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_write_by_dnode(dmu_tx_t
*tx
, dnode_t
*dn
, uint64_t off
, int len
)
324 ASSERT3U(len
, <=, DMU_MAX_ACCESS
);
325 ASSERT(len
== 0 || UINT64_MAX
- off
>= len
- 1);
327 txh
= dmu_tx_hold_dnode_impl(tx
, dn
, THT_WRITE
, off
, len
);
329 dmu_tx_count_write(txh
, off
, len
);
330 dmu_tx_count_dnode(txh
);
335 * This function marks the transaction as being a "net free". The end
336 * result is that refquotas will be disabled for this transaction, and
337 * this transaction will be able to use half of the pool space overhead
338 * (see dsl_pool_adjustedsize()). Therefore this function should only
339 * be called for transactions that we expect will not cause a net increase
340 * in the amount of space used (but it's OK if that is occasionally not true).
343 dmu_tx_mark_netfree(dmu_tx_t
*tx
)
345 tx
->tx_netfree
= B_TRUE
;
349 dmu_tx_hold_free_impl(dmu_tx_hold_t
*txh
, uint64_t off
, uint64_t len
)
351 dmu_tx_t
*tx
= txh
->txh_tx
;
352 dnode_t
*dn
= txh
->txh_dnode
;
355 ASSERT(tx
->tx_txg
== 0);
357 dmu_tx_count_dnode(txh
);
359 if (off
>= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
)
361 if (len
== DMU_OBJECT_END
)
362 len
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
- off
;
364 dmu_tx_count_dnode(txh
);
367 * For i/o error checking, we read the first and last level-0
368 * blocks if they are not aligned, and all the level-1 blocks.
370 * Note: dbuf_free_range() assumes that we have not instantiated
371 * any level-0 dbufs that will be completely freed. Therefore we must
372 * exercise care to not read or count the first and last blocks
373 * if they are blocksize-aligned.
375 if (dn
->dn_datablkshift
== 0) {
376 if (off
!= 0 || len
< dn
->dn_datablksz
)
377 dmu_tx_count_write(txh
, 0, dn
->dn_datablksz
);
379 /* first block will be modified if it is not aligned */
380 if (!IS_P2ALIGNED(off
, 1 << dn
->dn_datablkshift
))
381 dmu_tx_count_write(txh
, off
, 1);
382 /* last block will be modified if it is not aligned */
383 if (!IS_P2ALIGNED(off
+ len
, 1 << dn
->dn_datablkshift
))
384 dmu_tx_count_write(txh
, off
+ len
, 1);
388 * Check level-1 blocks.
390 if (dn
->dn_nlevels
> 1) {
391 int shift
= dn
->dn_datablkshift
+ dn
->dn_indblkshift
-
393 uint64_t start
= off
>> shift
;
394 uint64_t end
= (off
+ len
) >> shift
;
396 ASSERT(dn
->dn_indblkshift
!= 0);
399 * dnode_reallocate() can result in an object with indirect
400 * blocks having an odd data block size. In this case,
401 * just check the single block.
