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) 2012 by Delphix. All rights reserved.
26 #include <sys/dsl_pool.h>
27 #include <sys/dsl_dataset.h>
28 #include <sys/dsl_prop.h>
29 #include <sys/dsl_dir.h>
30 #include <sys/dsl_synctask.h>
31 #include <sys/dsl_scan.h>
32 #include <sys/dnode.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
38 #include <sys/zfs_context.h>
39 #include <sys/fs/zfs.h>
40 #include <sys/zfs_znode.h>
41 #include <sys/spa_impl.h>
42 #include <sys/dsl_deadlist.h>
43 #include <sys/bptree.h>
44 #include <sys/zfeature.h>
46 int zfs_no_write_throttle = 0;
47 int zfs_write_limit_shift = 3; /* 1/8th of physical memory */
48 int zfs_txg_synctime_ms = 1000; /* target millisecs to sync a txg */
49 int zfs_txg_history = 60; /* statistics for the last N txgs */
51 unsigned long zfs_write_limit_min = 32 << 20; /* min write limit is 32MB */
52 unsigned long zfs_write_limit_max = 0; /* max data payload per txg */
53 unsigned long zfs_write_limit_inflated = 0;
54 unsigned long zfs_write_limit_override = 0;
56 kmutex_t zfs_write_limit_lock;
58 static pgcnt_t old_physmem = 0;
61 dsl_pool_txg_history_update(kstat_t *ksp, int rw)
63 dsl_pool_t *dp = ksp->ks_private;
67 if (rw == KSTAT_WRITE)
71 kmem_free(ksp->ks_data, ksp->ks_data_size);
73 mutex_enter(&dp->dp_lock);
75 ksp->ks_ndata = dp->dp_txg_history_size;
76 ksp->ks_data_size = dp->dp_txg_history_size * sizeof(kstat_txg_t);
77 if (ksp->ks_data_size > 0)
78 ksp->ks_data = kmem_alloc(ksp->ks_data_size, KM_PUSHPAGE);
80 /* Traversed oldest to youngest for the most readable kstat output */
81 for (th = list_tail(&dp->dp_txg_history); th != NULL;
82 th = list_prev(&dp->dp_txg_history, th)) {
83 mutex_enter(&th->th_lock);
84 ASSERT3S(i + sizeof(kstat_txg_t), <=, ksp->ks_data_size);
85 memcpy(ksp->ks_data + i, &th->th_kstat, sizeof(kstat_txg_t));
86 i += sizeof(kstat_txg_t);
87 mutex_exit(&th->th_lock);
90 mutex_exit(&dp->dp_lock);
96 dsl_pool_txg_history_init(dsl_pool_t *dp, uint64_t txg)
98 char name[KSTAT_STRLEN];
100 list_create(&dp->dp_txg_history, sizeof (txg_history_t),
101 offsetof(txg_history_t, th_link));
102 dsl_pool_txg_history_add(dp, txg);
104 (void) snprintf(name, KSTAT_STRLEN, "txgs-%s", spa_name(dp->dp_spa));
105 dp->dp_txg_kstat = kstat_create("zfs", 0, name, "misc",
106 KSTAT_TYPE_TXG, 0, KSTAT_FLAG_VIRTUAL);
107 if (dp->dp_txg_kstat) {
108 dp->dp_txg_kstat->ks_data = NULL;
109 dp->dp_txg_kstat->ks_private = dp;
110 dp->dp_txg_kstat->ks_update = dsl_pool_txg_history_update;
111 kstat_install(dp->dp_txg_kstat);
116 dsl_pool_txg_history_destroy(dsl_pool_t *dp)
120 if (dp->dp_txg_kstat) {
121 if (dp->dp_txg_kstat->ks_data)
122 kmem_free(dp->dp_txg_kstat->ks_data,
123 dp->dp_txg_kstat->ks_data_size);
125 kstat_delete(dp->dp_txg_kstat);
128 mutex_enter(&dp->dp_lock);
129 while ((th = list_remove_head(&dp->dp_txg_history))) {
130 dp->dp_txg_history_size--;
131 mutex_destroy(&th->th_lock);
132 kmem_free(th, sizeof(txg_history_t));
135 ASSERT3U(dp->dp_txg_history_size, ==, 0);
136 list_destroy(&dp->dp_txg_history);
137 mutex_exit(&dp->dp_lock);
141 dsl_pool_txg_history_add(dsl_pool_t *dp, uint64_t txg)
143 txg_history_t *th, *rm;
145 th = kmem_zalloc(sizeof(txg_history_t), KM_SLEEP);
146 mutex_init(&th->th_lock, NULL, MUTEX_DEFAULT, NULL);
147 th->th_kstat.txg = txg;
148 th->th_kstat.state = TXG_STATE_OPEN;
149 th->th_kstat.birth = gethrtime();
151 mutex_enter(&dp->dp_lock);
153 list_insert_head(&dp->dp_txg_history, th);
154 dp->dp_txg_history_size++;
156 while (dp->dp_txg_history_size > zfs_txg_history) {
157 dp->dp_txg_history_size--;
158 rm = list_remove_tail(&dp->dp_txg_history);
159 mutex_destroy(&rm->th_lock);
160 kmem_free(rm, sizeof(txg_history_t));
163 mutex_exit(&dp->dp_lock);
169 * Traversed youngest to oldest because lookups are only done for open
170 * or syncing txgs which are guaranteed to be at the head of the list.
