(dp
->dp_meta_objset
,
2120 scn
->scn_phys
.scn_queue_obj
,
2121 dsl_dataset_phys(ds
)->ds_next_snap_obj
,
2123 zfs_dbgmsg("destroying ds %llu; in queue; "
2124 "replacing with %llu",
2125 (u_longlong_t
)ds
->ds_object
,
2126 (u_longlong_t
)dsl_dataset_phys(ds
)->
2129 zfs_dbgmsg("destroying ds %llu; in queue; removing",
2130 (u_longlong_t
)ds
->ds_object
);
2135 * dsl_scan_sync() should be called after this, and should sync
2136 * out our changed state, but just to be safe, do it here.
2138 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
2142 ds_snapshotted_bookmark(dsl_dataset_t
*ds
, zbookmark_phys_t
*scn_bookmark
)
2144 if (scn_bookmark
->zb_objset
== ds
->ds_object
) {
2145 scn_bookmark
->zb_objset
=
2146 dsl_dataset_phys(ds
)->ds_prev_snap_obj
;
2147 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
2148 "reset zb_objset to %llu",
2149 (u_longlong_t
)ds
->ds_object
,
2150 (u_longlong_t
)dsl_dataset_phys(ds
)->ds_prev_snap_obj
);
2155 * Called when a dataset is snapshotted. If we were currently traversing
2156 * this snapshot, we reset our bookmark to point at the newly created
2157 * snapshot. We also modify our work queue to remove the old snapshot and
2158 * replace with the new one.
2161 dsl_scan_ds_snapshotted(dsl_dataset_t
*ds
, dmu_tx_t
*tx
)
2163 dsl_pool_t
*dp
= ds
->ds_dir
->dd_pool
;
2164 dsl_scan_t
*scn
= dp
->dp_scan
;
2167 if (!dsl_scan_is_running(scn
))
2170 ASSERT(dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0);
2172 ds_snapshotted_bookmark(ds
, &scn
->scn_phys
.scn_bookmark
);
2173 ds_snapshotted_bookmark(ds
, &scn
->scn_phys_cached
.scn_bookmark
);
2175 if (scan_ds_queue_contains(scn
, ds
->ds_object
, &mintxg
)) {
2176 scan_ds_queue_remove(scn
, ds
->ds_object
);
2177 scan_ds_queue_insert(scn
,
2178 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, mintxg
);
2181 if (zap_lookup_int_key(dp
->dp_meta_objset
, scn
->scn_phys
.scn_queue_obj
,
2182 ds
->ds_object
, &mintxg
) == 0) {
2183 VERIFY3U(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2184 scn
->scn_phys
.scn_queue_obj
, ds
->ds_object
, tx
));
2185 VERIFY(zap_add_int_key(dp
->dp_meta_objset
,
2186 scn
->scn_phys
.scn_queue_obj
,
2187 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, mintxg
, tx
) == 0);
2188 zfs_dbgmsg("snapshotting ds %llu; in queue; "
2189 "replacing with %llu",
2190 (u_longlong_t
)ds
->ds_object
,
2191 (u_longlong_t
)dsl_dataset_phys(ds
)->ds_prev_snap_obj
);
2194 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
2198 ds_clone_swapped_bookmark(dsl_dataset_t
*ds1
, dsl_dataset_t
*ds2
,
2199 zbookmark_phys_t
*scn_bookmark
)
2201 if (scn_bookmark
->zb_objset
== ds1
->ds_object
) {
2202 scn_bookmark
->zb_objset
= ds2
->ds_object
;
2203 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2204 "reset zb_objset to %llu",
2205 (u_longlong_t
)ds1
->ds_object
,
2206 (u_longlong_t
)ds2
->ds_object
);
2207 } else if (scn_bookmark
->zb_objset
== ds2
->ds_object
) {
2208 scn_bookmark
->zb_objset
= ds1
->ds_object
;
2209 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2210 "reset zb_objset to %llu",
2211 (u_longlong_t
)ds2
->ds_object
,
2212 (u_longlong_t
)ds1
->ds_object
);
2217 * Called when an origin dataset and its clone are swapped. If we were
2218 * currently traversing the dataset, we need to switch to traversing the
2219 * newly promoted clone.
2222 dsl_scan_ds_clone_swapped(dsl_dataset_t
*ds1
, dsl_dataset_t
*ds2
, dmu_tx_t
*tx
)
2224 dsl_pool_t
*dp
= ds1
->ds_dir
->dd_pool
;
2225 dsl_scan_t
*scn
= dp
->dp_scan
;
2226 uint64_t mintxg1
, mintxg2
;
2227 boolean_t ds1_queued
, ds2_queued
;
2229 if (!dsl_scan_is_running(scn
))
2232 ds_clone_swapped_bookmark(ds1
, ds2
, &scn
->scn_phys
.scn_bookmark
);
2233 ds_clone_swapped_bookmark(ds1
, ds2
, &scn
->scn_phys_cached
.scn_bookmark
);
2236 * Handle the in-memory scan queue.
2238 ds1_queued
= scan_ds_queue_contains(scn
, ds1
->ds_object
, &mintxg1
);
2239 ds2_queued
= scan_ds_queue_contains(scn
, ds2
->ds_object
, &mintxg2
);
2241 /* Sanity checking. */
2243 ASSERT3U(mintxg1
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2244 ASSERT3U(mintxg1
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2247 ASSERT3U(mintxg2
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2248 ASSERT3U(mintxg2
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2251 if (ds1_queued
&& ds2_queued
) {
2253 * If both are queued, we don't need to do anything.
2254 * The swapping code below would not handle this case correctly,
2255 * since we can't insert ds2 if it is already there. That's
2256 * because scan_ds_queue_insert() prohibits a duplicate insert
2259 } else if (ds1_queued
) {
2260 scan_ds_queue_remove(scn
, ds1
->ds_object
);
2261 scan_ds_queue_insert(scn
, ds2
->ds_object
, mintxg1
);
2262 } else if (ds2_queued
) {
2263 scan_ds_queue_remove(scn
, ds2
->ds_object
);
2264 scan_ds_queue_insert(scn
, ds1
->ds_object
, mintxg2
);
2268 * Handle the on-disk scan queue.
2269 * The on-disk state is an out-of-date version of the in-memory state,
2270 * so the in-memory and on-disk values for ds1_queued and ds2_queued may
2271 * be different. Therefore we need to apply the swap logic to the
2272 * on-disk state independently of the in-memory state.
2274 ds1_queued
= zap_lookup_int_key(dp
->dp_meta_objset
,
2275 scn
->scn_phys
.scn_queue_obj
, ds1
->ds_object
, &mintxg1
) == 0;
2276 ds2_queued
= zap_lookup_int_key(dp
->dp_meta_objset
,
2277 scn
->scn_phys
.scn_queue_obj
, ds2
->ds_object
, &mintxg2
) == 0;
2279 /* Sanity checking. */
2281 ASSERT3U(mintxg1
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2282 ASSERT3U(mintxg1
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2285 ASSERT3U(mintxg2
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2286 ASSERT3U(mintxg2
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2289 if (ds1_queued
&& ds2_queued
) {
2291 * If both are queued, we don't need to do anything.
2292 * Alternatively, we could check for EEXIST from
2293 * zap_add_int_key() and back out to the original state, but
2294 * that would be more work than checking for this case upfront.
2296 } else if (ds1_queued
) {
2297 VERIFY3S(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2298 scn
->scn_phys
.scn_queue_obj
, ds1
->ds_object
, tx
));
2299 VERIFY3S(0, ==, zap_add_int_key(dp
->dp_meta_objset
,
2300 scn
->scn_phys
.scn_queue_obj
, ds2
->ds_object
, mintxg1
, tx
));
2301 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2302 "replacing with %llu",
2303 (u_longlong_t
)ds1
->ds_object
,
2304 (u_longlong_t
)ds2
->ds_object
);
2305 } else if (ds2_queued
) {
2306 VERIFY3S(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2307 scn
->scn_phys
.scn_queue_obj
, ds2
->ds_object
, tx
));
2308 VERIFY3S(0, ==, zap_add_int_key(dp
->dp_meta_objset
,
2309 scn
->scn_phys
.scn_queue_obj
, ds1
->ds_object
, mintxg2
, tx
));
2310 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2311 "replacing with %llu",
2312 (u_longlong_t
)ds2
->ds_object
,
2313 (u_longlong_t
)ds1
->ds_object
);
2316 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
2321 enqueue_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
2323 uint64_t originobj
= *(uint64_t *)arg
;
2326 dsl_scan_t
*scn
= dp
->dp_scan
;
2328 if (dsl_dir_phys(hds
->ds_dir
)->dd_origin_obj
!= originobj
)
2331 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
2335 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= originobj
) {
2336 dsl_dataset_t
*prev
;
2337 err
= dsl_dataset_hold_obj(dp
,
2338 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
2340 dsl_dataset_rele(ds
, FTAG
);
2345 scan_ds_queue_insert(scn
, ds
->ds_object
,
2346 dsl_dataset_phys(ds
)->ds_prev_snap_txg
);
2347 dsl_dataset_rele(ds
, FTAG
);
2352 dsl_scan_visitds(dsl_scan_t
*scn
, uint64_t dsobj
, dmu_tx_t
*tx
)
2354 dsl_pool_t
*dp
= scn
->scn_dp
;
2357 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
2359 if (scn
->scn_phys
.scn_cur_min_txg
>=
2360 scn
->scn_phys
.scn_max_txg
) {
2362 * This can happen if this snapshot was created after the
2363 * scan started, and we already completed a previous snapshot
2364 * that was created after the scan started. This snapshot
2365 * only references blocks with:
2367 * birth < our ds_creation_txg
2368 * cur_min_txg is no less than ds_creation_txg.
2369 * We have already visited these blocks.
2371 * birth > scn_max_txg
2372 * The scan requested not to visit these blocks.
2374 * Subsequent snapshots (and clones) can reference our
2375 * blocks, or blocks with even higher birth times.
2376 * Therefore we do not need to visit them either,
2377 * so we do not add them to the work queue.
2379 * Note that checking for cur_min_txg >= cur_max_txg
2380 * is not sufficient, because in that case we may need to
2381 * visit subsequent snapshots. This happens when min_txg > 0,
2382 * which raises cur_min_txg. In this case we will visit
2383 * this dataset but skip all of its blocks, because the
2384 * rootbp's birth time is < cur_min_txg. Then we will
2385 * add the next snapshots/clones to the work queue.
2387 char *dsname
= kmem_alloc(ZFS_MAX_DATASET_NAME_LEN
, KM_SLEEP
);
2388 dsl_dataset_name(ds
, dsname
);
2389 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2390 "cur_min_txg (%llu) >= max_txg (%llu)",
2391 (longlong_t
)dsobj
, dsname
,
2392 (longlong_t
)scn
->scn_phys
.scn_cur_min_txg
,
2393 (longlong_t
)scn
->scn_phys
.scn_max_txg
);
2394 kmem_free(dsname
, MAXNAMELEN
);
2400 * Only the ZIL in the head (non-snapshot) is valid. Even though
2401 * snapshots can have ZIL block pointers (which may be the same
2402 * BP as in the head), they must be ignored. In addition, $ORIGIN
2403 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2404 * need to look for a ZIL in it either. So we traverse the ZIL here,
2405 * rather than in scan_recurse(), because the regular snapshot
2406 * block-sharing rules don't apply to it.
