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 (c) 2015, Nexenta Systems, Inc. All rights reserved.
27 * Copyright 2019 Joyent, Inc.
30 #include <sys/dsl_scan.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_prop.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_synctask.h>
36 #include <sys/dnode.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/dmu_objset.h>
42 #include <sys/zfs_context.h>
43 #include <sys/fs/zfs.h>
44 #include <sys/zfs_znode.h>
45 #include <sys/spa_impl.h>
46 #include <sys/vdev_impl.h>
47 #include <sys/zil_impl.h>
48 #include <sys/zio_checksum.h>
51 #include <sys/sa_impl.h>
52 #include <sys/zfeature.h>
54 #include <sys/range_tree.h>
56 #include <sys/zfs_vfsops.h>
60 * Grand theory statement on scan queue sorting
62 * Scanning is implemented by recursively traversing all indirection levels
63 * in an object and reading all blocks referenced from said objects. This
64 * results in us approximately traversing the object from lowest logical
65 * offset to the highest. For best performance, we would want the logical
66 * blocks to be physically contiguous. However, this is frequently not the
67 * case with pools given the allocation patterns of copy-on-write filesystems.
68 * So instead, we put the I/Os into a reordering queue and issue them in a
69 * way that will most benefit physical disks (LBA-order).
73 * Ideally, we would want to scan all metadata and queue up all block I/O
74 * prior to starting to issue it, because that allows us to do an optimal
75 * sorting job. This can however consume large amounts of memory. Therefore
76 * we continuously monitor the size of the queues and constrain them to 5%
77 * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this
78 * limit, we clear out a few of the largest extents at the head of the queues
79 * to make room for more scanning. Hopefully, these extents will be fairly
80 * large and contiguous, allowing us to approach sequential I/O throughput
81 * even without a fully sorted tree.
83 * Metadata scanning takes place in dsl_scan_visit(), which is called from
84 * dsl_scan_sync() every spa_sync(). If we have either fully scanned all
85 * metadata on the pool, or we need to make room in memory because our
86 * queues are too large, dsl_scan_visit() is postponed and
87 * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies
88 * that metadata scanning and queued I/O issuing are mutually exclusive. This
89 * allows us to provide maximum sequential I/O throughput for the majority of
90 * I/O's issued since sequential I/O performance is significantly negatively
91 * impacted if it is interleaved with random I/O.
93 * Implementation Notes
95 * One side effect of the queued scanning algorithm is that the scanning code
96 * needs to be notified whenever a block is freed. This is needed to allow
97 * the scanning code to remove these I/Os from the issuing queue. Additionally,
98 * we do not attempt to queue gang blocks to be issued sequentially since this
99 * is very hard to do and would have an extremely limited performance benefit.
100 * Instead, we simply issue gang I/Os as soon as we find them using the legacy
103 * Backwards compatibility
105 * This new algorithm is backwards compatible with the legacy on-disk data
106 * structures (and therefore does not require a new feature flag).
107 * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan
108 * will stop scanning metadata (in logical order) and wait for all outstanding
109 * sorted I/O to complete. Once this is done, we write out a checkpoint
110 * bookmark, indicating that we have scanned everything logically before it.
111 * If the pool is imported on a machine without the new sorting algorithm,
112 * the scan simply resumes from the last checkpoint using the legacy algorithm.
115 typedef int (scan_cb_t
)(dsl_pool_t
*, const blkptr_t
*,
116 const zbookmark_phys_t
*);
118 static scan_cb_t dsl_scan_scrub_cb
;
120 static int scan_ds_queue_compare(const void *a
, const void *b
);
121 static int scan_prefetch_queue_compare(const void *a
, const void *b
);
122 static void scan_ds_queue_clear(dsl_scan_t
*scn
);
123 static void scan_ds_prefetch_queue_clear(dsl_scan_t
*scn
);
124 static boolean_t
scan_ds_queue_contains(dsl_scan_t
*scn
, uint64_t dsobj
,
126 static void scan_ds_queue_insert(dsl_scan_t
*scn
, uint64_t dsobj
, uint64_t txg
);
127 static void scan_ds_queue_remove(dsl_scan_t
*scn
, uint64_t dsobj
);
128 static void scan_ds_queue_sync(dsl_scan_t
*scn
, dmu_tx_t
*tx
);
129 static uint64_t dsl_scan_count_leaves(vdev_t
*vd
);
131 extern int zfs_vdev_async_write_active_min_dirty_percent
;
134 * By default zfs will check to ensure it is not over the hard memory
135 * limit before each txg. If finer-grained control of this is needed
136 * this value can be set to 1 to enable checking before scanning each
139 int zfs_scan_strict_mem_lim
= B_FALSE
;
142 * Maximum number of parallelly executed bytes per leaf vdev. We attempt
143 * to strike a balance here between keeping the vdev queues full of I/Os
144 * at all times and not overflowing the queues to cause long latency,
145 * which would cause long txg sync times. No matter what, we will not
146 * overload the drives with I/O, since that is protected by
147 * zfs_vdev_scrub_max_active.
149 unsigned long zfs_scan_vdev_limit
= 4 << 20;
151 int zfs_scan_issue_strategy
= 0;
152 int zfs_scan_legacy
= B_FALSE
; /* don't queue & sort zios, go direct */
153 unsigned long zfs_scan_max_ext_gap
= 2 << 20; /* in bytes */
156 * fill_weight is non-tunable at runtime, so we copy it at module init from
157 * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
158 * break queue sorting.
160 int zfs_scan_fill_weight
= 3;
161 static uint64_t fill_weight
;
163 /* See dsl_scan_should_clear() for details on the memory limit tunables */
164 uint64_t zfs_scan_mem_lim_min
= 16 << 20; /* bytes */
165 uint64_t zfs_scan_mem_lim_soft_max
= 128 << 20; /* bytes */
166 int zfs_scan_mem_lim_fact
= 20; /* fraction of physmem */
167 int zfs_scan_mem_lim_soft_fact
= 20; /* fraction of mem lim above */
169 int zfs_scrub_min_time_ms
= 1000; /* min millisecs to scrub per txg */
170 int zfs_obsolete_min_time_ms
= 500; /* min millisecs to obsolete per txg */
171 int zfs_free_min_time_ms
= 1000; /* min millisecs to free per txg */
172 int zfs_resilver_min_time_ms
= 3000; /* min millisecs to resilver per txg */
173 int zfs_scan_checkpoint_intval
= 7200; /* in seconds */
174 int zfs_scan_suspend_progress
= 0; /* set to prevent scans from progressing */
175 int zfs_no_scrub_io
= B_FALSE
; /* set to disable scrub i/o */
176 int zfs_no_scrub_prefetch
= B_FALSE
; /* set to disable scrub prefetch */
177 enum ddt_class zfs_scrub_ddt_class_max
= DDT_CLASS_DUPLICATE
;
178 /* max number of blocks to free in a single TXG */
179 unsigned long zfs_async_block_max_blocks
= ULONG_MAX
;
180 /* max number of dedup blocks to free in a single TXG */
181 unsigned long zfs_max_async_dedup_frees
= 100000;
183 int zfs_resilver_disable_defer
= 0; /* set to disable resilver deferring */
186 * We wait a few txgs after importing a pool to begin scanning so that
187 * the import / mounting code isn't held up by scrub / resilver IO.
188 * Unfortunately, it is a bit difficult to determine exactly how long
189 * this will take since userspace will trigger fs mounts asynchronously
190 * and the kernel will create zvol minors asynchronously. As a result,
191 * the value provided here is a bit arbitrary, but represents a
192 * reasonable estimate of how many txgs it will take to finish fully
195 #define SCAN_IMPORT_WAIT_TXGS 5
197 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
198 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
199 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
202 * Enable/disable the processing of the free_bpobj object.
204 int zfs_free_bpobj_enabled
= 1;
206 /* the order has to match pool_scan_type */
207 static scan_cb_t
*scan_funcs
[POOL_SCAN_FUNCS
] = {
209 dsl_scan_scrub_cb
, /* POOL_SCAN_SCRUB */
210 dsl_scan_scrub_cb
, /* POOL_SCAN_RESILVER */
213 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */
221 * This controls what conditions are placed on dsl_scan_sync_state():
222 * SYNC_OPTIONAL) write out scn_phys iff scn_bytes_pending == 0
223 * SYNC_MANDATORY) write out scn_phys always. scn_bytes_pending must be 0.
224 * SYNC_CACHED) if scn_bytes_pending == 0, write out scn_phys. Otherwise
225 * write out the scn_phys_cached version.
226 * See dsl_scan_sync_state for details.
235 * This struct represents the minimum information needed to reconstruct a
236 * zio for sequential scanning. This is useful because many of these will
237 * accumulate in the sequential IO queues before being issued, so saving
238 * memory matters here.
240 typedef struct scan_io
{
241 /* fields from blkptr_t */
242 uint64_t sio_blk_prop
;
243 uint64_t sio_phys_birth
;
245 zio_cksum_t sio_cksum
;
246 uint32_t sio_nr_dvas
;
248 /* fields from zio_t */
250 zbookmark_phys_t sio_zb
;
252 /* members for queue sorting */
254 avl_node_t sio_addr_node
; /* link into issuing queue */
255 list_node_t sio_list_node
; /* link for issuing to disk */
259 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp,
260 * depending on how many were in the original bp. Only the
261 * first DVA is really used for sorting and issuing purposes.
262 * The other DVAs (if provided) simply exist so that the zio
263 * layer can find additional copies to repair from in the
264 * event of an error. This array must go at the end of the
265 * struct to allow this for the variable number of elements.
270 #define SIO_SET_OFFSET(sio, x) DVA_SET_OFFSET(&(sio)->sio_dva[0], x)
271 #define SIO_SET_ASIZE(sio, x) DVA_SET_ASIZE(&(sio)->sio_dva[0], x)
272 #define SIO_GET_OFFSET(sio) DVA_GET_OFFSET(&(sio)->sio_dva[0])
273 #define SIO_GET_ASIZE(sio) DVA_GET_ASIZE(&(sio)->sio_dva[0])
274 #define SIO_GET_END_OFFSET(sio) \
275 (SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio))
276 #define SIO_GET_MUSED(sio) \
277 (sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t)))
279 struct dsl_scan_io_queue
{
280 dsl_scan_t
*q_scn
; /* associated dsl_scan_t */
281 vdev_t
*q_vd
; /* top-level vdev that this queue represents */
283 /* trees used for sorting I/Os and extents of I/Os */
284 range_tree_t
*q_exts_by_addr
;
285 zfs_btree_t q_exts_by_size
;
286 avl_tree_t q_sios_by_addr
;
287 uint64_t q_sio_memused
;
289 /* members for zio rate limiting */
290 uint64_t q_maxinflight_bytes
;
291 uint64_t q_inflight_bytes
;
292 kcondvar_t q_zio_cv
; /* used under vd->vdev_scan_io_queue_lock */
294 /* per txg statistics */
295 uint64_t q_total_seg_size_this_txg
;
296 uint64_t q_segs_this_txg
;
297 uint64_t q_total_zio_size_this_txg
;
298 uint64_t q_zios_this_txg
;
301 /* private data for dsl_scan_prefetch_cb() */
302 typedef struct scan_prefetch_ctx
{
303 zfs_refcount_t spc_refcnt
; /* refcount for memory management */
304 dsl_scan_t
*spc_scn
; /* dsl_scan_t for the pool */
305 boolean_t spc_root
; /* is this prefetch for an objset? */
306 uint8_t spc_indblkshift
; /* dn_indblkshift of current dnode */
307 uint16_t spc_datablkszsec
; /* dn_idatablkszsec of current dnode */
308 } scan_prefetch_ctx_t
;
310 /* private data for dsl_scan_prefetch() */
311 typedef struct scan_prefetch_issue_ctx
{
312 avl_node_t spic_avl_node
; /* link into scn->scn_prefetch_queue */
313 scan_prefetch_ctx_t
*spic_spc
; /* spc for the callback */
314 blkptr_t spic_bp
; /* bp to prefetch */
315 zbookmark_phys_t spic_zb
; /* bookmark to prefetch */
316 } scan_prefetch_issue_ctx_t
;
318 static void scan_exec_io(dsl_pool_t
*dp
, const blkptr_t
*bp
, int zio_flags
,
319 const zbookmark_phys_t
*zb
, dsl_scan_io_queue_t
*queue
);
320 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t
*queue
,
323 static dsl_scan_io_queue_t
*scan_io_queue_create(vdev_t
*vd
);
324 static void scan_io_queues_destroy(dsl_scan_t
*scn
);
326 static kmem_cache_t
*sio_cache
[SPA_DVAS_PER_BP
];
328 /* sio->sio_nr_dvas must be set so we know which cache to free from */
330 sio_free(scan_io_t
*sio
)
332 ASSERT3U(sio
->sio_nr_dvas
, >, 0);
333 ASSERT3U(sio
->sio_nr_dvas
, <=, SPA_DVAS_PER_BP
);
335 kmem_cache_free(sio_cache
[sio
->sio_nr_dvas
- 1], sio
);
338 /* It is up to the caller to set sio->sio_nr_dvas for freeing */
340 sio_alloc(unsigned short nr_dvas
)
342 ASSERT3U(nr_dvas
, >, 0);
343 ASSERT3U(nr_dvas
, <=, SPA_DVAS_PER_BP
);
345 return (kmem_cache_alloc(sio_cache
[nr_dvas
- 1], KM_SLEEP
));
352 * This is used in ext_size_compare() to weight segments
353 * based on how sparse they are. This cannot be changed
354 * mid-scan and the tree comparison functions don't currently
355 * have a mechanism for passing additional context to the
356 * compare functions. Thus we store this value globally and
357 * we only allow it to be set at module initialization time
359 fill_weight
= zfs_scan_fill_weight
;
361 for (int i
= 0; i
< SPA_DVAS_PER_BP
; i
++) {
364 (void) snprintf(name
, sizeof (name
), "sio_cache_%d", i
);
365 sio_cache
[i
] = kmem_cache_create(name
,
366 (sizeof (scan_io_t
) + ((i
+ 1) * sizeof (dva_t
))),
367 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
374 for (int i
= 0; i
< SPA_DVAS_PER_BP
; i
++) {
375 kmem_cache_destroy(sio_cache
[i
]);
379 static inline boolean_t
380 dsl_scan_is_running(const dsl_scan_t
*scn
)
382 return (scn
->scn_phys
.scn_state
== DSS_SCANNING
);
386 dsl_scan_resilvering(dsl_pool_t
*dp
)
388 return (dsl_scan_is_running(dp
->dp_scan
) &&
389 dp
->dp_scan
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
);
393 sio2bp(const scan_io_t
*sio
, blkptr_t
*bp
)
395 bzero(bp
, sizeof (*bp
));
396 bp
->blk_prop
= sio
->sio_blk_prop
;
397 bp
->blk_phys_birth
= sio
->sio_phys_birth
;
398 bp
->blk_birth
= sio
->sio_birth
;
399 bp
->blk_fill
= 1; /* we always only work with data pointers */
400 bp
->blk_cksum
= sio
->sio_cksum
;
402 ASSERT3U(sio
->sio_nr_dvas
, >, 0);
403 ASSERT3U(sio
->sio_nr_dvas
, <=, SPA_DVAS_PER_BP
);
405 bcopy(sio
->sio_dva
, bp
->blk_dva
, sio
->sio_nr_dvas
* sizeof (dva_t
));
409 bp2sio(const blkptr_t
*bp
, scan_io_t
*sio
, int dva_i
)
411 sio
->sio_blk_prop
= bp
->blk_prop
;
412 sio
->sio_phys_birth
= bp
->blk_phys_birth
;
413 sio
->sio_birth
= bp
->blk_birth
;
414 sio
->sio_cksum
= bp
->blk_cksum
;
415 sio
->sio_nr_dvas
= BP_GET_NDVAS(bp
);
418 * Copy the DVAs to the sio. We need all copies of the block so
419 * that the self healing code can use the alternate copies if the
420 * first is corrupted. We want the DVA at index dva_i to be first
421 * in the sio since this is the primary one that we want to issue.
423 for (int i
= 0, j
= dva_i
; i
< sio
->sio_nr_dvas
; i
++, j
++) {
424 sio
->sio_dva
[i
] = bp
->blk_dva
[j
% sio
->sio_nr_dvas
];
429 dsl_scan_init(dsl_pool_t
*dp
, uint64_t txg
)
433 spa_t
*spa
= dp
->dp_spa
;
436 scn
= dp
->dp_scan
= kmem_zalloc(sizeof (dsl_scan_t
), KM_SLEEP
);
440 * It's possible that we're resuming a scan after a reboot so
441 * make sure that the scan_async_destroying flag is initialized
444 ASSERT(!scn
->scn_async_destroying
);
445 scn
->scn_async_destroying
= spa_feature_is_active(dp
->dp_spa
,
446 SPA_FEATURE_ASYNC_DESTROY
);
449 * Calculate the max number of in-flight bytes for pool-wide
450 * scanning operations (minimum 1MB). Limits for the issuing
451 * phase are done per top-level vdev and are handled separately.
453 scn
->scn_maxinflight_bytes
= MAX(zfs_scan_vdev_limit
*
454 dsl_scan_count_leaves(spa
->spa_root_vdev
), 1ULL << 20);
456 avl_create(&scn
->scn_queue
, scan_ds_queue_compare
, sizeof (scan_ds_t
),
457 offsetof(scan_ds_t
, sds_node
));
458 avl_create(&scn
->scn_prefetch_queue
, scan_prefetch_queue_compare
,
459 sizeof (scan_prefetch_issue_ctx_t
),
460 offsetof(scan_prefetch_issue_ctx_t
, spic_avl_node
));
462 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
463 "scrub_func", sizeof (uint64_t), 1, &f
);
466 * There was an old-style scrub in progress. Restart a
467 * new-style scrub from the beginning.
469 scn
->scn_restart_txg
= txg
;
470 zfs_dbgmsg("old-style scrub was in progress; "
471 "restarting new-style scrub in txg %llu",
472 (longlong_t
)scn
->scn_restart_txg
);
475 * Load the queue obj from the old location so that it
476 * can be freed by dsl_scan_done().
478 (void) zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
479 "scrub_queue", sizeof (uint64_t), 1,
480 &scn
->scn_phys
.scn_queue_obj
);
482 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
483 DMU_POOL_SCAN
, sizeof (uint64_t), SCAN_PHYS_NUMINTS
,
486 * Detect if the pool contains the signature of #2094. If it
487 * does properly update the scn->scn_phys structure and notify
488 * the administrator by setting an errata for the pool.
490 if (err
== EOVERFLOW
) {
491 uint64_t zaptmp
[SCAN_PHYS_NUMINTS
+ 1];
492 VERIFY3S(SCAN_PHYS_NUMINTS
, ==, 24);
493 VERIFY3S(offsetof(dsl_scan_phys_t
, scn_flags
), ==,
494 (23 * sizeof (uint64_t)));
496 err
= zap_lookup(dp
->dp_meta_objset
,
497 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_SCAN
,
498 sizeof (uint64_t), SCAN_PHYS_NUMINTS
+ 1, &zaptmp
);
500 uint64_t overflow
= zaptmp
[SCAN_PHYS_NUMINTS
];
502 if (overflow
& ~DSL_SCAN_FLAGS_MASK
||
503 scn
->scn_async_destroying
) {
505 ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY
;
509 bcopy(zaptmp
, &scn
->scn_phys
,
510 SCAN_PHYS_NUMINTS
* sizeof (uint64_t));
511 scn
->scn_phys
.scn_flags
= overflow
;
513 /* Required scrub already in progress. */
514 if (scn
->scn_phys
.scn_state
== DSS_FINISHED
||
515 scn
->scn_phys
.scn_state
== DSS_CANCELED
)
517 ZPOOL_ERRATA_ZOL_2094_SCRUB
;
527 * We might be restarting after a reboot, so jump the issued
528 * counter to how far we've scanned. We know we're consistent
531 scn
->scn_issued_before_pass
= scn
->scn_phys
.scn_examined
;
533 if (dsl_scan_is_running(scn
) &&
534 spa_prev_software_version(dp
->dp_spa
) < SPA_VERSION_SCAN
) {
536 * A new-type scrub was in progress on an old
537 * pool, and the pool was accessed by old
538 * software. Restart from the beginning, since
539 * the old software may have changed the pool in
542 scn
->scn_restart_txg
= txg
;
543 zfs_dbgmsg("new-style scrub was modified "
544 "by old software; restarting in txg %llu",
545 (longlong_t
)scn
->scn_restart_txg
);
546 } else if (dsl_scan_resilvering(dp
)) {
548 * If a resilver is in progress and there are already
549 * errors, restart it instead of finishing this scan and
550 * then restarting it. If there haven't been any errors
551 * then remember that the incore DTL is valid.
