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 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
28 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_pool.h>
35 #include <sys/zap_impl.h>
38 #include <sys/sa_impl.h>
39 #include <sys/zfs_context.h>
40 #include <sys/varargs.h>
41 #include <sys/trace_dmu.h>
43 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
44 uint64_t arg1, uint64_t arg2);
46 dmu_tx_stats_t dmu_tx_stats = {
47 { "dmu_tx_assigned", KSTAT_DATA_UINT64 },
48 { "dmu_tx_delay", KSTAT_DATA_UINT64 },
49 { "dmu_tx_error", KSTAT_DATA_UINT64 },
50 { "dmu_tx_suspended", KSTAT_DATA_UINT64 },
51 { "dmu_tx_group", KSTAT_DATA_UINT64 },
52 { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64 },
53 { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64 },
54 { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64 },
55 { "dmu_tx_dirty_delay", KSTAT_DATA_UINT64 },
56 { "dmu_tx_dirty_over_max", KSTAT_DATA_UINT64 },
57 { "dmu_tx_quota", KSTAT_DATA_UINT64 },
60 static kstat_t *dmu_tx_ksp;
63 dmu_tx_create_dd(dsl_dir_t *dd)
65 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
68 tx->tx_pool = dd->dd_pool;
69 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
70 offsetof(dmu_tx_hold_t, txh_node));
71 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
72 offsetof(dmu_tx_callback_t, dcb_node));
73 tx->tx_start = gethrtime();
78 dmu_tx_create(objset_t *os)
80 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
86 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
88 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
90 txg_verify(dp->dp_spa, txg);
99 dmu_tx_is_syncing(dmu_tx_t *tx)
101 return (tx->tx_anyobj);
105 dmu_tx_private_ok(dmu_tx_t *tx)
107 return (tx->tx_anyobj);
110 static dmu_tx_hold_t *
111 dmu_tx_hold_dnode_impl(dmu_tx_t *tx, dnode_t *dn, enum dmu_tx_hold_type type,
112 uint64_t arg1, uint64_t arg2)
117 (void) refcount_add(&dn->dn_holds, tx);
118 if (tx->tx_txg != 0) {
119 mutex_enter(&dn->dn_mtx);
121 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
122 * problem, but there's no way for it to happen (for
125 ASSERT(dn->dn_assigned_txg == 0);
126 dn->dn_assigned_txg = tx->tx_txg;
127 (void) refcount_add(&dn->dn_tx_holds, tx);
128 mutex_exit(&dn->dn_mtx);
132 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
135 refcount_create(&txh->txh_space_towrite);
136 refcount_create(&txh->txh_memory_tohold);
137 txh->txh_type = type;
138 txh->txh_arg1 = arg1;
139 txh->txh_arg2 = arg2;
140 list_insert_tail(&tx->tx_holds, txh);
145 static dmu_tx_hold_t *
146 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
147 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
153 if (object != DMU_NEW_OBJECT) {
154 err = dnode_hold(os, object, FTAG, &dn);
160 txh = dmu_tx_hold_dnode_impl(tx, dn, type, arg1, arg2);
162 dnode_rele(dn, FTAG);
167 dmu_tx_add_new_object(dmu_tx_t *tx, dnode_t *dn)
170 * If we're syncing, they can manipulate any object anyhow, and
171 * the hold on the dnode_t can cause problems.
173 if (!dmu_tx_is_syncing(tx))
174 (void) dmu_tx_hold_dnode_impl(tx, dn, THT_NEWOBJECT, 0, 0);
178 * This function reads specified data from disk. The specified data will
179 * be needed to perform the transaction -- i.e, it will be read after
180 * we do dmu_tx_assign(). There are two reasons that we read the data now
181 * (before dmu_tx_assign()):
183 * 1. Reading it now has potentially better performance. The transaction
184 * has not yet been assigned, so the TXG is not held open, and also the
185 * caller typically has less locks held when calling dmu_tx_hold_*() than
186 * after the transaction has been assigned. This reduces the lock (and txg)
187 * hold times, thus reducing lock contention.