2408 if (!dsl_dataset_is_snapshot(ds
) &&
2409 (dp
->dp_origin_snap
== NULL
||
2410 ds
->ds_dir
!= dp
->dp_origin_snap
->ds_dir
)) {
2412 if (dmu_objset_from_ds(ds
, &os
) != 0) {
2415 dsl_scan_zil(dp
, &os
->os_zil_header
);
2419 * Iterate over the bps in this ds.
2421 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
2422 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
2423 dsl_scan_visit_rootbp(scn
, ds
, &dsl_dataset_phys(ds
)->ds_bp
, tx
);
2424 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
2426 char *dsname
= kmem_alloc(ZFS_MAX_DATASET_NAME_LEN
, KM_SLEEP
);
2427 dsl_dataset_name(ds
, dsname
);
2428 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2430 (longlong_t
)dsobj
, dsname
,
2431 (longlong_t
)scn
->scn_phys
.scn_cur_min_txg
,
2432 (longlong_t
)scn
->scn_phys
.scn_cur_max_txg
,
2433 (int)scn
->scn_suspending
);
2434 kmem_free(dsname
, ZFS_MAX_DATASET_NAME_LEN
);
2436 if (scn
->scn_suspending
)
2440 * We've finished this pass over this dataset.
2444 * If we did not completely visit this dataset, do another pass.
2446 if (scn
->scn_phys
.scn_flags
& DSF_VISIT_DS_AGAIN
) {
2447 zfs_dbgmsg("incomplete pass; visiting again");
2448 scn
->scn_phys
.scn_flags
&= ~DSF_VISIT_DS_AGAIN
;
2449 scan_ds_queue_insert(scn
, ds
->ds_object
,
2450 scn
->scn_phys
.scn_cur_max_txg
);
2455 * Add descendant datasets to work queue.
2457 if (dsl_dataset_phys(ds
)->ds_next_snap_obj
!= 0) {
2458 scan_ds_queue_insert(scn
,
2459 dsl_dataset_phys(ds
)->ds_next_snap_obj
,
2460 dsl_dataset_phys(ds
)->ds_creation_txg
);
2462 if (dsl_dataset_phys(ds
)->ds_num_children
> 1) {
2463 boolean_t usenext
= B_FALSE
;
2464 if (dsl_dataset_phys(ds
)->ds_next_clones_obj
!= 0) {
2467 * A bug in a previous version of the code could
2468 * cause upgrade_clones_cb() to not set
2469 * ds_next_snap_obj when it should, leading to a
2470 * missing entry. Therefore we can only use the
2471 * next_clones_obj when its count is correct.
2473 int err
= zap_count(dp
->dp_meta_objset
,
2474 dsl_dataset_phys(ds
)->ds_next_clones_obj
, &count
);
2476 count
== dsl_dataset_phys(ds
)->ds_num_children
- 1)
2483 for (zap_cursor_init(&zc
, dp
->dp_meta_objset
,
2484 dsl_dataset_phys(ds
)->ds_next_clones_obj
);
2485 zap_cursor_retrieve(&zc
, &za
) == 0;
2486 (void) zap_cursor_advance(&zc
)) {
2487 scan_ds_queue_insert(scn
,
2488 zfs_strtonum(za
.za_name
, NULL
),
2489 dsl_dataset_phys(ds
)->ds_creation_txg
);
2491 zap_cursor_fini(&zc
);
2493 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
2494 enqueue_clones_cb
, &ds
->ds_object
,
2500 dsl_dataset_rele(ds
, FTAG
);
2505 enqueue_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
2509 dsl_scan_t
*scn
= dp
->dp_scan
;
2511 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
2515 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
2516 dsl_dataset_t
*prev
;
2517 err
= dsl_dataset_hold_obj(dp
,
2518 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
2520 dsl_dataset_rele(ds
, FTAG
);
2525 * If this is a clone, we don't need to worry about it for now.
2527 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
) {
2528 dsl_dataset_rele(ds
, FTAG
);
2529 dsl_dataset_rele(prev
, FTAG
);
2532 dsl_dataset_rele(ds
, FTAG
);
2536 scan_ds_queue_insert(scn
, ds
->ds_object
,
2537 dsl_dataset_phys(ds
)->ds_prev_snap_txg
);
2538 dsl_dataset_rele(ds
, FTAG
);
2544 dsl_scan_ddt_entry(dsl_scan_t
*scn
, enum zio_checksum checksum
,
2545 ddt_entry_t
*dde
, dmu_tx_t
*tx
)
2547 const ddt_key_t
*ddk
= &dde
->dde_key
;
2548 ddt_phys_t
*ddp
= dde
->dde_phys
;
2550 zbookmark_phys_t zb
= { 0 };
2553 if (!dsl_scan_is_running(scn
))
2557 * This function is special because it is the only thing
2558 * that can add scan_io_t's to the vdev scan queues from
2559 * outside dsl_scan_sync(). For the most part this is ok
2560 * as long as it is called from within syncing context.
2561 * However, dsl_scan_sync() expects that no new sio's will
2562 * be added between when all the work for a scan is done
2563 * and the next txg when the scan is actually marked as
2564 * completed. This check ensures we do not issue new sio's
2565 * during this period.
2567 if (scn
->scn_done_txg
!= 0)
2570 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2571 if (ddp
->ddp_phys_birth
== 0 ||
2572 ddp
->ddp_phys_birth
> scn
->scn_phys
.scn_max_txg
)
2574 ddt_bp_create(checksum
, ddk
, ddp
, &bp
);
2576 scn
->scn_visited_this_txg
++;
2577 scan_funcs
[scn
->scn_phys
.scn_func
](scn
->scn_dp
, &bp
, &zb
);
2582 * Scrub/dedup interaction.
2584 * If there are N references to a deduped block, we don't want to scrub it
2585 * N times -- ideally, we should scrub it exactly once.
2587 * We leverage the fact that the dde's replication class (enum ddt_class)
2588 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2589 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2591 * To prevent excess scrubbing, the scrub begins by walking the DDT
2592 * to find all blocks with refcnt > 1, and scrubs each of these once.
2593 * Since there are two replication classes which contain blocks with
2594 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2595 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2597 * There would be nothing more to say if a block's refcnt couldn't change
2598 * during a scrub, but of course it can so we must account for changes
2599 * in a block's replication class.
2601 * Here's an example of what can occur:
2603 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2604 * when visited during the top-down scrub phase, it will be scrubbed twice.
2605 * This negates our scrub optimization, but is otherwise harmless.
2607 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2608 * on each visit during the top-down scrub phase, it will never be scrubbed.
2609 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2610 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2611 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2612 * while a scrub is in progress, it scrubs the block right then.
2615 dsl_scan_ddt(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
2617 ddt_bookmark_t
*ddb
= &scn
->scn_phys
.scn_ddt_bookmark
;
2622 bzero(&dde
, sizeof (ddt_entry_t
));
2624 while ((error
= ddt_walk(scn
->scn_dp
->dp_spa
, ddb
, &dde
)) == 0) {
2627 if (ddb
->ddb_class
> scn
->scn_phys
.scn_ddt_class_max
)
2629 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
2630 (longlong_t
)ddb
->ddb_class
,
2631 (longlong_t
)ddb
->ddb_type
,
2632 (longlong_t
)ddb
->ddb_checksum
,
2633 (longlong_t
)ddb
->ddb_cursor
);
2635 /* There should be no pending changes to the dedup table */
2636 ddt
= scn
->scn_dp
->dp_spa
->spa_ddt
[ddb
->ddb_checksum
];
2637 ASSERT(avl_first(&ddt
->ddt_tree
) == NULL
);
2639 dsl_scan_ddt_entry(scn
, ddb
->ddb_checksum
, &dde
, tx
);
2642 if (dsl_scan_check_suspend(scn
, NULL
))
2646 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; "
2647 "suspending=%u", (longlong_t
)n
,
2648 (int)scn
->scn_phys
.scn_ddt_class_max
, (int)scn
->scn_suspending
);
2650 ASSERT(error
== 0 || error
== ENOENT
);
2651 ASSERT(error
!= ENOENT
||
2652 ddb
->ddb_class
> scn
->scn_phys
.scn_ddt_class_max
);
2656 dsl_scan_ds_maxtxg(dsl_dataset_t
*ds
)
2658 uint64_t smt
= ds
->ds_dir
->dd_pool
->dp_scan
->scn_phys
.scn_max_txg
;
2659 if (ds
->ds_is_snapshot
)
2660 return (MIN(smt
, dsl_dataset_phys(ds
)->ds_creation_txg
));
2665 dsl_scan_visit(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
2668 dsl_pool_t
*dp
= scn
->scn_dp
;
2670 if (scn
->scn_phys
.scn_ddt_bookmark
.ddb_class
<=
2671 scn
->scn_phys
.scn_ddt_class_max
) {
2672 scn
->scn_phys
.scn_cur_min_txg
= scn
->scn_phys
.scn_min_txg
;
2673 scn
->scn_phys
.scn_cur_max_txg
= scn
->scn_phys
.scn_max_txg
;
2674 dsl_scan_ddt(scn
, tx
);
2675 if (scn
->scn_suspending
)
2679 if (scn
->scn_phys
.scn_bookmark
.zb_objset
== DMU_META_OBJSET
) {
2680 /* First do the MOS & ORIGIN */
2682 scn
->scn_phys
.scn_cur_min_txg
= scn
->scn_phys
.scn_min_txg
;
2683 scn
->scn_phys
.scn_cur_max_txg
= scn
->scn_phys
.scn_max_txg
;
2684 dsl_scan_visit_rootbp(scn
, NULL
,
2685 &dp
->dp_meta_rootbp
, tx
);
2686 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
2687 if (scn
->scn_suspending
)
2690 if (spa_version(dp
->dp_spa
) < SPA_VERSION_DSL_SCRUB
) {
2691 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
2692 enqueue_cb
, NULL
, DS_FIND_CHILDREN
));
2694 dsl_scan_visitds(scn
,
2695 dp
->dp_origin_snap
->ds_object
, tx
);
2697 ASSERT(!scn
->scn_suspending
);
2698 } else if (scn
->scn_phys
.scn_bookmark
.zb_objset
!=
2699 ZB_DESTROYED_OBJSET
) {
2700 uint64_t dsobj
= scn
->scn_phys
.scn_bookmark
.zb_objset
;
2702 * If we were suspended, continue from here. Note if the
2703 * ds we were suspended on was deleted, the zb_objset may
2704 * be -1, so we will skip this and find a new objset
2707 dsl_scan_visitds(scn
, dsobj
, tx
);
2708 if (scn
->scn_suspending
)
2713 * In case we suspended right at the end of the ds, zero the
2714 * bookmark so we don't think that we're still trying to resume.
2716 bzero(&scn
->scn_phys
.scn_bookmark
, sizeof (zbookmark_phys_t
));
2719 * Keep pulling things out of the dataset avl queue. Updates to the
2720 * persistent zap-object-as-queue happen only at checkpoints.
