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.
25 * Copyright (c) 2014 Integros [integros.com]
29 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dataset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_pool.h>
36 #include <sys/zap_impl.h>
39 #include <sys/sa_impl.h>
40 #include <sys/zfs_context.h>
41 #include <sys/varargs.h>
43 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
44 uint64_t arg1, uint64_t arg2);
48 dmu_tx_create_dd(dsl_dir_t *dd)
50 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
53 tx->tx_pool = dd->dd_pool;
54 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
55 offsetof(dmu_tx_hold_t, txh_node));
56 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
57 offsetof(dmu_tx_callback_t, dcb_node));
58 tx->tx_start = gethrtime();
63 dmu_tx_create(objset_t *os)
65 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
71 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
73 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
75 txg_verify(dp->dp_spa, txg);
84 dmu_tx_is_syncing(dmu_tx_t *tx)
86 return (tx->tx_anyobj);
90 dmu_tx_private_ok(dmu_tx_t *tx)
92 return (tx->tx_anyobj);
95 static dmu_tx_hold_t *
96 dmu_tx_hold_dnode_impl(dmu_tx_t *tx, dnode_t *dn, enum dmu_tx_hold_type type,
97 uint64_t arg1, uint64_t arg2)
102 (void) refcount_add(&dn->dn_holds, tx);
103 if (tx->tx_txg != 0) {
104 mutex_enter(&dn->dn_mtx);
106 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
107 * problem, but there's no way for it to happen (for
110 ASSERT(dn->dn_assigned_txg == 0);
111 dn->dn_assigned_txg = tx->tx_txg;
112 (void) refcount_add(&dn->dn_tx_holds, tx);
113 mutex_exit(&dn->dn_mtx);
117 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
120 refcount_create(&txh->txh_space_towrite);
121 refcount_create(&txh->txh_memory_tohold);
122 txh->txh_type = type;
123 txh->txh_arg1 = arg1;
124 txh->txh_arg2 = arg2;
125 list_insert_tail(&tx->tx_holds, txh);
130 static dmu_tx_hold_t *
131 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
132 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
138 if (object != DMU_NEW_OBJECT) {
139 err = dnode_hold(os, object, FTAG, &dn);
145 txh = dmu_tx_hold_dnode_impl(tx, dn, type, arg1, arg2);
147 dnode_rele(dn, FTAG);
152 dmu_tx_add_new_object(dmu_tx_t *tx, dnode_t *dn)
155 * If we're syncing, they can manipulate any object anyhow, and
156 * the hold on the dnode_t can cause problems.
158 if (!dmu_tx_is_syncing(tx))
159 (void) dmu_tx_hold_dnode_impl(tx, dn, THT_NEWOBJECT, 0, 0);
163 * This function reads specified data from disk. The specified data will
164 * be needed to perform the transaction -- i.e, it will be read after
165 * we do dmu_tx_assign(). There are two reasons that we read the data now
166 * (before dmu_tx_assign()):
168 * 1. Reading it now has potentially better performance. The transaction
169 * has not yet been assigned, so the TXG is not held open, and also the
170 * caller typically has less locks held when calling dmu_tx_hold_*() than
171 * after the transaction has been assigned. This reduces the lock (and txg)
172 * hold times, thus reducing lock contention.
174 * 2. It is easier for callers (primarily the ZPL) to handle i/o errors
175 * that are detected before they start making changes to the DMU state
176 * (i.e. now). Once the transaction has been assigned, and some DMU
177 * state has been changed, it can be difficult to recover from an i/o
178 * error (e.g. to undo the changes already made in memory at the DMU
179 * layer). Typically code to do so does not exist in the caller -- it
180 * assumes that the data has already been cached and thus i/o errors are
183 * It has been observed that the i/o initiated here can be a performance
184 * problem, and it appears to be optional, because we don't look at the
185 * data which is read. However, removing this read would only serve to
186 * move the work elsewhere (after the dmu_tx_assign()), where it may
187 * have a greater impact on performance (in addition to the impact on
188 * fault tolerance noted above).
