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, 2014 by Delphix. All rights reserved.
24 */
25
26#include <sys/dsl_scan.h>
27#include <sys/dsl_pool.h>
28#include <sys/dsl_dataset.h>
29#include <sys/dsl_prop.h>
30#include <sys/dsl_dir.h>
31#include <sys/dsl_synctask.h>
32#include <sys/dnode.h>
33#include <sys/dmu_tx.h>
34#include <sys/dmu_objset.h>
35#include <sys/arc.h>
36#include <sys/zap.h>
37#include <sys/zio.h>
38#include <sys/zfs_context.h>
39#include <sys/fs/zfs.h>
40#include <sys/zfs_znode.h>
41#include <sys/spa_impl.h>
42#include <sys/vdev_impl.h>
43#include <sys/zil_impl.h>
44#include <sys/zio_checksum.h>
45#include <sys/ddt.h>
46#include <sys/sa.h>
47#include <sys/sa_impl.h>
48#include <sys/zfeature.h>
49#ifdef _KERNEL
50#include <sys/zfs_vfsops.h>
51#endif
52
53typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
54 const zbookmark_phys_t *);
55
56static scan_cb_t dsl_scan_scrub_cb;
57static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
58static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
59
60unsigned int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */
61unsigned int zfs_resilver_delay = 2; /* number of ticks to delay resilver */
62unsigned int zfs_scrub_delay = 4; /* number of ticks to delay scrub */
63unsigned int zfs_scan_idle = 50; /* idle window in clock ticks */
64
65unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
66unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
67unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
68 per txg */
69boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
70boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
71
72SYSCTL_DECL(_vfs_zfs);
73TUNABLE_INT("vfs.zfs.top_maxinflight", &zfs_top_maxinflight);
74SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RW,
75 &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
76TUNABLE_INT("vfs.zfs.resilver_delay", &zfs_resilver_delay);
77SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RW,
78 &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
79TUNABLE_INT("vfs.zfs.scrub_delay", &zfs_scrub_delay);
80SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RW,
81 &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
82TUNABLE_INT("vfs.zfs.scan_idle", &zfs_scan_idle);
83SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RW,
84 &zfs_scan_idle, 0, "Idle scan window in clock ticks");
85TUNABLE_INT("vfs.zfs.scan_min_time_ms", &zfs_scan_min_time_ms);
86SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RW,
87 &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
88TUNABLE_INT("vfs.zfs.free_min_time_ms", &zfs_free_min_time_ms);
89SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RW,
90 &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
91TUNABLE_INT("vfs.zfs.resilver_min_time_ms", &zfs_resilver_min_time_ms);
92SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RW,
93 &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
94TUNABLE_INT("vfs.zfs.no_scrub_io", &zfs_no_scrub_io);
95SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RW,
96 &zfs_no_scrub_io, 0, "Disable scrub I/O");
97TUNABLE_INT("vfs.zfs.no_scrub_prefetch", &zfs_no_scrub_prefetch);
98SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RW,
99 &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
100
101enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
102/* max number of blocks to free in a single TXG */
103uint64_t zfs_free_max_blocks = UINT64_MAX;
104SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
105 &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
106
107
108#define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
109 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
110 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
111
112extern int zfs_txg_timeout;
113
114/* the order has to match pool_scan_type */
115static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
116 NULL,
117 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
118 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
119};
120
121int
122dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
123{
124 int err;
125 dsl_scan_t *scn;
126 spa_t *spa = dp->dp_spa;
127 uint64_t f;
128
129 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
130 scn->scn_dp = dp;
131
132 /*
133 * It's possible that we're resuming a scan after a reboot so
134 * make sure that the scan_async_destroying flag is initialized
135 * appropriately.
136 */
137 ASSERT(!scn->scn_async_destroying);
138 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
139 SPA_FEATURE_ASYNC_DESTROY);
140
141 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
142 "scrub_func", sizeof (uint64_t), 1, &f);
143 if (err == 0) {
144 /*
145 * There was an old-style scrub in progress. Restart a
146 * new-style scrub from the beginning.
147 */
148 scn->scn_restart_txg = txg;
149 zfs_dbgmsg("old-style scrub was in progress; "
150 "restarting new-style scrub in txg %llu",
151 scn->scn_restart_txg);
152
153 /*
154 * Load the queue obj from the old location so that it
155 * can be freed by dsl_scan_done().
156 */
157 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
158 "scrub_queue", sizeof (uint64_t), 1,
159 &scn->scn_phys.scn_queue_obj);
160 } else {
161 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
162 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
163 &scn->scn_phys);
164 if (err == ENOENT)
165 return (0);
166 else if (err)
167 return (err);
168
169 if (scn->scn_phys.scn_state == DSS_SCANNING &&
170 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
171 /*
172 * A new-type scrub was in progress on an old
173 * pool, and the pool was accessed by old
174 * software. Restart from the beginning, since
175 * the old software may have changed the pool in
176 * the meantime.
177 */
178 scn->scn_restart_txg = txg;
179 zfs_dbgmsg("new-style scrub was modified "
180 "by old software; restarting in txg %llu",
181 scn->scn_restart_txg);
182 }
183 }
184
185 spa_scan_stat_init(spa);
186 return (0);
187}
188
189void
190dsl_scan_fini(dsl_pool_t *dp)
191{
192 if (dp->dp_scan) {
193 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
194 dp->dp_scan = NULL;
195 }
196}
197
198/* ARGSUSED */
199static int
200dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
201{
202 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
203
204 if (scn->scn_phys.scn_state == DSS_SCANNING)
205 return (SET_ERROR(EBUSY));
206
207 return (0);
208}
209
210static void
211dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
212{
213 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
214 pool_scan_func_t *funcp = arg;
215 dmu_object_type_t ot = 0;
216 dsl_pool_t *dp = scn->scn_dp;
217 spa_t *spa = dp->dp_spa;
218
219 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
220 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
221 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
222 scn->scn_phys.scn_func = *funcp;
223 scn->scn_phys.scn_state = DSS_SCANNING;
224 scn->scn_phys.scn_min_txg = 0;
225 scn->scn_phys.scn_max_txg = tx->tx_txg;
226 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
227 scn->scn_phys.scn_start_time = gethrestime_sec();
228 scn->scn_phys.scn_errors = 0;
229 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
230 scn->scn_restart_txg = 0;
231 scn->scn_done_txg = 0;
232 spa_scan_stat_init(spa);
233
234 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
235 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
236
237 /* rewrite all disk labels */
238 vdev_config_dirty(spa->spa_root_vdev);
239
240 if (vdev_resilver_needed(spa->spa_root_vdev,
241 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
242 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
243 } else {
244 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
245 }
246
247 spa->spa_scrub_started = B_TRUE;
248 /*
249 * If this is an incremental scrub, limit the DDT scrub phase
250 * to just the auto-ditto class (for correctness); the rest
251 * of the scrub should go faster using top-down pruning.
252 */
253 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
254 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
255
256 }
257
258 /* back to the generic stuff */
259
260 if (dp->dp_blkstats == NULL) {
261 dp->dp_blkstats =
262 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
263 }
264 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
265
266 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
267 ot = DMU_OT_ZAP_OTHER;
268
269 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
270 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
271
272 dsl_scan_sync_state(scn, tx);
273
274 spa_history_log_internal(spa, "scan setup", tx,
275 "func=%u mintxg=%llu maxtxg=%llu",
276 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
277}
278
279/* ARGSUSED */
280static void
281dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
282{
283 static const char *old_names[] = {
284 "scrub_bookmark",
285 "scrub_ddt_bookmark",
286 "scrub_ddt_class_max",
287 "scrub_queue",
288 "scrub_min_txg",
289 "scrub_max_txg",
290 "scrub_func",
291 "scrub_errors",
292 NULL
293 };
294
295 dsl_pool_t *dp = scn->scn_dp;
296 spa_t *spa = dp->dp_spa;
297 int i;
298
299 /* Remove any remnants of an old-style scrub. */
300 for (i = 0; old_names[i]; i++) {
301 (void) zap_remove(dp->dp_meta_objset,
302 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
303 }
304
305 if (scn->scn_phys.scn_queue_obj != 0) {
306 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
307 scn->scn_phys.scn_queue_obj, tx));
308 scn->scn_phys.scn_queue_obj = 0;
309 }
310
311 /*
312 * If we were "restarted" from a stopped state, don't bother
313 * with anything else.
314 */
315 if (scn->scn_phys.scn_state != DSS_SCANNING)
316 return;
317
318 if (complete)
319 scn->scn_phys.scn_state = DSS_FINISHED;
320 else
321 scn->scn_phys.scn_state = DSS_CANCELED;
322
323 spa_history_log_internal(spa, "scan done", tx,
324 "complete=%u", complete);
325
326 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
327 mutex_enter(&spa->spa_scrub_lock);
328 while (spa->spa_scrub_inflight > 0) {
329 cv_wait(&spa->spa_scrub_io_cv,
330 &spa->spa_scrub_lock);
331 }
332 mutex_exit(&spa->spa_scrub_lock);
333 spa->spa_scrub_started = B_FALSE;
334 spa->spa_scrub_active = B_FALSE;
335
336 /*
337 * If the scrub/resilver completed, update all DTLs to
338 * reflect this. Whether it succeeded or not, vacate
339 * all temporary scrub DTLs.
340 */
341 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
342 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
343 if (complete) {
344 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
345 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
346 }
347 spa_errlog_rotate(spa);
348
349 /*
350 * We may have finished replacing a device.
351 * Let the async thread assess this and handle the detach.
352 */
353 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
354 }
355
356 scn->scn_phys.scn_end_time = gethrestime_sec();
357}
358
359/* ARGSUSED */
360static int
361dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
362{
363 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
364
365 if (scn->scn_phys.scn_state != DSS_SCANNING)
366 return (SET_ERROR(ENOENT));
367 return (0);
368}
369
370/* ARGSUSED */
371static void
372dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
373{
374 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
375
376 dsl_scan_done(scn, B_FALSE, tx);
377 dsl_scan_sync_state(scn, tx);
378}
379
380int
381dsl_scan_cancel(dsl_pool_t *dp)
382{
383 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
384 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
385}
386
387static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
388 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
389 dmu_objset_type_t ostype, dmu_tx_t *tx);
390static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
391 dmu_objset_type_t ostype,
392 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
393
394void
395dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
396{
397 zio_free(dp->dp_spa, txg, bp);
398}
399
400void
401dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
402{
403 ASSERT(dsl_pool_sync_context(dp));
404 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
405 pio->io_flags));
406}
407
408static uint64_t
409dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
410{
411 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
412 if (dsl_dataset_is_snapshot(ds))
413 return (MIN(smt, ds->ds_phys->ds_creation_txg));
414 return (smt);
415}
416
417static void
418dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
419{
420 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
421 DMU_POOL_DIRECTORY_OBJECT,
422 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
423 &scn->scn_phys, tx));
424}
425
| 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, 2014 by Delphix. All rights reserved.
24 */
25
26#include <sys/dsl_scan.h>
27#include <sys/dsl_pool.h>
28#include <sys/dsl_dataset.h>
29#include <sys/dsl_prop.h>
30#include <sys/dsl_dir.h>
31#include <sys/dsl_synctask.h>
32#include <sys/dnode.h>
33#include <sys/dmu_tx.h>
34#include <sys/dmu_objset.h>
35#include <sys/arc.h>
36#include <sys/zap.h>
37#include <sys/zio.h>
38#include <sys/zfs_context.h>
39#include <sys/fs/zfs.h>
40#include <sys/zfs_znode.h>
41#include <sys/spa_impl.h>
42#include <sys/vdev_impl.h>
43#include <sys/zil_impl.h>
44#include <sys/zio_checksum.h>
45#include <sys/ddt.h>
46#include <sys/sa.h>
47#include <sys/sa_impl.h>
48#include <sys/zfeature.h>
49#ifdef _KERNEL
50#include <sys/zfs_vfsops.h>
51#endif
52
53typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
54 const zbookmark_phys_t *);
55
56static scan_cb_t dsl_scan_scrub_cb;
57static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
58static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
59
60unsigned int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */
61unsigned int zfs_resilver_delay = 2; /* number of ticks to delay resilver */
62unsigned int zfs_scrub_delay = 4; /* number of ticks to delay scrub */
63unsigned int zfs_scan_idle = 50; /* idle window in clock ticks */
64
65unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
66unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
67unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
68 per txg */
69boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
70boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
71
72SYSCTL_DECL(_vfs_zfs);
73TUNABLE_INT("vfs.zfs.top_maxinflight", &zfs_top_maxinflight);
74SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RW,
75 &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
76TUNABLE_INT("vfs.zfs.resilver_delay", &zfs_resilver_delay);
77SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RW,
78 &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
79TUNABLE_INT("vfs.zfs.scrub_delay", &zfs_scrub_delay);
80SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RW,
81 &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
82TUNABLE_INT("vfs.zfs.scan_idle", &zfs_scan_idle);
83SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RW,
84 &zfs_scan_idle, 0, "Idle scan window in clock ticks");
85TUNABLE_INT("vfs.zfs.scan_min_time_ms", &zfs_scan_min_time_ms);
86SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RW,
87 &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
88TUNABLE_INT("vfs.zfs.free_min_time_ms", &zfs_free_min_time_ms);
89SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RW,
90 &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
91TUNABLE_INT("vfs.zfs.resilver_min_time_ms", &zfs_resilver_min_time_ms);
92SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RW,
93 &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
94TUNABLE_INT("vfs.zfs.no_scrub_io", &zfs_no_scrub_io);
95SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RW,
96 &zfs_no_scrub_io, 0, "Disable scrub I/O");
97TUNABLE_INT("vfs.zfs.no_scrub_prefetch", &zfs_no_scrub_prefetch);
98SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RW,
99 &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
100
101enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
102/* max number of blocks to free in a single TXG */
103uint64_t zfs_free_max_blocks = UINT64_MAX;
104SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
105 &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
106
107
108#define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
109 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
110 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
111
112extern int zfs_txg_timeout;
113
114/* the order has to match pool_scan_type */
115static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
116 NULL,
117 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */
118 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */
119};
120
121int
122dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
123{
124 int err;
125 dsl_scan_t *scn;
126 spa_t *spa = dp->dp_spa;
127 uint64_t f;
128
129 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
130 scn->scn_dp = dp;
131
132 /*
133 * It's possible that we're resuming a scan after a reboot so
134 * make sure that the scan_async_destroying flag is initialized
135 * appropriately.
