4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright 2016 Gary Mills
25 * Copyright (c) 2017 Datto Inc.
26 * Copyright 2017 Joyent, Inc.
29 #include <sys/dsl_scan.h>
30 #include <sys/dsl_pool.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_prop.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_synctask.h>
35 #include <sys/dnode.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/dmu_objset.h>
41 #include <sys/zfs_context.h>
42 #include <sys/fs/zfs.h>
43 #include <sys/zfs_znode.h>
44 #include <sys/spa_impl.h>
45 #include <sys/vdev_impl.h>
46 #include <sys/zil_impl.h>
47 #include <sys/zio_checksum.h>
50 #include <sys/sa_impl.h>
51 #include <sys/zfeature.h>
53 #include <sys/range_tree.h>
55 #include <sys/zfs_vfsops.h>
59 * Grand theory statement on scan queue sorting
61 * Scanning is implemented by recursively traversing all indirection levels
62 * in an object and reading all blocks referenced from said objects. This
63 * results in us approximately traversing the object from lowest logical
64 * offset to the highest. For best performance, we would want the logical
65 * blocks to be physically contiguous. However, this is frequently not the
66 * case with pools given the allocation patterns of copy-on-write filesystems.
67 * So instead, we put the I/Os into a reordering queue and issue them in a
68 * way that will most benefit physical disks (LBA-order).
72 * Ideally, we would want to scan all metadata and queue up all block I/O
73 * prior to starting to issue it, because that allows us to do an optimal
74 * sorting job. This can however consume large amounts of memory. Therefore
75 * we continuously monitor the size of the queues and constrain them to 5%
76 * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this
77 * limit, we clear out a few of the largest extents at the head of the queues
78 * to make room for more scanning. Hopefully, these extents will be fairly
79 * large and contiguous, allowing us to approach sequential I/O throughput
80 * even without a fully sorted tree.
82 * Metadata scanning takes place in dsl_scan_visit(), which is called from
83 * dsl_scan_sync() every spa_sync(). If we have either fully scanned all
84 * metadata on the pool, or we need to make room in memory because our
85 * queues are too large, dsl_scan_visit() is postponed and
86 * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies
87 * that metadata scanning and queued I/O issuing are mutually exclusive. This
88 * allows us to provide maximum sequential I/O throughput for the majority of
89 * I/O's issued since sequential I/O performance is significantly negatively
90 * impacted if it is interleaved with random I/O.
92 * Implementation Notes
94 * One side effect of the queued scanning algorithm is that the scanning code
95 * needs to be notified whenever a block is freed. This is needed to allow
96 * the scanning code to remove these I/Os from the issuing queue. Additionally,
97 * we do not attempt to queue gang blocks to be issued sequentially since this
98 * is very hard to do and would have an extremely limited performance benefit.
99 * Instead, we simply issue gang I/Os as soon as we find them using the legacy
102 * Backwards compatibility
104 * This new algorithm is backwards compatible with the legacy on-disk data
105 * structures (and therefore does not require a new feature flag).
106 * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan
107 * will stop scanning metadata (in logical order) and wait for all outstanding
108 * sorted I/O to complete. Once this is done, we write out a checkpoint
109 * bookmark, indicating that we have scanned everything logically before it.
110 * If the pool is imported on a machine without the new sorting algorithm,
111 * the scan simply resumes from the last checkpoint using the legacy algorithm.
114 typedef int (scan_cb_t
)(dsl_pool_t
*, const blkptr_t
*,
115 const zbookmark_phys_t
*);
117 static scan_cb_t dsl_scan_scrub_cb
;
119 static int scan_ds_queue_compare(const void *a
, const void *b
);
120 static int scan_prefetch_queue_compare(const void *a
, const void *b
);
121 static void scan_ds_queue_clear(dsl_scan_t
*scn
);
122 static void scan_ds_prefetch_queue_clear(dsl_scan_t
*scn
);
123 static boolean_t
scan_ds_queue_contains(dsl_scan_t
*scn
, uint64_t dsobj
,
125 static void scan_ds_queue_insert(dsl_scan_t
*scn
, uint64_t dsobj
, uint64_t txg
);
126 static void scan_ds_queue_remove(dsl_scan_t
*scn
, uint64_t dsobj
);
127 static void scan_ds_queue_sync(dsl_scan_t
*scn
, dmu_tx_t
*tx
);
128 static uint64_t dsl_scan_count_leaves(vdev_t
*vd
);
130 extern int zfs_vdev_async_write_active_min_dirty_percent
;
133 * By default zfs will check to ensure it is not over the hard memory
134 * limit before each txg. If finer-grained control of this is needed
135 * this value can be set to 1 to enable checking before scanning each
138 int zfs_scan_strict_mem_lim
= B_FALSE
;
141 * Maximum number of parallelly executed bytes per leaf vdev. We attempt
142 * to strike a balance here between keeping the vdev queues full of I/Os
143 * at all times and not overflowing the queues to cause long latency,
144 * which would cause long txg sync times. No matter what, we will not
145 * overload the drives with I/O, since that is protected by
146 * zfs_vdev_scrub_max_active.
148 unsigned long zfs_scan_vdev_limit
= 4 << 20;
150 int zfs_scan_issue_strategy
= 0;
151 int zfs_scan_legacy
= B_FALSE
; /* don't queue & sort zios, go direct */
152 unsigned long zfs_scan_max_ext_gap
= 2 << 20; /* in bytes */
155 * fill_weight is non-tunable at runtime, so we copy it at module init from
156 * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
157 * break queue sorting.
159 int zfs_scan_fill_weight
= 3;
160 static uint64_t fill_weight
;
162 /* See dsl_scan_should_clear() for details on the memory limit tunables */
163 uint64_t zfs_scan_mem_lim_min
= 16 << 20; /* bytes */
164 uint64_t zfs_scan_mem_lim_soft_max
= 128 << 20; /* bytes */
165 int zfs_scan_mem_lim_fact
= 20; /* fraction of physmem */
166 int zfs_scan_mem_lim_soft_fact
= 20; /* fraction of mem lim above */
168 int zfs_scrub_min_time_ms
= 1000; /* min millisecs to scrub per txg */
169 int zfs_obsolete_min_time_ms
= 500; /* min millisecs to obsolete per txg */
170 int zfs_free_min_time_ms
= 1000; /* min millisecs to free per txg */
171 int zfs_resilver_min_time_ms
= 3000; /* min millisecs to resilver per txg */
172 int zfs_scan_checkpoint_intval
= 7200; /* in seconds */
173 int zfs_scan_suspend_progress
= 0; /* set to prevent scans from progressing */
174 int zfs_no_scrub_io
= B_FALSE
; /* set to disable scrub i/o */
175 int zfs_no_scrub_prefetch
= B_FALSE
; /* set to disable scrub prefetch */
176 enum ddt_class zfs_scrub_ddt_class_max
= DDT_CLASS_DUPLICATE
;
177 /* max number of blocks to free in a single TXG */
178 unsigned long zfs_async_block_max_blocks
= 100000;
180 int zfs_resilver_disable_defer
= 0; /* set to disable resilver deferring */
183 * We wait a few txgs after importing a pool to begin scanning so that
184 * the import / mounting code isn't held up by scrub / resilver IO.
185 * Unfortunately, it is a bit difficult to determine exactly how long
186 * this will take since userspace will trigger fs mounts asynchronously
187 * and the kernel will create zvol minors asynchronously. As a result,
188 * the value provided here is a bit arbitrary, but represents a
189 * reasonable estimate of how many txgs it will take to finish fully
192 #define SCAN_IMPORT_WAIT_TXGS 5
194 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
195 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
196 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
199 * Enable/disable the processing of the free_bpobj object.
201 int zfs_free_bpobj_enabled
= 1;
203 /* the order has to match pool_scan_type */
204 static scan_cb_t
*scan_funcs
[POOL_SCAN_FUNCS
] = {
206 dsl_scan_scrub_cb
, /* POOL_SCAN_SCRUB */
207 dsl_scan_scrub_cb
, /* POOL_SCAN_RESILVER */
210 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */
218 * This controls what conditions are placed on dsl_scan_sync_state():
219 * SYNC_OPTIONAL) write out scn_phys iff scn_bytes_pending == 0
220 * SYNC_MANDATORY) write out scn_phys always. scn_bytes_pending must be 0.
221 * SYNC_CACHED) if scn_bytes_pending == 0, write out scn_phys. Otherwise
222 * write out the scn_phys_cached version.
223 * See dsl_scan_sync_state for details.
232 * This struct represents the minimum information needed to reconstruct a
233 * zio for sequential scanning. This is useful because many of these will
234 * accumulate in the sequential IO queues before being issued, so saving
235 * memory matters here.
237 typedef struct scan_io
{
238 /* fields from blkptr_t */
239 uint64_t sio_blk_prop
;
240 uint64_t sio_phys_birth
;
242 zio_cksum_t sio_cksum
;
243 uint32_t sio_nr_dvas
;
245 /* fields from zio_t */
247 zbookmark_phys_t sio_zb
;
249 /* members for queue sorting */
251 avl_node_t sio_addr_node
; /* link into issuing queue */
252 list_node_t sio_list_node
; /* link for issuing to disk */
256 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp,
257 * depending on how many were in the original bp. Only the
258 * first DVA is really used for sorting and issuing purposes.
259 * The other DVAs (if provided) simply exist so that the zio
260 * layer can find additional copies to repair from in the
261 * event of an error. This array must go at the end of the
262 * struct to allow this for the variable number of elements.
267 #define SIO_SET_OFFSET(sio, x) DVA_SET_OFFSET(&(sio)->sio_dva[0], x)
268 #define SIO_SET_ASIZE(sio, x) DVA_SET_ASIZE(&(sio)->sio_dva[0], x)
269 #define SIO_GET_OFFSET(sio) DVA_GET_OFFSET(&(sio)->sio_dva[0])
270 #define SIO_GET_ASIZE(sio) DVA_GET_ASIZE(&(sio)->sio_dva[0])
271 #define SIO_GET_END_OFFSET(sio) \
272 (SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio))
273 #define SIO_GET_MUSED(sio) \
274 (sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t)))
276 struct dsl_scan_io_queue
{
277 dsl_scan_t
*q_scn
; /* associated dsl_scan_t */
278 vdev_t
*q_vd
; /* top-level vdev that this queue represents */
280 /* trees used for sorting I/Os and extents of I/Os */
281 range_tree_t
*q_exts_by_addr
;
282 avl_tree_t q_exts_by_size
;
283 avl_tree_t q_sios_by_addr
;
284 uint64_t q_sio_memused
;
286 /* members for zio rate limiting */
287 uint64_t q_maxinflight_bytes
;
288 uint64_t q_inflight_bytes
;
289 kcondvar_t q_zio_cv
; /* used under vd->vdev_scan_io_queue_lock */
291 /* per txg statistics */
292 uint64_t q_total_seg_size_this_txg
;
293 uint64_t q_segs_this_txg
;
294 uint64_t q_total_zio_size_this_txg
;
295 uint64_t q_zios_this_txg
;
298 /* private data for dsl_scan_prefetch_cb() */
299 typedef struct scan_prefetch_ctx
{
300 zfs_refcount_t spc_refcnt
; /* refcount for memory management */
301 dsl_scan_t
*spc_scn
; /* dsl_scan_t for the pool */
302 boolean_t spc_root
; /* is this prefetch for an objset? */
303 uint8_t spc_indblkshift
; /* dn_indblkshift of current dnode */
304 uint16_t spc_datablkszsec
; /* dn_idatablkszsec of current dnode */
305 } scan_prefetch_ctx_t
;
307 /* private data for dsl_scan_prefetch() */
308 typedef struct scan_prefetch_issue_ctx
{
309 avl_node_t spic_avl_node
; /* link into scn->scn_prefetch_queue */
310 scan_prefetch_ctx_t
*spic_spc
; /* spc for the callback */
311 blkptr_t spic_bp
; /* bp to prefetch */
312 zbookmark_phys_t spic_zb
; /* bookmark to prefetch */
313 } scan_prefetch_issue_ctx_t
;
315 static void scan_exec_io(dsl_pool_t
*dp
, const blkptr_t
*bp
, int zio_flags
,
316 const zbookmark_phys_t
*zb
, dsl_scan_io_queue_t
*queue
);
317 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t
*queue
,
320 static dsl_scan_io_queue_t
*scan_io_queue_create(vdev_t
*vd
);
321 static void scan_io_queues_destroy(dsl_scan_t
*scn
);
323 static kmem_cache_t
*sio_cache
[SPA_DVAS_PER_BP
];
325 /* sio->sio_nr_dvas must be set so we know which cache to free from */
327 sio_free(scan_io_t
*sio
)
329 ASSERT3U(sio
->sio_nr_dvas
, >, 0);
330 ASSERT3U(sio
->sio_nr_dvas
, <=, SPA_DVAS_PER_BP
);
332 kmem_cache_free(sio_cache
[sio
->sio_nr_dvas
- 1], sio
);
335 /* It is up to the caller to set sio->sio_nr_dvas for freeing */
337 sio_alloc(unsigned short nr_dvas
)
339 ASSERT3U(nr_dvas
, >, 0);
340 ASSERT3U(nr_dvas
, <=, SPA_DVAS_PER_BP
);
342 return (kmem_cache_alloc(sio_cache
[nr_dvas
- 1], KM_SLEEP
));
349 * This is used in ext_size_compare() to weight segments
350 * based on how sparse they are. This cannot be changed
351 * mid-scan and the tree comparison functions don't currently
352 * have a mechanism for passing additional context to the
353 * compare functions. Thus we store this value globally and
354 * we only allow it to be set at module initialization time
356 fill_weight
= zfs_scan_fill_weight
;
358 for (int i
= 0; i
< SPA_DVAS_PER_BP
; i
++) {
361 (void) sprintf(name
, "sio_cache_%d", i
);
362 sio_cache
[i
] = kmem_cache_create(name
,
363 (sizeof (scan_io_t
) + ((i
+ 1) * sizeof (dva_t
))),
364 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
371 for (int i
= 0; i
< SPA_DVAS_PER_BP
; i
++) {
372 kmem_cache_destroy(sio_cache
[i
]);
376 static inline boolean_t
377 dsl_scan_is_running(const dsl_scan_t
*scn
)
379 return (scn
->scn_phys
.scn_state
== DSS_SCANNING
);
383 dsl_scan_resilvering(dsl_pool_t
*dp
)
385 return (dsl_scan_is_running(dp
->dp_scan
) &&
386 dp
->dp_scan
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
);
390 sio2bp(const scan_io_t
*sio
, blkptr_t
*bp
)
392 bzero(bp
, sizeof (*bp
));
393 bp
->blk_prop
= sio
->sio_blk_prop
;
394 bp
->blk_phys_birth
= sio
->sio_phys_birth
;
395 bp
->blk_birth
= sio
->sio_birth
;
396 bp
->blk_fill
= 1; /* we always only work with data pointers */
397 bp
->blk_cksum
= sio
->sio_cksum
;
399 ASSERT3U(sio
->sio_nr_dvas
, >, 0);
400 ASSERT3U(sio
->sio_nr_dvas
, <=, SPA_DVAS_PER_BP
);
402 bcopy(sio
->sio_dva
, bp
->blk_dva
, sio
->sio_nr_dvas
* sizeof (dva_t
));
406 bp2sio(const blkptr_t
*bp
, scan_io_t
*sio
, int dva_i
)
408 sio
->sio_blk_prop
= bp
->blk_prop
;
409 sio
->sio_phys_birth
= bp
->blk_phys_birth
;
410 sio
->sio_birth
= bp
->blk_birth
;
411 sio
->sio_cksum
= bp
->blk_cksum
;
412 sio
->sio_nr_dvas
= BP_GET_NDVAS(bp
);
415 * Copy the DVAs to the sio. We need all copies of the block so
416 * that the self healing code can use the alternate copies if the
417 * first is corrupted. We want the DVA at index dva_i to be first
418 * in the sio since this is the primary one that we want to issue.
420 for (int i
= 0, j
= dva_i
; i
< sio
->sio_nr_dvas
; i
++, j
++) {
421 sio
->sio_dva
[i
] = bp
->blk_dva
[j
% sio
->sio_nr_dvas
];
426 dsl_scan_init(dsl_pool_t
*dp
, uint64_t txg
)
430 spa_t
*spa
= dp
->dp_spa
;
433 scn
= dp
->dp_scan
= kmem_zalloc(sizeof (dsl_scan_t
), KM_SLEEP
);
437 * It's possible that we're resuming a scan after a reboot so
438 * make sure that the scan_async_destroying flag is initialized
441 ASSERT(!scn
->scn_async_destroying
);
442 scn
->scn_async_destroying
= spa_feature_is_active(dp
->dp_spa
,
443 SPA_FEATURE_ASYNC_DESTROY
);
446 * Calculate the max number of in-flight bytes for pool-wide
447 * scanning operations (minimum 1MB). Limits for the issuing
448 * phase are done per top-level vdev and are handled separately.
450 scn
->scn_maxinflight_bytes
= MAX(zfs_scan_vdev_limit
*
451 dsl_scan_count_leaves(spa
->spa_root_vdev
), 1ULL << 20);
453 avl_create(&scn
->scn_queue
, scan_ds_queue_compare
, sizeof (scan_ds_t
),
454 offsetof(scan_ds_t
, sds_node
));
455 avl_create(&scn
->scn_prefetch_queue
, scan_prefetch_queue_compare
,
456 sizeof (scan_prefetch_issue_ctx_t
),
457 offsetof(scan_prefetch_issue_ctx_t
, spic_avl_node
));
459 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
460 "scrub_func", sizeof (uint64_t), 1, &f
);
463 * There was an old-style scrub in progress. Restart a
464 * new-style scrub from the beginning.
466 scn
->scn_restart_txg
= txg
;
467 zfs_dbgmsg("old-style scrub was in progress; "
468 "restarting new-style scrub in txg %llu",
469 (longlong_t
)scn
->scn_restart_txg
);
472 * Load the queue obj from the old location so that it
473 * can be freed by dsl_scan_done().
475 (void) zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
476 "scrub_queue", sizeof (uint64_t), 1,
477 &scn
->scn_phys
.scn_queue_obj
);
479 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
480 DMU_POOL_SCAN
, sizeof (uint64_t), SCAN_PHYS_NUMINTS
,
483 * Detect if the pool contains the signature of #2094. If it
484 * does properly update the scn->scn_phys structure and notify
485 * the administrator by setting an errata for the pool.
487 if (err
== EOVERFLOW
) {
488 uint64_t zaptmp
[SCAN_PHYS_NUMINTS
+ 1];
489 VERIFY3S(SCAN_PHYS_NUMINTS
, ==, 24);
490 VERIFY3S(offsetof(dsl_scan_phys_t
, scn_flags
), ==,
491 (23 * sizeof (uint64_t)));
493 err
= zap_lookup(dp
->dp_meta_objset
,
494 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_SCAN
,
495 sizeof (uint64_t), SCAN_PHYS_NUMINTS
+ 1, &zaptmp
);
497 uint64_t overflow
= zaptmp
[SCAN_PHYS_NUMINTS
];
499 if (overflow
& ~DSL_SCAN_FLAGS_MASK
||
500 scn
->scn_async_destroying
) {
502 ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY
;
506 bcopy(zaptmp
, &scn
->scn_phys
,
507 SCAN_PHYS_NUMINTS
* sizeof (uint64_t));
508 scn
->scn_phys
.scn_flags
= overflow
;
510 /* Required scrub already in progress. */
511 if (scn
->scn_phys
.scn_state
== DSS_FINISHED
||
512 scn
->scn_phys
.scn_state
== DSS_CANCELED
)
514 ZPOOL_ERRATA_ZOL_2094_SCRUB
;
524 * We might be restarting after a reboot, so jump the issued
525 * counter to how far we've scanned. We know we're consistent
528 scn
->scn_issued_before_pass
= scn
->scn_phys
.scn_examined
;
530 if (dsl_scan_is_running(scn
) &&
531 spa_prev_software_version(dp
->dp_spa
) < SPA_VERSION_SCAN
) {
533 * A new-type scrub was in progress on an old
534 * pool, and the pool was accessed by old
535 * software. Restart from the beginning, since
536 * the old software may have changed the pool in
539 scn
->scn_restart_txg
= txg
;
540 zfs_dbgmsg("new-style scrub was modified "
541 "by old software; restarting in txg %llu",
542 (longlong_t
)scn
->scn_restart_txg
);
546 bcopy(&scn
->scn_phys
, &scn
->scn_phys_cached
, sizeof (scn
->scn_phys
));
548 /* reload the queue into the in-core state */
549 if (scn
->scn_phys
.scn_queue_obj
!= 0) {
553 for (zap_cursor_init(&zc
, dp
->dp_meta_objset
,
554 scn
->scn_phys
.scn_queue_obj
);
555 zap_cursor_retrieve(&zc
, &za
) == 0;
556 (void) zap_cursor_advance(&zc
)) {
557 scan_ds_queue_insert(scn
,
558 zfs_strtonum(za
.za_name
, NULL
),
559 za
.za_first_integer
);
561 zap_cursor_fini(&zc
);
564 spa_scan_stat_init(spa
);
569 dsl_scan_fini(dsl_pool_t
*dp
)
571 if (dp
->dp_scan
!= NULL
) {
572 dsl_scan_t
*scn
= dp
->dp_scan
;
574 if (scn
->scn_taskq
!= NULL
)
575 taskq_destroy(scn
->scn_taskq
);
577 scan_ds_queue_clear(scn
);
578 avl_destroy(&scn
->scn_queue
);
579 scan_ds_prefetch_queue_clear(scn
);
580 avl_destroy(&scn
->scn_prefetch_queue
);
582 kmem_free(dp
->dp_scan
, sizeof (dsl_scan_t
));
588 dsl_scan_restarting(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
590 return (scn
->scn_restart_txg
!= 0 &&
591 scn
->scn_restart_txg
<= tx
->tx_txg
);
595 dsl_scan_scrubbing(const dsl_pool_t
*dp
)
597 dsl_scan_phys_t
*scn_phys
= &dp
->dp_scan
->scn_phys
;
599 return (scn_phys
->scn_state
== DSS_SCANNING
&&
600 scn_phys
->scn_func
== POOL_SCAN_SCRUB
);
604 dsl_scan_is_paused_scrub(const dsl_scan_t
*scn
)
606 return (dsl_scan_scrubbing(scn
->scn_dp
) &&
607 scn
->scn_phys
.scn_flags
& DSF_SCRUB_PAUSED
);
611 * Writes out a persistent dsl_scan_phys_t record to the pool directory.
