xref: /freebsd/sys/contrib/openzfs/module/zfs/dnode.c (revision 4530e0c3e78d0616367d37273d6c1f47f627839b)
1 /*
2  * CDDL HEADER START
3  *
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.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or https://opensource.org/licenses/CDDL-1.0.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
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]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012, 2020 by Delphix. All rights reserved.
24  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/dbuf.h>
29 #include <sys/dnode.h>
30 #include <sys/dmu.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>
36 #include <sys/spa.h>
37 #include <sys/zio.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/range_tree.h>
40 #include <sys/trace_zfs.h>
41 #include <sys/zfs_project.h>
42 
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_free_interior_lock_retry",	KSTAT_DATA_UINT64 },
59 	{ "dnode_allocate",			KSTAT_DATA_UINT64 },
60 	{ "dnode_reallocate",			KSTAT_DATA_UINT64 },
61 	{ "dnode_buf_evict",			KSTAT_DATA_UINT64 },
62 	{ "dnode_alloc_next_chunk",		KSTAT_DATA_UINT64 },
63 	{ "dnode_alloc_race",			KSTAT_DATA_UINT64 },
64 	{ "dnode_alloc_next_block",		KSTAT_DATA_UINT64 },
65 	{ "dnode_move_invalid",			KSTAT_DATA_UINT64 },
66 	{ "dnode_move_recheck1",		KSTAT_DATA_UINT64 },
67 	{ "dnode_move_recheck2",		KSTAT_DATA_UINT64 },
68 	{ "dnode_move_special",			KSTAT_DATA_UINT64 },
69 	{ "dnode_move_handle",			KSTAT_DATA_UINT64 },
70 	{ "dnode_move_rwlock",			KSTAT_DATA_UINT64 },
71 	{ "dnode_move_active",			KSTAT_DATA_UINT64 },
72 };
73 
74 dnode_sums_t dnode_sums;
75 
76 static kstat_t *dnode_ksp;
77 static kmem_cache_t *dnode_cache;
78 
79 static dnode_phys_t dnode_phys_zero __maybe_unused;
80 
81 int zfs_default_bs = SPA_MINBLOCKSHIFT;
82 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
83 
84 #ifdef	_KERNEL
85 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
86 #endif /* _KERNEL */
87 
88 static int
89 dbuf_compare(const void *x1, const void *x2)
90 {
91 	const dmu_buf_impl_t *d1 = x1;
92 	const dmu_buf_impl_t *d2 = x2;
93 
94 	int cmp = TREE_CMP(d1->db_level, d2->db_level);
95 	if (likely(cmp))
96 		return (cmp);
97 
98 	cmp = TREE_CMP(d1->db_blkid, d2->db_blkid);
99 	if (likely(cmp))
100 		return (cmp);
101 
102 	if (d1->db_state == DB_SEARCH) {
103 		ASSERT3S(d2->db_state, !=, DB_SEARCH);
104 		return (-1);
105 	} else if (d2->db_state == DB_SEARCH) {
106 		ASSERT3S(d1->db_state, !=, DB_SEARCH);
107 		return (1);
108 	}
109 
110 	return (TREE_PCMP(d1, d2));
111 }
112 
113 static int
114 dnode_cons(void *arg, void *unused, int kmflag)
115 {
116 	(void) unused, (void) kmflag;
117 	dnode_t *dn = arg;
118 
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);
123 	cv_init(&dn->dn_nodnholds, NULL, CV_DEFAULT, NULL);
124 
125 	/*
126 	 * Every dbuf has a reference, and dropping a tracked reference is
127 	 * O(number of references), so don't track dn_holds.
128 	 */
129 	zfs_refcount_create_untracked(&dn->dn_holds);
130 	zfs_refcount_create(&dn->dn_tx_holds);
131 	list_link_init(&dn->dn_link);
132 
133 	memset(dn->dn_next_type, 0, sizeof (dn->dn_next_type));
134 	memset(dn->dn_next_nblkptr, 0, sizeof (dn->dn_next_nblkptr));
135 	memset(dn->dn_next_nlevels, 0, sizeof (dn->dn_next_nlevels));
136 	memset(dn->dn_next_indblkshift, 0, sizeof (dn->dn_next_indblkshift));
137 	memset(dn->dn_next_bonustype, 0, sizeof (dn->dn_next_bonustype));
138 	memset(dn->dn_rm_spillblk, 0, sizeof (dn->dn_rm_spillblk));
139 	memset(dn->dn_next_bonuslen, 0, sizeof (dn->dn_next_bonuslen));
140 	memset(dn->dn_next_blksz, 0, sizeof (dn->dn_next_blksz));
141 	memset(dn->dn_next_maxblkid, 0, sizeof (dn->dn_next_maxblkid));
142 
143 	for (int i = 0; i < TXG_SIZE; i++) {
144 		multilist_link_init(&dn->dn_dirty_link[i]);
145 		dn->dn_free_ranges[i] = NULL;
146 		list_create(&dn->dn_dirty_records[i],
147 		    sizeof (dbuf_dirty_record_t),
148 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
149 	}
150 
151 	dn->dn_allocated_txg = 0;
152 	dn->dn_free_txg = 0;
153 	dn->dn_assigned_txg = 0;
154 	dn->dn_dirty_txg = 0;
155 	dn->dn_dirtyctx = 0;
156 	dn->dn_dirtyctx_firstset = NULL;
157 	dn->dn_bonus = NULL;
158 	dn->dn_have_spill = B_FALSE;
159 	dn->dn_zio = NULL;
160 	dn->dn_oldused = 0;
161 	dn->dn_oldflags = 0;
162 	dn->dn_olduid = 0;
163 	dn->dn_oldgid = 0;
164 	dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
165 	dn->dn_newuid = 0;
166 	dn->dn_newgid = 0;
167 	dn->dn_newprojid = ZFS_DEFAULT_PROJID;
168 	dn->dn_id_flags = 0;
169 
170 	dn->dn_dbufs_count = 0;
171 	avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
172 	    offsetof(dmu_buf_impl_t, db_link));
173 
174 	dn->dn_moved = 0;
175 	return (0);
176 }
177 
178 static void
179 dnode_dest(void *arg, void *unused)
180 {
181 	(void) unused;
182 	dnode_t *dn = arg;
183 
184 	rw_destroy(&dn->dn_struct_rwlock);
185 	mutex_destroy(&dn->dn_mtx);
186 	mutex_destroy(&dn->dn_dbufs_mtx);
187 	cv_destroy(&dn->dn_notxholds);
188 	cv_destroy(&dn->dn_nodnholds);
189 	zfs_refcount_destroy(&dn->dn_holds);
190 	zfs_refcount_destroy(&dn->dn_tx_holds);
191 	ASSERT(!list_link_active(&dn->dn_link));
192 
193 	for (int i = 0; i < TXG_SIZE; i++) {
194 		ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
195 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
196 		list_destroy(&dn->dn_dirty_records[i]);
197 		ASSERT0(dn->dn_next_nblkptr[i]);
198 		ASSERT0(dn->dn_next_nlevels[i]);
199 		ASSERT0(dn->dn_next_indblkshift[i]);
200 		ASSERT0(dn->dn_next_bonustype[i]);
201 		ASSERT0(dn->dn_rm_spillblk[i]);
202 		ASSERT0(dn->dn_next_bonuslen[i]);
203 		ASSERT0(dn->dn_next_blksz[i]);
204 		ASSERT0(dn->dn_next_maxblkid[i]);
205 	}
206 
207 	ASSERT0(dn->dn_allocated_txg);
208 	ASSERT0(dn->dn_free_txg);
209 	ASSERT0(dn->dn_assigned_txg);
210 	ASSERT0(dn->dn_dirty_txg);
211 	ASSERT0(dn->dn_dirtyctx);
212 	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
213 	ASSERT3P(dn->dn_bonus, ==, NULL);
214 	ASSERT(!dn->dn_have_spill);
215 	ASSERT3P(dn->dn_zio, ==, NULL);
216 	ASSERT0(dn->dn_oldused);
217 	ASSERT0(dn->dn_oldflags);
218 	ASSERT0(dn->dn_olduid);
219 	ASSERT0(dn->dn_oldgid);
220 	ASSERT0(dn->dn_oldprojid);
221 	ASSERT0(dn->dn_newuid);
222 	ASSERT0(dn->dn_newgid);
223 	ASSERT0(dn->dn_newprojid);
224 	ASSERT0(dn->dn_id_flags);
225 
226 	ASSERT0(dn->dn_dbufs_count);
227 	avl_destroy(&dn->dn_dbufs);
228 }
229 
230 static int
231 dnode_kstats_update(kstat_t *ksp, int rw)
232 {
233 	dnode_stats_t *ds = ksp->ks_data;
234 
235 	if (rw == KSTAT_WRITE)
236 		return (EACCES);
237 	ds->dnode_hold_dbuf_hold.value.ui64 =
238 	    wmsum_value(&dnode_sums.dnode_hold_dbuf_hold);
239 	ds->dnode_hold_dbuf_read.value.ui64 =
240 	    wmsum_value(&dnode_sums.dnode_hold_dbuf_read);
241 	ds->dnode_hold_alloc_hits.value.ui64 =
242 	    wmsum_value(&dnode_sums.dnode_hold_alloc_hits);
243 	ds->dnode_hold_alloc_misses.value.ui64 =
244 	    wmsum_value(&dnode_sums.dnode_hold_alloc_misses);
245 	ds->dnode_hold_alloc_interior.value.ui64 =
246 	    wmsum_value(&dnode_sums.dnode_hold_alloc_interior);
247 	ds->dnode_hold_alloc_lock_retry.value.ui64 =
248 	    wmsum_value(&dnode_sums.dnode_hold_alloc_lock_retry);
249 	ds->dnode_hold_alloc_lock_misses.value.ui64 =
250 	    wmsum_value(&dnode_sums.dnode_hold_alloc_lock_misses);
251 	ds->dnode_hold_alloc_type_none.value.ui64 =
252 	    wmsum_value(&dnode_sums.dnode_hold_alloc_type_none);
253 	ds->dnode_hold_free_hits.value.ui64 =
254 	    wmsum_value(&dnode_sums.dnode_hold_free_hits);
255 	ds->dnode_hold_free_misses.value.ui64 =
256 	    wmsum_value(&dnode_sums.dnode_hold_free_misses);
257 	ds->dnode_hold_free_lock_misses.value.ui64 =
258 	    wmsum_value(&dnode_sums.dnode_hold_free_lock_misses);
259 	ds->dnode_hold_free_lock_retry.value.ui64 =
260 	    wmsum_value(&dnode_sums.dnode_hold_free_lock_retry);
261 	ds->dnode_hold_free_refcount.value.ui64 =
262 	    wmsum_value(&dnode_sums.dnode_hold_free_refcount);
263 	ds->dnode_hold_free_overflow.value.ui64 =
264 	    wmsum_value(&dnode_sums.dnode_hold_free_overflow);
265 	ds->dnode_free_interior_lock_retry.value.ui64 =
266 	    wmsum_value(&dnode_sums.dnode_free_interior_lock_retry);
267 	ds->dnode_allocate.value.ui64 =
268 	    wmsum_value(&dnode_sums.dnode_allocate);
269 	ds->dnode_reallocate.value.ui64 =
270 	    wmsum_value(&dnode_sums.dnode_reallocate);
271 	ds->dnode_buf_evict.value.ui64 =
272 	    wmsum_value(&dnode_sums.dnode_buf_evict);
273 	ds->dnode_alloc_next_chunk.value.ui64 =
274 	    wmsum_value(&dnode_sums.dnode_alloc_next_chunk);
275 	ds->dnode_alloc_race.value.ui64 =
276 	    wmsum_value(&dnode_sums.dnode_alloc_race);
277 	ds->dnode_alloc_next_block.value.ui64 =
278 	    wmsum_value(&dnode_sums.dnode_alloc_next_block);
279 	ds->dnode_move_invalid.value.ui64 =
280 	    wmsum_value(&dnode_sums.dnode_move_invalid);
281 	ds->dnode_move_recheck1.value.ui64 =
282 	    wmsum_value(&dnode_sums.dnode_move_recheck1);
283 	ds->dnode_move_recheck2.value.ui64 =
284 	    wmsum_value(&dnode_sums.dnode_move_recheck2);
285 	ds->dnode_move_special.value.ui64 =
286 	    wmsum_value(&dnode_sums.dnode_move_special);
287 	ds->dnode_move_handle.value.ui64 =
288 	    wmsum_value(&dnode_sums.dnode_move_handle);
289 	ds->dnode_move_rwlock.value.ui64 =
290 	    wmsum_value(&dnode_sums.dnode_move_rwlock);
291 	ds->dnode_move_active.value.ui64 =
292 	    wmsum_value(&dnode_sums.dnode_move_active);
293 	return (0);
294 }
295 
296 void
297 dnode_init(void)
298 {
299 	ASSERT(dnode_cache == NULL);
300 	dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
301 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
302 	kmem_cache_set_move(dnode_cache, dnode_move);
303 
304 	wmsum_init(&dnode_sums.dnode_hold_dbuf_hold, 0);
305 	wmsum_init(&dnode_sums.dnode_hold_dbuf_read, 0);
306 	wmsum_init(&dnode_sums.dnode_hold_alloc_hits, 0);
307 	wmsum_init(&dnode_sums.dnode_hold_alloc_misses, 0);
308 	wmsum_init(&dnode_sums.dnode_hold_alloc_interior, 0);
309 	wmsum_init(&dnode_sums.dnode_hold_alloc_lock_retry, 0);
310 	wmsum_init(&dnode_sums.dnode_hold_alloc_lock_misses, 0);
311 	wmsum_init(&dnode_sums.dnode_hold_alloc_type_none, 0);
312 	wmsum_init(&dnode_sums.dnode_hold_free_hits, 0);
313 	wmsum_init(&dnode_sums.dnode_hold_free_misses, 0);
314 	wmsum_init(&dnode_sums.dnode_hold_free_lock_misses, 0);
315 	wmsum_init(&dnode_sums.dnode_hold_free_lock_retry, 0);
316 	wmsum_init(&dnode_sums.dnode_hold_free_refcount, 0);
317 	wmsum_init(&dnode_sums.dnode_hold_free_overflow, 0);
318 	wmsum_init(&dnode_sums.dnode_free_interior_lock_retry, 0);
