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) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
27 #include <sys/dsl_pool.h>
28 #include <sys/dsl_dataset.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_synctask.h>
32 #include <sys/dsl_scan.h>
33 #include <sys/dnode.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
39 #include <sys/zfs_context.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/zfs_znode.h>
42 #include <sys/spa_impl.h>
43 #include <sys/dsl_deadlist.h>
44 #include <sys/bptree.h>
45 #include <sys/zfeature.h>
46 #include <sys/zil_impl.h>
47 #include <sys/dsl_userhold.h>
53 * ZFS must limit the rate of incoming writes to the rate at which it is able
54 * to sync data modifications to the backend storage. Throttling by too much
55 * creates an artificial limit; throttling by too little can only be sustained
56 * for short periods and would lead to highly lumpy performance. On a per-pool
57 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
58 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
59 * of dirty data decreases. When the amount of dirty data exceeds a
60 * predetermined threshold further modifications are blocked until the amount
61 * of dirty data decreases (as data is synced out).
63 * The limit on dirty data is tunable, and should be adjusted according to
64 * both the IO capacity and available memory of the system. The larger the
65 * window, the more ZFS is able to aggregate and amortize metadata (and data)
66 * changes. However, memory is a limited resource, and allowing for more dirty
67 * data comes at the cost of keeping other useful data in memory (for example
68 * ZFS data cached by the ARC).
72 * As buffers are modified dsl_pool_willuse_space() increments both the per-
73 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
74 * dirty space used; dsl_pool_dirty_space() decrements those values as data
75 * is synced out from dsl_pool_sync(). While only the poolwide value is
76 * relevant, the per-txg value is useful for debugging. The tunable
77 * zfs_dirty_data_max determines the dirty space limit. Once that value is
78 * exceeded, new writes are halted until space frees up.
80 * The zfs_dirty_data_sync tunable dictates the threshold at which we
81 * ensure that there is a txg syncing (see the comment in txg.c for a full
82 * description of transaction group stages).
84 * The IO scheduler uses both the dirty space limit and current amount of
85 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
86 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
88 * The delay is also calculated based on the amount of dirty data. See the
89 * comment above dmu_tx_delay() for details.
93 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
94 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
96 uint64_t zfs_dirty_data_max;
97 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
98 int zfs_dirty_data_max_percent = 10;
101 * If there is at least this much dirty data, push out a txg.
103 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
106 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
107 * and delay each transaction.
108 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
110 int zfs_delay_min_dirty_percent = 60;
113 * This controls how quickly the delay approaches infinity.
114 * Larger values cause it to delay less for a given amount of dirty data.
115 * Therefore larger values will cause there to be more dirty data for a
118 * For the smoothest delay, this value should be about 1 billion divided
119 * by the maximum number of operations per second. This will smoothly
120 * handle between 10x and 1/10th this number.
122 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
123 * multiply in dmu_tx_delay().
125 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
129 * XXX someday maybe turn these into #defines, and you have to tune it on a
130 * per-pool basis using zfs.conf.
134 SYSCTL_DECL(_vfs_zfs);
136 TUNABLE_INT("vfs.zfs.no_write_throttle", &zfs_no_write_throttle);
137 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_write_throttle, CTLFLAG_RDTUN,
138 &zfs_no_write_throttle, 0, "");
139 TUNABLE_INT("vfs.zfs.write_limit_shift", &zfs_write_limit_shift);
140 SYSCTL_INT(_vfs_zfs, OID_AUTO, write_limit_shift, CTLFLAG_RDTUN,
141 &zfs_write_limit_shift, 0, "2^N of physical memory");
142 SYSCTL_DECL(_vfs_zfs_txg);
143 TUNABLE_INT("vfs.zfs.txg.synctime_ms", &zfs_txg_synctime_ms);
144 SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, synctime_ms, CTLFLAG_RDTUN,
145 &zfs_txg_synctime_ms, 0, "Target milliseconds to sync a txg");
147 TUNABLE_QUAD("vfs.zfs.write_limit_min", &zfs_write_limit_min);
148 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_min, CTLFLAG_RDTUN,
149 &zfs_write_limit_min, 0, "Minimum write limit");
150 TUNABLE_QUAD("vfs.zfs.write_limit_max", &zfs_write_limit_max);
151 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_max, CTLFLAG_RDTUN,
152 &zfs_write_limit_max, 0, "Maximum data payload per txg");
153 TUNABLE_QUAD("vfs.zfs.write_limit_inflated", &zfs_write_limit_inflated);
154 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_inflated, CTLFLAG_RDTUN,
155 &zfs_write_limit_inflated, 0, "Maximum size of the dynamic write limit");
156 TUNABLE_QUAD("vfs.zfs.write_limit_override", &zfs_write_limit_override);
157 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_override, CTLFLAG_RDTUN,
158 &zfs_write_limit_override, 0,
159 "Force a txg if dirty buffers exceed this value (bytes)");
162 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
163 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
166 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
171 err = zap_lookup(dp->dp_meta_objset,
172 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
173 name, sizeof (obj), 1, &obj);