612 * Because we can be running in the block sorting algorithm, we do not always
613 * want to write out the record, only when it is "safe" to do so. This safety
614 * condition is achieved by making sure that the sorting queues are empty
615 * (scn_bytes_pending == 0). When this condition is not true, the sync'd state
616 * is inconsistent with how much actual scanning progress has been made. The
617 * kind of sync to be performed is specified by the sync_type argument. If the
618 * sync is optional, we only sync if the queues are empty. If the sync is
619 * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
620 * third possible state is a "cached" sync. This is done in response to:
621 * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
622 * destroyed, so we wouldn't be able to restart scanning from it.
623 * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
624 * superseded by a newer snapshot.
625 * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
626 * swapped with its clone.
627 * In all cases, a cached sync simply rewrites the last record we've written,
628 * just slightly modified. For the modifications that are performed to the
629 * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
630 * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
633 dsl_scan_sync_state(dsl_scan_t
*scn
, dmu_tx_t
*tx
, state_sync_type_t sync_type
)
636 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
638 ASSERT(sync_type
!= SYNC_MANDATORY
|| scn
->scn_bytes_pending
== 0);
639 if (scn
->scn_bytes_pending
== 0) {
640 for (i
= 0; i
< spa
->spa_root_vdev
->vdev_children
; i
++) {
641 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[i
];
642 dsl_scan_io_queue_t
*q
= vd
->vdev_scan_io_queue
;
647 mutex_enter(&vd
->vdev_scan_io_queue_lock
);
648 ASSERT3P(avl_first(&q
->q_sios_by_addr
), ==, NULL
);
649 ASSERT3P(avl_first(&q
->q_exts_by_size
), ==, NULL
);
650 ASSERT3P(range_tree_first(q
->q_exts_by_addr
), ==, NULL
);
651 mutex_exit(&vd
->vdev_scan_io_queue_lock
);
654 if (scn
->scn_phys
.scn_queue_obj
!= 0)
655 scan_ds_queue_sync(scn
, tx
);
656 VERIFY0(zap_update(scn
->scn_dp
->dp_meta_objset
,
657 DMU_POOL_DIRECTORY_OBJECT
,
658 DMU_POOL_SCAN
, sizeof (uint64_t), SCAN_PHYS_NUMINTS
,
659 &scn
->scn_phys
, tx
));
660 bcopy(&scn
->scn_phys
, &scn
->scn_phys_cached
,
661 sizeof (scn
->scn_phys
));
663 if (scn
->scn_checkpointing
)
664 zfs_dbgmsg("finish scan checkpoint");
666 scn
->scn_checkpointing
= B_FALSE
;
667 scn
->scn_last_checkpoint
= ddi_get_lbolt();
668 } else if (sync_type
== SYNC_CACHED
) {
669 VERIFY0(zap_update(scn
->scn_dp
->dp_meta_objset
,
670 DMU_POOL_DIRECTORY_OBJECT
,
671 DMU_POOL_SCAN
, sizeof (uint64_t), SCAN_PHYS_NUMINTS
,
672 &scn
->scn_phys_cached
, tx
));
678 dsl_scan_setup_check(void *arg
, dmu_tx_t
*tx
)
680 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
682 if (dsl_scan_is_running(scn
))
683 return (SET_ERROR(EBUSY
));
689 dsl_scan_setup_sync(void *arg
, dmu_tx_t
*tx
)
691 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
692 pool_scan_func_t
*funcp
= arg
;
693 dmu_object_type_t ot
= 0;
694 dsl_pool_t
*dp
= scn
->scn_dp
;
695 spa_t
*spa
= dp
->dp_spa
;
697 ASSERT(!dsl_scan_is_running(scn
));
698 ASSERT(*funcp
> POOL_SCAN_NONE
&& *funcp
< POOL_SCAN_FUNCS
);
699 bzero(&scn
->scn_phys
, sizeof (scn
->scn_phys
));
700 scn
->scn_phys
.scn_func
= *funcp
;
701 scn
->scn_phys
.scn_state
= DSS_SCANNING
;
702 scn
->scn_phys
.scn_min_txg
= 0;
703 scn
->scn_phys
.scn_max_txg
= tx
->tx_txg
;
704 scn
->scn_phys
.scn_ddt_class_max
= DDT_CLASSES
- 1; /* the entire DDT */
705 scn
->scn_phys
.scn_start_time
= gethrestime_sec();
706 scn
->scn_phys
.scn_errors
= 0;
707 scn
->scn_phys
.scn_to_examine
= spa
->spa_root_vdev
->vdev_stat
.vs_alloc
;
708 scn
->scn_issued_before_pass
= 0;
709 scn
->scn_restart_txg
= 0;
710 scn
->scn_done_txg
= 0;
711 scn
->scn_last_checkpoint
= 0;
712 scn
->scn_checkpointing
= B_FALSE
;
713 spa_scan_stat_init(spa
);
715 if (DSL_SCAN_IS_SCRUB_RESILVER(scn
)) {
716 scn
->scn_phys
.scn_ddt_class_max
= zfs_scrub_ddt_class_max
;
718 /* rewrite all disk labels */
719 vdev_config_dirty(spa
->spa_root_vdev
);
721 if (vdev_resilver_needed(spa
->spa_root_vdev
,
722 &scn
->scn_phys
.scn_min_txg
, &scn
->scn_phys
.scn_max_txg
)) {
723 spa_event_notify(spa
, NULL
, NULL
,
724 ESC_ZFS_RESILVER_START
);
726 spa_event_notify(spa
, NULL
, NULL
, ESC_ZFS_SCRUB_START
);
729 spa
->spa_scrub_started
= B_TRUE
;
731 * If this is an incremental scrub, limit the DDT scrub phase
732 * to just the auto-ditto class (for correctness); the rest
733 * of the scrub should go faster using top-down pruning.
735 if (scn
->scn_phys
.scn_min_txg
> TXG_INITIAL
)
736 scn
->scn_phys
.scn_ddt_class_max
= DDT_CLASS_DITTO
;
740 /* back to the generic stuff */
742 if (dp
->dp_blkstats
== NULL
) {
744 vmem_alloc(sizeof (zfs_all_blkstats_t
), KM_SLEEP
);
745 mutex_init(&dp
->dp_blkstats
->zab_lock
, NULL
,
746 MUTEX_DEFAULT
, NULL
);
748 bzero(&dp
->dp_blkstats
->zab_type
, sizeof (dp
->dp_blkstats
->zab_type
));
750 if (spa_version(spa
) < SPA_VERSION_DSL_SCRUB
)
751 ot
= DMU_OT_ZAP_OTHER
;
753 scn
->scn_phys
.scn_queue_obj
= zap_create(dp
->dp_meta_objset
,
754 ot
? ot
: DMU_OT_SCAN_QUEUE
, DMU_OT_NONE
, 0, tx
);
756 bcopy(&scn
->scn_phys
, &scn
->scn_phys_cached
, sizeof (scn
->scn_phys
));
758 dsl_scan_sync_state(scn
, tx
, SYNC_MANDATORY
);
760 spa_history_log_internal(spa
, "scan setup", tx
,
761 "func=%u mintxg=%llu maxtxg=%llu",
762 *funcp
, scn
->scn_phys
.scn_min_txg
, scn
->scn_phys
.scn_max_txg
);
766 * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
767 * Can also be called to resume a paused scrub.
770 dsl_scan(dsl_pool_t
*dp
, pool_scan_func_t func
)
772 spa_t
*spa
= dp
->dp_spa
;
773 dsl_scan_t
*scn
= dp
->dp_scan
;
776 * Purge all vdev caches and probe all devices. We do this here
777 * rather than in sync context because this requires a writer lock
778 * on the spa_config lock, which we can't do from sync context. The
779 * spa_scrub_reopen flag indicates that vdev_open() should not
780 * attempt to start another scrub.
782 spa_vdev_state_enter(spa
, SCL_NONE
);
783 spa
->spa_scrub_reopen
= B_TRUE
;
784 vdev_reopen(spa
->spa_root_vdev
);
785 spa
->spa_scrub_reopen
= B_FALSE
;
786 (void) spa_vdev_state_exit(spa
, NULL
, 0);
788 if (func
== POOL_SCAN_RESILVER
) {
789 dsl_resilver_restart(spa
->spa_dsl_pool
, 0);
793 if (func
== POOL_SCAN_SCRUB
&& dsl_scan_is_paused_scrub(scn
)) {
794 /* got scrub start cmd, resume paused scrub */
795 int err
= dsl_scrub_set_pause_resume(scn
->scn_dp
,
798 spa_event_notify(spa
, NULL
, NULL
, ESC_ZFS_SCRUB_RESUME
);
802 return (SET_ERROR(err
));
805 return (dsl_sync_task(spa_name(spa
), dsl_scan_setup_check
,
806 dsl_scan_setup_sync
, &func
, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED
));
810 * Sets the resilver defer flag to B_FALSE on all leaf devs under vd. Returns
811 * B_TRUE if we have devices that need to be resilvered and are available to
812 * accept resilver I/Os.
815 dsl_scan_clear_deferred(vdev_t
*vd
, dmu_tx_t
*tx
)
817 boolean_t resilver_needed
= B_FALSE
;
818 spa_t
*spa
= vd
->vdev_spa
;
820 for (int c
= 0; c
< vd
->vdev_children
; c
++) {
822 dsl_scan_clear_deferred(vd
->vdev_child
[c
], tx
);
825 if (vd
== spa
->spa_root_vdev
&&
826 spa_feature_is_active(spa
, SPA_FEATURE_RESILVER_DEFER
)) {
827 spa_feature_decr(spa
, SPA_FEATURE_RESILVER_DEFER
, tx
);
828 vdev_config_dirty(vd
);
829 spa
->spa_resilver_deferred
= B_FALSE
;
830 return (resilver_needed
);
833 if (!vdev_is_concrete(vd
) || vd
->vdev_aux
||
834 !vd
->vdev_ops
->vdev_op_leaf
)
835 return (resilver_needed
);
837 if (vd
->vdev_resilver_deferred
)
838 vd
->vdev_resilver_deferred
= B_FALSE
;
840 return (!vdev_is_dead(vd
) && !vd
->vdev_offline
&&
841 vdev_resilver_needed(vd
, NULL
, NULL
));
846 dsl_scan_done(dsl_scan_t
*scn
, boolean_t complete
, dmu_tx_t
*tx
)
848 static const char *old_names
[] = {
850 "scrub_ddt_bookmark",
851 "scrub_ddt_class_max",
860 dsl_pool_t
*dp
= scn
->scn_dp
;
861 spa_t
*spa
= dp
->dp_spa
;
864 /* Remove any remnants of an old-style scrub. */
865 for (i
= 0; old_names
[i
]; i
++) {
866 (void) zap_remove(dp
->dp_meta_objset
,
867 DMU_POOL_DIRECTORY_OBJECT
, old_names
[i
], tx
);
870 if (scn
->scn_phys
.scn_queue_obj
!= 0) {
871 VERIFY0(dmu_object_free(dp
->dp_meta_objset
,
872 scn
->scn_phys
.scn_queue_obj
, tx
));
873 scn
->scn_phys
.scn_queue_obj
= 0;
875 scan_ds_queue_clear(scn
);
876 scan_ds_prefetch_queue_clear(scn
);
878 scn
->scn_phys
.scn_flags
&= ~DSF_SCRUB_PAUSED
;
881 * If we were "restarted" from a stopped state, don't bother
882 * with anything else.
884 if (!dsl_scan_is_running(scn
)) {
885 ASSERT(!scn
->scn_is_sorted
);
889 if (scn
->scn_is_sorted
) {
890 scan_io_queues_destroy(scn
);
891 scn
->scn_is_sorted
= B_FALSE
;
893 if (scn
->scn_taskq
!= NULL
) {
894 taskq_destroy(scn
->scn_taskq
);
895 scn
->scn_taskq
= NULL
;
899 scn
->scn_phys
.scn_state
= complete
? DSS_FINISHED
: DSS_CANCELED
;
901 if (dsl_scan_restarting(scn
, tx
))
902 spa_history_log_internal(spa
, "scan aborted, restarting", tx
,
903 "errors=%llu", spa_get_errlog_size(spa
));
905 spa_history_log_internal(spa
, "scan cancelled", tx
,
906 "errors=%llu", spa_get_errlog_size(spa
));
908 spa_history_log_internal(spa
, "scan done", tx
,
909 "errors=%llu", spa_get_errlog_size(spa
));
911 if (DSL_SCAN_IS_SCRUB_RESILVER(scn
)) {
912 spa
->spa_scrub_started
= B_FALSE
;
913 spa
->spa_scrub_active
= B_FALSE
;
916 * If the scrub/resilver completed, update all DTLs to
917 * reflect this. Whether it succeeded or not, vacate
918 * all temporary scrub DTLs.
920 * As the scrub does not currently support traversing
921 * data that have been freed but are part of a checkpoint,
922 * we don't mark the scrub as done in the DTLs as faults
923 * may still exist in those vdevs.
926 !spa_feature_is_active(spa
, SPA_FEATURE_POOL_CHECKPOINT
)) {
927 vdev_dtl_reassess(spa
->spa_root_vdev
, tx
->tx_txg
,
928 scn
->scn_phys
.scn_max_txg
, B_TRUE
);
930 spa_event_notify(spa
, NULL
, NULL
,
931 scn
->scn_phys
.scn_min_txg
?
932 ESC_ZFS_RESILVER_FINISH
: ESC_ZFS_SCRUB_FINISH
);
934 vdev_dtl_reassess(spa
->spa_root_vdev
, tx
->tx_txg
,
937 spa_errlog_rotate(spa
);
940 * We may have finished replacing a device.
941 * Let the async thread assess this and handle the detach.
943 spa_async_request(spa
, SPA_ASYNC_RESILVER_DONE
);
946 * Clear any deferred_resilver flags in the config.
947 * If there are drives that need resilvering, kick
948 * off an asynchronous request to start resilver.
949 * dsl_scan_clear_deferred() may update the config
950 * before the resilver can restart. In the event of
951 * a crash during this period, the spa loading code
952 * will find the drives that need to be resilvered
953 * when the machine reboots and start the resilver then.
955 boolean_t resilver_needed
=
956 dsl_scan_clear_deferred(spa
->spa_root_vdev
, tx
);
957 if (resilver_needed
) {
958 spa_history_log_internal(spa
,
959 "starting deferred resilver", tx
,
960 "errors=%llu", spa_get_errlog_size(spa
));
961 spa_async_request(spa
, SPA_ASYNC_RESILVER
);
965 scn
->scn_phys
.scn_end_time
= gethrestime_sec();
967 if (spa
->spa_errata
== ZPOOL_ERRATA_ZOL_2094_SCRUB
)
970 ASSERT(!dsl_scan_is_running(scn
));
975 dsl_scan_cancel_check(void *arg
, dmu_tx_t
*tx
)
977 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
979 if (!dsl_scan_is_running(scn
))
980 return (SET_ERROR(ENOENT
));
986 dsl_scan_cancel_sync(void *arg
, dmu_tx_t
*tx
)
988 dsl_scan_t
*scn
= dmu_tx_pool(tx
)->dp_scan
;
990 dsl_scan_done(scn
, B_FALSE
, tx
);
991 dsl_scan_sync_state(scn
, tx
, SYNC_MANDATORY
);
992 spa_event_notify(scn
->scn_dp
->dp_spa
, NULL
, NULL
, ESC_ZFS_SCRUB_ABORT
);
996 dsl_scan_cancel(dsl_pool_t
*dp
)
998 return (dsl_sync_task(spa_name(dp
->dp_spa
), dsl_scan_cancel_check
,
999 dsl_scan_cancel_sync
, NULL
, 3, ZFS_SPACE_CHECK_RESERVED
));
1003 dsl_scrub_pause_resume_check(void *arg
, dmu_tx_t
*tx
)
1005 pool_scrub_cmd_t
*cmd
= arg
;
1006 dsl_pool_t
*dp
= dmu_tx_pool(tx
);
1007 dsl_scan_t
*scn
= dp
->dp_scan
;
1009 if (*cmd
== POOL_SCRUB_PAUSE
) {
1010 /* can't pause a scrub when there is no in-progress scrub */
1011 if (!dsl_scan_scrubbing(dp
))
1012 return (SET_ERROR(ENOENT
));
1014 /* can't pause a paused scrub */
1015 if (dsl_scan_is_paused_scrub(scn
))
1016 return (SET_ERROR(EBUSY
));
1017 } else if (*cmd
!= POOL_SCRUB_NORMAL
) {
1018 return (SET_ERROR(ENOTSUP
));
1025 dsl_scrub_pause_resume_sync(void *arg
, dmu_tx_t
*tx
)
1027 pool_scrub_cmd_t
*cmd
= arg
;
1028 dsl_pool_t
*dp
= dmu_tx_pool(tx
);
1029 spa_t
*spa
= dp
->dp_spa
;
1030 dsl_scan_t
*scn
= dp
->dp_scan
;
1032 if (*cmd
== POOL_SCRUB_PAUSE
) {
1033 /* can't pause a scrub when there is no in-progress scrub */
1034 spa
->spa_scan_pass_scrub_pause
= gethrestime_sec();
1035 scn
->scn_phys
.scn_flags
|= DSF_SCRUB_PAUSED
;
1036 scn
->scn_phys_cached
.scn_flags
|= DSF_SCRUB_PAUSED
;
1037 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
1038 spa_event_notify(spa
, NULL
, NULL
, ESC_ZFS_SCRUB_PAUSED
);
1040 ASSERT3U(*cmd
, ==, POOL_SCRUB_NORMAL
);
1041 if (dsl_scan_is_paused_scrub(scn
)) {
1043 * We need to keep track of how much time we spend
1044 * paused per pass so that we can adjust the scrub rate
1045 * shown in the output of 'zpool status'
1047 spa
->spa_scan_pass_scrub_spent_paused
+=
1048 gethrestime_sec() - spa
->spa_scan_pass_scrub_pause
;
1049 spa
->spa_scan_pass_scrub_pause
= 0;
1050 scn
->scn_phys
.scn_flags
&= ~DSF_SCRUB_PAUSED
;
1051 scn
->scn_phys_cached
.scn_flags
&= ~DSF_SCRUB_PAUSED
;
1052 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
1058 * Set scrub pause/resume state if it makes sense to do so
1061 dsl_scrub_set_pause_resume(const dsl_pool_t
*dp
, pool_scrub_cmd_t cmd
)
1063 return (dsl_sync_task(spa_name(dp
->dp_spa
),
1064 dsl_scrub_pause_resume_check
, dsl_scrub_pause_resume_sync
, &cmd
, 3,
1065 ZFS_SPACE_CHECK_RESERVED
));
1069 /* start a new scan, or restart an existing one. */
1071 dsl_resilver_restart(dsl_pool_t
*dp
, uint64_t txg
)
1075 tx
= dmu_tx_create_dd(dp
->dp_mos_dir
);
1076 VERIFY(0 == dmu_tx_assign(tx
, TXG_WAIT
));
1078 txg
= dmu_tx_get_txg(tx
);
1079 dp
->dp_scan
->scn_restart_txg
= txg
;
1082 dp
->dp_scan
->scn_restart_txg
= txg
;
1084 zfs_dbgmsg("restarting resilver txg=%llu", (longlong_t
)txg
);
1088 dsl_free(dsl_pool_t
*dp
, uint64_t txg
, const blkptr_t
*bp
)
1090 zio_free(dp
->dp_spa
, txg
, bp
);
1094 dsl_free_sync(zio_t
*pio
, dsl_pool_t
*dp
, uint64_t txg
, const blkptr_t
*bpp
)
1096 ASSERT(dsl_pool_sync_context(dp
));
1097 zio_nowait(zio_free_sync(pio
, dp
->dp_spa
, txg
, bpp
, pio
->io_flags
));
1101 scan_ds_queue_compare(const void *a
, const void *b
)
1103 const scan_ds_t
*sds_a
= a
, *sds_b
= b
;
1105 if (sds_a
->sds_dsobj
< sds_b
->sds_dsobj
)
1107 if (sds_a
->sds_dsobj
== sds_b
->sds_dsobj
)
1113 scan_ds_queue_clear(dsl_scan_t
*scn
)
1115 void *cookie
= NULL
;
1117 while ((sds
= avl_destroy_nodes(&scn
->scn_queue
, &cookie
)) != NULL
) {
1118 kmem_free(sds
, sizeof (*sds
));
1123 scan_ds_queue_contains(dsl_scan_t
*scn
, uint64_t dsobj
, uint64_t *txg
)
1125 scan_ds_t srch
, *sds
;
1127 srch
.sds_dsobj
= dsobj
;
1128 sds
= avl_find(&scn
->scn_queue
, &srch
, NULL
);
1129 if (sds
!= NULL
&& txg
!= NULL
)
1130 *txg
= sds
->sds_txg
;
1131 return (sds
!= NULL
);
1135 scan_ds_queue_insert(dsl_scan_t
*scn
, uint64_t dsobj
, uint64_t txg
)
1140 sds
= kmem_zalloc(sizeof (*sds
), KM_SLEEP
);
1141 sds
->sds_dsobj
= dsobj
;
1144 VERIFY3P(avl_find(&scn
->scn_queue
, sds
, &where
), ==, NULL
);
1145 avl_insert(&scn
->scn_queue
, sds
, where
);
1149 scan_ds_queue_remove(dsl_scan_t
*scn
, uint64_t dsobj
)
1151 scan_ds_t srch
, *sds
;
1153 srch
.sds_dsobj
= dsobj
;
1155 sds
= avl_find(&scn
->scn_queue
, &srch
, NULL
);
1156 VERIFY(sds
!= NULL
);
1157 avl_remove(&scn
->scn_queue
, sds
);
1158 kmem_free(sds
, sizeof (*sds
));
1162 scan_ds_queue_sync(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
1164 dsl_pool_t
*dp
= scn
->scn_dp
;
1165 spa_t
*spa
= dp
->dp_spa
;
1166 dmu_object_type_t ot
= (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
) ?
