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.
27 #include <sys/zfs_context.h>
29 #include <sys/dnode.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_dataset.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/range_tree.h>
40 #include <sys/trace_dnode.h>
41 #include <sys/zfs_project.h>
43 dnode_stats_t dnode_stats
= {
44 { "dnode_hold_dbuf_hold", KSTAT_DATA_UINT64
},
45 { "dnode_hold_dbuf_read", KSTAT_DATA_UINT64
},
46 { "dnode_hold_alloc_hits", KSTAT_DATA_UINT64
},
47 { "dnode_hold_alloc_misses", KSTAT_DATA_UINT64
},
48 { "dnode_hold_alloc_interior", KSTAT_DATA_UINT64
},
49 { "dnode_hold_alloc_lock_retry", KSTAT_DATA_UINT64
},
50 { "dnode_hold_alloc_lock_misses", KSTAT_DATA_UINT64
},
51 { "dnode_hold_alloc_type_none", KSTAT_DATA_UINT64
},
52 { "dnode_hold_free_hits", KSTAT_DATA_UINT64
},
53 { "dnode_hold_free_misses", KSTAT_DATA_UINT64
},
54 { "dnode_hold_free_lock_misses", KSTAT_DATA_UINT64
},
55 { "dnode_hold_free_lock_retry", KSTAT_DATA_UINT64
},
56 { "dnode_hold_free_overflow", KSTAT_DATA_UINT64
},
57 { "dnode_hold_free_refcount", KSTAT_DATA_UINT64
},
58 { "dnode_hold_free_txg", KSTAT_DATA_UINT64
},
59 { "dnode_free_interior_lock_retry", KSTAT_DATA_UINT64
},
60 { "dnode_allocate", KSTAT_DATA_UINT64
},
61 { "dnode_reallocate", KSTAT_DATA_UINT64
},
62 { "dnode_buf_evict", KSTAT_DATA_UINT64
},
63 { "dnode_alloc_next_chunk", KSTAT_DATA_UINT64
},
64 { "dnode_alloc_race", KSTAT_DATA_UINT64
},
65 { "dnode_alloc_next_block", KSTAT_DATA_UINT64
},
66 { "dnode_move_invalid", KSTAT_DATA_UINT64
},
67 { "dnode_move_recheck1", KSTAT_DATA_UINT64
},
68 { "dnode_move_recheck2", KSTAT_DATA_UINT64
},
69 { "dnode_move_special", KSTAT_DATA_UINT64
},
70 { "dnode_move_handle", KSTAT_DATA_UINT64
},
71 { "dnode_move_rwlock", KSTAT_DATA_UINT64
},
72 { "dnode_move_active", KSTAT_DATA_UINT64
},
75 static kstat_t
*dnode_ksp
;
76 static kmem_cache_t
*dnode_cache
;
78 ASSERTV(static dnode_phys_t dnode_phys_zero
);
80 int zfs_default_bs
= SPA_MINBLOCKSHIFT
;
81 int zfs_default_ibs
= DN_MAX_INDBLKSHIFT
;
84 static kmem_cbrc_t
dnode_move(void *, void *, size_t, void *);
88 dbuf_compare(const void *x1
, const void *x2
)
90 const dmu_buf_impl_t
*d1
= x1
;
91 const dmu_buf_impl_t
*d2
= x2
;
93 int cmp
= AVL_CMP(d1
->db_level
, d2
->db_level
);
97 cmp
= AVL_CMP(d1
->db_blkid
, d2
->db_blkid
);
101 if (d1
->db_state
== DB_SEARCH
) {
102 ASSERT3S(d2
->db_state
, !=, DB_SEARCH
);
104 } else if (d2
->db_state
== DB_SEARCH
) {
105 ASSERT3S(d1
->db_state
, !=, DB_SEARCH
);
109 return (AVL_PCMP(d1
, d2
));
114 dnode_cons(void *arg
, void *unused
, int kmflag
)
119 rw_init(&dn
->dn_struct_rwlock
, NULL
, RW_NOLOCKDEP
, NULL
);
120 mutex_init(&dn
->dn_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
121 mutex_init(&dn
->dn_dbufs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
122 cv_init(&dn
->dn_notxholds
, NULL
, CV_DEFAULT
, NULL
);
125 * Every dbuf has a reference, and dropping a tracked reference is
126 * O(number of references), so don't track dn_holds.
128 zfs_refcount_create_untracked(&dn
->dn_holds
);
129 zfs_refcount_create(&dn
->dn_tx_holds
);
130 list_link_init(&dn
->dn_link
);
132 bzero(&dn
->dn_next_nblkptr
[0], sizeof (dn
->dn_next_nblkptr
));
133 bzero(&dn
->dn_next_nlevels
[0], sizeof (dn
->dn_next_nlevels
));
134 bzero(&dn
->dn_next_indblkshift
[0], sizeof (dn
->dn_next_indblkshift
));
135 bzero(&dn
->dn_next_bonustype
[0], sizeof (dn
->dn_next_bonustype
));
136 bzero(&dn
->dn_rm_spillblk
[0], sizeof (dn
->dn_rm_spillblk
));
137 bzero(&dn
->dn_next_bonuslen
[0], sizeof (dn
->dn_next_bonuslen
));
138 bzero(&dn
->dn_next_blksz
[0], sizeof (dn
->dn_next_blksz
));
139 bzero(&dn
->dn_next_maxblkid
[0], sizeof (dn
->dn_next_maxblkid
));
141 for (i
= 0; i
< TXG_SIZE
; i
++) {
142 multilist_link_init(&dn
->dn_dirty_link
[i
]);
143 dn
->dn_free_ranges
[i
] = NULL
;
144 list_create(&dn
->dn_dirty_records
[i
],
145 sizeof (dbuf_dirty_record_t
),
146 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
149 dn
->dn_allocated_txg
= 0;
151 dn
->dn_assigned_txg
= 0;
152 dn
->dn_dirty_txg
= 0;
154 dn
->dn_dirtyctx_firstset
= NULL
;
156 dn
->dn_have_spill
= B_FALSE
;
162 dn
->dn_oldprojid
= ZFS_DEFAULT_PROJID
;
165 dn
->dn_newprojid
= ZFS_DEFAULT_PROJID
;
168 dn
->dn_dbufs_count
= 0;
169 avl_create(&dn
->dn_dbufs
, dbuf_compare
, sizeof (dmu_buf_impl_t
),
170 offsetof(dmu_buf_impl_t
, db_link
));
178 dnode_dest(void *arg
, void *unused
)
183 rw_destroy(&dn
->dn_struct_rwlock
);
184 mutex_destroy(&dn
->dn_mtx
);
185 mutex_destroy(&dn
->dn_dbufs_mtx
);
186 cv_destroy(&dn
->dn_notxholds
);
187 zfs_refcount_destroy(&dn
->dn_holds
);
188 zfs_refcount_destroy(&dn
->dn_tx_holds
);
189 ASSERT(!list_link_active(&dn
->dn_link
));
191 for (i
= 0; i
< TXG_SIZE
; i
++) {
192 ASSERT(!multilist_link_active(&dn
->dn_dirty_link
[i
]));
193 ASSERT3P(dn
->dn_free_ranges
[i
], ==, NULL
);
194 list_destroy(&dn
->dn_dirty_records
[i
]);
195 ASSERT0(dn
->dn_next_nblkptr
[i
]);
196 ASSERT0(dn
->dn_next_nlevels
[i
]);
197 ASSERT0(dn
->dn_next_indblkshift
[i
]);
198 ASSERT0(dn
->dn_next_bonustype
[i
]);
199 ASSERT0(dn
->dn_rm_spillblk
[i
]);
200 ASSERT0(dn
->dn_next_bonuslen
[i
]);
201 ASSERT0(dn
->dn_next_blksz
[i
]);
202 ASSERT0(dn
->dn_next_maxblkid
[i
]);
205 ASSERT0(dn
->dn_allocated_txg
);
206 ASSERT0(dn
->dn_free_txg
);
207 ASSERT0(dn
->dn_assigned_txg
);
208 ASSERT0(dn
->dn_dirty_txg
);
209 ASSERT0(dn
->dn_dirtyctx
);
210 ASSERT3P(dn
->dn_dirtyctx_firstset
, ==, NULL
);
211 ASSERT3P(dn
->dn_bonus
, ==, NULL
);
212 ASSERT(!dn
->dn_have_spill
);
213 ASSERT3P(dn
->dn_zio
, ==, NULL
);
214 ASSERT0(dn
->dn_oldused
);
215 ASSERT0(dn
->dn_oldflags
);
216 ASSERT0(dn
->dn_olduid
);
217 ASSERT0(dn
->dn_oldgid
);
218 ASSERT0(dn
->dn_oldprojid
);
219 ASSERT0(dn
->dn_newuid
);
220 ASSERT0(dn
->dn_newgid
);
221 ASSERT0(dn
->dn_newprojid
);
222 ASSERT0(dn
->dn_id_flags
);
224 ASSERT0(dn
->dn_dbufs_count
);
225 avl_destroy(&dn
->dn_dbufs
);
231 ASSERT(dnode_cache
== NULL
);
232 dnode_cache
= kmem_cache_create("dnode_t", sizeof (dnode_t
),
233 0, dnode_cons
, dnode_dest
, NULL
, NULL
, NULL
, 0);
234 kmem_cache_set_move(dnode_cache
, dnode_move
);
236 dnode_ksp
= kstat_create("zfs", 0, "dnodestats", "misc",
237 KSTAT_TYPE_NAMED
, sizeof (dnode_stats
) / sizeof (kstat_named_t
),
239 if (dnode_ksp
!= NULL
) {
240 dnode_ksp
->ks_data
= &dnode_stats
;
241 kstat_install(dnode_ksp
);
248 if (dnode_ksp
!= NULL
) {
249 kstat_delete(dnode_ksp
);
253 kmem_cache_destroy(dnode_cache
);
260 dnode_verify(dnode_t
*dn
)
262 int drop_struct_lock
= FALSE
;
265 ASSERT(dn
->dn_objset
);
266 ASSERT(dn
->dn_handle
->dnh_dnode
== dn
);
268 ASSERT(DMU_OT_IS_VALID(dn
->dn_phys
->dn_type
));
270 if (!(zfs_flags
& ZFS_DEBUG_DNODE_VERIFY
))
273 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
274 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
275 drop_struct_lock
= TRUE
;
277 if (dn
->dn_phys
->dn_type
!= DMU_OT_NONE
|| dn
->dn_allocated_txg
!= 0) {
279 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
280 ASSERT3U(dn
->dn_indblkshift
, <=, SPA_MAXBLOCKSHIFT
);
281 if (dn
->dn_datablkshift
) {
282 ASSERT3U(dn
->dn_datablkshift
, >=, SPA_MINBLOCKSHIFT
);
283 ASSERT3U(dn
->dn_datablkshift
, <=, SPA_MAXBLOCKSHIFT
);
284 ASSERT3U(1<<dn
->dn_datablkshift
, ==, dn
->dn_datablksz
);
286 ASSERT3U(dn
->dn_nlevels
, <=, 30);
287 ASSERT(DMU_OT_IS_VALID(dn
->dn_type
));
288 ASSERT3U(dn
->dn_nblkptr
, >=, 1);
289 ASSERT3U(dn
->dn_nblkptr
, <=, DN_MAX_NBLKPTR
);
290 ASSERT3U(dn
->dn_bonuslen
, <=, max_bonuslen
);
291 ASSERT3U(dn
->dn_datablksz
, ==,
292 dn
->dn_datablkszsec
<< SPA_MINBLOCKSHIFT
);
293 ASSERT3U(ISP2(dn
->dn_datablksz
), ==, dn
->dn_datablkshift
!= 0);
294 ASSERT3U((dn
->dn_nblkptr
- 1) * sizeof (blkptr_t
) +
295 dn
->dn_bonuslen
, <=, max_bonuslen
);
296 for (i
= 0; i
< TXG_SIZE
; i
++) {
297 ASSERT3U(dn
->dn_next_nlevels
[i
], <=, dn
->dn_nlevels
);
300 if (dn
->dn_phys
->dn_type
!= DMU_OT_NONE
)
301 ASSERT3U(dn
->dn_phys
->dn_nlevels
, <=, dn
->dn_nlevels
);
302 ASSERT(DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) || dn
->dn_dbuf
!= NULL
);
303 if (dn
->dn_dbuf
!= NULL
) {
304 ASSERT3P(dn
->dn_phys
, ==,
305 (dnode_phys_t
*)dn
->dn_dbuf
->db
.db_data
+
306 (dn
->dn_object
% (dn
->dn_dbuf
->db
.db_size
>> DNODE_SHIFT
)));
308 if (drop_struct_lock
)
309 rw_exit(&dn
->dn_struct_rwlock
);
314 dnode_byteswap(dnode_phys_t
*dnp
)
316 uint64_t *buf64
= (void*)&dnp
->dn_blkptr
;
319 if (dnp
->dn_type
== DMU_OT_NONE
) {
320 bzero(dnp
, sizeof (dnode_phys_t
));
324 dnp
->dn_datablkszsec
= BSWAP_16(dnp
->dn_datablkszsec
);
325 dnp
->dn_bonuslen
= BSWAP_16(dnp
->dn_bonuslen
);
326 dnp
->dn_extra_slots
= BSWAP_8(dnp
->dn_extra_slots
);
327 dnp
->dn_maxblkid
= BSWAP_64(dnp
->dn_maxblkid
);
328 dnp
->dn_used
= BSWAP_64(dnp
->dn_used
);
331 * dn_nblkptr is only one byte, so it's OK to read it in either
332 * byte order. We can't read dn_bouslen.
