1 /*
2 * CDDL HEADER START
3 *
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.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or https://opensource.org/licenses/CDDL-1.0.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2021 by Delphix. All rights reserved.
24 * Copyright 2016 Gary Mills
25 * Copyright (c) 2017, 2019, Datto Inc. All rights reserved.
26 * Copyright (c) 2015, Nexenta Systems, Inc. All rights reserved.
27 * Copyright 2019 Joyent, Inc.
28 */
29
30 #include <sys/dsl_scan.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_prop.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_synctask.h>
36 #include <sys/dnode.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/dmu_objset.h>
39 #include <sys/arc.h>
40 #include <sys/arc_impl.h>
41 #include <sys/zap.h>
42 #include <sys/zio.h>
43 #include <sys/zfs_context.h>
44 #include <sys/fs/zfs.h>
45 #include <sys/zfs_znode.h>
46 #include <sys/spa_impl.h>
47 #include <sys/vdev_impl.h>
48 #include <sys/zil_impl.h>
49 #include <sys/zio_checksum.h>
50 #include <sys/brt.h>
51 #include <sys/ddt.h>
52 #include <sys/sa.h>
53 #include <sys/sa_impl.h>
54 #include <sys/zfeature.h>
55 #include <sys/abd.h>
56 #include <sys/range_tree.h>
57 #include <sys/dbuf.h>
58 #ifdef _KERNEL
59 #include <sys/zfs_vfsops.h>
60 #endif
61
62 /*
63 * Grand theory statement on scan queue sorting
64 *
65 * Scanning is implemented by recursively traversing all indirection levels
66 * in an object and reading all blocks referenced from said objects. This
67 * results in us approximately traversing the object from lowest logical
68 * offset to the highest. For best performance, we would want the logical
69 * blocks to be physically contiguous. However, this is frequently not the
70 * case with pools given the allocation patterns of copy-on-write filesystems.
71 * So instead, we put the I/Os into a reordering queue and issue them in a
72 * way that will most benefit physical disks (LBA-order).
73 *
74 * Queue management:
75 *
76 * Ideally, we would want to scan all metadata and queue up all block I/O
77 * prior to starting to issue it, because that allows us to do an optimal
78 * sorting job. This can however consume large amounts of memory. Therefore
79 * we continuously monitor the size of the queues and constrain them to 5%
80 * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this
81 * limit, we clear out a few of the largest extents at the head of the queues
82 * to make room for more scanning. Hopefully, these extents will be fairly
83 * large and contiguous, allowing us to approach sequential I/O throughput
84 * even without a fully sorted tree.
85 *
86 * Metadata scanning takes place in dsl_scan_visit(), which is called from
87 * dsl_scan_sync() every spa_sync(). If we have either fully scanned all
88 * metadata on the pool, or we need to make room in memory because our
89 * queues are too large, dsl_scan_visit() is postponed and
90 * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies
91 * that metadata scanning and queued I/O issuing are mutually exclusive. This
92 * allows us to provide maximum sequential I/O throughput for the majority of
93 * I/O's issued since sequential I/O performance is significantly negatively
94 * impacted if it is interleaved with random I/O.
95 *
96 * Implementation Notes
97 *
98 * One side effect of the queued scanning algorithm is that the scanning code
99 * needs to be notified whenever a block is freed. This is needed to allow
100 * the scanning code to remove these I/Os from the issuing queue. Additionally,
101 * we do not attempt to queue gang blocks to be issued sequentially since this
102 * is very hard to do and would have an extremely limited performance benefit.
103 * Instead, we simply issue gang I/Os as soon as we find them using the legacy
104 * algorithm.
105 *
106 * Backwards compatibility
107 *
108 * This new algorithm is backwards compatible with the legacy on-disk data
109 * structures (and therefore does not require a new feature flag).
110 * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan
111 * will stop scanning metadata (in logical order) and wait for all outstanding
112 * sorted I/O to complete. Once this is done, we write out a checkpoint
113 * bookmark, indicating that we have scanned everything logically before it.
114 * If the pool is imported on a machine without the new sorting algorithm,
115 * the scan simply resumes from the last checkpoint using the legacy algorithm.
116 */
117
118 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
119 const zbookmark_phys_t *);
120
121 static scan_cb_t dsl_scan_scrub_cb;
122
123 static int scan_ds_queue_compare(const void *a, const void *b);
124 static int scan_prefetch_queue_compare(const void *a, const void *b);
125 static void scan_ds_queue_clear(dsl_scan_t *scn);
126 static void scan_ds_prefetch_queue_clear(dsl_scan_t *scn);
127 static boolean_t scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj,
128 uint64_t *txg);
129 static void scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg);
130 static void scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj);
131 static void scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx);
132 static uint64_t dsl_scan_count_data_disks(spa_t *spa);
133 static void read_by_block_level(dsl_scan_t *scn, zbookmark_phys_t zb);
134
135 extern uint_t zfs_vdev_async_write_active_min_dirty_percent;
136 static int zfs_scan_blkstats = 0;
137
138 /*
139 * 'zpool status' uses bytes processed per pass to report throughput and
140 * estimate time remaining. We define a pass to start when the scanning
141 * phase completes for a sequential resilver. Optionally, this value
142 * may be used to reset the pass statistics every N txgs to provide an
143 * estimated completion time based on currently observed performance.
144 */
145 static uint_t zfs_scan_report_txgs = 0;
146
147 /*
148 * By default zfs will check to ensure it is not over the hard memory
149 * limit before each txg. If finer-grained control of this is needed
150 * this value can be set to 1 to enable checking before scanning each
151 * block.
152 */
153 static int zfs_scan_strict_mem_lim = B_FALSE;
154
155 /*
156 * Maximum number of parallelly executed bytes per leaf vdev. We attempt
157 * to strike a balance here between keeping the vdev queues full of I/Os
158 * at all times and not overflowing the queues to cause long latency,
159 * which would cause long txg sync times. No matter what, we will not
160 * overload the drives with I/O, since that is protected by
161 * zfs_vdev_scrub_max_active.
162 */
163 static uint64_t zfs_scan_vdev_limit = 16 << 20;
164
165 static uint_t zfs_scan_issue_strategy = 0;
166
167 /* don't queue & sort zios, go direct */
168 static int zfs_scan_legacy = B_FALSE;
169 static uint64_t zfs_scan_max_ext_gap = 2 << 20; /* in bytes */
170
171 /*
172 * fill_weight is non-tunable at runtime, so we copy it at module init from
173 * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
174 * break queue sorting.
175 */
176 static uint_t zfs_scan_fill_weight = 3;
177 static uint64_t fill_weight;
178
179 /* See dsl_scan_should_clear() for details on the memory limit tunables */
180 static const uint64_t zfs_scan_mem_lim_min = 16 << 20; /* bytes */
181 static const uint64_t zfs_scan_mem_lim_soft_max = 128 << 20; /* bytes */
182
183
184 /* fraction of physmem */
185 static uint_t zfs_scan_mem_lim_fact = 20;
186
187 /* fraction of mem lim above */
188 static uint_t zfs_scan_mem_lim_soft_fact = 20;
189
190 /* minimum milliseconds to scrub per txg */
191 static uint_t zfs_scrub_min_time_ms = 1000;
192
193 /* minimum milliseconds to obsolete per txg */
194 static uint_t zfs_obsolete_min_time_ms = 500;
195
196 /* minimum milliseconds to free per txg */
197 static uint_t zfs_free_min_time_ms = 1000;
198
199 /* minimum milliseconds to resilver per txg */
200 static uint_t zfs_resilver_min_time_ms = 3000;
201
202 static uint_t zfs_scan_checkpoint_intval = 7200; /* in seconds */
203 int zfs_scan_suspend_progress = 0; /* set to prevent scans from progressing */
204 static int zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
205 static int zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
206 static const ddt_class_t zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
207 /* max number of blocks to free in a single TXG */
208 static uint64_t zfs_async_block_max_blocks = UINT64_MAX;
209 /* max number of dedup blocks to free in a single TXG */
210 static uint64_t zfs_max_async_dedup_frees = 100000;
211
212 /* set to disable resilver deferring */
213 static int zfs_resilver_disable_defer = B_FALSE;
214
215 /*
216 * We wait a few txgs after importing a pool to begin scanning so that
217 * the import / mounting code isn't held up by scrub / resilver IO.
218 * Unfortunately, it is a bit difficult to determine exactly how long
219 * this will take since userspace will trigger fs mounts asynchronously
220 * and the kernel will create zvol minors asynchronously. As a result,
221 * the value provided here is a bit arbitrary, but represents a
222 * reasonable estimate of how many txgs it will take to finish fully
223 * importing a pool
224 */
225 #define SCAN_IMPORT_WAIT_TXGS 5
226
227 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
228 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
229 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
230
231 /*
232 * Enable/disable the processing of the free_bpobj object.
233 */
234 static int zfs_free_bpobj_enabled = 1;
235
236 /* Error blocks to be scrubbed in one txg. */
237 static uint_t zfs_scrub_error_blocks_per_txg = 1 << 12;
238
239 /* the order has to match pool_scan_type */
240 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
241 NULL,
242 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
243 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
244 };
245
246 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */
247 typedef struct {
248 uint64_t sds_dsobj;
249 uint64_t sds_txg;
250 avl_node_t sds_node;
251 } scan_ds_t;
252
253 /*
254 * This controls what conditions are placed on dsl_scan_sync_state():
255 * SYNC_OPTIONAL) write out scn_phys iff scn_queues_pending == 0
256 * SYNC_MANDATORY) write out scn_phys always. scn_queues_pending must be 0.
257 * SYNC_CACHED) if scn_queues_pending == 0, write out scn_phys. Otherwise
258 * write out the scn_phys_cached version.
259 * See dsl_scan_sync_state for details.
260 */
261 typedef enum {
262 SYNC_OPTIONAL,
263 SYNC_MANDATORY,
264 SYNC_CACHED
265 } state_sync_type_t;
266
267 /*
268 * This struct represents the minimum information needed to reconstruct a
269 * zio for sequential scanning. This is useful because many of these will
270 * accumulate in the sequential IO queues before being issued, so saving
271 * memory matters here.
272 */
273 typedef struct scan_io {
274 /* fields from blkptr_t */
275 uint64_t sio_blk_prop;
276 uint64_t sio_phys_birth;
277 uint64_t sio_birth;
278 zio_cksum_t sio_cksum;
279 uint32_t sio_nr_dvas;
280
281 /* fields from zio_t */
282 uint32_t sio_flags;
283 zbookmark_phys_t sio_zb;
284
285 /* members for queue sorting */
286 union {
287 avl_node_t sio_addr_node; /* link into issuing queue */
288 list_node_t sio_list_node; /* link for issuing to disk */
289 } sio_nodes;
290
291 /*
292 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp,
293 * depending on how many were in the original bp. Only the
294 * first DVA is really used for sorting and issuing purposes.
295 * The other DVAs (if provided) simply exist so that the zio
296 * layer can find additional copies to repair from in the
297 * event of an error. This array must go at the end of the
298 * struct to allow this for the variable number of elements.
299 */
300 dva_t sio_dva[];
301 } scan_io_t;
302
303 #define SIO_SET_OFFSET(sio, x) DVA_SET_OFFSET(&(sio)->sio_dva[0], x)
304 #define SIO_SET_ASIZE(sio, x) DVA_SET_ASIZE(&(sio)->sio_dva[0], x)
305 #define SIO_GET_OFFSET(sio) DVA_GET_OFFSET(&(sio)->sio_dva[0])
306 #define SIO_GET_ASIZE(sio) DVA_GET_ASIZE(&(sio)->sio_dva[0])
307 #define SIO_GET_END_OFFSET(sio) \
308 (SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio))
309 #define SIO_GET_MUSED(sio) \
310 (sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t)))
311
312 struct dsl_scan_io_queue {
313 dsl_scan_t *q_scn; /* associated dsl_scan_t */
314 vdev_t *q_vd; /* top-level vdev that this queue represents */
315 zio_t *q_zio; /* scn_zio_root child for waiting on IO */
316
317 /* trees used for sorting I/Os and extents of I/Os */
318 range_tree_t *q_exts_by_addr;
319 zfs_btree_t q_exts_by_size;
320 avl_tree_t q_sios_by_addr;
321 uint64_t q_sio_memused;
322 uint64_t q_last_ext_addr;
323
324 /* members for zio rate limiting */
325 uint64_t q_maxinflight_bytes;
326 uint64_t q_inflight_bytes;
327 kcondvar_t q_zio_cv; /* used under vd->vdev_scan_io_queue_lock */
328
329 /* per txg statistics */
330 uint64_t q_total_seg_size_this_txg;
331 uint64_t q_segs_this_txg;
332 uint64_t q_total_zio_size_this_txg;
333 uint64_t q_zios_this_txg;
334 };
335
336 /* private data for dsl_scan_prefetch_cb() */
337 typedef struct scan_prefetch_ctx {
338 zfs_refcount_t spc_refcnt; /* refcount for memory management */
339 dsl_scan_t *spc_scn; /* dsl_scan_t for the pool */
340 boolean_t spc_root; /* is this prefetch for an objset? */
341 uint8_t spc_indblkshift; /* dn_indblkshift of current dnode */
342 uint16_t spc_datablkszsec; /* dn_idatablkszsec of current dnode */
343 } scan_prefetch_ctx_t;
344
345 /* private data for dsl_scan_prefetch() */
346 typedef struct scan_prefetch_issue_ctx {
347 avl_node_t spic_avl_node; /* link into scn->scn_prefetch_queue */
348 scan_prefetch_ctx_t *spic_spc; /* spc for the callback */
349 blkptr_t spic_bp; /* bp to prefetch */
350 zbookmark_phys_t spic_zb; /* bookmark to prefetch */
351 } scan_prefetch_issue_ctx_t;
352
353 static void scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
354 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue);
355 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue,
356 scan_io_t *sio);
357
358 static dsl_scan_io_queue_t *scan_io_queue_create(vdev_t *vd);
359 static void scan_io_queues_destroy(dsl_scan_t *scn);
360
361 static kmem_cache_t *sio_cache[SPA_DVAS_PER_BP];
362
363 /* sio->sio_nr_dvas must be set so we know which cache to free from */
364 static void
sio_free(scan_io_t * sio)365 sio_free(scan_io_t *sio)
366 {
367 ASSERT3U(sio->sio_nr_dvas, >, 0);
368 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
369
370 kmem_cache_free(sio_cache[sio->sio_nr_dvas - 1], sio);
371 }
372
373 /* It is up to the caller to set sio->sio_nr_dvas for freeing */
374 static scan_io_t *
sio_alloc(unsigned short nr_dvas)375 sio_alloc(unsigned short nr_dvas)
376 {
377 ASSERT3U(nr_dvas, >, 0);
378 ASSERT3U(nr_dvas, <=, SPA_DVAS_PER_BP);
379
380 return (kmem_cache_alloc(sio_cache[nr_dvas - 1], KM_SLEEP));
381 }
382
383 void
scan_init(void)384 scan_init(void)
385 {
386 /*
387 * This is used in ext_size_compare() to weight segments
388 * based on how sparse they are. This cannot be changed
389 * mid-scan and the tree comparison functions don't currently
390 * have a mechanism for passing additional context to the
391 * compare functions. Thus we store this value globally and
392 * we only allow it to be set at module initialization time
393 */
394 fill_weight = zfs_scan_fill_weight;
395
396 for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
397 char name[36];
398
399 (void) snprintf(name, sizeof (name), "sio_cache_%d", i);
400 sio_cache[i] = kmem_cache_create(name,
401 (sizeof (scan_io_t) + ((i + 1) * sizeof (dva_t))),
402 0, NULL, NULL, NULL, NULL, NULL, 0);
403 }
404 }
405
406 void
scan_fini(void)407 scan_fini(void)
408 {
409 for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
410 kmem_cache_destroy(sio_cache[i]);
411 }
412 }
413
414 static inline boolean_t
dsl_scan_is_running(const dsl_scan_t * scn)415 dsl_scan_is_running(const dsl_scan_t *scn)
416 {
417 return (scn->scn_phys.scn_state == DSS_SCANNING);
418 }
419
420 boolean_t
dsl_scan_resilvering(dsl_pool_t * dp)421 dsl_scan_resilvering(dsl_pool_t *dp)
422 {
423 return (dsl_scan_is_running(dp->dp_scan) &&
424 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
425 }
426
427 static inline void
sio2bp(const scan_io_t * sio,blkptr_t * bp)428 sio2bp(const scan_io_t *sio, blkptr_t *bp)
429 {
430 memset(bp, 0, sizeof (*bp));
431 bp->blk_prop = sio->sio_blk_prop;
432 BP_SET_PHYSICAL_BIRTH(bp, sio->sio_phys_birth);
433 BP_SET_LOGICAL_BIRTH(bp, sio->sio_birth);
434 bp->blk_fill = 1; /* we always only work with data pointers */
435 bp->blk_cksum = sio->sio_cksum;
436
437 ASSERT3U(sio->sio_nr_dvas, >, 0);
438 ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
439
440 memcpy(bp->blk_dva, sio->sio_dva, sio->sio_nr_dvas * sizeof (dva_t));
441 }
442
443 static inline void
bp2sio(const blkptr_t * bp,scan_io_t * sio,int dva_i)444 bp2sio(const blkptr_t *bp, scan_io_t *sio, int dva_i)
445 {
446 sio->sio_blk_prop = bp->blk_prop;
447 sio->sio_phys_birth = BP_GET_PHYSICAL_BIRTH(bp);
448 sio->sio_birth = BP_GET_LOGICAL_BIRTH(bp);
449 sio->sio_cksum = bp->blk_cksum;
450 sio->sio_nr_dvas = BP_GET_NDVAS(bp);
451
452 /*
453 * Copy the DVAs to the sio. We need all copies of the block so
454 * that the self healing code can use the alternate copies if the
455 * first is corrupted. We want the DVA at index dva_i to be first
456 * in the sio since this is the primary one that we want to issue.
457 */
458 for (int i = 0, j = dva_i; i < sio->sio_nr_dvas; i++, j++) {
459 sio->sio_dva[i] = bp->blk_dva[j % sio->sio_nr_dvas];
460 }
461 }
462
463 int
dsl_scan_init(dsl_pool_t * dp,uint64_t txg)464 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
465 {
466 int err;
467 dsl_scan_t *scn;
468 spa_t *spa = dp->dp_spa;
469 uint64_t f;
470
471 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
472 scn->scn_dp = dp;
473
474 /*
475 * It's possible that we're resuming a scan after a reboot so
476 * make sure that the scan_async_destroying flag is initialized
477 * appropriately.
478 */
479 ASSERT(!scn->scn_async_destroying);
480 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
481 SPA_FEATURE_ASYNC_DESTROY);
482
483 /*
484 * Calculate the max number of in-flight bytes for pool-wide
485 * scanning operations (minimum 1MB, maximum 1/4 of arc_c_max).
486 * Limits for the issuing phase are done per top-level vdev and
487 * are handled separately.
488 */
489 scn->scn_maxinflight_bytes = MIN(arc_c_max / 4, MAX(1ULL << 20,
490 zfs_scan_vdev_limit * dsl_scan_count_data_disks(spa)));
491
492 avl_create(&scn->scn_queue, scan_ds_queue_compare, sizeof (scan_ds_t),
493 offsetof(scan_ds_t, sds_node));
494 mutex_init(&scn->scn_queue_lock, NULL, MUTEX_DEFAULT, NULL);
495 avl_create(&scn->scn_prefetch_queue, scan_prefetch_queue_compare,
496 sizeof (scan_prefetch_issue_ctx_t),
497 offsetof(scan_prefetch_issue_ctx_t, spic_avl_node));
498
499 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
500 "scrub_func", sizeof (uint64_t), 1, &f);
501 if (err == 0) {
502 /*
503 * There was an old-style scrub in progress. Restart a
504 * new-style scrub from the beginning.
505 */
506 scn->scn_restart_txg = txg;
507 zfs_dbgmsg("old-style scrub was in progress for %s; "
508 "restarting new-style scrub in txg %llu",
509 spa->spa_name,
510 (longlong_t)scn->scn_restart_txg);
511
512 /*
513 * Load the queue obj from the old location so that it
514 * can be freed by dsl_scan_done().
515 */
516 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
517 "scrub_queue", sizeof (uint64_t), 1,
518 &scn->scn_phys.scn_queue_obj);
519 } else {
520 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
521 DMU_POOL_ERRORSCRUB, sizeof (uint64_t),
522 ERRORSCRUB_PHYS_NUMINTS, &scn->errorscrub_phys);
523
524 if (err != 0 && err != ENOENT)
525 return (err);
526
527 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
528 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
529 &scn->scn_phys);
530
531 /*
532 * Detect if the pool contains the signature of #2094. If it
533 * does properly update the scn->scn_phys structure and notify
534 * the administrator by setting an errata for the pool.
535 */
536 if (err == EOVERFLOW) {
537 uint64_t zaptmp[SCAN_PHYS_NUMINTS + 1];
538 VERIFY3S(SCAN_PHYS_NUMINTS, ==, 24);
539 VERIFY3S(offsetof(dsl_scan_phys_t, scn_flags), ==,
540 (23 * sizeof (uint64_t)));
541
542 err = zap_lookup(dp->dp_meta_objset,
543 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SCAN,
544 sizeof (uint64_t), SCAN_PHYS_NUMINTS + 1, &zaptmp);
545 if (err == 0) {
546 uint64_t overflow = zaptmp[SCAN_PHYS_NUMINTS];
547
548 if (overflow & ~DSL_SCAN_FLAGS_MASK ||
549 scn->scn_async_destroying) {
550 spa->spa_errata =
551 ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY;
552 return (EOVERFLOW);
553 }
554
555 memcpy(&scn->scn_phys, zaptmp,
556 SCAN_PHYS_NUMINTS * sizeof (uint64_t));
557 scn->scn_phys.scn_flags = overflow;
558
559 /* Required scrub already in progress. */
560 if (scn->scn_phys.scn_state == DSS_FINISHED ||
561 scn->scn_phys.scn_state == DSS_CANCELED)
562 spa->spa_errata =
563 ZPOOL_ERRATA_ZOL_2094_SCRUB;
564 }
565 }
566
567 if (err == ENOENT)
568 return (0);
569 else if (err)
570 return (err);
571
572 /*
573 * We might be restarting after a reboot, so jump the issued
574 * counter to how far we've scanned. We know we're consistent
575 * up to here.
576 */
577 scn->scn_issued_before_pass = scn->scn_phys.scn_examined -
578 scn->scn_phys.scn_skipped;
579
580 if (dsl_scan_is_running(scn) &&
581 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
582 /*
583 * A new-type scrub was in progress on an old
584 * pool, and the pool was accessed by old
585 * software. Restart from the beginning, since
586 * the old software may have changed the pool in
587 * the meantime.
588 */
589 scn->scn_restart_txg = txg;
590 zfs_dbgmsg("new-style scrub for %s was modified "
591 "by old software; restarting in txg %llu",
592 spa->spa_name,
593 (longlong_t)scn->scn_restart_txg);
594 } else if (dsl_scan_resilvering(dp)) {
595 /*
596 * If a resilver is in progress and there are already
597 * errors, restart it instead of finishing this scan and
598 * then restarting it. If there haven't been any errors
599 * then remember that the incore DTL is valid.