177 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
181 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
184 blkptr_t *bp = spa_get_rootblkptr(spa);
186 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
188 dp->dp_meta_rootbp = *bp;
189 rrw_init(&dp->dp_config_rwlock, B_TRUE);
192 txg_list_create(&dp->dp_dirty_datasets,
193 offsetof(dsl_dataset_t, ds_dirty_link));
194 txg_list_create(&dp->dp_dirty_zilogs,
195 offsetof(zilog_t, zl_dirty_link));
196 txg_list_create(&dp->dp_dirty_dirs,
197 offsetof(dsl_dir_t, dd_dirty_link));
198 txg_list_create(&dp->dp_sync_tasks,
199 offsetof(dsl_sync_task_t, dst_node));
201 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
202 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
204 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
211 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
214 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
216 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
217 &dp->dp_meta_objset);
227 dsl_pool_open(dsl_pool_t *dp)
234 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
235 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
236 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
237 &dp->dp_root_dir_obj);
241 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
242 NULL, dp, &dp->dp_root_dir);
246 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
250 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
251 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
254 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
257 err = dsl_dataset_hold_obj(dp,
258 ds->ds_phys->ds_prev_snap_obj, dp,
259 &dp->dp_origin_snap);
260 dsl_dataset_rele(ds, FTAG);
262 dsl_dir_rele(dd, dp);
267 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
268 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
273 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
274 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
277 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
278 dp->dp_meta_objset, obj));
281 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
282 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
283 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
289 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
290 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
291 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
292 &dp->dp_empty_bpobj);
297 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
298 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
299 &dp->dp_tmp_userrefs_obj);
305 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
308 rrw_exit(&dp->dp_config_rwlock, FTAG);
313 dsl_pool_close(dsl_pool_t *dp)
316 * Drop our references from dsl_pool_open().
318 * Since we held the origin_snap from "syncing" context (which
319 * includes pool-opening context), it actually only got a "ref"
320 * and not a hold, so just drop that here.
322 if (dp->dp_origin_snap)
323 dsl_dataset_rele(dp->dp_origin_snap, dp);
325 dsl_dir_rele(dp->dp_mos_dir, dp);
327 dsl_dir_rele(dp->dp_free_dir, dp);
329 dsl_dir_rele(dp->dp_root_dir, dp);
331 bpobj_close(&dp->dp_free_bpobj);
333 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
334 if (dp->dp_meta_objset)
335 dmu_objset_evict(dp->dp_meta_objset);
337 txg_list_destroy(&dp->dp_dirty_datasets);
338 txg_list_destroy(&dp->dp_dirty_zilogs);
339 txg_list_destroy(&dp->dp_sync_tasks);
340 txg_list_destroy(&dp->dp_dirty_dirs);
342 arc_flush(dp->dp_spa);
345 rrw_destroy(&dp->dp_config_rwlock);
346 mutex_destroy(&dp->dp_lock);
347 taskq_destroy(dp->dp_vnrele_taskq);
349 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
350 kmem_free(dp, sizeof (dsl_pool_t));
354 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
357 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
358 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
363 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
365 /* create and open the MOS (meta-objset) */
366 dp->dp_meta_objset = dmu_objset_create_impl(spa,
367 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
369 /* create the pool directory */
370 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
371 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
374 /* Initialize scan structures */
375 VERIFY0(dsl_scan_init(dp, txg));
377 /* create and open the root dir */
378 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
379 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
380 NULL, dp, &dp->dp_root_dir));
382 /* create and open the meta-objset dir */
383 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
384 VERIFY0(dsl_pool_open_special_dir(dp,
385 MOS_DIR_NAME, &dp->dp_mos_dir));
387 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
388 /* create and open the free dir */
389 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
391 VERIFY0(dsl_pool_open_special_dir(dp,
392 FREE_DIR_NAME, &dp->dp_free_dir));
394 /* create and open the free_bplist */
395 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
396 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
397 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
398 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
399 dp->dp_meta_objset, obj));
402 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
403 dsl_pool_create_origin(dp, tx);
405 /* create the root dataset */
406 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
408 /* create the root objset */
409 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
410 os = dmu_objset_create_impl(dp->dp_spa, ds,
411 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
413 zfs_create_fs(os, kcred, zplprops, tx);
415 dsl_dataset_rele(ds, FTAG);
419 rrw_exit(&dp->dp_config_rwlock, FTAG);
425 * Account for the meta-objset space in its placeholder dsl_dir.