334 ASSERT(dnp
->dn_indblkshift
<= SPA_MAXBLOCKSHIFT
);
335 ASSERT(dnp
->dn_nblkptr
<= DN_MAX_NBLKPTR
);
336 for (i
= 0; i
< dnp
->dn_nblkptr
* sizeof (blkptr_t
)/8; i
++)
337 buf64
[i
] = BSWAP_64(buf64
[i
]);
340 * OK to check dn_bonuslen for zero, because it won't matter if
341 * we have the wrong byte order. This is necessary because the
342 * dnode dnode is smaller than a regular dnode.
344 if (dnp
->dn_bonuslen
!= 0) {
346 * Note that the bonus length calculated here may be
347 * longer than the actual bonus buffer. This is because
348 * we always put the bonus buffer after the last block
349 * pointer (instead of packing it against the end of the
352 int off
= (dnp
->dn_nblkptr
-1) * sizeof (blkptr_t
);
353 int slots
= dnp
->dn_extra_slots
+ 1;
354 size_t len
= DN_SLOTS_TO_BONUSLEN(slots
) - off
;
355 dmu_object_byteswap_t byteswap
;
356 ASSERT(DMU_OT_IS_VALID(dnp
->dn_bonustype
));
357 byteswap
= DMU_OT_BYTESWAP(dnp
->dn_bonustype
);
358 dmu_ot_byteswap
[byteswap
].ob_func(dnp
->dn_bonus
+ off
, len
);
361 /* Swap SPILL block if we have one */
362 if (dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)
363 byteswap_uint64_array(DN_SPILL_BLKPTR(dnp
), sizeof (blkptr_t
));
367 dnode_buf_byteswap(void *vbuf
, size_t size
)
371 ASSERT3U(sizeof (dnode_phys_t
), ==, (1<<DNODE_SHIFT
));
372 ASSERT((size
& (sizeof (dnode_phys_t
)-1)) == 0);
375 dnode_phys_t
*dnp
= (void *)(((char *)vbuf
) + i
);
379 if (dnp
->dn_type
!= DMU_OT_NONE
)
380 i
+= dnp
->dn_extra_slots
* DNODE_MIN_SIZE
;
385 dnode_setbonuslen(dnode_t
*dn
, int newsize
, dmu_tx_t
*tx
)
387 ASSERT3U(zfs_refcount_count(&dn
->dn_holds
), >=, 1);
389 dnode_setdirty(dn
, tx
);
390 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
391 ASSERT3U(newsize
, <=, DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
392 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
));
393 dn
->dn_bonuslen
= newsize
;
395 dn
->dn_next_bonuslen
[tx
->tx_txg
& TXG_MASK
] = DN_ZERO_BONUSLEN
;
397 dn
->dn_next_bonuslen
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_bonuslen
;
398 rw_exit(&dn
->dn_struct_rwlock
);
402 dnode_setbonus_type(dnode_t
*dn
, dmu_object_type_t newtype
, dmu_tx_t
*tx
)
404 ASSERT3U(zfs_refcount_count(&dn
->dn_holds
), >=, 1);
405 dnode_setdirty(dn
, tx
);
406 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
407 dn
->dn_bonustype
= newtype
;
408 dn
->dn_next_bonustype
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_bonustype
;
409 rw_exit(&dn
->dn_struct_rwlock
);
413 dnode_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
415 ASSERT3U(zfs_refcount_count(&dn
->dn_holds
), >=, 1);
416 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
417 dnode_setdirty(dn
, tx
);
418 dn
->dn_rm_spillblk
[tx
->tx_txg
&TXG_MASK
] = DN_KILL_SPILLBLK
;
419 dn
->dn_have_spill
= B_FALSE
;
423 dnode_setdblksz(dnode_t
*dn
, int size
)
425 ASSERT0(P2PHASE(size
, SPA_MINBLOCKSIZE
));
426 ASSERT3U(size
, <=, SPA_MAXBLOCKSIZE
);
427 ASSERT3U(size
, >=, SPA_MINBLOCKSIZE
);
428 ASSERT3U(size
>> SPA_MINBLOCKSHIFT
, <,
429 1<<(sizeof (dn
->dn_phys
->dn_datablkszsec
) * 8));
430 dn
->dn_datablksz
= size
;
431 dn
->dn_datablkszsec
= size
>> SPA_MINBLOCKSHIFT
;
432 dn
->dn_datablkshift
= ISP2(size
) ? highbit64(size
- 1) : 0;
436 dnode_create(objset_t
*os
, dnode_phys_t
*dnp
, dmu_buf_impl_t
*db
,
437 uint64_t object
, dnode_handle_t
*dnh
)
441 dn
= kmem_cache_alloc(dnode_cache
, KM_SLEEP
);
442 ASSERT(!POINTER_IS_VALID(dn
->dn_objset
));
446 * Defer setting dn_objset until the dnode is ready to be a candidate
447 * for the dnode_move() callback.
449 dn
->dn_object
= object
;
454 if (dnp
->dn_datablkszsec
) {
455 dnode_setdblksz(dn
, dnp
->dn_datablkszsec
<< SPA_MINBLOCKSHIFT
);
457 dn
->dn_datablksz
= 0;
458 dn
->dn_datablkszsec
= 0;
459 dn
->dn_datablkshift
= 0;
461 dn
->dn_indblkshift
= dnp
->dn_indblkshift
;
462 dn
->dn_nlevels
= dnp
->dn_nlevels
;
463 dn
->dn_type
= dnp
->dn_type
;
464 dn
->dn_nblkptr
= dnp
->dn_nblkptr
;
465 dn
->dn_checksum
= dnp
->dn_checksum
;
466 dn
->dn_compress
= dnp
->dn_compress
;
467 dn
->dn_bonustype
= dnp
->dn_bonustype
;
468 dn
->dn_bonuslen
= dnp
->dn_bonuslen
;
469 dn
->dn_num_slots
= dnp
->dn_extra_slots
+ 1;
470 dn
->dn_maxblkid
= dnp
->dn_maxblkid
;
471 dn
->dn_have_spill
= ((dnp
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) != 0);
474 dmu_zfetch_init(&dn
->dn_zfetch
, dn
);
476 ASSERT(DMU_OT_IS_VALID(dn
->dn_phys
->dn_type
));
477 ASSERT(zrl_is_locked(&dnh
->dnh_zrlock
));
478 ASSERT(!DN_SLOT_IS_PTR(dnh
->dnh_dnode
));
480 mutex_enter(&os
->os_lock
);
483 * Exclude special dnodes from os_dnodes so an empty os_dnodes
484 * signifies that the special dnodes have no references from
485 * their children (the entries in os_dnodes). This allows
486 * dnode_destroy() to easily determine if the last child has
487 * been removed and then complete eviction of the objset.
489 if (!DMU_OBJECT_IS_SPECIAL(object
))
490 list_insert_head(&os
->os_dnodes
, dn
);
494 * Everything else must be valid before assigning dn_objset
495 * makes the dnode eligible for dnode_move().
500 mutex_exit(&os
->os_lock
);
502 arc_space_consume(sizeof (dnode_t
), ARC_SPACE_DNODE
);
508 * Caller must be holding the dnode handle, which is released upon return.
