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, 2017 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]
27 * Copyright 2016 Nexenta Systems, Inc. All rights reserved.
30 #include <sys/dsl_pool.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_prop.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_synctask.h>
35 #include <sys/dsl_scan.h>
36 #include <sys/dnode.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/dmu_objset.h>
42 #include <sys/zfs_context.h>
43 #include <sys/fs/zfs.h>
44 #include <sys/zfs_znode.h>
45 #include <sys/spa_impl.h>
46 #include <sys/dsl_deadlist.h>
47 #include <sys/bptree.h>
48 #include <sys/zfeature.h>
49 #include <sys/zil_impl.h>
50 #include <sys/dsl_userhold.h>
52 #if defined(__FreeBSD__) && defined(_KERNEL)
53 #include <sys/types.h>
54 #include <sys/sysctl.h>
61 * ZFS must limit the rate of incoming writes to the rate at which it is able
62 * to sync data modifications to the backend storage. Throttling by too much
63 * creates an artificial limit; throttling by too little can only be sustained
64 * for short periods and would lead to highly lumpy performance. On a per-pool
65 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
66 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
67 * of dirty data decreases. When the amount of dirty data exceeds a
68 * predetermined threshold further modifications are blocked until the amount
69 * of dirty data decreases (as data is synced out).
71 * The limit on dirty data is tunable, and should be adjusted according to
72 * both the IO capacity and available memory of the system. The larger the
73 * window, the more ZFS is able to aggregate and amortize metadata (and data)
74 * changes. However, memory is a limited resource, and allowing for more dirty
75 * data comes at the cost of keeping other useful data in memory (for example
76 * ZFS data cached by the ARC).
80 * As buffers are modified dsl_pool_willuse_space() increments both the per-
81 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
82 * dirty space used; dsl_pool_dirty_space() decrements those values as data
83 * is synced out from dsl_pool_sync(). While only the poolwide value is
84 * relevant, the per-txg value is useful for debugging. The tunable
85 * zfs_dirty_data_max determines the dirty space limit. Once that value is
86 * exceeded, new writes are halted until space frees up.
88 * The zfs_dirty_data_sync tunable dictates the threshold at which we
89 * ensure that there is a txg syncing (see the comment in txg.c for a full
90 * description of transaction group stages).
92 * The IO scheduler uses both the dirty space limit and current amount of
93 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
94 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
96 * The delay is also calculated based on the amount of dirty data. See the
97 * comment above dmu_tx_delay() for details.
101 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
102 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
104 uint64_t zfs_dirty_data_max;
105 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
106 int zfs_dirty_data_max_percent = 10;
109 * If there is at least this much dirty data, push out a txg.
111 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
114 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
115 * and delay each transaction.
116 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
118 int zfs_delay_min_dirty_percent = 60;
121 * This controls how quickly the delay approaches infinity.
122 * Larger values cause it to delay more for a given amount of dirty data.
123 * Therefore larger values will cause there to be less dirty data for a
126 * For the smoothest delay, this value should be about 1 billion divided
127 * by the maximum number of operations per second. This will smoothly
128 * handle between 10x and 1/10th this number.
130 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
131 * multiply in dmu_tx_delay().
133 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
136 * This determines the number of threads used by the dp_sync_taskq.