319 	wmsum_init(&dnode_sums.dnode_allocate, 0);
320 	wmsum_init(&dnode_sums.dnode_reallocate, 0);
321 	wmsum_init(&dnode_sums.dnode_buf_evict, 0);
322 	wmsum_init(&dnode_sums.dnode_alloc_next_chunk, 0);
323 	wmsum_init(&dnode_sums.dnode_alloc_race, 0);
324 	wmsum_init(&dnode_sums.dnode_alloc_next_block, 0);
325 	wmsum_init(&dnode_sums.dnode_move_invalid, 0);
326 	wmsum_init(&dnode_sums.dnode_move_recheck1, 0);
327 	wmsum_init(&dnode_sums.dnode_move_recheck2, 0);
328 	wmsum_init(&dnode_sums.dnode_move_special, 0);
329 	wmsum_init(&dnode_sums.dnode_move_handle, 0);
330 	wmsum_init(&dnode_sums.dnode_move_rwlock, 0);
331 	wmsum_init(&dnode_sums.dnode_move_active, 0);
332 
333 	dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
334 	    KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
335 	    KSTAT_FLAG_VIRTUAL);
336 	if (dnode_ksp != NULL) {
337 		dnode_ksp->ks_data = &dnode_stats;
338 		dnode_ksp->ks_update = dnode_kstats_update;
339 		kstat_install(dnode_ksp);
340 	}
341 }
342 
343 void
344 dnode_fini(void)
345 {
346 	if (dnode_ksp != NULL) {
347 		kstat_delete(dnode_ksp);
348 		dnode_ksp = NULL;
349 	}
350 
351 	wmsum_fini(&dnode_sums.dnode_hold_dbuf_hold);
352 	wmsum_fini(&dnode_sums.dnode_hold_dbuf_read);
353 	wmsum_fini(&dnode_sums.dnode_hold_alloc_hits);
354 	wmsum_fini(&dnode_sums.dnode_hold_alloc_misses);
355 	wmsum_fini(&dnode_sums.dnode_hold_alloc_interior);
356 	wmsum_fini(&dnode_sums.dnode_hold_alloc_lock_retry);
357 	wmsum_fini(&dnode_sums.dnode_hold_alloc_lock_misses);
358 	wmsum_fini(&dnode_sums.dnode_hold_alloc_type_none);
359 	wmsum_fini(&dnode_sums.dnode_hold_free_hits);
360 	wmsum_fini(&dnode_sums.dnode_hold_free_misses);
361 	wmsum_fini(&dnode_sums.dnode_hold_free_lock_misses);
362 	wmsum_fini(&dnode_sums.dnode_hold_free_lock_retry);
363 	wmsum_fini(&dnode_sums.dnode_hold_free_refcount);
364 	wmsum_fini(&dnode_sums.dnode_hold_free_overflow);
365 	wmsum_fini(&dnode_sums.dnode_free_interior_lock_retry);
366 	wmsum_fini(&dnode_sums.dnode_allocate);
367 	wmsum_fini(&dnode_sums.dnode_reallocate);
368 	wmsum_fini(&dnode_sums.dnode_buf_evict);
369 	wmsum_fini(&dnode_sums.dnode_alloc_next_chunk);
370 	wmsum_fini(&dnode_sums.dnode_alloc_race);
371 	wmsum_fini(&dnode_sums.dnode_alloc_next_block);
372 	wmsum_fini(&dnode_sums.dnode_move_invalid);
373 	wmsum_fini(&dnode_sums.dnode_move_recheck1);
374 	wmsum_fini(&dnode_sums.dnode_move_recheck2);
375 	wmsum_fini(&dnode_sums.dnode_move_special);
376 	wmsum_fini(&dnode_sums.dnode_move_handle);
377 	wmsum_fini(&dnode_sums.dnode_move_rwlock);
378 	wmsum_fini(&dnode_sums.dnode_move_active);
379 
380 	kmem_cache_destroy(dnode_cache);
381 	dnode_cache = NULL;
382 }
383 
384 
385 #ifdef ZFS_DEBUG
386 void
387 dnode_verify(dnode_t *dn)
388 {
389 	int drop_struct_lock = FALSE;
390 
391 	ASSERT(dn->dn_phys);
392 	ASSERT(dn->dn_objset);
393 	ASSERT(dn->dn_handle->dnh_dnode == dn);
394 
395 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
396 
397 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
398 		return;
399 
400 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
401 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
402 		drop_struct_lock = TRUE;
403 	}
404 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
405 		int i;
406 		int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
407 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
408 		if (dn->dn_datablkshift) {
409 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
410 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
411 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
412 		}
413 		ASSERT3U(dn->dn_nlevels, <=, 30);
414 		ASSERT(DMU_OT_IS_VALID(dn->dn_type));
415 		ASSERT3U(dn->dn_nblkptr, >=, 1);
416 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
417 		ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
418 		ASSERT3U(dn->dn_datablksz, ==,
419 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
420 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
421 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
422 		    dn->dn_bonuslen, <=, max_bonuslen);
423 		for (i = 0; i < TXG_SIZE; i++) {
424 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
425 		}
426 	}
427 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
428 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
429 	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
430 	if (dn->dn_dbuf != NULL) {
431 		ASSERT3P(dn->dn_phys, ==,
432 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
433 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
434 	}
435 	if (drop_struct_lock)
436 		rw_exit(&dn->dn_struct_rwlock);
437 }
438 #endif
439 
440 void
441 dnode_byteswap(dnode_phys_t *dnp)
442 {
443 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
444 	int i;
445 
446 	if (dnp->dn_type == DMU_OT_NONE) {
447 		memset(dnp, 0, sizeof (dnode_phys_t));
448 		return;
449 	}
450 
451 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
452 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
453 	dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
454 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
455 	dnp->dn_used = BSWAP_64(dnp->dn_used);
456 
457 	/*
458 	 * dn_nblkptr is only one byte, so it's OK to read it in either
459 	 * byte order.  We can't read dn_bouslen.
460 	 */
461 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
462 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
463 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
464 		buf64[i] = BSWAP_64(buf64[i]);
465 
466 	/*
467 	 * OK to check dn_bonuslen for zero, because it won't matter if
468 	 * we have the wrong byte order.  This is necessary because the
469 	 * dnode dnode is smaller than a regular dnode.
470 	 */
471 	if (dnp->dn_bonuslen != 0) {
472 		dmu_object_byteswap_t byteswap;
473 		ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
474 		byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
475 		dmu_ot_byteswap[byteswap].ob_func(DN_BONUS(dnp),
476 		    DN_MAX_BONUS_LEN(dnp));
477 	}
478 
479 	/* Swap SPILL block if we have one */
480 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
481 		byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
482 }
483 
484 void
485 dnode_buf_byteswap(void *vbuf, size_t size)
486 {
487 	int i = 0;
488 
489 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
490 	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
491 
492 	while (i < size) {
493 		dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
494 		dnode_byteswap(dnp);
495 
496 		i += DNODE_MIN_SIZE;
497 		if (dnp->dn_type != DMU_OT_NONE)
498 			i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
499 	}
500 }
501 
502 void
503 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
504 {
505 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
506 
507 	dnode_setdirty(dn, tx);
508 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
509 	ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
510 	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
511 
512 	if (newsize < dn->dn_bonuslen) {
513 		/* clear any data after the end of the new size */
514 		size_t diff = dn->dn_bonuslen - newsize;
515 		char *data_end = ((char *)dn->dn_bonus->db.db_data) + newsize;
516 		memset(data_end, 0, diff);
517 	}
518 
519 	dn->dn_bonuslen = newsize;
520 	if (newsize == 0)
521 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
522 	else
523 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
524 	rw_exit(&dn->dn_struct_rwlock);
525 }
526 
527 void
528 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
529 {
530 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
531 	dnode_setdirty(dn, tx);
532 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
533 	dn->dn_bonustype = newtype;
534 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
535 	rw_exit(&dn->dn_struct_rwlock);
536 }
537 
538 void
539 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
540 {
541 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
542 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
543 	dnode_setdirty(dn, tx);
544 	dn->dn_rm_spillblk[tx->tx_txg & TXG_MASK] = DN_KILL_SPILLBLK;
545 	dn->dn_have_spill = B_FALSE;
546 }
547 
548 static void
549 dnode_setdblksz(dnode_t *dn, int size)
550 {
551 	ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
552 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
553 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
554 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
555 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
556 	dn->dn_datablksz = size;
557 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
558 	dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
559 }
560 
561 static dnode_t *
562 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
563     uint64_t object, dnode_handle_t *dnh)
564 {
565 	dnode_t *dn;
566 
567 	dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
568 	dn->dn_moved = 0;
569 
570 	/*
571 	 * Defer setting dn_objset until the dnode is ready to be a candidate
572 	 * for the dnode_move() callback.
573 	 */
574 	dn->dn_object = object;
575 	dn->dn_dbuf = db;
576 	dn->dn_handle = dnh;
577 	dn->dn_phys = dnp;
578 
579 	if (dnp->dn_datablkszsec) {
580 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
581 	} else {
582 		dn->dn_datablksz = 0;
583 		dn->dn_datablkszsec = 0;
584 		dn->dn_datablkshift = 0;
585 	}
586 	dn->dn_indblkshift = dnp->dn_indblkshift;
587 	dn->dn_nlevels = dnp->dn_nlevels;
588 	dn->dn_type = dnp->dn_type;
589 	dn->dn_nblkptr = dnp->dn_nblkptr;
590 	dn->dn_checksum = dnp->dn_checksum;
591 	dn->dn_compress = dnp->dn_compress;
592 	dn->dn_bonustype = dnp->dn_bonustype;
593 	dn->dn_bonuslen = dnp->dn_bonuslen;
594 	dn->dn_num_slots = dnp->dn_extra_slots + 1;
595 	dn->dn_maxblkid = dnp->dn_maxblkid;
596 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
597 	dn->dn_id_flags = 0;
598 
599 	dmu_zfetch_init(&dn->dn_zfetch, dn);
600 
601 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
602 	ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
603 	ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
604 
605 	mutex_enter(&os->os_lock);
606 
607 	/*
608 	 * Exclude special dnodes from os_dnodes so an empty os_dnodes
609 	 * signifies that the special dnodes have no references from
610 	 * their children (the entries in os_dnodes).  This allows
611 	 * dnode_destroy() to easily determine if the last child has
612 	 * been removed and then complete eviction of the objset.
