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) zfs_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) zfs_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 zfs_refcount_create(&txh->txh_space_towrite);
121 zfs_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) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
218 if (zfs_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) zfs_refcount_add_many(&txh->txh_space_towrite, DNODE_MIN_SIZE,
288 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
293 ASSERT3U(len, <=, DMU_MAX_ACCESS);
294 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
296 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
297 object, THT_WRITE, off, len);
299 dmu_tx_count_write(txh, off, len);
300 dmu_tx_count_dnode(txh);
305 dmu_tx_hold_remap_l1indirect(dmu_tx_t *tx, uint64_t object)
309 ASSERT(tx->tx_txg == 0);
310 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
311 object, THT_WRITE, 0, 0);
315 dnode_t *dn = txh->txh_dnode;
316 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
317 1ULL << dn->dn_indblkshift, FTAG);
318 dmu_tx_count_dnode(txh);
322 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
327 ASSERT3U(len, <=, DMU_MAX_ACCESS);
328 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
330 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
332 dmu_tx_count_write(txh, off, len);
333 dmu_tx_count_dnode(txh);
338 * This function marks the transaction as being a "net free". The end
339 * result is that refquotas will be disabled for this transaction, and
340 * this transaction will be able to use half of the pool space overhead
341 * (see dsl_pool_adjustedsize()). Therefore this function should only
342 * be called for transactions that we expect will not cause a net increase
343 * in the amount of space used (but it's OK if that is occasionally not true).
346 dmu_tx_mark_netfree(dmu_tx_t *tx)
348 tx->tx_netfree = B_TRUE;
352 dmu_tx_hold_free_impl(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
359 ASSERT(tx->tx_txg == 0);
362 dmu_tx_count_dnode(txh);
364 if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
366 if (len == DMU_OBJECT_END)
367 len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
371 * For i/o error checking, we read the first and last level-0
372 * blocks if they are not aligned, and all the level-1 blocks.
374 * Note: dbuf_free_range() assumes that we have not instantiated
375 * any level-0 dbufs that will be completely freed. Therefore we must
376 * exercise care to not read or count the first and last blocks
377 * if they are blocksize-aligned.
379 if (dn->dn_datablkshift == 0) {
380 if (off != 0 || len < dn->dn_datablksz)
381 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
383 /* first block will be modified if it is not aligned */
384 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
385 dmu_tx_count_write(txh, off, 1);
386 /* last block will be modified if it is not aligned */
387 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
388 dmu_tx_count_write(txh, off + len, 1);
392 * Check level-1 blocks.
394 if (dn->dn_nlevels > 1) {
395 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
397 uint64_t start = off >> shift;
398 uint64_t end = (off + len) >> shift;
400 ASSERT(dn->dn_indblkshift != 0);
403 * dnode_reallocate() can result in an object with indirect
404 * blocks having an odd data block size. In this case,
405 * just check the single block.
407 if (dn->dn_datablkshift == 0)
410 zio_t *zio = zio_root(tx->tx_pool->dp_spa,
411 NULL, NULL, ZIO_FLAG_CANFAIL);
412 for (uint64_t i = start; i <= end; i++) {
413 uint64_t ibyte = i << shift;
414 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
416 if (err == ESRCH || i > end)
420 (void) zio_wait(zio);
424 (void) zfs_refcount_add_many(&txh->txh_memory_tohold,
425 1 << dn->dn_indblkshift, FTAG);
427 err = dmu_tx_check_ioerr(zio, dn, 1, i);
430 (void) zio_wait(zio);
443 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
447 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
448 object, THT_FREE, off, len);
450 (void) dmu_tx_hold_free_impl(txh, off, len);
454 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
458 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
460 (void) dmu_tx_hold_free_impl(txh, off, len);
464 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
466 dmu_tx_t *tx = txh->txh_tx;
470 ASSERT(tx->tx_txg == 0);
474 dmu_tx_count_dnode(txh);
477 * Modifying a almost-full microzap is around the worst case (128KB)
479 * If it is a fat zap, the worst case would be 7*16KB=112KB:
480 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
481 * - 4 new blocks written if adding:
482 * - 2 blocks for possibly split leaves,
483 * - 2 grown ptrtbl blocks
485 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
486 MZAP_MAX_BLKSZ, FTAG);
491 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
493 if (dn->dn_maxblkid == 0 || name == NULL) {
495 * This is a microzap (only one block), or we don't know
496 * the name. Check the first block for i/o errors.