600 */
601 if (scn->scn_phys.scn_errors > 0) {
602 scn->scn_restart_txg = txg;
603 zfs_dbgmsg("resilver can't excise DTL_MISSING "
604 "when finished; restarting on %s in txg "
605 "%llu",
606 spa->spa_name,
607 (u_longlong_t)scn->scn_restart_txg);
608 } else {
609 /* it's safe to excise DTL when finished */
610 spa->spa_scrub_started = B_TRUE;
611 }
612 }
613 }
614
615 memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys));
616
617 /* reload the queue into the in-core state */
618 if (scn->scn_phys.scn_queue_obj != 0) {
619 zap_cursor_t zc;
620 zap_attribute_t za;
621
622 for (zap_cursor_init(&zc, dp->dp_meta_objset,
623 scn->scn_phys.scn_queue_obj);
624 zap_cursor_retrieve(&zc, &za) == 0;
625 (void) zap_cursor_advance(&zc)) {
626 scan_ds_queue_insert(scn,
627 zfs_strtonum(za.za_name, NULL),
628 za.za_first_integer);
629 }
630 zap_cursor_fini(&zc);
631 }
632
633 spa_scan_stat_init(spa);
634 vdev_scan_stat_init(spa->spa_root_vdev);
635
636 return (0);
637 }
638
639 void
dsl_scan_fini(dsl_pool_t * dp)640 dsl_scan_fini(dsl_pool_t *dp)
641 {
642 if (dp->dp_scan != NULL) {
643 dsl_scan_t *scn = dp->dp_scan;
644
645 if (scn->scn_taskq != NULL)
646 taskq_destroy(scn->scn_taskq);
647
648 scan_ds_queue_clear(scn);
649 avl_destroy(&scn->scn_queue);
650 mutex_destroy(&scn->scn_queue_lock);
651 scan_ds_prefetch_queue_clear(scn);
652 avl_destroy(&scn->scn_prefetch_queue);
653
654 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
655 dp->dp_scan = NULL;
656 }
657 }
658
659 static boolean_t
dsl_scan_restarting(dsl_scan_t * scn,dmu_tx_t * tx)660 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
661 {
662 return (scn->scn_restart_txg != 0 &&
663 scn->scn_restart_txg <= tx->tx_txg);
664 }
665
666 boolean_t
dsl_scan_resilver_scheduled(dsl_pool_t * dp)667 dsl_scan_resilver_scheduled(dsl_pool_t *dp)
668 {
669 return ((dp->dp_scan && dp->dp_scan->scn_restart_txg != 0) ||
670 (spa_async_tasks(dp->dp_spa) & SPA_ASYNC_RESILVER));
671 }
672
673 boolean_t
dsl_scan_scrubbing(const dsl_pool_t * dp)674 dsl_scan_scrubbing(const dsl_pool_t *dp)
675 {
676 dsl_scan_phys_t *scn_phys = &dp->dp_scan->scn_phys;
677
678 return (scn_phys->scn_state == DSS_SCANNING &&
679 scn_phys->scn_func == POOL_SCAN_SCRUB);
680 }
681
682 boolean_t
dsl_errorscrubbing(const dsl_pool_t * dp)683 dsl_errorscrubbing(const dsl_pool_t *dp)
684 {
685 dsl_errorscrub_phys_t *errorscrub_phys = &dp->dp_scan->errorscrub_phys;
686
687 return (errorscrub_phys->dep_state == DSS_ERRORSCRUBBING &&
688 errorscrub_phys->dep_func == POOL_SCAN_ERRORSCRUB);
689 }
690
691 boolean_t
dsl_errorscrub_is_paused(const dsl_scan_t * scn)692 dsl_errorscrub_is_paused(const dsl_scan_t *scn)
693 {
694 return (dsl_errorscrubbing(scn->scn_dp) &&
695 scn->errorscrub_phys.dep_paused_flags);
696 }
697
698 boolean_t
dsl_scan_is_paused_scrub(const dsl_scan_t * scn)699 dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
700 {
701 return (dsl_scan_scrubbing(scn->scn_dp) &&
702 scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED);
703 }
704
705 static void
dsl_errorscrub_sync_state(dsl_scan_t * scn,dmu_tx_t * tx)706 dsl_errorscrub_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
707 {
708 scn->errorscrub_phys.dep_cursor =
709 zap_cursor_serialize(&scn->errorscrub_cursor);
710
711 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
712 DMU_POOL_DIRECTORY_OBJECT,
713 DMU_POOL_ERRORSCRUB, sizeof (uint64_t), ERRORSCRUB_PHYS_NUMINTS,
714 &scn->errorscrub_phys, tx));
715 }
716
717 static void
dsl_errorscrub_setup_sync(void * arg,dmu_tx_t * tx)718 dsl_errorscrub_setup_sync(void *arg, dmu_tx_t *tx)
719 {
720 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
721 pool_scan_func_t *funcp = arg;
722 dsl_pool_t *dp = scn->scn_dp;
723 spa_t *spa = dp->dp_spa;
724
725 ASSERT(!dsl_scan_is_running(scn));
726 ASSERT(!dsl_errorscrubbing(scn->scn_dp));
727 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
728
729 memset(&scn->errorscrub_phys, 0, sizeof (scn->errorscrub_phys));
730 scn->errorscrub_phys.dep_func = *funcp;
731 scn->errorscrub_phys.dep_state = DSS_ERRORSCRUBBING;
732 scn->errorscrub_phys.dep_start_time = gethrestime_sec();
733 scn->errorscrub_phys.dep_to_examine = spa_get_last_errlog_size(spa);
734 scn->errorscrub_phys.dep_examined = 0;
735 scn->errorscrub_phys.dep_errors = 0;
736 scn->errorscrub_phys.dep_cursor = 0;
737 zap_cursor_init_serialized(&scn->errorscrub_cursor,
738 spa->spa_meta_objset, spa->spa_errlog_last,
739 scn->errorscrub_phys.dep_cursor);
740
741 vdev_config_dirty(spa->spa_root_vdev);
742 spa_event_notify(spa, NULL, NULL, ESC_ZFS_ERRORSCRUB_START);
743
744 dsl_errorscrub_sync_state(scn, tx);
745
746 spa_history_log_internal(spa, "error scrub setup", tx,
747 "func=%u mintxg=%u maxtxg=%llu",
748 *funcp, 0, (u_longlong_t)tx->tx_txg);
749 }
750
751 static int
dsl_errorscrub_setup_check(void * arg,dmu_tx_t * tx)752 dsl_errorscrub_setup_check(void *arg, dmu_tx_t *tx)
753 {
754 (void) arg;
755 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
756
757 if (dsl_scan_is_running(scn) || (dsl_errorscrubbing(scn->scn_dp))) {
758 return (SET_ERROR(EBUSY));
759 }
760
761 if (spa_get_last_errlog_size(scn->scn_dp->dp_spa) == 0) {
762 return (ECANCELED);
763 }
764 return (0);
765 }
766
767 /*
768 * Writes out a persistent dsl_scan_phys_t record to the pool directory.
769 * Because we can be running in the block sorting algorithm, we do not always
770 * want to write out the record, only when it is "safe" to do so. This safety
771 * condition is achieved by making sure that the sorting queues are empty
772 * (scn_queues_pending == 0). When this condition is not true, the sync'd state
773 * is inconsistent with how much actual scanning progress has been made. The
774 * kind of sync to be performed is specified by the sync_type argument. If the
775 * sync is optional, we only sync if the queues are empty. If the sync is
776 * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
777 * third possible state is a "cached" sync. This is done in response to:
778 * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
779 * destroyed, so we wouldn't be able to restart scanning from it.
780 * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
781 * superseded by a newer snapshot.
782 * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
783 * swapped with its clone.
784 * In all cases, a cached sync simply rewrites the last record we've written,
785 * just slightly modified. For the modifications that are performed to the
786 * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
787 * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
788 */
789 static void
dsl_scan_sync_state(dsl_scan_t * scn,dmu_tx_t * tx,state_sync_type_t sync_type)790 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx, state_sync_type_t sync_type)
791 {
792 int i;
793 spa_t *spa = scn->scn_dp->dp_spa;
794
795 ASSERT(sync_type != SYNC_MANDATORY || scn->scn_queues_pending == 0);
796 if (scn->scn_queues_pending == 0) {
797 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
798 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
799 dsl_scan_io_queue_t *q = vd->vdev_scan_io_queue;
800
801 if (q == NULL)
802 continue;
803
804 mutex_enter(&vd->vdev_scan_io_queue_lock);
805 ASSERT3P(avl_first(&q->q_sios_by_addr), ==, NULL);
806 ASSERT3P(zfs_btree_first(&q->q_exts_by_size, NULL), ==,
807 NULL);
808 ASSERT3P(range_tree_first(q->q_exts_by_addr), ==, NULL);
809 mutex_exit(&vd->vdev_scan_io_queue_lock);
810 }
811
812 if (scn->scn_phys.scn_queue_obj != 0)
813 scan_ds_queue_sync(scn, tx);
814 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
815 DMU_POOL_DIRECTORY_OBJECT,
816 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
817 &scn->scn_phys, tx));
818 memcpy(&scn->scn_phys_cached, &scn->scn_phys,
819 sizeof (scn->scn_phys));
820
821 if (scn->scn_checkpointing)
822 zfs_dbgmsg("finish scan checkpoint for %s",
823 spa->spa_name);
824
825 scn->scn_checkpointing = B_FALSE;
826 scn->scn_last_checkpoint = ddi_get_lbolt();
827 } else if (sync_type == SYNC_CACHED) {
828 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
829 DMU_POOL_DIRECTORY_OBJECT,
830 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
831 &scn->scn_phys_cached, tx));
832 }
833 }
834
835 int
dsl_scan_setup_check(void * arg,dmu_tx_t * tx)836 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
837 {
838 (void) arg;
839 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
840 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
841
842 if (dsl_scan_is_running(scn) || vdev_rebuild_active(rvd) ||
843 dsl_errorscrubbing(scn->scn_dp))
844 return (SET_ERROR(EBUSY));
845
846 return (0);
847 }
848
849 void
dsl_scan_setup_sync(void * arg,dmu_tx_t * tx)850 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
851 {
852 (void) arg;
853 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
854 pool_scan_func_t *funcp = arg;
855 dmu_object_type_t ot = 0;
856 dsl_pool_t *dp = scn->scn_dp;
857 spa_t *spa = dp->dp_spa;
858
859 ASSERT(!dsl_scan_is_running(scn));
860 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
861 memset(&scn->scn_phys, 0, sizeof (scn->scn_phys));
862
863 /*
864 * If we are starting a fresh scrub, we erase the error scrub
865 * information from disk.
866 */
867 memset(&scn->errorscrub_phys, 0, sizeof (scn->errorscrub_phys));
868 dsl_errorscrub_sync_state(scn, tx);
869
870 scn->scn_phys.scn_func = *funcp;
871 scn->scn_phys.scn_state = DSS_SCANNING;
872 scn->scn_phys.scn_min_txg = 0;
873 scn->scn_phys.scn_max_txg = tx->tx_txg;
874 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
875 scn->scn_phys.scn_start_time = gethrestime_sec();
876 scn->scn_phys.scn_errors = 0;
877 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
878 scn->scn_issued_before_pass = 0;
879 scn->scn_restart_txg = 0;
880 scn->scn_done_txg = 0;
881 scn->scn_last_checkpoint = 0;
882 scn->scn_checkpointing = B_FALSE;
883 spa_scan_stat_init(spa);
884 vdev_scan_stat_init(spa->spa_root_vdev);
885
886 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
887 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
888
889 /* rewrite all disk labels */
890 vdev_config_dirty(spa->spa_root_vdev);
891
892 if (vdev_resilver_needed(spa->spa_root_vdev,
893 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
894 nvlist_t *aux = fnvlist_alloc();
895 fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE,
896 "healing");
897 spa_event_notify(spa, NULL, aux,
898 ESC_ZFS_RESILVER_START);
899 nvlist_free(aux);
900 } else {
901 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
902 }
903
904 spa->spa_scrub_started = B_TRUE;
905 /*
906 * If this is an incremental scrub, limit the DDT scrub phase
907 * to just the auto-ditto class (for correctness); the rest
908 * of the scrub should go faster using top-down pruning.
909 */
910 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
911 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
912
913 /*
914 * When starting a resilver clear any existing rebuild state.
915 * This is required to prevent stale rebuild status from
916 * being reported when a rebuild is run, then a resilver and
917 * finally a scrub. In which case only the scrub status
918 * should be reported by 'zpool status'.
919 */
920 if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
921 vdev_t *rvd = spa->spa_root_vdev;
922 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
923 vdev_t *vd = rvd->vdev_child[i];
924 vdev_rebuild_clear_sync(
925 (void *)(uintptr_t)vd->vdev_id, tx);
926 }
927 }
928 }
929
930 /* back to the generic stuff */
931
932 if (zfs_scan_blkstats) {
933 if (dp->dp_blkstats == NULL) {
934 dp->dp_blkstats =
935 vmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
936 }
937 memset(&dp->dp_blkstats->zab_type, 0,
938 sizeof (dp->dp_blkstats->zab_type));
939 } else {
940 if (dp->dp_blkstats) {
941 vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
942 dp->dp_blkstats = NULL;
943 }
944 }
945
946 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
947 ot = DMU_OT_ZAP_OTHER;
948
949 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
950 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
951
952 memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys));
953
954 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
955
956 spa_history_log_internal(spa, "scan setup", tx,
957 "func=%u mintxg=%llu maxtxg=%llu",
958 *funcp, (u_longlong_t)scn->scn_phys.scn_min_txg,
959 (u_longlong_t)scn->scn_phys.scn_max_txg);
960 }
961
962 /*
963 * Called by ZFS_IOC_POOL_SCRUB and ZFS_IOC_POOL_SCAN ioctl to start a scrub,
964 * error scrub or resilver. Can also be called to resume a paused scrub or
965 * error scrub.
966 */
967 int
dsl_scan(dsl_pool_t * dp,pool_scan_func_t func)968 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
969 {
970 spa_t *spa = dp->dp_spa;
971 dsl_scan_t *scn = dp->dp_scan;
972
973 /*
974 * Purge all vdev caches and probe all devices. We do this here
975 * rather than in sync context because this requires a writer lock
976 * on the spa_config lock, which we can't do from sync context. The
977 * spa_scrub_reopen flag indicates that vdev_open() should not
978 * attempt to start another scrub.
979 */
980 spa_vdev_state_enter(spa, SCL_NONE);
981 spa->spa_scrub_reopen = B_TRUE;
982 vdev_reopen(spa->spa_root_vdev);
983 spa->spa_scrub_reopen = B_FALSE;
984 (void) spa_vdev_state_exit(spa, NULL, 0);
985
986 if (func == POOL_SCAN_RESILVER) {
987 dsl_scan_restart_resilver(spa->spa_dsl_pool, 0);
988 return (0);
989 }
990
991 if (func == POOL_SCAN_ERRORSCRUB) {
992 if (dsl_errorscrub_is_paused(dp->dp_scan)) {
993 /*
994 * got error scrub start cmd, resume paused error scrub.
995 */
996 int err = dsl_scrub_set_pause_resume(scn->scn_dp,
997 POOL_SCRUB_NORMAL);
998 if (err == 0) {
999 spa_event_notify(spa, NULL, NULL,
1000 ESC_ZFS_ERRORSCRUB_RESUME);
1001 return (ECANCELED);
1002 }
1003 return (SET_ERROR(err));
1004 }
1005
1006 return (dsl_sync_task(spa_name(dp->dp_spa),
1007 dsl_errorscrub_setup_check, dsl_errorscrub_setup_sync,
1008 &func, 0, ZFS_SPACE_CHECK_RESERVED));
1009 }
1010
1011 if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
1012 /* got scrub start cmd, resume paused scrub */
1013 int err = dsl_scrub_set_pause_resume(scn->scn_dp,
1014 POOL_SCRUB_NORMAL);
1015 if (err == 0) {
1016 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
1017 return (SET_ERROR(ECANCELED));
1018 }
1019 return (SET_ERROR(err));
1020 }
1021
1022 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1023 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED));
1024 }
1025
1026 static void
dsl_errorscrub_done(dsl_scan_t * scn,boolean_t complete,dmu_tx_t * tx)1027 dsl_errorscrub_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
1028 {
1029 dsl_pool_t *dp = scn->scn_dp;
1030 spa_t *spa = dp->dp_spa;
1031
1032 if (complete) {
1033 spa_event_notify(spa, NULL, NULL, ESC_ZFS_ERRORSCRUB_FINISH);
1034 spa_history_log_internal(spa, "error scrub done", tx,
1035 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1036 } else {
1037 spa_history_log_internal(spa, "error scrub canceled", tx,
1038 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1039 }
1040
1041 scn->errorscrub_phys.dep_state = complete ? DSS_FINISHED : DSS_CANCELED;
1042 spa->spa_scrub_active = B_FALSE;
1043 spa_errlog_rotate(spa);
1044 scn->errorscrub_phys.dep_end_time = gethrestime_sec();
1045 zap_cursor_fini(&scn->errorscrub_cursor);
1046
1047 if (spa->spa_errata == ZPOOL_ERRATA_ZOL_2094_SCRUB)
1048 spa->spa_errata = 0;
1049
1050 ASSERT(!dsl_errorscrubbing(scn->scn_dp));
1051 }
1052
1053 static void
dsl_scan_done(dsl_scan_t * scn,boolean_t complete,dmu_tx_t * tx)1054 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
1055 {
1056 static const char *old_names[] = {
1057 "scrub_bookmark",
1058 "scrub_ddt_bookmark",
1059 "scrub_ddt_class_max",
1060 "scrub_queue",
1061 "scrub_min_txg",
1062 "scrub_max_txg",
1063 "scrub_func",
1064 "scrub_errors",
1065 NULL
1066 };
1067
1068 dsl_pool_t *dp = scn->scn_dp;
1069 spa_t *spa = dp->dp_spa;
1070 int i;
1071
1072 /* Remove any remnants of an old-style scrub. */
1073 for (i = 0; old_names[i]; i++) {
1074 (void) zap_remove(dp->dp_meta_objset,
1075 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
1076 }
1077
1078 if (scn->scn_phys.scn_queue_obj != 0) {
1079 VERIFY0(dmu_object_free(dp->dp_meta_objset,
1080 scn->scn_phys.scn_queue_obj, tx));
1081 scn->scn_phys.scn_queue_obj = 0;
1082 }
1083 scan_ds_queue_clear(scn);
1084 scan_ds_prefetch_queue_clear(scn);
1085
1086 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
1087
1088 /*
1089 * If we were "restarted" from a stopped state, don't bother
1090 * with anything else.
1091 */
1092 if (!dsl_scan_is_running(scn)) {
1093 ASSERT(!scn->scn_is_sorted);
1094 return;
1095 }
1096
1097 if (scn->scn_is_sorted) {
1098 scan_io_queues_destroy(scn);
1099 scn->scn_is_sorted = B_FALSE;
1100
1101 if (scn->scn_taskq != NULL) {
1102 taskq_destroy(scn->scn_taskq);
1103 scn->scn_taskq = NULL;
1104 }
1105 }
1106
1107 scn->scn_phys.scn_state = complete ? DSS_FINISHED : DSS_CANCELED;
1108
1109 spa_notify_waiters(spa);
1110
1111 if (dsl_scan_restarting(scn, tx))
1112 spa_history_log_internal(spa, "scan aborted, restarting", tx,
1113 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1114 else if (!complete)
1115 spa_history_log_internal(spa, "scan cancelled", tx,
1116 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1117 else
1118 spa_history_log_internal(spa, "scan done", tx,
1119 "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1120
1121 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1122 spa->spa_scrub_active = B_FALSE;
1123
1124 /*
1125 * If the scrub/resilver completed, update all DTLs to
1126 * reflect this. Whether it succeeded or not, vacate
1127 * all temporary scrub DTLs.
1128 *
1129 * As the scrub does not currently support traversing
1130 * data that have been freed but are part of a checkpoint,
1131 * we don't mark the scrub as done in the DTLs as faults
1132 * may still exist in those vdevs.
1133 */
1134 if (complete &&
1135 !spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
1136 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
1137 scn->scn_phys.scn_max_txg, B_TRUE, B_FALSE);
1138
1139 if (scn->scn_phys.scn_min_txg) {
1140 nvlist_t *aux = fnvlist_alloc();
1141 fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE,
1142 "healing");
1143 spa_event_notify(spa, NULL, aux,
1144 ESC_ZFS_RESILVER_FINISH);
1145 nvlist_free(aux);
1146 } else {
1147 spa_event_notify(spa, NULL, NULL,
1148 ESC_ZFS_SCRUB_FINISH);
1149 }
1150 } else {
1151 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
1152 0, B_TRUE, B_FALSE);
1153 }
1154 spa_errlog_rotate(spa);
1155
1156 /*
1157 * Don't clear flag until after vdev_dtl_reassess to ensure that
1158 * DTL_MISSING will get updated when possible.
1159 */
1160 spa->spa_scrub_started = B_FALSE;
1161
1162 /*
1163 * We may have finished replacing a device.
1164 * Let the async thread assess this and handle the detach.
1165 */
1166 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
1167
1168 /*
1169 * Clear any resilver_deferred flags in the config.
1170 * If there are drives that need resilvering, kick
1171 * off an asynchronous request to start resilver.
1172 * vdev_clear_resilver_deferred() may update the config
1173 * before the resilver can restart. In the event of
1174 * a crash during this period, the spa loading code
1175 * will find the drives that need to be resilvered
1176 * and start the resilver then.
1177 */
1178 if (spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER) &&
1179 vdev_clear_resilver_deferred(spa->spa_root_vdev, tx)) {
1180 spa_history_log_internal(spa,
1181 "starting deferred resilver", tx, "errors=%llu",
1182 (u_longlong_t)spa_approx_errlog_size(spa));
1183 spa_async_request(spa, SPA_ASYNC_RESILVER);
1184 }
1185
1186 /* Clear recent error events (i.e. duplicate events tracking) */
1187 if (complete)
1188 zfs_ereport_clear(spa, NULL);
1189 }
1190
1191 scn->scn_phys.scn_end_time = gethrestime_sec();
1192
1193 if (spa->spa_errata == ZPOOL_ERRATA_ZOL_2094_SCRUB)
1194 spa->spa_errata = 0;
1195
1196 ASSERT(!dsl_scan_is_running(scn));
1197 }
1198
1199 static int
dsl_errorscrub_pause_resume_check(void * arg,dmu_tx_t * tx)1200 dsl_errorscrub_pause_resume_check(void *arg, dmu_tx_t *tx)
1201 {
1202 pool_scrub_cmd_t *cmd = arg;
1203 dsl_pool_t *dp = dmu_tx_pool(tx);
1204 dsl_scan_t *scn = dp->dp_scan;
1205
1206 if (*cmd == POOL_SCRUB_PAUSE) {
1207 /*
1208 * can't pause a error scrub when there is no in-progress
1209 * error scrub.
1210 */
1211 if (!dsl_errorscrubbing(dp))
1212 return (SET_ERROR(ENOENT));
1213
1214 /* can't pause a paused error scrub */
1215 if (dsl_errorscrub_is_paused(scn))
1216 return (SET_ERROR(EBUSY));
1217 } else if (*cmd != POOL_SCRUB_NORMAL) {
1218 return (SET_ERROR(ENOTSUP));
1219 }
1220
1221 return (0);
1222 }
1223
1224 static void
dsl_errorscrub_pause_resume_sync(void * arg,dmu_tx_t * tx)1225 dsl_errorscrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
1226 {
1227 pool_scrub_cmd_t *cmd = arg;
1228 dsl_pool_t *dp = dmu_tx_pool(tx);
1229 spa_t *spa = dp->dp_spa;
1230 dsl_scan_t *scn = dp->dp_scan;
1231
1232 if (*cmd == POOL_SCRUB_PAUSE) {
1233 spa->spa_scan_pass_errorscrub_pause = gethrestime_sec();
1234 scn->errorscrub_phys.dep_paused_flags = B_TRUE;
1235 dsl_errorscrub_sync_state(scn, tx);
1236 spa_event_notify(spa, NULL, NULL, ESC_ZFS_ERRORSCRUB_PAUSED);
1237 } else {
1238 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
1239 if (dsl_errorscrub_is_paused(scn)) {
1240 /*
1241 * We need to keep track of how much time we spend
1242 * paused per pass so that we can adjust the error scrub
1243 * rate shown in the output of 'zpool status'.
1244 */
1245 spa->spa_scan_pass_errorscrub_spent_paused +=
1246 gethrestime_sec() -
1247 spa->spa_scan_pass_errorscrub_pause;
1248
1249 spa->spa_scan_pass_errorscrub_pause = 0;
1250 scn->errorscrub_phys.dep_paused_flags = B_FALSE;
1251
1252 zap_cursor_init_serialized(
1253 &scn->errorscrub_cursor,
1254 spa->spa_meta_objset, spa->spa_errlog_last,
1255 scn->errorscrub_phys.dep_cursor);
1256
1257 dsl_errorscrub_sync_state(scn, tx);
1258 }
1259 }
1260 }
1261
1262 static int
dsl_errorscrub_cancel_check(void * arg,dmu_tx_t * tx)1263 dsl_errorscrub_cancel_check(void *arg, dmu_tx_t *tx)
1264 {
1265 (void) arg;
1266 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1267 /* can't cancel a error scrub when there is no one in-progress */
1268 if (!dsl_errorscrubbing(scn->scn_dp))
1269 return (SET_ERROR(ENOENT));
1270 return (0);
1271 }
1272
1273 static void
dsl_errorscrub_cancel_sync(void * arg,dmu_tx_t * tx)1274 dsl_errorscrub_cancel_sync(void *arg, dmu_tx_t *tx)
1275 {
1276 (void) arg;
1277 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1278
1279 dsl_errorscrub_done(scn, B_FALSE, tx);
1280 dsl_errorscrub_sync_state(scn, tx);
1281 spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL,
1282 ESC_ZFS_ERRORSCRUB_ABORT);
1283 }
1284
1285 static int
dsl_scan_cancel_check(void * arg,dmu_tx_t * tx)1286 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
1287 {
1288 (void) arg;
1289 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1290
1291 if (!dsl_scan_is_running(scn))
1292 return (SET_ERROR(ENOENT));
1293 return (0);
1294 }
1295
1296 static void
dsl_scan_cancel_sync(void * arg,dmu_tx_t * tx)1297 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
1298 {
1299 (void) arg;
1300 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1301
1302 dsl_scan_done(scn, B_FALSE, tx);
1303 dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
1304 spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
1305 }
1306
1307 int
dsl_scan_cancel(dsl_pool_t * dp)1308 dsl_scan_cancel(dsl_pool_t *dp)
1309 {
1310 if (dsl_errorscrubbing(dp)) {
1311 return (dsl_sync_task(spa_name(dp->dp_spa),
1312 dsl_errorscrub_cancel_check, dsl_errorscrub_cancel_sync,
1313 NULL, 3, ZFS_SPACE_CHECK_RESERVED));
1314 }
1315 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
1316 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
1317 }
1318
1319 static int
dsl_scrub_pause_resume_check(void * arg,dmu_tx_t * tx)1320 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
1321 {
1322 pool_scrub_cmd_t *cmd = arg;
1323 dsl_pool_t *dp = dmu_tx_pool(tx);
1324 dsl_scan_t *scn = dp->dp_scan;
1325
1326 if (*cmd == POOL_SCRUB_PAUSE) {
1327 /* can't pause a scrub when there is no in-progress scrub */
1328 if (!dsl_scan_scrubbing(dp))
1329 return (SET_ERROR(ENOENT));
1330
1331 /* can't pause a paused scrub */
1332 if (dsl_scan_is_paused_scrub(scn))
1333 return (SET_ERROR(EBUSY));
1334 } else if (*cmd != POOL_SCRUB_NORMAL) {
1335 return (SET_ERROR(ENOTSUP));
1336 }
1337
1338 return (0);
1339 }
1340
1341 static void
dsl_scrub_pause_resume_sync(void * arg,dmu_tx_t * tx)1342 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
1343 {
1344 pool_scrub_cmd_t *cmd = arg;
1345 dsl_pool_t *dp = dmu_tx_pool(tx);
1346 spa_t *spa = dp->dp_spa;
1347 dsl_scan_t *scn = dp->dp_scan;
1348
1349 if (*cmd == POOL_SCRUB_PAUSE) {
1350 /* can't pause a scrub when there is no in-progress scrub */
1351 spa->spa_scan_pass_scrub_pause = gethrestime_sec();
1352 scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
1353 scn->scn_phys_cached.scn_flags |= DSF_SCRUB_PAUSED;
1354 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1355 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
1356 spa_notify_waiters(spa);
1357 } else {
1358 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
1359 if (dsl_scan_is_paused_scrub(scn)) {
1360 /*
1361 * We need to keep track of how much time we spend
1362 * paused per pass so that we can adjust the scrub rate
1363 * shown in the output of 'zpool status'
1364 */
1365 spa->spa_scan_pass_scrub_spent_paused +=
1366 gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
1367 spa->spa_scan_pass_scrub_pause = 0;
1368 scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
1369 scn->scn_phys_cached.scn_flags &= ~DSF_SCRUB_PAUSED;
1370 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1371 }
1372 }
1373 }
1374
1375 /*
1376 * Set scrub pause/resume state if it makes sense to do so
1377 */
1378 int
dsl_scrub_set_pause_resume(const dsl_pool_t * dp,pool_scrub_cmd_t cmd)1379 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
1380 {
1381 if (dsl_errorscrubbing(dp)) {
1382 return (dsl_sync_task(spa_name(dp->dp_spa),
1383 dsl_errorscrub_pause_resume_check,
1384 dsl_errorscrub_pause_resume_sync, &cmd, 3,
1385 ZFS_SPACE_CHECK_RESERVED));
1386 }
1387 return (dsl_sync_task(spa_name(dp->dp_spa),
1388 dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
1389 ZFS_SPACE_CHECK_RESERVED));
1390 }
1391
1392
1393 /* start a new scan, or restart an existing one. */
1394 void
dsl_scan_restart_resilver(dsl_pool_t * dp,uint64_t txg)1395 dsl_scan_restart_resilver(dsl_pool_t *dp, uint64_t txg)
1396 {
1397 if (txg == 0) {
1398 dmu_tx_t *tx;
1399 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1400 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1401
1402 txg = dmu_tx_get_txg(tx);
1403 dp->dp_scan->scn_restart_txg = txg;
1404 dmu_tx_commit(tx);
1405 } else {
1406 dp->dp_scan->scn_restart_txg = txg;
1407 }
1408 zfs_dbgmsg("restarting resilver for %s at txg=%llu",
1409 dp->dp_spa->spa_name, (longlong_t)txg);
1410 }
1411
1412 void
dsl_free(dsl_pool_t * dp,uint64_t txg,const blkptr_t * bp)1413 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
1414 {
1415 zio_free(dp->dp_spa, txg, bp);
1416 }
1417
1418 void
dsl_free_sync(zio_t * pio,dsl_pool_t * dp,uint64_t txg,const blkptr_t * bpp)1419 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
1420 {
1421 ASSERT(dsl_pool_sync_context(dp));
1422 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
1423 }
1424
1425 static int
scan_ds_queue_compare(const void * a,const void * b)1426 scan_ds_queue_compare(const void *a, const void *b)
1427 {
1428 const scan_ds_t *sds_a = a, *sds_b = b;
1429
1430 if (sds_a->sds_dsobj < sds_b->sds_dsobj)
1431 return (-1);
1432 if (sds_a->sds_dsobj == sds_b->sds_dsobj)
1433 return (0);
1434 return (1);
1435 }
1436
1437 static void
scan_ds_queue_clear(dsl_scan_t * scn)1438 scan_ds_queue_clear(dsl_scan_t *scn)
1439 {
1440 void *cookie = NULL;
1441 scan_ds_t *sds;
1442 while ((sds = avl_destroy_nodes(&scn->scn_queue, &cookie)) != NULL) {
1443 kmem_free(sds, sizeof (*sds));
1444 }
1445 }
1446
1447 static boolean_t
scan_ds_queue_contains(dsl_scan_t * scn,uint64_t dsobj,uint64_t * txg)1448 scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, uint64_t *txg)
1449 {
1450 scan_ds_t srch, *sds;
1451
1452 srch.sds_dsobj = dsobj;
1453 sds = avl_find(&scn->scn_queue, &srch, NULL);
1454 if (sds != NULL && txg != NULL)
1455 *txg = sds->sds_txg;
1456 return (sds != NULL);
1457 }
1458
1459 static void
scan_ds_queue_insert(dsl_scan_t * scn,uint64_t dsobj,uint64_t txg)1460 scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg)
1461 {
1462 scan_ds_t *sds;
1463 avl_index_t where;
1464
1465 sds = kmem_zalloc(sizeof (*sds), KM_SLEEP);
1466 sds->sds_dsobj = dsobj;
1467 sds->sds_txg = txg;
1468
1469 VERIFY3P(avl_find(&scn->scn_queue, sds, &where), ==, NULL);
1470 avl_insert(&scn->scn_queue, sds, where);
1471 }
1472
1473 static void
scan_ds_queue_remove(dsl_scan_t * scn,uint64_t dsobj)1474 scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj)
1475 {
1476 scan_ds_t srch, *sds;
1477
1478 srch.sds_dsobj = dsobj;
1479
1480 sds = avl_find(&scn->scn_queue, &srch, NULL);
1481 VERIFY(sds != NULL);
1482 avl_remove(&scn->scn_queue, sds);
1483 kmem_free(sds, sizeof (*sds));
1484 }
1485
1486 static void
scan_ds_queue_sync(dsl_scan_t * scn,dmu_tx_t * tx)1487 scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx)
1488 {
1489 dsl_pool_t *dp = scn->scn_dp;
1490 spa_t *spa = dp->dp_spa;
1491 dmu_object_type_t ot = (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) ?
