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, 2016 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 #if defined(__FreeBSD__) && defined(_KERNEL)
138 extern int zfs_vdev_async_write_active_max_dirty_percent;
140 SYSCTL_DECL(_vfs_zfs);
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 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max_max, CTLFLAG_RDTUN,
148 &zfs_dirty_data_max_max, 0,
149 "The absolute cap on dirty_data_max when auto calculating");
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 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_sync, CTLFLAG_RWTUN,
158 &zfs_dirty_data_sync, 0,
159 "Force a txg if the number of dirty buffer bytes exceed this value");
161 static int sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS);
162 /* No zfs_delay_min_dirty_percent tunable due to limit requirements */
163 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_min_dirty_percent,
164 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(int),
165 sysctl_zfs_delay_min_dirty_percent, "I",
166 "The limit of outstanding dirty data before transations are delayed");
168 static int sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS);
169 /* No zfs_delay_scale tunable due to limit requirements */
170 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_scale,
171 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
172 sysctl_zfs_delay_scale, "QU",
173 "Controls how quickly the delay approaches infinity");
176 sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS)
180 val = zfs_dirty_data_max_percent;
181 err = sysctl_handle_int(oidp, &val, 0, req);
182 if (err != 0 || req->newptr == NULL)
185 if (val < 0 || val > 100)
188 zfs_dirty_data_max_percent = val;
194 sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS)
198 val = zfs_delay_min_dirty_percent;
199 err = sysctl_handle_int(oidp, &val, 0, req);
200 if (err != 0 || req->newptr == NULL)
203 if (val < zfs_vdev_async_write_active_max_dirty_percent)
206 zfs_delay_min_dirty_percent = val;
212 sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS)
217 val = zfs_delay_scale;
218 err = sysctl_handle_64(oidp, &val, 0, req);
219 if (err != 0 || req->newptr == NULL)
222 if (val > UINT64_MAX / zfs_dirty_data_max)
225 zfs_delay_scale = val;
231 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
232 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
235 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
240 err = zap_lookup(dp->dp_meta_objset,
241 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
242 name, sizeof (obj), 1, &obj);
246 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
250 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
253 blkptr_t *bp = spa_get_rootblkptr(spa);
255 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
257 dp->dp_meta_rootbp = *bp;
258 rrw_init(&dp->dp_config_rwlock, B_TRUE);
261 txg_list_create(&dp->dp_dirty_datasets,
262 offsetof(dsl_dataset_t, ds_dirty_link));
263 txg_list_create(&dp->dp_dirty_zilogs,
264 offsetof(zilog_t, zl_dirty_link));
265 txg_list_create(&dp->dp_dirty_dirs,
266 offsetof(dsl_dir_t, dd_dirty_link));
267 txg_list_create(&dp->dp_sync_tasks,
268 offsetof(dsl_sync_task_t, dst_node));
270 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
271 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
273 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
280 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
283 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
285 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
286 &dp->dp_meta_objset);
296 dsl_pool_open(dsl_pool_t *dp)
303 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
304 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
305 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
306 &dp->dp_root_dir_obj);
310 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
311 NULL, dp, &dp->dp_root_dir);
315 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
319 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
320 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
323 err = dsl_dataset_hold_obj(dp,
324 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
326 err = dsl_dataset_hold_obj(dp,
327 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
328 &dp->dp_origin_snap);
329 dsl_dataset_rele(ds, FTAG);
331 dsl_dir_rele(dd, dp);
336 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
337 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
342 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
343 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
346 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
347 dp->dp_meta_objset, obj));
351 * Note: errors ignored, because the leak dir will not exist if we
352 * have not encountered a leak yet.
354 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
357 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
358 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
359 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
365 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
366 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
367 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
368 &dp->dp_empty_bpobj);
373 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
374 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
375 &dp->dp_tmp_userrefs_obj);
381 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
384 rrw_exit(&dp->dp_config_rwlock, FTAG);
389 dsl_pool_close(dsl_pool_t *dp)
392 * Drop our references from dsl_pool_open().
394 * Since we held the origin_snap from "syncing" context (which
395 * includes pool-opening context), it actually only got a "ref"
396 * and not a hold, so just drop that here.