553 if (scn
->scn_phys
.scn_errors
> 0) {
554 scn
->scn_restart_txg
= txg
;
555 zfs_dbgmsg("resilver can't excise DTL_MISSING "
556 "when finished; restarting in txg %llu",
557 (u_longlong_t
)scn
->scn_restart_txg
);
559 /* it's safe to excise DTL when finished */
560 spa
->spa_scrub_started
= B_TRUE
;
565 bcopy(&scn
->scn_phys
, &scn
->scn_phys_cached
, sizeof (scn
->scn_phys
));
567 /* reload the queue into the in-core state */
568 if (scn
->scn_phys
.scn_queue_obj
!= 0) {
572 for (zap_cursor_init(&zc
, dp
->dp_meta_objset
,
573 scn
->scn_phys
.scn_queue_obj
);
574 zap_cursor_retrieve(&zc
, &za
) == 0;
575 (void) zap_cursor_advance(&zc
)) {
576 scan_ds_queue_insert(scn
,
577 zfs_strtonum(za
.za_name
, NULL
),
578 za
.za_first_integer
);
580 zap_cursor_fini(&zc
);
583 spa_scan_stat_init(spa
);
588 dsl_scan_fini(dsl_pool_t
*dp
)
590 if (dp
->dp_scan
!= NULL
) {
591 dsl_scan_t
*scn
= dp
->dp_scan
;
593 if (scn
->scn_taskq
!= NULL
)
594 taskq_destroy(scn
->scn_taskq
);
596 scan_ds_queue_clear(scn
);
597 avl_destroy(&scn
->scn_queue
);
598 scan_ds_prefetch_queue_clear(scn
);
599 avl_destroy(&scn
->scn_prefetch_queue
);
601 kmem_free(dp
->dp_scan
, sizeof (dsl_scan_t
));
607 dsl_scan_restarting(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
609 return (scn
->scn_restart_txg
!= 0 &&
610 scn
->scn_restart_txg
<= tx
->tx_txg
);
614 dsl_scan_resilver_scheduled(dsl_pool_t
*dp
)
616 return ((dp
->dp_scan
&& dp
->dp_scan
->scn_restart_txg
!= 0) ||
617 (spa_async_tasks(dp
->dp_spa
) & SPA_ASYNC_RESILVER
));
621 dsl_scan_scrubbing(const dsl_pool_t
*dp
)
623 dsl_scan_phys_t
*scn_phys
= &dp
->dp_scan
->scn_phys
;
625 return (scn_phys
->scn_state
== DSS_SCANNING
&&
626 scn_phys
->scn_func
== POOL_SCAN_SCRUB
);
630 dsl_scan_is_paused_scrub(const dsl_scan_t
*scn
)
632 return (dsl_scan_scrubbing(scn
->scn_dp
) &&
633 scn
->scn_phys
.scn_flags
& DSF_SCRUB_PAUSED
);
637 * Writes out a persistent dsl_scan_phys_t record to the pool directory.
638 * Because we can be running in the block sorting algorithm, we do not always
639 * want to write out the record, only when it is "safe" to do so. This safety
640 * condition is achieved by making sure that the sorting queues are empty
641 * (scn_bytes_pending == 0). When this condition is not true, the sync'd state
642 * is inconsistent with how much actual scanning progress has been made. The
643 * kind of sync to be performed is specified by the sync_type argument. If the
644 * sync is optional, we only sync if the queues are empty. If the sync is
645 * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
646 * third possible state is a "cached" sync. This is done in response to:
647 * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
648 * destroyed, so we wouldn't be able to restart scanning from it.
649 * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
650 * superseded by a newer snapshot.
651 * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
652 * swapped with its clone.
653 * In all cases, a cached sync simply rewrites the last record we've written,
654 * just slightly modified. For the modifications that are performed to the
655 * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
656 * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
659 dsl_scan_sync_state(dsl_scan_t
*scn
, dmu_tx_t
*tx
, state_sync_type_t sync_type
)
662 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
664 ASSERT(sync_type
!= SYNC_MANDATORY
|| scn
->scn_bytes_pending
== 0);
665 if (scn
->scn_bytes_pending
== 0) {
666 for (i
= 0; i
< spa
->spa_root_vdev
->vdev_children
; i
++) {
667 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[i
];
668 dsl_scan_io_queue_t
*q
= vd
->vdev_scan_io_queue
;
673 mutex_enter(&vd
->vdev_scan_io_queue_lock
);
674 ASSERT3P(avl_first(&q
->q_sios_by_addr
), ==, NULL
);
675 ASSERT3P(zfs_btree_first(&q
->q_exts_by_size
, NULL
), ==,
677 ASSERT3P(range_tree_first(q
->q_exts_by_addr
), ==, NULL
);
678 mutex_exit(&vd
->vdev_scan_io_queue_lock
);
681 if (scn
->scn_phys
.scn_queue_obj
!= 0)
682 scan_ds_queue_sync(scn
, tx
);
683 VERIFY0(zap_update(scn
->scn_dp
->dp_meta_objset
,
684 DMU_POOL_DIRECTORY_OBJECT
,
685 DMU_POOL_SCAN
, sizeof (uint64_t), SCAN_PHYS_NUMINTS
,
686 &scn
->scn_phys
, tx
));
687 bcopy(&scn
->scn_phys
, &scn
->scn_phys_cached
,
688 sizeof (scn
->scn_phys
));
690 if (scn
->scn_checkpointing
)
691 zfs_dbgmsg("finish scan checkpoint");
693 scn
->scn_checkpointing
= B_FALSE
;
694 scn
->scn_last_checkpoint
= ddi_get_lbolt();
695 } else if (sync_type
== SYNC_CACHED
) {
696 VERIFY0(zap_update(scn
->scn_dp
->dp_meta_objset
,
697 DMU_POOL_DIRECTORY_OBJECT
,
698 DMU_POOL_SCAN
, sizeof (uint64_t), SCAN_PHYS_NUMINTS
,
699 &scn
->scn_phys_cached
, tx
));
705 dsl_scan_setup_check(void *arg
, dmu_tx_t
*tx
)
707 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
708 vdev_t
*rvd
= scn
->scn_dp
->dp_spa
->spa_root_vdev
;
710 if (dsl_scan_is_running(scn
) || vdev_rebuild_active(rvd
))
711 return (SET_ERROR(EBUSY
));
717 dsl_scan_setup_sync(void *arg
, dmu_tx_t
*tx
)
719 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
720 pool_scan_func_t
*funcp
= arg
;
721 dmu_object_type_t ot
= 0;
722 dsl_pool_t
*dp
= scn
->scn_dp
;
723 spa_t
*spa
= dp
->dp_spa
;
725 ASSERT(!dsl_scan_is_running(scn
));
726 ASSERT(*funcp
> POOL_SCAN_NONE
&& *funcp
< POOL_SCAN_FUNCS
);
727 bzero(&scn
->scn_phys
, sizeof (scn
->scn_phys
));
728 scn
->scn_phys
.scn_func
= *funcp
;
729 scn
->scn_phys
.scn_state
= DSS_SCANNING
;
730 scn
->scn_phys
.scn_min_txg
= 0;
731 scn
->scn_phys
.scn_max_txg
= tx
->tx_txg
;
732 scn
->scn_phys
.scn_ddt_class_max
= DDT_CLASSES
- 1; /* the entire DDT */
733 scn
->scn_phys
.scn_start_time
= gethrestime_sec();
734 scn
->scn_phys
.scn_errors
= 0;
735 scn
->scn_phys
.scn_to_examine
= spa
->spa_root_vdev
->vdev_stat
.vs_alloc
;
736 scn
->scn_issued_before_pass
= 0;
737 scn
->scn_restart_txg
= 0;
738 scn
->scn_done_txg
= 0;
739 scn
->scn_last_checkpoint
= 0;
740 scn
->scn_checkpointing
= B_FALSE
;
741 spa_scan_stat_init(spa
);
743 if (DSL_SCAN_IS_SCRUB_RESILVER(scn
)) {
744 scn
->scn_phys
.scn_ddt_class_max
= zfs_scrub_ddt_class_max
;
746 /* rewrite all disk labels */
747 vdev_config_dirty(spa
->spa_root_vdev
);
749 if (vdev_resilver_needed(spa
->spa_root_vdev
,
750 &scn
->scn_phys
.scn_min_txg
, &scn
->scn_phys
.scn_max_txg
)) {
751 nvlist_t
*aux
= fnvlist_alloc();
752 fnvlist_add_string(aux
, ZFS_EV_RESILVER_TYPE
,
754 spa_event_notify(spa
, NULL
, aux
,
755 ESC_ZFS_RESILVER_START
);
758 spa_event_notify(spa
, NULL
, NULL
, ESC_ZFS_SCRUB_START
);
761 spa
->spa_scrub_started
= B_TRUE
;
763 * If this is an incremental scrub, limit the DDT scrub phase
764 * to just the auto-ditto class (for correctness); the rest
765 * of the scrub should go faster using top-down pruning.
767 if (scn
->scn_phys
.scn_min_txg
> TXG_INITIAL
)
768 scn
->scn_phys
.scn_ddt_class_max
= DDT_CLASS_DITTO
;
771 * When starting a resilver clear any existing rebuild state.
772 * This is required to prevent stale rebuild status from
773 * being reported when a rebuild is run, then a resilver and
774 * finally a scrub. In which case only the scrub status
775 * should be reported by 'zpool status'.
777 if (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) {
778 vdev_t
*rvd
= spa
->spa_root_vdev
;
779 for (uint64_t i
= 0; i
< rvd
->vdev_children
; i
++) {
780 vdev_t
*vd
= rvd
->vdev_child
[i
];
781 vdev_rebuild_clear_sync(
782 (void *)(uintptr_t)vd
->vdev_id
, tx
);
787 /* back to the generic stuff */
789 if (dp
->dp_blkstats
== NULL
) {
791 vmem_alloc(sizeof (zfs_all_blkstats_t
), KM_SLEEP
);
792 mutex_init(&dp
->dp_blkstats
->zab_lock
, NULL
,
793 MUTEX_DEFAULT
, NULL
);
795 bzero(&dp
->dp_blkstats
->zab_type
, sizeof (dp
->dp_blkstats
->zab_type
));
797 if (spa_version(spa
) < SPA_VERSION_DSL_SCRUB
)
798 ot
= DMU_OT_ZAP_OTHER
;
800 scn
->scn_phys
.scn_queue_obj
= zap_create(dp
->dp_meta_objset
,
801 ot
? ot
: DMU_OT_SCAN_QUEUE
, DMU_OT_NONE
, 0, tx
);
803 bcopy(&scn
->scn_phys
, &scn
->scn_phys_cached
, sizeof (scn
->scn_phys
));
805 dsl_scan_sync_state(scn
, tx
, SYNC_MANDATORY
);
807 spa_history_log_internal(spa
, "scan setup", tx
,
808 "func=%u mintxg=%llu maxtxg=%llu",
809 *funcp
, (u_longlong_t
)scn
->scn_phys
.scn_min_txg
,
810 (u_longlong_t
)scn
->scn_phys
.scn_max_txg
);
814 * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
815 * Can also be called to resume a paused scrub.
818 dsl_scan(dsl_pool_t
*dp
, pool_scan_func_t func
)
820 spa_t
*spa
= dp
->dp_spa
;
821 dsl_scan_t
*scn
= dp
->dp_scan
;
824 * Purge all vdev caches and probe all devices. We do this here
825 * rather than in sync context because this requires a writer lock
826 * on the spa_config lock, which we can't do from sync context. The
827 * spa_scrub_reopen flag indicates that vdev_open() should not
828 * attempt to start another scrub.
830 spa_vdev_state_enter(spa
, SCL_NONE
);
831 spa
->spa_scrub_reopen
= B_TRUE
;
832 vdev_reopen(spa
->spa_root_vdev
);
833 spa
->spa_scrub_reopen
= B_FALSE
;
834 (void) spa_vdev_state_exit(spa
, NULL
, 0);
836 if (func
== POOL_SCAN_RESILVER
) {
837 dsl_scan_restart_resilver(spa
->spa_dsl_pool
, 0);
841 if (func
== POOL_SCAN_SCRUB
&& dsl_scan_is_paused_scrub(scn
)) {
842 /* got scrub start cmd, resume paused scrub */
843 int err
= dsl_scrub_set_pause_resume(scn
->scn_dp
,
846 spa_event_notify(spa
, NULL
, NULL
, ESC_ZFS_SCRUB_RESUME
);
847 return (SET_ERROR(ECANCELED
));
850 return (SET_ERROR(err
));
853 return (dsl_sync_task(spa_name(spa
), dsl_scan_setup_check
,
854 dsl_scan_setup_sync
, &func
, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED
));
859 dsl_scan_done(dsl_scan_t
*scn
, boolean_t complete
, dmu_tx_t
*tx
)
861 static const char *old_names
[] = {
863 "scrub_ddt_bookmark",
864 "scrub_ddt_class_max",
873 dsl_pool_t
*dp
= scn
->scn_dp
;
874 spa_t
*spa
= dp
->dp_spa
;
877 /* Remove any remnants of an old-style scrub. */
878 for (i
= 0; old_names
[i
]; i
++) {
879 (void) zap_remove(dp
->dp_meta_objset
,
880 DMU_POOL_DIRECTORY_OBJECT
, old_names
[i
], tx
);
883 if (scn
->scn_phys
.scn_queue_obj
!= 0) {
884 VERIFY0(dmu_object_free(dp
->dp_meta_objset
,
885 scn
->scn_phys
.scn_queue_obj
, tx
));
886 scn
->scn_phys
.scn_queue_obj
= 0;
888 scan_ds_queue_clear(scn
);
889 scan_ds_prefetch_queue_clear(scn
);
891 scn
->scn_phys
.scn_flags
&= ~DSF_SCRUB_PAUSED
;
894 * If we were "restarted" from a stopped state, don't bother
895 * with anything else.
897 if (!dsl_scan_is_running(scn
)) {
898 ASSERT(!scn
->scn_is_sorted
);
902 if (scn
->scn_is_sorted
) {
903 scan_io_queues_destroy(scn
);
904 scn
->scn_is_sorted
= B_FALSE
;
906 if (scn
->scn_taskq
!= NULL
) {
907 taskq_destroy(scn
->scn_taskq
);
908 scn
->scn_taskq
= NULL
;
912 scn
->scn_phys
.scn_state
= complete
? DSS_FINISHED
: DSS_CANCELED
;
914 spa_notify_waiters(spa
);
916 if (dsl_scan_restarting(scn
, tx
))
917 spa_history_log_internal(spa
, "scan aborted, restarting", tx
,
918 "errors=%llu", (u_longlong_t
)spa_get_errlog_size(spa
));
920 spa_history_log_internal(spa
, "scan cancelled", tx
,
921 "errors=%llu", (u_longlong_t
)spa_get_errlog_size(spa
));
923 spa_history_log_internal(spa
, "scan done", tx
,
924 "errors=%llu", (u_longlong_t
)spa_get_errlog_size(spa
));
926 if (DSL_SCAN_IS_SCRUB_RESILVER(scn
)) {
927 spa
->spa_scrub_active
= B_FALSE
;
930 * If the scrub/resilver completed, update all DTLs to
931 * reflect this. Whether it succeeded or not, vacate
932 * all temporary scrub DTLs.
934 * As the scrub does not currently support traversing
935 * data that have been freed but are part of a checkpoint,
936 * we don't mark the scrub as done in the DTLs as faults
937 * may still exist in those vdevs.
940 !spa_feature_is_active(spa
, SPA_FEATURE_POOL_CHECKPOINT
)) {
941 vdev_dtl_reassess(spa
->spa_root_vdev
, tx
->tx_txg
,
942 scn
->scn_phys
.scn_max_txg
, B_TRUE
, B_FALSE
);
944 if (scn
->scn_phys
.scn_min_txg
) {
945 nvlist_t
*aux
= fnvlist_alloc();
946 fnvlist_add_string(aux
, ZFS_EV_RESILVER_TYPE
,
948 spa_event_notify(spa
, NULL
, aux
,
949 ESC_ZFS_RESILVER_FINISH
);
952 spa_event_notify(spa
, NULL
, NULL
,
953 ESC_ZFS_SCRUB_FINISH
);
956 vdev_dtl_reassess(spa
->spa_root_vdev
, tx
->tx_txg
,
959 spa_errlog_rotate(spa
);
962 * Don't clear flag until after vdev_dtl_reassess to ensure that
963 * DTL_MISSING will get updated when possible.
965 spa
->spa_scrub_started
= B_FALSE
;
968 * We may have finished replacing a device.
969 * Let the async thread assess this and handle the detach.
971 spa_async_request(spa
, SPA_ASYNC_RESILVER_DONE
);
974 * Clear any resilver_deferred flags in the config.
975 * If there are drives that need resilvering, kick
976 * off an asynchronous request to start resilver.
977 * vdev_clear_resilver_deferred() may update the config
978 * before the resilver can restart. In the event of
979 * a crash during this period, the spa loading code
980 * will find the drives that need to be resilvered
981 * and start the resilver then.
983 if (spa_feature_is_enabled(spa
, SPA_FEATURE_RESILVER_DEFER
) &&
984 vdev_clear_resilver_deferred(spa
->spa_root_vdev
, tx
)) {
985 spa_history_log_internal(spa
,
986 "starting deferred resilver", tx
, "errors=%llu",
987 (u_longlong_t
)spa_get_errlog_size(spa
));
988 spa_async_request(spa
, SPA_ASYNC_RESILVER
);
992 scn
->scn_phys
.scn_end_time
= gethrestime_sec();
994 if (spa
->spa_errata
== ZPOOL_ERRATA_ZOL_2094_SCRUB
)
997 ASSERT(!dsl_scan_is_running(scn
));
1002 dsl_scan_cancel_check(void *arg
, dmu_tx_t
*tx
)
1004 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
1006 if (!dsl_scan_is_running(scn
))
1007 return (SET_ERROR(ENOENT
));
1013 dsl_scan_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1015 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
1017 dsl_scan_done(scn
, B_FALSE
, tx
);
1018 dsl_scan_sync_state(scn
, tx
, SYNC_MANDATORY
);
1019 spa_event_notify(scn
->scn_dp
->dp_spa
, NULL
, NULL
, ESC_ZFS_SCRUB_ABORT
);
1023 dsl_scan_cancel(dsl_pool_t
*dp
)
1025 return (dsl_sync_task(spa_name(dp
->dp_spa
), dsl_scan_cancel_check
,
1026 dsl_scan_cancel_sync
, NULL
, 3, ZFS_SPACE_CHECK_RESERVED
));
1030 dsl_scrub_pause_resume_check(void *arg
, dmu_tx_t
*tx
)
1032 pool_scrub_cmd_t
*cmd
= arg
;
1033 dsl_pool_t
*dp
= dmu_tx_pool(tx
);
1034 dsl_scan_t
*scn
= dp
->dp_scan
;
1036 if (*cmd
== POOL_SCRUB_PAUSE
) {
1037 /* can't pause a scrub when there is no in-progress scrub */
1038 if (!dsl_scan_scrubbing(dp
))
1039 return (SET_ERROR(ENOENT
));
1041 /* can't pause a paused scrub */
1042 if (dsl_scan_is_paused_scrub(scn
))
1043 return (SET_ERROR(EBUSY
));
1044 } else if (*cmd
!= POOL_SCRUB_NORMAL
) {
1045 return (SET_ERROR(ENOTSUP
));
1052 dsl_scrub_pause_resume_sync(void *arg
, dmu_tx_t
*tx
)
1054 pool_scrub_cmd_t
*cmd
= arg
;
1055 dsl_pool_t
*dp
= dmu_tx_pool(tx
);
1056 spa_t
*spa
= dp
->dp_spa
;
1057 dsl_scan_t
*scn
= dp
->dp_scan
;
1059 if (*cmd
== POOL_SCRUB_PAUSE
) {
1060 /* can't pause a scrub when there is no in-progress scrub */
1061 spa
->spa_scan_pass_scrub_pause
= gethrestime_sec();
1062 scn
->scn_phys
.scn_flags
|= DSF_SCRUB_PAUSED
;
1063 scn
->scn_phys_cached
.scn_flags
|= DSF_SCRUB_PAUSED
;
1064 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
1065 spa_event_notify(spa
, NULL
, NULL
, ESC_ZFS_SCRUB_PAUSED
);
1066 spa_notify_waiters(spa
);
1068 ASSERT3U(*cmd
, ==, POOL_SCRUB_NORMAL
);
1069 if (dsl_scan_is_paused_scrub(scn
)) {
1071 * We need to keep track of how much time we spend
1072 * paused per pass so that we can adjust the scrub rate
1073 * shown in the output of 'zpool status'
1075 spa
->spa_scan_pass_scrub_spent_paused
+=
1076 gethrestime_sec() - spa
->spa_scan_pass_scrub_pause
;
1077 spa
->spa_scan_pass_scrub_pause
= 0;
1078 scn
->scn_phys
.scn_flags
&= ~DSF_SCRUB_PAUSED
;
1079 scn
->scn_phys_cached
.scn_flags
&= ~DSF_SCRUB_PAUSED
;
1080 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
1086 * Set scrub pause/resume state if it makes sense to do so
1089 dsl_scrub_set_pause_resume(const dsl_pool_t
*dp
, pool_scrub_cmd_t cmd
)
1091 return (dsl_sync_task(spa_name(dp
->dp_spa
),
1092 dsl_scrub_pause_resume_check
, dsl_scrub_pause_resume_sync
, &cmd
, 3,
1093 ZFS_SPACE_CHECK_RESERVED
));
1097 /* start a new scan, or restart an existing one. */
1099 dsl_scan_restart_resilver(dsl_pool_t
*dp
, uint64_t txg
)
1103 tx
= dmu_tx_create_dd(dp
->dp_mos_dir
);
1104 VERIFY(0 == dmu_tx_assign(tx
, TXG_WAIT
));
1106 txg
= dmu_tx_get_txg(tx
);
1107 dp
->dp_scan
->scn_restart_txg
= txg
;
1110 dp
->dp_scan
->scn_restart_txg
= txg
;
1112 zfs_dbgmsg("restarting resilver txg=%llu", (longlong_t
)txg
);
1116 dsl_free(dsl_pool_t
*dp
, uint64_t txg
, const blkptr_t
*bp
)
1118 zio_free(dp
->dp_spa
, txg
, bp
);
1122 dsl_free_sync(zio_t
*pio
, dsl_pool_t
*dp
, uint64_t txg
, const blkptr_t
*bpp
)
1124 ASSERT(dsl_pool_sync_context(dp
));
1125 zio_nowait(zio_free_sync(pio
, dp
->dp_spa
, txg
, bpp
, pio
->io_flags
));
1129 scan_ds_queue_compare(const void *a
, const void *b
)
1131 const scan_ds_t
*sds_a
= a
, *sds_b
= b
;
1133 if (sds_a
->sds_dsobj
< sds_b
->sds_dsobj
)
1135 if (sds_a
->sds_dsobj
== sds_b
->sds_dsobj
)
1141 scan_ds_queue_clear(dsl_scan_t
*scn
)
1143 void *cookie
= NULL
;
1145 while ((sds
= avl_destroy_nodes(&scn
->scn_queue
, &cookie
)) != NULL
) {
1146 kmem_free(sds
, sizeof (*sds
));
1151 scan_ds_queue_contains(dsl_scan_t
*scn
, uint64_t dsobj
, uint64_t *txg
)
1153 scan_ds_t srch
, *sds
;
1155 srch
.sds_dsobj
= dsobj
;
1156 sds
= avl_find(&scn
->scn_queue
, &srch
, NULL
);
1157 if (sds
!= NULL
&& txg
!= NULL
)
1158 *txg
= sds
->sds_txg
;
1159 return (sds
!= NULL
);
1163 scan_ds_queue_insert(dsl_scan_t
*scn
, uint64_t dsobj
, uint64_t txg
)
1168 sds
= kmem_zalloc(sizeof (*sds
), KM_SLEEP
);
1169 sds
->sds_dsobj
= dsobj
;
1172 VERIFY3P(avl_find(&scn
->scn_queue
, sds
, &where
), ==, NULL
);
1173 avl_insert(&scn
->scn_queue
, sds
, where
);
1177 scan_ds_queue_remove(dsl_scan_t
*scn
, uint64_t dsobj
)
1179 scan_ds_t srch
, *sds
;
1181 srch
.sds_dsobj
= dsobj
;
1183 sds
= avl_find(&scn
->scn_queue
, &srch
, NULL
);
1184 VERIFY(sds
!= NULL
);
1185 avl_remove(&scn
->scn_queue
, sds
);
1186 kmem_free(sds
, sizeof (*sds
));
1190 scan_ds_queue_sync(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
1192 dsl_pool_t
*dp
= scn
->scn_dp
;
1193 spa_t
*spa
= dp
->dp_spa
;
1194 dmu_object_type_t ot
= (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
) ?
