4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26 * Copyright 2016 Nexenta Systems, Inc. All rights reserved.
29 #include <sys/dsl_pool.h>
30 #include <sys/dsl_dataset.h>
31 #include <sys/dsl_prop.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dsl_synctask.h>
34 #include <sys/dsl_scan.h>
35 #include <sys/dnode.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/dmu_objset.h>
41 #include <sys/zfs_context.h>
42 #include <sys/fs/zfs.h>
43 #include <sys/zfs_znode.h>
44 #include <sys/spa_impl.h>
45 #include <sys/dsl_deadlist.h>
46 #include <sys/bptree.h>
47 #include <sys/zfeature.h>
48 #include <sys/zil_impl.h>
49 #include <sys/dsl_userhold.h>
50 #include <sys/trace_txg.h>
56 * ZFS must limit the rate of incoming writes to the rate at which it is able
57 * to sync data modifications to the backend storage. Throttling by too much
58 * creates an artificial limit; throttling by too little can only be sustained
59 * for short periods and would lead to highly lumpy performance. On a per-pool
60 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
61 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
62 * of dirty data decreases. When the amount of dirty data exceeds a
63 * predetermined threshold further modifications are blocked until the amount
64 * of dirty data decreases (as data is synced out).
66 * The limit on dirty data is tunable, and should be adjusted according to
67 * both the IO capacity and available memory of the system. The larger the
68 * window, the more ZFS is able to aggregate and amortize metadata (and data)
69 * changes. However, memory is a limited resource, and allowing for more dirty
70 * data comes at the cost of keeping other useful data in memory (for example
71 * ZFS data cached by the ARC).
75 * As buffers are modified dsl_pool_willuse_space() increments both the per-
76 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
77 * dirty space used; dsl_pool_dirty_space() decrements those values as data
78 * is synced out from dsl_pool_sync(). While only the poolwide value is
79 * relevant, the per-txg value is useful for debugging. The tunable
80 * zfs_dirty_data_max determines the dirty space limit. Once that value is
81 * exceeded, new writes are halted until space frees up.
83 * The zfs_dirty_data_sync tunable dictates the threshold at which we
84 * ensure that there is a txg syncing (see the comment in txg.c for a full
85 * description of transaction group stages).
87 * The IO scheduler uses both the dirty space limit and current amount of
88 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
89 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
91 * The delay is also calculated based on the amount of dirty data. See the
92 * comment above dmu_tx_delay() for details.
96 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
97 * capped at zfs_dirty_data_max_max. It can also be overridden with a module
100 unsigned long zfs_dirty_data_max = 0;
101 unsigned long zfs_dirty_data_max_max = 0;
102 int zfs_dirty_data_max_percent = 10;
103 int zfs_dirty_data_max_max_percent = 25;
106 * If there is at least this much dirty data, push out a txg.
108 unsigned long zfs_dirty_data_sync = 64 * 1024 * 1024;
111 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
112 * and delay each transaction.
113 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
115 int zfs_delay_min_dirty_percent = 60;
118 * This controls how quickly the delay approaches infinity.
119 * Larger values cause it to delay more for a given amount of dirty data.
120 * Therefore larger values will cause there to be less dirty data for a
123 * For the smoothest delay, this value should be about 1 billion divided
124 * by the maximum number of operations per second. This will smoothly
125 * handle between 10x and 1/10th this number.
127 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
128 * multiply in dmu_tx_delay().
130 unsigned long zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
133 * This determines the number of threads used by the dp_sync_taskq.
135 int zfs_sync_taskq_batch_pct = 75;
138 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
143 err = zap_lookup(dp->dp_meta_objset,
144 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
145 name, sizeof (obj), 1, &obj);
149 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
153 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
156 blkptr_t *bp = spa_get_rootblkptr(spa);
158 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
160 dp->dp_meta_rootbp = *bp;
161 rrw_init(&dp->dp_config_rwlock, B_TRUE);
164 txg_list_create(&dp->dp_dirty_datasets, spa,
165 offsetof(dsl_dataset_t, ds_dirty_link));
166 txg_list_create(&dp->dp_dirty_zilogs, spa,
167 offsetof(zilog_t, zl_dirty_link));
168 txg_list_create(&dp->dp_dirty_dirs, spa,
169 offsetof(dsl_dir_t, dd_dirty_link));
170 txg_list_create(&dp->dp_sync_tasks, spa,
171 offsetof(dsl_sync_task_t, dst_node));
173 dp->dp_sync_taskq = taskq_create("dp_sync_taskq",
174 zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX,
175 TASKQ_THREADS_CPU_PCT);
177 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
178 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
180 dp->dp_iput_taskq = taskq_create("z_iput", max_ncpus, defclsyspri,
181 max_ncpus * 8, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
187 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
190 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
193 * Initialize the caller's dsl_pool_t structure before we actually open
194 * the meta objset. This is done because a self-healing write zio may
195 * be issued as part of dmu_objset_open_impl() and the spa needs its
196 * dsl_pool_t initialized in order to handle the write.
