4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/zfs_context.h>
30 #include <sys/dnode.h>
32 #include <sys/dmu_impl.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_dataset.h>
39 #include <sys/dmu_zfetch.h>
40 #include <sys/range_tree.h>
42 static kmem_cache_t
*dnode_cache
;
44 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
45 * turned on when DEBUG is also defined.
52 #define DNODE_STAT_ADD(stat) ((stat)++)
54 #define DNODE_STAT_ADD(stat) /* nothing */
55 #endif /* DNODE_STATS */
57 static dnode_phys_t dnode_phys_zero
;
59 int zfs_default_bs
= SPA_MINBLOCKSHIFT
;
60 int zfs_default_ibs
= DN_MAX_INDBLKSHIFT
;
62 static kmem_cbrc_t
dnode_move(void *, void *, size_t, void *);
65 dbuf_compare(const void *x1
, const void *x2
)
67 const dmu_buf_impl_t
*d1
= x1
;
68 const dmu_buf_impl_t
*d2
= x2
;
70 if (d1
->db_level
< d2
->db_level
) {
73 if (d1
->db_level
> d2
->db_level
) {
77 if (d1
->db_blkid
< d2
->db_blkid
) {
80 if (d1
->db_blkid
> d2
->db_blkid
) {
84 if (d1
->db_state
== DB_SEARCH
) {
85 ASSERT3S(d2
->db_state
, !=, DB_SEARCH
);
87 } else if (d2
->db_state
== DB_SEARCH
) {
88 ASSERT3S(d1
->db_state
, !=, DB_SEARCH
);
92 if ((uintptr_t)d1
< (uintptr_t)d2
) {
95 if ((uintptr_t)d1
> (uintptr_t)d2
) {
103 dnode_cons(void *arg
, void *unused
, int kmflag
)
108 rw_init(&dn
->dn_struct_rwlock
, NULL
, RW_DEFAULT
, NULL
);
109 mutex_init(&dn
->dn_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
110 mutex_init(&dn
->dn_dbufs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
111 cv_init(&dn
->dn_notxholds
, NULL
, CV_DEFAULT
, NULL
);
114 * Every dbuf has a reference, and dropping a tracked reference is
115 * O(number of references), so don't track dn_holds.
117 refcount_create_untracked(&dn
->dn_holds
);
118 refcount_create(&dn
->dn_tx_holds
);
119 list_link_init(&dn
->dn_link
);
121 bzero(&dn
->dn_next_nblkptr
[0], sizeof (dn
->dn_next_nblkptr
));
122 bzero(&dn
->dn_next_nlevels
[0], sizeof (dn
->dn_next_nlevels
));
123 bzero(&dn
->dn_next_indblkshift
[0], sizeof (dn
->dn_next_indblkshift
));
124 bzero(&dn
->dn_next_bonustype
[0], sizeof (dn
->dn_next_bonustype
));
125 bzero(&dn
->dn_rm_spillblk
[0], sizeof (dn
->dn_rm_spillblk
));
126 bzero(&dn
->dn_next_bonuslen
[0], sizeof (dn
->dn_next_bonuslen
));
127 bzero(&dn
->dn_next_blksz
[0], sizeof (dn
->dn_next_blksz
));
129 for (i
= 0; i
< TXG_SIZE
; i
++) {
130 list_link_init(&dn
->dn_dirty_link
[i
]);
131 dn
->dn_free_ranges
[i
] = NULL
;
132 list_create(&dn
->dn_dirty_records
[i
],
133 sizeof (dbuf_dirty_record_t
),
134 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
137 dn
->dn_allocated_txg
= 0;
139 dn
->dn_assigned_txg
= 0;
141 dn
->dn_dirtyctx_firstset
= NULL
;
143 dn
->dn_have_spill
= B_FALSE
;
153 dn
->dn_dbufs_count
= 0;
154 avl_create(&dn
->dn_dbufs
, dbuf_compare
, sizeof (dmu_buf_impl_t
),
155 offsetof(dmu_buf_impl_t
, db_link
));
163 dnode_dest(void *arg
, void *unused
)
168 rw_destroy(&dn
->dn_struct_rwlock
);
169 mutex_destroy(&dn
->dn_mtx
);
170 mutex_destroy(&dn
->dn_dbufs_mtx
);
171 cv_destroy(&dn
->dn_notxholds
);
172 refcount_destroy(&dn
->dn_holds
);
173 refcount_destroy(&dn
->dn_tx_holds
);
174 ASSERT(!list_link_active(&dn
->dn_link
));
176 for (i
= 0; i
< TXG_SIZE
; i
++) {
177 ASSERT(!list_link_active(&dn
->dn_dirty_link
[i
]));
178 ASSERT3P(dn
->dn_free_ranges
[i
], ==, NULL
);
179 list_destroy(&dn
->dn_dirty_records
[i
]);
180 ASSERT0(dn
->dn_next_nblkptr
[i
]);
181 ASSERT0(dn
->dn_next_nlevels
[i
]);
182 ASSERT0(dn
->dn_next_indblkshift
[i
]);
183 ASSERT0(dn
->dn_next_bonustype
[i
]);
184 ASSERT0(dn
->dn_rm_spillblk
[i
]);
185 ASSERT0(dn
->dn_next_bonuslen
[i
]);
186 ASSERT0(dn
->dn_next_blksz
[i
]);
189 ASSERT0(dn
->dn_allocated_txg
);
190 ASSERT0(dn
->dn_free_txg
);
191 ASSERT0(dn
->dn_assigned_txg
);
192 ASSERT0(dn
->dn_dirtyctx
);
193 ASSERT3P(dn
->dn_dirtyctx_firstset
, ==, NULL
);
194 ASSERT3P(dn
->dn_bonus
, ==, NULL
);
195 ASSERT(!dn
->dn_have_spill
);
196 ASSERT3P(dn
->dn_zio
, ==, NULL
);
197 ASSERT0(dn
->dn_oldused
);
198 ASSERT0(dn
->dn_oldflags
);
199 ASSERT0(dn
->dn_olduid
);
200 ASSERT0(dn
->dn_oldgid
);
201 ASSERT0(dn
->dn_newuid
);
202 ASSERT0(dn
->dn_newgid
);
203 ASSERT0(dn
->dn_id_flags
);
205 ASSERT0(dn
->dn_dbufs_count
);
206 avl_destroy(&dn
->dn_dbufs
);
212 ASSERT(dnode_cache
== NULL
);
213 dnode_cache
= kmem_cache_create("dnode_t",
215 0, dnode_cons
, dnode_dest
, NULL
, NULL
, NULL
, 0);
216 kmem_cache_set_move(dnode_cache
, dnode_move
);
222 kmem_cache_destroy(dnode_cache
);
229 dnode_verify(dnode_t
*dn
)
231 int drop_struct_lock
= FALSE
;
234 ASSERT(dn
->dn_objset
);
235 ASSERT(dn
->dn_handle
->dnh_dnode
== dn
);
237 ASSERT(DMU_OT_IS_VALID(dn
->dn_phys
->dn_type
));
239 if (!(zfs_flags
& ZFS_DEBUG_DNODE_VERIFY
))
242 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
243 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
244 drop_struct_lock
= TRUE
;
246 if (dn
->dn_phys
->dn_type
!= DMU_OT_NONE
|| dn
->dn_allocated_txg
!= 0) {
248 ASSERT3U(dn
->dn_indblkshift
, >=, 0);
249 ASSERT3U(dn
->dn_indblkshift
, <=, SPA_MAXBLOCKSHIFT
);
250 if (dn
->dn_datablkshift
) {
251 ASSERT3U(dn
->dn_datablkshift
, >=, SPA_MINBLOCKSHIFT
);
252 ASSERT3U(dn
->dn_datablkshift
, <=, SPA_MAXBLOCKSHIFT
);
253 ASSERT3U(1<<dn
->dn_datablkshift
, ==, dn
->dn_datablksz
);
255 ASSERT3U(dn
->dn_nlevels
, <=, 30);
256 ASSERT(DMU_OT_IS_VALID(dn
->dn_type
));
257 ASSERT3U(dn
->dn_nblkptr
, >=, 1);
258 ASSERT3U(dn
->dn_nblkptr
, <=, DN_MAX_NBLKPTR
);
259 ASSERT3U(dn
->dn_bonuslen
, <=, DN_MAX_BONUSLEN
);
260 ASSERT3U(dn
->dn_datablksz
, ==,
261 dn
->dn_datablkszsec
<< SPA_MINBLOCKSHIFT
);
262 ASSERT3U(ISP2(dn
->dn_datablksz
), ==, dn
->dn_datablkshift
!= 0);
263 ASSERT3U((dn
->dn_nblkptr
- 1) * sizeof (blkptr_t
) +
264 dn
->dn_bonuslen
, <=, DN_MAX_BONUSLEN
);
265 for (i
= 0; i
< TXG_SIZE
; i
++) {
266 ASSERT3U(dn
->dn_next_nlevels
[i
], <=, dn
->dn_nlevels
);
269 if (dn
->dn_phys
->dn_type
!= DMU_OT_NONE
)
270 ASSERT3U(dn
->dn_phys
->dn_nlevels
, <=, dn
->dn_nlevels
);
271 ASSERT(DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) || dn
->dn_dbuf
!= NULL
);
272 if (dn
->dn_dbuf
!= NULL
) {
273 ASSERT3P(dn
->dn_phys
, ==,
274 (dnode_phys_t
*)dn
->dn_dbuf
->db
.db_data
+
275 (dn
->dn_object
% (dn
->dn_dbuf
->db
.db_size
>> DNODE_SHIFT
)));
277 if (drop_struct_lock
)
278 rw_exit(&dn
->dn_struct_rwlock
);
283 dnode_byteswap(dnode_phys_t
*dnp
)
285 uint64_t *buf64
= (void*)&dnp
->dn_blkptr
;
288 if (dnp
->dn_type
== DMU_OT_NONE
) {
289 bzero(dnp
, sizeof (dnode_phys_t
));
293 dnp
->dn_datablkszsec
= BSWAP_16(dnp
->dn_datablkszsec
);
294 dnp
->dn_bonuslen
= BSWAP_16(dnp
->dn_bonuslen
);
295 dnp
->dn_maxblkid
= BSWAP_64(dnp
->dn_maxblkid
);
296 dnp
->dn_used
= BSWAP_64(dnp
->dn_used
);
299 * dn_nblkptr is only one byte, so it's OK to read it in either
300 * byte order. We can't read dn_bouslen.