613 	 */
614 	if (!DMU_OBJECT_IS_SPECIAL(object))
615 		list_insert_head(&os->os_dnodes, dn);
616 	membar_producer();
617 
618 	/*
619 	 * Everything else must be valid before assigning dn_objset
620 	 * makes the dnode eligible for dnode_move().
621 	 */
622 	dn->dn_objset = os;
623 
624 	dnh->dnh_dnode = dn;
625 	mutex_exit(&os->os_lock);
626 
627 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
628 
629 	return (dn);
630 }
631 
632 /*
633  * Caller must be holding the dnode handle, which is released upon return.
634  */
635 static void
636 dnode_destroy(dnode_t *dn)
637 {
638 	objset_t *os = dn->dn_objset;
639 	boolean_t complete_os_eviction = B_FALSE;
640 
641 	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
642 
643 	mutex_enter(&os->os_lock);
644 	POINTER_INVALIDATE(&dn->dn_objset);
645 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
646 		list_remove(&os->os_dnodes, dn);
647 		complete_os_eviction =
648 		    list_is_empty(&os->os_dnodes) &&
649 		    list_link_active(&os->os_evicting_node);
650 	}
651 	mutex_exit(&os->os_lock);
652 
653 	/* the dnode can no longer move, so we can release the handle */
654 	if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
655 		zrl_remove(&dn->dn_handle->dnh_zrlock);
656 
657 	dn->dn_allocated_txg = 0;
658 	dn->dn_free_txg = 0;
659 	dn->dn_assigned_txg = 0;
660 	dn->dn_dirty_txg = 0;
661 
662 	dn->dn_dirtyctx = 0;
663 	dn->dn_dirtyctx_firstset = NULL;
664 	if (dn->dn_bonus != NULL) {
665 		mutex_enter(&dn->dn_bonus->db_mtx);
666 		dbuf_destroy(dn->dn_bonus);
667 		dn->dn_bonus = NULL;
668 	}
669 	dn->dn_zio = NULL;
670 
671 	dn->dn_have_spill = B_FALSE;
672 	dn->dn_oldused = 0;
673 	dn->dn_oldflags = 0;
674 	dn->dn_olduid = 0;
675 	dn->dn_oldgid = 0;
676 	dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
677 	dn->dn_newuid = 0;
678 	dn->dn_newgid = 0;
679 	dn->dn_newprojid = ZFS_DEFAULT_PROJID;
680 	dn->dn_id_flags = 0;
681 
682 	dmu_zfetch_fini(&dn->dn_zfetch);
683 	kmem_cache_free(dnode_cache, dn);
684 	arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
685 
686 	if (complete_os_eviction)
687 		dmu_objset_evict_done(os);
688 }
689 
690 void
691 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
692     dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
693 {
694 	int i;
695 
696 	ASSERT3U(dn_slots, >, 0);
697 	ASSERT3U(dn_slots << DNODE_SHIFT, <=,
698 	    spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
699 	ASSERT3U(blocksize, <=,
700 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
701 	if (blocksize == 0)
702 		blocksize = 1 << zfs_default_bs;
703 	else
704 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
705 
706 	if (ibs == 0)
707 		ibs = zfs_default_ibs;
708 
709 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
710 
711 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
712 	    dn->dn_objset, (u_longlong_t)dn->dn_object,
713 	    (u_longlong_t)tx->tx_txg, blocksize, ibs, dn_slots);
714 	DNODE_STAT_BUMP(dnode_allocate);
715 
716 	ASSERT(dn->dn_type == DMU_OT_NONE);
717 	ASSERT0(memcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)));
718 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
719 	ASSERT(ot != DMU_OT_NONE);
720 	ASSERT(DMU_OT_IS_VALID(ot));
721 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
722 	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
723 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
724 	ASSERT(DMU_OT_IS_VALID(bonustype));
725 	ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
726 	ASSERT(dn->dn_type == DMU_OT_NONE);
727 	ASSERT0(dn->dn_maxblkid);
728 	ASSERT0(dn->dn_allocated_txg);
729 	ASSERT0(dn->dn_assigned_txg);
730 	ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
731 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
732 	ASSERT(avl_is_empty(&dn->dn_dbufs));
733 
734 	for (i = 0; i < TXG_SIZE; i++) {
735 		ASSERT0(dn->dn_next_nblkptr[i]);
736 		ASSERT0(dn->dn_next_nlevels[i]);
737 		ASSERT0(dn->dn_next_indblkshift[i]);
738 		ASSERT0(dn->dn_next_bonuslen[i]);
739 		ASSERT0(dn->dn_next_bonustype[i]);
740 		ASSERT0(dn->dn_rm_spillblk[i]);
741 		ASSERT0(dn->dn_next_blksz[i]);
742 		ASSERT0(dn->dn_next_maxblkid[i]);
743 		ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
744 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
745 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
746 	}
747 
748 	dn->dn_type = ot;
749 	dnode_setdblksz(dn, blocksize);
750 	dn->dn_indblkshift = ibs;
751 	dn->dn_nlevels = 1;
752 	dn->dn_num_slots = dn_slots;
753 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
754 		dn->dn_nblkptr = 1;
755 	else {
756 		dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
757 		    1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
758 		    SPA_BLKPTRSHIFT));
759 	}
760 
761 	dn->dn_bonustype = bonustype;
762 	dn->dn_bonuslen = bonuslen;
763 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
764 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
765 	dn->dn_dirtyctx = 0;
766 
767 	dn->dn_free_txg = 0;
768 	dn->dn_dirtyctx_firstset = NULL;
769 	dn->dn_dirty_txg = 0;
770 
771 	dn->dn_allocated_txg = tx->tx_txg;
772 	dn->dn_id_flags = 0;
773 
774 	dnode_setdirty(dn, tx);
775 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
776 	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
777 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
778 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
779 }
780 
781 void
782 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
783     dmu_object_type_t bonustype, int bonuslen, int dn_slots,
784     boolean_t keep_spill, dmu_tx_t *tx)
785 {
786 	int nblkptr;
787 
788 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
789 	ASSERT3U(blocksize, <=,
790 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
791 	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
792 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
793 	ASSERT(tx->tx_txg != 0);
794 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
795 	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
796 	    (bonustype == DMU_OT_SA && bonuslen == 0));
797 	ASSERT(DMU_OT_IS_VALID(bonustype));
798 	ASSERT3U(bonuslen, <=,
799 	    DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
800 	ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
801 
802 	dnode_free_interior_slots(dn);
803 	DNODE_STAT_BUMP(dnode_reallocate);
804 
805 	/* clean up any unreferenced dbufs */
806 	dnode_evict_dbufs(dn);
807 
808 	dn->dn_id_flags = 0;
809 
810 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
811 	dnode_setdirty(dn, tx);
812 	if (dn->dn_datablksz != blocksize) {
813 		/* change blocksize */
814 		ASSERT0(dn->dn_maxblkid);
815 		ASSERT(BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
816 		    dnode_block_freed(dn, 0));
817 
818 		dnode_setdblksz(dn, blocksize);
819 		dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = blocksize;
820 	}
821 	if (dn->dn_bonuslen != bonuslen)
822 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = bonuslen;
823 
824 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
825 		nblkptr = 1;
826 	else
827 		nblkptr = MIN(DN_MAX_NBLKPTR,
828 		    1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
829 		    SPA_BLKPTRSHIFT));
830 	if (dn->dn_bonustype != bonustype)
831 		dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = bonustype;
832 	if (dn->dn_nblkptr != nblkptr)
833 		dn->dn_next_nblkptr[tx->tx_txg & TXG_MASK] = nblkptr;
834 	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR && !keep_spill) {
835 		dbuf_rm_spill(dn, tx);
836 		dnode_rm_spill(dn, tx);
837 	}
838 
839 	rw_exit(&dn->dn_struct_rwlock);
840 
841 	/* change type */
842 	dn->dn_type = ot;
843 
844 	/* change bonus size and type */
845 	mutex_enter(&dn->dn_mtx);
846 	dn->dn_bonustype = bonustype;
847 	dn->dn_bonuslen = bonuslen;
848 	dn->dn_num_slots = dn_slots;
849 	dn->dn_nblkptr = nblkptr;
850 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
851 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
852 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
853 
854 	/* fix up the bonus db_size */
855 	if (dn->dn_bonus) {
856 		dn->dn_bonus->db.db_size =
857 		    DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
858 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
859 		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
860 	}
861 
862 	dn->dn_allocated_txg = tx->tx_txg;
863 	mutex_exit(&dn->dn_mtx);
864 }
865 
866 #ifdef	_KERNEL
867 static void
868 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
869 {
870 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
871 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
872 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
873 
874 	/* Copy fields. */
875 	ndn->dn_objset = odn->dn_objset;
876 	ndn->dn_object = odn->dn_object;
877 	ndn->dn_dbuf = odn->dn_dbuf;
878 	ndn->dn_handle = odn->dn_handle;
879 	ndn->dn_phys = odn->dn_phys;
880 	ndn->dn_type = odn->dn_type;
881 	ndn->dn_bonuslen = odn->dn_bonuslen;
882 	ndn->dn_bonustype = odn->dn_bonustype;
883 	ndn->dn_nblkptr = odn->dn_nblkptr;
884 	ndn->dn_checksum = odn->dn_checksum;
885 	ndn->dn_compress = odn->dn_compress;
886 	ndn->dn_nlevels = odn->dn_nlevels;
887 	ndn->dn_indblkshift = odn->dn_indblkshift;
888 	ndn->dn_datablkshift = odn->dn_datablkshift;
889 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
890 	ndn->dn_datablksz = odn->dn_datablksz;
891 	ndn->dn_maxblkid = odn->dn_maxblkid;
892 	ndn->dn_num_slots = odn->dn_num_slots;
893 	memcpy(ndn->dn_next_type, odn->dn_next_type,
894 	    sizeof (odn->dn_next_type));
895 	memcpy(ndn->dn_next_nblkptr, odn->dn_next_nblkptr,
896 	    sizeof (odn->dn_next_nblkptr));
897 	memcpy(ndn->dn_next_nlevels, odn->dn_next_nlevels,
898 	    sizeof (odn->dn_next_nlevels));
899 	memcpy(ndn->dn_next_indblkshift, odn->dn_next_indblkshift,
900 	    sizeof (odn->dn_next_indblkshift));
901 	memcpy(ndn->dn_next_bonustype, odn->dn_next_bonustype,
902 	    sizeof (odn->dn_next_bonustype));
903 	memcpy(ndn->dn_rm_spillblk, odn->dn_rm_spillblk,
904 	    sizeof (odn->dn_rm_spillblk));
905 	memcpy(ndn->dn_next_bonuslen, odn->dn_next_bonuslen,
906 	    sizeof (odn->dn_next_bonuslen));
907 	memcpy(ndn->dn_next_blksz, odn->dn_next_blksz,
908 	    sizeof (odn->dn_next_blksz));
909 	memcpy(ndn->dn_next_maxblkid, odn->dn_next_maxblkid,
910 	    sizeof (odn->dn_next_maxblkid));
911 	for (int i = 0; i < TXG_SIZE; i++) {
912 		list_move_tail(&ndn->dn_dirty_records[i],
913 		    &odn->dn_dirty_records[i]);
914 	}
915 	memcpy(ndn->dn_free_ranges, odn->dn_free_ranges,
916 	    sizeof (odn->dn_free_ranges));
917 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
918 	ndn->dn_free_txg = odn->dn_free_txg;
919 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
920 	ndn->dn_dirty_txg = odn->dn_dirty_txg;
921 	ndn->dn_dirtyctx = odn->dn_dirtyctx;
922 	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
923 	ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0);
924 	zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
925 	ASSERT(avl_is_empty(&ndn->dn_dbufs));
926 	avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
927 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
928 	ndn->dn_bonus = odn->dn_bonus;
929 	ndn->dn_have_spill = odn->dn_have_spill;
930 	ndn->dn_zio = odn->dn_zio;
931 	ndn->dn_oldused = odn->dn_oldused;
932 	ndn->dn_oldflags = odn->dn_oldflags;
933 	ndn->dn_olduid = odn->dn_olduid;
934 	ndn->dn_oldgid = odn->dn_oldgid;
935 	ndn->dn_oldprojid = odn->dn_oldprojid;
936 	ndn->dn_newuid = odn->dn_newuid;
937 	ndn->dn_newgid = odn->dn_newgid;
938 	ndn->dn_newprojid = odn->dn_newprojid;
939 	ndn->dn_id_flags = odn->dn_id_flags;
940 	dmu_zfetch_init(&ndn->dn_zfetch, ndn);
941 
942 	/*
943 	 * Update back pointers. Updating the handle fixes the back pointer of
944 	 * every descendant dbuf as well as the bonus dbuf.