1195 DMU_OT_SCAN_QUEUE
: DMU_OT_ZAP_OTHER
;
1197 ASSERT0(scn
->scn_bytes_pending
);
1198 ASSERT(scn
->scn_phys
.scn_queue_obj
!= 0);
1200 VERIFY0(dmu_object_free(dp
->dp_meta_objset
,
1201 scn
->scn_phys
.scn_queue_obj
, tx
));
1202 scn
->scn_phys
.scn_queue_obj
= zap_create(dp
->dp_meta_objset
, ot
,
1203 DMU_OT_NONE
, 0, tx
);
1204 for (scan_ds_t
*sds
= avl_first(&scn
->scn_queue
);
1205 sds
!= NULL
; sds
= AVL_NEXT(&scn
->scn_queue
, sds
)) {
1206 VERIFY0(zap_add_int_key(dp
->dp_meta_objset
,
1207 scn
->scn_phys
.scn_queue_obj
, sds
->sds_dsobj
,
1213 * Computes the memory limit state that we're currently in. A sorted scan
1214 * needs quite a bit of memory to hold the sorting queue, so we need to
1215 * reasonably constrain the size so it doesn't impact overall system
1216 * performance. We compute two limits:
1217 * 1) Hard memory limit: if the amount of memory used by the sorting
1218 * queues on a pool gets above this value, we stop the metadata
1219 * scanning portion and start issuing the queued up and sorted
1220 * I/Os to reduce memory usage.
1221 * This limit is calculated as a fraction of physmem (by default 5%).
1222 * We constrain the lower bound of the hard limit to an absolute
1223 * minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1224 * the upper bound to 5% of the total pool size - no chance we'll
1225 * ever need that much memory, but just to keep the value in check.
1226 * 2) Soft memory limit: once we hit the hard memory limit, we start
1227 * issuing I/O to reduce queue memory usage, but we don't want to
1228 * completely empty out the queues, since we might be able to find I/Os
1229 * that will fill in the gaps of our non-sequential IOs at some point
1230 * in the future. So we stop the issuing of I/Os once the amount of
1231 * memory used drops below the soft limit (at which point we stop issuing
1232 * I/O and start scanning metadata again).
1234 * This limit is calculated by subtracting a fraction of the hard
1235 * limit from the hard limit. By default this fraction is 5%, so
1236 * the soft limit is 95% of the hard limit. We cap the size of the
1237 * difference between the hard and soft limits at an absolute
1238 * maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1239 * sufficient to not cause too frequent switching between the
1240 * metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1241 * worth of queues is about 1.2 GiB of on-pool data, so scanning
1242 * that should take at least a decent fraction of a second).
1245 dsl_scan_should_clear(dsl_scan_t
*scn
)
1247 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1248 vdev_t
*rvd
= scn
->scn_dp
->dp_spa
->spa_root_vdev
;
1249 uint64_t alloc
, mlim_hard
, mlim_soft
, mused
;
1251 alloc
= metaslab_class_get_alloc(spa_normal_class(spa
));
1252 alloc
+= metaslab_class_get_alloc(spa_special_class(spa
));
1253 alloc
+= metaslab_class_get_alloc(spa_dedup_class(spa
));
1255 mlim_hard
= MAX((physmem
/ zfs_scan_mem_lim_fact
) * PAGESIZE
,
1256 zfs_scan_mem_lim_min
);
1257 mlim_hard
= MIN(mlim_hard
, alloc
/ 20);
1258 mlim_soft
= mlim_hard
- MIN(mlim_hard
/ zfs_scan_mem_lim_soft_fact
,
1259 zfs_scan_mem_lim_soft_max
);
1261 for (uint64_t i
= 0; i
< rvd
->vdev_children
; i
++) {
1262 vdev_t
*tvd
= rvd
->vdev_child
[i
];
1263 dsl_scan_io_queue_t
*queue
;
1265 mutex_enter(&tvd
->vdev_scan_io_queue_lock
);
1266 queue
= tvd
->vdev_scan_io_queue
;
1267 if (queue
!= NULL
) {
1268 /* # extents in exts_by_size = # in exts_by_addr */
1269 mused
+= zfs_btree_numnodes(&queue
->q_exts_by_size
) *
1270 sizeof (range_seg_gap_t
) + queue
->q_sio_memused
;
1272 mutex_exit(&tvd
->vdev_scan_io_queue_lock
);
1275 dprintf("current scan memory usage: %llu bytes\n", (longlong_t
)mused
);
1278 ASSERT0(scn
->scn_bytes_pending
);
1281 * If we are above our hard limit, we need to clear out memory.
1282 * If we are below our soft limit, we need to accumulate sequential IOs.
1283 * Otherwise, we should keep doing whatever we are currently doing.
1285 if (mused
>= mlim_hard
)
1287 else if (mused
< mlim_soft
)
1290 return (scn
->scn_clearing
);
1294 dsl_scan_check_suspend(dsl_scan_t
*scn
, const zbookmark_phys_t
*zb
)
1296 /* we never skip user/group accounting objects */
1297 if (zb
&& (int64_t)zb
->zb_object
< 0)
1300 if (scn
->scn_suspending
)
1301 return (B_TRUE
); /* we're already suspending */
1303 if (!ZB_IS_ZERO(&scn
->scn_phys
.scn_bookmark
))
1304 return (B_FALSE
); /* we're resuming */
1306 /* We only know how to resume from level-0 and objset blocks. */
1307 if (zb
&& (zb
->zb_level
!= 0 && zb
->zb_level
!= ZB_ROOT_LEVEL
))
1312 * - we have scanned for at least the minimum time (default 1 sec
1313 * for scrub, 3 sec for resilver), and either we have sufficient
1314 * dirty data that we are starting to write more quickly
1315 * (default 30%), someone is explicitly waiting for this txg
1316 * to complete, or we have used up all of the time in the txg
1317 * timeout (default 5 sec).
1319 * - the spa is shutting down because this pool is being exported
1320 * or the machine is rebooting.
1322 * - the scan queue has reached its memory use limit
1324 uint64_t curr_time_ns
= gethrtime();
1325 uint64_t scan_time_ns
= curr_time_ns
- scn
->scn_sync_start_time
;
1326 uint64_t sync_time_ns
= curr_time_ns
-
1327 scn
->scn_dp
->dp_spa
->spa_sync_starttime
;
1328 int dirty_pct
= scn
->scn_dp
->dp_dirty_total
* 100 / zfs_dirty_data_max
;
1329 int mintime
= (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) ?
1330 zfs_resilver_min_time_ms
: zfs_scrub_min_time_ms
;
1332 if ((NSEC2MSEC(scan_time_ns
) > mintime
&&
1333 (dirty_pct
>= zfs_vdev_async_write_active_min_dirty_percent
||
1334 txg_sync_waiting(scn
->scn_dp
) ||
1335 NSEC2SEC(sync_time_ns
) >= zfs_txg_timeout
)) ||
1336 spa_shutting_down(scn
->scn_dp
->dp_spa
) ||
1337 (zfs_scan_strict_mem_lim
&& dsl_scan_should_clear(scn
))) {
1338 if (zb
&& zb
->zb_level
== ZB_ROOT_LEVEL
) {
1339 dprintf("suspending at first available bookmark "
1340 "%llx/%llx/%llx/%llx\n",
1341 (longlong_t
)zb
->zb_objset
,
1342 (longlong_t
)zb
->zb_object
,
1343 (longlong_t
)zb
->zb_level
,
1344 (longlong_t
)zb
->zb_blkid
);
1345 SET_BOOKMARK(&scn
->scn_phys
.scn_bookmark
,
1346 zb
->zb_objset
, 0, 0, 0);
1347 } else if (zb
!= NULL
) {
1348 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1349 (longlong_t
)zb
->zb_objset
,
1350 (longlong_t
)zb
->zb_object
,
1351 (longlong_t
)zb
->zb_level
,
1352 (longlong_t
)zb
->zb_blkid
);
1353 scn
->scn_phys
.scn_bookmark
= *zb
;
1356 dsl_scan_phys_t
*scnp
= &scn
->scn_phys
;
1357 dprintf("suspending at at DDT bookmark "
1358 "%llx/%llx/%llx/%llx\n",
1359 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_class
,
1360 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_type
,
1361 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_checksum
,
1362 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_cursor
);
1365 scn
->scn_suspending
= B_TRUE
;
1371 typedef struct zil_scan_arg
{
1373 zil_header_t
*zsa_zh
;
1378 dsl_scan_zil_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *arg
,
1381 zil_scan_arg_t
*zsa
= arg
;
1382 dsl_pool_t
*dp
= zsa
->zsa_dp
;
1383 dsl_scan_t
*scn
= dp
->dp_scan
;
1384 zil_header_t
*zh
= zsa
->zsa_zh
;
1385 zbookmark_phys_t zb
;
1387 ASSERT(!BP_IS_REDACTED(bp
));
1388 if (BP_IS_HOLE(bp
) || bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
)
1392 * One block ("stubby") can be allocated a long time ago; we
1393 * want to visit that one because it has been allocated
1394 * (on-disk) even if it hasn't been claimed (even though for
1395 * scrub there's nothing to do to it).
1397 if (claim_txg
== 0 && bp
->blk_birth
>= spa_min_claim_txg(dp
->dp_spa
))
1400 SET_BOOKMARK(&zb
, zh
->zh_log
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1401 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
1403 VERIFY(0 == scan_funcs
[scn
->scn_phys
.scn_func
](dp
, bp
, &zb
));
1409 dsl_scan_zil_record(zilog_t
*zilog
, const lr_t
*lrc
, void *arg
,
1412 if (lrc
->lrc_txtype
== TX_WRITE
) {
1413 zil_scan_arg_t
*zsa
= arg
;
1414 dsl_pool_t
*dp
= zsa
->zsa_dp
;
1415 dsl_scan_t
*scn
= dp
->dp_scan
;
1416 zil_header_t
*zh
= zsa
->zsa_zh
;
1417 const lr_write_t
*lr
= (const lr_write_t
*)lrc
;
1418 const blkptr_t
*bp
= &lr
->lr_blkptr
;
1419 zbookmark_phys_t zb
;
1421 ASSERT(!BP_IS_REDACTED(bp
));
1422 if (BP_IS_HOLE(bp
) ||
1423 bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
)
1427 * birth can be < claim_txg if this record's txg is
1428 * already txg sync'ed (but this log block contains
1429 * other records that are not synced)
1431 if (claim_txg
== 0 || bp
->blk_birth
< claim_txg
)
1434 SET_BOOKMARK(&zb
, zh
->zh_log
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1435 lr
->lr_foid
, ZB_ZIL_LEVEL
,
1436 lr
->lr_offset
/ BP_GET_LSIZE(bp
));
1438 VERIFY(0 == scan_funcs
[scn
->scn_phys
.scn_func
](dp
, bp
, &zb
));
1444 dsl_scan_zil(dsl_pool_t
*dp
, zil_header_t
*zh
)
1446 uint64_t claim_txg
= zh
->zh_claim_txg
;
1447 zil_scan_arg_t zsa
= { dp
, zh
};
1450 ASSERT(spa_writeable(dp
->dp_spa
));
1453 * We only want to visit blocks that have been claimed but not yet
1454 * replayed (or, in read-only mode, blocks that *would* be claimed).
1459 zilog
= zil_alloc(dp
->dp_meta_objset
, zh
);
1461 (void) zil_parse(zilog
, dsl_scan_zil_block
, dsl_scan_zil_record
, &zsa
,
1462 claim_txg
, B_FALSE
);
1468 * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1469 * here is to sort the AVL tree by the order each block will be needed.
1472 scan_prefetch_queue_compare(const void *a
, const void *b
)
1474 const scan_prefetch_issue_ctx_t
*spic_a
= a
, *spic_b
= b
;
1475 const scan_prefetch_ctx_t
*spc_a
= spic_a
->spic_spc
;
1476 const scan_prefetch_ctx_t
*spc_b
= spic_b
->spic_spc
;
1478 return (zbookmark_compare(spc_a
->spc_datablkszsec
,
1479 spc_a
->spc_indblkshift
, spc_b
->spc_datablkszsec
,
1480 spc_b
->spc_indblkshift
, &spic_a
->spic_zb
, &spic_b
->spic_zb
));
1484 scan_prefetch_ctx_rele(scan_prefetch_ctx_t
*spc
, void *tag
)
1486 if (zfs_refcount_remove(&spc
->spc_refcnt
, tag
) == 0) {
1487 zfs_refcount_destroy(&spc
->spc_refcnt
);
1488 kmem_free(spc
, sizeof (scan_prefetch_ctx_t
));
1492 static scan_prefetch_ctx_t
*
1493 scan_prefetch_ctx_create(dsl_scan_t
*scn
, dnode_phys_t
*dnp
, void *tag
)
1495 scan_prefetch_ctx_t
*spc
;
1497 spc
= kmem_alloc(sizeof (scan_prefetch_ctx_t
), KM_SLEEP
);
1498 zfs_refcount_create(&spc
->spc_refcnt
);
1499 zfs_refcount_add(&spc
->spc_refcnt
, tag
);
1502 spc
->spc_datablkszsec
= dnp
->dn_datablkszsec
;
1503 spc
->spc_indblkshift
= dnp
->dn_indblkshift
;
1504 spc
->spc_root
= B_FALSE
;
1506 spc
->spc_datablkszsec
= 0;
1507 spc
->spc_indblkshift
= 0;
1508 spc
->spc_root
= B_TRUE
;
1515 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t
*spc
, void *tag
)
1517 zfs_refcount_add(&spc
->spc_refcnt
, tag
);
1521 scan_ds_prefetch_queue_clear(dsl_scan_t
*scn
)
1523 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1524 void *cookie
= NULL
;
1525 scan_prefetch_issue_ctx_t
*spic
= NULL
;
1527 mutex_enter(&spa
->spa_scrub_lock
);
1528 while ((spic
= avl_destroy_nodes(&scn
->scn_prefetch_queue
,
1529 &cookie
)) != NULL
) {
1530 scan_prefetch_ctx_rele(spic
->spic_spc
, scn
);
1531 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1533 mutex_exit(&spa
->spa_scrub_lock
);
1537 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t
*spc
,
1538 const zbookmark_phys_t
*zb
)
1540 zbookmark_phys_t
*last_zb
= &spc
->spc_scn
->scn_prefetch_bookmark
;
1541 dnode_phys_t tmp_dnp
;
1542 dnode_phys_t
*dnp
= (spc
->spc_root
) ? NULL
: &tmp_dnp
;
1544 if (zb
->zb_objset
!= last_zb
->zb_objset
)
1546 if ((int64_t)zb
->zb_object
< 0)
1549 tmp_dnp
.dn_datablkszsec
= spc
->spc_datablkszsec
;
1550 tmp_dnp
.dn_indblkshift
= spc
->spc_indblkshift
;
1552 if (zbookmark_subtree_completed(dnp
, zb
, last_zb
))
1559 dsl_scan_prefetch(scan_prefetch_ctx_t
*spc
, blkptr_t
*bp
, zbookmark_phys_t
*zb
)
1562 dsl_scan_t
*scn
= spc
->spc_scn
;
1563 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1564 scan_prefetch_issue_ctx_t
*spic
;
1566 if (zfs_no_scrub_prefetch
|| BP_IS_REDACTED(bp
))
1569 if (BP_IS_HOLE(bp
) || bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
||
1570 (BP_GET_LEVEL(bp
) == 0 && BP_GET_TYPE(bp
) != DMU_OT_DNODE
&&
1571 BP_GET_TYPE(bp
) != DMU_OT_OBJSET
))
1574 if (dsl_scan_check_prefetch_resume(spc
, zb
))
1577 scan_prefetch_ctx_add_ref(spc
, scn
);
1578 spic
= kmem_alloc(sizeof (scan_prefetch_issue_ctx_t
), KM_SLEEP
);
1579 spic
->spic_spc
= spc
;
1580 spic
->spic_bp
= *bp
;
1581 spic
->spic_zb
= *zb
;
1584 * Add the IO to the queue of blocks to prefetch. This allows us to
1585 * prioritize blocks that we will need first for the main traversal
1588 mutex_enter(&spa
->spa_scrub_lock
);
1589 if (avl_find(&scn
->scn_prefetch_queue
, spic
, &idx
) != NULL
) {
1590 /* this block is already queued for prefetch */
1591 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1592 scan_prefetch_ctx_rele(spc
, scn
);
1593 mutex_exit(&spa
->spa_scrub_lock
);
1597 avl_insert(&scn
->scn_prefetch_queue
, spic
, idx
);
1598 cv_broadcast(&spa
->spa_scrub_io_cv
);
1599 mutex_exit(&spa
->spa_scrub_lock
);
1603 dsl_scan_prefetch_dnode(dsl_scan_t
*scn
, dnode_phys_t
*dnp
,
1604 uint64_t objset
, uint64_t object
)
1607 zbookmark_phys_t zb
;
1608 scan_prefetch_ctx_t
*spc
;
1610 if (dnp
->dn_nblkptr
== 0 && !(dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1613 SET_BOOKMARK(&zb
, objset
, object
, 0, 0);
1615 spc
= scan_prefetch_ctx_create(scn
, dnp
, FTAG
);
1617 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++) {
1618 zb
.zb_level
= BP_GET_LEVEL(&dnp
->dn_blkptr
[i
]);
1620 dsl_scan_prefetch(spc
, &dnp
->dn_blkptr
[i
], &zb
);
1623 if (dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) {
1625 zb
.zb_blkid
= DMU_SPILL_BLKID
;
1626 dsl_scan_prefetch(spc
, DN_SPILL_BLKPTR(dnp
), &zb
);
1629 scan_prefetch_ctx_rele(spc
, FTAG
);
1633 dsl_scan_prefetch_cb(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1634 arc_buf_t
*buf
, void *private)
1636 scan_prefetch_ctx_t
*spc
= private;
1637 dsl_scan_t
*scn
= spc
->spc_scn
;
1638 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1640 /* broadcast that the IO has completed for rate limiting purposes */
1641 mutex_enter(&spa
->spa_scrub_lock
);
1642 ASSERT3U(spa
->spa_scrub_inflight
, >=, BP_GET_PSIZE(bp
));
1643 spa
->spa_scrub_inflight
-= BP_GET_PSIZE(bp
);
1644 cv_broadcast(&spa
->spa_scrub_io_cv
);
1645 mutex_exit(&spa
->spa_scrub_lock
);
1647 /* if there was an error or we are done prefetching, just cleanup */
1648 if (buf
== NULL
|| scn
->scn_prefetch_stop
)
1651 if (BP_GET_LEVEL(bp
) > 0) {
1654 int epb
= BP_GET_LSIZE(bp
) >> SPA_BLKPTRSHIFT
;
1655 zbookmark_phys_t czb
;
1657 for (i
= 0, cbp
= buf
->b_data
; i
< epb
; i
++, cbp
++) {
1658 SET_BOOKMARK(&czb
, zb
->zb_objset
, zb
->zb_object
,
1659 zb
->zb_level
- 1, zb
->zb_blkid
* epb
+ i
);
1660 dsl_scan_prefetch(spc
, cbp
, &czb
);
1662 } else if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
) {
1665 int epb
= BP_GET_LSIZE(bp
) >> DNODE_SHIFT
;
1667 for (i
= 0, cdnp
= buf
->b_data
; i
< epb
;
1668 i
+= cdnp
->dn_extra_slots
+ 1,
1669 cdnp
+= cdnp
->dn_extra_slots
+ 1) {
1670 dsl_scan_prefetch_dnode(scn
, cdnp
,
1671 zb
->zb_objset
, zb
->zb_blkid
* epb
+ i
);
1673 } else if (BP_GET_TYPE(bp
) == DMU_OT_OBJSET
) {
1674 objset_phys_t
*osp
= buf
->b_data
;
1676 dsl_scan_prefetch_dnode(scn
, &osp
->os_meta_dnode
,
1677 zb
->zb_objset
, DMU_META_DNODE_OBJECT
);
1679 if (OBJSET_BUF_HAS_USERUSED(buf
)) {
1680 dsl_scan_prefetch_dnode(scn
,
1681 &osp
->os_groupused_dnode
, zb
->zb_objset
,
1682 DMU_GROUPUSED_OBJECT
);
1683 dsl_scan_prefetch_dnode(scn
,
1684 &osp
->os_userused_dnode
, zb
->zb_objset
,
1685 DMU_USERUSED_OBJECT
);
1691 arc_buf_destroy(buf
, private);
1692 scan_prefetch_ctx_rele(spc
, scn
);
1697 dsl_scan_prefetch_thread(void *arg
)
1699 dsl_scan_t
*scn
= arg
;
1700 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1701 scan_prefetch_issue_ctx_t
*spic
;
1703 /* loop until we are told to stop */
1704 while (!scn
->scn_prefetch_stop
) {
1705 arc_flags_t flags
= ARC_FLAG_NOWAIT
|
1706 ARC_FLAG_PRESCIENT_PREFETCH
| ARC_FLAG_PREFETCH
;
1707 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SCAN_THREAD
;
1709 mutex_enter(&spa
->spa_scrub_lock
);
1712 * Wait until we have an IO to issue and are not above our
1713 * maximum in flight limit.
