4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
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/dsl_scan.h>
33 #include <sys/dnode.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
39 #include <sys/zfs_context.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/zfs_znode.h>
42 #include <sys/spa_impl.h>
43 #include <sys/dsl_deadlist.h>
44 #include <sys/bptree.h>
45 #include <sys/zfeature.h>
46 #include <sys/zil_impl.h>
47 #include <sys/dsl_userhold.h>
50 #include <sys/sysctl.h>
51 #include <sys/types.h>
58 * ZFS must limit the rate of incoming writes to the rate at which it is able
59 * to sync data modifications to the backend storage. Throttling by too much
60 * creates an artificial limit; throttling by too little can only be sustained
61 * for short periods and would lead to highly lumpy performance. On a per-pool
62 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
63 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
64 * of dirty data decreases. When the amount of dirty data exceeds a
65 * predetermined threshold further modifications are blocked until the amount
66 * of dirty data decreases (as data is synced out).
68 * The limit on dirty data is tunable, and should be adjusted according to
69 * both the IO capacity and available memory of the system. The larger the
70 * window, the more ZFS is able to aggregate and amortize metadata (and data)
71 * changes. However, memory is a limited resource, and allowing for more dirty
72 * data comes at the cost of keeping other useful data in memory (for example
73 * ZFS data cached by the ARC).
77 * As buffers are modified dsl_pool_willuse_space() increments both the per-
78 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
79 * dirty space used; dsl_pool_dirty_space() decrements those values as data
80 * is synced out from dsl_pool_sync(). While only the poolwide value is
81 * relevant, the per-txg value is useful for debugging. The tunable
82 * zfs_dirty_data_max determines the dirty space limit. Once that value is
83 * exceeded, new writes are halted until space frees up.
85 * The zfs_dirty_data_sync tunable dictates the threshold at which we
86 * ensure that there is a txg syncing (see the comment in txg.c for a full
87 * description of transaction group stages).
89 * The IO scheduler uses both the dirty space limit and current amount of
90 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
91 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
93 * The delay is also calculated based on the amount of dirty data. See the
94 * comment above dmu_tx_delay() for details.
98 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
99 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
101 uint64_t zfs_dirty_data_max;
102 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
103 int zfs_dirty_data_max_percent = 10;
106 * If there is at least this much dirty data, push out a txg.
108 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
111 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
112 * and delay each transaction.
113 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
115 int zfs_delay_min_dirty_percent = 60;
118 * This controls how quickly the delay approaches infinity.
119 * Larger values cause it to delay more for a given amount of dirty data.
120 * Therefore larger values will cause there to be less dirty data for a
123 * For the smoothest delay, this value should be about 1 billion divided
124 * by the maximum number of operations per second. This will smoothly
125 * handle between 10x and 1/10th this number.
127 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
128 * multiply in dmu_tx_delay().