189 * 2. It is easier for callers (primarily the ZPL) to handle i/o errors
190 * that are detected before they start making changes to the DMU state
191 * (i.e. now). Once the transaction has been assigned, and some DMU
192 * state has been changed, it can be difficult to recover from an i/o
193 * error (e.g. to undo the changes already made in memory at the DMU
194 * layer). Typically code to do so does not exist in the caller -- it
195 * assumes that the data has already been cached and thus i/o errors are
198 * It has been observed that the i/o initiated here can be a performance
199 * problem, and it appears to be optional, because we don't look at the
200 * data which is read. However, removing this read would only serve to
201 * move the work elsewhere (after the dmu_tx_assign()), where it may
202 * have a greater impact on performance (in addition to the impact on
203 * fault tolerance noted above).
206 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
211 rw_enter(&dn->dn_struct_rwlock, RW_READER);
212 db = dbuf_hold_level(dn, level, blkid, FTAG);
213 rw_exit(&dn->dn_struct_rwlock);
215 return (SET_ERROR(EIO));
216 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
223 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
225 dnode_t *dn = txh->txh_dnode;
231 (void) refcount_add_many(&txh->txh_space_towrite, len, FTAG);
233 if (refcount_count(&txh->txh_space_towrite) > 2 * DMU_MAX_ACCESS)
234 err = SET_ERROR(EFBIG);
240 * For i/o error checking, read the blocks that will be needed
241 * to perform the write: the first and last level-0 blocks (if
242 * they are not aligned, i.e. if they are partial-block writes),
243 * and all the level-1 blocks.
245 if (dn->dn_maxblkid == 0) {
246 if (off < dn->dn_datablksz &&
247 (off > 0 || len < dn->dn_datablksz)) {
248 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
250 txh->txh_tx->tx_err = err;
254 zio_t *zio = zio_root(dn->dn_objset->os_spa,
255 NULL, NULL, ZIO_FLAG_CANFAIL);
257 /* first level-0 block */
258 uint64_t start = off >> dn->dn_datablkshift;
259 if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
260 err = dmu_tx_check_ioerr(zio, dn, 0, start);
262 txh->txh_tx->tx_err = err;
266 /* last level-0 block */
267 uint64_t end = (off + len - 1) >> dn->dn_datablkshift;
268 if (end != start && end <= dn->dn_maxblkid &&
269 P2PHASE(off + len, dn->dn_datablksz)) {
270 err = dmu_tx_check_ioerr(zio, dn, 0, end);
272 txh->txh_tx->tx_err = err;
277 if (dn->dn_nlevels > 1) {
278 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
279 for (uint64_t i = (start >> shft) + 1;
280 i < end >> shft; i++) {
281 err = dmu_tx_check_ioerr(zio, dn, 1, i);
283 txh->txh_tx->tx_err = err;
290 txh->txh_tx->tx_err = err;
296 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
298 (void) refcount_add_many(&txh->txh_space_towrite, DNODE_MIN_SIZE, FTAG);
302 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
307 ASSERT3U(len, <=, DMU_MAX_ACCESS);
308 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
310 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
311 object, THT_WRITE, off, len);
313 dmu_tx_count_write(txh, off, len);
314 dmu_tx_count_dnode(txh);
319 dmu_tx_hold_remap_l1indirect(dmu_tx_t *tx, uint64_t object)
323 ASSERT(tx->tx_txg == 0);
324 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
325 object, THT_WRITE, 0, 0);
329 dnode_t *dn = txh->txh_dnode;
330 (void) refcount_add_many(&txh->txh_space_towrite,
331 1ULL << dn->dn_indblkshift, FTAG);
332 dmu_tx_count_dnode(txh);
336 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
341 ASSERT3U(len, <=, DMU_MAX_ACCESS);
342 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
344 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
346 dmu_tx_count_write(txh, off, len);
347 dmu_tx_count_dnode(txh);
352 * This function marks the transaction as being a "net free". The end
353 * result is that refquotas will be disabled for this transaction, and
354 * this transaction will be able to use half of the pool space overhead
355 * (see dsl_pool_adjustedsize()). Therefore this function should only
356 * be called for transactions that we expect will not cause a net increase
357 * in the amount of space used (but it's OK if that is occasionally not true).