302 ASSERT(dnp
->dn_indblkshift
<= SPA_MAXBLOCKSHIFT
);
303 ASSERT(dnp
->dn_nblkptr
<= DN_MAX_NBLKPTR
);
304 for (i
= 0; i
< dnp
->dn_nblkptr
* sizeof (blkptr_t
)/8; i
++)
305 buf64
[i
] = BSWAP_64(buf64
[i
]);
308 * OK to check dn_bonuslen for zero, because it won't matter if
309 * we have the wrong byte order. This is necessary because the
310 * dnode dnode is smaller than a regular dnode.
312 if (dnp
->dn_bonuslen
!= 0) {
314 * Note that the bonus length calculated here may be
315 * longer than the actual bonus buffer. This is because
316 * we always put the bonus buffer after the last block
317 * pointer (instead of packing it against the end of the
320 int off
= (dnp
->dn_nblkptr
-1) * sizeof (blkptr_t
);
321 size_t len
= DN_MAX_BONUSLEN
- off
;
322 ASSERT(DMU_OT_IS_VALID(dnp
->dn_bonustype
));
323 dmu_object_byteswap_t byteswap
=
324 DMU_OT_BYTESWAP(dnp
->dn_bonustype
);
325 dmu_ot_byteswap
[byteswap
].ob_func(dnp
->dn_bonus
+ off
, len
);
328 /* Swap SPILL block if we have one */
329 if (dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)
330 byteswap_uint64_array(&dnp
->dn_spill
, sizeof (blkptr_t
));
335 dnode_buf_byteswap(void *vbuf
, size_t size
)
337 dnode_phys_t
*buf
= vbuf
;
340 ASSERT3U(sizeof (dnode_phys_t
), ==, (1<<DNODE_SHIFT
));
341 ASSERT((size
& (sizeof (dnode_phys_t
)-1)) == 0);
343 size
>>= DNODE_SHIFT
;
344 for (i
= 0; i
< size
; i
++) {
351 dnode_setbonuslen(dnode_t
*dn
, int newsize
, dmu_tx_t
*tx
)
353 ASSERT3U(refcount_count(&dn
->dn_holds
), >=, 1);
355 dnode_setdirty(dn
, tx
);
356 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
357 ASSERT3U(newsize
, <=, DN_MAX_BONUSLEN
-
358 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
));
359 dn
->dn_bonuslen
= newsize
;
361 dn
->dn_next_bonuslen
[tx
->tx_txg
& TXG_MASK
] = DN_ZERO_BONUSLEN
;
363 dn
->dn_next_bonuslen
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_bonuslen
;
364 rw_exit(&dn
->dn_struct_rwlock
);
368 dnode_setbonus_type(dnode_t
*dn
, dmu_object_type_t newtype
, dmu_tx_t
*tx
)
370 ASSERT3U(refcount_count(&dn
->dn_holds
), >=, 1);
371 dnode_setdirty(dn
, tx
);
372 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
373 dn
->dn_bonustype
= newtype
;
374 dn
->dn_next_bonustype
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_bonustype
;
375 rw_exit(&dn
->dn_struct_rwlock
);
379 dnode_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
381 ASSERT3U(refcount_count(&dn
->dn_holds
), >=, 1);
382 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
383 dnode_setdirty(dn
, tx
);
384 dn
->dn_rm_spillblk
[tx
->tx_txg
&TXG_MASK
] = DN_KILL_SPILLBLK
;
385 dn
->dn_have_spill
= B_FALSE
;
389 dnode_setdblksz(dnode_t
*dn
, int size
)
391 ASSERT0(P2PHASE(size
, SPA_MINBLOCKSIZE
));
392 ASSERT3U(size
, <=, SPA_MAXBLOCKSIZE
);
393 ASSERT3U(size
, >=, SPA_MINBLOCKSIZE
);
394 ASSERT3U(size
>> SPA_MINBLOCKSHIFT
, <,
395 1<<(sizeof (dn
->dn_phys
->dn_datablkszsec
) * 8));
396 dn
->dn_datablksz
= size
;
397 dn
->dn_datablkszsec
= size
>> SPA_MINBLOCKSHIFT
;
398 dn
->dn_datablkshift
= ISP2(size
) ? highbit64(size
- 1) : 0;
402 dnode_create(objset_t
*os
, dnode_phys_t
*dnp
, dmu_buf_impl_t
*db
,
403 uint64_t object
, dnode_handle_t
*dnh
)
407 dn
= kmem_cache_alloc(dnode_cache
, KM_SLEEP
);
408 ASSERT(!POINTER_IS_VALID(dn
->dn_objset
));
412 * Defer setting dn_objset until the dnode is ready to be a candidate
413 * for the dnode_move() callback.
415 dn
->dn_object
= object
;
420 if (dnp
->dn_datablkszsec
) {
421 dnode_setdblksz(dn
, dnp
->dn_datablkszsec
<< SPA_MINBLOCKSHIFT
);
423 dn
->dn_datablksz
= 0;
424 dn
->dn_datablkszsec
= 0;
425 dn
->dn_datablkshift
= 0;
427 dn
->dn_indblkshift
= dnp
->dn_indblkshift
;
428 dn
->dn_nlevels
= dnp
->dn_nlevels
;
429 dn
->dn_type
= dnp
->dn_type
;
430 dn
->dn_nblkptr
= dnp
->dn_nblkptr
;
431 dn
->dn_checksum
= dnp
->dn_checksum
;
432 dn
->dn_compress
= dnp
->dn_compress
;
433 dn
->dn_bonustype
= dnp
->dn_bonustype
;
434 dn
->dn_bonuslen
= dnp
->dn_bonuslen
;
435 dn
->dn_maxblkid
= dnp
->dn_maxblkid
;
436 dn
->dn_have_spill
= ((dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) != 0);
439 dmu_zfetch_init(&dn
->dn_zfetch
, dn
);
441 ASSERT(DMU_OT_IS_VALID(dn
->dn_phys
->dn_type
));
443 mutex_enter(&os
->os_lock
);
444 if (dnh
->dnh_dnode
!= NULL
) {
445 /* Lost the allocation race. */
446 mutex_exit(&os
->os_lock
);
447 kmem_cache_free(dnode_cache
, dn
);
448 return (dnh
->dnh_dnode
);
452 * Exclude special dnodes from os_dnodes so an empty os_dnodes
453 * signifies that the special dnodes have no references from
454 * their children (the entries in os_dnodes). This allows
455 * dnode_destroy() to easily determine if the last child has
456 * been removed and then complete eviction of the objset.
458 if (!DMU_OBJECT_IS_SPECIAL(object
))
459 list_insert_head(&os
->os_dnodes
, dn
);
463 * Everything else must be valid before assigning dn_objset
464 * makes the dnode eligible for dnode_move().
469 mutex_exit(&os
->os_lock
);
471 arc_space_consume(sizeof (dnode_t
), ARC_SPACE_OTHER
);
476 * Caller must be holding the dnode handle, which is released upon return.