171 * The txg_history_t structure will be returned locked.
174 dsl_pool_txg_history_get(dsl_pool_t *dp, uint64_t txg)
178 mutex_enter(&dp->dp_lock);
179 for (th = list_head(&dp->dp_txg_history); th != NULL;
180 th = list_next(&dp->dp_txg_history, th)) {
181 if (th->th_kstat.txg == txg) {
182 mutex_enter(&th->th_lock);
186 mutex_exit(&dp->dp_lock);
192 dsl_pool_txg_history_put(txg_history_t *th)
194 mutex_exit(&th->th_lock);
198 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
203 err = zap_lookup(dp->dp_meta_objset,
204 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
205 name, sizeof (obj), 1, &obj);
209 return (dsl_dir_open_obj(dp, obj, name, dp, ddp));
213 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
216 blkptr_t *bp = spa_get_rootblkptr(spa);
218 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
220 dp->dp_meta_rootbp = *bp;
221 rw_init(&dp->dp_config_rwlock, NULL, RW_DEFAULT, NULL);
222 dp->dp_write_limit = zfs_write_limit_min;
225 txg_list_create(&dp->dp_dirty_datasets,
226 offsetof(dsl_dataset_t, ds_dirty_link));
227 txg_list_create(&dp->dp_dirty_dirs,
228 offsetof(dsl_dir_t, dd_dirty_link));
229 txg_list_create(&dp->dp_sync_tasks,
230 offsetof(dsl_sync_task_group_t, dstg_node));
231 list_create(&dp->dp_synced_datasets, sizeof (dsl_dataset_t),
232 offsetof(dsl_dataset_t, ds_synced_link));
234 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
236 dp->dp_iput_taskq = taskq_create("zfs_iput_taskq", 1, minclsyspri,
239 dsl_pool_txg_history_init(dp, txg);
245 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
248 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
250 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
251 &dp->dp_meta_objset);
261 dsl_pool_open(dsl_pool_t *dp)
268 rw_enter(&dp->dp_config_rwlock, RW_WRITER);
269 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
270 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
271 &dp->dp_root_dir_obj);
275 err = dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
276 NULL, dp, &dp->dp_root_dir);
280 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
284 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
285 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
288 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
291 err = dsl_dataset_hold_obj(dp,
292 ds->ds_phys->ds_prev_snap_obj, dp,
293 &dp->dp_origin_snap);
294 dsl_dataset_rele(ds, FTAG);
296 dsl_dir_close(dd, dp);
301 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
302 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
307 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
308 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
311 VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
312 dp->dp_meta_objset, obj));
315 if (spa_feature_is_active(dp->dp_spa,
316 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
317 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
318 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
324 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
325 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
326 &dp->dp_tmp_userrefs_obj);
332 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
335 rw_exit(&dp->dp_config_rwlock);
340 dsl_pool_close(dsl_pool_t *dp)
342 /* drop our references from dsl_pool_open() */
345 * Since we held the origin_snap from "syncing" context (which
346 * includes pool-opening context), it actually only got a "ref"
347 * and not a hold, so just drop that here.