200 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
201 &dp->dp_meta_objset);
211 dsl_pool_open(dsl_pool_t *dp)
218 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
219 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
220 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
221 &dp->dp_root_dir_obj);
225 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
226 NULL, dp, &dp->dp_root_dir);
230 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
234 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
235 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
238 err = dsl_dataset_hold_obj(dp,
239 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
241 err = dsl_dataset_hold_obj(dp,
242 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
243 &dp->dp_origin_snap);
244 dsl_dataset_rele(ds, FTAG);
246 dsl_dir_rele(dd, dp);
251 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
252 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
257 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
258 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
261 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
262 dp->dp_meta_objset, obj));
266 * Note: errors ignored, because the leak dir will not exist if we
267 * have not encountered a leak yet.
269 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
272 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
273 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
274 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
280 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
281 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
282 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
283 &dp->dp_empty_bpobj);
288 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
289 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
290 &dp->dp_tmp_userrefs_obj);
296 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
299 rrw_exit(&dp->dp_config_rwlock, FTAG);
304 dsl_pool_close(dsl_pool_t *dp)
307 * Drop our references from dsl_pool_open().
309 * Since we held the origin_snap from "syncing" context (which
310 * includes pool-opening context), it actually only got a "ref"
311 * and not a hold, so just drop that here.
313 if (dp->dp_origin_snap)
314 dsl_dataset_rele(dp->dp_origin_snap, dp);
316 dsl_dir_rele(dp->dp_mos_dir, dp);
318 dsl_dir_rele(dp->dp_free_dir, dp);
320 dsl_dir_rele(dp->dp_leak_dir, dp);
322 dsl_dir_rele(dp->dp_root_dir, dp);
324 bpobj_close(&dp->dp_free_bpobj);
326 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
327 if (dp->dp_meta_objset)
328 dmu_objset_evict(dp->dp_meta_objset);
330 txg_list_destroy(&dp->dp_dirty_datasets);
331 txg_list_destroy(&dp->dp_dirty_zilogs);
332 txg_list_destroy(&dp->dp_sync_tasks);
333 txg_list_destroy(&dp->dp_dirty_dirs);
335 taskq_destroy(dp->dp_sync_taskq);
338 * We can't set retry to TRUE since we're explicitly specifying
339 * a spa to flush. This is good enough; any missed buffers for
340 * this spa won't cause trouble, and they'll eventually fall
341 * out of the ARC just like any other unused buffer.
343 arc_flush(dp->dp_spa, FALSE);
347 dmu_buf_user_evict_wait();
349 rrw_destroy(&dp->dp_config_rwlock);
350 mutex_destroy(&dp->dp_lock);
351 cv_destroy(&dp->dp_spaceavail_cv);
352 taskq_destroy(dp->dp_iput_taskq);
354 vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
355 kmem_free(dp, sizeof (dsl_pool_t));
359 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
362 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
363 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
368 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
370 /* create and open the MOS (meta-objset) */
371 dp->dp_meta_objset = dmu_objset_create_impl(spa,
372 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
374 /* create the pool directory */
375 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
376 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
379 /* Initialize scan structures */
380 VERIFY0(dsl_scan_init(dp, txg));
382 /* create and open the root dir */
383 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
384 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
385 NULL, dp, &dp->dp_root_dir));
387 /* create and open the meta-objset dir */
388 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
389 VERIFY0(dsl_pool_open_special_dir(dp,
390 MOS_DIR_NAME, &dp->dp_mos_dir));
392 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
393 /* create and open the free dir */
394 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
396 VERIFY0(dsl_pool_open_special_dir(dp,
397 FREE_DIR_NAME, &dp->dp_free_dir));
399 /* create and open the free_bplist */
400 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
401 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
402 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
403 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
404 dp->dp_meta_objset, obj));
407 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
408 dsl_pool_create_origin(dp, tx);
410 /* create the root dataset */
411 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
413 /* create the root objset */
414 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
415 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
416 VERIFY(NULL != (os = dmu_objset_create_impl(dp->dp_spa, ds,
417 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx)));
418 rrw_exit(&ds->ds_bp_rwlock, FTAG);
420 zfs_create_fs(os, kcred, zplprops, tx);
422 dsl_dataset_rele(ds, FTAG);
426 rrw_exit(&dp->dp_config_rwlock, FTAG);
432 * Account for the meta-objset space in its placeholder dsl_dir.