1492 DMU_OT_SCAN_QUEUE : DMU_OT_ZAP_OTHER;
1493
1494 ASSERT0(scn->scn_queues_pending);
1495 ASSERT(scn->scn_phys.scn_queue_obj != 0);
1496
1497 VERIFY0(dmu_object_free(dp->dp_meta_objset,
1498 scn->scn_phys.scn_queue_obj, tx));
1499 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, ot,
1500 DMU_OT_NONE, 0, tx);
1501 for (scan_ds_t *sds = avl_first(&scn->scn_queue);
1502 sds != NULL; sds = AVL_NEXT(&scn->scn_queue, sds)) {
1503 VERIFY0(zap_add_int_key(dp->dp_meta_objset,
1504 scn->scn_phys.scn_queue_obj, sds->sds_dsobj,
1505 sds->sds_txg, tx));
1506 }
1507 }
1508
1509 /*
1510 * Computes the memory limit state that we're currently in. A sorted scan
1511 * needs quite a bit of memory to hold the sorting queue, so we need to
1512 * reasonably constrain the size so it doesn't impact overall system
1513 * performance. We compute two limits:
1514 * 1) Hard memory limit: if the amount of memory used by the sorting
1515 * queues on a pool gets above this value, we stop the metadata
1516 * scanning portion and start issuing the queued up and sorted
1517 * I/Os to reduce memory usage.
1518 * This limit is calculated as a fraction of physmem (by default 5%).
1519 * We constrain the lower bound of the hard limit to an absolute
1520 * minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1521 * the upper bound to 5% of the total pool size - no chance we'll
1522 * ever need that much memory, but just to keep the value in check.
1523 * 2) Soft memory limit: once we hit the hard memory limit, we start
1524 * issuing I/O to reduce queue memory usage, but we don't want to
1525 * completely empty out the queues, since we might be able to find I/Os
1526 * that will fill in the gaps of our non-sequential IOs at some point
1527 * in the future. So we stop the issuing of I/Os once the amount of
1528 * memory used drops below the soft limit (at which point we stop issuing
1529 * I/O and start scanning metadata again).
1530 *
1531 * This limit is calculated by subtracting a fraction of the hard
1532 * limit from the hard limit. By default this fraction is 5%, so
1533 * the soft limit is 95% of the hard limit. We cap the size of the
1534 * difference between the hard and soft limits at an absolute
1535 * maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1536 * sufficient to not cause too frequent switching between the
1537 * metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1538 * worth of queues is about 1.2 GiB of on-pool data, so scanning
1539 * that should take at least a decent fraction of a second).
1540 */
1541 static boolean_t
dsl_scan_should_clear(dsl_scan_t * scn)1542 dsl_scan_should_clear(dsl_scan_t *scn)
1543 {
1544 spa_t *spa = scn->scn_dp->dp_spa;
1545 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
1546 uint64_t alloc, mlim_hard, mlim_soft, mused;
1547
1548 alloc = metaslab_class_get_alloc(spa_normal_class(spa));
1549 alloc += metaslab_class_get_alloc(spa_special_class(spa));
1550 alloc += metaslab_class_get_alloc(spa_dedup_class(spa));
1551
1552 mlim_hard = MAX((physmem / zfs_scan_mem_lim_fact) * PAGESIZE,
1553 zfs_scan_mem_lim_min);
1554 mlim_hard = MIN(mlim_hard, alloc / 20);
1555 mlim_soft = mlim_hard - MIN(mlim_hard / zfs_scan_mem_lim_soft_fact,
1556 zfs_scan_mem_lim_soft_max);
1557 mused = 0;
1558 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1559 vdev_t *tvd = rvd->vdev_child[i];
1560 dsl_scan_io_queue_t *queue;
1561
1562 mutex_enter(&tvd->vdev_scan_io_queue_lock);
1563 queue = tvd->vdev_scan_io_queue;
1564 if (queue != NULL) {
1565 /*
1566 * # of extents in exts_by_addr = # in exts_by_size.
1567 * B-tree efficiency is ~75%, but can be as low as 50%.
1568 */
1569 mused += zfs_btree_numnodes(&queue->q_exts_by_size) *
1570 ((sizeof (range_seg_gap_t) + sizeof (uint64_t)) *
1571 3 / 2) + queue->q_sio_memused;
1572 }
1573 mutex_exit(&tvd->vdev_scan_io_queue_lock);
1574 }
1575
1576 dprintf("current scan memory usage: %llu bytes\n", (longlong_t)mused);
1577
1578 if (mused == 0)
1579 ASSERT0(scn->scn_queues_pending);
1580
1581 /*
1582 * If we are above our hard limit, we need to clear out memory.
1583 * If we are below our soft limit, we need to accumulate sequential IOs.
1584 * Otherwise, we should keep doing whatever we are currently doing.
1585 */
1586 if (mused >= mlim_hard)
1587 return (B_TRUE);
1588 else if (mused < mlim_soft)
1589 return (B_FALSE);
1590 else
1591 return (scn->scn_clearing);
1592 }
1593
1594 static boolean_t
dsl_scan_check_suspend(dsl_scan_t * scn,const zbookmark_phys_t * zb)1595 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1596 {
1597 /* we never skip user/group accounting objects */
1598 if (zb && (int64_t)zb->zb_object < 0)
1599 return (B_FALSE);
1600
1601 if (scn->scn_suspending)
1602 return (B_TRUE); /* we're already suspending */
1603
1604 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
1605 return (B_FALSE); /* we're resuming */
1606
1607 /* We only know how to resume from level-0 and objset blocks. */
1608 if (zb && (zb->zb_level != 0 && zb->zb_level != ZB_ROOT_LEVEL))
1609 return (B_FALSE);
1610
1611 /*
1612 * We suspend if:
1613 * - we have scanned for at least the minimum time (default 1 sec
1614 * for scrub, 3 sec for resilver), and either we have sufficient
1615 * dirty data that we are starting to write more quickly
1616 * (default 30%), someone is explicitly waiting for this txg
1617 * to complete, or we have used up all of the time in the txg
1618 * timeout (default 5 sec).
1619 * or
1620 * - the spa is shutting down because this pool is being exported
1621 * or the machine is rebooting.
1622 * or
1623 * - the scan queue has reached its memory use limit
1624 */
1625 uint64_t curr_time_ns = gethrtime();
1626 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
1627 uint64_t sync_time_ns = curr_time_ns -
1628 scn->scn_dp->dp_spa->spa_sync_starttime;
1629 uint64_t dirty_min_bytes = zfs_dirty_data_max *
1630 zfs_vdev_async_write_active_min_dirty_percent / 100;
1631 uint_t mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
1632 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
1633
1634 if ((NSEC2MSEC(scan_time_ns) > mintime &&
1635 (scn->scn_dp->dp_dirty_total >= dirty_min_bytes ||
1636 txg_sync_waiting(scn->scn_dp) ||
1637 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1638 spa_shutting_down(scn->scn_dp->dp_spa) ||
1639 (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn))) {
1640 if (zb && zb->zb_level == ZB_ROOT_LEVEL) {
1641 dprintf("suspending at first available bookmark "
1642 "%llx/%llx/%llx/%llx\n",
1643 (longlong_t)zb->zb_objset,
1644 (longlong_t)zb->zb_object,
1645 (longlong_t)zb->zb_level,
1646 (longlong_t)zb->zb_blkid);
1647 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
1648 zb->zb_objset, 0, 0, 0);
1649 } else if (zb != NULL) {
1650 dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1651 (longlong_t)zb->zb_objset,
1652 (longlong_t)zb->zb_object,
1653 (longlong_t)zb->zb_level,
1654 (longlong_t)zb->zb_blkid);
1655 scn->scn_phys.scn_bookmark = *zb;
1656 } else {
1657 #ifdef ZFS_DEBUG
1658 dsl_scan_phys_t *scnp = &scn->scn_phys;
1659 dprintf("suspending at at DDT bookmark "
1660 "%llx/%llx/%llx/%llx\n",
1661 (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
1662 (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
1663 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
1664 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
1665 #endif
1666 }
1667 scn->scn_suspending = B_TRUE;
1668 return (B_TRUE);
1669 }
1670 return (B_FALSE);
1671 }
1672
1673 static boolean_t
dsl_error_scrub_check_suspend(dsl_scan_t * scn,const zbookmark_phys_t * zb)1674 dsl_error_scrub_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1675 {
1676 /*
1677 * We suspend if:
1678 * - we have scrubbed for at least the minimum time (default 1 sec
1679 * for error scrub), someone is explicitly waiting for this txg
1680 * to complete, or we have used up all of the time in the txg
1681 * timeout (default 5 sec).
1682 * or
1683 * - the spa is shutting down because this pool is being exported
1684 * or the machine is rebooting.
1685 */
1686 uint64_t curr_time_ns = gethrtime();
1687 uint64_t error_scrub_time_ns = curr_time_ns - scn->scn_sync_start_time;
1688 uint64_t sync_time_ns = curr_time_ns -
1689 scn->scn_dp->dp_spa->spa_sync_starttime;
1690 int mintime = zfs_scrub_min_time_ms;
1691
1692 if ((NSEC2MSEC(error_scrub_time_ns) > mintime &&
1693 (txg_sync_waiting(scn->scn_dp) ||
1694 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1695 spa_shutting_down(scn->scn_dp->dp_spa)) {
1696 if (zb) {
1697 dprintf("error scrub suspending at bookmark "
1698 "%llx/%llx/%llx/%llx\n",
1699 (longlong_t)zb->zb_objset,
1700 (longlong_t)zb->zb_object,
1701 (longlong_t)zb->zb_level,
1702 (longlong_t)zb->zb_blkid);
1703 }
1704 return (B_TRUE);
1705 }
1706 return (B_FALSE);
1707 }
1708
1709 typedef struct zil_scan_arg {
1710 dsl_pool_t *zsa_dp;
1711 zil_header_t *zsa_zh;
1712 } zil_scan_arg_t;
1713
1714 static int
dsl_scan_zil_block(zilog_t * zilog,const blkptr_t * bp,void * arg,uint64_t claim_txg)1715 dsl_scan_zil_block(zilog_t *zilog, const blkptr_t *bp, void *arg,
1716 uint64_t claim_txg)
1717 {
1718 (void) zilog;
1719 zil_scan_arg_t *zsa = arg;
1720 dsl_pool_t *dp = zsa->zsa_dp;
1721 dsl_scan_t *scn = dp->dp_scan;
1722 zil_header_t *zh = zsa->zsa_zh;
1723 zbookmark_phys_t zb;
1724
1725 ASSERT(!BP_IS_REDACTED(bp));
1726 if (BP_IS_HOLE(bp) ||
1727 BP_GET_LOGICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_min_txg)
1728 return (0);
1729
1730 /*
1731 * One block ("stubby") can be allocated a long time ago; we
1732 * want to visit that one because it has been allocated
1733 * (on-disk) even if it hasn't been claimed (even though for
1734 * scrub there's nothing to do to it).
1735 */
1736 if (claim_txg == 0 &&
1737 BP_GET_LOGICAL_BIRTH(bp) >= spa_min_claim_txg(dp->dp_spa))
1738 return (0);
1739
1740 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1741 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
1742
1743 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1744 return (0);
1745 }
1746
1747 static int
dsl_scan_zil_record(zilog_t * zilog,const lr_t * lrc,void * arg,uint64_t claim_txg)1748 dsl_scan_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg,
1749 uint64_t claim_txg)
1750 {
1751 (void) zilog;
1752 if (lrc->lrc_txtype == TX_WRITE) {
1753 zil_scan_arg_t *zsa = arg;
1754 dsl_pool_t *dp = zsa->zsa_dp;
1755 dsl_scan_t *scn = dp->dp_scan;
1756 zil_header_t *zh = zsa->zsa_zh;
1757 const lr_write_t *lr = (const lr_write_t *)lrc;
1758 const blkptr_t *bp = &lr->lr_blkptr;
1759 zbookmark_phys_t zb;
1760
1761 ASSERT(!BP_IS_REDACTED(bp));
1762 if (BP_IS_HOLE(bp) ||
1763 BP_GET_LOGICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_min_txg)
1764 return (0);
1765
1766 /*
1767 * birth can be < claim_txg if this record's txg is
1768 * already txg sync'ed (but this log block contains
1769 * other records that are not synced)
1770 */
1771 if (claim_txg == 0 || BP_GET_LOGICAL_BIRTH(bp) < claim_txg)
1772 return (0);
1773
1774 ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
1775 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1776 lr->lr_foid, ZB_ZIL_LEVEL,
1777 lr->lr_offset / BP_GET_LSIZE(bp));
1778
1779 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1780 }
1781 return (0);
1782 }
1783
1784 static void
dsl_scan_zil(dsl_pool_t * dp,zil_header_t * zh)1785 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
1786 {
1787 uint64_t claim_txg = zh->zh_claim_txg;
1788 zil_scan_arg_t zsa = { dp, zh };
1789 zilog_t *zilog;
1790
1791 ASSERT(spa_writeable(dp->dp_spa));
1792
1793 /*
1794 * We only want to visit blocks that have been claimed but not yet
1795 * replayed (or, in read-only mode, blocks that *would* be claimed).
1796 */
1797 if (claim_txg == 0)
1798 return;
1799
1800 zilog = zil_alloc(dp->dp_meta_objset, zh);
1801
1802 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
1803 claim_txg, B_FALSE);
1804
1805 zil_free(zilog);
1806 }
1807
1808 /*
1809 * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1810 * here is to sort the AVL tree by the order each block will be needed.
1811 */
1812 static int
scan_prefetch_queue_compare(const void * a,const void * b)1813 scan_prefetch_queue_compare(const void *a, const void *b)
1814 {
1815 const scan_prefetch_issue_ctx_t *spic_a = a, *spic_b = b;
1816 const scan_prefetch_ctx_t *spc_a = spic_a->spic_spc;
1817 const scan_prefetch_ctx_t *spc_b = spic_b->spic_spc;
1818
1819 return (zbookmark_compare(spc_a->spc_datablkszsec,
1820 spc_a->spc_indblkshift, spc_b->spc_datablkszsec,
1821 spc_b->spc_indblkshift, &spic_a->spic_zb, &spic_b->spic_zb));
1822 }
1823
1824 static void
scan_prefetch_ctx_rele(scan_prefetch_ctx_t * spc,const void * tag)1825 scan_prefetch_ctx_rele(scan_prefetch_ctx_t *spc, const void *tag)
1826 {
1827 if (zfs_refcount_remove(&spc->spc_refcnt, tag) == 0) {
1828 zfs_refcount_destroy(&spc->spc_refcnt);
1829 kmem_free(spc, sizeof (scan_prefetch_ctx_t));
1830 }
1831 }
1832
1833 static scan_prefetch_ctx_t *
scan_prefetch_ctx_create(dsl_scan_t * scn,dnode_phys_t * dnp,const void * tag)1834 scan_prefetch_ctx_create(dsl_scan_t *scn, dnode_phys_t *dnp, const void *tag)
1835 {
1836 scan_prefetch_ctx_t *spc;
1837
1838 spc = kmem_alloc(sizeof (scan_prefetch_ctx_t), KM_SLEEP);
1839 zfs_refcount_create(&spc->spc_refcnt);
1840 zfs_refcount_add(&spc->spc_refcnt, tag);
1841 spc->spc_scn = scn;
1842 if (dnp != NULL) {
1843 spc->spc_datablkszsec = dnp->dn_datablkszsec;
1844 spc->spc_indblkshift = dnp->dn_indblkshift;
1845 spc->spc_root = B_FALSE;
1846 } else {
1847 spc->spc_datablkszsec = 0;
1848 spc->spc_indblkshift = 0;
1849 spc->spc_root = B_TRUE;
1850 }
1851
1852 return (spc);
1853 }
1854
1855 static void
scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t * spc,const void * tag)1856 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t *spc, const void *tag)
1857 {
1858 zfs_refcount_add(&spc->spc_refcnt, tag);
1859 }
1860
1861 static void
scan_ds_prefetch_queue_clear(dsl_scan_t * scn)1862 scan_ds_prefetch_queue_clear(dsl_scan_t *scn)
1863 {
1864 spa_t *spa = scn->scn_dp->dp_spa;
1865 void *cookie = NULL;
1866 scan_prefetch_issue_ctx_t *spic = NULL;
1867
1868 mutex_enter(&spa->spa_scrub_lock);
1869 while ((spic = avl_destroy_nodes(&scn->scn_prefetch_queue,
1870 &cookie)) != NULL) {
1871 scan_prefetch_ctx_rele(spic->spic_spc, scn);
1872 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1873 }
1874 mutex_exit(&spa->spa_scrub_lock);
1875 }
1876
1877 static boolean_t
dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t * spc,const zbookmark_phys_t * zb)1878 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t *spc,
1879 const zbookmark_phys_t *zb)
1880 {
1881 zbookmark_phys_t *last_zb = &spc->spc_scn->scn_prefetch_bookmark;
1882 dnode_phys_t tmp_dnp;
1883 dnode_phys_t *dnp = (spc->spc_root) ? NULL : &tmp_dnp;
1884
1885 if (zb->zb_objset != last_zb->zb_objset)
1886 return (B_TRUE);
1887 if ((int64_t)zb->zb_object < 0)
1888 return (B_FALSE);
1889
1890 tmp_dnp.dn_datablkszsec = spc->spc_datablkszsec;
1891 tmp_dnp.dn_indblkshift = spc->spc_indblkshift;
1892
1893 if (zbookmark_subtree_completed(dnp, zb, last_zb))
1894 return (B_TRUE);
1895
1896 return (B_FALSE);
1897 }
1898
1899 static void
dsl_scan_prefetch(scan_prefetch_ctx_t * spc,blkptr_t * bp,zbookmark_phys_t * zb)1900 dsl_scan_prefetch(scan_prefetch_ctx_t *spc, blkptr_t *bp, zbookmark_phys_t *zb)
1901 {
1902 avl_index_t idx;
1903 dsl_scan_t *scn = spc->spc_scn;
1904 spa_t *spa = scn->scn_dp->dp_spa;
1905 scan_prefetch_issue_ctx_t *spic;
1906
1907 if (zfs_no_scrub_prefetch || BP_IS_REDACTED(bp))
1908 return;
1909
1910 if (BP_IS_HOLE(bp) ||
1911 BP_GET_LOGICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_min_txg ||
1912 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE &&
1913 BP_GET_TYPE(bp) != DMU_OT_OBJSET))
1914 return;
1915
1916 if (dsl_scan_check_prefetch_resume(spc, zb))
1917 return;
1918
1919 scan_prefetch_ctx_add_ref(spc, scn);
1920 spic = kmem_alloc(sizeof (scan_prefetch_issue_ctx_t), KM_SLEEP);
1921 spic->spic_spc = spc;
1922 spic->spic_bp = *bp;
1923 spic->spic_zb = *zb;
1924
1925 /*
1926 * Add the IO to the queue of blocks to prefetch. This allows us to
1927 * prioritize blocks that we will need first for the main traversal
1928 * thread.
1929 */
1930 mutex_enter(&spa->spa_scrub_lock);
1931 if (avl_find(&scn->scn_prefetch_queue, spic, &idx) != NULL) {
1932 /* this block is already queued for prefetch */
1933 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1934 scan_prefetch_ctx_rele(spc, scn);
1935 mutex_exit(&spa->spa_scrub_lock);
1936 return;
1937 }
1938
1939 avl_insert(&scn->scn_prefetch_queue, spic, idx);
1940 cv_broadcast(&spa->spa_scrub_io_cv);
1941 mutex_exit(&spa->spa_scrub_lock);
1942 }
1943
1944 static void
dsl_scan_prefetch_dnode(dsl_scan_t * scn,dnode_phys_t * dnp,uint64_t objset,uint64_t object)1945 dsl_scan_prefetch_dnode(dsl_scan_t *scn, dnode_phys_t *dnp,
1946 uint64_t objset, uint64_t object)
1947 {
1948 int i;
1949 zbookmark_phys_t zb;
1950 scan_prefetch_ctx_t *spc;
1951
1952 if (dnp->dn_nblkptr == 0 && !(dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1953 return;
1954
1955 SET_BOOKMARK(&zb, objset, object, 0, 0);
1956
1957 spc = scan_prefetch_ctx_create(scn, dnp, FTAG);
1958
1959 for (i = 0; i < dnp->dn_nblkptr; i++) {
1960 zb.zb_level = BP_GET_LEVEL(&dnp->dn_blkptr[i]);
1961 zb.zb_blkid = i;
1962 dsl_scan_prefetch(spc, &dnp->dn_blkptr[i], &zb);
1963 }
1964
1965 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1966 zb.zb_level = 0;
1967 zb.zb_blkid = DMU_SPILL_BLKID;
1968 dsl_scan_prefetch(spc, DN_SPILL_BLKPTR(dnp), &zb);
1969 }
1970
1971 scan_prefetch_ctx_rele(spc, FTAG);
1972 }
1973
1974 static void
dsl_scan_prefetch_cb(zio_t * zio,const zbookmark_phys_t * zb,const blkptr_t * bp,arc_buf_t * buf,void * private)1975 dsl_scan_prefetch_cb(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1976 arc_buf_t *buf, void *private)
1977 {
1978 (void) zio;
1979 scan_prefetch_ctx_t *spc = private;
1980 dsl_scan_t *scn = spc->spc_scn;
1981 spa_t *spa = scn->scn_dp->dp_spa;
1982
1983 /* broadcast that the IO has completed for rate limiting purposes */
1984 mutex_enter(&spa->spa_scrub_lock);
1985 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
1986 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
1987 cv_broadcast(&spa->spa_scrub_io_cv);
1988 mutex_exit(&spa->spa_scrub_lock);
1989
1990 /* if there was an error or we are done prefetching, just cleanup */
1991 if (buf == NULL || scn->scn_prefetch_stop)
1992 goto out;
1993
1994 if (BP_GET_LEVEL(bp) > 0) {
1995 int i;
1996 blkptr_t *cbp;
1997 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1998 zbookmark_phys_t czb;
1999
2000 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
2001 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
2002 zb->zb_level - 1, zb->zb_blkid * epb + i);
2003 dsl_scan_prefetch(spc, cbp, &czb);
2004 }
2005 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
2006 dnode_phys_t *cdnp;
2007 int i;
2008 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
2009
2010 for (i = 0, cdnp = buf->b_data; i < epb;
2011 i += cdnp->dn_extra_slots + 1,
2012 cdnp += cdnp->dn_extra_slots + 1) {
2013 dsl_scan_prefetch_dnode(scn, cdnp,
2014 zb->zb_objset, zb->zb_blkid * epb + i);
2015 }
2016 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
2017 objset_phys_t *osp = buf->b_data;
2018
2019 dsl_scan_prefetch_dnode(scn, &osp->os_meta_dnode,
2020 zb->zb_objset, DMU_META_DNODE_OBJECT);
2021
2022 if (OBJSET_BUF_HAS_USERUSED(buf)) {
2023 if (OBJSET_BUF_HAS_PROJECTUSED(buf)) {
2024 dsl_scan_prefetch_dnode(scn,
2025 &osp->os_projectused_dnode, zb->zb_objset,
2026 DMU_PROJECTUSED_OBJECT);
2027 }
2028 dsl_scan_prefetch_dnode(scn,
2029 &osp->os_groupused_dnode, zb->zb_objset,
2030 DMU_GROUPUSED_OBJECT);
2031 dsl_scan_prefetch_dnode(scn,
2032 &osp->os_userused_dnode, zb->zb_objset,
2033 DMU_USERUSED_OBJECT);
2034 }
2035 }
2036
2037 out:
2038 if (buf != NULL)
2039 arc_buf_destroy(buf, private);
2040 scan_prefetch_ctx_rele(spc, scn);
2041 }
2042
2043 static void
dsl_scan_prefetch_thread(void * arg)2044 dsl_scan_prefetch_thread(void *arg)
2045 {
2046 dsl_scan_t *scn = arg;
2047 spa_t *spa = scn->scn_dp->dp_spa;
2048 scan_prefetch_issue_ctx_t *spic;
2049
2050 /* loop until we are told to stop */
2051 while (!scn->scn_prefetch_stop) {
2052 arc_flags_t flags = ARC_FLAG_NOWAIT |
2053 ARC_FLAG_PRESCIENT_PREFETCH | ARC_FLAG_PREFETCH;
2054 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
2055
2056 mutex_enter(&spa->spa_scrub_lock);
2057
2058 /*
2059 * Wait until we have an IO to issue and are not above our
2060 * maximum in flight limit.