1715 while (!scn
->scn_prefetch_stop
&&
1716 (avl_numnodes(&scn
->scn_prefetch_queue
) == 0 ||
1717 spa
->spa_scrub_inflight
>= scn
->scn_maxinflight_bytes
)) {
1718 cv_wait(&spa
->spa_scrub_io_cv
, &spa
->spa_scrub_lock
);
1721 /* recheck if we should stop since we waited for the cv */
1722 if (scn
->scn_prefetch_stop
) {
1723 mutex_exit(&spa
->spa_scrub_lock
);
1727 /* remove the prefetch IO from the tree */
1728 spic
= avl_first(&scn
->scn_prefetch_queue
);
1729 spa
->spa_scrub_inflight
+= BP_GET_PSIZE(&spic
->spic_bp
);
1730 avl_remove(&scn
->scn_prefetch_queue
, spic
);
1732 mutex_exit(&spa
->spa_scrub_lock
);
1734 if (BP_IS_PROTECTED(&spic
->spic_bp
)) {
1735 ASSERT(BP_GET_TYPE(&spic
->spic_bp
) == DMU_OT_DNODE
||
1736 BP_GET_TYPE(&spic
->spic_bp
) == DMU_OT_OBJSET
);
1737 ASSERT3U(BP_GET_LEVEL(&spic
->spic_bp
), ==, 0);
1738 zio_flags
|= ZIO_FLAG_RAW
;
1741 /* issue the prefetch asynchronously */
1742 (void) arc_read(scn
->scn_zio_root
, scn
->scn_dp
->dp_spa
,
1743 &spic
->spic_bp
, dsl_scan_prefetch_cb
, spic
->spic_spc
,
1744 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, &spic
->spic_zb
);
1746 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1749 ASSERT(scn
->scn_prefetch_stop
);
1751 /* free any prefetches we didn't get to complete */
1752 mutex_enter(&spa
->spa_scrub_lock
);
1753 while ((spic
= avl_first(&scn
->scn_prefetch_queue
)) != NULL
) {
1754 avl_remove(&scn
->scn_prefetch_queue
, spic
);
1755 scan_prefetch_ctx_rele(spic
->spic_spc
, scn
);
1756 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1758 ASSERT0(avl_numnodes(&scn
->scn_prefetch_queue
));
1759 mutex_exit(&spa
->spa_scrub_lock
);
1763 dsl_scan_check_resume(dsl_scan_t
*scn
, const dnode_phys_t
*dnp
,
1764 const zbookmark_phys_t
*zb
)
1767 * We never skip over user/group accounting objects (obj<0)
1769 if (!ZB_IS_ZERO(&scn
->scn_phys
.scn_bookmark
) &&
1770 (int64_t)zb
->zb_object
>= 0) {
1772 * If we already visited this bp & everything below (in
1773 * a prior txg sync), don't bother doing it again.
1775 if (zbookmark_subtree_completed(dnp
, zb
,
1776 &scn
->scn_phys
.scn_bookmark
))
1780 * If we found the block we're trying to resume from, or
1781 * we went past it to a different object, zero it out to
1782 * indicate that it's OK to start checking for suspending
1785 if (bcmp(zb
, &scn
->scn_phys
.scn_bookmark
, sizeof (*zb
)) == 0 ||
1786 zb
->zb_object
> scn
->scn_phys
.scn_bookmark
.zb_object
) {
1787 dprintf("resuming at %llx/%llx/%llx/%llx\n",
1788 (longlong_t
)zb
->zb_objset
,
1789 (longlong_t
)zb
->zb_object
,
1790 (longlong_t
)zb
->zb_level
,
1791 (longlong_t
)zb
->zb_blkid
);
1792 bzero(&scn
->scn_phys
.scn_bookmark
, sizeof (*zb
));
1798 static void dsl_scan_visitbp(blkptr_t
*bp
, const zbookmark_phys_t
*zb
,
1799 dnode_phys_t
*dnp
, dsl_dataset_t
*ds
, dsl_scan_t
*scn
,
1800 dmu_objset_type_t ostype
, dmu_tx_t
*tx
);
1801 inline __attribute__((always_inline
)) static void dsl_scan_visitdnode(
1802 dsl_scan_t
*, dsl_dataset_t
*ds
, dmu_objset_type_t ostype
,
1803 dnode_phys_t
*dnp
, uint64_t object
, dmu_tx_t
*tx
);
1806 * Return nonzero on i/o error.
1807 * Return new buf to write out in *bufp.
1809 inline __attribute__((always_inline
)) static int
1810 dsl_scan_recurse(dsl_scan_t
*scn
, dsl_dataset_t
*ds
, dmu_objset_type_t ostype
,
1811 dnode_phys_t
*dnp
, const blkptr_t
*bp
,
1812 const zbookmark_phys_t
*zb
, dmu_tx_t
*tx
)
1814 dsl_pool_t
*dp
= scn
->scn_dp
;
1815 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SCAN_THREAD
;
1818 ASSERT(!BP_IS_REDACTED(bp
));
1820 if (BP_GET_LEVEL(bp
) > 0) {
1821 arc_flags_t flags
= ARC_FLAG_WAIT
;
1824 int epb
= BP_GET_LSIZE(bp
) >> SPA_BLKPTRSHIFT
;
1827 err
= arc_read(NULL
, dp
->dp_spa
, bp
, arc_getbuf_func
, &buf
,
1828 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, zb
);
1830 scn
->scn_phys
.scn_errors
++;
1833 for (i
= 0, cbp
= buf
->b_data
; i
< epb
; i
++, cbp
++) {
1834 zbookmark_phys_t czb
;
1836 SET_BOOKMARK(&czb
, zb
->zb_objset
, zb
->zb_object
,
1838 zb
->zb_blkid
* epb
+ i
);
1839 dsl_scan_visitbp(cbp
, &czb
, dnp
,
1840 ds
, scn
, ostype
, tx
);
1842 arc_buf_destroy(buf
, &buf
);
1843 } else if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
) {
1844 arc_flags_t flags
= ARC_FLAG_WAIT
;
1847 int epb
= BP_GET_LSIZE(bp
) >> DNODE_SHIFT
;
1850 if (BP_IS_PROTECTED(bp
)) {
1851 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
1852 zio_flags
|= ZIO_FLAG_RAW
;
1855 err
= arc_read(NULL
, dp
->dp_spa
, bp
, arc_getbuf_func
, &buf
,
1856 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, zb
);
1858 scn
->scn_phys
.scn_errors
++;
1861 for (i
= 0, cdnp
= buf
->b_data
; i
< epb
;
1862 i
+= cdnp
->dn_extra_slots
+ 1,
1863 cdnp
+= cdnp
->dn_extra_slots
+ 1) {
1864 dsl_scan_visitdnode(scn
, ds
, ostype
,
1865 cdnp
, zb
->zb_blkid
* epb
+ i
, tx
);
1868 arc_buf_destroy(buf
, &buf
);
1869 } else if (BP_GET_TYPE(bp
) == DMU_OT_OBJSET
) {
1870 arc_flags_t flags
= ARC_FLAG_WAIT
;
1874 err
= arc_read(NULL
, dp
->dp_spa
, bp
, arc_getbuf_func
, &buf
,
1875 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, zb
);
1877 scn
->scn_phys
.scn_errors
++;
1883 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1884 &osp
->os_meta_dnode
, DMU_META_DNODE_OBJECT
, tx
);
1886 if (OBJSET_BUF_HAS_USERUSED(buf
)) {
1888 * We also always visit user/group/project accounting
1889 * objects, and never skip them, even if we are
1890 * suspending. This is necessary so that the
1891 * space deltas from this txg get integrated.
1893 if (OBJSET_BUF_HAS_PROJECTUSED(buf
))
1894 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1895 &osp
->os_projectused_dnode
,
1896 DMU_PROJECTUSED_OBJECT
, tx
);
1897 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1898 &osp
->os_groupused_dnode
,
1899 DMU_GROUPUSED_OBJECT
, tx
);
1900 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1901 &osp
->os_userused_dnode
,
1902 DMU_USERUSED_OBJECT
, tx
);
1904 arc_buf_destroy(buf
, &buf
);
1910 inline __attribute__((always_inline
)) static void
1911 dsl_scan_visitdnode(dsl_scan_t
*scn
, dsl_dataset_t
*ds
,
1912 dmu_objset_type_t ostype
, dnode_phys_t
*dnp
,
1913 uint64_t object
, dmu_tx_t
*tx
)
1917 for (j
= 0; j
< dnp
->dn_nblkptr
; j
++) {
1918 zbookmark_phys_t czb
;
1920 SET_BOOKMARK(&czb
, ds
? ds
->ds_object
: 0, object
,
1921 dnp
->dn_nlevels
- 1, j
);
1922 dsl_scan_visitbp(&dnp
->dn_blkptr
[j
],
1923 &czb
, dnp
, ds
, scn
, ostype
, tx
);
1926 if (dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) {
1927 zbookmark_phys_t czb
;
1928 SET_BOOKMARK(&czb
, ds
? ds
->ds_object
: 0, object
,
1929 0, DMU_SPILL_BLKID
);
1930 dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp
),
1931 &czb
, dnp
, ds
, scn
, ostype
, tx
);
1936 * The arguments are in this order because mdb can only print the
1937 * first 5; we want them to be useful.
1940 dsl_scan_visitbp(blkptr_t
*bp
, const zbookmark_phys_t
*zb
,
1941 dnode_phys_t
*dnp
, dsl_dataset_t
*ds
, dsl_scan_t
*scn
,
1942 dmu_objset_type_t ostype
, dmu_tx_t
*tx
)
1944 dsl_pool_t
*dp
= scn
->scn_dp
;
1945 blkptr_t
*bp_toread
= NULL
;
1947 if (dsl_scan_check_suspend(scn
, zb
))
1950 if (dsl_scan_check_resume(scn
, dnp
, zb
))
1953 scn
->scn_visited_this_txg
++;
1956 * This debugging is commented out to conserve stack space. This
1957 * function is called recursively and the debugging adds several
1958 * bytes to the stack for each call. It can be commented back in
1959 * if required to debug an issue in dsl_scan_visitbp().
1962 * "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
1963 * ds, ds ? ds->ds_object : 0,
1964 * zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1968 if (BP_IS_HOLE(bp
)) {
1969 scn
->scn_holes_this_txg
++;
1973 if (BP_IS_REDACTED(bp
)) {
1974 ASSERT(dsl_dataset_feature_is_active(ds
,
1975 SPA_FEATURE_REDACTED_DATASETS
));
1979 if (bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
) {
1980 scn
->scn_lt_min_this_txg
++;
1984 bp_toread
= kmem_alloc(sizeof (blkptr_t
), KM_SLEEP
);
1987 if (dsl_scan_recurse(scn
, ds
, ostype
, dnp
, bp_toread
, zb
, tx
) != 0)
1991 * If dsl_scan_ddt() has already visited this block, it will have
1992 * already done any translations or scrubbing, so don't call the
1995 if (ddt_class_contains(dp
->dp_spa
,
1996 scn
->scn_phys
.scn_ddt_class_max
, bp
)) {
1997 scn
->scn_ddt_contained_this_txg
++;
2002 * If this block is from the future (after cur_max_txg), then we
2003 * are doing this on behalf of a deleted snapshot, and we will
2004 * revisit the future block on the next pass of this dataset.
2005 * Don't scan it now unless we need to because something
2006 * under it was modified.
2008 if (BP_PHYSICAL_BIRTH(bp
) > scn
->scn_phys
.scn_cur_max_txg
) {
2009 scn
->scn_gt_max_this_txg
++;
2013 scan_funcs
[scn
->scn_phys
.scn_func
](dp
, bp
, zb
);
2016 kmem_free(bp_toread
, sizeof (blkptr_t
));
2020 dsl_scan_visit_rootbp(dsl_scan_t
*scn
, dsl_dataset_t
*ds
, blkptr_t
*bp
,
2023 zbookmark_phys_t zb
;
2024 scan_prefetch_ctx_t
*spc
;
2026 SET_BOOKMARK(&zb
, ds
? ds
->ds_object
: DMU_META_OBJSET
,
2027 ZB_ROOT_OBJECT
, ZB_ROOT_LEVEL
, ZB_ROOT_BLKID
);
2029 if (ZB_IS_ZERO(&scn
->scn_phys
.scn_bookmark
)) {
2030 SET_BOOKMARK(&scn
->scn_prefetch_bookmark
,
2031 zb
.zb_objset
, 0, 0, 0);
2033 scn
->scn_prefetch_bookmark
= scn
->scn_phys
.scn_bookmark
;
2036 scn
->scn_objsets_visited_this_txg
++;
2038 spc
= scan_prefetch_ctx_create(scn
, NULL
, FTAG
);
2039 dsl_scan_prefetch(spc
, bp
, &zb
);
2040 scan_prefetch_ctx_rele(spc
, FTAG
);
2042 dsl_scan_visitbp(bp
, &zb
, NULL
, ds
, scn
, DMU_OST_NONE
, tx
);
2044 dprintf_ds(ds
, "finished scan%s", "");
2048 ds_destroyed_scn_phys(dsl_dataset_t
*ds
, dsl_scan_phys_t
*scn_phys
)
2050 if (scn_phys
->scn_bookmark
.zb_objset
== ds
->ds_object
) {
2051 if (ds
->ds_is_snapshot
) {
2054 * - scn_cur_{min,max}_txg stays the same.
2055 * - Setting the flag is not really necessary if
2056 * scn_cur_max_txg == scn_max_txg, because there
2057 * is nothing after this snapshot that we care
2058 * about. However, we set it anyway and then
2059 * ignore it when we retraverse it in
2060 * dsl_scan_visitds().
2062 scn_phys
->scn_bookmark
.zb_objset
=
2063 dsl_dataset_phys(ds
)->ds_next_snap_obj
;
2064 zfs_dbgmsg("destroying ds %llu; currently traversing; "
2065 "reset zb_objset to %llu",
2066 (u_longlong_t
)ds
->ds_object
,
2067 (u_longlong_t
)dsl_dataset_phys(ds
)->
2069 scn_phys
->scn_flags
|= DSF_VISIT_DS_AGAIN
;
2071 SET_BOOKMARK(&scn_phys
->scn_bookmark
,
2072 ZB_DESTROYED_OBJSET
, 0, 0, 0);
2073 zfs_dbgmsg("destroying ds %llu; currently traversing; "
2074 "reset bookmark to -1,0,0,0",
2075 (u_longlong_t
)ds
->ds_object
);
2081 * Invoked when a dataset is destroyed. We need to make sure that:
2083 * 1) If it is the dataset that was currently being scanned, we write
2084 * a new dsl_scan_phys_t and marking the objset reference in it
2086 * 2) Remove it from the work queue, if it was present.
2088 * If the dataset was actually a snapshot, instead of marking the dataset
2089 * as destroyed, we instead substitute the next snapshot in line.
2092 dsl_scan_ds_destroyed(dsl_dataset_t
*ds
, dmu_tx_t
*tx
)
2094 dsl_pool_t
*dp
= ds
->ds_dir
->dd_pool
;
2095 dsl_scan_t
*scn
= dp
->dp_scan
;
2098 if (!dsl_scan_is_running(scn
))
2101 ds_destroyed_scn_phys(ds
, &scn
->scn_phys
);
2102 ds_destroyed_scn_phys(ds
, &scn
->scn_phys_cached
);
2104 if (scan_ds_queue_contains(scn
, ds
->ds_object
, &mintxg
)) {
2105 scan_ds_queue_remove(scn
, ds
->ds_object
);
2106 if (ds
->ds_is_snapshot
)
2107 scan_ds_queue_insert(scn
,
2108 dsl_dataset_phys(ds
)->ds_next_snap_obj
, mintxg
);
2111 if (zap_lookup_int_key(dp
->dp_meta_objset
, scn
->scn_phys
.scn_queue_obj
,
2112 ds
->ds_object
, &mintxg
) == 0) {
2113 ASSERT3U(dsl_dataset_phys(ds
)->ds_num_children
, <=, 1);
2114 VERIFY3U(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2115 scn
->scn_phys
.scn_queue_obj
, ds
->ds_object
, tx
));
2116 if (ds
->ds_is_snapshot
) {
2118 * We keep the same mintxg; it could be >
2119 * ds_creation_txg if the previous snapshot was
2122 VERIFY(zap_add_int_key(dp
->dp_meta_objset
,
2123 scn
->scn_phys
.scn_queue_obj
,
2124 dsl_dataset_phys(ds
)->ds_next_snap_obj
,
2126 zfs_dbgmsg("destroying ds %llu; in queue; "
2127 "replacing with %llu",
2128 (u_longlong_t
)ds
->ds_object
,
2129 (u_longlong_t
)dsl_dataset_phys(ds
)->
2132 zfs_dbgmsg("destroying ds %llu; in queue; removing",
2133 (u_longlong_t
)ds
->ds_object
);
2138 * dsl_scan_sync() should be called after this, and should sync
2139 * out our changed state, but just to be safe, do it here.
2141 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
2145 ds_snapshotted_bookmark(dsl_dataset_t
*ds
, zbookmark_phys_t
*scn_bookmark
)
2147 if (scn_bookmark
->zb_objset
== ds
->ds_object
) {
2148 scn_bookmark
->zb_objset
=
2149 dsl_dataset_phys(ds
)->ds_prev_snap_obj
;
2150 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
2151 "reset zb_objset to %llu",
2152 (u_longlong_t
)ds
->ds_object
,
2153 (u_longlong_t
)dsl_dataset_phys(ds
)->ds_prev_snap_obj
);
2158 * Called when a dataset is snapshotted. If we were currently traversing
2159 * this snapshot, we reset our bookmark to point at the newly created
2160 * snapshot. We also modify our work queue to remove the old snapshot and
2161 * replace with the new one.
2164 dsl_scan_ds_snapshotted(dsl_dataset_t
*ds
, dmu_tx_t
*tx
)
2166 dsl_pool_t
*dp
= ds
->ds_dir
->dd_pool
;
2167 dsl_scan_t
*scn
= dp
->dp_scan
;
2170 if (!dsl_scan_is_running(scn
))
2173 ASSERT(dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0);
2175 ds_snapshotted_bookmark(ds
, &scn
->scn_phys
.scn_bookmark
);
2176 ds_snapshotted_bookmark(ds
, &scn
->scn_phys_cached
.scn_bookmark
);
2178 if (scan_ds_queue_contains(scn
, ds
->ds_object
, &mintxg
)) {
2179 scan_ds_queue_remove(scn
, ds
->ds_object
);
2180 scan_ds_queue_insert(scn
,
2181 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, mintxg
);
2184 if (zap_lookup_int_key(dp
->dp_meta_objset
, scn
->scn_phys
.scn_queue_obj
,
2185 ds
->ds_object
, &mintxg
) == 0) {
2186 VERIFY3U(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2187 scn
->scn_phys
.scn_queue_obj
, ds
->ds_object
, tx
));
2188 VERIFY(zap_add_int_key(dp
->dp_meta_objset
,
2189 scn
->scn_phys
.scn_queue_obj
,
2190 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, mintxg
, tx
) == 0);
2191 zfs_dbgmsg("snapshotting ds %llu; in queue; "
2192 "replacing with %llu",
2193 (u_longlong_t
)ds
->ds_object
,
2194 (u_longlong_t
)dsl_dataset_phys(ds
)->ds_prev_snap_obj
);
2197 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
2201 ds_clone_swapped_bookmark(dsl_dataset_t
*ds1
, dsl_dataset_t
*ds2
,
2202 zbookmark_phys_t
*scn_bookmark
)
2204 if (scn_bookmark
->zb_objset
== ds1
->ds_object
) {
2205 scn_bookmark
->zb_objset
= ds2
->ds_object
;
2206 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2207 "reset zb_objset to %llu",
2208 (u_longlong_t
)ds1
->ds_object
,
2209 (u_longlong_t
)ds2
->ds_object
);
2210 } else if (scn_bookmark
->zb_objset
== ds2
->ds_object
) {
2211 scn_bookmark
->zb_objset
= ds1
->ds_object
;
2212 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2213 "reset zb_objset to %llu",
2214 (u_longlong_t
)ds2
->ds_object
,
2215 (u_longlong_t
)ds1
->ds_object
);
2220 * Called when an origin dataset and its clone are swapped. If we were
2221 * currently traversing the dataset, we need to switch to traversing the
2222 * newly promoted clone.