138 int zfs_sync_taskq_batch_pct = 75;
140 #if defined(__FreeBSD__) && defined(_KERNEL)
142 extern int zfs_vdev_async_write_active_max_dirty_percent;
144 SYSCTL_DECL(_vfs_zfs);
146 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max, CTLFLAG_RWTUN,
147 &zfs_dirty_data_max, 0,
148 "The maximum amount of dirty data in bytes after which new writes are "
149 "halted until space becomes available");
151 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max_max, CTLFLAG_RDTUN,
152 &zfs_dirty_data_max_max, 0,
153 "The absolute cap on dirty_data_max when auto calculating");
155 static int sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS);
156 SYSCTL_PROC(_vfs_zfs, OID_AUTO, dirty_data_max_percent,
157 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
158 sysctl_zfs_dirty_data_max_percent, "I",
159 "The percent of physical memory used to auto calculate dirty_data_max");
161 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_sync, CTLFLAG_RWTUN,
162 &zfs_dirty_data_sync, 0,
163 "Force a txg if the number of dirty buffer bytes exceed this value");
165 static int sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS);
166 /* No zfs_delay_min_dirty_percent tunable due to limit requirements */
167 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_min_dirty_percent,
168 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(int),
169 sysctl_zfs_delay_min_dirty_percent, "I",
170 "The limit of outstanding dirty data before transactions are delayed");
172 static int sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS);
173 /* No zfs_delay_scale tunable due to limit requirements */
174 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_scale,
175 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
176 sysctl_zfs_delay_scale, "QU",
177 "Controls how quickly the delay approaches infinity");
180 sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS)
184 val = zfs_dirty_data_max_percent;
185 err = sysctl_handle_int(oidp, &val, 0, req);
186 if (err != 0 || req->newptr == NULL)
189 if (val < 0 || val > 100)
192 zfs_dirty_data_max_percent = val;
198 sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS)
202 val = zfs_delay_min_dirty_percent;
203 err = sysctl_handle_int(oidp, &val, 0, req);
204 if (err != 0 || req->newptr == NULL)
207 if (val < zfs_vdev_async_write_active_max_dirty_percent)
210 zfs_delay_min_dirty_percent = val;
216 sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS)
221 val = zfs_delay_scale;
222 err = sysctl_handle_64(oidp, &val, 0, req);
223 if (err != 0 || req->newptr == NULL)
226 if (val > UINT64_MAX / zfs_dirty_data_max)
229 zfs_delay_scale = val;
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 dsl_dir_phys(dp->dp_root_dir)->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, spa,
263 offsetof(dsl_dataset_t, ds_dirty_link));
264 txg_list_create(&dp->dp_dirty_zilogs, spa,
265 offsetof(zilog_t, zl_dirty_link));
266 txg_list_create(&dp->dp_dirty_dirs, spa,
267 offsetof(dsl_dir_t, dd_dirty_link));
268 txg_list_create(&dp->dp_sync_tasks, spa,
269 offsetof(dsl_sync_task_t, dst_node));
271 dp->dp_sync_taskq = taskq_create("dp_sync_taskq",
272 zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX,
273 TASKQ_THREADS_CPU_PCT);
275 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
276 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
278 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
285 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
288 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
290 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
291 &dp->dp_meta_objset);
301 dsl_pool_open(dsl_pool_t *dp)
308 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
309 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
310 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
311 &dp->dp_root_dir_obj);
315 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
316 NULL, dp, &dp->dp_root_dir);
320 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
324 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
325 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
328 err = dsl_dataset_hold_obj(dp,
329 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
331 err = dsl_dataset_hold_obj(dp,
332 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
333 &dp->dp_origin_snap);
334 dsl_dataset_rele(ds, FTAG);
336 dsl_dir_rele(dd, dp);
341 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
342 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
347 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
348 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
351 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
352 dp->dp_meta_objset, obj));
356 * Note: errors ignored, because the leak dir will not exist if we
357 * have not encountered a leak yet.
359 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
362 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
363 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
364 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
370 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
371 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
372 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
373 &dp->dp_empty_bpobj);
378 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
379 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
380 &dp->dp_tmp_userrefs_obj);
386 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
389 rrw_exit(&dp->dp_config_rwlock, FTAG);
394 dsl_pool_close(dsl_pool_t *dp)
397 * Drop our references from dsl_pool_open().
399 * Since we held the origin_snap from "syncing" context (which
400 * includes pool-opening context), it actually only got a "ref"
401 * and not a hold, so just drop that here.
403 if (dp->dp_origin_snap)
404 dsl_dataset_rele(dp->dp_origin_snap, dp);
406 dsl_dir_rele(dp->dp_mos_dir, dp);
408 dsl_dir_rele(dp->dp_free_dir, dp);
410 dsl_dir_rele(dp->dp_leak_dir, dp);
412 dsl_dir_rele(dp->dp_root_dir, dp);
414 bpobj_close(&dp->dp_free_bpobj);
416 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
417 if (dp->dp_meta_objset)
418 dmu_objset_evict(dp->dp_meta_objset);
420 txg_list_destroy(&dp->dp_dirty_datasets);
421 txg_list_destroy(&dp->dp_dirty_zilogs);
422 txg_list_destroy(&dp->dp_sync_tasks);
423 txg_list_destroy(&dp->dp_dirty_dirs);
425 taskq_destroy(dp->dp_sync_taskq);
428 * We can't set retry to TRUE since we're explicitly specifying
429 * a spa to flush. This is good enough; any missed buffers for
430 * this spa won't cause trouble, and they'll eventually fall
431 * out of the ARC just like any other unused buffer.