191 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
196 rw_enter(&dn->dn_struct_rwlock, RW_READER);
197 db = dbuf_hold_level(dn, level, blkid, FTAG);
198 rw_exit(&dn->dn_struct_rwlock);
200 return (SET_ERROR(EIO));
201 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
208 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
210 dnode_t *dn = txh->txh_dnode;
216 (void) refcount_add_many(&txh->txh_space_towrite, len, FTAG);
218 if (refcount_count(&txh->txh_space_towrite) > 2 * DMU_MAX_ACCESS)
219 err = SET_ERROR(EFBIG);
225 * For i/o error checking, read the blocks that will be needed
226 * to perform the write: the first and last level-0 blocks (if
227 * they are not aligned, i.e. if they are partial-block writes),
228 * and all the level-1 blocks.
230 if (dn->dn_maxblkid == 0) {
231 if (off < dn->dn_datablksz &&
232 (off > 0 || len < dn->dn_datablksz)) {
233 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
235 txh->txh_tx->tx_err = err;
239 zio_t *zio = zio_root(dn->dn_objset->os_spa,
240 NULL, NULL, ZIO_FLAG_CANFAIL);
242 /* first level-0 block */
243 uint64_t start = off >> dn->dn_datablkshift;
244 if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
245 err = dmu_tx_check_ioerr(zio, dn, 0, start);
247 txh->txh_tx->tx_err = err;
251 /* last level-0 block */
252 uint64_t end = (off + len - 1) >> dn->dn_datablkshift;
253 if (end != start && end <= dn->dn_maxblkid &&
254 P2PHASE(off + len, dn->dn_datablksz)) {
255 err = dmu_tx_check_ioerr(zio, dn, 0, end);
257 txh->txh_tx->tx_err = err;
262 if (dn->dn_nlevels > 1) {
263 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
264 for (uint64_t i = (start >> shft) + 1;
265 i < end >> shft; i++) {
266 err = dmu_tx_check_ioerr(zio, dn, 1, i);
268 txh->txh_tx->tx_err = err;
275 txh->txh_tx->tx_err = err;
281 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
283 (void) refcount_add_many(&txh->txh_space_towrite, DNODE_MIN_SIZE, FTAG);
287 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
292 ASSERT3U(len, <=, DMU_MAX_ACCESS);
293 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
295 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
296 object, THT_WRITE, off, len);
298 dmu_tx_count_write(txh, off, len);
299 dmu_tx_count_dnode(txh);
304 dmu_tx_hold_remap_l1indirect(dmu_tx_t *tx, uint64_t object)
308 ASSERT(tx->tx_txg == 0);
309 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
310 object, THT_WRITE, 0, 0);
314 dnode_t *dn = txh->txh_dnode;
315 (void) refcount_add_many(&txh->txh_space_towrite,
316 1ULL << dn->dn_indblkshift, FTAG);
317 dmu_tx_count_dnode(txh);
321 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
326 ASSERT3U(len, <=, DMU_MAX_ACCESS);
327 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
329 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
331 dmu_tx_count_write(txh, off, len);
332 dmu_tx_count_dnode(txh);
337 * This function marks the transaction as being a "net free". The end
338 * result is that refquotas will be disabled for this transaction, and
339 * this transaction will be able to use half of the pool space overhead
340 * (see dsl_pool_adjustedsize()). Therefore this function should only
341 * be called for transactions that we expect will not cause a net increase
342 * in the amount of space used (but it's OK if that is occasionally not true).
345 dmu_tx_mark_netfree(dmu_tx_t *tx)
347 tx->tx_netfree = B_TRUE;
351 dmu_tx_hold_free_impl(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
358 ASSERT(tx->tx_txg == 0);
361 dmu_tx_count_dnode(txh);
363 if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
365 if (len == DMU_OBJECT_END)
366 len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
370 * For i/o error checking, we read the first and last level-0
371 * blocks if they are not aligned, and all the level-1 blocks.
373 * Note: dbuf_free_range() assumes that we have not instantiated
374 * any level-0 dbufs that will be completely freed. Therefore we must
375 * exercise care to not read or count the first and last blocks
376 * if they are blocksize-aligned.
378 if (dn->dn_datablkshift == 0) {
379 if (off != 0 || len < dn->dn_datablksz)
380 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
382 /* first block will be modified if it is not aligned */
383 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
384 dmu_tx_count_write(txh, off, 1);
385 /* last block will be modified if it is not aligned */
386 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
387 dmu_tx_count_write(txh, off + len, 1);
391 * Check level-1 blocks.