2225 dsl_scan_ds_clone_swapped(dsl_dataset_t
*ds1
, dsl_dataset_t
*ds2
, dmu_tx_t
*tx
)
2227 dsl_pool_t
*dp
= ds1
->ds_dir
->dd_pool
;
2228 dsl_scan_t
*scn
= dp
->dp_scan
;
2229 uint64_t mintxg1
, mintxg2
;
2230 boolean_t ds1_queued
, ds2_queued
;
2232 if (!dsl_scan_is_running(scn
))
2235 ds_clone_swapped_bookmark(ds1
, ds2
, &scn
->scn_phys
.scn_bookmark
);
2236 ds_clone_swapped_bookmark(ds1
, ds2
, &scn
->scn_phys_cached
.scn_bookmark
);
2239 * Handle the in-memory scan queue.
2241 ds1_queued
= scan_ds_queue_contains(scn
, ds1
->ds_object
, &mintxg1
);
2242 ds2_queued
= scan_ds_queue_contains(scn
, ds2
->ds_object
, &mintxg2
);
2244 /* Sanity checking. */
2246 ASSERT3U(mintxg1
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2247 ASSERT3U(mintxg1
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2250 ASSERT3U(mintxg2
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2251 ASSERT3U(mintxg2
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2254 if (ds1_queued
&& ds2_queued
) {
2256 * If both are queued, we don't need to do anything.
2257 * The swapping code below would not handle this case correctly,
2258 * since we can't insert ds2 if it is already there. That's
2259 * because scan_ds_queue_insert() prohibits a duplicate insert
2262 } else if (ds1_queued
) {
2263 scan_ds_queue_remove(scn
, ds1
->ds_object
);
2264 scan_ds_queue_insert(scn
, ds2
->ds_object
, mintxg1
);
2265 } else if (ds2_queued
) {
2266 scan_ds_queue_remove(scn
, ds2
->ds_object
);
2267 scan_ds_queue_insert(scn
, ds1
->ds_object
, mintxg2
);
2271 * Handle the on-disk scan queue.
2272 * The on-disk state is an out-of-date version of the in-memory state,
2273 * so the in-memory and on-disk values for ds1_queued and ds2_queued may
2274 * be different. Therefore we need to apply the swap logic to the
2275 * on-disk state independently of the in-memory state.
2277 ds1_queued
= zap_lookup_int_key(dp
->dp_meta_objset
,
2278 scn
->scn_phys
.scn_queue_obj
, ds1
->ds_object
, &mintxg1
) == 0;
2279 ds2_queued
= zap_lookup_int_key(dp
->dp_meta_objset
,
2280 scn
->scn_phys
.scn_queue_obj
, ds2
->ds_object
, &mintxg2
) == 0;
2282 /* Sanity checking. */
2284 ASSERT3U(mintxg1
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2285 ASSERT3U(mintxg1
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2288 ASSERT3U(mintxg2
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2289 ASSERT3U(mintxg2
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2292 if (ds1_queued
&& ds2_queued
) {
2294 * If both are queued, we don't need to do anything.
2295 * Alternatively, we could check for EEXIST from
2296 * zap_add_int_key() and back out to the original state, but
2297 * that would be more work than checking for this case upfront.
2299 } else if (ds1_queued
) {
2300 VERIFY3S(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2301 scn
->scn_phys
.scn_queue_obj
, ds1
->ds_object
, tx
));
2302 VERIFY3S(0, ==, zap_add_int_key(dp
->dp_meta_objset
,
2303 scn
->scn_phys
.scn_queue_obj
, ds2
->ds_object
, mintxg1
, tx
));
2304 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2305 "replacing with %llu",
2306 (u_longlong_t
)ds1
->ds_object
,
2307 (u_longlong_t
)ds2
->ds_object
);
2308 } else if (ds2_queued
) {
2309 VERIFY3S(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2310 scn
->scn_phys
.scn_queue_obj
, ds2
->ds_object
, tx
));
2311 VERIFY3S(0, ==, zap_add_int_key(dp
->dp_meta_objset
,
2312 scn
->scn_phys
.scn_queue_obj
, ds1
->ds_object
, mintxg2
, tx
));
2313 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2314 "replacing with %llu",
2315 (u_longlong_t
)ds2
->ds_object
,
2316 (u_longlong_t
)ds1
->ds_object
);
2319 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
2324 enqueue_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
2326 uint64_t originobj
= *(uint64_t *)arg
;
2329 dsl_scan_t
*scn
= dp
->dp_scan
;
2331 if (dsl_dir_phys(hds
->ds_dir
)->dd_origin_obj
!= originobj
)
2334 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
2338 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= originobj
) {
2339 dsl_dataset_t
*prev
;
2340 err
= dsl_dataset_hold_obj(dp
,
2341 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
2343 dsl_dataset_rele(ds
, FTAG
);
2348 scan_ds_queue_insert(scn
, ds
->ds_object
,
2349 dsl_dataset_phys(ds
)->ds_prev_snap_txg
);
2350 dsl_dataset_rele(ds
, FTAG
);
2355 dsl_scan_visitds(dsl_scan_t
*scn
, uint64_t dsobj
, dmu_tx_t
*tx
)
2357 dsl_pool_t
*dp
= scn
->scn_dp
;
2360 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
2362 if (scn
->scn_phys
.scn_cur_min_txg
>=
2363 scn
->scn_phys
.scn_max_txg
) {
2365 * This can happen if this snapshot was created after the
2366 * scan started, and we already completed a previous snapshot
2367 * that was created after the scan started. This snapshot
2368 * only references blocks with:
2370 * birth < our ds_creation_txg
2371 * cur_min_txg is no less than ds_creation_txg.
2372 * We have already visited these blocks.
2374 * birth > scn_max_txg
2375 * The scan requested not to visit these blocks.
2377 * Subsequent snapshots (and clones) can reference our
2378 * blocks, or blocks with even higher birth times.
2379 * Therefore we do not need to visit them either,
2380 * so we do not add them to the work queue.
2382 * Note that checking for cur_min_txg >= cur_max_txg
2383 * is not sufficient, because in that case we may need to
2384 * visit subsequent snapshots. This happens when min_txg > 0,
2385 * which raises cur_min_txg. In this case we will visit
2386 * this dataset but skip all of its blocks, because the
2387 * rootbp's birth time is < cur_min_txg. Then we will
2388 * add the next snapshots/clones to the work queue.
2390 char *dsname
= kmem_alloc(ZFS_MAX_DATASET_NAME_LEN
, KM_SLEEP
);
2391 dsl_dataset_name(ds
, dsname
);
2392 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2393 "cur_min_txg (%llu) >= max_txg (%llu)",
2394 (longlong_t
)dsobj
, dsname
,
2395 (longlong_t
)scn
->scn_phys
.scn_cur_min_txg
,
2396 (longlong_t
)scn
->scn_phys
.scn_max_txg
);
2397 kmem_free(dsname
, MAXNAMELEN
);
2403 * Only the ZIL in the head (non-snapshot) is valid. Even though
2404 * snapshots can have ZIL block pointers (which may be the same
2405 * BP as in the head), they must be ignored. In addition, $ORIGIN
2406 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2407 * need to look for a ZIL in it either. So we traverse the ZIL here,
2408 * rather than in scan_recurse(), because the regular snapshot
2409 * block-sharing rules don't apply to it.
2411 if (!dsl_dataset_is_snapshot(ds
) &&
2412 (dp
->dp_origin_snap
== NULL
||
2413 ds
->ds_dir
!= dp
->dp_origin_snap
->ds_dir
)) {
2415 if (dmu_objset_from_ds(ds
, &os
) != 0) {
2418 dsl_scan_zil(dp
, &os
->os_zil_header
);
2422 * Iterate over the bps in this ds.
2424 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
2425 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
2426 dsl_scan_visit_rootbp(scn
, ds
, &dsl_dataset_phys(ds
)->ds_bp
, tx
);
2427 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
2429 char *dsname
= kmem_alloc(ZFS_MAX_DATASET_NAME_LEN
, KM_SLEEP
);
2430 dsl_dataset_name(ds
, dsname
);
2431 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2433 (longlong_t
)dsobj
, dsname
,
2434 (longlong_t
)scn
->scn_phys
.scn_cur_min_txg
,
2435 (longlong_t
)scn
->scn_phys
.scn_cur_max_txg
,
2436 (int)scn
->scn_suspending
);
2437 kmem_free(dsname
, ZFS_MAX_DATASET_NAME_LEN
);
2439 if (scn
->scn_suspending
)
2443 * We've finished this pass over this dataset.
2447 * If we did not completely visit this dataset, do another pass.
2449 if (scn
->scn_phys
.scn_flags
& DSF_VISIT_DS_AGAIN
) {
2450 zfs_dbgmsg("incomplete pass; visiting again");
2451 scn
->scn_phys
.scn_flags
&= ~DSF_VISIT_DS_AGAIN
;
2452 scan_ds_queue_insert(scn
, ds
->ds_object
,
2453 scn
->scn_phys
.scn_cur_max_txg
);
2458 * Add descendant datasets to work queue.
2460 if (dsl_dataset_phys(ds
)->ds_next_snap_obj
!= 0) {
2461 scan_ds_queue_insert(scn
,
2462 dsl_dataset_phys(ds
)->ds_next_snap_obj
,
2463 dsl_dataset_phys(ds
)->ds_creation_txg
);
2465 if (dsl_dataset_phys(ds
)->ds_num_children
> 1) {
2466 boolean_t usenext
= B_FALSE
;
2467 if (dsl_dataset_phys(ds
)->ds_next_clones_obj
!= 0) {
2470 * A bug in a previous version of the code could
2471 * cause upgrade_clones_cb() to not set
2472 * ds_next_snap_obj when it should, leading to a
2473 * missing entry. Therefore we can only use the
2474 * next_clones_obj when its count is correct.
2476 int err
= zap_count(dp
->dp_meta_objset
,
2477 dsl_dataset_phys(ds
)->ds_next_clones_obj
, &count
);
2479 count
== dsl_dataset_phys(ds
)->ds_num_children
- 1)
2486 for (zap_cursor_init(&zc
, dp
->dp_meta_objset
,
2487 dsl_dataset_phys(ds
)->ds_next_clones_obj
);
2488 zap_cursor_retrieve(&zc
, &za
) == 0;
2489 (void) zap_cursor_advance(&zc
)) {
2490 scan_ds_queue_insert(scn
,
2491 zfs_strtonum(za
.za_name
, NULL
),
2492 dsl_dataset_phys(ds
)->ds_creation_txg
);
2494 zap_cursor_fini(&zc
);
2496 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
2497 enqueue_clones_cb
, &ds
->ds_object
,
2503 dsl_dataset_rele(ds
, FTAG
);
2508 enqueue_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
2512 dsl_scan_t
*scn
= dp
->dp_scan
;
2514 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
2518 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
2519 dsl_dataset_t
*prev
;
2520 err
= dsl_dataset_hold_obj(dp
,
2521 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
2523 dsl_dataset_rele(ds
, FTAG
);
2528 * If this is a clone, we don't need to worry about it for now.
2530 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
) {
2531 dsl_dataset_rele(ds
, FTAG
);
2532 dsl_dataset_rele(prev
, FTAG
);
2535 dsl_dataset_rele(ds
, FTAG
);
2539 scan_ds_queue_insert(scn
, ds
->ds_object
,
2540 dsl_dataset_phys(ds
)->ds_prev_snap_txg
);
2541 dsl_dataset_rele(ds
, FTAG
);
2547 dsl_scan_ddt_entry(dsl_scan_t
*scn
, enum zio_checksum checksum
,
2548 ddt_entry_t
*dde
, dmu_tx_t
*tx
)
2550 const ddt_key_t
*ddk
= &dde
->dde_key
;
2551 ddt_phys_t
*ddp
= dde
->dde_phys
;
2553 zbookmark_phys_t zb
= { 0 };
2556 if (!dsl_scan_is_running(scn
))
2560 * This function is special because it is the only thing
2561 * that can add scan_io_t's to the vdev scan queues from
2562 * outside dsl_scan_sync(). For the most part this is ok
2563 * as long as it is called from within syncing context.
2564 * However, dsl_scan_sync() expects that no new sio's will
2565 * be added between when all the work for a scan is done
2566 * and the next txg when the scan is actually marked as
2567 * completed. This check ensures we do not issue new sio's
2568 * during this period.
2570 if (scn
->scn_done_txg
!= 0)
2573 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2574 if (ddp
->ddp_phys_birth
== 0 ||
2575 ddp
->ddp_phys_birth
> scn
->scn_phys
.scn_max_txg
)
2577 ddt_bp_create(checksum
, ddk
, ddp
, &bp
);
2579 scn
->scn_visited_this_txg
++;
2580 scan_funcs
[scn
->scn_phys
.scn_func
](scn
->scn_dp
, &bp
, &zb
);
2585 * Scrub/dedup interaction.
2587 * If there are N references to a deduped block, we don't want to scrub it
2588 * N times -- ideally, we should scrub it exactly once.
2590 * We leverage the fact that the dde's replication class (enum ddt_class)
2591 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2592 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2594 * To prevent excess scrubbing, the scrub begins by walking the DDT
2595 * to find all blocks with refcnt > 1, and scrubs each of these once.
2596 * Since there are two replication classes which contain blocks with
2597 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2598 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2600 * There would be nothing more to say if a block's refcnt couldn't change
2601 * during a scrub, but of course it can so we must account for changes
2602 * in a block's replication class.
2604 * Here's an example of what can occur:
2606 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2607 * when visited during the top-down scrub phase, it will be scrubbed twice.
2608 * This negates our scrub optimization, but is otherwise harmless.
2610 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2611 * on each visit during the top-down scrub phase, it will never be scrubbed.
2612 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2613 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2614 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2615 * while a scrub is in progress, it scrubs the block right then.
2618 dsl_scan_ddt(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
2620 ddt_bookmark_t
*ddb
= &scn
->scn_phys
.scn_ddt_bookmark
;
2625 bzero(&dde
, sizeof (ddt_entry_t
));
2627 while ((error
= ddt_walk(scn
->scn_dp
->dp_spa
, ddb
, &dde
)) == 0) {
2630 if (ddb
->ddb_class
> scn
->scn_phys
.scn_ddt_class_max
)
2632 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
2633 (longlong_t
)ddb
->ddb_class
,
2634 (longlong_t
)ddb
->ddb_type
,
2635 (longlong_t
)ddb
->ddb_checksum
,
2636 (longlong_t
)ddb
->ddb_cursor
);
2638 /* There should be no pending changes to the dedup table */
2639 ddt
= scn
->scn_dp
->dp_spa
->spa_ddt
[ddb
->ddb_checksum
];
2640 ASSERT(avl_first(&ddt
->ddt_tree
) == NULL
);
2642 dsl_scan_ddt_entry(scn
, ddb
->ddb_checksum
, &dde
, tx
);
2645 if (dsl_scan_check_suspend(scn
, NULL
))
2649 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; "
2650 "suspending=%u", (longlong_t
)n
,
2651 (int)scn
->scn_phys
.scn_ddt_class_max
, (int)scn
->scn_suspending
);
2653 ASSERT(error
== 0 || error
== ENOENT
);
2654 ASSERT(error
!= ENOENT
||
2655 ddb
->ddb_class
> scn
->scn_phys
.scn_ddt_class_max
);
2659 dsl_scan_ds_maxtxg(dsl_dataset_t
*ds
)
2661 uint64_t smt
= ds
->ds_dir
->dd_pool
->dp_scan
->scn_phys
.scn_max_txg
;
2662 if (ds
->ds_is_snapshot
)
2663 return (MIN(smt
, dsl_dataset_phys(ds
)->ds_creation_txg
));
2668 dsl_scan_visit(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
2671 dsl_pool_t
*dp
= scn
->scn_dp
;
2673 if (scn
->scn_phys
.scn_ddt_bookmark
.ddb_class
<=
2674 scn
->scn_phys
.scn_ddt_class_max
) {
2675 scn
->scn_phys
.scn_cur_min_txg
= scn
->scn_phys
.scn_min_txg
;
2676 scn
->scn_phys
.scn_cur_max_txg
= scn
->scn_phys
.scn_max_txg
;
2677 dsl_scan_ddt(scn
, tx
);
2678 if (scn
->scn_suspending
)
2682 if (scn
->scn_phys
.scn_bookmark
.zb_objset
== DMU_META_OBJSET
) {
2683 /* First do the MOS & ORIGIN */
2685 scn
->scn_phys
.scn_cur_min_txg
= scn
->scn_phys
.scn_min_txg
;
2686 scn
->scn_phys
.scn_cur_max_txg
= scn
->scn_phys
.scn_max_txg
;
2687 dsl_scan_visit_rootbp(scn
, NULL
,
2688 &dp
->dp_meta_rootbp
, tx
);
2689 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
2690 if (scn
->scn_suspending
)
2693 if (spa_version(dp
->dp_spa
) < SPA_VERSION_DSL_SCRUB
) {
2694 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
2695 enqueue_cb
, NULL
, DS_FIND_CHILDREN
));
2697 dsl_scan_visitds(scn
,
2698 dp
->dp_origin_snap
->ds_object
, tx
);
2700 ASSERT(!scn
->scn_suspending
);
2701 } else if (scn
->scn_phys
.scn_bookmark
.zb_objset
!=
2702 ZB_DESTROYED_OBJSET
) {
2703 uint64_t dsobj
= scn
->scn_phys
.scn_bookmark
.zb_objset
;
2705 * If we were suspended, continue from here. Note if the
2706 * ds we were suspended on was deleted, the zb_objset may
2707 * be -1, so we will skip this and find a new objset
2710 dsl_scan_visitds(scn
, dsobj
, tx
);
2711 if (scn
->scn_suspending
)
2716 * In case we suspended right at the end of the ds, zero the
2717 * bookmark so we don't think that we're still trying to resume.
2719 bzero(&scn
->scn_phys
.scn_bookmark
, sizeof (zbookmark_phys_t
));
2722 * Keep pulling things out of the dataset avl queue. Updates to the
2723 * persistent zap-object-as-queue happen only at checkpoints.
2725 while ((sds
= avl_first(&scn
->scn_queue
)) != NULL
) {
2727 uint64_t dsobj
= sds
->sds_dsobj
;
2728 uint64_t txg
= sds
->sds_txg
;
2730 /* dequeue and free the ds from the queue */
2731 scan_ds_queue_remove(scn
, dsobj
);
2734 /* set up min / max txg */
2735 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
2737 scn
->scn_phys
.scn_cur_min_txg
=
2738 MAX(scn
->scn_phys
.scn_min_txg
, txg
);
2740 scn
->scn_phys
.scn_cur_min_txg
=
2741 MAX(scn
->scn_phys
.scn_min_txg
,
2742 dsl_dataset_phys(ds
)->ds_prev_snap_txg
);
2744 scn
->scn_phys
.scn_cur_max_txg
= dsl_scan_ds_maxtxg(ds
);
2745 dsl_dataset_rele(ds
, FTAG
);
2747 dsl_scan_visitds(scn
, dsobj
, tx
);
2748 if (scn
->scn_suspending
)
2752 /* No more objsets to fetch, we're done */
2753 scn
->scn_phys
.scn_bookmark
.zb_objset
= ZB_DESTROYED_OBJSET
;
2754 ASSERT0(scn
->scn_suspending
);
2758 dsl_scan_count_leaves(vdev_t
*vd
)
2760 uint64_t i
, leaves
= 0;
2762 /* we only count leaves that belong to the main pool and are readable */
2763 if (vd
->vdev_islog
|| vd
->vdev_isspare
||
2764 vd
->vdev_isl2cache
|| !vdev_readable(vd
))
2767 if (vd
->vdev_ops
->vdev_op_leaf
)
2770 for (i
= 0; i
< vd
->vdev_children
; i
++) {
2771 leaves
+= dsl_scan_count_leaves(vd
->vdev_child
[i
]);
2778 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t
*q
, const blkptr_t
*bp
)
2781 uint64_t cur_size
= 0;
2783 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
2784 cur_size
+= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
2787 q
->q_total_zio_size_this_txg
+= cur_size
;
2788 q
->q_zios_this_txg
++;
2792 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t
*q
, uint64_t start
,
2795 q
->q_total_seg_size_this_txg
+= end
- start
;
2796 q
->q_segs_this_txg
++;
2800 scan_io_queue_check_suspend(dsl_scan_t
*scn
)
2802 /* See comment in dsl_scan_check_suspend() */
2803 uint64_t curr_time_ns
= gethrtime();
2804 uint64_t scan_time_ns
= curr_time_ns
- scn
->scn_sync_start_time
;
2805 uint64_t sync_time_ns
= curr_time_ns
-
2806 scn
->scn_dp
->dp_spa
->spa_sync_starttime
;
2807 int dirty_pct
= scn
->scn_dp
->dp_dirty_total
* 100 / zfs_dirty_data_max
;
2808 int mintime
= (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) ?
2809 zfs_resilver_min_time_ms
: zfs_scrub_min_time_ms
;
2811 return ((NSEC2MSEC(scan_time_ns
) > mintime
&&
2812 (dirty_pct
>= zfs_vdev_async_write_active_min_dirty_percent
||
2813 txg_sync_waiting(scn
->scn_dp
) ||
2814 NSEC2SEC(sync_time_ns
) >= zfs_txg_timeout
)) ||
2815 spa_shutting_down(scn
->scn_dp
->dp_spa
));
2819 * Given a list of scan_io_t's in io_list, this issues the I/Os out to
2820 * disk. This consumes the io_list and frees the scan_io_t's. This is
2821 * called when emptying queues, either when we're up against the memory
2822 * limit or when we have finished scanning. Returns B_TRUE if we stopped
2823 * processing the list before we finished. Any sios that were not issued
2824 * will remain in the io_list.
