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, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26 * Copyright (c) 2014 Integros [integros.com]
29 #include <sys/dsl_pool.h>
30 #include <sys/dsl_dataset.h>
31 #include <sys/dsl_prop.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dsl_synctask.h>
34 #include <sys/dsl_scan.h>
35 #include <sys/dnode.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/dmu_objset.h>
41 #include <sys/zfs_context.h>
42 #include <sys/fs/zfs.h>
43 #include <sys/zfs_znode.h>
44 #include <sys/spa_impl.h>
45 #include <sys/dsl_deadlist.h>
46 #include <sys/bptree.h>
47 #include <sys/zfeature.h>
48 #include <sys/zil_impl.h>
49 #include <sys/dsl_userhold.h>
52 #include <sys/types.h>
53 #include <sys/sysctl.h>
60 * ZFS must limit the rate of incoming writes to the rate at which it is able
61 * to sync data modifications to the backend storage. Throttling by too much
62 * creates an artificial limit; throttling by too little can only be sustained
63 * for short periods and would lead to highly lumpy performance. On a per-pool
64 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
65 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
66 * of dirty data decreases. When the amount of dirty data exceeds a
67 * predetermined threshold further modifications are blocked until the amount
68 * of dirty data decreases (as data is synced out).
70 * The limit on dirty data is tunable, and should be adjusted according to
71 * both the IO capacity and available memory of the system. The larger the
72 * window, the more ZFS is able to aggregate and amortize metadata (and data)
73 * changes. However, memory is a limited resource, and allowing for more dirty
74 * data comes at the cost of keeping other useful data in memory (for example
75 * ZFS data cached by the ARC).
79 * As buffers are modified dsl_pool_willuse_space() increments both the per-
80 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
81 * dirty space used; dsl_pool_dirty_space() decrements those values as data
82 * is synced out from dsl_pool_sync(). While only the poolwide value is
83 * relevant, the per-txg value is useful for debugging. The tunable
84 * zfs_dirty_data_max determines the dirty space limit. Once that value is
85 * exceeded, new writes are halted until space frees up.
87 * The zfs_dirty_data_sync tunable dictates the threshold at which we
88 * ensure that there is a txg syncing (see the comment in txg.c for a full
89 * description of transaction group stages).
91 * The IO scheduler uses both the dirty space limit and current amount of
92 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
93 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
95 * The delay is also calculated based on the amount of dirty data. See the
96 * comment above dmu_tx_delay() for details.
100 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
101 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
103 uint64_t zfs_dirty_data_max;
104 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
105 int zfs_dirty_data_max_percent = 10;
108 * If there is at least this much dirty data, push out a txg.
110 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
113 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
114 * and delay each transaction.
115 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
117 int zfs_delay_min_dirty_percent = 60;
120 * This controls how quickly the delay approaches infinity.
121 * Larger values cause it to delay more for a given amount of dirty data.
122 * Therefore larger values will cause there to be less dirty data for a
125 * For the smoothest delay, this value should be about 1 billion divided
126 * by the maximum number of operations per second. This will smoothly
127 * handle between 10x and 1/10th this number.
129 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
130 * multiply in dmu_tx_delay().