393 if (dn->dn_nlevels > 1) {
394 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
396 uint64_t start = off >> shift;
397 uint64_t end = (off + len) >> shift;
399 ASSERT(dn->dn_indblkshift != 0);
402 * dnode_reallocate() can result in an object with indirect
403 * blocks having an odd data block size. In this case,
404 * just check the single block.
406 if (dn->dn_datablkshift == 0)
409 zio_t *zio = zio_root(tx->tx_pool->dp_spa,
410 NULL, NULL, ZIO_FLAG_CANFAIL);
411 for (uint64_t i = start; i <= end; i++) {
412 uint64_t ibyte = i << shift;
413 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
415 if (err == ESRCH || i > end)
419 (void) zio_wait(zio);
423 (void) refcount_add_many(&txh->txh_memory_tohold,
424 1 << dn->dn_indblkshift, FTAG);
426 err = dmu_tx_check_ioerr(zio, dn, 1, i);
429 (void) zio_wait(zio);
442 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
446 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
447 object, THT_FREE, off, len);
449 (void) dmu_tx_hold_free_impl(txh, off, len);
453 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
457 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
459 (void) dmu_tx_hold_free_impl(txh, off, len);
463 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
465 dmu_tx_t *tx = txh->txh_tx;
469 ASSERT(tx->tx_txg == 0);
473 dmu_tx_count_dnode(txh);
476 * Modifying a almost-full microzap is around the worst case (128KB)
478 * If it is a fat zap, the worst case would be 7*16KB=112KB:
479 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
480 * - 4 new blocks written if adding:
481 * - 2 blocks for possibly split leaves,
482 * - 2 grown ptrtbl blocks
484 (void) refcount_add_many(&txh->txh_space_towrite,
485 MZAP_MAX_BLKSZ, FTAG);
490 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
492 if (dn->dn_maxblkid == 0 || name == NULL) {
494 * This is a microzap (only one block), or we don't know
495 * the name. Check the first block for i/o errors.
497 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
503 * Access the name so that we'll check for i/o errors to
504 * the leaf blocks, etc. We ignore ENOENT, as this name
507 err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
508 if (err == EIO || err == ECKSUM || err == ENXIO) {
515 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
521 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
522 object, THT_ZAP, add, (uintptr_t)name);
524 dmu_tx_hold_zap_impl(txh, name);
528 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
535 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
537 dmu_tx_hold_zap_impl(txh, name);
541 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
545 ASSERT(tx->tx_txg == 0);
547 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
548 object, THT_BONUS, 0, 0);
550 dmu_tx_count_dnode(txh);
554 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
560 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
562 dmu_tx_count_dnode(txh);
566 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
569 ASSERT(tx->tx_txg == 0);
571 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
572 DMU_NEW_OBJECT, THT_SPACE, space, 0);
574 (void) refcount_add_many(&txh->txh_space_towrite, space, FTAG);
579 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
581 boolean_t match_object = B_FALSE;
582 boolean_t match_offset = B_FALSE;
585 dnode_t *dn = DB_DNODE(db);
586 ASSERT(tx->tx_txg != 0);
587 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
588 ASSERT3U(dn->dn_object, ==, db->db.db_object);
595 /* XXX No checking on the meta dnode for now */
596 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
601 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
602 txh = list_next(&tx->tx_holds, txh)) {
603 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
604 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
606 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
607 int datablkshift = dn->dn_datablkshift ?
608 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
609 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
610 int shift = datablkshift + epbs * db->db_level;
611 uint64_t beginblk = shift >= 64 ? 0 :
612 (txh->txh_arg1 >> shift);
613 uint64_t endblk = shift >= 64 ? 0 :
614 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
615 uint64_t blkid = db->db_blkid;
617 /* XXX txh_arg2 better not be zero... */
619 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
620 txh->txh_type, beginblk, endblk);
622 switch (txh->txh_type) {
624 if (blkid >= beginblk && blkid <= endblk)
627 * We will let this hold work for the bonus
628 * or spill buffer so that we don't need to
629 * hold it when creating a new object.