136 */
137 ASSERT(!scn->scn_async_destroying);
138 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
139 SPA_FEATURE_ASYNC_DESTROY);
140
141 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
142 "scrub_func", sizeof (uint64_t), 1, &f);
143 if (err == 0) {
144 /*
145 * There was an old-style scrub in progress. Restart a
146 * new-style scrub from the beginning.
147 */
148 scn->scn_restart_txg = txg;
149 zfs_dbgmsg("old-style scrub was in progress; "
150 "restarting new-style scrub in txg %llu",
151 scn->scn_restart_txg);
152
153 /*
154 * Load the queue obj from the old location so that it
155 * can be freed by dsl_scan_done().
156 */
157 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
158 "scrub_queue", sizeof (uint64_t), 1,
159 &scn->scn_phys.scn_queue_obj);
160 } else {
161 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
162 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
163 &scn->scn_phys);
164 if (err == ENOENT)
165 return (0);
166 else if (err)
167 return (err);
168
169 if (scn->scn_phys.scn_state == DSS_SCANNING &&
170 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
171 /*
172 * A new-type scrub was in progress on an old
173 * pool, and the pool was accessed by old
174 * software. Restart from the beginning, since
175 * the old software may have changed the pool in
176 * the meantime.
177 */
178 scn->scn_restart_txg = txg;
179 zfs_dbgmsg("new-style scrub was modified "
180 "by old software; restarting in txg %llu",
181 scn->scn_restart_txg);
182 }
183 }
184
185 spa_scan_stat_init(spa);
186 return (0);
187}
188
189void
190dsl_scan_fini(dsl_pool_t *dp)
191{
192 if (dp->dp_scan) {
193 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
194 dp->dp_scan = NULL;
195 }
196}
197
198/* ARGSUSED */
199static int
200dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
201{
202 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
203
204 if (scn->scn_phys.scn_state == DSS_SCANNING)
205 return (SET_ERROR(EBUSY));
206
207 return (0);
208}
209
210static void
211dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
212{
213 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
214 pool_scan_func_t *funcp = arg;
215 dmu_object_type_t ot = 0;
216 dsl_pool_t *dp = scn->scn_dp;
217 spa_t *spa = dp->dp_spa;
218
219 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
220 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
221 bzero(&scn->scn_phys, sizeof (scn->scn_phys));
222 scn->scn_phys.scn_func = *funcp;
223 scn->scn_phys.scn_state = DSS_SCANNING;
224 scn->scn_phys.scn_min_txg = 0;
225 scn->scn_phys.scn_max_txg = tx->tx_txg;
226 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
227 scn->scn_phys.scn_start_time = gethrestime_sec();
228 scn->scn_phys.scn_errors = 0;
229 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
230 scn->scn_restart_txg = 0;
231 scn->scn_done_txg = 0;
232 spa_scan_stat_init(spa);
233
234 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
235 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
236
237 /* rewrite all disk labels */
238 vdev_config_dirty(spa->spa_root_vdev);
239
240 if (vdev_resilver_needed(spa->spa_root_vdev,
241 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
242 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
243 } else {
244 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
245 }
246
247 spa->spa_scrub_started = B_TRUE;
248 /*
249 * If this is an incremental scrub, limit the DDT scrub phase
250 * to just the auto-ditto class (for correctness); the rest
251 * of the scrub should go faster using top-down pruning.
252 */
253 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
254 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
255
256 }
257
258 /* back to the generic stuff */
259
260 if (dp->dp_blkstats == NULL) {
261 dp->dp_blkstats =
262 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
263 }
264 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
265
266 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
267 ot = DMU_OT_ZAP_OTHER;
268
269 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
270 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
271
272 dsl_scan_sync_state(scn, tx);
273
274 spa_history_log_internal(spa, "scan setup", tx,
275 "func=%u mintxg=%llu maxtxg=%llu",
276 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
277}
278
279/* ARGSUSED */
280static void
281dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
282{
283 static const char *old_names[] = {
284 "scrub_bookmark",
285 "scrub_ddt_bookmark",
286 "scrub_ddt_class_max",
287 "scrub_queue",
288 "scrub_min_txg",
289 "scrub_max_txg",
290 "scrub_func",
291 "scrub_errors",
292 NULL
293 };
294
295 dsl_pool_t *dp = scn->scn_dp;
296 spa_t *spa = dp->dp_spa;
297 int i;
298
299 /* Remove any remnants of an old-style scrub. */
300 for (i = 0; old_names[i]; i++) {
301 (void) zap_remove(dp->dp_meta_objset,
302 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
303 }
304
305 if (scn->scn_phys.scn_queue_obj != 0) {
306 VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
307 scn->scn_phys.scn_queue_obj, tx));
308 scn->scn_phys.scn_queue_obj = 0;
309 }
310
311 /*
312 * If we were "restarted" from a stopped state, don't bother
313 * with anything else.
314 */
315 if (scn->scn_phys.scn_state != DSS_SCANNING)
316 return;
317
318 if (complete)
319 scn->scn_phys.scn_state = DSS_FINISHED;
320 else
321 scn->scn_phys.scn_state = DSS_CANCELED;
322
323 spa_history_log_internal(spa, "scan done", tx,
324 "complete=%u", complete);
325
326 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
327 mutex_enter(&spa->spa_scrub_lock);
328 while (spa->spa_scrub_inflight > 0) {
329 cv_wait(&spa->spa_scrub_io_cv,
330 &spa->spa_scrub_lock);
331 }
332 mutex_exit(&spa->spa_scrub_lock);
333 spa->spa_scrub_started = B_FALSE;
334 spa->spa_scrub_active = B_FALSE;
335
336 /*
337 * If the scrub/resilver completed, update all DTLs to
338 * reflect this. Whether it succeeded or not, vacate
339 * all temporary scrub DTLs.
340 */
341 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
342 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
343 if (complete) {
344 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
345 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
346 }
347 spa_errlog_rotate(spa);
348
349 /*
350 * We may have finished replacing a device.
351 * Let the async thread assess this and handle the detach.
352 */
353 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
354 }
355
356 scn->scn_phys.scn_end_time = gethrestime_sec();
357}
358
359/* ARGSUSED */
360static int
361dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
362{
363 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
364
365 if (scn->scn_phys.scn_state != DSS_SCANNING)
366 return (SET_ERROR(ENOENT));
367 return (0);
368}
369
370/* ARGSUSED */
371static void
372dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
373{
374 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
375
376 dsl_scan_done(scn, B_FALSE, tx);
377 dsl_scan_sync_state(scn, tx);
378}
379
380int
381dsl_scan_cancel(dsl_pool_t *dp)
382{
383 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
384 dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
385}
386
387static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
388 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
389 dmu_objset_type_t ostype, dmu_tx_t *tx);
390static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
391 dmu_objset_type_t ostype,
392 dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
393
394void
395dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
396{
397 zio_free(dp->dp_spa, txg, bp);
398}
399
400void
401dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
402{
403 ASSERT(dsl_pool_sync_context(dp));
404 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
405 pio->io_flags));
406}
407
408static uint64_t
409dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
410{
411 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
412 if (dsl_dataset_is_snapshot(ds))
413 return (MIN(smt, ds->ds_phys->ds_creation_txg));
414 return (smt);
415}
416
417static void
418dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
419{
420 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
421 DMU_POOL_DIRECTORY_OBJECT,
422 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
423 &scn->scn_phys, tx));
424}
425
|
452 spa_shutting_down(scn->scn_dp->dp_spa)) {
453 if (zb) {
454 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
455 (longlong_t)zb->zb_objset,
456 (longlong_t)zb->zb_object,
457 (longlong_t)zb->zb_level,
458 (longlong_t)zb->zb_blkid);
459 scn->scn_phys.scn_bookmark = *zb;
460 }
461 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
462 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
463 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
464 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
465 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
466 scn->scn_pausing = B_TRUE;
467 return (B_TRUE);
468 }
469 return (B_FALSE);
470}
471
472typedef struct zil_scan_arg {
473 dsl_pool_t *zsa_dp;
474 zil_header_t *zsa_zh;
475} zil_scan_arg_t;
476
477/* ARGSUSED */
478static int
479dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
480{
481 zil_scan_arg_t *zsa = arg;
482 dsl_pool_t *dp = zsa->zsa_dp;
483 dsl_scan_t *scn = dp->dp_scan;
484 zil_header_t *zh = zsa->zsa_zh;
485 zbookmark_phys_t zb;
486
487 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
488 return (0);
489
490 /*
491 * One block ("stubby") can be allocated a long time ago; we
492 * want to visit that one because it has been allocated
493 * (on-disk) even if it hasn't been claimed (even though for
494 * scrub there's nothing to do to it).
495 */
496 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
497 return (0);
498
499 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
500 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
501
502 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
503 return (0);
504}
505
506/* ARGSUSED */
507static int
508dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
509{
510 if (lrc->lrc_txtype == TX_WRITE) {
511 zil_scan_arg_t *zsa = arg;
512 dsl_pool_t *dp = zsa->zsa_dp;
513 dsl_scan_t *scn = dp->dp_scan;
514 zil_header_t *zh = zsa->zsa_zh;
515 lr_write_t *lr = (lr_write_t *)lrc;
516 blkptr_t *bp = &lr->lr_blkptr;
517 zbookmark_phys_t zb;
518
519 if (BP_IS_HOLE(bp) ||
520 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
521 return (0);
522
523 /*
524 * birth can be < claim_txg if this record's txg is
525 * already txg sync'ed (but this log block contains
526 * other records that are not synced)
527 */
528 if (claim_txg == 0 || bp->blk_birth < claim_txg)
529 return (0);
530
531 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
532 lr->lr_foid, ZB_ZIL_LEVEL,
533 lr->lr_offset / BP_GET_LSIZE(bp));
534
535 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
536 }
537 return (0);
538}
539
540static void
541dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
542{
543 uint64_t claim_txg = zh->zh_claim_txg;
544 zil_scan_arg_t zsa = { dp, zh };
545 zilog_t *zilog;
546
547 /*
548 * We only want to visit blocks that have been claimed but not yet
549 * replayed (or, in read-only mode, blocks that *would* be claimed).
550 */
551 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
552 return;
553
554 zilog = zil_alloc(dp->dp_meta_objset, zh);
555
556 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
557 claim_txg);
558
559 zil_free(zilog);
560}
561
562/* ARGSUSED */
563static void
564dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
565 uint64_t objset, uint64_t object, uint64_t blkid)
566{
567 zbookmark_phys_t czb;
568 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
569
570 if (zfs_no_scrub_prefetch)
571 return;
572
573 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
574 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
575 return;
576
577 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
578
579 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
580 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
581 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
582}
583
584static boolean_t
585dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
586 const zbookmark_phys_t *zb)
587{
588 /*
589 * We never skip over user/group accounting objects (obj<0)
590 */
591 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
592 (int64_t)zb->zb_object >= 0) {
593 /*
594 * If we already visited this bp & everything below (in
595 * a prior txg sync), don't bother doing it again.
596 */
597 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
598 return (B_TRUE);
599
600 /*
601 * If we found the block we're trying to resume from, or
602 * we went past it to a different object, zero it out to
603 * indicate that it's OK to start checking for pausing
604 * again.