132 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
137 extern int zfs_vdev_async_write_active_max_dirty_percent;
139 SYSCTL_DECL(_vfs_zfs);
141 TUNABLE_QUAD("vfs.zfs.dirty_data_max", &zfs_dirty_data_max);
142 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max, CTLFLAG_RWTUN,
143 &zfs_dirty_data_max, 0,
144 "The maximum amount of dirty data in bytes after which new writes are "
145 "halted until space becomes available");
147 TUNABLE_QUAD("vfs.zfs.dirty_data_max_max", &zfs_dirty_data_max_max);
148 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max_max, CTLFLAG_RDTUN,
149 &zfs_dirty_data_max_max, 0,
150 "The absolute cap on dirty_data_max when auto calculating");
152 TUNABLE_INT("vfs.zfs.dirty_data_max_percent", &zfs_dirty_data_max_percent);
153 static int sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS);
154 SYSCTL_PROC(_vfs_zfs, OID_AUTO, dirty_data_max_percent,
155 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
156 sysctl_zfs_dirty_data_max_percent, "I",
157 "The percent of physical memory used to auto calculate dirty_data_max");
159 TUNABLE_QUAD("vfs.zfs.dirty_data_sync", &zfs_dirty_data_sync);
160 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_sync, CTLFLAG_RWTUN,
161 &zfs_dirty_data_sync, 0,
162 "Force a txg if the number of dirty buffer bytes exceed this value");
164 static int sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS);
165 /* No zfs_delay_min_dirty_percent tunable due to limit requirements */
166 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_min_dirty_percent,
167 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(int),
168 sysctl_zfs_delay_min_dirty_percent, "I",
169 "The limit of outstanding dirty data before transations are delayed");
171 static int sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS);
172 /* No zfs_delay_scale tunable due to limit requirements */
173 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_scale,
174 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
175 sysctl_zfs_delay_scale, "QU",
176 "Controls how quickly the delay approaches infinity");
179 sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS)
183 val = zfs_dirty_data_max_percent;
184 err = sysctl_handle_int(oidp, &val, 0, req);
185 if (err != 0 || req->newptr == NULL)
188 if (val < 0 || val > 100)
191 zfs_dirty_data_max_percent = val;
197 sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS)
201 val = zfs_delay_min_dirty_percent;
202 err = sysctl_handle_int(oidp, &val, 0, req);
203 if (err != 0 || req->newptr == NULL)
206 if (val < zfs_vdev_async_write_active_max_dirty_percent)
209 zfs_delay_min_dirty_percent = val;
215 sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS)
220 val = zfs_delay_scale;
221 err = sysctl_handle_64(oidp, &val, 0, req);
222 if (err != 0 || req->newptr == NULL)
225 if (val > UINT64_MAX / zfs_dirty_data_max)
228 zfs_delay_scale = val;
234 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
235 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
238 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
243 err = zap_lookup(dp->dp_meta_objset,
244 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
245 name, sizeof (obj), 1, &obj);
249 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
253 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
256 blkptr_t *bp = spa_get_rootblkptr(spa);
258 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
260 dp->dp_meta_rootbp = *bp;
261 rrw_init(&dp->dp_config_rwlock, B_TRUE);
264 txg_list_create(&dp->dp_dirty_datasets,
265 offsetof(dsl_dataset_t, ds_dirty_link));
266 txg_list_create(&dp->dp_dirty_zilogs,
267 offsetof(zilog_t, zl_dirty_link));
268 txg_list_create(&dp->dp_dirty_dirs,
269 offsetof(dsl_dir_t, dd_dirty_link));
270 txg_list_create(&dp->dp_sync_tasks,
271 offsetof(dsl_sync_task_t, dst_node));
273 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
274 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
276 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
283 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
286 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
288 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
289 &dp->dp_meta_objset);
299 dsl_pool_open(dsl_pool_t *dp)
306 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
307 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
308 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
309 &dp->dp_root_dir_obj);
313 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
314 NULL, dp, &dp->dp_root_dir);
318 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
322 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
323 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
326 err = dsl_dataset_hold_obj(dp,
327 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
329 err = dsl_dataset_hold_obj(dp,
330 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
331 &dp->dp_origin_snap);
332 dsl_dataset_rele(ds, FTAG);
334 dsl_dir_rele(dd, dp);
339 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
340 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
345 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
346 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
349 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
350 dp->dp_meta_objset, obj));
354 * Note: errors ignored, because the leak dir will not exist if we
355 * have not encountered a leak yet.
357 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
360 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
361 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
362 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
368 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
369 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
370 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
371 &dp->dp_empty_bpobj);
376 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
377 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
378 &dp->dp_tmp_userrefs_obj);
384 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
387 rrw_exit(&dp->dp_config_rwlock, FTAG);
392 dsl_pool_close(dsl_pool_t *dp)
395 * Drop our references from dsl_pool_open().
397 * Since we held the origin_snap from "syncing" context (which
398 * includes pool-opening context), it actually only got a "ref"
399 * and not a hold, so just drop that here.