433 arc_flush(dp->dp_spa, FALSE);
437 dmu_buf_user_evict_wait();
439 rrw_destroy(&dp->dp_config_rwlock);
440 mutex_destroy(&dp->dp_lock);
441 taskq_destroy(dp->dp_vnrele_taskq);
443 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
444 kmem_free(dp, sizeof (dsl_pool_t));
448 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
451 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
452 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
457 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
459 /* create and open the MOS (meta-objset) */
460 dp->dp_meta_objset = dmu_objset_create_impl(spa,
461 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
463 /* create the pool directory */
464 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
465 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
468 /* Initialize scan structures */
469 VERIFY0(dsl_scan_init(dp, txg));
471 /* create and open the root dir */
472 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
473 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
474 NULL, dp, &dp->dp_root_dir));
476 /* create and open the meta-objset dir */
477 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
478 VERIFY0(dsl_pool_open_special_dir(dp,
479 MOS_DIR_NAME, &dp->dp_mos_dir));
481 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
482 /* create and open the free dir */
483 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
485 VERIFY0(dsl_pool_open_special_dir(dp,
486 FREE_DIR_NAME, &dp->dp_free_dir));
488 /* create and open the free_bplist */
489 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
490 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
491 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
492 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
493 dp->dp_meta_objset, obj));
496 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
497 dsl_pool_create_origin(dp, tx);
499 /* create the root dataset */
500 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
502 /* create the root objset */
503 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
504 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
505 os = dmu_objset_create_impl(dp->dp_spa, ds,
506 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
507 rrw_exit(&ds->ds_bp_rwlock, FTAG);
509 zfs_create_fs(os, kcred, zplprops, tx);
511 dsl_dataset_rele(ds, FTAG);
515 rrw_exit(&dp->dp_config_rwlock, FTAG);
521 * Account for the meta-objset space in its placeholder dsl_dir.
524 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
525 int64_t used, int64_t comp, int64_t uncomp)
527 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
528 mutex_enter(&dp->dp_lock);
529 dp->dp_mos_used_delta += used;
530 dp->dp_mos_compressed_delta += comp;
531 dp->dp_mos_uncompressed_delta += uncomp;
532 mutex_exit(&dp->dp_lock);
536 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
538 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
539 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
540 VERIFY0(zio_wait(zio));
541 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
542 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
546 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
548 ASSERT(MUTEX_HELD(&dp->dp_lock));
551 ASSERT3U(-delta, <=, dp->dp_dirty_total);
553 dp->dp_dirty_total += delta;
556 * Note: we signal even when increasing dp_dirty_total.
557 * This ensures forward progress -- each thread wakes the next waiter.
559 if (dp->dp_dirty_total < zfs_dirty_data_max)
560 cv_signal(&dp->dp_spaceavail_cv);
564 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
570 objset_t *mos = dp->dp_meta_objset;
571 list_t synced_datasets;
573 list_create(&synced_datasets, sizeof (dsl_dataset_t),
574 offsetof(dsl_dataset_t, ds_synced_link));
576 tx = dmu_tx_create_assigned(dp, txg);
579 * Write out all dirty blocks of dirty datasets.
581 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
582 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
584 * We must not sync any non-MOS datasets twice, because
585 * we may have taken a snapshot of them. However, we
586 * may sync newly-created datasets on pass 2.
588 ASSERT(!list_link_active(&ds->ds_synced_link));
589 list_insert_tail(&synced_datasets, ds);
590 dsl_dataset_sync(ds, zio, tx);
592 VERIFY0(zio_wait(zio));
595 * We have written all of the accounted dirty data, so our
596 * dp_space_towrite should now be zero. However, some seldom-used
597 * code paths do not adhere to this (e.g. dbuf_undirty(), also
598 * rounding error in dbuf_write_physdone).