130 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
135 extern int zfs_vdev_async_write_active_max_dirty_percent;
137 SYSCTL_DECL(_vfs_zfs);
139 TUNABLE_QUAD("vfs.zfs.dirty_data_max", &zfs_dirty_data_max);
140 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max, CTLFLAG_RWTUN,
141 &zfs_dirty_data_max, 0,
142 "The maximum amount of dirty data in bytes after which new writes are "
143 "halted until space becomes available");
145 TUNABLE_QUAD("vfs.zfs.dirty_data_max_max", &zfs_dirty_data_max_max);
146 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max_max, CTLFLAG_RDTUN,
147 &zfs_dirty_data_max_max, 0,
148 "The absolute cap on dirty_data_max when auto calculating");
150 TUNABLE_INT("vfs.zfs.dirty_data_max_percent", &zfs_dirty_data_max_percent);
151 static int sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS);
152 SYSCTL_PROC(_vfs_zfs, OID_AUTO, dirty_data_max_percent,
153 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
154 sysctl_zfs_dirty_data_max_percent, "I",
155 "The percent of physical memory used to auto calculate dirty_data_max");
157 TUNABLE_QUAD("vfs.zfs.dirty_data_sync", &zfs_dirty_data_sync);
158 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_sync, CTLFLAG_RWTUN,
159 &zfs_dirty_data_sync, 0,
160 "Force a txg if the number of dirty buffer bytes exceed this value");
162 static int sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS);
163 /* No zfs_delay_min_dirty_percent tunable due to limit requirements */
164 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_min_dirty_percent,
165 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(int),
166 sysctl_zfs_delay_min_dirty_percent, "I",
167 "The limit of outstanding dirty data before transations are delayed");
169 static int sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS);
170 /* No zfs_delay_scale tunable due to limit requirements */
171 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_scale,
172 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
173 sysctl_zfs_delay_scale, "QU",
174 "Controls how quickly the delay approaches infinity");
177 sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS)
181 val = zfs_dirty_data_max_percent;
182 err = sysctl_handle_int(oidp, &val, 0, req);
183 if (err != 0 || req->newptr == NULL)
186 if (val < 0 || val > 100)
189 zfs_dirty_data_max_percent = val;
195 sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS)
199 val = zfs_delay_min_dirty_percent;
200 err = sysctl_handle_int(oidp, &val, 0, req);
201 if (err != 0 || req->newptr == NULL)
204 if (val < zfs_vdev_async_write_active_max_dirty_percent)
207 zfs_delay_min_dirty_percent = val;
213 sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS)
218 val = zfs_delay_scale;
219 err = sysctl_handle_64(oidp, &val, 0, req);
220 if (err != 0 || req->newptr == NULL)
223 if (val > UINT64_MAX / zfs_dirty_data_max)
226 zfs_delay_scale = val;
232 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
233 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
236 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
241 err = zap_lookup(dp->dp_meta_objset,
242 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
243 name, sizeof (obj), 1, &obj);
247 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
251 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
254 blkptr_t *bp = spa_get_rootblkptr(spa);
256 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
258 dp->dp_meta_rootbp = *bp;
259 rrw_init(&dp->dp_config_rwlock, B_TRUE);
262 txg_list_create(&dp->dp_dirty_datasets,
263 offsetof(dsl_dataset_t, ds_dirty_link));
264 txg_list_create(&dp->dp_dirty_zilogs,
265 offsetof(zilog_t, zl_dirty_link));
266 txg_list_create(&dp->dp_dirty_dirs,
267 offsetof(dsl_dir_t, dd_dirty_link));
268 txg_list_create(&dp->dp_sync_tasks,
269 offsetof(dsl_sync_task_t, dst_node));
271 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
272 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
274 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
281 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
284 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
286 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
287 &dp->dp_meta_objset);
297 dsl_pool_open(dsl_pool_t *dp)
304 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
305 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
306 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
307 &dp->dp_root_dir_obj);
311 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
312 NULL, dp, &dp->dp_root_dir);
316 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
320 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
321 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
324 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
327 err = dsl_dataset_hold_obj(dp,
328 ds->ds_phys->ds_prev_snap_obj, dp,
329 &dp->dp_origin_snap);
330 dsl_dataset_rele(ds, FTAG);
332 dsl_dir_rele(dd, dp);
337 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
338 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
343 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
344 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
347 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
348 dp->dp_meta_objset, obj));
352 * Note: errors ignored, because the leak dir will not exist if we
353 * have not encountered a leak yet.
355 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
358 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
359 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
360 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
366 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
367 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
368 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
369 &dp->dp_empty_bpobj);
374 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
375 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
376 &dp->dp_tmp_userrefs_obj);
382 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
385 rrw_exit(&dp->dp_config_rwlock, FTAG);
390 dsl_pool_close(dsl_pool_t *dp)
393 * Drop our references from dsl_pool_open().
395 * Since we held the origin_snap from "syncing" context (which
396 * includes pool-opening context), it actually only got a "ref"
397 * and not a hold, so just drop that here.