401 if (dp->dp_origin_snap)
402 dsl_dataset_rele(dp->dp_origin_snap, dp);
404 dsl_dir_rele(dp->dp_mos_dir, dp);
406 dsl_dir_rele(dp->dp_free_dir, dp);
408 dsl_dir_rele(dp->dp_leak_dir, dp);
410 dsl_dir_rele(dp->dp_root_dir, dp);
412 bpobj_close(&dp->dp_free_bpobj);
414 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
415 if (dp->dp_meta_objset)
416 dmu_objset_evict(dp->dp_meta_objset);
418 txg_list_destroy(&dp->dp_dirty_datasets);
419 txg_list_destroy(&dp->dp_dirty_zilogs);
420 txg_list_destroy(&dp->dp_sync_tasks);
421 txg_list_destroy(&dp->dp_dirty_dirs);
424 * We can't set retry to TRUE since we're explicitly specifying
425 * a spa to flush. This is good enough; any missed buffers for
426 * this spa won't cause trouble, and they'll eventually fall
427 * out of the ARC just like any other unused buffer.
429 arc_flush(dp->dp_spa, FALSE);
433 dmu_buf_user_evict_wait();
435 rrw_destroy(&dp->dp_config_rwlock);
436 mutex_destroy(&dp->dp_lock);
437 taskq_destroy(dp->dp_vnrele_taskq);
439 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
440 kmem_free(dp, sizeof (dsl_pool_t));
444 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
447 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
448 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
453 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
455 /* create and open the MOS (meta-objset) */
456 dp->dp_meta_objset = dmu_objset_create_impl(spa,
457 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
459 /* create the pool directory */
460 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
461 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
464 /* Initialize scan structures */
465 VERIFY0(dsl_scan_init(dp, txg));
467 /* create and open the root dir */
468 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
469 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
470 NULL, dp, &dp->dp_root_dir));
472 /* create and open the meta-objset dir */
473 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
474 VERIFY0(dsl_pool_open_special_dir(dp,
475 MOS_DIR_NAME, &dp->dp_mos_dir));
477 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
478 /* create and open the free dir */
479 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
481 VERIFY0(dsl_pool_open_special_dir(dp,
482 FREE_DIR_NAME, &dp->dp_free_dir));
484 /* create and open the free_bplist */
485 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
486 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
487 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
488 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
489 dp->dp_meta_objset, obj));
492 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
493 dsl_pool_create_origin(dp, tx);
495 /* create the root dataset */
496 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
498 /* create the root objset */
499 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
500 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
501 os = dmu_objset_create_impl(dp->dp_spa, ds,
502 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
503 rrw_exit(&ds->ds_bp_rwlock, FTAG);
505 zfs_create_fs(os, kcred, zplprops, tx);
507 dsl_dataset_rele(ds, FTAG);
511 rrw_exit(&dp->dp_config_rwlock, FTAG);
517 * Account for the meta-objset space in its placeholder dsl_dir.
520 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
521 int64_t used, int64_t comp, int64_t uncomp)
523 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
524 mutex_enter(&dp->dp_lock);
525 dp->dp_mos_used_delta += used;
526 dp->dp_mos_compressed_delta += comp;
527 dp->dp_mos_uncompressed_delta += uncomp;
528 mutex_exit(&dp->dp_lock);
532 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
534 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
535 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
536 VERIFY0(zio_wait(zio));
537 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
538 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
542 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
544 ASSERT(MUTEX_HELD(&dp->dp_lock));
547 ASSERT3U(-delta, <=, dp->dp_dirty_total);
549 dp->dp_dirty_total += delta;
552 * Note: we signal even when increasing dp_dirty_total.
553 * This ensures forward progress -- each thread wakes the next waiter.
555 if (dp->dp_dirty_total <= zfs_dirty_data_max)
556 cv_signal(&dp->dp_spaceavail_cv);
560 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
566 objset_t *mos = dp->dp_meta_objset;
567 list_t synced_datasets;
569 list_create(&synced_datasets, sizeof (dsl_dataset_t),
570 offsetof(dsl_dataset_t, ds_synced_link));
572 tx = dmu_tx_create_assigned(dp, txg);
575 * Write out all dirty blocks of dirty datasets.
