xref: /freebsd/sys/contrib/openzfs/module/zfs/dnode.c (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
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 http://www.opensolaris.org/os/licensing.
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 static kstat_t *dnode_ksp;
75 static kmem_cache_t *dnode_cache;
76 
77 static dnode_phys_t dnode_phys_zero __maybe_unused;
78 
79 int zfs_default_bs = SPA_MINBLOCKSHIFT;
80 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
81 
82 #ifdef	_KERNEL
83 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
84 #endif /* _KERNEL */
85 
86 static int
87 dbuf_compare(const void *x1, const void *x2)
88 {
89 	const dmu_buf_impl_t *d1 = x1;
90 	const dmu_buf_impl_t *d2 = x2;
91 
92 	int cmp = TREE_CMP(d1->db_level, d2->db_level);
93 	if (likely(cmp))
94 		return (cmp);
95 
96 	cmp = TREE_CMP(d1->db_blkid, d2->db_blkid);
97 	if (likely(cmp))
98 		return (cmp);
99 
100 	if (d1->db_state == DB_SEARCH) {
101 		ASSERT3S(d2->db_state, !=, DB_SEARCH);
102 		return (-1);
103 	} else if (d2->db_state == DB_SEARCH) {
104 		ASSERT3S(d1->db_state, !=, DB_SEARCH);
105 		return (1);
106 	}
107 
108 	return (TREE_PCMP(d1, d2));
109 }
110 
111 /* ARGSUSED */
112 static int
113 dnode_cons(void *arg, void *unused, int kmflag)
114 {
115 	dnode_t *dn = arg;
116 	int i;
117 
118 	rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL);
119 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
120 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
121 	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
122 	cv_init(&dn->dn_nodnholds, NULL, CV_DEFAULT, NULL);
123 
124 	/*
125 	 * Every dbuf has a reference, and dropping a tracked reference is
126 	 * O(number of references), so don't track dn_holds.
127 	 */
128 	zfs_refcount_create_untracked(&dn->dn_holds);
129 	zfs_refcount_create(&dn->dn_tx_holds);
130 	list_link_init(&dn->dn_link);
131 
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));
140 
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));
147 	}
148 
149 	dn->dn_allocated_txg = 0;
150 	dn->dn_free_txg = 0;
151 	dn->dn_assigned_txg = 0;
152 	dn->dn_dirty_txg = 0;
153 	dn->dn_dirtyctx = 0;
154 	dn->dn_dirtyctx_firstset = NULL;
155 	dn->dn_bonus = NULL;
156 	dn->dn_have_spill = B_FALSE;
157 	dn->dn_zio = NULL;
158 	dn->dn_oldused = 0;
159 	dn->dn_oldflags = 0;
160 	dn->dn_olduid = 0;
161 	dn->dn_oldgid = 0;
162 	dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
163 	dn->dn_newuid = 0;
164 	dn->dn_newgid = 0;
165 	dn->dn_newprojid = ZFS_DEFAULT_PROJID;
166 	dn->dn_id_flags = 0;
167 
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));
171 
172 	dn->dn_moved = 0;
173 	return (0);
174 }
175 
176 /* ARGSUSED */
177 static void
178 dnode_dest(void *arg, void *unused)
179 {
180 	int i;
181 	dnode_t *dn = arg;
182 
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 	cv_destroy(&dn->dn_nodnholds);
188 	zfs_refcount_destroy(&dn->dn_holds);
189 	zfs_refcount_destroy(&dn->dn_tx_holds);
190 	ASSERT(!list_link_active(&dn->dn_link));
191 
192 	for (i = 0; i < TXG_SIZE; i++) {
193 		ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
194 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
195 		list_destroy(&dn->dn_dirty_records[i]);
196 		ASSERT0(dn->dn_next_nblkptr[i]);
197 		ASSERT0(dn->dn_next_nlevels[i]);
198 		ASSERT0(dn->dn_next_indblkshift[i]);
199 		ASSERT0(dn->dn_next_bonustype[i]);
200 		ASSERT0(dn->dn_rm_spillblk[i]);
201 		ASSERT0(dn->dn_next_bonuslen[i]);
202 		ASSERT0(dn->dn_next_blksz[i]);
203 		ASSERT0(dn->dn_next_maxblkid[i]);
204 	}
205 
206 	ASSERT0(dn->dn_allocated_txg);
207 	ASSERT0(dn->dn_free_txg);
208 	ASSERT0(dn->dn_assigned_txg);
209 	ASSERT0(dn->dn_dirty_txg);
210 	ASSERT0(dn->dn_dirtyctx);
211 	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
212 	ASSERT3P(dn->dn_bonus, ==, NULL);
213 	ASSERT(!dn->dn_have_spill);
214 	ASSERT3P(dn->dn_zio, ==, NULL);
215 	ASSERT0(dn->dn_oldused);
216 	ASSERT0(dn->dn_oldflags);
217 	ASSERT0(dn->dn_olduid);
218 	ASSERT0(dn->dn_oldgid);
219 	ASSERT0(dn->dn_oldprojid);
220 	ASSERT0(dn->dn_newuid);
221 	ASSERT0(dn->dn_newgid);
222 	ASSERT0(dn->dn_newprojid);
223 	ASSERT0(dn->dn_id_flags);
224 
225 	ASSERT0(dn->dn_dbufs_count);
226 	avl_destroy(&dn->dn_dbufs);
227 }
228 
229 void
230 dnode_init(void)
231 {
232 	ASSERT(dnode_cache == NULL);
233 	dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
234 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
235 	kmem_cache_set_move(dnode_cache, dnode_move);
236 
237 	dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
238 	    KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
239 	    KSTAT_FLAG_VIRTUAL);
240 	if (dnode_ksp != NULL) {
241 		dnode_ksp->ks_data = &dnode_stats;
242 		kstat_install(dnode_ksp);
243 	}
244 }
245 
246 void
247 dnode_fini(void)
248 {
249 	if (dnode_ksp != NULL) {
250 		kstat_delete(dnode_ksp);
251 		dnode_ksp = NULL;
252 	}
253 
254 	kmem_cache_destroy(dnode_cache);
255 	dnode_cache = NULL;
256 }
257 
258 
259 #ifdef ZFS_DEBUG
260 void
261 dnode_verify(dnode_t *dn)
262 {
263 	int drop_struct_lock = FALSE;
264 
265 	ASSERT(dn->dn_phys);
266 	ASSERT(dn->dn_objset);
267 	ASSERT(dn->dn_handle->dnh_dnode == dn);
268 
269 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
270 
271 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
272 		return;
273 
274 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
275 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
276 		drop_struct_lock = TRUE;
277 	}
278 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
279 		int i;
280 		int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
281 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
282 		if (dn->dn_datablkshift) {
283 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
284 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
285 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
286 		}
287 		ASSERT3U(dn->dn_nlevels, <=, 30);
288 		ASSERT(DMU_OT_IS_VALID(dn->dn_type));
289 		ASSERT3U(dn->dn_nblkptr, >=, 1);
290 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
291 		ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
292 		ASSERT3U(dn->dn_datablksz, ==,
293 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
294 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
295 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
296 		    dn->dn_bonuslen, <=, max_bonuslen);
297 		for (i = 0; i < TXG_SIZE; i++) {
298 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
299 		}
300 	}
301 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
302 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
303 	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
304 	if (dn->dn_dbuf != NULL) {
305 		ASSERT3P(dn->dn_phys, ==,
306 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
307 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
308 	}
309 	if (drop_struct_lock)
310 		rw_exit(&dn->dn_struct_rwlock);
311 }
312 #endif
313 
314 void
315 dnode_byteswap(dnode_phys_t *dnp)
316 {
317 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
318 	int i;
319 
320 	if (dnp->dn_type == DMU_OT_NONE) {
321 		bzero(dnp, sizeof (dnode_phys_t));
322 		return;
323 	}
324 
325 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
326 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
327 	dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
328 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
329 	dnp->dn_used = BSWAP_64(dnp->dn_used);
330 
331 	/*
332 	 * dn_nblkptr is only one byte, so it's OK to read it in either
333 	 * byte order.  We can't read dn_bouslen.
334 	 */
335 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
336 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
337 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
338 		buf64[i] = BSWAP_64(buf64[i]);
339 
340 	/*
341 	 * OK to check dn_bonuslen for zero, because it won't matter if
342 	 * we have the wrong byte order.  This is necessary because the
343 	 * dnode dnode is smaller than a regular dnode.
344 	 */
345 	if (dnp->dn_bonuslen != 0) {
346 		/*
347 		 * Note that the bonus length calculated here may be
348 		 * longer than the actual bonus buffer.  This is because
349 		 * we always put the bonus buffer after the last block
350 		 * pointer (instead of packing it against the end of the
351 		 * dnode buffer).