2722 while ((sds
= avl_first(&scn
->scn_queue
)) != NULL
) {
2724 uint64_t dsobj
= sds
->sds_dsobj
;
2725 uint64_t txg
= sds
->sds_txg
;
2727 /* dequeue and free the ds from the queue */
2728 scan_ds_queue_remove(scn
, dsobj
);
2731 /* set up min / max txg */
2732 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
2734 scn
->scn_phys
.scn_cur_min_txg
=
2735 MAX(scn
->scn_phys
.scn_min_txg
, txg
);
2737 scn
->scn_phys
.scn_cur_min_txg
=
2738 MAX(scn
->scn_phys
.scn_min_txg
,
2739 dsl_dataset_phys(ds
)->ds_prev_snap_txg
);
2741 scn
->scn_phys
.scn_cur_max_txg
= dsl_scan_ds_maxtxg(ds
);
2742 dsl_dataset_rele(ds
, FTAG
);
2744 dsl_scan_visitds(scn
, dsobj
, tx
);
2745 if (scn
->scn_suspending
)
2749 /* No more objsets to fetch, we're done */
2750 scn
->scn_phys
.scn_bookmark
.zb_objset
= ZB_DESTROYED_OBJSET
;
2751 ASSERT0(scn
->scn_suspending
);
2755 dsl_scan_count_leaves(vdev_t
*vd
)
2757 uint64_t i
, leaves
= 0;
2759 /* we only count leaves that belong to the main pool and are readable */
2760 if (vd
->vdev_islog
|| vd
->vdev_isspare
||
2761 vd
->vdev_isl2cache
|| !vdev_readable(vd
))
2764 if (vd
->vdev_ops
->vdev_op_leaf
)
2767 for (i
= 0; i
< vd
->vdev_children
; i
++) {
2768 leaves
+= dsl_scan_count_leaves(vd
->vdev_child
[i
]);
2775 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t
*q
, const blkptr_t
*bp
)
2778 uint64_t cur_size
= 0;
2780 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
2781 cur_size
+= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
2784 q
->q_total_zio_size_this_txg
+= cur_size
;
2785 q
->q_zios_this_txg
++;
2789 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t
*q
, uint64_t start
,
2792 q
->q_total_seg_size_this_txg
+= end
- start
;
2793 q
->q_segs_this_txg
++;
2797 scan_io_queue_check_suspend(dsl_scan_t
*scn
)
2799 /* See comment in dsl_scan_check_suspend() */
2800 uint64_t curr_time_ns
= gethrtime();
2801 uint64_t scan_time_ns
= curr_time_ns
- scn
->scn_sync_start_time
;
2802 uint64_t sync_time_ns
= curr_time_ns
-
2803 scn
->scn_dp
->dp_spa
->spa_sync_starttime
;
2804 int dirty_pct
= scn
->scn_dp
->dp_dirty_total
* 100 / zfs_dirty_data_max
;
2805 int mintime
= (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) ?
2806 zfs_resilver_min_time_ms
: zfs_scrub_min_time_ms
;
2808 return ((NSEC2MSEC(scan_time_ns
) > mintime
&&
2809 (dirty_pct
>= zfs_vdev_async_write_active_min_dirty_percent
||
2810 txg_sync_waiting(scn
->scn_dp
) ||
2811 NSEC2SEC(sync_time_ns
) >= zfs_txg_timeout
)) ||
2812 spa_shutting_down(scn
->scn_dp
->dp_spa
));
2816 * Given a list of scan_io_t's in io_list, this issues the I/Os out to
2817 * disk. This consumes the io_list and frees the scan_io_t's. This is
2818 * called when emptying queues, either when we're up against the memory
2819 * limit or when we have finished scanning. Returns B_TRUE if we stopped
2820 * processing the list before we finished. Any sios that were not issued
2821 * will remain in the io_list.
2824 scan_io_queue_issue(dsl_scan_io_queue_t
*queue
, list_t
*io_list
)
2826 dsl_scan_t
*scn
= queue
->q_scn
;
2828 int64_t bytes_issued
= 0;
2829 boolean_t suspended
= B_FALSE
;
2831 while ((sio
= list_head(io_list
)) != NULL
) {
2834 if (scan_io_queue_check_suspend(scn
)) {
2840 bytes_issued
+= SIO_GET_ASIZE(sio
);
2841 scan_exec_io(scn
->scn_dp
, &bp
, sio
->sio_flags
,
2842 &sio
->sio_zb
, queue
);
2843 (void) list_remove_head(io_list
);
2844 scan_io_queues_update_zio_stats(queue
, &bp
);
2848 atomic_add_64(&scn
->scn_bytes_pending
, -bytes_issued
);
2854 * This function removes sios from an IO queue which reside within a given
2855 * range_seg_t and inserts them (in offset order) into a list. Note that
2856 * we only ever return a maximum of 32 sios at once. If there are more sios
2857 * to process within this segment that did not make it onto the list we
2858 * return B_TRUE and otherwise B_FALSE.
2861 scan_io_queue_gather(dsl_scan_io_queue_t
*queue
, range_seg_t
*rs
, list_t
*list
)
2863 scan_io_t
*srch_sio
, *sio
, *next_sio
;
2865 uint_t num_sios
= 0;
2866 int64_t bytes_issued
= 0;
2869 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
2871 srch_sio
= sio_alloc(1);
2872 srch_sio
->sio_nr_dvas
= 1;
2873 SIO_SET_OFFSET(srch_sio
, rs_get_start(rs
, queue
->q_exts_by_addr
));
2876 * The exact start of the extent might not contain any matching zios,
2877 * so if that's the case, examine the next one in the tree.
2879 sio
= avl_find(&queue
->q_sios_by_addr
, srch_sio
, &idx
);
2883 sio
= avl_nearest(&queue
->q_sios_by_addr
, idx
, AVL_AFTER
);
2885 while (sio
!= NULL
&& SIO_GET_OFFSET(sio
) < rs_get_end(rs
,
2886 queue
->q_exts_by_addr
) && num_sios
<= 32) {
2887 ASSERT3U(SIO_GET_OFFSET(sio
), >=, rs_get_start(rs
,
2888 queue
->q_exts_by_addr
));
2889 ASSERT3U(SIO_GET_END_OFFSET(sio
), <=, rs_get_end(rs
,
2890 queue
->q_exts_by_addr
));
2892 next_sio
= AVL_NEXT(&queue
->q_sios_by_addr
, sio
);
2893 avl_remove(&queue
->q_sios_by_addr
, sio
);
2894 queue
->q_sio_memused
-= SIO_GET_MUSED(sio
);
2896 bytes_issued
+= SIO_GET_ASIZE(sio
);
2898 list_insert_tail(list
, sio
);
2903 * We limit the number of sios we process at once to 32 to avoid
2904 * biting off more than we can chew. If we didn't take everything
2905 * in the segment we update it to reflect the work we were able to
2906 * complete. Otherwise, we remove it from the range tree entirely.
2908 if (sio
!= NULL
&& SIO_GET_OFFSET(sio
) < rs_get_end(rs
,
2909 queue
->q_exts_by_addr
)) {
2910 range_tree_adjust_fill(queue
->q_exts_by_addr
, rs
,
2912 range_tree_resize_segment(queue
->q_exts_by_addr
, rs
,
2913 SIO_GET_OFFSET(sio
), rs_get_end(rs
,
2914 queue
->q_exts_by_addr
) - SIO_GET_OFFSET(sio
));
2918 uint64_t rstart
= rs_get_start(rs
, queue
->q_exts_by_addr
);
2919 uint64_t rend
= rs_get_end(rs
, queue
->q_exts_by_addr
);
2920 range_tree_remove(queue
->q_exts_by_addr
, rstart
, rend
- rstart
);
2926 * This is called from the queue emptying thread and selects the next
2927 * extent from which we are to issue I/Os. The behavior of this function
2928 * depends on the state of the scan, the current memory consumption and
2929 * whether or not we are performing a scan shutdown.
2930 * 1) We select extents in an elevator algorithm (LBA-order) if the scan
2931 * needs to perform a checkpoint
2932 * 2) We select the largest available extent if we are up against the
2934 * 3) Otherwise we don't select any extents.
2936 static range_seg_t
*
2937 scan_io_queue_fetch_ext(dsl_scan_io_queue_t
*queue
)
2939 dsl_scan_t
*scn
= queue
->q_scn
;
2940 range_tree_t
*rt
= queue
->q_exts_by_addr
;
2942 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
2943 ASSERT(scn
->scn_is_sorted
);
2945 /* handle tunable overrides */
2946 if (scn
->scn_checkpointing
|| scn
->scn_clearing
) {
2947 if (zfs_scan_issue_strategy
== 1) {
2948 return (range_tree_first(rt
));
2949 } else if (zfs_scan_issue_strategy
== 2) {
2951 * We need to get the original entry in the by_addr
2952 * tree so we can modify it.
2954 range_seg_t
*size_rs
=
2955 zfs_btree_first(&queue
->q_exts_by_size
, NULL
);
2956 if (size_rs
== NULL
)
2958 uint64_t start
= rs_get_start(size_rs
, rt
);
2959 uint64_t size
= rs_get_end(size_rs
, rt
) - start
;
2960 range_seg_t
*addr_rs
= range_tree_find(rt
, start
,
2962 ASSERT3P(addr_rs
, !=, NULL
);
2963 ASSERT3U(rs_get_start(size_rs
, rt
), ==,
2964 rs_get_start(addr_rs
, rt
));
2965 ASSERT3U(rs_get_end(size_rs
, rt
), ==,
2966 rs_get_end(addr_rs
, rt
));
2972 * During normal clearing, we want to issue our largest segments
2973 * first, keeping IO as sequential as possible, and leaving the
2974 * smaller extents for later with the hope that they might eventually
2975 * grow to larger sequential segments. However, when the scan is
2976 * checkpointing, no new extents will be added to the sorting queue,
2977 * so the way we are sorted now is as good as it will ever get.
2978 * In this case, we instead switch to issuing extents in LBA order.
2980 if (scn
->scn_checkpointing
) {
2981 return (range_tree_first(rt
));
2982 } else if (scn
->scn_clearing
) {
2984 * We need to get the original entry in the by_addr
2985 * tree so we can modify it.
2987 range_seg_t
*size_rs
= zfs_btree_first(&queue
->q_exts_by_size
,
2989 if (size_rs
== NULL
)
2991 uint64_t start
= rs_get_start(size_rs
, rt
);
2992 uint64_t size
= rs_get_end(size_rs
, rt
) - start
;
2993 range_seg_t
*addr_rs
= range_tree_find(rt
, start
, size
);
2994 ASSERT3P(addr_rs
, !=, NULL
);
2995 ASSERT3U(rs_get_start(size_rs
, rt
), ==, rs_get_start(addr_rs
,
2997 ASSERT3U(rs_get_end(size_rs
, rt
), ==, rs_get_end(addr_rs
, rt
));
3005 scan_io_queues_run_one(void *arg
)
3007 dsl_scan_io_queue_t
*queue
= arg
;
3008 kmutex_t
*q_lock
= &queue
->q_vd
->vdev_scan_io_queue_lock
;
3009 boolean_t suspended
= B_FALSE
;
3010 range_seg_t
*rs
= NULL
;
3011 scan_io_t
*sio
= NULL
;
3013 uint64_t bytes_per_leaf
= zfs_scan_vdev_limit
;
3014 uint64_t nr_leaves
= dsl_scan_count_leaves(queue
->q_vd
);
3016 ASSERT(queue
->q_scn
->scn_is_sorted
);
3018 list_create(&sio_list
, sizeof (scan_io_t
),
3019 offsetof(scan_io_t
, sio_nodes
.sio_list_node
));
3020 mutex_enter(q_lock
);
3022 /* calculate maximum in-flight bytes for this txg (min 1MB) */
3023 queue
->q_maxinflight_bytes
=
3024 MAX(nr_leaves
* bytes_per_leaf
, 1ULL << 20);
3026 /* reset per-queue scan statistics for this txg */
3027 queue
->q_total_seg_size_this_txg
= 0;
3028 queue
->q_segs_this_txg
= 0;
3029 queue
->q_total_zio_size_this_txg
= 0;
3030 queue
->q_zios_this_txg
= 0;
3032 /* loop until we run out of time or sios */
3033 while ((rs
= scan_io_queue_fetch_ext(queue
)) != NULL
) {
3034 uint64_t seg_start
= 0, seg_end
= 0;
3035 boolean_t more_left
= B_TRUE
;
3037 ASSERT(list_is_empty(&sio_list
));
3039 /* loop while we still have sios left to process in this rs */
3041 scan_io_t
*first_sio
, *last_sio
;
3044 * We have selected which extent needs to be
3045 * processed next. Gather up the corresponding sios.