945 	 */
946 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
947 	ndn->dn_handle->dnh_dnode = ndn;
948 
949 	/*
950 	 * Invalidate the original dnode by clearing all of its back pointers.
951 	 */
952 	odn->dn_dbuf = NULL;
953 	odn->dn_handle = NULL;
954 	avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
955 	    offsetof(dmu_buf_impl_t, db_link));
956 	odn->dn_dbufs_count = 0;
957 	odn->dn_bonus = NULL;
958 	dmu_zfetch_fini(&odn->dn_zfetch);
959 
960 	/*
961 	 * Set the low bit of the objset pointer to ensure that dnode_move()
962 	 * recognizes the dnode as invalid in any subsequent callback.
963 	 */
964 	POINTER_INVALIDATE(&odn->dn_objset);
965 
966 	/*
967 	 * Satisfy the destructor.
968 	 */
969 	for (int i = 0; i < TXG_SIZE; i++) {
970 		list_create(&odn->dn_dirty_records[i],
971 		    sizeof (dbuf_dirty_record_t),
972 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
973 		odn->dn_free_ranges[i] = NULL;
974 		odn->dn_next_nlevels[i] = 0;
975 		odn->dn_next_indblkshift[i] = 0;
976 		odn->dn_next_bonustype[i] = 0;
977 		odn->dn_rm_spillblk[i] = 0;
978 		odn->dn_next_bonuslen[i] = 0;
979 		odn->dn_next_blksz[i] = 0;
980 	}
981 	odn->dn_allocated_txg = 0;
982 	odn->dn_free_txg = 0;
983 	odn->dn_assigned_txg = 0;
984 	odn->dn_dirty_txg = 0;
985 	odn->dn_dirtyctx = 0;
986 	odn->dn_dirtyctx_firstset = NULL;
987 	odn->dn_have_spill = B_FALSE;
988 	odn->dn_zio = NULL;
989 	odn->dn_oldused = 0;
990 	odn->dn_oldflags = 0;
991 	odn->dn_olduid = 0;
992 	odn->dn_oldgid = 0;
993 	odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
994 	odn->dn_newuid = 0;
995 	odn->dn_newgid = 0;
996 	odn->dn_newprojid = ZFS_DEFAULT_PROJID;
997 	odn->dn_id_flags = 0;
998 
999 	/*
1000 	 * Mark the dnode.
1001 	 */
1002 	ndn->dn_moved = 1;
1003 	odn->dn_moved = (uint8_t)-1;
1004 }
1005 
1006 static kmem_cbrc_t
1007 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
1008 {
1009 	dnode_t *odn = buf, *ndn = newbuf;
1010 	objset_t *os;
1011 	int64_t refcount;
1012 	uint32_t dbufs;
1013 
1014 	/*
1015 	 * The dnode is on the objset's list of known dnodes if the objset
1016 	 * pointer is valid. We set the low bit of the objset pointer when
1017 	 * freeing the dnode to invalidate it, and the memory patterns written
1018 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
1019 	 * A newly created dnode sets the objset pointer last of all to indicate
1020 	 * that the dnode is known and in a valid state to be moved by this
1021 	 * function.
1022 	 */
1023 	os = odn->dn_objset;
1024 	if (!POINTER_IS_VALID(os)) {
1025 		DNODE_STAT_BUMP(dnode_move_invalid);
1026 		return (KMEM_CBRC_DONT_KNOW);
1027 	}
1028 
1029 	/*
1030 	 * Ensure that the objset does not go away during the move.
1031 	 */
1032 	rw_enter(&os_lock, RW_WRITER);
1033 	if (os != odn->dn_objset) {
1034 		rw_exit(&os_lock);
1035 		DNODE_STAT_BUMP(dnode_move_recheck1);
1036 		return (KMEM_CBRC_DONT_KNOW);
1037 	}
1038 
1039 	/*
1040 	 * If the dnode is still valid, then so is the objset. We know that no
1041 	 * valid objset can be freed while we hold os_lock, so we can safely
1042 	 * ensure that the objset remains in use.
1043 	 */
1044 	mutex_enter(&os->os_lock);
1045 
1046 	/*
1047 	 * Recheck the objset pointer in case the dnode was removed just before
1048 	 * acquiring the lock.
1049 	 */
1050 	if (os != odn->dn_objset) {
1051 		mutex_exit(&os->os_lock);
1052 		rw_exit(&os_lock);
1053 		DNODE_STAT_BUMP(dnode_move_recheck2);
1054 		return (KMEM_CBRC_DONT_KNOW);
1055 	}
1056 
1057 	/*
1058 	 * At this point we know that as long as we hold os->os_lock, the dnode
1059 	 * cannot be freed and fields within the dnode can be safely accessed.
1060 	 * The objset listing this dnode cannot go away as long as this dnode is
1061 	 * on its list.
1062 	 */
1063 	rw_exit(&os_lock);
1064 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
1065 		mutex_exit(&os->os_lock);
1066 		DNODE_STAT_BUMP(dnode_move_special);
1067 		return (KMEM_CBRC_NO);
1068 	}
1069 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
1070 
1071 	/*
1072 	 * Lock the dnode handle to prevent the dnode from obtaining any new
1073 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
1074 	 * from accessing the dnode, so that we can discount their holds. The
1075 	 * handle is safe to access because we know that while the dnode cannot
1076 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
1077 	 * safely move any dnode referenced only by dbufs.
1078 	 */
1079 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
1080 		mutex_exit(&os->os_lock);
1081 		DNODE_STAT_BUMP(dnode_move_handle);
1082 		return (KMEM_CBRC_LATER);
1083 	}
1084 
1085 	/*
1086 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
1087 	 * We need to guarantee that there is a hold for every dbuf in order to
1088 	 * determine whether the dnode is actively referenced. Falsely matching
1089 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
1090 	 * that a thread already having an active dnode hold is about to add a
1091 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
1092 	 * progress.
1093 	 */
1094 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
1095 		zrl_exit(&odn->dn_handle->dnh_zrlock);
1096 		mutex_exit(&os->os_lock);
1097 		DNODE_STAT_BUMP(dnode_move_rwlock);
1098 		return (KMEM_CBRC_LATER);
1099 	}
1100 
1101 	/*
1102 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
1103 	 * case, the dbuf count is decremented under the handle lock before the
1104 	 * dbuf's hold is released. This order ensures that if we count the hold
1105 	 * after the dbuf is removed but before its hold is released, we will
1106 	 * treat the unmatched hold as active and exit safely. If we count the
1107 	 * hold before the dbuf is removed, the hold is discounted, and the
1108 	 * removal is blocked until the move completes.
1109 	 */
1110 	refcount = zfs_refcount_count(&odn->dn_holds);
1111 	ASSERT(refcount >= 0);
1112 	dbufs = DN_DBUFS_COUNT(odn);
1113 
1114 	/* We can't have more dbufs than dnode holds. */
1115 	ASSERT3U(dbufs, <=, refcount);
1116 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1117 	    uint32_t, dbufs);
1118 
1119 	if (refcount > dbufs) {
1120 		rw_exit(&odn->dn_struct_rwlock);
1121 		zrl_exit(&odn->dn_handle->dnh_zrlock);
1122 		mutex_exit(&os->os_lock);
1123 		DNODE_STAT_BUMP(dnode_move_active);
1124 		return (KMEM_CBRC_LATER);
1125 	}
1126 
1127 	rw_exit(&odn->dn_struct_rwlock);
1128 
1129 	/*
1130 	 * At this point we know that anyone with a hold on the dnode is not
1131 	 * actively referencing it. The dnode is known and in a valid state to
1132 	 * move. We're holding the locks needed to execute the critical section.
1133 	 */
1134 	dnode_move_impl(odn, ndn);
1135 
1136 	list_link_replace(&odn->dn_link, &ndn->dn_link);
1137 	/* If the dnode was safe to move, the refcount cannot have changed. */
1138 	ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1139 	ASSERT(dbufs == DN_DBUFS_COUNT(ndn));
1140 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1141 	mutex_exit(&os->os_lock);
1142 
1143 	return (KMEM_CBRC_YES);
1144 }
1145 #endif	/* _KERNEL */
1146 
1147 static void
1148 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1149 {
1150 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1151 
1152 	for (int i = idx; i < idx + slots; i++) {
1153 		dnode_handle_t *dnh = &children->dnc_children[i];
1154 		zrl_add(&dnh->dnh_zrlock);
1155 	}
1156 }
1157 
1158 static void
1159 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1160 {
1161 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1162 
1163 	for (int i = idx; i < idx + slots; i++) {
1164 		dnode_handle_t *dnh = &children->dnc_children[i];
1165 
1166 		if (zrl_is_locked(&dnh->dnh_zrlock))
1167 			zrl_exit(&dnh->dnh_zrlock);
1168 		else
1169 			zrl_remove(&dnh->dnh_zrlock);
1170 	}
1171 }
1172 
1173 static int
1174 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1175 {
1176 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1177 
1178 	for (int i = idx; i < idx + slots; i++) {
1179 		dnode_handle_t *dnh = &children->dnc_children[i];
1180 
1181 		if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1182 			for (int j = idx; j < i; j++) {
1183 				dnh = &children->dnc_children[j];
1184 				zrl_exit(&dnh->dnh_zrlock);
1185 			}
1186 
1187 			return (0);
1188 		}
1189 	}
1190 
1191 	return (1);
1192 }
1193 
1194 static void
1195 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1196 {
1197 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1198 
1199 	for (int i = idx; i < idx + slots; i++) {
1200 		dnode_handle_t *dnh = &children->dnc_children[i];
1201 		dnh->dnh_dnode = ptr;
1202 	}
1203 }
1204 
1205 static boolean_t
1206 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1207 {
1208 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1209 
1210 	/*
1211 	 * If all dnode slots are either already free or
1212 	 * evictable return B_TRUE.
1213 	 */
1214 	for (int i = idx; i < idx + slots; i++) {
1215 		dnode_handle_t *dnh = &children->dnc_children[i];
1216 		dnode_t *dn = dnh->dnh_dnode;
1217 
1218 		if (dn == DN_SLOT_FREE) {
1219 			continue;
1220 		} else if (DN_SLOT_IS_PTR(dn)) {
1221 			mutex_enter(&dn->dn_mtx);
1222 			boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1223 			    zfs_refcount_is_zero(&dn->dn_holds) &&
1224 			    !DNODE_IS_DIRTY(dn));
1225 			mutex_exit(&dn->dn_mtx);
1226 
1227 			if (!can_free)
1228 				return (B_FALSE);
1229 			else
1230 				continue;
1231 		} else {
1232 			return (B_FALSE);
1233 		}
1234 	}
1235 
1236 	return (B_TRUE);
1237 }
1238 
1239 static void
1240 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1241 {
1242 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1243 
1244 	for (int i = idx; i < idx + slots; i++) {
1245 		dnode_handle_t *dnh = &children->dnc_children[i];
1246 
1247 		ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1248 
1249 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1250 			ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1251 			dnode_destroy(dnh->dnh_dnode);
1252 			dnh->dnh_dnode = DN_SLOT_FREE;
1253 		}
1254 	}
1255 }
1256 
1257 void
1258 dnode_free_interior_slots(dnode_t *dn)
1259 {
1260 	dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1261 	int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1262 	int idx = (dn->dn_object & (epb - 1)) + 1;
1263 	int slots = dn->dn_num_slots - 1;
1264 
1265 	if (slots == 0)
1266 		return;
1267 
1268 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1269 
1270 	while (!dnode_slots_tryenter(children, idx, slots)) {
1271 		DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1272 		kpreempt(KPREEMPT_SYNC);
1273 	}
1274 
1275 	dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1276 	dnode_slots_rele(children, idx, slots);
1277 }
1278 
1279 void
1280 dnode_special_close(dnode_handle_t *dnh)
1281 {
1282 	dnode_t *dn = dnh->dnh_dnode;
1283 
1284 	/*
1285 	 * Ensure dnode_rele_and_unlock() has released dn_mtx, after final
1286 	 * zfs_refcount_remove()
1287 	 */
1288 	mutex_enter(&dn->dn_mtx);
1289 	if (zfs_refcount_count(&dn->dn_holds) > 0)
1290 		cv_wait(&dn->dn_nodnholds, &dn->dn_mtx);
1291 	mutex_exit(&dn->dn_mtx);
1292 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 0);
1293 
1294 	ASSERT(dn->dn_dbuf == NULL ||
1295 	    dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1296 	zrl_add(&dnh->dnh_zrlock);
1297 	dnode_destroy(dn); /* implicit zrl_remove() */
1298 	zrl_destroy(&dnh->dnh_zrlock);
1299 	dnh->dnh_dnode = NULL;
1300 }
1301 
1302 void
1303 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1304     dnode_handle_t *dnh)
1305 {
1306 	dnode_t *dn;
1307 
1308 	zrl_init(&dnh->dnh_zrlock);
1309 	VERIFY3U(1, ==, zrl_tryenter(&dnh->dnh_zrlock));
1310 
1311 	dn = dnode_create(os, dnp, NULL, object, dnh);
1312 	DNODE_VERIFY(dn);
1313 
1314 	zrl_exit(&dnh->dnh_zrlock);
1315 }
1316 
1317 static void
1318 dnode_buf_evict_async(void *dbu)
1319 {
1320 	dnode_children_t *dnc = dbu;
1321 
1322 	DNODE_STAT_BUMP(dnode_buf_evict);
1323 
1324 	for (int i = 0; i < dnc->dnc_count; i++) {
1325 		dnode_handle_t *dnh = &dnc->dnc_children[i];
1326 		dnode_t *dn;
1327 
1328 		/*
1329 		 * The dnode handle lock guards against the dnode moving to
1330 		 * another valid address, so there is no need here to guard
1331 		 * against changes to or from NULL.