352 		 */
353 		int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
354 		int slots = dnp->dn_extra_slots + 1;
355 		size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
356 		dmu_object_byteswap_t byteswap;
357 		ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
358 		byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
359 		dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
360 	}
361 
362 	/* Swap SPILL block if we have one */
363 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
364 		byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
365 }
366 
367 void
368 dnode_buf_byteswap(void *vbuf, size_t size)
369 {
370 	int i = 0;
371 
372 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
373 	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
374 
375 	while (i < size) {
376 		dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
377 		dnode_byteswap(dnp);
378 
379 		i += DNODE_MIN_SIZE;
380 		if (dnp->dn_type != DMU_OT_NONE)
381 			i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
382 	}
383 }
384 
385 void
386 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
387 {
388 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
389 
390 	dnode_setdirty(dn, tx);
391 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
392 	ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
393 	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
394 
395 	if (newsize < dn->dn_bonuslen) {
396 		/* clear any data after the end of the new size */
397 		size_t diff = dn->dn_bonuslen - newsize;
398 		char *data_end = ((char *)dn->dn_bonus->db.db_data) + newsize;
399 		bzero(data_end, diff);
400 	}
401 
402 	dn->dn_bonuslen = newsize;
403 	if (newsize == 0)
404 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
405 	else
406 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
407 	rw_exit(&dn->dn_struct_rwlock);
408 }
409 
410 void
411 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
412 {
413 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
414 	dnode_setdirty(dn, tx);
415 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
416 	dn->dn_bonustype = newtype;
417 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
418 	rw_exit(&dn->dn_struct_rwlock);
419 }
420 
421 void
422 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
423 {
424 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
425 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
426 	dnode_setdirty(dn, tx);
427 	dn->dn_rm_spillblk[tx->tx_txg & TXG_MASK] = DN_KILL_SPILLBLK;
428 	dn->dn_have_spill = B_FALSE;
429 }
430 
431 static void
432 dnode_setdblksz(dnode_t *dn, int size)
433 {
434 	ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
435 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
436 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
437 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
438 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
439 	dn->dn_datablksz = size;
440 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
441 	dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
442 }
443 
444 static dnode_t *
445 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
446     uint64_t object, dnode_handle_t *dnh)
447 {
448 	dnode_t *dn;
449 
450 	dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
451 	dn->dn_moved = 0;
452 
453 	/*
454 	 * Defer setting dn_objset until the dnode is ready to be a candidate
455 	 * for the dnode_move() callback.
456 	 */
457 	dn->dn_object = object;
458 	dn->dn_dbuf = db;
459 	dn->dn_handle = dnh;
460 	dn->dn_phys = dnp;
461 
462 	if (dnp->dn_datablkszsec) {
463 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
464 	} else {
465 		dn->dn_datablksz = 0;
466 		dn->dn_datablkszsec = 0;
467 		dn->dn_datablkshift = 0;
468 	}
469 	dn->dn_indblkshift = dnp->dn_indblkshift;
470 	dn->dn_nlevels = dnp->dn_nlevels;
471 	dn->dn_type = dnp->dn_type;
472 	dn->dn_nblkptr = dnp->dn_nblkptr;
473 	dn->dn_checksum = dnp->dn_checksum;
474 	dn->dn_compress = dnp->dn_compress;
475 	dn->dn_bonustype = dnp->dn_bonustype;
476 	dn->dn_bonuslen = dnp->dn_bonuslen;
477 	dn->dn_num_slots = dnp->dn_extra_slots + 1;
478 	dn->dn_maxblkid = dnp->dn_maxblkid;
479 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
480 	dn->dn_id_flags = 0;
481 
482 	dmu_zfetch_init(&dn->dn_zfetch, dn);
483 
484 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
485 	ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
486 	ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
487 
488 	mutex_enter(&os->os_lock);
489 
490 	/*
491 	 * Exclude special dnodes from os_dnodes so an empty os_dnodes
492 	 * signifies that the special dnodes have no references from
493 	 * their children (the entries in os_dnodes).  This allows
494 	 * dnode_destroy() to easily determine if the last child has
495 	 * been removed and then complete eviction of the objset.
496 	 */
497 	if (!DMU_OBJECT_IS_SPECIAL(object))
498 		list_insert_head(&os->os_dnodes, dn);
499 	membar_producer();
500 
501 	/*
502 	 * Everything else must be valid before assigning dn_objset
503 	 * makes the dnode eligible for dnode_move().
504 	 */
505 	dn->dn_objset = os;
506 
507 	dnh->dnh_dnode = dn;
508 	mutex_exit(&os->os_lock);
509 
510 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
511 
512 	return (dn);
513 }
514 
515 /*
516  * Caller must be holding the dnode handle, which is released upon return.
517  */
518 static void
519 dnode_destroy(dnode_t *dn)
520 {
521 	objset_t *os = dn->dn_objset;
522 	boolean_t complete_os_eviction = B_FALSE;
523 
524 	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
525 
526 	mutex_enter(&os->os_lock);
527 	POINTER_INVALIDATE(&dn->dn_objset);
528 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
529 		list_remove(&os->os_dnodes, dn);
530 		complete_os_eviction =
531 		    list_is_empty(&os->os_dnodes) &&
532 		    list_link_active(&os->os_evicting_node);
533 	}
534 	mutex_exit(&os->os_lock);
535 
536 	/* the dnode can no longer move, so we can release the handle */
537 	if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
538 		zrl_remove(&dn->dn_handle->dnh_zrlock);
539 
540 	dn->dn_allocated_txg = 0;
541 	dn->dn_free_txg = 0;
542 	dn->dn_assigned_txg = 0;
543 	dn->dn_dirty_txg = 0;
544 
545 	dn->dn_dirtyctx = 0;
546 	dn->dn_dirtyctx_firstset = NULL;
547 	if (dn->dn_bonus != NULL) {
548 		mutex_enter(&dn->dn_bonus->db_mtx);
549 		dbuf_destroy(dn->dn_bonus);
550 		dn->dn_bonus = NULL;
551 	}
552 	dn->dn_zio = NULL;
553 
554 	dn->dn_have_spill = B_FALSE;
555 	dn->dn_oldused = 0;
556 	dn->dn_oldflags = 0;
557 	dn->dn_olduid = 0;
558 	dn->dn_oldgid = 0;
559 	dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
560 	dn->dn_newuid = 0;
561 	dn->dn_newgid = 0;
562 	dn->dn_newprojid = ZFS_DEFAULT_PROJID;
563 	dn->dn_id_flags = 0;
564 
565 	dmu_zfetch_fini(&dn->dn_zfetch);
566 	kmem_cache_free(dnode_cache, dn);
567 	arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
568 
569 	if (complete_os_eviction)
570 		dmu_objset_evict_done(os);
571 }
572 
573 void
574 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
575     dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
576 {
577 	int i;
578 
579 	ASSERT3U(dn_slots, >, 0);
580 	ASSERT3U(dn_slots << DNODE_SHIFT, <=,
581 	    spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
582 	ASSERT3U(blocksize, <=,
583 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
584 	if (blocksize == 0)
585 		blocksize = 1 << zfs_default_bs;
586 	else
587 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
588 
589 	if (ibs == 0)
590 		ibs = zfs_default_ibs;
591 
592 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
593 
594 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
595 	    dn->dn_objset, dn->dn_object, tx->tx_txg, blocksize, ibs, dn_slots);
596 	DNODE_STAT_BUMP(dnode_allocate);
597 
598 	ASSERT(dn->dn_type == DMU_OT_NONE);
599 	ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
600 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
601 	ASSERT(ot != DMU_OT_NONE);
602 	ASSERT(DMU_OT_IS_VALID(ot));
603 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
604 	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
605 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
606 	ASSERT(DMU_OT_IS_VALID(bonustype));
607 	ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
608 	ASSERT(dn->dn_type == DMU_OT_NONE);
609 	ASSERT0(dn->dn_maxblkid);
610 	ASSERT0(dn->dn_allocated_txg);
611 	ASSERT0(dn->dn_assigned_txg);
612 	ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
613 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
614 	ASSERT(avl_is_empty(&dn->dn_dbufs));
615 
616 	for (i = 0; i < TXG_SIZE; i++) {
617 		ASSERT0(dn->dn_next_nblkptr[i]);
618 		ASSERT0(dn->dn_next_nlevels[i]);
619 		ASSERT0(dn->dn_next_indblkshift[i]);
620 		ASSERT0(dn->dn_next_bonuslen[i]);
621 		ASSERT0(dn->dn_next_bonustype[i]);
622 		ASSERT0(dn->dn_rm_spillblk[i]);
623 		ASSERT0(dn->dn_next_blksz[i]);
624 		ASSERT0(dn->dn_next_maxblkid[i]);
625 		ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
626 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
627 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
628 	}
629 
630 	dn->dn_type = ot;
631 	dnode_setdblksz(dn, blocksize);
632 	dn->dn_indblkshift = ibs;
633 	dn->dn_nlevels = 1;
634 	dn->dn_num_slots = dn_slots;
635 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
636 		dn->dn_nblkptr = 1;
637 	else {
638 		dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
639 		    1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
640 		    SPA_BLKPTRSHIFT));
641 	}
642 
643 	dn->dn_bonustype = bonustype;
644 	dn->dn_bonuslen = bonuslen;
645 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
