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,
282 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
283 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
284 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
290 if (spa_feature_is_active(dp->dp_spa,
291 &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) {
292 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
293 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
294 &dp->dp_empty_bpobj);
299 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
300 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
301 &dp->dp_tmp_userrefs_obj);
307 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
310 rrw_exit(&dp->dp_config_rwlock, FTAG);
315 dsl_pool_close(dsl_pool_t *dp)
318 * Drop our references from dsl_pool_open().
320 * Since we held the origin_snap from "syncing" context (which
321 * includes pool-opening context), it actually only got a "ref"
322 * and not a hold, so just drop that here.
324 if (dp->dp_origin_snap)
325 dsl_dataset_rele(dp->dp_origin_snap, dp);
327 dsl_dir_rele(dp->dp_mos_dir, dp);
329 dsl_dir_rele(dp->dp_free_dir, dp);
331 dsl_dir_rele(dp->dp_root_dir, dp);
333 bpobj_close(&dp->dp_free_bpobj);
335 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
336 if (dp->dp_meta_objset)
337 dmu_objset_evict(dp->dp_meta_objset);
339 txg_list_destroy(&dp->dp_dirty_datasets);
340 txg_list_destroy(&dp->dp_dirty_zilogs);
341 txg_list_destroy(&dp->dp_sync_tasks);
342 txg_list_destroy(&dp->dp_dirty_dirs);
344 arc_flush(dp->dp_spa);
347 rrw_destroy(&dp->dp_config_rwlock);
348 mutex_destroy(&dp->dp_lock);
349 taskq_destroy(dp->dp_vnrele_taskq);
351 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
352 kmem_free(dp, sizeof (dsl_pool_t));
356 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
359 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
360 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
365 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
367 /* create and open the MOS (meta-objset) */
368 dp->dp_meta_objset = dmu_objset_create_impl(spa,
369 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
371 /* create the pool directory */
372 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
373 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
376 /* Initialize scan structures */
377 VERIFY0(dsl_scan_init(dp, txg));
379 /* create and open the root dir */
380 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
381 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
382 NULL, dp, &dp->dp_root_dir));
384 /* create and open the meta-objset dir */
385 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
386 VERIFY0(dsl_pool_open_special_dir(dp,
387 MOS_DIR_NAME, &dp->dp_mos_dir));
389 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
390 /* create and open the free dir */
391 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
393 VERIFY0(dsl_pool_open_special_dir(dp,
394 FREE_DIR_NAME, &dp->dp_free_dir));
396 /* create and open the free_bplist */
397 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
398 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
399 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
400 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
401 dp->dp_meta_objset, obj));
404 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
405 dsl_pool_create_origin(dp, tx);
407 /* create the root dataset */
408 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
410 /* create the root objset */
411 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
412 os = dmu_objset_create_impl(dp->dp_spa, ds,
413 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
415 zfs_create_fs(os, kcred, zplprops, tx);
417 dsl_dataset_rele(ds, FTAG);
421 rrw_exit(&dp->dp_config_rwlock, FTAG);
427 * Account for the meta-objset space in its placeholder dsl_dir.
430 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
431 int64_t used, int64_t comp, int64_t uncomp)
433 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
434 mutex_enter(&dp->dp_lock);
435 dp->dp_mos_used_delta += used;
436 dp->dp_mos_compressed_delta += comp;
437 dp->dp_mos_uncompressed_delta += uncomp;
438 mutex_exit(&dp->dp_lock);
442 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
444 dsl_deadlist_t *dl = arg;
445 dsl_deadlist_insert(dl, bp, tx);
450 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
452 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
453 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
454 VERIFY0(zio_wait(zio));
455 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
456 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
460 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
462 ASSERT(MUTEX_HELD(&dp->dp_lock));
465 ASSERT3U(-delta, <=, dp->dp_dirty_total);
467 dp->dp_dirty_total += delta;
470 * Note: we signal even when increasing dp_dirty_total.
471 * This ensures forward progress -- each thread wakes the next waiter.