631 if (blkid == DMU_BONUS_BLKID ||
632 blkid == DMU_SPILL_BLKID)
635 * They might have to increase nlevels,
636 * thus dirtying the new TLIBs. Or the
637 * might have to change the block size,
638 * thus dirying the new lvl=0 blk=0.
645 * We will dirty all the level 1 blocks in
646 * the free range and perhaps the first and
647 * last level 0 block.
649 if (blkid >= beginblk && (blkid <= endblk ||
650 txh->txh_arg2 == DMU_OBJECT_END))
654 if (blkid == DMU_SPILL_BLKID)
658 if (blkid == DMU_BONUS_BLKID)
668 ASSERT(!"bad txh_type");
671 if (match_object && match_offset) {
677 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
678 (u_longlong_t)db->db.db_object, db->db_level,
679 (u_longlong_t)db->db_blkid);
684 * If we can't do 10 iops, something is wrong. Let us go ahead
685 * and hit zfs_dirty_data_max.
687 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
688 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
691 * We delay transactions when we've determined that the backend storage
692 * isn't able to accommodate the rate of incoming writes.
694 * If there is already a transaction waiting, we delay relative to when
695 * that transaction finishes waiting. This way the calculated min_time
696 * is independent of the number of threads concurrently executing
699 * If we are the only waiter, wait relative to when the transaction
700 * started, rather than the current time. This credits the transaction for
701 * "time already served", e.g. reading indirect blocks.
703 * The minimum time for a transaction to take is calculated as:
704 * min_time = scale * (dirty - min) / (max - dirty)
705 * min_time is then capped at zfs_delay_max_ns.
707 * The delay has two degrees of freedom that can be adjusted via tunables.
708 * The percentage of dirty data at which we start to delay is defined by
709 * zfs_delay_min_dirty_percent. This should typically be at or above
710 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
711 * delay after writing at full speed has failed to keep up with the incoming
712 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
713 * speaking, this variable determines the amount of delay at the midpoint of
717 * 10ms +-------------------------------------------------------------*+
733 * 2ms + (midpoint) * +
736 * | zfs_delay_scale ----------> ******** |
737 * 0 +-------------------------------------*********----------------+
738 * 0% <- zfs_dirty_data_max -> 100%
740 * Note that since the delay is added to the outstanding time remaining on the
741 * most recent transaction, the delay is effectively the inverse of IOPS.
742 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
743 * was chosen such that small changes in the amount of accumulated dirty data
744 * in the first 3/4 of the curve yield relatively small differences in the
747 * The effects can be easier to understand when the amount of delay is
748 * represented on a log scale:
751 * 100ms +-------------------------------------------------------------++
760 * + zfs_delay_scale ----------> ***** +
771 * +--------------------------------------------------------------+
772 * 0% <- zfs_dirty_data_max -> 100%
774 * Note here that only as the amount of dirty data approaches its limit does
775 * the delay start to increase rapidly. The goal of a properly tuned system
776 * should be to keep the amount of dirty data out of that range by first
777 * ensuring that the appropriate limits are set for the I/O scheduler to reach
778 * optimal throughput on the backend storage, and then by changing the value
779 * of zfs_delay_scale to increase the steepness of the curve.
782 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
784 dsl_pool_t *dp = tx->tx_pool;
785 uint64_t delay_min_bytes =
786 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
787 hrtime_t wakeup, min_tx_time, now;
789 if (dirty <= delay_min_bytes)
793 * The caller has already waited until we are under the max.
794 * We make them pass us the amount of dirty data so we don't
795 * have to handle the case of it being >= the max, which could
796 * cause a divide-by-zero if it's == the max.