646 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
647 	dn->dn_dirtyctx = 0;
648 
649 	dn->dn_free_txg = 0;
650 	dn->dn_dirtyctx_firstset = NULL;
651 	dn->dn_dirty_txg = 0;
652 
653 	dn->dn_allocated_txg = tx->tx_txg;
654 	dn->dn_id_flags = 0;
655 
656 	dnode_setdirty(dn, tx);
657 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
658 	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
659 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
660 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
661 }
662 
663 void
664 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
665     dmu_object_type_t bonustype, int bonuslen, int dn_slots,
666     boolean_t keep_spill, dmu_tx_t *tx)
667 {
668 	int nblkptr;
669 
670 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
671 	ASSERT3U(blocksize, <=,
672 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
673 	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
674 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
675 	ASSERT(tx->tx_txg != 0);
676 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
677 	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
678 	    (bonustype == DMU_OT_SA && bonuslen == 0));
679 	ASSERT(DMU_OT_IS_VALID(bonustype));
680 	ASSERT3U(bonuslen, <=,
681 	    DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
682 	ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
683 
684 	dnode_free_interior_slots(dn);
685 	DNODE_STAT_BUMP(dnode_reallocate);
686 
687 	/* clean up any unreferenced dbufs */
688 	dnode_evict_dbufs(dn);
689 
690 	dn->dn_id_flags = 0;
691 
692 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
693 	dnode_setdirty(dn, tx);
694 	if (dn->dn_datablksz != blocksize) {
695 		/* change blocksize */
696 		ASSERT0(dn->dn_maxblkid);
697 		ASSERT(BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
698 		    dnode_block_freed(dn, 0));
699 
700 		dnode_setdblksz(dn, blocksize);
701 		dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = blocksize;
702 	}
703 	if (dn->dn_bonuslen != bonuslen)
704 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = bonuslen;
705 
706 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
707 		nblkptr = 1;
708 	else
709 		nblkptr = MIN(DN_MAX_NBLKPTR,
710 		    1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
711 		    SPA_BLKPTRSHIFT));
712 	if (dn->dn_bonustype != bonustype)
713 		dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = bonustype;
714 	if (dn->dn_nblkptr != nblkptr)
715 		dn->dn_next_nblkptr[tx->tx_txg & TXG_MASK] = nblkptr;
716 	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR && !keep_spill) {
717 		dbuf_rm_spill(dn, tx);
718 		dnode_rm_spill(dn, tx);
719 	}
720 
721 	rw_exit(&dn->dn_struct_rwlock);
722 
723 	/* change type */
724 	dn->dn_type = ot;
725 
726 	/* change bonus size and type */
727 	mutex_enter(&dn->dn_mtx);
728 	dn->dn_bonustype = bonustype;
729 	dn->dn_bonuslen = bonuslen;
730 	dn->dn_num_slots = dn_slots;
731 	dn->dn_nblkptr = nblkptr;
732 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
733 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
734 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
735 
736 	/* fix up the bonus db_size */
737 	if (dn->dn_bonus) {
738 		dn->dn_bonus->db.db_size =
739 		    DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
740 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
741 		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
742 	}
743 
744 	dn->dn_allocated_txg = tx->tx_txg;
745 	mutex_exit(&dn->dn_mtx);
746 }
747 
748 #ifdef	_KERNEL
749 static void
750 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
751 {
752 	int i;
753 
754 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
755 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
756 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
757 	ASSERT(!MUTEX_HELD(&odn->dn_zfetch.zf_lock));
758 
759 	/* Copy fields. */
760 	ndn->dn_objset = odn->dn_objset;
761 	ndn->dn_object = odn->dn_object;
762 	ndn->dn_dbuf = odn->dn_dbuf;
763 	ndn->dn_handle = odn->dn_handle;
764 	ndn->dn_phys = odn->dn_phys;
765 	ndn->dn_type = odn->dn_type;
766 	ndn->dn_bonuslen = odn->dn_bonuslen;
767 	ndn->dn_bonustype = odn->dn_bonustype;
768 	ndn->dn_nblkptr = odn->dn_nblkptr;
769 	ndn->dn_checksum = odn->dn_checksum;
770 	ndn->dn_compress = odn->dn_compress;
771 	ndn->dn_nlevels = odn->dn_nlevels;
772 	ndn->dn_indblkshift = odn->dn_indblkshift;
773 	ndn->dn_datablkshift = odn->dn_datablkshift;
774 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
775 	ndn->dn_datablksz = odn->dn_datablksz;
776 	ndn->dn_maxblkid = odn->dn_maxblkid;
777 	ndn->dn_num_slots = odn->dn_num_slots;
778 	bcopy(&odn->dn_next_type[0], &ndn->dn_next_type[0],
779 	    sizeof (odn->dn_next_type));
780 	bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
781 	    sizeof (odn->dn_next_nblkptr));
782 	bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
783 	    sizeof (odn->dn_next_nlevels));
784 	bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
785 	    sizeof (odn->dn_next_indblkshift));
786 	bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
787 	    sizeof (odn->dn_next_bonustype));
788 	bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
789 	    sizeof (odn->dn_rm_spillblk));
790 	bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
791 	    sizeof (odn->dn_next_bonuslen));
792 	bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
793 	    sizeof (odn->dn_next_blksz));
794 	bcopy(&odn->dn_next_maxblkid[0], &ndn->dn_next_maxblkid[0],
795 	    sizeof (odn->dn_next_maxblkid));
796 	for (i = 0; i < TXG_SIZE; i++) {
797 		list_move_tail(&ndn->dn_dirty_records[i],
798 		    &odn->dn_dirty_records[i]);
799 	}
800 	bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
801 	    sizeof (odn->dn_free_ranges));
802 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
803 	ndn->dn_free_txg = odn->dn_free_txg;
804 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
805 	ndn->dn_dirty_txg = odn->dn_dirty_txg;
806 	ndn->dn_dirtyctx = odn->dn_dirtyctx;
807 	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
808 	ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0);
809 	zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
810 	ASSERT(avl_is_empty(&ndn->dn_dbufs));
811 	avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
812 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
813 	ndn->dn_bonus = odn->dn_bonus;
814 	ndn->dn_have_spill = odn->dn_have_spill;
815 	ndn->dn_zio = odn->dn_zio;
816 	ndn->dn_oldused = odn->dn_oldused;
817 	ndn->dn_oldflags = odn->dn_oldflags;
818 	ndn->dn_olduid = odn->dn_olduid;
819 	ndn->dn_oldgid = odn->dn_oldgid;
820 	ndn->dn_oldprojid = odn->dn_oldprojid;
821 	ndn->dn_newuid = odn->dn_newuid;
822 	ndn->dn_newgid = odn->dn_newgid;
823 	ndn->dn_newprojid = odn->dn_newprojid;
824 	ndn->dn_id_flags = odn->dn_id_flags;
825 	dmu_zfetch_init(&ndn->dn_zfetch, NULL);
826 	list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
827 	ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
828 
829 	/*
830 	 * Update back pointers. Updating the handle fixes the back pointer of
831 	 * every descendant dbuf as well as the bonus dbuf.
832 	 */
833 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
834 	ndn->dn_handle->dnh_dnode = ndn;
835 	if (ndn->dn_zfetch.zf_dnode == odn) {
836 		ndn->dn_zfetch.zf_dnode = ndn;
837 	}
838 
839 	/*
840 	 * Invalidate the original dnode by clearing all of its back pointers.
841 	 */
842 	odn->dn_dbuf = NULL;
843 	odn->dn_handle = NULL;
844 	avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
845 	    offsetof(dmu_buf_impl_t, db_link));
846 	odn->dn_dbufs_count = 0;
847 	odn->dn_bonus = NULL;
848 	dmu_zfetch_fini(&odn->dn_zfetch);
849 
850 	/*
851 	 * Set the low bit of the objset pointer to ensure that dnode_move()
852 	 * recognizes the dnode as invalid in any subsequent callback.
853 	 */
854 	POINTER_INVALIDATE(&odn->dn_objset);
855 
856 	/*
857 	 * Satisfy the destructor.
858 	 */
859 	for (i = 0; i < TXG_SIZE; i++) {
860 		list_create(&odn->dn_dirty_records[i],
861 		    sizeof (dbuf_dirty_record_t),
862 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
863 		odn->dn_free_ranges[i] = NULL;
864 		odn->dn_next_nlevels[i] = 0;
865 		odn->dn_next_indblkshift[i] = 0;
866 		odn->dn_next_bonustype[i] = 0;
867 		odn->dn_rm_spillblk[i] = 0;
868 		odn->dn_next_bonuslen[i] = 0;
869 		odn->dn_next_blksz[i] = 0;
870 	}
871 	odn->dn_allocated_txg = 0;
872 	odn->dn_free_txg = 0;
873 	odn->dn_assigned_txg = 0;
874 	odn->dn_dirty_txg = 0;
875 	odn->dn_dirtyctx = 0;
876 	odn->dn_dirtyctx_firstset = NULL;
877 	odn->dn_have_spill = B_FALSE;
878 	odn->dn_zio = NULL;
879 	odn->dn_oldused = 0;
880 	odn->dn_oldflags = 0;
881 	odn->dn_olduid = 0;
882 	odn->dn_oldgid = 0;
883 	odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
884 	odn->dn_newuid = 0;
885 	odn->dn_newgid = 0;
886 	odn->dn_newprojid = ZFS_DEFAULT_PROJID;
887 	odn->dn_id_flags = 0;
888 
889 	/*
890 	 * Mark the dnode.
891 	 */
892 	ndn->dn_moved = 1;
893 	odn->dn_moved = (uint8_t)-1;
894 }
895 
896 /*ARGSUSED*/
897 static kmem_cbrc_t
898 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
899 {
900 	dnode_t *odn = buf, *ndn = newbuf;
901 	objset_t *os;
902 	int64_t refcount;
903 	uint32_t dbufs;
904 
905 	/*
906 	 * The dnode is on the objset's list of known dnodes if the objset
907 	 * pointer is valid. We set the low bit of the objset pointer when
908 	 * freeing the dnode to invalidate it, and the memory patterns written
909 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
910 	 * A newly created dnode sets the objset pointer last of all to indicate
911 	 * that the dnode is known and in a valid state to be moved by this
912 	 * function.