2827 scan_io_queue_issue(dsl_scan_io_queue_t
*queue
, list_t
*io_list
)
2829 dsl_scan_t
*scn
= queue
->q_scn
;
2831 int64_t bytes_issued
= 0;
2832 boolean_t suspended
= B_FALSE
;
2834 while ((sio
= list_head(io_list
)) != NULL
) {
2837 if (scan_io_queue_check_suspend(scn
)) {
2843 bytes_issued
+= SIO_GET_ASIZE(sio
);
2844 scan_exec_io(scn
->scn_dp
, &bp
, sio
->sio_flags
,
2845 &sio
->sio_zb
, queue
);
2846 (void) list_remove_head(io_list
);
2847 scan_io_queues_update_zio_stats(queue
, &bp
);
2851 atomic_add_64(&scn
->scn_bytes_pending
, -bytes_issued
);
2857 * This function removes sios from an IO queue which reside within a given
2858 * range_seg_t and inserts them (in offset order) into a list. Note that
2859 * we only ever return a maximum of 32 sios at once. If there are more sios
2860 * to process within this segment that did not make it onto the list we
2861 * return B_TRUE and otherwise B_FALSE.
2864 scan_io_queue_gather(dsl_scan_io_queue_t
*queue
, range_seg_t
*rs
, list_t
*list
)
2866 scan_io_t
*srch_sio
, *sio
, *next_sio
;
2868 uint_t num_sios
= 0;
2869 int64_t bytes_issued
= 0;
2872 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
2874 srch_sio
= sio_alloc(1);
2875 srch_sio
->sio_nr_dvas
= 1;
2876 SIO_SET_OFFSET(srch_sio
, rs_get_start(rs
, queue
->q_exts_by_addr
));
2879 * The exact start of the extent might not contain any matching zios,
2880 * so if that's the case, examine the next one in the tree.
2882 sio
= avl_find(&queue
->q_sios_by_addr
, srch_sio
, &idx
);
2886 sio
= avl_nearest(&queue
->q_sios_by_addr
, idx
, AVL_AFTER
);
2888 while (sio
!= NULL
&& SIO_GET_OFFSET(sio
) < rs_get_end(rs
,
2889 queue
->q_exts_by_addr
) && num_sios
<= 32) {
2890 ASSERT3U(SIO_GET_OFFSET(sio
), >=, rs_get_start(rs
,
2891 queue
->q_exts_by_addr
));
2892 ASSERT3U(SIO_GET_END_OFFSET(sio
), <=, rs_get_end(rs
,
2893 queue
->q_exts_by_addr
));
2895 next_sio
= AVL_NEXT(&queue
->q_sios_by_addr
, sio
);
2896 avl_remove(&queue
->q_sios_by_addr
, sio
);
2897 queue
->q_sio_memused
-= SIO_GET_MUSED(sio
);
2899 bytes_issued
+= SIO_GET_ASIZE(sio
);
2901 list_insert_tail(list
, sio
);
2906 * We limit the number of sios we process at once to 32 to avoid
2907 * biting off more than we can chew. If we didn't take everything
2908 * in the segment we update it to reflect the work we were able to
2909 * complete. Otherwise, we remove it from the range tree entirely.
2911 if (sio
!= NULL
&& SIO_GET_OFFSET(sio
) < rs_get_end(rs
,
2912 queue
->q_exts_by_addr
)) {
2913 range_tree_adjust_fill(queue
->q_exts_by_addr
, rs
,
2915 range_tree_resize_segment(queue
->q_exts_by_addr
, rs
,
2916 SIO_GET_OFFSET(sio
), rs_get_end(rs
,
2917 queue
->q_exts_by_addr
) - SIO_GET_OFFSET(sio
));
2921 uint64_t rstart
= rs_get_start(rs
, queue
->q_exts_by_addr
);
2922 uint64_t rend
= rs_get_end(rs
, queue
->q_exts_by_addr
);
2923 range_tree_remove(queue
->q_exts_by_addr
, rstart
, rend
- rstart
);
2929 * This is called from the queue emptying thread and selects the next
2930 * extent from which we are to issue I/Os. The behavior of this function
2931 * depends on the state of the scan, the current memory consumption and
2932 * whether or not we are performing a scan shutdown.
2933 * 1) We select extents in an elevator algorithm (LBA-order) if the scan
2934 * needs to perform a checkpoint
2935 * 2) We select the largest available extent if we are up against the
2937 * 3) Otherwise we don't select any extents.
2939 static range_seg_t
*
2940 scan_io_queue_fetch_ext(dsl_scan_io_queue_t
*queue
)
2942 dsl_scan_t
*scn
= queue
->q_scn
;
2943 range_tree_t
*rt
= queue
->q_exts_by_addr
;
2945 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
2946 ASSERT(scn
->scn_is_sorted
);
2948 /* handle tunable overrides */
2949 if (scn
->scn_checkpointing
|| scn
->scn_clearing
) {
2950 if (zfs_scan_issue_strategy
== 1) {
2951 return (range_tree_first(rt
));
2952 } else if (zfs_scan_issue_strategy
== 2) {
2954 * We need to get the original entry in the by_addr
2955 * tree so we can modify it.
2957 range_seg_t
*size_rs
=
2958 zfs_btree_first(&queue
->q_exts_by_size
, NULL
);
2959 if (size_rs
== NULL
)
2961 uint64_t start
= rs_get_start(size_rs
, rt
);
2962 uint64_t size
= rs_get_end(size_rs
, rt
) - start
;
2963 range_seg_t
*addr_rs
= range_tree_find(rt
, start
,
2965 ASSERT3P(addr_rs
, !=, NULL
);
2966 ASSERT3U(rs_get_start(size_rs
, rt
), ==,
2967 rs_get_start(addr_rs
, rt
));
2968 ASSERT3U(rs_get_end(size_rs
, rt
), ==,
2969 rs_get_end(addr_rs
, rt
));
2975 * During normal clearing, we want to issue our largest segments
2976 * first, keeping IO as sequential as possible, and leaving the
2977 * smaller extents for later with the hope that they might eventually
2978 * grow to larger sequential segments. However, when the scan is
2979 * checkpointing, no new extents will be added to the sorting queue,
2980 * so the way we are sorted now is as good as it will ever get.
2981 * In this case, we instead switch to issuing extents in LBA order.
2983 if (scn
->scn_checkpointing
) {
2984 return (range_tree_first(rt
));
2985 } else if (scn
->scn_clearing
) {
2987 * We need to get the original entry in the by_addr
2988 * tree so we can modify it.
2990 range_seg_t
*size_rs
= zfs_btree_first(&queue
->q_exts_by_size
,
2992 if (size_rs
== NULL
)
2994 uint64_t start
= rs_get_start(size_rs
, rt
);
2995 uint64_t size
= rs_get_end(size_rs
, rt
) - start
;
2996 range_seg_t
*addr_rs
= range_tree_find(rt
, start
, size
);
2997 ASSERT3P(addr_rs
, !=, NULL
);
2998 ASSERT3U(rs_get_start(size_rs
, rt
), ==, rs_get_start(addr_rs
,
3000 ASSERT3U(rs_get_end(size_rs
, rt
), ==, rs_get_end(addr_rs
, rt
));
3008 scan_io_queues_run_one(void *arg
)
3010 dsl_scan_io_queue_t
*queue
= arg
;
3011 kmutex_t
*q_lock
= &queue
->q_vd
->vdev_scan_io_queue_lock
;
3012 boolean_t suspended
= B_FALSE
;
3013 range_seg_t
*rs
= NULL
;
3014 scan_io_t
*sio
= NULL
;
3016 uint64_t bytes_per_leaf
= zfs_scan_vdev_limit
;
3017 uint64_t nr_leaves
= dsl_scan_count_leaves(queue
->q_vd
);
3019 ASSERT(queue
->q_scn
->scn_is_sorted
);
3021 list_create(&sio_list
, sizeof (scan_io_t
),
3022 offsetof(scan_io_t
, sio_nodes
.sio_list_node
));
3023 mutex_enter(q_lock
);
3025 /* calculate maximum in-flight bytes for this txg (min 1MB) */
3026 queue
->q_maxinflight_bytes
=
3027 MAX(nr_leaves
* bytes_per_leaf
, 1ULL << 20);
3029 /* reset per-queue scan statistics for this txg */
3030 queue
->q_total_seg_size_this_txg
= 0;
3031 queue
->q_segs_this_txg
= 0;
3032 queue
->q_total_zio_size_this_txg
= 0;
3033 queue
->q_zios_this_txg
= 0;
3035 /* loop until we run out of time or sios */
3036 while ((rs
= scan_io_queue_fetch_ext(queue
)) != NULL
) {
3037 uint64_t seg_start
= 0, seg_end
= 0;
3038 boolean_t more_left
= B_TRUE
;
3040 ASSERT(list_is_empty(&sio_list
));
3042 /* loop while we still have sios left to process in this rs */
3044 scan_io_t
*first_sio
, *last_sio
;
3047 * We have selected which extent needs to be
3048 * processed next. Gather up the corresponding sios.
3050 more_left
= scan_io_queue_gather(queue
, rs
, &sio_list
);
3051 ASSERT(!list_is_empty(&sio_list
));
3052 first_sio
= list_head(&sio_list
);
3053 last_sio
= list_tail(&sio_list
);
3055 seg_end
= SIO_GET_END_OFFSET(last_sio
);
3057 seg_start
= SIO_GET_OFFSET(first_sio
);
3060 * Issuing sios can take a long time so drop the
3061 * queue lock. The sio queue won't be updated by
3062 * other threads since we're in syncing context so
3063 * we can be sure that our trees will remain exactly
3067 suspended
= scan_io_queue_issue(queue
, &sio_list
);
3068 mutex_enter(q_lock
);
3074 /* update statistics for debugging purposes */
3075 scan_io_queues_update_seg_stats(queue
, seg_start
, seg_end
);
3082 * If we were suspended in the middle of processing,
3083 * requeue any unfinished sios and exit.
3085 while ((sio
= list_head(&sio_list
)) != NULL
) {
3086 list_remove(&sio_list
, sio
);
3087 scan_io_queue_insert_impl(queue
, sio
);
3091 list_destroy(&sio_list
);
3095 * Performs an emptying run on all scan queues in the pool. This just
3096 * punches out one thread per top-level vdev, each of which processes
3097 * only that vdev's scan queue. We can parallelize the I/O here because
3098 * we know that each queue's I/Os only affect its own top-level vdev.
3100 * This function waits for the queue runs to complete, and must be
3101 * called from dsl_scan_sync (or in general, syncing context).
3104 scan_io_queues_run(dsl_scan_t
*scn
)
3106 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3108 ASSERT(scn
->scn_is_sorted
);
3109 ASSERT(spa_config_held(spa
, SCL_CONFIG
, RW_READER
));
3111 if (scn
->scn_bytes_pending
== 0)
3114 if (scn
->scn_taskq
== NULL
) {
3115 int nthreads
= spa
->spa_root_vdev
->vdev_children
;
3118 * We need to make this taskq *always* execute as many
3119 * threads in parallel as we have top-level vdevs and no
3120 * less, otherwise strange serialization of the calls to
3121 * scan_io_queues_run_one can occur during spa_sync runs
3122 * and that significantly impacts performance.
3124 scn
->scn_taskq
= taskq_create("dsl_scan_iss", nthreads
,
3125 minclsyspri
, nthreads
, nthreads
, TASKQ_PREPOPULATE
);
3128 for (uint64_t i
= 0; i
< spa
->spa_root_vdev
->vdev_children
; i
++) {
3129 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[i
];
3131 mutex_enter(&vd
->vdev_scan_io_queue_lock
);
3132 if (vd
->vdev_scan_io_queue
!= NULL
) {
3133 VERIFY(taskq_dispatch(scn
->scn_taskq
,
3134 scan_io_queues_run_one
, vd
->vdev_scan_io_queue
,
3135 TQ_SLEEP
) != TASKQID_INVALID
);
3137 mutex_exit(&vd
->vdev_scan_io_queue_lock
);
3141 * Wait for the queues to finish issuing their IOs for this run
3142 * before we return. There may still be IOs in flight at this
3145 taskq_wait(scn
->scn_taskq
);
3149 dsl_scan_async_block_should_pause(dsl_scan_t
*scn
)
3151 uint64_t elapsed_nanosecs
;
3156 if (zfs_async_block_max_blocks
!= 0 &&
3157 scn
->scn_visited_this_txg
>= zfs_async_block_max_blocks
) {
3161 if (zfs_max_async_dedup_frees
!= 0 &&
3162 scn
->scn_dedup_frees_this_txg
>= zfs_max_async_dedup_frees
) {
3166 elapsed_nanosecs
= gethrtime() - scn
->scn_sync_start_time
;
3167 return (elapsed_nanosecs
/ NANOSEC
> zfs_txg_timeout
||
3168 (NSEC2MSEC(elapsed_nanosecs
) > scn
->scn_async_block_min_time_ms
&&
3169 txg_sync_waiting(scn
->scn_dp
)) ||
3170 spa_shutting_down(scn
->scn_dp
->dp_spa
));
3174 dsl_scan_free_block_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
3176 dsl_scan_t
*scn
= arg
;
3178 if (!scn
->scn_is_bptree
||
3179 (BP_GET_LEVEL(bp
) == 0 && BP_GET_TYPE(bp
) != DMU_OT_OBJSET
)) {
3180 if (dsl_scan_async_block_should_pause(scn
))
3181 return (SET_ERROR(ERESTART
));
3184 zio_nowait(zio_free_sync(scn
->scn_zio_root
, scn
->scn_dp
->dp_spa
,
3185 dmu_tx_get_txg(tx
), bp
, 0));
3186 dsl_dir_diduse_space(tx
->tx_pool
->dp_free_dir
, DD_USED_HEAD
,
3187 -bp_get_dsize_sync(scn
->scn_dp
->dp_spa
, bp
),
3188 -BP_GET_PSIZE(bp
), -BP_GET_UCSIZE(bp
), tx
);
3189 scn
->scn_visited_this_txg
++;
3190 if (BP_GET_DEDUP(bp
))
3191 scn
->scn_dedup_frees_this_txg
++;
3196 dsl_scan_update_stats(dsl_scan_t
*scn
)
3198 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3200 uint64_t seg_size_total
= 0, zio_size_total
= 0;
3201 uint64_t seg_count_total
= 0, zio_count_total
= 0;
3203 for (i
= 0; i
< spa
->spa_root_vdev
->vdev_children
; i
++) {
3204 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[i
];
3205 dsl_scan_io_queue_t
*queue
= vd
->vdev_scan_io_queue
;
3210 seg_size_total
+= queue
->q_total_seg_size_this_txg
;
3211 zio_size_total
+= queue
->q_total_zio_size_this_txg
;
3212 seg_count_total
+= queue
->q_segs_this_txg
;
3213 zio_count_total
+= queue
->q_zios_this_txg
;
3216 if (seg_count_total
== 0 || zio_count_total
== 0) {
3217 scn
->scn_avg_seg_size_this_txg
= 0;
3218 scn
->scn_avg_zio_size_this_txg
= 0;
3219 scn
->scn_segs_this_txg
= 0;
3220 scn
->scn_zios_this_txg
= 0;
3224 scn
->scn_avg_seg_size_this_txg
= seg_size_total
/ seg_count_total
;
3225 scn
->scn_avg_zio_size_this_txg
= zio_size_total
/ zio_count_total
;
3226 scn
->scn_segs_this_txg
= seg_count_total
;
3227 scn
->scn_zios_this_txg
= zio_count_total
;
3231 bpobj_dsl_scan_free_block_cb(void *arg
, const blkptr_t
*bp
, boolean_t bp_freed
,
3235 return (dsl_scan_free_block_cb(arg
, bp
, tx
));
3239 dsl_scan_obsolete_block_cb(void *arg
, const blkptr_t
*bp
, boolean_t bp_freed
,
3243 dsl_scan_t
*scn
= arg
;
3244 const dva_t
*dva
= &bp
->blk_dva
[0];
3246 if (dsl_scan_async_block_should_pause(scn
))
3247 return (SET_ERROR(ERESTART
));
3249 spa_vdev_indirect_mark_obsolete(scn
->scn_dp
->dp_spa
,
3250 DVA_GET_VDEV(dva
), DVA_GET_OFFSET(dva
),
3251 DVA_GET_ASIZE(dva
), tx
);
3252 scn
->scn_visited_this_txg
++;
3257 dsl_scan_active(dsl_scan_t
*scn
)
3259 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3260 uint64_t used
= 0, comp
, uncomp
;
3261 boolean_t clones_left
;
3263 if (spa
->spa_load_state
!= SPA_LOAD_NONE
)
3265 if (spa_shutting_down(spa
))
3267 if ((dsl_scan_is_running(scn
) && !dsl_scan_is_paused_scrub(scn
)) ||
3268 (scn
->scn_async_destroying
&& !scn
->scn_async_stalled
))
3271 if (spa_version(scn
->scn_dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
3272 (void) bpobj_space(&scn
->scn_dp
->dp_free_bpobj
,
3273 &used
, &comp
, &uncomp
);
3275 clones_left
= spa_livelist_delete_check(spa
);
3276 return ((used
!= 0) || (clones_left
));
3280 dsl_scan_check_deferred(vdev_t
*vd
)
3282 boolean_t need_resilver
= B_FALSE
;
3284 for (int c
= 0; c
< vd
->vdev_children
; c
++) {
3286 dsl_scan_check_deferred(vd
->vdev_child
[c
]);
3289 if (!vdev_is_concrete(vd
) || vd
->vdev_aux
||
3290 !vd
->vdev_ops
->vdev_op_leaf
)
3291 return (need_resilver
);
3293 if (!vd
->vdev_resilver_deferred
)
3294 need_resilver
= B_TRUE
;
3296 return (need_resilver
);
3300 dsl_scan_need_resilver(spa_t
*spa
, const dva_t
*dva
, size_t psize
,
3301 uint64_t phys_birth
)
3305 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(dva
));
3307 if (vd
->vdev_ops
== &vdev_indirect_ops
) {
3309 * The indirect vdev can point to multiple
3310 * vdevs. For simplicity, always create
3311 * the resilver zio_t. zio_vdev_io_start()
3312 * will bypass the child resilver i/o's if
3313 * they are on vdevs that don't have DTL's.
3318 if (DVA_GET_GANG(dva
)) {
3320 * Gang members may be spread across multiple
3321 * vdevs, so the best estimate we have is the
3322 * scrub range, which has already been checked.
3323 * XXX -- it would be better to change our
3324 * allocation policy to ensure that all
3325 * gang members reside on the same vdev.
3331 * Check if the top-level vdev must resilver this offset.
3332 * When the offset does not intersect with a dirty leaf DTL
3333 * then it may be possible to skip the resilver IO. The psize
3334 * is provided instead of asize to simplify the check for RAIDZ.
3336 if (!vdev_dtl_need_resilver(vd
, dva
, psize
, phys_birth
))
3340 * Check that this top-level vdev has a device under it which
3341 * is resilvering and is not deferred.
3343 if (!dsl_scan_check_deferred(vd
))
3350 dsl_process_async_destroys(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
3352 dsl_scan_t
*scn
= dp
->dp_scan
;
3353 spa_t
*spa
= dp
->dp_spa
;
3356 if (spa_suspend_async_destroy(spa
))
3359 if (zfs_free_bpobj_enabled
&&
3360 spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
3361 scn
->scn_is_bptree
= B_FALSE
;
3362 scn
->scn_async_block_min_time_ms
= zfs_free_min_time_ms
;
3363 scn
->scn_zio_root
= zio_root(spa
, NULL
,
3364 NULL
, ZIO_FLAG_MUSTSUCCEED
);
3365 err
= bpobj_iterate(&dp
->dp_free_bpobj
,
3366 bpobj_dsl_scan_free_block_cb
, scn
, tx
);
3367 VERIFY0(zio_wait(scn
->scn_zio_root
));
3368 scn
->scn_zio_root
= NULL
;
3370 if (err
!= 0 && err
!= ERESTART
)
3371 zfs_panic_recover("error %u from bpobj_iterate()", err
);
3374 if (err
== 0 && spa_feature_is_active(spa
, SPA_FEATURE_ASYNC_DESTROY
)) {
3375 ASSERT(scn
->scn_async_destroying
);
3376 scn
->scn_is_bptree
= B_TRUE
;
3377 scn
->scn_zio_root
= zio_root(spa
, NULL
,
3378 NULL
, ZIO_FLAG_MUSTSUCCEED
);
3379 err
= bptree_iterate(dp
->dp_meta_objset
,
3380 dp
->dp_bptree_obj
, B_TRUE
, dsl_scan_free_block_cb
, scn
, tx
);
3381 VERIFY0(zio_wait(scn
->scn_zio_root
));
3382 scn
->scn_zio_root
= NULL
;
3384 if (err
== EIO
|| err
== ECKSUM
) {
3386 } else if (err
!= 0 && err
!= ERESTART
) {
3387 zfs_panic_recover("error %u from "
3388 "traverse_dataset_destroyed()", err
);
3391 if (bptree_is_empty(dp
->dp_meta_objset
, dp
->dp_bptree_obj
)) {
3392 /* finished; deactivate async destroy feature */
3393 spa_feature_decr(spa
, SPA_FEATURE_ASYNC_DESTROY
, tx
);
3394 ASSERT(!spa_feature_is_active(spa
,
3395 SPA_FEATURE_ASYNC_DESTROY
));
3396 VERIFY0(zap_remove(dp
->dp_meta_objset
,
3397 DMU_POOL_DIRECTORY_OBJECT
,
3398 DMU_POOL_BPTREE_OBJ
, tx
));
3399 VERIFY0(bptree_free(dp
->dp_meta_objset
,
3400 dp
->dp_bptree_obj
, tx
));
3401 dp
->dp_bptree_obj
= 0;
3402 scn
->scn_async_destroying
= B_FALSE
;
3403 scn
->scn_async_stalled
= B_FALSE
;
3406 * If we didn't make progress, mark the async
3407 * destroy as stalled, so that we will not initiate
3408 * a spa_sync() on its behalf. Note that we only
3409 * check this if we are not finished, because if the
3410 * bptree had no blocks for us to visit, we can
3411 * finish without "making progress".