349 if (dp->dp_origin_snap)
350 dsl_dataset_drop_ref(dp->dp_origin_snap, dp);
352 dsl_dir_close(dp->dp_mos_dir, dp);
354 dsl_dir_close(dp->dp_free_dir, dp);
356 dsl_dir_close(dp->dp_root_dir, dp);
358 bpobj_close(&dp->dp_free_bpobj);
360 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
361 if (dp->dp_meta_objset)
362 dmu_objset_evict(dp->dp_meta_objset);
364 txg_list_destroy(&dp->dp_dirty_datasets);
365 txg_list_destroy(&dp->dp_sync_tasks);
366 txg_list_destroy(&dp->dp_dirty_dirs);
367 list_destroy(&dp->dp_synced_datasets);
369 arc_flush(dp->dp_spa);
372 dsl_pool_txg_history_destroy(dp);
373 rw_destroy(&dp->dp_config_rwlock);
374 mutex_destroy(&dp->dp_lock);
375 taskq_destroy(dp->dp_iput_taskq);
377 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
378 kmem_free(dp, sizeof (dsl_pool_t));
382 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
385 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
386 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
391 /* create and open the MOS (meta-objset) */
392 dp->dp_meta_objset = dmu_objset_create_impl(spa,
393 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
395 /* create the pool directory */
396 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
397 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
398 ASSERT3U(err, ==, 0);
400 /* Initialize scan structures */
401 VERIFY3U(0, ==, dsl_scan_init(dp, txg));
403 /* create and open the root dir */
404 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
405 VERIFY(0 == dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
406 NULL, dp, &dp->dp_root_dir));
408 /* create and open the meta-objset dir */
409 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
410 VERIFY(0 == dsl_pool_open_special_dir(dp,
411 MOS_DIR_NAME, &dp->dp_mos_dir));
413 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
414 /* create and open the free dir */
415 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
417 VERIFY(0 == dsl_pool_open_special_dir(dp,
418 FREE_DIR_NAME, &dp->dp_free_dir));
420 /* create and open the free_bplist */
421 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
422 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
423 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
424 VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
425 dp->dp_meta_objset, obj));
428 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
429 dsl_pool_create_origin(dp, tx);
431 /* create the root dataset */
432 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
434 /* create the root objset */
435 VERIFY(0 == dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
436 VERIFY(NULL != (os = dmu_objset_create_impl(dp->dp_spa, ds,
437 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx)));
439 zfs_create_fs(os, kcred, zplprops, tx);
441 dsl_dataset_rele(ds, FTAG);
449 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
451 dsl_deadlist_t *dl = arg;
452 dsl_pool_t *dp = dmu_objset_pool(dl->dl_os);
453 rw_enter(&dp->dp_config_rwlock, RW_READER);
454 dsl_deadlist_insert(dl, bp, tx);
455 rw_exit(&dp->dp_config_rwlock);
460 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
466 dsl_sync_task_group_t *dstg;
467 objset_t *mos = dp->dp_meta_objset;
468 hrtime_t start, write_time;
469 uint64_t data_written;
473 * We need to copy dp_space_towrite() before doing
474 * dsl_sync_task_group_sync(), because
475 * dsl_dataset_snapshot_reserve_space() will increase
476 * dp_space_towrite but not actually write anything.
478 data_written = dp->dp_space_towrite[txg & TXG_MASK];
480 tx = dmu_tx_create_assigned(dp, txg);
482 dp->dp_read_overhead = 0;
485 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
486 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg))) {
488 * We must not sync any non-MOS datasets twice, because
489 * we may have taken a snapshot of them. However, we
490 * may sync newly-created datasets on pass 2.
492 ASSERT(!list_link_active(&ds->ds_synced_link));
493 list_insert_tail(&dp->dp_synced_datasets, ds);
494 dsl_dataset_sync(ds, zio, tx);
496 DTRACE_PROBE(pool_sync__1setup);
499 write_time = gethrtime() - start;
501 DTRACE_PROBE(pool_sync__2rootzio);
503 for (ds = list_head(&dp->dp_synced_datasets); ds;
504 ds = list_next(&dp->dp_synced_datasets, ds))
505 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
508 * Sync the datasets again to push out the changes due to
509 * userspace updates. This must be done before we process the
510 * sync tasks, because that could cause a snapshot of a dataset
511 * whose ds_bp will be rewritten when we do this 2nd sync.