605 */
606 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
607 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
608 dprintf("resuming at %llx/%llx/%llx/%llx\n",
609 (longlong_t)zb->zb_objset,
610 (longlong_t)zb->zb_object,
611 (longlong_t)zb->zb_level,
612 (longlong_t)zb->zb_blkid);
613 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
614 }
615 }
616 return (B_FALSE);
617}
618
619/*
620 * Return nonzero on i/o error.
621 * Return new buf to write out in *bufp.
622 */
623static int
624dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
625 dnode_phys_t *dnp, const blkptr_t *bp,
626 const zbookmark_phys_t *zb, dmu_tx_t *tx)
627{
628 dsl_pool_t *dp = scn->scn_dp;
629 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
630 int err;
631
632 if (BP_GET_LEVEL(bp) > 0) {
633 uint32_t flags = ARC_WAIT;
634 int i;
635 blkptr_t *cbp;
636 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
637 arc_buf_t *buf;
638
639 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
640 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
641 if (err) {
642 scn->scn_phys.scn_errors++;
643 return (err);
644 }
645 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
646 dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
647 zb->zb_object, zb->zb_blkid * epb + i);
648 }
649 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
650 zbookmark_phys_t czb;
651
652 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
653 zb->zb_level - 1,
654 zb->zb_blkid * epb + i);
655 dsl_scan_visitbp(cbp, &czb, dnp,
656 ds, scn, ostype, tx);
657 }
658 (void) arc_buf_remove_ref(buf, &buf);
659 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
660 uint32_t flags = ARC_WAIT;
661 dnode_phys_t *cdnp;
662 int i, j;
663 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
664 arc_buf_t *buf;
665
666 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
667 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
668 if (err) {
669 scn->scn_phys.scn_errors++;
670 return (err);
671 }
672 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
673 for (j = 0; j < cdnp->dn_nblkptr; j++) {
674 blkptr_t *cbp = &cdnp->dn_blkptr[j];
675 dsl_scan_prefetch(scn, buf, cbp,
676 zb->zb_objset, zb->zb_blkid * epb + i, j);
677 }
678 }
679 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
680 dsl_scan_visitdnode(scn, ds, ostype,
681 cdnp, zb->zb_blkid * epb + i, tx);
682 }
683
684 (void) arc_buf_remove_ref(buf, &buf);
685 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
686 uint32_t flags = ARC_WAIT;
687 objset_phys_t *osp;
688 arc_buf_t *buf;
689
690 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
691 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
692 if (err) {
693 scn->scn_phys.scn_errors++;
694 return (err);
695 }
696
697 osp = buf->b_data;
698
699 dsl_scan_visitdnode(scn, ds, osp->os_type,
700 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
701
702 if (OBJSET_BUF_HAS_USERUSED(buf)) {
703 /*
704 * We also always visit user/group accounting
705 * objects, and never skip them, even if we are
706 * pausing. This is necessary so that the space
707 * deltas from this txg get integrated.
708 */
709 dsl_scan_visitdnode(scn, ds, osp->os_type,
710 &osp->os_groupused_dnode,
711 DMU_GROUPUSED_OBJECT, tx);
712 dsl_scan_visitdnode(scn, ds, osp->os_type,
713 &osp->os_userused_dnode,
714 DMU_USERUSED_OBJECT, tx);
715 }
716 (void) arc_buf_remove_ref(buf, &buf);
717 }
718
719 return (0);
720}
721
722static void
723dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
724 dmu_objset_type_t ostype, dnode_phys_t *dnp,
725 uint64_t object, dmu_tx_t *tx)
726{
727 int j;
728
729 for (j = 0; j < dnp->dn_nblkptr; j++) {
730 zbookmark_phys_t czb;
731
732 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
733 dnp->dn_nlevels - 1, j);
734 dsl_scan_visitbp(&dnp->dn_blkptr[j],
735 &czb, dnp, ds, scn, ostype, tx);
736 }
737
738 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
739 zbookmark_phys_t czb;
740 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
741 0, DMU_SPILL_BLKID);
742 dsl_scan_visitbp(&dnp->dn_spill,
743 &czb, dnp, ds, scn, ostype, tx);
744 }
745}
746
747/*
748 * The arguments are in this order because mdb can only print the
749 * first 5; we want them to be useful.
750 */
751static void
752dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
753 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
754 dmu_objset_type_t ostype, dmu_tx_t *tx)
755{
756 dsl_pool_t *dp = scn->scn_dp;
757 arc_buf_t *buf = NULL;
758 blkptr_t bp_toread = *bp;
759
760 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
761
762 if (dsl_scan_check_pause(scn, zb))
763 return;
764
765 if (dsl_scan_check_resume(scn, dnp, zb))
766 return;
767
768 if (BP_IS_HOLE(bp))
769 return;
770
771 scn->scn_visited_this_txg++;
772
773 dprintf_bp(bp,
774 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
775 ds, ds ? ds->ds_object : 0,
776 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
777 bp);
778
779 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
780 return;
781
782 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
783 return;
784
785 /*
786 * If dsl_scan_ddt() has aready visited this block, it will have
787 * already done any translations or scrubbing, so don't call the
788 * callback again.
789 */
790 if (ddt_class_contains(dp->dp_spa,
791 scn->scn_phys.scn_ddt_class_max, bp)) {
792 ASSERT(buf == NULL);
793 return;
794 }
795
796 /*
797 * If this block is from the future (after cur_max_txg), then we
798 * are doing this on behalf of a deleted snapshot, and we will
799 * revisit the future block on the next pass of this dataset.
800 * Don't scan it now unless we need to because something
801 * under it was modified.
802 */
803 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
804 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
805 }
806}
807
808static void
809dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
810 dmu_tx_t *tx)
811{
812 zbookmark_phys_t zb;
813
814 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
815 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
816 dsl_scan_visitbp(bp, &zb, NULL,
817 ds, scn, DMU_OST_NONE, tx);
818
819 dprintf_ds(ds, "finished scan%s", "");
820}
821
822void
823dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
824{
825 dsl_pool_t *dp = ds->ds_dir->dd_pool;
826 dsl_scan_t *scn = dp->dp_scan;
827 uint64_t mintxg;
828
829 if (scn->scn_phys.scn_state != DSS_SCANNING)
830 return;
831
832 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
833 if (dsl_dataset_is_snapshot(ds)) {
834 /* Note, scn_cur_{min,max}_txg stays the same. */
835 scn->scn_phys.scn_bookmark.zb_objset =
836 ds->ds_phys->ds_next_snap_obj;
837 zfs_dbgmsg("destroying ds %llu; currently traversing; "
838 "reset zb_objset to %llu",
839 (u_longlong_t)ds->ds_object,
840 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
841 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
842 } else {
843 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
844 ZB_DESTROYED_OBJSET, 0, 0, 0);
845 zfs_dbgmsg("destroying ds %llu; currently traversing; "
846 "reset bookmark to -1,0,0,0",
847 (u_longlong_t)ds->ds_object);
848 }
849 } else if (zap_lookup_int_key(dp->dp_meta_objset,
850 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
851 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
852 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
853 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
854 if (dsl_dataset_is_snapshot(ds)) {
855 /*
856 * We keep the same mintxg; it could be >
857 * ds_creation_txg if the previous snapshot was
858 * deleted too.
859 */
860 VERIFY(zap_add_int_key(dp->dp_meta_objset,
861 scn->scn_phys.scn_queue_obj,
862 ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
863 zfs_dbgmsg("destroying ds %llu; in queue; "
864 "replacing with %llu",
865 (u_longlong_t)ds->ds_object,
866 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
867 } else {
868 zfs_dbgmsg("destroying ds %llu; in queue; removing",
869 (u_longlong_t)ds->ds_object);
870 }
871 } else {
872 zfs_dbgmsg("destroying ds %llu; ignoring",
873 (u_longlong_t)ds->ds_object);
874 }
875
876 /*
877 * dsl_scan_sync() should be called after this, and should sync
878 * out our changed state, but just to be safe, do it here.
879 */
880 dsl_scan_sync_state(scn, tx);
881}
882
883void
884dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
885{
886 dsl_pool_t *dp = ds->ds_dir->dd_pool;
887 dsl_scan_t *scn = dp->dp_scan;
888 uint64_t mintxg;
889
890 if (scn->scn_phys.scn_state != DSS_SCANNING)
891 return;
892
893 ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
894
895 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
896 scn->scn_phys.scn_bookmark.zb_objset =
897 ds->ds_phys->ds_prev_snap_obj;
898 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
899 "reset zb_objset to %llu",
900 (u_longlong_t)ds->ds_object,
901 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
902 } else if (zap_lookup_int_key(dp->dp_meta_objset,
903 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
904 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
905 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
906 VERIFY(zap_add_int_key(dp->dp_meta_objset,
907 scn->scn_phys.scn_queue_obj,
908 ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
909 zfs_dbgmsg("snapshotting ds %llu; in queue; "
910 "replacing with %llu",
911 (u_longlong_t)ds->ds_object,
912 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
913 }
914 dsl_scan_sync_state(scn, tx);
915}
916
917void
918dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
919{
920 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
921 dsl_scan_t *scn = dp->dp_scan;
922 uint64_t mintxg;
923
924 if (scn->scn_phys.scn_state != DSS_SCANNING)
925 return;
926
927 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
928 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
929 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
930 "reset zb_objset to %llu",
931 (u_longlong_t)ds1->ds_object,
932 (u_longlong_t)ds2->ds_object);
933 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
934 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
935 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
936 "reset zb_objset to %llu",
937 (u_longlong_t)ds2->ds_object,
938 (u_longlong_t)ds1->ds_object);
939 }
940
941 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
942 ds1->ds_object, &mintxg) == 0) {
943 int err;
944
945 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
946 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
947 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
948 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
949 err = zap_add_int_key(dp->dp_meta_objset,
950 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
951 VERIFY(err == 0 || err == EEXIST);
952 if (err == EEXIST) {
953 /* Both were there to begin with */
954 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
955 scn->scn_phys.scn_queue_obj,
956 ds1->ds_object, mintxg, tx));
957 }
958 zfs_dbgmsg("clone_swap ds %llu; in queue; "
959 "replacing with %llu",
960 (u_longlong_t)ds1->ds_object,
961 (u_longlong_t)ds2->ds_object);
962 } else if (zap_lookup_int_key(dp->dp_meta_objset,
963 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
964 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
965 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
966 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
967 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
968 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
969 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
970 zfs_dbgmsg("clone_swap ds %llu; in queue; "
971 "replacing with %llu",
972 (u_longlong_t)ds2->ds_object,
973 (u_longlong_t)ds1->ds_object);
974 }
975
976 dsl_scan_sync_state(scn, tx);
977}
978
979struct enqueue_clones_arg {
980 dmu_tx_t *tx;
981 uint64_t originobj;
982};
983
984/* ARGSUSED */
985static int
986enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
987{
988 struct enqueue_clones_arg *eca = arg;
989 dsl_dataset_t *ds;
990 int err;
991 dsl_scan_t *scn = dp->dp_scan;
992
993 if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
994 return (0);
995
996 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
997 if (err)
998 return (err);
999
1000 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1001 dsl_dataset_t *prev;
1002 err = dsl_dataset_hold_obj(dp,
1003 ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1004
1005 dsl_dataset_rele(ds, FTAG);
1006 if (err)
1007 return (err);
1008 ds = prev;
1009 }
1010 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1011 scn->scn_phys.scn_queue_obj, ds->ds_object,
1012 ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1013 dsl_dataset_rele(ds, FTAG);
1014 return (0);
1015}
1016
1017static void
1018dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1019{
1020 dsl_pool_t *dp = scn->scn_dp;
1021 dsl_dataset_t *ds;
1022 objset_t *os;
1023
1024 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1025
1026 if (dmu_objset_from_ds(ds, &os))
1027 goto out;
1028
1029 /*
1030 * Only the ZIL in the head (non-snapshot) is valid. Even though
1031 * snapshots can have ZIL block pointers (which may be the same
1032 * BP as in the head), they must be ignored. So we traverse the
1033 * ZIL here, rather than in scan_recurse(), because the regular
1034 * snapshot block-sharing rules don't apply to it.
1035 */
1036 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1037 dsl_scan_zil(dp, &os->os_zil_header);
1038
1039 /*
1040 * Iterate over the bps in this ds.
1041 */
1042 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1043 dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1044
1045 char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1046 dsl_dataset_name(ds, dsname);
1047 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1048 "pausing=%u",
1049 (longlong_t)dsobj, dsname,
1050 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1051 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1052 (int)scn->scn_pausing);
1053 kmem_free(dsname, ZFS_MAXNAMELEN);
1054
1055 if (scn->scn_pausing)
1056 goto out;
1057
1058 /*
1059 * We've finished this pass over this dataset.
1060 */
1061
1062 /*
1063 * If we did not completely visit this dataset, do another pass.
1064 */
1065 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1066 zfs_dbgmsg("incomplete pass; visiting again");
1067 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1068 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1069 scn->scn_phys.scn_queue_obj, ds->ds_object,
1070 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1071 goto out;
1072 }
1073
1074 /*
1075 * Add descendent datasets to work queue.