479 dnode_destroy(dnode_t
*dn
)
481 objset_t
*os
= dn
->dn_objset
;
482 boolean_t complete_os_eviction
= B_FALSE
;
484 ASSERT((dn
->dn_id_flags
& DN_ID_NEW_EXIST
) == 0);
486 mutex_enter(&os
->os_lock
);
487 POINTER_INVALIDATE(&dn
->dn_objset
);
488 if (!DMU_OBJECT_IS_SPECIAL(dn
->dn_object
)) {
489 list_remove(&os
->os_dnodes
, dn
);
490 complete_os_eviction
=
491 list_is_empty(&os
->os_dnodes
) &&
492 list_link_active(&os
->os_evicting_node
);
494 mutex_exit(&os
->os_lock
);
496 /* the dnode can no longer move, so we can release the handle */
497 zrl_remove(&dn
->dn_handle
->dnh_zrlock
);
499 dn
->dn_allocated_txg
= 0;
501 dn
->dn_assigned_txg
= 0;
504 if (dn
->dn_dirtyctx_firstset
!= NULL
) {
505 kmem_free(dn
->dn_dirtyctx_firstset
, 1);
506 dn
->dn_dirtyctx_firstset
= NULL
;
508 if (dn
->dn_bonus
!= NULL
) {
509 mutex_enter(&dn
->dn_bonus
->db_mtx
);
510 dbuf_destroy(dn
->dn_bonus
);
515 dn
->dn_have_spill
= B_FALSE
;
524 dmu_zfetch_fini(&dn
->dn_zfetch
);
525 kmem_cache_free(dnode_cache
, dn
);
526 arc_space_return(sizeof (dnode_t
), ARC_SPACE_OTHER
);
528 if (complete_os_eviction
)
529 dmu_objset_evict_done(os
);
533 dnode_allocate(dnode_t
*dn
, dmu_object_type_t ot
, int blocksize
, int ibs
,
534 dmu_object_type_t bonustype
, int bonuslen
, dmu_tx_t
*tx
)
538 ASSERT3U(blocksize
, <=,
539 spa_maxblocksize(dmu_objset_spa(dn
->dn_objset
)));
541 blocksize
= 1 << zfs_default_bs
;
543 blocksize
= P2ROUNDUP(blocksize
, SPA_MINBLOCKSIZE
);
546 ibs
= zfs_default_ibs
;
548 ibs
= MIN(MAX(ibs
, DN_MIN_INDBLKSHIFT
), DN_MAX_INDBLKSHIFT
);
550 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn
->dn_objset
,
551 dn
->dn_object
, tx
->tx_txg
, blocksize
, ibs
);
553 ASSERT(dn
->dn_type
== DMU_OT_NONE
);
554 ASSERT(bcmp(dn
->dn_phys
, &dnode_phys_zero
, sizeof (dnode_phys_t
)) == 0);
555 ASSERT(dn
->dn_phys
->dn_type
== DMU_OT_NONE
);
556 ASSERT(ot
!= DMU_OT_NONE
);
557 ASSERT(DMU_OT_IS_VALID(ot
));
558 ASSERT((bonustype
== DMU_OT_NONE
&& bonuslen
== 0) ||
559 (bonustype
== DMU_OT_SA
&& bonuslen
== 0) ||
560 (bonustype
!= DMU_OT_NONE
&& bonuslen
!= 0));
561 ASSERT(DMU_OT_IS_VALID(bonustype
));
562 ASSERT3U(bonuslen
, <=, DN_MAX_BONUSLEN
);
563 ASSERT(dn
->dn_type
== DMU_OT_NONE
);
564 ASSERT0(dn
->dn_maxblkid
);
565 ASSERT0(dn
->dn_allocated_txg
);
566 ASSERT0(dn
->dn_assigned_txg
);
567 ASSERT(refcount_is_zero(&dn
->dn_tx_holds
));
568 ASSERT3U(refcount_count(&dn
->dn_holds
), <=, 1);
569 ASSERT(avl_is_empty(&dn
->dn_dbufs
));
571 for (i
= 0; i
< TXG_SIZE
; i
++) {
572 ASSERT0(dn
->dn_next_nblkptr
[i
]);
573 ASSERT0(dn
->dn_next_nlevels
[i
]);
574 ASSERT0(dn
->dn_next_indblkshift
[i
]);
575 ASSERT0(dn
->dn_next_bonuslen
[i
]);
576 ASSERT0(dn
->dn_next_bonustype
[i
]);
577 ASSERT0(dn
->dn_rm_spillblk
[i
]);
578 ASSERT0(dn
->dn_next_blksz
[i
]);
579 ASSERT(!list_link_active(&dn
->dn_dirty_link
[i
]));
580 ASSERT3P(list_head(&dn
->dn_dirty_records
[i
]), ==, NULL
);
581 ASSERT3P(dn
->dn_free_ranges
[i
], ==, NULL
);
585 dnode_setdblksz(dn
, blocksize
);
586 dn
->dn_indblkshift
= ibs
;
588 if (bonustype
== DMU_OT_SA
) /* Maximize bonus space for SA */
592 ((DN_MAX_BONUSLEN
- bonuslen
) >> SPA_BLKPTRSHIFT
);
593 dn
->dn_bonustype
= bonustype
;
594 dn
->dn_bonuslen
= bonuslen
;
595 dn
->dn_checksum
= ZIO_CHECKSUM_INHERIT
;
596 dn
->dn_compress
= ZIO_COMPRESS_INHERIT
;
600 if (dn
->dn_dirtyctx_firstset
) {
601 kmem_free(dn
->dn_dirtyctx_firstset
, 1);
602 dn
->dn_dirtyctx_firstset
= NULL
;
605 dn
->dn_allocated_txg
= tx
->tx_txg
;
608 dnode_setdirty(dn
, tx
);
609 dn
->dn_next_indblkshift
[tx
->tx_txg
& TXG_MASK
] = ibs
;
610 dn
->dn_next_bonuslen
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_bonuslen
;
611 dn
->dn_next_bonustype
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_bonustype
;
612 dn
->dn_next_blksz
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_datablksz
;
616 dnode_reallocate(dnode_t
*dn
, dmu_object_type_t ot
, int blocksize
,
617 dmu_object_type_t bonustype
, int bonuslen
, dmu_tx_t
*tx
)
621 ASSERT3U(blocksize
, >=, SPA_MINBLOCKSIZE
);
622 ASSERT3U(blocksize
, <=,
623 spa_maxblocksize(dmu_objset_spa(dn
->dn_objset
)));
624 ASSERT0(blocksize
% SPA_MINBLOCKSIZE
);
625 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
|| dmu_tx_private_ok(tx
));
626 ASSERT(tx
->tx_txg
!= 0);
627 ASSERT((bonustype
== DMU_OT_NONE
&& bonuslen
== 0) ||
628 (bonustype
!= DMU_OT_NONE
&& bonuslen
!= 0) ||
629 (bonustype
== DMU_OT_SA
&& bonuslen
== 0));
630 ASSERT(DMU_OT_IS_VALID(bonustype
));
631 ASSERT3U(bonuslen
, <=, DN_MAX_BONUSLEN
);
633 /* clean up any unreferenced dbufs */
634 dnode_evict_dbufs(dn
);
638 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
639 dnode_setdirty(dn
, tx
);
640 if (dn
->dn_datablksz
!= blocksize
) {
641 /* change blocksize */
642 ASSERT(dn
->dn_maxblkid
== 0 &&
643 (BP_IS_HOLE(&dn
->dn_phys
->dn_blkptr
[0]) ||
644 dnode_block_freed(dn
, 0)));
645 dnode_setdblksz(dn
, blocksize
);
646 dn
->dn_next_blksz
[tx
->tx_txg
&TXG_MASK
] = blocksize
;
648 if (dn
->dn_bonuslen
!= bonuslen
)
649 dn
->dn_next_bonuslen
[tx
->tx_txg
&TXG_MASK
] = bonuslen
;
651 if (bonustype
== DMU_OT_SA
) /* Maximize bonus space for SA */
654 nblkptr
= 1 + ((DN_MAX_BONUSLEN
- bonuslen
) >> SPA_BLKPTRSHIFT
);
655 if (dn
->dn_bonustype
!= bonustype
)
656 dn
->dn_next_bonustype
[tx
->tx_txg
&TXG_MASK
] = bonustype
;
657 if (dn
->dn_nblkptr
!= nblkptr
)
658 dn
->dn_next_nblkptr
[tx
->tx_txg
&TXG_MASK
] = nblkptr
;
659 if (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) {
660 dbuf_rm_spill(dn
, tx
);
661 dnode_rm_spill(dn
, tx
);
663 rw_exit(&dn
->dn_struct_rwlock
);
668 /* change bonus size and type */
669 mutex_enter(&dn
->dn_mtx
);
670 dn
->dn_bonustype
= bonustype
;
671 dn
->dn_bonuslen
= bonuslen
;
672 dn
->dn_nblkptr
= nblkptr
;
673 dn
->dn_checksum
= ZIO_CHECKSUM_INHERIT
;
674 dn
->dn_compress
= ZIO_COMPRESS_INHERIT
;
675 ASSERT3U(dn
->dn_nblkptr
, <=, DN_MAX_NBLKPTR
);
677 /* fix up the bonus db_size */
679 dn
->dn_bonus
->db
.db_size
=
680 DN_MAX_BONUSLEN
- (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
681 ASSERT(dn
->dn_bonuslen
<= dn
->dn_bonus
->db
.db_size
);
684 dn
->dn_allocated_txg
= tx
->tx_txg
;
685 mutex_exit(&dn
->dn_mtx
);
690 uint64_t dms_dnode_invalid
;
691 uint64_t dms_dnode_recheck1
;
692 uint64_t dms_dnode_recheck2
;
693 uint64_t dms_dnode_special
;
694 uint64_t dms_dnode_handle
;
695 uint64_t dms_dnode_rwlock
;
696 uint64_t dms_dnode_active
;
698 #endif /* DNODE_STATS */
701 dnode_move_impl(dnode_t
*odn
, dnode_t
*ndn
)
705 ASSERT(!RW_LOCK_HELD(&odn
->dn_struct_rwlock
));
706 ASSERT(MUTEX_NOT_HELD(&odn
->dn_mtx
));
707 ASSERT(MUTEX_NOT_HELD(&odn
->dn_dbufs_mtx
));
708 ASSERT(!RW_LOCK_HELD(&odn
->dn_zfetch
.zf_rwlock
));
711 ndn
->dn_objset
= odn
->dn_objset
;
712 ndn
->dn_object
= odn
->dn_object
;
713 ndn
->dn_dbuf
= odn
->dn_dbuf
;
714 ndn
->dn_handle
= odn
->dn_handle
;
715 ndn
->dn_phys
= odn
->dn_phys
;
716 ndn
->dn_type
= odn
->dn_type
;
717 ndn
->dn_bonuslen
= odn
->dn_bonuslen
;
718 ndn
->dn_bonustype
= odn
->dn_bonustype
;
719 ndn
->dn_nblkptr
= odn
->dn_nblkptr
;
720 ndn
->dn_checksum
= odn
->dn_checksum
;
721 ndn
->dn_compress
= odn
->dn_compress
;
722 ndn
->dn_nlevels
= odn
->dn_nlevels
;
723 ndn
->dn_indblkshift
= odn
->dn_indblkshift
;
724 ndn
->dn_datablkshift
= odn
->dn_datablkshift
;
725 ndn
->dn_datablkszsec
= odn