403 if (dn
->dn_datablkshift
== 0)
406 zio_t
*zio
= zio_root(tx
->tx_pool
->dp_spa
,
407 NULL
, NULL
, ZIO_FLAG_CANFAIL
);
408 for (uint64_t i
= start
; i
<= end
; i
++) {
409 uint64_t ibyte
= i
<< shift
;
410 err
= dnode_next_offset(dn
, 0, &ibyte
, 2, 1, 0);
412 if (err
== ESRCH
|| i
> end
)
416 (void) zio_wait(zio
);
420 (void) refcount_add_many(&txh
->txh_memory_tohold
,
421 1 << dn
->dn_indblkshift
, FTAG
);
423 err
= dmu_tx_check_ioerr(zio
, dn
, 1, i
);
426 (void) zio_wait(zio
);
439 dmu_tx_hold_free(dmu_tx_t
*tx
, uint64_t object
, uint64_t off
, uint64_t len
)
443 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
444 object
, THT_FREE
, off
, len
);
446 (void) dmu_tx_hold_free_impl(txh
, off
, len
);
450 dmu_tx_hold_free_by_dnode(dmu_tx_t
*tx
, dnode_t
*dn
, uint64_t off
, uint64_t len
)
454 txh
= dmu_tx_hold_dnode_impl(tx
, dn
, THT_FREE
, off
, len
);
456 (void) dmu_tx_hold_free_impl(txh
, off
, len
);
460 dmu_tx_hold_zap_impl(dmu_tx_hold_t
*txh
, const char *name
)
462 dmu_tx_t
*tx
= txh
->txh_tx
;
463 dnode_t
*dn
= txh
->txh_dnode
;
466 ASSERT(tx
->tx_txg
== 0);
468 dmu_tx_count_dnode(txh
);
471 * Modifying a almost-full microzap is around the worst case (128KB)
473 * If it is a fat zap, the worst case would be 7*16KB=112KB:
474 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
475 * - 4 new blocks written if adding:
476 * - 2 blocks for possibly split leaves,
477 * - 2 grown ptrtbl blocks
479 (void) refcount_add_many(&txh
->txh_space_towrite
,
480 MZAP_MAX_BLKSZ
, FTAG
);
485 ASSERT3U(DMU_OT_BYTESWAP(dn
->dn_type
), ==, DMU_BSWAP_ZAP
);
487 if (dn
->dn_maxblkid
== 0 || name
== NULL
) {
489 * This is a microzap (only one block), or we don't know
490 * the name. Check the first block for i/o errors.
492 err
= dmu_tx_check_ioerr(NULL
, dn
, 0, 0);
498 * Access the name so that we'll check for i/o errors to
499 * the leaf blocks, etc. We ignore ENOENT, as this name
502 err
= zap_lookup_by_dnode(dn
, name
, 8, 0, NULL
);
503 if (err
== EIO
|| err
== ECKSUM
|| err
== ENXIO
) {
510 dmu_tx_hold_zap(dmu_tx_t
*tx
, uint64_t object
, int add
, const char *name
)
516 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
517 object
, THT_ZAP
, add
, (uintptr_t)name
);
519 dmu_tx_hold_zap_impl(txh
, name
);
523 dmu_tx_hold_zap_by_dnode(dmu_tx_t
*tx
, dnode_t
*dn
, int add
, const char *name
)
530 txh
= dmu_tx_hold_dnode_impl(tx
, dn
, THT_ZAP
, add
, (uintptr_t)name
);
532 dmu_tx_hold_zap_impl(txh
, name
);
536 dmu_tx_hold_bonus(dmu_tx_t
*tx
, uint64_t object
)
540 ASSERT(tx
->tx_txg
== 0);
542 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
543 object
, THT_BONUS
, 0, 0);
545 dmu_tx_count_dnode(txh
);
549 dmu_tx_hold_bonus_by_dnode(dmu_tx_t
*tx
, dnode_t
*dn
)
555 txh
= dmu_tx_hold_dnode_impl(tx
, dn
, THT_BONUS
, 0, 0);
557 dmu_tx_count_dnode(txh
);
561 dmu_tx_hold_space(dmu_tx_t
*tx
, uint64_t space
)
565 ASSERT(tx
->tx_txg
== 0);
567 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
568 DMU_NEW_OBJECT
, THT_SPACE
, space
, 0);
570 (void) refcount_add_many(&txh
->txh_space_towrite
, space
, FTAG
);
575 dmu_tx_dirty_buf(dmu_tx_t
*tx
, dmu_buf_impl_t
*db
)
577 boolean_t match_object
= B_FALSE
;
578 boolean_t match_offset
= B_FALSE
;
581 dnode_t
*dn
= DB_DNODE(db
);
582 ASSERT(tx
->tx_txg
!= 0);
583 ASSERT(tx
->tx_objset
== NULL
|| dn
->dn_objset
== tx
->tx_objset
);
584 ASSERT3U(dn
->dn_object
, ==, db
->db
.db_object
);
591 /* XXX No checking on the meta dnode for now */
592 if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
597 for (dmu_tx_hold_t
*txh
= list_head(&tx
->tx_holds
); txh
!= NULL
;
598 txh
= list_next(&tx
->tx_holds
, txh
)) {
599 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
600 if (txh
->txh_dnode
== dn
&& txh
->txh_type
!= THT_NEWOBJECT
)
602 if (txh
->txh_dnode
== NULL
|| txh
->txh_dnode
== dn
) {
603 int datablkshift
= dn
->dn_datablkshift
?