913 	 */
914 	os = odn->dn_objset;
915 	if (!POINTER_IS_VALID(os)) {
916 		DNODE_STAT_BUMP(dnode_move_invalid);
917 		return (KMEM_CBRC_DONT_KNOW);
918 	}
919 
920 	/*
921 	 * Ensure that the objset does not go away during the move.
922 	 */
923 	rw_enter(&os_lock, RW_WRITER);
924 	if (os != odn->dn_objset) {
925 		rw_exit(&os_lock);
926 		DNODE_STAT_BUMP(dnode_move_recheck1);
927 		return (KMEM_CBRC_DONT_KNOW);
928 	}
929 
930 	/*
931 	 * If the dnode is still valid, then so is the objset. We know that no
932 	 * valid objset can be freed while we hold os_lock, so we can safely
933 	 * ensure that the objset remains in use.
934 	 */
935 	mutex_enter(&os->os_lock);
936 
937 	/*
938 	 * Recheck the objset pointer in case the dnode was removed just before
939 	 * acquiring the lock.
940 	 */
941 	if (os != odn->dn_objset) {
942 		mutex_exit(&os->os_lock);
943 		rw_exit(&os_lock);
944 		DNODE_STAT_BUMP(dnode_move_recheck2);
945 		return (KMEM_CBRC_DONT_KNOW);
946 	}
947 
948 	/*
949 	 * At this point we know that as long as we hold os->os_lock, the dnode
950 	 * cannot be freed and fields within the dnode can be safely accessed.
951 	 * The objset listing this dnode cannot go away as long as this dnode is
952 	 * on its list.
953 	 */
954 	rw_exit(&os_lock);
955 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
956 		mutex_exit(&os->os_lock);
957 		DNODE_STAT_BUMP(dnode_move_special);
958 		return (KMEM_CBRC_NO);
959 	}
960 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
961 
962 	/*
963 	 * Lock the dnode handle to prevent the dnode from obtaining any new
964 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
965 	 * from accessing the dnode, so that we can discount their holds. The
966 	 * handle is safe to access because we know that while the dnode cannot
967 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
968 	 * safely move any dnode referenced only by dbufs.
969 	 */
970 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
971 		mutex_exit(&os->os_lock);
972 		DNODE_STAT_BUMP(dnode_move_handle);
973 		return (KMEM_CBRC_LATER);
974 	}
975 
976 	/*
977 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
978 	 * We need to guarantee that there is a hold for every dbuf in order to
979 	 * determine whether the dnode is actively referenced. Falsely matching
980 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
981 	 * that a thread already having an active dnode hold is about to add a
982 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
983 	 * progress.
984 	 */
985 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
986 		zrl_exit(&odn->dn_handle->dnh_zrlock);
987 		mutex_exit(&os->os_lock);
988 		DNODE_STAT_BUMP(dnode_move_rwlock);
989 		return (KMEM_CBRC_LATER);
990 	}
991 
992 	/*
993 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
994 	 * case, the dbuf count is decremented under the handle lock before the
995 	 * dbuf's hold is released. This order ensures that if we count the hold
996 	 * after the dbuf is removed but before its hold is released, we will
997 	 * treat the unmatched hold as active and exit safely. If we count the
998 	 * hold before the dbuf is removed, the hold is discounted, and the
999 	 * removal is blocked until the move completes.
1000 	 */
1001 	refcount = zfs_refcount_count(&odn->dn_holds);
1002 	ASSERT(refcount >= 0);
1003 	dbufs = DN_DBUFS_COUNT(odn);
1004 
1005 	/* We can't have more dbufs than dnode holds. */
1006 	ASSERT3U(dbufs, <=, refcount);
1007 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1008 	    uint32_t, dbufs);
1009 
1010 	if (refcount > dbufs) {
1011 		rw_exit(&odn->dn_struct_rwlock);
1012 		zrl_exit(&odn->dn_handle->dnh_zrlock);
1013 		mutex_exit(&os->os_lock);
1014 		DNODE_STAT_BUMP(dnode_move_active);
1015 		return (KMEM_CBRC_LATER);
1016 	}
1017 
1018 	rw_exit(&odn->dn_struct_rwlock);
1019 
1020 	/*
1021 	 * At this point we know that anyone with a hold on the dnode is not
1022 	 * actively referencing it. The dnode is known and in a valid state to
1023 	 * move. We're holding the locks needed to execute the critical section.
1024 	 */
1025 	dnode_move_impl(odn, ndn);
1026 
1027 	list_link_replace(&odn->dn_link, &ndn->dn_link);
1028 	/* If the dnode was safe to move, the refcount cannot have changed. */
1029 	ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1030 	ASSERT(dbufs == DN_DBUFS_COUNT(ndn));
1031 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1032 	mutex_exit(&os->os_lock);
1033 
1034 	return (KMEM_CBRC_YES);
1035 }
1036 #endif	/* _KERNEL */
1037 
1038 static void
1039 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1040 {
1041 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1042 
1043 	for (int i = idx; i < idx + slots; i++) {
1044 		dnode_handle_t *dnh = &children->dnc_children[i];
1045 		zrl_add(&dnh->dnh_zrlock);
1046 	}
1047 }
1048 
1049 static void
1050 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1051 {
1052 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1053 
1054 	for (int i = idx; i < idx + slots; i++) {
1055 		dnode_handle_t *dnh = &children->dnc_children[i];
1056 
1057 		if (zrl_is_locked(&dnh->dnh_zrlock))
1058 			zrl_exit(&dnh->dnh_zrlock);
1059 		else
1060 			zrl_remove(&dnh->dnh_zrlock);
1061 	}
1062 }
1063 
1064 static int
1065 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1066 {
1067 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1068 
1069 	for (int i = idx; i < idx + slots; i++) {
1070 		dnode_handle_t *dnh = &children->dnc_children[i];
1071 
1072 		if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1073 			for (int j = idx; j < i; j++) {
1074 				dnh = &children->dnc_children[j];
1075 				zrl_exit(&dnh->dnh_zrlock);
1076 			}
1077 
1078 			return (0);
1079 		}
1080 	}
1081 
1082 	return (1);
1083 }
1084 
1085 static void
1086 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1087 {
1088 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1089 
1090 	for (int i = idx; i < idx + slots; i++) {
1091 		dnode_handle_t *dnh = &children->dnc_children[i];
1092 		dnh->dnh_dnode = ptr;
1093 	}
1094 }
1095 
1096 static boolean_t
1097 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1098 {
1099 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1100 
1101 	/*
1102 	 * If all dnode slots are either already free or
1103 	 * evictable return B_TRUE.
1104 	 */
1105 	for (int i = idx; i < idx + slots; i++) {
1106 		dnode_handle_t *dnh = &children->dnc_children[i];
1107 		dnode_t *dn = dnh->dnh_dnode;
1108 
1109 		if (dn == DN_SLOT_FREE) {
1110 			continue;
1111 		} else if (DN_SLOT_IS_PTR(dn)) {
1112 			mutex_enter(&dn->dn_mtx);
1113 			boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1114 			    zfs_refcount_is_zero(&dn->dn_holds) &&
1115 			    !DNODE_IS_DIRTY(dn));
1116 			mutex_exit(&dn->dn_mtx);
1117 
1118 			if (!can_free)
1119 				return (B_FALSE);
1120 			else
1121 				continue;
1122 		} else {
1123 			return (B_FALSE);
1124 		}
1125 	}
1126 
1127 	return (B_TRUE);
1128 }
1129 
1130 static void
1131 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1132 {
1133 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1134 
1135 	for (int i = idx; i < idx + slots; i++) {
1136 		dnode_handle_t *dnh = &children->dnc_children[i];
1137 
1138 		ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1139 
1140 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1141 			ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1142 			dnode_destroy(dnh->dnh_dnode);
1143 			dnh->dnh_dnode = DN_SLOT_FREE;
1144 		}
1145 	}
1146 }
1147 
1148 void
1149 dnode_free_interior_slots(dnode_t *dn)
1150 {
1151 	dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1152 	int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1153 	int idx = (dn->dn_object & (epb - 1)) + 1;
1154 	int slots = dn->dn_num_slots - 1;
1155 
1156 	if (slots == 0)
1157 		return;
1158 
1159 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1160 
1161 	while (!dnode_slots_tryenter(children, idx, slots)) {
1162 		DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1163 		cond_resched();
1164 	}
1165 
1166 	dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1167 	dnode_slots_rele(children, idx, slots);
1168 }
1169 
1170 void
1171 dnode_special_close(dnode_handle_t *dnh)
1172 {
1173 	dnode_t *dn = dnh->dnh_dnode;
1174 
1175 	/*
1176 	 * Ensure dnode_rele_and_unlock() has released dn_mtx, after final
1177 	 * zfs_refcount_remove()
1178 	 */
1179 	mutex_enter(&dn->dn_mtx);
1180 	if (zfs_refcount_count(&dn->dn_holds) > 0)
1181 		cv_wait(&dn->dn_nodnholds, &dn->dn_mtx);
1182 	mutex_exit(&dn->dn_mtx);
1183 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 0);
1184 
1185 	ASSERT(dn->dn_dbuf == NULL ||
1186 	    dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1187 	zrl_add(&dnh->dnh_zrlock);
1188 	dnode_destroy(dn); /* implicit zrl_remove() */
1189 	zrl_destroy(&dnh->dnh_zrlock);
1190 	dnh->dnh_dnode = NULL;
1191 }
1192 
1193 void
1194 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1195     dnode_handle_t *dnh)
1196 {
1197 	dnode_t *dn;
1198 
1199 	zrl_init(&dnh->dnh_zrlock);
1200 	VERIFY3U(1, ==, zrl_tryenter(&dnh->dnh_zrlock));
1201 
1202 	dn = dnode_create(os, dnp, NULL, object, dnh);
1203 	DNODE_VERIFY(dn);
1204 
1205 	zrl_exit(&dnh->dnh_zrlock);
1206 }
1207 
1208 static void
1209 dnode_buf_evict_async(void *dbu)
1210 {
1211 	dnode_children_t *dnc = dbu;
1212 
1213 	DNODE_STAT_BUMP(dnode_buf_evict);
1214 
1215 	for (int i = 0; i < dnc->dnc_count; i++) {
1216 		dnode_handle_t *dnh = &dnc->dnc_children[i];
1217 		dnode_t *dn;
1218 
1219 		/*
1220 		 * The dnode handle lock guards against the dnode moving to
1221 		 * another valid address, so there is no need here to guard
1222 		 * against changes to or from NULL.