428 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
429 int64_t used, int64_t comp, int64_t uncomp)
431 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
432 mutex_enter(&dp->dp_lock);
433 dp->dp_mos_used_delta += used;
434 dp->dp_mos_compressed_delta += comp;
435 dp->dp_mos_uncompressed_delta += uncomp;
436 mutex_exit(&dp->dp_lock);
440 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
442 dsl_deadlist_t *dl = arg;
443 dsl_deadlist_insert(dl, bp, tx);
448 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
450 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
451 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
452 VERIFY0(zio_wait(zio));
453 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
454 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
458 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
460 ASSERT(MUTEX_HELD(&dp->dp_lock));
463 ASSERT3U(-delta, <=, dp->dp_dirty_total);
465 dp->dp_dirty_total += delta;
468 * Note: we signal even when increasing dp_dirty_total.
469 * This ensures forward progress -- each thread wakes the next waiter.
471 if (dp->dp_dirty_total <= zfs_dirty_data_max)
472 cv_signal(&dp->dp_spaceavail_cv);
476 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
482 objset_t *mos = dp->dp_meta_objset;
483 list_t synced_datasets;
485 list_create(&synced_datasets, sizeof (dsl_dataset_t),
486 offsetof(dsl_dataset_t, ds_synced_link));
488 tx = dmu_tx_create_assigned(dp, txg);
491 * Write out all dirty blocks of dirty datasets.
493 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
494 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
496 * We must not sync any non-MOS datasets twice, because
497 * we may have taken a snapshot of them. However, we
498 * may sync newly-created datasets on pass 2.
500 ASSERT(!list_link_active(&ds->ds_synced_link));
501 list_insert_tail(&synced_datasets, ds);
502 dsl_dataset_sync(ds, zio, tx);
504 VERIFY0(zio_wait(zio));
507 * We have written all of the accounted dirty data, so our
508 * dp_space_towrite should now be zero. However, some seldom-used
509 * code paths do not adhere to this (e.g. dbuf_undirty(), also
510 * rounding error in dbuf_write_physdone).
511 * Shore up the accounting of any dirtied space now.
513 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
516 * After the data blocks have been written (ensured by the zio_wait()
517 * above), update the user/group space accounting.
519 for (ds = list_head(&synced_datasets); ds != NULL;
520 ds = list_next(&synced_datasets, ds)) {
521 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
525 * Sync the datasets again to push out the changes due to
526 * userspace updates. This must be done before we process the
527 * sync tasks, so that any snapshots will have the correct
528 * user accounting information (and we won't get confused
529 * about which blocks are part of the snapshot).
531 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
532 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
533 ASSERT(list_link_active(&ds->ds_synced_link));
534 dmu_buf_rele(ds->ds_dbuf, ds);
535 dsl_dataset_sync(ds, zio, tx);
537 VERIFY0(zio_wait(zio));
540 * Now that the datasets have been completely synced, we can
541 * clean up our in-memory structures accumulated while syncing:
543 * - move dead blocks from the pending deadlist to the on-disk deadlist
544 * - release hold from dsl_dataset_dirty()
546 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
547 objset_t *os = ds->ds_objset;
548 bplist_iterate(&ds->ds_pending_deadlist,
549 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
550 ASSERT(!dmu_objset_is_dirty(os, txg));
551 dmu_buf_rele(ds->ds_dbuf, ds);
553 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
554 dsl_dir_sync(dd, tx);
558 * The MOS's space is accounted for in the pool/$MOS
559 * (dp_mos_dir). We can't modify the mos while we're syncing
560 * it, so we remember the deltas and apply them here.
562 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
563 dp->dp_mos_uncompressed_delta != 0) {
564 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
565 dp->dp_mos_used_delta,
566 dp->dp_mos_compressed_delta,
567 dp->dp_mos_uncompressed_delta, tx);
568 dp->dp_mos_used_delta = 0;
569 dp->dp_mos_compressed_delta = 0;
570 dp->dp_mos_uncompressed_delta = 0;
573 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
574 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
575 dsl_pool_sync_mos(dp, tx);
579 * If we modify a dataset in the same txg that we want to destroy it,
580 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
581 * dsl_dir_destroy_check() will fail if there are unexpected holds.
582 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
583 * and clearing the hold on it) before we process the sync_tasks.