3413 scn
->scn_async_stalled
=
3414 (scn
->scn_visited_this_txg
== 0);
3417 if (scn
->scn_visited_this_txg
) {
3418 zfs_dbgmsg("freed %llu blocks in %llums from "
3419 "free_bpobj/bptree txg %llu; err=%u",
3420 (longlong_t
)scn
->scn_visited_this_txg
,
3422 NSEC2MSEC(gethrtime() - scn
->scn_sync_start_time
),
3423 (longlong_t
)tx
->tx_txg
, err
);
3424 scn
->scn_visited_this_txg
= 0;
3425 scn
->scn_dedup_frees_this_txg
= 0;
3428 * Write out changes to the DDT that may be required as a
3429 * result of the blocks freed. This ensures that the DDT
3430 * is clean when a scrub/resilver runs.
3432 ddt_sync(spa
, tx
->tx_txg
);
3436 if (dp
->dp_free_dir
!= NULL
&& !scn
->scn_async_destroying
&&
3437 zfs_free_leak_on_eio
&&
3438 (dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
!= 0 ||
3439 dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
!= 0 ||
3440 dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
!= 0)) {
3442 * We have finished background destroying, but there is still
3443 * some space left in the dp_free_dir. Transfer this leaked
3444 * space to the dp_leak_dir.
3446 if (dp
->dp_leak_dir
== NULL
) {
3447 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
3448 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
3450 VERIFY0(dsl_pool_open_special_dir(dp
,
3451 LEAK_DIR_NAME
, &dp
->dp_leak_dir
));
3452 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
3454 dsl_dir_diduse_space(dp
->dp_leak_dir
, DD_USED_HEAD
,
3455 dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
,
3456 dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
,
3457 dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
, tx
);
3458 dsl_dir_diduse_space(dp
->dp_free_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
);
3464 if (dp
->dp_free_dir
!= NULL
&& !scn
->scn_async_destroying
&&
3465 !spa_livelist_delete_check(spa
)) {
3466 /* finished; verify that space accounting went to zero */
3467 ASSERT0(dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
);
3468 ASSERT0(dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
);
3469 ASSERT0(dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
);
3472 spa_notify_waiters(spa
);
3474 EQUIV(bpobj_is_open(&dp
->dp_obsolete_bpobj
),
3475 0 == zap_contains(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
3476 DMU_POOL_OBSOLETE_BPOBJ
));
3477 if (err
== 0 && bpobj_is_open(&dp
->dp_obsolete_bpobj
)) {
3478 ASSERT(spa_feature_is_active(dp
->dp_spa
,
3479 SPA_FEATURE_OBSOLETE_COUNTS
));
3481 scn
->scn_is_bptree
= B_FALSE
;
3482 scn
->scn_async_block_min_time_ms
= zfs_obsolete_min_time_ms
;
3483 err
= bpobj_iterate(&dp
->dp_obsolete_bpobj
,
3484 dsl_scan_obsolete_block_cb
, scn
, tx
);
3485 if (err
!= 0 && err
!= ERESTART
)
3486 zfs_panic_recover("error %u from bpobj_iterate()", err
);
3488 if (bpobj_is_empty(&dp
->dp_obsolete_bpobj
))
3489 dsl_pool_destroy_obsolete_bpobj(dp
, tx
);
3495 * This is the primary entry point for scans that is called from syncing
3496 * context. Scans must happen entirely during syncing context so that we
3497 * can guarantee that blocks we are currently scanning will not change out
3498 * from under us. While a scan is active, this function controls how quickly
3499 * transaction groups proceed, instead of the normal handling provided by
3500 * txg_sync_thread().
3503 dsl_scan_sync(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
3506 dsl_scan_t
*scn
= dp
->dp_scan
;
3507 spa_t
*spa
= dp
->dp_spa
;
3508 state_sync_type_t sync_type
= SYNC_OPTIONAL
;
3510 if (spa
->spa_resilver_deferred
&&
3511 !spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_RESILVER_DEFER
))
3512 spa_feature_incr(spa
, SPA_FEATURE_RESILVER_DEFER
, tx
);
3515 * Check for scn_restart_txg before checking spa_load_state, so
3516 * that we can restart an old-style scan while the pool is being
3517 * imported (see dsl_scan_init). We also restart scans if there
3518 * is a deferred resilver and the user has manually disabled
3519 * deferred resilvers via the tunable.
3521 if (dsl_scan_restarting(scn
, tx
) ||
3522 (spa
->spa_resilver_deferred
&& zfs_resilver_disable_defer
)) {
3523 pool_scan_func_t func
= POOL_SCAN_SCRUB
;
3524 dsl_scan_done(scn
, B_FALSE
, tx
);
3525 if (vdev_resilver_needed(spa
->spa_root_vdev
, NULL
, NULL
))
3526 func
= POOL_SCAN_RESILVER
;
3527 zfs_dbgmsg("restarting scan func=%u txg=%llu",
3528 func
, (longlong_t
)tx
->tx_txg
);
3529 dsl_scan_setup_sync(&func
, tx
);
3533 * Only process scans in sync pass 1.
3535 if (spa_sync_pass(spa
) > 1)
3539 * If the spa is shutting down, then stop scanning. This will
3540 * ensure that the scan does not dirty any new data during the
3543 if (spa_shutting_down(spa
))
3547 * If the scan is inactive due to a stalled async destroy, try again.
3549 if (!scn
->scn_async_stalled
&& !dsl_scan_active(scn
))
3552 /* reset scan statistics */
3553 scn
->scn_visited_this_txg
= 0;
3554 scn
->scn_dedup_frees_this_txg
= 0;
3555 scn
->scn_holes_this_txg
= 0;
3556 scn
->scn_lt_min_this_txg
= 0;
3557 scn
->scn_gt_max_this_txg
= 0;
3558 scn
->scn_ddt_contained_this_txg
= 0;
3559 scn
->scn_objsets_visited_this_txg
= 0;
3560 scn
->scn_avg_seg_size_this_txg
= 0;
3561 scn
->scn_segs_this_txg
= 0;
3562 scn
->scn_avg_zio_size_this_txg
= 0;
3563 scn
->scn_zios_this_txg
= 0;
3564 scn
->scn_suspending
= B_FALSE
;
3565 scn
->scn_sync_start_time
= gethrtime();
3566 spa
->spa_scrub_active
= B_TRUE
;
3569 * First process the async destroys. If we suspend, don't do
3570 * any scrubbing or resilvering. This ensures that there are no
3571 * async destroys while we are scanning, so the scan code doesn't
3572 * have to worry about traversing it. It is also faster to free the
3573 * blocks than to scrub them.
3575 err
= dsl_process_async_destroys(dp
, tx
);
3579 if (!dsl_scan_is_running(scn
) || dsl_scan_is_paused_scrub(scn
))
3583 * Wait a few txgs after importing to begin scanning so that
3584 * we can get the pool imported quickly.
3586 if (spa
->spa_syncing_txg
< spa
->spa_first_txg
+ SCAN_IMPORT_WAIT_TXGS
)
3590 * zfs_scan_suspend_progress can be set to disable scan progress.
3591 * We don't want to spin the txg_sync thread, so we add a delay
3592 * here to simulate the time spent doing a scan. This is mostly
3593 * useful for testing and debugging.
3595 if (zfs_scan_suspend_progress
) {
3596 uint64_t scan_time_ns
= gethrtime() - scn
->scn_sync_start_time
;
3597 int mintime
= (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) ?
3598 zfs_resilver_min_time_ms
: zfs_scrub_min_time_ms
;
3600 while (zfs_scan_suspend_progress
&&
3601 !txg_sync_waiting(scn
->scn_dp
) &&
3602 !spa_shutting_down(scn
->scn_dp
->dp_spa
) &&
3603 NSEC2MSEC(scan_time_ns
) < mintime
) {
3605 scan_time_ns
= gethrtime() - scn
->scn_sync_start_time
;
3611 * It is possible to switch from unsorted to sorted at any time,
3612 * but afterwards the scan will remain sorted unless reloaded from
3613 * a checkpoint after a reboot.
3615 if (!zfs_scan_legacy
) {
3616 scn
->scn_is_sorted
= B_TRUE
;
3617 if (scn
->scn_last_checkpoint
== 0)
3618 scn
->scn_last_checkpoint
= ddi_get_lbolt();
3622 * For sorted scans, determine what kind of work we will be doing
3623 * this txg based on our memory limitations and whether or not we
3624 * need to perform a checkpoint.
3626 if (scn
->scn_is_sorted
) {
3628 * If we are over our checkpoint interval, set scn_clearing
3629 * so that we can begin checkpointing immediately. The
3630 * checkpoint allows us to save a consistent bookmark
3631 * representing how much data we have scrubbed so far.
3632 * Otherwise, use the memory limit to determine if we should
3633 * scan for metadata or start issue scrub IOs. We accumulate
3634 * metadata until we hit our hard memory limit at which point
3635 * we issue scrub IOs until we are at our soft memory limit.
3637 if (scn
->scn_checkpointing
||
3638 ddi_get_lbolt() - scn
->scn_last_checkpoint
>
3639 SEC_TO_TICK(zfs_scan_checkpoint_intval
)) {
3640 if (!scn
->scn_checkpointing
)
3641 zfs_dbgmsg("begin scan checkpoint");
3643 scn
->scn_checkpointing
= B_TRUE
;
3644 scn
->scn_clearing
= B_TRUE
;
3646 boolean_t should_clear
= dsl_scan_should_clear(scn
);
3647 if (should_clear
&& !scn
->scn_clearing
) {
3648 zfs_dbgmsg("begin scan clearing");
3649 scn
->scn_clearing
= B_TRUE
;
3650 } else if (!should_clear
&& scn
->scn_clearing
) {
3651 zfs_dbgmsg("finish scan clearing");
3652 scn
->scn_clearing
= B_FALSE
;
3656 ASSERT0(scn
->scn_checkpointing
);
3657 ASSERT0(scn
->scn_clearing
);
3660 if (!scn
->scn_clearing
&& scn
->scn_done_txg
== 0) {
3661 /* Need to scan metadata for more blocks to scrub */
3662 dsl_scan_phys_t
*scnp
= &scn
->scn_phys
;
3663 taskqid_t prefetch_tqid
;
3664 uint64_t bytes_per_leaf
= zfs_scan_vdev_limit
;
3665 uint64_t nr_leaves
= dsl_scan_count_leaves(spa
->spa_root_vdev
);
3668 * Recalculate the max number of in-flight bytes for pool-wide
3669 * scanning operations (minimum 1MB). Limits for the issuing
3670 * phase are done per top-level vdev and are handled separately.
3672 scn
->scn_maxinflight_bytes
=
3673 MAX(nr_leaves
* bytes_per_leaf
, 1ULL << 20);
3675 if (scnp
->scn_ddt_bookmark
.ddb_class
<=
3676 scnp
->scn_ddt_class_max
) {
3677 ASSERT(ZB_IS_ZERO(&scnp
->scn_bookmark
));
3678 zfs_dbgmsg("doing scan sync txg %llu; "
3679 "ddt bm=%llu/%llu/%llu/%llx",
3680 (longlong_t
)tx
->tx_txg
,
3681 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_class
,
3682 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_type
,
3683 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_checksum
,
3684 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_cursor
);
3686 zfs_dbgmsg("doing scan sync txg %llu; "
3687 "bm=%llu/%llu/%llu/%llu",
3688 (longlong_t
)tx
->tx_txg
,
3689 (longlong_t
)scnp
->scn_bookmark
.zb_objset
,
3690 (longlong_t
)scnp
->scn_bookmark
.zb_object
,
3691 (longlong_t
)scnp
->scn_bookmark
.zb_level
,
3692 (longlong_t
)scnp
->scn_bookmark
.zb_blkid
);
3695 scn
->scn_zio_root
= zio_root(dp
->dp_spa
, NULL
,
3696 NULL
, ZIO_FLAG_CANFAIL
);
3698 scn
->scn_prefetch_stop
= B_FALSE
;
3699 prefetch_tqid
= taskq_dispatch(dp
->dp_sync_taskq
,
3700 dsl_scan_prefetch_thread
, scn
, TQ_SLEEP
);
3701 ASSERT(prefetch_tqid
!= TASKQID_INVALID
);
3703 dsl_pool_config_enter(dp
, FTAG
);
3704 dsl_scan_visit(scn
, tx
);
3705 dsl_pool_config_exit(dp
, FTAG
);
3707 mutex_enter(&dp
->dp_spa
->spa_scrub_lock
);
3708 scn
->scn_prefetch_stop
= B_TRUE
;
3709 cv_broadcast(&spa
->spa_scrub_io_cv
);
3710 mutex_exit(&dp
->dp_spa
->spa_scrub_lock
);
3712 taskq_wait_id(dp
->dp_sync_taskq
, prefetch_tqid
);
3713 (void) zio_wait(scn
->scn_zio_root
);
3714 scn
->scn_zio_root
= NULL
;
3716 zfs_dbgmsg("scan visited %llu blocks in %llums "
3717 "(%llu os's, %llu holes, %llu < mintxg, "
3718 "%llu in ddt, %llu > maxtxg)",
3719 (longlong_t
)scn
->scn_visited_this_txg
,
3720 (longlong_t
)NSEC2MSEC(gethrtime() -
3721 scn
->scn_sync_start_time
),
3722 (longlong_t
)scn
->scn_objsets_visited_this_txg
,
3723 (longlong_t
)scn
->scn_holes_this_txg
,
3724 (longlong_t
)scn
->scn_lt_min_this_txg
,
3725 (longlong_t
)scn
->scn_ddt_contained_this_txg
,
3726 (longlong_t
)scn
->scn_gt_max_this_txg
);
3728 if (!scn
->scn_suspending
) {
3729 ASSERT0(avl_numnodes(&scn
->scn_queue
));
3730 scn
->scn_done_txg
= tx
->tx_txg
+ 1;
3731 if (scn
->scn_is_sorted
) {
3732 scn
->scn_checkpointing
= B_TRUE
;
3733 scn
->scn_clearing
= B_TRUE
;
3735 zfs_dbgmsg("scan complete txg %llu",
3736 (longlong_t
)tx
->tx_txg
);
3738 } else if (scn
->scn_is_sorted
&& scn
->scn_bytes_pending
!= 0) {
3739 ASSERT(scn
->scn_clearing
);
3741 /* need to issue scrubbing IOs from per-vdev queues */
3742 scn
->scn_zio_root
= zio_root(dp
->dp_spa
, NULL
,
3743 NULL
, ZIO_FLAG_CANFAIL
);
3744 scan_io_queues_run(scn
);
3745 (void) zio_wait(scn
->scn_zio_root
);
3746 scn
->scn_zio_root
= NULL
;
3748 /* calculate and dprintf the current memory usage */
3749 (void) dsl_scan_should_clear(scn
);
3750 dsl_scan_update_stats(scn
);
3752 zfs_dbgmsg("scan issued %llu blocks (%llu segs) in %llums "
3753 "(avg_block_size = %llu, avg_seg_size = %llu)",
3754 (longlong_t
)scn
->scn_zios_this_txg
,
3755 (longlong_t
)scn
->scn_segs_this_txg
,
3756 (longlong_t
)NSEC2MSEC(gethrtime() -
3757 scn
->scn_sync_start_time
),
3758 (longlong_t
)scn
->scn_avg_zio_size_this_txg
,
3759 (longlong_t
)scn
->scn_avg_seg_size_this_txg
);
3760 } else if (scn
->scn_done_txg
!= 0 && scn
->scn_done_txg
<= tx
->tx_txg
) {
3761 /* Finished with everything. Mark the scrub as complete */
3762 zfs_dbgmsg("scan issuing complete txg %llu",
3763 (longlong_t
)tx
->tx_txg
);
3764 ASSERT3U(scn
->scn_done_txg
, !=, 0);
3765 ASSERT0(spa
->spa_scrub_inflight
);
3766 ASSERT0(scn
->scn_bytes_pending
);
3767 dsl_scan_done(scn
, B_TRUE
, tx
);
3768 sync_type
= SYNC_MANDATORY
;
3771 dsl_scan_sync_state(scn
, tx
, sync_type
);
3775 count_block(dsl_scan_t
*scn
, zfs_all_blkstats_t
*zab
, const blkptr_t
*bp
)
3780 * Don't count embedded bp's, since we already did the work of
3781 * scanning these when we scanned the containing block.
3783 if (BP_IS_EMBEDDED(bp
))
3787 * Update the spa's stats on how many bytes we have issued.
3788 * Sequential scrubs create a zio for each DVA of the bp. Each
3789 * of these will include all DVAs for repair purposes, but the
3790 * zio code will only try the first one unless there is an issue.
3791 * Therefore, we should only count the first DVA for these IOs.
3793 if (scn
->scn_is_sorted
) {
3794 atomic_add_64(&scn
->scn_dp
->dp_spa
->spa_scan_pass_issued
,
3795 DVA_GET_ASIZE(&bp
->blk_dva
[0]));
3797 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3799 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
3800 atomic_add_64(&spa
->spa_scan_pass_issued
,
3801 DVA_GET_ASIZE(&bp
->blk_dva
[i
]));
3806 * If we resume after a reboot, zab will be NULL; don't record
3807 * incomplete stats in that case.
3812 mutex_enter(&zab
->zab_lock
);
3814 for (i
= 0; i
< 4; i
++) {
3815 int l
= (i
< 2) ? BP_GET_LEVEL(bp
) : DN_MAX_LEVELS
;
3816 int t
= (i
& 1) ? BP_GET_TYPE(bp
) : DMU_OT_TOTAL
;
3818 if (t
& DMU_OT_NEWTYPE
)
3820 zfs_blkstat_t
*zb
= &zab
->zab_type
[l
][t
];
3824 zb
->zb_asize
+= BP_GET_ASIZE(bp
);
3825 zb
->zb_lsize
+= BP_GET_LSIZE(bp
);
3826 zb
->zb_psize
+= BP_GET_PSIZE(bp
);
3827 zb
->zb_gangs
+= BP_COUNT_GANG(bp
);
3829 switch (BP_GET_NDVAS(bp
)) {
3831 if (DVA_GET_VDEV(&bp
->blk_dva
[0]) ==
3832 DVA_GET_VDEV(&bp
->blk_dva
[1]))
3833 zb
->zb_ditto_2_of_2_samevdev
++;
3836 equal
= (DVA_GET_VDEV(&bp
->blk_dva
[0]) ==
3837 DVA_GET_VDEV(&bp
->blk_dva
[1])) +
3838 (DVA_GET_VDEV(&bp
->blk_dva
[0]) ==
3839 DVA_GET_VDEV(&bp
->blk_dva
[2])) +
3840 (DVA_GET_VDEV(&bp
->blk_dva
[1]) ==
3841 DVA_GET_VDEV(&bp
->blk_dva
[2]));
3843 zb
->zb_ditto_2_of_3_samevdev
++;
3844 else if (equal
== 3)
3845 zb
->zb_ditto_3_of_3_samevdev
++;
3850 mutex_exit(&zab
->zab_lock
);
3854 scan_io_queue_insert_impl(dsl_scan_io_queue_t
*queue
, scan_io_t
*sio
)
3857 int64_t asize
= SIO_GET_ASIZE(sio
);
3858 dsl_scan_t
*scn
= queue
->q_scn
;
3860 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
3862 if (avl_find(&queue
->q_sios_by_addr
, sio
, &idx
) != NULL
) {
3863 /* block is already scheduled for reading */
3864 atomic_add_64(&scn
->scn_bytes_pending
, -asize
);
3868 avl_insert(&queue
->q_sios_by_addr
, sio
, idx
);
3869 queue
->q_sio_memused
+= SIO_GET_MUSED(sio
);
3870 range_tree_add(queue
->q_exts_by_addr
, SIO_GET_OFFSET(sio
), asize
);
3874 * Given all the info we got from our metadata scanning process, we
3875 * construct a scan_io_t and insert it into the scan sorting queue. The
3876 * I/O must already be suitable for us to process. This is controlled
3877 * by dsl_scan_enqueue().
3880 scan_io_queue_insert(dsl_scan_io_queue_t
*queue
, const blkptr_t
*bp
, int dva_i
,
3881 int zio_flags
, const zbookmark_phys_t
*zb
)
3883 dsl_scan_t
*scn
= queue
->q_scn
;
3884 scan_io_t
*sio
= sio_alloc(BP_GET_NDVAS(bp
));
3886 ASSERT0(BP_IS_GANG(bp
));
3887 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
3889 bp2sio(bp
, sio
, dva_i
);
3890 sio
->sio_flags
= zio_flags
;
3894 * Increment the bytes pending counter now so that we can't
3895 * get an integer underflow in case the worker processes the
3896 * zio before we get to incrementing this counter.