473 if (dp->dp_dirty_total <= zfs_dirty_data_max)
474 cv_signal(&dp->dp_spaceavail_cv);
478 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
484 objset_t *mos = dp->dp_meta_objset;
485 list_t synced_datasets;
487 list_create(&synced_datasets, sizeof (dsl_dataset_t),
488 offsetof(dsl_dataset_t, ds_synced_link));
490 tx = dmu_tx_create_assigned(dp, txg);
493 * Write out all dirty blocks of dirty datasets.
495 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
496 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
498 * We must not sync any non-MOS datasets twice, because
499 * we may have taken a snapshot of them. However, we
500 * may sync newly-created datasets on pass 2.
502 ASSERT(!list_link_active(&ds->ds_synced_link));
503 list_insert_tail(&synced_datasets, ds);
504 dsl_dataset_sync(ds, zio, tx);
506 VERIFY0(zio_wait(zio));
509 * We have written all of the accounted dirty data, so our
510 * dp_space_towrite should now be zero. However, some seldom-used
511 * code paths do not adhere to this (e.g. dbuf_undirty(), also
512 * rounding error in dbuf_write_physdone).
513 * Shore up the accounting of any dirtied space now.
515 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
518 * After the data blocks have been written (ensured by the zio_wait()
519 * above), update the user/group space accounting.
521 for (ds = list_head(&synced_datasets); ds != NULL;
522 ds = list_next(&synced_datasets, ds)) {
523 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
527 * Sync the datasets again to push out the changes due to
528 * userspace updates. This must be done before we process the
529 * sync tasks, so that any snapshots will have the correct
530 * user accounting information (and we won't get confused
531 * about which blocks are part of the snapshot).
533 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
534 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
535 ASSERT(list_link_active(&ds->ds_synced_link));
536 dmu_buf_rele(ds->ds_dbuf, ds);
537 dsl_dataset_sync(ds, zio, tx);
539 VERIFY0(zio_wait(zio));
542 * Now that the datasets have been completely synced, we can
543 * clean up our in-memory structures accumulated while syncing:
545 * - move dead blocks from the pending deadlist to the on-disk deadlist
546 * - release hold from dsl_dataset_dirty()
548 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
549 objset_t *os = ds->ds_objset;
550 bplist_iterate(&ds->ds_pending_deadlist,
551 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
552 ASSERT(!dmu_objset_is_dirty(os, txg));
553 dmu_buf_rele(ds->ds_dbuf, ds);
555 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
556 dsl_dir_sync(dd, tx);
560 * The MOS's space is accounted for in the pool/$MOS
561 * (dp_mos_dir). We can't modify the mos while we're syncing
562 * it, so we remember the deltas and apply them here.
564 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
565 dp->dp_mos_uncompressed_delta != 0) {
566 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
567 dp->dp_mos_used_delta,
568 dp->dp_mos_compressed_delta,
569 dp->dp_mos_uncompressed_delta, tx);
570 dp->dp_mos_used_delta = 0;
571 dp->dp_mos_compressed_delta = 0;
572 dp->dp_mos_uncompressed_delta = 0;
575 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
576 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
577 dsl_pool_sync_mos(dp, tx);
581 * If we modify a dataset in the same txg that we want to destroy it,
582 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
583 * dsl_dir_destroy_check() will fail if there are unexpected holds.
584 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
585 * and clearing the hold on it) before we process the sync_tasks.
586 * The MOS data dirtied by the sync_tasks will be synced on the next
589 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
590 dsl_sync_task_t *dst;
592 * No more sync tasks should have been added while we
595 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
596 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
597 dsl_sync_task_sync(dst, tx);
602 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
606 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
610 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
611 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
612 zil_clean(zilog, txg);
613 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
614 dmu_buf_rele(ds->ds_dbuf, zilog);
616 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
620 * TRUE if the current thread is the tx_sync_thread or if we
621 * are being called from SPA context during pool initialization.
624 dsl_pool_sync_context(dsl_pool_t *dp)
626 return (curthread == dp->dp_tx.tx_sync_thread ||
627 spa_is_initializing(dp->dp_spa));
631 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
633 uint64_t space, resv;
636 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
638 * XXX The intent log is not accounted for, so it must fit
641 * If we're trying to assess whether it's OK to do a free,
642 * cut the reservation in half to allow forward progress
643 * (e.g. make it possible to rm(1) files from a full pool).