1332 		 */
1333 		if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1334 			zrl_destroy(&dnh->dnh_zrlock);
1335 			dnh->dnh_dnode = DN_SLOT_UNINIT;
1336 			continue;
1337 		}
1338 
1339 		zrl_add(&dnh->dnh_zrlock);
1340 		dn = dnh->dnh_dnode;
1341 		/*
1342 		 * If there are holds on this dnode, then there should
1343 		 * be holds on the dnode's containing dbuf as well; thus
1344 		 * it wouldn't be eligible for eviction and this function
1345 		 * would not have been called.
1346 		 */
1347 		ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1348 		ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1349 
1350 		dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1351 		zrl_destroy(&dnh->dnh_zrlock);
1352 		dnh->dnh_dnode = DN_SLOT_UNINIT;
1353 	}
1354 	kmem_free(dnc, sizeof (dnode_children_t) +
1355 	    dnc->dnc_count * sizeof (dnode_handle_t));
1356 }
1357 
1358 /*
1359  * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1360  * to ensure the hole at the specified object offset is large enough to
1361  * hold the dnode being created. The slots parameter is also used to ensure
1362  * a dnode does not span multiple dnode blocks. In both of these cases, if
1363  * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1364  * are only possible when using DNODE_MUST_BE_FREE.
1365  *
1366  * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1367  * dnode_hold_impl() will check if the requested dnode is already consumed
1368  * as an extra dnode slot by an large dnode, in which case it returns
1369  * ENOENT.
1370  *
1371  * If the DNODE_DRY_RUN flag is set, we don't actually hold the dnode, just
1372  * return whether the hold would succeed or not. tag and dnp should set to
1373  * NULL in this case.
1374  *
1375  * errors:
1376  * EINVAL - Invalid object number or flags.
1377  * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1378  * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1379  *        - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1380  *        - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1381  * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1382  *        - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1383  * EIO    - I/O error when reading the meta dnode dbuf.
1384  *
1385  * succeeds even for free dnodes.
1386  */
1387 int
1388 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1389     const void *tag, dnode_t **dnp)
1390 {
1391 	int epb, idx, err;
1392 	int drop_struct_lock = FALSE;
1393 	int type;
1394 	uint64_t blk;
1395 	dnode_t *mdn, *dn;
1396 	dmu_buf_impl_t *db;
1397 	dnode_children_t *dnc;
1398 	dnode_phys_t *dn_block;
1399 	dnode_handle_t *dnh;
1400 
1401 	ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1402 	ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1403 	IMPLY(flag & DNODE_DRY_RUN, (tag == NULL) && (dnp == NULL));
1404 
1405 	/*
1406 	 * If you are holding the spa config lock as writer, you shouldn't
1407 	 * be asking the DMU to do *anything* unless it's the root pool
1408 	 * which may require us to read from the root filesystem while
1409 	 * holding some (not all) of the locks as writer.
1410 	 */
1411 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1412 	    (spa_is_root(os->os_spa) &&
1413 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1414 
1415 	ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1416 
1417 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1418 	    object == DMU_PROJECTUSED_OBJECT) {
1419 		if (object == DMU_USERUSED_OBJECT)
1420 			dn = DMU_USERUSED_DNODE(os);
1421 		else if (object == DMU_GROUPUSED_OBJECT)
1422 			dn = DMU_GROUPUSED_DNODE(os);
1423 		else
1424 			dn = DMU_PROJECTUSED_DNODE(os);
1425 		if (dn == NULL)
1426 			return (SET_ERROR(ENOENT));
1427 		type = dn->dn_type;
1428 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1429 			return (SET_ERROR(ENOENT));
1430 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1431 			return (SET_ERROR(EEXIST));
1432 		DNODE_VERIFY(dn);
1433 		/* Don't actually hold if dry run, just return 0 */
1434 		if (!(flag & DNODE_DRY_RUN)) {
1435 			(void) zfs_refcount_add(&dn->dn_holds, tag);
1436 			*dnp = dn;
1437 		}
1438 		return (0);
1439 	}
1440 
1441 	if (object == 0 || object >= DN_MAX_OBJECT)
1442 		return (SET_ERROR(EINVAL));
1443 
1444 	mdn = DMU_META_DNODE(os);
1445 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1446 
1447 	DNODE_VERIFY(mdn);
1448 
1449 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1450 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1451 		drop_struct_lock = TRUE;
1452 	}
1453 
1454 	blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1455 	db = dbuf_hold(mdn, blk, FTAG);
1456 	if (drop_struct_lock)
1457 		rw_exit(&mdn->dn_struct_rwlock);
1458 	if (db == NULL) {
1459 		DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1460 		return (SET_ERROR(EIO));
1461 	}
1462 
1463 	/*
1464 	 * We do not need to decrypt to read the dnode so it doesn't matter
1465 	 * if we get the encrypted or decrypted version.
1466 	 */
1467 	err = dbuf_read(db, NULL, DB_RF_CANFAIL |
1468 	    DB_RF_NO_DECRYPT | DB_RF_NOPREFETCH);
1469 	if (err) {
1470 		DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1471 		dbuf_rele(db, FTAG);
1472 		return (err);
1473 	}
1474 
1475 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1476 	epb = db->db.db_size >> DNODE_SHIFT;
1477 
1478 	idx = object & (epb - 1);
1479 	dn_block = (dnode_phys_t *)db->db.db_data;
1480 
1481 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1482 	dnc = dmu_buf_get_user(&db->db);
1483 	dnh = NULL;
1484 	if (dnc == NULL) {
1485 		dnode_children_t *winner;
1486 		int skip = 0;
1487 
1488 		dnc = kmem_zalloc(sizeof (dnode_children_t) +
1489 		    epb * sizeof (dnode_handle_t), KM_SLEEP);
1490 		dnc->dnc_count = epb;
1491 		dnh = &dnc->dnc_children[0];
1492 
1493 		/* Initialize dnode slot status from dnode_phys_t */
1494 		for (int i = 0; i < epb; i++) {
1495 			zrl_init(&dnh[i].dnh_zrlock);
1496 
1497 			if (skip) {
1498 				skip--;
1499 				continue;
1500 			}
1501 
1502 			if (dn_block[i].dn_type != DMU_OT_NONE) {
1503 				int interior = dn_block[i].dn_extra_slots;
1504 
1505 				dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1506 				dnode_set_slots(dnc, i + 1, interior,
1507 				    DN_SLOT_INTERIOR);
1508 				skip = interior;
1509 			} else {
1510 				dnh[i].dnh_dnode = DN_SLOT_FREE;
1511 				skip = 0;
1512 			}
1513 		}
1514 
1515 		dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1516 		    dnode_buf_evict_async, NULL);
1517 		winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1518 		if (winner != NULL) {
1519 
1520 			for (int i = 0; i < epb; i++)
1521 				zrl_destroy(&dnh[i].dnh_zrlock);
1522 
1523 			kmem_free(dnc, sizeof (dnode_children_t) +
1524 			    epb * sizeof (dnode_handle_t));
1525 			dnc = winner;
1526 		}
1527 	}
1528 
1529 	ASSERT(dnc->dnc_count == epb);
1530 
1531 	if (flag & DNODE_MUST_BE_ALLOCATED) {
1532 		slots = 1;
1533 
1534 		dnode_slots_hold(dnc, idx, slots);
1535 		dnh = &dnc->dnc_children[idx];
1536 
1537 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1538 			dn = dnh->dnh_dnode;
1539 		} else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1540 			DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1541 			dnode_slots_rele(dnc, idx, slots);
1542 			dbuf_rele(db, FTAG);
1543 			return (SET_ERROR(EEXIST));
1544 		} else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1545 			DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1546 			dnode_slots_rele(dnc, idx, slots);
1547 			dbuf_rele(db, FTAG);
1548 			return (SET_ERROR(ENOENT));
1549 		} else {
1550 			dnode_slots_rele(dnc, idx, slots);
1551 			while (!dnode_slots_tryenter(dnc, idx, slots)) {
1552 				DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1553 				kpreempt(KPREEMPT_SYNC);
1554 			}
1555 
1556 			/*
1557 			 * Someone else won the race and called dnode_create()
1558 			 * after we checked DN_SLOT_IS_PTR() above but before
1559 			 * we acquired the lock.
1560 			 */
1561 			if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1562 				DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1563 				dn = dnh->dnh_dnode;
1564 			} else {
1565 				dn = dnode_create(os, dn_block + idx, db,
1566 				    object, dnh);
1567 			}
1568 		}
1569 
1570 		mutex_enter(&dn->dn_mtx);
1571 		if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1572 			DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1573 			mutex_exit(&dn->dn_mtx);
1574 			dnode_slots_rele(dnc, idx, slots);
1575 			dbuf_rele(db, FTAG);
1576 			return (SET_ERROR(ENOENT));
1577 		}
1578 
1579 		/* Don't actually hold if dry run, just return 0 */
1580 		if (flag & DNODE_DRY_RUN) {
1581 			mutex_exit(&dn->dn_mtx);
1582 			dnode_slots_rele(dnc, idx, slots);
1583 			dbuf_rele(db, FTAG);
1584 			return (0);
1585 		}
1586 
1587 		DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1588 	} else if (flag & DNODE_MUST_BE_FREE) {
1589 
1590 		if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1591 			DNODE_STAT_BUMP(dnode_hold_free_overflow);
1592 			dbuf_rele(db, FTAG);
1593 			return (SET_ERROR(ENOSPC));
1594 		}
1595 
1596 		dnode_slots_hold(dnc, idx, slots);
1597 
1598 		if (!dnode_check_slots_free(dnc, idx, slots)) {
1599 			DNODE_STAT_BUMP(dnode_hold_free_misses);
1600 			dnode_slots_rele(dnc, idx, slots);
1601 			dbuf_rele(db, FTAG);
1602 			return (SET_ERROR(ENOSPC));
1603 		}
1604 
1605 		dnode_slots_rele(dnc, idx, slots);
1606 		while (!dnode_slots_tryenter(dnc, idx, slots)) {
1607 			DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1608 			kpreempt(KPREEMPT_SYNC);
1609 		}
1610 
1611 		if (!dnode_check_slots_free(dnc, idx, slots)) {
1612 			DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1613 			dnode_slots_rele(dnc, idx, slots);
1614 			dbuf_rele(db, FTAG);
1615 			return (SET_ERROR(ENOSPC));
1616 		}
1617 
1618 		/*
1619 		 * Allocated but otherwise free dnodes which would
1620 		 * be in the interior of a multi-slot dnodes need
1621 		 * to be freed.  Single slot dnodes can be safely
1622 		 * re-purposed as a performance optimization.