498 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
504 * Access the name so that we'll check for i/o errors to
505 * the leaf blocks, etc. We ignore ENOENT, as this name
508 err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
509 if (err == EIO || err == ECKSUM || err == ENXIO) {
516 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
522 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
523 object, THT_ZAP, add, (uintptr_t)name);
525 dmu_tx_hold_zap_impl(txh, name);
529 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
536 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
538 dmu_tx_hold_zap_impl(txh, name);
542 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
546 ASSERT(tx->tx_txg == 0);
548 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
549 object, THT_BONUS, 0, 0);
551 dmu_tx_count_dnode(txh);
555 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
561 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
563 dmu_tx_count_dnode(txh);
567 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
570 ASSERT(tx->tx_txg == 0);
572 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
573 DMU_NEW_OBJECT, THT_SPACE, space, 0);
575 (void) zfs_refcount_add_many(&txh->txh_space_towrite, space, FTAG);
580 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
582 boolean_t match_object = B_FALSE;
583 boolean_t match_offset = B_FALSE;
586 dnode_t *dn = DB_DNODE(db);
587 ASSERT(tx->tx_txg != 0);
588 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
589 ASSERT3U(dn->dn_object, ==, db->db.db_object);
596 /* XXX No checking on the meta dnode for now */
597 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
602 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
603 txh = list_next(&tx->tx_holds, txh)) {
604 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
605 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
607 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
608 int datablkshift = dn->dn_datablkshift ?
609 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
610 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
611 int shift = datablkshift + epbs * db->db_level;
612 uint64_t beginblk = shift >= 64 ? 0 :
613 (txh->txh_arg1 >> shift);
614 uint64_t endblk = shift >= 64 ? 0 :
615 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
616 uint64_t blkid = db->db_blkid;
618 /* XXX txh_arg2 better not be zero... */
620 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
621 txh->txh_type, beginblk, endblk);
623 switch (txh->txh_type) {
625 if (blkid >= beginblk && blkid <= endblk)
628 * We will let this hold work for the bonus
629 * or spill buffer so that we don't need to
630 * hold it when creating a new object.
632 if (blkid == DMU_BONUS_BLKID ||
633 blkid == DMU_SPILL_BLKID)
636 * They might have to increase nlevels,
637 * thus dirtying the new TLIBs. Or the
638 * might have to change the block size,
639 * thus dirying the new lvl=0 blk=0.
646 * We will dirty all the level 1 blocks in
647 * the free range and perhaps the first and
648 * last level 0 block.
650 if (blkid >= beginblk && (blkid <= endblk ||
651 txh->txh_arg2 == DMU_OBJECT_END))
655 if (blkid == DMU_SPILL_BLKID)
659 if (blkid == DMU_BONUS_BLKID)
669 ASSERT(!"bad txh_type");
672 if (match_object && match_offset) {
678 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
679 (u_longlong_t)db->db.db_object, db->db_level,
680 (u_longlong_t)db->db_blkid);
685 * If we can't do 10 iops, something is wrong. Let us go ahead
686 * and hit zfs_dirty_data_max.
688 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
689 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
692 * We delay transactions when we've determined that the backend storage
693 * isn't able to accommodate the rate of incoming writes.