1076 */
1077 if (ds->ds_phys->ds_next_snap_obj != 0) {
1078 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1079 scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1080 ds->ds_phys->ds_creation_txg, tx) == 0);
1081 }
1082 if (ds->ds_phys->ds_num_children > 1) {
1083 boolean_t usenext = B_FALSE;
1084 if (ds->ds_phys->ds_next_clones_obj != 0) {
1085 uint64_t count;
1086 /*
1087 * A bug in a previous version of the code could
1088 * cause upgrade_clones_cb() to not set
1089 * ds_next_snap_obj when it should, leading to a
1090 * missing entry. Therefore we can only use the
1091 * next_clones_obj when its count is correct.
1092 */
1093 int err = zap_count(dp->dp_meta_objset,
1094 ds->ds_phys->ds_next_clones_obj, &count);
1095 if (err == 0 &&
1096 count == ds->ds_phys->ds_num_children - 1)
1097 usenext = B_TRUE;
1098 }
1099
1100 if (usenext) {
1101 VERIFY0(zap_join_key(dp->dp_meta_objset,
1102 ds->ds_phys->ds_next_clones_obj,
1103 scn->scn_phys.scn_queue_obj,
1104 ds->ds_phys->ds_creation_txg, tx));
1105 } else {
1106 struct enqueue_clones_arg eca;
1107 eca.tx = tx;
1108 eca.originobj = ds->ds_object;
1109
1110 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1111 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1112 }
1113 }
1114
1115out:
1116 dsl_dataset_rele(ds, FTAG);
1117}
1118
1119/* ARGSUSED */
1120static int
1121enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1122{
1123 dmu_tx_t *tx = arg;
1124 dsl_dataset_t *ds;
1125 int err;
1126 dsl_scan_t *scn = dp->dp_scan;
1127
1128 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1129 if (err)
1130 return (err);
1131
1132 while (ds->ds_phys->ds_prev_snap_obj != 0) {
1133 dsl_dataset_t *prev;
1134 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1135 FTAG, &prev);
1136 if (err) {
1137 dsl_dataset_rele(ds, FTAG);
1138 return (err);
1139 }
1140
1141 /*
1142 * If this is a clone, we don't need to worry about it for now.
1143 */
1144 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1145 dsl_dataset_rele(ds, FTAG);
1146 dsl_dataset_rele(prev, FTAG);
1147 return (0);
1148 }
1149 dsl_dataset_rele(ds, FTAG);
1150 ds = prev;
1151 }
1152
1153 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1154 ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1155 dsl_dataset_rele(ds, FTAG);
1156 return (0);
1157}
1158
1159/*
1160 * Scrub/dedup interaction.
1161 *
1162 * If there are N references to a deduped block, we don't want to scrub it
1163 * N times -- ideally, we should scrub it exactly once.
1164 *
1165 * We leverage the fact that the dde's replication class (enum ddt_class)
1166 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1167 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1168 *
1169 * To prevent excess scrubbing, the scrub begins by walking the DDT
1170 * to find all blocks with refcnt > 1, and scrubs each of these once.
1171 * Since there are two replication classes which contain blocks with
1172 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1173 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1174 *
1175 * There would be nothing more to say if a block's refcnt couldn't change
1176 * during a scrub, but of course it can so we must account for changes
1177 * in a block's replication class.
1178 *
1179 * Here's an example of what can occur:
1180 *
1181 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1182 * when visited during the top-down scrub phase, it will be scrubbed twice.
1183 * This negates our scrub optimization, but is otherwise harmless.
1184 *
1185 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1186 * on each visit during the top-down scrub phase, it will never be scrubbed.
1187 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1188 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1189 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1190 * while a scrub is in progress, it scrubs the block right then.
1191 */
1192static void
1193dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1194{
1195 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1196 ddt_entry_t dde = { 0 };
1197 int error;
1198 uint64_t n = 0;
1199
1200 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1201 ddt_t *ddt;
1202
1203 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1204 break;
1205 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1206 (longlong_t)ddb->ddb_class,
1207 (longlong_t)ddb->ddb_type,
1208 (longlong_t)ddb->ddb_checksum,
1209 (longlong_t)ddb->ddb_cursor);
1210
1211 /* There should be no pending changes to the dedup table */
1212 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1213 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1214
1215 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1216 n++;
1217
1218 if (dsl_scan_check_pause(scn, NULL))
1219 break;
1220 }
1221
1222 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1223 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1224 (int)scn->scn_pausing);
1225
1226 ASSERT(error == 0 || error == ENOENT);
1227 ASSERT(error != ENOENT ||
1228 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1229}
1230
1231/* ARGSUSED */
1232void
1233dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1234 ddt_entry_t *dde, dmu_tx_t *tx)
1235{
1236 const ddt_key_t *ddk = &dde->dde_key;
1237 ddt_phys_t *ddp = dde->dde_phys;
1238 blkptr_t bp;
1239 zbookmark_phys_t zb = { 0 };
1240
1241 if (scn->scn_phys.scn_state != DSS_SCANNING)
1242 return;
1243
1244 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1245 if (ddp->ddp_phys_birth == 0 ||
1246 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1247 continue;
1248 ddt_bp_create(checksum, ddk, ddp, &bp);
1249
1250 scn->scn_visited_this_txg++;
1251 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1252 }
1253}
1254
1255static void
1256dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1257{
1258 dsl_pool_t *dp = scn->scn_dp;
1259 zap_cursor_t zc;
1260 zap_attribute_t za;
1261
1262 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1263 scn->scn_phys.scn_ddt_class_max) {
1264 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1265 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1266 dsl_scan_ddt(scn, tx);
1267 if (scn->scn_pausing)
1268 return;
1269 }
1270
1271 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1272 /* First do the MOS & ORIGIN */
1273
1274 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1275 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1276 dsl_scan_visit_rootbp(scn, NULL,
1277 &dp->dp_meta_rootbp, tx);
1278 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1279 if (scn->scn_pausing)
1280 return;
1281
1282 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1283 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1284 enqueue_cb, tx, DS_FIND_CHILDREN));
1285 } else {
1286 dsl_scan_visitds(scn,
1287 dp->dp_origin_snap->ds_object, tx);
1288 }
1289 ASSERT(!scn->scn_pausing);
1290 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1291 ZB_DESTROYED_OBJSET) {
1292 /*
1293 * If we were paused, continue from here. Note if the
1294 * ds we were paused on was deleted, the zb_objset may
1295 * be -1, so we will skip this and find a new objset
1296 * below.
1297 */
1298 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1299 if (scn->scn_pausing)
1300 return;
1301 }
1302
1303 /*
1304 * In case we were paused right at the end of the ds, zero the
1305 * bookmark so we don't think that we're still trying to resume.
1306 */
1307 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1308
1309 /* keep pulling things out of the zap-object-as-queue */
1310 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1311 scn->scn_phys.scn_queue_obj),
1312 zap_cursor_retrieve(&zc, &za) == 0) {
1313 dsl_dataset_t *ds;
1314 uint64_t dsobj;
1315
1316 dsobj = strtonum(za.za_name, NULL);
1317 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1318 scn->scn_phys.scn_queue_obj, dsobj, tx));
1319
1320 /* Set up min/max txg */
1321 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1322 if (za.za_first_integer != 0) {
1323 scn->scn_phys.scn_cur_min_txg =
1324 MAX(scn->scn_phys.scn_min_txg,
1325 za.za_first_integer);
1326 } else {
1327 scn->scn_phys.scn_cur_min_txg =
1328 MAX(scn->scn_phys.scn_min_txg,
1329 ds->ds_phys->ds_prev_snap_txg);
1330 }
1331 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1332 dsl_dataset_rele(ds, FTAG);
1333
1334 dsl_scan_visitds(scn, dsobj, tx);
1335 zap_cursor_fini(&zc);
1336 if (scn->scn_pausing)
1337 return;
1338 }
1339 zap_cursor_fini(&zc);
1340}
1341
1342static boolean_t
1343dsl_scan_free_should_pause(dsl_scan_t *scn)
1344{
1345 uint64_t elapsed_nanosecs;
1346
1347 if (zfs_recover)
1348 return (B_FALSE);
1349
1350 if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1351 return (B_TRUE);
1352
1353 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1354 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1355 (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1356 txg_sync_waiting(scn->scn_dp)) ||
1357 spa_shutting_down(scn->scn_dp->dp_spa));
1358}
1359
1360static int
1361dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1362{
1363 dsl_scan_t *scn = arg;
1364
1365 if (!scn->scn_is_bptree ||
1366 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1367 if (dsl_scan_free_should_pause(scn))
1368 return (SET_ERROR(ERESTART));
1369 }
1370
1371 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1372 dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1373 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1374 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1375 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1376 scn->scn_visited_this_txg++;
1377 return (0);
1378}
1379
1380boolean_t
1381dsl_scan_active(dsl_scan_t *scn)
1382{
1383 spa_t *spa = scn->scn_dp->dp_spa;
1384 uint64_t used = 0, comp, uncomp;
1385
1386 if (spa->spa_load_state != SPA_LOAD_NONE)
1387 return (B_FALSE);
1388 if (spa_shutting_down(spa))
1389 return (B_FALSE);
1390 if (scn->scn_phys.scn_state == DSS_SCANNING ||
1391 (scn->scn_async_destroying && !scn->scn_async_stalled))
1392 return (B_TRUE);
1393
1394 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1395 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1396 &used, &comp, &uncomp);
1397 }
1398 return (used != 0);
1399}
1400
1401void
1402dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1403{
1404 dsl_scan_t *scn = dp->dp_scan;
1405 spa_t *spa = dp->dp_spa;
1406 int err = 0;
1407
1408 /*
1409 * Check for scn_restart_txg before checking spa_load_state, so
1410 * that we can restart an old-style scan while the pool is being
1411 * imported (see dsl_scan_init).
1412 */
1413 if (scn->scn_restart_txg != 0 &&
1414 scn->scn_restart_txg <= tx->tx_txg) {
1415 pool_scan_func_t func = POOL_SCAN_SCRUB;
1416 dsl_scan_done(scn, B_FALSE, tx);
1417 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1418 func = POOL_SCAN_RESILVER;
1419 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1420 func, tx->tx_txg);
1421 dsl_scan_setup_sync(&func, tx);
1422 }
1423
1424 /*
1425 * If the scan is inactive due to a stalled async destroy, try again.
1426 */
1427 if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1428 spa_sync_pass(dp->dp_spa) > 1)
1429 return;
1430
1431 scn->scn_visited_this_txg = 0;
1432 scn->scn_pausing = B_FALSE;
1433 scn->scn_sync_start_time = gethrtime();
1434 spa->spa_scrub_active = B_TRUE;
1435
1436 /*
1437 * First process the async destroys. If we pause, don't do
1438 * any scrubbing or resilvering. This ensures that there are no
1439 * async destroys while we are scanning, so the scan code doesn't
1440 * have to worry about traversing it. It is also faster to free the
1441 * blocks than to scrub them.
1442 */
1443 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1444 scn->scn_is_bptree = B_FALSE;
1445 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1446 NULL, ZIO_FLAG_MUSTSUCCEED);
1447 err = bpobj_iterate(&dp->dp_free_bpobj,
1448 dsl_scan_free_block_cb, scn, tx);
1449 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1450
1451 if (err != 0 && err != ERESTART)
1452 zfs_panic_recover("error %u from bpobj_iterate()", err);
1453 }
1454
1455 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1456 ASSERT(scn->scn_async_destroying);
1457 scn->scn_is_bptree = B_TRUE;
1458 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1459 NULL, ZIO_FLAG_MUSTSUCCEED);
1460 err = bptree_iterate(dp->dp_meta_objset,
1461 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1462 VERIFY0(zio_wait(scn->scn_zio_root));
1463
1464 if (err == EIO || err == ECKSUM) {
1465 err = 0;
1466 } else if (err != 0 && err != ERESTART) {
1467 zfs_panic_recover("error %u from "
1468 "traverse_dataset_destroyed()", err);
1469 }
1470
1471 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1472 /* finished; deactivate async destroy feature */
1473 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1474 ASSERT(!spa_feature_is_active(spa,
1475 SPA_FEATURE_ASYNC_DESTROY));
1476 VERIFY0(zap_remove(dp->dp_meta_objset,
1477 DMU_POOL_DIRECTORY_OBJECT,
1478 DMU_POOL_BPTREE_OBJ, tx));
1479 VERIFY0(bptree_free(dp->dp_meta_objset,
1480 dp->dp_bptree_obj, tx));
1481 dp->dp_bptree_obj = 0;
1482 scn->scn_async_destroying = B_FALSE;
1483 } else {
1484 /*
1485 * If we didn't make progress, mark the async destroy as
1486 * stalled, so that we will not initiate a spa_sync() on
1487 * its behalf.