1167 DMU_OT_SCAN_QUEUE
: DMU_OT_ZAP_OTHER
;
1169 ASSERT0(scn
->scn_bytes_pending
);
1170 ASSERT(scn
->scn_phys
.scn_queue_obj
!= 0);
1172 VERIFY0(dmu_object_free(dp
->dp_meta_objset
,
1173 scn
->scn_phys
.scn_queue_obj
, tx
));
1174 scn
->scn_phys
.scn_queue_obj
= zap_create(dp
->dp_meta_objset
, ot
,
1175 DMU_OT_NONE
, 0, tx
);
1176 for (scan_ds_t
*sds
= avl_first(&scn
->scn_queue
);
1177 sds
!= NULL
; sds
= AVL_NEXT(&scn
->scn_queue
, sds
)) {
1178 VERIFY0(zap_add_int_key(dp
->dp_meta_objset
,
1179 scn
->scn_phys
.scn_queue_obj
, sds
->sds_dsobj
,
1185 * Computes the memory limit state that we're currently in. A sorted scan
1186 * needs quite a bit of memory to hold the sorting queue, so we need to
1187 * reasonably constrain the size so it doesn't impact overall system
1188 * performance. We compute two limits:
1189 * 1) Hard memory limit: if the amount of memory used by the sorting
1190 * queues on a pool gets above this value, we stop the metadata
1191 * scanning portion and start issuing the queued up and sorted
1192 * I/Os to reduce memory usage.
1193 * This limit is calculated as a fraction of physmem (by default 5%).
1194 * We constrain the lower bound of the hard limit to an absolute
1195 * minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1196 * the upper bound to 5% of the total pool size - no chance we'll
1197 * ever need that much memory, but just to keep the value in check.
1198 * 2) Soft memory limit: once we hit the hard memory limit, we start
1199 * issuing I/O to reduce queue memory usage, but we don't want to
1200 * completely empty out the queues, since we might be able to find I/Os
1201 * that will fill in the gaps of our non-sequential IOs at some point
1202 * in the future. So we stop the issuing of I/Os once the amount of
1203 * memory used drops below the soft limit (at which point we stop issuing
1204 * I/O and start scanning metadata again).
1206 * This limit is calculated by subtracting a fraction of the hard
1207 * limit from the hard limit. By default this fraction is 5%, so
1208 * the soft limit is 95% of the hard limit. We cap the size of the
1209 * difference between the hard and soft limits at an absolute
1210 * maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1211 * sufficient to not cause too frequent switching between the
1212 * metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1213 * worth of queues is about 1.2 GiB of on-pool data, so scanning
1214 * that should take at least a decent fraction of a second).
1217 dsl_scan_should_clear(dsl_scan_t
*scn
)
1219 vdev_t
*rvd
= scn
->scn_dp
->dp_spa
->spa_root_vdev
;
1220 uint64_t mlim_hard
, mlim_soft
, mused
;
1221 uint64_t alloc
= metaslab_class_get_alloc(spa_normal_class(
1222 scn
->scn_dp
->dp_spa
));
1224 mlim_hard
= MAX((physmem
/ zfs_scan_mem_lim_fact
) * PAGESIZE
,
1225 zfs_scan_mem_lim_min
);
1226 mlim_hard
= MIN(mlim_hard
, alloc
/ 20);
1227 mlim_soft
= mlim_hard
- MIN(mlim_hard
/ zfs_scan_mem_lim_soft_fact
,
1228 zfs_scan_mem_lim_soft_max
);
1230 for (uint64_t i
= 0; i
< rvd
->vdev_children
; i
++) {
1231 vdev_t
*tvd
= rvd
->vdev_child
[i
];
1232 dsl_scan_io_queue_t
*queue
;
1234 mutex_enter(&tvd
->vdev_scan_io_queue_lock
);
1235 queue
= tvd
->vdev_scan_io_queue
;
1236 if (queue
!= NULL
) {
1237 /* # extents in exts_by_size = # in exts_by_addr */
1238 mused
+= avl_numnodes(&queue
->q_exts_by_size
) *
1239 sizeof (range_seg_t
) + queue
->q_sio_memused
;
1241 mutex_exit(&tvd
->vdev_scan_io_queue_lock
);
1244 dprintf("current scan memory usage: %llu bytes\n", (longlong_t
)mused
);
1247 ASSERT0(scn
->scn_bytes_pending
);
1250 * If we are above our hard limit, we need to clear out memory.
1251 * If we are below our soft limit, we need to accumulate sequential IOs.
1252 * Otherwise, we should keep doing whatever we are currently doing.
1254 if (mused
>= mlim_hard
)
1256 else if (mused
< mlim_soft
)
1259 return (scn
->scn_clearing
);
1263 dsl_scan_check_suspend(dsl_scan_t
*scn
, const zbookmark_phys_t
*zb
)
1265 /* we never skip user/group accounting objects */
1266 if (zb
&& (int64_t)zb
->zb_object
< 0)
1269 if (scn
->scn_suspending
)
1270 return (B_TRUE
); /* we're already suspending */
1272 if (!ZB_IS_ZERO(&scn
->scn_phys
.scn_bookmark
))
1273 return (B_FALSE
); /* we're resuming */
1275 /* We only know how to resume from level-0 blocks. */
1276 if (zb
&& zb
->zb_level
!= 0)
1281 * - we have scanned for at least the minimum time (default 1 sec
1282 * for scrub, 3 sec for resilver), and either we have sufficient
1283 * dirty data that we are starting to write more quickly
1284 * (default 30%), someone is explicitly waiting for this txg
1285 * to complete, or we have used up all of the time in the txg
1286 * timeout (default 5 sec).
1288 * - the spa is shutting down because this pool is being exported
1289 * or the machine is rebooting.
1291 * - the scan queue has reached its memory use limit
1293 uint64_t curr_time_ns
= gethrtime();
1294 uint64_t scan_time_ns
= curr_time_ns
- scn
->scn_sync_start_time
;
1295 uint64_t sync_time_ns
= curr_time_ns
-
1296 scn
->scn_dp
->dp_spa
->spa_sync_starttime
;
1297 int dirty_pct
= scn
->scn_dp
->dp_dirty_total
* 100 / zfs_dirty_data_max
;
1298 int mintime
= (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) ?
1299 zfs_resilver_min_time_ms
: zfs_scrub_min_time_ms
;
1301 if ((NSEC2MSEC(scan_time_ns
) > mintime
&&
1302 (dirty_pct
>= zfs_vdev_async_write_active_min_dirty_percent
||
1303 txg_sync_waiting(scn
->scn_dp
) ||
1304 NSEC2SEC(sync_time_ns
) >= zfs_txg_timeout
)) ||
1305 spa_shutting_down(scn
->scn_dp
->dp_spa
) ||
1306 (zfs_scan_strict_mem_lim
&& dsl_scan_should_clear(scn
))) {
1308 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1309 (longlong_t
)zb
->zb_objset
,
1310 (longlong_t
)zb
->zb_object
,
1311 (longlong_t
)zb
->zb_level
,
1312 (longlong_t
)zb
->zb_blkid
);
1313 scn
->scn_phys
.scn_bookmark
= *zb
;
1316 dsl_scan_phys_t
*scnp
= &scn
->scn_phys
;
1317 dprintf("suspending at at DDT bookmark "
1318 "%llx/%llx/%llx/%llx\n",
1319 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_class
,
1320 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_type
,
1321 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_checksum
,
1322 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_cursor
);
1325 scn
->scn_suspending
= B_TRUE
;
1331 typedef struct zil_scan_arg
{
1333 zil_header_t
*zsa_zh
;
1338 dsl_scan_zil_block(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
1340 zil_scan_arg_t
*zsa
= arg
;
1341 dsl_pool_t
*dp
= zsa
->zsa_dp
;
1342 dsl_scan_t
*scn
= dp
->dp_scan
;
1343 zil_header_t
*zh
= zsa
->zsa_zh
;
1344 zbookmark_phys_t zb
;
1346 ASSERT(!BP_IS_REDACTED(bp
));
1347 if (BP_IS_HOLE(bp
) || bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
)
1351 * One block ("stubby") can be allocated a long time ago; we
1352 * want to visit that one because it has been allocated
1353 * (on-disk) even if it hasn't been claimed (even though for
1354 * scrub there's nothing to do to it).
1356 if (claim_txg
== 0 && bp
->blk_birth
>= spa_min_claim_txg(dp
->dp_spa
))
1359 SET_BOOKMARK(&zb
, zh
->zh_log
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1360 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
1362 VERIFY(0 == scan_funcs
[scn
->scn_phys
.scn_func
](dp
, bp
, &zb
));
1368 dsl_scan_zil_record(zilog_t
*zilog
, lr_t
*lrc
, void *arg
, uint64_t claim_txg
)
1370 if (lrc
->lrc_txtype
== TX_WRITE
) {
1371 zil_scan_arg_t
*zsa
= arg
;
1372 dsl_pool_t
*dp
= zsa
->zsa_dp
;
1373 dsl_scan_t
*scn
= dp
->dp_scan
;
1374 zil_header_t
*zh
= zsa
->zsa_zh
;
1375 lr_write_t
*lr
= (lr_write_t
*)lrc
;
1376 blkptr_t
*bp
= &lr
->lr_blkptr
;
1377 zbookmark_phys_t zb
;
1379 ASSERT(!BP_IS_REDACTED(bp
));
1380 if (BP_IS_HOLE(bp
) ||
1381 bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
)
1385 * birth can be < claim_txg if this record's txg is
1386 * already txg sync'ed (but this log block contains
1387 * other records that are not synced)
1389 if (claim_txg
== 0 || bp
->blk_birth
< claim_txg
)
1392 SET_BOOKMARK(&zb
, zh
->zh_log
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1393 lr
->lr_foid
, ZB_ZIL_LEVEL
,
1394 lr
->lr_offset
/ BP_GET_LSIZE(bp
));
1396 VERIFY(0 == scan_funcs
[scn
->scn_phys
.scn_func
](dp
, bp
, &zb
));
1402 dsl_scan_zil(dsl_pool_t
*dp
, zil_header_t
*zh
)
1404 uint64_t claim_txg
= zh
->zh_claim_txg
;
1405 zil_scan_arg_t zsa
= { dp
, zh
};
1408 ASSERT(spa_writeable(dp
->dp_spa
));
1411 * We only want to visit blocks that have been claimed but not yet
1412 * replayed (or, in read-only mode, blocks that *would* be claimed).
1417 zilog
= zil_alloc(dp
->dp_meta_objset
, zh
);
1419 (void) zil_parse(zilog
, dsl_scan_zil_block
, dsl_scan_zil_record
, &zsa
,
1420 claim_txg
, B_FALSE
);
1426 * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1427 * here is to sort the AVL tree by the order each block will be needed.
1430 scan_prefetch_queue_compare(const void *a
, const void *b
)
1432 const scan_prefetch_issue_ctx_t
*spic_a
= a
, *spic_b
= b
;
1433 const scan_prefetch_ctx_t
*spc_a
= spic_a
->spic_spc
;
1434 const scan_prefetch_ctx_t
*spc_b
= spic_b
->spic_spc
;
1436 return (zbookmark_compare(spc_a
->spc_datablkszsec
,
1437 spc_a
->spc_indblkshift
, spc_b
->spc_datablkszsec
,
1438 spc_b
->spc_indblkshift
, &spic_a
->spic_zb
, &spic_b
->spic_zb
));
1442 scan_prefetch_ctx_rele(scan_prefetch_ctx_t
*spc
, void *tag
)
1444 if (zfs_refcount_remove(&spc
->spc_refcnt
, tag
) == 0) {
1445 zfs_refcount_destroy(&spc
->spc_refcnt
);
1446 kmem_free(spc
, sizeof (scan_prefetch_ctx_t
));
1450 static scan_prefetch_ctx_t
*
1451 scan_prefetch_ctx_create(dsl_scan_t
*scn
, dnode_phys_t
*dnp
, void *tag
)
1453 scan_prefetch_ctx_t
*spc
;
1455 spc
= kmem_alloc(sizeof (scan_prefetch_ctx_t
), KM_SLEEP
);
1456 zfs_refcount_create(&spc
->spc_refcnt
);
1457 zfs_refcount_add(&spc
->spc_refcnt
, tag
);
1460 spc
->spc_datablkszsec
= dnp
->dn_datablkszsec
;
1461 spc
->spc_indblkshift
= dnp
->dn_indblkshift
;
1462 spc
->spc_root
= B_FALSE
;
1464 spc
->spc_datablkszsec
= 0;
1465 spc
->spc_indblkshift
= 0;
1466 spc
->spc_root
= B_TRUE
;
1473 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t
*spc
, void *tag
)
1475 zfs_refcount_add(&spc
->spc_refcnt
, tag
);
1479 scan_ds_prefetch_queue_clear(dsl_scan_t
*scn
)
1481 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1482 void *cookie
= NULL
;
1483 scan_prefetch_issue_ctx_t
*spic
= NULL
;
1485 mutex_enter(&spa
->spa_scrub_lock
);
1486 while ((spic
= avl_destroy_nodes(&scn
->scn_prefetch_queue
,
1487 &cookie
)) != NULL
) {
1488 scan_prefetch_ctx_rele(spic
->spic_spc
, scn
);
1489 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1491 mutex_exit(&spa
->spa_scrub_lock
);
1495 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t
*spc
,
1496 const zbookmark_phys_t
*zb
)
1498 zbookmark_phys_t
*last_zb
= &spc
->spc_scn
->scn_prefetch_bookmark
;
1499 dnode_phys_t tmp_dnp
;
1500 dnode_phys_t
*dnp
= (spc
->spc_root
) ? NULL
: &tmp_dnp
;
1502 if (zb
->zb_objset
!= last_zb
->zb_objset
)
1504 if ((int64_t)zb
->zb_object
< 0)
1507 tmp_dnp
.dn_datablkszsec
= spc
->spc_datablkszsec
;
1508 tmp_dnp
.dn_indblkshift
= spc
->spc_indblkshift
;
1510 if (zbookmark_subtree_completed(dnp
, zb
, last_zb
))
1517 dsl_scan_prefetch(scan_prefetch_ctx_t
*spc
, blkptr_t
*bp
, zbookmark_phys_t
*zb
)
1520 dsl_scan_t
*scn
= spc
->spc_scn
;
1521 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1522 scan_prefetch_issue_ctx_t
*spic
;
1524 if (zfs_no_scrub_prefetch
|| BP_IS_REDACTED(bp
))
1527 if (BP_IS_HOLE(bp
) || bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
||
1528 (BP_GET_LEVEL(bp
) == 0 && BP_GET_TYPE(bp
) != DMU_OT_DNODE
&&
1529 BP_GET_TYPE(bp
) != DMU_OT_OBJSET
))
1532 if (dsl_scan_check_prefetch_resume(spc
, zb
))
1535 scan_prefetch_ctx_add_ref(spc
, scn
);
1536 spic
= kmem_alloc(sizeof (scan_prefetch_issue_ctx_t
), KM_SLEEP
);
1537 spic
->spic_spc
= spc
;
1538 spic
->spic_bp
= *bp
;
1539 spic
->spic_zb
= *zb
;
1542 * Add the IO to the queue of blocks to prefetch. This allows us to
1543 * prioritize blocks that we will need first for the main traversal
1546 mutex_enter(&spa
->spa_scrub_lock
);
1547 if (avl_find(&scn
->scn_prefetch_queue
, spic
, &idx
) != NULL
) {
1548 /* this block is already queued for prefetch */
1549 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1550 scan_prefetch_ctx_rele(spc
, scn
);
1551 mutex_exit(&spa
->spa_scrub_lock
);
1555 avl_insert(&scn
->scn_prefetch_queue
, spic
, idx
);
1556 cv_broadcast(&spa
->spa_scrub_io_cv
);
1557 mutex_exit(&spa
->spa_scrub_lock
);
1561 dsl_scan_prefetch_dnode(dsl_scan_t
*scn
, dnode_phys_t
*dnp
,
1562 uint64_t objset
, uint64_t object
)
1565 zbookmark_phys_t zb
;
1566 scan_prefetch_ctx_t
*spc
;
1568 if (dnp
->dn_nblkptr
== 0 && !(dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1571 SET_BOOKMARK(&zb
, objset
, object
, 0, 0);
1573 spc
= scan_prefetch_ctx_create(scn
, dnp
, FTAG
);
1575 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++) {
1576 zb
.zb_level
= BP_GET_LEVEL(&dnp
->dn_blkptr
[i
]);
1578 dsl_scan_prefetch(spc
, &dnp
->dn_blkptr
[i
], &zb
);
1581 if (dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) {
1583 zb
.zb_blkid
= DMU_SPILL_BLKID
;
1584 dsl_scan_prefetch(spc
, DN_SPILL_BLKPTR(dnp
), &zb
);
1587 scan_prefetch_ctx_rele(spc
, FTAG
);
1591 dsl_scan_prefetch_cb(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1592 arc_buf_t
*buf
, void *private)
1594 scan_prefetch_ctx_t
*spc
= private;
1595 dsl_scan_t
*scn
= spc
->spc_scn
;
1596 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1598 /* broadcast that the IO has completed for rate limiting purposes */
1599 mutex_enter(&spa
->spa_scrub_lock
);
1600 ASSERT3U(spa
->spa_scrub_inflight
, >=, BP_GET_PSIZE(bp
));
1601 spa
->spa_scrub_inflight
-= BP_GET_PSIZE(bp
);
1602 cv_broadcast(&spa
->spa_scrub_io_cv
);
1603 mutex_exit(&spa
->spa_scrub_lock
);
1605 /* if there was an error or we are done prefetching, just cleanup */
1606 if (buf
== NULL
|| scn
->scn_prefetch_stop
)
1609 if (BP_GET_LEVEL(bp
) > 0) {
1612 int epb
= BP_GET_LSIZE(bp
) >> SPA_BLKPTRSHIFT
;
1613 zbookmark_phys_t czb
;
1615 for (i
= 0, cbp
= buf
->b_data
; i
< epb
; i
++, cbp
++) {
1616 SET_BOOKMARK(&czb
, zb
->zb_objset
, zb
->zb_object
,
1617 zb
->zb_level
- 1, zb
->zb_blkid
* epb
+ i
);
1618 dsl_scan_prefetch(spc
, cbp
, &czb
);
1620 } else if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
) {
1623 int epb
= BP_GET_LSIZE(bp
) >> DNODE_SHIFT
;
1625 for (i
= 0, cdnp
= buf
->b_data
; i
< epb
;
1626 i
+= cdnp
->dn_extra_slots
+ 1,
1627 cdnp
+= cdnp
->dn_extra_slots
+ 1) {
1628 dsl_scan_prefetch_dnode(scn
, cdnp
,
1629 zb
->zb_objset
, zb
->zb_blkid
* epb
+ i
);
1631 } else if (BP_GET_TYPE(bp
) == DMU_OT_OBJSET
) {
1632 objset_phys_t
*osp
= buf
->b_data
;
1634 dsl_scan_prefetch_dnode(scn
, &osp
->os_meta_dnode
,
1635 zb
->zb_objset
, DMU_META_DNODE_OBJECT
);
1637 if (OBJSET_BUF_HAS_USERUSED(buf
)) {
1638 dsl_scan_prefetch_dnode(scn
,
1639 &osp
->os_groupused_dnode
, zb
->zb_objset
,
1640 DMU_GROUPUSED_OBJECT
);
1641 dsl_scan_prefetch_dnode(scn
,
1642 &osp
->os_userused_dnode
, zb
->zb_objset
,
1643 DMU_USERUSED_OBJECT
);
1649 arc_buf_destroy(buf
, private);
1650 scan_prefetch_ctx_rele(spc
, scn
);
1655 dsl_scan_prefetch_thread(void *arg
)
1657 dsl_scan_t
*scn
= arg
;
1658 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
1659 scan_prefetch_issue_ctx_t
*spic
;
1661 /* loop until we are told to stop */
1662 while (!scn
->scn_prefetch_stop
) {
1663 arc_flags_t flags
= ARC_FLAG_NOWAIT
|
1664 ARC_FLAG_PRESCIENT_PREFETCH
| ARC_FLAG_PREFETCH
;
1665 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SCAN_THREAD
;
1667 mutex_enter(&spa
->spa_scrub_lock
);
1670 * Wait until we have an IO to issue and are not above our
1671 * maximum in flight limit.