398 if (dp->dp_origin_snap)
399 dsl_dataset_rele(dp->dp_origin_snap, dp);
401 dsl_dir_rele(dp->dp_mos_dir, dp);
403 dsl_dir_rele(dp->dp_free_dir, dp);
405 dsl_dir_rele(dp->dp_leak_dir, dp);
407 dsl_dir_rele(dp->dp_root_dir, dp);
409 bpobj_close(&dp->dp_free_bpobj);
411 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
412 if (dp->dp_meta_objset)
413 dmu_objset_evict(dp->dp_meta_objset);
415 txg_list_destroy(&dp->dp_dirty_datasets);
416 txg_list_destroy(&dp->dp_dirty_zilogs);
417 txg_list_destroy(&dp->dp_sync_tasks);
418 txg_list_destroy(&dp->dp_dirty_dirs);
421 * We can't set retry to TRUE since we're explicitly specifying
422 * a spa to flush. This is good enough; any missed buffers for
423 * this spa won't cause trouble, and they'll eventually fall
424 * out of the ARC just like any other unused buffer.
426 arc_flush(dp->dp_spa, FALSE);
430 dmu_buf_user_evict_wait();
432 rrw_destroy(&dp->dp_config_rwlock);
433 mutex_destroy(&dp->dp_lock);
434 taskq_destroy(dp->dp_vnrele_taskq);
436 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
437 kmem_free(dp, sizeof (dsl_pool_t));
441 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
444 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
445 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
450 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
452 /* create and open the MOS (meta-objset) */
453 dp->dp_meta_objset = dmu_objset_create_impl(spa,
454 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
456 /* create the pool directory */
457 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
458 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
461 /* Initialize scan structures */
462 VERIFY0(dsl_scan_init(dp, txg));
464 /* create and open the root dir */
465 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
466 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
467 NULL, dp, &dp->dp_root_dir));
469 /* create and open the meta-objset dir */
470 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
471 VERIFY0(dsl_pool_open_special_dir(dp,
472 MOS_DIR_NAME, &dp->dp_mos_dir));
474 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
475 /* create and open the free dir */
476 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
478 VERIFY0(dsl_pool_open_special_dir(dp,
479 FREE_DIR_NAME, &dp->dp_free_dir));
481 /* create and open the free_bplist */
482 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
483 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
484 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
485 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
486 dp->dp_meta_objset, obj));
489 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
490 dsl_pool_create_origin(dp, tx);
492 /* create the root dataset */
493 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
495 /* create the root objset */
496 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
497 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
498 os = dmu_objset_create_impl(dp->dp_spa, ds,
499 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
500 rrw_exit(&ds->ds_bp_rwlock, FTAG);
502 zfs_create_fs(os, kcred, zplprops, tx);
504 dsl_dataset_rele(ds, FTAG);
508 rrw_exit(&dp->dp_config_rwlock, FTAG);
514 * Account for the meta-objset space in its placeholder dsl_dir.
517 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
518 int64_t used, int64_t comp, int64_t uncomp)
520 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
521 mutex_enter(&dp->dp_lock);
522 dp->dp_mos_used_delta += used;
523 dp->dp_mos_compressed_delta += comp;
524 dp->dp_mos_uncompressed_delta += uncomp;
525 mutex_exit(&dp->dp_lock);
529 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
531 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
532 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
533 VERIFY0(zio_wait(zio));
534 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
535 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
539 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
541 ASSERT(MUTEX_HELD(&dp->dp_lock));
544 ASSERT3U(-delta, <=, dp->dp_dirty_total);
546 dp->dp_dirty_total += delta;
549 * Note: we signal even when increasing dp_dirty_total.
550 * This ensures forward progress -- each thread wakes the next waiter.
552 if (dp->dp_dirty_total <= zfs_dirty_data_max)
553 cv_signal(&dp->dp_spaceavail_cv);
557 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
563 objset_t *mos = dp->dp_meta_objset;
564 list_t synced_datasets;
566 list_create(&synced_datasets, sizeof (dsl_dataset_t),
567 offsetof(dsl_dataset_t, ds_synced_link));
569 tx = dmu_tx_create_assigned(dp, txg);
572 * Write out all dirty blocks of dirty datasets.
574 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
575 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
577 * We must not sync any non-MOS datasets twice, because
578 * we may have taken a snapshot of them. However, we
579 * may sync newly-created datasets on pass 2.