1488 */
1489 scn->scn_async_stalled =
1490 (scn->scn_visited_this_txg == 0);
1491 }
1492 }
1493 if (scn->scn_visited_this_txg) {
1494 zfs_dbgmsg("freed %llu blocks in %llums from "
1495 "free_bpobj/bptree txg %llu; err=%d",
1496 (longlong_t)scn->scn_visited_this_txg,
1497 (longlong_t)
1498 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1499 (longlong_t)tx->tx_txg, err);
1500 scn->scn_visited_this_txg = 0;
1501
1502 /*
1503 * Write out changes to the DDT that may be required as a
1504 * result of the blocks freed. This ensures that the DDT
1505 * is clean when a scrub/resilver runs.
1506 */
1507 ddt_sync(spa, tx->tx_txg);
1508 }
1509 if (err != 0)
1510 return;
1511 if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1512 (dp->dp_free_dir->dd_phys->dd_used_bytes != 0 ||
1513 dp->dp_free_dir->dd_phys->dd_compressed_bytes != 0 ||
1514 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes != 0)) {
1515 /*
1516 * We have finished background destroying, but there is still
1517 * some space left in the dp_free_dir. Transfer this leaked
1518 * space to the dp_leak_dir.
1519 */
1520 if (dp->dp_leak_dir == NULL) {
1521 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1522 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1523 LEAK_DIR_NAME, tx);
1524 VERIFY0(dsl_pool_open_special_dir(dp,
1525 LEAK_DIR_NAME, &dp->dp_leak_dir));
1526 rrw_exit(&dp->dp_config_rwlock, FTAG);
1527 }
1528 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1529 dp->dp_free_dir->dd_phys->dd_used_bytes,
1530 dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1531 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1532 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1533 -dp->dp_free_dir->dd_phys->dd_used_bytes,
1534 -dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1535 -dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1536 }
1537 if (!scn->scn_async_destroying) {
1538 /* finished; verify that space accounting went to zero */
1539 ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes);
1540 ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes);
1541 ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes);
1542 }
1543
1544 if (scn->scn_phys.scn_state != DSS_SCANNING)
1545 return;
1546
1547 if (scn->scn_done_txg == tx->tx_txg) {
1548 ASSERT(!scn->scn_pausing);
1549 /* finished with scan. */
1550 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1551 dsl_scan_done(scn, B_TRUE, tx);
1552 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1553 dsl_scan_sync_state(scn, tx);
1554 return;
1555 }
1556
1557 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1558 scn->scn_phys.scn_ddt_class_max) {
1559 zfs_dbgmsg("doing scan sync txg %llu; "
1560 "ddt bm=%llu/%llu/%llu/%llx",
1561 (longlong_t)tx->tx_txg,
1562 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1563 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1564 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1565 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1566 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1567 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1568 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1569 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1570 } else {
1571 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1572 (longlong_t)tx->tx_txg,
1573 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1574 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1575 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1576 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1577 }
1578
1579 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1580 NULL, ZIO_FLAG_CANFAIL);
1581 dsl_pool_config_enter(dp, FTAG);
1582 dsl_scan_visit(scn, tx);
1583 dsl_pool_config_exit(dp, FTAG);
1584 (void) zio_wait(scn->scn_zio_root);
1585 scn->scn_zio_root = NULL;
1586
1587 zfs_dbgmsg("visited %llu blocks in %llums",
1588 (longlong_t)scn->scn_visited_this_txg,
1589 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1590
1591 if (!scn->scn_pausing) {
1592 scn->scn_done_txg = tx->tx_txg + 1;
1593 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1594 tx->tx_txg, scn->scn_done_txg);
1595 }
1596
1597 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1598 mutex_enter(&spa->spa_scrub_lock);
1599 while (spa->spa_scrub_inflight > 0) {
1600 cv_wait(&spa->spa_scrub_io_cv,
1601 &spa->spa_scrub_lock);
1602 }
1603 mutex_exit(&spa->spa_scrub_lock);
1604 }
1605
1606 dsl_scan_sync_state(scn, tx);
1607}
1608
1609/*
1610 * This will start a new scan, or restart an existing one.
1611 */
1612void
1613dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1614{
1615 if (txg == 0) {
1616 dmu_tx_t *tx;
1617 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1618 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1619
1620 txg = dmu_tx_get_txg(tx);
1621 dp->dp_scan->scn_restart_txg = txg;
1622 dmu_tx_commit(tx);
1623 } else {
1624 dp->dp_scan->scn_restart_txg = txg;
1625 }
1626 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1627}
1628
1629boolean_t
1630dsl_scan_resilvering(dsl_pool_t *dp)
1631{
1632 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1633 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1634}
1635
1636/*
1637 * scrub consumers
1638 */
1639
1640static void
1641count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1642{
1643 int i;
1644
1645 /*
1646 * If we resume after a reboot, zab will be NULL; don't record
1647 * incomplete stats in that case.
1648 */
1649 if (zab == NULL)
1650 return;
1651
1652 for (i = 0; i < 4; i++) {
1653 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1654 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1655 if (t & DMU_OT_NEWTYPE)
1656 t = DMU_OT_OTHER;
1657 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1658 int equal;
1659
1660 zb->zb_count++;
1661 zb->zb_asize += BP_GET_ASIZE(bp);
1662 zb->zb_lsize += BP_GET_LSIZE(bp);
1663 zb->zb_psize += BP_GET_PSIZE(bp);
1664 zb->zb_gangs += BP_COUNT_GANG(bp);
1665
1666 switch (BP_GET_NDVAS(bp)) {
1667 case 2:
1668 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1669 DVA_GET_VDEV(&bp->blk_dva[1]))
1670 zb->zb_ditto_2_of_2_samevdev++;
1671 break;
1672 case 3:
1673 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1674 DVA_GET_VDEV(&bp->blk_dva[1])) +
1675 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1676 DVA_GET_VDEV(&bp->blk_dva[2])) +
1677 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1678 DVA_GET_VDEV(&bp->blk_dva[2]));
1679 if (equal == 1)
1680 zb->zb_ditto_2_of_3_samevdev++;
1681 else if (equal == 3)
1682 zb->zb_ditto_3_of_3_samevdev++;
1683 break;
1684 }
1685 }
1686}
1687
1688static void
1689dsl_scan_scrub_done(zio_t *zio)
1690{
1691 spa_t *spa = zio->io_spa;
1692
1693 zio_data_buf_free(zio->io_data, zio->io_size);
1694
1695 mutex_enter(&spa->spa_scrub_lock);
1696 spa->spa_scrub_inflight--;
1697 cv_broadcast(&spa->spa_scrub_io_cv);
1698
1699 if (zio->io_error && (zio->io_error != ECKSUM ||
1700 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1701 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1702 }
1703 mutex_exit(&spa->spa_scrub_lock);
1704}
1705
1706static int
1707dsl_scan_scrub_cb(dsl_pool_t *dp,
1708 const blkptr_t *bp, const zbookmark_phys_t *zb)
1709{
1710 dsl_scan_t *scn = dp->dp_scan;
1711 size_t size = BP_GET_PSIZE(bp);
1712 spa_t *spa = dp->dp_spa;
1713 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1714 boolean_t needs_io;
1715 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1716 unsigned int scan_delay = 0;
1717
1718 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1719 phys_birth >= scn->scn_phys.scn_max_txg)
1720 return (0);
1721
1722 count_block(dp->dp_blkstats, bp);
1723
1724 if (BP_IS_EMBEDDED(bp))
1725 return (0);
1726
1727 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1728 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1729 zio_flags |= ZIO_FLAG_SCRUB;
1730 needs_io = B_TRUE;
1731 scan_delay = zfs_scrub_delay;
1732 } else {
1733 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1734 zio_flags |= ZIO_FLAG_RESILVER;
1735 needs_io = B_FALSE;
1736 scan_delay = zfs_resilver_delay;
1737 }
1738
1739 /* If it's an intent log block, failure is expected. */
1740 if (zb->zb_level == ZB_ZIL_LEVEL)
1741 zio_flags |= ZIO_FLAG_SPECULATIVE;
1742
1743 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1744 vdev_t *vd = vdev_lookup_top(spa,
1745 DVA_GET_VDEV(&bp->blk_dva[d]));
1746
1747 /*
1748 * Keep track of how much data we've examined so that
1749 * zpool(1M) status can make useful progress reports.
1750 */
1751 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1752 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1753
1754 /* if it's a resilver, this may not be in the target range */
1755 if (!needs_io) {
1756 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1757 /*
1758 * Gang members may be spread across multiple
1759 * vdevs, so the best estimate we have is the
1760 * scrub range, which has already been checked.
1761 * XXX -- it would be better to change our
1762 * allocation policy to ensure that all
1763 * gang members reside on the same vdev.
1764 */
1765 needs_io = B_TRUE;
1766 } else {
1767 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1768 phys_birth, 1);
1769 }
1770 }
1771 }
1772
1773 if (needs_io && !zfs_no_scrub_io) {
1774 vdev_t *rvd = spa->spa_root_vdev;
1775 uint64_t maxinflight = rvd->vdev_children *
1776 MAX(zfs_top_maxinflight, 1);
1777 void *data = zio_data_buf_alloc(size);
1778
1779 mutex_enter(&spa->spa_scrub_lock);
1780 while (spa->spa_scrub_inflight >= maxinflight)
1781 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1782 spa->spa_scrub_inflight++;
1783 mutex_exit(&spa->spa_scrub_lock);
1784
1785 /*
1786 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1787 * then throttle our workload to limit the impact of a scan.
1788 */
1789 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1790 delay(MAX((int)scan_delay, 0));
1791
1792 zio_nowait(zio_read(NULL, spa, bp, data, size,
1793 dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1794 zio_flags, zb));
1795 }
1796
1797 /* do not relocate this block */
1798 return (0);
1799}
1800
1801int
1802dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1803{
1804 spa_t *spa = dp->dp_spa;
1805
1806 /*
1807 * Purge all vdev caches and probe all devices. We do this here
1808 * rather than in sync context because this requires a writer lock
1809 * on the spa_config lock, which we can't do from sync context. The
1810 * spa_scrub_reopen flag indicates that vdev_open() should not
1811 * attempt to start another scrub.
1812 */
1813 spa_vdev_state_enter(spa, SCL_NONE);
1814 spa->spa_scrub_reopen = B_TRUE;
1815 vdev_reopen(spa->spa_root_vdev);
1816 spa->spa_scrub_reopen = B_FALSE;
1817 (void) spa_vdev_state_exit(spa, NULL, 0);
1818
1819 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1820 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1821}
| 467 spa_shutting_down(scn->scn_dp->dp_spa)) {
468 if (zb) {
469 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
470 (longlong_t)zb->zb_objset,
471 (longlong_t)zb->zb_object,
472 (longlong_t)zb->zb_level,
473 (longlong_t)zb->zb_blkid);
474 scn->scn_phys.scn_bookmark = *zb;
475 }
476 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
477 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
478 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
479 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
480 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
481 scn->scn_pausing = B_TRUE;
482 return (B_TRUE);
483 }
484 return (B_FALSE);
485}
486
487typedef struct zil_scan_arg {
488 dsl_pool_t *zsa_dp;
489 zil_header_t *zsa_zh;
490} zil_scan_arg_t;
491
492/* ARGSUSED */
493static int
494dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
495{
496 zil_scan_arg_t *zsa = arg;
497 dsl_pool_t *dp = zsa->zsa_dp;
498 dsl_scan_t *scn = dp->dp_scan;
499 zil_header_t *zh = zsa->zsa_zh;
500 zbookmark_phys_t zb;
501
502 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
503 return (0);
504
505 /*
506 * One block ("stubby") can be allocated a long time ago; we
507 * want to visit that one because it has been allocated
508 * (on-disk) even if it hasn't been claimed (even though for
509 * scrub there's nothing to do to it).
510 */
511 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
512 return (0);
513
514 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
515 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
516
517 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
518 return (0);
519}
520
521/* ARGSUSED */
522static int
523dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
524{
525 if (lrc->lrc_txtype == TX_WRITE) {
526 zil_scan_arg_t *zsa = arg;
527 dsl_pool_t *dp = zsa->zsa_dp;
528 dsl_scan_t *scn = dp->dp_scan;
529 zil_header_t *zh = zsa->zsa_zh;
530 lr_write_t *lr = (lr_write_t *)lrc;
531 blkptr_t *bp = &lr->lr_blkptr;
532 zbookmark_phys_t zb;
533
534 if (BP_IS_HOLE(bp) ||
535 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
536 return (0);
537
538 /*
539 * birth can be < claim_txg if this record's txg is
540 * already txg sync'ed (but this log block contains
541 * other records that are not synced)
542 */
543 if (claim_txg == 0 || bp->blk_birth < claim_txg)
544 return (0);
545
546 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
547 lr->lr_foid, ZB_ZIL_LEVEL,
548 lr->lr_offset / BP_GET_LSIZE(bp));
549
550 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
551 }
552 return (0);
553}
554
555static void
556dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
557{
558 uint64_t claim_txg = zh->zh_claim_txg;
559 zil_scan_arg_t zsa = { dp, zh };
560 zilog_t *zilog;
561
562 /*
563 * We only want to visit blocks that have been claimed but not yet
564 * replayed (or, in read-only mode, blocks that *would* be claimed).
