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