798 ASSERT3U(dirty, <, zfs_dirty_data_max);
801 min_tx_time = zfs_delay_scale *
802 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
803 if (now > tx->tx_start + min_tx_time)
806 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
808 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
809 uint64_t, min_tx_time);
811 mutex_enter(&dp->dp_lock);
812 wakeup = MAX(tx->tx_start + min_tx_time,
813 dp->dp_last_wakeup + min_tx_time);
814 dp->dp_last_wakeup = wakeup;
815 mutex_exit(&dp->dp_lock);
819 mutex_enter(&curthread->t_delay_lock);
820 while (cv_timedwait_hires(&curthread->t_delay_cv,
821 &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
822 CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
824 mutex_exit(&curthread->t_delay_lock);
826 pause_sbt("dmu_tx_delay", nstosbt(wakeup),
827 nstosbt(zfs_delay_resolution_ns), C_ABSOLUTE);
830 hrtime_t delta = wakeup - gethrtime();
832 ts.tv_sec = delta / NANOSEC;
833 ts.tv_nsec = delta % NANOSEC;
834 (void) nanosleep(&ts, NULL);
839 * This routine attempts to assign the transaction to a transaction group.
840 * To do so, we must determine if there is sufficient free space on disk.
842 * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
843 * on it), then it is assumed that there is sufficient free space,
844 * unless there's insufficient slop space in the pool (see the comment
845 * above spa_slop_shift in spa_misc.c).
847 * If it is not a "netfree" transaction, then if the data already on disk
848 * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
849 * ENOSPC. Otherwise, if the current rough estimate of pending changes,
850 * plus the rough estimate of this transaction's changes, may exceed the
851 * allowed usage, then this will fail with ERESTART, which will cause the
852 * caller to wait for the pending changes to be written to disk (by waiting
853 * for the next TXG to open), and then check the space usage again.
855 * The rough estimate of pending changes is comprised of the sum of:
857 * - this transaction's holds' txh_space_towrite
859 * - dd_tempreserved[], which is the sum of in-flight transactions'
860 * holds' txh_space_towrite (i.e. those transactions that have called
861 * dmu_tx_assign() but not yet called dmu_tx_commit()).
863 * - dd_space_towrite[], which is the amount of dirtied dbufs.
865 * Note that all of these values are inflated by spa_get_worst_case_asize(),
866 * which means that we may get ERESTART well before we are actually in danger
867 * of running out of space, but this also mitigates any small inaccuracies
868 * in the rough estimate (e.g. txh_space_towrite doesn't take into account
869 * indirect blocks, and dd_space_towrite[] doesn't take into account changes
872 * Note that due to this algorithm, it is possible to exceed the allowed
873 * usage by one transaction. Also, as we approach the allowed usage,
874 * we will allow a very limited amount of changes into each TXG, thus
875 * decreasing performance.
878 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
880 spa_t *spa = tx->tx_pool->dp_spa;
887 if (spa_suspended(spa)) {
889 * If the user has indicated a blocking failure mode
890 * then return ERESTART which will block in dmu_tx_wait().
891 * Otherwise, return EIO so that an error can get
892 * propagated back to the VOP calls.
894 * Note that we always honor the txg_how flag regardless
895 * of the failuremode setting.
897 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
898 !(txg_how & TXG_WAIT))
899 return (SET_ERROR(EIO));
901 return (SET_ERROR(ERESTART));
904 if (!tx->tx_dirty_delayed &&
905 dsl_pool_need_dirty_delay(tx->tx_pool)) {
906 tx->tx_wait_dirty = B_TRUE;
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 return (SET_ERROR(ERESTART));
931 if (dn->dn_assigned_txg == 0)
932 dn->dn_assigned_txg = tx->tx_txg;
933 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
934 (void) refcount_add(&dn->dn_tx_holds, tx);
935 mutex_exit(&dn->dn_mtx);
937 towrite += refcount_count(&txh->txh_space_towrite);
938 tohold += refcount_count(&txh->txh_memory_tohold);
941 /* needed allocation: worst-case estimate of write space */
942 uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite);
943 /* calculate memory footprint estimate */
944 uint64_t memory = towrite + tohold;
946 if (tx->tx_dir != NULL && asize != 0) {
947 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
948 asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx);
957 dmu_tx_unassign(dmu_tx_t *tx)
962 txg_rele_to_quiesce(&tx->tx_txgh);
965 * Walk the transaction's hold list, removing the hold on the
966 * associated dnode, and notifying waiters if the refcount drops to 0.