599 * Shore up the accounting of any dirtied space now.
601 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
604 * Update the long range free counter after
605 * we're done syncing user data
607 mutex_enter(&dp->dp_lock);
608 ASSERT(spa_sync_pass(dp->dp_spa) == 1 ||
609 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0);
610 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0;
611 mutex_exit(&dp->dp_lock);
614 * After the data blocks have been written (ensured by the zio_wait()
615 * above), update the user/group space accounting. This happens
616 * in tasks dispatched to dp_sync_taskq, so wait for them before
619 for (ds = list_head(&synced_datasets); ds != NULL;
620 ds = list_next(&synced_datasets, ds)) {
621 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
623 taskq_wait(dp->dp_sync_taskq);
626 * Sync the datasets again to push out the changes due to
627 * userspace updates. This must be done before we process the
628 * sync tasks, so that any snapshots will have the correct
629 * user accounting information (and we won't get confused
630 * about which blocks are part of the snapshot).
632 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
633 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
634 ASSERT(list_link_active(&ds->ds_synced_link));
635 dmu_buf_rele(ds->ds_dbuf, ds);
636 dsl_dataset_sync(ds, zio, tx);
638 VERIFY0(zio_wait(zio));
641 * Now that the datasets have been completely synced, we can
642 * clean up our in-memory structures accumulated while syncing:
644 * - move dead blocks from the pending deadlist to the on-disk deadlist
645 * - release hold from dsl_dataset_dirty()
647 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
648 dsl_dataset_sync_done(ds, tx);
650 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
651 dsl_dir_sync(dd, tx);
655 * The MOS's space is accounted for in the pool/$MOS
656 * (dp_mos_dir). We can't modify the mos while we're syncing
657 * it, so we remember the deltas and apply them here.
659 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
660 dp->dp_mos_uncompressed_delta != 0) {
661 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
662 dp->dp_mos_used_delta,
663 dp->dp_mos_compressed_delta,
664 dp->dp_mos_uncompressed_delta, tx);
665 dp->dp_mos_used_delta = 0;
666 dp->dp_mos_compressed_delta = 0;
667 dp->dp_mos_uncompressed_delta = 0;
670 if (!multilist_is_empty(mos->os_dirty_dnodes[txg & TXG_MASK])) {
671 dsl_pool_sync_mos(dp, tx);
675 * If we modify a dataset in the same txg that we want to destroy it,
676 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
677 * dsl_dir_destroy_check() will fail if there are unexpected holds.
678 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
679 * and clearing the hold on it) before we process the sync_tasks.
680 * The MOS data dirtied by the sync_tasks will be synced on the next
683 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
684 dsl_sync_task_t *dst;
686 * No more sync tasks should have been added while we
689 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
690 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
691 dsl_sync_task_sync(dst, tx);
696 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
700 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
704 while (zilog = txg_list_head(&dp->dp_dirty_zilogs, txg)) {
705 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
707 * We don't remove the zilog from the dp_dirty_zilogs
708 * list until after we've cleaned it. This ensures that
709 * callers of zilog_is_dirty() receive an accurate
710 * answer when they are racing with the spa sync thread.
712 zil_clean(zilog, txg);
713 (void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg);
714 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
715 dmu_buf_rele(ds->ds_dbuf, zilog);
717 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
721 * TRUE if the current thread is the tx_sync_thread or if we
722 * are being called from SPA context during pool initialization.
725 dsl_pool_sync_context(dsl_pool_t *dp)
727 return (curthread == dp->dp_tx.tx_sync_thread ||
728 spa_is_initializing(dp->dp_spa) ||
729 taskq_member(dp->dp_sync_taskq, curthread));
733 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
735 uint64_t space, resv;
738 * If we're trying to assess whether it's OK to do a free,
739 * cut the reservation in half to allow forward progress
740 * (e.g. make it possible to rm(1) files from a full pool).