584 * The MOS data dirtied by the sync_tasks will be synced on the next
587 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
588 dsl_sync_task_t *dst;
590 * No more sync tasks should have been added while we
593 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
594 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
595 dsl_sync_task_sync(dst, tx);
600 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
604 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
608 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
609 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
610 zil_clean(zilog, txg);
611 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
612 dmu_buf_rele(ds->ds_dbuf, zilog);
614 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
618 * TRUE if the current thread is the tx_sync_thread or if we
619 * are being called from SPA context during pool initialization.
622 dsl_pool_sync_context(dsl_pool_t *dp)
624 return (curthread == dp->dp_tx.tx_sync_thread ||
625 spa_is_initializing(dp->dp_spa));
629 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
631 uint64_t space, resv;
634 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
636 * XXX The intent log is not accounted for, so it must fit
639 * If we're trying to assess whether it's OK to do a free,
640 * cut the reservation in half to allow forward progress
641 * (e.g. make it possible to rm(1) files from a full pool).
643 space = spa_get_dspace(dp->dp_spa);
644 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
648 return (space - resv);
652 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
654 uint64_t delay_min_bytes =
655 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
658 mutex_enter(&dp->dp_lock);
659 if (dp->dp_dirty_total > zfs_dirty_data_sync)
661 rv = (dp->dp_dirty_total > delay_min_bytes);
662 mutex_exit(&dp->dp_lock);
667 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
670 mutex_enter(&dp->dp_lock);
671 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
672 dsl_pool_dirty_delta(dp, space);
673 mutex_exit(&dp->dp_lock);
678 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
680 ASSERT3S(space, >=, 0);
683 mutex_enter(&dp->dp_lock);
684 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
685 /* XXX writing something we didn't dirty? */
686 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
688 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
689 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
690 ASSERT3U(dp->dp_dirty_total, >=, space);
691 dsl_pool_dirty_delta(dp, -space);
692 mutex_exit(&dp->dp_lock);
697 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
700 dsl_dataset_t *ds, *prev = NULL;
703 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
707 while (ds->ds_phys->ds_prev_snap_obj != 0) {
708 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
711 dsl_dataset_rele(ds, FTAG);
715 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
717 dsl_dataset_rele(ds, FTAG);
723 prev = dp->dp_origin_snap;
726 * The $ORIGIN can't have any data, or the accounting
729 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
731 /* The origin doesn't get attached to itself */
732 if (ds->ds_object == prev->ds_object) {
733 dsl_dataset_rele(ds, FTAG);
737 dmu_buf_will_dirty(ds->ds_dbuf, tx);
738 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
739 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
741 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
742 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
744 dmu_buf_will_dirty(prev->ds_dbuf, tx);
745 prev->ds_phys->ds_num_children++;
747 if (ds->ds_phys->ds_next_snap_obj == 0) {
748 ASSERT(ds->ds_prev == NULL);
749 VERIFY0(dsl_dataset_hold_obj(dp,
750 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
754 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
755 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
757 if (prev->ds_phys->ds_next_clones_obj == 0) {
758 dmu_buf_will_dirty(prev->ds_dbuf, tx);
759 prev->ds_phys->ds_next_clones_obj =
760 zap_create(dp->dp_meta_objset,
761 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
763 VERIFY0(zap_add_int(dp->dp_meta_objset,
764 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
766 dsl_dataset_rele(ds, FTAG);
767 if (prev != dp->dp_origin_snap)
768 dsl_dataset_rele(prev, FTAG);
773 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
775 ASSERT(dmu_tx_is_syncing(tx));
776 ASSERT(dp->dp_origin_snap != NULL);
778 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
779 tx, DS_FIND_CHILDREN));
784 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
787 objset_t *mos = dp->dp_meta_objset;
789 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
790 dsl_dataset_t *origin;
792 VERIFY0(dsl_dataset_hold_obj(dp,
793 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
795 if (origin->ds_dir->dd_phys->dd_clones == 0) {
796 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
797 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
798 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
801 VERIFY0(zap_add_int(dp->dp_meta_objset,
802 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
804 dsl_dataset_rele(origin, FTAG);
810 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
812 ASSERT(dmu_tx_is_syncing(tx));
815 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
816 VERIFY0(dsl_pool_open_special_dir(dp,
817 FREE_DIR_NAME, &dp->dp_free_dir));
820 * We can't use bpobj_alloc(), because spa_version() still
821 * returns the old version, and we need a new-version bpobj with
822 * subobj support. So call dmu_object_alloc() directly.