645 space = spa_get_dspace(dp->dp_spa);
646 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
650 return (space - resv);
654 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
656 uint64_t delay_min_bytes =
657 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
660 mutex_enter(&dp->dp_lock);
661 if (dp->dp_dirty_total > zfs_dirty_data_sync)
663 rv = (dp->dp_dirty_total > delay_min_bytes);
664 mutex_exit(&dp->dp_lock);
669 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
672 mutex_enter(&dp->dp_lock);
673 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
674 dsl_pool_dirty_delta(dp, space);
675 mutex_exit(&dp->dp_lock);
680 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
682 ASSERT3S(space, >=, 0);
685 mutex_enter(&dp->dp_lock);
686 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
687 /* XXX writing something we didn't dirty? */
688 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
690 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
691 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
692 ASSERT3U(dp->dp_dirty_total, >=, space);
693 dsl_pool_dirty_delta(dp, -space);
694 mutex_exit(&dp->dp_lock);
699 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
702 dsl_dataset_t *ds, *prev = NULL;
705 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
709 while (ds->ds_phys->ds_prev_snap_obj != 0) {
710 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
713 dsl_dataset_rele(ds, FTAG);
717 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
719 dsl_dataset_rele(ds, FTAG);
725 prev = dp->dp_origin_snap;
728 * The $ORIGIN can't have any data, or the accounting
731 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
733 /* The origin doesn't get attached to itself */
734 if (ds->ds_object == prev->ds_object) {
735 dsl_dataset_rele(ds, FTAG);
739 dmu_buf_will_dirty(ds->ds_dbuf, tx);
740 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
741 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
743 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
744 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
746 dmu_buf_will_dirty(prev->ds_dbuf, tx);
747 prev->ds_phys->ds_num_children++;
749 if (ds->ds_phys->ds_next_snap_obj == 0) {
750 ASSERT(ds->ds_prev == NULL);
751 VERIFY0(dsl_dataset_hold_obj(dp,
752 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
756 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
757 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
759 if (prev->ds_phys->ds_next_clones_obj == 0) {
760 dmu_buf_will_dirty(prev->ds_dbuf, tx);
761 prev->ds_phys->ds_next_clones_obj =
762 zap_create(dp->dp_meta_objset,
763 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
765 VERIFY0(zap_add_int(dp->dp_meta_objset,
766 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
768 dsl_dataset_rele(ds, FTAG);
769 if (prev != dp->dp_origin_snap)
770 dsl_dataset_rele(prev, FTAG);
775 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
777 ASSERT(dmu_tx_is_syncing(tx));
778 ASSERT(dp->dp_origin_snap != NULL);
780 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
781 tx, DS_FIND_CHILDREN));
786 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
789 objset_t *mos = dp->dp_meta_objset;
791 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
792 dsl_dataset_t *origin;
794 VERIFY0(dsl_dataset_hold_obj(dp,
795 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
797 if (origin->ds_dir->dd_phys->dd_clones == 0) {
798 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
799 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
800 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
803 VERIFY0(zap_add_int(dp->dp_meta_objset,
804 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
806 dsl_dataset_rele(origin, FTAG);
812 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
814 ASSERT(dmu_tx_is_syncing(tx));
817 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
818 VERIFY0(dsl_pool_open_special_dir(dp,
819 FREE_DIR_NAME, &dp->dp_free_dir));
822 * We can't use bpobj_alloc(), because spa_version() still
823 * returns the old version, and we need a new-version bpobj with
824 * subobj support. So call dmu_object_alloc() directly.