2061 */
2062 while (!scn->scn_prefetch_stop &&
2063 (avl_numnodes(&scn->scn_prefetch_queue) == 0 ||
2064 spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)) {
2065 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
2066 }
2067
2068 /* recheck if we should stop since we waited for the cv */
2069 if (scn->scn_prefetch_stop) {
2070 mutex_exit(&spa->spa_scrub_lock);
2071 break;
2072 }
2073
2074 /* remove the prefetch IO from the tree */
2075 spic = avl_first(&scn->scn_prefetch_queue);
2076 spa->spa_scrub_inflight += BP_GET_PSIZE(&spic->spic_bp);
2077 avl_remove(&scn->scn_prefetch_queue, spic);
2078
2079 mutex_exit(&spa->spa_scrub_lock);
2080
2081 if (BP_IS_PROTECTED(&spic->spic_bp)) {
2082 ASSERT(BP_GET_TYPE(&spic->spic_bp) == DMU_OT_DNODE ||
2083 BP_GET_TYPE(&spic->spic_bp) == DMU_OT_OBJSET);
2084 ASSERT3U(BP_GET_LEVEL(&spic->spic_bp), ==, 0);
2085 zio_flags |= ZIO_FLAG_RAW;
2086 }
2087
2088 /* We don't need data L1 buffer since we do not prefetch L0. */
2089 blkptr_t *bp = &spic->spic_bp;
2090 if (BP_GET_LEVEL(bp) == 1 && BP_GET_TYPE(bp) != DMU_OT_DNODE &&
2091 BP_GET_TYPE(bp) != DMU_OT_OBJSET)
2092 flags |= ARC_FLAG_NO_BUF;
2093
2094 /* issue the prefetch asynchronously */
2095 (void) arc_read(scn->scn_zio_root, spa, bp,
2096 dsl_scan_prefetch_cb, spic->spic_spc, ZIO_PRIORITY_SCRUB,
2097 zio_flags, &flags, &spic->spic_zb);
2098
2099 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
2100 }
2101
2102 ASSERT(scn->scn_prefetch_stop);
2103
2104 /* free any prefetches we didn't get to complete */
2105 mutex_enter(&spa->spa_scrub_lock);
2106 while ((spic = avl_first(&scn->scn_prefetch_queue)) != NULL) {
2107 avl_remove(&scn->scn_prefetch_queue, spic);
2108 scan_prefetch_ctx_rele(spic->spic_spc, scn);
2109 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
2110 }
2111 ASSERT0(avl_numnodes(&scn->scn_prefetch_queue));
2112 mutex_exit(&spa->spa_scrub_lock);
2113 }
2114
2115 static boolean_t
dsl_scan_check_resume(dsl_scan_t * scn,const dnode_phys_t * dnp,const zbookmark_phys_t * zb)2116 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
2117 const zbookmark_phys_t *zb)
2118 {
2119 /*
2120 * We never skip over user/group accounting objects (obj<0)
2121 */
2122 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
2123 (int64_t)zb->zb_object >= 0) {
2124 /*
2125 * If we already visited this bp & everything below (in
2126 * a prior txg sync), don't bother doing it again.
2127 */
2128 if (zbookmark_subtree_completed(dnp, zb,
2129 &scn->scn_phys.scn_bookmark))
2130 return (B_TRUE);
2131
2132 /*
2133 * If we found the block we're trying to resume from, or
2134 * we went past it, zero it out to indicate that it's OK
2135 * to start checking for suspending again.
2136 */
2137 if (zbookmark_subtree_tbd(dnp, zb,
2138 &scn->scn_phys.scn_bookmark)) {
2139 dprintf("resuming at %llx/%llx/%llx/%llx\n",
2140 (longlong_t)zb->zb_objset,
2141 (longlong_t)zb->zb_object,
2142 (longlong_t)zb->zb_level,
2143 (longlong_t)zb->zb_blkid);
2144 memset(&scn->scn_phys.scn_bookmark, 0, sizeof (*zb));
2145 }
2146 }
2147 return (B_FALSE);
2148 }
2149
2150 static void dsl_scan_visitbp(const blkptr_t *bp, const zbookmark_phys_t *zb,
2151 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
2152 dmu_objset_type_t ostype, dmu_tx_t *tx);
2153 inline __attribute__((always_inline)) static void dsl_scan_visitdnode(
2154 dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
2155 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
2156
2157 /*
2158 * Return nonzero on i/o error.
2159 * Return new buf to write out in *bufp.
2160 */
2161 inline __attribute__((always_inline)) static int
dsl_scan_recurse(dsl_scan_t * scn,dsl_dataset_t * ds,dmu_objset_type_t ostype,dnode_phys_t * dnp,const blkptr_t * bp,const zbookmark_phys_t * zb,dmu_tx_t * tx)2162 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
2163 dnode_phys_t *dnp, const blkptr_t *bp,
2164 const zbookmark_phys_t *zb, dmu_tx_t *tx)
2165 {
2166 dsl_pool_t *dp = scn->scn_dp;
2167 spa_t *spa = dp->dp_spa;
2168 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
2169 int err;
2170
2171 ASSERT(!BP_IS_REDACTED(bp));
2172
2173 /*
2174 * There is an unlikely case of encountering dnodes with contradicting
2175 * dn_bonuslen and DNODE_FLAG_SPILL_BLKPTR flag before in files created
2176 * or modified before commit 4254acb was merged. As it is not possible
2177 * to know which of the two is correct, report an error.
2178 */
2179 if (dnp != NULL &&
2180 dnp->dn_bonuslen > DN_MAX_BONUS_LEN(dnp)) {
2181 scn->scn_phys.scn_errors++;
2182 spa_log_error(spa, zb, BP_GET_LOGICAL_BIRTH(bp));
2183 return (SET_ERROR(EINVAL));
2184 }
2185
2186 if (BP_GET_LEVEL(bp) > 0) {
2187 arc_flags_t flags = ARC_FLAG_WAIT;
2188 int i;
2189 blkptr_t *cbp;
2190 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
2191 arc_buf_t *buf;
2192
2193 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
2194 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
2195 if (err) {
2196 scn->scn_phys.scn_errors++;
2197 return (err);
2198 }
2199 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
2200 zbookmark_phys_t czb;
2201
2202 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
2203 zb->zb_level - 1,
2204 zb->zb_blkid * epb + i);
2205 dsl_scan_visitbp(cbp, &czb, dnp,
2206 ds, scn, ostype, tx);
2207 }
2208 arc_buf_destroy(buf, &buf);
2209 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
2210 arc_flags_t flags = ARC_FLAG_WAIT;
2211 dnode_phys_t *cdnp;
2212 int i;
2213 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
2214 arc_buf_t *buf;
2215
2216 if (BP_IS_PROTECTED(bp)) {
2217 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
2218 zio_flags |= ZIO_FLAG_RAW;
2219 }
2220
2221 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
2222 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
2223 if (err) {
2224 scn->scn_phys.scn_errors++;
2225 return (err);
2226 }
2227 for (i = 0, cdnp = buf->b_data; i < epb;
2228 i += cdnp->dn_extra_slots + 1,
2229 cdnp += cdnp->dn_extra_slots + 1) {
2230 dsl_scan_visitdnode(scn, ds, ostype,
2231 cdnp, zb->zb_blkid * epb + i, tx);
2232 }
2233
2234 arc_buf_destroy(buf, &buf);
2235 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
2236 arc_flags_t flags = ARC_FLAG_WAIT;
2237 objset_phys_t *osp;
2238 arc_buf_t *buf;
2239
2240 err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
2241 ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
2242 if (err) {
2243 scn->scn_phys.scn_errors++;
2244 return (err);
2245 }
2246
2247 osp = buf->b_data;
2248
2249 dsl_scan_visitdnode(scn, ds, osp->os_type,
2250 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
2251
2252 if (OBJSET_BUF_HAS_USERUSED(buf)) {
2253 /*
2254 * We also always visit user/group/project accounting
2255 * objects, and never skip them, even if we are
2256 * suspending. This is necessary so that the
2257 * space deltas from this txg get integrated.
2258 */
2259 if (OBJSET_BUF_HAS_PROJECTUSED(buf))
2260 dsl_scan_visitdnode(scn, ds, osp->os_type,
2261 &osp->os_projectused_dnode,
2262 DMU_PROJECTUSED_OBJECT, tx);
2263 dsl_scan_visitdnode(scn, ds, osp->os_type,
2264 &osp->os_groupused_dnode,
2265 DMU_GROUPUSED_OBJECT, tx);
2266 dsl_scan_visitdnode(scn, ds, osp->os_type,
2267 &osp->os_userused_dnode,
2268 DMU_USERUSED_OBJECT, tx);
2269 }
2270 arc_buf_destroy(buf, &buf);
2271 } else if (!zfs_blkptr_verify(spa, bp,
2272 BLK_CONFIG_NEEDED, BLK_VERIFY_LOG)) {
2273 /*
2274 * Sanity check the block pointer contents, this is handled
2275 * by arc_read() for the cases above.
2276 */
2277 scn->scn_phys.scn_errors++;
2278 spa_log_error(spa, zb, BP_GET_LOGICAL_BIRTH(bp));
2279 return (SET_ERROR(EINVAL));
2280 }
2281
2282 return (0);
2283 }
2284
2285 inline __attribute__((always_inline)) static void
dsl_scan_visitdnode(dsl_scan_t * scn,dsl_dataset_t * ds,dmu_objset_type_t ostype,dnode_phys_t * dnp,uint64_t object,dmu_tx_t * tx)2286 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
2287 dmu_objset_type_t ostype, dnode_phys_t *dnp,
2288 uint64_t object, dmu_tx_t *tx)
2289 {
2290 int j;
2291
2292 for (j = 0; j < dnp->dn_nblkptr; j++) {
2293 zbookmark_phys_t czb;
2294
2295 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
2296 dnp->dn_nlevels - 1, j);
2297 dsl_scan_visitbp(&dnp->dn_blkptr[j],
2298 &czb, dnp, ds, scn, ostype, tx);
2299 }
2300
2301 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
2302 zbookmark_phys_t czb;
2303 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
2304 0, DMU_SPILL_BLKID);
2305 dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp),
2306 &czb, dnp, ds, scn, ostype, tx);
2307 }
2308 }
2309
2310 /*
2311 * The arguments are in this order because mdb can only print the
2312 * first 5; we want them to be useful.
2313 */
2314 static void
dsl_scan_visitbp(const blkptr_t * bp,const zbookmark_phys_t * zb,dnode_phys_t * dnp,dsl_dataset_t * ds,dsl_scan_t * scn,dmu_objset_type_t ostype,dmu_tx_t * tx)2315 dsl_scan_visitbp(const blkptr_t *bp, const zbookmark_phys_t *zb,
2316 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
2317 dmu_objset_type_t ostype, dmu_tx_t *tx)
2318 {
2319 dsl_pool_t *dp = scn->scn_dp;
2320
2321 if (dsl_scan_check_suspend(scn, zb))
2322 return;
2323
2324 if (dsl_scan_check_resume(scn, dnp, zb))
2325 return;
2326
2327 scn->scn_visited_this_txg++;
2328
2329 if (BP_IS_HOLE(bp)) {
2330 scn->scn_holes_this_txg++;
2331 return;
2332 }
2333
2334 if (BP_IS_REDACTED(bp)) {
2335 ASSERT(dsl_dataset_feature_is_active(ds,
2336 SPA_FEATURE_REDACTED_DATASETS));
2337 return;
2338 }
2339
2340 /*
2341 * Check if this block contradicts any filesystem flags.
2342 */
2343 spa_feature_t f = SPA_FEATURE_LARGE_BLOCKS;
2344 if (BP_GET_LSIZE(bp) > SPA_OLD_MAXBLOCKSIZE)
2345 ASSERT(dsl_dataset_feature_is_active(ds, f));
2346
2347 f = zio_checksum_to_feature(BP_GET_CHECKSUM(bp));
2348 if (f != SPA_FEATURE_NONE)
2349 ASSERT(dsl_dataset_feature_is_active(ds, f));
2350
2351 f = zio_compress_to_feature(BP_GET_COMPRESS(bp));
2352 if (f != SPA_FEATURE_NONE)
2353 ASSERT(dsl_dataset_feature_is_active(ds, f));
2354
2355 if (BP_GET_LOGICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_min_txg) {
2356 scn->scn_lt_min_this_txg++;
2357 return;
2358 }
2359
2360 if (dsl_scan_recurse(scn, ds, ostype, dnp, bp, zb, tx) != 0)
2361 return;
2362
2363 /*
2364 * If dsl_scan_ddt() has already visited this block, it will have
2365 * already done any translations or scrubbing, so don't call the
2366 * callback again.
2367 */
2368 if (ddt_class_contains(dp->dp_spa,
2369 scn->scn_phys.scn_ddt_class_max, bp)) {
2370 scn->scn_ddt_contained_this_txg++;
2371 return;
2372 }
2373
2374 /*
2375 * If this block is from the future (after cur_max_txg), then we
2376 * are doing this on behalf of a deleted snapshot, and we will
2377 * revisit the future block on the next pass of this dataset.
2378 * Don't scan it now unless we need to because something
2379 * under it was modified.
2380 */
2381 if (BP_GET_BIRTH(bp) > scn->scn_phys.scn_cur_max_txg) {
2382 scn->scn_gt_max_this_txg++;
2383 return;
2384 }
2385
2386 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
2387 }
2388
2389 static void
dsl_scan_visit_rootbp(dsl_scan_t * scn,dsl_dataset_t * ds,blkptr_t * bp,dmu_tx_t * tx)2390 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
2391 dmu_tx_t *tx)
2392 {
2393 zbookmark_phys_t zb;
2394 scan_prefetch_ctx_t *spc;
2395
2396 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
2397 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
2398
2399 if (ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) {
2400 SET_BOOKMARK(&scn->scn_prefetch_bookmark,
2401 zb.zb_objset, 0, 0, 0);
2402 } else {
2403 scn->scn_prefetch_bookmark = scn->scn_phys.scn_bookmark;
2404 }
2405
2406 scn->scn_objsets_visited_this_txg++;
2407
2408 spc = scan_prefetch_ctx_create(scn, NULL, FTAG);
2409 dsl_scan_prefetch(spc, bp, &zb);
2410 scan_prefetch_ctx_rele(spc, FTAG);
2411
2412 dsl_scan_visitbp(bp, &zb, NULL, ds, scn, DMU_OST_NONE, tx);
2413
2414 dprintf_ds(ds, "finished scan%s", "");
2415 }
2416
2417 static void
ds_destroyed_scn_phys(dsl_dataset_t * ds,dsl_scan_phys_t * scn_phys)2418 ds_destroyed_scn_phys(dsl_dataset_t *ds, dsl_scan_phys_t *scn_phys)
2419 {
2420 if (scn_phys->scn_bookmark.zb_objset == ds->ds_object) {
2421 if (ds->ds_is_snapshot) {
2422 /*
2423 * Note:
2424 * - scn_cur_{min,max}_txg stays the same.
2425 * - Setting the flag is not really necessary if
2426 * scn_cur_max_txg == scn_max_txg, because there
2427 * is nothing after this snapshot that we care
2428 * about. However, we set it anyway and then
2429 * ignore it when we retraverse it in
2430 * dsl_scan_visitds().
2431 */
2432 scn_phys->scn_bookmark.zb_objset =
2433 dsl_dataset_phys(ds)->ds_next_snap_obj;
2434 zfs_dbgmsg("destroying ds %llu on %s; currently "
2435 "traversing; reset zb_objset to %llu",
2436 (u_longlong_t)ds->ds_object,
2437 ds->ds_dir->dd_pool->dp_spa->spa_name,
2438 (u_longlong_t)dsl_dataset_phys(ds)->
2439 ds_next_snap_obj);
2440 scn_phys->scn_flags |= DSF_VISIT_DS_AGAIN;
2441 } else {
2442 SET_BOOKMARK(&scn_phys->scn_bookmark,
2443 ZB_DESTROYED_OBJSET, 0, 0, 0);
2444 zfs_dbgmsg("destroying ds %llu on %s; currently "
2445 "traversing; reset bookmark to -1,0,0,0",
2446 (u_longlong_t)ds->ds_object,
2447 ds->ds_dir->dd_pool->dp_spa->spa_name);
2448 }
2449 }
2450 }
2451
2452 /*
2453 * Invoked when a dataset is destroyed. We need to make sure that:
2454 *
2455 * 1) If it is the dataset that was currently being scanned, we write
2456 * a new dsl_scan_phys_t and marking the objset reference in it
2457 * as destroyed.
2458 * 2) Remove it from the work queue, if it was present.
2459 *
2460 * If the dataset was actually a snapshot, instead of marking the dataset
2461 * as destroyed, we instead substitute the next snapshot in line.
2462 */
2463 void
dsl_scan_ds_destroyed(dsl_dataset_t * ds,dmu_tx_t * tx)2464 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
2465 {
2466 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2467 dsl_scan_t *scn = dp->dp_scan;
2468 uint64_t mintxg;
2469
2470 if (!dsl_scan_is_running(scn))
2471 return;
2472
2473 ds_destroyed_scn_phys(ds, &scn->scn_phys);
2474 ds_destroyed_scn_phys(ds, &scn->scn_phys_cached);
2475
2476 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2477 scan_ds_queue_remove(scn, ds->ds_object);
2478 if (ds->ds_is_snapshot)
2479 scan_ds_queue_insert(scn,
2480 dsl_dataset_phys(ds)->ds_next_snap_obj, mintxg);
2481 }
2482
2483 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2484 ds->ds_object, &mintxg) == 0) {
2485 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
2486 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2487 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2488 if (ds->ds_is_snapshot) {
2489 /*
2490 * We keep the same mintxg; it could be >
2491 * ds_creation_txg if the previous snapshot was
2492 * deleted too.
2493 */
2494 VERIFY(zap_add_int_key(dp->dp_meta_objset,
2495 scn->scn_phys.scn_queue_obj,
2496 dsl_dataset_phys(ds)->ds_next_snap_obj,
2497 mintxg, tx) == 0);
2498 zfs_dbgmsg("destroying ds %llu on %s; in queue; "
2499 "replacing with %llu",
2500 (u_longlong_t)ds->ds_object,
2501 dp->dp_spa->spa_name,
2502 (u_longlong_t)dsl_dataset_phys(ds)->
2503 ds_next_snap_obj);
2504 } else {
2505 zfs_dbgmsg("destroying ds %llu on %s; in queue; "
2506 "removing",
2507 (u_longlong_t)ds->ds_object,
2508 dp->dp_spa->spa_name);
2509 }
2510 }
2511
2512 /*
2513 * dsl_scan_sync() should be called after this, and should sync
2514 * out our changed state, but just to be safe, do it here.
2515 */
2516 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2517 }
2518
2519 static void
ds_snapshotted_bookmark(dsl_dataset_t * ds,zbookmark_phys_t * scn_bookmark)2520 ds_snapshotted_bookmark(dsl_dataset_t *ds, zbookmark_phys_t *scn_bookmark)
2521 {
2522 if (scn_bookmark->zb_objset == ds->ds_object) {
2523 scn_bookmark->zb_objset =
2524 dsl_dataset_phys(ds)->ds_prev_snap_obj;
2525 zfs_dbgmsg("snapshotting ds %llu on %s; currently traversing; "
2526 "reset zb_objset to %llu",
2527 (u_longlong_t)ds->ds_object,
2528 ds->ds_dir->dd_pool->dp_spa->spa_name,
2529 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2530 }
2531 }
2532
2533 /*
2534 * Called when a dataset is snapshotted. If we were currently traversing
2535 * this snapshot, we reset our bookmark to point at the newly created
2536 * snapshot. We also modify our work queue to remove the old snapshot and
2537 * replace with the new one.
2538 */
2539 void
dsl_scan_ds_snapshotted(dsl_dataset_t * ds,dmu_tx_t * tx)2540 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
2541 {
2542 dsl_pool_t *dp = ds->ds_dir->dd_pool;
2543 dsl_scan_t *scn = dp->dp_scan;
2544 uint64_t mintxg;
2545
2546 if (!dsl_scan_is_running(scn))
2547 return;
2548
2549 ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
2550
2551 ds_snapshotted_bookmark(ds, &scn->scn_phys.scn_bookmark);
2552 ds_snapshotted_bookmark(ds, &scn->scn_phys_cached.scn_bookmark);
2553
2554 if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2555 scan_ds_queue_remove(scn, ds->ds_object);
2556 scan_ds_queue_insert(scn,
2557 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg);
2558 }
2559
2560 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2561 ds->ds_object, &mintxg) == 0) {
2562 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2563 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2564 VERIFY(zap_add_int_key(dp->dp_meta_objset,
2565 scn->scn_phys.scn_queue_obj,
2566 dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
2567 zfs_dbgmsg("snapshotting ds %llu on %s; in queue; "
2568 "replacing with %llu",
2569 (u_longlong_t)ds->ds_object,
2570 dp->dp_spa->spa_name,
2571 (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2572 }
2573
2574 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2575 }
2576
2577 static void
ds_clone_swapped_bookmark(dsl_dataset_t * ds1,dsl_dataset_t * ds2,zbookmark_phys_t * scn_bookmark)2578 ds_clone_swapped_bookmark(dsl_dataset_t *ds1, dsl_dataset_t *ds2,
2579 zbookmark_phys_t *scn_bookmark)
2580 {
2581 if (scn_bookmark->zb_objset == ds1->ds_object) {
2582 scn_bookmark->zb_objset = ds2->ds_object;
2583 zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; "
2584 "reset zb_objset to %llu",
2585 (u_longlong_t)ds1->ds_object,
2586 ds1->ds_dir->dd_pool->dp_spa->spa_name,
2587 (u_longlong_t)ds2->ds_object);
2588 } else if (scn_bookmark->zb_objset == ds2->ds_object) {
2589 scn_bookmark->zb_objset = ds1->ds_object;
2590 zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; "
2591 "reset zb_objset to %llu",
2592 (u_longlong_t)ds2->ds_object,
2593 ds2->ds_dir->dd_pool->dp_spa->spa_name,
2594 (u_longlong_t)ds1->ds_object);
2595 }
2596 }
2597
2598 /*
2599 * Called when an origin dataset and its clone are swapped. If we were
2600 * currently traversing the dataset, we need to switch to traversing the
2601 * newly promoted clone.
2602 */
2603 void
dsl_scan_ds_clone_swapped(dsl_dataset_t * ds1,dsl_dataset_t * ds2,dmu_tx_t * tx)2604 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
2605 {
2606 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
2607 dsl_scan_t *scn = dp->dp_scan;
2608 uint64_t mintxg1, mintxg2;
2609 boolean_t ds1_queued, ds2_queued;
2610
2611 if (!dsl_scan_is_running(scn))
2612 return;
2613
2614 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys.scn_bookmark);
2615 ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys_cached.scn_bookmark);
2616
2617 /*
2618 * Handle the in-memory scan queue.
2619 */
2620 ds1_queued = scan_ds_queue_contains(scn, ds1->ds_object, &mintxg1);
2621 ds2_queued = scan_ds_queue_contains(scn, ds2->ds_object, &mintxg2);
2622
2623 /* Sanity checking. */
2624 if (ds1_queued) {
2625 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2626 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2627 }
2628 if (ds2_queued) {
2629 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2630 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2631 }
2632
2633 if (ds1_queued && ds2_queued) {
2634 /*
2635 * If both are queued, we don't need to do anything.
2636 * The swapping code below would not handle this case correctly,
2637 * since we can't insert ds2 if it is already there. That's
2638 * because scan_ds_queue_insert() prohibits a duplicate insert
2639 * and panics.
2640 */
2641 } else if (ds1_queued) {
2642 scan_ds_queue_remove(scn, ds1->ds_object);
2643 scan_ds_queue_insert(scn, ds2->ds_object, mintxg1);
2644 } else if (ds2_queued) {
2645 scan_ds_queue_remove(scn, ds2->ds_object);
2646 scan_ds_queue_insert(scn, ds1->ds_object, mintxg2);
2647 }
2648
2649 /*
2650 * Handle the on-disk scan queue.
2651 * The on-disk state is an out-of-date version of the in-memory state,
2652 * so the in-memory and on-disk values for ds1_queued and ds2_queued may
2653 * be different. Therefore we need to apply the swap logic to the
2654 * on-disk state independently of the in-memory state.