565 */
566 if (claim_txg == 0 && spa_writeable(dp->dp_spa))
567 return;
568
569 zilog = zil_alloc(dp->dp_meta_objset, zh);
570
571 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
572 claim_txg);
573
574 zil_free(zilog);
575}
576
577/* ARGSUSED */
578static void
579dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
580 uint64_t objset, uint64_t object, uint64_t blkid)
581{
582 zbookmark_phys_t czb;
583 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
584
585 if (zfs_no_scrub_prefetch)
586 return;
587
588 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
589 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
590 return;
591
592 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
593
594 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
595 NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
596 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
597}
598
599static boolean_t
600dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
601 const zbookmark_phys_t *zb)
602{
603 /*
604 * We never skip over user/group accounting objects (obj<0)
605 */
606 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
607 (int64_t)zb->zb_object >= 0) {
608 /*
609 * If we already visited this bp & everything below (in
610 * a prior txg sync), don't bother doing it again.
611 */
612 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
613 return (B_TRUE);
614
615 /*
616 * If we found the block we're trying to resume from, or
617 * we went past it to a different object, zero it out to
618 * indicate that it's OK to start checking for pausing
619 * again.
620 */
621 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
622 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
623 dprintf("resuming at %llx/%llx/%llx/%llx\n",
624 (longlong_t)zb->zb_objset,
625 (longlong_t)zb->zb_object,
626 (longlong_t)zb->zb_level,
627 (longlong_t)zb->zb_blkid);
628 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
629 }
630 }
631 return (B_FALSE);
632}
633
634/*
635 * Return nonzero on i/o error.
636 * Return new buf to write out in *bufp.
637 */
638static int
639dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
640 dnode_phys_t *dnp, const blkptr_t *bp,
641 const zbookmark_phys_t *zb, dmu_tx_t *tx)
642{
643 dsl_pool_t *dp = scn->scn_dp;
644 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
645 int err;
646
647 if (BP_GET_LEVEL(bp) > 0) {
648 uint32_t flags = ARC_WAIT;
649 int i;
650 blkptr_t *cbp;
651 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
652 arc_buf_t *buf;
653
654 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
655 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
656 if (err) {
657 scn->scn_phys.scn_errors++;
658 return (err);
659 }
660 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
661 dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
662 zb->zb_object, zb->zb_blkid * epb + i);
663 }
664 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
665 zbookmark_phys_t czb;
666
667 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
668 zb->zb_level - 1,
669 zb->zb_blkid * epb + i);
670 dsl_scan_visitbp(cbp, &czb, dnp,
671 ds, scn, ostype, tx);
672 }
673 (void) arc_buf_remove_ref(buf, &buf);
674 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
675 uint32_t flags = ARC_WAIT;
676 dnode_phys_t *cdnp;
677 int i, j;
678 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
679 arc_buf_t *buf;
680
681 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
682 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
683 if (err) {
684 scn->scn_phys.scn_errors++;
685 return (err);
686 }
687 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
688 for (j = 0; j < cdnp->dn_nblkptr; j++) {
689 blkptr_t *cbp = &cdnp->dn_blkptr[j];
690 dsl_scan_prefetch(scn, buf, cbp,
691 zb->zb_objset, zb->zb_blkid * epb + i, j);
692 }
693 }
694 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
695 dsl_scan_visitdnode(scn, ds, ostype,
696 cdnp, zb->zb_blkid * epb + i, tx);
697 }
698
699 (void) arc_buf_remove_ref(buf, &buf);
700 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
701 uint32_t flags = ARC_WAIT;
702 objset_phys_t *osp;
703 arc_buf_t *buf;
704
705 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
706 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
707 if (err) {
708 scn->scn_phys.scn_errors++;
709 return (err);
710 }
711
712 osp = buf->b_data;
713
714 dsl_scan_visitdnode(scn, ds, osp->os_type,
715 &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
716
717 if (OBJSET_BUF_HAS_USERUSED(buf)) {
718 /*
719 * We also always visit user/group accounting
720 * objects, and never skip them, even if we are
721 * pausing. This is necessary so that the space
722 * deltas from this txg get integrated.
723 */
724 dsl_scan_visitdnode(scn, ds, osp->os_type,
725 &osp->os_groupused_dnode,
726 DMU_GROUPUSED_OBJECT, tx);
727 dsl_scan_visitdnode(scn, ds, osp->os_type,
728 &osp->os_userused_dnode,
729 DMU_USERUSED_OBJECT, tx);
730 }
731 (void) arc_buf_remove_ref(buf, &buf);
732 }
733
734 return (0);
735}
736
737static void
738dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
739 dmu_objset_type_t ostype, dnode_phys_t *dnp,
740 uint64_t object, dmu_tx_t *tx)
741{
742 int j;
743
744 for (j = 0; j < dnp->dn_nblkptr; j++) {
745 zbookmark_phys_t czb;
746
747 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
748 dnp->dn_nlevels - 1, j);
749 dsl_scan_visitbp(&dnp->dn_blkptr[j],
750 &czb, dnp, ds, scn, ostype, tx);
751 }
752
753 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
754 zbookmark_phys_t czb;
755 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
756 0, DMU_SPILL_BLKID);
757 dsl_scan_visitbp(&dnp->dn_spill,
758 &czb, dnp, ds, scn, ostype, tx);
759 }
760}
761
762/*
763 * The arguments are in this order because mdb can only print the
764 * first 5; we want them to be useful.
765 */
766static void
767dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
768 dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
769 dmu_objset_type_t ostype, dmu_tx_t *tx)
770{
771 dsl_pool_t *dp = scn->scn_dp;
772 arc_buf_t *buf = NULL;
773 blkptr_t bp_toread = *bp;
774
775 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */
776
777 if (dsl_scan_check_pause(scn, zb))
778 return;
779
780 if (dsl_scan_check_resume(scn, dnp, zb))
781 return;
782
783 if (BP_IS_HOLE(bp))
784 return;
785
786 scn->scn_visited_this_txg++;
787
788 dprintf_bp(bp,
789 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
790 ds, ds ? ds->ds_object : 0,
791 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
792 bp);
793
794 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
795 return;
796
797 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
798 return;
799
800 /*
801 * If dsl_scan_ddt() has aready visited this block, it will have
802 * already done any translations or scrubbing, so don't call the
803 * callback again.
804 */
805 if (ddt_class_contains(dp->dp_spa,
806 scn->scn_phys.scn_ddt_class_max, bp)) {
807 ASSERT(buf == NULL);
808 return;
809 }
810
811 /*
812 * If this block is from the future (after cur_max_txg), then we
813 * are doing this on behalf of a deleted snapshot, and we will
814 * revisit the future block on the next pass of this dataset.
815 * Don't scan it now unless we need to because something
816 * under it was modified.
817 */
818 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
819 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
820 }
821}
822
823static void
824dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
825 dmu_tx_t *tx)
826{
827 zbookmark_phys_t zb;
828
829 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
830 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
831 dsl_scan_visitbp(bp, &zb, NULL,
832 ds, scn, DMU_OST_NONE, tx);
833
834 dprintf_ds(ds, "finished scan%s", "");
835}
836
837void
838dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
839{
840 dsl_pool_t *dp = ds->ds_dir->dd_pool;
841 dsl_scan_t *scn = dp->dp_scan;
842 uint64_t mintxg;
843
844 if (scn->scn_phys.scn_state != DSS_SCANNING)
845 return;
846
847 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
848 if (dsl_dataset_is_snapshot(ds)) {
849 /* Note, scn_cur_{min,max}_txg stays the same. */
850 scn->scn_phys.scn_bookmark.zb_objset =
851 ds->ds_phys->ds_next_snap_obj;
852 zfs_dbgmsg("destroying ds %llu; currently traversing; "
853 "reset zb_objset to %llu",
854 (u_longlong_t)ds->ds_object,
855 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
856 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
857 } else {
858 SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
859 ZB_DESTROYED_OBJSET, 0, 0, 0);
860 zfs_dbgmsg("destroying ds %llu; currently traversing; "
861 "reset bookmark to -1,0,0,0",
862 (u_longlong_t)ds->ds_object);
863 }
864 } else if (zap_lookup_int_key(dp->dp_meta_objset,
865 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
866 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
867 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
868 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
869 if (dsl_dataset_is_snapshot(ds)) {
870 /*
871 * We keep the same mintxg; it could be >
872 * ds_creation_txg if the previous snapshot was
873 * deleted too.
874 */
875 VERIFY(zap_add_int_key(dp->dp_meta_objset,
876 scn->scn_phys.scn_queue_obj,
877 ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
878 zfs_dbgmsg("destroying ds %llu; in queue; "
879 "replacing with %llu",
880 (u_longlong_t)ds->ds_object,
881 (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
882 } else {
883 zfs_dbgmsg("destroying ds %llu; in queue; removing",
884 (u_longlong_t)ds->ds_object);
885 }
886 } else {
887 zfs_dbgmsg("destroying ds %llu; ignoring",
888 (u_longlong_t)ds->ds_object);
889 }
890
891 /*
892 * dsl_scan_sync() should be called after this, and should sync
893 * out our changed state, but just to be safe, do it here.
894 */
895 dsl_scan_sync_state(scn, tx);
896}
897
898void
899dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
900{
901 dsl_pool_t *dp = ds->ds_dir->dd_pool;
902 dsl_scan_t *scn = dp->dp_scan;
903 uint64_t mintxg;
904
905 if (scn->scn_phys.scn_state != DSS_SCANNING)
906 return;
907
908 ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
909
910 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
911 scn->scn_phys.scn_bookmark.zb_objset =
912 ds->ds_phys->ds_prev_snap_obj;
913 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
914 "reset zb_objset to %llu",
915 (u_longlong_t)ds->ds_object,
916 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
917 } else if (zap_lookup_int_key(dp->dp_meta_objset,
918 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
919 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
920 scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
921 VERIFY(zap_add_int_key(dp->dp_meta_objset,
922 scn->scn_phys.scn_queue_obj,
923 ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
924 zfs_dbgmsg("snapshotting ds %llu; in queue; "
925 "replacing with %llu",
926 (u_longlong_t)ds->ds_object,
927 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
928 }
929 dsl_scan_sync_state(scn, tx);
930}
931
932void
933dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
934{
935 dsl_pool_t *dp = ds1->ds_dir->dd_pool;
936 dsl_scan_t *scn = dp->dp_scan;
937 uint64_t mintxg;
938
939 if (scn->scn_phys.scn_state != DSS_SCANNING)
940 return;
941
942 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
943 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
944 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
945 "reset zb_objset to %llu",
946 (u_longlong_t)ds1->ds_object,
947 (u_longlong_t)ds2->ds_object);
948 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
949 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
950 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
951 "reset zb_objset to %llu",
952 (u_longlong_t)ds2->ds_object,
953 (u_longlong_t)ds1->ds_object);
954 }
955
956 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
957 ds1->ds_object, &mintxg) == 0) {
958 int err;
959
960 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
961 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
962 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
963 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
964 err = zap_add_int_key(dp->dp_meta_objset,
965 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
966 VERIFY(err == 0 || err == EEXIST);
967 if (err == EEXIST) {
968 /* Both were there to begin with */
969 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
970 scn->scn_phys.scn_queue_obj,
971 ds1->ds_object, mintxg, tx));
972 }
973 zfs_dbgmsg("clone_swap ds %llu; in queue; "
974 "replacing with %llu",
975 (u_longlong_t)ds1->ds_object,
976 (u_longlong_t)ds2->ds_object);
977 } else if (zap_lookup_int_key(dp->dp_meta_objset,
978 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
979 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
980 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
981 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
982 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
983 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
984 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
985 zfs_dbgmsg("clone_swap ds %llu; in queue; "
986 "replacing with %llu",
987 (u_longlong_t)ds2->ds_object,
988 (u_longlong_t)ds1->ds_object);
989 }
990
991 dsl_scan_sync_state(scn, tx);
992}
993
994struct enqueue_clones_arg {
995 dmu_tx_t *tx;
996 uint64_t originobj;
997};
998
999/* ARGSUSED */
1000static int
1001enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1002{
1003 struct enqueue_clones_arg *eca = arg;
1004 dsl_dataset_t *ds;
1005 int err;
1006 dsl_scan_t *scn = dp->dp_scan;
1007
1008 if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
1009 return (0);
1010
1011 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1012 if (err)
1013 return (err);
1014
1015 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
1016 dsl_dataset_t *prev;
1017 err = dsl_dataset_hold_obj(dp,
1018 ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
1019
1020 dsl_dataset_rele(ds, FTAG);
1021 if (err)
1022 return (err);
1023 ds = prev;
1024 }
1025 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1026 scn->scn_phys.scn_queue_obj, ds->ds_object,
1027 ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
1028 dsl_dataset_rele(ds, FTAG);
1029 return (0);
1030}
1031
1032static void
1033dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1034{
1035 dsl_pool_t *dp = scn->scn_dp;
1036 dsl_dataset_t *ds;
1037 objset_t *os;
1038
1039 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1040
1041 if (dmu_objset_from_ds(ds, &os))
1042 goto out;
1043
1044 /*
1045 * Only the ZIL in the head (non-snapshot) is valid. Even though
1046 * snapshots can have ZIL block pointers (which may be the same
1047 * BP as in the head), they must be ignored. So we traverse the
1048 * ZIL here, rather than in scan_recurse(), because the regular
1049 * snapshot block-sharing rules don't apply to it.