968 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
969 txh != tx->tx_needassign_txh;
970 txh = list_next(&tx->tx_holds, txh)) {
971 dnode_t *dn = txh->txh_dnode;
975 mutex_enter(&dn->dn_mtx);
976 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
978 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
979 dn->dn_assigned_txg = 0;
980 cv_broadcast(&dn->dn_notxholds);
982 mutex_exit(&dn->dn_mtx);
985 txg_rele_to_sync(&tx->tx_txgh);
987 tx->tx_lasttried_txg = tx->tx_txg;
992 * Assign tx to a transaction group; txg_how is a bitmask:
994 * If TXG_WAIT is set and the currently open txg is full, this function
995 * will wait until there's a new txg. This should be used when no locks
996 * are being held. With this bit set, this function will only fail if
997 * we're truly out of space (or over quota).
999 * If TXG_WAIT is *not* set and we can't assign into the currently open
1000 * txg without blocking, this function will return immediately with
1001 * ERESTART. This should be used whenever locks are being held. On an
1002 * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
1005 * If TXG_NOTHROTTLE is set, this indicates that this tx should not be
1006 * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
1007 * details on the throttle). This is used by the VFS operations, after
1008 * they have already called dmu_tx_wait() (though most likely on a
1012 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1016 ASSERT(tx->tx_txg == 0);
1017 ASSERT0(txg_how & ~(TXG_WAIT | TXG_NOTHROTTLE));
1018 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1020 /* If we might wait, we must not hold the config lock. */
1021 IMPLY((txg_how & TXG_WAIT), !dsl_pool_config_held(tx->tx_pool));
1023 if ((txg_how & TXG_NOTHROTTLE))
1024 tx->tx_dirty_delayed = B_TRUE;
1026 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1027 dmu_tx_unassign(tx);
1029 if (err != ERESTART || !(txg_how & TXG_WAIT))
1035 txg_rele_to_quiesce(&tx->tx_txgh);
1041 dmu_tx_wait(dmu_tx_t *tx)
1043 spa_t *spa = tx->tx_pool->dp_spa;
1044 dsl_pool_t *dp = tx->tx_pool;
1046 ASSERT(tx->tx_txg == 0);
1047 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1049 if (tx->tx_wait_dirty) {
1051 * dmu_tx_try_assign() has determined that we need to wait
1052 * because we've consumed much or all of the dirty buffer
1055 mutex_enter(&dp->dp_lock);
1056 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1057 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1058 uint64_t dirty = dp->dp_dirty_total;
1059 mutex_exit(&dp->dp_lock);
1061 dmu_tx_delay(tx, dirty);
1063 tx->tx_wait_dirty = B_FALSE;
1066 * Note: setting tx_dirty_delayed only has effect if the
1067 * caller used TX_WAIT. Otherwise they are going to
1068 * destroy this tx and try again. The common case,
1069 * zfs_write(), uses TX_WAIT.
1071 tx->tx_dirty_delayed = B_TRUE;
1072 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1074 * If the pool is suspended we need to wait until it
1075 * is resumed. Note that it's possible that the pool
1076 * has become active after this thread has tried to
1077 * obtain a tx. If that's the case then tx_lasttried_txg
1078 * would not have been set.
1080 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1081 } else if (tx->tx_needassign_txh) {
1083 * A dnode is assigned to the quiescing txg. Wait for its
1084 * transaction to complete.
1086 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1088 mutex_enter(&dn->dn_mtx);
1089 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1090 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1091 mutex_exit(&dn->dn_mtx);
1092 tx->tx_needassign_txh = NULL;
1095 * If we have a lot of dirty data just wait until we sync
1096 * out a TXG at which point we'll hopefully have synced
1097 * a portion of the changes.
1099 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1104 dmu_tx_destroy(dmu_tx_t *tx)
1108 while ((txh = list_head(&tx->tx_holds)) != NULL) {
1109 dnode_t *dn = txh->txh_dnode;
1111 list_remove(&tx->tx_holds, txh);
1112 refcount_destroy_many(&txh->txh_space_towrite,
1113 refcount_count(&txh->txh_space_towrite));
1114 refcount_destroy_many(&txh->txh_memory_tohold,
1115 refcount_count(&txh->txh_memory_tohold));
1116 kmem_free(txh, sizeof (dmu_tx_hold_t));
1121 list_destroy(&tx->tx_callbacks);
1122 list_destroy(&tx->tx_holds);
1123 kmem_free(tx, sizeof (dmu_tx_t));
1127 dmu_tx_commit(dmu_tx_t *tx)
1129 ASSERT(tx->tx_txg != 0);
1132 * Go through the transaction's hold list and remove holds on
1133 * associated dnodes, notifying waiters if no holds remain.