360 dmu_tx_mark_netfree(dmu_tx_t *tx)
362 tx->tx_netfree = B_TRUE;
366 dmu_tx_hold_free_impl(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
368 dmu_tx_t *tx = txh->txh_tx;
369 dnode_t *dn = txh->txh_dnode;
372 ASSERT(tx->tx_txg == 0);
374 dmu_tx_count_dnode(txh);
376 if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
378 if (len == DMU_OBJECT_END)
379 len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
381 dmu_tx_count_dnode(txh);
384 * For i/o error checking, we read the first and last level-0
385 * blocks if they are not aligned, and all the level-1 blocks.
387 * Note: dbuf_free_range() assumes that we have not instantiated
388 * any level-0 dbufs that will be completely freed. Therefore we must
389 * exercise care to not read or count the first and last blocks
390 * if they are blocksize-aligned.
392 if (dn->dn_datablkshift == 0) {
393 if (off != 0 || len < dn->dn_datablksz)
394 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
396 /* first block will be modified if it is not aligned */
397 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
398 dmu_tx_count_write(txh, off, 1);
399 /* last block will be modified if it is not aligned */
400 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
401 dmu_tx_count_write(txh, off + len, 1);
405 * Check level-1 blocks.
407 if (dn->dn_nlevels > 1) {
408 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
410 uint64_t start = off >> shift;
411 uint64_t end = (off + len) >> shift;
413 ASSERT(dn->dn_indblkshift != 0);
416 * dnode_reallocate() can result in an object with indirect
417 * blocks having an odd data block size. In this case,
418 * just check the single block.
420 if (dn->dn_datablkshift == 0)
423 zio_t *zio = zio_root(tx->tx_pool->dp_spa,
424 NULL, NULL, ZIO_FLAG_CANFAIL);
425 for (uint64_t i = start; i <= end; i++) {
426 uint64_t ibyte = i << shift;
427 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
429 if (err == ESRCH || i > end)
433 (void) zio_wait(zio);
437 (void) refcount_add_many(&txh->txh_memory_tohold,
438 1 << dn->dn_indblkshift, FTAG);
440 err = dmu_tx_check_ioerr(zio, dn, 1, i);
443 (void) zio_wait(zio);
456 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
460 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
461 object, THT_FREE, off, len);
463 (void) dmu_tx_hold_free_impl(txh, off, len);
467 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
471 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
473 (void) dmu_tx_hold_free_impl(txh, off, len);
477 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
479 dmu_tx_t *tx = txh->txh_tx;
480 dnode_t *dn = txh->txh_dnode;
483 ASSERT(tx->tx_txg == 0);
485 dmu_tx_count_dnode(txh);
488 * Modifying a almost-full microzap is around the worst case (128KB)
490 * If it is a fat zap, the worst case would be 7*16KB=112KB:
491 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
492 * - 4 new blocks written if adding:
493 * - 2 blocks for possibly split leaves,
494 * - 2 grown ptrtbl blocks
496 (void) refcount_add_many(&txh->txh_space_towrite,
497 MZAP_MAX_BLKSZ, FTAG);
502 ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
504 if (dn->dn_maxblkid == 0 || name == NULL) {
506 * This is a microzap (only one block), or we don't know
507 * the name. Check the first block for i/o errors.