742 space = spa_get_dspace(dp->dp_spa);
743 resv = spa_get_slop_space(dp->dp_spa);
747 return (space - resv);
751 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
753 uint64_t delay_min_bytes =
754 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
757 mutex_enter(&dp->dp_lock);
758 if (dp->dp_dirty_total > zfs_dirty_data_sync)
760 rv = (dp->dp_dirty_total > delay_min_bytes);
761 mutex_exit(&dp->dp_lock);
766 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
769 mutex_enter(&dp->dp_lock);
770 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
771 dsl_pool_dirty_delta(dp, space);
772 mutex_exit(&dp->dp_lock);
777 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
779 ASSERT3S(space, >=, 0);
782 mutex_enter(&dp->dp_lock);
783 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
784 /* XXX writing something we didn't dirty? */
785 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
787 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
788 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
789 ASSERT3U(dp->dp_dirty_total, >=, space);
790 dsl_pool_dirty_delta(dp, -space);
791 mutex_exit(&dp->dp_lock);
796 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
799 dsl_dataset_t *ds, *prev = NULL;
802 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
806 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
807 err = dsl_dataset_hold_obj(dp,
808 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
810 dsl_dataset_rele(ds, FTAG);
814 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
816 dsl_dataset_rele(ds, FTAG);
822 prev = dp->dp_origin_snap;
825 * The $ORIGIN can't have any data, or the accounting
828 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
829 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
830 rrw_exit(&ds->ds_bp_rwlock, FTAG);
832 /* The origin doesn't get attached to itself */
833 if (ds->ds_object == prev->ds_object) {
834 dsl_dataset_rele(ds, FTAG);
838 dmu_buf_will_dirty(ds->ds_dbuf, tx);
839 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
840 dsl_dataset_phys(ds)->ds_prev_snap_txg =
841 dsl_dataset_phys(prev)->ds_creation_txg;
843 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
844 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
846 dmu_buf_will_dirty(prev->ds_dbuf, tx);
847 dsl_dataset_phys(prev)->ds_num_children++;
849 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
850 ASSERT(ds->ds_prev == NULL);
851 VERIFY0(dsl_dataset_hold_obj(dp,
852 dsl_dataset_phys(ds)->ds_prev_snap_obj,
857 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
858 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
860 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
861 dmu_buf_will_dirty(prev->ds_dbuf, tx);
862 dsl_dataset_phys(prev)->ds_next_clones_obj =
863 zap_create(dp->dp_meta_objset,
864 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
866 VERIFY0(zap_add_int(dp->dp_meta_objset,
867 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
869 dsl_dataset_rele(ds, FTAG);
870 if (prev != dp->dp_origin_snap)
871 dsl_dataset_rele(prev, FTAG);
876 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
878 ASSERT(dmu_tx_is_syncing(tx));
879 ASSERT(dp->dp_origin_snap != NULL);
881 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
882 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
887 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
890 objset_t *mos = dp->dp_meta_objset;
892 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
893 dsl_dataset_t *origin;
895 VERIFY0(dsl_dataset_hold_obj(dp,
896 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
898 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
899 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
900 dsl_dir_phys(origin->ds_dir)->dd_clones =
901 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
905 VERIFY0(zap_add_int(dp->dp_meta_objset,
906 dsl_dir_phys(origin->ds_dir)->dd_clones,
909 dsl_dataset_rele(origin, FTAG);
915 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
917 ASSERT(dmu_tx_is_syncing(tx));
920 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
921 VERIFY0(dsl_pool_open_special_dir(dp,
922 FREE_DIR_NAME, &dp->dp_free_dir));
925 * We can't use bpobj_alloc(), because spa_version() still
926 * returns the old version, and we need a new-version bpobj with
927 * subobj support. So call dmu_object_alloc() directly.
929 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
930 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
931 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
932 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
933 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
935 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
936 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
940 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
945 ASSERT(dmu_tx_is_syncing(tx));
946 ASSERT(dp->dp_origin_snap == NULL);
947 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
949 /* create the origin dir, ds, & snap-ds */
950 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
952 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
953 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
954 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
955 dp, &dp->dp_origin_snap));
956 dsl_dataset_rele(ds, FTAG);
960 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
962 return (dp->dp_vnrele_taskq);
966 * Walk through the pool-wide zap object of temporary snapshot user holds
970 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
974 objset_t *mos = dp->dp_meta_objset;
975 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
980 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
982 holds = fnvlist_alloc();
984 for (zap_cursor_init(&zc, mos, zapobj);
985 zap_cursor_retrieve(&zc, &za) == 0;
986 zap_cursor_advance(&zc)) {
990 htag = strchr(za.za_name, '-');
993 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
994 tags = fnvlist_alloc();
995 fnvlist_add_boolean(tags, htag);
996 fnvlist_add_nvlist(holds, za.za_name, tags);
999 fnvlist_add_boolean(tags, htag);
1002 dsl_dataset_user_release_tmp(dp, holds);
1003 fnvlist_free(holds);
1004 zap_cursor_fini(&zc);
1008 * Create the pool-wide zap object for storing temporary snapshot holds.