511 dnode_destroy(dnode_t
*dn
)
513 objset_t
*os
= dn
->dn_objset
;
514 boolean_t complete_os_eviction
= B_FALSE
;
516 ASSERT((dn
->dn_id_flags
& DN_ID_NEW_EXIST
) == 0);
518 mutex_enter(&os
->os_lock
);
519 POINTER_INVALIDATE(&dn
->dn_objset
);
520 if (!DMU_OBJECT_IS_SPECIAL(dn
->dn_object
)) {
521 list_remove(&os
->os_dnodes
, dn
);
522 complete_os_eviction
=
523 list_is_empty(&os
->os_dnodes
) &&
524 list_link_active(&os
->os_evicting_node
);
526 mutex_exit(&os
->os_lock
);
528 /* the dnode can no longer move, so we can release the handle */
529 if (!zrl_is_locked(&dn
->dn_handle
->dnh_zrlock
))
530 zrl_remove(&dn
->dn_handle
->dnh_zrlock
);
532 dn
->dn_allocated_txg
= 0;
534 dn
->dn_assigned_txg
= 0;
535 dn
->dn_dirty_txg
= 0;
538 if (dn
->dn_dirtyctx_firstset
!= NULL
) {
539 kmem_free(dn
->dn_dirtyctx_firstset
, 1);
540 dn
->dn_dirtyctx_firstset
= NULL
;
542 if (dn
->dn_bonus
!= NULL
) {
543 mutex_enter(&dn
->dn_bonus
->db_mtx
);
544 dbuf_destroy(dn
->dn_bonus
);
549 dn
->dn_have_spill
= B_FALSE
;
554 dn
->dn_oldprojid
= ZFS_DEFAULT_PROJID
;
557 dn
->dn_newprojid
= ZFS_DEFAULT_PROJID
;
560 dmu_zfetch_fini(&dn
->dn_zfetch
);
561 kmem_cache_free(dnode_cache
, dn
);
562 arc_space_return(sizeof (dnode_t
), ARC_SPACE_DNODE
);
564 if (complete_os_eviction
)
565 dmu_objset_evict_done(os
);
569 dnode_allocate(dnode_t
*dn
, dmu_object_type_t ot
, int blocksize
, int ibs
,
570 dmu_object_type_t bonustype
, int bonuslen
, int dn_slots
, dmu_tx_t
*tx
)
574 ASSERT3U(dn_slots
, >, 0);
575 ASSERT3U(dn_slots
<< DNODE_SHIFT
, <=,
576 spa_maxdnodesize(dmu_objset_spa(dn
->dn_objset
)));
577 ASSERT3U(blocksize
, <=,
578 spa_maxblocksize(dmu_objset_spa(dn
->dn_objset
)));
580 blocksize
= 1 << zfs_default_bs
;
582 blocksize
= P2ROUNDUP(blocksize
, SPA_MINBLOCKSIZE
);
585 ibs
= zfs_default_ibs
;
587 ibs
= MIN(MAX(ibs
, DN_MIN_INDBLKSHIFT
), DN_MAX_INDBLKSHIFT
);
589 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
590 dn
->dn_objset
, dn
->dn_object
, tx
->tx_txg
, blocksize
, ibs
, dn_slots
);
591 DNODE_STAT_BUMP(dnode_allocate
);
593 ASSERT(dn
->dn_type
== DMU_OT_NONE
);
594 ASSERT(bcmp(dn
->dn_phys
, &dnode_phys_zero
, sizeof (dnode_phys_t
)) == 0);
595 ASSERT(dn
->dn_phys
->dn_type
== DMU_OT_NONE
);
596 ASSERT(ot
!= DMU_OT_NONE
);
597 ASSERT(DMU_OT_IS_VALID(ot
));
598 ASSERT((bonustype
== DMU_OT_NONE
&& bonuslen
== 0) ||
599 (bonustype
== DMU_OT_SA
&& bonuslen
== 0) ||
600 (bonustype
!= DMU_OT_NONE
&& bonuslen
!= 0));
601 ASSERT(DMU_OT_IS_VALID(bonustype
));
602 ASSERT3U(bonuslen
, <=, DN_SLOTS_TO_BONUSLEN(dn_slots
));
603 ASSERT(dn
->dn_type
== DMU_OT_NONE
);
604 ASSERT0(dn
->dn_maxblkid
);
605 ASSERT0(dn
->dn_allocated_txg
);
606 ASSERT0(dn
->dn_assigned_txg
);
607 ASSERT0(dn
->dn_dirty_txg
);
608 ASSERT(zfs_refcount_is_zero(&dn
->dn_tx_holds
));
609 ASSERT3U(zfs_refcount_count(&dn
->dn_holds
), <=, 1);
610 ASSERT(avl_is_empty(&dn
->dn_dbufs
));
612 for (i
= 0; i
< TXG_SIZE
; i
++) {
613 ASSERT0(dn
->dn_next_nblkptr
[i
]);
614 ASSERT0(dn
->dn_next_nlevels
[i
]);
615 ASSERT0(dn
->dn_next_indblkshift
[i
]);
616 ASSERT0(dn
->dn_next_bonuslen
[i
]);
617 ASSERT0(dn
->dn_next_bonustype
[i
]);
618 ASSERT0(dn
->dn_rm_spillblk
[i
]);
619 ASSERT0(dn
->dn_next_blksz
[i
]);
620 ASSERT0(dn
->dn_next_maxblkid
[i
]);
621 ASSERT(!multilist_link_active(&dn
->dn_dirty_link
[i
]));
622 ASSERT3P(list_head(&dn
->dn_dirty_records
[i
]), ==, NULL
);
623 ASSERT3P(dn
->dn_free_ranges
[i
], ==, NULL
);
627 dnode_setdblksz(dn
, blocksize
);
628 dn
->dn_indblkshift
= ibs
;
630 dn
->dn_num_slots
= dn_slots
;
631 if (bonustype
== DMU_OT_SA
) /* Maximize bonus space for SA */
634 dn
->dn_nblkptr
= MIN(DN_MAX_NBLKPTR
,
635 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots
) - bonuslen
) >>
639 dn
->dn_bonustype
= bonustype
;
640 dn
->dn_bonuslen
= bonuslen
;
641 dn
->dn_checksum
= ZIO_CHECKSUM_INHERIT
;
642 dn
->dn_compress
= ZIO_COMPRESS_INHERIT
;
646 if (dn
->dn_dirtyctx_firstset
) {
647 kmem_free(dn
->dn_dirtyctx_firstset
, 1);
648 dn
->dn_dirtyctx_firstset
= NULL
;
651 dn
->dn_allocated_txg
= tx
->tx_txg
;
654 dnode_setdirty(dn
, tx
);
655 dn
->dn_next_indblkshift
[tx
->tx_txg
& TXG_MASK
] = ibs
;
656 dn
->dn_next_bonuslen
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_bonuslen
;
657 dn
->dn_next_bonustype
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_bonustype
;
658 dn
->dn_next_blksz
[tx
->tx_txg
& TXG_MASK
] = dn
->dn_datablksz
;
662 dnode_reallocate(dnode_t
*dn
, dmu_object_type_t ot
, int blocksize
,
663 dmu_object_type_t bonustype
, int bonuslen
, int dn_slots
, dmu_tx_t
*tx
)
667 ASSERT3U(blocksize
, >=, SPA_MINBLOCKSIZE
);
668 ASSERT3U(blocksize
, <=,
669 spa_maxblocksize(dmu_objset_spa(dn
->dn_objset
)));
670 ASSERT0(blocksize
% SPA_MINBLOCKSIZE
);
671 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
|| dmu_tx_private_ok(tx
));
672 ASSERT(tx
->tx_txg
!= 0);
673 ASSERT((bonustype
== DMU_OT_NONE
&& bonuslen
== 0) ||
674 (bonustype
!= DMU_OT_NONE
&& bonuslen
!= 0) ||
675 (bonustype
== DMU_OT_SA
&& bonuslen
== 0));
676 ASSERT(DMU_OT_IS_VALID(bonustype
));
677 ASSERT3U(bonuslen
, <=,
678 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn
->dn_objset
))));
679 ASSERT3U(bonuslen
, <=, DN_BONUS_SIZE(dn_slots
<< DNODE_SHIFT
));
681 dnode_free_interior_slots(dn
);
682 DNODE_STAT_BUMP(dnode_reallocate
);
684 /* clean up any unreferenced dbufs */
685 dnode_evict_dbufs(dn
);
689 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
690 dnode_setdirty(dn
, tx
);
691 if (dn
->dn_datablksz
!= blocksize
) {
692 /* change blocksize */
693 ASSERT(dn
->dn_maxblkid
== 0 &&
694 (BP_IS_HOLE(&dn
->dn_phys
->dn_blkptr
[0]) ||
695 dnode_block_freed(dn
, 0)));
696 dnode_setdblksz(dn
, blocksize
);
697 dn
->dn_next_blksz
[tx
->tx_txg
&TXG_MASK
] = blocksize
;
699 if (dn
->dn_bonuslen
!= bonuslen
)
700 dn
->dn_next_bonuslen
[tx
->tx_txg
&TXG_MASK
] = bonuslen
;
702 if (bonustype
== DMU_OT_SA
) /* Maximize bonus space for SA */
705 nblkptr
= MIN(DN_MAX_NBLKPTR
,
706 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots
) - bonuslen
) >>
708 if (dn
->dn_bonustype
!= bonustype
)
709 dn
->dn_next_bonustype
[tx
->tx_txg
&TXG_MASK
] = bonustype
;
710 if (dn
->dn_nblkptr
!= nblkptr
)
711 dn
->dn_next_nblkptr
[tx
->tx_txg
&TXG_MASK
] = nblkptr
;
712 if (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
) {
713 dbuf_rm_spill(dn
, tx
);
714 dnode_rm_spill(dn
, tx
);
716 rw_exit(&dn
->dn_struct_rwlock
);
721 /* change bonus size and type */
722 mutex_enter(&dn
->dn_mtx
);
723 dn
->dn_bonustype
= bonustype
;
724 dn
->dn_bonuslen
= bonuslen
;
725 dn
->dn_num_slots
= dn_slots
;
726 dn
->dn_nblkptr
= nblkptr
;
727 dn
->dn_checksum
= ZIO_CHECKSUM_INHERIT
;
728 dn
->dn_compress
= ZIO_COMPRESS_INHERIT
;
729 ASSERT3U(dn
->dn_nblkptr
, <=, DN_MAX_NBLKPTR
);
731 /* fix up the bonus db_size */
733 dn
->dn_bonus
->db
.db_size
=
734 DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
735 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
736 ASSERT(dn
->dn_bonuslen
<= dn
->dn_bonus
->db
.db_size
);
739 dn
->dn_allocated_txg
= tx
->tx_txg
;
740 mutex_exit(&dn
->dn_mtx
);
745 dnode_move_impl(dnode_t
*odn
, dnode_t
*ndn
)
749 ASSERT(!RW_LOCK_HELD(&odn
->dn_struct_rwlock
));
750 ASSERT(MUTEX_NOT_HELD(&odn
->dn_mtx
));
751 ASSERT(MUTEX_NOT_HELD(&odn
->dn_dbufs_mtx
));
752 ASSERT(!RW_LOCK_HELD(&odn
->dn_zfetch
.zf_rwlock
));
755 ndn
->dn_objset
= odn
->dn_objset
;
756 ndn
->dn_object
= odn
->dn_object
;
757 ndn
->dn_dbuf
= odn
->dn_dbuf
;
758 ndn
->dn_handle
= odn
->dn_handle
;
759 ndn
->dn_phys
= odn
->dn_phys
;
760 ndn
->dn_type
= odn
->dn_type
;
761 ndn
->dn_bonuslen
= odn
->dn_bonuslen
;
762 ndn
->dn_bonustype
= odn
->dn_bonustype
;
763 ndn
->dn_nblkptr
= odn
->dn_nblkptr
;
764 ndn
->dn_checksum
= odn
->dn_checksum
;
765 ndn
->dn_compress
= odn
->dn_compress
;
766 ndn
->dn_nlevels
= odn
->dn_nlevels
;
767 ndn
->dn_indblkshift
= odn
->dn_indblkshift
;
768 ndn
->dn_datablkshift
= odn
->dn_datablkshift
;
769 ndn
->dn_datablkszsec
= odn
->dn_datablkszsec
;
770 ndn
->dn_datablksz
= odn
->dn_datablksz
;
771 ndn
->dn_maxblkid
= odn
->dn_maxblkid
;
772 ndn
->dn_num_slots
= odn
->dn_num_slots
;
773 bcopy(&odn
->dn_next_type
[0], &ndn
->dn_next_type
[0],
774 sizeof (odn
->dn_next_type
));
775 bcopy(&odn
->dn_next_nblkptr
[0], &ndn
->dn_next_nblkptr
[0],
776 sizeof (odn
->dn_next_nblkptr
));
777 bcopy(&odn
->dn_next_nlevels
[0], &ndn
->dn_next_nlevels
[0],
778 sizeof (odn
->dn_next_nlevels
));
779 bcopy(&odn
->dn_next_indblkshift
[0], &ndn
->dn_next_indblkshift
[0],
780 sizeof (odn
->dn_next_indblkshift
));
781 bcopy(&odn
->dn_next_bonustype
[0], &ndn
->dn_next_bonustype
[0],
782 sizeof (odn
->dn_next_bonustype
));
783 bcopy(&odn
->dn_rm_spillblk
[0], &ndn
->dn_rm_spillblk
[0],
784 sizeof (odn
->dn_rm_spillblk
));
785 bcopy(&odn
->dn_next_bonuslen
[0], &ndn
->dn_next_bonuslen
[0],
786 sizeof (odn
->dn_next_bonuslen
));
787 bcopy(&odn
->dn_next_blksz
[0], &ndn
->dn_next_blksz
[0],
788 sizeof (odn
->dn_next_blksz
));
789 bcopy(&odn
->dn_next_maxblkid
[0], &ndn
->dn_next_maxblkid
[0],
790 sizeof (odn
->dn_next_maxblkid
));
791 for (i
= 0; i
< TXG_SIZE
; i
++) {
792 list_move_tail(&ndn
->dn_dirty_records
[i
],
793 &odn
->dn_dirty_records
[i
]);
795 bcopy(&odn
->dn_free_ranges
[0], &ndn
->dn_free_ranges
[0],
796 sizeof (odn
->dn_free_ranges
));
797 ndn
->dn_allocated_txg
= odn
->dn_allocated_txg
;
798 ndn
->dn_free_txg
= odn
->dn_free_txg
;
799 ndn
->dn_assigned_txg
= odn
->dn_assigned_txg
;
800 ndn
->dn_dirty_txg
= odn
->dn_dirty_txg
;
801 ndn
->dn_dirtyctx
= odn
->dn_dirtyctx
;
802 ndn
->dn_dirtyctx_firstset
= odn
->dn_dirtyctx_firstset
;
803 ASSERT(zfs_refcount_count(&odn
->dn_tx_holds
) == 0);
804 zfs_refcount_transfer(&ndn
->dn_holds
, &odn
->dn_holds
);
805 ASSERT(avl_is_empty(&ndn
->dn_dbufs
));
806 avl_swap(&ndn
->dn_dbufs
, &odn
->dn_dbufs
);
807 ndn
->dn_dbufs_count
= odn
->dn_dbufs_count
;
808 ndn
->dn_bonus
= odn
->dn_bonus
;
809 ndn
->dn_have_spill
= odn
->dn_have_spill
;
810 ndn
->dn_zio
= odn
->dn_zio
;
811 ndn
->dn_oldused
= odn
->dn_oldused
;
812 ndn
->dn_oldflags
= odn
->dn_oldflags
;
813 ndn
->dn_olduid
= odn
->dn_olduid
;
814 ndn
->dn_oldgid
= odn
->dn_oldgid
;
815 ndn
->dn_oldprojid
= odn
->dn_oldprojid
;
816 ndn
->dn_newuid
= odn
->dn_newuid
;
817 ndn
->dn_newgid
= odn
->dn_newgid
;
818 ndn
->dn_newprojid
= odn
->dn_newprojid
;
819 ndn
->dn_id_flags
= odn
->dn_id_flags
;
820 dmu_zfetch_init(&ndn
->dn_zfetch
, NULL
);
821 list_move_tail(&ndn
->dn_zfetch
.zf_stream
, &odn
->dn_zfetch
.zf_stream
);
822 ndn
->dn_zfetch
.zf_dnode
= odn
->dn_zfetch
.zf_dnode
;
825 * Update back pointers. Updating the handle fixes the back pointer of
826 * every descendant dbuf as well as the bonus dbuf.