2655 */
2656 ds1_queued = zap_lookup_int_key(dp->dp_meta_objset,
2657 scn->scn_phys.scn_queue_obj, ds1->ds_object, &mintxg1) == 0;
2658 ds2_queued = zap_lookup_int_key(dp->dp_meta_objset,
2659 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg2) == 0;
2660
2661 /* Sanity checking. */
2662 if (ds1_queued) {
2663 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2664 ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2665 }
2666 if (ds2_queued) {
2667 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2668 ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2669 }
2670
2671 if (ds1_queued && ds2_queued) {
2672 /*
2673 * If both are queued, we don't need to do anything.
2674 * Alternatively, we could check for EEXIST from
2675 * zap_add_int_key() and back out to the original state, but
2676 * that would be more work than checking for this case upfront.
2677 */
2678 } else if (ds1_queued) {
2679 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2680 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
2681 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2682 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg1, tx));
2683 zfs_dbgmsg("clone_swap ds %llu on %s; in queue; "
2684 "replacing with %llu",
2685 (u_longlong_t)ds1->ds_object,
2686 dp->dp_spa->spa_name,
2687 (u_longlong_t)ds2->ds_object);
2688 } else if (ds2_queued) {
2689 VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2690 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
2691 VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2692 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg2, tx));
2693 zfs_dbgmsg("clone_swap ds %llu on %s; in queue; "
2694 "replacing with %llu",
2695 (u_longlong_t)ds2->ds_object,
2696 dp->dp_spa->spa_name,
2697 (u_longlong_t)ds1->ds_object);
2698 }
2699
2700 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2701 }
2702
2703 static int
enqueue_clones_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)2704 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2705 {
2706 uint64_t originobj = *(uint64_t *)arg;
2707 dsl_dataset_t *ds;
2708 int err;
2709 dsl_scan_t *scn = dp->dp_scan;
2710
2711 if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != originobj)
2712 return (0);
2713
2714 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2715 if (err)
2716 return (err);
2717
2718 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != originobj) {
2719 dsl_dataset_t *prev;
2720 err = dsl_dataset_hold_obj(dp,
2721 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2722
2723 dsl_dataset_rele(ds, FTAG);
2724 if (err)
2725 return (err);
2726 ds = prev;
2727 }
2728 mutex_enter(&scn->scn_queue_lock);
2729 scan_ds_queue_insert(scn, ds->ds_object,
2730 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2731 mutex_exit(&scn->scn_queue_lock);
2732 dsl_dataset_rele(ds, FTAG);
2733 return (0);
2734 }
2735
2736 static void
dsl_scan_visitds(dsl_scan_t * scn,uint64_t dsobj,dmu_tx_t * tx)2737 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
2738 {
2739 dsl_pool_t *dp = scn->scn_dp;
2740 dsl_dataset_t *ds;
2741
2742 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2743
2744 if (scn->scn_phys.scn_cur_min_txg >=
2745 scn->scn_phys.scn_max_txg) {
2746 /*
2747 * This can happen if this snapshot was created after the
2748 * scan started, and we already completed a previous snapshot
2749 * that was created after the scan started. This snapshot
2750 * only references blocks with:
2751 *
2752 * birth < our ds_creation_txg
2753 * cur_min_txg is no less than ds_creation_txg.
2754 * We have already visited these blocks.
2755 * or
2756 * birth > scn_max_txg
2757 * The scan requested not to visit these blocks.
2758 *
2759 * Subsequent snapshots (and clones) can reference our
2760 * blocks, or blocks with even higher birth times.
2761 * Therefore we do not need to visit them either,
2762 * so we do not add them to the work queue.
2763 *
2764 * Note that checking for cur_min_txg >= cur_max_txg
2765 * is not sufficient, because in that case we may need to
2766 * visit subsequent snapshots. This happens when min_txg > 0,
2767 * which raises cur_min_txg. In this case we will visit
2768 * this dataset but skip all of its blocks, because the
2769 * rootbp's birth time is < cur_min_txg. Then we will
2770 * add the next snapshots/clones to the work queue.
2771 */
2772 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2773 dsl_dataset_name(ds, dsname);
2774 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2775 "cur_min_txg (%llu) >= max_txg (%llu)",
2776 (longlong_t)dsobj, dsname,
2777 (longlong_t)scn->scn_phys.scn_cur_min_txg,
2778 (longlong_t)scn->scn_phys.scn_max_txg);
2779 kmem_free(dsname, MAXNAMELEN);
2780
2781 goto out;
2782 }
2783
2784 /*
2785 * Only the ZIL in the head (non-snapshot) is valid. Even though
2786 * snapshots can have ZIL block pointers (which may be the same
2787 * BP as in the head), they must be ignored. In addition, $ORIGIN
2788 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2789 * need to look for a ZIL in it either. So we traverse the ZIL here,
2790 * rather than in scan_recurse(), because the regular snapshot
2791 * block-sharing rules don't apply to it.
2792 */
2793 if (!dsl_dataset_is_snapshot(ds) &&
2794 (dp->dp_origin_snap == NULL ||
2795 ds->ds_dir != dp->dp_origin_snap->ds_dir)) {
2796 objset_t *os;
2797 if (dmu_objset_from_ds(ds, &os) != 0) {
2798 goto out;
2799 }
2800 dsl_scan_zil(dp, &os->os_zil_header);
2801 }
2802
2803 /*
2804 * Iterate over the bps in this ds.
2805 */
2806 dmu_buf_will_dirty(ds->ds_dbuf, tx);
2807 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
2808 dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
2809 rrw_exit(&ds->ds_bp_rwlock, FTAG);
2810
2811 char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2812 dsl_dataset_name(ds, dsname);
2813 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2814 "suspending=%u",
2815 (longlong_t)dsobj, dsname,
2816 (longlong_t)scn->scn_phys.scn_cur_min_txg,
2817 (longlong_t)scn->scn_phys.scn_cur_max_txg,
2818 (int)scn->scn_suspending);
2819 kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
2820
2821 if (scn->scn_suspending)
2822 goto out;
2823
2824 /*
2825 * We've finished this pass over this dataset.
2826 */
2827
2828 /*
2829 * If we did not completely visit this dataset, do another pass.
2830 */
2831 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
2832 zfs_dbgmsg("incomplete pass on %s; visiting again",
2833 dp->dp_spa->spa_name);
2834 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
2835 scan_ds_queue_insert(scn, ds->ds_object,
2836 scn->scn_phys.scn_cur_max_txg);
2837 goto out;
2838 }
2839
2840 /*
2841 * Add descendant datasets to work queue.
2842 */
2843 if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
2844 scan_ds_queue_insert(scn,
2845 dsl_dataset_phys(ds)->ds_next_snap_obj,
2846 dsl_dataset_phys(ds)->ds_creation_txg);
2847 }
2848 if (dsl_dataset_phys(ds)->ds_num_children > 1) {
2849 boolean_t usenext = B_FALSE;
2850 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
2851 uint64_t count;
2852 /*
2853 * A bug in a previous version of the code could
2854 * cause upgrade_clones_cb() to not set
2855 * ds_next_snap_obj when it should, leading to a
2856 * missing entry. Therefore we can only use the
2857 * next_clones_obj when its count is correct.
2858 */
2859 int err = zap_count(dp->dp_meta_objset,
2860 dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
2861 if (err == 0 &&
2862 count == dsl_dataset_phys(ds)->ds_num_children - 1)
2863 usenext = B_TRUE;
2864 }
2865
2866 if (usenext) {
2867 zap_cursor_t zc;
2868 zap_attribute_t za;
2869 for (zap_cursor_init(&zc, dp->dp_meta_objset,
2870 dsl_dataset_phys(ds)->ds_next_clones_obj);
2871 zap_cursor_retrieve(&zc, &za) == 0;
2872 (void) zap_cursor_advance(&zc)) {
2873 scan_ds_queue_insert(scn,
2874 zfs_strtonum(za.za_name, NULL),
2875 dsl_dataset_phys(ds)->ds_creation_txg);
2876 }
2877 zap_cursor_fini(&zc);
2878 } else {
2879 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2880 enqueue_clones_cb, &ds->ds_object,
2881 DS_FIND_CHILDREN));
2882 }
2883 }
2884
2885 out:
2886 dsl_dataset_rele(ds, FTAG);
2887 }
2888
2889 static int
enqueue_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)2890 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2891 {
2892 (void) arg;
2893 dsl_dataset_t *ds;
2894 int err;
2895 dsl_scan_t *scn = dp->dp_scan;
2896
2897 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2898 if (err)
2899 return (err);
2900
2901 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
2902 dsl_dataset_t *prev;
2903 err = dsl_dataset_hold_obj(dp,
2904 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2905 if (err) {
2906 dsl_dataset_rele(ds, FTAG);
2907 return (err);
2908 }
2909
2910 /*
2911 * If this is a clone, we don't need to worry about it for now.
2912 */
2913 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
2914 dsl_dataset_rele(ds, FTAG);
2915 dsl_dataset_rele(prev, FTAG);
2916 return (0);
2917 }
2918 dsl_dataset_rele(ds, FTAG);
2919 ds = prev;
2920 }
2921
2922 mutex_enter(&scn->scn_queue_lock);
2923 scan_ds_queue_insert(scn, ds->ds_object,
2924 dsl_dataset_phys(ds)->ds_prev_snap_txg);
2925 mutex_exit(&scn->scn_queue_lock);
2926 dsl_dataset_rele(ds, FTAG);
2927 return (0);
2928 }
2929
2930 void
dsl_scan_ddt_entry(dsl_scan_t * scn,enum zio_checksum checksum,ddt_entry_t * dde,dmu_tx_t * tx)2931 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
2932 ddt_entry_t *dde, dmu_tx_t *tx)
2933 {
2934 (void) tx;
2935 const ddt_key_t *ddk = &dde->dde_key;
2936 ddt_phys_t *ddp = dde->dde_phys;
2937 blkptr_t bp;
2938 zbookmark_phys_t zb = { 0 };
2939
2940 if (!dsl_scan_is_running(scn))
2941 return;
2942
2943 /*
2944 * This function is special because it is the only thing
2945 * that can add scan_io_t's to the vdev scan queues from
2946 * outside dsl_scan_sync(). For the most part this is ok
2947 * as long as it is called from within syncing context.
2948 * However, dsl_scan_sync() expects that no new sio's will
2949 * be added between when all the work for a scan is done
2950 * and the next txg when the scan is actually marked as
2951 * completed. This check ensures we do not issue new sio's
2952 * during this period.
2953 */
2954 if (scn->scn_done_txg != 0)
2955 return;
2956
2957 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2958 if (ddp->ddp_phys_birth == 0 ||
2959 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
2960 continue;
2961 ddt_bp_create(checksum, ddk, ddp, &bp);
2962
2963 scn->scn_visited_this_txg++;
2964 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
2965 }
2966 }
2967
2968 /*
2969 * Scrub/dedup interaction.
2970 *
2971 * If there are N references to a deduped block, we don't want to scrub it
2972 * N times -- ideally, we should scrub it exactly once.
2973 *
2974 * We leverage the fact that the dde's replication class (ddt_class_t)
2975 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2976 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2977 *
2978 * To prevent excess scrubbing, the scrub begins by walking the DDT
2979 * to find all blocks with refcnt > 1, and scrubs each of these once.
2980 * Since there are two replication classes which contain blocks with
2981 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2982 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2983 *
2984 * There would be nothing more to say if a block's refcnt couldn't change
2985 * during a scrub, but of course it can so we must account for changes
2986 * in a block's replication class.
2987 *
2988 * Here's an example of what can occur:
2989 *
2990 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2991 * when visited during the top-down scrub phase, it will be scrubbed twice.
2992 * This negates our scrub optimization, but is otherwise harmless.
2993 *
2994 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2995 * on each visit during the top-down scrub phase, it will never be scrubbed.
2996 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2997 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2998 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2999 * while a scrub is in progress, it scrubs the block right then.
3000 */
3001 static void
dsl_scan_ddt(dsl_scan_t * scn,dmu_tx_t * tx)3002 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
3003 {
3004 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
3005 ddt_entry_t dde = {{{{0}}}};
3006 int error;
3007 uint64_t n = 0;
3008
3009 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
3010 ddt_t *ddt;
3011
3012 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
3013 break;
3014 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
3015 (longlong_t)ddb->ddb_class,
3016 (longlong_t)ddb->ddb_type,
3017 (longlong_t)ddb->ddb_checksum,
3018 (longlong_t)ddb->ddb_cursor);
3019
3020 /* There should be no pending changes to the dedup table */
3021 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
3022 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
3023
3024 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
3025 n++;
3026
3027 if (dsl_scan_check_suspend(scn, NULL))
3028 break;
3029 }
3030
3031 zfs_dbgmsg("scanned %llu ddt entries on %s with class_max = %u; "
3032 "suspending=%u", (longlong_t)n, scn->scn_dp->dp_spa->spa_name,
3033 (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
3034
3035 ASSERT(error == 0 || error == ENOENT);
3036 ASSERT(error != ENOENT ||
3037 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
3038 }
3039
3040 static uint64_t
dsl_scan_ds_maxtxg(dsl_dataset_t * ds)3041 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
3042 {
3043 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
3044 if (ds->ds_is_snapshot)
3045 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
3046 return (smt);
3047 }
3048
3049 static void
dsl_scan_visit(dsl_scan_t * scn,dmu_tx_t * tx)3050 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
3051 {
3052 scan_ds_t *sds;
3053 dsl_pool_t *dp = scn->scn_dp;
3054
3055 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
3056 scn->scn_phys.scn_ddt_class_max) {
3057 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
3058 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
3059 dsl_scan_ddt(scn, tx);
3060 if (scn->scn_suspending)
3061 return;
3062 }
3063
3064 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
3065 /* First do the MOS & ORIGIN */
3066
3067 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
3068 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
3069 dsl_scan_visit_rootbp(scn, NULL,
3070 &dp->dp_meta_rootbp, tx);
3071 if (scn->scn_suspending)
3072 return;
3073
3074 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
3075 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
3076 enqueue_cb, NULL, DS_FIND_CHILDREN));
3077 } else {
3078 dsl_scan_visitds(scn,
3079 dp->dp_origin_snap->ds_object, tx);
3080 }
3081 ASSERT(!scn->scn_suspending);
3082 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
3083 ZB_DESTROYED_OBJSET) {
3084 uint64_t dsobj = scn->scn_phys.scn_bookmark.zb_objset;
3085 /*
3086 * If we were suspended, continue from here. Note if the
3087 * ds we were suspended on was deleted, the zb_objset may
3088 * be -1, so we will skip this and find a new objset
3089 * below.
3090 */
3091 dsl_scan_visitds(scn, dsobj, tx);
3092 if (scn->scn_suspending)
3093 return;
3094 }
3095
3096 /*
3097 * In case we suspended right at the end of the ds, zero the
3098 * bookmark so we don't think that we're still trying to resume.
3099 */
3100 memset(&scn->scn_phys.scn_bookmark, 0, sizeof (zbookmark_phys_t));
3101
3102 /*
3103 * Keep pulling things out of the dataset avl queue. Updates to the
3104 * persistent zap-object-as-queue happen only at checkpoints.
3105 */
3106 while ((sds = avl_first(&scn->scn_queue)) != NULL) {
3107 dsl_dataset_t *ds;
3108 uint64_t dsobj = sds->sds_dsobj;
3109 uint64_t txg = sds->sds_txg;
3110
3111 /* dequeue and free the ds from the queue */
3112 scan_ds_queue_remove(scn, dsobj);
3113 sds = NULL;
3114
3115 /* set up min / max txg */
3116 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
3117 if (txg != 0) {
3118 scn->scn_phys.scn_cur_min_txg =
3119 MAX(scn->scn_phys.scn_min_txg, txg);
3120 } else {
3121 scn->scn_phys.scn_cur_min_txg =
3122 MAX(scn->scn_phys.scn_min_txg,
3123 dsl_dataset_phys(ds)->ds_prev_snap_txg);
3124 }
3125 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
3126 dsl_dataset_rele(ds, FTAG);
3127
3128 dsl_scan_visitds(scn, dsobj, tx);
3129 if (scn->scn_suspending)
3130 return;
3131 }
3132
3133 /* No more objsets to fetch, we're done */
3134 scn->scn_phys.scn_bookmark.zb_objset = ZB_DESTROYED_OBJSET;
3135 ASSERT0(scn->scn_suspending);
3136 }
3137
3138 static uint64_t
dsl_scan_count_data_disks(spa_t * spa)3139 dsl_scan_count_data_disks(spa_t *spa)
3140 {
3141 vdev_t *rvd = spa->spa_root_vdev;
3142 uint64_t i, leaves = 0;
3143
3144 for (i = 0; i < rvd->vdev_children; i++) {
3145 vdev_t *vd = rvd->vdev_child[i];
3146 if (vd->vdev_islog || vd->vdev_isspare || vd->vdev_isl2cache)
3147 continue;
3148 leaves += vdev_get_ndisks(vd) - vdev_get_nparity(vd);
3149 }
3150 return (leaves);
3151 }
3152
3153 static void
scan_io_queues_update_zio_stats(dsl_scan_io_queue_t * q,const blkptr_t * bp)3154 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t *q, const blkptr_t *bp)
3155 {
3156 int i;
3157 uint64_t cur_size = 0;
3158
3159 for (i = 0; i < BP_GET_NDVAS(bp); i++) {
3160 cur_size += DVA_GET_ASIZE(&bp->blk_dva[i]);
3161 }
3162
3163 q->q_total_zio_size_this_txg += cur_size;
3164 q->q_zios_this_txg++;
3165 }
3166
3167 static void
scan_io_queues_update_seg_stats(dsl_scan_io_queue_t * q,uint64_t start,uint64_t end)3168 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t *q, uint64_t start,
3169 uint64_t end)
3170 {
3171 q->q_total_seg_size_this_txg += end - start;
3172 q->q_segs_this_txg++;
3173 }
3174
3175 static boolean_t
scan_io_queue_check_suspend(dsl_scan_t * scn)3176 scan_io_queue_check_suspend(dsl_scan_t *scn)
3177 {
3178 /* See comment in dsl_scan_check_suspend() */
3179 uint64_t curr_time_ns = gethrtime();
3180 uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
3181 uint64_t sync_time_ns = curr_time_ns -
3182 scn->scn_dp->dp_spa->spa_sync_starttime;
3183 uint64_t dirty_min_bytes = zfs_dirty_data_max *
3184 zfs_vdev_async_write_active_min_dirty_percent / 100;
3185 uint_t mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
3186 zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
3187
3188 return ((NSEC2MSEC(scan_time_ns) > mintime &&
3189 (scn->scn_dp->dp_dirty_total >= dirty_min_bytes ||
3190 txg_sync_waiting(scn->scn_dp) ||
3191 NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
3192 spa_shutting_down(scn->scn_dp->dp_spa));
3193 }
3194
3195 /*
3196 * Given a list of scan_io_t's in io_list, this issues the I/Os out to
3197 * disk. This consumes the io_list and frees the scan_io_t's. This is
3198 * called when emptying queues, either when we're up against the memory
3199 * limit or when we have finished scanning. Returns B_TRUE if we stopped
3200 * processing the list before we finished. Any sios that were not issued
3201 * will remain in the io_list.
3202 */
3203 static boolean_t
scan_io_queue_issue(dsl_scan_io_queue_t * queue,list_t * io_list)3204 scan_io_queue_issue(dsl_scan_io_queue_t *queue, list_t *io_list)
3205 {
3206 dsl_scan_t *scn = queue->q_scn;
3207 scan_io_t *sio;
3208 boolean_t suspended = B_FALSE;
3209
3210 while ((sio = list_head(io_list)) != NULL) {
3211 blkptr_t bp;
3212
3213 if (scan_io_queue_check_suspend(scn)) {
3214 suspended = B_TRUE;
3215 break;
3216 }
3217
3218 sio2bp(sio, &bp);
3219 scan_exec_io(scn->scn_dp, &bp, sio->sio_flags,
3220 &sio->sio_zb, queue);
3221 (void) list_remove_head(io_list);
3222 scan_io_queues_update_zio_stats(queue, &bp);
3223 sio_free(sio);
3224 }
3225 return (suspended);
3226 }
3227
3228 /*
3229 * This function removes sios from an IO queue which reside within a given
3230 * range_seg_t and inserts them (in offset order) into a list. Note that
3231 * we only ever return a maximum of 32 sios at once. If there are more sios
3232 * to process within this segment that did not make it onto the list we
3233 * return B_TRUE and otherwise B_FALSE.
3234 */
3235 static boolean_t
scan_io_queue_gather(dsl_scan_io_queue_t * queue,range_seg_t * rs,list_t * list)3236 scan_io_queue_gather(dsl_scan_io_queue_t *queue, range_seg_t *rs, list_t *list)
3237 {
3238 scan_io_t *srch_sio, *sio, *next_sio;
3239 avl_index_t idx;
3240 uint_t num_sios = 0;
3241 int64_t bytes_issued = 0;
3242
3243 ASSERT(rs != NULL);
3244 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3245
3246 srch_sio = sio_alloc(1);
3247 srch_sio->sio_nr_dvas = 1;
3248 SIO_SET_OFFSET(srch_sio, rs_get_start(rs, queue->q_exts_by_addr));
3249
3250 /*
3251 * The exact start of the extent might not contain any matching zios,
3252 * so if that's the case, examine the next one in the tree.
3253 */
3254 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
3255 sio_free(srch_sio);
3256
3257 if (sio == NULL)
3258 sio = avl_nearest(&queue->q_sios_by_addr, idx, AVL_AFTER);
3259
3260 while (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
3261 queue->q_exts_by_addr) && num_sios <= 32) {
3262 ASSERT3U(SIO_GET_OFFSET(sio), >=, rs_get_start(rs,
3263 queue->q_exts_by_addr));
3264 ASSERT3U(SIO_GET_END_OFFSET(sio), <=, rs_get_end(rs,
3265 queue->q_exts_by_addr));
3266
3267 next_sio = AVL_NEXT(&queue->q_sios_by_addr, sio);
3268 avl_remove(&queue->q_sios_by_addr, sio);
3269 if (avl_is_empty(&queue->q_sios_by_addr))
3270 atomic_add_64(&queue->q_scn->scn_queues_pending, -1);
3271 queue->q_sio_memused -= SIO_GET_MUSED(sio);
3272
3273 bytes_issued += SIO_GET_ASIZE(sio);
3274 num_sios++;
3275 list_insert_tail(list, sio);
3276 sio = next_sio;
3277 }
3278
3279 /*
3280 * We limit the number of sios we process at once to 32 to avoid
3281 * biting off more than we can chew. If we didn't take everything
3282 * in the segment we update it to reflect the work we were able to
3283 * complete. Otherwise, we remove it from the range tree entirely.
3284 */
3285 if (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
3286 queue->q_exts_by_addr)) {
3287 range_tree_adjust_fill(queue->q_exts_by_addr, rs,
3288 -bytes_issued);
3289 range_tree_resize_segment(queue->q_exts_by_addr, rs,
3290 SIO_GET_OFFSET(sio), rs_get_end(rs,
3291 queue->q_exts_by_addr) - SIO_GET_OFFSET(sio));
3292 queue->q_last_ext_addr = SIO_GET_OFFSET(sio);
3293 return (B_TRUE);
3294 } else {
3295 uint64_t rstart = rs_get_start(rs, queue->q_exts_by_addr);
3296 uint64_t rend = rs_get_end(rs, queue->q_exts_by_addr);
3297 range_tree_remove(queue->q_exts_by_addr, rstart, rend - rstart);
3298 queue->q_last_ext_addr = -1;
3299 return (B_FALSE);
3300 }
3301 }
3302
3303 /*
3304 * This is called from the queue emptying thread and selects the next
3305 * extent from which we are to issue I/Os. The behavior of this function
3306 * depends on the state of the scan, the current memory consumption and
3307 * whether or not we are performing a scan shutdown.
3308 * 1) We select extents in an elevator algorithm (LBA-order) if the scan
3309 * needs to perform a checkpoint
3310 * 2) We select the largest available extent if we are up against the
3311 * memory limit.
3312 * 3) Otherwise we don't select any extents.
3313 */
3314 static range_seg_t *
scan_io_queue_fetch_ext(dsl_scan_io_queue_t * queue)3315 scan_io_queue_fetch_ext(dsl_scan_io_queue_t *queue)
3316 {
3317 dsl_scan_t *scn = queue->q_scn;
3318 range_tree_t *rt = queue->q_exts_by_addr;
3319
3320 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3321 ASSERT(scn->scn_is_sorted);
3322
3323 if (!scn->scn_checkpointing && !scn->scn_clearing)
3324 return (NULL);
3325
3326 /*
3327 * During normal clearing, we want to issue our largest segments
3328 * first, keeping IO as sequential as possible, and leaving the
3329 * smaller extents for later with the hope that they might eventually
3330 * grow to larger sequential segments. However, when the scan is
3331 * checkpointing, no new extents will be added to the sorting queue,
3332 * so the way we are sorted now is as good as it will ever get.
3333 * In this case, we instead switch to issuing extents in LBA order.
3334 */
3335 if ((zfs_scan_issue_strategy < 1 && scn->scn_checkpointing) ||
3336 zfs_scan_issue_strategy == 1)
3337 return (range_tree_first(rt));
3338
3339 /*
3340 * Try to continue previous extent if it is not completed yet. After
3341 * shrink in scan_io_queue_gather() it may no longer be the best, but
3342 * otherwise we leave shorter remnant every txg.
3343 */
3344 uint64_t start;
3345 uint64_t size = 1ULL << rt->rt_shift;
3346 range_seg_t *addr_rs;
3347 if (queue->q_last_ext_addr != -1) {
3348 start = queue->q_last_ext_addr;
3349 addr_rs = range_tree_find(rt, start, size);
3350 if (addr_rs != NULL)
3351 return (addr_rs);
3352 }
3353
3354 /*
3355 * Nothing to continue, so find new best extent.
3356 */
3357 uint64_t *v = zfs_btree_first(&queue->q_exts_by_size, NULL);
3358 if (v == NULL)
3359 return (NULL);
3360 queue->q_last_ext_addr = start = *v << rt->rt_shift;
3361
3362 /*
3363 * We need to get the original entry in the by_addr tree so we can
3364 * modify it.