1623 		 */
1624 		if (slots > 1)
1625 			dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1626 
1627 		dnh = &dnc->dnc_children[idx];
1628 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1629 			dn = dnh->dnh_dnode;
1630 		} else {
1631 			dn = dnode_create(os, dn_block + idx, db,
1632 			    object, dnh);
1633 		}
1634 
1635 		mutex_enter(&dn->dn_mtx);
1636 		if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1637 			DNODE_STAT_BUMP(dnode_hold_free_refcount);
1638 			mutex_exit(&dn->dn_mtx);
1639 			dnode_slots_rele(dnc, idx, slots);
1640 			dbuf_rele(db, FTAG);
1641 			return (SET_ERROR(EEXIST));
1642 		}
1643 
1644 		/* Don't actually hold if dry run, just return 0 */
1645 		if (flag & DNODE_DRY_RUN) {
1646 			mutex_exit(&dn->dn_mtx);
1647 			dnode_slots_rele(dnc, idx, slots);
1648 			dbuf_rele(db, FTAG);
1649 			return (0);
1650 		}
1651 
1652 		dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1653 		DNODE_STAT_BUMP(dnode_hold_free_hits);
1654 	} else {
1655 		dbuf_rele(db, FTAG);
1656 		return (SET_ERROR(EINVAL));
1657 	}
1658 
1659 	ASSERT0(dn->dn_free_txg);
1660 
1661 	if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1662 		dbuf_add_ref(db, dnh);
1663 
1664 	mutex_exit(&dn->dn_mtx);
1665 
1666 	/* Now we can rely on the hold to prevent the dnode from moving. */
1667 	dnode_slots_rele(dnc, idx, slots);
1668 
1669 	DNODE_VERIFY(dn);
1670 	ASSERT3P(dnp, !=, NULL);
1671 	ASSERT3P(dn->dn_dbuf, ==, db);
1672 	ASSERT3U(dn->dn_object, ==, object);
1673 	dbuf_rele(db, FTAG);
1674 
1675 	*dnp = dn;
1676 	return (0);
1677 }
1678 
1679 /*
1680  * Return held dnode if the object is allocated, NULL if not.
1681  */
1682 int
1683 dnode_hold(objset_t *os, uint64_t object, const void *tag, dnode_t **dnp)
1684 {
1685 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1686 	    dnp));
1687 }
1688 
1689 /*
1690  * Can only add a reference if there is already at least one
1691  * reference on the dnode.  Returns FALSE if unable to add a
1692  * new reference.
1693  */
1694 boolean_t
1695 dnode_add_ref(dnode_t *dn, const void *tag)
1696 {
1697 	mutex_enter(&dn->dn_mtx);
1698 	if (zfs_refcount_is_zero(&dn->dn_holds)) {
1699 		mutex_exit(&dn->dn_mtx);
1700 		return (FALSE);
1701 	}
1702 	VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1703 	mutex_exit(&dn->dn_mtx);
1704 	return (TRUE);
1705 }
1706 
1707 void
1708 dnode_rele(dnode_t *dn, const void *tag)
1709 {
1710 	mutex_enter(&dn->dn_mtx);
1711 	dnode_rele_and_unlock(dn, tag, B_FALSE);
1712 }
1713 
1714 void
1715 dnode_rele_and_unlock(dnode_t *dn, const void *tag, boolean_t evicting)
1716 {
1717 	uint64_t refs;
1718 	/* Get while the hold prevents the dnode from moving. */
1719 	dmu_buf_impl_t *db = dn->dn_dbuf;
1720 	dnode_handle_t *dnh = dn->dn_handle;
1721 
1722 	refs = zfs_refcount_remove(&dn->dn_holds, tag);
1723 	if (refs == 0)
1724 		cv_broadcast(&dn->dn_nodnholds);
1725 	mutex_exit(&dn->dn_mtx);
1726 	/* dnode could get destroyed at this point, so don't use it anymore */
1727 
1728 	/*
1729 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1730 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1731 	 * prevent the dnode from moving, since releasing the last hold could
1732 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1733 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1734 	 * other direct or indirect hold on the dnode must first drop the dnode
1735 	 * handle.
1736 	 */
1737 #ifdef ZFS_DEBUG
1738 	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1739 #endif
1740 
1741 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1742 	if (refs == 0 && db != NULL) {
1743 		/*
1744 		 * Another thread could add a hold to the dnode handle in
1745 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1746 		 * hold on the parent dbuf prevents the handle from being
1747 		 * destroyed, the hold on the handle is OK. We can't yet assert
1748 		 * that the handle has zero references, but that will be
1749 		 * asserted anyway when the handle gets destroyed.
1750 		 */
1751 		mutex_enter(&db->db_mtx);
1752 		dbuf_rele_and_unlock(db, dnh, evicting);
1753 	}
1754 }
1755 
1756 /*
1757  * Test whether we can create a dnode at the specified location.
1758  */
1759 int
1760 dnode_try_claim(objset_t *os, uint64_t object, int slots)
1761 {
1762 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_FREE | DNODE_DRY_RUN,
1763 	    slots, NULL, NULL));
1764 }
1765 
1766 /*
1767  * Checks if the dnode contains any uncommitted dirty records.
1768  */
1769 boolean_t
1770 dnode_is_dirty(dnode_t *dn)
1771 {
1772 	mutex_enter(&dn->dn_mtx);
1773 
1774 	for (int i = 0; i < TXG_SIZE; i++) {
1775 		if (multilist_link_active(&dn->dn_dirty_link[i])) {
1776 			mutex_exit(&dn->dn_mtx);
1777 			return (B_TRUE);
1778 		}
1779 	}
1780 
1781 	mutex_exit(&dn->dn_mtx);
1782 
1783 	return (B_FALSE);
1784 }
1785 
1786 void
1787 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1788 {
1789 	objset_t *os = dn->dn_objset;
1790 	uint64_t txg = tx->tx_txg;
1791 
1792 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1793 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1794 		return;
1795 	}
1796 
1797 	DNODE_VERIFY(dn);
1798 
1799 #ifdef ZFS_DEBUG
1800 	mutex_enter(&dn->dn_mtx);
1801 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1802 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1803 	mutex_exit(&dn->dn_mtx);
1804 #endif
1805 
1806 	/*
1807 	 * Determine old uid/gid when necessary
1808 	 */
1809 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1810 
1811 	multilist_t *dirtylist = &os->os_dirty_dnodes[txg & TXG_MASK];
1812 	multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1813 
1814 	/*
1815 	 * If we are already marked dirty, we're done.
1816 	 */
1817 	if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1818 		multilist_sublist_unlock(mls);
1819 		return;
1820 	}
1821 
1822 	ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1823 	    !avl_is_empty(&dn->dn_dbufs));
1824 	ASSERT(dn->dn_datablksz != 0);
1825 	ASSERT0(dn->dn_next_bonuslen[txg & TXG_MASK]);
1826 	ASSERT0(dn->dn_next_blksz[txg & TXG_MASK]);
1827 	ASSERT0(dn->dn_next_bonustype[txg & TXG_MASK]);
1828 
1829 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1830 	    (u_longlong_t)dn->dn_object, (u_longlong_t)txg);
1831 
1832 	multilist_sublist_insert_head(mls, dn);
1833 
1834 	multilist_sublist_unlock(mls);
1835 
1836 	/*
1837 	 * The dnode maintains a hold on its containing dbuf as
1838 	 * long as there are holds on it.  Each instantiated child
1839 	 * dbuf maintains a hold on the dnode.  When the last child
1840 	 * drops its hold, the dnode will drop its hold on the
1841 	 * containing dbuf. We add a "dirty hold" here so that the
1842 	 * dnode will hang around after we finish processing its
1843 	 * children.
1844 	 */
1845 	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1846 
1847 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1848 
1849 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1850 }
1851 
1852 void
1853 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1854 {
1855 	mutex_enter(&dn->dn_mtx);
1856 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1857 		mutex_exit(&dn->dn_mtx);
1858 		return;
1859 	}
1860 	dn->dn_free_txg = tx->tx_txg;
1861 	mutex_exit(&dn->dn_mtx);
1862 
1863 	dnode_setdirty(dn, tx);
1864 }
1865 
1866 /*
1867  * Try to change the block size for the indicated dnode.  This can only
1868  * succeed if there are no blocks allocated or dirty beyond first block
1869  */
1870 int
1871 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1872 {
1873 	dmu_buf_impl_t *db;
1874 	int err;
1875 
1876 	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1877 	if (size == 0)
1878 		size = SPA_MINBLOCKSIZE;
1879 	else
1880 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1881 
1882 	if (ibs == dn->dn_indblkshift)
1883 		ibs = 0;
1884 
1885 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1886 		return (0);
1887 
1888 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1889 
1890 	/* Check for any allocated blocks beyond the first */
1891 	if (dn->dn_maxblkid != 0)
1892 		goto fail;
1893 
1894 	mutex_enter(&dn->dn_dbufs_mtx);
1895 	for (db = avl_first(&dn->dn_dbufs); db != NULL;
1896 	    db = AVL_NEXT(&dn->dn_dbufs, db)) {
1897 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1898 		    db->db_blkid != DMU_SPILL_BLKID) {
1899 			mutex_exit(&dn->dn_dbufs_mtx);
1900 			goto fail;
1901 		}
1902 	}
1903 	mutex_exit(&dn->dn_dbufs_mtx);
1904 
1905 	if (ibs && dn->dn_nlevels != 1)
1906 		goto fail;
1907 
1908 	/* resize the old block */
1909 	err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1910 	if (err == 0) {
1911 		dbuf_new_size(db, size, tx);
1912 	} else if (err != ENOENT) {
1913 		goto fail;
1914 	}
1915 
1916 	dnode_setdblksz(dn, size);
1917 	dnode_setdirty(dn, tx);
1918 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1919 	if (ibs) {
1920 		dn->dn_indblkshift = ibs;
1921 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1922 	}
1923 	/* release after we have fixed the blocksize in the dnode */
1924 	if (db)
1925 		dbuf_rele(db, FTAG);
1926 
1927 	rw_exit(&dn->dn_struct_rwlock);
1928 	return (0);
1929 
1930 fail:
1931 	rw_exit(&dn->dn_struct_rwlock);
1932 	return (SET_ERROR(ENOTSUP));
1933 }
1934 
1935 static void
1936 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1937 {
1938 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1939 	int old_nlevels = dn->dn_nlevels;
1940 	dmu_buf_impl_t *db;
1941 	list_t *list;
1942 	dbuf_dirty_record_t *new, *dr, *dr_next;
1943 
1944 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1945 
1946 	ASSERT3U(new_nlevels, >, dn->dn_nlevels);
1947 	dn->dn_nlevels = new_nlevels;
1948 
1949 	ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1950 	dn->dn_next_nlevels[txgoff] = new_nlevels;
1951 
1952 	/* dirty the left indirects */
1953 	db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1954 	ASSERT(db != NULL);
1955 	new = dbuf_dirty(db, tx);
1956 	dbuf_rele(db, FTAG);
1957 
1958 	/* transfer the dirty records to the new indirect */
1959 	mutex_enter(&dn->dn_mtx);
1960 	mutex_enter(&new->dt.di.dr_mtx);
1961 	list = &dn->dn_dirty_records[txgoff];
1962 	for (dr = list_head(list); dr; dr = dr_next) {
1963 		dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1964 
1965 		IMPLY(dr->dr_dbuf == NULL, old_nlevels == 1);
1966 		if (dr->dr_dbuf == NULL ||
1967 		    (dr->dr_dbuf->db_level == old_nlevels - 1 &&
1968 		    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1969 		    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID)) {
1970 			list_remove(&dn->dn_dirty_records[txgoff], dr);
1971 			list_insert_tail(&new->dt.di.dr_children, dr);
1972 			dr->dr_parent = new;
1973 		}
1974 	}
1975 	mutex_exit(&new->dt.di.dr_mtx);
1976 	mutex_exit(&dn->dn_mtx);
1977 }
1978 
1979 int
1980 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
1981 {
1982 	int ret = 0;
1983 
1984 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1985 
1986 	if (dn->dn_nlevels == nlevels) {
1987 		ret = 0;
1988 		goto out;
1989 	} else if (nlevels < dn->dn_nlevels) {
1990 		ret = SET_ERROR(EINVAL);
1991 		goto out;
1992 	}
1993 
1994 	dnode_set_nlevels_impl(dn, nlevels, tx);
1995 
1996 out:
1997 	rw_exit(&dn->dn_struct_rwlock);
1998 	return (ret);
1999 }
2000 
2001 /* read-holding callers must not rely on the lock being continuously held */
2002 void
2003 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read,
2004     boolean_t force)
2005 {
2006 	int epbs, new_nlevels;
2007 	uint64_t sz;
2008 
2009 	ASSERT(blkid != DMU_BONUS_BLKID);
2010 
2011 	ASSERT(have_read ?
2012 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
2013 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
2014 
2015 	/*
2016 	 * if we have a read-lock, check to see if we need to do any work
2017 	 * before upgrading to a write-lock.