828 ASSERT(ndn
->dn_handle
->dnh_dnode
== odn
);
829 ndn
->dn_handle
->dnh_dnode
= ndn
;
830 if (ndn
->dn_zfetch
.zf_dnode
== odn
) {
831 ndn
->dn_zfetch
.zf_dnode
= ndn
;
835 * Invalidate the original dnode by clearing all of its back pointers.
838 odn
->dn_handle
= NULL
;
839 avl_create(&odn
->dn_dbufs
, dbuf_compare
, sizeof (dmu_buf_impl_t
),
840 offsetof(dmu_buf_impl_t
, db_link
));
841 odn
->dn_dbufs_count
= 0;
842 odn
->dn_bonus
= NULL
;
843 odn
->dn_zfetch
.zf_dnode
= NULL
;
846 * Set the low bit of the objset pointer to ensure that dnode_move()
847 * recognizes the dnode as invalid in any subsequent callback.
849 POINTER_INVALIDATE(&odn
->dn_objset
);
852 * Satisfy the destructor.
854 for (i
= 0; i
< TXG_SIZE
; i
++) {
855 list_create(&odn
->dn_dirty_records
[i
],
856 sizeof (dbuf_dirty_record_t
),
857 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
858 odn
->dn_free_ranges
[i
] = NULL
;
859 odn
->dn_next_nlevels
[i
] = 0;
860 odn
->dn_next_indblkshift
[i
] = 0;
861 odn
->dn_next_bonustype
[i
] = 0;
862 odn
->dn_rm_spillblk
[i
] = 0;
863 odn
->dn_next_bonuslen
[i
] = 0;
864 odn
->dn_next_blksz
[i
] = 0;
866 odn
->dn_allocated_txg
= 0;
867 odn
->dn_free_txg
= 0;
868 odn
->dn_assigned_txg
= 0;
869 odn
->dn_dirty_txg
= 0;
870 odn
->dn_dirtyctx
= 0;
871 odn
->dn_dirtyctx_firstset
= NULL
;
872 odn
->dn_have_spill
= B_FALSE
;
875 odn
->dn_oldflags
= 0;
878 odn
->dn_oldprojid
= ZFS_DEFAULT_PROJID
;
881 odn
->dn_newprojid
= ZFS_DEFAULT_PROJID
;
882 odn
->dn_id_flags
= 0;
888 odn
->dn_moved
= (uint8_t)-1;
893 dnode_move(void *buf
, void *newbuf
, size_t size
, void *arg
)
895 dnode_t
*odn
= buf
, *ndn
= newbuf
;
901 * The dnode is on the objset's list of known dnodes if the objset
902 * pointer is valid. We set the low bit of the objset pointer when
903 * freeing the dnode to invalidate it, and the memory patterns written
904 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
905 * A newly created dnode sets the objset pointer last of all to indicate
906 * that the dnode is known and in a valid state to be moved by this
910 if (!POINTER_IS_VALID(os
)) {
911 DNODE_STAT_BUMP(dnode_move_invalid
);
912 return (KMEM_CBRC_DONT_KNOW
);
916 * Ensure that the objset does not go away during the move.
918 rw_enter(&os_lock
, RW_WRITER
);
919 if (os
!= odn
->dn_objset
) {
921 DNODE_STAT_BUMP(dnode_move_recheck1
);
922 return (KMEM_CBRC_DONT_KNOW
);
926 * If the dnode is still valid, then so is the objset. We know that no
927 * valid objset can be freed while we hold os_lock, so we can safely
928 * ensure that the objset remains in use.
930 mutex_enter(&os
->os_lock
);
933 * Recheck the objset pointer in case the dnode was removed just before
934 * acquiring the lock.
936 if (os
!= odn
->dn_objset
) {
937 mutex_exit(&os
->os_lock
);
939 DNODE_STAT_BUMP(dnode_move_recheck2
);
940 return (KMEM_CBRC_DONT_KNOW
);
944 * At this point we know that as long as we hold os->os_lock, the dnode
945 * cannot be freed and fields within the dnode can be safely accessed.
946 * The objset listing this dnode cannot go away as long as this dnode is
950 if (DMU_OBJECT_IS_SPECIAL(odn
->dn_object
)) {
951 mutex_exit(&os
->os_lock
);
952 DNODE_STAT_BUMP(dnode_move_special
);
953 return (KMEM_CBRC_NO
);
955 ASSERT(odn
->dn_dbuf
!= NULL
); /* only "special" dnodes have no parent */
958 * Lock the dnode handle to prevent the dnode from obtaining any new
959 * holds. This also prevents the descendant dbufs and the bonus dbuf
960 * from accessing the dnode, so that we can discount their holds. The
961 * handle is safe to access because we know that while the dnode cannot
962 * go away, neither can its handle. Once we hold dnh_zrlock, we can
963 * safely move any dnode referenced only by dbufs.
965 if (!zrl_tryenter(&odn
->dn_handle
->dnh_zrlock
)) {
966 mutex_exit(&os
->os_lock
);
967 DNODE_STAT_BUMP(dnode_move_handle
);
968 return (KMEM_CBRC_LATER
);
972 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
973 * We need to guarantee that there is a hold for every dbuf in order to
974 * determine whether the dnode is actively referenced. Falsely matching
975 * a dbuf to an active hold would lead to an unsafe move. It's possible
976 * that a thread already having an active dnode hold is about to add a
977 * dbuf, and we can't compare hold and dbuf counts while the add is in
980 if (!rw_tryenter(&odn
->dn_struct_rwlock
, RW_WRITER
)) {
981 zrl_exit(&odn
->dn_handle
->dnh_zrlock
);
982 mutex_exit(&os
->os_lock
);
983 DNODE_STAT_BUMP(dnode_move_rwlock
);
984 return (KMEM_CBRC_LATER
);
988 * A dbuf may be removed (evicted) without an active dnode hold. In that
989 * case, the dbuf count is decremented under the handle lock before the
990 * dbuf's hold is released. This order ensures that if we count the hold
991 * after the dbuf is removed but before its hold is released, we will
992 * treat the unmatched hold as active and exit safely. If we count the
993 * hold before the dbuf is removed, the hold is discounted, and the
994 * removal is blocked until the move completes.
996 refcount
= zfs_refcount_count(&odn
->dn_holds
);
997 ASSERT(refcount
>= 0);
998 dbufs
= odn
->dn_dbufs_count
;
1000 /* We can't have more dbufs than dnode holds. */
1001 ASSERT3U(dbufs
, <=, refcount
);
1002 DTRACE_PROBE3(dnode__move
, dnode_t
*, odn
, int64_t, refcount
,
1005 if (refcount
> dbufs
) {
1006 rw_exit(&odn
->dn_struct_rwlock
);
1007 zrl_exit(&odn
->dn_handle
->dnh_zrlock
);
1008 mutex_exit(&os
->os_lock
);
1009 DNODE_STAT_BUMP(dnode_move_active
);
1010 return (KMEM_CBRC_LATER
);
1013 rw_exit(&odn
->dn_struct_rwlock
);
1016 * At this point we know that anyone with a hold on the dnode is not
1017 * actively referencing it. The dnode is known and in a valid state to
1018 * move. We're holding the locks needed to execute the critical section.
1020 dnode_move_impl(odn
, ndn
);
1022 list_link_replace(&odn
->dn_link
, &ndn
->dn_link
);
1023 /* If the dnode was safe to move, the refcount cannot have changed. */
1024 ASSERT(refcount
== zfs_refcount_count(&ndn
->dn_holds
));
1025 ASSERT(dbufs
== ndn
->dn_dbufs_count
);
1026 zrl_exit(&ndn
->dn_handle
->dnh_zrlock
); /* handle has moved */
1027 mutex_exit(&os
->os_lock
);
1029 return (KMEM_CBRC_YES
);
1031 #endif /* _KERNEL */
1034 dnode_slots_hold(dnode_children_t
*children
, int idx
, int slots
)
1036 ASSERT3S(idx
+ slots
, <=, DNODES_PER_BLOCK
);
1038 for (int i
= idx
; i
< idx
+ slots
; i
++) {
1039 dnode_handle_t
*dnh
= &children
->dnc_children
[i
];
1040 zrl_add(&dnh
->dnh_zrlock
);
1045 dnode_slots_rele(dnode_children_t
*children
, int idx
, int slots
)
1047 ASSERT3S(idx
+ slots
, <=, DNODES_PER_BLOCK
);
1049 for (int i
= idx
; i
< idx
+ slots
; i
++) {
1050 dnode_handle_t
*dnh
= &children
->dnc_children
[i
];
1052 if (zrl_is_locked(&dnh
->dnh_zrlock
))
1053 zrl_exit(&dnh
->dnh_zrlock
);
1055 zrl_remove(&dnh
->dnh_zrlock
);
1060 dnode_slots_tryenter(dnode_children_t
*children
, int idx
, int slots
)
1062 ASSERT3S(idx
+ slots
, <=, DNODES_PER_BLOCK
);
1064 for (int i
= idx
; i
< idx
+ slots
; i
++) {
1065 dnode_handle_t
*dnh
= &children
->dnc_children
[i
];
1067 if (!zrl_tryenter(&dnh
->dnh_zrlock
)) {
1068 for (int j
= idx
; j
< i
; j
++) {
1069 dnh
= &children
->dnc_children
[j
];
1070 zrl_exit(&dnh
->dnh_zrlock
);
1081 dnode_set_slots(dnode_children_t
*children
, int idx
, int slots
, void *ptr
)
1083 ASSERT3S(idx
+ slots
, <=, DNODES_PER_BLOCK
);
1085 for (int i
= idx
; i
< idx
+ slots
; i
++) {
1086 dnode_handle_t
*dnh
= &children
->dnc_children
[i
];
1087 dnh
->dnh_dnode
= ptr
;
1092 dnode_check_slots_free(dnode_children_t
*children
, int idx
, int slots
)
1094 ASSERT3S(idx
+ slots
, <=, DNODES_PER_BLOCK
);
1097 * If all dnode slots are either already free or
1098 * evictable return B_TRUE.