581 ASSERT(!list_link_active(&ds->ds_synced_link));
582 list_insert_tail(&synced_datasets, ds);
583 dsl_dataset_sync(ds, zio, tx);
585 VERIFY0(zio_wait(zio));
588 * We have written all of the accounted dirty data, so our
589 * dp_space_towrite should now be zero. However, some seldom-used
590 * code paths do not adhere to this (e.g. dbuf_undirty(), also
591 * rounding error in dbuf_write_physdone).
592 * Shore up the accounting of any dirtied space now.
594 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
597 * Update the long range free counter after
598 * we're done syncing user data
600 mutex_enter(&dp->dp_lock);
601 ASSERT(spa_sync_pass(dp->dp_spa) == 1 ||
602 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0);
603 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0;
604 mutex_exit(&dp->dp_lock);
607 * After the data blocks have been written (ensured by the zio_wait()
608 * above), update the user/group space accounting.
610 for (ds = list_head(&synced_datasets); ds != NULL;
611 ds = list_next(&synced_datasets, ds)) {
612 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
616 * Sync the datasets again to push out the changes due to
617 * userspace updates. This must be done before we process the
618 * sync tasks, so that any snapshots will have the correct
619 * user accounting information (and we won't get confused
620 * about which blocks are part of the snapshot).
622 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
623 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
624 ASSERT(list_link_active(&ds->ds_synced_link));
625 dmu_buf_rele(ds->ds_dbuf, ds);
626 dsl_dataset_sync(ds, zio, tx);
628 VERIFY0(zio_wait(zio));
631 * Now that the datasets have been completely synced, we can
632 * clean up our in-memory structures accumulated while syncing:
634 * - move dead blocks from the pending deadlist to the on-disk deadlist
635 * - release hold from dsl_dataset_dirty()
637 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
638 dsl_dataset_sync_done(ds, tx);
640 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
641 dsl_dir_sync(dd, tx);
645 * The MOS's space is accounted for in the pool/$MOS
646 * (dp_mos_dir). We can't modify the mos while we're syncing
647 * it, so we remember the deltas and apply them here.
649 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
650 dp->dp_mos_uncompressed_delta != 0) {
651 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
652 dp->dp_mos_used_delta,
653 dp->dp_mos_compressed_delta,
654 dp->dp_mos_uncompressed_delta, tx);
655 dp->dp_mos_used_delta = 0;
656 dp->dp_mos_compressed_delta = 0;
657 dp->dp_mos_uncompressed_delta = 0;
660 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
661 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
662 dsl_pool_sync_mos(dp, tx);
666 * If we modify a dataset in the same txg that we want to destroy it,
667 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
668 * dsl_dir_destroy_check() will fail if there are unexpected holds.
669 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
670 * and clearing the hold on it) before we process the sync_tasks.
671 * The MOS data dirtied by the sync_tasks will be synced on the next
674 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
675 dsl_sync_task_t *dst;
677 * No more sync tasks should have been added while we
680 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
681 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
682 dsl_sync_task_sync(dst, tx);
687 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
691 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
695 while (zilog = txg_list_head(&dp->dp_dirty_zilogs, txg)) {
696 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
698 * We don't remove the zilog from the dp_dirty_zilogs
699 * list until after we've cleaned it. This ensures that
700 * callers of zilog_is_dirty() receive an accurate
701 * answer when they are racing with the spa sync thread.
703 zil_clean(zilog, txg);
704 (void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg);
705 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
706 dmu_buf_rele(ds->ds_dbuf, zilog);
708 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
712 * TRUE if the current thread is the tx_sync_thread or if we
713 * are being called from SPA context during pool initialization.
716 dsl_pool_sync_context(dsl_pool_t *dp)
718 return (curthread == dp->dp_tx.tx_sync_thread ||
719 spa_is_initializing(dp->dp_spa));
723 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
725 uint64_t space, resv;
728 * If we're trying to assess whether it's OK to do a free,
729 * cut the reservation in half to allow forward progress
730 * (e.g. make it possible to rm(1) files from a full pool).