1223 		 */
1224 		if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1225 			zrl_destroy(&dnh->dnh_zrlock);
1226 			dnh->dnh_dnode = DN_SLOT_UNINIT;
1227 			continue;
1228 		}
1229 
1230 		zrl_add(&dnh->dnh_zrlock);
1231 		dn = dnh->dnh_dnode;
1232 		/*
1233 		 * If there are holds on this dnode, then there should
1234 		 * be holds on the dnode's containing dbuf as well; thus
1235 		 * it wouldn't be eligible for eviction and this function
1236 		 * would not have been called.
1237 		 */
1238 		ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1239 		ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1240 
1241 		dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1242 		zrl_destroy(&dnh->dnh_zrlock);
1243 		dnh->dnh_dnode = DN_SLOT_UNINIT;
1244 	}
1245 	kmem_free(dnc, sizeof (dnode_children_t) +
1246 	    dnc->dnc_count * sizeof (dnode_handle_t));
1247 }
1248 
1249 /*
1250  * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1251  * to ensure the hole at the specified object offset is large enough to
1252  * hold the dnode being created. The slots parameter is also used to ensure
1253  * a dnode does not span multiple dnode blocks. In both of these cases, if
1254  * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1255  * are only possible when using DNODE_MUST_BE_FREE.
1256  *
1257  * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1258  * dnode_hold_impl() will check if the requested dnode is already consumed
1259  * as an extra dnode slot by an large dnode, in which case it returns
1260  * ENOENT.
1261  *
1262  * If the DNODE_DRY_RUN flag is set, we don't actually hold the dnode, just
1263  * return whether the hold would succeed or not. tag and dnp should set to
1264  * NULL in this case.
1265  *
1266  * errors:
1267  * EINVAL - Invalid object number or flags.
1268  * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1269  * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1270  *        - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1271  *        - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1272  * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1273  *        - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1274  * EIO    - I/O error when reading the meta dnode dbuf.
1275  *
1276  * succeeds even for free dnodes.
1277  */
1278 int
1279 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1280     void *tag, dnode_t **dnp)
1281 {
1282 	int epb, idx, err;
1283 	int drop_struct_lock = FALSE;
1284 	int type;
1285 	uint64_t blk;
1286 	dnode_t *mdn, *dn;
1287 	dmu_buf_impl_t *db;
1288 	dnode_children_t *dnc;
1289 	dnode_phys_t *dn_block;
1290 	dnode_handle_t *dnh;
1291 
1292 	ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1293 	ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1294 	IMPLY(flag & DNODE_DRY_RUN, (tag == NULL) && (dnp == NULL));
1295 
1296 	/*
1297 	 * If you are holding the spa config lock as writer, you shouldn't
1298 	 * be asking the DMU to do *anything* unless it's the root pool
1299 	 * which may require us to read from the root filesystem while
1300 	 * holding some (not all) of the locks as writer.
1301 	 */
1302 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1303 	    (spa_is_root(os->os_spa) &&
1304 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1305 
1306 	ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1307 
1308 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1309 	    object == DMU_PROJECTUSED_OBJECT) {
1310 		if (object == DMU_USERUSED_OBJECT)
1311 			dn = DMU_USERUSED_DNODE(os);
1312 		else if (object == DMU_GROUPUSED_OBJECT)
1313 			dn = DMU_GROUPUSED_DNODE(os);
1314 		else
1315 			dn = DMU_PROJECTUSED_DNODE(os);
1316 		if (dn == NULL)
1317 			return (SET_ERROR(ENOENT));
1318 		type = dn->dn_type;
1319 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1320 			return (SET_ERROR(ENOENT));
1321 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1322 			return (SET_ERROR(EEXIST));
1323 		DNODE_VERIFY(dn);
1324 		/* Don't actually hold if dry run, just return 0 */
1325 		if (!(flag & DNODE_DRY_RUN)) {
1326 			(void) zfs_refcount_add(&dn->dn_holds, tag);
1327 			*dnp = dn;
1328 		}
1329 		return (0);
1330 	}
1331 
1332 	if (object == 0 || object >= DN_MAX_OBJECT)
1333 		return (SET_ERROR(EINVAL));
1334 
1335 	mdn = DMU_META_DNODE(os);
1336 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1337 
1338 	DNODE_VERIFY(mdn);
1339 
1340 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1341 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1342 		drop_struct_lock = TRUE;
1343 	}
1344 
1345 	blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1346 	db = dbuf_hold(mdn, blk, FTAG);
1347 	if (drop_struct_lock)
1348 		rw_exit(&mdn->dn_struct_rwlock);
1349 	if (db == NULL) {
1350 		DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1351 		return (SET_ERROR(EIO));
1352 	}
1353 
1354 	/*
1355 	 * We do not need to decrypt to read the dnode so it doesn't matter
1356 	 * if we get the encrypted or decrypted version.
1357 	 */
1358 	err = dbuf_read(db, NULL, DB_RF_CANFAIL |
1359 	    DB_RF_NO_DECRYPT | DB_RF_NOPREFETCH);
1360 	if (err) {
1361 		DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1362 		dbuf_rele(db, FTAG);
1363 		return (err);
1364 	}
1365 
1366 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1367 	epb = db->db.db_size >> DNODE_SHIFT;
1368 
1369 	idx = object & (epb - 1);
1370 	dn_block = (dnode_phys_t *)db->db.db_data;
1371 
1372 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1373 	dnc = dmu_buf_get_user(&db->db);
1374 	dnh = NULL;
1375 	if (dnc == NULL) {
1376 		dnode_children_t *winner;
1377 		int skip = 0;
1378 
1379 		dnc = kmem_zalloc(sizeof (dnode_children_t) +
1380 		    epb * sizeof (dnode_handle_t), KM_SLEEP);
1381 		dnc->dnc_count = epb;
1382 		dnh = &dnc->dnc_children[0];
1383 
1384 		/* Initialize dnode slot status from dnode_phys_t */
1385 		for (int i = 0; i < epb; i++) {
1386 			zrl_init(&dnh[i].dnh_zrlock);
1387 
1388 			if (skip) {
1389 				skip--;
1390 				continue;
1391 			}
1392 
1393 			if (dn_block[i].dn_type != DMU_OT_NONE) {
1394 				int interior = dn_block[i].dn_extra_slots;
1395 
1396 				dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1397 				dnode_set_slots(dnc, i + 1, interior,
1398 				    DN_SLOT_INTERIOR);
1399 				skip = interior;
1400 			} else {
1401 				dnh[i].dnh_dnode = DN_SLOT_FREE;
1402 				skip = 0;
1403 			}
1404 		}
1405 
1406 		dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1407 		    dnode_buf_evict_async, NULL);
1408 		winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1409 		if (winner != NULL) {
1410 
1411 			for (int i = 0; i < epb; i++)
1412 				zrl_destroy(&dnh[i].dnh_zrlock);
1413 
1414 			kmem_free(dnc, sizeof (dnode_children_t) +
1415 			    epb * sizeof (dnode_handle_t));
1416 			dnc = winner;
1417 		}
1418 	}
1419 
1420 	ASSERT(dnc->dnc_count == epb);
1421 
1422 	if (flag & DNODE_MUST_BE_ALLOCATED) {
1423 		slots = 1;
1424 
1425 		dnode_slots_hold(dnc, idx, slots);
1426 		dnh = &dnc->dnc_children[idx];
1427 
1428 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1429 			dn = dnh->dnh_dnode;
1430 		} else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1431 			DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1432 			dnode_slots_rele(dnc, idx, slots);
1433 			dbuf_rele(db, FTAG);
1434 			return (SET_ERROR(EEXIST));
1435 		} else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1436 			DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1437 			dnode_slots_rele(dnc, idx, slots);
1438 			dbuf_rele(db, FTAG);
1439 			return (SET_ERROR(ENOENT));
1440 		} else {
1441 			dnode_slots_rele(dnc, idx, slots);
1442 			while (!dnode_slots_tryenter(dnc, idx, slots)) {
1443 				DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1444 				cond_resched();
1445 			}
1446 
1447 			/*
1448 			 * Someone else won the race and called dnode_create()
1449 			 * after we checked DN_SLOT_IS_PTR() above but before
1450 			 * we acquired the lock.