509 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
515 * Access the name so that we'll check for i/o errors to
516 * the leaf blocks, etc. We ignore ENOENT, as this name
519 err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
520 if (err == EIO || err == ECKSUM || err == ENXIO) {
527 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
533 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
534 object, THT_ZAP, add, (uintptr_t)name);
536 dmu_tx_hold_zap_impl(txh, name);
540 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
547 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
549 dmu_tx_hold_zap_impl(txh, name);
553 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
557 ASSERT(tx->tx_txg == 0);
559 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
560 object, THT_BONUS, 0, 0);
562 dmu_tx_count_dnode(txh);
566 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
572 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
574 dmu_tx_count_dnode(txh);
578 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
582 ASSERT(tx->tx_txg == 0);
584 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
585 DMU_NEW_OBJECT, THT_SPACE, space, 0);
587 (void) refcount_add_many(&txh->txh_space_towrite, space, FTAG);
592 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
594 boolean_t match_object = B_FALSE;
595 boolean_t match_offset = B_FALSE;
598 dnode_t *dn = DB_DNODE(db);
599 ASSERT(tx->tx_txg != 0);
600 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
601 ASSERT3U(dn->dn_object, ==, db->db.db_object);
608 /* XXX No checking on the meta dnode for now */
609 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
614 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
615 txh = list_next(&tx->tx_holds, txh)) {
616 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
617 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
619 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
620 int datablkshift = dn->dn_datablkshift ?
621 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
622 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
623 int shift = datablkshift + epbs * db->db_level;
624 uint64_t beginblk = shift >= 64 ? 0 :
625 (txh->txh_arg1 >> shift);
626 uint64_t endblk = shift >= 64 ? 0 :
627 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
628 uint64_t blkid = db->db_blkid;
630 /* XXX txh_arg2 better not be zero... */
632 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
633 txh->txh_type, beginblk, endblk);
635 switch (txh->txh_type) {
637 if (blkid >= beginblk && blkid <= endblk)
640 * We will let this hold work for the bonus
641 * or spill buffer so that we don't need to
642 * hold it when creating a new object.
644 if (blkid == DMU_BONUS_BLKID ||
645 blkid == DMU_SPILL_BLKID)
648 * They might have to increase nlevels,
649 * thus dirtying the new TLIBs. Or the
650 * might have to change the block size,
651 * thus dirying the new lvl=0 blk=0.
658 * We will dirty all the level 1 blocks in
659 * the free range and perhaps the first and
660 * last level 0 block.
662 if (blkid >= beginblk && (blkid <= endblk ||
663 txh->txh_arg2 == DMU_OBJECT_END))
667 if (blkid == DMU_SPILL_BLKID)
671 if (blkid == DMU_BONUS_BLKID)
681 cmn_err(CE_PANIC, "bad txh_type %d",
685 if (match_object && match_offset) {
691 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
692 (u_longlong_t)db->db.db_object, db->db_level,
693 (u_longlong_t)db->db_blkid);
698 * If we can't do 10 iops, something is wrong. Let us go ahead
699 * and hit zfs_dirty_data_max.
701 hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
702 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
705 * We delay transactions when we've determined that the backend storage
706 * isn't able to accommodate the rate of incoming writes.
708 * If there is already a transaction waiting, we delay relative to when
709 * that transaction finishes waiting. This way the calculated min_time
710 * is independent of the number of threads concurrently executing
713 * If we are the only waiter, wait relative to when the transaction
714 * started, rather than the current time. This credits the transaction for
715 * "time already served", e.g. reading indirect blocks.
717 * The minimum time for a transaction to take is calculated as:
718 * min_time = scale * (dirty - min) / (max - dirty)
719 * min_time is then capped at zfs_delay_max_ns.
721 * The delay has two degrees of freedom that can be adjusted via tunables.
722 * The percentage of dirty data at which we start to delay is defined by
723 * zfs_delay_min_dirty_percent. This should typically be at or above
724 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
725 * delay after writing at full speed has failed to keep up with the incoming
726 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
727 * speaking, this variable determines the amount of delay at the midpoint of
731 * 10ms +-------------------------------------------------------------*+
747 * 2ms + (midpoint) * +
750 * | zfs_delay_scale ----------> ******** |
751 * 0 +-------------------------------------*********----------------+
752 * 0% <- zfs_dirty_data_max -> 100%
754 * Note that since the delay is added to the outstanding time remaining on the
755 * most recent transaction, the delay is effectively the inverse of IOPS.