->dn_datablkszsec
;
726 ndn
->dn_datablksz
= odn
->dn_datablksz
;
727 ndn
->dn_maxblkid
= odn
->dn_maxblkid
;
728 bcopy(&odn
->dn_next_nblkptr
[0], &ndn
->dn_next_nblkptr
[0],
729 sizeof (odn
->dn_next_nblkptr
));
730 bcopy(&odn
->dn_next_nlevels
[0], &ndn
->dn_next_nlevels
[0],
731 sizeof (odn
->dn_next_nlevels
));
732 bcopy(&odn
->dn_next_indblkshift
[0], &ndn
->dn_next_indblkshift
[0],
733 sizeof (odn
->dn_next_indblkshift
));
734 bcopy(&odn
->dn_next_bonustype
[0], &ndn
->dn_next_bonustype
[0],
735 sizeof (odn
->dn_next_bonustype
));
736 bcopy(&odn
->dn_rm_spillblk
[0], &ndn
->dn_rm_spillblk
[0],
737 sizeof (odn
->dn_rm_spillblk
));
738 bcopy(&odn
->dn_next_bonuslen
[0], &ndn
->dn_next_bonuslen
[0],
739 sizeof (odn
->dn_next_bonuslen
));
740 bcopy(&odn
->dn_next_blksz
[0], &ndn
->dn_next_blksz
[0],
741 sizeof (odn
->dn_next_blksz
));
742 for (i
= 0; i
< TXG_SIZE
; i
++) {
743 list_move_tail(&ndn
->dn_dirty_records
[i
],
744 &odn
->dn_dirty_records
[i
]);
746 bcopy(&odn
->dn_free_ranges
[0], &ndn
->dn_free_ranges
[0],
747 sizeof (odn
->dn_free_ranges
));
748 ndn
->dn_allocated_txg
= odn
->dn_allocated_txg
;
749 ndn
->dn_free_txg
= odn
->dn_free_txg
;
750 ndn
->dn_assigned_txg
= odn
->dn_assigned_txg
;
751 ndn
->dn_dirtyctx
= odn
->dn_dirtyctx
;
752 ndn
->dn_dirtyctx_firstset
= odn
->dn_dirtyctx_firstset
;
753 ASSERT(refcount_count(&odn
->dn_tx_holds
) == 0);
754 refcount_transfer(&ndn
->dn_holds
, &odn
->dn_holds
);
755 ASSERT(avl_is_empty(&ndn
->dn_dbufs
));
756 avl_swap(&ndn
->dn_dbufs
, &odn
->dn_dbufs
);
757 ndn
->dn_dbufs_count
= odn
->dn_dbufs_count
;
758 ndn
->dn_bonus
= odn
->dn_bonus
;
759 ndn
->dn_have_spill
= odn
->dn_have_spill
;
760 ndn
->dn_zio
= odn
->dn_zio
;
761 ndn
->dn_oldused
= odn
->dn_oldused
;
762 ndn
->dn_oldflags
= odn
->dn_oldflags
;
763 ndn
->dn_olduid
= odn
->dn_olduid
;
764 ndn
->dn_oldgid
= odn
->dn_oldgid
;
765 ndn
->dn_newuid
= odn
->dn_newuid
;
766 ndn
->dn_newgid
= odn
->dn_newgid
;
767 ndn
->dn_id_flags
= odn
->dn_id_flags
;
768 dmu_zfetch_init(&ndn
->dn_zfetch
, NULL
);
769 list_move_tail(&ndn
->dn_zfetch
.zf_stream
, &odn
->dn_zfetch
.zf_stream
);
770 ndn
->dn_zfetch
.zf_dnode
= odn
->dn_zfetch
.zf_dnode
;
773 * Update back pointers. Updating the handle fixes the back pointer of
774 * every descendant dbuf as well as the bonus dbuf.
776 ASSERT(ndn
->dn_handle
->dnh_dnode
== odn
);
777 ndn
->dn_handle
->dnh_dnode
= ndn
;
778 if (ndn
->dn_zfetch
.zf_dnode
== odn
) {
779 ndn
->dn_zfetch
.zf_dnode
= ndn
;
783 * Invalidate the original dnode by clearing all of its back pointers.
786 odn
->dn_handle
= NULL
;
787 avl_create(&odn
->dn_dbufs
, dbuf_compare
, sizeof (dmu_buf_impl_t
),
788 offsetof(dmu_buf_impl_t
, db_link
));
789 odn
->dn_dbufs_count
= 0;
790 odn
->dn_bonus
= NULL
;
791 odn
->dn_zfetch
.zf_dnode
= NULL
;
794 * Set the low bit of the objset pointer to ensure that dnode_move()
795 * recognizes the dnode as invalid in any subsequent callback.
797 POINTER_INVALIDATE(&odn
->dn_objset
);
800 * Satisfy the destructor.
802 for (i
= 0; i
< TXG_SIZE
; i
++) {
803 list_create(&odn
->dn_dirty_records
[i
],
804 sizeof (dbuf_dirty_record_t
),
805 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
806 odn
->dn_free_ranges
[i
] = NULL
;
807 odn
->dn_next_nlevels
[i
] = 0;
808 odn
->dn_next_indblkshift
[i
] = 0;
809 odn
->dn_next_bonustype
[i
] = 0;
810 odn
->dn_rm_spillblk
[i
] = 0;
811 odn
->dn_next_bonuslen
[i
] = 0;
812 odn
->dn_next_blksz
[i
] = 0;
814 odn
->dn_allocated_txg
= 0;
815 odn
->dn_free_txg
= 0;
816 odn
->dn_assigned_txg
= 0;
817 odn
->dn_dirtyctx
= 0;
818 odn
->dn_dirtyctx_firstset
= NULL
;
819 odn
->dn_have_spill
= B_FALSE
;
822 odn
->dn_oldflags
= 0;
827 odn
->dn_id_flags
= 0;
833 odn
->dn_moved
= (uint8_t)-1;
839 dnode_move(void *buf
, void *newbuf
, size_t size
, void *arg
)
841 dnode_t
*odn
= buf
, *ndn
= newbuf
;
847 * The dnode is on the objset's list of known dnodes if the objset
848 * pointer is valid. We set the low bit of the objset pointer when
849 * freeing the dnode to invalidate it, and the memory patterns written
850 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
851 * A newly created dnode sets the objset pointer last of all to indicate
852 * that the dnode is known and in a valid state to be moved by this
856 if (!POINTER_IS_VALID(os
)) {
857 DNODE_STAT_ADD(dnode_move_stats
.dms_dnode_invalid
);
858 return (KMEM_CBRC_DONT_KNOW
);
862 * Ensure that the objset does not go away during the move.
864 rw_enter(&os_lock
, RW_WRITER
);
865 if (os
!= odn
->dn_objset
) {
867 DNODE_STAT_ADD(dnode_move_stats
.dms_dnode_recheck1
);
868 return (KMEM_CBRC_DONT_KNOW
);
872 * If the dnode is still valid, then so is the objset. We know that no
873 * valid objset can be freed while we hold os_lock, so we can safely
874 * ensure that the objset remains in use.
876 mutex_enter(&os
->os_lock
);
879 * Recheck the objset pointer in case the dnode was removed just before
880 * acquiring the lock.
882 if (os
!= odn
->dn_objset
) {
883 mutex_exit(&os
->os_lock
);
885 DNODE_STAT_ADD(dnode_move_stats
.dms_dnode_recheck2
);
886 return (KMEM_CBRC_DONT_KNOW
);
890 * At this point we know that as long as we hold os->os_lock, the dnode
891 * cannot be freed and fields within the dnode can be safely accessed.
892 * The objset listing this dnode cannot go away as long as this dnode is
896 if (DMU_OBJECT_IS_SPECIAL(odn
->dn_object
)) {
897 mutex_exit(&os
->os_lock
);
898 DNODE_STAT_ADD(dnode_move_stats
.dms_dnode_special
);
899 return (KMEM_CBRC_NO
);
901 ASSERT(odn
->dn_dbuf
!= NULL
); /* only "special" dnodes have no parent */
904 * Lock the dnode handle to prevent the dnode from obtaining any new
905 * holds. This also prevents the descendant dbufs and the bonus dbuf
906 * from accessing the dnode, so that we can discount their holds. The
907 * handle is safe to access because we know that while the dnode cannot
908 * go away, neither can its handle. Once we hold dnh_zrlock, we can
909 * safely move any dnode referenced only by dbufs.
911 if (!zrl_tryenter(&odn
->dn_handle
->dnh_zrlock
)) {
912 mutex_exit(&os
->os_lock
);
913 DNODE_STAT_ADD(dnode_move_stats
.dms_dnode_handle
);
914 return (KMEM_CBRC_LATER
);
918 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
919 * We need to guarantee that there is a hold for every dbuf in order to
920 * determine whether the dnode is actively referenced. Falsely matching
921 * a dbuf to an active hold would lead to an unsafe move. It's possible
922 * that a thread already having an active dnode hold is about to add a
923 * dbuf, and we can't compare hold and dbuf counts while the add is in
926 if (!rw_tryenter(&odn
->dn_struct_rwlock
, RW_WRITER
)) {
927 zrl_exit(&odn
->dn_handle
->dnh_zrlock
);
928 mutex_exit(&os
->os_lock
);
929 DNODE_STAT_ADD(dnode_move_stats
.dms_dnode_rwlock
);
930 return (KMEM_CBRC_LATER
);
934 * A dbuf may be removed (evicted) without an active dnode hold. In that
935 * case, the dbuf count is decremented under the handle lock before the
936 * dbuf's hold is released. This order ensures that if we count the hold
937 * after the dbuf is removed but before its hold is released, we will
938 * treat the unmatched hold as active and exit safely. If we count the
939 * hold before the dbuf is removed, the hold is discounted, and the
940 * removal is blocked until the move completes.