732 space = spa_get_dspace(dp->dp_spa);
733 resv = spa_get_slop_space(dp->dp_spa);
737 return (space - resv);
741 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
743 uint64_t delay_min_bytes =
744 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
747 mutex_enter(&dp->dp_lock);
748 if (dp->dp_dirty_total > zfs_dirty_data_sync)
750 rv = (dp->dp_dirty_total > delay_min_bytes);
751 mutex_exit(&dp->dp_lock);
756 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
759 mutex_enter(&dp->dp_lock);
760 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
761 dsl_pool_dirty_delta(dp, space);
762 mutex_exit(&dp->dp_lock);
767 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
769 ASSERT3S(space, >=, 0);
772 mutex_enter(&dp->dp_lock);
773 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
774 /* XXX writing something we didn't dirty? */
775 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
777 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
778 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
779 ASSERT3U(dp->dp_dirty_total, >=, space);
780 dsl_pool_dirty_delta(dp, -space);
781 mutex_exit(&dp->dp_lock);
786 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
789 dsl_dataset_t *ds, *prev = NULL;
792 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
796 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
797 err = dsl_dataset_hold_obj(dp,
798 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
800 dsl_dataset_rele(ds, FTAG);
804 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
806 dsl_dataset_rele(ds, FTAG);
812 prev = dp->dp_origin_snap;
815 * The $ORIGIN can't have any data, or the accounting
818 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
819 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
820 rrw_exit(&ds->ds_bp_rwlock, FTAG);
822 /* The origin doesn't get attached to itself */
823 if (ds->ds_object == prev->ds_object) {
824 dsl_dataset_rele(ds, FTAG);
828 dmu_buf_will_dirty(ds->ds_dbuf, tx);
829 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
830 dsl_dataset_phys(ds)->ds_prev_snap_txg =
831 dsl_dataset_phys(prev)->ds_creation_txg;
833 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
834 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
836 dmu_buf_will_dirty(prev->ds_dbuf, tx);
837 dsl_dataset_phys(prev)->ds_num_children++;
839 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
840 ASSERT(ds->ds_prev == NULL);
841 VERIFY0(dsl_dataset_hold_obj(dp,
842 dsl_dataset_phys(ds)->ds_prev_snap_obj,
847 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
848 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
850 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
851 dmu_buf_will_dirty(prev->ds_dbuf, tx);
852 dsl_dataset_phys(prev)->ds_next_clones_obj =
853 zap_create(dp->dp_meta_objset,
854 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
856 VERIFY0(zap_add_int(dp->dp_meta_objset,
857 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
859 dsl_dataset_rele(ds, FTAG);
860 if (prev != dp->dp_origin_snap)
861 dsl_dataset_rele(prev, FTAG);
866 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
868 ASSERT(dmu_tx_is_syncing(tx));
869 ASSERT(dp->dp_origin_snap != NULL);
871 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
872 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
877 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
880 objset_t *mos = dp->dp_meta_objset;
882 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
883 dsl_dataset_t *origin;
885 VERIFY0(dsl_dataset_hold_obj(dp,
886 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
888 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
889 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
890 dsl_dir_phys(origin->ds_dir)->dd_clones =
891 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
895 VERIFY0(zap_add_int(dp->dp_meta_objset,
896 dsl_dir_phys(origin->ds_dir)->dd_clones,
899 dsl_dataset_rele(origin, FTAG);
905 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
907 ASSERT(dmu_tx_is_syncing(tx));
910 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
911 VERIFY0(dsl_pool_open_special_dir(dp,
912 FREE_DIR_NAME, &dp->dp_free_dir));
915 * We can't use bpobj_alloc(), because spa_version() still
916 * returns the old version, and we need a new-version bpobj with
917 * subobj support. So call dmu_object_alloc() directly.