399 if (dp->dp_origin_snap)
400 dsl_dataset_rele(dp->dp_origin_snap, dp);
402 dsl_dir_rele(dp->dp_mos_dir, dp);
404 dsl_dir_rele(dp->dp_free_dir, dp);
406 dsl_dir_rele(dp->dp_leak_dir, dp);
408 dsl_dir_rele(dp->dp_root_dir, dp);
410 bpobj_close(&dp->dp_free_bpobj);
412 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
413 if (dp->dp_meta_objset)
414 dmu_objset_evict(dp->dp_meta_objset);
416 txg_list_destroy(&dp->dp_dirty_datasets);
417 txg_list_destroy(&dp->dp_dirty_zilogs);
418 txg_list_destroy(&dp->dp_sync_tasks);
419 txg_list_destroy(&dp->dp_dirty_dirs);
421 arc_flush(dp->dp_spa);
424 rrw_destroy(&dp->dp_config_rwlock);
425 mutex_destroy(&dp->dp_lock);
426 taskq_destroy(dp->dp_vnrele_taskq);
428 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
429 kmem_free(dp, sizeof (dsl_pool_t));
433 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
436 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
437 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
442 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
444 /* create and open the MOS (meta-objset) */
445 dp->dp_meta_objset = dmu_objset_create_impl(spa,
446 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
448 /* create the pool directory */
449 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
450 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
453 /* Initialize scan structures */
454 VERIFY0(dsl_scan_init(dp, txg));
456 /* create and open the root dir */
457 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
458 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
459 NULL, dp, &dp->dp_root_dir));
461 /* create and open the meta-objset dir */
462 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
463 VERIFY0(dsl_pool_open_special_dir(dp,
464 MOS_DIR_NAME, &dp->dp_mos_dir));
466 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
467 /* create and open the free dir */
468 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
470 VERIFY0(dsl_pool_open_special_dir(dp,
471 FREE_DIR_NAME, &dp->dp_free_dir));
473 /* create and open the free_bplist */
474 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
475 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
476 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
477 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
478 dp->dp_meta_objset, obj));
481 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
482 dsl_pool_create_origin(dp, tx);
484 /* create the root dataset */
485 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
487 /* create the root objset */
488 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
489 os = dmu_objset_create_impl(dp->dp_spa, ds,
490 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
492 zfs_create_fs(os, kcred, zplprops, tx);
494 dsl_dataset_rele(ds, FTAG);
498 rrw_exit(&dp->dp_config_rwlock, FTAG);
504 * Account for the meta-objset space in its placeholder dsl_dir.
507 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
508 int64_t used, int64_t comp, int64_t uncomp)
510 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
511 mutex_enter(&dp->dp_lock);
512 dp->dp_mos_used_delta += used;
513 dp->dp_mos_compressed_delta += comp;
514 dp->dp_mos_uncompressed_delta += uncomp;
515 mutex_exit(&dp->dp_lock);
519 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
521 dsl_deadlist_t *dl = arg;
522 dsl_deadlist_insert(dl, bp, tx);
527 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
529 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
530 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
531 VERIFY0(zio_wait(zio));
532 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
533 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
537 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
539 ASSERT(MUTEX_HELD(&dp->dp_lock));
542 ASSERT3U(-delta, <=, dp->dp_dirty_total);
544 dp->dp_dirty_total += delta;
547 * Note: we signal even when increasing dp_dirty_total.
548 * This ensures forward progress -- each thread wakes the next waiter.
550 if (dp->dp_dirty_total <= zfs_dirty_data_max)
551 cv_signal(&dp->dp_spaceavail_cv);
555 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
561 objset_t *mos = dp->dp_meta_objset;
562 list_t synced_datasets;
564 list_create(&synced_datasets, sizeof (dsl_dataset_t),
565 offsetof(dsl_dataset_t, ds_synced_link));
567 tx = dmu_tx_create_assigned(dp, txg);
570 * Write out all dirty blocks of dirty datasets.
572 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
573 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
575 * We must not sync any non-MOS datasets twice, because
576 * we may have taken a snapshot of them. However, we
577 * may sync newly-created datasets on pass 2.