1673 while (!scn
->scn_prefetch_stop
&&
1674 (avl_numnodes(&scn
->scn_prefetch_queue
) == 0 ||
1675 spa
->spa_scrub_inflight
>= scn
->scn_maxinflight_bytes
)) {
1676 cv_wait(&spa
->spa_scrub_io_cv
, &spa
->spa_scrub_lock
);
1679 /* recheck if we should stop since we waited for the cv */
1680 if (scn
->scn_prefetch_stop
) {
1681 mutex_exit(&spa
->spa_scrub_lock
);
1685 /* remove the prefetch IO from the tree */
1686 spic
= avl_first(&scn
->scn_prefetch_queue
);
1687 spa
->spa_scrub_inflight
+= BP_GET_PSIZE(&spic
->spic_bp
);
1688 avl_remove(&scn
->scn_prefetch_queue
, spic
);
1690 mutex_exit(&spa
->spa_scrub_lock
);
1692 if (BP_IS_PROTECTED(&spic
->spic_bp
)) {
1693 ASSERT(BP_GET_TYPE(&spic
->spic_bp
) == DMU_OT_DNODE
||
1694 BP_GET_TYPE(&spic
->spic_bp
) == DMU_OT_OBJSET
);
1695 ASSERT3U(BP_GET_LEVEL(&spic
->spic_bp
), ==, 0);
1696 zio_flags
|= ZIO_FLAG_RAW
;
1699 /* issue the prefetch asynchronously */
1700 (void) arc_read(scn
->scn_zio_root
, scn
->scn_dp
->dp_spa
,
1701 &spic
->spic_bp
, dsl_scan_prefetch_cb
, spic
->spic_spc
,
1702 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, &spic
->spic_zb
);
1704 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1707 ASSERT(scn
->scn_prefetch_stop
);
1709 /* free any prefetches we didn't get to complete */
1710 mutex_enter(&spa
->spa_scrub_lock
);
1711 while ((spic
= avl_first(&scn
->scn_prefetch_queue
)) != NULL
) {
1712 avl_remove(&scn
->scn_prefetch_queue
, spic
);
1713 scan_prefetch_ctx_rele(spic
->spic_spc
, scn
);
1714 kmem_free(spic
, sizeof (scan_prefetch_issue_ctx_t
));
1716 ASSERT0(avl_numnodes(&scn
->scn_prefetch_queue
));
1717 mutex_exit(&spa
->spa_scrub_lock
);
1721 dsl_scan_check_resume(dsl_scan_t
*scn
, const dnode_phys_t
*dnp
,
1722 const zbookmark_phys_t
*zb
)
1725 * We never skip over user/group accounting objects (obj<0)
1727 if (!ZB_IS_ZERO(&scn
->scn_phys
.scn_bookmark
) &&
1728 (int64_t)zb
->zb_object
>= 0) {
1730 * If we already visited this bp & everything below (in
1731 * a prior txg sync), don't bother doing it again.
1733 if (zbookmark_subtree_completed(dnp
, zb
,
1734 &scn
->scn_phys
.scn_bookmark
))
1738 * If we found the block we're trying to resume from, or
1739 * we went past it to a different object, zero it out to
1740 * indicate that it's OK to start checking for suspending
1743 if (bcmp(zb
, &scn
->scn_phys
.scn_bookmark
, sizeof (*zb
)) == 0 ||
1744 zb
->zb_object
> scn
->scn_phys
.scn_bookmark
.zb_object
) {
1745 dprintf("resuming at %llx/%llx/%llx/%llx\n",
1746 (longlong_t
)zb
->zb_objset
,
1747 (longlong_t
)zb
->zb_object
,
1748 (longlong_t
)zb
->zb_level
,
1749 (longlong_t
)zb
->zb_blkid
);
1750 bzero(&scn
->scn_phys
.scn_bookmark
, sizeof (*zb
));
1756 static void dsl_scan_visitbp(blkptr_t
*bp
, const zbookmark_phys_t
*zb
,
1757 dnode_phys_t
*dnp
, dsl_dataset_t
*ds
, dsl_scan_t
*scn
,
1758 dmu_objset_type_t ostype
, dmu_tx_t
*tx
);
1759 inline __attribute__((always_inline
)) static void dsl_scan_visitdnode(
1760 dsl_scan_t
*, dsl_dataset_t
*ds
, dmu_objset_type_t ostype
,
1761 dnode_phys_t
*dnp
, uint64_t object
, dmu_tx_t
*tx
);
1764 * Return nonzero on i/o error.
1765 * Return new buf to write out in *bufp.
1767 inline __attribute__((always_inline
)) static int
1768 dsl_scan_recurse(dsl_scan_t
*scn
, dsl_dataset_t
*ds
, dmu_objset_type_t ostype
,
1769 dnode_phys_t
*dnp
, const blkptr_t
*bp
,
1770 const zbookmark_phys_t
*zb
, dmu_tx_t
*tx
)
1772 dsl_pool_t
*dp
= scn
->scn_dp
;
1773 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SCAN_THREAD
;
1776 ASSERT(!BP_IS_REDACTED(bp
));
1778 if (BP_GET_LEVEL(bp
) > 0) {
1779 arc_flags_t flags
= ARC_FLAG_WAIT
;
1782 int epb
= BP_GET_LSIZE(bp
) >> SPA_BLKPTRSHIFT
;
1785 err
= arc_read(NULL
, dp
->dp_spa
, bp
, arc_getbuf_func
, &buf
,
1786 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, zb
);
1788 scn
->scn_phys
.scn_errors
++;
1791 for (i
= 0, cbp
= buf
->b_data
; i
< epb
; i
++, cbp
++) {
1792 zbookmark_phys_t czb
;
1794 SET_BOOKMARK(&czb
, zb
->zb_objset
, zb
->zb_object
,
1796 zb
->zb_blkid
* epb
+ i
);
1797 dsl_scan_visitbp(cbp
, &czb
, dnp
,
1798 ds
, scn
, ostype
, tx
);
1800 arc_buf_destroy(buf
, &buf
);
1801 } else if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
) {
1802 arc_flags_t flags
= ARC_FLAG_WAIT
;
1805 int epb
= BP_GET_LSIZE(bp
) >> DNODE_SHIFT
;
1808 if (BP_IS_PROTECTED(bp
)) {
1809 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
1810 zio_flags
|= ZIO_FLAG_RAW
;
1813 err
= arc_read(NULL
, dp
->dp_spa
, bp
, arc_getbuf_func
, &buf
,
1814 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, zb
);
1816 scn
->scn_phys
.scn_errors
++;
1819 for (i
= 0, cdnp
= buf
->b_data
; i
< epb
;
1820 i
+= cdnp
->dn_extra_slots
+ 1,
1821 cdnp
+= cdnp
->dn_extra_slots
+ 1) {
1822 dsl_scan_visitdnode(scn
, ds
, ostype
,
1823 cdnp
, zb
->zb_blkid
* epb
+ i
, tx
);
1826 arc_buf_destroy(buf
, &buf
);
1827 } else if (BP_GET_TYPE(bp
) == DMU_OT_OBJSET
) {
1828 arc_flags_t flags
= ARC_FLAG_WAIT
;
1832 err
= arc_read(NULL
, dp
->dp_spa
, bp
, arc_getbuf_func
, &buf
,
1833 ZIO_PRIORITY_SCRUB
, zio_flags
, &flags
, zb
);
1835 scn
->scn_phys
.scn_errors
++;
1841 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1842 &osp
->os_meta_dnode
, DMU_META_DNODE_OBJECT
, tx
);
1844 if (OBJSET_BUF_HAS_USERUSED(buf
)) {
1846 * We also always visit user/group/project accounting
1847 * objects, and never skip them, even if we are
1848 * suspending. This is necessary so that the
1849 * space deltas from this txg get integrated.
1851 if (OBJSET_BUF_HAS_PROJECTUSED(buf
))
1852 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1853 &osp
->os_projectused_dnode
,
1854 DMU_PROJECTUSED_OBJECT
, tx
);
1855 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1856 &osp
->os_groupused_dnode
,
1857 DMU_GROUPUSED_OBJECT
, tx
);
1858 dsl_scan_visitdnode(scn
, ds
, osp
->os_type
,
1859 &osp
->os_userused_dnode
,
1860 DMU_USERUSED_OBJECT
, tx
);
1862 arc_buf_destroy(buf
, &buf
);
1868 inline __attribute__((always_inline
)) static void
1869 dsl_scan_visitdnode(dsl_scan_t
*scn
, dsl_dataset_t
*ds
,
1870 dmu_objset_type_t ostype
, dnode_phys_t
*dnp
,
1871 uint64_t object
, dmu_tx_t
*tx
)
1875 for (j
= 0; j
< dnp
->dn_nblkptr
; j
++) {
1876 zbookmark_phys_t czb
;
1878 SET_BOOKMARK(&czb
, ds
? ds
->ds_object
: 0, object
,
1879 dnp
->dn_nlevels
- 1, j
);
1880 dsl_scan_visitbp(&dnp
->dn_blkptr
[j
],
1881 &czb
, dnp
, ds
, scn
, ostype
, tx
);
1884 if (dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) {
1885 zbookmark_phys_t czb
;
1886 SET_BOOKMARK(&czb
, ds
? ds
->ds_object
: 0, object
,
1887 0, DMU_SPILL_BLKID
);
1888 dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp
),
1889 &czb
, dnp
, ds
, scn
, ostype
, tx
);
1894 * The arguments are in this order because mdb can only print the
1895 * first 5; we want them to be useful.
1898 dsl_scan_visitbp(blkptr_t
*bp
, const zbookmark_phys_t
*zb
,
1899 dnode_phys_t
*dnp
, dsl_dataset_t
*ds
, dsl_scan_t
*scn
,
1900 dmu_objset_type_t ostype
, dmu_tx_t
*tx
)
1902 dsl_pool_t
*dp
= scn
->scn_dp
;
1903 blkptr_t
*bp_toread
= NULL
;
1905 if (dsl_scan_check_suspend(scn
, zb
))
1908 if (dsl_scan_check_resume(scn
, dnp
, zb
))
1911 scn
->scn_visited_this_txg
++;
1914 * This debugging is commented out to conserve stack space. This
1915 * function is called recursively and the debugging adds several
1916 * bytes to the stack for each call. It can be commented back in
1917 * if required to debug an issue in dsl_scan_visitbp().
1920 * "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
1921 * ds, ds ? ds->ds_object : 0,
1922 * zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1926 if (BP_IS_HOLE(bp
)) {
1927 scn
->scn_holes_this_txg
++;
1931 if (BP_IS_REDACTED(bp
)) {
1932 ASSERT(dsl_dataset_feature_is_active(ds
,
1933 SPA_FEATURE_REDACTED_DATASETS
));
1937 if (bp
->blk_birth
<= scn
->scn_phys
.scn_cur_min_txg
) {
1938 scn
->scn_lt_min_this_txg
++;
1942 bp_toread
= kmem_alloc(sizeof (blkptr_t
), KM_SLEEP
);
1945 if (dsl_scan_recurse(scn
, ds
, ostype
, dnp
, bp_toread
, zb
, tx
) != 0)
1949 * If dsl_scan_ddt() has already visited this block, it will have
1950 * already done any translations or scrubbing, so don't call the
1953 if (ddt_class_contains(dp
->dp_spa
,
1954 scn
->scn_phys
.scn_ddt_class_max
, bp
)) {
1955 scn
->scn_ddt_contained_this_txg
++;
1960 * If this block is from the future (after cur_max_txg), then we
1961 * are doing this on behalf of a deleted snapshot, and we will
1962 * revisit the future block on the next pass of this dataset.
1963 * Don't scan it now unless we need to because something
1964 * under it was modified.
1966 if (BP_PHYSICAL_BIRTH(bp
) > scn
->scn_phys
.scn_cur_max_txg
) {
1967 scn
->scn_gt_max_this_txg
++;
1971 scan_funcs
[scn
->scn_phys
.scn_func
](dp
, bp
, zb
);
1974 kmem_free(bp_toread
, sizeof (blkptr_t
));
1978 dsl_scan_visit_rootbp(dsl_scan_t
*scn
, dsl_dataset_t
*ds
, blkptr_t
*bp
,
1981 zbookmark_phys_t zb
;
1982 scan_prefetch_ctx_t
*spc
;
1984 SET_BOOKMARK(&zb
, ds
? ds
->ds_object
: DMU_META_OBJSET
,
1985 ZB_ROOT_OBJECT
, ZB_ROOT_LEVEL
, ZB_ROOT_BLKID
);
1987 if (ZB_IS_ZERO(&scn
->scn_phys
.scn_bookmark
)) {
1988 SET_BOOKMARK(&scn
->scn_prefetch_bookmark
,
1989 zb
.zb_objset
, 0, 0, 0);
1991 scn
->scn_prefetch_bookmark
= scn
->scn_phys
.scn_bookmark
;
1994 scn
->scn_objsets_visited_this_txg
++;
1996 spc
= scan_prefetch_ctx_create(scn
, NULL
, FTAG
);
1997 dsl_scan_prefetch(spc
, bp
, &zb
);
1998 scan_prefetch_ctx_rele(spc
, FTAG
);
2000 dsl_scan_visitbp(bp
, &zb
, NULL
, ds
, scn
, DMU_OST_NONE
, tx
);
2002 dprintf_ds(ds
, "finished scan%s", "");
2006 ds_destroyed_scn_phys(dsl_dataset_t
*ds
, dsl_scan_phys_t
*scn_phys
)
2008 if (scn_phys
->scn_bookmark
.zb_objset
== ds
->ds_object
) {
2009 if (ds
->ds_is_snapshot
) {
2012 * - scn_cur_{min,max}_txg stays the same.
2013 * - Setting the flag is not really necessary if
2014 * scn_cur_max_txg == scn_max_txg, because there
2015 * is nothing after this snapshot that we care
2016 * about. However, we set it anyway and then
2017 * ignore it when we retraverse it in
2018 * dsl_scan_visitds().
2020 scn_phys
->scn_bookmark
.zb_objset
=
2021 dsl_dataset_phys(ds
)->ds_next_snap_obj
;
2022 zfs_dbgmsg("destroying ds %llu; currently traversing; "
2023 "reset zb_objset to %llu",
2024 (u_longlong_t
)ds
->ds_object
,
2025 (u_longlong_t
)dsl_dataset_phys(ds
)->
2027 scn_phys
->scn_flags
|= DSF_VISIT_DS_AGAIN
;
2029 SET_BOOKMARK(&scn_phys
->scn_bookmark
,
2030 ZB_DESTROYED_OBJSET
, 0, 0, 0);
2031 zfs_dbgmsg("destroying ds %llu; currently traversing; "
2032 "reset bookmark to -1,0,0,0",
2033 (u_longlong_t
)ds
->ds_object
);
2039 * Invoked when a dataset is destroyed. We need to make sure that:
2041 * 1) If it is the dataset that was currently being scanned, we write
2042 * a new dsl_scan_phys_t and marking the objset reference in it
2044 * 2) Remove it from the work queue, if it was present.
2046 * If the dataset was actually a snapshot, instead of marking the dataset
2047 * as destroyed, we instead substitute the next snapshot in line.
2050 dsl_scan_ds_destroyed(dsl_dataset_t
*ds
, dmu_tx_t
*tx
)
2052 dsl_pool_t
*dp
= ds
->ds_dir
->dd_pool
;
2053 dsl_scan_t
*scn
= dp
->dp_scan
;
2056 if (!dsl_scan_is_running(scn
))
2059 ds_destroyed_scn_phys(ds
, &scn
->scn_phys
);
2060 ds_destroyed_scn_phys(ds
, &scn
->scn_phys_cached
);
2062 if (scan_ds_queue_contains(scn
, ds
->ds_object
, &mintxg
)) {
2063 scan_ds_queue_remove(scn
, ds
->ds_object
);
2064 if (ds
->ds_is_snapshot
)
2065 scan_ds_queue_insert(scn
,
2066 dsl_dataset_phys(ds
)->ds_next_snap_obj
, mintxg
);
2069 if (zap_lookup_int_key(dp
->dp_meta_objset
, scn
->scn_phys
.scn_queue_obj
,
2070 ds
->ds_object
, &mintxg
) == 0) {
2071 ASSERT3U(dsl_dataset_phys(ds
)->ds_num_children
, <=, 1);
2072 VERIFY3U(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2073 scn
->scn_phys
.scn_queue_obj
, ds
->ds_object
, tx
));
2074 if (ds
->ds_is_snapshot
) {
2076 * We keep the same mintxg; it could be >
2077 * ds_creation_txg if the previous snapshot was
2080 VERIFY(zap_add_int_key(dp
->dp_meta_objset
,
2081 scn
->scn_phys
.scn_queue_obj
,
2082 dsl_dataset_phys(ds
)->ds_next_snap_obj
,
2084 zfs_dbgmsg("destroying ds %llu; in queue; "
2085 "replacing with %llu",
2086 (u_longlong_t
)ds
->ds_object
,
2087 (u_longlong_t
)dsl_dataset_phys(ds
)->
2090 zfs_dbgmsg("destroying ds %llu; in queue; removing",
2091 (u_longlong_t
)ds
->ds_object
);
2096 * dsl_scan_sync() should be called after this, and should sync
2097 * out our changed state, but just to be safe, do it here.
2099 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
2103 ds_snapshotted_bookmark(dsl_dataset_t
*ds
, zbookmark_phys_t
*scn_bookmark
)
2105 if (scn_bookmark
->zb_objset
== ds
->ds_object
) {
2106 scn_bookmark
->zb_objset
=
2107 dsl_dataset_phys(ds
)->ds_prev_snap_obj
;
2108 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
2109 "reset zb_objset to %llu",
2110 (u_longlong_t
)ds
->ds_object
,
2111 (u_longlong_t
)dsl_dataset_phys(ds
)->ds_prev_snap_obj
);
2116 * Called when a dataset is snapshotted. If we were currently traversing
2117 * this snapshot, we reset our bookmark to point at the newly created
2118 * snapshot. We also modify our work queue to remove the old snapshot and
2119 * replace with the new one.