3365 */
3366 addr_rs = range_tree_find(rt, start, size);
3367 ASSERT3P(addr_rs, !=, NULL);
3368 ASSERT3U(rs_get_start(addr_rs, rt), ==, start);
3369 ASSERT3U(rs_get_end(addr_rs, rt), >, start);
3370 return (addr_rs);
3371 }
3372
3373 static void
scan_io_queues_run_one(void * arg)3374 scan_io_queues_run_one(void *arg)
3375 {
3376 dsl_scan_io_queue_t *queue = arg;
3377 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
3378 boolean_t suspended = B_FALSE;
3379 range_seg_t *rs;
3380 scan_io_t *sio;
3381 zio_t *zio;
3382 list_t sio_list;
3383
3384 ASSERT(queue->q_scn->scn_is_sorted);
3385
3386 list_create(&sio_list, sizeof (scan_io_t),
3387 offsetof(scan_io_t, sio_nodes.sio_list_node));
3388 zio = zio_null(queue->q_scn->scn_zio_root, queue->q_scn->scn_dp->dp_spa,
3389 NULL, NULL, NULL, ZIO_FLAG_CANFAIL);
3390 mutex_enter(q_lock);
3391 queue->q_zio = zio;
3392
3393 /* Calculate maximum in-flight bytes for this vdev. */
3394 queue->q_maxinflight_bytes = MAX(1, zfs_scan_vdev_limit *
3395 (vdev_get_ndisks(queue->q_vd) - vdev_get_nparity(queue->q_vd)));
3396
3397 /* reset per-queue scan statistics for this txg */
3398 queue->q_total_seg_size_this_txg = 0;
3399 queue->q_segs_this_txg = 0;
3400 queue->q_total_zio_size_this_txg = 0;
3401 queue->q_zios_this_txg = 0;
3402
3403 /* loop until we run out of time or sios */
3404 while ((rs = scan_io_queue_fetch_ext(queue)) != NULL) {
3405 uint64_t seg_start = 0, seg_end = 0;
3406 boolean_t more_left;
3407
3408 ASSERT(list_is_empty(&sio_list));
3409
3410 /* loop while we still have sios left to process in this rs */
3411 do {
3412 scan_io_t *first_sio, *last_sio;
3413
3414 /*
3415 * We have selected which extent needs to be
3416 * processed next. Gather up the corresponding sios.
3417 */
3418 more_left = scan_io_queue_gather(queue, rs, &sio_list);
3419 ASSERT(!list_is_empty(&sio_list));
3420 first_sio = list_head(&sio_list);
3421 last_sio = list_tail(&sio_list);
3422
3423 seg_end = SIO_GET_END_OFFSET(last_sio);
3424 if (seg_start == 0)
3425 seg_start = SIO_GET_OFFSET(first_sio);
3426
3427 /*
3428 * Issuing sios can take a long time so drop the
3429 * queue lock. The sio queue won't be updated by
3430 * other threads since we're in syncing context so
3431 * we can be sure that our trees will remain exactly
3432 * as we left them.
3433 */
3434 mutex_exit(q_lock);
3435 suspended = scan_io_queue_issue(queue, &sio_list);
3436 mutex_enter(q_lock);
3437
3438 if (suspended)
3439 break;
3440 } while (more_left);
3441
3442 /* update statistics for debugging purposes */
3443 scan_io_queues_update_seg_stats(queue, seg_start, seg_end);
3444
3445 if (suspended)
3446 break;
3447 }
3448
3449 /*
3450 * If we were suspended in the middle of processing,
3451 * requeue any unfinished sios and exit.
3452 */
3453 while ((sio = list_remove_head(&sio_list)) != NULL)
3454 scan_io_queue_insert_impl(queue, sio);
3455
3456 queue->q_zio = NULL;
3457 mutex_exit(q_lock);
3458 zio_nowait(zio);
3459 list_destroy(&sio_list);
3460 }
3461
3462 /*
3463 * Performs an emptying run on all scan queues in the pool. This just
3464 * punches out one thread per top-level vdev, each of which processes
3465 * only that vdev's scan queue. We can parallelize the I/O here because
3466 * we know that each queue's I/Os only affect its own top-level vdev.
3467 *
3468 * This function waits for the queue runs to complete, and must be
3469 * called from dsl_scan_sync (or in general, syncing context).
3470 */
3471 static void
scan_io_queues_run(dsl_scan_t * scn)3472 scan_io_queues_run(dsl_scan_t *scn)
3473 {
3474 spa_t *spa = scn->scn_dp->dp_spa;
3475
3476 ASSERT(scn->scn_is_sorted);
3477 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
3478
3479 if (scn->scn_queues_pending == 0)
3480 return;
3481
3482 if (scn->scn_taskq == NULL) {
3483 int nthreads = spa->spa_root_vdev->vdev_children;
3484
3485 /*
3486 * We need to make this taskq *always* execute as many
3487 * threads in parallel as we have top-level vdevs and no
3488 * less, otherwise strange serialization of the calls to
3489 * scan_io_queues_run_one can occur during spa_sync runs
3490 * and that significantly impacts performance.
3491 */
3492 scn->scn_taskq = taskq_create("dsl_scan_iss", nthreads,
3493 minclsyspri, nthreads, nthreads, TASKQ_PREPOPULATE);
3494 }
3495
3496 for (uint64_t i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3497 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3498
3499 mutex_enter(&vd->vdev_scan_io_queue_lock);
3500 if (vd->vdev_scan_io_queue != NULL) {
3501 VERIFY(taskq_dispatch(scn->scn_taskq,
3502 scan_io_queues_run_one, vd->vdev_scan_io_queue,
3503 TQ_SLEEP) != TASKQID_INVALID);
3504 }
3505 mutex_exit(&vd->vdev_scan_io_queue_lock);
3506 }
3507
3508 /*
3509 * Wait for the queues to finish issuing their IOs for this run
3510 * before we return. There may still be IOs in flight at this
3511 * point.
3512 */
3513 taskq_wait(scn->scn_taskq);
3514 }
3515
3516 static boolean_t
dsl_scan_async_block_should_pause(dsl_scan_t * scn)3517 dsl_scan_async_block_should_pause(dsl_scan_t *scn)
3518 {
3519 uint64_t elapsed_nanosecs;
3520
3521 if (zfs_recover)
3522 return (B_FALSE);
3523
3524 if (zfs_async_block_max_blocks != 0 &&
3525 scn->scn_visited_this_txg >= zfs_async_block_max_blocks) {
3526 return (B_TRUE);
3527 }
3528
3529 if (zfs_max_async_dedup_frees != 0 &&
3530 scn->scn_dedup_frees_this_txg >= zfs_max_async_dedup_frees) {
3531 return (B_TRUE);
3532 }
3533
3534 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
3535 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
3536 (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms &&
3537 txg_sync_waiting(scn->scn_dp)) ||
3538 spa_shutting_down(scn->scn_dp->dp_spa));
3539 }
3540
3541 static int
dsl_scan_free_block_cb(void * arg,const blkptr_t * bp,dmu_tx_t * tx)3542 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
3543 {
3544 dsl_scan_t *scn = arg;
3545
3546 if (!scn->scn_is_bptree ||
3547 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
3548 if (dsl_scan_async_block_should_pause(scn))
3549 return (SET_ERROR(ERESTART));
3550 }
3551
3552 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
3553 dmu_tx_get_txg(tx), bp, 0));
3554 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
3555 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
3556 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
3557 scn->scn_visited_this_txg++;
3558 if (BP_GET_DEDUP(bp))
3559 scn->scn_dedup_frees_this_txg++;
3560 return (0);
3561 }
3562
3563 static void
dsl_scan_update_stats(dsl_scan_t * scn)3564 dsl_scan_update_stats(dsl_scan_t *scn)
3565 {
3566 spa_t *spa = scn->scn_dp->dp_spa;
3567 uint64_t i;
3568 uint64_t seg_size_total = 0, zio_size_total = 0;
3569 uint64_t seg_count_total = 0, zio_count_total = 0;
3570
3571 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3572 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3573 dsl_scan_io_queue_t *queue = vd->vdev_scan_io_queue;
3574
3575 if (queue == NULL)
3576 continue;
3577
3578 seg_size_total += queue->q_total_seg_size_this_txg;
3579 zio_size_total += queue->q_total_zio_size_this_txg;
3580 seg_count_total += queue->q_segs_this_txg;
3581 zio_count_total += queue->q_zios_this_txg;
3582 }
3583
3584 if (seg_count_total == 0 || zio_count_total == 0) {
3585 scn->scn_avg_seg_size_this_txg = 0;
3586 scn->scn_avg_zio_size_this_txg = 0;
3587 scn->scn_segs_this_txg = 0;
3588 scn->scn_zios_this_txg = 0;
3589 return;
3590 }
3591
3592 scn->scn_avg_seg_size_this_txg = seg_size_total / seg_count_total;
3593 scn->scn_avg_zio_size_this_txg = zio_size_total / zio_count_total;
3594 scn->scn_segs_this_txg = seg_count_total;
3595 scn->scn_zios_this_txg = zio_count_total;
3596 }
3597
3598 static int
bpobj_dsl_scan_free_block_cb(void * arg,const blkptr_t * bp,boolean_t bp_freed,dmu_tx_t * tx)3599 bpobj_dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3600 dmu_tx_t *tx)
3601 {
3602 ASSERT(!bp_freed);
3603 return (dsl_scan_free_block_cb(arg, bp, tx));
3604 }
3605
3606 static int
dsl_scan_obsolete_block_cb(void * arg,const blkptr_t * bp,boolean_t bp_freed,dmu_tx_t * tx)3607 dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3608 dmu_tx_t *tx)
3609 {
3610 ASSERT(!bp_freed);
3611 dsl_scan_t *scn = arg;
3612 const dva_t *dva = &bp->blk_dva[0];
3613
3614 if (dsl_scan_async_block_should_pause(scn))
3615 return (SET_ERROR(ERESTART));
3616
3617 spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa,
3618 DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva),
3619 DVA_GET_ASIZE(dva), tx);
3620 scn->scn_visited_this_txg++;
3621 return (0);
3622 }
3623
3624 boolean_t
dsl_scan_active(dsl_scan_t * scn)3625 dsl_scan_active(dsl_scan_t *scn)
3626 {
3627 spa_t *spa = scn->scn_dp->dp_spa;
3628 uint64_t used = 0, comp, uncomp;
3629 boolean_t clones_left;
3630
3631 if (spa->spa_load_state != SPA_LOAD_NONE)
3632 return (B_FALSE);
3633 if (spa_shutting_down(spa))
3634 return (B_FALSE);
3635 if ((dsl_scan_is_running(scn) && !dsl_scan_is_paused_scrub(scn)) ||
3636 (scn->scn_async_destroying && !scn->scn_async_stalled))
3637 return (B_TRUE);
3638
3639 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
3640 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
3641 &used, &comp, &uncomp);
3642 }
3643 clones_left = spa_livelist_delete_check(spa);
3644 return ((used != 0) || (clones_left));
3645 }
3646
3647 boolean_t
dsl_errorscrub_active(dsl_scan_t * scn)3648 dsl_errorscrub_active(dsl_scan_t *scn)
3649 {
3650 spa_t *spa = scn->scn_dp->dp_spa;
3651 if (spa->spa_load_state != SPA_LOAD_NONE)
3652 return (B_FALSE);
3653 if (spa_shutting_down(spa))
3654 return (B_FALSE);
3655 if (dsl_errorscrubbing(scn->scn_dp))
3656 return (B_TRUE);
3657 return (B_FALSE);
3658 }
3659
3660 static boolean_t
dsl_scan_check_deferred(vdev_t * vd)3661 dsl_scan_check_deferred(vdev_t *vd)
3662 {
3663 boolean_t need_resilver = B_FALSE;
3664
3665 for (int c = 0; c < vd->vdev_children; c++) {
3666 need_resilver |=
3667 dsl_scan_check_deferred(vd->vdev_child[c]);
3668 }
3669
3670 if (!vdev_is_concrete(vd) || vd->vdev_aux ||
3671 !vd->vdev_ops->vdev_op_leaf)
3672 return (need_resilver);
3673
3674 if (!vd->vdev_resilver_deferred)
3675 need_resilver = B_TRUE;
3676
3677 return (need_resilver);
3678 }
3679
3680 static boolean_t
dsl_scan_need_resilver(spa_t * spa,const dva_t * dva,size_t psize,uint64_t phys_birth)3681 dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize,
3682 uint64_t phys_birth)
3683 {
3684 vdev_t *vd;
3685
3686 vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
3687
3688 if (vd->vdev_ops == &vdev_indirect_ops) {
3689 /*
3690 * The indirect vdev can point to multiple
3691 * vdevs. For simplicity, always create
3692 * the resilver zio_t. zio_vdev_io_start()
3693 * will bypass the child resilver i/o's if
3694 * they are on vdevs that don't have DTL's.
3695 */
3696 return (B_TRUE);
3697 }
3698
3699 if (DVA_GET_GANG(dva)) {
3700 /*
3701 * Gang members may be spread across multiple
3702 * vdevs, so the best estimate we have is the
3703 * scrub range, which has already been checked.
3704 * XXX -- it would be better to change our
3705 * allocation policy to ensure that all
3706 * gang members reside on the same vdev.
3707 */
3708 return (B_TRUE);
3709 }
3710
3711 /*
3712 * Check if the top-level vdev must resilver this offset.
3713 * When the offset does not intersect with a dirty leaf DTL
3714 * then it may be possible to skip the resilver IO. The psize
3715 * is provided instead of asize to simplify the check for RAIDZ.
3716 */
3717 if (!vdev_dtl_need_resilver(vd, dva, psize, phys_birth))
3718 return (B_FALSE);
3719
3720 /*
3721 * Check that this top-level vdev has a device under it which
3722 * is resilvering and is not deferred.
3723 */
3724 if (!dsl_scan_check_deferred(vd))
3725 return (B_FALSE);
3726
3727 return (B_TRUE);
3728 }
3729
3730 static int
dsl_process_async_destroys(dsl_pool_t * dp,dmu_tx_t * tx)3731 dsl_process_async_destroys(dsl_pool_t *dp, dmu_tx_t *tx)
3732 {
3733 dsl_scan_t *scn = dp->dp_scan;
3734 spa_t *spa = dp->dp_spa;
3735 int err = 0;
3736
3737 if (spa_suspend_async_destroy(spa))
3738 return (0);
3739
3740 if (zfs_free_bpobj_enabled &&
3741 spa_version(spa) >= SPA_VERSION_DEADLISTS) {
3742 scn->scn_is_bptree = B_FALSE;
3743 scn->scn_async_block_min_time_ms = zfs_free_min_time_ms;
3744 scn->scn_zio_root = zio_root(spa, NULL,
3745 NULL, ZIO_FLAG_MUSTSUCCEED);
3746 err = bpobj_iterate(&dp->dp_free_bpobj,
3747 bpobj_dsl_scan_free_block_cb, scn, tx);
3748 VERIFY0(zio_wait(scn->scn_zio_root));
3749 scn->scn_zio_root = NULL;
3750
3751 if (err != 0 && err != ERESTART)
3752 zfs_panic_recover("error %u from bpobj_iterate()", err);
3753 }
3754
3755 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
3756 ASSERT(scn->scn_async_destroying);
3757 scn->scn_is_bptree = B_TRUE;
3758 scn->scn_zio_root = zio_root(spa, NULL,
3759 NULL, ZIO_FLAG_MUSTSUCCEED);
3760 err = bptree_iterate(dp->dp_meta_objset,
3761 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
3762 VERIFY0(zio_wait(scn->scn_zio_root));
3763 scn->scn_zio_root = NULL;
3764
3765 if (err == EIO || err == ECKSUM) {
3766 err = 0;
3767 } else if (err != 0 && err != ERESTART) {
3768 zfs_panic_recover("error %u from "
3769 "traverse_dataset_destroyed()", err);
3770 }
3771
3772 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
3773 /* finished; deactivate async destroy feature */
3774 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
3775 ASSERT(!spa_feature_is_active(spa,
3776 SPA_FEATURE_ASYNC_DESTROY));
3777 VERIFY0(zap_remove(dp->dp_meta_objset,
3778 DMU_POOL_DIRECTORY_OBJECT,
3779 DMU_POOL_BPTREE_OBJ, tx));
3780 VERIFY0(bptree_free(dp->dp_meta_objset,
3781 dp->dp_bptree_obj, tx));
3782 dp->dp_bptree_obj = 0;
3783 scn->scn_async_destroying = B_FALSE;
3784 scn->scn_async_stalled = B_FALSE;
3785 } else {
3786 /*
3787 * If we didn't make progress, mark the async
3788 * destroy as stalled, so that we will not initiate
3789 * a spa_sync() on its behalf. Note that we only
3790 * check this if we are not finished, because if the
3791 * bptree had no blocks for us to visit, we can
3792 * finish without "making progress".
3793 */
3794 scn->scn_async_stalled =
3795 (scn->scn_visited_this_txg == 0);
3796 }
3797 }
3798 if (scn->scn_visited_this_txg) {
3799 zfs_dbgmsg("freed %llu blocks in %llums from "
3800 "free_bpobj/bptree on %s in txg %llu; err=%u",
3801 (longlong_t)scn->scn_visited_this_txg,
3802 (longlong_t)
3803 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
3804 spa->spa_name, (longlong_t)tx->tx_txg, err);
3805 scn->scn_visited_this_txg = 0;
3806 scn->scn_dedup_frees_this_txg = 0;
3807
3808 /*
3809 * Write out changes to the DDT and the BRT that may be required
3810 * as a result of the blocks freed. This ensures that the DDT
3811 * and the BRT are clean when a scrub/resilver runs.
3812 */
3813 ddt_sync(spa, tx->tx_txg);
3814 brt_sync(spa, tx->tx_txg);
3815 }
3816 if (err != 0)
3817 return (err);
3818 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3819 zfs_free_leak_on_eio &&
3820 (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
3821 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
3822 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
3823 /*
3824 * We have finished background destroying, but there is still
3825 * some space left in the dp_free_dir. Transfer this leaked
3826 * space to the dp_leak_dir.
3827 */
3828 if (dp->dp_leak_dir == NULL) {
3829 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
3830 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
3831 LEAK_DIR_NAME, tx);
3832 VERIFY0(dsl_pool_open_special_dir(dp,
3833 LEAK_DIR_NAME, &dp->dp_leak_dir));
3834 rrw_exit(&dp->dp_config_rwlock, FTAG);
3835 }
3836 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
3837 dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3838 dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3839 dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3840 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
3841 -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3842 -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3843 -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3844 }
3845
3846 if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3847 !spa_livelist_delete_check(spa)) {
3848 /* finished; verify that space accounting went to zero */
3849 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
3850 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
3851 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
3852 }
3853
3854 spa_notify_waiters(spa);
3855
3856 EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj),
3857 0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3858 DMU_POOL_OBSOLETE_BPOBJ));
3859 if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) {
3860 ASSERT(spa_feature_is_active(dp->dp_spa,
3861 SPA_FEATURE_OBSOLETE_COUNTS));
3862
3863 scn->scn_is_bptree = B_FALSE;
3864 scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms;
3865 err = bpobj_iterate(&dp->dp_obsolete_bpobj,
3866 dsl_scan_obsolete_block_cb, scn, tx);
3867 if (err != 0 && err != ERESTART)
3868 zfs_panic_recover("error %u from bpobj_iterate()", err);
3869
3870 if (bpobj_is_empty(&dp->dp_obsolete_bpobj))
3871 dsl_pool_destroy_obsolete_bpobj(dp, tx);
3872 }
3873 return (0);
3874 }
3875
3876 static void
name_to_bookmark(char * buf,zbookmark_phys_t * zb)3877 name_to_bookmark(char *buf, zbookmark_phys_t *zb)
3878 {
3879 zb->zb_objset = zfs_strtonum(buf, &buf);
3880 ASSERT(*buf == ':');
3881 zb->zb_object = zfs_strtonum(buf + 1, &buf);
3882 ASSERT(*buf == ':');
3883 zb->zb_level = (int)zfs_strtonum(buf + 1, &buf);
3884 ASSERT(*buf == ':');
3885 zb->zb_blkid = zfs_strtonum(buf + 1, &buf);
3886 ASSERT(*buf == '\0');
3887 }
3888
3889 static void
name_to_object(char * buf,uint64_t * obj)3890 name_to_object(char *buf, uint64_t *obj)
3891 {
3892 *obj = zfs_strtonum(buf, &buf);
3893 ASSERT(*buf == '\0');
3894 }
3895
3896 static void
read_by_block_level(dsl_scan_t * scn,zbookmark_phys_t zb)3897 read_by_block_level(dsl_scan_t *scn, zbookmark_phys_t zb)
3898 {
3899 dsl_pool_t *dp = scn->scn_dp;
3900 dsl_dataset_t *ds;
3901 objset_t *os;
3902 if (dsl_dataset_hold_obj(dp, zb.zb_objset, FTAG, &ds) != 0)
3903 return;
3904
3905 if (dmu_objset_from_ds(ds, &os) != 0) {
3906 dsl_dataset_rele(ds, FTAG);
3907 return;
3908 }
3909
3910 /*
3911 * If the key is not loaded dbuf_dnode_findbp() will error out with
3912 * EACCES. However in that case dnode_hold() will eventually call
3913 * dbuf_read()->zio_wait() which may call spa_log_error(). This will
3914 * lead to a deadlock due to us holding the mutex spa_errlist_lock.
3915 * Avoid this by checking here if the keys are loaded, if not return.
3916 * If the keys are not loaded the head_errlog feature is meaningless
3917 * as we cannot figure out the birth txg of the block pointer.
3918 */
3919 if (dsl_dataset_get_keystatus(ds->ds_dir) ==
3920 ZFS_KEYSTATUS_UNAVAILABLE) {
3921 dsl_dataset_rele(ds, FTAG);
3922 return;
3923 }
3924
3925 dnode_t *dn;
3926 blkptr_t bp;
3927
3928 if (dnode_hold(os, zb.zb_object, FTAG, &dn) != 0) {
3929 dsl_dataset_rele(ds, FTAG);
3930 return;
3931 }
3932
3933 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3934 int error = dbuf_dnode_findbp(dn, zb.zb_level, zb.zb_blkid, &bp, NULL,
3935 NULL);
3936
3937 if (error) {
3938 rw_exit(&dn->dn_struct_rwlock);
3939 dnode_rele(dn, FTAG);
3940 dsl_dataset_rele(ds, FTAG);
3941 return;
3942 }
3943
3944 if (!error && BP_IS_HOLE(&bp)) {
3945 rw_exit(&dn->dn_struct_rwlock);
3946 dnode_rele(dn, FTAG);
3947 dsl_dataset_rele(ds, FTAG);
3948 return;
3949 }
3950
3951 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW |
3952 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCRUB;
3953
3954 /* If it's an intent log block, failure is expected. */
3955 if (zb.zb_level == ZB_ZIL_LEVEL)
3956 zio_flags |= ZIO_FLAG_SPECULATIVE;
3957
3958 ASSERT(!BP_IS_EMBEDDED(&bp));
3959 scan_exec_io(dp, &bp, zio_flags, &zb, NULL);
3960 rw_exit(&dn->dn_struct_rwlock);
3961 dnode_rele(dn, FTAG);
3962 dsl_dataset_rele(ds, FTAG);
3963 }
3964
3965 /*
3966 * We keep track of the scrubbed error blocks in "count". This will be used
3967 * when deciding whether we exceeded zfs_scrub_error_blocks_per_txg. This
3968 * function is modelled after check_filesystem().
3969 */
3970 static int
scrub_filesystem(spa_t * spa,uint64_t fs,zbookmark_err_phys_t * zep,int * count)3971 scrub_filesystem(spa_t *spa, uint64_t fs, zbookmark_err_phys_t *zep,
3972 int *count)
3973 {
3974 dsl_dataset_t *ds;
3975 dsl_pool_t *dp = spa->spa_dsl_pool;
3976 dsl_scan_t *scn = dp->dp_scan;
3977
3978 int error = dsl_dataset_hold_obj(dp, fs, FTAG, &ds);
3979 if (error != 0)
3980 return (error);
3981
3982 uint64_t latest_txg;
3983 uint64_t txg_to_consider = spa->spa_syncing_txg;
3984 boolean_t check_snapshot = B_TRUE;
3985
3986 error = find_birth_txg(ds, zep, &latest_txg);
3987
3988 /*
3989 * If find_birth_txg() errors out, then err on the side of caution and
3990 * proceed. In worst case scenario scrub all objects. If zep->zb_birth
3991 * is 0 (e.g. in case of encryption with unloaded keys) also proceed to
3992 * scrub all objects.
3993 */
3994 if (error == 0 && zep->zb_birth == latest_txg) {
3995 /* Block neither free nor re written. */
3996 zbookmark_phys_t zb;
3997 zep_to_zb(fs, zep, &zb);
3998 scn->scn_zio_root = zio_root(spa, NULL, NULL,
3999 ZIO_FLAG_CANFAIL);
4000 /* We have already acquired the config lock for spa */
4001 read_by_block_level(scn, zb);
4002
4003 (void) zio_wait(scn->scn_zio_root);
4004 scn->scn_zio_root = NULL;
4005
4006 scn->errorscrub_phys.dep_examined++;
4007 scn->errorscrub_phys.dep_to_examine--;
4008 (*count)++;
4009 if ((*count) == zfs_scrub_error_blocks_per_txg ||
4010 dsl_error_scrub_check_suspend(scn, &zb)) {
4011 dsl_dataset_rele(ds, FTAG);
4012 return (SET_ERROR(EFAULT));
4013 }
4014
4015 check_snapshot = B_FALSE;
4016 } else if (error == 0) {
4017 txg_to_consider = latest_txg;
4018 }
4019
4020 /*
4021 * Retrieve the number of snapshots if the dataset is not a snapshot.
4022 */
4023 uint64_t snap_count = 0;
4024 if (dsl_dataset_phys(ds)->ds_snapnames_zapobj != 0) {
4025
4026 error = zap_count(spa->spa_meta_objset,
4027 dsl_dataset_phys(ds)->ds_snapnames_zapobj, &snap_count);
4028
4029 if (error != 0) {
4030 dsl_dataset_rele(ds, FTAG);
4031 return (error);
4032 }
4033 }
4034
4035 if (snap_count == 0) {
4036 /* Filesystem without snapshots. */
4037 dsl_dataset_rele(ds, FTAG);
4038 return (0);
4039 }
4040
4041 uint64_t snap_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
4042 uint64_t snap_obj_txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
4043
4044 dsl_dataset_rele(ds, FTAG);
4045
4046 /* Check only snapshots created from this file system. */
4047 while (snap_obj != 0 && zep->zb_birth < snap_obj_txg &&
4048 snap_obj_txg <= txg_to_consider) {
4049
4050 error = dsl_dataset_hold_obj(dp, snap_obj, FTAG, &ds);
4051 if (error != 0)
4052 return (error);
4053
4054 if (dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj != fs) {
4055 snap_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
4056 snap_obj_txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
4057 dsl_dataset_rele(ds, FTAG);
4058 continue;
4059 }
4060
4061 boolean_t affected = B_TRUE;
4062 if (check_snapshot) {
4063 uint64_t blk_txg;
4064 error = find_birth_txg(ds, zep, &blk_txg);
4065
4066 /*
4067 * Scrub the snapshot also when zb_birth == 0 or when
4068 * find_birth_txg() returns an error.