3047 more_left
= scan_io_queue_gather(queue
, rs
, &sio_list
);
3048 ASSERT(!list_is_empty(&sio_list
));
3049 first_sio
= list_head(&sio_list
);
3050 last_sio
= list_tail(&sio_list
);
3052 seg_end
= SIO_GET_END_OFFSET(last_sio
);
3054 seg_start
= SIO_GET_OFFSET(first_sio
);
3057 * Issuing sios can take a long time so drop the
3058 * queue lock. The sio queue won't be updated by
3059 * other threads since we're in syncing context so
3060 * we can be sure that our trees will remain exactly
3064 suspended
= scan_io_queue_issue(queue
, &sio_list
);
3065 mutex_enter(q_lock
);
3071 /* update statistics for debugging purposes */
3072 scan_io_queues_update_seg_stats(queue
, seg_start
, seg_end
);
3079 * If we were suspended in the middle of processing,
3080 * requeue any unfinished sios and exit.
3082 while ((sio
= list_head(&sio_list
)) != NULL
) {
3083 list_remove(&sio_list
, sio
);
3084 scan_io_queue_insert_impl(queue
, sio
);
3088 list_destroy(&sio_list
);
3092 * Performs an emptying run on all scan queues in the pool. This just
3093 * punches out one thread per top-level vdev, each of which processes
3094 * only that vdev's scan queue. We can parallelize the I/O here because
3095 * we know that each queue's I/Os only affect its own top-level vdev.
3097 * This function waits for the queue runs to complete, and must be
3098 * called from dsl_scan_sync (or in general, syncing context).
3101 scan_io_queues_run(dsl_scan_t
*scn
)
3103 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3105 ASSERT(scn
->scn_is_sorted
);
3106 ASSERT(spa_config_held(spa
, SCL_CONFIG
, RW_READER
));
3108 if (scn
->scn_bytes_pending
== 0)
3111 if (scn
->scn_taskq
== NULL
) {
3112 int nthreads
= spa
->spa_root_vdev
->vdev_children
;
3115 * We need to make this taskq *always* execute as many
3116 * threads in parallel as we have top-level vdevs and no
3117 * less, otherwise strange serialization of the calls to
3118 * scan_io_queues_run_one can occur during spa_sync runs
3119 * and that significantly impacts performance.
3121 scn
->scn_taskq
= taskq_create("dsl_scan_iss", nthreads
,
3122 minclsyspri
, nthreads
, nthreads
, TASKQ_PREPOPULATE
);
3125 for (uint64_t i
= 0; i
< spa
->spa_root_vdev
->vdev_children
; i
++) {
3126 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[i
];
3128 mutex_enter(&vd
->vdev_scan_io_queue_lock
);
3129 if (vd
->vdev_scan_io_queue
!= NULL
) {
3130 VERIFY(taskq_dispatch(scn
->scn_taskq
,
3131 scan_io_queues_run_one
, vd
->vdev_scan_io_queue
,
3132 TQ_SLEEP
) != TASKQID_INVALID
);
3134 mutex_exit(&vd
->vdev_scan_io_queue_lock
);
3138 * Wait for the queues to finish issuing their IOs for this run
3139 * before we return. There may still be IOs in flight at this
3142 taskq_wait(scn
->scn_taskq
);
3146 dsl_scan_async_block_should_pause(dsl_scan_t
*scn
)
3148 uint64_t elapsed_nanosecs
;
3153 if (zfs_async_block_max_blocks
!= 0 &&
3154 scn
->scn_visited_this_txg
>= zfs_async_block_max_blocks
) {
3158 if (zfs_max_async_dedup_frees
!= 0 &&
3159 scn
->scn_dedup_frees_this_txg
>= zfs_max_async_dedup_frees
) {
3163 elapsed_nanosecs
= gethrtime() - scn
->scn_sync_start_time
;
3164 return (elapsed_nanosecs
/ NANOSEC
> zfs_txg_timeout
||
3165 (NSEC2MSEC(elapsed_nanosecs
) > scn
->scn_async_block_min_time_ms
&&
3166 txg_sync_waiting(scn
->scn_dp
)) ||
3167 spa_shutting_down(scn
->scn_dp
->dp_spa
));
3171 dsl_scan_free_block_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
3173 dsl_scan_t
*scn
= arg
;
3175 if (!scn
->scn_is_bptree
||
3176 (BP_GET_LEVEL(bp
) == 0 && BP_GET_TYPE(bp
) != DMU_OT_OBJSET
)) {
3177 if (dsl_scan_async_block_should_pause(scn
))
3178 return (SET_ERROR(ERESTART
));
3181 zio_nowait(zio_free_sync(scn
->scn_zio_root
, scn
->scn_dp
->dp_spa
,
3182 dmu_tx_get_txg(tx
), bp
, 0));
3183 dsl_dir_diduse_space(tx
->tx_pool
->dp_free_dir
, DD_USED_HEAD
,
3184 -bp_get_dsize_sync(scn
->scn_dp
->dp_spa
, bp
),
3185 -BP_GET_PSIZE(bp
), -BP_GET_UCSIZE(bp
), tx
);
3186 scn
->scn_visited_this_txg
++;
3187 if (BP_GET_DEDUP(bp
))
3188 scn
->scn_dedup_frees_this_txg
++;
3193 dsl_scan_update_stats(dsl_scan_t
*scn
)
3195 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3197 uint64_t seg_size_total
= 0, zio_size_total
= 0;
3198 uint64_t seg_count_total
= 0, zio_count_total
= 0;
3200 for (i
= 0; i
< spa
->spa_root_vdev
->vdev_children
; i
++) {
3201 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[i
];
3202 dsl_scan_io_queue_t
*queue
= vd
->vdev_scan_io_queue
;
3207 seg_size_total
+= queue
->q_total_seg_size_this_txg
;
3208 zio_size_total
+= queue
->q_total_zio_size_this_txg
;
3209 seg_count_total
+= queue
->q_segs_this_txg
;
3210 zio_count_total
+= queue
->q_zios_this_txg
;
3213 if (seg_count_total
== 0 || zio_count_total
== 0) {
3214 scn
->scn_avg_seg_size_this_txg
= 0;
3215 scn
->scn_avg_zio_size_this_txg
= 0;
3216 scn
->scn_segs_this_txg
= 0;
3217 scn
->scn_zios_this_txg
= 0;
3221 scn
->scn_avg_seg_size_this_txg
= seg_size_total
/ seg_count_total
;
3222 scn
->scn_avg_zio_size_this_txg
= zio_size_total
/ zio_count_total
;
3223 scn
->scn_segs_this_txg
= seg_count_total
;
3224 scn
->scn_zios_this_txg
= zio_count_total
;
3228 bpobj_dsl_scan_free_block_cb(void *arg
, const blkptr_t
*bp
, boolean_t bp_freed
,
3232 return (dsl_scan_free_block_cb(arg
, bp
, tx
));
3236 dsl_scan_obsolete_block_cb(void *arg
, const blkptr_t
*bp
, boolean_t bp_freed
,
3240 dsl_scan_t
*scn
= arg
;
3241 const dva_t
*dva
= &bp
->blk_dva
[0];
3243 if (dsl_scan_async_block_should_pause(scn
))
3244 return (SET_ERROR(ERESTART
));
3246 spa_vdev_indirect_mark_obsolete(scn
->scn_dp
->dp_spa
,
3247 DVA_GET_VDEV(dva
), DVA_GET_OFFSET(dva
),
3248 DVA_GET_ASIZE(dva
), tx
);
3249 scn
->scn_visited_this_txg
++;
3254 dsl_scan_active(dsl_scan_t
*scn
)
3256 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3257 uint64_t used
= 0, comp
, uncomp
;
3258 boolean_t clones_left
;
3260 if (spa
->spa_load_state
!= SPA_LOAD_NONE
)
3262 if (spa_shutting_down(spa
))
3264 if ((dsl_scan_is_running(scn
) && !dsl_scan_is_paused_scrub(scn
)) ||
3265 (scn
->scn_async_destroying
&& !scn
->scn_async_stalled
))
3268 if (spa_version(scn
->scn_dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
3269 (void) bpobj_space(&scn
->scn_dp
->dp_free_bpobj
,
3270 &used
, &comp
, &uncomp
);
3272 clones_left
= spa_livelist_delete_check(spa
);
3273 return ((used
!= 0) || (clones_left
));
3277 dsl_scan_check_deferred(vdev_t
*vd
)
3279 boolean_t need_resilver
= B_FALSE
;
3281 for (int c
= 0; c
< vd
->vdev_children
; c
++) {
3283 dsl_scan_check_deferred(vd
->vdev_child
[c
]);
3286 if (!vdev_is_concrete(vd
) || vd
->vdev_aux
||
3287 !vd
->vdev_ops
->vdev_op_leaf
)
3288 return (need_resilver
);
3290 if (!vd
->vdev_resilver_deferred
)
3291 need_resilver
= B_TRUE
;
3293 return (need_resilver
);
3297 dsl_scan_need_resilver(spa_t
*spa
, const dva_t
*dva
, size_t psize
,
3298 uint64_t phys_birth
)
3302 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(dva
));
3304 if (vd
->vdev_ops
== &vdev_indirect_ops
) {
3306 * The indirect vdev can point to multiple
3307 * vdevs. For simplicity, always create
3308 * the resilver zio_t. zio_vdev_io_start()
3309 * will bypass the child resilver i/o's if
3310 * they are on vdevs that don't have DTL's.
3315 if (DVA_GET_GANG(dva
)) {
3317 * Gang members may be spread across multiple
3318 * vdevs, so the best estimate we have is the
3319 * scrub range, which has already been checked.
3320 * XXX -- it would be better to change our
3321 * allocation policy to ensure that all
3322 * gang members reside on the same vdev.
3328 * Check if the txg falls within the range which must be
3329 * resilvered. DVAs outside this range can always be skipped.
3331 if (!vdev_dtl_contains(vd
, DTL_PARTIAL
, phys_birth
, 1))
3335 * Check if the top-level vdev must resilver this offset.
3336 * When the offset does not intersect with a dirty leaf DTL
3337 * then it may be possible to skip the resilver IO. The psize
3338 * is provided instead of asize to simplify the check for RAIDZ.
3340 if (!vdev_dtl_need_resilver(vd
, DVA_GET_OFFSET(dva
), psize
))
3344 * Check that this top-level vdev has a device under it which
3345 * is resilvering and is not deferred.