2122 dsl_scan_ds_snapshotted(dsl_dataset_t
*ds
, dmu_tx_t
*tx
)
2124 dsl_pool_t
*dp
= ds
->ds_dir
->dd_pool
;
2125 dsl_scan_t
*scn
= dp
->dp_scan
;
2128 if (!dsl_scan_is_running(scn
))
2131 ASSERT(dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0);
2133 ds_snapshotted_bookmark(ds
, &scn
->scn_phys
.scn_bookmark
);
2134 ds_snapshotted_bookmark(ds
, &scn
->scn_phys_cached
.scn_bookmark
);
2136 if (scan_ds_queue_contains(scn
, ds
->ds_object
, &mintxg
)) {
2137 scan_ds_queue_remove(scn
, ds
->ds_object
);
2138 scan_ds_queue_insert(scn
,
2139 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, mintxg
);
2142 if (zap_lookup_int_key(dp
->dp_meta_objset
, scn
->scn_phys
.scn_queue_obj
,
2143 ds
->ds_object
, &mintxg
) == 0) {
2144 VERIFY3U(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2145 scn
->scn_phys
.scn_queue_obj
, ds
->ds_object
, tx
));
2146 VERIFY(zap_add_int_key(dp
->dp_meta_objset
,
2147 scn
->scn_phys
.scn_queue_obj
,
2148 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, mintxg
, tx
) == 0);
2149 zfs_dbgmsg("snapshotting ds %llu; in queue; "
2150 "replacing with %llu",
2151 (u_longlong_t
)ds
->ds_object
,
2152 (u_longlong_t
)dsl_dataset_phys(ds
)->ds_prev_snap_obj
);
2155 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
2159 ds_clone_swapped_bookmark(dsl_dataset_t
*ds1
, dsl_dataset_t
*ds2
,
2160 zbookmark_phys_t
*scn_bookmark
)
2162 if (scn_bookmark
->zb_objset
== ds1
->ds_object
) {
2163 scn_bookmark
->zb_objset
= ds2
->ds_object
;
2164 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2165 "reset zb_objset to %llu",
2166 (u_longlong_t
)ds1
->ds_object
,
2167 (u_longlong_t
)ds2
->ds_object
);
2168 } else if (scn_bookmark
->zb_objset
== ds2
->ds_object
) {
2169 scn_bookmark
->zb_objset
= ds1
->ds_object
;
2170 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2171 "reset zb_objset to %llu",
2172 (u_longlong_t
)ds2
->ds_object
,
2173 (u_longlong_t
)ds1
->ds_object
);
2178 * Called when a parent dataset and its clone are swapped. If we were
2179 * currently traversing the dataset, we need to switch to traversing the
2180 * newly promoted parent.
2183 dsl_scan_ds_clone_swapped(dsl_dataset_t
*ds1
, dsl_dataset_t
*ds2
, dmu_tx_t
*tx
)
2185 dsl_pool_t
*dp
= ds1
->ds_dir
->dd_pool
;
2186 dsl_scan_t
*scn
= dp
->dp_scan
;
2189 if (!dsl_scan_is_running(scn
))
2192 ds_clone_swapped_bookmark(ds1
, ds2
, &scn
->scn_phys
.scn_bookmark
);
2193 ds_clone_swapped_bookmark(ds1
, ds2
, &scn
->scn_phys_cached
.scn_bookmark
);
2195 if (scan_ds_queue_contains(scn
, ds1
->ds_object
, &mintxg
)) {
2196 scan_ds_queue_remove(scn
, ds1
->ds_object
);
2197 scan_ds_queue_insert(scn
, ds2
->ds_object
, mintxg
);
2199 if (scan_ds_queue_contains(scn
, ds2
->ds_object
, &mintxg
)) {
2200 scan_ds_queue_remove(scn
, ds2
->ds_object
);
2201 scan_ds_queue_insert(scn
, ds1
->ds_object
, mintxg
);
2204 if (zap_lookup_int_key(dp
->dp_meta_objset
, scn
->scn_phys
.scn_queue_obj
,
2205 ds1
->ds_object
, &mintxg
) == 0) {
2207 ASSERT3U(mintxg
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2208 ASSERT3U(mintxg
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2209 VERIFY3U(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2210 scn
->scn_phys
.scn_queue_obj
, ds1
->ds_object
, tx
));
2211 err
= zap_add_int_key(dp
->dp_meta_objset
,
2212 scn
->scn_phys
.scn_queue_obj
, ds2
->ds_object
, mintxg
, tx
);
2213 VERIFY(err
== 0 || err
== EEXIST
);
2214 if (err
== EEXIST
) {
2215 /* Both were there to begin with */
2216 VERIFY(0 == zap_add_int_key(dp
->dp_meta_objset
,
2217 scn
->scn_phys
.scn_queue_obj
,
2218 ds1
->ds_object
, mintxg
, tx
));
2220 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2221 "replacing with %llu",
2222 (u_longlong_t
)ds1
->ds_object
,
2223 (u_longlong_t
)ds2
->ds_object
);
2225 if (zap_lookup_int_key(dp
->dp_meta_objset
, scn
->scn_phys
.scn_queue_obj
,
2226 ds2
->ds_object
, &mintxg
) == 0) {
2227 ASSERT3U(mintxg
, ==, dsl_dataset_phys(ds1
)->ds_prev_snap_txg
);
2228 ASSERT3U(mintxg
, ==, dsl_dataset_phys(ds2
)->ds_prev_snap_txg
);
2229 VERIFY3U(0, ==, zap_remove_int(dp
->dp_meta_objset
,
2230 scn
->scn_phys
.scn_queue_obj
, ds2
->ds_object
, tx
));
2231 VERIFY(0 == zap_add_int_key(dp
->dp_meta_objset
,
2232 scn
->scn_phys
.scn_queue_obj
, ds1
->ds_object
, mintxg
, tx
));
2233 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2234 "replacing with %llu",
2235 (u_longlong_t
)ds2
->ds_object
,
2236 (u_longlong_t
)ds1
->ds_object
);
2239 dsl_scan_sync_state(scn
, tx
, SYNC_CACHED
);
2244 enqueue_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
2246 uint64_t originobj
= *(uint64_t *)arg
;
2249 dsl_scan_t
*scn
= dp
->dp_scan
;
2251 if (dsl_dir_phys(hds
->ds_dir
)->dd_origin_obj
!= originobj
)
2254 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
2258 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= originobj
) {
2259 dsl_dataset_t
*prev
;
2260 err
= dsl_dataset_hold_obj(dp
,
2261 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
2263 dsl_dataset_rele(ds
, FTAG
);
2268 scan_ds_queue_insert(scn
, ds
->ds_object
,
2269 dsl_dataset_phys(ds
)->ds_prev_snap_txg
);
2270 dsl_dataset_rele(ds
, FTAG
);
2275 dsl_scan_visitds(dsl_scan_t
*scn
, uint64_t dsobj
, dmu_tx_t
*tx
)
2277 dsl_pool_t
*dp
= scn
->scn_dp
;
2280 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
2282 if (scn
->scn_phys
.scn_cur_min_txg
>=
2283 scn
->scn_phys
.scn_max_txg
) {
2285 * This can happen if this snapshot was created after the
2286 * scan started, and we already completed a previous snapshot
2287 * that was created after the scan started. This snapshot
2288 * only references blocks with:
2290 * birth < our ds_creation_txg
2291 * cur_min_txg is no less than ds_creation_txg.
2292 * We have already visited these blocks.
2294 * birth > scn_max_txg
2295 * The scan requested not to visit these blocks.
2297 * Subsequent snapshots (and clones) can reference our
2298 * blocks, or blocks with even higher birth times.
2299 * Therefore we do not need to visit them either,
2300 * so we do not add them to the work queue.
2302 * Note that checking for cur_min_txg >= cur_max_txg
2303 * is not sufficient, because in that case we may need to
2304 * visit subsequent snapshots. This happens when min_txg > 0,
2305 * which raises cur_min_txg. In this case we will visit
2306 * this dataset but skip all of its blocks, because the
2307 * rootbp's birth time is < cur_min_txg. Then we will
2308 * add the next snapshots/clones to the work queue.
2310 char *dsname
= kmem_alloc(ZFS_MAX_DATASET_NAME_LEN
, KM_SLEEP
);
2311 dsl_dataset_name(ds
, dsname
);
2312 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2313 "cur_min_txg (%llu) >= max_txg (%llu)",
2314 (longlong_t
)dsobj
, dsname
,
2315 (longlong_t
)scn
->scn_phys
.scn_cur_min_txg
,
2316 (longlong_t
)scn
->scn_phys
.scn_max_txg
);
2317 kmem_free(dsname
, MAXNAMELEN
);
2323 * Only the ZIL in the head (non-snapshot) is valid. Even though
2324 * snapshots can have ZIL block pointers (which may be the same
2325 * BP as in the head), they must be ignored. In addition, $ORIGIN
2326 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2327 * need to look for a ZIL in it either. So we traverse the ZIL here,
2328 * rather than in scan_recurse(), because the regular snapshot
2329 * block-sharing rules don't apply to it.
2331 if (!dsl_dataset_is_snapshot(ds
) &&
2332 (dp
->dp_origin_snap
== NULL
||
2333 ds
->ds_dir
!= dp
->dp_origin_snap
->ds_dir
)) {
2335 if (dmu_objset_from_ds(ds
, &os
) != 0) {
2338 dsl_scan_zil(dp
, &os
->os_zil_header
);
2342 * Iterate over the bps in this ds.
2344 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
2345 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
2346 dsl_scan_visit_rootbp(scn
, ds
, &dsl_dataset_phys(ds
)->ds_bp
, tx
);
2347 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
2349 char *dsname
= kmem_alloc(ZFS_MAX_DATASET_NAME_LEN
, KM_SLEEP
);
2350 dsl_dataset_name(ds
, dsname
);
2351 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2353 (longlong_t
)dsobj
, dsname
,
2354 (longlong_t
)scn
->scn_phys
.scn_cur_min_txg
,
2355 (longlong_t
)scn
->scn_phys
.scn_cur_max_txg
,
2356 (int)scn
->scn_suspending
);
2357 kmem_free(dsname
, ZFS_MAX_DATASET_NAME_LEN
);
2359 if (scn
->scn_suspending
)
2363 * We've finished this pass over this dataset.
2367 * If we did not completely visit this dataset, do another pass.
2369 if (scn
->scn_phys
.scn_flags
& DSF_VISIT_DS_AGAIN
) {
2370 zfs_dbgmsg("incomplete pass; visiting again");
2371 scn
->scn_phys
.scn_flags
&= ~DSF_VISIT_DS_AGAIN
;
2372 scan_ds_queue_insert(scn
, ds
->ds_object
,
2373 scn
->scn_phys
.scn_cur_max_txg
);
2378 * Add descendant datasets to work queue.
2380 if (dsl_dataset_phys(ds
)->ds_next_snap_obj
!= 0) {
2381 scan_ds_queue_insert(scn
,
2382 dsl_dataset_phys(ds
)->ds_next_snap_obj
,
2383 dsl_dataset_phys(ds
)->ds_creation_txg
);
2385 if (dsl_dataset_phys(ds
)->ds_num_children
> 1) {
2386 boolean_t usenext
= B_FALSE
;
2387 if (dsl_dataset_phys(ds
)->ds_next_clones_obj
!= 0) {
2390 * A bug in a previous version of the code could
2391 * cause upgrade_clones_cb() to not set
2392 * ds_next_snap_obj when it should, leading to a
2393 * missing entry. Therefore we can only use the
2394 * next_clones_obj when its count is correct.
2396 int err
= zap_count(dp
->dp_meta_objset
,
2397 dsl_dataset_phys(ds
)->ds_next_clones_obj
, &count
);
2399 count
== dsl_dataset_phys(ds
)->ds_num_children
- 1)
2406 for (zap_cursor_init(&zc
, dp
->dp_meta_objset
,
2407 dsl_dataset_phys(ds
)->ds_next_clones_obj
);
2408 zap_cursor_retrieve(&zc
, &za
) == 0;
2409 (void) zap_cursor_advance(&zc
)) {
2410 scan_ds_queue_insert(scn
,
2411 zfs_strtonum(za
.za_name
, NULL
),
2412 dsl_dataset_phys(ds
)->ds_creation_txg
);
2414 zap_cursor_fini(&zc
);
2416 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
2417 enqueue_clones_cb
, &ds
->ds_object
,
2423 dsl_dataset_rele(ds
, FTAG
);
2428 enqueue_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
2432 dsl_scan_t
*scn
= dp
->dp_scan
;
2434 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
2438 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
2439 dsl_dataset_t
*prev
;
2440 err
= dsl_dataset_hold_obj(dp
,
2441 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
2443 dsl_dataset_rele(ds
, FTAG
);
2448 * If this is a clone, we don't need to worry about it for now.
2450 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
) {
2451 dsl_dataset_rele(ds
, FTAG
);
2452 dsl_dataset_rele(prev
, FTAG
);
2455 dsl_dataset_rele(ds
, FTAG
);
2459 scan_ds_queue_insert(scn
, ds
->ds_object
,
2460 dsl_dataset_phys(ds
)->ds_prev_snap_txg
);
2461 dsl_dataset_rele(ds
, FTAG
);
2467 dsl_scan_ddt_entry(dsl_scan_t
*scn
, enum zio_checksum checksum
,
2468 ddt_entry_t
*dde
, dmu_tx_t
*tx
)
2470 const ddt_key_t
*ddk
= &dde
->dde_key
;
2471 ddt_phys_t
*ddp
= dde
->dde_phys
;
2473 zbookmark_phys_t zb
= { 0 };
2476 if (!dsl_scan_is_running(scn
))
2480 * This function is special because it is the only thing
2481 * that can add scan_io_t's to the vdev scan queues from
2482 * outside dsl_scan_sync(). For the most part this is ok
2483 * as long as it is called from within syncing context.
2484 * However, dsl_scan_sync() expects that no new sio's will
2485 * be added between when all the work for a scan is done
2486 * and the next txg when the scan is actually marked as
2487 * completed. This check ensures we do not issue new sio's
2488 * during this period.
2490 if (scn
->scn_done_txg
!= 0)
2493 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2494 if (ddp
->ddp_phys_birth
== 0 ||
2495 ddp
->ddp_phys_birth
> scn
->scn_phys
.scn_max_txg
)
2497 ddt_bp_create(checksum
, ddk
, ddp
, &bp
);
2499 scn
->scn_visited_this_txg
++;
2500 scan_funcs
[scn
->scn_phys
.scn_func
](scn
->scn_dp
, &bp
, &zb
);
2505 * Scrub/dedup interaction.
2507 * If there are N references to a deduped block, we don't want to scrub it
2508 * N times -- ideally, we should scrub it exactly once.
2510 * We leverage the fact that the dde's replication class (enum ddt_class)
2511 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2512 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2514 * To prevent excess scrubbing, the scrub begins by walking the DDT
2515 * to find all blocks with refcnt > 1, and scrubs each of these once.
2516 * Since there are two replication classes which contain blocks with
2517 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2518 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2520 * There would be nothing more to say if a block's refcnt couldn't change
2521 * during a scrub, but of course it can so we must account for changes
2522 * in a block's replication class.
2524 * Here's an example of what can occur:
2526 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2527 * when visited during the top-down scrub phase, it will be scrubbed twice.
2528 * This negates our scrub optimization, but is otherwise harmless.
2530 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2531 * on each visit during the top-down scrub phase, it will never be scrubbed.
2532 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2533 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2534 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2535 * while a scrub is in progress, it scrubs the block right then.
2538 dsl_scan_ddt(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
2540 ddt_bookmark_t
*ddb
= &scn
->scn_phys
.scn_ddt_bookmark
;
2545 bzero(&dde
, sizeof (ddt_entry_t
));
2547 while ((error
= ddt_walk(scn
->scn_dp
->dp_spa
, ddb
, &dde
)) == 0) {
2550 if (ddb
->ddb_class
> scn
->scn_phys
.scn_ddt_class_max
)
2552 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
2553 (longlong_t
)ddb
->ddb_class
,
2554 (longlong_t
)ddb
->ddb_type
,
2555 (longlong_t
)ddb
->ddb_checksum
,
2556 (longlong_t
)ddb
->ddb_cursor
);
2558 /* There should be no pending changes to the dedup table */
2559 ddt
= scn
->scn_dp
->dp_spa
->spa_ddt
[ddb
->ddb_checksum
];
2560 ASSERT(avl_first(&ddt
->ddt_tree
) == NULL
);
2562 dsl_scan_ddt_entry(scn
, ddb
->ddb_checksum
, &dde
, tx
);
2565 if (dsl_scan_check_suspend(scn
, NULL
))
2569 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; "
2570 "suspending=%u", (longlong_t
)n
,
2571 (int)scn
->scn_phys
.scn_ddt_class_max
, (int)scn
->scn_suspending
);
2573 ASSERT(error
== 0 || error
== ENOENT
);
2574 ASSERT(error
!= ENOENT
||
2575 ddb
->ddb_class
> scn
->scn_phys
.scn_ddt_class_max
);
2579 dsl_scan_ds_maxtxg(dsl_dataset_t
*ds
)
2581 uint64_t smt
= ds
->ds_dir
->dd_pool
->dp_scan
->scn_phys
.scn_max_txg
;
2582 if (ds
->ds_is_snapshot
)
2583 return (MIN(smt
, dsl_dataset_phys(ds
)->ds_creation_txg
));
2588 dsl_scan_visit(dsl_scan_t
*scn
, dmu_tx_t
*tx
)
2591 dsl_pool_t
*dp
= scn
->scn_dp
;
2593 if (scn
->scn_phys
.scn_ddt_bookmark
.ddb_class
<=
2594 scn
->scn_phys
.scn_ddt_class_max
) {
2595 scn
->scn_phys
.scn_cur_min_txg
= scn
->scn_phys
.scn_min_txg
;
2596 scn
->scn_phys
.scn_cur_max_txg
= scn
->scn_phys
.scn_max_txg
;
2597 dsl_scan_ddt(scn
, tx
);
2598 if (scn
->scn_suspending
)
2602 if (scn
->scn_phys
.scn_bookmark
.zb_objset
== DMU_META_OBJSET
) {
2603 /* First do the MOS & ORIGIN */
2605 scn
->scn_phys
.scn_cur_min_txg
= scn
->scn_phys
.scn_min_txg
;
2606 scn
->scn_phys
.scn_cur_max_txg
= scn
->scn_phys
.scn_max_txg
;
2607 dsl_scan_visit_rootbp(scn
, NULL
,
2608 &dp
->dp_meta_rootbp
, tx
);
2609 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
2610 if (scn
->scn_suspending
)
2613 if (spa_version(dp
->dp_spa
) < SPA_VERSION_DSL_SCRUB
) {
2614 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
2615 enqueue_cb
, NULL
, DS_FIND_CHILDREN
));
2617 dsl_scan_visitds(scn
,
2618 dp
->dp_origin_snap
->ds_object
, tx
);
2620 ASSERT(!scn
->scn_suspending
);
2621 } else if (scn
->scn_phys
.scn_bookmark
.zb_objset
!=
2622 ZB_DESTROYED_OBJSET
) {
2623 uint64_t dsobj
= scn
->scn_phys
.scn_bookmark
.zb_objset
;
2625 * If we were suspended, continue from here. Note if the
2626 * ds we were suspended on was deleted, the zb_objset may
2627 * be -1, so we will skip this and find a new objset
2630 dsl_scan_visitds(scn
, dsobj
, tx
);
2631 if (scn
->scn_suspending
)
2636 * In case we suspended right at the end of the ds, zero the
2637 * bookmark so we don't think that we're still trying to resume.
2639 bzero(&scn
->scn_phys
.scn_bookmark
, sizeof (zbookmark_phys_t
));
2642 * Keep pulling things out of the dataset avl queue. Updates to the
2643 * persistent zap-object-as-queue happen only at checkpoints.
2645 while ((sds
= avl_first(&scn
->scn_queue
)) != NULL
) {
2647 uint64_t dsobj
= sds
->sds_dsobj
;
2648 uint64_t txg
= sds
->sds_txg
;
2650 /* dequeue and free the ds from the queue */
2651 scan_ds_queue_remove(scn
, dsobj
);
2654 /* set up min / max txg */
2655 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
2657 scn
->scn_phys
.scn_cur_min_txg
=
2658 MAX(scn
->scn_phys
.scn_min_txg
, txg
);
2660 scn
->scn_phys
.scn_cur_min_txg
=
2661 MAX(scn
->scn_phys
.scn_min_txg
,
2662 dsl_dataset_phys(ds
)->ds_prev_snap_txg
);
2664 scn
->scn_phys
.scn_cur_max_txg
= dsl_scan_ds_maxtxg(ds
);
2665 dsl_dataset_rele(ds
, FTAG
);
2667 dsl_scan_visitds(scn
, dsobj
, tx
);
2668 if (scn
->scn_suspending
)
2672 /* No more objsets to fetch, we're done */
2673 scn
->scn_phys
.scn_bookmark
.zb_objset
= ZB_DESTROYED_OBJSET
;
2674 ASSERT0(scn
->scn_suspending
);
2678 dsl_scan_count_leaves(vdev_t
*vd
)
2680 uint64_t i
, leaves
= 0;
2682 /* we only count leaves that belong to the main pool and are readable */
2683 if (vd
->vdev_islog
|| vd
->vdev_isspare
||
2684 vd
->vdev_isl2cache
|| !vdev_readable(vd
))
2687 if (vd
->vdev_ops
->vdev_op_leaf
)
2690 for (i
= 0; i
< vd
->vdev_children
; i
++) {
2691 leaves
+= dsl_scan_count_leaves(vd
->vdev_child
[i
]);
2698 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t
*q
, const blkptr_t
*bp
)
2701 uint64_t cur_size
= 0;
2703 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
2704 cur_size
+= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
2707 q
->q_total_zio_size_this_txg
+= cur_size
;
2708 q
->q_zios_this_txg
++;
2712 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t
*q
, uint64_t start
,
2715 q
->q_total_seg_size_this_txg
+= end
- start
;
2716 q
->q_segs_this_txg
++;
2720 scan_io_queue_check_suspend(dsl_scan_t
*scn
)
2722 /* See comment in dsl_scan_check_suspend() */
2723 uint64_t curr_time_ns
= gethrtime();
2724 uint64_t scan_time_ns
= curr_time_ns
- scn
->scn_sync_start_time
;
2725 uint64_t sync_time_ns
= curr_time_ns
-
2726 scn
->scn_dp
->dp_spa
->spa_sync_starttime
;
2727 int dirty_pct
= scn
->scn_dp
->dp_dirty_total
* 100 / zfs_dirty_data_max
;
2728 int mintime
= (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) ?