1100 for (int i
= idx
; i
< idx
+ slots
; i
++) {
1101 dnode_handle_t
*dnh
= &children
->dnc_children
[i
];
1102 dnode_t
*dn
= dnh
->dnh_dnode
;
1104 if (dn
== DN_SLOT_FREE
) {
1106 } else if (DN_SLOT_IS_PTR(dn
)) {
1107 mutex_enter(&dn
->dn_mtx
);
1108 boolean_t can_free
= (dn
->dn_type
== DMU_OT_NONE
&&
1109 !DNODE_IS_DIRTY(dn
));
1110 mutex_exit(&dn
->dn_mtx
);
1125 dnode_reclaim_slots(dnode_children_t
*children
, int idx
, int slots
)
1127 ASSERT3S(idx
+ slots
, <=, DNODES_PER_BLOCK
);
1129 for (int i
= idx
; i
< idx
+ slots
; i
++) {
1130 dnode_handle_t
*dnh
= &children
->dnc_children
[i
];
1132 ASSERT(zrl_is_locked(&dnh
->dnh_zrlock
));
1134 if (DN_SLOT_IS_PTR(dnh
->dnh_dnode
)) {
1135 ASSERT3S(dnh
->dnh_dnode
->dn_type
, ==, DMU_OT_NONE
);
1136 dnode_destroy(dnh
->dnh_dnode
);
1137 dnh
->dnh_dnode
= DN_SLOT_FREE
;
1143 dnode_free_interior_slots(dnode_t
*dn
)
1145 dnode_children_t
*children
= dmu_buf_get_user(&dn
->dn_dbuf
->db
);
1146 int epb
= dn
->dn_dbuf
->db
.db_size
>> DNODE_SHIFT
;
1147 int idx
= (dn
->dn_object
& (epb
- 1)) + 1;
1148 int slots
= dn
->dn_num_slots
- 1;
1153 ASSERT3S(idx
+ slots
, <=, DNODES_PER_BLOCK
);
1155 while (!dnode_slots_tryenter(children
, idx
, slots
))
1156 DNODE_STAT_BUMP(dnode_free_interior_lock_retry
);
1158 dnode_set_slots(children
, idx
, slots
, DN_SLOT_FREE
);
1159 dnode_slots_rele(children
, idx
, slots
);
1163 dnode_special_close(dnode_handle_t
*dnh
)
1165 dnode_t
*dn
= dnh
->dnh_dnode
;
1168 * Wait for final references to the dnode to clear. This can
1169 * only happen if the arc is asynchronously evicting state that
1170 * has a hold on this dnode while we are trying to evict this
1173 while (zfs_refcount_count(&dn
->dn_holds
) > 0)
1175 ASSERT(dn
->dn_dbuf
== NULL
||
1176 dmu_buf_get_user(&dn
->dn_dbuf
->db
) == NULL
);
1177 zrl_add(&dnh
->dnh_zrlock
);
1178 dnode_destroy(dn
); /* implicit zrl_remove() */
1179 zrl_destroy(&dnh
->dnh_zrlock
);
1180 dnh
->dnh_dnode
= NULL
;
1184 dnode_special_open(objset_t
*os
, dnode_phys_t
*dnp
, uint64_t object
,
1185 dnode_handle_t
*dnh
)
1189 zrl_init(&dnh
->dnh_zrlock
);
1190 zrl_tryenter(&dnh
->dnh_zrlock
);
1192 dn
= dnode_create(os
, dnp
, NULL
, object
, dnh
);
1195 zrl_exit(&dnh
->dnh_zrlock
);
1199 dnode_buf_evict_async(void *dbu
)
1201 dnode_children_t
*dnc
= dbu
;
1203 DNODE_STAT_BUMP(dnode_buf_evict
);
1205 for (int i
= 0; i
< dnc
->dnc_count
; i
++) {
1206 dnode_handle_t
*dnh
= &dnc
->dnc_children
[i
];
1210 * The dnode handle lock guards against the dnode moving to
1211 * another valid address, so there is no need here to guard
1212 * against changes to or from NULL.
1214 if (!DN_SLOT_IS_PTR(dnh
->dnh_dnode
)) {
1215 zrl_destroy(&dnh
->dnh_zrlock
);
1216 dnh
->dnh_dnode
= DN_SLOT_UNINIT
;
1220 zrl_add(&dnh
->dnh_zrlock
);
1221 dn
= dnh
->dnh_dnode
;
1223 * If there are holds on this dnode, then there should
1224 * be holds on the dnode's containing dbuf as well; thus
1225 * it wouldn't be eligible for eviction and this function
1226 * would not have been called.
1228 ASSERT(zfs_refcount_is_zero(&dn
->dn_holds
));
1229 ASSERT(zfs_refcount_is_zero(&dn
->dn_tx_holds
));
1231 dnode_destroy(dn
); /* implicit zrl_remove() for first slot */
1232 zrl_destroy(&dnh
->dnh_zrlock
);
1233 dnh
->dnh_dnode
= DN_SLOT_UNINIT
;
1235 kmem_free(dnc
, sizeof (dnode_children_t
) +
1236 dnc
->dnc_count
* sizeof (dnode_handle_t
));
1240 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1241 * to ensure the hole at the specified object offset is large enough to
1242 * hold the dnode being created. The slots parameter is also used to ensure
1243 * a dnode does not span multiple dnode blocks. In both of these cases, if
1244 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1245 * are only possible when using DNODE_MUST_BE_FREE.
1247 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1248 * dnode_hold_impl() will check if the requested dnode is already consumed
1249 * as an extra dnode slot by an large dnode, in which case it returns
1253 * EINVAL - Invalid object number or flags.
1254 * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1255 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1256 * - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1257 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1258 * EIO - I/O error when reading the meta dnode dbuf.
1260 * succeeds even for free dnodes.
1263 dnode_hold_impl(objset_t
*os
, uint64_t object
, int flag
, int slots
,
1264 void *tag
, dnode_t
**dnp
)
1267 int drop_struct_lock
= FALSE
;
1272 dnode_children_t
*dnc
;
1273 dnode_phys_t
*dn_block
;
1274 dnode_handle_t
*dnh
;
1276 ASSERT(!(flag
& DNODE_MUST_BE_ALLOCATED
) || (slots
== 0));
1277 ASSERT(!(flag
& DNODE_MUST_BE_FREE
) || (slots
> 0));
1280 * If you are holding the spa config lock as writer, you shouldn't
1281 * be asking the DMU to do *anything* unless it's the root pool
1282 * which may require us to read from the root filesystem while
1283 * holding some (not all) of the locks as writer.
1285 ASSERT(spa_config_held(os
->os_spa
, SCL_ALL
, RW_WRITER
) == 0 ||
1286 (spa_is_root(os
->os_spa
) &&
1287 spa_config_held(os
->os_spa
, SCL_STATE
, RW_WRITER
)));
1289 ASSERT((flag
& DNODE_MUST_BE_ALLOCATED
) || (flag
& DNODE_MUST_BE_FREE
));
1291 if (object
== DMU_USERUSED_OBJECT
|| object
== DMU_GROUPUSED_OBJECT
||
1292 object
== DMU_PROJECTUSED_OBJECT
) {
1293 if (object
== DMU_USERUSED_OBJECT
)
1294 dn
= DMU_USERUSED_DNODE(os
);
1295 else if (object
== DMU_GROUPUSED_OBJECT
)
1296 dn
= DMU_GROUPUSED_DNODE(os
);
1298 dn
= DMU_PROJECTUSED_DNODE(os
);
1300 return (SET_ERROR(ENOENT
));
1302 if ((flag
& DNODE_MUST_BE_ALLOCATED
) && type
== DMU_OT_NONE
)
1303 return (SET_ERROR(ENOENT
));
1304 if ((flag
& DNODE_MUST_BE_FREE
) && type
!= DMU_OT_NONE
)
1305 return (SET_ERROR(EEXIST
));
1307 (void) zfs_refcount_add(&dn
->dn_holds
, tag
);
1312 if (object
== 0 || object
>= DN_MAX_OBJECT
)
1313 return (SET_ERROR(EINVAL
));
1315 mdn
= DMU_META_DNODE(os
);
1316 ASSERT(mdn
->dn_object
== DMU_META_DNODE_OBJECT
);
1320 if (!RW_WRITE_HELD(&mdn
->dn_struct_rwlock
)) {
1321 rw_enter(&mdn
->dn_struct_rwlock
, RW_READER
);
1322 drop_struct_lock
= TRUE
;
1325 blk
= dbuf_whichblock(mdn
, 0, object
* sizeof (dnode_phys_t
));
1327 db
= dbuf_hold(mdn
, blk
, FTAG
);
1328 if (drop_struct_lock
)
1329 rw_exit(&mdn
->dn_struct_rwlock
);
1331 DNODE_STAT_BUMP(dnode_hold_dbuf_hold
);
1332 return (SET_ERROR(EIO
));
1336 * We do not need to decrypt to read the dnode so it doesn't matter
1337 * if we get the encrypted or decrypted version.
1339 err
= dbuf_read(db
, NULL
, DB_RF_CANFAIL
| DB_RF_NO_DECRYPT
);
1341 DNODE_STAT_BUMP(dnode_hold_dbuf_read
);
1342 dbuf_rele(db
, FTAG
);
1346 ASSERT3U(db
->db
.db_size
, >=, 1<<DNODE_SHIFT
);
1347 epb
= db
->db
.db_size
>> DNODE_SHIFT
;
1349 idx
= object
& (epb
- 1);
1350 dn_block
= (dnode_phys_t
*)db
->db
.db_data
;
1352 ASSERT(DB_DNODE(db
)->dn_type
== DMU_OT_DNODE
);
1353 dnc
= dmu_buf_get_user(&db
->db
);
1356 dnode_children_t
*winner
;
1359 dnc
= kmem_zalloc(sizeof (dnode_children_t
) +
1360 epb
* sizeof (dnode_handle_t
), KM_SLEEP
);
1361 dnc
->dnc_count
= epb
;
1362 dnh
= &dnc
->dnc_children
[0];
1364 /* Initialize dnode slot status from dnode_phys_t */
1365 for (int i
= 0; i
< epb
; i
++) {
1366 zrl_init(&dnh
[i
].dnh_zrlock
);
1373 if (dn_block
[i
].dn_type
!= DMU_OT_NONE
) {
1374 int interior
= dn_block
[i
].dn_extra_slots
;
1376 dnode_set_slots(dnc
, i
, 1, DN_SLOT_ALLOCATED
);
1377 dnode_set_slots(dnc
, i
+ 1, interior
,
1381 dnh
[i
].dnh_dnode
= DN_SLOT_FREE
;
1386 dmu_buf_init_user(&dnc
->dnc_dbu
, NULL
,
1387 dnode_buf_evict_async
, NULL
);
1388 winner
= dmu_buf_set_user(&db
->db
, &dnc
->dnc_dbu
);
1389 if (winner
!= NULL
) {
1391 for (int i
= 0; i
< epb
; i
++)
1392 zrl_destroy(&dnh
[i
].dnh_zrlock
);
1394 kmem_free(dnc
, sizeof (dnode_children_t
) +
1395 epb
* sizeof (dnode_handle_t
));
1400 ASSERT(dnc
->dnc_count
== epb
);
1401 dn
= DN_SLOT_UNINIT
;
1403 if (flag
& DNODE_MUST_BE_ALLOCATED
) {
1406 while (dn
== DN_SLOT_UNINIT
) {
1407 dnode_slots_hold(dnc
, idx
, slots
);
1408 dnh
= &dnc
->dnc_children
[idx
];
1410 if (DN_SLOT_IS_PTR(dnh
->dnh_dnode
)) {
1411 dn
= dnh
->dnh_dnode
;
1413 } else if (dnh
->dnh_dnode
== DN_SLOT_INTERIOR
) {
1414 DNODE_STAT_BUMP(dnode_hold_alloc_interior
);
1415 dnode_slots_rele(dnc
, idx
, slots
);
1416 dbuf_rele(db
, FTAG
);
1417 return (SET_ERROR(EEXIST
));
1418 } else if (dnh
->dnh_dnode
!= DN_SLOT_ALLOCATED
) {
1419 DNODE_STAT_BUMP(dnode_hold_alloc_misses
);
1420 dnode_slots_rele(dnc
, idx
, slots
);
1421 dbuf_rele(db
, FTAG
);
1422 return (SET_ERROR(ENOENT
));
1425 dnode_slots_rele(dnc
, idx
, slots
);
1426 if (!dnode_slots_tryenter(dnc
, idx
, slots
)) {
1427 DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry
);
1432 * Someone else won the race and called dnode_create()
1433 * after we checked DN_SLOT_IS_PTR() above but before
1434 * we acquired the lock.