604 dn
->dn_datablkshift
: SPA_MAXBLOCKSHIFT
;
605 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
606 int shift
= datablkshift
+ epbs
* db
->db_level
;
607 uint64_t beginblk
= shift
>= 64 ? 0 :
608 (txh
->txh_arg1
>> shift
);
609 uint64_t endblk
= shift
>= 64 ? 0 :
610 ((txh
->txh_arg1
+ txh
->txh_arg2
- 1) >> shift
);
611 uint64_t blkid
= db
->db_blkid
;
613 /* XXX txh_arg2 better not be zero... */
615 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
616 txh
->txh_type
, beginblk
, endblk
);
618 switch (txh
->txh_type
) {
620 if (blkid
>= beginblk
&& blkid
<= endblk
)
623 * We will let this hold work for the bonus
624 * or spill buffer so that we don't need to
625 * hold it when creating a new object.
627 if (blkid
== DMU_BONUS_BLKID
||
628 blkid
== DMU_SPILL_BLKID
)
631 * They might have to increase nlevels,
632 * thus dirtying the new TLIBs. Or the
633 * might have to change the block size,
634 * thus dirying the new lvl=0 blk=0.
641 * We will dirty all the level 1 blocks in
642 * the free range and perhaps the first and
643 * last level 0 block.
645 if (blkid
>= beginblk
&& (blkid
<= endblk
||
646 txh
->txh_arg2
== DMU_OBJECT_END
))
650 if (blkid
== DMU_SPILL_BLKID
)
654 if (blkid
== DMU_BONUS_BLKID
)
664 cmn_err(CE_PANIC
, "bad txh_type %d",
668 if (match_object
&& match_offset
) {
674 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
675 (u_longlong_t
)db
->db
.db_object
, db
->db_level
,
676 (u_longlong_t
)db
->db_blkid
);
681 * If we can't do 10 iops, something is wrong. Let us go ahead
682 * and hit zfs_dirty_data_max.
684 hrtime_t zfs_delay_max_ns
= 100 * MICROSEC
; /* 100 milliseconds */
685 int zfs_delay_resolution_ns
= 100 * 1000; /* 100 microseconds */
688 * We delay transactions when we've determined that the backend storage
689 * isn't able to accommodate the rate of incoming writes.
691 * If there is already a transaction waiting, we delay relative to when
692 * that transaction finishes waiting. This way the calculated min_time
693 * is independent of the number of threads concurrently executing
696 * If we are the only waiter, wait relative to when the transaction
697 * started, rather than the current time. This credits the transaction for
698 * "time already served", e.g. reading indirect blocks.
700 * The minimum time for a transaction to take is calculated as:
701 * min_time = scale * (dirty - min) / (max - dirty)
702 * min_time is then capped at zfs_delay_max_ns.
704 * The delay has two degrees of freedom that can be adjusted via tunables.
705 * The percentage of dirty data at which we start to delay is defined by
706 * zfs_delay_min_dirty_percent. This should typically be at or above
707 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
708 * delay after writing at full speed has failed to keep up with the incoming
709 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
710 * speaking, this variable determines the amount of delay at the midpoint of
714 * 10ms +-------------------------------------------------------------*+
730 * 2ms + (midpoint) * +
733 * | zfs_delay_scale ----------> ******** |
734 * 0 +-------------------------------------*********----------------+
735 * 0% <- zfs_dirty_data_max -> 100%
737 * Note that since the delay is added to the outstanding time remaining on the
738 * most recent transaction, the delay is effectively the inverse of IOPS.
739 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
740 * was chosen such that small changes in the amount of accumulated dirty data
741 * in the first 3/4 of the curve yield relatively small differences in the
744 * The effects can be easier to understand when the amount of delay is
745 * represented on a log scale:
748 * 100ms +-------------------------------------------------------------++
757 * + zfs_delay_scale ----------> ***** +
768 * +--------------------------------------------------------------+
769 * 0% <- zfs_dirty_data_max -> 100%
771 * Note here that only as the amount of dirty data approaches its limit does
772 * the delay start to increase rapidly. The goal of a properly tuned system
773 * should be to keep the amount of dirty data out of that range by first
774 * ensuring that the appropriate limits are set for the I/O scheduler to reach
775 * optimal throughput on the backend storage, and then by changing the value
776 * of zfs_delay_scale to increase the steepness of the curve.