435 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
436 int64_t used, int64_t comp, int64_t uncomp)
438 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
439 mutex_enter(&dp->dp_lock);
440 dp->dp_mos_used_delta += used;
441 dp->dp_mos_compressed_delta += comp;
442 dp->dp_mos_uncompressed_delta += uncomp;
443 mutex_exit(&dp->dp_lock);
447 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
449 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
450 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
451 VERIFY0(zio_wait(zio));
452 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
453 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
457 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
459 ASSERT(MUTEX_HELD(&dp->dp_lock));
462 ASSERT3U(-delta, <=, dp->dp_dirty_total);
464 dp->dp_dirty_total += delta;
467 * Note: we signal even when increasing dp_dirty_total.
468 * This ensures forward progress -- each thread wakes the next waiter.
470 if (dp->dp_dirty_total < zfs_dirty_data_max)
471 cv_signal(&dp->dp_spaceavail_cv);
475 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
481 objset_t *mos = dp->dp_meta_objset;
482 list_t synced_datasets;
484 list_create(&synced_datasets, sizeof (dsl_dataset_t),
485 offsetof(dsl_dataset_t, ds_synced_link));
487 tx = dmu_tx_create_assigned(dp, txg);
490 * Write out all dirty blocks of dirty datasets.
492 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
493 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
495 * We must not sync any non-MOS datasets twice, because
496 * we may have taken a snapshot of them. However, we
497 * may sync newly-created datasets on pass 2.
499 ASSERT(!list_link_active(&ds->ds_synced_link));
500 list_insert_tail(&synced_datasets, ds);
501 dsl_dataset_sync(ds, zio, tx);
503 VERIFY0(zio_wait(zio));
506 * We have written all of the accounted dirty data, so our
507 * dp_space_towrite should now be zero. However, some seldom-used
508 * code paths do not adhere to this (e.g. dbuf_undirty(), also
509 * rounding error in dbuf_write_physdone).
510 * Shore up the accounting of any dirtied space now.
512 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
515 * Update the long range free counter after
516 * we're done syncing user data
518 mutex_enter(&dp->dp_lock);
519 ASSERT(spa_sync_pass(dp->dp_spa) == 1 ||
520 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0);
521 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0;
522 mutex_exit(&dp->dp_lock);
525 * After the data blocks have been written (ensured by the zio_wait()
526 * above), update the user/group space accounting. This happens
527 * in tasks dispatched to dp_sync_taskq, so wait for them before
530 for (ds = list_head(&synced_datasets); ds != NULL;
531 ds = list_next(&synced_datasets, ds)) {
532 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
534 taskq_wait(dp->dp_sync_taskq);
537 * Sync the datasets again to push out the changes due to
538 * userspace updates. This must be done before we process the
539 * sync tasks, so that any snapshots will have the correct
540 * user accounting information (and we won't get confused
541 * about which blocks are part of the snapshot).
543 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
544 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
545 ASSERT(list_link_active(&ds->ds_synced_link));
546 dmu_buf_rele(ds->ds_dbuf, ds);
547 dsl_dataset_sync(ds, zio, tx);
549 VERIFY0(zio_wait(zio));
552 * Now that the datasets have been completely synced, we can
553 * clean up our in-memory structures accumulated while syncing:
555 * - move dead blocks from the pending deadlist to the on-disk deadlist
556 * - release hold from dsl_dataset_dirty()
558 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
559 dsl_dataset_sync_done(ds, tx);
562 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
563 dsl_dir_sync(dd, tx);
567 * The MOS's space is accounted for in the pool/$MOS
568 * (dp_mos_dir). We can't modify the mos while we're syncing
569 * it, so we remember the deltas and apply them here.