579 ASSERT(!list_link_active(&ds->ds_synced_link));
580 list_insert_tail(&synced_datasets, ds);
581 dsl_dataset_sync(ds, zio, tx);
583 VERIFY0(zio_wait(zio));
586 * We have written all of the accounted dirty data, so our
587 * dp_space_towrite should now be zero. However, some seldom-used
588 * code paths do not adhere to this (e.g. dbuf_undirty(), also
589 * rounding error in dbuf_write_physdone).
590 * Shore up the accounting of any dirtied space now.
592 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
595 * After the data blocks have been written (ensured by the zio_wait()
596 * above), update the user/group space accounting.
598 for (ds = list_head(&synced_datasets); ds != NULL;
599 ds = list_next(&synced_datasets, ds)) {
600 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
604 * Sync the datasets again to push out the changes due to
605 * userspace updates. This must be done before we process the
606 * sync tasks, so that any snapshots will have the correct
607 * user accounting information (and we won't get confused
608 * about which blocks are part of the snapshot).
610 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
611 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
612 ASSERT(list_link_active(&ds->ds_synced_link));
613 dmu_buf_rele(ds->ds_dbuf, ds);
614 dsl_dataset_sync(ds, zio, tx);
616 VERIFY0(zio_wait(zio));
619 * Now that the datasets have been completely synced, we can
620 * clean up our in-memory structures accumulated while syncing:
622 * - move dead blocks from the pending deadlist to the on-disk deadlist
623 * - release hold from dsl_dataset_dirty()
625 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
626 objset_t *os = ds->ds_objset;
627 bplist_iterate(&ds->ds_pending_deadlist,
628 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
629 ASSERT(!dmu_objset_is_dirty(os, txg));
630 dmu_buf_rele(ds->ds_dbuf, ds);
632 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
633 dsl_dir_sync(dd, tx);
637 * The MOS's space is accounted for in the pool/$MOS
638 * (dp_mos_dir). We can't modify the mos while we're syncing
639 * it, so we remember the deltas and apply them here.
641 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
642 dp->dp_mos_uncompressed_delta != 0) {
643 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
644 dp->dp_mos_used_delta,
645 dp->dp_mos_compressed_delta,
646 dp->dp_mos_uncompressed_delta, tx);
647 dp->dp_mos_used_delta = 0;
648 dp->dp_mos_compressed_delta = 0;
649 dp->dp_mos_uncompressed_delta = 0;
652 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
653 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
654 dsl_pool_sync_mos(dp, tx);
658 * If we modify a dataset in the same txg that we want to destroy it,
659 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
660 * dsl_dir_destroy_check() will fail if there are unexpected holds.
661 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
662 * and clearing the hold on it) before we process the sync_tasks.
663 * The MOS data dirtied by the sync_tasks will be synced on the next
666 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
667 dsl_sync_task_t *dst;
669 * No more sync tasks should have been added while we
672 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
673 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
674 dsl_sync_task_sync(dst, tx);
679 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
683 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
687 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
688 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
689 zil_clean(zilog, txg);
690 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
691 dmu_buf_rele(ds->ds_dbuf, zilog);
693 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
697 * TRUE if the current thread is the tx_sync_thread or if we
698 * are being called from SPA context during pool initialization.
701 dsl_pool_sync_context(dsl_pool_t *dp)
703 return (curthread == dp->dp_tx.tx_sync_thread ||
704 spa_is_initializing(dp->dp_spa));
708 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
710 uint64_t space, resv;
713 * If we're trying to assess whether it's OK to do a free,
714 * cut the reservation in half to allow forward progress
715 * (e.g. make it possible to rm(1) files from a full pool).