756 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
757 * was chosen such that small changes in the amount of accumulated dirty data
758 * in the first 3/4 of the curve yield relatively small differences in the
761 * The effects can be easier to understand when the amount of delay is
762 * represented on a log scale:
765 * 100ms +-------------------------------------------------------------++
774 * + zfs_delay_scale ----------> ***** +
785 * +--------------------------------------------------------------+
786 * 0% <- zfs_dirty_data_max -> 100%
788 * Note here that only as the amount of dirty data approaches its limit does
789 * the delay start to increase rapidly. The goal of a properly tuned system
790 * should be to keep the amount of dirty data out of that range by first
791 * ensuring that the appropriate limits are set for the I/O scheduler to reach
792 * optimal throughput on the backend storage, and then by changing the value
793 * of zfs_delay_scale to increase the steepness of the curve.
796 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
798 dsl_pool_t *dp = tx->tx_pool;
799 uint64_t delay_min_bytes =
800 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
801 hrtime_t wakeup, min_tx_time, now;
803 if (dirty <= delay_min_bytes)
807 * The caller has already waited until we are under the max.
808 * We make them pass us the amount of dirty data so we don't
809 * have to handle the case of it being >= the max, which could
810 * cause a divide-by-zero if it's == the max.
812 ASSERT3U(dirty, <, zfs_dirty_data_max);
815 min_tx_time = zfs_delay_scale *
816 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
817 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
818 if (now > tx->tx_start + min_tx_time)
821 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
822 uint64_t, min_tx_time);
824 mutex_enter(&dp->dp_lock);
825 wakeup = MAX(tx->tx_start + min_tx_time,
826 dp->dp_last_wakeup + min_tx_time);
827 dp->dp_last_wakeup = wakeup;
828 mutex_exit(&dp->dp_lock);
830 zfs_sleep_until(wakeup);
834 * This routine attempts to assign the transaction to a transaction group.
835 * To do so, we must determine if there is sufficient free space on disk.
837 * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
838 * on it), then it is assumed that there is sufficient free space,
839 * unless there's insufficient slop space in the pool (see the comment
840 * above spa_slop_shift in spa_misc.c).
842 * If it is not a "netfree" transaction, then if the data already on disk
843 * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
844 * ENOSPC. Otherwise, if the current rough estimate of pending changes,
845 * plus the rough estimate of this transaction's changes, may exceed the
846 * allowed usage, then this will fail with ERESTART, which will cause the
847 * caller to wait for the pending changes to be written to disk (by waiting
848 * for the next TXG to open), and then check the space usage again.
850 * The rough estimate of pending changes is comprised of the sum of:
852 * - this transaction's holds' txh_space_towrite
854 * - dd_tempreserved[], which is the sum of in-flight transactions'
855 * holds' txh_space_towrite (i.e. those transactions that have called
856 * dmu_tx_assign() but not yet called dmu_tx_commit()).
858 * - dd_space_towrite[], which is the amount of dirtied dbufs.
860 * Note that all of these values are inflated by spa_get_worst_case_asize(),
861 * which means that we may get ERESTART well before we are actually in danger
862 * of running out of space, but this also mitigates any small inaccuracies
863 * in the rough estimate (e.g. txh_space_towrite doesn't take into account
864 * indirect blocks, and dd_space_towrite[] doesn't take into account changes
867 * Note that due to this algorithm, it is possible to exceed the allowed
868 * usage by one transaction. Also, as we approach the allowed usage,
869 * we will allow a very limited amount of changes into each TXG, thus
870 * decreasing performance.
873 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
875 spa_t *spa = tx->tx_pool->dp_spa;
880 DMU_TX_STAT_BUMP(dmu_tx_error);
884 if (spa_suspended(spa)) {
885 DMU_TX_STAT_BUMP(dmu_tx_suspended);
888 * If the user has indicated a blocking failure mode
889 * then return ERESTART which will block in dmu_tx_wait().
890 * Otherwise, return EIO so that an error can get
891 * propagated back to the VOP calls.
893 * Note that we always honor the txg_how flag regardless
894 * of the failuremode setting.
896 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
897 !(txg_how & TXG_WAIT))
898 return (SET_ERROR(EIO));
900 return (SET_ERROR(ERESTART));
903 if (!tx->tx_dirty_delayed &&
904 dsl_pool_need_dirty_delay(tx->tx_pool)) {
905 tx->tx_wait_dirty = B_TRUE;
906 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
907 return (SET_ERROR(ERESTART));
910 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
911 tx->tx_needassign_txh = NULL;
914 * NB: No error returns are allowed after txg_hold_open, but
915 * before processing the dnode holds, due to the
916 * dmu_tx_unassign() logic.