4069 */
4070 affected = (error == 0 && zep->zb_birth == blk_txg) ||
4071 (error != 0) || (zep->zb_birth == 0);
4072 }
4073
4074 /* Scrub snapshots. */
4075 if (affected) {
4076 zbookmark_phys_t zb;
4077 zep_to_zb(snap_obj, zep, &zb);
4078 scn->scn_zio_root = zio_root(spa, NULL, NULL,
4079 ZIO_FLAG_CANFAIL);
4080 /* We have already acquired the config lock for spa */
4081 read_by_block_level(scn, zb);
4082
4083 (void) zio_wait(scn->scn_zio_root);
4084 scn->scn_zio_root = NULL;
4085
4086 scn->errorscrub_phys.dep_examined++;
4087 scn->errorscrub_phys.dep_to_examine--;
4088 (*count)++;
4089 if ((*count) == zfs_scrub_error_blocks_per_txg ||
4090 dsl_error_scrub_check_suspend(scn, &zb)) {
4091 dsl_dataset_rele(ds, FTAG);
4092 return (EFAULT);
4093 }
4094 }
4095 snap_obj_txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
4096 snap_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
4097 dsl_dataset_rele(ds, FTAG);
4098 }
4099 return (0);
4100 }
4101
4102 void
dsl_errorscrub_sync(dsl_pool_t * dp,dmu_tx_t * tx)4103 dsl_errorscrub_sync(dsl_pool_t *dp, dmu_tx_t *tx)
4104 {
4105 spa_t *spa = dp->dp_spa;
4106 dsl_scan_t *scn = dp->dp_scan;
4107
4108 /*
4109 * Only process scans in sync pass 1.
4110 */
4111
4112 if (spa_sync_pass(spa) > 1)
4113 return;
4114
4115 /*
4116 * If the spa is shutting down, then stop scanning. This will
4117 * ensure that the scan does not dirty any new data during the
4118 * shutdown phase.
4119 */
4120 if (spa_shutting_down(spa))
4121 return;
4122
4123 if (!dsl_errorscrub_active(scn) || dsl_errorscrub_is_paused(scn)) {
4124 return;
4125 }
4126
4127 if (dsl_scan_resilvering(scn->scn_dp)) {
4128 /* cancel the error scrub if resilver started */
4129 dsl_scan_cancel(scn->scn_dp);
4130 return;
4131 }
4132
4133 spa->spa_scrub_active = B_TRUE;
4134 scn->scn_sync_start_time = gethrtime();
4135
4136 /*
4137 * zfs_scan_suspend_progress can be set to disable scrub progress.
4138 * See more detailed comment in dsl_scan_sync().
4139 */
4140 if (zfs_scan_suspend_progress) {
4141 uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time;
4142 int mintime = zfs_scrub_min_time_ms;
4143
4144 while (zfs_scan_suspend_progress &&
4145 !txg_sync_waiting(scn->scn_dp) &&
4146 !spa_shutting_down(scn->scn_dp->dp_spa) &&
4147 NSEC2MSEC(scan_time_ns) < mintime) {
4148 delay(hz);
4149 scan_time_ns = gethrtime() - scn->scn_sync_start_time;
4150 }
4151 return;
4152 }
4153
4154 int i = 0;
4155 zap_attribute_t *za;
4156 zbookmark_phys_t *zb;
4157 boolean_t limit_exceeded = B_FALSE;
4158
4159 za = kmem_zalloc(sizeof (zap_attribute_t), KM_SLEEP);
4160 zb = kmem_zalloc(sizeof (zbookmark_phys_t), KM_SLEEP);
4161
4162 if (!spa_feature_is_enabled(spa, SPA_FEATURE_HEAD_ERRLOG)) {
4163 for (; zap_cursor_retrieve(&scn->errorscrub_cursor, za) == 0;
4164 zap_cursor_advance(&scn->errorscrub_cursor)) {
4165 name_to_bookmark(za->za_name, zb);
4166
4167 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
4168 NULL, ZIO_FLAG_CANFAIL);
4169 dsl_pool_config_enter(dp, FTAG);
4170 read_by_block_level(scn, *zb);
4171 dsl_pool_config_exit(dp, FTAG);
4172
4173 (void) zio_wait(scn->scn_zio_root);
4174 scn->scn_zio_root = NULL;
4175
4176 scn->errorscrub_phys.dep_examined += 1;
4177 scn->errorscrub_phys.dep_to_examine -= 1;
4178 i++;
4179 if (i == zfs_scrub_error_blocks_per_txg ||
4180 dsl_error_scrub_check_suspend(scn, zb)) {
4181 limit_exceeded = B_TRUE;
4182 break;
4183 }
4184 }
4185
4186 if (!limit_exceeded)
4187 dsl_errorscrub_done(scn, B_TRUE, tx);
4188
4189 dsl_errorscrub_sync_state(scn, tx);
4190 kmem_free(za, sizeof (*za));
4191 kmem_free(zb, sizeof (*zb));
4192 return;
4193 }
4194
4195 int error = 0;
4196 for (; zap_cursor_retrieve(&scn->errorscrub_cursor, za) == 0;
4197 zap_cursor_advance(&scn->errorscrub_cursor)) {
4198
4199 zap_cursor_t *head_ds_cursor;
4200 zap_attribute_t *head_ds_attr;
4201 zbookmark_err_phys_t head_ds_block;
4202
4203 head_ds_cursor = kmem_zalloc(sizeof (zap_cursor_t), KM_SLEEP);
4204 head_ds_attr = kmem_zalloc(sizeof (zap_attribute_t), KM_SLEEP);
4205
4206 uint64_t head_ds_err_obj = za->za_first_integer;
4207 uint64_t head_ds;
4208 name_to_object(za->za_name, &head_ds);
4209 boolean_t config_held = B_FALSE;
4210 uint64_t top_affected_fs;
4211
4212 for (zap_cursor_init(head_ds_cursor, spa->spa_meta_objset,
4213 head_ds_err_obj); zap_cursor_retrieve(head_ds_cursor,
4214 head_ds_attr) == 0; zap_cursor_advance(head_ds_cursor)) {
4215
4216 name_to_errphys(head_ds_attr->za_name, &head_ds_block);
4217
4218 /*
4219 * In case we are called from spa_sync the pool
4220 * config is already held.
4221 */
4222 if (!dsl_pool_config_held(dp)) {
4223 dsl_pool_config_enter(dp, FTAG);
4224 config_held = B_TRUE;
4225 }
4226
4227 error = find_top_affected_fs(spa,
4228 head_ds, &head_ds_block, &top_affected_fs);
4229 if (error)
4230 break;
4231
4232 error = scrub_filesystem(spa, top_affected_fs,
4233 &head_ds_block, &i);
4234
4235 if (error == SET_ERROR(EFAULT)) {
4236 limit_exceeded = B_TRUE;
4237 break;
4238 }
4239 }
4240
4241 zap_cursor_fini(head_ds_cursor);
4242 kmem_free(head_ds_cursor, sizeof (*head_ds_cursor));
4243 kmem_free(head_ds_attr, sizeof (*head_ds_attr));
4244
4245 if (config_held)
4246 dsl_pool_config_exit(dp, FTAG);
4247 }
4248
4249 kmem_free(za, sizeof (*za));
4250 kmem_free(zb, sizeof (*zb));
4251 if (!limit_exceeded)
4252 dsl_errorscrub_done(scn, B_TRUE, tx);
4253
4254 dsl_errorscrub_sync_state(scn, tx);
4255 }
4256
4257 /*
4258 * This is the primary entry point for scans that is called from syncing
4259 * context. Scans must happen entirely during syncing context so that we
4260 * can guarantee that blocks we are currently scanning will not change out
4261 * from under us. While a scan is active, this function controls how quickly
4262 * transaction groups proceed, instead of the normal handling provided by
4263 * txg_sync_thread().
4264 */
4265 void
dsl_scan_sync(dsl_pool_t * dp,dmu_tx_t * tx)4266 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
4267 {
4268 int err = 0;
4269 dsl_scan_t *scn = dp->dp_scan;
4270 spa_t *spa = dp->dp_spa;
4271 state_sync_type_t sync_type = SYNC_OPTIONAL;
4272
4273 if (spa->spa_resilver_deferred &&
4274 !spa_feature_is_active(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
4275 spa_feature_incr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
4276
4277 /*
4278 * Check for scn_restart_txg before checking spa_load_state, so
4279 * that we can restart an old-style scan while the pool is being
4280 * imported (see dsl_scan_init). We also restart scans if there
4281 * is a deferred resilver and the user has manually disabled
4282 * deferred resilvers via the tunable.
4283 */
4284 if (dsl_scan_restarting(scn, tx) ||
4285 (spa->spa_resilver_deferred && zfs_resilver_disable_defer)) {
4286 pool_scan_func_t func = POOL_SCAN_SCRUB;
4287 dsl_scan_done(scn, B_FALSE, tx);
4288 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
4289 func = POOL_SCAN_RESILVER;
4290 zfs_dbgmsg("restarting scan func=%u on %s txg=%llu",
4291 func, dp->dp_spa->spa_name, (longlong_t)tx->tx_txg);
4292 dsl_scan_setup_sync(&func, tx);
4293 }
4294
4295 /*
4296 * Only process scans in sync pass 1.
4297 */
4298 if (spa_sync_pass(spa) > 1)
4299 return;
4300
4301 /*
4302 * If the spa is shutting down, then stop scanning. This will
4303 * ensure that the scan does not dirty any new data during the
4304 * shutdown phase.
4305 */
4306 if (spa_shutting_down(spa))
4307 return;
4308
4309 /*
4310 * If the scan is inactive due to a stalled async destroy, try again.
4311 */
4312 if (!scn->scn_async_stalled && !dsl_scan_active(scn))
4313 return;
4314
4315 /* reset scan statistics */
4316 scn->scn_visited_this_txg = 0;
4317 scn->scn_dedup_frees_this_txg = 0;
4318 scn->scn_holes_this_txg = 0;
4319 scn->scn_lt_min_this_txg = 0;
4320 scn->scn_gt_max_this_txg = 0;
4321 scn->scn_ddt_contained_this_txg = 0;
4322 scn->scn_objsets_visited_this_txg = 0;
4323 scn->scn_avg_seg_size_this_txg = 0;
4324 scn->scn_segs_this_txg = 0;
4325 scn->scn_avg_zio_size_this_txg = 0;
4326 scn->scn_zios_this_txg = 0;
4327 scn->scn_suspending = B_FALSE;
4328 scn->scn_sync_start_time = gethrtime();
4329 spa->spa_scrub_active = B_TRUE;
4330
4331 /*
4332 * First process the async destroys. If we suspend, don't do
4333 * any scrubbing or resilvering. This ensures that there are no
4334 * async destroys while we are scanning, so the scan code doesn't
4335 * have to worry about traversing it. It is also faster to free the
4336 * blocks than to scrub them.
4337 */
4338 err = dsl_process_async_destroys(dp, tx);
4339 if (err != 0)
4340 return;
4341
4342 if (!dsl_scan_is_running(scn) || dsl_scan_is_paused_scrub(scn))
4343 return;
4344
4345 /*
4346 * Wait a few txgs after importing to begin scanning so that
4347 * we can get the pool imported quickly.
4348 */
4349 if (spa->spa_syncing_txg < spa->spa_first_txg + SCAN_IMPORT_WAIT_TXGS)
4350 return;
4351
4352 /*
4353 * zfs_scan_suspend_progress can be set to disable scan progress.
4354 * We don't want to spin the txg_sync thread, so we add a delay
4355 * here to simulate the time spent doing a scan. This is mostly
4356 * useful for testing and debugging.
4357 */
4358 if (zfs_scan_suspend_progress) {
4359 uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time;
4360 uint_t mintime = (scn->scn_phys.scn_func ==
4361 POOL_SCAN_RESILVER) ? zfs_resilver_min_time_ms :
4362 zfs_scrub_min_time_ms;
4363
4364 while (zfs_scan_suspend_progress &&
4365 !txg_sync_waiting(scn->scn_dp) &&
4366 !spa_shutting_down(scn->scn_dp->dp_spa) &&
4367 NSEC2MSEC(scan_time_ns) < mintime) {
4368 delay(hz);
4369 scan_time_ns = gethrtime() - scn->scn_sync_start_time;
4370 }
4371 return;
4372 }
4373
4374 /*
4375 * Disabled by default, set zfs_scan_report_txgs to report
4376 * average performance over the last zfs_scan_report_txgs TXGs.
4377 */
4378 if (zfs_scan_report_txgs != 0 &&
4379 tx->tx_txg % zfs_scan_report_txgs == 0) {
4380 scn->scn_issued_before_pass += spa->spa_scan_pass_issued;
4381 spa_scan_stat_init(spa);
4382 }
4383
4384 /*
4385 * It is possible to switch from unsorted to sorted at any time,
4386 * but afterwards the scan will remain sorted unless reloaded from
4387 * a checkpoint after a reboot.
4388 */
4389 if (!zfs_scan_legacy) {
4390 scn->scn_is_sorted = B_TRUE;
4391 if (scn->scn_last_checkpoint == 0)
4392 scn->scn_last_checkpoint = ddi_get_lbolt();
4393 }
4394
4395 /*
4396 * For sorted scans, determine what kind of work we will be doing
4397 * this txg based on our memory limitations and whether or not we
4398 * need to perform a checkpoint.
4399 */
4400 if (scn->scn_is_sorted) {
4401 /*
4402 * If we are over our checkpoint interval, set scn_clearing
4403 * so that we can begin checkpointing immediately. The
4404 * checkpoint allows us to save a consistent bookmark
4405 * representing how much data we have scrubbed so far.
4406 * Otherwise, use the memory limit to determine if we should
4407 * scan for metadata or start issue scrub IOs. We accumulate
4408 * metadata until we hit our hard memory limit at which point
4409 * we issue scrub IOs until we are at our soft memory limit.
4410 */
4411 if (scn->scn_checkpointing ||
4412 ddi_get_lbolt() - scn->scn_last_checkpoint >
4413 SEC_TO_TICK(zfs_scan_checkpoint_intval)) {
4414 if (!scn->scn_checkpointing)
4415 zfs_dbgmsg("begin scan checkpoint for %s",
4416 spa->spa_name);
4417
4418 scn->scn_checkpointing = B_TRUE;
4419 scn->scn_clearing = B_TRUE;
4420 } else {
4421 boolean_t should_clear = dsl_scan_should_clear(scn);
4422 if (should_clear && !scn->scn_clearing) {
4423 zfs_dbgmsg("begin scan clearing for %s",
4424 spa->spa_name);
4425 scn->scn_clearing = B_TRUE;
4426 } else if (!should_clear && scn->scn_clearing) {
4427 zfs_dbgmsg("finish scan clearing for %s",
4428 spa->spa_name);
4429 scn->scn_clearing = B_FALSE;
4430 }
4431 }
4432 } else {
4433 ASSERT0(scn->scn_checkpointing);
4434 ASSERT0(scn->scn_clearing);
4435 }
4436
4437 if (!scn->scn_clearing && scn->scn_done_txg == 0) {
4438 /* Need to scan metadata for more blocks to scrub */
4439 dsl_scan_phys_t *scnp = &scn->scn_phys;
4440 taskqid_t prefetch_tqid;
4441
4442 /*
4443 * Calculate the max number of in-flight bytes for pool-wide
4444 * scanning operations (minimum 1MB, maximum 1/4 of arc_c_max).
4445 * Limits for the issuing phase are done per top-level vdev and
4446 * are handled separately.
4447 */
4448 scn->scn_maxinflight_bytes = MIN(arc_c_max / 4, MAX(1ULL << 20,
4449 zfs_scan_vdev_limit * dsl_scan_count_data_disks(spa)));
4450
4451 if (scnp->scn_ddt_bookmark.ddb_class <=
4452 scnp->scn_ddt_class_max) {
4453 ASSERT(ZB_IS_ZERO(&scnp->scn_bookmark));
4454 zfs_dbgmsg("doing scan sync for %s txg %llu; "
4455 "ddt bm=%llu/%llu/%llu/%llx",
4456 spa->spa_name,
4457 (longlong_t)tx->tx_txg,
4458 (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
4459 (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
4460 (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
4461 (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
4462 } else {
4463 zfs_dbgmsg("doing scan sync for %s txg %llu; "
4464 "bm=%llu/%llu/%llu/%llu",
4465 spa->spa_name,
4466 (longlong_t)tx->tx_txg,
4467 (longlong_t)scnp->scn_bookmark.zb_objset,
4468 (longlong_t)scnp->scn_bookmark.zb_object,
4469 (longlong_t)scnp->scn_bookmark.zb_level,
4470 (longlong_t)scnp->scn_bookmark.zb_blkid);
4471 }
4472
4473 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
4474 NULL, ZIO_FLAG_CANFAIL);
4475
4476 scn->scn_prefetch_stop = B_FALSE;
4477 prefetch_tqid = taskq_dispatch(dp->dp_sync_taskq,
4478 dsl_scan_prefetch_thread, scn, TQ_SLEEP);
4479 ASSERT(prefetch_tqid != TASKQID_INVALID);
4480
4481 dsl_pool_config_enter(dp, FTAG);
4482 dsl_scan_visit(scn, tx);
4483 dsl_pool_config_exit(dp, FTAG);
4484
4485 mutex_enter(&dp->dp_spa->spa_scrub_lock);
4486 scn->scn_prefetch_stop = B_TRUE;
4487 cv_broadcast(&spa->spa_scrub_io_cv);
4488 mutex_exit(&dp->dp_spa->spa_scrub_lock);
4489
4490 taskq_wait_id(dp->dp_sync_taskq, prefetch_tqid);
4491 (void) zio_wait(scn->scn_zio_root);
4492 scn->scn_zio_root = NULL;
4493
4494 zfs_dbgmsg("scan visited %llu blocks of %s in %llums "
4495 "(%llu os's, %llu holes, %llu < mintxg, "
4496 "%llu in ddt, %llu > maxtxg)",
4497 (longlong_t)scn->scn_visited_this_txg,
4498 spa->spa_name,
4499 (longlong_t)NSEC2MSEC(gethrtime() -
4500 scn->scn_sync_start_time),
4501 (longlong_t)scn->scn_objsets_visited_this_txg,
4502 (longlong_t)scn->scn_holes_this_txg,
4503 (longlong_t)scn->scn_lt_min_this_txg,
4504 (longlong_t)scn->scn_ddt_contained_this_txg,
4505 (longlong_t)scn->scn_gt_max_this_txg);
4506
4507 if (!scn->scn_suspending) {
4508 ASSERT0(avl_numnodes(&scn->scn_queue));
4509 scn->scn_done_txg = tx->tx_txg + 1;
4510 if (scn->scn_is_sorted) {
4511 scn->scn_checkpointing = B_TRUE;
4512 scn->scn_clearing = B_TRUE;
4513 scn->scn_issued_before_pass +=
4514 spa->spa_scan_pass_issued;
4515 spa_scan_stat_init(spa);
4516 }
4517 zfs_dbgmsg("scan complete for %s txg %llu",
4518 spa->spa_name,
4519 (longlong_t)tx->tx_txg);
4520 }
4521 } else if (scn->scn_is_sorted && scn->scn_queues_pending != 0) {
4522 ASSERT(scn->scn_clearing);
4523
4524 /* need to issue scrubbing IOs from per-vdev queues */
4525 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
4526 NULL, ZIO_FLAG_CANFAIL);
4527 scan_io_queues_run(scn);
4528 (void) zio_wait(scn->scn_zio_root);
4529 scn->scn_zio_root = NULL;
4530
4531 /* calculate and dprintf the current memory usage */
4532 (void) dsl_scan_should_clear(scn);
4533 dsl_scan_update_stats(scn);
4534
4535 zfs_dbgmsg("scan issued %llu blocks for %s (%llu segs) "
4536 "in %llums (avg_block_size = %llu, avg_seg_size = %llu)",
4537 (longlong_t)scn->scn_zios_this_txg,
4538 spa->spa_name,
4539 (longlong_t)scn->scn_segs_this_txg,
4540 (longlong_t)NSEC2MSEC(gethrtime() -
4541 scn->scn_sync_start_time),
4542 (longlong_t)scn->scn_avg_zio_size_this_txg,
4543 (longlong_t)scn->scn_avg_seg_size_this_txg);
4544 } else if (scn->scn_done_txg != 0 && scn->scn_done_txg <= tx->tx_txg) {
4545 /* Finished with everything. Mark the scrub as complete */
4546 zfs_dbgmsg("scan issuing complete txg %llu for %s",
4547 (longlong_t)tx->tx_txg,
4548 spa->spa_name);
4549 ASSERT3U(scn->scn_done_txg, !=, 0);
4550 ASSERT0(spa->spa_scrub_inflight);
4551 ASSERT0(scn->scn_queues_pending);
4552 dsl_scan_done(scn, B_TRUE, tx);
4553 sync_type = SYNC_MANDATORY;
4554 }
4555
4556 dsl_scan_sync_state(scn, tx, sync_type);
4557 }
4558
4559 static void
count_block_issued(spa_t * spa,const blkptr_t * bp,boolean_t all)4560 count_block_issued(spa_t *spa, const blkptr_t *bp, boolean_t all)
4561 {
4562 /*
4563 * Don't count embedded bp's, since we already did the work of
4564 * scanning these when we scanned the containing block.
4565 */
4566 if (BP_IS_EMBEDDED(bp))
4567 return;
4568
4569 /*
4570 * Update the spa's stats on how many bytes we have issued.
4571 * Sequential scrubs create a zio for each DVA of the bp. Each
4572 * of these will include all DVAs for repair purposes, but the
4573 * zio code will only try the first one unless there is an issue.
4574 * Therefore, we should only count the first DVA for these IOs.
4575 */
4576 atomic_add_64(&spa->spa_scan_pass_issued,
4577 all ? BP_GET_ASIZE(bp) : DVA_GET_ASIZE(&bp->blk_dva[0]));
4578 }
4579
4580 static void
count_block_skipped(dsl_scan_t * scn,const blkptr_t * bp,boolean_t all)4581 count_block_skipped(dsl_scan_t *scn, const blkptr_t *bp, boolean_t all)
4582 {
4583 if (BP_IS_EMBEDDED(bp))
4584 return;
4585 atomic_add_64(&scn->scn_phys.scn_skipped,
4586 all ? BP_GET_ASIZE(bp) : DVA_GET_ASIZE(&bp->blk_dva[0]));
4587 }
4588
4589 static void
count_block(zfs_all_blkstats_t * zab,const blkptr_t * bp)4590 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
4591 {
4592 /*
4593 * If we resume after a reboot, zab will be NULL; don't record
4594 * incomplete stats in that case.
4595 */
4596 if (zab == NULL)
4597 return;
4598
4599 for (int i = 0; i < 4; i++) {
4600 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
4601 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
4602
4603 if (t & DMU_OT_NEWTYPE)
4604 t = DMU_OT_OTHER;
4605 zfs_blkstat_t *zb = &zab->zab_type[l][t];
4606 int equal;
4607
4608 zb->zb_count++;
4609 zb->zb_asize += BP_GET_ASIZE(bp);
4610 zb->zb_lsize += BP_GET_LSIZE(bp);
4611 zb->zb_psize += BP_GET_PSIZE(bp);
4612 zb->zb_gangs += BP_COUNT_GANG(bp);
4613
4614 switch (BP_GET_NDVAS(bp)) {
4615 case 2:
4616 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
4617 DVA_GET_VDEV(&bp->blk_dva[1]))
4618 zb->zb_ditto_2_of_2_samevdev++;
4619 break;
4620 case 3:
4621 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
4622 DVA_GET_VDEV(&bp->blk_dva[1])) +
4623 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
4624 DVA_GET_VDEV(&bp->blk_dva[2])) +
4625 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
4626 DVA_GET_VDEV(&bp->blk_dva[2]));
4627 if (equal == 1)
4628 zb->zb_ditto_2_of_3_samevdev++;
4629 else if (equal == 3)
4630 zb->zb_ditto_3_of_3_samevdev++;
4631 break;
4632 }
4633 }
4634 }
4635
4636 static void
scan_io_queue_insert_impl(dsl_scan_io_queue_t * queue,scan_io_t * sio)4637 scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue, scan_io_t *sio)
4638 {
4639 avl_index_t idx;
4640 dsl_scan_t *scn = queue->q_scn;
4641
4642 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
4643
4644 if (unlikely(avl_is_empty(&queue->q_sios_by_addr)))
4645 atomic_add_64(&scn->scn_queues_pending, 1);
4646 if (avl_find(&queue->q_sios_by_addr, sio, &idx) != NULL) {
4647 /* block is already scheduled for reading */
4648 sio_free(sio);
4649 return;
4650 }
4651 avl_insert(&queue->q_sios_by_addr, sio, idx);
4652 queue->q_sio_memused += SIO_GET_MUSED(sio);
4653 range_tree_add(queue->q_exts_by_addr, SIO_GET_OFFSET(sio),
4654 SIO_GET_ASIZE(sio));
4655 }
4656
4657 /*
4658 * Given all the info we got from our metadata scanning process, we
4659 * construct a scan_io_t and insert it into the scan sorting queue. The
4660 * I/O must already be suitable for us to process. This is controlled
4661 * by dsl_scan_enqueue().
4662 */
4663 static void
scan_io_queue_insert(dsl_scan_io_queue_t * queue,const blkptr_t * bp,int dva_i,int zio_flags,const zbookmark_phys_t * zb)4664 scan_io_queue_insert(dsl_scan_io_queue_t *queue, const blkptr_t *bp, int dva_i,
4665 int zio_flags, const zbookmark_phys_t *zb)
4666 {
4667 scan_io_t *sio = sio_alloc(BP_GET_NDVAS(bp));
4668
4669 ASSERT0(BP_IS_GANG(bp));
4670 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
4671
4672 bp2sio(bp, sio, dva_i);
4673 sio->sio_flags = zio_flags;
4674 sio->sio_zb = *zb;
4675
4676 queue->q_last_ext_addr = -1;
4677 scan_io_queue_insert_impl(queue, sio);
4678 }
4679
4680 /*
4681 * Given a set of I/O parameters as discovered by the metadata traversal
4682 * process, attempts to place the I/O into the sorted queues (if allowed),
4683 * or immediately executes the I/O.