3347 if (!dsl_scan_check_deferred(vd
))
3354 dsl_process_async_destroys(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
3356 dsl_scan_t
*scn
= dp
->dp_scan
;
3357 spa_t
*spa
= dp
->dp_spa
;
3360 if (spa_suspend_async_destroy(spa
))
3363 if (zfs_free_bpobj_enabled
&&
3364 spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
3365 scn
->scn_is_bptree
= B_FALSE
;
3366 scn
->scn_async_block_min_time_ms
= zfs_free_min_time_ms
;
3367 scn
->scn_zio_root
= zio_root(spa
, NULL
,
3368 NULL
, ZIO_FLAG_MUSTSUCCEED
);
3369 err
= bpobj_iterate(&dp
->dp_free_bpobj
,
3370 bpobj_dsl_scan_free_block_cb
, scn
, tx
);
3371 VERIFY0(zio_wait(scn
->scn_zio_root
));
3372 scn
->scn_zio_root
= NULL
;
3374 if (err
!= 0 && err
!= ERESTART
)
3375 zfs_panic_recover("error %u from bpobj_iterate()", err
);
3378 if (err
== 0 && spa_feature_is_active(spa
, SPA_FEATURE_ASYNC_DESTROY
)) {
3379 ASSERT(scn
->scn_async_destroying
);
3380 scn
->scn_is_bptree
= B_TRUE
;
3381 scn
->scn_zio_root
= zio_root(spa
, NULL
,
3382 NULL
, ZIO_FLAG_MUSTSUCCEED
);
3383 err
= bptree_iterate(dp
->dp_meta_objset
,
3384 dp
->dp_bptree_obj
, B_TRUE
, dsl_scan_free_block_cb
, scn
, tx
);
3385 VERIFY0(zio_wait(scn
->scn_zio_root
));
3386 scn
->scn_zio_root
= NULL
;
3388 if (err
== EIO
|| err
== ECKSUM
) {
3390 } else if (err
!= 0 && err
!= ERESTART
) {
3391 zfs_panic_recover("error %u from "
3392 "traverse_dataset_destroyed()", err
);
3395 if (bptree_is_empty(dp
->dp_meta_objset
, dp
->dp_bptree_obj
)) {
3396 /* finished; deactivate async destroy feature */
3397 spa_feature_decr(spa
, SPA_FEATURE_ASYNC_DESTROY
, tx
);
3398 ASSERT(!spa_feature_is_active(spa
,
3399 SPA_FEATURE_ASYNC_DESTROY
));
3400 VERIFY0(zap_remove(dp
->dp_meta_objset
,
3401 DMU_POOL_DIRECTORY_OBJECT
,
3402 DMU_POOL_BPTREE_OBJ
, tx
));
3403 VERIFY0(bptree_free(dp
->dp_meta_objset
,
3404 dp
->dp_bptree_obj
, tx
));
3405 dp
->dp_bptree_obj
= 0;
3406 scn
->scn_async_destroying
= B_FALSE
;
3407 scn
->scn_async_stalled
= B_FALSE
;
3410 * If we didn't make progress, mark the async
3411 * destroy as stalled, so that we will not initiate
3412 * a spa_sync() on its behalf. Note that we only
3413 * check this if we are not finished, because if the
3414 * bptree had no blocks for us to visit, we can
3415 * finish without "making progress".
3417 scn
->scn_async_stalled
=
3418 (scn
->scn_visited_this_txg
== 0);
3421 if (scn
->scn_visited_this_txg
) {
3422 zfs_dbgmsg("freed %llu blocks in %llums from "
3423 "free_bpobj/bptree txg %llu; err=%u",
3424 (longlong_t
)scn
->scn_visited_this_txg
,
3426 NSEC2MSEC(gethrtime() - scn
->scn_sync_start_time
),
3427 (longlong_t
)tx
->tx_txg
, err
);
3428 scn
->scn_visited_this_txg
= 0;
3429 scn
->scn_dedup_frees_this_txg
= 0;
3432 * Write out changes to the DDT that may be required as a
3433 * result of the blocks freed. This ensures that the DDT
3434 * is clean when a scrub/resilver runs.
3436 ddt_sync(spa
, tx
->tx_txg
);
3440 if (dp
->dp_free_dir
!= NULL
&& !scn
->scn_async_destroying
&&
3441 zfs_free_leak_on_eio
&&
3442 (dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
!= 0 ||
3443 dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
!= 0 ||
3444 dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
!= 0)) {
3446 * We have finished background destroying, but there is still
3447 * some space left in the dp_free_dir. Transfer this leaked
3448 * space to the dp_leak_dir.
3450 if (dp
->dp_leak_dir
== NULL
) {
3451 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
3452 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
3454 VERIFY0(dsl_pool_open_special_dir(dp
,
3455 LEAK_DIR_NAME
, &dp
->dp_leak_dir
));
3456 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
3458 dsl_dir_diduse_space(dp
->dp_leak_dir
, DD_USED_HEAD
,
3459 dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
,
3460 dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
,
3461 dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
, tx
);
3462 dsl_dir_diduse_space(dp
->dp_free_dir
, DD_USED_HEAD
,
3463 -dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
,
3464 -dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
,
3465 -dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
, tx
);
3468 if (dp
->dp_free_dir
!= NULL
&& !scn
->scn_async_destroying
&&
3469 !spa_livelist_delete_check(spa
)) {
3470 /* finished; verify that space accounting went to zero */
3471 ASSERT0(dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
);
3472 ASSERT0(dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
);
3473 ASSERT0(dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
);
3476 spa_notify_waiters(spa
);
3478 EQUIV(bpobj_is_open(&dp
->dp_obsolete_bpobj
),
3479 0 == zap_contains(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
3480 DMU_POOL_OBSOLETE_BPOBJ
));
3481 if (err
== 0 && bpobj_is_open(&dp
->dp_obsolete_bpobj
)) {
3482 ASSERT(spa_feature_is_active(dp
->dp_spa
,
3483 SPA_FEATURE_OBSOLETE_COUNTS
));
3485 scn
->scn_is_bptree
= B_FALSE
;
3486 scn
->scn_async_block_min_time_ms
= zfs_obsolete_min_time_ms
;
3487 err
= bpobj_iterate(&dp
->dp_obsolete_bpobj
,
3488 dsl_scan_obsolete_block_cb
, scn
, tx
);
3489 if (err
!= 0 && err
!= ERESTART
)
3490 zfs_panic_recover("error %u from bpobj_iterate()", err
);
3492 if (bpobj_is_empty(&dp
->dp_obsolete_bpobj
))
3493 dsl_pool_destroy_obsolete_bpobj(dp
, tx
);
3499 * This is the primary entry point for scans that is called from syncing
3500 * context. Scans must happen entirely during syncing context so that we
3501 * can guarantee that blocks we are currently scanning will not change out
3502 * from under us. While a scan is active, this function controls how quickly
3503 * transaction groups proceed, instead of the normal handling provided by
3504 * txg_sync_thread().
3507 dsl_scan_sync(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
3510 dsl_scan_t
*scn
= dp
->dp_scan
;
3511 spa_t
*spa
= dp
->dp_spa
;
3512 state_sync_type_t sync_type
= SYNC_OPTIONAL
;
3514 if (spa
->spa_resilver_deferred
&&
3515 !spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_RESILVER_DEFER
))
3516 spa_feature_incr(spa
, SPA_FEATURE_RESILVER_DEFER
, tx
);
3519 * Check for scn_restart_txg before checking spa_load_state, so
3520 * that we can restart an old-style scan while the pool is being
3521 * imported (see dsl_scan_init). We also restart scans if there
3522 * is a deferred resilver and the user has manually disabled
3523 * deferred resilvers via the tunable.
3525 if (dsl_scan_restarting(scn
, tx
) ||
3526 (spa
->spa_resilver_deferred
&& zfs_resilver_disable_defer
)) {
3527 pool_scan_func_t func
= POOL_SCAN_SCRUB
;
3528 dsl_scan_done(scn
, B_FALSE
, tx
);
3529 if (vdev_resilver_needed(spa
->spa_root_vdev
, NULL
, NULL
))
3530 func
= POOL_SCAN_RESILVER
;
3531 zfs_dbgmsg("restarting scan func=%u txg=%llu",
3532 func
, (longlong_t
)tx
->tx_txg
);
3533 dsl_scan_setup_sync(&func
, tx
);
3537 * Only process scans in sync pass 1.
3539 if (spa_sync_pass(spa
) > 1)
3543 * If the spa is shutting down, then stop scanning. This will
3544 * ensure that the scan does not dirty any new data during the
3547 if (spa_shutting_down(spa
))
3551 * If the scan is inactive due to a stalled async destroy, try again.
3553 if (!scn
->scn_async_stalled
&& !dsl_scan_active(scn
))
3556 /* reset scan statistics */
3557 scn
->scn_visited_this_txg
= 0;
3558 scn
->scn_dedup_frees_this_txg
= 0;
3559 scn
->scn_holes_this_txg
= 0;
3560 scn
->scn_lt_min_this_txg
= 0;
3561 scn
->scn_gt_max_this_txg
= 0;
3562 scn
->scn_ddt_contained_this_txg
= 0;
3563 scn
->scn_objsets_visited_this_txg
= 0;
3564 scn
->scn_avg_seg_size_this_txg
= 0;
3565 scn
->scn_segs_this_txg
= 0;
3566 scn
->scn_avg_zio_size_this_txg
= 0;
3567 scn
->scn_zios_this_txg
= 0;
3568 scn
->scn_suspending
= B_FALSE
;
3569 scn
->scn_sync_start_time
= gethrtime();
3570 spa
->spa_scrub_active
= B_TRUE
;
3573 * First process the async destroys. If we suspend, don't do
3574 * any scrubbing or resilvering. This ensures that there are no
3575 * async destroys while we are scanning, so the scan code doesn't
3576 * have to worry about traversing it. It is also faster to free the
3577 * blocks than to scrub them.
3579 err
= dsl_process_async_destroys(dp
, tx
);
3583 if (!dsl_scan_is_running(scn
) || dsl_scan_is_paused_scrub(scn
))
3587 * Wait a few txgs after importing to begin scanning so that
3588 * we can get the pool imported quickly.
3590 if (spa
->spa_syncing_txg
< spa
->spa_first_txg
+ SCAN_IMPORT_WAIT_TXGS
)
3594 * zfs_scan_suspend_progress can be set to disable scan progress.
3595 * We don't want to spin the txg_sync thread, so we add a delay
3596 * here to simulate the time spent doing a scan. This is mostly
3597 * useful for testing and debugging.
3599 if (zfs_scan_suspend_progress
) {
3600 uint64_t scan_time_ns
= gethrtime() - scn
->scn_sync_start_time
;
3601 int mintime
= (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) ?
3602 zfs_resilver_min_time_ms
: zfs_scrub_min_time_ms
;
3604 while (zfs_scan_suspend_progress
&&
3605 !txg_sync_waiting(scn
->scn_dp
) &&
3606 !spa_shutting_down(scn
->scn_dp
->dp_spa
) &&
3607 NSEC2MSEC(scan_time_ns
) < mintime
) {
3609 scan_time_ns
= gethrtime() - scn
->scn_sync_start_time
;
3615 * It is possible to switch from unsorted to sorted at any time,
3616 * but afterwards the scan will remain sorted unless reloaded from
3617 * a checkpoint after a reboot.
3619 if (!zfs_scan_legacy
) {
3620 scn
->scn_is_sorted
= B_TRUE
;
3621 if (scn
->scn_last_checkpoint
== 0)
3622 scn
->scn_last_checkpoint
= ddi_get_lbolt();
3626 * For sorted scans, determine what kind of work we will be doing
3627 * this txg based on our memory limitations and whether or not we
3628 * need to perform a checkpoint.