571 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
572 dp->dp_mos_uncompressed_delta != 0) {
573 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
574 dp->dp_mos_used_delta,
575 dp->dp_mos_compressed_delta,
576 dp->dp_mos_uncompressed_delta, tx);
577 dp->dp_mos_used_delta = 0;
578 dp->dp_mos_compressed_delta = 0;
579 dp->dp_mos_uncompressed_delta = 0;
582 if (!multilist_is_empty(mos->os_dirty_dnodes[txg & TXG_MASK])) {
583 dsl_pool_sync_mos(dp, tx);
587 * If we modify a dataset in the same txg that we want to destroy it,
588 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
589 * dsl_dir_destroy_check() will fail if there are unexpected holds.
590 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
591 * and clearing the hold on it) before we process the sync_tasks.
592 * The MOS data dirtied by the sync_tasks will be synced on the next
595 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
596 dsl_sync_task_t *dst;
598 * No more sync tasks should have been added while we
601 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
602 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
603 dsl_sync_task_sync(dst, tx);
608 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
612 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
616 while ((zilog = txg_list_head(&dp->dp_dirty_zilogs, txg))) {
617 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
619 * We don't remove the zilog from the dp_dirty_zilogs
620 * list until after we've cleaned it. This ensures that
621 * callers of zilog_is_dirty() receive an accurate
622 * answer when they are racing with the spa sync thread.
624 zil_clean(zilog, txg);
625 (void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg);
626 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
627 dmu_buf_rele(ds->ds_dbuf, zilog);
629 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
633 * TRUE if the current thread is the tx_sync_thread or if we
634 * are being called from SPA context during pool initialization.
637 dsl_pool_sync_context(dsl_pool_t *dp)
639 return (curthread == dp->dp_tx.tx_sync_thread ||
640 spa_is_initializing(dp->dp_spa) ||
641 taskq_member(dp->dp_sync_taskq, curthread));
645 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
647 uint64_t space, resv;
650 * If we're trying to assess whether it's OK to do a free,
651 * cut the reservation in half to allow forward progress
652 * (e.g. make it possible to rm(1) files from a full pool).
654 space = spa_get_dspace(dp->dp_spa);
655 resv = spa_get_slop_space(dp->dp_spa);
659 return (space - resv);
663 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
665 uint64_t delay_min_bytes =
666 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
669 mutex_enter(&dp->dp_lock);
670 if (dp->dp_dirty_total > zfs_dirty_data_sync)
672 rv = (dp->dp_dirty_total > delay_min_bytes);
673 mutex_exit(&dp->dp_lock);
678 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
681 mutex_enter(&dp->dp_lock);
682 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
683 dsl_pool_dirty_delta(dp, space);
684 mutex_exit(&dp->dp_lock);
689 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
691 ASSERT3S(space, >=, 0);
695 mutex_enter(&dp->dp_lock);
696 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
697 /* XXX writing something we didn't dirty? */
698 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
700 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
701 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
702 ASSERT3U(dp->dp_dirty_total, >=, space);
703 dsl_pool_dirty_delta(dp, -space);
704 mutex_exit(&dp->dp_lock);
709 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
712 dsl_dataset_t *ds, *prev = NULL;
715 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
719 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
720 err = dsl_dataset_hold_obj(dp,
721 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
723 dsl_dataset_rele(ds, FTAG);
727 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
729 dsl_dataset_rele(ds, FTAG);
735 prev = dp->dp_origin_snap;
738 * The $ORIGIN can't have any data, or the accounting
741 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
742 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
743 rrw_exit(&ds->ds_bp_rwlock, FTAG);
745 /* The origin doesn't get attached to itself */
746 if (ds->ds_object == prev->ds_object) {
747 dsl_dataset_rele(ds, FTAG);
751 dmu_buf_will_dirty(ds->ds_dbuf, tx);
752 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
753 dsl_dataset_phys(ds)->ds_prev_snap_txg =
754 dsl_dataset_phys(prev)->ds_creation_txg;
756 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
757 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
759 dmu_buf_will_dirty(prev->ds_dbuf, tx);
760 dsl_dataset_phys(prev)->ds_num_children++;
762 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
763 ASSERT(ds->ds_prev == NULL);
764 VERIFY0(dsl_dataset_hold_obj(dp,
765 dsl_dataset_phys(ds)->ds_prev_snap_obj,