826 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
827 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
828 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
829 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
830 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
832 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
833 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
837 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
842 ASSERT(dmu_tx_is_syncing(tx));
843 ASSERT(dp->dp_origin_snap == NULL);
844 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
846 /* create the origin dir, ds, & snap-ds */
847 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
849 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
850 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
851 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
852 dp, &dp->dp_origin_snap));
853 dsl_dataset_rele(ds, FTAG);
857 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
859 return (dp->dp_vnrele_taskq);
863 * Walk through the pool-wide zap object of temporary snapshot user holds
867 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
871 objset_t *mos = dp->dp_meta_objset;
872 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
877 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
879 holds = fnvlist_alloc();
881 for (zap_cursor_init(&zc, mos, zapobj);
882 zap_cursor_retrieve(&zc, &za) == 0;
883 zap_cursor_advance(&zc)) {
887 htag = strchr(za.za_name, '-');
890 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
891 tags = fnvlist_alloc();
892 fnvlist_add_boolean(tags, htag);
893 fnvlist_add_nvlist(holds, za.za_name, tags);
896 fnvlist_add_boolean(tags, htag);
899 dsl_dataset_user_release_tmp(dp, holds);
901 zap_cursor_fini(&zc);
905 * Create the pool-wide zap object for storing temporary snapshot holds.
908 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
910 objset_t *mos = dp->dp_meta_objset;
912 ASSERT(dp->dp_tmp_userrefs_obj == 0);
913 ASSERT(dmu_tx_is_syncing(tx));
915 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
916 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
920 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
921 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
923 objset_t *mos = dp->dp_meta_objset;
924 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
928 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
929 ASSERT(dmu_tx_is_syncing(tx));
932 * If the pool was created prior to SPA_VERSION_USERREFS, the
933 * zap object for temporary holds might not exist yet.
937 dsl_pool_user_hold_create_obj(dp, tx);
938 zapobj = dp->dp_tmp_userrefs_obj;
940 return (SET_ERROR(ENOENT));
944 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
946 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
948 error = zap_remove(mos, zapobj, name, tx);
955 * Add a temporary hold for the given dataset object and tag.
958 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
959 uint64_t now, dmu_tx_t *tx)
961 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
965 * Release a temporary hold for the given dataset object and tag.
968 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
971 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
976 * DSL Pool Configuration Lock
978 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
979 * creation / destruction / rename / property setting). It must be held for
980 * read to hold a dataset or dsl_dir. I.e. you must call
981 * dsl_pool_config_enter() or dsl_pool_hold() before calling
982 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
983 * must be held continuously until all datasets and dsl_dirs are released.
985 * The only exception to this rule is that if a "long hold" is placed on
986 * a dataset, then the dp_config_rwlock may be dropped while the dataset
987 * is still held. The long hold will prevent the dataset from being
988 * destroyed -- the destroy will fail with EBUSY. A long hold can be
989 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
990 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
992 * Legitimate long-holders (including owners) should be long-running, cancelable
993 * tasks that should cause "zfs destroy" to fail. This includes DMU
994 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
995 * "zfs send", and "zfs diff". There are several other long-holders whose
996 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
998 * The usual formula for long-holding would be:
1000 * dsl_dataset_hold()
1001 * ... perform checks ...
1002 * dsl_dataset_long_hold()
1004 * ... perform long-running task ...
1005 * dsl_dataset_long_rele()
1006 * dsl_dataset_rele()
1008 * Note that when the long hold is released, the dataset is still held but
1009 * the pool is not held. The dataset may change arbitrarily during this time
1010 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1011 * dataset except release it.
1013 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1014 * or modifying operations.
1016 * Modifying operations should generally use dsl_sync_task(). The synctask
1017 * infrastructure enforces proper locking strategy with respect to the
1018 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1020 * Read-only operations will manually hold the pool, then the dataset, obtain
1021 * information from the dataset, then release the pool and dataset.
1022 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1027 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1032 error = spa_open(name, &spa, tag);
1034 *dp = spa_get_dsl(spa);
1035 dsl_pool_config_enter(*dp, tag);
1041 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1043 dsl_pool_config_exit(dp, tag);
1044 spa_close(dp->dp_spa, tag);
1048 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1051 * We use a "reentrant" reader-writer lock, but not reentrantly.
1053 * The rrwlock can (with the track_all flag) track all reading threads,
1054 * which is very useful for debugging which code path failed to release
1055 * the lock, and for verifying that the *current* thread does hold
1058 * (Unlike a rwlock, which knows that N threads hold it for
1059 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1060 * if any thread holds it for read, even if this thread doesn't).
1062 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1063 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1067 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1069 rrw_exit(&dp->dp_config_rwlock, tag);
1073 dsl_pool_config_held(dsl_pool_t *dp)
1075 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));