779 dmu_tx_delay(dmu_tx_t
*tx
, uint64_t dirty
)
781 dsl_pool_t
*dp
= tx
->tx_pool
;
782 uint64_t delay_min_bytes
=
783 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
784 hrtime_t wakeup
, min_tx_time
, now
;
786 if (dirty
<= delay_min_bytes
)
790 * The caller has already waited until we are under the max.
791 * We make them pass us the amount of dirty data so we don't
792 * have to handle the case of it being >= the max, which could
793 * cause a divide-by-zero if it's == the max.
795 ASSERT3U(dirty
, <, zfs_dirty_data_max
);
798 min_tx_time
= zfs_delay_scale
*
799 (dirty
- delay_min_bytes
) / (zfs_dirty_data_max
- dirty
);
800 min_tx_time
= MIN(min_tx_time
, zfs_delay_max_ns
);
801 if (now
> tx
->tx_start
+ min_tx_time
)
804 DTRACE_PROBE3(delay__mintime
, dmu_tx_t
*, tx
, uint64_t, dirty
,
805 uint64_t, min_tx_time
);
807 mutex_enter(&dp
->dp_lock
);
808 wakeup
= MAX(tx
->tx_start
+ min_tx_time
,
809 dp
->dp_last_wakeup
+ min_tx_time
);
810 dp
->dp_last_wakeup
= wakeup
;
811 mutex_exit(&dp
->dp_lock
);
813 zfs_sleep_until(wakeup
);
817 * This routine attempts to assign the transaction to a transaction group.
818 * To do so, we must determine if there is sufficient free space on disk.
820 * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
821 * on it), then it is assumed that there is sufficient free space,
822 * unless there's insufficient slop space in the pool (see the comment
823 * above spa_slop_shift in spa_misc.c).
825 * If it is not a "netfree" transaction, then if the data already on disk
826 * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
827 * ENOSPC. Otherwise, if the current rough estimate of pending changes,
828 * plus the rough estimate of this transaction's changes, may exceed the
829 * allowed usage, then this will fail with ERESTART, which will cause the
830 * caller to wait for the pending changes to be written to disk (by waiting
831 * for the next TXG to open), and then check the space usage again.
833 * The rough estimate of pending changes is comprised of the sum of:
835 * - this transaction's holds' txh_space_towrite
837 * - dd_tempreserved[], which is the sum of in-flight transactions'
838 * holds' txh_space_towrite (i.e. those transactions that have called
839 * dmu_tx_assign() but not yet called dmu_tx_commit()).
841 * - dd_space_towrite[], which is the amount of dirtied dbufs.
843 * Note that all of these values are inflated by spa_get_worst_case_asize(),
844 * which means that we may get ERESTART well before we are actually in danger
845 * of running out of space, but this also mitigates any small inaccuracies
846 * in the rough estimate (e.g. txh_space_towrite doesn't take into account
847 * indirect blocks, and dd_space_towrite[] doesn't take into account changes
850 * Note that due to this algorithm, it is possible to exceed the allowed
851 * usage by one transaction. Also, as we approach the allowed usage,
852 * we will allow a very limited amount of changes into each TXG, thus
853 * decreasing performance.
856 dmu_tx_try_assign(dmu_tx_t
*tx
, uint64_t txg_how
)
858 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
863 DMU_TX_STAT_BUMP(dmu_tx_error
);
867 if (spa_suspended(spa
)) {
868 DMU_TX_STAT_BUMP(dmu_tx_suspended
);
871 * If the user has indicated a blocking failure mode
872 * then return ERESTART which will block in dmu_tx_wait().
873 * Otherwise, return EIO so that an error can get
874 * propagated back to the VOP calls.
876 * Note that we always honor the txg_how flag regardless
877 * of the failuremode setting.