770 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
771 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
773 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
774 dmu_buf_will_dirty(prev->ds_dbuf, tx);
775 dsl_dataset_phys(prev)->ds_next_clones_obj =
776 zap_create(dp->dp_meta_objset,
777 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
779 VERIFY0(zap_add_int(dp->dp_meta_objset,
780 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
782 dsl_dataset_rele(ds, FTAG);
783 if (prev != dp->dp_origin_snap)
784 dsl_dataset_rele(prev, FTAG);
789 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
791 ASSERT(dmu_tx_is_syncing(tx));
792 ASSERT(dp->dp_origin_snap != NULL);
794 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
795 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
800 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
803 objset_t *mos = dp->dp_meta_objset;
805 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
806 dsl_dataset_t *origin;
808 VERIFY0(dsl_dataset_hold_obj(dp,
809 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
811 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
812 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
813 dsl_dir_phys(origin->ds_dir)->dd_clones =
814 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
818 VERIFY0(zap_add_int(dp->dp_meta_objset,
819 dsl_dir_phys(origin->ds_dir)->dd_clones,
822 dsl_dataset_rele(origin, FTAG);
828 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
832 ASSERT(dmu_tx_is_syncing(tx));
834 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
835 VERIFY0(dsl_pool_open_special_dir(dp,
836 FREE_DIR_NAME, &dp->dp_free_dir));
839 * We can't use bpobj_alloc(), because spa_version() still
840 * returns the old version, and we need a new-version bpobj with
841 * subobj support. So call dmu_object_alloc() directly.
843 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
844 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
845 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
846 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
847 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
849 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
850 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
854 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
859 ASSERT(dmu_tx_is_syncing(tx));
860 ASSERT(dp->dp_origin_snap == NULL);
861 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
863 /* create the origin dir, ds, & snap-ds */
864 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
866 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
867 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
868 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
869 dp, &dp->dp_origin_snap));
870 dsl_dataset_rele(ds, FTAG);
874 dsl_pool_iput_taskq(dsl_pool_t *dp)
876 return (dp->dp_iput_taskq);
880 * Walk through the pool-wide zap object of temporary snapshot user holds
884 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
888 objset_t *mos = dp->dp_meta_objset;
889 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
894 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
896 holds = fnvlist_alloc();
898 for (zap_cursor_init(&zc, mos, zapobj);
899 zap_cursor_retrieve(&zc, &za) == 0;
900 zap_cursor_advance(&zc)) {
904 htag = strchr(za.za_name, '-');
907 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
908 tags = fnvlist_alloc();
909 fnvlist_add_boolean(tags, htag);
910 fnvlist_add_nvlist(holds, za.za_name, tags);
913 fnvlist_add_boolean(tags, htag);
916 dsl_dataset_user_release_tmp(dp, holds);
918 zap_cursor_fini(&zc);
922 * Create the pool-wide zap object for storing temporary snapshot holds.
925 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
927 objset_t *mos = dp->dp_meta_objset;
929 ASSERT(dp->dp_tmp_userrefs_obj == 0);
930 ASSERT(dmu_tx_is_syncing(tx));
932 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
933 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
937 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
938 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
940 objset_t *mos = dp->dp_meta_objset;
941 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
945 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
946 ASSERT(dmu_tx_is_syncing(tx));
949 * If the pool was created prior to SPA_VERSION_USERREFS, the
950 * zap object for temporary holds might not exist yet.
954 dsl_pool_user_hold_create_obj(dp, tx);
955 zapobj = dp->dp_tmp_userrefs_obj;
957 return (SET_ERROR(ENOENT));
961 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
963 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
965 error = zap_remove(mos, zapobj, name, tx);
972 * Add a temporary hold for the given dataset object and tag.
975 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
976 uint64_t now, dmu_tx_t *tx)
978 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
982 * Release a temporary hold for the given dataset object and tag.