3898 atomic_add_64(&scn
->scn_bytes_pending
, SIO_GET_ASIZE(sio
));
3900 scan_io_queue_insert_impl(queue
, sio
);
3904 * Given a set of I/O parameters as discovered by the metadata traversal
3905 * process, attempts to place the I/O into the sorted queues (if allowed),
3906 * or immediately executes the I/O.
3909 dsl_scan_enqueue(dsl_pool_t
*dp
, const blkptr_t
*bp
, int zio_flags
,
3910 const zbookmark_phys_t
*zb
)
3912 spa_t
*spa
= dp
->dp_spa
;
3914 ASSERT(!BP_IS_EMBEDDED(bp
));
3917 * Gang blocks are hard to issue sequentially, so we just issue them
3918 * here immediately instead of queuing them.
3920 if (!dp
->dp_scan
->scn_is_sorted
|| BP_IS_GANG(bp
)) {
3921 scan_exec_io(dp
, bp
, zio_flags
, zb
, NULL
);
3925 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
3929 dva
= bp
->blk_dva
[i
];
3930 vdev
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
));
3931 ASSERT(vdev
!= NULL
);
3933 mutex_enter(&vdev
->vdev_scan_io_queue_lock
);
3934 if (vdev
->vdev_scan_io_queue
== NULL
)
3935 vdev
->vdev_scan_io_queue
= scan_io_queue_create(vdev
);
3936 ASSERT(dp
->dp_scan
!= NULL
);
3937 scan_io_queue_insert(vdev
->vdev_scan_io_queue
, bp
,
3939 mutex_exit(&vdev
->vdev_scan_io_queue_lock
);
3944 dsl_scan_scrub_cb(dsl_pool_t
*dp
,
3945 const blkptr_t
*bp
, const zbookmark_phys_t
*zb
)
3947 dsl_scan_t
*scn
= dp
->dp_scan
;
3948 spa_t
*spa
= dp
->dp_spa
;
3949 uint64_t phys_birth
= BP_PHYSICAL_BIRTH(bp
);
3950 size_t psize
= BP_GET_PSIZE(bp
);
3951 boolean_t needs_io
= B_FALSE
;
3952 int zio_flags
= ZIO_FLAG_SCAN_THREAD
| ZIO_FLAG_RAW
| ZIO_FLAG_CANFAIL
;
3955 if (phys_birth
<= scn
->scn_phys
.scn_min_txg
||
3956 phys_birth
>= scn
->scn_phys
.scn_max_txg
) {
3957 count_block(scn
, dp
->dp_blkstats
, bp
);
3961 /* Embedded BP's have phys_birth==0, so we reject them above. */
3962 ASSERT(!BP_IS_EMBEDDED(bp
));
3964 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn
));
3965 if (scn
->scn_phys
.scn_func
== POOL_SCAN_SCRUB
) {
3966 zio_flags
|= ZIO_FLAG_SCRUB
;
3969 ASSERT3U(scn
->scn_phys
.scn_func
, ==, POOL_SCAN_RESILVER
);
3970 zio_flags
|= ZIO_FLAG_RESILVER
;
3974 /* If it's an intent log block, failure is expected. */
3975 if (zb
->zb_level
== ZB_ZIL_LEVEL
)
3976 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
3978 for (int d
= 0; d
< BP_GET_NDVAS(bp
); d
++) {
3979 const dva_t
*dva
= &bp
->blk_dva
[d
];
3982 * Keep track of how much data we've examined so that
3983 * zpool(8) status can make useful progress reports.
3985 scn
->scn_phys
.scn_examined
+= DVA_GET_ASIZE(dva
);
3986 spa
->spa_scan_pass_exam
+= DVA_GET_ASIZE(dva
);
3988 /* if it's a resilver, this may not be in the target range */
3990 needs_io
= dsl_scan_need_resilver(spa
, dva
, psize
,
3994 if (needs_io
&& !zfs_no_scrub_io
) {
3995 dsl_scan_enqueue(dp
, bp
, zio_flags
, zb
);
3997 count_block(scn
, dp
->dp_blkstats
, bp
);
4000 /* do not relocate this block */
4005 dsl_scan_scrub_done(zio_t
*zio
)
4007 spa_t
*spa
= zio
->io_spa
;
4008 blkptr_t
*bp
= zio
->io_bp
;
4009 dsl_scan_io_queue_t
*queue
= zio
->io_private
;
4011 abd_free(zio
->io_abd
);
4013 if (queue
== NULL
) {
4014 mutex_enter(&spa
->spa_scrub_lock
);
4015 ASSERT3U(spa
->spa_scrub_inflight
, >=, BP_GET_PSIZE(bp
));
4016 spa
->spa_scrub_inflight
-= BP_GET_PSIZE(bp
);
4017 cv_broadcast(&spa
->spa_scrub_io_cv
);
4018 mutex_exit(&spa
->spa_scrub_lock
);
4020 mutex_enter(&queue
->q_vd
->vdev_scan_io_queue_lock
);
4021 ASSERT3U(queue
->q_inflight_bytes
, >=, BP_GET_PSIZE(bp
));
4022 queue
->q_inflight_bytes
-= BP_GET_PSIZE(bp
);
4023 cv_broadcast(&queue
->q_zio_cv
);
4024 mutex_exit(&queue
->q_vd
->vdev_scan_io_queue_lock
);
4027 if (zio
->io_error
&& (zio
->io_error
!= ECKSUM
||
4028 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
))) {
4029 atomic_inc_64(&spa
->spa_dsl_pool
->dp_scan
->scn_phys
.scn_errors
);
4034 * Given a scanning zio's information, executes the zio. The zio need
4035 * not necessarily be only sortable, this function simply executes the
4036 * zio, no matter what it is. The optional queue argument allows the
4037 * caller to specify that they want per top level vdev IO rate limiting
4038 * instead of the legacy global limiting.
4041 scan_exec_io(dsl_pool_t
*dp
, const blkptr_t
*bp
, int zio_flags
,
4042 const zbookmark_phys_t
*zb
, dsl_scan_io_queue_t
*queue
)
4044 spa_t
*spa
= dp
->dp_spa
;
4045 dsl_scan_t
*scn
= dp
->dp_scan
;
4046 size_t size
= BP_GET_PSIZE(bp
);
4047 abd_t
*data
= abd_alloc_for_io(size
, B_FALSE
);
4049 ASSERT3U(scn
->scn_maxinflight_bytes
, >, 0);
4051 if (queue
== NULL
) {
4052 mutex_enter(&spa
->spa_scrub_lock
);
4053 while (spa
->spa_scrub_inflight
>= scn
->scn_maxinflight_bytes
)
4054 cv_wait(&spa
->spa_scrub_io_cv
, &spa
->spa_scrub_lock
);
4055 spa
->spa_scrub_inflight
+= BP_GET_PSIZE(bp
);
4056 mutex_exit(&spa
->spa_scrub_lock
);
4058 kmutex_t
*q_lock
= &queue
->q_vd
->vdev_scan_io_queue_lock
;
4060 mutex_enter(q_lock
);
4061 while (queue
->q_inflight_bytes
>= queue
->q_maxinflight_bytes
)
4062 cv_wait(&queue
->q_zio_cv
, q_lock
);
4063 queue
->q_inflight_bytes
+= BP_GET_PSIZE(bp
);
4067 count_block(scn
, dp
->dp_blkstats
, bp
);
4068 zio_nowait(zio_read(scn
->scn_zio_root
, spa
, bp
, data
, size
,
4069 dsl_scan_scrub_done
, queue
, ZIO_PRIORITY_SCRUB
, zio_flags
, zb
));
4073 * This is the primary extent sorting algorithm. We balance two parameters:
4074 * 1) how many bytes of I/O are in an extent
4075 * 2) how well the extent is filled with I/O (as a fraction of its total size)
4076 * Since we allow extents to have gaps between their constituent I/Os, it's
4077 * possible to have a fairly large extent that contains the same amount of
4078 * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
4079 * The algorithm sorts based on a score calculated from the extent's size,
4080 * the relative fill volume (in %) and a "fill weight" parameter that controls
4081 * the split between whether we prefer larger extents or more well populated
4084 * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
4087 * 1) assume extsz = 64 MiB
4088 * 2) assume fill = 32 MiB (extent is half full)
4089 * 3) assume fill_weight = 3
4090 * 4) SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
4091 * SCORE = 32M + (50 * 3 * 32M) / 100
4092 * SCORE = 32M + (4800M / 100)
4095 * | +--- final total relative fill-based score
4096 * +--------- final total fill-based score
4099 * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
4100 * extents that are more completely filled (in a 3:2 ratio) vs just larger.
4101 * Note that as an optimization, we replace multiplication and division by
4102 * 100 with bitshifting by 7 (which effectively multiplies and divides by 128).
4105 ext_size_compare(const void *x
, const void *y
)
4107 const range_seg_gap_t
*rsa
= x
, *rsb
= y
;
4109 uint64_t sa
= rsa
->rs_end
- rsa
->rs_start
;
4110 uint64_t sb
= rsb
->rs_end
- rsb
->rs_start
;
4111 uint64_t score_a
, score_b
;
4113 score_a
= rsa
->rs_fill
+ ((((rsa
->rs_fill
<< 7) / sa
) *
4114 fill_weight
* rsa
->rs_fill
) >> 7);
4115 score_b
= rsb
->rs_fill
+ ((((rsb
->rs_fill
<< 7) / sb
) *
4116 fill_weight
* rsb
->rs_fill
) >> 7);
4118 if (score_a
> score_b
)
4120 if (score_a
== score_b
) {
4121 if (rsa
->rs_start
< rsb
->rs_start
)
4123 if (rsa
->rs_start
== rsb
->rs_start
)
4131 * Comparator for the q_sios_by_addr tree. Sorting is simply performed
4132 * based on LBA-order (from lowest to highest).
4135 sio_addr_compare(const void *x
, const void *y
)
4137 const scan_io_t
*a
= x
, *b
= y
;
4139 return (TREE_CMP(SIO_GET_OFFSET(a
), SIO_GET_OFFSET(b
)));
4142 /* IO queues are created on demand when they are needed. */
4143 static dsl_scan_io_queue_t
*
4144 scan_io_queue_create(vdev_t
*vd
)
4146 dsl_scan_t
*scn
= vd
->vdev_spa
->spa_dsl_pool
->dp_scan
;
4147 dsl_scan_io_queue_t
*q
= kmem_zalloc(sizeof (*q
), KM_SLEEP
);
4151 q
->q_sio_memused
= 0;
4152 cv_init(&q
->q_zio_cv
, NULL
, CV_DEFAULT
, NULL
);
4153 q
->q_exts_by_addr
= range_tree_create_impl(&rt_btree_ops
, RANGE_SEG_GAP
,
4154 &q
->q_exts_by_size
, 0, 0, ext_size_compare
, zfs_scan_max_ext_gap
);
4155 avl_create(&q
->q_sios_by_addr
, sio_addr_compare
,
4156 sizeof (scan_io_t
), offsetof(scan_io_t
, sio_nodes
.sio_addr_node
));
4162 * Destroys a scan queue and all segments and scan_io_t's contained in it.
4163 * No further execution of I/O occurs, anything pending in the queue is
4164 * simply freed without being executed.
4167 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t
*queue
)
4169 dsl_scan_t
*scn
= queue
->q_scn
;
4171 void *cookie
= NULL
;
4172 int64_t bytes_dequeued
= 0;
4174 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
4176 while ((sio
= avl_destroy_nodes(&queue
->q_sios_by_addr
, &cookie
)) !=
4178 ASSERT(range_tree_contains(queue
->q_exts_by_addr
,
4179 SIO_GET_OFFSET(sio
), SIO_GET_ASIZE(sio
)));
4180 bytes_dequeued
+= SIO_GET_ASIZE(sio
);
4181 queue
->q_sio_memused
-= SIO_GET_MUSED(sio
);
4185 ASSERT0(queue
->q_sio_memused
);
4186 atomic_add_64(&scn
->scn_bytes_pending
, -bytes_dequeued
);
4187 range_tree_vacate(queue
->q_exts_by_addr
, NULL
, queue
);
4188 range_tree_destroy(queue
->q_exts_by_addr
);
4189 avl_destroy(&queue
->q_sios_by_addr
);
4190 cv_destroy(&queue
->q_zio_cv
);
4192 kmem_free(queue
, sizeof (*queue
));
4196 * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
4197 * called on behalf of vdev_top_transfer when creating or destroying
4198 * a mirror vdev due to zpool attach/detach.
4201 dsl_scan_io_queue_vdev_xfer(vdev_t
*svd
, vdev_t
*tvd
)
4203 mutex_enter(&svd
->vdev_scan_io_queue_lock
);
4204 mutex_enter(&tvd
->vdev_scan_io_queue_lock
);
4206 VERIFY3P(tvd
->vdev_scan_io_queue
, ==, NULL
);
4207 tvd
->vdev_scan_io_queue
= svd
->vdev_scan_io_queue
;
4208 svd
->vdev_scan_io_queue
= NULL
;
4209 if (tvd
->vdev_scan_io_queue
!= NULL
)
4210 tvd
->vdev_scan_io_queue
->q_vd
= tvd
;
4212 mutex_exit(&tvd
->vdev_scan_io_queue_lock
);
4213 mutex_exit(&svd
->vdev_scan_io_queue_lock
);
4217 scan_io_queues_destroy(dsl_scan_t
*scn
)
4219 vdev_t
*rvd
= scn
->scn_dp
->dp_spa
->spa_root_vdev
;
4221 for (uint64_t i
= 0; i
< rvd
->vdev_children
; i
++) {
4222 vdev_t
*tvd
= rvd
->vdev_child
[i
];
4224 mutex_enter(&tvd
->vdev_scan_io_queue_lock
);
4225 if (tvd
->vdev_scan_io_queue
!= NULL
)
4226 dsl_scan_io_queue_destroy(tvd
->vdev_scan_io_queue
);
4227 tvd
->vdev_scan_io_queue
= NULL
;
4228 mutex_exit(&tvd
->vdev_scan_io_queue_lock
);
4233 dsl_scan_freed_dva(spa_t
*spa
, const blkptr_t
*bp
, int dva_i
)
4235 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
4236 dsl_scan_t
*scn
= dp
->dp_scan
;
4239 dsl_scan_io_queue_t
*queue
;
4240 scan_io_t
*srch_sio
, *sio
;
4242 uint64_t start
, size
;
4244 vdev
= vdev_lookup_top(spa
, DVA_GET_VDEV(&bp
->blk_dva
[dva_i
]));
4245 ASSERT(vdev
!= NULL
);
4246 q_lock
= &vdev
->vdev_scan_io_queue_lock
;
4247 queue
= vdev
->vdev_scan_io_queue
;
4249 mutex_enter(q_lock
);
4250 if (queue
== NULL
) {
4255 srch_sio
= sio_alloc(BP_GET_NDVAS(bp
));
4256 bp2sio(bp
, srch_sio
, dva_i
);
4257 start
= SIO_GET_OFFSET(srch_sio
);
4258 size
= SIO_GET_ASIZE(srch_sio
);
4261 * We can find the zio in two states:
4262 * 1) Cold, just sitting in the queue of zio's to be issued at
4263 * some point in the future. In this case, all we do is
4264 * remove the zio from the q_sios_by_addr tree, decrement
4265 * its data volume from the containing range_seg_t and
4266 * resort the q_exts_by_size tree to reflect that the
4267 * range_seg_t has lost some of its 'fill'. We don't shorten
4268 * the range_seg_t - this is usually rare enough not to be
4269 * worth the extra hassle of trying keep track of precise
4270 * extent boundaries.
4271 * 2) Hot, where the zio is currently in-flight in
4272 * dsl_scan_issue_ios. In this case, we can't simply
4273 * reach in and stop the in-flight zio's, so we instead
4274 * block the caller. Eventually, dsl_scan_issue_ios will
4275 * be done with issuing the zio's it gathered and will
4278 sio
= avl_find(&queue
->q_sios_by_addr
, srch_sio
, &idx
);
4282 int64_t asize
= SIO_GET_ASIZE(sio
);
4285 /* Got it while it was cold in the queue */
4286 ASSERT3U(start
, ==, SIO_GET_OFFSET(sio
));
4287 ASSERT3U(size
, ==, asize
);
4288 avl_remove(&queue
->q_sios_by_addr
, sio
);
4289 queue
->q_sio_memused
-= SIO_GET_MUSED(sio
);
4291 ASSERT(range_tree_contains(queue
->q_exts_by_addr
, start
, size
));
4292 range_tree_remove_fill(queue
->q_exts_by_addr
, start
, size
);
4295 * We only update scn_bytes_pending in the cold path,
4296 * otherwise it will already have been accounted for as
4297 * part of the zio's execution.
4299 atomic_add_64(&scn
->scn_bytes_pending
, -asize
);
4301 /* count the block as though we issued it */
4302 sio2bp(sio
, &tmpbp
);
4303 count_block(scn
, dp
->dp_blkstats
, &tmpbp
);
4311 * Callback invoked when a zio_free() zio is executing. This needs to be
4312 * intercepted to prevent the zio from deallocating a particular portion
4313 * of disk space and it then getting reallocated and written to, while we
4314 * still have it queued up for processing.
4317 dsl_scan_freed(spa_t
*spa
, const blkptr_t
*bp
)
4319 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
4320 dsl_scan_t
*scn
= dp
->dp_scan
;
4322 ASSERT(!BP_IS_EMBEDDED(bp
));
4323 ASSERT(scn
!= NULL
);
4324 if (!dsl_scan_is_running(scn
))
4327 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++)
4328 dsl_scan_freed_dva(spa
, bp
, i
);
4332 * Check if a vdev needs resilvering (non-empty DTL), if so, and resilver has
4333 * not started, start it. Otherwise, only restart if max txg in DTL range is
4334 * greater than the max txg in the current scan. If the DTL max is less than
4335 * the scan max, then the vdev has not missed any new data since the resilver
4336 * started, so a restart is not needed.
4339 dsl_scan_assess_vdev(dsl_pool_t
*dp
, vdev_t
*vd
)
4343 if (!vdev_resilver_needed(vd
, &min
, &max
))
4346 if (!dsl_scan_resilvering(dp
)) {
4347 spa_async_request(dp
->dp_spa
, SPA_ASYNC_RESILVER
);
4351 if (max
<= dp
->dp_scan
->scn_phys
.scn_max_txg
)
4354 /* restart is needed, check if it can be deferred */
4355 if (spa_feature_is_enabled(dp
->dp_spa
, SPA_FEATURE_RESILVER_DEFER
))
4356 vdev_defer_resilver(vd
);
4358 spa_async_request(dp
->dp_spa
, SPA_ASYNC_RESILVER
);
4362 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_vdev_limit
, ULONG
, ZMOD_RW
,
4363 "Max bytes in flight per leaf vdev for scrubs and resilvers");
4365 ZFS_MODULE_PARAM(zfs
, zfs_
, scrub_min_time_ms
, INT
, ZMOD_RW
,
4366 "Min millisecs to scrub per txg");
4368 ZFS_MODULE_PARAM(zfs
, zfs_
, obsolete_min_time_ms
, INT
, ZMOD_RW
,
4369 "Min millisecs to obsolete per txg");
4371 ZFS_MODULE_PARAM(zfs
, zfs_
, free_min_time_ms
, INT
, ZMOD_RW
,
4372 "Min millisecs to free per txg");
4374 ZFS_MODULE_PARAM(zfs
, zfs_
, resilver_min_time_ms
, INT
, ZMOD_RW
,
4375 "Min millisecs to resilver per txg");
4377 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_suspend_progress
, INT
, ZMOD_RW
,
4378 "Set to prevent scans from progressing");
4380 ZFS_MODULE_PARAM(zfs
, zfs_
, no_scrub_io
, INT
, ZMOD_RW
,
4381 "Set to disable scrub I/O");
4383 ZFS_MODULE_PARAM(zfs
, zfs_
, no_scrub_prefetch
, INT
, ZMOD_RW
,
4384 "Set to disable scrub prefetching");
4386 ZFS_MODULE_PARAM(zfs
, zfs_
, async_block_max_blocks
, ULONG
, ZMOD_RW
,
4387 "Max number of blocks freed in one txg");
4389 ZFS_MODULE_PARAM(zfs
, zfs_
, max_async_dedup_frees
, ULONG
, ZMOD_RW
,
4390 "Max number of dedup blocks freed in one txg");
4392 ZFS_MODULE_PARAM(zfs
, zfs_
, free_bpobj_enabled
, INT
, ZMOD_RW
,
4393 "Enable processing of the free_bpobj");
4395 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_mem_lim_fact
, INT
, ZMOD_RW
,
4396 "Fraction of RAM for scan hard limit");
4398 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_issue_strategy
, INT
, ZMOD_RW
,
4399 "IO issuing strategy during scrubbing. "
4400 "0 = default, 1 = LBA, 2 = size");
4402 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_legacy
, INT
, ZMOD_RW
,
4403 "Scrub using legacy non-sequential method");
4405 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_checkpoint_intval
, INT
, ZMOD_RW
,
4406 "Scan progress on-disk checkpointing interval");
4408 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_max_ext_gap
, ULONG
, ZMOD_RW
,
4409 "Max gap in bytes between sequential scrub / resilver I/Os");
4411 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_mem_lim_soft_fact
, INT
, ZMOD_RW
,
4412 "Fraction of hard limit used as soft limit");
4414 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_strict_mem_lim
, INT
, ZMOD_RW
,
4415 "Tunable to attempt to reduce lock contention");
4417 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_fill_weight
, INT
, ZMOD_RW
,
4418 "Tunable to adjust bias towards more filled segments during scans");
4420 ZFS_MODULE_PARAM(zfs
, zfs_
, resilver_disable_defer
, INT
, ZMOD_RW
,
4421 "Process all resilvers immediately");