1436 if (DN_SLOT_IS_PTR(dnh
->dnh_dnode
)) {
1437 DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses
);
1438 dn
= dnh
->dnh_dnode
;
1440 dn
= dnode_create(os
, dn_block
+ idx
, db
,
1445 mutex_enter(&dn
->dn_mtx
);
1446 if (dn
->dn_type
== DMU_OT_NONE
) {
1447 DNODE_STAT_BUMP(dnode_hold_alloc_type_none
);
1448 mutex_exit(&dn
->dn_mtx
);
1449 dnode_slots_rele(dnc
, idx
, slots
);
1450 dbuf_rele(db
, FTAG
);
1451 return (SET_ERROR(ENOENT
));
1454 DNODE_STAT_BUMP(dnode_hold_alloc_hits
);
1455 } else if (flag
& DNODE_MUST_BE_FREE
) {
1457 if (idx
+ slots
- 1 >= DNODES_PER_BLOCK
) {
1458 DNODE_STAT_BUMP(dnode_hold_free_overflow
);
1459 dbuf_rele(db
, FTAG
);
1460 return (SET_ERROR(ENOSPC
));
1463 while (dn
== DN_SLOT_UNINIT
) {
1464 dnode_slots_hold(dnc
, idx
, slots
);
1466 if (!dnode_check_slots_free(dnc
, idx
, slots
)) {
1467 DNODE_STAT_BUMP(dnode_hold_free_misses
);
1468 dnode_slots_rele(dnc
, idx
, slots
);
1469 dbuf_rele(db
, FTAG
);
1470 return (SET_ERROR(ENOSPC
));
1473 dnode_slots_rele(dnc
, idx
, slots
);
1474 if (!dnode_slots_tryenter(dnc
, idx
, slots
)) {
1475 DNODE_STAT_BUMP(dnode_hold_free_lock_retry
);
1479 if (!dnode_check_slots_free(dnc
, idx
, slots
)) {
1480 DNODE_STAT_BUMP(dnode_hold_free_lock_misses
);
1481 dnode_slots_rele(dnc
, idx
, slots
);
1482 dbuf_rele(db
, FTAG
);
1483 return (SET_ERROR(ENOSPC
));
1487 * Allocated but otherwise free dnodes which would
1488 * be in the interior of a multi-slot dnodes need
1489 * to be freed. Single slot dnodes can be safely
1490 * re-purposed as a performance optimization.
1493 dnode_reclaim_slots(dnc
, idx
+ 1, slots
- 1);
1495 dnh
= &dnc
->dnc_children
[idx
];
1496 if (DN_SLOT_IS_PTR(dnh
->dnh_dnode
)) {
1497 dn
= dnh
->dnh_dnode
;
1499 dn
= dnode_create(os
, dn_block
+ idx
, db
,
1504 mutex_enter(&dn
->dn_mtx
);
1505 if (!zfs_refcount_is_zero(&dn
->dn_holds
)) {
1506 DNODE_STAT_BUMP(dnode_hold_free_refcount
);
1507 mutex_exit(&dn
->dn_mtx
);
1508 dnode_slots_rele(dnc
, idx
, slots
);
1509 dbuf_rele(db
, FTAG
);
1510 return (SET_ERROR(EEXIST
));
1513 dnode_set_slots(dnc
, idx
+ 1, slots
- 1, DN_SLOT_INTERIOR
);
1514 DNODE_STAT_BUMP(dnode_hold_free_hits
);
1516 dbuf_rele(db
, FTAG
);
1517 return (SET_ERROR(EINVAL
));
1520 if (dn
->dn_free_txg
) {
1521 DNODE_STAT_BUMP(dnode_hold_free_txg
);
1523 mutex_exit(&dn
->dn_mtx
);
1524 dnode_slots_rele(dnc
, idx
, slots
);
1525 dbuf_rele(db
, FTAG
);
1526 return (SET_ERROR((flag
& DNODE_MUST_BE_ALLOCATED
) ?
1530 if (zfs_refcount_add(&dn
->dn_holds
, tag
) == 1)
1531 dbuf_add_ref(db
, dnh
);
1533 mutex_exit(&dn
->dn_mtx
);
1535 /* Now we can rely on the hold to prevent the dnode from moving. */
1536 dnode_slots_rele(dnc
, idx
, slots
);
1539 ASSERT3P(dn
->dn_dbuf
, ==, db
);
1540 ASSERT3U(dn
->dn_object
, ==, object
);
1541 dbuf_rele(db
, FTAG
);
1548 * Return held dnode if the object is allocated, NULL if not.
1551 dnode_hold(objset_t
*os
, uint64_t object
, void *tag
, dnode_t
**dnp
)
1553 return (dnode_hold_impl(os
, object
, DNODE_MUST_BE_ALLOCATED
, 0, tag
,
1558 * Can only add a reference if there is already at least one
1559 * reference on the dnode. Returns FALSE if unable to add a
1563 dnode_add_ref(dnode_t
*dn
, void *tag
)
1565 mutex_enter(&dn
->dn_mtx
);
1566 if (zfs_refcount_is_zero(&dn
->dn_holds
)) {
1567 mutex_exit(&dn
->dn_mtx
);
1570 VERIFY(1 < zfs_refcount_add(&dn
->dn_holds
, tag
));
1571 mutex_exit(&dn
->dn_mtx
);
1576 dnode_rele(dnode_t
*dn
, void *tag
)
1578 mutex_enter(&dn
->dn_mtx
);
1579 dnode_rele_and_unlock(dn
, tag
, B_FALSE
);
1583 dnode_rele_and_unlock(dnode_t
*dn
, void *tag
, boolean_t evicting
)
1586 /* Get while the hold prevents the dnode from moving. */
1587 dmu_buf_impl_t
*db
= dn
->dn_dbuf
;
1588 dnode_handle_t
*dnh
= dn
->dn_handle
;
1590 refs
= zfs_refcount_remove(&dn
->dn_holds
, tag
);
1591 mutex_exit(&dn
->dn_mtx
);
1594 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1595 * indirectly by dbuf_rele() while relying on the dnode handle to
1596 * prevent the dnode from moving, since releasing the last hold could
1597 * result in the dnode's parent dbuf evicting its dnode handles. For
1598 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1599 * other direct or indirect hold on the dnode must first drop the dnode
1602 ASSERT(refs
> 0 || dnh
->dnh_zrlock
.zr_owner
!= curthread
);
1604 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1605 if (refs
== 0 && db
!= NULL
) {
1607 * Another thread could add a hold to the dnode handle in
1608 * dnode_hold_impl() while holding the parent dbuf. Since the
1609 * hold on the parent dbuf prevents the handle from being
1610 * destroyed, the hold on the handle is OK. We can't yet assert
1611 * that the handle has zero references, but that will be
1612 * asserted anyway when the handle gets destroyed.
1614 mutex_enter(&db
->db_mtx
);
1615 dbuf_rele_and_unlock(db
, dnh
, evicting
);
1620 dnode_setdirty(dnode_t
*dn
, dmu_tx_t
*tx
)
1622 objset_t
*os
= dn
->dn_objset
;
1623 uint64_t txg
= tx
->tx_txg
;
1625 if (DMU_OBJECT_IS_SPECIAL(dn
->dn_object
)) {
1626 dsl_dataset_dirty(os
->os_dsl_dataset
, tx
);
1633 mutex_enter(&dn
->dn_mtx
);
1634 ASSERT(dn
->dn_phys
->dn_type
|| dn
->dn_allocated_txg
);
1635 ASSERT(dn
->dn_free_txg
== 0 || dn
->dn_free_txg
>= txg
);
1636 mutex_exit(&dn
->dn_mtx
);
1640 * Determine old uid/gid when necessary
1642 dmu_objset_userquota_get_ids(dn
, B_TRUE
, tx
);
1644 multilist_t
*dirtylist
= os
->os_dirty_dnodes
[txg
& TXG_MASK
];
1645 multilist_sublist_t
*mls
= multilist_sublist_lock_obj(dirtylist
, dn
);
1648 * If we are already marked dirty, we're done.