2729 zfs_resilver_min_time_ms
: zfs_scrub_min_time_ms
;
2731 return ((NSEC2MSEC(scan_time_ns
) > mintime
&&
2732 (dirty_pct
>= zfs_vdev_async_write_active_min_dirty_percent
||
2733 txg_sync_waiting(scn
->scn_dp
) ||
2734 NSEC2SEC(sync_time_ns
) >= zfs_txg_timeout
)) ||
2735 spa_shutting_down(scn
->scn_dp
->dp_spa
));
2739 * Given a list of scan_io_t's in io_list, this issues the I/Os out to
2740 * disk. This consumes the io_list and frees the scan_io_t's. This is
2741 * called when emptying queues, either when we're up against the memory
2742 * limit or when we have finished scanning. Returns B_TRUE if we stopped
2743 * processing the list before we finished. Any sios that were not issued
2744 * will remain in the io_list.
2747 scan_io_queue_issue(dsl_scan_io_queue_t
*queue
, list_t
*io_list
)
2749 dsl_scan_t
*scn
= queue
->q_scn
;
2751 int64_t bytes_issued
= 0;
2752 boolean_t suspended
= B_FALSE
;
2754 while ((sio
= list_head(io_list
)) != NULL
) {
2757 if (scan_io_queue_check_suspend(scn
)) {
2763 bytes_issued
+= SIO_GET_ASIZE(sio
);
2764 scan_exec_io(scn
->scn_dp
, &bp
, sio
->sio_flags
,
2765 &sio
->sio_zb
, queue
);
2766 (void) list_remove_head(io_list
);
2767 scan_io_queues_update_zio_stats(queue
, &bp
);
2771 atomic_add_64(&scn
->scn_bytes_pending
, -bytes_issued
);
2777 * This function removes sios from an IO queue which reside within a given
2778 * range_seg_t and inserts them (in offset order) into a list. Note that
2779 * we only ever return a maximum of 32 sios at once. If there are more sios
2780 * to process within this segment that did not make it onto the list we
2781 * return B_TRUE and otherwise B_FALSE.
2784 scan_io_queue_gather(dsl_scan_io_queue_t
*queue
, range_seg_t
*rs
, list_t
*list
)
2786 scan_io_t
*srch_sio
, *sio
, *next_sio
;
2788 uint_t num_sios
= 0;
2789 int64_t bytes_issued
= 0;
2792 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
2794 srch_sio
= sio_alloc(1);
2795 srch_sio
->sio_nr_dvas
= 1;
2796 SIO_SET_OFFSET(srch_sio
, rs
->rs_start
);
2799 * The exact start of the extent might not contain any matching zios,
2800 * so if that's the case, examine the next one in the tree.
2802 sio
= avl_find(&queue
->q_sios_by_addr
, srch_sio
, &idx
);
2806 sio
= avl_nearest(&queue
->q_sios_by_addr
, idx
, AVL_AFTER
);
2808 while (sio
!= NULL
&&
2809 SIO_GET_OFFSET(sio
) < rs
->rs_end
&& num_sios
<= 32) {
2810 ASSERT3U(SIO_GET_OFFSET(sio
), >=, rs
->rs_start
);
2811 ASSERT3U(SIO_GET_END_OFFSET(sio
), <=, rs
->rs_end
);
2813 next_sio
= AVL_NEXT(&queue
->q_sios_by_addr
, sio
);
2814 avl_remove(&queue
->q_sios_by_addr
, sio
);
2815 queue
->q_sio_memused
-= SIO_GET_MUSED(sio
);
2817 bytes_issued
+= SIO_GET_ASIZE(sio
);
2819 list_insert_tail(list
, sio
);
2824 * We limit the number of sios we process at once to 32 to avoid
2825 * biting off more than we can chew. If we didn't take everything
2826 * in the segment we update it to reflect the work we were able to
2827 * complete. Otherwise, we remove it from the range tree entirely.
2829 if (sio
!= NULL
&& SIO_GET_OFFSET(sio
) < rs
->rs_end
) {
2830 range_tree_adjust_fill(queue
->q_exts_by_addr
, rs
,
2832 range_tree_resize_segment(queue
->q_exts_by_addr
, rs
,
2833 SIO_GET_OFFSET(sio
), rs
->rs_end
- SIO_GET_OFFSET(sio
));
2837 range_tree_remove(queue
->q_exts_by_addr
, rs
->rs_start
,
2838 rs
->rs_end
- rs
->rs_start
);
2844 * This is called from the queue emptying thread and selects the next
2845 * extent from which we are to issue I/Os. The behavior of this function
2846 * depends on the state of the scan, the current memory consumption and
2847 * whether or not we are performing a scan shutdown.
2848 * 1) We select extents in an elevator algorithm (LBA-order) if the scan
2849 * needs to perform a checkpoint
2850 * 2) We select the largest available extent if we are up against the
2852 * 3) Otherwise we don't select any extents.
2854 static range_seg_t
*
2855 scan_io_queue_fetch_ext(dsl_scan_io_queue_t
*queue
)
2857 dsl_scan_t
*scn
= queue
->q_scn
;
2859 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
2860 ASSERT(scn
->scn_is_sorted
);
2862 /* handle tunable overrides */
2863 if (scn
->scn_checkpointing
|| scn
->scn_clearing
) {
2864 if (zfs_scan_issue_strategy
== 1) {
2865 return (range_tree_first(queue
->q_exts_by_addr
));
2866 } else if (zfs_scan_issue_strategy
== 2) {
2867 return (avl_first(&queue
->q_exts_by_size
));
2872 * During normal clearing, we want to issue our largest segments
2873 * first, keeping IO as sequential as possible, and leaving the
2874 * smaller extents for later with the hope that they might eventually
2875 * grow to larger sequential segments. However, when the scan is
2876 * checkpointing, no new extents will be added to the sorting queue,
2877 * so the way we are sorted now is as good as it will ever get.
2878 * In this case, we instead switch to issuing extents in LBA order.
2880 if (scn
->scn_checkpointing
) {
2881 return (range_tree_first(queue
->q_exts_by_addr
));
2882 } else if (scn
->scn_clearing
) {
2883 return (avl_first(&queue
->q_exts_by_size
));
2890 scan_io_queues_run_one(void *arg
)
2892 dsl_scan_io_queue_t
*queue
= arg
;
2893 kmutex_t
*q_lock
= &queue
->q_vd
->vdev_scan_io_queue_lock
;
2894 boolean_t suspended
= B_FALSE
;
2895 range_seg_t
*rs
= NULL
;
2896 scan_io_t
*sio
= NULL
;
2898 uint64_t bytes_per_leaf
= zfs_scan_vdev_limit
;
2899 uint64_t nr_leaves
= dsl_scan_count_leaves(queue
->q_vd
);
2901 ASSERT(queue
->q_scn
->scn_is_sorted
);
2903 list_create(&sio_list
, sizeof (scan_io_t
),
2904 offsetof(scan_io_t
, sio_nodes
.sio_list_node
));
2905 mutex_enter(q_lock
);
2907 /* calculate maximum in-flight bytes for this txg (min 1MB) */
2908 queue
->q_maxinflight_bytes
=
2909 MAX(nr_leaves
* bytes_per_leaf
, 1ULL << 20);
2911 /* reset per-queue scan statistics for this txg */
2912 queue
->q_total_seg_size_this_txg
= 0;
2913 queue
->q_segs_this_txg
= 0;
2914 queue
->q_total_zio_size_this_txg
= 0;
2915 queue
->q_zios_this_txg
= 0;
2917 /* loop until we run out of time or sios */
2918 while ((rs
= scan_io_queue_fetch_ext(queue
)) != NULL
) {
2919 uint64_t seg_start
= 0, seg_end
= 0;
2920 boolean_t more_left
= B_TRUE
;
2922 ASSERT(list_is_empty(&sio_list
));
2924 /* loop while we still have sios left to process in this rs */
2926 scan_io_t
*first_sio
, *last_sio
;
2929 * We have selected which extent needs to be
2930 * processed next. Gather up the corresponding sios.
2932 more_left
= scan_io_queue_gather(queue
, rs
, &sio_list
);
2933 ASSERT(!list_is_empty(&sio_list
));
2934 first_sio
= list_head(&sio_list
);
2935 last_sio
= list_tail(&sio_list
);
2937 seg_end
= SIO_GET_END_OFFSET(last_sio
);
2939 seg_start
= SIO_GET_OFFSET(first_sio
);
2942 * Issuing sios can take a long time so drop the
2943 * queue lock. The sio queue won't be updated by
2944 * other threads since we're in syncing context so
2945 * we can be sure that our trees will remain exactly
2949 suspended
= scan_io_queue_issue(queue
, &sio_list
);
2950 mutex_enter(q_lock
);
2956 /* update statistics for debugging purposes */
2957 scan_io_queues_update_seg_stats(queue
, seg_start
, seg_end
);
2964 * If we were suspended in the middle of processing,
2965 * requeue any unfinished sios and exit.
2967 while ((sio
= list_head(&sio_list
)) != NULL
) {
2968 list_remove(&sio_list
, sio
);
2969 scan_io_queue_insert_impl(queue
, sio
);
2973 list_destroy(&sio_list
);
2977 * Performs an emptying run on all scan queues in the pool. This just
2978 * punches out one thread per top-level vdev, each of which processes
2979 * only that vdev's scan queue. We can parallelize the I/O here because
2980 * we know that each queue's I/Os only affect its own top-level vdev.
2982 * This function waits for the queue runs to complete, and must be
2983 * called from dsl_scan_sync (or in general, syncing context).
2986 scan_io_queues_run(dsl_scan_t
*scn
)
2988 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
2990 ASSERT(scn
->scn_is_sorted
);
2991 ASSERT(spa_config_held(spa
, SCL_CONFIG
, RW_READER
));
2993 if (scn
->scn_bytes_pending
== 0)
2996 if (scn
->scn_taskq
== NULL
) {
2997 int nthreads
= spa
->spa_root_vdev
->vdev_children
;
3000 * We need to make this taskq *always* execute as many
3001 * threads in parallel as we have top-level vdevs and no
3002 * less, otherwise strange serialization of the calls to
3003 * scan_io_queues_run_one can occur during spa_sync runs
3004 * and that significantly impacts performance.
3006 scn
->scn_taskq
= taskq_create("dsl_scan_iss", nthreads
,
3007 minclsyspri
, nthreads
, nthreads
, TASKQ_PREPOPULATE
);
3010 for (uint64_t i
= 0; i
< spa
->spa_root_vdev
->vdev_children
; i
++) {
3011 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[i
];
3013 mutex_enter(&vd
->vdev_scan_io_queue_lock
);
3014 if (vd
->vdev_scan_io_queue
!= NULL
) {
3015 VERIFY(taskq_dispatch(scn
->scn_taskq
,
3016 scan_io_queues_run_one
, vd
->vdev_scan_io_queue
,
3017 TQ_SLEEP
) != TASKQID_INVALID
);
3019 mutex_exit(&vd
->vdev_scan_io_queue_lock
);
3023 * Wait for the queues to finish issuing their IOs for this run
3024 * before we return. There may still be IOs in flight at this
3027 taskq_wait(scn
->scn_taskq
);
3031 dsl_scan_async_block_should_pause(dsl_scan_t
*scn
)
3033 uint64_t elapsed_nanosecs
;
3038 if (zfs_async_block_max_blocks
!= 0 &&
3039 scn
->scn_visited_this_txg
>= zfs_async_block_max_blocks
) {
3043 elapsed_nanosecs
= gethrtime() - scn
->scn_sync_start_time
;
3044 return (elapsed_nanosecs
/ NANOSEC
> zfs_txg_timeout
||
3045 (NSEC2MSEC(elapsed_nanosecs
) > scn
->scn_async_block_min_time_ms
&&
3046 txg_sync_waiting(scn
->scn_dp
)) ||
3047 spa_shutting_down(scn
->scn_dp
->dp_spa
));
3051 dsl_scan_free_block_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
3053 dsl_scan_t
*scn
= arg
;
3055 if (!scn
->scn_is_bptree
||
3056 (BP_GET_LEVEL(bp
) == 0 && BP_GET_TYPE(bp
) != DMU_OT_OBJSET
)) {
3057 if (dsl_scan_async_block_should_pause(scn
))
3058 return (SET_ERROR(ERESTART
));
3061 zio_nowait(zio_free_sync(scn
->scn_zio_root
, scn
->scn_dp
->dp_spa
,
3062 dmu_tx_get_txg(tx
), bp
, 0));
3063 dsl_dir_diduse_space(tx
->tx_pool
->dp_free_dir
, DD_USED_HEAD
,
3064 -bp_get_dsize_sync(scn
->scn_dp
->dp_spa
, bp
),
3065 -BP_GET_PSIZE(bp
), -BP_GET_UCSIZE(bp
), tx
);
3066 scn
->scn_visited_this_txg
++;
3071 dsl_scan_update_stats(dsl_scan_t
*scn
)
3073 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3075 uint64_t seg_size_total
= 0, zio_size_total
= 0;
3076 uint64_t seg_count_total
= 0, zio_count_total
= 0;
3078 for (i
= 0; i
< spa
->spa_root_vdev
->vdev_children
; i
++) {
3079 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[i
];
3080 dsl_scan_io_queue_t
*queue
= vd
->vdev_scan_io_queue
;
3085 seg_size_total
+= queue
->q_total_seg_size_this_txg
;
3086 zio_size_total
+= queue
->q_total_zio_size_this_txg
;
3087 seg_count_total
+= queue
->q_segs_this_txg
;
3088 zio_count_total
+= queue
->q_zios_this_txg
;
3091 if (seg_count_total
== 0 || zio_count_total
== 0) {
3092 scn
->scn_avg_seg_size_this_txg
= 0;
3093 scn
->scn_avg_zio_size_this_txg
= 0;
3094 scn
->scn_segs_this_txg
= 0;
3095 scn
->scn_zios_this_txg
= 0;
3099 scn
->scn_avg_seg_size_this_txg
= seg_size_total
/ seg_count_total
;
3100 scn
->scn_avg_zio_size_this_txg
= zio_size_total
/ zio_count_total
;
3101 scn
->scn_segs_this_txg
= seg_count_total
;
3102 scn
->scn_zios_this_txg
= zio_count_total
;
3106 bpobj_dsl_scan_free_block_cb(void *arg
, const blkptr_t
*bp
, boolean_t bp_freed
,
3110 return (dsl_scan_free_block_cb(arg
, bp
, tx
));
3114 dsl_scan_obsolete_block_cb(void *arg
, const blkptr_t
*bp
, boolean_t bp_freed
,
3118 dsl_scan_t
*scn
= arg
;
3119 const dva_t
*dva
= &bp
->blk_dva
[0];
3121 if (dsl_scan_async_block_should_pause(scn
))
3122 return (SET_ERROR(ERESTART
));
3124 spa_vdev_indirect_mark_obsolete(scn
->scn_dp
->dp_spa
,
3125 DVA_GET_VDEV(dva
), DVA_GET_OFFSET(dva
),
3126 DVA_GET_ASIZE(dva
), tx
);
3127 scn
->scn_visited_this_txg
++;
3132 dsl_scan_active(dsl_scan_t
*scn
)
3134 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3135 uint64_t used
= 0, comp
, uncomp
;
3136 boolean_t clones_left
;
3138 if (spa
->spa_load_state
!= SPA_LOAD_NONE
)
3140 if (spa_shutting_down(spa
))
3142 if ((dsl_scan_is_running(scn
) && !dsl_scan_is_paused_scrub(scn
)) ||
3143 (scn
->scn_async_destroying
&& !scn
->scn_async_stalled
))
3146 if (spa_version(scn
->scn_dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
3147 (void) bpobj_space(&scn
->scn_dp
->dp_free_bpobj
,
3148 &used
, &comp
, &uncomp
);
3150 clones_left
= spa_livelist_delete_check(spa
);
3151 return ((used
!= 0) || (clones_left
));
3155 dsl_scan_check_deferred(vdev_t
*vd
)
3157 boolean_t need_resilver
= B_FALSE
;
3159 for (int c
= 0; c
< vd
->vdev_children
; c
++) {
3161 dsl_scan_check_deferred(vd
->vdev_child
[c
]);
3164 if (!vdev_is_concrete(vd
) || vd
->vdev_aux
||
3165 !vd
->vdev_ops
->vdev_op_leaf
)
3166 return (need_resilver
);
3168 if (!vd
->vdev_resilver_deferred
)
3169 need_resilver
= B_TRUE
;
3171 return (need_resilver
);
3175 dsl_scan_need_resilver(spa_t
*spa
, const dva_t
*dva
, size_t psize
,
3176 uint64_t phys_birth
)
3180 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(dva
));
3182 if (vd
->vdev_ops
== &vdev_indirect_ops
) {
3184 * The indirect vdev can point to multiple
3185 * vdevs. For simplicity, always create
3186 * the resilver zio_t. zio_vdev_io_start()
3187 * will bypass the child resilver i/o's if
3188 * they are on vdevs that don't have DTL's.
3193 if (DVA_GET_GANG(dva
)) {
3195 * Gang members may be spread across multiple
3196 * vdevs, so the best estimate we have is the
3197 * scrub range, which has already been checked.
3198 * XXX -- it would be better to change our
3199 * allocation policy to ensure that all
3200 * gang members reside on the same vdev.
3206 * Check if the txg falls within the range which must be
3207 * resilvered. DVAs outside this range can always be skipped.
3209 if (!vdev_dtl_contains(vd
, DTL_PARTIAL
, phys_birth
, 1))
3213 * Check if the top-level vdev must resilver this offset.
3214 * When the offset does not intersect with a dirty leaf DTL
3215 * then it may be possible to skip the resilver IO. The psize
3216 * is provided instead of asize to simplify the check for RAIDZ.
3218 if (!vdev_dtl_need_resilver(vd
, DVA_GET_OFFSET(dva
), psize
))
3222 * Check that this top-level vdev has a device under it which
3223 * is resilvering and is not deferred.
3225 if (!dsl_scan_check_deferred(vd
))
3232 dsl_process_async_destroys(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
3234 dsl_scan_t
*scn
= dp
->dp_scan
;
3235 spa_t
*spa
= dp
->dp_spa
;
3238 if (spa_suspend_async_destroy(spa
))
3241 if (zfs_free_bpobj_enabled
&&
3242 spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
3243 scn
->scn_is_bptree
= B_FALSE
;
3244 scn
->scn_async_block_min_time_ms
= zfs_free_min_time_ms
;
3245 scn
->scn_zio_root
= zio_root(spa
, NULL
,
3246 NULL
, ZIO_FLAG_MUSTSUCCEED
);
3247 err
= bpobj_iterate(&dp
->dp_free_bpobj
,
3248 bpobj_dsl_scan_free_block_cb
, scn
, tx
);
3249 VERIFY0(zio_wait(scn
->scn_zio_root
));
3250 scn
->scn_zio_root
= NULL
;
3252 if (err
!= 0 && err
!= ERESTART
)
3253 zfs_panic_recover("error %u from bpobj_iterate()", err
);
3256 if (err
== 0 && spa_feature_is_active(spa
, SPA_FEATURE_ASYNC_DESTROY
)) {
3257 ASSERT(scn
->scn_async_destroying
);
3258 scn
->scn_is_bptree
= B_TRUE
;
3259 scn
->scn_zio_root
= zio_root(spa
, NULL
,
3260 NULL
, ZIO_FLAG_MUSTSUCCEED
);
3261 err
= bptree_iterate(dp
->dp_meta_objset
,
3262 dp
->dp_bptree_obj
, B_TRUE
, dsl_scan_free_block_cb
, scn
, tx
);
3263 VERIFY0(zio_wait(scn
->scn_zio_root
));
3264 scn
->scn_zio_root
= NULL
;
3266 if (err
== EIO
|| err
== ECKSUM
) {
3268 } else if (err
!= 0 && err
!= ERESTART
) {
3269 zfs_panic_recover("error %u from "
3270 "traverse_dataset_destroyed()", err
);
3273 if (bptree_is_empty(dp
->dp_meta_objset
, dp
->dp_bptree_obj
)) {
3274 /* finished; deactivate async destroy feature */
3275 spa_feature_decr(spa
, SPA_FEATURE_ASYNC_DESTROY
, tx
);
3276 ASSERT(!spa_feature_is_active(spa
,
3277 SPA_FEATURE_ASYNC_DESTROY
));
3278 VERIFY0(zap_remove(dp
->dp_meta_objset
,
3279 DMU_POOL_DIRECTORY_OBJECT
,
3280 DMU_POOL_BPTREE_OBJ
, tx
));
3281 VERIFY0(bptree_free(dp
->dp_meta_objset
,
3282 dp
->dp_bptree_obj
, tx
));
3283 dp
->dp_bptree_obj
= 0;
3284 scn
->scn_async_destroying
= B_FALSE
;
3285 scn
->scn_async_stalled
= B_FALSE
;
3288 * If we didn't make progress, mark the async
3289 * destroy as stalled, so that we will not initiate
3290 * a spa_sync() on its behalf. Note that we only
3291 * check this if we are not finished, because if the
3292 * bptree had no blocks for us to visit, we can
3293 * finish without "making progress".