577 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
578 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
580 * We must not sync any non-MOS datasets twice, because
581 * we may have taken a snapshot of them. However, we
582 * may sync newly-created datasets on pass 2.
584 ASSERT(!list_link_active(&ds->ds_synced_link));
585 list_insert_tail(&synced_datasets, ds);
586 dsl_dataset_sync(ds, zio, tx);
588 VERIFY0(zio_wait(zio));
591 * We have written all of the accounted dirty data, so our
592 * dp_space_towrite should now be zero. However, some seldom-used
593 * code paths do not adhere to this (e.g. dbuf_undirty(), also
594 * rounding error in dbuf_write_physdone).
595 * Shore up the accounting of any dirtied space now.
597 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
600 * After the data blocks have been written (ensured by the zio_wait()
601 * above), update the user/group space accounting.
603 for (ds = list_head(&synced_datasets); ds != NULL;
604 ds = list_next(&synced_datasets, ds)) {
605 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
609 * Sync the datasets again to push out the changes due to
610 * userspace updates. This must be done before we process the
611 * sync tasks, so that any snapshots will have the correct
612 * user accounting information (and we won't get confused
613 * about which blocks are part of the snapshot).
615 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
616 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
617 ASSERT(list_link_active(&ds->ds_synced_link));
618 dmu_buf_rele(ds->ds_dbuf, ds);
619 dsl_dataset_sync(ds, zio, tx);
621 VERIFY0(zio_wait(zio));
624 * Now that the datasets have been completely synced, we can
625 * clean up our in-memory structures accumulated while syncing:
627 * - move dead blocks from the pending deadlist to the on-disk deadlist
628 * - release hold from dsl_dataset_dirty()
630 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
631 dsl_dataset_sync_done(ds, tx);
633 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
634 dsl_dir_sync(dd, tx);
638 * The MOS's space is accounted for in the pool/$MOS
639 * (dp_mos_dir). We can't modify the mos while we're syncing
640 * it, so we remember the deltas and apply them here.
642 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
643 dp->dp_mos_uncompressed_delta != 0) {
644 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
645 dp->dp_mos_used_delta,
646 dp->dp_mos_compressed_delta,
647 dp->dp_mos_uncompressed_delta, tx);
648 dp->dp_mos_used_delta = 0;
649 dp->dp_mos_compressed_delta = 0;
650 dp->dp_mos_uncompressed_delta = 0;
653 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
654 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
655 dsl_pool_sync_mos(dp, tx);
659 * If we modify a dataset in the same txg that we want to destroy it,
660 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
661 * dsl_dir_destroy_check() will fail if there are unexpected holds.
662 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
663 * and clearing the hold on it) before we process the sync_tasks.
664 * The MOS data dirtied by the sync_tasks will be synced on the next
667 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
668 dsl_sync_task_t *dst;
670 * No more sync tasks should have been added while we
673 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
674 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
675 dsl_sync_task_sync(dst, tx);
680 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
684 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
688 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
689 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
690 zil_clean(zilog, txg);
691 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
692 dmu_buf_rele(ds->ds_dbuf, zilog);
694 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
698 * TRUE if the current thread is the tx_sync_thread or if we
699 * are being called from SPA context during pool initialization.
702 dsl_pool_sync_context(dsl_pool_t *dp)
704 return (curthread == dp->dp_tx.tx_sync_thread ||
705 spa_is_initializing(dp->dp_spa));
709 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
711 uint64_t space, resv;
714 * If we're trying to assess whether it's OK to do a free,
715 * cut the reservation in half to allow forward progress
716 * (e.g. make it possible to rm(1) files from a full pool).