695 * If there is already a transaction waiting, we delay relative to when
696 * that transaction finishes waiting. This way the calculated min_time
697 * is independent of the number of threads concurrently executing
700 * If we are the only waiter, wait relative to when the transaction
701 * started, rather than the current time. This credits the transaction for
702 * "time already served", e.g. reading indirect blocks.
704 * The minimum time for a transaction to take is calculated as:
705 * min_time = scale * (dirty - min) / (max - dirty)
706 * min_time is then capped at zfs_delay_max_ns.
708 * The delay has two degrees of freedom that can be adjusted via tunables.
709 * The percentage of dirty data at which we start to delay is defined by
710 * zfs_delay_min_dirty_percent. This should typically be at or above
711 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
712 * delay after writing at full speed has failed to keep up with the incoming
713 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
714 * speaking, this variable determines the amount of delay at the midpoint of
718 * 10ms +-------------------------------------------------------------*+
734 * 2ms + (midpoint) * +
737 * | zfs_delay_scale ----------> ******** |
738 * 0 +-------------------------------------*********----------------+
739 * 0% <- zfs_dirty_data_max -> 100%
741 * Note that since the delay is added to the outstanding time remaining on the
742 * most recent transaction, the delay is effectively the inverse of IOPS.
743 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
744 * was chosen such that small changes in the amount of accumulated dirty data
745 * in the first 3/4 of the curve yield relatively small differences in the
748 * The effects can be easier to understand when the amount of delay is
749 * represented on a log scale:
752 * 100ms +-------------------------------------------------------------++
761 * + zfs_delay_scale ----------> ***** +
772 * +--------------------------------------------------------------+
773 * 0% <- zfs_dirty_data_max -> 100%
775 * Note here that only as the amount of dirty data approaches its limit does
776 * the delay start to increase rapidly. The goal of a properly tuned system
777 * should be to keep the amount of dirty data out of that range by first
778 * ensuring that the appropriate limits are set for the I/O scheduler to reach
779 * optimal throughput on the backend storage, and then by changing the value
780 * of zfs_delay_scale to increase the steepness of the curve.
783 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
785 dsl_pool_t *dp = tx->tx_pool;
786 uint64_t delay_min_bytes =
787 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
788 hrtime_t wakeup, min_tx_time, now;
790 if (dirty <= delay_min_bytes)
794 * The caller has already waited until we are under the max.
795 * We make them pass us the amount of dirty data so we don't
796 * have to handle the case of it being >= the max, which could
797 * cause a divide-by-zero if it's == the max.
799 ASSERT3U(dirty, <, zfs_dirty_data_max);
802 min_tx_time = zfs_delay_scale *
803 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
804 if (now > tx->tx_start + min_tx_time)
807 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
809 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
810 uint64_t, min_tx_time);
812 mutex_enter(&dp->dp_lock);
813 wakeup = MAX(tx->tx_start + min_tx_time,
814 dp->dp_last_wakeup + min_tx_time);
815 dp->dp_last_wakeup = wakeup;
816 mutex_exit(&dp->dp_lock);
820 mutex_enter(&curthread->t_delay_lock);
821 while (cv_timedwait_hires(&curthread->t_delay_cv,
822 &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
823 CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
825 mutex_exit(&curthread->t_delay_lock);
827 pause_sbt("dmu_tx_delay", nstosbt(wakeup),
828 nstosbt(zfs_delay_resolution_ns), C_ABSOLUTE);
831 hrtime_t delta = wakeup - gethrtime();
833 ts.tv_sec = delta / NANOSEC;
834 ts.tv_nsec = delta % NANOSEC;
835 (void) nanosleep(&ts, NULL);
840 * This routine attempts to assign the transaction to a transaction group.
841 * To do so, we must determine if there is sufficient free space on disk.
843 * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
844 * on it), then it is assumed that there is sufficient free space,
845 * unless there's insufficient slop space in the pool (see the comment
846 * above spa_slop_shift in spa_misc.c).