879 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_CONTINUE
&&
880 !(txg_how
& TXG_WAIT
))
881 return (SET_ERROR(EIO
));
883 return (SET_ERROR(ERESTART
));
886 if (!tx
->tx_dirty_delayed
&&
887 dsl_pool_need_dirty_delay(tx
->tx_pool
)) {
888 tx
->tx_wait_dirty
= B_TRUE
;
889 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay
);
890 return (SET_ERROR(ERESTART
));
893 tx
->tx_txg
= txg_hold_open(tx
->tx_pool
, &tx
->tx_txgh
);
894 tx
->tx_needassign_txh
= NULL
;
897 * NB: No error returns are allowed after txg_hold_open, but
898 * before processing the dnode holds, due to the
899 * dmu_tx_unassign() logic.
902 uint64_t towrite
= 0;
904 for (dmu_tx_hold_t
*txh
= list_head(&tx
->tx_holds
); txh
!= NULL
;
905 txh
= list_next(&tx
->tx_holds
, txh
)) {
906 dnode_t
*dn
= txh
->txh_dnode
;
908 mutex_enter(&dn
->dn_mtx
);
909 if (dn
->dn_assigned_txg
== tx
->tx_txg
- 1) {
910 mutex_exit(&dn
->dn_mtx
);
911 tx
->tx_needassign_txh
= txh
;
912 DMU_TX_STAT_BUMP(dmu_tx_group
);
913 return (SET_ERROR(ERESTART
));
915 if (dn
->dn_assigned_txg
== 0)
916 dn
->dn_assigned_txg
= tx
->tx_txg
;
917 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
918 (void) refcount_add(&dn
->dn_tx_holds
, tx
);
919 mutex_exit(&dn
->dn_mtx
);
921 towrite
+= refcount_count(&txh
->txh_space_towrite
);
922 tohold
+= refcount_count(&txh
->txh_memory_tohold
);
925 /* needed allocation: worst-case estimate of write space */
926 uint64_t asize
= spa_get_worst_case_asize(tx
->tx_pool
->dp_spa
, towrite
);
927 /* calculate memory footprint estimate */
928 uint64_t memory
= towrite
+ tohold
;
930 if (tx
->tx_dir
!= NULL
&& asize
!= 0) {
931 int err
= dsl_dir_tempreserve_space(tx
->tx_dir
, memory
,
932 asize
, tx
->tx_netfree
, &tx
->tx_tempreserve_cookie
, tx
);
937 DMU_TX_STAT_BUMP(dmu_tx_assigned
);
943 dmu_tx_unassign(dmu_tx_t
*tx
)
948 txg_rele_to_quiesce(&tx
->tx_txgh
);
951 * Walk the transaction's hold list, removing the hold on the
952 * associated dnode, and notifying waiters if the refcount drops to 0.
954 for (dmu_tx_hold_t
*txh
= list_head(&tx
->tx_holds
);
955 txh
&& txh
!= tx
->tx_needassign_txh
;
956 txh
= list_next(&tx
->tx_holds
, txh
)) {
957 dnode_t
*dn
= txh
->txh_dnode
;
961 mutex_enter(&dn
->dn_mtx
);
962 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
964 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
965 dn
->dn_assigned_txg
= 0;
966 cv_broadcast(&dn
->dn_notxholds
);
968 mutex_exit(&dn
->dn_mtx
);
971 txg_rele_to_sync(&tx
->tx_txgh
);
973 tx
->tx_lasttried_txg
= tx
->tx_txg
;
978 * Assign tx to a transaction group; txg_how is a bitmask:
980 * If TXG_WAIT is set and the currently open txg is full, this function
981 * will wait until there's a new txg. This should be used when no locks
982 * are being held. With this bit set, this function will only fail if
983 * we're truly out of space (or over quota).