1011 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
1013 objset_t *mos = dp->dp_meta_objset;
1015 ASSERT(dp->dp_tmp_userrefs_obj == 0);
1016 ASSERT(dmu_tx_is_syncing(tx));
1018 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
1019 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
1023 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
1024 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
1026 objset_t *mos = dp->dp_meta_objset;
1027 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1031 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1032 ASSERT(dmu_tx_is_syncing(tx));
1035 * If the pool was created prior to SPA_VERSION_USERREFS, the
1036 * zap object for temporary holds might not exist yet.
1040 dsl_pool_user_hold_create_obj(dp, tx);
1041 zapobj = dp->dp_tmp_userrefs_obj;
1043 return (SET_ERROR(ENOENT));
1047 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
1049 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
1051 error = zap_remove(mos, zapobj, name, tx);
1058 * Add a temporary hold for the given dataset object and tag.
1061 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1062 uint64_t now, dmu_tx_t *tx)
1064 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
1068 * Release a temporary hold for the given dataset object and tag.
1071 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1074 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
1079 * DSL Pool Configuration Lock
1081 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1082 * creation / destruction / rename / property setting). It must be held for
1083 * read to hold a dataset or dsl_dir. I.e. you must call
1084 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1085 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1086 * must be held continuously until all datasets and dsl_dirs are released.
1088 * The only exception to this rule is that if a "long hold" is placed on
1089 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1090 * is still held. The long hold will prevent the dataset from being
1091 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1092 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1093 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1095 * Legitimate long-holders (including owners) should be long-running, cancelable
1096 * tasks that should cause "zfs destroy" to fail. This includes DMU
1097 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1098 * "zfs send", and "zfs diff". There are several other long-holders whose
1099 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1101 * The usual formula for long-holding would be:
1103 * dsl_dataset_hold()
1104 * ... perform checks ...
1105 * dsl_dataset_long_hold()
1107 * ... perform long-running task ...
1108 * dsl_dataset_long_rele()
1109 * dsl_dataset_rele()
1111 * Note that when the long hold is released, the dataset is still held but
1112 * the pool is not held. The dataset may change arbitrarily during this time
1113 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1114 * dataset except release it.
1116 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1117 * or modifying operations.
1119 * Modifying operations should generally use dsl_sync_task(). The synctask
1120 * infrastructure enforces proper locking strategy with respect to the
1121 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1123 * Read-only operations will manually hold the pool, then the dataset, obtain
1124 * information from the dataset, then release the pool and dataset.
1125 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1130 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1135 error = spa_open(name, &spa, tag);
1137 *dp = spa_get_dsl(spa);
1138 dsl_pool_config_enter(*dp, tag);
1144 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1146 dsl_pool_config_exit(dp, tag);
1147 spa_close(dp->dp_spa, tag);
1151 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1154 * We use a "reentrant" reader-writer lock, but not reentrantly.
1156 * The rrwlock can (with the track_all flag) track all reading threads,
1157 * which is very useful for debugging which code path failed to release
1158 * the lock, and for verifying that the *current* thread does hold
1161 * (Unlike a rwlock, which knows that N threads hold it for
1162 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1163 * if any thread holds it for read, even if this thread doesn't).
1165 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1166 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1170 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1172 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1173 rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1177 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1179 rrw_exit(&dp->dp_config_rwlock, tag);
1183 dsl_pool_config_held(dsl_pool_t *dp)
1185 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1189 dsl_pool_config_held_writer(dsl_pool_t *dp)
1191 return (RRW_WRITE_HELD(&dp->dp_config_rwlock));