2018 	 */
2019 	if (have_read) {
2020 		if (blkid <= dn->dn_maxblkid)
2021 			return;
2022 
2023 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
2024 			rw_exit(&dn->dn_struct_rwlock);
2025 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2026 		}
2027 	}
2028 
2029 	/*
2030 	 * Raw sends (indicated by the force flag) require that we take the
2031 	 * given blkid even if the value is lower than the current value.
2032 	 */
2033 	if (!force && blkid <= dn->dn_maxblkid)
2034 		goto out;
2035 
2036 	/*
2037 	 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff]
2038 	 * to indicate that this field is set. This allows us to set the
2039 	 * maxblkid to 0 on an existing object in dnode_sync().
2040 	 */
2041 	dn->dn_maxblkid = blkid;
2042 	dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] =
2043 	    blkid | DMU_NEXT_MAXBLKID_SET;
2044 
2045 	/*
2046 	 * Compute the number of levels necessary to support the new maxblkid.
2047 	 * Raw sends will ensure nlevels is set correctly for us.
2048 	 */
2049 	new_nlevels = 1;
2050 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2051 	for (sz = dn->dn_nblkptr;
2052 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
2053 		new_nlevels++;
2054 
2055 	ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
2056 
2057 	if (!force) {
2058 		if (new_nlevels > dn->dn_nlevels)
2059 			dnode_set_nlevels_impl(dn, new_nlevels, tx);
2060 	} else {
2061 		ASSERT3U(dn->dn_nlevels, >=, new_nlevels);
2062 	}
2063 
2064 out:
2065 	if (have_read)
2066 		rw_downgrade(&dn->dn_struct_rwlock);
2067 }
2068 
2069 static void
2070 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
2071 {
2072 	dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
2073 	if (db != NULL) {
2074 		dmu_buf_will_dirty(&db->db, tx);
2075 		dbuf_rele(db, FTAG);
2076 	}
2077 }
2078 
2079 /*
2080  * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
2081  * and end_blkid.
2082  */
2083 static void
2084 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
2085     dmu_tx_t *tx)
2086 {
2087 	dmu_buf_impl_t *db_search;
2088 	dmu_buf_impl_t *db;
2089 	avl_index_t where;
2090 
2091 	db_search = kmem_zalloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
2092 
2093 	mutex_enter(&dn->dn_dbufs_mtx);
2094 
2095 	db_search->db_level = 1;
2096 	db_search->db_blkid = start_blkid + 1;
2097 	db_search->db_state = DB_SEARCH;
2098 	for (;;) {
2099 
2100 		db = avl_find(&dn->dn_dbufs, db_search, &where);
2101 		if (db == NULL)
2102 			db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
2103 
2104 		if (db == NULL || db->db_level != 1 ||
2105 		    db->db_blkid >= end_blkid) {
2106 			break;
2107 		}
2108 
2109 		/*
2110 		 * Setup the next blkid we want to search for.
2111 		 */
2112 		db_search->db_blkid = db->db_blkid + 1;
2113 		ASSERT3U(db->db_blkid, >=, start_blkid);
2114 
2115 		/*
2116 		 * If the dbuf transitions to DB_EVICTING while we're trying
2117 		 * to dirty it, then we will be unable to discover it in
2118 		 * the dbuf hash table. This will result in a call to
2119 		 * dbuf_create() which needs to acquire the dn_dbufs_mtx
2120 		 * lock. To avoid a deadlock, we drop the lock before
2121 		 * dirtying the level-1 dbuf.
2122 		 */
2123 		mutex_exit(&dn->dn_dbufs_mtx);
2124 		dnode_dirty_l1(dn, db->db_blkid, tx);
2125 		mutex_enter(&dn->dn_dbufs_mtx);
2126 	}
2127 
2128 #ifdef ZFS_DEBUG
2129 	/*
2130 	 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
2131 	 */
2132 	db_search->db_level = 1;
2133 	db_search->db_blkid = start_blkid + 1;
2134 	db_search->db_state = DB_SEARCH;
2135 	db = avl_find(&dn->dn_dbufs, db_search, &where);
2136 	if (db == NULL)
2137 		db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
2138 	for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
2139 		if (db->db_level != 1 || db->db_blkid >= end_blkid)
2140 			break;
2141 		if (db->db_state != DB_EVICTING)
2142 			ASSERT(db->db_dirtycnt > 0);
2143 	}
2144 #endif
2145 	kmem_free(db_search, sizeof (dmu_buf_impl_t));
2146 	mutex_exit(&dn->dn_dbufs_mtx);
2147 }
2148 
2149 void
2150 dnode_set_dirtyctx(dnode_t *dn, dmu_tx_t *tx, const void *tag)
2151 {
2152 	/*
2153 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
2154 	 * initialize the objset.
2155 	 */
2156 	if (dn->dn_dirtyctx == DN_UNDIRTIED) {
2157 		dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2158 
2159 		if (ds != NULL) {
2160 			rrw_enter(&ds->ds_bp_rwlock, RW_READER, tag);
2161 		}
2162 		if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
2163 			if (dmu_tx_is_syncing(tx))
2164 				dn->dn_dirtyctx = DN_DIRTY_SYNC;
2165 			else
2166 				dn->dn_dirtyctx = DN_DIRTY_OPEN;
2167 			dn->dn_dirtyctx_firstset = tag;
2168 		}
2169 		if (ds != NULL) {
2170 			rrw_exit(&ds->ds_bp_rwlock, tag);
2171 		}
2172 	}
2173 }
2174 
2175 static void
2176 dnode_partial_zero(dnode_t *dn, uint64_t off, uint64_t blkoff, uint64_t len,
2177     dmu_tx_t *tx)
2178 {
2179 	dmu_buf_impl_t *db;
2180 	int res;
2181 
2182 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2183 	res = dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off), TRUE, FALSE,
2184 	    FTAG, &db);
2185 	rw_exit(&dn->dn_struct_rwlock);
2186 	if (res == 0) {
2187 		db_lock_type_t dblt;
2188 		boolean_t dirty;
2189 
2190 		dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
2191 		/* don't dirty if not on disk and not dirty */
2192 		dirty = !list_is_empty(&db->db_dirty_records) ||
2193 		    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr));
2194 		dmu_buf_unlock_parent(db, dblt, FTAG);
2195 		if (dirty) {
2196 			caddr_t data;
2197 
2198 			dmu_buf_will_dirty(&db->db, tx);
2199 			data = db->db.db_data;
2200 			memset(data + blkoff, 0, len);
2201 		}
2202 		dbuf_rele(db, FTAG);
2203 	}
2204 }
2205 
2206 void
2207 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
2208 {
2209 	uint64_t blkoff, blkid, nblks;
2210 	int blksz, blkshift, head, tail;
2211 	int trunc = FALSE;
2212 	int epbs;
2213 
2214 	blksz = dn->dn_datablksz;
2215 	blkshift = dn->dn_datablkshift;
2216 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2217 
2218 	if (len == DMU_OBJECT_END) {
2219 		len = UINT64_MAX - off;
2220 		trunc = TRUE;
2221 	}
2222 
2223 	/*
2224 	 * First, block align the region to free:
2225 	 */
2226 	if (ISP2(blksz)) {
2227 		head = P2NPHASE(off, blksz);
2228 		blkoff = P2PHASE(off, blksz);
2229 		if ((off >> blkshift) > dn->dn_maxblkid)
2230 			return;
2231 	} else {
2232 		ASSERT(dn->dn_maxblkid == 0);
2233 		if (off == 0 && len >= blksz) {
2234 			/*
2235 			 * Freeing the whole block; fast-track this request.
2236 			 */
2237 			blkid = 0;
2238 			nblks = 1;
2239 			if (dn->dn_nlevels > 1) {
2240 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2241 				dnode_dirty_l1(dn, 0, tx);
2242 				rw_exit(&dn->dn_struct_rwlock);
2243 			}
2244 			goto done;
2245 		} else if (off >= blksz) {
2246 			/* Freeing past end-of-data */
2247 			return;
2248 		} else {
2249 			/* Freeing part of the block. */
2250 			head = blksz - off;
2251 			ASSERT3U(head, >, 0);
2252 		}
2253 		blkoff = off;
2254 	}
2255 	/* zero out any partial block data at the start of the range */
2256 	if (head) {
2257 		ASSERT3U(blkoff + head, ==, blksz);
2258 		if (len < head)
2259 			head = len;
2260 		dnode_partial_zero(dn, off, blkoff, head, tx);
2261 		off += head;
2262 		len -= head;
2263 	}
2264 
2265 	/* If the range was less than one block, we're done */
2266 	if (len == 0)
2267 		return;
2268 
2269 	/* If the remaining range is past end of file, we're done */
2270 	if ((off >> blkshift) > dn->dn_maxblkid)
2271 		return;
2272 
2273 	ASSERT(ISP2(blksz));
2274 	if (trunc)
2275 		tail = 0;
2276 	else
2277 		tail = P2PHASE(len, blksz);
2278 
2279 	ASSERT0(P2PHASE(off, blksz));
2280 	/* zero out any partial block data at the end of the range */
2281 	if (tail) {
2282 		if (len < tail)
2283 			tail = len;
2284 		dnode_partial_zero(dn, off + len, 0, tail, tx);
2285 		len -= tail;
2286 	}
2287 
2288 	/* If the range did not include a full block, we are done */
2289 	if (len == 0)
2290 		return;
2291 
2292 	ASSERT(IS_P2ALIGNED(off, blksz));
2293 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2294 	blkid = off >> blkshift;
2295 	nblks = len >> blkshift;
2296 	if (trunc)
2297 		nblks += 1;
2298 
2299 	/*
2300 	 * Dirty all the indirect blocks in this range.  Note that only
2301 	 * the first and last indirect blocks can actually be written
2302 	 * (if they were partially freed) -- they must be dirtied, even if
2303 	 * they do not exist on disk yet.  The interior blocks will
2304 	 * be freed by free_children(), so they will not actually be written.
2305 	 * Even though these interior blocks will not be written, we
2306 	 * dirty them for two reasons:
2307 	 *
2308 	 *  - It ensures that the indirect blocks remain in memory until
2309 	 *    syncing context.  (They have already been prefetched by
2310 	 *    dmu_tx_hold_free(), so we don't have to worry about reading
2311 	 *    them serially here.)
2312 	 *
2313 	 *  - The dirty space accounting will put pressure on the txg sync
2314 	 *    mechanism to begin syncing, and to delay transactions if there
2315 	 *    is a large amount of freeing.  Even though these indirect
2316 	 *    blocks will not be written, we could need to write the same
2317 	 *    amount of space if we copy the freed BPs into deadlists.
2318 	 */
2319 	if (dn->dn_nlevels > 1) {
2320 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2321 		uint64_t first, last;
2322 
2323 		first = blkid >> epbs;
2324 		dnode_dirty_l1(dn, first, tx);
2325 		if (trunc)
2326 			last = dn->dn_maxblkid >> epbs;
2327 		else
2328 			last = (blkid + nblks - 1) >> epbs;
2329 		if (last != first)
2330 			dnode_dirty_l1(dn, last, tx);
2331 
2332 		dnode_dirty_l1range(dn, first, last, tx);
2333 
2334 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2335 		    SPA_BLKPTRSHIFT;
2336 		for (uint64_t i = first + 1; i < last; i++) {
2337 			/*
2338 			 * Set i to the blockid of the next non-hole
2339 			 * level-1 indirect block at or after i.  Note
2340 			 * that dnode_next_offset() operates in terms of
2341 			 * level-0-equivalent bytes.