985 * If TXG_WAIT is *not* set and we can't assign into the currently open
986 * txg without blocking, this function will return immediately with
987 * ERESTART. This should be used whenever locks are being held. On an
988 * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
991 * If TXG_NOTHROTTLE is set, this indicates that this tx should not be
992 * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
993 * details on the throttle). This is used by the VFS operations, after
994 * they have already called dmu_tx_wait() (though most likely on a
998 dmu_tx_assign(dmu_tx_t
*tx
, uint64_t txg_how
)
1002 ASSERT(tx
->tx_txg
== 0);
1003 ASSERT0(txg_how
& ~(TXG_WAIT
| TXG_NOTHROTTLE
));
1004 ASSERT(!dsl_pool_sync_context(tx
->tx_pool
));
1006 /* If we might wait, we must not hold the config lock. */
1007 IMPLY((txg_how
& TXG_WAIT
), !dsl_pool_config_held(tx
->tx_pool
));
1009 if ((txg_how
& TXG_NOTHROTTLE
))
1010 tx
->tx_dirty_delayed
= B_TRUE
;
1012 while ((err
= dmu_tx_try_assign(tx
, txg_how
)) != 0) {
1013 dmu_tx_unassign(tx
);
1015 if (err
!= ERESTART
|| !(txg_how
& TXG_WAIT
))
1021 txg_rele_to_quiesce(&tx
->tx_txgh
);
1027 dmu_tx_wait(dmu_tx_t
*tx
)
1029 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
1030 dsl_pool_t
*dp
= tx
->tx_pool
;
1033 ASSERT(tx
->tx_txg
== 0);
1034 ASSERT(!dsl_pool_config_held(tx
->tx_pool
));
1036 before
= gethrtime();
1038 if (tx
->tx_wait_dirty
) {
1042 * dmu_tx_try_assign() has determined that we need to wait
1043 * because we've consumed much or all of the dirty buffer
1046 mutex_enter(&dp
->dp_lock
);
1047 if (dp
->dp_dirty_total
>= zfs_dirty_data_max
)
1048 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max
);
1049 while (dp
->dp_dirty_total
>= zfs_dirty_data_max
)
1050 cv_wait(&dp
->dp_spaceavail_cv
, &dp
->dp_lock
);
1051 dirty
= dp
->dp_dirty_total
;
1052 mutex_exit(&dp
->dp_lock
);
1054 dmu_tx_delay(tx
, dirty
);
1056 tx
->tx_wait_dirty
= B_FALSE
;
1059 * Note: setting tx_dirty_delayed only has effect if the
1060 * caller used TX_WAIT. Otherwise they are going to
1061 * destroy this tx and try again. The common case,
1062 * zfs_write(), uses TX_WAIT.
1064 tx
->tx_dirty_delayed
= B_TRUE
;
1065 } else if (spa_suspended(spa
) || tx
->tx_lasttried_txg
== 0) {
1067 * If the pool is suspended we need to wait until it
1068 * is resumed. Note that it's possible that the pool
1069 * has become active after this thread has tried to
1070 * obtain a tx. If that's the case then tx_lasttried_txg
1071 * would not have been set.
1073 txg_wait_synced(dp
, spa_last_synced_txg(spa
) + 1);
1074 } else if (tx
->tx_needassign_txh
) {
1075 dnode_t
*dn
= tx
->tx_needassign_txh
->txh_dnode
;
1077 mutex_enter(&dn
->dn_mtx
);
1078 while (dn
->dn_assigned_txg
== tx
->tx_lasttried_txg
- 1)
1079 cv_wait(&dn
->dn_notxholds
, &dn
->dn_mtx
);
1080 mutex_exit(&dn
->dn_mtx
);
1081 tx
->tx_needassign_txh
= NULL
;
1084 * A dnode is assigned to the quiescing txg. Wait for its
1085 * transaction to complete.
1087 txg_wait_open(tx
->tx_pool
, tx
->tx_lasttried_txg
+ 1);
1090 spa_tx_assign_add_nsecs(spa
, gethrtime() - before
);
1094 dmu_tx_destroy(dmu_tx_t
*tx
)
1098 while ((txh
= list_head(&tx
->tx_holds
)) != NULL
) {
1099 dnode_t
*dn
= txh
->txh_dnode
;
1101 list_remove(&tx
->tx_holds
, txh
);
1102 refcount_destroy_many(&txh
->txh_space_towrite
,
1103 refcount_count(&txh
->txh_space_towrite
));
1104 refcount_destroy_many(&txh
->txh_memory_tohold
,
1105 refcount_count(&txh
->txh_memory_tohold
));
1106 kmem_free(txh
, sizeof (dmu_tx_hold_t
));
1111 list_destroy(&tx
->tx_callbacks
);
1112 list_destroy(&tx
->tx_holds
);
1113 kmem_free(tx
, sizeof (dmu_tx_t
));
1117 dmu_tx_commit(dmu_tx_t
*tx
)
1119 ASSERT(tx
->tx_txg
!= 0);
1122 * Go through the transaction's hold list and remove holds on
1123 * associated dnodes, notifying waiters if no holds remain.