848 * If it is not a "netfree" transaction, then if the data already on disk
849 * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
850 * ENOSPC. Otherwise, if the current rough estimate of pending changes,
851 * plus the rough estimate of this transaction's changes, may exceed the
852 * allowed usage, then this will fail with ERESTART, which will cause the
853 * caller to wait for the pending changes to be written to disk (by waiting
854 * for the next TXG to open), and then check the space usage again.
856 * The rough estimate of pending changes is comprised of the sum of:
858 * - this transaction's holds' txh_space_towrite
860 * - dd_tempreserved[], which is the sum of in-flight transactions'
861 * holds' txh_space_towrite (i.e. those transactions that have called
862 * dmu_tx_assign() but not yet called dmu_tx_commit()).
864 * - dd_space_towrite[], which is the amount of dirtied dbufs.
866 * Note that all of these values are inflated by spa_get_worst_case_asize(),
867 * which means that we may get ERESTART well before we are actually in danger
868 * of running out of space, but this also mitigates any small inaccuracies
869 * in the rough estimate (e.g. txh_space_towrite doesn't take into account
870 * indirect blocks, and dd_space_towrite[] doesn't take into account changes
873 * Note that due to this algorithm, it is possible to exceed the allowed
874 * usage by one transaction. Also, as we approach the allowed usage,
875 * we will allow a very limited amount of changes into each TXG, thus
876 * decreasing performance.
879 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
881 spa_t *spa = tx->tx_pool->dp_spa;
888 if (spa_suspended(spa)) {
890 * If the user has indicated a blocking failure mode
891 * then return ERESTART which will block in dmu_tx_wait().
892 * Otherwise, return EIO so that an error can get
893 * propagated back to the VOP calls.
895 * Note that we always honor the txg_how flag regardless
896 * of the failuremode setting.
898 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
899 !(txg_how & TXG_WAIT))
900 return (SET_ERROR(EIO));
902 return (SET_ERROR(ERESTART));
905 if (!tx->tx_dirty_delayed &&
906 dsl_pool_need_dirty_delay(tx->tx_pool)) {
907 tx->tx_wait_dirty = B_TRUE;
908 return (SET_ERROR(ERESTART));
911 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
912 tx->tx_needassign_txh = NULL;
915 * NB: No error returns are allowed after txg_hold_open, but
916 * before processing the dnode holds, due to the
917 * dmu_tx_unassign() logic.
920 uint64_t towrite = 0;
922 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
923 txh = list_next(&tx->tx_holds, txh)) {
924 dnode_t *dn = txh->txh_dnode;
926 mutex_enter(&dn->dn_mtx);
927 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
928 mutex_exit(&dn->dn_mtx);
929 tx->tx_needassign_txh = txh;
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) zfs_refcount_add(&dn->dn_tx_holds, tx);
936 mutex_exit(&dn->dn_mtx);
938 towrite += zfs_refcount_count(&txh->txh_space_towrite);
939 tohold += zfs_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);
958 dmu_tx_unassign(dmu_tx_t *tx)
963 txg_rele_to_quiesce(&tx->tx_txgh);
966 * Walk the transaction's hold list, removing the hold on the
967 * associated dnode, and notifying waiters if the refcount drops to 0.
969 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
970 txh != tx->tx_needassign_txh;
971 txh = list_next(&tx->tx_holds, txh)) {
972 dnode_t *dn = txh->txh_dnode;
976 mutex_enter(&dn->dn_mtx);
977 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
979 if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
980 dn->dn_assigned_txg = 0;
981 cv_broadcast(&dn->dn_notxholds);
983 mutex_exit(&dn->dn_mtx);
986 txg_rele_to_sync(&tx->tx_txgh);
988 tx->tx_lasttried_txg = tx->tx_txg;
993 * Assign tx to a transaction group; txg_how is a bitmask:
995 * If TXG_WAIT is set and the currently open txg is full, this function
996 * will wait until there's a new txg. This should be used when no locks
997 * are being held. With this bit set, this function will only fail if
998 * we're truly out of space (or over quota).