3295 scn
->scn_async_stalled
=
3296 (scn
->scn_visited_this_txg
== 0);
3299 if (scn
->scn_visited_this_txg
) {
3300 zfs_dbgmsg("freed %llu blocks in %llums from "
3301 "free_bpobj/bptree txg %llu; err=%u",
3302 (longlong_t
)scn
->scn_visited_this_txg
,
3304 NSEC2MSEC(gethrtime() - scn
->scn_sync_start_time
),
3305 (longlong_t
)tx
->tx_txg
, err
);
3306 scn
->scn_visited_this_txg
= 0;
3309 * Write out changes to the DDT that may be required as a
3310 * result of the blocks freed. This ensures that the DDT
3311 * is clean when a scrub/resilver runs.
3313 ddt_sync(spa
, tx
->tx_txg
);
3317 if (dp
->dp_free_dir
!= NULL
&& !scn
->scn_async_destroying
&&
3318 zfs_free_leak_on_eio
&&
3319 (dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
!= 0 ||
3320 dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
!= 0 ||
3321 dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
!= 0)) {
3323 * We have finished background destroying, but there is still
3324 * some space left in the dp_free_dir. Transfer this leaked
3325 * space to the dp_leak_dir.
3327 if (dp
->dp_leak_dir
== NULL
) {
3328 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
3329 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
3331 VERIFY0(dsl_pool_open_special_dir(dp
,
3332 LEAK_DIR_NAME
, &dp
->dp_leak_dir
));
3333 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
3335 dsl_dir_diduse_space(dp
->dp_leak_dir
, DD_USED_HEAD
,
3336 dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
,
3337 dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
,
3338 dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
, tx
);
3339 dsl_dir_diduse_space(dp
->dp_free_dir
, DD_USED_HEAD
,
3340 -dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
,
3341 -dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
,
3342 -dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
, tx
);
3345 if (dp
->dp_free_dir
!= NULL
&& !scn
->scn_async_destroying
&&
3346 !spa_livelist_delete_check(spa
)) {
3347 /* finished; verify that space accounting went to zero */
3348 ASSERT0(dsl_dir_phys(dp
->dp_free_dir
)->dd_used_bytes
);
3349 ASSERT0(dsl_dir_phys(dp
->dp_free_dir
)->dd_compressed_bytes
);
3350 ASSERT0(dsl_dir_phys(dp
->dp_free_dir
)->dd_uncompressed_bytes
);
3353 EQUIV(bpobj_is_open(&dp
->dp_obsolete_bpobj
),
3354 0 == zap_contains(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
3355 DMU_POOL_OBSOLETE_BPOBJ
));
3356 if (err
== 0 && bpobj_is_open(&dp
->dp_obsolete_bpobj
)) {
3357 ASSERT(spa_feature_is_active(dp
->dp_spa
,
3358 SPA_FEATURE_OBSOLETE_COUNTS
));
3360 scn
->scn_is_bptree
= B_FALSE
;
3361 scn
->scn_async_block_min_time_ms
= zfs_obsolete_min_time_ms
;
3362 err
= bpobj_iterate(&dp
->dp_obsolete_bpobj
,
3363 dsl_scan_obsolete_block_cb
, scn
, tx
);
3364 if (err
!= 0 && err
!= ERESTART
)
3365 zfs_panic_recover("error %u from bpobj_iterate()", err
);
3367 if (bpobj_is_empty(&dp
->dp_obsolete_bpobj
))
3368 dsl_pool_destroy_obsolete_bpobj(dp
, tx
);
3374 * This is the primary entry point for scans that is called from syncing
3375 * context. Scans must happen entirely during syncing context so that we
3376 * can guarantee that blocks we are currently scanning will not change out
3377 * from under us. While a scan is active, this function controls how quickly
3378 * transaction groups proceed, instead of the normal handling provided by
3379 * txg_sync_thread().
3382 dsl_scan_sync(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
3385 dsl_scan_t
*scn
= dp
->dp_scan
;
3386 spa_t
*spa
= dp
->dp_spa
;
3387 state_sync_type_t sync_type
= SYNC_OPTIONAL
;
3389 if (spa
->spa_resilver_deferred
&&
3390 !spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_RESILVER_DEFER
))
3391 spa_feature_incr(spa
, SPA_FEATURE_RESILVER_DEFER
, tx
);
3394 * Check for scn_restart_txg before checking spa_load_state, so
3395 * that we can restart an old-style scan while the pool is being
3396 * imported (see dsl_scan_init). We also restart scans if there
3397 * is a deferred resilver and the user has manually disabled
3398 * deferred resilvers via the tunable.
3400 if (dsl_scan_restarting(scn
, tx
) ||
3401 (spa
->spa_resilver_deferred
&& zfs_resilver_disable_defer
)) {
3402 pool_scan_func_t func
= POOL_SCAN_SCRUB
;
3403 dsl_scan_done(scn
, B_FALSE
, tx
);
3404 if (vdev_resilver_needed(spa
->spa_root_vdev
, NULL
, NULL
))
3405 func
= POOL_SCAN_RESILVER
;
3406 zfs_dbgmsg("restarting scan func=%u txg=%llu",
3407 func
, (longlong_t
)tx
->tx_txg
);
3408 dsl_scan_setup_sync(&func
, tx
);
3412 * Only process scans in sync pass 1.
3414 if (spa_sync_pass(spa
) > 1)
3418 * If the spa is shutting down, then stop scanning. This will
3419 * ensure that the scan does not dirty any new data during the
3422 if (spa_shutting_down(spa
))
3426 * If the scan is inactive due to a stalled async destroy, try again.
3428 if (!scn
->scn_async_stalled
&& !dsl_scan_active(scn
))
3431 /* reset scan statistics */
3432 scn
->scn_visited_this_txg
= 0;
3433 scn
->scn_holes_this_txg
= 0;
3434 scn
->scn_lt_min_this_txg
= 0;
3435 scn
->scn_gt_max_this_txg
= 0;
3436 scn
->scn_ddt_contained_this_txg
= 0;
3437 scn
->scn_objsets_visited_this_txg
= 0;
3438 scn
->scn_avg_seg_size_this_txg
= 0;
3439 scn
->scn_segs_this_txg
= 0;
3440 scn
->scn_avg_zio_size_this_txg
= 0;
3441 scn
->scn_zios_this_txg
= 0;
3442 scn
->scn_suspending
= B_FALSE
;
3443 scn
->scn_sync_start_time
= gethrtime();
3444 spa
->spa_scrub_active
= B_TRUE
;
3447 * First process the async destroys. If we suspend, don't do
3448 * any scrubbing or resilvering. This ensures that there are no
3449 * async destroys while we are scanning, so the scan code doesn't
3450 * have to worry about traversing it. It is also faster to free the
3451 * blocks than to scrub them.
3453 err
= dsl_process_async_destroys(dp
, tx
);
3457 if (!dsl_scan_is_running(scn
) || dsl_scan_is_paused_scrub(scn
))
3461 * Wait a few txgs after importing to begin scanning so that
3462 * we can get the pool imported quickly.
3464 if (spa
->spa_syncing_txg
< spa
->spa_first_txg
+ SCAN_IMPORT_WAIT_TXGS
)
3468 * zfs_scan_suspend_progress can be set to disable scan progress.
3469 * We don't want to spin the txg_sync thread, so we add a delay
3470 * here to simulate the time spent doing a scan. This is mostly
3471 * useful for testing and debugging.
3473 if (zfs_scan_suspend_progress
) {
3474 uint64_t scan_time_ns
= gethrtime() - scn
->scn_sync_start_time
;
3475 int mintime
= (scn
->scn_phys
.scn_func
== POOL_SCAN_RESILVER
) ?
3476 zfs_resilver_min_time_ms
: zfs_scrub_min_time_ms
;
3478 while (zfs_scan_suspend_progress
&&
3479 !txg_sync_waiting(scn
->scn_dp
) &&
3480 !spa_shutting_down(scn
->scn_dp
->dp_spa
) &&
3481 NSEC2MSEC(scan_time_ns
) < mintime
) {
3483 scan_time_ns
= gethrtime() - scn
->scn_sync_start_time
;
3489 * It is possible to switch from unsorted to sorted at any time,
3490 * but afterwards the scan will remain sorted unless reloaded from
3491 * a checkpoint after a reboot.
3493 if (!zfs_scan_legacy
) {
3494 scn
->scn_is_sorted
= B_TRUE
;
3495 if (scn
->scn_last_checkpoint
== 0)
3496 scn
->scn_last_checkpoint
= ddi_get_lbolt();
3500 * For sorted scans, determine what kind of work we will be doing
3501 * this txg based on our memory limitations and whether or not we
3502 * need to perform a checkpoint.
3504 if (scn
->scn_is_sorted
) {
3506 * If we are over our checkpoint interval, set scn_clearing
3507 * so that we can begin checkpointing immediately. The
3508 * checkpoint allows us to save a consistent bookmark
3509 * representing how much data we have scrubbed so far.
3510 * Otherwise, use the memory limit to determine if we should
3511 * scan for metadata or start issue scrub IOs. We accumulate
3512 * metadata until we hit our hard memory limit at which point
3513 * we issue scrub IOs until we are at our soft memory limit.
3515 if (scn
->scn_checkpointing
||
3516 ddi_get_lbolt() - scn
->scn_last_checkpoint
>
3517 SEC_TO_TICK(zfs_scan_checkpoint_intval
)) {
3518 if (!scn
->scn_checkpointing
)
3519 zfs_dbgmsg("begin scan checkpoint");
3521 scn
->scn_checkpointing
= B_TRUE
;
3522 scn
->scn_clearing
= B_TRUE
;
3524 boolean_t should_clear
= dsl_scan_should_clear(scn
);
3525 if (should_clear
&& !scn
->scn_clearing
) {
3526 zfs_dbgmsg("begin scan clearing");
3527 scn
->scn_clearing
= B_TRUE
;
3528 } else if (!should_clear
&& scn
->scn_clearing
) {
3529 zfs_dbgmsg("finish scan clearing");
3530 scn
->scn_clearing
= B_FALSE
;
3534 ASSERT0(scn
->scn_checkpointing
);
3535 ASSERT0(scn
->scn_clearing
);
3538 if (!scn
->scn_clearing
&& scn
->scn_done_txg
== 0) {
3539 /* Need to scan metadata for more blocks to scrub */
3540 dsl_scan_phys_t
*scnp
= &scn
->scn_phys
;
3541 taskqid_t prefetch_tqid
;
3542 uint64_t bytes_per_leaf
= zfs_scan_vdev_limit
;
3543 uint64_t nr_leaves
= dsl_scan_count_leaves(spa
->spa_root_vdev
);
3546 * Recalculate the max number of in-flight bytes for pool-wide
3547 * scanning operations (minimum 1MB). Limits for the issuing
3548 * phase are done per top-level vdev and are handled separately.
3550 scn
->scn_maxinflight_bytes
=
3551 MAX(nr_leaves
* bytes_per_leaf
, 1ULL << 20);
3553 if (scnp
->scn_ddt_bookmark
.ddb_class
<=
3554 scnp
->scn_ddt_class_max
) {
3555 ASSERT(ZB_IS_ZERO(&scnp
->scn_bookmark
));
3556 zfs_dbgmsg("doing scan sync txg %llu; "
3557 "ddt bm=%llu/%llu/%llu/%llx",
3558 (longlong_t
)tx
->tx_txg
,
3559 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_class
,
3560 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_type
,
3561 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_checksum
,
3562 (longlong_t
)scnp
->scn_ddt_bookmark
.ddb_cursor
);
3564 zfs_dbgmsg("doing scan sync txg %llu; "
3565 "bm=%llu/%llu/%llu/%llu",
3566 (longlong_t
)tx
->tx_txg
,
3567 (longlong_t
)scnp
->scn_bookmark
.zb_objset
,
3568 (longlong_t
)scnp
->scn_bookmark
.zb_object
,
3569 (longlong_t
)scnp
->scn_bookmark
.zb_level
,
3570 (longlong_t
)scnp
->scn_bookmark
.zb_blkid
);
3573 scn
->scn_zio_root
= zio_root(dp
->dp_spa
, NULL
,
3574 NULL
, ZIO_FLAG_CANFAIL
);
3576 scn
->scn_prefetch_stop
= B_FALSE
;
3577 prefetch_tqid
= taskq_dispatch(dp
->dp_sync_taskq
,
3578 dsl_scan_prefetch_thread
, scn
, TQ_SLEEP
);
3579 ASSERT(prefetch_tqid
!= TASKQID_INVALID
);
3581 dsl_pool_config_enter(dp
, FTAG
);
3582 dsl_scan_visit(scn
, tx
);
3583 dsl_pool_config_exit(dp
, FTAG
);
3585 mutex_enter(&dp
->dp_spa
->spa_scrub_lock
);
3586 scn
->scn_prefetch_stop
= B_TRUE
;
3587 cv_broadcast(&spa
->spa_scrub_io_cv
);
3588 mutex_exit(&dp
->dp_spa
->spa_scrub_lock
);
3590 taskq_wait_id(dp
->dp_sync_taskq
, prefetch_tqid
);
3591 (void) zio_wait(scn
->scn_zio_root
);
3592 scn
->scn_zio_root
= NULL
;
3594 zfs_dbgmsg("scan visited %llu blocks in %llums "
3595 "(%llu os's, %llu holes, %llu < mintxg, "
3596 "%llu in ddt, %llu > maxtxg)",
3597 (longlong_t
)scn
->scn_visited_this_txg
,
3598 (longlong_t
)NSEC2MSEC(gethrtime() -
3599 scn
->scn_sync_start_time
),
3600 (longlong_t
)scn
->scn_objsets_visited_this_txg
,
3601 (longlong_t
)scn
->scn_holes_this_txg
,
3602 (longlong_t
)scn
->scn_lt_min_this_txg
,
3603 (longlong_t
)scn
->scn_ddt_contained_this_txg
,
3604 (longlong_t
)scn
->scn_gt_max_this_txg
);
3606 if (!scn
->scn_suspending
) {
3607 ASSERT0(avl_numnodes(&scn
->scn_queue
));
3608 scn
->scn_done_txg
= tx
->tx_txg
+ 1;
3609 if (scn
->scn_is_sorted
) {
3610 scn
->scn_checkpointing
= B_TRUE
;
3611 scn
->scn_clearing
= B_TRUE
;
3613 zfs_dbgmsg("scan complete txg %llu",
3614 (longlong_t
)tx
->tx_txg
);
3616 } else if (scn
->scn_is_sorted
&& scn
->scn_bytes_pending
!= 0) {
3617 ASSERT(scn
->scn_clearing
);
3619 /* need to issue scrubbing IOs from per-vdev queues */
3620 scn
->scn_zio_root
= zio_root(dp
->dp_spa
, NULL
,
3621 NULL
, ZIO_FLAG_CANFAIL
);
3622 scan_io_queues_run(scn
);
3623 (void) zio_wait(scn
->scn_zio_root
);
3624 scn
->scn_zio_root
= NULL
;
3626 /* calculate and dprintf the current memory usage */
3627 (void) dsl_scan_should_clear(scn
);
3628 dsl_scan_update_stats(scn
);
3630 zfs_dbgmsg("scan issued %llu blocks (%llu segs) in %llums "
3631 "(avg_block_size = %llu, avg_seg_size = %llu)",
3632 (longlong_t
)scn
->scn_zios_this_txg
,
3633 (longlong_t
)scn
->scn_segs_this_txg
,
3634 (longlong_t
)NSEC2MSEC(gethrtime() -
3635 scn
->scn_sync_start_time
),
3636 (longlong_t
)scn
->scn_avg_zio_size_this_txg
,
3637 (longlong_t
)scn
->scn_avg_seg_size_this_txg
);
3638 } else if (scn
->scn_done_txg
!= 0 && scn
->scn_done_txg
<= tx
->tx_txg
) {
3639 /* Finished with everything. Mark the scrub as complete */
3640 zfs_dbgmsg("scan issuing complete txg %llu",
3641 (longlong_t
)tx
->tx_txg
);
3642 ASSERT3U(scn
->scn_done_txg
, !=, 0);
3643 ASSERT0(spa
->spa_scrub_inflight
);
3644 ASSERT0(scn
->scn_bytes_pending
);
3645 dsl_scan_done(scn
, B_TRUE
, tx
);
3646 sync_type
= SYNC_MANDATORY
;
3649 dsl_scan_sync_state(scn
, tx
, sync_type
);
3653 count_block(dsl_scan_t
*scn
, zfs_all_blkstats_t
*zab
, const blkptr_t
*bp
)
3658 * Don't count embedded bp's, since we already did the work of
3659 * scanning these when we scanned the containing block.
3661 if (BP_IS_EMBEDDED(bp
))
3665 * Update the spa's stats on how many bytes we have issued.
3666 * Sequential scrubs create a zio for each DVA of the bp. Each
3667 * of these will include all DVAs for repair purposes, but the
3668 * zio code will only try the first one unless there is an issue.
3669 * Therefore, we should only count the first DVA for these IOs.
3671 if (scn
->scn_is_sorted
) {
3672 atomic_add_64(&scn
->scn_dp
->dp_spa
->spa_scan_pass_issued
,
3673 DVA_GET_ASIZE(&bp
->blk_dva
[0]));
3675 spa_t
*spa
= scn
->scn_dp
->dp_spa
;
3677 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
3678 atomic_add_64(&spa
->spa_scan_pass_issued
,
3679 DVA_GET_ASIZE(&bp
->blk_dva
[i
]));
3684 * If we resume after a reboot, zab will be NULL; don't record
3685 * incomplete stats in that case.
3690 mutex_enter(&zab
->zab_lock
);
3692 for (i
= 0; i
< 4; i
++) {
3693 int l
= (i
< 2) ? BP_GET_LEVEL(bp
) : DN_MAX_LEVELS
;
3694 int t
= (i
& 1) ? BP_GET_TYPE(bp
) : DMU_OT_TOTAL
;
3696 if (t
& DMU_OT_NEWTYPE
)
3698 zfs_blkstat_t
*zb
= &zab
->zab_type
[l
][t
];
3702 zb
->zb_asize
+= BP_GET_ASIZE(bp
);
3703 zb
->zb_lsize
+= BP_GET_LSIZE(bp
);
3704 zb
->zb_psize
+= BP_GET_PSIZE(bp
);
3705 zb
->zb_gangs
+= BP_COUNT_GANG(bp
);
3707 switch (BP_GET_NDVAS(bp
)) {
3709 if (DVA_GET_VDEV(&bp
->blk_dva
[0]) ==
3710 DVA_GET_VDEV(&bp
->blk_dva
[1]))
3711 zb
->zb_ditto_2_of_2_samevdev
++;
3714 equal
= (DVA_GET_VDEV(&bp
->blk_dva
[0]) ==
3715 DVA_GET_VDEV(&bp
->blk_dva
[1])) +
3716 (DVA_GET_VDEV(&bp
->blk_dva
[0]) ==
3717 DVA_GET_VDEV(&bp
->blk_dva
[2])) +
3718 (DVA_GET_VDEV(&bp
->blk_dva
[1]) ==
3719 DVA_GET_VDEV(&bp
->blk_dva
[2]));
3721 zb
->zb_ditto_2_of_3_samevdev
++;
3722 else if (equal
== 3)
3723 zb
->zb_ditto_3_of_3_samevdev
++;
3728 mutex_exit(&zab
->zab_lock
);
3732 scan_io_queue_insert_impl(dsl_scan_io_queue_t
*queue
, scan_io_t
*sio
)
3735 int64_t asize
= SIO_GET_ASIZE(sio
);
3736 dsl_scan_t
*scn
= queue
->q_scn
;
3738 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
3740 if (avl_find(&queue
->q_sios_by_addr
, sio
, &idx
) != NULL
) {
3741 /* block is already scheduled for reading */
3742 atomic_add_64(&scn
->scn_bytes_pending
, -asize
);
3746 avl_insert(&queue
->q_sios_by_addr
, sio
, idx
);
3747 queue
->q_sio_memused
+= SIO_GET_MUSED(sio
);
3748 range_tree_add(queue
->q_exts_by_addr
, SIO_GET_OFFSET(sio
), asize
);
3752 * Given all the info we got from our metadata scanning process, we
3753 * construct a scan_io_t and insert it into the scan sorting queue. The
3754 * I/O must already be suitable for us to process. This is controlled
3755 * by dsl_scan_enqueue().