1451 			 */
1452 			if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1453 				DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1454 				dn = dnh->dnh_dnode;
1455 			} else {
1456 				dn = dnode_create(os, dn_block + idx, db,
1457 				    object, dnh);
1458 			}
1459 		}
1460 
1461 		mutex_enter(&dn->dn_mtx);
1462 		if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1463 			DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1464 			mutex_exit(&dn->dn_mtx);
1465 			dnode_slots_rele(dnc, idx, slots);
1466 			dbuf_rele(db, FTAG);
1467 			return (SET_ERROR(ENOENT));
1468 		}
1469 
1470 		/* Don't actually hold if dry run, just return 0 */
1471 		if (flag & DNODE_DRY_RUN) {
1472 			mutex_exit(&dn->dn_mtx);
1473 			dnode_slots_rele(dnc, idx, slots);
1474 			dbuf_rele(db, FTAG);
1475 			return (0);
1476 		}
1477 
1478 		DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1479 	} else if (flag & DNODE_MUST_BE_FREE) {
1480 
1481 		if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1482 			DNODE_STAT_BUMP(dnode_hold_free_overflow);
1483 			dbuf_rele(db, FTAG);
1484 			return (SET_ERROR(ENOSPC));
1485 		}
1486 
1487 		dnode_slots_hold(dnc, idx, slots);
1488 
1489 		if (!dnode_check_slots_free(dnc, idx, slots)) {
1490 			DNODE_STAT_BUMP(dnode_hold_free_misses);
1491 			dnode_slots_rele(dnc, idx, slots);
1492 			dbuf_rele(db, FTAG);
1493 			return (SET_ERROR(ENOSPC));
1494 		}
1495 
1496 		dnode_slots_rele(dnc, idx, slots);
1497 		while (!dnode_slots_tryenter(dnc, idx, slots)) {
1498 			DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1499 			cond_resched();
1500 		}
1501 
1502 		if (!dnode_check_slots_free(dnc, idx, slots)) {
1503 			DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1504 			dnode_slots_rele(dnc, idx, slots);
1505 			dbuf_rele(db, FTAG);
1506 			return (SET_ERROR(ENOSPC));
1507 		}
1508 
1509 		/*
1510 		 * Allocated but otherwise free dnodes which would
1511 		 * be in the interior of a multi-slot dnodes need
1512 		 * to be freed.  Single slot dnodes can be safely
1513 		 * re-purposed as a performance optimization.
1514 		 */
1515 		if (slots > 1)
1516 			dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1517 
1518 		dnh = &dnc->dnc_children[idx];
1519 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1520 			dn = dnh->dnh_dnode;
1521 		} else {
1522 			dn = dnode_create(os, dn_block + idx, db,
1523 			    object, dnh);
1524 		}
1525 
1526 		mutex_enter(&dn->dn_mtx);
1527 		if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1528 			DNODE_STAT_BUMP(dnode_hold_free_refcount);
1529 			mutex_exit(&dn->dn_mtx);
1530 			dnode_slots_rele(dnc, idx, slots);
1531 			dbuf_rele(db, FTAG);
1532 			return (SET_ERROR(EEXIST));
1533 		}
1534 
1535 		/* Don't actually hold if dry run, just return 0 */
1536 		if (flag & DNODE_DRY_RUN) {
1537 			mutex_exit(&dn->dn_mtx);
1538 			dnode_slots_rele(dnc, idx, slots);
1539 			dbuf_rele(db, FTAG);
1540 			return (0);
1541 		}
1542 
1543 		dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1544 		DNODE_STAT_BUMP(dnode_hold_free_hits);
1545 	} else {
1546 		dbuf_rele(db, FTAG);
1547 		return (SET_ERROR(EINVAL));
1548 	}
1549 
1550 	ASSERT0(dn->dn_free_txg);
1551 
1552 	if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1553 		dbuf_add_ref(db, dnh);
1554 
1555 	mutex_exit(&dn->dn_mtx);
1556 
1557 	/* Now we can rely on the hold to prevent the dnode from moving. */
1558 	dnode_slots_rele(dnc, idx, slots);
1559 
1560 	DNODE_VERIFY(dn);
1561 	ASSERT3P(dnp, !=, NULL);
1562 	ASSERT3P(dn->dn_dbuf, ==, db);
1563 	ASSERT3U(dn->dn_object, ==, object);
1564 	dbuf_rele(db, FTAG);
1565 
1566 	*dnp = dn;
1567 	return (0);
1568 }
1569 
1570 /*
1571  * Return held dnode if the object is allocated, NULL if not.
1572  */
1573 int
1574 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1575 {
1576 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1577 	    dnp));
1578 }
1579 
1580 /*
1581  * Can only add a reference if there is already at least one
1582  * reference on the dnode.  Returns FALSE if unable to add a
1583  * new reference.
1584  */
1585 boolean_t
1586 dnode_add_ref(dnode_t *dn, void *tag)
1587 {
1588 	mutex_enter(&dn->dn_mtx);
1589 	if (zfs_refcount_is_zero(&dn->dn_holds)) {
1590 		mutex_exit(&dn->dn_mtx);
1591 		return (FALSE);
1592 	}
1593 	VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1594 	mutex_exit(&dn->dn_mtx);
1595 	return (TRUE);
1596 }
1597 
1598 void
1599 dnode_rele(dnode_t *dn, void *tag)
1600 {
1601 	mutex_enter(&dn->dn_mtx);
1602 	dnode_rele_and_unlock(dn, tag, B_FALSE);
1603 }
1604 
1605 void
1606 dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1607 {
1608 	uint64_t refs;
1609 	/* Get while the hold prevents the dnode from moving. */
1610 	dmu_buf_impl_t *db = dn->dn_dbuf;
1611 	dnode_handle_t *dnh = dn->dn_handle;
1612 
1613 	refs = zfs_refcount_remove(&dn->dn_holds, tag);
1614 	if (refs == 0)
1615 		cv_broadcast(&dn->dn_nodnholds);
1616 	mutex_exit(&dn->dn_mtx);
1617 	/* dnode could get destroyed at this point, so don't use it anymore */
1618 
1619 	/*
1620 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1621 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1622 	 * prevent the dnode from moving, since releasing the last hold could
1623 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1624 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1625 	 * other direct or indirect hold on the dnode must first drop the dnode
1626 	 * handle.
1627 	 */
1628 	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1629 
1630 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1631 	if (refs == 0 && db != NULL) {
1632 		/*
1633 		 * Another thread could add a hold to the dnode handle in
1634 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1635 		 * hold on the parent dbuf prevents the handle from being
1636 		 * destroyed, the hold on the handle is OK. We can't yet assert
1637 		 * that the handle has zero references, but that will be
1638 		 * asserted anyway when the handle gets destroyed.
1639 		 */
1640 		mutex_enter(&db->db_mtx);
1641 		dbuf_rele_and_unlock(db, dnh, evicting);
1642 	}
1643 }
1644 
1645 /*
1646  * Test whether we can create a dnode at the specified location.
1647  */
1648 int
1649 dnode_try_claim(objset_t *os, uint64_t object, int slots)
1650 {
1651 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_FREE | DNODE_DRY_RUN,
1652 	    slots, NULL, NULL));
1653 }
1654 
1655 void
1656 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1657 {
1658 	objset_t *os = dn->dn_objset;
1659 	uint64_t txg = tx->tx_txg;
1660 
1661 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1662 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1663 		return;
1664 	}
1665 
1666 	DNODE_VERIFY(dn);
1667 
1668 #ifdef ZFS_DEBUG
1669 	mutex_enter(&dn->dn_mtx);
1670 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1671 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1672 	mutex_exit(&dn->dn_mtx);
1673 #endif
1674 
1675 	/*
1676 	 * Determine old uid/gid when necessary
1677 	 */
1678 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1679 
1680 	multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1681 	multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1682 
1683 	/*
1684 	 * If we are already marked dirty, we're done.
1685 	 */
1686 	if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1687 		multilist_sublist_unlock(mls);
1688 		return;
1689 	}
1690 
1691 	ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1692 	    !avl_is_empty(&dn->dn_dbufs));
1693 	ASSERT(dn->dn_datablksz != 0);
1694 	ASSERT0(dn->dn_next_bonuslen[txg & TXG_MASK]);
1695 	ASSERT0(dn->dn_next_blksz[txg & TXG_MASK]);
1696 	ASSERT0(dn->dn_next_bonustype[txg & TXG_MASK]);
1697 
1698 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1699 	    dn->dn_object, txg);
1700 
1701 	multilist_sublist_insert_head(mls, dn);
1702 
1703 	multilist_sublist_unlock(mls);
1704 
1705 	/*
1706 	 * The dnode maintains a hold on its containing dbuf as
1707 	 * long as there are holds on it.  Each instantiated child
1708 	 * dbuf maintains a hold on the dnode.  When the last child
1709 	 * drops its hold, the dnode will drop its hold on the
1710 	 * containing dbuf. We add a "dirty hold" here so that the
1711 	 * dnode will hang around after we finish processing its
1712 	 * children.