919 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
920 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
921 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
922 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
923 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
925 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
926 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
930 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
935 ASSERT(dmu_tx_is_syncing(tx));
936 ASSERT(dp->dp_origin_snap == NULL);
937 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
939 /* create the origin dir, ds, & snap-ds */
940 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
942 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
943 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
944 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
945 dp, &dp->dp_origin_snap));
946 dsl_dataset_rele(ds, FTAG);
950 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
952 return (dp->dp_vnrele_taskq);
956 * Walk through the pool-wide zap object of temporary snapshot user holds
960 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
964 objset_t *mos = dp->dp_meta_objset;
965 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
970 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
972 holds = fnvlist_alloc();
974 for (zap_cursor_init(&zc, mos, zapobj);
975 zap_cursor_retrieve(&zc, &za) == 0;
976 zap_cursor_advance(&zc)) {
980 htag = strchr(za.za_name, '-');
983 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
984 tags = fnvlist_alloc();
985 fnvlist_add_boolean(tags, htag);
986 fnvlist_add_nvlist(holds, za.za_name, tags);
989 fnvlist_add_boolean(tags, htag);
992 dsl_dataset_user_release_tmp(dp, holds);
994 zap_cursor_fini(&zc);
998 * Create the pool-wide zap object for storing temporary snapshot holds.
1001 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
1003 objset_t *mos = dp->dp_meta_objset;
1005 ASSERT(dp->dp_tmp_userrefs_obj == 0);
1006 ASSERT(dmu_tx_is_syncing(tx));
1008 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
1009 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
1013 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
1014 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
1016 objset_t *mos = dp->dp_meta_objset;
1017 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1021 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1022 ASSERT(dmu_tx_is_syncing(tx));
1025 * If the pool was created prior to SPA_VERSION_USERREFS, the
1026 * zap object for temporary holds might not exist yet.
1030 dsl_pool_user_hold_create_obj(dp, tx);
1031 zapobj = dp->dp_tmp_userrefs_obj;
1033 return (SET_ERROR(ENOENT));
1037 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
1039 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
1041 error = zap_remove(mos, zapobj, name, tx);
1048 * Add a temporary hold for the given dataset object and tag.
1051 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1052 uint64_t now, dmu_tx_t *tx)
1054 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
1058 * Release a temporary hold for the given dataset object and tag.
1061 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1064 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
1069 * DSL Pool Configuration Lock
1071 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1072 * creation / destruction / rename / property setting). It must be held for
1073 * read to hold a dataset or dsl_dir. I.e. you must call
1074 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1075 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1076 * must be held continuously until all datasets and dsl_dirs are released.
1078 * The only exception to this rule is that if a "long hold" is placed on
1079 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1080 * is still held. The long hold will prevent the dataset from being
1081 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1082 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1083 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1085 * Legitimate long-holders (including owners) should be long-running, cancelable
1086 * tasks that should cause "zfs destroy" to fail. This includes DMU
1087 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1088 * "zfs send", and "zfs diff". There are several other long-holders whose
1089 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1091 * The usual formula for long-holding would be:
1093 * dsl_dataset_hold()
1094 * ... perform checks ...
1095 * dsl_dataset_long_hold()
1097 * ... perform long-running task ...
1098 * dsl_dataset_long_rele()
1099 * dsl_dataset_rele()
1101 * Note that when the long hold is released, the dataset is still held but
1102 * the pool is not held. The dataset may change arbitrarily during this time
1103 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1104 * dataset except release it.
1106 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1107 * or modifying operations.
1109 * Modifying operations should generally use dsl_sync_task(). The synctask
1110 * infrastructure enforces proper locking strategy with respect to the
1111 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1113 * Read-only operations will manually hold the pool, then the dataset, obtain
1114 * information from the dataset, then release the pool and dataset.
1115 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1120 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1125 error = spa_open(name, &spa, tag);
1127 *dp = spa_get_dsl(spa);
1128 dsl_pool_config_enter(*dp, tag);
1134 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1136 dsl_pool_config_exit(dp, tag);
1137 spa_close(dp->dp_spa, tag);
1141 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1144 * We use a "reentrant" reader-writer lock, but not reentrantly.
1146 * The rrwlock can (with the track_all flag) track all reading threads,
1147 * which is very useful for debugging which code path failed to release
1148 * the lock, and for verifying that the *current* thread does hold
1151 * (Unlike a rwlock, which knows that N threads hold it for
1152 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1153 * if any thread holds it for read, even if this thread doesn't).
1155 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1156 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1160 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1162 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1163 rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1167 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1169 rrw_exit(&dp->dp_config_rwlock, tag);
1173 dsl_pool_config_held(dsl_pool_t *dp)
1175 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1179 dsl_pool_config_held_writer(dsl_pool_t *dp)
1181 return (RRW_WRITE_HELD(&dp->dp_config_rwlock));