513 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
514 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg))) {
515 ASSERT(list_link_active(&ds->ds_synced_link));
516 dmu_buf_rele(ds->ds_dbuf, ds);
517 dsl_dataset_sync(ds, zio, tx);
522 * Move dead blocks from the pending deadlist to the on-disk
525 for (ds = list_head(&dp->dp_synced_datasets); ds;
526 ds = list_next(&dp->dp_synced_datasets, ds)) {
527 bplist_iterate(&ds->ds_pending_deadlist,
528 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
531 while ((dstg = txg_list_remove(&dp->dp_sync_tasks, txg))) {
533 * No more sync tasks should have been added while we
536 ASSERT(spa_sync_pass(dp->dp_spa) == 1);
537 dsl_sync_task_group_sync(dstg, tx);
539 DTRACE_PROBE(pool_sync__3task);
542 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)))
543 dsl_dir_sync(dd, tx);
544 write_time += gethrtime() - start;
547 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
548 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
549 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
550 dmu_objset_sync(mos, zio, tx);
553 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
554 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
556 write_time += gethrtime() - start;
557 DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
558 hrtime_t, dp->dp_read_overhead);
559 write_time -= dp->dp_read_overhead;
563 dp->dp_space_towrite[txg & TXG_MASK] = 0;
564 ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
567 * If the write limit max has not been explicitly set, set it
568 * to a fraction of available physical memory (default 1/8th).
569 * Note that we must inflate the limit because the spa
570 * inflates write sizes to account for data replication.
571 * Check this each sync phase to catch changing memory size.
573 if (physmem != old_physmem && zfs_write_limit_shift) {
574 mutex_enter(&zfs_write_limit_lock);
575 old_physmem = physmem;
576 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
577 zfs_write_limit_inflated = MAX(zfs_write_limit_min,
578 spa_get_asize(dp->dp_spa, zfs_write_limit_max));
579 mutex_exit(&zfs_write_limit_lock);
583 * Attempt to keep the sync time consistent by adjusting the
584 * amount of write traffic allowed into each transaction group.
585 * Weight the throughput calculation towards the current value:
586 * thru = 3/4 old_thru + 1/4 new_thru
588 * Note: write_time is in nanosecs, so write_time/MICROSEC
591 ASSERT(zfs_write_limit_min > 0);
592 if (data_written > zfs_write_limit_min / 8 && write_time > MICROSEC) {
593 uint64_t throughput = data_written / (write_time / MICROSEC);
595 if (dp->dp_throughput)
596 dp->dp_throughput = throughput / 4 +
597 3 * dp->dp_throughput / 4;
599 dp->dp_throughput = throughput;
600 dp->dp_write_limit = MIN(zfs_write_limit_inflated,
601 MAX(zfs_write_limit_min,
602 dp->dp_throughput * zfs_txg_synctime_ms));
607 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
612 while ((ds = list_head(&dp->dp_synced_datasets))) {
613 list_remove(&dp->dp_synced_datasets, ds);
615 zil_clean(os->os_zil, txg);
616 ASSERT(!dmu_objset_is_dirty(os, txg));
617 dmu_buf_rele(ds->ds_dbuf, ds);
619 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
623 * TRUE if the current thread is the tx_sync_thread or if we
624 * are being called from SPA context during pool initialization.
627 dsl_pool_sync_context(dsl_pool_t *dp)
629 return (curthread == dp->dp_tx.tx_sync_thread ||
630 spa_is_initializing(dp->dp_spa));
634 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
636 uint64_t space, resv;
639 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
641 * XXX The intent log is not accounted for, so it must fit
644 * If we're trying to assess whether it's OK to do a free,
645 * cut the reservation in half to allow forward progress
646 * (e.g. make it possible to rm(1) files from a full pool).
648 space = spa_get_dspace(dp->dp_spa);
649 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
653 return (space - resv);
657 dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
659 uint64_t reserved = 0;
660 uint64_t write_limit = (zfs_write_limit_override ?
661 zfs_write_limit_override : dp->dp_write_limit);
663 if (zfs_no_write_throttle) {
664 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
670 * Check to see if we have exceeded the maximum allowed IO for
671 * this transaction group. We can do this without locks since
672 * a little slop here is ok. Note that we do the reserved check
673 * with only half the requested reserve: this is because the
674 * reserve requests are worst-case, and we really don't want to
675 * throttle based off of worst-case estimates.