1125 for (dmu_tx_hold_t
*txh
= list_head(&tx
->tx_holds
); txh
!= NULL
;
1126 txh
= list_next(&tx
->tx_holds
, txh
)) {
1127 dnode_t
*dn
= txh
->txh_dnode
;
1132 mutex_enter(&dn
->dn_mtx
);
1133 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1135 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
1136 dn
->dn_assigned_txg
= 0;
1137 cv_broadcast(&dn
->dn_notxholds
);
1139 mutex_exit(&dn
->dn_mtx
);
1142 if (tx
->tx_tempreserve_cookie
)
1143 dsl_dir_tempreserve_clear(tx
->tx_tempreserve_cookie
, tx
);
1145 if (!list_is_empty(&tx
->tx_callbacks
))
1146 txg_register_callbacks(&tx
->tx_txgh
, &tx
->tx_callbacks
);
1148 if (tx
->tx_anyobj
== FALSE
)
1149 txg_rele_to_sync(&tx
->tx_txgh
);
1155 dmu_tx_abort(dmu_tx_t
*tx
)
1157 ASSERT(tx
->tx_txg
== 0);
1160 * Call any registered callbacks with an error code.
1162 if (!list_is_empty(&tx
->tx_callbacks
))
1163 dmu_tx_do_callbacks(&tx
->tx_callbacks
, ECANCELED
);
1169 dmu_tx_get_txg(dmu_tx_t
*tx
)
1171 ASSERT(tx
->tx_txg
!= 0);
1172 return (tx
->tx_txg
);
1176 dmu_tx_pool(dmu_tx_t
*tx
)
1178 ASSERT(tx
->tx_pool
!= NULL
);
1179 return (tx
->tx_pool
);
1183 dmu_tx_callback_register(dmu_tx_t
*tx
, dmu_tx_callback_func_t
*func
, void *data
)
1185 dmu_tx_callback_t
*dcb
;
1187 dcb
= kmem_alloc(sizeof (dmu_tx_callback_t
), KM_SLEEP
);
1189 dcb
->dcb_func
= func
;
1190 dcb
->dcb_data
= data
;
1192 list_insert_tail(&tx
->tx_callbacks
, dcb
);
1196 * Call all the commit callbacks on a list, with a given error code.
1199 dmu_tx_do_callbacks(list_t
*cb_list
, int error
)
1201 dmu_tx_callback_t
*dcb
;
1203 while ((dcb
= list_tail(cb_list
)) != NULL
) {
1204 list_remove(cb_list
, dcb
);
1205 dcb
->dcb_func(dcb
->dcb_data
, error
);
1206 kmem_free(dcb
, sizeof (dmu_tx_callback_t
));
1211 * Interface to hold a bunch of attributes.
1212 * used for creating new files.
1213 * attrsize is the total size of all attributes
1214 * to be added during object creation
1216 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1220 * hold necessary attribute name for attribute registration.
1221 * should be a very rare case where this is needed. If it does
1222 * happen it would only happen on the first write to the file system.
1225 dmu_tx_sa_registration_hold(sa_os_t
*sa
, dmu_tx_t
*tx
)
1227 if (!sa
->sa_need_attr_registration
)
1230 for (int i
= 0; i
!= sa
->sa_num_attrs
; i
++) {
1231 if (!sa
->sa_attr_table
[i
].sa_registered
) {
1232 if (sa
->sa_reg_attr_obj
)
1233 dmu_tx_hold_zap(tx
, sa
->sa_reg_attr_obj
,
1234 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1236 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
,
1237 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1243 dmu_tx_hold_spill(dmu_tx_t
*tx
, uint64_t object
)
1247 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, object
,
1250 (void) refcount_add_many(&txh
->txh_space_towrite
,
1251 SPA_OLD_MAXBLOCKSIZE
, FTAG
);
1255 dmu_tx_hold_sa_create(dmu_tx_t
*tx
, int attrsize
)
1257 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1259 dmu_tx_hold_bonus(tx
, DMU_NEW_OBJECT
);
1261 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1264 if (tx
->tx_objset
->os_sa
->sa_layout_attr_obj
) {
1265 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1267 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1268 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1269 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1270 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1273 dmu_tx_sa_registration_hold(sa
, tx
);
1275 if (attrsize
<= DN_OLD_MAX_BONUSLEN
&& !sa
->sa_force_spill
)
1278 (void) dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, DMU_NEW_OBJECT
,
1285 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1287 * variable_size is the total size of all variable sized attributes
1288 * passed to this function. It is not the total size of all
1289 * variable size attributes that *may* exist on this object.