1050 */
1051 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1052 dsl_scan_zil(dp, &os->os_zil_header);
1053
1054 /*
1055 * Iterate over the bps in this ds.
1056 */
1057 dmu_buf_will_dirty(ds->ds_dbuf, tx);
1058 dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1059
1060 char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1061 dsl_dataset_name(ds, dsname);
1062 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1063 "pausing=%u",
1064 (longlong_t)dsobj, dsname,
1065 (longlong_t)scn->scn_phys.scn_cur_min_txg,
1066 (longlong_t)scn->scn_phys.scn_cur_max_txg,
1067 (int)scn->scn_pausing);
1068 kmem_free(dsname, ZFS_MAXNAMELEN);
1069
1070 if (scn->scn_pausing)
1071 goto out;
1072
1073 /*
1074 * We've finished this pass over this dataset.
1075 */
1076
1077 /*
1078 * If we did not completely visit this dataset, do another pass.
1079 */
1080 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1081 zfs_dbgmsg("incomplete pass; visiting again");
1082 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1083 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1084 scn->scn_phys.scn_queue_obj, ds->ds_object,
1085 scn->scn_phys.scn_cur_max_txg, tx) == 0);
1086 goto out;
1087 }
1088
1089 /*
1090 * Add descendent datasets to work queue.
1091 */
1092 if (ds->ds_phys->ds_next_snap_obj != 0) {
1093 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1094 scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1095 ds->ds_phys->ds_creation_txg, tx) == 0);
1096 }
1097 if (ds->ds_phys->ds_num_children > 1) {
1098 boolean_t usenext = B_FALSE;
1099 if (ds->ds_phys->ds_next_clones_obj != 0) {
1100 uint64_t count;
1101 /*
1102 * A bug in a previous version of the code could
1103 * cause upgrade_clones_cb() to not set
1104 * ds_next_snap_obj when it should, leading to a
1105 * missing entry. Therefore we can only use the
1106 * next_clones_obj when its count is correct.
1107 */
1108 int err = zap_count(dp->dp_meta_objset,
1109 ds->ds_phys->ds_next_clones_obj, &count);
1110 if (err == 0 &&
1111 count == ds->ds_phys->ds_num_children - 1)
1112 usenext = B_TRUE;
1113 }
1114
1115 if (usenext) {
1116 VERIFY0(zap_join_key(dp->dp_meta_objset,
1117 ds->ds_phys->ds_next_clones_obj,
1118 scn->scn_phys.scn_queue_obj,
1119 ds->ds_phys->ds_creation_txg, tx));
1120 } else {
1121 struct enqueue_clones_arg eca;
1122 eca.tx = tx;
1123 eca.originobj = ds->ds_object;
1124
1125 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1126 enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1127 }
1128 }
1129
1130out:
1131 dsl_dataset_rele(ds, FTAG);
1132}
1133
1134/* ARGSUSED */
1135static int
1136enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1137{
1138 dmu_tx_t *tx = arg;
1139 dsl_dataset_t *ds;
1140 int err;
1141 dsl_scan_t *scn = dp->dp_scan;
1142
1143 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1144 if (err)
1145 return (err);
1146
1147 while (ds->ds_phys->ds_prev_snap_obj != 0) {
1148 dsl_dataset_t *prev;
1149 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1150 FTAG, &prev);
1151 if (err) {
1152 dsl_dataset_rele(ds, FTAG);
1153 return (err);
1154 }
1155
1156 /*
1157 * If this is a clone, we don't need to worry about it for now.
1158 */
1159 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1160 dsl_dataset_rele(ds, FTAG);
1161 dsl_dataset_rele(prev, FTAG);
1162 return (0);
1163 }
1164 dsl_dataset_rele(ds, FTAG);
1165 ds = prev;
1166 }
1167
1168 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1169 ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1170 dsl_dataset_rele(ds, FTAG);
1171 return (0);
1172}
1173
1174/*
1175 * Scrub/dedup interaction.
1176 *
1177 * If there are N references to a deduped block, we don't want to scrub it
1178 * N times -- ideally, we should scrub it exactly once.
1179 *
1180 * We leverage the fact that the dde's replication class (enum ddt_class)
1181 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1182 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1183 *
1184 * To prevent excess scrubbing, the scrub begins by walking the DDT
1185 * to find all blocks with refcnt > 1, and scrubs each of these once.
1186 * Since there are two replication classes which contain blocks with
1187 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1188 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1189 *
1190 * There would be nothing more to say if a block's refcnt couldn't change
1191 * during a scrub, but of course it can so we must account for changes
1192 * in a block's replication class.
1193 *
1194 * Here's an example of what can occur:
1195 *
1196 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1197 * when visited during the top-down scrub phase, it will be scrubbed twice.
1198 * This negates our scrub optimization, but is otherwise harmless.
1199 *
1200 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1201 * on each visit during the top-down scrub phase, it will never be scrubbed.
1202 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1203 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1204 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1205 * while a scrub is in progress, it scrubs the block right then.
1206 */
1207static void
1208dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1209{
1210 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1211 ddt_entry_t dde = { 0 };
1212 int error;
1213 uint64_t n = 0;
1214
1215 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1216 ddt_t *ddt;
1217
1218 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1219 break;
1220 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1221 (longlong_t)ddb->ddb_class,
1222 (longlong_t)ddb->ddb_type,
1223 (longlong_t)ddb->ddb_checksum,
1224 (longlong_t)ddb->ddb_cursor);
1225
1226 /* There should be no pending changes to the dedup table */
1227 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1228 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1229
1230 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1231 n++;
1232
1233 if (dsl_scan_check_pause(scn, NULL))
1234 break;
1235 }
1236
1237 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1238 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1239 (int)scn->scn_pausing);
1240
1241 ASSERT(error == 0 || error == ENOENT);
1242 ASSERT(error != ENOENT ||
1243 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1244}
1245
1246/* ARGSUSED */
1247void
1248dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1249 ddt_entry_t *dde, dmu_tx_t *tx)
1250{
1251 const ddt_key_t *ddk = &dde->dde_key;
1252 ddt_phys_t *ddp = dde->dde_phys;
1253 blkptr_t bp;
1254 zbookmark_phys_t zb = { 0 };
1255
1256 if (scn->scn_phys.scn_state != DSS_SCANNING)
1257 return;
1258
1259 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1260 if (ddp->ddp_phys_birth == 0 ||
1261 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1262 continue;
1263 ddt_bp_create(checksum, ddk, ddp, &bp);
1264
1265 scn->scn_visited_this_txg++;
1266 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1267 }
1268}
1269
1270static void
1271dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1272{
1273 dsl_pool_t *dp = scn->scn_dp;
1274 zap_cursor_t zc;
1275 zap_attribute_t za;
1276
1277 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1278 scn->scn_phys.scn_ddt_class_max) {
1279 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1280 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1281 dsl_scan_ddt(scn, tx);
1282 if (scn->scn_pausing)
1283 return;
1284 }
1285
1286 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1287 /* First do the MOS & ORIGIN */
1288
1289 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1290 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1291 dsl_scan_visit_rootbp(scn, NULL,
1292 &dp->dp_meta_rootbp, tx);
1293 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1294 if (scn->scn_pausing)
1295 return;
1296
1297 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1298 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1299 enqueue_cb, tx, DS_FIND_CHILDREN));
1300 } else {
1301 dsl_scan_visitds(scn,
1302 dp->dp_origin_snap->ds_object, tx);
1303 }
1304 ASSERT(!scn->scn_pausing);
1305 } else if (scn->scn_phys.scn_bookmark.zb_objset !=
1306 ZB_DESTROYED_OBJSET) {
1307 /*
1308 * If we were paused, continue from here. Note if the
1309 * ds we were paused on was deleted, the zb_objset may
1310 * be -1, so we will skip this and find a new objset
1311 * below.
1312 */
1313 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1314 if (scn->scn_pausing)
1315 return;
1316 }
1317
1318 /*
1319 * In case we were paused right at the end of the ds, zero the
1320 * bookmark so we don't think that we're still trying to resume.
1321 */
1322 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1323
1324 /* keep pulling things out of the zap-object-as-queue */
1325 while (zap_cursor_init(&zc, dp->dp_meta_objset,
1326 scn->scn_phys.scn_queue_obj),
1327 zap_cursor_retrieve(&zc, &za) == 0) {
1328 dsl_dataset_t *ds;
1329 uint64_t dsobj;
1330
1331 dsobj = strtonum(za.za_name, NULL);
1332 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1333 scn->scn_phys.scn_queue_obj, dsobj, tx));
1334
1335 /* Set up min/max txg */
1336 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1337 if (za.za_first_integer != 0) {
1338 scn->scn_phys.scn_cur_min_txg =
1339 MAX(scn->scn_phys.scn_min_txg,
1340 za.za_first_integer);
1341 } else {
1342 scn->scn_phys.scn_cur_min_txg =
1343 MAX(scn->scn_phys.scn_min_txg,
1344 ds->ds_phys->ds_prev_snap_txg);
1345 }
1346 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1347 dsl_dataset_rele(ds, FTAG);
1348
1349 dsl_scan_visitds(scn, dsobj, tx);
1350 zap_cursor_fini(&zc);
1351 if (scn->scn_pausing)
1352 return;
1353 }
1354 zap_cursor_fini(&zc);
1355}
1356
1357static boolean_t
1358dsl_scan_free_should_pause(dsl_scan_t *scn)
1359{
1360 uint64_t elapsed_nanosecs;
1361
1362 if (zfs_recover)
1363 return (B_FALSE);
1364
1365 if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1366 return (B_TRUE);
1367
1368 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1369 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1370 (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1371 txg_sync_waiting(scn->scn_dp)) ||
1372 spa_shutting_down(scn->scn_dp->dp_spa));
1373}
1374
1375static int
1376dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1377{
1378 dsl_scan_t *scn = arg;
1379
1380 if (!scn->scn_is_bptree ||
1381 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1382 if (dsl_scan_free_should_pause(scn))
1383 return (SET_ERROR(ERESTART));
1384 }
1385
1386 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1387 dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1388 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1389 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1390 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1391 scn->scn_visited_this_txg++;
1392 return (0);
1393}
1394
1395boolean_t
1396dsl_scan_active(dsl_scan_t *scn)
1397{
1398 spa_t *spa = scn->scn_dp->dp_spa;
1399 uint64_t used = 0, comp, uncomp;
1400
1401 if (spa->spa_load_state != SPA_LOAD_NONE)
1402 return (B_FALSE);
1403 if (spa_shutting_down(spa))
1404 return (B_FALSE);
1405 if (scn->scn_phys.scn_state == DSS_SCANNING ||
1406 (scn->scn_async_destroying && !scn->scn_async_stalled))
1407 return (B_TRUE);
1408
1409 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1410 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1411 &used, &comp, &uncomp);
1412 }
1413 return (used != 0);
1414}
1415
1416void
1417dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1418{
1419 dsl_scan_t *scn = dp->dp_scan;
1420 spa_t *spa = dp->dp_spa;
1421 int err = 0;
1422
1423 /*
1424 * Check for scn_restart_txg before checking spa_load_state, so
1425 * that we can restart an old-style scan while the pool is being
1426 * imported (see dsl_scan_init).
1427 */
1428 if (scn->scn_restart_txg != 0 &&
1429 scn->scn_restart_txg <= tx->tx_txg) {
1430 pool_scan_func_t func = POOL_SCAN_SCRUB;
1431 dsl_scan_done(scn, B_FALSE, tx);
1432 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1433 func = POOL_SCAN_RESILVER;
1434 zfs_dbgmsg("restarting scan func=%u txg=%llu",
1435 func, tx->tx_txg);
1436 dsl_scan_setup_sync(&func, tx);
1437 }
1438
1439 /*
1440 * If the scan is inactive due to a stalled async destroy, try again.
1441 */
1442 if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1443 spa_sync_pass(dp->dp_spa) > 1)
1444 return;
1445
1446 scn->scn_visited_this_txg = 0;
1447 scn->scn_pausing = B_FALSE;
1448 scn->scn_sync_start_time = gethrtime();
1449 spa->spa_scrub_active = B_TRUE;
1450
1451 /*
1452 * First process the async destroys. If we pause, don't do
1453 * any scrubbing or resilvering. This ensures that there are no
1454 * async destroys while we are scanning, so the scan code doesn't
1455 * have to worry about traversing it. It is also faster to free the
1456 * blocks than to scrub them.