3630 if (scn
->scn_is_sorted
) {
3632 * If we are over our checkpoint interval, set scn_clearing
3633 * so that we can begin checkpointing immediately. The
3634 * checkpoint allows us to save a consistent bookmark
3635 * representing how much data we have scrubbed so far.
3636 * Otherwise, use the memory limit to determine if we should
3637 * scan for metadata or start issue scrub IOs. We accumulate
3638 * metadata until we hit our hard memory limit at which point
3639 * we issue scrub IOs until we are at our soft memory limit.
3641 if (scn
->scn_checkpointing
||
3642 ddi_get_lbolt() - scn
->scn_last_checkpoint
>
3643 SEC_TO_TICK(zfs_scan_checkpoint_intval
)) {
3644 if (!scn
->scn_checkpointing
)
3645 zfs_dbgmsg("begin scan checkpoint");
3647 scn
->scn_checkpointing
= B_TRUE
;
3648 scn
->scn_clearing
= B_TRUE
;
3650 boolean_t should_clear
= dsl_scan_should_clear(scn
);
3651 if (should_clear
&& !scn
->scn_clearing
) {
3652 zfs_dbgmsg("begin scan clearing");
3653 scn
->scn_clearing
= B_TRUE
;
3654 } else if (!should_clear
&& scn
->scn_clearing
) {
3655 zfs_dbgmsg("finish scan clearing");
3656 scn
->scn_clearing
= B_FALSE
;
3660 ASSERT0(scn
->scn_checkpointing
);
3661 ASSERT0(scn
->scn_clearing
);
3664 if (!scn
->scn_clearing
&& scn
->scn_done_txg
== 0) {
3665 /* Need to scan metadata for more blocks to scrub */
3666 dsl_scan_phys_t
*scnp
= &scn
->scn_phys
;
3667 taskqid_t prefetch_tqid
;
3668 uint64_t bytes_per_leaf
= zfs_scan_vdev_limit
;
3669 uint64_t nr_leaves
= dsl_scan_count_leaves(spa
->spa_root_vdev
);
3672 * Recalculate the max number of in-flight bytes for pool-wide
3673 * scanning operations (minimum 1MB). Limits for the issuing
3674 * phase are done per top-level vdev and are handled separately.
3676 scn
->scn_maxinflight_bytes
=
3677 MAX(nr_leaves
* bytes_per_leaf
, 1ULL << 20);
3679 if (scnp
->scn_ddt_bookmark
.ddb_class
<=
3680 scnp
->scn_ddt_class_max
) {
3681 ASSERT(ZB_IS_ZERO(&scnp
->scn_bookmark
));
3682 zfs_dbgmsg("doing scan sync txg %llu; "
3683 "ddt bm=%llu/%llu/%llu/%llx",
3684 (longlong_t
)tx
->tx_txg
,
3685 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_class
,
3686 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_type
,
3687 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_checksum
,
3688 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_cursor
);
3690 zfs_dbgmsg("doing scan sync txg %llu; "
3691 "bm=%llu/%llu/%llu/%llu",
3692 (longlong_t
)tx
->tx_txg
,
3693 (longlong_t
)scnp
->scn_bookmark
.zb_objset
,
3694 (longlong_t
)scnp
->scn_bookmark
.zb_object
,
3695 (longlong_t
)scnp
->scn_bookmark
.zb_level
,
3696 (longlong_t
)scnp
->scn_bookmark
.zb_blkid
);
3699 scn
->scn_zio_root
= zio_root(dp
->dp_spa
, NULL
,
3700 NULL
, ZIO_FLAG_CANFAIL
);
3702 scn
->scn_prefetch_stop
= B_FALSE
;
3703 prefetch_tqid
= taskq_dispatch(dp
->dp_sync_taskq
,
3704 dsl_scan_prefetch_thread
, scn
, TQ_SLEEP
);
3705 ASSERT(prefetch_tqid
!= TASKQID_INVALID
);
3707 dsl_pool_config_enter(dp
, FTAG
);
3708 dsl_scan_visit(scn
, tx
);
3709 dsl_pool_config_exit(dp
, FTAG
);
3711 mutex_enter(&dp
->dp_spa
->spa_scrub_lock
);
3712 scn
->scn_prefetch_stop
= B_TRUE
;
3713 cv_broadcast(&spa
->spa_scrub_io_cv
);
3714 mutex_exit(&dp
->dp_spa
->spa_scrub_lock
);
3716 taskq_wait_id(dp
->dp_sync_taskq
, prefetch_tqid
);
3717 (void) zio_wait(scn
->scn_zio_root
);
3718 scn
->scn_zio_root
= NULL
;
3720 zfs_dbgmsg("scan visited %llu blocks in %llums "
3721 "(%llu os's, %llu holes, %llu < mintxg, "
3722 "%llu in ddt, %llu > maxtxg)",
3723 (longlong_t
)scn
->scn_visited_this_txg
,
3724 (longlong_t
)NSEC2MSEC(gethrtime() -
3725 scn
->scn_sync_start_time
),
3726 (longlong_t
)scn
->scn_objsets_visited_this_txg
,
3727 (longlong_t
)scn
->scn_holes_this_txg
,
3728 (longlong_t
)scn
->scn_lt_min_this_txg
,
3729 (longlong_t
)scn
->scn_ddt_contained_this_txg
,
3730 (longlong_t
)scn
->scn_gt_max_this_txg
);
3732 if (!scn
->scn_suspending
) {
3733 ASSERT0(avl_numnodes(&scn
->scn_queue
));
3734 scn
->scn_done_txg
= tx
->tx_txg
+ 1;
3735 if (scn
->scn_is_sorted
) {
3736 scn
->scn_checkpointing
= B_TRUE
;
3737 scn
->scn_clearing
= B_TRUE
;
3739 zfs_dbgmsg("scan complete txg %llu",
3740 (longlong_t
)tx
->tx_txg
);
3742 } else if (scn
->scn_is_sorted
&& scn
->scn_bytes_pending
!= 0) {
3743 ASSERT(scn
->scn_clearing
);
3745 /* need to issue scrubbing IOs from per-vdev queues */
3746 scn
->scn_zio_root
= zio_root(dp
->dp_spa
, NULL
,
3747 NULL
, ZIO_FLAG_CANFAIL
);
3748 scan_io_queues_run(scn
);
3749 (void) zio_wait(scn
->scn_zio_root
);
3750 scn
->scn_zio_root
= NULL
;
3752 /* calculate and dprintf the current memory usage */
3753 (void) dsl_scan_should_clear(scn
);
3754 dsl_scan_update_stats(scn
);
3756 zfs_dbgmsg("scan issued %llu blocks (%llu segs) in %llums "
3757 "(avg_block_size = %llu, avg_seg_size = %llu)",
3758 (longlong_t
)scn
->scn_zios_this_txg
,
3759 (longlong_t
)scn
->scn_segs_this_txg
,
3760 (longlong_t
)NSEC2MSEC(gethrtime() -
3761 scn
->scn_sync_start_time
),
3762 (longlong_t
)scn
->scn_avg_zio_size_this_txg
,
3763 (longlong_t
)scn
->scn_avg_seg_size_this_txg
);
3764 } else if (scn
->scn_done_txg
!= 0 && scn
->scn_done_txg
<= tx
->tx_txg
) {
3765 /* Finished with everything. Mark the scrub as complete */
3766 zfs_dbgmsg("scan issuing complete txg %llu",
3767 (longlong_t
)tx
->tx_txg
);
3768 ASSERT3U(scn
->scn_done_txg
, !=, 0);
3769 ASSERT0(spa
->spa_scrub_inflight
);
3770 ASSERT0(scn
->scn_bytes_pending
);
3771 dsl_scan_done(scn
, B_TRUE
, tx
);
3772 sync_type
= SYNC_MANDATORY
;
3775 dsl_scan_sync_state(scn
, tx
, sync_type
);
3779 count_block(dsl_scan_t
*scn
, zfs_all_blkstats_t
*zab
, const blkptr_t
*bp
)
3784 * Don't count embedded bp's, since we already did the work of
3785 * scanning these when we scanned the containing block.
3787 if (BP_IS_EMBEDDED(bp
))
3791 * Update the spa's stats on how many bytes we have issued.
3792 * Sequential scrubs create a zio for each DVA of the bp. Each
3793 * of these will include all DVAs for repair purposes, but the
3794 * zio code will only try the first one unless there is an issue.
3795 * Therefore, we should only count the first DVA for these IOs.
3797 if (scn
->scn_is_sorted
) {
3798 atomic_add_64(&scn
->scn_dp
->dp_spa
->spa_scan_pass_issued
,
3799 DVA_GET_ASIZE(&bp
->blk_dva
[0]));
3801 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3803 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
3804 atomic_add_64(&spa
->spa_scan_pass_issued
,
3805 DVA_GET_ASIZE(&bp
->blk_dva
[i
]));
3810 * If we resume after a reboot, zab will be NULL; don't record
3811 * incomplete stats in that case.
3816 mutex_enter(&zab
->zab_lock
);
3818 for (i
= 0; i
< 4; i
++) {
3819 int l
= (i
< 2) ? BP_GET_LEVEL(bp
) : DN_MAX_LEVELS
;
3820 int t
= (i
& 1) ? BP_GET_TYPE(bp
) : DMU_OT_TOTAL
;
3822 if (t
& DMU_OT_NEWTYPE
)
3824 zfs_blkstat_t
*zb
= &zab
->zab_type
[l
][t
];
3828 zb
->zb_asize
+= BP_GET_ASIZE(bp
);
3829 zb
->zb_lsize
+= BP_GET_LSIZE(bp
);
3830 zb
->zb_psize
+= BP_GET_PSIZE(bp
);
3831 zb
->zb_gangs
+= BP_COUNT_GANG(bp
);
3833 switch (BP_GET_NDVAS(bp
)) {
3835 if (DVA_GET_VDEV(&bp
->blk_dva
[0]) ==
3836 DVA_GET_VDEV(&bp
->blk_dva
[1]))
3837 zb
->zb_ditto_2_of_2_samevdev
++;
3840 equal
= (DVA_GET_VDEV(&bp
->blk_dva
[0]) ==
3841 DVA_GET_VDEV(&bp
->blk_dva
[1])) +
3842 (DVA_GET_VDEV(&bp
->blk_dva
[0]) ==
3843 DVA_GET_VDEV(&bp
->blk_dva
[2])) +
3844 (DVA_GET_VDEV(&bp
->blk_dva
[1]) ==
3845 DVA_GET_VDEV(&bp
->blk_dva
[2]));
3847 zb
->zb_ditto_2_of_3_samevdev
++;
3848 else if (equal
== 3)
3849 zb
->zb_ditto_3_of_3_samevdev
++;
3854 mutex_exit(&zab
->zab_lock
);
3858 scan_io_queue_insert_impl(dsl_scan_io_queue_t
*queue
, scan_io_t
*sio
)
3861 int64_t asize
= SIO_GET_ASIZE(sio
);
3862 dsl_scan_t
*scn
= queue
->q_scn
;
3864 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
3866 if (avl_find(&queue
->q_sios_by_addr
, sio
, &idx
) != NULL
) {
3867 /* block is already scheduled for reading */
3868 atomic_add_64(&scn
->scn_bytes_pending
, -asize
);
3872 avl_insert(&queue
->q_sios_by_addr
, sio
, idx
);
3873 queue
->q_sio_memused
+= SIO_GET_MUSED(sio
);
3874 range_tree_add(queue
->q_exts_by_addr
, SIO_GET_OFFSET(sio
), asize
);
3878 * Given all the info we got from our metadata scanning process, we
3879 * construct a scan_io_t and insert it into the scan sorting queue. The
3880 * I/O must already be suitable for us to process. This is controlled
3881 * by dsl_scan_enqueue().