942 refcount
= refcount_count(&odn
->dn_holds
);
943 ASSERT(refcount
>= 0);
944 dbufs
= odn
->dn_dbufs_count
;
946 /* We can't have more dbufs than dnode holds. */
947 ASSERT3U(dbufs
, <=, refcount
);
948 DTRACE_PROBE3(dnode__move
, dnode_t
*, odn
, int64_t, refcount
,
951 if (refcount
> dbufs
) {
952 rw_exit(&odn
->dn_struct_rwlock
);
953 zrl_exit(&odn
->dn_handle
->dnh_zrlock
);
954 mutex_exit(&os
->os_lock
);
955 DNODE_STAT_ADD(dnode_move_stats
.dms_dnode_active
);
956 return (KMEM_CBRC_LATER
);
959 rw_exit(&odn
->dn_struct_rwlock
);
962 * At this point we know that anyone with a hold on the dnode is not
963 * actively referencing it. The dnode is known and in a valid state to
964 * move. We're holding the locks needed to execute the critical section.
966 dnode_move_impl(odn
, ndn
);
968 list_link_replace(&odn
->dn_link
, &ndn
->dn_link
);
969 /* If the dnode was safe to move, the refcount cannot have changed. */
970 ASSERT(refcount
== refcount_count(&ndn
->dn_holds
));
971 ASSERT(dbufs
== ndn
->dn_dbufs_count
);
972 zrl_exit(&ndn
->dn_handle
->dnh_zrlock
); /* handle has moved */
973 mutex_exit(&os
->os_lock
);
975 return (KMEM_CBRC_YES
);
980 dnode_special_close(dnode_handle_t
*dnh
)
982 dnode_t
*dn
= dnh
->dnh_dnode
;
985 * Wait for final references to the dnode to clear. This can
986 * only happen if the arc is asyncronously evicting state that
987 * has a hold on this dnode while we are trying to evict this
990 while (refcount_count(&dn
->dn_holds
) > 0)
992 ASSERT(dn
->dn_dbuf
== NULL
||
993 dmu_buf_get_user(&dn
->dn_dbuf
->db
) == NULL
);
994 zrl_add(&dnh
->dnh_zrlock
);
995 dnode_destroy(dn
); /* implicit zrl_remove() */
996 zrl_destroy(&dnh
->dnh_zrlock
);
997 dnh
->dnh_dnode
= NULL
;
1001 dnode_special_open(objset_t
*os
, dnode_phys_t
*dnp
, uint64_t object
,
1002 dnode_handle_t
*dnh
)
1006 dn
= dnode_create(os
, dnp
, NULL
, object
, dnh
);
1007 zrl_init(&dnh
->dnh_zrlock
);
1012 dnode_buf_evict_async(void *dbu
)
1014 dnode_children_t
*children_dnodes
= dbu
;
1017 for (i
= 0; i
< children_dnodes
->dnc_count
; i
++) {
1018 dnode_handle_t
*dnh
= &children_dnodes
->dnc_children
[i
];
1022 * The dnode handle lock guards against the dnode moving to
1023 * another valid address, so there is no need here to guard
1024 * against changes to or from NULL.
1026 if (dnh
->dnh_dnode
== NULL
) {
1027 zrl_destroy(&dnh
->dnh_zrlock
);
1031 zrl_add(&dnh
->dnh_zrlock
);
1032 dn
= dnh
->dnh_dnode
;
1034 * If there are holds on this dnode, then there should
1035 * be holds on the dnode's containing dbuf as well; thus
1036 * it wouldn't be eligible for eviction and this function
1037 * would not have been called.
1039 ASSERT(refcount_is_zero(&dn
->dn_holds
));
1040 ASSERT(refcount_is_zero(&dn
->dn_tx_holds
));
1042 dnode_destroy(dn
); /* implicit zrl_remove() */
1043 zrl_destroy(&dnh
->dnh_zrlock
);
1044 dnh
->dnh_dnode
= NULL
;
1046 kmem_free(children_dnodes
, sizeof (dnode_children_t
) +
1047 children_dnodes
->dnc_count
* sizeof (dnode_handle_t
));
1052 * EINVAL - invalid object number.
1054 * succeeds even for free dnodes.
1057 dnode_hold_impl(objset_t
*os
, uint64_t object
, int flag
,
1058 void *tag
, dnode_t
**dnp
)
1061 int drop_struct_lock
= FALSE
;
1066 dnode_children_t
*children_dnodes
;
1067 dnode_handle_t
*dnh
;
1070 * If you are holding the spa config lock as writer, you shouldn't
1071 * be asking the DMU to do *anything* unless it's the root pool
1072 * which may require us to read from the root filesystem while
1073 * holding some (not all) of the locks as writer.
1075 ASSERT(spa_config_held(os
->os_spa
, SCL_ALL
, RW_WRITER
) == 0 ||
1076 (spa_is_root(os
->os_spa
) &&
1077 spa_config_held(os
->os_spa
, SCL_STATE
, RW_WRITER
)));
1079 ASSERT((flag
& DNODE_MUST_BE_ALLOCATED
) || (flag
& DNODE_MUST_BE_FREE
));
1081 if (object
== DMU_USERUSED_OBJECT
|| object
== DMU_GROUPUSED_OBJECT
) {
1082 dn
= (object
== DMU_USERUSED_OBJECT
) ?
1083 DMU_USERUSED_DNODE(os
) : DMU_GROUPUSED_DNODE(os
);
1085 return (SET_ERROR(ENOENT
));
1087 if ((flag
& DNODE_MUST_BE_ALLOCATED
) && type
== DMU_OT_NONE
)
1088 return (SET_ERROR(ENOENT
));
1089 if ((flag
& DNODE_MUST_BE_FREE
) && type
!= DMU_OT_NONE
)
1090 return (SET_ERROR(EEXIST
));
1092 (void) refcount_add(&dn
->dn_holds
, tag
);
1097 if (object
== 0 || object
>= DN_MAX_OBJECT
)
1098 return (SET_ERROR(EINVAL
));
1100 mdn
= DMU_META_DNODE(os
);
1101 ASSERT(mdn
->dn_object
== DMU_META_DNODE_OBJECT
);
1105 if (!RW_WRITE_HELD(&mdn
->dn_struct_rwlock
)) {
1106 rw_enter(&mdn
->dn_struct_rwlock
, RW_READER
);
1107 drop_struct_lock
= TRUE
;
1110 blk
= dbuf_whichblock(mdn
, 0, object
* sizeof (dnode_phys_t
));
1112 db
= dbuf_hold(mdn
, blk
, FTAG
);
1113 if (drop_struct_lock
)
1114 rw_exit(&mdn
->dn_struct_rwlock
);
1116 return (SET_ERROR(EIO
));
1117 err
= dbuf_read(db
, NULL
, DB_RF_CANFAIL
);
1119 dbuf_rele(db
, FTAG
);
1123 ASSERT3U(db
->db
.db_size
, >=, 1<<DNODE_SHIFT
);
1124 epb
= db
->db
.db_size
>> DNODE_SHIFT
;
1126 idx
= object
& (epb
-1);
1128 ASSERT(DB_DNODE(db
)->dn_type
== DMU_OT_DNODE
);
1129 children_dnodes
= dmu_buf_get_user(&db
->db
);
1130 if (children_dnodes
== NULL
) {
1132 dnode_children_t
*winner
;
1133 children_dnodes
= kmem_zalloc(sizeof (dnode_children_t
) +
1134 epb
* sizeof (dnode_handle_t
), KM_SLEEP
);
1135 children_dnodes
->dnc_count
= epb
;
1136 dnh
= &children_dnodes
->dnc_children
[0];
1137 for (i
= 0; i
< epb
; i
++) {
1138 zrl_init(&dnh
[i
].dnh_zrlock
);
1140 dmu_buf_init_user(&children_dnodes
->dnc_dbu
, NULL
,
1141 dnode_buf_evict_async
, NULL
);
1142 winner
= dmu_buf_set_user(&db
->db
, &children_dnodes
->dnc_dbu
);
1143 if (winner
!= NULL
) {
1145 for (i
= 0; i
< epb
; i
++) {
1146 zrl_destroy(&dnh
[i
].dnh_zrlock
);
1149 kmem_free(children_dnodes
, sizeof (dnode_children_t
) +
1150 epb
* sizeof (dnode_handle_t
));
1151 children_dnodes
= winner
;
1154 ASSERT(children_dnodes
->dnc_count
== epb
);
1156 dnh
= &children_dnodes
->dnc_children
[idx
];
1157 zrl_add(&dnh
->dnh_zrlock
);
1158 dn
= dnh
->dnh_dnode
;
1160 dnode_phys_t
*phys
= (dnode_phys_t
*)db
->db
.db_data
+idx
;
1162 dn
= dnode_create(os
, phys
, db
, object
, dnh
);
1165 mutex_enter(&dn
->dn_mtx
);
1167 if (dn
->dn_free_txg
||
1168 ((flag
& DNODE_MUST_BE_ALLOCATED
) && type
== DMU_OT_NONE
) ||
1169 ((flag
& DNODE_MUST_BE_FREE
) &&
1170 (type
!= DMU_OT_NONE
|| !refcount_is_zero(&dn
->dn_holds
)))) {
1171 mutex_exit(&dn
->dn_mtx
);
1172 zrl_remove(&dnh
->dnh_zrlock
);
1173 dbuf_rele(db
, FTAG
);
1174 return ((flag
& DNODE_MUST_BE_ALLOCATED
) ? ENOENT
: EEXIST
);
1176 if (refcount_add(&dn
->dn_holds
, tag
) == 1)
1177 dbuf_add_ref(db
, dnh
);
1178 mutex_exit(&dn
->dn_mtx
);
1180 /* Now we can rely on the hold to prevent the dnode from moving. */
1181 zrl_remove(&dnh
->dnh_zrlock
);
1184 ASSERT3P(dn
->dn_dbuf
, ==, db
);
1185 ASSERT3U(dn
->dn_object
, ==, object
);
1186 dbuf_rele(db
, FTAG
);
1193 * Return held dnode if the object is allocated, NULL if not.