1000 * If TXG_WAIT is *not* set and we can't assign into the currently open
1001 * txg without blocking, this function will return immediately with
1002 * ERESTART. This should be used whenever locks are being held. On an
1003 * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
1006 * If TXG_NOTHROTTLE is set, this indicates that this tx should not be
1007 * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
1008 * details on the throttle). This is used by the VFS operations, after
1009 * they have already called dmu_tx_wait() (though most likely on a
1013 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1017 ASSERT(tx->tx_txg == 0);
1018 ASSERT0(txg_how & ~(TXG_WAIT | TXG_NOTHROTTLE));
1019 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1021 /* If we might wait, we must not hold the config lock. */
1022 IMPLY((txg_how & TXG_WAIT), !dsl_pool_config_held(tx->tx_pool));
1024 if ((txg_how & TXG_NOTHROTTLE))
1025 tx->tx_dirty_delayed = B_TRUE;
1027 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1028 dmu_tx_unassign(tx);
1030 if (err != ERESTART || !(txg_how & TXG_WAIT))
1036 txg_rele_to_quiesce(&tx->tx_txgh);
1042 dmu_tx_wait(dmu_tx_t *tx)
1044 spa_t *spa = tx->tx_pool->dp_spa;
1045 dsl_pool_t *dp = tx->tx_pool;
1047 ASSERT(tx->tx_txg == 0);
1048 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1050 if (tx->tx_wait_dirty) {
1052 * dmu_tx_try_assign() has determined that we need to wait
1053 * because we've consumed much or all of the dirty buffer
1056 mutex_enter(&dp->dp_lock);
1057 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1058 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1059 uint64_t dirty = dp->dp_dirty_total;
1060 mutex_exit(&dp->dp_lock);
1062 dmu_tx_delay(tx, dirty);
1064 tx->tx_wait_dirty = B_FALSE;
1067 * Note: setting tx_dirty_delayed only has effect if the
1068 * caller used TX_WAIT. Otherwise they are going to
1069 * destroy this tx and try again. The common case,
1070 * zfs_write(), uses TX_WAIT.
1072 tx->tx_dirty_delayed = B_TRUE;
1073 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1075 * If the pool is suspended we need to wait until it
1076 * is resumed. Note that it's possible that the pool
1077 * has become active after this thread has tried to
1078 * obtain a tx. If that's the case then tx_lasttried_txg
1079 * would not have been set.
1081 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1082 } else if (tx->tx_needassign_txh) {
1084 * A dnode is assigned to the quiescing txg. Wait for its
1085 * transaction to complete.
1087 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1089 mutex_enter(&dn->dn_mtx);
1090 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1091 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1092 mutex_exit(&dn->dn_mtx);
1093 tx->tx_needassign_txh = NULL;
1096 * If we have a lot of dirty data just wait until we sync
1097 * out a TXG at which point we'll hopefully have synced
1098 * a portion of the changes.
1100 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1105 dmu_tx_destroy(dmu_tx_t *tx)
1109 while ((txh = list_head(&tx->tx_holds)) != NULL) {
1110 dnode_t *dn = txh->txh_dnode;
1112 list_remove(&tx->tx_holds, txh);
1113 zfs_refcount_destroy_many(&txh->txh_space_towrite,
1114 zfs_refcount_count(&txh->txh_space_towrite));
1115 zfs_refcount_destroy_many(&txh->txh_memory_tohold,
1116 zfs_refcount_count(&txh->txh_memory_tohold));
1117 kmem_free(txh, sizeof (dmu_tx_hold_t));
1122 list_destroy(&tx->tx_callbacks);
1123 list_destroy(&tx->tx_holds);
1124 kmem_free(tx, sizeof (dmu_tx_t));
1128 dmu_tx_commit(dmu_tx_t *tx)
1130 ASSERT(tx->tx_txg != 0);
1133 * Go through the transaction's hold list and remove holds on
1134 * associated dnodes, notifying waiters if no holds remain.