1650 if (multilist_link_active(&dn
->dn_dirty_link
[txg
& TXG_MASK
])) {
1651 multilist_sublist_unlock(mls
);
1655 ASSERT(!zfs_refcount_is_zero(&dn
->dn_holds
) ||
1656 !avl_is_empty(&dn
->dn_dbufs
));
1657 ASSERT(dn
->dn_datablksz
!= 0);
1658 ASSERT0(dn
->dn_next_bonuslen
[txg
&TXG_MASK
]);
1659 ASSERT0(dn
->dn_next_blksz
[txg
&TXG_MASK
]);
1660 ASSERT0(dn
->dn_next_bonustype
[txg
&TXG_MASK
]);
1662 dprintf_ds(os
->os_dsl_dataset
, "obj=%llu txg=%llu\n",
1663 dn
->dn_object
, txg
);
1665 multilist_sublist_insert_head(mls
, dn
);
1667 multilist_sublist_unlock(mls
);
1670 * The dnode maintains a hold on its containing dbuf as
1671 * long as there are holds on it. Each instantiated child
1672 * dbuf maintains a hold on the dnode. When the last child
1673 * drops its hold, the dnode will drop its hold on the
1674 * containing dbuf. We add a "dirty hold" here so that the
1675 * dnode will hang around after we finish processing its
1678 VERIFY(dnode_add_ref(dn
, (void *)(uintptr_t)tx
->tx_txg
));
1680 (void) dbuf_dirty(dn
->dn_dbuf
, tx
);
1682 dsl_dataset_dirty(os
->os_dsl_dataset
, tx
);
1686 dnode_free(dnode_t
*dn
, dmu_tx_t
*tx
)
1688 mutex_enter(&dn
->dn_mtx
);
1689 if (dn
->dn_type
== DMU_OT_NONE
|| dn
->dn_free_txg
) {
1690 mutex_exit(&dn
->dn_mtx
);
1693 dn
->dn_free_txg
= tx
->tx_txg
;
1694 mutex_exit(&dn
->dn_mtx
);
1696 dnode_setdirty(dn
, tx
);
1700 * Try to change the block size for the indicated dnode. This can only
1701 * succeed if there are no blocks allocated or dirty beyond first block
1704 dnode_set_blksz(dnode_t
*dn
, uint64_t size
, int ibs
, dmu_tx_t
*tx
)
1709 ASSERT3U(size
, <=, spa_maxblocksize(dmu_objset_spa(dn
->dn_objset
)));
1711 size
= SPA_MINBLOCKSIZE
;
1713 size
= P2ROUNDUP(size
, SPA_MINBLOCKSIZE
);
1715 if (ibs
== dn
->dn_indblkshift
)
1718 if (size
>> SPA_MINBLOCKSHIFT
== dn
->dn_datablkszsec
&& ibs
== 0)
1721 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1723 /* Check for any allocated blocks beyond the first */
1724 if (dn
->dn_maxblkid
!= 0)
1727 mutex_enter(&dn
->dn_dbufs_mtx
);
1728 for (db
= avl_first(&dn
->dn_dbufs
); db
!= NULL
;
1729 db
= AVL_NEXT(&dn
->dn_dbufs
, db
)) {
1730 if (db
->db_blkid
!= 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1731 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1732 mutex_exit(&dn
->dn_dbufs_mtx
);
1736 mutex_exit(&dn
->dn_dbufs_mtx
);
1738 if (ibs
&& dn
->dn_nlevels
!= 1)
1741 /* resize the old block */
1742 err
= dbuf_hold_impl(dn
, 0, 0, TRUE
, FALSE
, FTAG
, &db
);
1744 dbuf_new_size(db
, size
, tx
);
1745 else if (err
!= ENOENT
)
1748 dnode_setdblksz(dn
, size
);
1749 dnode_setdirty(dn
, tx
);
1750 dn
->dn_next_blksz
[tx
->tx_txg
&TXG_MASK
] = size
;
1752 dn
->dn_indblkshift
= ibs
;
1753 dn
->dn_next_indblkshift
[tx
->tx_txg
&TXG_MASK
] = ibs
;
1755 /* rele after we have fixed the blocksize in the dnode */
1757 dbuf_rele(db
, FTAG
);
1759 rw_exit(&dn
->dn_struct_rwlock
);
1763 rw_exit(&dn
->dn_struct_rwlock
);
1764 return (SET_ERROR(ENOTSUP
));
1768 dnode_set_nlevels_impl(dnode_t
*dn
, int new_nlevels
, dmu_tx_t
*tx
)
1770 uint64_t txgoff
= tx
->tx_txg
& TXG_MASK
;
1771 int old_nlevels
= dn
->dn_nlevels
;
1774 dbuf_dirty_record_t
*new, *dr
, *dr_next
;
1776 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1778 dn
->dn_nlevels
= new_nlevels
;
1780 ASSERT3U(new_nlevels
, >, dn
->dn_next_nlevels
[txgoff
]);
1781 dn
->dn_next_nlevels
[txgoff
] = new_nlevels
;
1783 /* dirty the left indirects */
1784 db
= dbuf_hold_level(dn
, old_nlevels
, 0, FTAG
);
1786 new = dbuf_dirty(db
, tx
);
1787 dbuf_rele(db
, FTAG
);
1789 /* transfer the dirty records to the new indirect */
1790 mutex_enter(&dn
->dn_mtx
);
1791 mutex_enter(&new->dt
.di
.dr_mtx
);
1792 list
= &dn
->dn_dirty_records
[txgoff
];
1793 for (dr
= list_head(list
); dr
; dr
= dr_next
) {
1794 dr_next
= list_next(&dn
->dn_dirty_records
[txgoff
], dr
);
1795 if (dr
->dr_dbuf
->db_level
!= new_nlevels
-1 &&
1796 dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
1797 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
1798 ASSERT(dr
->dr_dbuf
->db_level
== old_nlevels
-1);
1799 list_remove(&dn
->dn_dirty_records
[txgoff
], dr
);
1800 list_insert_tail(&new->dt
.di
.dr_children
, dr
);
1801 dr
->dr_parent
= new;
1804 mutex_exit(&new->dt
.di
.dr_mtx
);
1805 mutex_exit(&dn
->dn_mtx
);
1809 dnode_set_nlevels(dnode_t
*dn
, int nlevels
, dmu_tx_t
*tx
)
1813 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1815 if (dn
->dn_nlevels
== nlevels
) {
1818 } else if (nlevels
< dn
->dn_nlevels
) {
1819 ret
= SET_ERROR(EINVAL
);
1823 dnode_set_nlevels_impl(dn
, nlevels
, tx
);
1826 rw_exit(&dn
->dn_struct_rwlock
);
1830 /* read-holding callers must not rely on the lock being continuously held */
1832 dnode_new_blkid(dnode_t
*dn
, uint64_t blkid
, dmu_tx_t
*tx
, boolean_t have_read
)
1834 int epbs
, new_nlevels
;
1837 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1840 RW_READ_HELD(&dn
->dn_struct_rwlock
) :
1841 RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1844 * if we have a read-lock, check to see if we need to do any work
1845 * before upgrading to a write-lock.
1848 if (blkid
<= dn
->dn_maxblkid
)
1851 if (!rw_tryupgrade(&dn
->dn_struct_rwlock
)) {
1852 rw_exit(&dn
->dn_struct_rwlock
);
1853 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1857 if (blkid
<= dn
->dn_maxblkid
)
1860 dn
->dn_maxblkid
= blkid
;
1861 dn
->dn_next_maxblkid
[tx
->tx_txg
& TXG_MASK
] = blkid
;
1864 * Compute the number of levels necessary to support the new maxblkid.
1867 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1868 for (sz
= dn
->dn_nblkptr
;
1869 sz
<= blkid
&& sz
>= dn
->dn_nblkptr
; sz
<<= epbs
)
1872 ASSERT3U(new_nlevels
, <=, DN_MAX_LEVELS
);
1874 if (new_nlevels
> dn
->dn_nlevels
)
1875 dnode_set_nlevels_impl(dn
, new_nlevels
, tx
);
1879 rw_downgrade(&dn
->dn_struct_rwlock
);
1883 dnode_dirty_l1(dnode_t
*dn
, uint64_t l1blkid
, dmu_tx_t
*tx
)
1885 dmu_buf_impl_t
*db
= dbuf_hold_level(dn
, 1, l1blkid
, FTAG
);
1887 dmu_buf_will_dirty(&db
->db
, tx
);
1888 dbuf_rele(db
, FTAG
);
1893 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1897 dnode_dirty_l1range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1900 dmu_buf_impl_t db_search
;
1904 mutex_enter(&dn
->dn_dbufs_mtx
);
1906 db_search
.db_level
= 1;
1907 db_search
.db_blkid
= start_blkid
+ 1;
1908 db_search
.db_state
= DB_SEARCH
;
1911 db
= avl_find(&dn
->dn_dbufs
, &db_search
, &where
);
1913 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1915 if (db
== NULL
|| db
->db_level
!= 1 ||
1916 db
->db_blkid
>= end_blkid
) {
1921 * Setup the next blkid we want to search for.
1923 db_search
.db_blkid
= db
->db_blkid
+ 1;
1924 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1927 * If the dbuf transitions to DB_EVICTING while we're trying
1928 * to dirty it, then we will be unable to discover it in
1929 * the dbuf hash table. This will result in a call to
1930 * dbuf_create() which needs to acquire the dn_dbufs_mtx
1931 * lock. To avoid a deadlock, we drop the lock before
1932 * dirtying the level-1 dbuf.
1934 mutex_exit(&dn
->dn_dbufs_mtx
);
1935 dnode_dirty_l1(dn
, db
->db_blkid
, tx
);
1936 mutex_enter(&dn
->dn_dbufs_mtx
);
1941 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
1943 db_search
.db_level
= 1;
1944 db_search
.db_blkid
= start_blkid
+ 1;
1945 db_search
.db_state
= DB_SEARCH
;
1946 db
= avl_find(&dn
->dn_dbufs
, &db_search
, &where
);
1948 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1949 for (; db
!= NULL
; db
= AVL_NEXT(&dn
->dn_dbufs
, db
)) {
1950 if (db
->db_level
!= 1 || db
->db_blkid
>= end_blkid
)
1952 ASSERT(db
->db_dirtycnt
> 0);
1955 mutex_exit(&dn
->dn_dbufs_mtx
);
1959 dnode_free_range(dnode_t
*dn
, uint64_t off
, uint64_t len
, dmu_tx_t
*tx
)
1962 uint64_t blkoff
, blkid
, nblks
;
1963 int blksz
, blkshift
, head
, tail
;
1967 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1968 blksz
= dn
->dn_datablksz
;
1969 blkshift
= dn
->dn_datablkshift
;
1970 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1972 if (len
== DMU_OBJECT_END
) {
1973 len
= UINT64_MAX
- off
;
1978 * First, block align the region to free:
1981 head
= P2NPHASE(off
, blksz
);
1982 blkoff
= P2PHASE(off
, blksz
);
1983 if ((off
>> blkshift
) > dn
->dn_maxblkid
)
1986 ASSERT(dn
->dn_maxblkid
== 0);
1987 if (off
== 0 && len
>= blksz
) {
1989 * Freeing the whole block; fast-track this request.
1993 if (dn
->dn_nlevels
> 1)
1994 dnode_dirty_l1(dn
, 0, tx
);
1996 } else if (off
>= blksz
) {
1997 /* Freeing past end-of-data */
2000 /* Freeing part of the block. */
2002 ASSERT3U(head
, >, 0);
2006 /* zero out any partial block data at the start of the range */
2008 ASSERT3U(blkoff
+ head
, ==, blksz
);
2011 if (dbuf_hold_impl(dn
, 0, dbuf_whichblock(dn
, 0, off
),
2012 TRUE
, FALSE
, FTAG
, &db
) == 0) {
2015 /* don't dirty if it isn't on disk and isn't dirty */
2016 if (db
->db_last_dirty
||
2017 (db
->db_blkptr
&& !BP_IS_HOLE(db
->db_blkptr
))) {
2018 rw_exit(&dn
->dn_struct_rwlock
);
2019 dmu_buf_will_dirty(&db
->db
, tx
);
2020 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2021 data
= db
->db
.db_data
;
2022 bzero(data
+ blkoff
, head
);
2024 dbuf_rele(db
, FTAG
);
2030 /* If the range was less than one block, we're done */
2034 /* If the remaining range is past end of file, we're done */
2035 if ((off
>> blkshift
) > dn
->dn_maxblkid
)
2038 ASSERT(ISP2(blksz
));
2042 tail
= P2PHASE(len
, blksz
);
2044 ASSERT0(P2PHASE(off
, blksz
));
2045 /* zero out any partial block data at the end of the range */
2049 if (dbuf_hold_impl(dn
, 0, dbuf_whichblock(dn
, 0, off
+len
),
2050 TRUE
, FALSE
, FTAG
, &db
) == 0) {
2051 /* don't dirty if not on disk and not dirty */
2052 if (db
->db_last_dirty
||
2053 (db
->db_blkptr
&& !BP_IS_HOLE(db
->db_blkptr
))) {
2054 rw_exit(&dn
->dn_struct_rwlock
);
2055 dmu_buf_will_dirty(&db
->db
, tx
);
2056 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2057 bzero(db
->db
.db_data
, tail
);
2059 dbuf_rele(db
, FTAG
);
2064 /* If the range did not include a full block, we are done */
2068 ASSERT(IS_P2ALIGNED(off
, blksz
));
2069 ASSERT(trunc
|| IS_P2ALIGNED(len
, blksz
));
2070 blkid
= off
>> blkshift
;
2071 nblks
= len
>> blkshift
;
2076 * Dirty all the indirect blocks in this range. Note that only
2077 * the first and last indirect blocks can actually be written
2078 * (if they were partially freed) -- they must be dirtied, even if
2079 * they do not exist on disk yet. The interior blocks will
2080 * be freed by free_children(), so they will not actually be written.
2081 * Even though these interior blocks will not be written, we
2082 * dirty them for two reasons:
2084 * - It ensures that the indirect blocks remain in memory until
2085 * syncing context. (They have already been prefetched by
2086 * dmu_tx_hold_free(), so we don't have to worry about reading
2087 * them serially here.)
2089 * - The dirty space accounting will put pressure on the txg sync
2090 * mechanism to begin syncing, and to delay transactions if there
2091 * is a large amount of freeing. Even though these indirect
2092 * blocks will not be written, we could need to write the same
2093 * amount of space if we copy the freed BPs into deadlists.
2095 if (dn
->dn_nlevels
> 1) {
2096 uint64_t first
, last
;
2098 first
= blkid
>> epbs
;
2099 dnode_dirty_l1(dn
, first
, tx
);
2101 last
= dn
->dn_maxblkid
>> epbs
;
2103 last
= (blkid
+ nblks
- 1) >> epbs
;
2105 dnode_dirty_l1(dn
, last
, tx
);
2107 dnode_dirty_l1range(dn
, first
, last
, tx
);
2109 int shift
= dn
->dn_datablkshift
+ dn
->dn_indblkshift
-
2111 for (uint64_t i
= first
+ 1; i
< last
; i
++) {
2113 * Set i to the blockid of the next non-hole
2114 * level-1 indirect block at or after i. Note
2115 * that dnode_next_offset() operates in terms of
2116 * level-0-equivalent bytes.