1713 	 */
1714 	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1715 
1716 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1717 
1718 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1719 }
1720 
1721 void
1722 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1723 {
1724 	mutex_enter(&dn->dn_mtx);
1725 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1726 		mutex_exit(&dn->dn_mtx);
1727 		return;
1728 	}
1729 	dn->dn_free_txg = tx->tx_txg;
1730 	mutex_exit(&dn->dn_mtx);
1731 
1732 	dnode_setdirty(dn, tx);
1733 }
1734 
1735 /*
1736  * Try to change the block size for the indicated dnode.  This can only
1737  * succeed if there are no blocks allocated or dirty beyond first block
1738  */
1739 int
1740 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1741 {
1742 	dmu_buf_impl_t *db;
1743 	int err;
1744 
1745 	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1746 	if (size == 0)
1747 		size = SPA_MINBLOCKSIZE;
1748 	else
1749 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1750 
1751 	if (ibs == dn->dn_indblkshift)
1752 		ibs = 0;
1753 
1754 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1755 		return (0);
1756 
1757 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1758 
1759 	/* Check for any allocated blocks beyond the first */
1760 	if (dn->dn_maxblkid != 0)
1761 		goto fail;
1762 
1763 	mutex_enter(&dn->dn_dbufs_mtx);
1764 	for (db = avl_first(&dn->dn_dbufs); db != NULL;
1765 	    db = AVL_NEXT(&dn->dn_dbufs, db)) {
1766 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1767 		    db->db_blkid != DMU_SPILL_BLKID) {
1768 			mutex_exit(&dn->dn_dbufs_mtx);
1769 			goto fail;
1770 		}
1771 	}
1772 	mutex_exit(&dn->dn_dbufs_mtx);
1773 
1774 	if (ibs && dn->dn_nlevels != 1)
1775 		goto fail;
1776 
1777 	/* resize the old block */
1778 	err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1779 	if (err == 0) {
1780 		dbuf_new_size(db, size, tx);
1781 	} else if (err != ENOENT) {
1782 		goto fail;
1783 	}
1784 
1785 	dnode_setdblksz(dn, size);
1786 	dnode_setdirty(dn, tx);
1787 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1788 	if (ibs) {
1789 		dn->dn_indblkshift = ibs;
1790 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1791 	}
1792 	/* release after we have fixed the blocksize in the dnode */
1793 	if (db)
1794 		dbuf_rele(db, FTAG);
1795 
1796 	rw_exit(&dn->dn_struct_rwlock);
1797 	return (0);
1798 
1799 fail:
1800 	rw_exit(&dn->dn_struct_rwlock);
1801 	return (SET_ERROR(ENOTSUP));
1802 }
1803 
1804 static void
1805 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1806 {
1807 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1808 	int old_nlevels = dn->dn_nlevels;
1809 	dmu_buf_impl_t *db;
1810 	list_t *list;
1811 	dbuf_dirty_record_t *new, *dr, *dr_next;
1812 
1813 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1814 
1815 	ASSERT3U(new_nlevels, >, dn->dn_nlevels);
1816 	dn->dn_nlevels = new_nlevels;
1817 
1818 	ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1819 	dn->dn_next_nlevels[txgoff] = new_nlevels;
1820 
1821 	/* dirty the left indirects */
1822 	db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1823 	ASSERT(db != NULL);
1824 	new = dbuf_dirty(db, tx);
1825 	dbuf_rele(db, FTAG);
1826 
1827 	/* transfer the dirty records to the new indirect */
1828 	mutex_enter(&dn->dn_mtx);
1829 	mutex_enter(&new->dt.di.dr_mtx);
1830 	list = &dn->dn_dirty_records[txgoff];
1831 	for (dr = list_head(list); dr; dr = dr_next) {
1832 		dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1833 
1834 		IMPLY(dr->dr_dbuf == NULL, old_nlevels == 1);
1835 		if (dr->dr_dbuf == NULL ||
1836 		    (dr->dr_dbuf->db_level == old_nlevels - 1 &&
1837 		    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1838 		    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID)) {
1839 			list_remove(&dn->dn_dirty_records[txgoff], dr);
1840 			list_insert_tail(&new->dt.di.dr_children, dr);
1841 			dr->dr_parent = new;
1842 		}
1843 	}
1844 	mutex_exit(&new->dt.di.dr_mtx);
1845 	mutex_exit(&dn->dn_mtx);
1846 }
1847 
1848 int
1849 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
1850 {
1851 	int ret = 0;
1852 
1853 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1854 
1855 	if (dn->dn_nlevels == nlevels) {
1856 		ret = 0;
1857 		goto out;
1858 	} else if (nlevels < dn->dn_nlevels) {
1859 		ret = SET_ERROR(EINVAL);
1860 		goto out;
1861 	}
1862 
1863 	dnode_set_nlevels_impl(dn, nlevels, tx);
1864 
1865 out:
1866 	rw_exit(&dn->dn_struct_rwlock);
1867 	return (ret);
1868 }
1869 
1870 /* read-holding callers must not rely on the lock being continuously held */
1871 void
1872 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read,
1873     boolean_t force)
1874 {
1875 	int epbs, new_nlevels;
1876 	uint64_t sz;
1877 
1878 	ASSERT(blkid != DMU_BONUS_BLKID);
1879 
1880 	ASSERT(have_read ?
1881 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
1882 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
1883 
1884 	/*
1885 	 * if we have a read-lock, check to see if we need to do any work
1886 	 * before upgrading to a write-lock.
1887 	 */
1888 	if (have_read) {
1889 		if (blkid <= dn->dn_maxblkid)
1890 			return;
1891 
1892 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1893 			rw_exit(&dn->dn_struct_rwlock);
1894 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1895 		}
1896 	}
1897 
1898 	/*
1899 	 * Raw sends (indicated by the force flag) require that we take the
1900 	 * given blkid even if the value is lower than the current value.
1901 	 */
1902 	if (!force && blkid <= dn->dn_maxblkid)
1903 		goto out;
1904 
1905 	/*
1906 	 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff]
1907 	 * to indicate that this field is set. This allows us to set the
1908 	 * maxblkid to 0 on an existing object in dnode_sync().
1909 	 */
1910 	dn->dn_maxblkid = blkid;
1911 	dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] =
1912 	    blkid | DMU_NEXT_MAXBLKID_SET;
1913 
1914 	/*
1915 	 * Compute the number of levels necessary to support the new maxblkid.
1916 	 * Raw sends will ensure nlevels is set correctly for us.
1917 	 */
1918 	new_nlevels = 1;
1919 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1920 	for (sz = dn->dn_nblkptr;
1921 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1922 		new_nlevels++;
1923 
1924 	ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
1925 
1926 	if (!force) {
1927 		if (new_nlevels > dn->dn_nlevels)
1928 			dnode_set_nlevels_impl(dn, new_nlevels, tx);
1929 	} else {
1930 		ASSERT3U(dn->dn_nlevels, >=, new_nlevels);
1931 	}
1932 
1933 out:
1934 	if (have_read)
1935 		rw_downgrade(&dn->dn_struct_rwlock);
1936 }
1937 
1938 static void
1939 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1940 {
1941 	dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1942 	if (db != NULL) {
1943 		dmu_buf_will_dirty(&db->db, tx);
1944 		dbuf_rele(db, FTAG);
1945 	}
1946 }
1947 
1948 /*
1949  * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1950  * and end_blkid.
1951  */
1952 static void
1953 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1954     dmu_tx_t *tx)
1955 {
1956 	dmu_buf_impl_t *db_search;
1957 	dmu_buf_impl_t *db;
1958 	avl_index_t where;
1959 
1960 	db_search = kmem_zalloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1961 
1962 	mutex_enter(&dn->dn_dbufs_mtx);
1963 
1964 	db_search->db_level = 1;
1965 	db_search->db_blkid = start_blkid + 1;
1966 	db_search->db_state = DB_SEARCH;
1967 	for (;;) {
1968 
1969 		db = avl_find(&dn->dn_dbufs, db_search, &where);
1970 		if (db == NULL)
1971 			db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1972 
1973 		if (db == NULL || db->db_level != 1 ||
1974 		    db->db_blkid >= end_blkid) {
1975 			break;
1976 		}
1977 
1978 		/*
1979 		 * Setup the next blkid we want to search for.
1980 		 */
1981 		db_search->db_blkid = db->db_blkid + 1;
1982 		ASSERT3U(db->db_blkid, >=, start_blkid);
1983 
1984 		/*
1985 		 * If the dbuf transitions to DB_EVICTING while we're trying
1986 		 * to dirty it, then we will be unable to discover it in
1987 		 * the dbuf hash table. This will result in a call to
1988 		 * dbuf_create() which needs to acquire the dn_dbufs_mtx
1989 		 * lock. To avoid a deadlock, we drop the lock before
1990 		 * dirtying the level-1 dbuf.