1457 */
1458 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1459 scn->scn_is_bptree = B_FALSE;
1460 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1461 NULL, ZIO_FLAG_MUSTSUCCEED);
1462 err = bpobj_iterate(&dp->dp_free_bpobj,
1463 dsl_scan_free_block_cb, scn, tx);
1464 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1465
1466 if (err != 0 && err != ERESTART)
1467 zfs_panic_recover("error %u from bpobj_iterate()", err);
1468 }
1469
1470 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1471 ASSERT(scn->scn_async_destroying);
1472 scn->scn_is_bptree = B_TRUE;
1473 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1474 NULL, ZIO_FLAG_MUSTSUCCEED);
1475 err = bptree_iterate(dp->dp_meta_objset,
1476 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1477 VERIFY0(zio_wait(scn->scn_zio_root));
1478
1479 if (err == EIO || err == ECKSUM) {
1480 err = 0;
1481 } else if (err != 0 && err != ERESTART) {
1482 zfs_panic_recover("error %u from "
1483 "traverse_dataset_destroyed()", err);
1484 }
1485
1486 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1487 /* finished; deactivate async destroy feature */
1488 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1489 ASSERT(!spa_feature_is_active(spa,
1490 SPA_FEATURE_ASYNC_DESTROY));
1491 VERIFY0(zap_remove(dp->dp_meta_objset,
1492 DMU_POOL_DIRECTORY_OBJECT,
1493 DMU_POOL_BPTREE_OBJ, tx));
1494 VERIFY0(bptree_free(dp->dp_meta_objset,
1495 dp->dp_bptree_obj, tx));
1496 dp->dp_bptree_obj = 0;
1497 scn->scn_async_destroying = B_FALSE;
1498 } else {
1499 /*
1500 * If we didn't make progress, mark the async destroy as
1501 * stalled, so that we will not initiate a spa_sync() on
1502 * its behalf.
1503 */
1504 scn->scn_async_stalled =
1505 (scn->scn_visited_this_txg == 0);
1506 }
1507 }
1508 if (scn->scn_visited_this_txg) {
1509 zfs_dbgmsg("freed %llu blocks in %llums from "
1510 "free_bpobj/bptree txg %llu; err=%d",
1511 (longlong_t)scn->scn_visited_this_txg,
1512 (longlong_t)
1513 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1514 (longlong_t)tx->tx_txg, err);
1515 scn->scn_visited_this_txg = 0;
1516
1517 /*
1518 * Write out changes to the DDT that may be required as a
1519 * result of the blocks freed. This ensures that the DDT
1520 * is clean when a scrub/resilver runs.
1521 */
1522 ddt_sync(spa, tx->tx_txg);
1523 }
1524 if (err != 0)
1525 return;
1526 if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1527 (dp->dp_free_dir->dd_phys->dd_used_bytes != 0 ||
1528 dp->dp_free_dir->dd_phys->dd_compressed_bytes != 0 ||
1529 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes != 0)) {
1530 /*
1531 * We have finished background destroying, but there is still
1532 * some space left in the dp_free_dir. Transfer this leaked
1533 * space to the dp_leak_dir.
1534 */
1535 if (dp->dp_leak_dir == NULL) {
1536 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1537 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1538 LEAK_DIR_NAME, tx);
1539 VERIFY0(dsl_pool_open_special_dir(dp,
1540 LEAK_DIR_NAME, &dp->dp_leak_dir));
1541 rrw_exit(&dp->dp_config_rwlock, FTAG);
1542 }
1543 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1544 dp->dp_free_dir->dd_phys->dd_used_bytes,
1545 dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1546 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1547 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1548 -dp->dp_free_dir->dd_phys->dd_used_bytes,
1549 -dp->dp_free_dir->dd_phys->dd_compressed_bytes,
1550 -dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
1551 }
1552 if (!scn->scn_async_destroying) {
1553 /* finished; verify that space accounting went to zero */
1554 ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes);
1555 ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes);
1556 ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes);
1557 }
1558
1559 if (scn->scn_phys.scn_state != DSS_SCANNING)
1560 return;
1561
1562 if (scn->scn_done_txg == tx->tx_txg) {
1563 ASSERT(!scn->scn_pausing);
1564 /* finished with scan. */
1565 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1566 dsl_scan_done(scn, B_TRUE, tx);
1567 ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1568 dsl_scan_sync_state(scn, tx);
1569 return;
1570 }
1571
1572 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1573 scn->scn_phys.scn_ddt_class_max) {
1574 zfs_dbgmsg("doing scan sync txg %llu; "
1575 "ddt bm=%llu/%llu/%llu/%llx",
1576 (longlong_t)tx->tx_txg,
1577 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1578 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1579 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1580 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1581 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1582 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1583 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1584 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1585 } else {
1586 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1587 (longlong_t)tx->tx_txg,
1588 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1589 (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1590 (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1591 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1592 }
1593
1594 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1595 NULL, ZIO_FLAG_CANFAIL);
1596 dsl_pool_config_enter(dp, FTAG);
1597 dsl_scan_visit(scn, tx);
1598 dsl_pool_config_exit(dp, FTAG);
1599 (void) zio_wait(scn->scn_zio_root);
1600 scn->scn_zio_root = NULL;
1601
1602 zfs_dbgmsg("visited %llu blocks in %llums",
1603 (longlong_t)scn->scn_visited_this_txg,
1604 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1605
1606 if (!scn->scn_pausing) {
1607 scn->scn_done_txg = tx->tx_txg + 1;
1608 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1609 tx->tx_txg, scn->scn_done_txg);
1610 }
1611
1612 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1613 mutex_enter(&spa->spa_scrub_lock);
1614 while (spa->spa_scrub_inflight > 0) {
1615 cv_wait(&spa->spa_scrub_io_cv,
1616 &spa->spa_scrub_lock);
1617 }
1618 mutex_exit(&spa->spa_scrub_lock);
1619 }
1620
1621 dsl_scan_sync_state(scn, tx);
1622}
1623
1624/*
1625 * This will start a new scan, or restart an existing one.
1626 */
1627void
1628dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1629{
1630 if (txg == 0) {
1631 dmu_tx_t *tx;
1632 tx = dmu_tx_create_dd(dp->dp_mos_dir);
1633 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1634
1635 txg = dmu_tx_get_txg(tx);
1636 dp->dp_scan->scn_restart_txg = txg;
1637 dmu_tx_commit(tx);
1638 } else {
1639 dp->dp_scan->scn_restart_txg = txg;
1640 }
1641 zfs_dbgmsg("restarting resilver txg=%llu", txg);
1642}
1643
1644boolean_t
1645dsl_scan_resilvering(dsl_pool_t *dp)
1646{
1647 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1648 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1649}
1650
1651/*
1652 * scrub consumers
1653 */
1654
1655static void
1656count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1657{
1658 int i;
1659
1660 /*
1661 * If we resume after a reboot, zab will be NULL; don't record
1662 * incomplete stats in that case.
1663 */
1664 if (zab == NULL)
1665 return;
1666
1667 for (i = 0; i < 4; i++) {
1668 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1669 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1670 if (t & DMU_OT_NEWTYPE)
1671 t = DMU_OT_OTHER;
1672 zfs_blkstat_t *zb = &zab->zab_type[l][t];
1673 int equal;
1674
1675 zb->zb_count++;
1676 zb->zb_asize += BP_GET_ASIZE(bp);
1677 zb->zb_lsize += BP_GET_LSIZE(bp);
1678 zb->zb_psize += BP_GET_PSIZE(bp);
1679 zb->zb_gangs += BP_COUNT_GANG(bp);
1680
1681 switch (BP_GET_NDVAS(bp)) {
1682 case 2:
1683 if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1684 DVA_GET_VDEV(&bp->blk_dva[1]))
1685 zb->zb_ditto_2_of_2_samevdev++;
1686 break;
1687 case 3:
1688 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1689 DVA_GET_VDEV(&bp->blk_dva[1])) +
1690 (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1691 DVA_GET_VDEV(&bp->blk_dva[2])) +
1692 (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1693 DVA_GET_VDEV(&bp->blk_dva[2]));
1694 if (equal == 1)
1695 zb->zb_ditto_2_of_3_samevdev++;
1696 else if (equal == 3)
1697 zb->zb_ditto_3_of_3_samevdev++;
1698 break;
1699 }
1700 }
1701}
1702
1703static void
1704dsl_scan_scrub_done(zio_t *zio)
1705{
1706 spa_t *spa = zio->io_spa;
1707
1708 zio_data_buf_free(zio->io_data, zio->io_size);
1709
1710 mutex_enter(&spa->spa_scrub_lock);
1711 spa->spa_scrub_inflight--;
1712 cv_broadcast(&spa->spa_scrub_io_cv);
1713
1714 if (zio->io_error && (zio->io_error != ECKSUM ||
1715 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1716 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1717 }
1718 mutex_exit(&spa->spa_scrub_lock);
1719}
1720
1721static int
1722dsl_scan_scrub_cb(dsl_pool_t *dp,
1723 const blkptr_t *bp, const zbookmark_phys_t *zb)
1724{
1725 dsl_scan_t *scn = dp->dp_scan;
1726 size_t size = BP_GET_PSIZE(bp);
1727 spa_t *spa = dp->dp_spa;
1728 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1729 boolean_t needs_io;
1730 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1731 unsigned int scan_delay = 0;
1732
1733 if (phys_birth <= scn->scn_phys.scn_min_txg ||
1734 phys_birth >= scn->scn_phys.scn_max_txg)
1735 return (0);
1736
1737 count_block(dp->dp_blkstats, bp);
1738
1739 if (BP_IS_EMBEDDED(bp))
1740 return (0);
1741
1742 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1743 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1744 zio_flags |= ZIO_FLAG_SCRUB;
1745 needs_io = B_TRUE;
1746 scan_delay = zfs_scrub_delay;
1747 } else {
1748 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1749 zio_flags |= ZIO_FLAG_RESILVER;
1750 needs_io = B_FALSE;
1751 scan_delay = zfs_resilver_delay;
1752 }
1753
1754 /* If it's an intent log block, failure is expected. */
1755 if (zb->zb_level == ZB_ZIL_LEVEL)
1756 zio_flags |= ZIO_FLAG_SPECULATIVE;
1757
1758 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1759 vdev_t *vd = vdev_lookup_top(spa,
1760 DVA_GET_VDEV(&bp->blk_dva[d]));
1761
1762 /*
1763 * Keep track of how much data we've examined so that
1764 * zpool(1M) status can make useful progress reports.
1765 */
1766 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1767 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1768
1769 /* if it's a resilver, this may not be in the target range */
1770 if (!needs_io) {
1771 if (DVA_GET_GANG(&bp->blk_dva[d])) {
1772 /*
1773 * Gang members may be spread across multiple
1774 * vdevs, so the best estimate we have is the
1775 * scrub range, which has already been checked.
1776 * XXX -- it would be better to change our
1777 * allocation policy to ensure that all
1778 * gang members reside on the same vdev.
1779 */
1780 needs_io = B_TRUE;
1781 } else {
1782 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1783 phys_birth, 1);
1784 }
1785 }
1786 }
1787
1788 if (needs_io && !zfs_no_scrub_io) {
1789 vdev_t *rvd = spa->spa_root_vdev;
1790 uint64_t maxinflight = rvd->vdev_children *
1791 MAX(zfs_top_maxinflight, 1);
1792 void *data = zio_data_buf_alloc(size);
1793
1794 mutex_enter(&spa->spa_scrub_lock);
1795 while (spa->spa_scrub_inflight >= maxinflight)
1796 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1797 spa->spa_scrub_inflight++;
1798 mutex_exit(&spa->spa_scrub_lock);
1799
1800 /*
1801 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1802 * then throttle our workload to limit the impact of a scan.
1803 */
1804 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1805 delay(MAX((int)scan_delay, 0));
1806
1807 zio_nowait(zio_read(NULL, spa, bp, data, size,
1808 dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1809 zio_flags, zb));
1810 }
1811
1812 /* do not relocate this block */
1813 return (0);
1814}
1815
1816int
1817dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1818{
1819 spa_t *spa = dp->dp_spa;
1820
1821 /*
1822 * Purge all vdev caches and probe all devices. We do this here
1823 * rather than in sync context because this requires a writer lock
1824 * on the spa_config lock, which we can't do from sync context. The
1825 * spa_scrub_reopen flag indicates that vdev_open() should not
1826 * attempt to start another scrub.
1827 */
1828 spa_vdev_state_enter(spa, SCL_NONE);
1829 spa->spa_scrub_reopen = B_TRUE;
1830 vdev_reopen(spa->spa_root_vdev);
1831 spa->spa_scrub_reopen = B_FALSE;
1832 (void) spa_vdev_state_exit(spa, NULL, 0);
1833
1834 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1835 dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1836}
|