677 if (write_limit > 0) {
678 reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
679 + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
681 if (reserved && reserved > write_limit) {
682 DMU_TX_STAT_BUMP(dmu_tx_write_limit);
687 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
690 * If this transaction group is over 7/8ths capacity, delay
691 * the caller 1 clock tick. This will slow down the "fill"
692 * rate until the sync process can catch up with us.
694 if (reserved && reserved > (write_limit - (write_limit >> 3)))
695 txg_delay(dp, tx->tx_txg, 1);
701 dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
703 ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
704 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
708 dsl_pool_memory_pressure(dsl_pool_t *dp)
710 uint64_t space_inuse = 0;
713 if (dp->dp_write_limit == zfs_write_limit_min)
716 for (i = 0; i < TXG_SIZE; i++) {
717 space_inuse += dp->dp_space_towrite[i];
718 space_inuse += dp->dp_tempreserved[i];
720 dp->dp_write_limit = MAX(zfs_write_limit_min,
721 MIN(dp->dp_write_limit, space_inuse / 4));
725 dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
728 mutex_enter(&dp->dp_lock);
729 dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
730 mutex_exit(&dp->dp_lock);
736 upgrade_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
739 dsl_dataset_t *ds, *prev = NULL;
741 dsl_pool_t *dp = spa_get_dsl(spa);
743 err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
747 while (ds->ds_phys->ds_prev_snap_obj != 0) {
748 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
751 dsl_dataset_rele(ds, FTAG);
755 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
757 dsl_dataset_rele(ds, FTAG);
763 prev = dp->dp_origin_snap;
766 * The $ORIGIN can't have any data, or the accounting
769 ASSERT(prev->ds_phys->ds_bp.blk_birth == 0);
771 /* The origin doesn't get attached to itself */
772 if (ds->ds_object == prev->ds_object) {
773 dsl_dataset_rele(ds, FTAG);
777 dmu_buf_will_dirty(ds->ds_dbuf, tx);
778 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
779 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
781 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
782 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
784 dmu_buf_will_dirty(prev->ds_dbuf, tx);
785 prev->ds_phys->ds_num_children++;
787 if (ds->ds_phys->ds_next_snap_obj == 0) {
788 ASSERT(ds->ds_prev == NULL);
789 VERIFY(0 == dsl_dataset_hold_obj(dp,
790 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
794 ASSERT(ds->ds_dir->dd_phys->dd_origin_obj == prev->ds_object);
795 ASSERT(ds->ds_phys->ds_prev_snap_obj == prev->ds_object);
797 if (prev->ds_phys->ds_next_clones_obj == 0) {
798 dmu_buf_will_dirty(prev->ds_dbuf, tx);
799 prev->ds_phys->ds_next_clones_obj =
800 zap_create(dp->dp_meta_objset,
801 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
803 VERIFY(0 == zap_add_int(dp->dp_meta_objset,
804 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
806 dsl_dataset_rele(ds, FTAG);
807 if (prev != dp->dp_origin_snap)
808 dsl_dataset_rele(prev, FTAG);
813 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
815 ASSERT(dmu_tx_is_syncing(tx));
816 ASSERT(dp->dp_origin_snap != NULL);
818 VERIFY3U(0, ==, dmu_objset_find_spa(dp->dp_spa, NULL, upgrade_clones_cb,
819 tx, DS_FIND_CHILDREN));
824 upgrade_dir_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
828 dsl_pool_t *dp = spa_get_dsl(spa);
829 objset_t *mos = dp->dp_meta_objset;
831 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
833 if (ds->ds_dir->dd_phys->dd_origin_obj) {
834 dsl_dataset_t *origin;
836 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
837 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
839 if (origin->ds_dir->dd_phys->dd_clones == 0) {
840 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
841 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
842 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
845 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
846 origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
848 dsl_dataset_rele(origin, FTAG);
851 dsl_dataset_rele(ds, FTAG);
856 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
860 ASSERT(dmu_tx_is_syncing(tx));
862 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
863 VERIFY(0 == dsl_pool_open_special_dir(dp,
864 FREE_DIR_NAME, &dp->dp_free_dir));
867 * We can't use bpobj_alloc(), because spa_version() still
868 * returns the old version, and we need a new-version bpobj with
869 * subobj support. So call dmu_object_alloc() directly.