1292 dmu_tx_hold_sa(dmu_tx_t
*tx
, sa_handle_t
*hdl
, boolean_t may_grow
)
1295 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1297 ASSERT(hdl
!= NULL
);
1299 object
= sa_handle_object(hdl
);
1301 dmu_tx_hold_bonus(tx
, object
);
1303 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1306 if (tx
->tx_objset
->os_sa
->sa_reg_attr_obj
== 0 ||
1307 tx
->tx_objset
->os_sa
->sa_layout_attr_obj
== 0) {
1308 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1309 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1310 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1311 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1314 dmu_tx_sa_registration_hold(sa
, tx
);
1316 if (may_grow
&& tx
->tx_objset
->os_sa
->sa_layout_attr_obj
)
1317 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1319 if (sa
->sa_force_spill
|| may_grow
|| hdl
->sa_spill
) {
1320 ASSERT(tx
->tx_txg
== 0);
1321 dmu_tx_hold_spill(tx
, object
);
1323 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)hdl
->sa_bonus
;
1328 if (dn
->dn_have_spill
) {
1329 ASSERT(tx
->tx_txg
== 0);
1330 dmu_tx_hold_spill(tx
, object
);
1339 dmu_tx_ksp
= kstat_create("zfs", 0, "dmu_tx", "misc",
1340 KSTAT_TYPE_NAMED
, sizeof (dmu_tx_stats
) / sizeof (kstat_named_t
),
1341 KSTAT_FLAG_VIRTUAL
);
1343 if (dmu_tx_ksp
!= NULL
) {
1344 dmu_tx_ksp
->ks_data
= &dmu_tx_stats
;
1345 kstat_install(dmu_tx_ksp
);
1352 if (dmu_tx_ksp
!= NULL
) {
1353 kstat_delete(dmu_tx_ksp
);
1358 #if defined(_KERNEL) && defined(HAVE_SPL)
1359 EXPORT_SYMBOL(dmu_tx_create
);
1360 EXPORT_SYMBOL(dmu_tx_hold_write
);
1361 EXPORT_SYMBOL(dmu_tx_hold_write_by_dnode
);
1362 EXPORT_SYMBOL(dmu_tx_hold_free
);
1363 EXPORT_SYMBOL(dmu_tx_hold_free_by_dnode
);
1364 EXPORT_SYMBOL(dmu_tx_hold_zap
);
1365 EXPORT_SYMBOL(dmu_tx_hold_zap_by_dnode
);
1366 EXPORT_SYMBOL(dmu_tx_hold_bonus
);
1367 EXPORT_SYMBOL(dmu_tx_hold_bonus_by_dnode
);
1368 EXPORT_SYMBOL(dmu_tx_abort
);
1369 EXPORT_SYMBOL(dmu_tx_assign
);
1370 EXPORT_SYMBOL(dmu_tx_wait
);
1371 EXPORT_SYMBOL(dmu_tx_commit
);
1372 EXPORT_SYMBOL(dmu_tx_mark_netfree
);
1373 EXPORT_SYMBOL(dmu_tx_get_txg
);
1374 EXPORT_SYMBOL(dmu_tx_callback_register
);
1375 EXPORT_SYMBOL(dmu_tx_do_callbacks
);
1376 EXPORT_SYMBOL(dmu_tx_hold_spill
);
1377 EXPORT_SYMBOL(dmu_tx_hold_sa_create
);
1378 EXPORT_SYMBOL(dmu_tx_hold_sa
);