2342 			 */
2343 			uint64_t ibyte = i << shift;
2344 			int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2345 			    &ibyte, 2, 1, 0);
2346 			i = ibyte >> shift;
2347 			if (i >= last)
2348 				break;
2349 
2350 			/*
2351 			 * Normally we should not see an error, either
2352 			 * from dnode_next_offset() or dbuf_hold_level()
2353 			 * (except for ESRCH from dnode_next_offset).
2354 			 * If there is an i/o error, then when we read
2355 			 * this block in syncing context, it will use
2356 			 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2357 			 * to the "failmode" property.  dnode_next_offset()
2358 			 * doesn't have a flag to indicate MUSTSUCCEED.
2359 			 */
2360 			if (err != 0)
2361 				break;
2362 
2363 			dnode_dirty_l1(dn, i, tx);
2364 		}
2365 		rw_exit(&dn->dn_struct_rwlock);
2366 	}
2367 
2368 done:
2369 	/*
2370 	 * Add this range to the dnode range list.
2371 	 * We will finish up this free operation in the syncing phase.
2372 	 */
2373 	mutex_enter(&dn->dn_mtx);
2374 	{
2375 		int txgoff = tx->tx_txg & TXG_MASK;
2376 		if (dn->dn_free_ranges[txgoff] == NULL) {
2377 			dn->dn_free_ranges[txgoff] = range_tree_create(NULL,
2378 			    RANGE_SEG64, NULL, 0, 0);
2379 		}
2380 		range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2381 		range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2382 	}
2383 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2384 	    (u_longlong_t)blkid, (u_longlong_t)nblks,
2385 	    (u_longlong_t)tx->tx_txg);
2386 	mutex_exit(&dn->dn_mtx);
2387 
2388 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2389 	dnode_setdirty(dn, tx);
2390 }
2391 
2392 static boolean_t
2393 dnode_spill_freed(dnode_t *dn)
2394 {
2395 	int i;
2396 
2397 	mutex_enter(&dn->dn_mtx);
2398 	for (i = 0; i < TXG_SIZE; i++) {
2399 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2400 			break;
2401 	}
2402 	mutex_exit(&dn->dn_mtx);
2403 	return (i < TXG_SIZE);
2404 }
2405 
2406 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2407 uint64_t
2408 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2409 {
2410 	int i;
2411 
2412 	if (blkid == DMU_BONUS_BLKID)
2413 		return (FALSE);
2414 
2415 	if (dn->dn_free_txg)
2416 		return (TRUE);
2417 
2418 	if (blkid == DMU_SPILL_BLKID)
2419 		return (dnode_spill_freed(dn));
2420 
2421 	mutex_enter(&dn->dn_mtx);
2422 	for (i = 0; i < TXG_SIZE; i++) {
2423 		if (dn->dn_free_ranges[i] != NULL &&
2424 		    range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2425 			break;
2426 	}
2427 	mutex_exit(&dn->dn_mtx);
2428 	return (i < TXG_SIZE);
2429 }
2430 
2431 /* call from syncing context when we actually write/free space for this dnode */
2432 void
2433 dnode_diduse_space(dnode_t *dn, int64_t delta)
2434 {
2435 	uint64_t space;
2436 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2437 	    dn, dn->dn_phys,
2438 	    (u_longlong_t)dn->dn_phys->dn_used,
2439 	    (longlong_t)delta);
2440 
2441 	mutex_enter(&dn->dn_mtx);
2442 	space = DN_USED_BYTES(dn->dn_phys);
2443 	if (delta > 0) {
2444 		ASSERT3U(space + delta, >=, space); /* no overflow */
2445 	} else {
2446 		ASSERT3U(space, >=, -delta); /* no underflow */
2447 	}
2448 	space += delta;
2449 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2450 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2451 		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2452 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2453 	} else {
2454 		dn->dn_phys->dn_used = space;
2455 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2456 	}
2457 	mutex_exit(&dn->dn_mtx);
2458 }
2459 
2460 /*
2461  * Scans a block at the indicated "level" looking for a hole or data,
2462  * depending on 'flags'.
2463  *
2464  * If level > 0, then we are scanning an indirect block looking at its
2465  * pointers.  If level == 0, then we are looking at a block of dnodes.
2466  *
2467  * If we don't find what we are looking for in the block, we return ESRCH.
2468  * Otherwise, return with *offset pointing to the beginning (if searching
2469  * forwards) or end (if searching backwards) of the range covered by the
2470  * block pointer we matched on (or dnode).
2471  *
2472  * The basic search algorithm used below by dnode_next_offset() is to
2473  * use this function to search up the block tree (widen the search) until
2474  * we find something (i.e., we don't return ESRCH) and then search back
2475  * down the tree (narrow the search) until we reach our original search
2476  * level.
2477  */
2478 static int
2479 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2480     int lvl, uint64_t blkfill, uint64_t txg)
2481 {
2482 	dmu_buf_impl_t *db = NULL;
2483 	void *data = NULL;
2484 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2485 	uint64_t epb = 1ULL << epbs;
2486 	uint64_t minfill, maxfill;
2487 	boolean_t hole;
2488 	int i, inc, error, span;
2489 
2490 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2491 
2492 	hole = ((flags & DNODE_FIND_HOLE) != 0);
2493 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2494 	ASSERT(txg == 0 || !hole);
2495 
2496 	if (lvl == dn->dn_phys->dn_nlevels) {
2497 		error = 0;
2498 		epb = dn->dn_phys->dn_nblkptr;
2499 		data = dn->dn_phys->dn_blkptr;
2500 	} else {
2501 		uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2502 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2503 		if (error) {
2504 			if (error != ENOENT)
2505 				return (error);
2506 			if (hole)
2507 				return (0);
2508 			/*
2509 			 * This can only happen when we are searching up
2510 			 * the block tree for data.  We don't really need to
2511 			 * adjust the offset, as we will just end up looking
2512 			 * at the pointer to this block in its parent, and its
2513 			 * going to be unallocated, so we will skip over it.
2514 			 */
2515 			return (SET_ERROR(ESRCH));
2516 		}
2517 		error = dbuf_read(db, NULL,
2518 		    DB_RF_CANFAIL | DB_RF_HAVESTRUCT |
2519 		    DB_RF_NO_DECRYPT | DB_RF_NOPREFETCH);
2520 		if (error) {
2521 			dbuf_rele(db, FTAG);
2522 			return (error);
2523 		}
2524 		data = db->db.db_data;
2525 		rw_enter(&db->db_rwlock, RW_READER);
2526 	}
2527 
2528 	if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2529 	    db->db_blkptr->blk_birth <= txg ||
2530 	    BP_IS_HOLE(db->db_blkptr))) {
2531 		/*
2532 		 * This can only happen when we are searching up the tree
2533 		 * and these conditions mean that we need to keep climbing.
2534 		 */
2535 		error = SET_ERROR(ESRCH);
2536 	} else if (lvl == 0) {
2537 		dnode_phys_t *dnp = data;
2538 
2539 		ASSERT(dn->dn_type == DMU_OT_DNODE);
2540 		ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2541 
2542 		for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2543 		    i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2544 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2545 				break;
2546 		}
2547 
2548 		if (i == blkfill)
2549 			error = SET_ERROR(ESRCH);
2550 
2551 		*offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2552 		    (i << DNODE_SHIFT);
2553 	} else {
2554 		blkptr_t *bp = data;
2555 		uint64_t start = *offset;
2556 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
2557 		minfill = 0;
2558 		maxfill = blkfill << ((lvl - 1) * epbs);
2559 
2560 		if (hole)
2561 			maxfill--;
2562 		else
2563 			minfill++;
2564 
2565 		if (span >= 8 * sizeof (*offset)) {
2566 			/* This only happens on the highest indirection level */
2567 			ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1);
2568 			*offset = 0;
2569 		} else {
2570 			*offset = *offset >> span;
2571 		}
2572 
2573 		for (i = BF64_GET(*offset, 0, epbs);
2574 		    i >= 0 && i < epb; i += inc) {
2575 			if (BP_GET_FILL(&bp[i]) >= minfill &&
2576 			    BP_GET_FILL(&bp[i]) <= maxfill &&
2577 			    (hole || bp[i].blk_birth > txg))
2578 				break;
2579 			if (inc > 0 || *offset > 0)
2580 				*offset += inc;
2581 		}
2582 
2583 		if (span >= 8 * sizeof (*offset)) {
2584 			*offset = start;
2585 		} else {
2586 			*offset = *offset << span;
2587 		}
2588 
2589 		if (inc < 0) {
2590 			/* traversing backwards; position offset at the end */
2591 			ASSERT3U(*offset, <=, start);
2592 			*offset = MIN(*offset + (1ULL << span) - 1, start);
2593 		} else if (*offset < start) {
2594 			*offset = start;
2595 		}
2596 		if (i < 0 || i >= epb)
2597 			error = SET_ERROR(ESRCH);
2598 	}
2599 
2600 	if (db != NULL) {
2601 		rw_exit(&db->db_rwlock);
2602 		dbuf_rele(db, FTAG);
2603 	}
2604 
2605 	return (error);
2606 }
2607 
2608 /*
2609  * Find the next hole, data, or sparse region at or after *offset.
2610  * The value 'blkfill' tells us how many items we expect to find
2611  * in an L0 data block; this value is 1 for normal objects,
2612  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2613  * DNODES_PER_BLOCK when searching for sparse regions thereof.
2614  *
2615  * Examples:
2616  *
2617  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2618  *	Finds the next/previous hole/data in a file.
2619  *	Used in dmu_offset_next().
2620  *
2621  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2622  *	Finds the next free/allocated dnode an objset's meta-dnode.
2623  *	Only finds objects that have new contents since txg (ie.
2624  *	bonus buffer changes and content removal are ignored).
2625  *	Used in dmu_object_next().
2626  *
2627  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2628  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
2629  *	Used in dmu_object_alloc().
2630  */
2631 int
2632 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2633     int minlvl, uint64_t blkfill, uint64_t txg)
2634 {
2635 	uint64_t initial_offset = *offset;
2636 	int lvl, maxlvl;
2637 	int error = 0;
2638 
2639 	if (!(flags & DNODE_FIND_HAVELOCK))
2640 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
2641 
2642 	if (dn->dn_phys->dn_nlevels == 0) {
2643 		error = SET_ERROR(ESRCH);
2644 		goto out;
2645 	}
2646 
2647 	if (dn->dn_datablkshift == 0) {
2648 		if (*offset < dn->dn_datablksz) {
2649 			if (flags & DNODE_FIND_HOLE)
2650 				*offset = dn->dn_datablksz;
2651 		} else {
2652 			error = SET_ERROR(ESRCH);
2653 		}
2654 		goto out;
2655 	}
2656 
2657 	maxlvl = dn->dn_phys->dn_nlevels;
2658 
2659 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2660 		error = dnode_next_offset_level(dn,
2661 		    flags, offset, lvl, blkfill, txg);
2662 		if (error != ESRCH)
2663 			break;
2664 	}
2665 
2666 	while (error == 0 && --lvl >= minlvl) {
2667 		error = dnode_next_offset_level(dn,
2668 		    flags, offset, lvl, blkfill, txg);
2669 	}
2670 
2671 	/*
2672 	 * There's always a "virtual hole" at the end of the object, even
2673 	 * if all BP's which physically exist are non-holes.
2674 	 */
2675 	if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2676 	    minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2677 		error = 0;
2678 	}
2679 
2680 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2681 	    initial_offset < *offset : initial_offset > *offset))
2682 		error = SET_ERROR(ESRCH);
2683 out:
2684 	if (!(flags & DNODE_FIND_HAVELOCK))
2685 		rw_exit(&dn->dn_struct_rwlock);
2686 
2687 	return (error);
2688 }
2689 
2690 #if defined(_KERNEL)
2691 EXPORT_SYMBOL(dnode_hold);
2692 EXPORT_SYMBOL(dnode_rele);
2693 EXPORT_SYMBOL(dnode_set_nlevels);
2694 EXPORT_SYMBOL(dnode_set_blksz);
2695 EXPORT_SYMBOL(dnode_free_range);
2696 EXPORT_SYMBOL(dnode_evict_dbufs);
2697 EXPORT_SYMBOL(dnode_evict_bonus);
2698 #endif
2699 
2700 ZFS_MODULE_PARAM(zfs, zfs_, default_bs, INT, ZMOD_RW,
2701 	"Default dnode block shift");
2702 ZFS_MODULE_PARAM(zfs, zfs_, default_ibs, INT, ZMOD_RW,
2703 	"Default dnode indirect block shift");
2704