824 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
825 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
826 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
827 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
828 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
830 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
831 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
835 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
840 ASSERT(dmu_tx_is_syncing(tx));
841 ASSERT(dp->dp_origin_snap == NULL);
842 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
844 /* create the origin dir, ds, & snap-ds */
845 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
847 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
848 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
849 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
850 dp, &dp->dp_origin_snap));
851 dsl_dataset_rele(ds, FTAG);
855 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
857 return (dp->dp_vnrele_taskq);
861 * Walk through the pool-wide zap object of temporary snapshot user holds
865 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
869 objset_t *mos = dp->dp_meta_objset;
870 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
875 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
877 holds = fnvlist_alloc();
879 for (zap_cursor_init(&zc, mos, zapobj);
880 zap_cursor_retrieve(&zc, &za) == 0;
881 zap_cursor_advance(&zc)) {
885 htag = strchr(za.za_name, '-');
888 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
889 tags = fnvlist_alloc();
890 fnvlist_add_boolean(tags, htag);
891 fnvlist_add_nvlist(holds, za.za_name, tags);
894 fnvlist_add_boolean(tags, htag);
897 dsl_dataset_user_release_tmp(dp, holds);
899 zap_cursor_fini(&zc);
903 * Create the pool-wide zap object for storing temporary snapshot holds.
906 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
908 objset_t *mos = dp->dp_meta_objset;
910 ASSERT(dp->dp_tmp_userrefs_obj == 0);
911 ASSERT(dmu_tx_is_syncing(tx));
913 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
914 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
918 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
919 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
921 objset_t *mos = dp->dp_meta_objset;
922 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
926 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
927 ASSERT(dmu_tx_is_syncing(tx));
930 * If the pool was created prior to SPA_VERSION_USERREFS, the
931 * zap object for temporary holds might not exist yet.
935 dsl_pool_user_hold_create_obj(dp, tx);
936 zapobj = dp->dp_tmp_userrefs_obj;
938 return (SET_ERROR(ENOENT));
942 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
944 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
946 error = zap_remove(mos, zapobj, name, tx);
953 * Add a temporary hold for the given dataset object and tag.
956 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
957 uint64_t now, dmu_tx_t *tx)
959 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
963 * Release a temporary hold for the given dataset object and tag.
966 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
969 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
974 * DSL Pool Configuration Lock
976 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
977 * creation / destruction / rename / property setting). It must be held for
978 * read to hold a dataset or dsl_dir. I.e. you must call
979 * dsl_pool_config_enter() or dsl_pool_hold() before calling
980 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
981 * must be held continuously until all datasets and dsl_dirs are released.
983 * The only exception to this rule is that if a "long hold" is placed on
984 * a dataset, then the dp_config_rwlock may be dropped while the dataset
985 * is still held. The long hold will prevent the dataset from being
986 * destroyed -- the destroy will fail with EBUSY. A long hold can be
987 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
988 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
990 * Legitimate long-holders (including owners) should be long-running, cancelable
991 * tasks that should cause "zfs destroy" to fail. This includes DMU
992 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
993 * "zfs send", and "zfs diff". There are several other long-holders whose
994 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
996 * The usual formula for long-holding would be:
999 * ... perform checks ...
1000 * dsl_dataset_long_hold()
1002 * ... perform long-running task ...
1003 * dsl_dataset_long_rele()
1004 * dsl_dataset_rele()
1006 * Note that when the long hold is released, the dataset is still held but
1007 * the pool is not held. The dataset may change arbitrarily during this time
1008 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1009 * dataset except release it.
1011 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1012 * or modifying operations.
1014 * Modifying operations should generally use dsl_sync_task(). The synctask
1015 * infrastructure enforces proper locking strategy with respect to the
1016 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1018 * Read-only operations will manually hold the pool, then the dataset, obtain
1019 * information from the dataset, then release the pool and dataset.
1020 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1025 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1030 error = spa_open(name, &spa, tag);
1032 *dp = spa_get_dsl(spa);
1033 dsl_pool_config_enter(*dp, tag);
1039 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1041 dsl_pool_config_exit(dp, tag);
1042 spa_close(dp->dp_spa, tag);
1046 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1049 * We use a "reentrant" reader-writer lock, but not reentrantly.
1051 * The rrwlock can (with the track_all flag) track all reading threads,
1052 * which is very useful for debugging which code path failed to release
1053 * the lock, and for verifying that the *current* thread does hold
1056 * (Unlike a rwlock, which knows that N threads hold it for
1057 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1058 * if any thread holds it for read, even if this thread doesn't).
1060 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1061 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1065 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1067 rrw_exit(&dp->dp_config_rwlock, tag);
1071 dsl_pool_config_held(dsl_pool_t *dp)
1073 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));