3884 scan_io_queue_insert(dsl_scan_io_queue_t
*queue
, const blkptr_t
*bp
, int dva_i
,
3885 int zio_flags
, const zbookmark_phys_t
*zb
)
3887 dsl_scan_t
*scn
= queue
->q_scn
;
3888 scan_io_t
*sio
= sio_alloc(BP_GET_NDVAS(bp
));
3890 ASSERT0(BP_IS_GANG(bp
));
3891 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
3893 bp2sio(bp
, sio
, dva_i
);
3894 sio
->sio_flags
= zio_flags
;
3898 * Increment the bytes pending counter now so that we can't
3899 * get an integer underflow in case the worker processes the
3900 * zio before we get to incrementing this counter.
3902 atomic_add_64(&scn
->scn_bytes_pending
, SIO_GET_ASIZE(sio
));
3904 scan_io_queue_insert_impl(queue
, sio
);
3908 * Given a set of I/O parameters as discovered by the metadata traversal
3909 * process, attempts to place the I/O into the sorted queues (if allowed),
3910 * or immediately executes the I/O.
3913 dsl_scan_enqueue(dsl_pool_t
*dp
, const blkptr_t
*bp
, int zio_flags
,
3914 const zbookmark_phys_t
*zb
)
3916 spa_t
*spa
= dp
->dp_spa
;
3918 ASSERT(!BP_IS_EMBEDDED(bp
));
3921 * Gang blocks are hard to issue sequentially, so we just issue them
3922 * here immediately instead of queuing them.
3924 if (!dp
->dp_scan
->scn_is_sorted
|| BP_IS_GANG(bp
)) {
3925 scan_exec_io(dp
, bp
, zio_flags
, zb
, NULL
);
3929 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
3933 dva
= bp
->blk_dva
[i
];
3934 vdev
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
));
3935 ASSERT(vdev
!= NULL
);
3937 mutex_enter(&vdev
->vdev_scan_io_queue_lock
);
3938 if (vdev
->vdev_scan_io_queue
== NULL
)
3939 vdev
->vdev_scan_io_queue
= scan_io_queue_create(vdev
);
3940 ASSERT(dp
->dp_scan
!= NULL
);
3941 scan_io_queue_insert(vdev
->vdev_scan_io_queue
, bp
,
3943 mutex_exit(&vdev
->vdev_scan_io_queue_lock
);
3948 dsl_scan_scrub_cb(dsl_pool_t
*dp
,
3949 const blkptr_t
*bp
, const zbookmark_phys_t
*zb
)
3951 dsl_scan_t
*scn
= dp
->dp_scan
;
3952 spa_t
*spa
= dp
->dp_spa
;
3953 uint64_t phys_birth
= BP_PHYSICAL_BIRTH(bp
);
3954 size_t psize
= BP_GET_PSIZE(bp
);
3955 boolean_t needs_io
= B_FALSE
;
3956 int zio_flags
= ZIO_FLAG_SCAN_THREAD
| ZIO_FLAG_RAW
| ZIO_FLAG_CANFAIL
;
3959 if (phys_birth
<= scn
->scn_phys
.scn_min_txg
||
3960 phys_birth
>= scn
->scn_phys
.scn_max_txg
) {
3961 count_block(scn
, dp
->dp_blkstats
, bp
);
3965 /* Embedded BP's have phys_birth==0, so we reject them above. */
3966 ASSERT(!BP_IS_EMBEDDED(bp
));
3968 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn
));
3969 if (scn
->scn_phys
.scn_func
== POOL_SCAN_SCRUB
) {
3970 zio_flags
|= ZIO_FLAG_SCRUB
;
3973 ASSERT3U(scn
->scn_phys
.scn_func
, ==, POOL_SCAN_RESILVER
);
3974 zio_flags
|= ZIO_FLAG_RESILVER
;
3978 /* If it's an intent log block, failure is expected. */
3979 if (zb
->zb_level
== ZB_ZIL_LEVEL
)
3980 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
3982 for (int d
= 0; d
< BP_GET_NDVAS(bp
); d
++) {
3983 const dva_t
*dva
= &bp
->blk_dva
[d
];
3986 * Keep track of how much data we've examined so that
3987 * zpool(1M) status can make useful progress reports.
3989 scn
->scn_phys
.scn_examined
+= DVA_GET_ASIZE(dva
);
3990 spa
->spa_scan_pass_exam
+= DVA_GET_ASIZE(dva
);
3992 /* if it's a resilver, this may not be in the target range */
3994 needs_io
= dsl_scan_need_resilver(spa
, dva
, psize
,
3998 if (needs_io
&& !zfs_no_scrub_io
) {
3999 dsl_scan_enqueue(dp
, bp
, zio_flags
, zb
);
4001 count_block(scn
, dp
->dp_blkstats
, bp
);
4004 /* do not relocate this block */
4009 dsl_scan_scrub_done(zio_t
*zio
)
4011 spa_t
*spa
= zio
->io_spa
;
4012 blkptr_t
*bp
= zio
->io_bp
;
4013 dsl_scan_io_queue_t
*queue
= zio
->io_private
;
4015 abd_free(zio
->io_abd
);
4017 if (queue
== NULL
) {
4018 mutex_enter(&spa
->spa_scrub_lock
);
4019 ASSERT3U(spa
->spa_scrub_inflight
, >=, BP_GET_PSIZE(bp
));
4020 spa
->spa_scrub_inflight
-= BP_GET_PSIZE(bp
);
4021 cv_broadcast(&spa
->spa_scrub_io_cv
);
4022 mutex_exit(&spa
->spa_scrub_lock
);
4024 mutex_enter(&queue
->q_vd
->vdev_scan_io_queue_lock
);
4025 ASSERT3U(queue
->q_inflight_bytes
, >=, BP_GET_PSIZE(bp
));
4026 queue
->q_inflight_bytes
-= BP_GET_PSIZE(bp
);
4027 cv_broadcast(&queue
->q_zio_cv
);
4028 mutex_exit(&queue
->q_vd
->vdev_scan_io_queue_lock
);
4031 if (zio
->io_error
&& (zio
->io_error
!= ECKSUM
||
4032 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
))) {
4033 atomic_inc_64(&spa
->spa_dsl_pool
->dp_scan
->scn_phys
.scn_errors
);
4038 * Given a scanning zio's information, executes the zio. The zio need
4039 * not necessarily be only sortable, this function simply executes the
4040 * zio, no matter what it is. The optional queue argument allows the
4041 * caller to specify that they want per top level vdev IO rate limiting
4042 * instead of the legacy global limiting.
4045 scan_exec_io(dsl_pool_t
*dp
, const blkptr_t
*bp
, int zio_flags
,
4046 const zbookmark_phys_t
*zb
, dsl_scan_io_queue_t
*queue
)
4048 spa_t
*spa
= dp
->dp_spa
;
4049 dsl_scan_t
*scn
= dp
->dp_scan
;
4050 size_t size
= BP_GET_PSIZE(bp
);
4051 abd_t
*data
= abd_alloc_for_io(size
, B_FALSE
);
4053 ASSERT3U(scn
->scn_maxinflight_bytes
, >, 0);
4055 if (queue
== NULL
) {
4056 mutex_enter(&spa
->spa_scrub_lock
);
4057 while (spa
->spa_scrub_inflight
>= scn
->scn_maxinflight_bytes
)
4058 cv_wait(&spa
->spa_scrub_io_cv
, &spa
->spa_scrub_lock
);
4059 spa
->spa_scrub_inflight
+= BP_GET_PSIZE(bp
);
4060 mutex_exit(&spa
->spa_scrub_lock
);
4062 kmutex_t
*q_lock
= &queue
->q_vd
->vdev_scan_io_queue_lock
;
4064 mutex_enter(q_lock
);
4065 while (queue
->q_inflight_bytes
>= queue
->q_maxinflight_bytes
)
4066 cv_wait(&queue
->q_zio_cv
, q_lock
);
4067 queue
->q_inflight_bytes
+= BP_GET_PSIZE(bp
);
4071 count_block(scn
, dp
->dp_blkstats
, bp
);
4072 zio_nowait(zio_read(scn
->scn_zio_root
, spa
, bp
, data
, size
,
4073 dsl_scan_scrub_done
, queue
, ZIO_PRIORITY_SCRUB
, zio_flags
, zb
));
4077 * This is the primary extent sorting algorithm. We balance two parameters:
4078 * 1) how many bytes of I/O are in an extent
4079 * 2) how well the extent is filled with I/O (as a fraction of its total size)
4080 * Since we allow extents to have gaps between their constituent I/Os, it's
4081 * possible to have a fairly large extent that contains the same amount of
4082 * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
4083 * The algorithm sorts based on a score calculated from the extent's size,
4084 * the relative fill volume (in %) and a "fill weight" parameter that controls
4085 * the split between whether we prefer larger extents or more well populated
4088 * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
4091 * 1) assume extsz = 64 MiB
4092 * 2) assume fill = 32 MiB (extent is half full)
4093 * 3) assume fill_weight = 3
4094 * 4) SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
4095 * SCORE = 32M + (50 * 3 * 32M) / 100
4096 * SCORE = 32M + (4800M / 100)
4099 * | +--- final total relative fill-based score
4100 * +--------- final total fill-based score
4103 * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
4104 * extents that are more completely filled (in a 3:2 ratio) vs just larger.
4105 * Note that as an optimization, we replace multiplication and division by
4106 * 100 with bitshifting by 7 (which effectively multiplies and divides by 128).
4109 ext_size_compare(const void *x
, const void *y
)
4111 const range_seg_gap_t
*rsa
= x
, *rsb
= y
;
4113 uint64_t sa
= rsa
->rs_end
- rsa
->rs_start
;
4114 uint64_t sb
= rsb
->rs_end
- rsb
->rs_start
;
4115 uint64_t score_a
, score_b
;
4117 score_a
= rsa
->rs_fill
+ ((((rsa
->rs_fill
<< 7) / sa
) *
4118 fill_weight
* rsa
->rs_fill
) >> 7);
4119 score_b
= rsb
->rs_fill
+ ((((rsb
->rs_fill
<< 7) / sb
) *
4120 fill_weight
* rsb
->rs_fill
) >> 7);
4122 if (score_a
> score_b
)
4124 if (score_a
== score_b
) {
4125 if (rsa
->rs_start
< rsb
->rs_start
)
4127 if (rsa
->rs_start
== rsb
->rs_start
)
4135 * Comparator for the q_sios_by_addr tree. Sorting is simply performed
4136 * based on LBA-order (from lowest to highest).
4139 sio_addr_compare(const void *x
, const void *y
)
4141 const scan_io_t
*a
= x
, *b
= y
;
4143 return (TREE_CMP(SIO_GET_OFFSET(a
), SIO_GET_OFFSET(b
)));
4146 /* IO queues are created on demand when they are needed. */
4147 static dsl_scan_io_queue_t
*
4148 scan_io_queue_create(vdev_t
*vd
)
4150 dsl_scan_t
*scn
= vd
->vdev_spa
->spa_dsl_pool
->dp_scan
;
4151 dsl_scan_io_queue_t
*q
= kmem_zalloc(sizeof (*q
), KM_SLEEP
);
4155 q
->q_sio_memused
= 0;
4156 cv_init(&q
->q_zio_cv
, NULL
, CV_DEFAULT
, NULL
);
4157 q
->q_exts_by_addr
= range_tree_create_impl(&rt_btree_ops
, RANGE_SEG_GAP
,
4158 &q
->q_exts_by_size
, 0, 0, ext_size_compare
, zfs_scan_max_ext_gap
);
4159 avl_create(&q
->q_sios_by_addr
, sio_addr_compare
,
4160 sizeof (scan_io_t
), offsetof(scan_io_t
, sio_nodes
.sio_addr_node
));
4166 * Destroys a scan queue and all segments and scan_io_t's contained in it.