718 space = spa_get_dspace(dp->dp_spa);
719 resv = spa_get_slop_space(dp->dp_spa);
723 return (space - resv);
727 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
729 uint64_t delay_min_bytes =
730 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
733 mutex_enter(&dp->dp_lock);
734 if (dp->dp_dirty_total > zfs_dirty_data_sync)
736 rv = (dp->dp_dirty_total > delay_min_bytes);
737 mutex_exit(&dp->dp_lock);
742 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
745 mutex_enter(&dp->dp_lock);
746 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
747 dsl_pool_dirty_delta(dp, space);
748 mutex_exit(&dp->dp_lock);
753 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
755 ASSERT3S(space, >=, 0);
758 mutex_enter(&dp->dp_lock);
759 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
760 /* XXX writing something we didn't dirty? */
761 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
763 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
764 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
765 ASSERT3U(dp->dp_dirty_total, >=, space);
766 dsl_pool_dirty_delta(dp, -space);
767 mutex_exit(&dp->dp_lock);
772 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
775 dsl_dataset_t *ds, *prev = NULL;
778 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
782 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
783 err = dsl_dataset_hold_obj(dp,
784 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
786 dsl_dataset_rele(ds, FTAG);
790 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
792 dsl_dataset_rele(ds, FTAG);
798 prev = dp->dp_origin_snap;
801 * The $ORIGIN can't have any data, or the accounting
804 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
805 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
806 rrw_exit(&ds->ds_bp_rwlock, FTAG);
808 /* The origin doesn't get attached to itself */
809 if (ds->ds_object == prev->ds_object) {
810 dsl_dataset_rele(ds, FTAG);
814 dmu_buf_will_dirty(ds->ds_dbuf, tx);
815 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
816 dsl_dataset_phys(ds)->ds_prev_snap_txg =
817 dsl_dataset_phys(prev)->ds_creation_txg;
819 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
820 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
822 dmu_buf_will_dirty(prev->ds_dbuf, tx);
823 dsl_dataset_phys(prev)->ds_num_children++;
825 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
826 ASSERT(ds->ds_prev == NULL);
827 VERIFY0(dsl_dataset_hold_obj(dp,
828 dsl_dataset_phys(ds)->ds_prev_snap_obj,
833 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
834 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
836 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
837 dmu_buf_will_dirty(prev->ds_dbuf, tx);
838 dsl_dataset_phys(prev)->ds_next_clones_obj =
839 zap_create(dp->dp_meta_objset,
840 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
842 VERIFY0(zap_add_int(dp->dp_meta_objset,
843 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
845 dsl_dataset_rele(ds, FTAG);
846 if (prev != dp->dp_origin_snap)
847 dsl_dataset_rele(prev, FTAG);
852 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
854 ASSERT(dmu_tx_is_syncing(tx));
855 ASSERT(dp->dp_origin_snap != NULL);
857 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
858 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
863 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
866 objset_t *mos = dp->dp_meta_objset;
868 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
869 dsl_dataset_t *origin;
871 VERIFY0(dsl_dataset_hold_obj(dp,
872 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
874 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
875 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
876 dsl_dir_phys(origin->ds_dir)->dd_clones =
877 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
881 VERIFY0(zap_add_int(dp->dp_meta_objset,
882 dsl_dir_phys(origin->ds_dir)->dd_clones,
885 dsl_dataset_rele(origin, FTAG);
891 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
893 ASSERT(dmu_tx_is_syncing(tx));
896 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
897 VERIFY0(dsl_pool_open_special_dir(dp,
898 FREE_DIR_NAME, &dp->dp_free_dir));
901 * We can't use bpobj_alloc(), because spa_version() still
902 * returns the old version, and we need a new-version bpobj with
903 * subobj support. So call dmu_object_alloc() directly.
905 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
906 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
907 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
908 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
909 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
911 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
912 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
916 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
921 ASSERT(dmu_tx_is_syncing(tx));
922 ASSERT(dp->dp_origin_snap == NULL);
923 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
925 /* create the origin dir, ds, & snap-ds */
926 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
928 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
929 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
930 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
931 dp, &dp->dp_origin_snap));
932 dsl_dataset_rele(ds, FTAG);
936 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
938 return (dp->dp_vnrele_taskq);
942 * Walk through the pool-wide zap object of temporary snapshot user holds
946 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
950 objset_t *mos = dp->dp_meta_objset;
951 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
956 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
958 holds = fnvlist_alloc();
960 for (zap_cursor_init(&zc, mos, zapobj);
961 zap_cursor_retrieve(&zc, &za) == 0;
962 zap_cursor_advance(&zc)) {
966 htag = strchr(za.za_name, '-');
969 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
970 tags = fnvlist_alloc();
971 fnvlist_add_boolean(tags, htag);
972 fnvlist_add_nvlist(holds, za.za_name, tags);
975 fnvlist_add_boolean(tags, htag);
978 dsl_dataset_user_release_tmp(dp, holds);
980 zap_cursor_fini(&zc);
984 * Create the pool-wide zap object for storing temporary snapshot holds.