2118 uint64_t ibyte
= i
<< shift
;
2119 int err
= dnode_next_offset(dn
, DNODE_FIND_HAVELOCK
,
2126 * Normally we should not see an error, either
2127 * from dnode_next_offset() or dbuf_hold_level()
2128 * (except for ESRCH from dnode_next_offset).
2129 * If there is an i/o error, then when we read
2130 * this block in syncing context, it will use
2131 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2132 * to the "failmode" property. dnode_next_offset()
2133 * doesn't have a flag to indicate MUSTSUCCEED.
2138 dnode_dirty_l1(dn
, i
, tx
);
2144 * Add this range to the dnode range list.
2145 * We will finish up this free operation in the syncing phase.
2147 mutex_enter(&dn
->dn_mtx
);
2149 int txgoff
= tx
->tx_txg
& TXG_MASK
;
2150 if (dn
->dn_free_ranges
[txgoff
] == NULL
) {
2151 dn
->dn_free_ranges
[txgoff
] = range_tree_create(NULL
, NULL
);
2153 range_tree_clear(dn
->dn_free_ranges
[txgoff
], blkid
, nblks
);
2154 range_tree_add(dn
->dn_free_ranges
[txgoff
], blkid
, nblks
);
2156 dprintf_dnode(dn
, "blkid=%llu nblks=%llu txg=%llu\n",
2157 blkid
, nblks
, tx
->tx_txg
);
2158 mutex_exit(&dn
->dn_mtx
);
2160 dbuf_free_range(dn
, blkid
, blkid
+ nblks
- 1, tx
);
2161 dnode_setdirty(dn
, tx
);
2164 rw_exit(&dn
->dn_struct_rwlock
);
2168 dnode_spill_freed(dnode_t
*dn
)
2172 mutex_enter(&dn
->dn_mtx
);
2173 for (i
= 0; i
< TXG_SIZE
; i
++) {
2174 if (dn
->dn_rm_spillblk
[i
] == DN_KILL_SPILLBLK
)
2177 mutex_exit(&dn
->dn_mtx
);
2178 return (i
< TXG_SIZE
);
2181 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2183 dnode_block_freed(dnode_t
*dn
, uint64_t blkid
)
2185 void *dp
= spa_get_dsl(dn
->dn_objset
->os_spa
);
2188 if (blkid
== DMU_BONUS_BLKID
)
2192 * If we're in the process of opening the pool, dp will not be
2193 * set yet, but there shouldn't be anything dirty.
2198 if (dn
->dn_free_txg
)
2201 if (blkid
== DMU_SPILL_BLKID
)
2202 return (dnode_spill_freed(dn
));
2204 mutex_enter(&dn
->dn_mtx
);
2205 for (i
= 0; i
< TXG_SIZE
; i
++) {
2206 if (dn
->dn_free_ranges
[i
] != NULL
&&
2207 range_tree_contains(dn
->dn_free_ranges
[i
], blkid
, 1))
2210 mutex_exit(&dn
->dn_mtx
);
2211 return (i
< TXG_SIZE
);
2214 /* call from syncing context when we actually write/free space for this dnode */
2216 dnode_diduse_space(dnode_t
*dn
, int64_t delta
)
2219 dprintf_dnode(dn
, "dn=%p dnp=%p used=%llu delta=%lld\n",
2221 (u_longlong_t
)dn
->dn_phys
->dn_used
,
2224 mutex_enter(&dn
->dn_mtx
);
2225 space
= DN_USED_BYTES(dn
->dn_phys
);
2227 ASSERT3U(space
+ delta
, >=, space
); /* no overflow */
2229 ASSERT3U(space
, >=, -delta
); /* no underflow */
2232 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_DNODE_BYTES
) {
2233 ASSERT((dn
->dn_phys
->dn_flags
& DNODE_FLAG_USED_BYTES
) == 0);
2234 ASSERT0(P2PHASE(space
, 1<<DEV_BSHIFT
));
2235 dn
->dn_phys
->dn_used
= space
>> DEV_BSHIFT
;
2237 dn
->dn_phys
->dn_used
= space
;
2238 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_USED_BYTES
;
2240 mutex_exit(&dn
->dn_mtx
);
2244 * Scans a block at the indicated "level" looking for a hole or data,
2245 * depending on 'flags'.
2247 * If level > 0, then we are scanning an indirect block looking at its
2248 * pointers. If level == 0, then we are looking at a block of dnodes.
2250 * If we don't find what we are looking for in the block, we return ESRCH.
2251 * Otherwise, return with *offset pointing to the beginning (if searching
2252 * forwards) or end (if searching backwards) of the range covered by the
2253 * block pointer we matched on (or dnode).
2255 * The basic search algorithm used below by dnode_next_offset() is to
2256 * use this function to search up the block tree (widen the search) until
2257 * we find something (i.e., we don't return ESRCH) and then search back
2258 * down the tree (narrow the search) until we reach our original search
2262 dnode_next_offset_level(dnode_t
*dn
, int flags
, uint64_t *offset
,
2263 int lvl
, uint64_t blkfill
, uint64_t txg
)
2265 dmu_buf_impl_t
*db
= NULL
;
2267 uint64_t epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2268 uint64_t epb
= 1ULL << epbs
;
2269 uint64_t minfill
, maxfill
;
2271 int i
, inc
, error
, span
;
2273 hole
= ((flags
& DNODE_FIND_HOLE
) != 0);
2274 inc
= (flags
& DNODE_FIND_BACKWARDS
) ? -1 : 1;
2275 ASSERT(txg
== 0 || !hole
);
2277 if (lvl
== dn
->dn_phys
->dn_nlevels
) {
2279 epb
= dn
->dn_phys
->dn_nblkptr
;
2280 data
= dn
->dn_phys
->dn_blkptr
;
2282 uint64_t blkid
= dbuf_whichblock(dn
, lvl
, *offset
);
2283 error
= dbuf_hold_impl(dn
, lvl
, blkid
, TRUE
, FALSE
, FTAG
, &db
);
2285 if (error
!= ENOENT
)
2290 * This can only happen when we are searching up
2291 * the block tree for data. We don't really need to
2292 * adjust the offset, as we will just end up looking
2293 * at the pointer to this block in its parent, and its
2294 * going to be unallocated, so we will skip over it.
2296 return (SET_ERROR(ESRCH
));
2298 error
= dbuf_read(db
, NULL
,
2299 DB_RF_CANFAIL
| DB_RF_HAVESTRUCT
| DB_RF_NO_DECRYPT
);
2301 dbuf_rele(db
, FTAG
);
2304 data
= db
->db
.db_data
;
2308 if (db
!= NULL
&& txg
!= 0 && (db
->db_blkptr
== NULL
||
2309 db
->db_blkptr
->blk_birth
<= txg
||
2310 BP_IS_HOLE(db
->db_blkptr
))) {
2312 * This can only happen when we are searching up the tree
2313 * and these conditions mean that we need to keep climbing.
2315 error
= SET_ERROR(ESRCH
);
2316 } else if (lvl
== 0) {
2317 dnode_phys_t
*dnp
= data
;
2319 ASSERT(dn
->dn_type
== DMU_OT_DNODE
);
2320 ASSERT(!(flags
& DNODE_FIND_BACKWARDS
));
2322 for (i
= (*offset
>> DNODE_SHIFT
) & (blkfill
- 1);
2323 i
< blkfill
; i
+= dnp
[i
].dn_extra_slots
+ 1) {
2324 if ((dnp
[i
].dn_type
== DMU_OT_NONE
) == hole
)
2329 error
= SET_ERROR(ESRCH
);
2331 *offset
= (*offset
& ~(DNODE_BLOCK_SIZE
- 1)) +
2334 blkptr_t
*bp
= data
;
2335 uint64_t start
= *offset
;
2336 span
= (lvl
- 1) * epbs
+ dn
->dn_datablkshift
;
2338 maxfill
= blkfill
<< ((lvl
- 1) * epbs
);
2345 if (span
>= 8 * sizeof (*offset
)) {
2346 /* This only happens on the highest indirection level */
2347 ASSERT3U((lvl
- 1), ==, dn
->dn_phys
->dn_nlevels
- 1);
2350 *offset
= *offset
>> span
;
2353 for (i
= BF64_GET(*offset
, 0, epbs
);
2354 i
>= 0 && i
< epb
; i
+= inc
) {
2355 if (BP_GET_FILL(&bp
[i
]) >= minfill
&&
2356 BP_GET_FILL(&bp
[i
]) <= maxfill
&&
2357 (hole
|| bp
[i
].blk_birth
> txg
))
2359 if (inc
> 0 || *offset
> 0)
2363 if (span
>= 8 * sizeof (*offset
)) {
2366 *offset
= *offset
<< span
;
2370 /* traversing backwards; position offset at the end */
2371 ASSERT3U(*offset
, <=, start
);
2372 *offset
= MIN(*offset
+ (1ULL << span
) - 1, start
);
2373 } else if (*offset
< start
) {
2376 if (i
< 0 || i
>= epb
)
2377 error
= SET_ERROR(ESRCH
);
2381 dbuf_rele(db
, FTAG
);
2387 * Find the next hole, data, or sparse region at or after *offset.
2388 * The value 'blkfill' tells us how many items we expect to find
2389 * in an L0 data block; this value is 1 for normal objects,
2390 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2391 * DNODES_PER_BLOCK when searching for sparse regions thereof.
2395 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2396 * Finds the next/previous hole/data in a file.
2397 * Used in dmu_offset_next().
2399 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2400 * Finds the next free/allocated dnode an objset's meta-dnode.
2401 * Only finds objects that have new contents since txg (ie.
2402 * bonus buffer changes and content removal are ignored).
2403 * Used in dmu_object_next().
2405 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2406 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
2407 * Used in dmu_object_alloc().
2410 dnode_next_offset(dnode_t
*dn
, int flags
, uint64_t *offset
,
2411 int minlvl
, uint64_t blkfill
, uint64_t txg
)
2413 uint64_t initial_offset
= *offset
;
2417 if (!(flags
& DNODE_FIND_HAVELOCK
))
2418 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2420 if (dn
->dn_phys
->dn_nlevels
== 0) {
2421 error
= SET_ERROR(ESRCH
);
2425 if (dn
->dn_datablkshift
== 0) {
2426 if (*offset
< dn
->dn_datablksz
) {
2427 if (flags
& DNODE_FIND_HOLE
)
2428 *offset
= dn
->dn_datablksz
;
2430 error
= SET_ERROR(ESRCH
);
2435 maxlvl
= dn
->dn_phys
->dn_nlevels
;
2437 for (lvl
= minlvl
; lvl
<= maxlvl
; lvl
++) {
2438 error
= dnode_next_offset_level(dn
,
2439 flags
, offset
, lvl
, blkfill
, txg
);
2444 while (error
== 0 && --lvl
>= minlvl
) {
2445 error
= dnode_next_offset_level(dn
,
2446 flags
, offset
, lvl
, blkfill
, txg
);
2450 * There's always a "virtual hole" at the end of the object, even
2451 * if all BP's which physically exist are non-holes.
2453 if ((flags
& DNODE_FIND_HOLE
) && error
== ESRCH
&& txg
== 0 &&
2454 minlvl
== 1 && blkfill
== 1 && !(flags
& DNODE_FIND_BACKWARDS
)) {
2458 if (error
== 0 && (flags
& DNODE_FIND_BACKWARDS
?
2459 initial_offset
< *offset
: initial_offset
> *offset
))
2460 error
= SET_ERROR(ESRCH
);
2462 if (!(flags
& DNODE_FIND_HAVELOCK
))
2463 rw_exit(&dn
->dn_struct_rwlock
);