1196 dnode_hold(objset_t
*os
, uint64_t object
, void *tag
, dnode_t
**dnp
)
1198 return (dnode_hold_impl(os
, object
, DNODE_MUST_BE_ALLOCATED
, tag
, dnp
));
1202 * Can only add a reference if there is already at least one
1203 * reference on the dnode. Returns FALSE if unable to add a
1207 dnode_add_ref(dnode_t
*dn
, void *tag
)
1209 mutex_enter(&dn
->dn_mtx
);
1210 if (refcount_is_zero(&dn
->dn_holds
)) {
1211 mutex_exit(&dn
->dn_mtx
);
1214 VERIFY(1 < refcount_add(&dn
->dn_holds
, tag
));
1215 mutex_exit(&dn
->dn_mtx
);
1220 dnode_rele(dnode_t
*dn
, void *tag
)
1222 mutex_enter(&dn
->dn_mtx
);
1223 dnode_rele_and_unlock(dn
, tag
);
1227 dnode_rele_and_unlock(dnode_t
*dn
, void *tag
)
1230 /* Get while the hold prevents the dnode from moving. */
1231 dmu_buf_impl_t
*db
= dn
->dn_dbuf
;
1232 dnode_handle_t
*dnh
= dn
->dn_handle
;
1234 refs
= refcount_remove(&dn
->dn_holds
, tag
);
1235 mutex_exit(&dn
->dn_mtx
);
1238 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1239 * indirectly by dbuf_rele() while relying on the dnode handle to
1240 * prevent the dnode from moving, since releasing the last hold could
1241 * result in the dnode's parent dbuf evicting its dnode handles. For
1242 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1243 * other direct or indirect hold on the dnode must first drop the dnode
1246 ASSERT(refs
> 0 || dnh
->dnh_zrlock
.zr_owner
!= curthread
);
1248 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1249 if (refs
== 0 && db
!= NULL
) {
1251 * Another thread could add a hold to the dnode handle in
1252 * dnode_hold_impl() while holding the parent dbuf. Since the
1253 * hold on the parent dbuf prevents the handle from being
1254 * destroyed, the hold on the handle is OK. We can't yet assert
1255 * that the handle has zero references, but that will be
1256 * asserted anyway when the handle gets destroyed.
1263 dnode_setdirty(dnode_t
*dn
, dmu_tx_t
*tx
)
1265 objset_t
*os
= dn
->dn_objset
;
1266 uint64_t txg
= tx
->tx_txg
;
1268 if (DMU_OBJECT_IS_SPECIAL(dn
->dn_object
)) {
1269 dsl_dataset_dirty(os
->os_dsl_dataset
, tx
);
1276 mutex_enter(&dn
->dn_mtx
);
1277 ASSERT(dn
->dn_phys
->dn_type
|| dn
->dn_allocated_txg
);
1278 ASSERT(dn
->dn_free_txg
== 0 || dn
->dn_free_txg
>= txg
);
1279 mutex_exit(&dn
->dn_mtx
);
1283 * Determine old uid/gid when necessary
1285 dmu_objset_userquota_get_ids(dn
, B_TRUE
, tx
);
1287 multilist_t
*dirtylist
= os
->os_dirty_dnodes
[txg
& TXG_MASK
];
1288 multilist_sublist_t
*mls
= multilist_sublist_lock_obj(dirtylist
, dn
);
1291 * If we are already marked dirty, we're done.
1293 if (list_link_active(&dn
->dn_dirty_link
[txg
& TXG_MASK
])) {
1294 multilist_sublist_unlock(mls
);
1298 ASSERT(!refcount_is_zero(&dn
->dn_holds
) ||
1299 !avl_is_empty(&dn
->dn_dbufs
));
1300 ASSERT(dn
->dn_datablksz
!= 0);
1301 ASSERT0(dn
->dn_next_bonuslen
[txg
&TXG_MASK
]);
1302 ASSERT0(dn
->dn_next_blksz
[txg
&TXG_MASK
]);
1303 ASSERT0(dn
->dn_next_bonustype
[txg
&TXG_MASK
]);
1305 dprintf_ds(os
->os_dsl_dataset
, "obj=%llu txg=%llu\n",
1306 dn
->dn_object
, txg
);
1308 multilist_sublist_insert_head(mls
, dn
);
1310 multilist_sublist_unlock(mls
);
1313 * The dnode maintains a hold on its containing dbuf as
1314 * long as there are holds on it. Each instantiated child
1315 * dbuf maintains a hold on the dnode. When the last child
1316 * drops its hold, the dnode will drop its hold on the
1317 * containing dbuf. We add a "dirty hold" here so that the
1318 * dnode will hang around after we finish processing its
1321 VERIFY(dnode_add_ref(dn
, (void *)(uintptr_t)tx
->tx_txg
));
1323 (void) dbuf_dirty(dn
->dn_dbuf
, tx
);
1325 dsl_dataset_dirty(os
->os_dsl_dataset
, tx
);
1329 dnode_free(dnode_t
*dn
, dmu_tx_t
*tx
)
1331 mutex_enter(&dn
->dn_mtx
);
1332 if (dn
->dn_type
== DMU_OT_NONE
|| dn
->dn_free_txg
) {
1333 mutex_exit(&dn
->dn_mtx
);
1336 dn
->dn_free_txg
= tx
->tx_txg
;
1337 mutex_exit(&dn
->dn_mtx
);
1339 dnode_setdirty(dn
, tx
);
1343 * Try to change the block size for the indicated dnode. This can only
1344 * succeed if there are no blocks allocated or dirty beyond first block
1347 dnode_set_blksz(dnode_t
*dn
, uint64_t size
, int ibs
, dmu_tx_t
*tx
)
1352 ASSERT3U(size
, <=, spa_maxblocksize(dmu_objset_spa(dn
->dn_objset
)));
1354 size
= SPA_MINBLOCKSIZE
;
1356 size
= P2ROUNDUP(size
, SPA_MINBLOCKSIZE
);
1358 if (ibs
== dn
->dn_indblkshift
)
1361 if (size
>> SPA_MINBLOCKSHIFT
== dn
->dn_datablkszsec
&& ibs
== 0)
1364 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1366 /* Check for any allocated blocks beyond the first */
1367 if (dn
->dn_maxblkid
!= 0)
1370 mutex_enter(&dn
->dn_dbufs_mtx
);
1371 for (db
= avl_first(&dn
->dn_dbufs
); db
!= NULL
;
1372 db
= AVL_NEXT(&dn
->dn_dbufs
, db
)) {
1373 if (db
->db_blkid
!= 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1374 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1375 mutex_exit(&dn
->dn_dbufs_mtx
);
1379 mutex_exit(&dn
->dn_dbufs_mtx
);
1381 if (ibs
&& dn
->dn_nlevels
!= 1)
1384 /* resize the old block */
1385 err
= dbuf_hold_impl(dn
, 0, 0, TRUE
, FALSE
, FTAG
, &db
);
1387 dbuf_new_size(db
, size
, tx
);
1388 else if (err
!= ENOENT
)
1391 dnode_setdblksz(dn
, size
);
1392 dnode_setdirty(dn
, tx
);
1393 dn
->dn_next_blksz
[tx
->tx_txg
&TXG_MASK
] = size
;
1395 dn
->dn_indblkshift
= ibs
;
1396 dn
->dn_next_indblkshift
[tx
->tx_txg
&TXG_MASK
] = ibs
;
1398 /* rele after we have fixed the blocksize in the dnode */
1400 dbuf_rele(db
, FTAG
);
1402 rw_exit(&dn
->dn_struct_rwlock
);
1406 rw_exit(&dn
->dn_struct_rwlock
);
1407 return (SET_ERROR(ENOTSUP
));
1410 /* read-holding callers must not rely on the lock being continuously held */
1412 dnode_new_blkid(dnode_t
*dn
, uint64_t blkid
, dmu_tx_t
*tx
, boolean_t have_read
)
1414 uint64_t txgoff
= tx
->tx_txg
& TXG_MASK
;
1415 int epbs
, new_nlevels
;
1418 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1421 RW_READ_HELD(&dn
->dn_struct_rwlock
) :
1422 RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1425 * if we have a read-lock, check to see if we need to do any work
1426 * before upgrading to a write-lock.