919 uint64_t towrite = 0;
921 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
922 txh = list_next(&tx->tx_holds, txh)) {
923 dnode_t *dn = txh->txh_dnode;
925 mutex_enter(&dn->dn_mtx);
926 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
927 mutex_exit(&dn->dn_mtx);
928 tx->tx_needassign_txh = txh;
929 DMU_TX_STAT_BUMP(dmu_tx_group);
930 return (SET_ERROR(ERESTART));
932 if (dn->dn_assigned_txg == 0)
933 dn->dn_assigned_txg = tx->tx_txg;
934 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
935 (void) refcount_add(&dn->dn_tx_holds, tx);
936 mutex_exit(&dn->dn_mtx);
938 towrite += refcount_count(&txh->txh_space_towrite);
939 tohold += refcount_count(&txh->txh_memory_tohold);
942 /* needed allocation: worst-case estimate of write space */
943 uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite);
944 /* calculate memory footprint estimate */
945 uint64_t memory = towrite + tohold;
947 if (tx->tx_dir != NULL && asize != 0) {
948 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
949 asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx);
954 DMU_TX_STAT_BUMP(dmu_tx_assigned);
960 dmu_tx_unassign(dmu_tx_t *tx)
965 txg_rele_to_quiesce(&tx->tx_txgh);
968 * Walk the transaction's hold list, removing the hold on the
969 * associated dnode, and notifying waiters if the refcount drops to 0.
971 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
972 txh && txh != tx->tx_needassign_txh;
973 txh = list_next(&tx->tx_holds, txh)) {
974 dnode_t *dn = txh->txh_dnode;
978 mutex_enter(&dn->dn_mtx);
979 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
981 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
982 dn->dn_assigned_txg = 0;
983 cv_broadcast(&dn->dn_notxholds);
985 mutex_exit(&dn->dn_mtx);
988 txg_rele_to_sync(&tx->tx_txgh);
990 tx->tx_lasttried_txg = tx->tx_txg;
995 * Assign tx to a transaction group; txg_how is a bitmask:
997 * If TXG_WAIT is set and the currently open txg is full, this function
998 * will wait until there's a new txg. This should be used when no locks
999 * are being held. With this bit set, this function will only fail if
1000 * we're truly out of space (or over quota).
1002 * If TXG_WAIT is *not* set and we can't assign into the currently open
1003 * txg without blocking, this function will return immediately with
1004 * ERESTART. This should be used whenever locks are being held. On an
1005 * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
1008 * If TXG_NOTHROTTLE is set, this indicates that this tx should not be
1009 * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
1010 * details on the throttle). This is used by the VFS operations, after
1011 * they have already called dmu_tx_wait() (though most likely on a
1015 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1019 ASSERT(tx->tx_txg == 0);
1020 ASSERT0(txg_how & ~(TXG_WAIT | TXG_NOTHROTTLE));
1021 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1023 /* If we might wait, we must not hold the config lock. */
1024 IMPLY((txg_how & TXG_WAIT), !dsl_pool_config_held(tx->tx_pool));
1026 if ((txg_how & TXG_NOTHROTTLE))
1027 tx->tx_dirty_delayed = B_TRUE;
1029 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1030 dmu_tx_unassign(tx);
1032 if (err != ERESTART || !(txg_how & TXG_WAIT))
1038 txg_rele_to_quiesce(&tx->tx_txgh);
1044 dmu_tx_wait(dmu_tx_t *tx)
1046 spa_t *spa = tx->tx_pool->dp_spa;
1047 dsl_pool_t *dp = tx->tx_pool;
1050 ASSERT(tx->tx_txg == 0);
1051 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1053 before = gethrtime();
1055 if (tx->tx_wait_dirty) {
1059 * dmu_tx_try_assign() has determined that we need to wait
1060 * because we've consumed much or all of the dirty buffer
1063 mutex_enter(&dp->dp_lock);
1064 if (dp->dp_dirty_total >= zfs_dirty_data_max)
1065 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
1066 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1067 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1068 dirty = dp->dp_dirty_total;
1069 mutex_exit(&dp->dp_lock);
1071 dmu_tx_delay(tx, dirty);
1073 tx->tx_wait_dirty = B_FALSE;
1076 * Note: setting tx_dirty_delayed only has effect if the
1077 * caller used TX_WAIT. Otherwise they are going to
1078 * destroy this tx and try again. The common case,
1079 * zfs_write(), uses TX_WAIT.