3758 scan_io_queue_insert(dsl_scan_io_queue_t
*queue
, const blkptr_t
*bp
, int dva_i
,
3759 int zio_flags
, const zbookmark_phys_t
*zb
)
3761 dsl_scan_t
*scn
= queue
->q_scn
;
3762 scan_io_t
*sio
= sio_alloc(BP_GET_NDVAS(bp
));
3764 ASSERT0(BP_IS_GANG(bp
));
3765 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
3767 bp2sio(bp
, sio
, dva_i
);
3768 sio
->sio_flags
= zio_flags
;
3772 * Increment the bytes pending counter now so that we can't
3773 * get an integer underflow in case the worker processes the
3774 * zio before we get to incrementing this counter.
3776 atomic_add_64(&scn
->scn_bytes_pending
, SIO_GET_ASIZE(sio
));
3778 scan_io_queue_insert_impl(queue
, sio
);
3782 * Given a set of I/O parameters as discovered by the metadata traversal
3783 * process, attempts to place the I/O into the sorted queues (if allowed),
3784 * or immediately executes the I/O.
3787 dsl_scan_enqueue(dsl_pool_t
*dp
, const blkptr_t
*bp
, int zio_flags
,
3788 const zbookmark_phys_t
*zb
)
3790 spa_t
*spa
= dp
->dp_spa
;
3792 ASSERT(!BP_IS_EMBEDDED(bp
));
3795 * Gang blocks are hard to issue sequentially, so we just issue them
3796 * here immediately instead of queuing them.
3798 if (!dp
->dp_scan
->scn_is_sorted
|| BP_IS_GANG(bp
)) {
3799 scan_exec_io(dp
, bp
, zio_flags
, zb
, NULL
);
3803 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
3807 dva
= bp
->blk_dva
[i
];
3808 vdev
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
));
3809 ASSERT(vdev
!= NULL
);
3811 mutex_enter(&vdev
->vdev_scan_io_queue_lock
);
3812 if (vdev
->vdev_scan_io_queue
== NULL
)
3813 vdev
->vdev_scan_io_queue
= scan_io_queue_create(vdev
);
3814 ASSERT(dp
->dp_scan
!= NULL
);
3815 scan_io_queue_insert(vdev
->vdev_scan_io_queue
, bp
,
3817 mutex_exit(&vdev
->vdev_scan_io_queue_lock
);
3822 dsl_scan_scrub_cb(dsl_pool_t
*dp
,
3823 const blkptr_t
*bp
, const zbookmark_phys_t
*zb
)
3825 dsl_scan_t
*scn
= dp
->dp_scan
;
3826 spa_t
*spa
= dp
->dp_spa
;
3827 uint64_t phys_birth
= BP_PHYSICAL_BIRTH(bp
);
3828 size_t psize
= BP_GET_PSIZE(bp
);
3829 boolean_t needs_io
= B_FALSE
;
3830 int zio_flags
= ZIO_FLAG_SCAN_THREAD
| ZIO_FLAG_RAW
| ZIO_FLAG_CANFAIL
;
3833 if (phys_birth
<= scn
->scn_phys
.scn_min_txg
||
3834 phys_birth
>= scn
->scn_phys
.scn_max_txg
) {
3835 count_block(scn
, dp
->dp_blkstats
, bp
);
3839 /* Embedded BP's have phys_birth==0, so we reject them above. */
3840 ASSERT(!BP_IS_EMBEDDED(bp
));
3842 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn
));
3843 if (scn
->scn_phys
.scn_func
== POOL_SCAN_SCRUB
) {
3844 zio_flags
|= ZIO_FLAG_SCRUB
;
3847 ASSERT3U(scn
->scn_phys
.scn_func
, ==, POOL_SCAN_RESILVER
);
3848 zio_flags
|= ZIO_FLAG_RESILVER
;
3852 /* If it's an intent log block, failure is expected. */
3853 if (zb
->zb_level
== ZB_ZIL_LEVEL
)
3854 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
3856 for (int d
= 0; d
< BP_GET_NDVAS(bp
); d
++) {
3857 const dva_t
*dva
= &bp
->blk_dva
[d
];
3860 * Keep track of how much data we've examined so that
3861 * zpool(1M) status can make useful progress reports.
3863 scn
->scn_phys
.scn_examined
+= DVA_GET_ASIZE(dva
);
3864 spa
->spa_scan_pass_exam
+= DVA_GET_ASIZE(dva
);
3866 /* if it's a resilver, this may not be in the target range */
3868 needs_io
= dsl_scan_need_resilver(spa
, dva
, psize
,
3872 if (needs_io
&& !zfs_no_scrub_io
) {
3873 dsl_scan_enqueue(dp
, bp
, zio_flags
, zb
);
3875 count_block(scn
, dp
->dp_blkstats
, bp
);
3878 /* do not relocate this block */
3883 dsl_scan_scrub_done(zio_t
*zio
)
3885 spa_t
*spa
= zio
->io_spa
;
3886 blkptr_t
*bp
= zio
->io_bp
;
3887 dsl_scan_io_queue_t
*queue
= zio
->io_private
;
3889 abd_free(zio
->io_abd
);
3891 if (queue
== NULL
) {
3892 mutex_enter(&spa
->spa_scrub_lock
);
3893 ASSERT3U(spa
->spa_scrub_inflight
, >=, BP_GET_PSIZE(bp
));
3894 spa
->spa_scrub_inflight
-= BP_GET_PSIZE(bp
);
3895 cv_broadcast(&spa
->spa_scrub_io_cv
);
3896 mutex_exit(&spa
->spa_scrub_lock
);
3898 mutex_enter(&queue
->q_vd
->vdev_scan_io_queue_lock
);
3899 ASSERT3U(queue
->q_inflight_bytes
, >=, BP_GET_PSIZE(bp
));
3900 queue
->q_inflight_bytes
-= BP_GET_PSIZE(bp
);
3901 cv_broadcast(&queue
->q_zio_cv
);
3902 mutex_exit(&queue
->q_vd
->vdev_scan_io_queue_lock
);
3905 if (zio
->io_error
&& (zio
->io_error
!= ECKSUM
||
3906 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
))) {
3907 atomic_inc_64(&spa
->spa_dsl_pool
->dp_scan
->scn_phys
.scn_errors
);
3912 * Given a scanning zio's information, executes the zio. The zio need
3913 * not necessarily be only sortable, this function simply executes the
3914 * zio, no matter what it is. The optional queue argument allows the
3915 * caller to specify that they want per top level vdev IO rate limiting
3916 * instead of the legacy global limiting.
3919 scan_exec_io(dsl_pool_t
*dp
, const blkptr_t
*bp
, int zio_flags
,
3920 const zbookmark_phys_t
*zb
, dsl_scan_io_queue_t
*queue
)
3922 spa_t
*spa
= dp
->dp_spa
;
3923 dsl_scan_t
*scn
= dp
->dp_scan
;
3924 size_t size
= BP_GET_PSIZE(bp
);
3925 abd_t
*data
= abd_alloc_for_io(size
, B_FALSE
);
3927 ASSERT3U(scn
->scn_maxinflight_bytes
, >, 0);
3929 if (queue
== NULL
) {
3930 mutex_enter(&spa
->spa_scrub_lock
);
3931 while (spa
->spa_scrub_inflight
>= scn
->scn_maxinflight_bytes
)
3932 cv_wait(&spa
->spa_scrub_io_cv
, &spa
->spa_scrub_lock
);
3933 spa
->spa_scrub_inflight
+= BP_GET_PSIZE(bp
);
3934 mutex_exit(&spa
->spa_scrub_lock
);
3936 kmutex_t
*q_lock
= &queue
->q_vd
->vdev_scan_io_queue_lock
;
3938 mutex_enter(q_lock
);
3939 while (queue
->q_inflight_bytes
>= queue
->q_maxinflight_bytes
)
3940 cv_wait(&queue
->q_zio_cv
, q_lock
);
3941 queue
->q_inflight_bytes
+= BP_GET_PSIZE(bp
);
3945 count_block(scn
, dp
->dp_blkstats
, bp
);
3946 zio_nowait(zio_read(scn
->scn_zio_root
, spa
, bp
, data
, size
,
3947 dsl_scan_scrub_done
, queue
, ZIO_PRIORITY_SCRUB
, zio_flags
, zb
));
3951 * This is the primary extent sorting algorithm. We balance two parameters:
3952 * 1) how many bytes of I/O are in an extent
3953 * 2) how well the extent is filled with I/O (as a fraction of its total size)
3954 * Since we allow extents to have gaps between their constituent I/Os, it's
3955 * possible to have a fairly large extent that contains the same amount of
3956 * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
3957 * The algorithm sorts based on a score calculated from the extent's size,
3958 * the relative fill volume (in %) and a "fill weight" parameter that controls
3959 * the split between whether we prefer larger extents or more well populated
3962 * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
3965 * 1) assume extsz = 64 MiB
3966 * 2) assume fill = 32 MiB (extent is half full)
3967 * 3) assume fill_weight = 3
3968 * 4) SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
3969 * SCORE = 32M + (50 * 3 * 32M) / 100
3970 * SCORE = 32M + (4800M / 100)
3973 * | +--- final total relative fill-based score
3974 * +--------- final total fill-based score
3977 * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
3978 * extents that are more completely filled (in a 3:2 ratio) vs just larger.
3979 * Note that as an optimization, we replace multiplication and division by
3980 * 100 with bitshifting by 7 (which effectively multiplies and divides by 128).
3983 ext_size_compare(const void *x
, const void *y
)
3985 const range_seg_t
*rsa
= x
, *rsb
= y
;
3986 uint64_t sa
= rsa
->rs_end
- rsa
->rs_start
,
3987 sb
= rsb
->rs_end
- rsb
->rs_start
;
3988 uint64_t score_a
, score_b
;
3990 score_a
= rsa
->rs_fill
+ ((((rsa
->rs_fill
<< 7) / sa
) *
3991 fill_weight
* rsa
->rs_fill
) >> 7);
3992 score_b
= rsb
->rs_fill
+ ((((rsb
->rs_fill
<< 7) / sb
) *
3993 fill_weight
* rsb
->rs_fill
) >> 7);
3995 if (score_a
> score_b
)
3997 if (score_a
== score_b
) {
3998 if (rsa
->rs_start
< rsb
->rs_start
)
4000 if (rsa
->rs_start
== rsb
->rs_start
)
4008 * Comparator for the q_sios_by_addr tree. Sorting is simply performed
4009 * based on LBA-order (from lowest to highest).
4012 sio_addr_compare(const void *x
, const void *y
)
4014 const scan_io_t
*a
= x
, *b
= y
;
4016 return (AVL_CMP(SIO_GET_OFFSET(a
), SIO_GET_OFFSET(b
)));
4019 /* IO queues are created on demand when they are needed. */
4020 static dsl_scan_io_queue_t
*
4021 scan_io_queue_create(vdev_t
*vd
)
4023 dsl_scan_t
*scn
= vd
->vdev_spa
->spa_dsl_pool
->dp_scan
;
4024 dsl_scan_io_queue_t
*q
= kmem_zalloc(sizeof (*q
), KM_SLEEP
);
4028 q
->q_sio_memused
= 0;
4029 cv_init(&q
->q_zio_cv
, NULL
, CV_DEFAULT
, NULL
);
4030 q
->q_exts_by_addr
= range_tree_create_impl(&rt_avl_ops
,
4031 &q
->q_exts_by_size
, ext_size_compare
, zfs_scan_max_ext_gap
);
4032 avl_create(&q
->q_sios_by_addr
, sio_addr_compare
,
4033 sizeof (scan_io_t
), offsetof(scan_io_t
, sio_nodes
.sio_addr_node
));
4039 * Destroys a scan queue and all segments and scan_io_t's contained in it.
4040 * No further execution of I/O occurs, anything pending in the queue is
4041 * simply freed without being executed.
4044 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t
*queue
)
4046 dsl_scan_t
*scn
= queue
->q_scn
;
4048 void *cookie
= NULL
;
4049 int64_t bytes_dequeued
= 0;
4051 ASSERT(MUTEX_HELD(&queue
->q_vd
->vdev_scan_io_queue_lock
));
4053 while ((sio
= avl_destroy_nodes(&queue
->q_sios_by_addr
, &cookie
)) !=
4055 ASSERT(range_tree_contains(queue
->q_exts_by_addr
,
4056 SIO_GET_OFFSET(sio
), SIO_GET_ASIZE(sio
)));
4057 bytes_dequeued
+= SIO_GET_ASIZE(sio
);
4058 queue
->q_sio_memused
-= SIO_GET_MUSED(sio
);
4062 ASSERT0(queue
->q_sio_memused
);
4063 atomic_add_64(&scn
->scn_bytes_pending
, -bytes_dequeued
);
4064 range_tree_vacate(queue
->q_exts_by_addr
, NULL
, queue
);
4065 range_tree_destroy(queue
->q_exts_by_addr
);
4066 avl_destroy(&queue
->q_sios_by_addr
);
4067 cv_destroy(&queue
->q_zio_cv
);
4069 kmem_free(queue
, sizeof (*queue
));
4073 * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
4074 * called on behalf of vdev_top_transfer when creating or destroying
4075 * a mirror vdev due to zpool attach/detach.
4078 dsl_scan_io_queue_vdev_xfer(vdev_t
*svd
, vdev_t
*tvd
)
4080 mutex_enter(&svd
->vdev_scan_io_queue_lock
);
4081 mutex_enter(&tvd
->vdev_scan_io_queue_lock
);
4083 VERIFY3P(tvd
->vdev_scan_io_queue
, ==, NULL
);
4084 tvd
->vdev_scan_io_queue
= svd
->vdev_scan_io_queue
;
4085 svd
->vdev_scan_io_queue
= NULL
;
4086 if (tvd
->vdev_scan_io_queue
!= NULL
)
4087 tvd
->vdev_scan_io_queue
->q_vd
= tvd
;
4089 mutex_exit(&tvd
->vdev_scan_io_queue_lock
);
4090 mutex_exit(&svd
->vdev_scan_io_queue_lock
);
4094 scan_io_queues_destroy(dsl_scan_t
*scn
)
4096 vdev_t
*rvd
= scn
->scn_dp
->dp_spa
->spa_root_vdev
;
4098 for (uint64_t i
= 0; i
< rvd
->vdev_children
; i
++) {
4099 vdev_t
*tvd
= rvd
->vdev_child
[i
];
4101 mutex_enter(&tvd
->vdev_scan_io_queue_lock
);
4102 if (tvd
->vdev_scan_io_queue
!= NULL
)
4103 dsl_scan_io_queue_destroy(tvd
->vdev_scan_io_queue
);
4104 tvd
->vdev_scan_io_queue
= NULL
;
4105 mutex_exit(&tvd
->vdev_scan_io_queue_lock
);
4110 dsl_scan_freed_dva(spa_t
*spa
, const blkptr_t
*bp
, int dva_i
)
4112 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
4113 dsl_scan_t
*scn
= dp
->dp_scan
;
4116 dsl_scan_io_queue_t
*queue
;
4117 scan_io_t
*srch_sio
, *sio
;
4119 uint64_t start
, size
;
4121 vdev
= vdev_lookup_top(spa
, DVA_GET_VDEV(&bp
->blk_dva
[dva_i
]));
4122 ASSERT(vdev
!= NULL
);
4123 q_lock
= &vdev
->vdev_scan_io_queue_lock
;
4124 queue
= vdev
->vdev_scan_io_queue
;
4126 mutex_enter(q_lock
);
4127 if (queue
== NULL
) {
4132 srch_sio
= sio_alloc(BP_GET_NDVAS(bp
));
4133 bp2sio(bp
, srch_sio
, dva_i
);
4134 start
= SIO_GET_OFFSET(srch_sio
);
4135 size
= SIO_GET_ASIZE(srch_sio
);
4138 * We can find the zio in two states:
4139 * 1) Cold, just sitting in the queue of zio's to be issued at
4140 * some point in the future. In this case, all we do is
4141 * remove the zio from the q_sios_by_addr tree, decrement
4142 * its data volume from the containing range_seg_t and
4143 * resort the q_exts_by_size tree to reflect that the
4144 * range_seg_t has lost some of its 'fill'. We don't shorten
4145 * the range_seg_t - this is usually rare enough not to be
4146 * worth the extra hassle of trying keep track of precise
4147 * extent boundaries.
4148 * 2) Hot, where the zio is currently in-flight in
4149 * dsl_scan_issue_ios. In this case, we can't simply
4150 * reach in and stop the in-flight zio's, so we instead
4151 * block the caller. Eventually, dsl_scan_issue_ios will
4152 * be done with issuing the zio's it gathered and will
4155 sio
= avl_find(&queue
->q_sios_by_addr
, srch_sio
, &idx
);
4159 int64_t asize
= SIO_GET_ASIZE(sio
);
4162 /* Got it while it was cold in the queue */
4163 ASSERT3U(start
, ==, SIO_GET_OFFSET(sio
));
4164 ASSERT3U(size
, ==, asize
);
4165 avl_remove(&queue
->q_sios_by_addr
, sio
);
4166 queue
->q_sio_memused
-= SIO_GET_MUSED(sio
);
4168 ASSERT(range_tree_contains(queue
->q_exts_by_addr
, start
, size
));
4169 range_tree_remove_fill(queue
->q_exts_by_addr
, start
, size
);
4172 * We only update scn_bytes_pending in the cold path,
4173 * otherwise it will already have been accounted for as
4174 * part of the zio's execution.
4176 atomic_add_64(&scn
->scn_bytes_pending
, -asize
);
4178 /* count the block as though we issued it */
4179 sio2bp(sio
, &tmpbp
);
4180 count_block(scn
, dp
->dp_blkstats
, &tmpbp
);
4188 * Callback invoked when a zio_free() zio is executing. This needs to be
4189 * intercepted to prevent the zio from deallocating a particular portion
4190 * of disk space and it then getting reallocated and written to, while we
4191 * still have it queued up for processing.
4194 dsl_scan_freed(spa_t
*spa
, const blkptr_t
*bp
)
4196 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
4197 dsl_scan_t
*scn
= dp
->dp_scan
;
4199 ASSERT(!BP_IS_EMBEDDED(bp
));
4200 ASSERT(scn
!= NULL
);
4201 if (!dsl_scan_is_running(scn
))
4204 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++)
4205 dsl_scan_freed_dva(spa
, bp
, i
);
4209 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_vdev_limit
, ULONG
, ZMOD_RW
,
4210 "Max bytes in flight per leaf vdev for scrubs and resilvers");
4212 ZFS_MODULE_PARAM(zfs
, zfs_
, scrub_min_time_ms
, INT
, ZMOD_RW
,
4213 "Min millisecs to scrub per txg");
4215 ZFS_MODULE_PARAM(zfs
, zfs_
, obsolete_min_time_ms
, INT
, ZMOD_RW
,
4216 "Min millisecs to obsolete per txg");
4218 ZFS_MODULE_PARAM(zfs
, zfs_
, free_min_time_ms
, INT
, ZMOD_RW
,
4219 "Min millisecs to free per txg");
4221 ZFS_MODULE_PARAM(zfs
, zfs_
, resilver_min_time_ms
, INT
, ZMOD_RW
,
4222 "Min millisecs to resilver per txg");
4224 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_suspend_progress
, INT
, ZMOD_RW
,
4225 "Set to prevent scans from progressing");
4227 ZFS_MODULE_PARAM(zfs
, zfs_
, no_scrub_io
, INT
, ZMOD_RW
,
4228 "Set to disable scrub I/O");
4230 ZFS_MODULE_PARAM(zfs
, zfs_
, no_scrub_prefetch
, INT
, ZMOD_RW
,
4231 "Set to disable scrub prefetching");
4233 ZFS_MODULE_PARAM(zfs
, zfs_
, async_block_max_blocks
, ULONG
, ZMOD_RW
,
4234 "Max number of blocks freed in one txg");
4236 ZFS_MODULE_PARAM(zfs
, zfs_
, free_bpobj_enabled
, INT
, ZMOD_RW
,
4237 "Enable processing of the free_bpobj");
4239 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_mem_lim_fact
, INT
, ZMOD_RW
,
4240 "Fraction of RAM for scan hard limit");
4242 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_issue_strategy
, INT
, ZMOD_RW
,
4243 "IO issuing strategy during scrubbing. "
4244 "0 = default, 1 = LBA, 2 = size");
4246 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_legacy
, INT
, ZMOD_RW
,
4247 "Scrub using legacy non-sequential method");
4249 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_checkpoint_intval
, INT
, ZMOD_RW
,
4250 "Scan progress on-disk checkpointing interval");
4252 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_max_ext_gap
, ULONG
, ZMOD_RW
,
4253 "Max gap in bytes between sequential scrub / resilver I/Os");
4255 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_mem_lim_soft_fact
, INT
, ZMOD_RW
,
4256 "Fraction of hard limit used as soft limit");
4258 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_strict_mem_lim
, INT
, ZMOD_RW
,
4259 "Tunable to attempt to reduce lock contention");
4261 ZFS_MODULE_PARAM(zfs
, zfs_
, scan_fill_weight
, INT
, ZMOD_RW
,
4262 "Tunable to adjust bias towards more filled segments during scans");
4264 ZFS_MODULE_PARAM(zfs
, zfs_
, resilver_disable_defer
, INT
, ZMOD_RW
,
4265 "Process all resilvers immediately");