1429 if (blkid
<= dn
->dn_maxblkid
)
1432 if (!rw_tryupgrade(&dn
->dn_struct_rwlock
)) {
1433 rw_exit(&dn
->dn_struct_rwlock
);
1434 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1438 if (blkid
<= dn
->dn_maxblkid
)
1441 dn
->dn_maxblkid
= blkid
;
1444 * Compute the number of levels necessary to support the new maxblkid.
1447 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1448 for (sz
= dn
->dn_nblkptr
;
1449 sz
<= blkid
&& sz
>= dn
->dn_nblkptr
; sz
<<= epbs
)
1452 if (new_nlevels
> dn
->dn_nlevels
) {
1453 int old_nlevels
= dn
->dn_nlevels
;
1456 dbuf_dirty_record_t
*new, *dr
, *dr_next
;
1458 dn
->dn_nlevels
= new_nlevels
;
1460 ASSERT3U(new_nlevels
, >, dn
->dn_next_nlevels
[txgoff
]);
1461 dn
->dn_next_nlevels
[txgoff
] = new_nlevels
;
1463 /* dirty the left indirects */
1464 db
= dbuf_hold_level(dn
, old_nlevels
, 0, FTAG
);
1466 new = dbuf_dirty(db
, tx
);
1467 dbuf_rele(db
, FTAG
);
1469 /* transfer the dirty records to the new indirect */
1470 mutex_enter(&dn
->dn_mtx
);
1471 mutex_enter(&new->dt
.di
.dr_mtx
);
1472 list
= &dn
->dn_dirty_records
[txgoff
];
1473 for (dr
= list_head(list
); dr
; dr
= dr_next
) {
1474 dr_next
= list_next(&dn
->dn_dirty_records
[txgoff
], dr
);
1475 if (dr
->dr_dbuf
->db_level
!= new_nlevels
-1 &&
1476 dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
1477 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
1478 ASSERT(dr
->dr_dbuf
->db_level
== old_nlevels
-1);
1479 list_remove(&dn
->dn_dirty_records
[txgoff
], dr
);
1480 list_insert_tail(&new->dt
.di
.dr_children
, dr
);
1481 dr
->dr_parent
= new;
1484 mutex_exit(&new->dt
.di
.dr_mtx
);
1485 mutex_exit(&dn
->dn_mtx
);
1490 rw_downgrade(&dn
->dn_struct_rwlock
);
1494 dnode_dirty_l1(dnode_t
*dn
, uint64_t l1blkid
, dmu_tx_t
*tx
)
1496 dmu_buf_impl_t
*db
= dbuf_hold_level(dn
, 1, l1blkid
, FTAG
);
1498 dmu_buf_will_dirty(&db
->db
, tx
);
1499 dbuf_rele(db
, FTAG
);
1504 dnode_free_range(dnode_t
*dn
, uint64_t off
, uint64_t len
, dmu_tx_t
*tx
)
1507 uint64_t blkoff
, blkid
, nblks
;
1508 int blksz
, blkshift
, head
, tail
;
1512 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1513 blksz
= dn
->dn_datablksz
;
1514 blkshift
= dn
->dn_datablkshift
;
1515 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1517 if (len
== DMU_OBJECT_END
) {
1518 len
= UINT64_MAX
- off
;
1523 * First, block align the region to free:
1526 head
= P2NPHASE(off
, blksz
);
1527 blkoff
= P2PHASE(off
, blksz
);
1528 if ((off
>> blkshift
) > dn
->dn_maxblkid
)
1531 ASSERT(dn
->dn_maxblkid
== 0);
1532 if (off
== 0 && len
>= blksz
) {
1534 * Freeing the whole block; fast-track this request.
1535 * Note that we won't dirty any indirect blocks,
1536 * which is fine because we will be freeing the entire
1537 * file and thus all indirect blocks will be freed
1538 * by free_children().
1543 } else if (off
>= blksz
) {
1544 /* Freeing past end-of-data */
1547 /* Freeing part of the block. */
1549 ASSERT3U(head
, >, 0);
1553 /* zero out any partial block data at the start of the range */
1555 ASSERT3U(blkoff
+ head
, ==, blksz
);
1558 if (dbuf_hold_impl(dn
, 0, dbuf_whichblock(dn
, 0, off
),
1559 TRUE
, FALSE
, FTAG
, &db
) == 0) {
1562 /* don't dirty if it isn't on disk and isn't dirty */
1563 if (db
->db_last_dirty
||
1564 (db
->db_blkptr
&& !BP_IS_HOLE(db
->db_blkptr
))) {
1565 rw_exit(&dn
->dn_struct_rwlock
);
1566 dmu_buf_will_dirty(&db
->db
, tx
);
1567 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1568 data
= db
->db
.db_data
;
1569 bzero(data
+ blkoff
, head
);
1571 dbuf_rele(db
, FTAG
);
1577 /* If the range was less than one block, we're done */
1581 /* If the remaining range is past end of file, we're done */
1582 if ((off
>> blkshift
) > dn
->dn_maxblkid
)
1585 ASSERT(ISP2(blksz
));
1589 tail
= P2PHASE(len
, blksz
);
1591 ASSERT0(P2PHASE(off
, blksz
));
1592 /* zero out any partial block data at the end of the range */
1596 if (dbuf_hold_impl(dn
, 0, dbuf_whichblock(dn
, 0, off
+len
),
1597 TRUE
, FALSE
, FTAG
, &db
) == 0) {
1598 /* don't dirty if not on disk and not dirty */
1599 if (db
->db_last_dirty
||
1600 (db
->db_blkptr
&& !BP_IS_HOLE(db
->db_blkptr
))) {
1601 rw_exit(&dn
->dn_struct_rwlock
);
1602 dmu_buf_will_dirty(&db
->db
, tx
);
1603 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1604 bzero(db
->db
.db_data
, tail
);
1606 dbuf_rele(db
, FTAG
);
1611 /* If the range did not include a full block, we are done */
1615 ASSERT(IS_P2ALIGNED(off
, blksz
));
1616 ASSERT(trunc
|| IS_P2ALIGNED(len
, blksz
));
1617 blkid
= off
>> blkshift
;
1618 nblks
= len
>> blkshift
;
1623 * Dirty all the indirect blocks in this range. Note that only
1624 * the first and last indirect blocks can actually be written
1625 * (if they were partially freed) -- they must be dirtied, even if
1626 * they do not exist on disk yet. The interior blocks will
1627 * be freed by free_children(), so they will not actually be written.
1628 * Even though these interior blocks will not be written, we
1629 * dirty them for two reasons:
1631 * - It ensures that the indirect blocks remain in memory until
1632 * syncing context. (They have already been prefetched by
1633 * dmu_tx_hold_free(), so we don't have to worry about reading
1634 * them serially here.)
1636 * - The dirty space accounting will put pressure on the txg sync
1637 * mechanism to begin syncing, and to delay transactions if there
1638 * is a large amount of freeing. Even though these indirect
1639 * blocks will not be written, we could need to write the same
1640 * amount of space if we copy the freed BPs into deadlists.
1642 if (dn
->dn_nlevels
> 1) {
1643 uint64_t first
, last
;
1645 first
= blkid
>> epbs
;
1646 dnode_dirty_l1(dn
, first
, tx
);
1648 last
= dn
->dn_maxblkid
>> epbs
;
1650 last
= (blkid
+ nblks
- 1) >> epbs
;
1652 dnode_dirty_l1(dn
, last
, tx
);
1654 int shift
= dn
->dn_datablkshift
+ dn
->dn_indblkshift
-
1656 for (uint64_t i
= first
+ 1; i
< last
; i
++) {
1658 * Set i to the blockid of the next non-hole
1659 * level-1 indirect block at or after i. Note
1660 * that dnode_next_offset() operates in terms of
1661 * level-0-equivalent bytes.
1663 uint64_t ibyte
= i
<< shift
;
1664 int err
= dnode_next_offset(dn
, DNODE_FIND_HAVELOCK
,
1671 * Normally we should not see an error, either
1672 * from dnode_next_offset() or dbuf_hold_level()
1673 * (except for ESRCH from dnode_next_offset).
1674 * If there is an i/o error, then when we read
1675 * this block in syncing context, it will use
1676 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1677 * to the "failmode" property. dnode_next_offset()
1678 * doesn't have a flag to indicate MUSTSUCCEED.
1683 dnode_dirty_l1(dn
, i
, tx
);
1689 * Add this range to the dnode range list.
1690 * We will finish up this free operation in the syncing phase.
1692 mutex_enter(&dn
->dn_mtx
);
1693 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1694 if (dn
->dn_free_ranges
[txgoff
] == NULL
) {
1695 dn
->dn_free_ranges
[txgoff
] = range_tree_create(NULL
, NULL
);
1697 range_tree_clear(dn
->dn_free_ranges
[txgoff
], blkid
, nblks
);
1698 range_tree_add(dn
->dn_free_ranges
[txgoff
], blkid
, nblks
);
1699 dprintf_dnode(dn
, "blkid=%llu nblks=%llu txg=%llu\n",
1700 blkid
, nblks
, tx
->tx_txg
);
1701 mutex_exit(&dn
->dn_mtx
);
1703 dbuf_free_range(dn
, blkid
, blkid
+ nblks
- 1, tx
);
1704 dnode_setdirty(dn
, tx
);
1707 rw_exit(&dn
->dn_struct_rwlock
);
1711 dnode_spill_freed(dnode_t
*dn
)
1715 mutex_enter(&dn
->dn_mtx
);
1716 for (i
= 0; i
< TXG_SIZE
; i
++) {
1717 if (dn
->dn_rm_spillblk
[i
] == DN_KILL_SPILLBLK
)
1720 mutex_exit(&dn
->dn_mtx
);
1721 return (i
< TXG_SIZE
);
1724 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1726 dnode_block_freed(dnode_t
*dn
, uint64_t blkid
)
1728 void *dp
= spa_get_dsl(dn
->dn_objset
->os_spa
);
1731 if (blkid
== DMU_BONUS_BLKID
)
1735 * If we're in the process of opening the pool, dp will not be
1736 * set yet, but there shouldn't be anything dirty.