1081 tx->tx_dirty_delayed = B_TRUE;
1082 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1084 * If the pool is suspended we need to wait until it
1085 * is resumed. Note that it's possible that the pool
1086 * has become active after this thread has tried to
1087 * obtain a tx. If that's the case then tx_lasttried_txg
1088 * would not have been set.
1090 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1091 } else if (tx->tx_needassign_txh) {
1092 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1094 mutex_enter(&dn->dn_mtx);
1095 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1096 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1097 mutex_exit(&dn->dn_mtx);
1098 tx->tx_needassign_txh = NULL;
1101 * A dnode is assigned to the quiescing txg. Wait for its
1102 * transaction to complete.
1104 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1107 spa_tx_assign_add_nsecs(spa, gethrtime() - before);
1111 dmu_tx_destroy(dmu_tx_t *tx)
1115 while ((txh = list_head(&tx->tx_holds)) != NULL) {
1116 dnode_t *dn = txh->txh_dnode;
1118 list_remove(&tx->tx_holds, txh);
1119 refcount_destroy_many(&txh->txh_space_towrite,
1120 refcount_count(&txh->txh_space_towrite));
1121 refcount_destroy_many(&txh->txh_memory_tohold,
1122 refcount_count(&txh->txh_memory_tohold));
1123 kmem_free(txh, sizeof (dmu_tx_hold_t));
1128 list_destroy(&tx->tx_callbacks);
1129 list_destroy(&tx->tx_holds);
1130 kmem_free(tx, sizeof (dmu_tx_t));
1134 dmu_tx_commit(dmu_tx_t *tx)
1136 ASSERT(tx->tx_txg != 0);
1139 * Go through the transaction's hold list and remove holds on
1140 * associated dnodes, notifying waiters if no holds remain.
1142 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1143 txh = list_next(&tx->tx_holds, txh)) {
1144 dnode_t *dn = txh->txh_dnode;
1149 mutex_enter(&dn->dn_mtx);
1150 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1152 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1153 dn->dn_assigned_txg = 0;
1154 cv_broadcast(&dn->dn_notxholds);
1156 mutex_exit(&dn->dn_mtx);
1159 if (tx->tx_tempreserve_cookie)
1160 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1162 if (!list_is_empty(&tx->tx_callbacks))
1163 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1165 if (tx->tx_anyobj == FALSE)
1166 txg_rele_to_sync(&tx->tx_txgh);
1172 dmu_tx_abort(dmu_tx_t *tx)
1174 ASSERT(tx->tx_txg == 0);
1177 * Call any registered callbacks with an error code.
1179 if (!list_is_empty(&tx->tx_callbacks))
1180 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1186 dmu_tx_get_txg(dmu_tx_t *tx)
1188 ASSERT(tx->tx_txg != 0);
1189 return (tx->tx_txg);
1193 dmu_tx_pool(dmu_tx_t *tx)
1195 ASSERT(tx->tx_pool != NULL);
1196 return (tx->tx_pool);
1200 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1202 dmu_tx_callback_t *dcb;
1204 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1206 dcb->dcb_func = func;
1207 dcb->dcb_data = data;
1209 list_insert_tail(&tx->tx_callbacks, dcb);
1213 * Call all the commit callbacks on a list, with a given error code.
1216 dmu_tx_do_callbacks(list_t *cb_list, int error)
1218 dmu_tx_callback_t *dcb;
1220 while ((dcb = list_tail(cb_list)) != NULL) {
1221 list_remove(cb_list, dcb);
1222 dcb->dcb_func(dcb->dcb_data, error);
1223 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1228 * Interface to hold a bunch of attributes.