871 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
872 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
873 VERIFY3U(0, ==, zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
874 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
875 VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
876 dp->dp_meta_objset, obj));
878 VERIFY3U(0, ==, dmu_objset_find_spa(dp->dp_spa, NULL,
879 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
883 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
888 ASSERT(dmu_tx_is_syncing(tx));
889 ASSERT(dp->dp_origin_snap == NULL);
891 /* create the origin dir, ds, & snap-ds */
892 rw_enter(&dp->dp_config_rwlock, RW_WRITER);
893 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
895 VERIFY(0 == dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
896 dsl_dataset_snapshot_sync(ds, ORIGIN_DIR_NAME, tx);
897 VERIFY(0 == dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
898 dp, &dp->dp_origin_snap));
899 dsl_dataset_rele(ds, FTAG);
900 rw_exit(&dp->dp_config_rwlock);
904 dsl_pool_iput_taskq(dsl_pool_t *dp)
906 return (dp->dp_iput_taskq);
910 * Walk through the pool-wide zap object of temporary snapshot user holds
914 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
918 objset_t *mos = dp->dp_meta_objset;
919 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
923 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
925 for (zap_cursor_init(&zc, mos, zapobj);
926 zap_cursor_retrieve(&zc, &za) == 0;
927 zap_cursor_advance(&zc)) {
931 htag = strchr(za.za_name, '-');
934 dsobj = strtonum(za.za_name, NULL);
935 (void) dsl_dataset_user_release_tmp(dp, dsobj, htag, B_FALSE);
937 zap_cursor_fini(&zc);
941 * Create the pool-wide zap object for storing temporary snapshot holds.
944 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
946 objset_t *mos = dp->dp_meta_objset;
948 ASSERT(dp->dp_tmp_userrefs_obj == 0);
949 ASSERT(dmu_tx_is_syncing(tx));
951 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
952 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
956 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
957 const char *tag, uint64_t *now, dmu_tx_t *tx, boolean_t holding)
959 objset_t *mos = dp->dp_meta_objset;
960 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
964 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
965 ASSERT(dmu_tx_is_syncing(tx));
968 * If the pool was created prior to SPA_VERSION_USERREFS, the
969 * zap object for temporary holds might not exist yet.
973 dsl_pool_user_hold_create_obj(dp, tx);
974 zapobj = dp->dp_tmp_userrefs_obj;
980 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
982 error = zap_add(mos, zapobj, name, 8, 1, now, tx);
984 error = zap_remove(mos, zapobj, name, tx);
991 * Add a temporary hold for the given dataset object and tag.
994 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
995 uint64_t *now, dmu_tx_t *tx)
997 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
1001 * Release a temporary hold for the given dataset object and tag.
1004 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1007 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
1011 #if defined(_KERNEL) && defined(HAVE_SPL)
1012 module_param(zfs_no_write_throttle, int, 0644);
1013 MODULE_PARM_DESC(zfs_no_write_throttle, "Disable write throttling");
1015 module_param(zfs_write_limit_shift, int, 0444);
1016 MODULE_PARM_DESC(zfs_write_limit_shift, "log2(fraction of memory) per txg");
1018 module_param(zfs_txg_synctime_ms, int, 0644);
1019 MODULE_PARM_DESC(zfs_txg_synctime_ms, "Target milliseconds between txg sync");
1021 module_param(zfs_txg_history, int, 0644);
1022 MODULE_PARM_DESC(zfs_txg_history, "Historic statistics for the last N txgs");
1024 module_param(zfs_write_limit_min, ulong, 0444);
1025 MODULE_PARM_DESC(zfs_write_limit_min, "Min txg write limit");
1027 module_param(zfs_write_limit_max, ulong, 0444);
1028 MODULE_PARM_DESC(zfs_write_limit_max, "Max txg write limit");
1030 module_param(zfs_write_limit_inflated, ulong, 0444);
1031 MODULE_PARM_DESC(zfs_write_limit_inflated, "Inflated txg write limit");
1033 module_param(zfs_write_limit_override, ulong, 0444);
1034 MODULE_PARM_DESC(zfs_write_limit_override, "Override txg write limit");