987 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
989 objset_t *mos = dp->dp_meta_objset;
991 ASSERT(dp->dp_tmp_userrefs_obj == 0);
992 ASSERT(dmu_tx_is_syncing(tx));
994 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
995 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
999 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
1000 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
1002 objset_t *mos = dp->dp_meta_objset;
1003 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1007 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1008 ASSERT(dmu_tx_is_syncing(tx));
1011 * If the pool was created prior to SPA_VERSION_USERREFS, the
1012 * zap object for temporary holds might not exist yet.
1016 dsl_pool_user_hold_create_obj(dp, tx);
1017 zapobj = dp->dp_tmp_userrefs_obj;
1019 return (SET_ERROR(ENOENT));
1023 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
1025 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
1027 error = zap_remove(mos, zapobj, name, tx);
1034 * Add a temporary hold for the given dataset object and tag.
1037 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1038 uint64_t now, dmu_tx_t *tx)
1040 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
1044 * Release a temporary hold for the given dataset object and tag.
1047 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1050 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
1055 * DSL Pool Configuration Lock
1057 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1058 * creation / destruction / rename / property setting). It must be held for
1059 * read to hold a dataset or dsl_dir. I.e. you must call
1060 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1061 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1062 * must be held continuously until all datasets and dsl_dirs are released.
1064 * The only exception to this rule is that if a "long hold" is placed on
1065 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1066 * is still held. The long hold will prevent the dataset from being
1067 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1068 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1069 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1071 * Legitimate long-holders (including owners) should be long-running, cancelable
1072 * tasks that should cause "zfs destroy" to fail. This includes DMU
1073 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1074 * "zfs send", and "zfs diff". There are several other long-holders whose
1075 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1077 * The usual formula for long-holding would be:
1079 * dsl_dataset_hold()
1080 * ... perform checks ...
1081 * dsl_dataset_long_hold()
1083 * ... perform long-running task ...
1084 * dsl_dataset_long_rele()
1085 * dsl_dataset_rele()
1087 * Note that when the long hold is released, the dataset is still held but
1088 * the pool is not held. The dataset may change arbitrarily during this time
1089 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1090 * dataset except release it.
1092 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1093 * or modifying operations.
1095 * Modifying operations should generally use dsl_sync_task(). The synctask
1096 * infrastructure enforces proper locking strategy with respect to the
1097 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1099 * Read-only operations will manually hold the pool, then the dataset, obtain
1100 * information from the dataset, then release the pool and dataset.
1101 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1106 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1111 error = spa_open(name, &spa, tag);
1113 *dp = spa_get_dsl(spa);
1114 dsl_pool_config_enter(*dp, tag);
1120 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1122 dsl_pool_config_exit(dp, tag);
1123 spa_close(dp->dp_spa, tag);
1127 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1130 * We use a "reentrant" reader-writer lock, but not reentrantly.
1132 * The rrwlock can (with the track_all flag) track all reading threads,
1133 * which is very useful for debugging which code path failed to release
1134 * the lock, and for verifying that the *current* thread does hold
1137 * (Unlike a rwlock, which knows that N threads hold it for
1138 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1139 * if any thread holds it for read, even if this thread doesn't).
1141 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1142 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1146 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1148 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1149 rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1153 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1155 rrw_exit(&dp->dp_config_rwlock, tag);
1159 dsl_pool_config_held(dsl_pool_t *dp)
1161 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1165 dsl_pool_config_held_writer(dsl_pool_t *dp)
1167 return (RRW_WRITE_HELD(&dp->dp_config_rwlock));