1741 if (dn
->dn_free_txg
)
1744 if (blkid
== DMU_SPILL_BLKID
)
1745 return (dnode_spill_freed(dn
));
1747 mutex_enter(&dn
->dn_mtx
);
1748 for (i
= 0; i
< TXG_SIZE
; i
++) {
1749 if (dn
->dn_free_ranges
[i
] != NULL
&&
1750 range_tree_contains(dn
->dn_free_ranges
[i
], blkid
, 1))
1753 mutex_exit(&dn
->dn_mtx
);
1754 return (i
< TXG_SIZE
);
1757 /* call from syncing context when we actually write/free space for this dnode */
1759 dnode_diduse_space(dnode_t
*dn
, int64_t delta
)
1762 dprintf_dnode(dn
, "dn=%p dnp=%p used=%llu delta=%lld\n",
1764 (u_longlong_t
)dn
->dn_phys
->dn_used
,
1767 mutex_enter(&dn
->dn_mtx
);
1768 space
= DN_USED_BYTES(dn
->dn_phys
);
1770 ASSERT3U(space
+ delta
, >=, space
); /* no overflow */
1772 ASSERT3U(space
, >=, -delta
); /* no underflow */
1775 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_DNODE_BYTES
) {
1776 ASSERT((dn
->dn_phys
->dn_flags
& DNODE_FLAG_USED_BYTES
) == 0);
1777 ASSERT0(P2PHASE(space
, 1<<DEV_BSHIFT
));
1778 dn
->dn_phys
->dn_used
= space
>> DEV_BSHIFT
;
1780 dn
->dn_phys
->dn_used
= space
;
1781 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_USED_BYTES
;
1783 mutex_exit(&dn
->dn_mtx
);
1787 * Scans a block at the indicated "level" looking for a hole or data,
1788 * depending on 'flags'.
1790 * If level > 0, then we are scanning an indirect block looking at its
1791 * pointers. If level == 0, then we are looking at a block of dnodes.
1793 * If we don't find what we are looking for in the block, we return ESRCH.
1794 * Otherwise, return with *offset pointing to the beginning (if searching
1795 * forwards) or end (if searching backwards) of the range covered by the
1796 * block pointer we matched on (or dnode).
1798 * The basic search algorithm used below by dnode_next_offset() is to
1799 * use this function to search up the block tree (widen the search) until
1800 * we find something (i.e., we don't return ESRCH) and then search back
1801 * down the tree (narrow the search) until we reach our original search
1805 dnode_next_offset_level(dnode_t
*dn
, int flags
, uint64_t *offset
,
1806 int lvl
, uint64_t blkfill
, uint64_t txg
)
1808 dmu_buf_impl_t
*db
= NULL
;
1810 uint64_t epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1811 uint64_t epb
= 1ULL << epbs
;
1812 uint64_t minfill
, maxfill
;
1814 int i
, inc
, error
, span
;
1816 dprintf("probing object %llu offset %llx level %d of %u\n",
1817 dn
->dn_object
, *offset
, lvl
, dn
->dn_phys
->dn_nlevels
);
1819 hole
= ((flags
& DNODE_FIND_HOLE
) != 0);
1820 inc
= (flags
& DNODE_FIND_BACKWARDS
) ? -1 : 1;
1821 ASSERT(txg
== 0 || !hole
);
1823 if (lvl
== dn
->dn_phys
->dn_nlevels
) {
1825 epb
= dn
->dn_phys
->dn_nblkptr
;
1826 data
= dn
->dn_phys
->dn_blkptr
;
1828 uint64_t blkid
= dbuf_whichblock(dn
, lvl
, *offset
);
1829 error
= dbuf_hold_impl(dn
, lvl
, blkid
, TRUE
, FALSE
, FTAG
, &db
);
1831 if (error
!= ENOENT
)
1836 * This can only happen when we are searching up
1837 * the block tree for data. We don't really need to
1838 * adjust the offset, as we will just end up looking
1839 * at the pointer to this block in its parent, and its
1840 * going to be unallocated, so we will skip over it.
1842 return (SET_ERROR(ESRCH
));
1844 error
= dbuf_read(db
, NULL
, DB_RF_CANFAIL
| DB_RF_HAVESTRUCT
);
1846 dbuf_rele(db
, FTAG
);
1849 data
= db
->db
.db_data
;
1853 if (db
!= NULL
&& txg
!= 0 && (db
->db_blkptr
== NULL
||
1854 db
->db_blkptr
->blk_birth
<= txg
||
1855 BP_IS_HOLE(db
->db_blkptr
))) {
1857 * This can only happen when we are searching up the tree
1858 * and these conditions mean that we need to keep climbing.
1860 error
= SET_ERROR(ESRCH
);
1861 } else if (lvl
== 0) {
1862 dnode_phys_t
*dnp
= data
;
1864 ASSERT(dn
->dn_type
== DMU_OT_DNODE
);
1866 for (i
= (*offset
>> span
) & (blkfill
- 1);
1867 i
>= 0 && i
< blkfill
; i
+= inc
) {
1868 if ((dnp
[i
].dn_type
== DMU_OT_NONE
) == hole
)
1870 *offset
+= (1ULL << span
) * inc
;
1872 if (i
< 0 || i
== blkfill
)
1873 error
= SET_ERROR(ESRCH
);
1875 blkptr_t
*bp
= data
;
1876 uint64_t start
= *offset
;
1877 span
= (lvl
- 1) * epbs
+ dn
->dn_datablkshift
;
1879 maxfill
= blkfill
<< ((lvl
- 1) * epbs
);
1886 *offset
= *offset
>> span
;
1887 for (i
= BF64_GET(*offset
, 0, epbs
);
1888 i
>= 0 && i
< epb
; i
+= inc
) {
1889 if (BP_GET_FILL(&bp
[i
]) >= minfill
&&
1890 BP_GET_FILL(&bp
[i
]) <= maxfill
&&
1891 (hole
|| bp
[i
].blk_birth
> txg
))
1893 if (inc
> 0 || *offset
> 0)
1896 *offset
= *offset
<< span
;
1898 /* traversing backwards; position offset at the end */
1899 ASSERT3U(*offset
, <=, start
);
1900 *offset
= MIN(*offset
+ (1ULL << span
) - 1, start
);
1901 } else if (*offset
< start
) {
1904 if (i
< 0 || i
>= epb
)
1905 error
= SET_ERROR(ESRCH
);
1909 dbuf_rele(db
, FTAG
);
1915 * Find the next hole, data, or sparse region at or after *offset.
1916 * The value 'blkfill' tells us how many items we expect to find
1917 * in an L0 data block; this value is 1 for normal objects,
1918 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1919 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1923 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1924 * Finds the next/previous hole/data in a file.
1925 * Used in dmu_offset_next().
1927 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1928 * Finds the next free/allocated dnode an objset's meta-dnode.
1929 * Only finds objects that have new contents since txg (ie.
1930 * bonus buffer changes and content removal are ignored).
1931 * Used in dmu_object_next().
1933 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1934 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1935 * Used in dmu_object_alloc().
1938 dnode_next_offset(dnode_t
*dn
, int flags
, uint64_t *offset
,
1939 int minlvl
, uint64_t blkfill
, uint64_t txg
)
1941 uint64_t initial_offset
= *offset
;
1945 if (!(flags
& DNODE_FIND_HAVELOCK
))
1946 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1948 if (dn
->dn_phys
->dn_nlevels
== 0) {
1949 error
= SET_ERROR(ESRCH
);
1953 if (dn
->dn_datablkshift
== 0) {
1954 if (*offset
< dn
->dn_datablksz
) {
1955 if (flags
& DNODE_FIND_HOLE
)
1956 *offset
= dn
->dn_datablksz
;
1958 error
= SET_ERROR(ESRCH
);
1963 maxlvl
= dn
->dn_phys
->dn_nlevels
;
1965 for (lvl
= minlvl
; lvl
<= maxlvl
; lvl
++) {
1966 error
= dnode_next_offset_level(dn
,
1967 flags
, offset
, lvl
, blkfill
, txg
);
1972 while (error
== 0 && --lvl
>= minlvl
) {
1973 error
= dnode_next_offset_level(dn
,
1974 flags
, offset
, lvl
, blkfill
, txg
);
1978 * There's always a "virtual hole" at the end of the object, even
1979 * if all BP's which physically exist are non-holes.
1981 if ((flags
& DNODE_FIND_HOLE
) && error
== ESRCH
&& txg
== 0 &&
1982 minlvl
== 1 && blkfill
== 1 && !(flags
& DNODE_FIND_BACKWARDS
)) {
1986 if (error
== 0 && (flags
& DNODE_FIND_BACKWARDS
?
1987 initial_offset
< *offset
: initial_offset
> *offset
))
1988 error
= SET_ERROR(ESRCH
);
1990 if (!(flags
& DNODE_FIND_HAVELOCK
))
1991 rw_exit(&dn
->dn_struct_rwlock
);