985 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
988 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
993 * DSL Pool Configuration Lock
995 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
996 * creation / destruction / rename / property setting). It must be held for
997 * read to hold a dataset or dsl_dir. I.e. you must call
998 * dsl_pool_config_enter() or dsl_pool_hold() before calling
999 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1000 * must be held continuously until all datasets and dsl_dirs are released.
1002 * The only exception to this rule is that if a "long hold" is placed on
1003 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1004 * is still held. The long hold will prevent the dataset from being
1005 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1006 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1007 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1009 * Legitimate long-holders (including owners) should be long-running, cancelable
1010 * tasks that should cause "zfs destroy" to fail. This includes DMU
1011 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1012 * "zfs send", and "zfs diff". There are several other long-holders whose
1013 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1015 * The usual formula for long-holding would be:
1017 * dsl_dataset_hold()
1018 * ... perform checks ...
1019 * dsl_dataset_long_hold()
1021 * ... perform long-running task ...
1022 * dsl_dataset_long_rele()
1023 * dsl_dataset_rele()
1025 * Note that when the long hold is released, the dataset is still held but
1026 * the pool is not held. The dataset may change arbitrarily during this time
1027 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1028 * dataset except release it.
1030 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1031 * or modifying operations.
1033 * Modifying operations should generally use dsl_sync_task(). The synctask
1034 * infrastructure enforces proper locking strategy with respect to the
1035 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1037 * Read-only operations will manually hold the pool, then the dataset, obtain
1038 * information from the dataset, then release the pool and dataset.
1039 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1044 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1049 error = spa_open(name, &spa, tag);
1051 *dp = spa_get_dsl(spa);
1052 dsl_pool_config_enter(*dp, tag);
1058 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1060 dsl_pool_config_exit(dp, tag);
1061 spa_close(dp->dp_spa, tag);
1065 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1068 * We use a "reentrant" reader-writer lock, but not reentrantly.
1070 * The rrwlock can (with the track_all flag) track all reading threads,
1071 * which is very useful for debugging which code path failed to release
1072 * the lock, and for verifying that the *current* thread does hold
1075 * (Unlike a rwlock, which knows that N threads hold it for
1076 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1077 * if any thread holds it for read, even if this thread doesn't).
1079 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1080 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1084 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1086 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1087 rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1091 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1093 rrw_exit(&dp->dp_config_rwlock, tag);
1097 dsl_pool_config_held(dsl_pool_t *dp)
1099 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1103 dsl_pool_config_held_writer(dsl_pool_t *dp)
1105 return (RRW_WRITE_HELD(&dp->dp_config_rwlock));
1108 #if defined(_KERNEL) && defined(HAVE_SPL)
1109 EXPORT_SYMBOL(dsl_pool_config_enter);
1110 EXPORT_SYMBOL(dsl_pool_config_exit);
1113 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */
1114 module_param(zfs_dirty_data_max_percent, int, 0444);
1115 MODULE_PARM_DESC(zfs_dirty_data_max_percent, "percent of ram can be dirty");
1117 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */
1118 module_param(zfs_dirty_data_max_max_percent, int, 0444);
1119 MODULE_PARM_DESC(zfs_dirty_data_max_max_percent,
1120 "zfs_dirty_data_max upper bound as % of RAM");
1122 module_param(zfs_delay_min_dirty_percent, int, 0644);
1123 MODULE_PARM_DESC(zfs_delay_min_dirty_percent, "transaction delay threshold");
1125 module_param(zfs_dirty_data_max, ulong, 0644);
1126 MODULE_PARM_DESC(zfs_dirty_data_max, "determines the dirty space limit");
1128 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */
1129 module_param(zfs_dirty_data_max_max, ulong, 0444);
1130 MODULE_PARM_DESC(zfs_dirty_data_max_max,
1131 "zfs_dirty_data_max upper bound in bytes");
1133 module_param(zfs_dirty_data_sync, ulong, 0644);
1134 MODULE_PARM_DESC(zfs_dirty_data_sync, "sync txg when this much dirty data");
1136 module_param(zfs_delay_scale, ulong, 0644);
1137 MODULE_PARM_DESC(zfs_delay_scale, "how quickly delay approaches infinity");
1139 module_param(zfs_sync_taskq_batch_pct, int, 0644);
1140 MODULE_PARM_DESC(zfs_sync_taskq_batch_pct,
1141 "max percent of CPUs that are used to sync dirty data");