717 space = spa_get_dspace(dp->dp_spa);
718 resv = spa_get_slop_space(dp->dp_spa);
722 return (space - resv);
726 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
728 uint64_t delay_min_bytes =
729 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
732 mutex_enter(&dp->dp_lock);
733 if (dp->dp_dirty_total > zfs_dirty_data_sync)
735 rv = (dp->dp_dirty_total > delay_min_bytes);
736 mutex_exit(&dp->dp_lock);
741 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
744 mutex_enter(&dp->dp_lock);
745 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
746 dsl_pool_dirty_delta(dp, space);
747 mutex_exit(&dp->dp_lock);
752 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
754 ASSERT3S(space, >=, 0);
757 mutex_enter(&dp->dp_lock);
758 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
759 /* XXX writing something we didn't dirty? */
760 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
762 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
763 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
764 ASSERT3U(dp->dp_dirty_total, >=, space);
765 dsl_pool_dirty_delta(dp, -space);
766 mutex_exit(&dp->dp_lock);
771 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
774 dsl_dataset_t *ds, *prev = NULL;
777 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
781 while (ds->ds_phys->ds_prev_snap_obj != 0) {
782 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
785 dsl_dataset_rele(ds, FTAG);
789 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
791 dsl_dataset_rele(ds, FTAG);
797 prev = dp->dp_origin_snap;
800 * The $ORIGIN can't have any data, or the accounting
803 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
805 /* The origin doesn't get attached to itself */
806 if (ds->ds_object == prev->ds_object) {
807 dsl_dataset_rele(ds, FTAG);
811 dmu_buf_will_dirty(ds->ds_dbuf, tx);
812 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
813 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
815 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
816 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
818 dmu_buf_will_dirty(prev->ds_dbuf, tx);
819 prev->ds_phys->ds_num_children++;
821 if (ds->ds_phys->ds_next_snap_obj == 0) {
822 ASSERT(ds->ds_prev == NULL);
823 VERIFY0(dsl_dataset_hold_obj(dp,
824 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
828 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
829 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
831 if (prev->ds_phys->ds_next_clones_obj == 0) {
832 dmu_buf_will_dirty(prev->ds_dbuf, tx);
833 prev->ds_phys->ds_next_clones_obj =
834 zap_create(dp->dp_meta_objset,
835 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
837 VERIFY0(zap_add_int(dp->dp_meta_objset,
838 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
840 dsl_dataset_rele(ds, FTAG);
841 if (prev != dp->dp_origin_snap)
842 dsl_dataset_rele(prev, FTAG);
847 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
849 ASSERT(dmu_tx_is_syncing(tx));
850 ASSERT(dp->dp_origin_snap != NULL);
852 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
853 tx, DS_FIND_CHILDREN));
858 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
861 objset_t *mos = dp->dp_meta_objset;
863 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
864 dsl_dataset_t *origin;
866 VERIFY0(dsl_dataset_hold_obj(dp,
867 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
869 if (origin->ds_dir->dd_phys->dd_clones == 0) {
870 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
871 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
872 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
875 VERIFY0(zap_add_int(dp->dp_meta_objset,
876 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
878 dsl_dataset_rele(origin, FTAG);
884 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
886 ASSERT(dmu_tx_is_syncing(tx));
889 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
890 VERIFY0(dsl_pool_open_special_dir(dp,
891 FREE_DIR_NAME, &dp->dp_free_dir));
894 * We can't use bpobj_alloc(), because spa_version() still
895 * returns the old version, and we need a new-version bpobj with
896 * subobj support. So call dmu_object_alloc() directly.
898 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
899 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
900 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
901 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
902 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
904 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
905 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
909 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
914 ASSERT(dmu_tx_is_syncing(tx));
915 ASSERT(dp->dp_origin_snap == NULL);
916 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
918 /* create the origin dir, ds, & snap-ds */
919 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
921 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
922 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
923 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
924 dp, &dp->dp_origin_snap));
925 dsl_dataset_rele(ds, FTAG);
929 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
931 return (dp->dp_vnrele_taskq);
935 * Walk through the pool-wide zap object of temporary snapshot user holds
939 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
943 objset_t *mos = dp->dp_meta_objset;
944 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
949 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
951 holds = fnvlist_alloc();
953 for (zap_cursor_init(&zc, mos, zapobj);
954 zap_cursor_retrieve(&zc, &za) == 0;
955 zap_cursor_advance(&zc)) {
959 htag = strchr(za.za_name, '-');
962 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
963 tags = fnvlist_alloc();
964 fnvlist_add_boolean(tags, htag);
965 fnvlist_add_nvlist(holds, za.za_name, tags);
968 fnvlist_add_boolean(tags, htag);
971 dsl_dataset_user_release_tmp(dp, holds);
973 zap_cursor_fini(&zc);
977 * Create the pool-wide zap object for storing temporary snapshot holds.