4167 * No further execution of I/O occurs, anything pending in the queue is
4168 * simply freed without being executed.
4171 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t
*queue
)
4173 dsl_scan_t
*scn
= queue
->q_scn
;
4175 void *cookie
= NULL
;
4176 int64_t bytes_dequeued
= 0;
4178 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
4180 while ((sio
= avl_destroy_nodes(&queue
->q_sios_by_addr
, &cookie
)) !=
4182 ASSERT(range_tree_contains(queue
->q_exts_by_addr
,
4183 SIO_GET_OFFSET(sio
), SIO_GET_ASIZE(sio
)));
4184 bytes_dequeued
+= SIO_GET_ASIZE(sio
);
4185 queue
->q_sio_memused
-= SIO_GET_MUSED(sio
);
4189 ASSERT0(queue
->q_sio_memused
);
4190 atomic_add_64(&scn
->scn_bytes_pending
, -bytes_dequeued
);
4191 range_tree_vacate(queue
->q_exts_by_addr
, NULL
, queue
);
4192 range_tree_destroy(queue
->q_exts_by_addr
);
4193 avl_destroy(&queue
->q_sios_by_addr
);
4194 cv_destroy(&queue
->q_zio_cv
);
4196 kmem_free(queue
, sizeof (*queue
));
4200 * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
4201 * called on behalf of vdev_top_transfer when creating or destroying
4202 * a mirror vdev due to zpool attach/detach.
4205 dsl_scan_io_queue_vdev_xfer(vdev_t
*svd
, vdev_t
*tvd
)
4207 mutex_enter(&svd
->vdev_scan_io_queue_lock
);
4208 mutex_enter(&tvd
->vdev_scan_io_queue_lock
);
4210 VERIFY3P(tvd
->vdev_scan_io_queue
, ==, NULL
);
4211 tvd
->vdev_scan_io_queue
= svd
->vdev_scan_io_queue
;
4212 svd
->vdev_scan_io_queue
= NULL
;
4213 if (tvd
->vdev_scan_io_queue
!= NULL
)
4214 tvd
->vdev_scan_io_queue
->q_vd
= tvd
;
4216 mutex_exit(&tvd
->vdev_scan_io_queue_lock
);
4217 mutex_exit(&svd
->vdev_scan_io_queue_lock
);
4221 scan_io_queues_destroy(dsl_scan_t
*scn
)
4223 vdev_t
*rvd
= scn
->scn_dp
->dp_spa
->spa_root_vdev
;
4225 for (uint64_t i
= 0; i
< rvd
->vdev_children
; i
++) {
4226 vdev_t
*tvd
= rvd
->vdev_child
[i
];
4228 mutex_enter(&tvd
->vdev_scan_io_queue_lock
);
4229 if (tvd
->vdev_scan_io_queue
!= NULL
)
4230 dsl_scan_io_queue_destroy(tvd
->vdev_scan_io_queue
);
4231 tvd
->vdev_scan_io_queue
= NULL
;
4232 mutex_exit(&tvd
->vdev_scan_io_queue_lock
);
4237 dsl_scan_freed_dva(spa_t
*spa
, const blkptr_t
*bp
, int dva_i
)
4239 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
4240 dsl_scan_t
*scn
= dp
->dp_scan
;
4243 dsl_scan_io_queue_t
*queue
;
4244 scan_io_t
*srch_sio
, *sio
;
4246 uint64_t start
, size
;
4248 vdev
= vdev_lookup_top(spa
, DVA_GET_VDEV(&bp
->blk_dva
[dva_i
]));
4249 ASSERT(vdev
!= NULL
);
4250 q_lock
= &vdev
->vdev_scan_io_queue_lock
;
4251 queue
= vdev
->vdev_scan_io_queue
;
4253 mutex_enter(q_lock
);
4254 if (queue
== NULL
) {
4259 srch_sio
= sio_alloc(BP_GET_NDVAS(bp
));
4260 bp2sio(bp
, srch_sio
, dva_i
);
4261 start
= SIO_GET_OFFSET(srch_sio
);
4262 size
= SIO_GET_ASIZE(srch_sio
);
4265 * We can find the zio in two states:
4266 * 1) Cold, just sitting in the queue of zio's to be issued at
4267 * some point in the future. In this case, all we do is
4268 * remove the zio from the q_sios_by_addr tree, decrement
4269 * its data volume from the containing range_seg_t and
4270 * resort the q_exts_by_size tree to reflect that the
4271 * range_seg_t has lost some of its 'fill'. We don't shorten
4272 * the range_seg_t - this is usually rare enough not to be
4273 * worth the extra hassle of trying keep track of precise
4274 * extent boundaries.
4275 * 2) Hot, where the zio is currently in-flight in
4276 * dsl_scan_issue_ios. In this case, we can't simply
4277 * reach in and stop the in-flight zio's, so we instead
4278 * block the caller. Eventually, dsl_scan_issue_ios will
4279 * be done with issuing the zio's it gathered and will
4282 sio
= avl_find(&queue
->q_sios_by_addr
, srch_sio
, &idx
);
4286 int64_t asize
= SIO_GET_ASIZE(sio
);
4289 /* Got it while it was cold in the queue */
4290 ASSERT3U(start
, ==, SIO_GET_OFFSET(sio
));
4291 ASSERT3U(size
, ==, asize
);
4292 avl_remove(&queue
->q_sios_by_addr
, sio
);
4293 queue
->q_sio_memused
-= SIO_GET_MUSED(sio
);
4295 ASSERT(range_tree_contains(queue
->q_exts_by_addr
, start
, size
));
4296 range_tree_remove_fill(queue
->q_exts_by_addr
, start
, size
);
4299 * We only update scn_bytes_pending in the cold path,
4300 * otherwise it will already have been accounted for as
4301 * part of the zio's execution.
4303 atomic_add_64(&scn
->scn_bytes_pending
, -asize
);
4305 /* count the block as though we issued it */
4306 sio2bp(sio
, &tmpbp
);
4307 count_block(scn
, dp
->dp_blkstats
, &tmpbp
);
4315 * Callback invoked when a zio_free() zio is executing. This needs to be
4316 * intercepted to prevent the zio from deallocating a particular portion
4317 * of disk space and it then getting reallocated and written to, while we
4318 * still have it queued up for processing.
4321 dsl_scan_freed(spa_t
*spa
, const blkptr_t
*bp
)
4323 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
4324 dsl_scan_t
*scn
= dp
->dp_scan
;
4326 ASSERT(!BP_IS_EMBEDDED(bp
));
4327 ASSERT(scn
!= NULL
);
4328 if (!dsl_scan_is_running(scn
))
4331 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++)
4332 dsl_scan_freed_dva(spa
, bp
, i
);
4336 * Check if a vdev needs resilvering (non-empty DTL), if so, and resilver has
4337 * not started, start it. Otherwise, only restart if max txg in DTL range is
4338 * greater than the max txg in the current scan. If the DTL max is less than
4339 * the scan max, then the vdev has not missed any new data since the resilver
4340 * started, so a restart is not needed.
4343 dsl_scan_assess_vdev(dsl_pool_t
*dp
, vdev_t
*vd
)
4347 if (!vdev_resilver_needed(vd
, &min
, &max
))
4350 if (!dsl_scan_resilvering(dp
)) {
4351 spa_async_request(dp
->dp_spa
, SPA_ASYNC_RESILVER
);
4355 if (max
<= dp
->dp_scan
->scn_phys
.scn_max_txg
)
4358 /* restart is needed, check if it can be deferred */
4359 if (spa_feature_is_enabled(dp
->dp_spa
, SPA_FEATURE_RESILVER_DEFER
))
4360 vdev_defer_resilver(vd
);
4362 spa_async_request(dp
->dp_spa
, SPA_ASYNC_RESILVER
);
4366 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_vdev_limit
, ULONG
, ZMOD_RW
,
4367 "Max bytes in flight per leaf vdev for scrubs and resilvers");
4369 ZFS_MODULE_PARAM(zfs
, zfs_
, scrub_min_time_ms
, INT
, ZMOD_RW
,
4370 "Min millisecs to scrub per txg");
4372 ZFS_MODULE_PARAM(zfs
, zfs_
, obsolete_min_time_ms
, INT
, ZMOD_RW
,
4373 "Min millisecs to obsolete per txg");
4375 ZFS_MODULE_PARAM(zfs
, zfs_
, free_min_time_ms
, INT
, ZMOD_RW
,
4376 "Min millisecs to free per txg");
4378 ZFS_MODULE_PARAM(zfs
, zfs_
, resilver_min_time_ms
, INT
, ZMOD_RW
,
4379 "Min millisecs to resilver per txg");
4381 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_suspend_progress
, INT
, ZMOD_RW
,
4382 "Set to prevent scans from progressing");
4384 ZFS_MODULE_PARAM(zfs
, zfs_
, no_scrub_io
, INT
, ZMOD_RW
,
4385 "Set to disable scrub I/O");
4387 ZFS_MODULE_PARAM(zfs
, zfs_
, no_scrub_prefetch
, INT
, ZMOD_RW
,
4388 "Set to disable scrub prefetching");
4390 ZFS_MODULE_PARAM(zfs
, zfs_
, async_block_max_blocks
, ULONG
, ZMOD_RW
,
4391 "Max number of blocks freed in one txg");
4393 ZFS_MODULE_PARAM(zfs
, zfs_
, max_async_dedup_frees
, ULONG
, ZMOD_RW
,
4394 "Max number of dedup blocks freed in one txg");
4396 ZFS_MODULE_PARAM(zfs
, zfs_
, free_bpobj_enabled
, INT
, ZMOD_RW
,
4397 "Enable processing of the free_bpobj");
4399 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_mem_lim_fact
, INT
, ZMOD_RW
,
4400 "Fraction of RAM for scan hard limit");
4402 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_issue_strategy
, INT
, ZMOD_RW
,
4403 "IO issuing strategy during scrubbing. "
4404 "0 = default, 1 = LBA, 2 = size");
4406 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_legacy
, INT
, ZMOD_RW
,
4407 "Scrub using legacy non-sequential method");
4409 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_checkpoint_intval
, INT
, ZMOD_RW
,
4410 "Scan progress on-disk checkpointing interval");
4412 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_max_ext_gap
, ULONG
, ZMOD_RW
,
4413 "Max gap in bytes between sequential scrub / resilver I/Os");
4415 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_mem_lim_soft_fact
, INT
, ZMOD_RW
,
4416 "Fraction of hard limit used as soft limit");
4418 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_strict_mem_lim
, INT
, ZMOD_RW
,
4419 "Tunable to attempt to reduce lock contention");
4421 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_fill_weight
, INT
, ZMOD_RW
,
4422 "Tunable to adjust bias towards more filled segments during scans");
4424 ZFS_MODULE_PARAM(zfs
, zfs_
, resilver_disable_defer
, INT
, ZMOD_RW
,
4425 "Process all resilvers immediately");