1136 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1137 txh = list_next(&tx->tx_holds, txh)) {
1138 dnode_t *dn = txh->txh_dnode;
1143 mutex_enter(&dn->dn_mtx);
1144 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1146 if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1147 dn->dn_assigned_txg = 0;
1148 cv_broadcast(&dn->dn_notxholds);
1150 mutex_exit(&dn->dn_mtx);
1153 if (tx->tx_tempreserve_cookie)
1154 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1156 if (!list_is_empty(&tx->tx_callbacks))
1157 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1159 if (tx->tx_anyobj == FALSE)
1160 txg_rele_to_sync(&tx->tx_txgh);
1166 dmu_tx_abort(dmu_tx_t *tx)
1168 ASSERT(tx->tx_txg == 0);
1171 * Call any registered callbacks with an error code.
1173 if (!list_is_empty(&tx->tx_callbacks))
1174 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1180 dmu_tx_get_txg(dmu_tx_t *tx)
1182 ASSERT(tx->tx_txg != 0);
1183 return (tx->tx_txg);
1187 dmu_tx_pool(dmu_tx_t *tx)
1189 ASSERT(tx->tx_pool != NULL);
1190 return (tx->tx_pool);
1194 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1196 dmu_tx_callback_t *dcb;
1198 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1200 dcb->dcb_func = func;
1201 dcb->dcb_data = data;
1203 list_insert_tail(&tx->tx_callbacks, dcb);
1207 * Call all the commit callbacks on a list, with a given error code.
1210 dmu_tx_do_callbacks(list_t *cb_list, int error)
1212 dmu_tx_callback_t *dcb;
1214 while ((dcb = list_head(cb_list)) != NULL) {
1215 list_remove(cb_list, dcb);
1216 dcb->dcb_func(dcb->dcb_data, error);
1217 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1222 * Interface to hold a bunch of attributes.
1223 * used for creating new files.
1224 * attrsize is the total size of all attributes
1225 * to be added during object creation
1227 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1231 * hold necessary attribute name for attribute registration.
1232 * should be a very rare case where this is needed. If it does
1233 * happen it would only happen on the first write to the file system.
1236 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1238 if (!sa->sa_need_attr_registration)
1241 for (int i = 0; i != sa->sa_num_attrs; i++) {
1242 if (!sa->sa_attr_table[i].sa_registered) {
1243 if (sa->sa_reg_attr_obj)
1244 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1245 B_TRUE, sa->sa_attr_table[i].sa_name);
1247 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1248 B_TRUE, sa->sa_attr_table[i].sa_name);
1254 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1258 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1261 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
1262 SPA_OLD_MAXBLOCKSIZE, FTAG);
1266 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1268 sa_os_t *sa = tx->tx_objset->os_sa;
1270 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1272 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1275 if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1276 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1278 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1279 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1280 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1281 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1284 dmu_tx_sa_registration_hold(sa, tx);
1286 if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill)
1289 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1296 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1298 * variable_size is the total size of all variable sized attributes
1299 * passed to this function. It is not the total size of all
1300 * variable size attributes that *may* exist on this object.
1303 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1306 sa_os_t *sa = tx->tx_objset->os_sa;
1308 ASSERT(hdl != NULL);
1310 object = sa_handle_object(hdl);
1312 dmu_tx_hold_bonus(tx, object);
1314 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1317 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1318 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1319 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1320 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1321 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1322 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1325 dmu_tx_sa_registration_hold(sa, tx);
1327 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1328 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1330 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1331 ASSERT(tx->tx_txg == 0);
1332 dmu_tx_hold_spill(tx, object);
1334 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1339 if (dn->dn_have_spill) {
1340 ASSERT(tx->tx_txg == 0);
1341 dmu_tx_hold_spill(tx, object);