980 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
982 objset_t *mos = dp->dp_meta_objset;
984 ASSERT(dp->dp_tmp_userrefs_obj == 0);
985 ASSERT(dmu_tx_is_syncing(tx));
987 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
988 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
992 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
993 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
995 objset_t *mos = dp->dp_meta_objset;
996 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1000 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1001 ASSERT(dmu_tx_is_syncing(tx));
1004 * If the pool was created prior to SPA_VERSION_USERREFS, the
1005 * zap object for temporary holds might not exist yet.
1009 dsl_pool_user_hold_create_obj(dp, tx);
1010 zapobj = dp->dp_tmp_userrefs_obj;
1012 return (SET_ERROR(ENOENT));
1016 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
1018 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
1020 error = zap_remove(mos, zapobj, name, tx);
1027 * Add a temporary hold for the given dataset object and tag.
1030 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1031 uint64_t now, dmu_tx_t *tx)
1033 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
1037 * Release a temporary hold for the given dataset object and tag.
1040 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1043 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
1048 * DSL Pool Configuration Lock
1050 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1051 * creation / destruction / rename / property setting). It must be held for
1052 * read to hold a dataset or dsl_dir. I.e. you must call
1053 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1054 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1055 * must be held continuously until all datasets and dsl_dirs are released.
1057 * The only exception to this rule is that if a "long hold" is placed on
1058 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1059 * is still held. The long hold will prevent the dataset from being
1060 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1061 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1062 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1064 * Legitimate long-holders (including owners) should be long-running, cancelable
1065 * tasks that should cause "zfs destroy" to fail. This includes DMU
1066 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1067 * "zfs send", and "zfs diff". There are several other long-holders whose
1068 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1070 * The usual formula for long-holding would be:
1072 * dsl_dataset_hold()
1073 * ... perform checks ...
1074 * dsl_dataset_long_hold()
1076 * ... perform long-running task ...
1077 * dsl_dataset_long_rele()
1078 * dsl_dataset_rele()
1080 * Note that when the long hold is released, the dataset is still held but
1081 * the pool is not held. The dataset may change arbitrarily during this time
1082 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1083 * dataset except release it.
1085 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1086 * or modifying operations.
1088 * Modifying operations should generally use dsl_sync_task(). The synctask
1089 * infrastructure enforces proper locking strategy with respect to the
1090 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1092 * Read-only operations will manually hold the pool, then the dataset, obtain
1093 * information from the dataset, then release the pool and dataset.
1094 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1099 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1104 error = spa_open(name, &spa, tag);
1106 *dp = spa_get_dsl(spa);
1107 dsl_pool_config_enter(*dp, tag);
1113 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1115 dsl_pool_config_exit(dp, tag);
1116 spa_close(dp->dp_spa, tag);
1120 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1123 * We use a "reentrant" reader-writer lock, but not reentrantly.
1125 * The rrwlock can (with the track_all flag) track all reading threads,
1126 * which is very useful for debugging which code path failed to release
1127 * the lock, and for verifying that the *current* thread does hold
1130 * (Unlike a rwlock, which knows that N threads hold it for
1131 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1132 * if any thread holds it for read, even if this thread doesn't).
1134 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1135 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1139 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1141 rrw_exit(&dp->dp_config_rwlock, tag);
1145 dsl_pool_config_held(dsl_pool_t *dp)
1147 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));