4684 */
4685 static void
dsl_scan_enqueue(dsl_pool_t * dp,const blkptr_t * bp,int zio_flags,const zbookmark_phys_t * zb)4686 dsl_scan_enqueue(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
4687 const zbookmark_phys_t *zb)
4688 {
4689 spa_t *spa = dp->dp_spa;
4690
4691 ASSERT(!BP_IS_EMBEDDED(bp));
4692
4693 /*
4694 * Gang blocks are hard to issue sequentially, so we just issue them
4695 * here immediately instead of queuing them.
4696 */
4697 if (!dp->dp_scan->scn_is_sorted || BP_IS_GANG(bp)) {
4698 scan_exec_io(dp, bp, zio_flags, zb, NULL);
4699 return;
4700 }
4701
4702 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
4703 dva_t dva;
4704 vdev_t *vdev;
4705
4706 dva = bp->blk_dva[i];
4707 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&dva));
4708 ASSERT(vdev != NULL);
4709
4710 mutex_enter(&vdev->vdev_scan_io_queue_lock);
4711 if (vdev->vdev_scan_io_queue == NULL)
4712 vdev->vdev_scan_io_queue = scan_io_queue_create(vdev);
4713 ASSERT(dp->dp_scan != NULL);
4714 scan_io_queue_insert(vdev->vdev_scan_io_queue, bp,
4715 i, zio_flags, zb);
4716 mutex_exit(&vdev->vdev_scan_io_queue_lock);
4717 }
4718 }
4719
4720 static int
dsl_scan_scrub_cb(dsl_pool_t * dp,const blkptr_t * bp,const zbookmark_phys_t * zb)4721 dsl_scan_scrub_cb(dsl_pool_t *dp,
4722 const blkptr_t *bp, const zbookmark_phys_t *zb)
4723 {
4724 dsl_scan_t *scn = dp->dp_scan;
4725 spa_t *spa = dp->dp_spa;
4726 uint64_t phys_birth = BP_GET_BIRTH(bp);
4727 size_t psize = BP_GET_PSIZE(bp);
4728 boolean_t needs_io = B_FALSE;
4729 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
4730
4731 count_block(dp->dp_blkstats, bp);
4732 if (phys_birth <= scn->scn_phys.scn_min_txg ||
4733 phys_birth >= scn->scn_phys.scn_max_txg) {
4734 count_block_skipped(scn, bp, B_TRUE);
4735 return (0);
4736 }
4737
4738 /* Embedded BP's have phys_birth==0, so we reject them above. */
4739 ASSERT(!BP_IS_EMBEDDED(bp));
4740
4741 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
4742 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
4743 zio_flags |= ZIO_FLAG_SCRUB;
4744 needs_io = B_TRUE;
4745 } else {
4746 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
4747 zio_flags |= ZIO_FLAG_RESILVER;
4748 needs_io = B_FALSE;
4749 }
4750
4751 /* If it's an intent log block, failure is expected. */
4752 if (zb->zb_level == ZB_ZIL_LEVEL)
4753 zio_flags |= ZIO_FLAG_SPECULATIVE;
4754
4755 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
4756 const dva_t *dva = &bp->blk_dva[d];
4757
4758 /*
4759 * Keep track of how much data we've examined so that
4760 * zpool(8) status can make useful progress reports.
4761 */
4762 uint64_t asize = DVA_GET_ASIZE(dva);
4763 scn->scn_phys.scn_examined += asize;
4764 spa->spa_scan_pass_exam += asize;
4765
4766 /* if it's a resilver, this may not be in the target range */
4767 if (!needs_io)
4768 needs_io = dsl_scan_need_resilver(spa, dva, psize,
4769 phys_birth);
4770 }
4771
4772 if (needs_io && !zfs_no_scrub_io) {
4773 dsl_scan_enqueue(dp, bp, zio_flags, zb);
4774 } else {
4775 count_block_skipped(scn, bp, B_TRUE);
4776 }
4777
4778 /* do not relocate this block */
4779 return (0);
4780 }
4781
4782 static void
dsl_scan_scrub_done(zio_t * zio)4783 dsl_scan_scrub_done(zio_t *zio)
4784 {
4785 spa_t *spa = zio->io_spa;
4786 blkptr_t *bp = zio->io_bp;
4787 dsl_scan_io_queue_t *queue = zio->io_private;
4788
4789 abd_free(zio->io_abd);
4790
4791 if (queue == NULL) {
4792 mutex_enter(&spa->spa_scrub_lock);
4793 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
4794 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
4795 cv_broadcast(&spa->spa_scrub_io_cv);
4796 mutex_exit(&spa->spa_scrub_lock);
4797 } else {
4798 mutex_enter(&queue->q_vd->vdev_scan_io_queue_lock);
4799 ASSERT3U(queue->q_inflight_bytes, >=, BP_GET_PSIZE(bp));
4800 queue->q_inflight_bytes -= BP_GET_PSIZE(bp);
4801 cv_broadcast(&queue->q_zio_cv);
4802 mutex_exit(&queue->q_vd->vdev_scan_io_queue_lock);
4803 }
4804
4805 if (zio->io_error && (zio->io_error != ECKSUM ||
4806 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
4807 if (dsl_errorscrubbing(spa->spa_dsl_pool) &&
4808 !dsl_errorscrub_is_paused(spa->spa_dsl_pool->dp_scan)) {
4809 atomic_inc_64(&spa->spa_dsl_pool->dp_scan
4810 ->errorscrub_phys.dep_errors);
4811 } else {
4812 atomic_inc_64(&spa->spa_dsl_pool->dp_scan->scn_phys
4813 .scn_errors);
4814 }
4815 }
4816 }
4817
4818 /*
4819 * Given a scanning zio's information, executes the zio. The zio need
4820 * not necessarily be only sortable, this function simply executes the
4821 * zio, no matter what it is. The optional queue argument allows the
4822 * caller to specify that they want per top level vdev IO rate limiting
4823 * instead of the legacy global limiting.
4824 */
4825 static void
scan_exec_io(dsl_pool_t * dp,const blkptr_t * bp,int zio_flags,const zbookmark_phys_t * zb,dsl_scan_io_queue_t * queue)4826 scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
4827 const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue)
4828 {
4829 spa_t *spa = dp->dp_spa;
4830 dsl_scan_t *scn = dp->dp_scan;
4831 size_t size = BP_GET_PSIZE(bp);
4832 abd_t *data = abd_alloc_for_io(size, B_FALSE);
4833 zio_t *pio;
4834
4835 if (queue == NULL) {
4836 ASSERT3U(scn->scn_maxinflight_bytes, >, 0);
4837 mutex_enter(&spa->spa_scrub_lock);
4838 while (spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)
4839 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
4840 spa->spa_scrub_inflight += BP_GET_PSIZE(bp);
4841 mutex_exit(&spa->spa_scrub_lock);
4842 pio = scn->scn_zio_root;
4843 } else {
4844 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
4845
4846 ASSERT3U(queue->q_maxinflight_bytes, >, 0);
4847 mutex_enter(q_lock);
4848 while (queue->q_inflight_bytes >= queue->q_maxinflight_bytes)
4849 cv_wait(&queue->q_zio_cv, q_lock);
4850 queue->q_inflight_bytes += BP_GET_PSIZE(bp);
4851 pio = queue->q_zio;
4852 mutex_exit(q_lock);
4853 }
4854
4855 ASSERT(pio != NULL);
4856 count_block_issued(spa, bp, queue == NULL);
4857 zio_nowait(zio_read(pio, spa, bp, data, size, dsl_scan_scrub_done,
4858 queue, ZIO_PRIORITY_SCRUB, zio_flags, zb));
4859 }
4860
4861 /*
4862 * This is the primary extent sorting algorithm. We balance two parameters:
4863 * 1) how many bytes of I/O are in an extent
4864 * 2) how well the extent is filled with I/O (as a fraction of its total size)
4865 * Since we allow extents to have gaps between their constituent I/Os, it's
4866 * possible to have a fairly large extent that contains the same amount of
4867 * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
4868 * The algorithm sorts based on a score calculated from the extent's size,
4869 * the relative fill volume (in %) and a "fill weight" parameter that controls
4870 * the split between whether we prefer larger extents or more well populated
4871 * extents:
4872 *
4873 * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
4874 *
4875 * Example:
4876 * 1) assume extsz = 64 MiB
4877 * 2) assume fill = 32 MiB (extent is half full)
4878 * 3) assume fill_weight = 3
4879 * 4) SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
4880 * SCORE = 32M + (50 * 3 * 32M) / 100
4881 * SCORE = 32M + (4800M / 100)
4882 * SCORE = 32M + 48M
4883 * ^ ^
4884 * | +--- final total relative fill-based score
4885 * +--------- final total fill-based score
4886 * SCORE = 80M
4887 *
4888 * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
4889 * extents that are more completely filled (in a 3:2 ratio) vs just larger.
4890 * Note that as an optimization, we replace multiplication and division by
4891 * 100 with bitshifting by 7 (which effectively multiplies and divides by 128).
4892 *
4893 * Since we do not care if one extent is only few percent better than another,
4894 * compress the score into 6 bits via binary logarithm AKA highbit64() and
4895 * put into otherwise unused due to ashift high bits of offset. This allows
4896 * to reduce q_exts_by_size B-tree elements to only 64 bits and compare them
4897 * with single operation. Plus it makes scrubs more sequential and reduces
4898 * chances that minor extent change move it within the B-tree.
4899 */
4900 __attribute__((always_inline)) inline
4901 static int
ext_size_compare(const void * x,const void * y)4902 ext_size_compare(const void *x, const void *y)
4903 {
4904 const uint64_t *a = x, *b = y;
4905
4906 return (TREE_CMP(*a, *b));
4907 }
4908
ZFS_BTREE_FIND_IN_BUF_FUNC(ext_size_find_in_buf,uint64_t,ext_size_compare)4909 ZFS_BTREE_FIND_IN_BUF_FUNC(ext_size_find_in_buf, uint64_t,
4910 ext_size_compare)
4911
4912 static void
4913 ext_size_create(range_tree_t *rt, void *arg)
4914 {
4915 (void) rt;
4916 zfs_btree_t *size_tree = arg;
4917
4918 zfs_btree_create(size_tree, ext_size_compare, ext_size_find_in_buf,
4919 sizeof (uint64_t));
4920 }
4921
4922 static void
ext_size_destroy(range_tree_t * rt,void * arg)4923 ext_size_destroy(range_tree_t *rt, void *arg)
4924 {
4925 (void) rt;
4926 zfs_btree_t *size_tree = arg;
4927 ASSERT0(zfs_btree_numnodes(size_tree));
4928
4929 zfs_btree_destroy(size_tree);
4930 }
4931
4932 static uint64_t
ext_size_value(range_tree_t * rt,range_seg_gap_t * rsg)4933 ext_size_value(range_tree_t *rt, range_seg_gap_t *rsg)
4934 {
4935 (void) rt;
4936 uint64_t size = rsg->rs_end - rsg->rs_start;
4937 uint64_t score = rsg->rs_fill + ((((rsg->rs_fill << 7) / size) *
4938 fill_weight * rsg->rs_fill) >> 7);
4939 ASSERT3U(rt->rt_shift, >=, 8);
4940 return (((uint64_t)(64 - highbit64(score)) << 56) | rsg->rs_start);
4941 }
4942
4943 static void
ext_size_add(range_tree_t * rt,range_seg_t * rs,void * arg)4944 ext_size_add(range_tree_t *rt, range_seg_t *rs, void *arg)
4945 {
4946 zfs_btree_t *size_tree = arg;
4947 ASSERT3U(rt->rt_type, ==, RANGE_SEG_GAP);
4948 uint64_t v = ext_size_value(rt, (range_seg_gap_t *)rs);
4949 zfs_btree_add(size_tree, &v);
4950 }
4951
4952 static void
ext_size_remove(range_tree_t * rt,range_seg_t * rs,void * arg)4953 ext_size_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
4954 {
4955 zfs_btree_t *size_tree = arg;
4956 ASSERT3U(rt->rt_type, ==, RANGE_SEG_GAP);
4957 uint64_t v = ext_size_value(rt, (range_seg_gap_t *)rs);
4958 zfs_btree_remove(size_tree, &v);
4959 }
4960
4961 static void
ext_size_vacate(range_tree_t * rt,void * arg)4962 ext_size_vacate(range_tree_t *rt, void *arg)
4963 {
4964 zfs_btree_t *size_tree = arg;
4965 zfs_btree_clear(size_tree);
4966 zfs_btree_destroy(size_tree);
4967
4968 ext_size_create(rt, arg);
4969 }
4970
4971 static const range_tree_ops_t ext_size_ops = {
4972 .rtop_create = ext_size_create,
4973 .rtop_destroy = ext_size_destroy,
4974 .rtop_add = ext_size_add,
4975 .rtop_remove = ext_size_remove,
4976 .rtop_vacate = ext_size_vacate
4977 };
4978
4979 /*
4980 * Comparator for the q_sios_by_addr tree. Sorting is simply performed
4981 * based on LBA-order (from lowest to highest).
4982 */
4983 static int
sio_addr_compare(const void * x,const void * y)4984 sio_addr_compare(const void *x, const void *y)
4985 {
4986 const scan_io_t *a = x, *b = y;
4987
4988 return (TREE_CMP(SIO_GET_OFFSET(a), SIO_GET_OFFSET(b)));
4989 }
4990
4991 /* IO queues are created on demand when they are needed. */
4992 static dsl_scan_io_queue_t *
scan_io_queue_create(vdev_t * vd)4993 scan_io_queue_create(vdev_t *vd)
4994 {
4995 dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan;
4996 dsl_scan_io_queue_t *q = kmem_zalloc(sizeof (*q), KM_SLEEP);
4997
4998 q->q_scn = scn;
4999 q->q_vd = vd;
5000 q->q_sio_memused = 0;
5001 q->q_last_ext_addr = -1;
5002 cv_init(&q->q_zio_cv, NULL, CV_DEFAULT, NULL);
5003 q->q_exts_by_addr = range_tree_create_gap(&ext_size_ops, RANGE_SEG_GAP,
5004 &q->q_exts_by_size, 0, vd->vdev_ashift, zfs_scan_max_ext_gap);
5005 avl_create(&q->q_sios_by_addr, sio_addr_compare,
5006 sizeof (scan_io_t), offsetof(scan_io_t, sio_nodes.sio_addr_node));
5007
5008 return (q);
5009 }
5010
5011 /*
5012 * Destroys a scan queue and all segments and scan_io_t's contained in it.
5013 * No further execution of I/O occurs, anything pending in the queue is
5014 * simply freed without being executed.
5015 */
5016 void
dsl_scan_io_queue_destroy(dsl_scan_io_queue_t * queue)5017 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t *queue)
5018 {
5019 dsl_scan_t *scn = queue->q_scn;
5020 scan_io_t *sio;
5021 void *cookie = NULL;
5022
5023 ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
5024
5025 if (!avl_is_empty(&queue->q_sios_by_addr))
5026 atomic_add_64(&scn->scn_queues_pending, -1);
5027 while ((sio = avl_destroy_nodes(&queue->q_sios_by_addr, &cookie)) !=
5028 NULL) {
5029 ASSERT(range_tree_contains(queue->q_exts_by_addr,
5030 SIO_GET_OFFSET(sio), SIO_GET_ASIZE(sio)));
5031 queue->q_sio_memused -= SIO_GET_MUSED(sio);
5032 sio_free(sio);
5033 }
5034
5035 ASSERT0(queue->q_sio_memused);
5036 range_tree_vacate(queue->q_exts_by_addr, NULL, queue);
5037 range_tree_destroy(queue->q_exts_by_addr);
5038 avl_destroy(&queue->q_sios_by_addr);
5039 cv_destroy(&queue->q_zio_cv);
5040
5041 kmem_free(queue, sizeof (*queue));
5042 }
5043
5044 /*
5045 * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
5046 * called on behalf of vdev_top_transfer when creating or destroying
5047 * a mirror vdev due to zpool attach/detach.
5048 */
5049 void
dsl_scan_io_queue_vdev_xfer(vdev_t * svd,vdev_t * tvd)5050 dsl_scan_io_queue_vdev_xfer(vdev_t *svd, vdev_t *tvd)
5051 {
5052 mutex_enter(&svd->vdev_scan_io_queue_lock);
5053 mutex_enter(&tvd->vdev_scan_io_queue_lock);
5054
5055 VERIFY3P(tvd->vdev_scan_io_queue, ==, NULL);
5056 tvd->vdev_scan_io_queue = svd->vdev_scan_io_queue;
5057 svd->vdev_scan_io_queue = NULL;
5058 if (tvd->vdev_scan_io_queue != NULL)
5059 tvd->vdev_scan_io_queue->q_vd = tvd;
5060
5061 mutex_exit(&tvd->vdev_scan_io_queue_lock);
5062 mutex_exit(&svd->vdev_scan_io_queue_lock);
5063 }
5064
5065 static void
scan_io_queues_destroy(dsl_scan_t * scn)5066 scan_io_queues_destroy(dsl_scan_t *scn)
5067 {
5068 vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
5069
5070 for (uint64_t i = 0; i < rvd->vdev_children; i++) {
5071 vdev_t *tvd = rvd->vdev_child[i];
5072
5073 mutex_enter(&tvd->vdev_scan_io_queue_lock);
5074 if (tvd->vdev_scan_io_queue != NULL)
5075 dsl_scan_io_queue_destroy(tvd->vdev_scan_io_queue);
5076 tvd->vdev_scan_io_queue = NULL;
5077 mutex_exit(&tvd->vdev_scan_io_queue_lock);
5078 }
5079 }
5080
5081 static void
dsl_scan_freed_dva(spa_t * spa,const blkptr_t * bp,int dva_i)5082 dsl_scan_freed_dva(spa_t *spa, const blkptr_t *bp, int dva_i)
5083 {
5084 dsl_pool_t *dp = spa->spa_dsl_pool;
5085 dsl_scan_t *scn = dp->dp_scan;
5086 vdev_t *vdev;
5087 kmutex_t *q_lock;
5088 dsl_scan_io_queue_t *queue;
5089 scan_io_t *srch_sio, *sio;
5090 avl_index_t idx;
5091 uint64_t start, size;
5092
5093 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&bp->blk_dva[dva_i]));
5094 ASSERT(vdev != NULL);
5095 q_lock = &vdev->vdev_scan_io_queue_lock;
5096 queue = vdev->vdev_scan_io_queue;
5097
5098 mutex_enter(q_lock);
5099 if (queue == NULL) {
5100 mutex_exit(q_lock);
5101 return;
5102 }
5103
5104 srch_sio = sio_alloc(BP_GET_NDVAS(bp));
5105 bp2sio(bp, srch_sio, dva_i);
5106 start = SIO_GET_OFFSET(srch_sio);
5107 size = SIO_GET_ASIZE(srch_sio);
5108
5109 /*
5110 * We can find the zio in two states:
5111 * 1) Cold, just sitting in the queue of zio's to be issued at
5112 * some point in the future. In this case, all we do is
5113 * remove the zio from the q_sios_by_addr tree, decrement
5114 * its data volume from the containing range_seg_t and
5115 * resort the q_exts_by_size tree to reflect that the
5116 * range_seg_t has lost some of its 'fill'. We don't shorten
5117 * the range_seg_t - this is usually rare enough not to be
5118 * worth the extra hassle of trying keep track of precise
5119 * extent boundaries.
5120 * 2) Hot, where the zio is currently in-flight in
5121 * dsl_scan_issue_ios. In this case, we can't simply
5122 * reach in and stop the in-flight zio's, so we instead
5123 * block the caller. Eventually, dsl_scan_issue_ios will
5124 * be done with issuing the zio's it gathered and will
5125 * signal us.
5126 */
5127 sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
5128 sio_free(srch_sio);
5129
5130 if (sio != NULL) {
5131 blkptr_t tmpbp;
5132
5133 /* Got it while it was cold in the queue */
5134 ASSERT3U(start, ==, SIO_GET_OFFSET(sio));
5135 ASSERT3U(size, ==, SIO_GET_ASIZE(sio));
5136 avl_remove(&queue->q_sios_by_addr, sio);
5137 if (avl_is_empty(&queue->q_sios_by_addr))
5138 atomic_add_64(&scn->scn_queues_pending, -1);
5139 queue->q_sio_memused -= SIO_GET_MUSED(sio);
5140
5141 ASSERT(range_tree_contains(queue->q_exts_by_addr, start, size));
5142 range_tree_remove_fill(queue->q_exts_by_addr, start, size);
5143
5144 /* count the block as though we skipped it */
5145 sio2bp(sio, &tmpbp);
5146 count_block_skipped(scn, &tmpbp, B_FALSE);
5147
5148 sio_free(sio);
5149 }
5150 mutex_exit(q_lock);
5151 }
5152
5153 /*
5154 * Callback invoked when a zio_free() zio is executing. This needs to be
5155 * intercepted to prevent the zio from deallocating a particular portion
5156 * of disk space and it then getting reallocated and written to, while we
5157 * still have it queued up for processing.
5158 */
5159 void
dsl_scan_freed(spa_t * spa,const blkptr_t * bp)5160 dsl_scan_freed(spa_t *spa, const blkptr_t *bp)
5161 {
5162 dsl_pool_t *dp = spa->spa_dsl_pool;
5163 dsl_scan_t *scn = dp->dp_scan;
5164
5165 ASSERT(!BP_IS_EMBEDDED(bp));
5166 ASSERT(scn != NULL);
5167 if (!dsl_scan_is_running(scn))
5168 return;
5169
5170 for (int i = 0; i < BP_GET_NDVAS(bp); i++)
5171 dsl_scan_freed_dva(spa, bp, i);
5172 }
5173
5174 /*
5175 * Check if a vdev needs resilvering (non-empty DTL), if so, and resilver has
5176 * not started, start it. Otherwise, only restart if max txg in DTL range is
5177 * greater than the max txg in the current scan. If the DTL max is less than
5178 * the scan max, then the vdev has not missed any new data since the resilver
5179 * started, so a restart is not needed.
5180 */
5181 void
dsl_scan_assess_vdev(dsl_pool_t * dp,vdev_t * vd)5182 dsl_scan_assess_vdev(dsl_pool_t *dp, vdev_t *vd)
5183 {
5184 uint64_t min, max;
5185
5186 if (!vdev_resilver_needed(vd, &min, &max))
5187 return;
5188
5189 if (!dsl_scan_resilvering(dp)) {
5190 spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER);
5191 return;
5192 }
5193
5194 if (max <= dp->dp_scan->scn_phys.scn_max_txg)
5195 return;
5196
5197 /* restart is needed, check if it can be deferred */
5198 if (spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
5199 vdev_defer_resilver(vd);
5200 else
5201 spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER);
5202 }
5203
5204 ZFS_MODULE_PARAM(zfs, zfs_, scan_vdev_limit, U64, ZMOD_RW,
5205 "Max bytes in flight per leaf vdev for scrubs and resilvers");
5206
5207 ZFS_MODULE_PARAM(zfs, zfs_, scrub_min_time_ms, UINT, ZMOD_RW,
5208 "Min millisecs to scrub per txg");
5209
5210 ZFS_MODULE_PARAM(zfs, zfs_, obsolete_min_time_ms, UINT, ZMOD_RW,
5211 "Min millisecs to obsolete per txg");
5212
5213 ZFS_MODULE_PARAM(zfs, zfs_, free_min_time_ms, UINT, ZMOD_RW,
5214 "Min millisecs to free per txg");
5215
5216 ZFS_MODULE_PARAM(zfs, zfs_, resilver_min_time_ms, UINT, ZMOD_RW,
5217 "Min millisecs to resilver per txg");
5218
5219 ZFS_MODULE_PARAM(zfs, zfs_, scan_suspend_progress, INT, ZMOD_RW,
5220 "Set to prevent scans from progressing");
5221
5222 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_io, INT, ZMOD_RW,
5223 "Set to disable scrub I/O");
5224
5225 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_prefetch, INT, ZMOD_RW,
5226 "Set to disable scrub prefetching");
5227
5228 ZFS_MODULE_PARAM(zfs, zfs_, async_block_max_blocks, U64, ZMOD_RW,
5229 "Max number of blocks freed in one txg");
5230
5231 ZFS_MODULE_PARAM(zfs, zfs_, max_async_dedup_frees, U64, ZMOD_RW,
5232 "Max number of dedup blocks freed in one txg");
5233
5234 ZFS_MODULE_PARAM(zfs, zfs_, free_bpobj_enabled, INT, ZMOD_RW,
5235 "Enable processing of the free_bpobj");
5236
5237 ZFS_MODULE_PARAM(zfs, zfs_, scan_blkstats, INT, ZMOD_RW,
5238 "Enable block statistics calculation during scrub");
5239
5240 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_fact, UINT, ZMOD_RW,
5241 "Fraction of RAM for scan hard limit");
5242
5243 ZFS_MODULE_PARAM(zfs, zfs_, scan_issue_strategy, UINT, ZMOD_RW,
5244 "IO issuing strategy during scrubbing. 0 = default, 1 = LBA, 2 = size");
5245
5246 ZFS_MODULE_PARAM(zfs, zfs_, scan_legacy, INT, ZMOD_RW,
5247 "Scrub using legacy non-sequential method");
5248
5249 ZFS_MODULE_PARAM(zfs, zfs_, scan_checkpoint_intval, UINT, ZMOD_RW,
5250 "Scan progress on-disk checkpointing interval");
5251
5252 ZFS_MODULE_PARAM(zfs, zfs_, scan_max_ext_gap, U64, ZMOD_RW,
5253 "Max gap in bytes between sequential scrub / resilver I/Os");
5254
5255 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_soft_fact, UINT, ZMOD_RW,
5256 "Fraction of hard limit used as soft limit");
5257
5258 ZFS_MODULE_PARAM(zfs, zfs_, scan_strict_mem_lim, INT, ZMOD_RW,
5259 "Tunable to attempt to reduce lock contention");
5260
5261 ZFS_MODULE_PARAM(zfs, zfs_, scan_fill_weight, UINT, ZMOD_RW,
5262 "Tunable to adjust bias towards more filled segments during scans");
5263
5264 ZFS_MODULE_PARAM(zfs, zfs_, scan_report_txgs, UINT, ZMOD_RW,
5265 "Tunable to report resilver performance over the last N txgs");
5266
5267 ZFS_MODULE_PARAM(zfs, zfs_, resilver_disable_defer, INT, ZMOD_RW,
5268 "Process all resilvers immediately");
5269
5270 ZFS_MODULE_PARAM(zfs, zfs_, scrub_error_blocks_per_txg, UINT, ZMOD_RW,
5271 "Error blocks to be scrubbed in one txg");
5272 /* END CSTYLED */
5273