1229 * used for creating new files.
1230 * attrsize is the total size of all attributes
1231 * to be added during object creation
1233 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1237 * hold necessary attribute name for attribute registration.
1238 * should be a very rare case where this is needed. If it does
1239 * happen it would only happen on the first write to the file system.
1242 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1244 if (!sa->sa_need_attr_registration)
1247 for (int i = 0; i != sa->sa_num_attrs; i++) {
1248 if (!sa->sa_attr_table[i].sa_registered) {
1249 if (sa->sa_reg_attr_obj)
1250 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1251 B_TRUE, sa->sa_attr_table[i].sa_name);
1253 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1254 B_TRUE, sa->sa_attr_table[i].sa_name);
1260 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1264 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1267 (void) refcount_add_many(&txh->txh_space_towrite,
1268 SPA_OLD_MAXBLOCKSIZE, FTAG);
1272 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1274 sa_os_t *sa = tx->tx_objset->os_sa;
1276 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1278 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1281 if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1282 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1284 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1285 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1286 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1287 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1290 dmu_tx_sa_registration_hold(sa, tx);
1292 if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill)
1295 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1302 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1304 * variable_size is the total size of all variable sized attributes
1305 * passed to this function. It is not the total size of all
1306 * variable size attributes that *may* exist on this object.
1309 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1312 sa_os_t *sa = tx->tx_objset->os_sa;
1314 ASSERT(hdl != NULL);
1316 object = sa_handle_object(hdl);
1318 dmu_tx_hold_bonus(tx, object);
1320 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1323 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1324 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1325 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1326 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1327 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1328 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1331 dmu_tx_sa_registration_hold(sa, tx);
1333 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1334 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1336 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1337 ASSERT(tx->tx_txg == 0);
1338 dmu_tx_hold_spill(tx, object);
1340 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1345 if (dn->dn_have_spill) {
1346 ASSERT(tx->tx_txg == 0);
1347 dmu_tx_hold_spill(tx, object);
1356 dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
1357 KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
1358 KSTAT_FLAG_VIRTUAL);
1360 if (dmu_tx_ksp != NULL) {
1361 dmu_tx_ksp->ks_data = &dmu_tx_stats;
1362 kstat_install(dmu_tx_ksp);
1369 if (dmu_tx_ksp != NULL) {
1370 kstat_delete(dmu_tx_ksp);
1375 #if defined(_KERNEL) && defined(HAVE_SPL)
1376 EXPORT_SYMBOL(dmu_tx_create);
1377 EXPORT_SYMBOL(dmu_tx_hold_write);
1378 EXPORT_SYMBOL(dmu_tx_hold_write_by_dnode);
1379 EXPORT_SYMBOL(dmu_tx_hold_free);
1380 EXPORT_SYMBOL(dmu_tx_hold_free_by_dnode);
1381 EXPORT_SYMBOL(dmu_tx_hold_zap);
1382 EXPORT_SYMBOL(dmu_tx_hold_zap_by_dnode);
1383 EXPORT_SYMBOL(dmu_tx_hold_bonus);
1384 EXPORT_SYMBOL(dmu_tx_hold_bonus_by_dnode);
1385 EXPORT_SYMBOL(dmu_tx_abort);
1386 EXPORT_SYMBOL(dmu_tx_assign);
1387 EXPORT_SYMBOL(dmu_tx_wait);
1388 EXPORT_SYMBOL(dmu_tx_commit);
1389 EXPORT_SYMBOL(dmu_tx_mark_netfree);
1390 EXPORT_SYMBOL(dmu_tx_get_txg);
1391 EXPORT_SYMBOL(dmu_tx_callback_register);
1392 EXPORT_SYMBOL(dmu_tx_do_callbacks);
1393 EXPORT_SYMBOL(dmu_tx_hold_spill);
1394 EXPORT_SYMBOL(dmu_tx_hold_sa_create);
1395 EXPORT_SYMBOL(dmu_tx_hold_sa);