1135 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1136 txh = list_next(&tx->tx_holds, txh)) {
1137 dnode_t *dn = txh->txh_dnode;
1142 mutex_enter(&dn->dn_mtx);
1143 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1145 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1146 dn->dn_assigned_txg = 0;
1147 cv_broadcast(&dn->dn_notxholds);
1149 mutex_exit(&dn->dn_mtx);
1152 if (tx->tx_tempreserve_cookie)
1153 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1155 if (!list_is_empty(&tx->tx_callbacks))
1156 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1158 if (tx->tx_anyobj == FALSE)
1159 txg_rele_to_sync(&tx->tx_txgh);
1165 dmu_tx_abort(dmu_tx_t *tx)
1167 ASSERT(tx->tx_txg == 0);
1170 * Call any registered callbacks with an error code.
1172 if (!list_is_empty(&tx->tx_callbacks))
1173 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1179 dmu_tx_get_txg(dmu_tx_t *tx)
1181 ASSERT(tx->tx_txg != 0);
1182 return (tx->tx_txg);
1186 dmu_tx_pool(dmu_tx_t *tx)
1188 ASSERT(tx->tx_pool != NULL);
1189 return (tx->tx_pool);
1193 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1195 dmu_tx_callback_t *dcb;
1197 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1199 dcb->dcb_func = func;
1200 dcb->dcb_data = data;
1202 list_insert_tail(&tx->tx_callbacks, dcb);
1206 * Call all the commit callbacks on a list, with a given error code.
1209 dmu_tx_do_callbacks(list_t *cb_list, int error)
1211 dmu_tx_callback_t *dcb;
1213 while ((dcb = list_head(cb_list)) != NULL) {
1214 list_remove(cb_list, dcb);
1215 dcb->dcb_func(dcb->dcb_data, error);
1216 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1221 * Interface to hold a bunch of attributes.
1222 * used for creating new files.
1223 * attrsize is the total size of all attributes
1224 * to be added during object creation
1226 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1230 * hold necessary attribute name for attribute registration.
1231 * should be a very rare case where this is needed. If it does
1232 * happen it would only happen on the first write to the file system.
1235 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1237 if (!sa->sa_need_attr_registration)
1240 for (int i = 0; i != sa->sa_num_attrs; i++) {
1241 if (!sa->sa_attr_table[i].sa_registered) {
1242 if (sa->sa_reg_attr_obj)
1243 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1244 B_TRUE, sa->sa_attr_table[i].sa_name);
1246 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1247 B_TRUE, sa->sa_attr_table[i].sa_name);
1253 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1255 dmu_tx_hold_t *txh = dmu_tx_hold_object_impl(tx,
1256 tx->tx_objset, object, THT_SPILL, 0, 0);
1258 (void) refcount_add_many(&txh->txh_space_towrite,
1259 SPA_OLD_MAXBLOCKSIZE, FTAG);
1263 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1265 sa_os_t *sa = tx->tx_objset->os_sa;
1267 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1269 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1272 if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1273 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1275 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1276 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1277 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1278 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1281 dmu_tx_sa_registration_hold(sa, tx);
1283 if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill)
1286 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1293 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1295 * variable_size is the total size of all variable sized attributes
1296 * passed to this function. It is not the total size of all
1297 * variable size attributes that *may* exist on this object.
1300 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1303 sa_os_t *sa = tx->tx_objset->os_sa;
1305 ASSERT(hdl != NULL);
1307 object = sa_handle_object(hdl);
1309 dmu_tx_hold_bonus(tx, object);
1311 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1314 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1315 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1316 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1317 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1318 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1319 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1322 dmu_tx_sa_registration_hold(sa, tx);
1324 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1325 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1327 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1328 ASSERT(tx->tx_txg == 0);
1329 dmu_tx_hold_spill(tx, object);
1331 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1336 if (dn->dn_have_spill) {
1337 ASSERT(tx->tx_txg == 0);
1338 dmu_tx_hold_spill(tx, object);