1991 		 */
1992 		mutex_exit(&dn->dn_dbufs_mtx);
1993 		dnode_dirty_l1(dn, db->db_blkid, tx);
1994 		mutex_enter(&dn->dn_dbufs_mtx);
1995 	}
1996 
1997 #ifdef ZFS_DEBUG
1998 	/*
1999 	 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
2000 	 */
2001 	db_search->db_level = 1;
2002 	db_search->db_blkid = start_blkid + 1;
2003 	db_search->db_state = DB_SEARCH;
2004 	db = avl_find(&dn->dn_dbufs, db_search, &where);
2005 	if (db == NULL)
2006 		db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
2007 	for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
2008 		if (db->db_level != 1 || db->db_blkid >= end_blkid)
2009 			break;
2010 		if (db->db_state != DB_EVICTING)
2011 			ASSERT(db->db_dirtycnt > 0);
2012 	}
2013 #endif
2014 	kmem_free(db_search, sizeof (dmu_buf_impl_t));
2015 	mutex_exit(&dn->dn_dbufs_mtx);
2016 }
2017 
2018 void
2019 dnode_set_dirtyctx(dnode_t *dn, dmu_tx_t *tx, void *tag)
2020 {
2021 	/*
2022 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
2023 	 * initialize the objset.
2024 	 */
2025 	if (dn->dn_dirtyctx == DN_UNDIRTIED) {
2026 		dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2027 
2028 		if (ds != NULL) {
2029 			rrw_enter(&ds->ds_bp_rwlock, RW_READER, tag);
2030 		}
2031 		if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
2032 			if (dmu_tx_is_syncing(tx))
2033 				dn->dn_dirtyctx = DN_DIRTY_SYNC;
2034 			else
2035 				dn->dn_dirtyctx = DN_DIRTY_OPEN;
2036 			dn->dn_dirtyctx_firstset = tag;
2037 		}
2038 		if (ds != NULL) {
2039 			rrw_exit(&ds->ds_bp_rwlock, tag);
2040 		}
2041 	}
2042 }
2043 
2044 void
2045 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
2046 {
2047 	dmu_buf_impl_t *db;
2048 	uint64_t blkoff, blkid, nblks;
2049 	int blksz, blkshift, head, tail;
2050 	int trunc = FALSE;
2051 	int epbs;
2052 
2053 	blksz = dn->dn_datablksz;
2054 	blkshift = dn->dn_datablkshift;
2055 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2056 
2057 	if (len == DMU_OBJECT_END) {
2058 		len = UINT64_MAX - off;
2059 		trunc = TRUE;
2060 	}
2061 
2062 	/*
2063 	 * First, block align the region to free:
2064 	 */
2065 	if (ISP2(blksz)) {
2066 		head = P2NPHASE(off, blksz);
2067 		blkoff = P2PHASE(off, blksz);
2068 		if ((off >> blkshift) > dn->dn_maxblkid)
2069 			return;
2070 	} else {
2071 		ASSERT(dn->dn_maxblkid == 0);
2072 		if (off == 0 && len >= blksz) {
2073 			/*
2074 			 * Freeing the whole block; fast-track this request.
2075 			 */
2076 			blkid = 0;
2077 			nblks = 1;
2078 			if (dn->dn_nlevels > 1) {
2079 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2080 				dnode_dirty_l1(dn, 0, tx);
2081 				rw_exit(&dn->dn_struct_rwlock);
2082 			}
2083 			goto done;
2084 		} else if (off >= blksz) {
2085 			/* Freeing past end-of-data */
2086 			return;
2087 		} else {
2088 			/* Freeing part of the block. */
2089 			head = blksz - off;
2090 			ASSERT3U(head, >, 0);
2091 		}
2092 		blkoff = off;
2093 	}
2094 	/* zero out any partial block data at the start of the range */
2095 	if (head) {
2096 		int res;
2097 		ASSERT3U(blkoff + head, ==, blksz);
2098 		if (len < head)
2099 			head = len;
2100 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
2101 		res = dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
2102 		    TRUE, FALSE, FTAG, &db);
2103 		rw_exit(&dn->dn_struct_rwlock);
2104 		if (res == 0) {
2105 			caddr_t data;
2106 			boolean_t dirty;
2107 
2108 			db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER,
2109 			    FTAG);
2110 			/* don't dirty if it isn't on disk and isn't dirty */
2111 			dirty = !list_is_empty(&db->db_dirty_records) ||
2112 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr));
2113 			dmu_buf_unlock_parent(db, dblt, FTAG);
2114 			if (dirty) {
2115 				dmu_buf_will_dirty(&db->db, tx);
2116 				data = db->db.db_data;
2117 				bzero(data + blkoff, head);
2118 			}
2119 			dbuf_rele(db, FTAG);
2120 		}
2121 		off += head;
2122 		len -= head;
2123 	}
2124 
2125 	/* If the range was less than one block, we're done */
2126 	if (len == 0)
2127 		return;
2128 
2129 	/* If the remaining range is past end of file, we're done */
2130 	if ((off >> blkshift) > dn->dn_maxblkid)
2131 		return;
2132 
2133 	ASSERT(ISP2(blksz));
2134 	if (trunc)
2135 		tail = 0;
2136 	else
2137 		tail = P2PHASE(len, blksz);
2138 
2139 	ASSERT0(P2PHASE(off, blksz));
2140 	/* zero out any partial block data at the end of the range */
2141 	if (tail) {
2142 		int res;
2143 		if (len < tail)
2144 			tail = len;
2145 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
2146 		res = dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
2147 		    TRUE, FALSE, FTAG, &db);
2148 		rw_exit(&dn->dn_struct_rwlock);
2149 		if (res == 0) {
2150 			boolean_t dirty;
2151 			/* don't dirty if not on disk and not dirty */
2152 			db_lock_type_t type = dmu_buf_lock_parent(db, RW_READER,
2153 			    FTAG);
2154 			dirty = !list_is_empty(&db->db_dirty_records) ||
2155 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr));
2156 			dmu_buf_unlock_parent(db, type, FTAG);
2157 			if (dirty) {
2158 				dmu_buf_will_dirty(&db->db, tx);
2159 				bzero(db->db.db_data, tail);
2160 			}
2161 			dbuf_rele(db, FTAG);
2162 		}
2163 		len -= tail;
2164 	}
2165 
2166 	/* If the range did not include a full block, we are done */
2167 	if (len == 0)
2168 		return;
2169 
2170 	ASSERT(IS_P2ALIGNED(off, blksz));
2171 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2172 	blkid = off >> blkshift;
2173 	nblks = len >> blkshift;
2174 	if (trunc)
2175 		nblks += 1;
2176 
2177 	/*
2178 	 * Dirty all the indirect blocks in this range.  Note that only
2179 	 * the first and last indirect blocks can actually be written
2180 	 * (if they were partially freed) -- they must be dirtied, even if
2181 	 * they do not exist on disk yet.  The interior blocks will
2182 	 * be freed by free_children(), so they will not actually be written.
2183 	 * Even though these interior blocks will not be written, we
2184 	 * dirty them for two reasons:
2185 	 *
2186 	 *  - It ensures that the indirect blocks remain in memory until
2187 	 *    syncing context.  (They have already been prefetched by
2188 	 *    dmu_tx_hold_free(), so we don't have to worry about reading
2189 	 *    them serially here.)
2190 	 *
2191 	 *  - The dirty space accounting will put pressure on the txg sync
2192 	 *    mechanism to begin syncing, and to delay transactions if there
2193 	 *    is a large amount of freeing.  Even though these indirect
2194 	 *    blocks will not be written, we could need to write the same
2195 	 *    amount of space if we copy the freed BPs into deadlists.
2196 	 */
2197 	if (dn->dn_nlevels > 1) {
2198 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2199 		uint64_t first, last;
2200 
2201 		first = blkid >> epbs;
2202 		dnode_dirty_l1(dn, first, tx);
2203 		if (trunc)
2204 			last = dn->dn_maxblkid >> epbs;
2205 		else
2206 			last = (blkid + nblks - 1) >> epbs;
2207 		if (last != first)
2208 			dnode_dirty_l1(dn, last, tx);
2209 
2210 		dnode_dirty_l1range(dn, first, last, tx);
2211 
2212 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2213 		    SPA_BLKPTRSHIFT;
2214 		for (uint64_t i = first + 1; i < last; i++) {
2215 			/*
2216 			 * Set i to the blockid of the next non-hole
2217 			 * level-1 indirect block at or after i.  Note
2218 			 * that dnode_next_offset() operates in terms of
2219 			 * level-0-equivalent bytes.
2220 			 */
2221 			uint64_t ibyte = i << shift;
2222 			int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2223 			    &ibyte, 2, 1, 0);
2224 			i = ibyte >> shift;
2225 			if (i >= last)
2226 				break;
2227 
2228 			/*
2229 			 * Normally we should not see an error, either
2230 			 * from dnode_next_offset() or dbuf_hold_level()
2231 			 * (except for ESRCH from dnode_next_offset).
2232 			 * If there is an i/o error, then when we read
2233 			 * this block in syncing context, it will use
2234 			 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2235 			 * to the "failmode" property.  dnode_next_offset()
2236 			 * doesn't have a flag to indicate MUSTSUCCEED.
2237 			 */
2238 			if (err != 0)
2239 				break;
2240 
2241 			dnode_dirty_l1(dn, i, tx);
2242 		}
2243 		rw_exit(&dn->dn_struct_rwlock);
2244 	}
2245 
2246 done:
2247 	/*
2248 	 * Add this range to the dnode range list.
2249 	 * We will finish up this free operation in the syncing phase.
2250 	 */
2251 	mutex_enter(&dn->dn_mtx);
2252 	{
2253 		int txgoff = tx->tx_txg & TXG_MASK;
2254 		if (dn->dn_free_ranges[txgoff] == NULL) {
2255 			dn->dn_free_ranges[txgoff] = range_tree_create(NULL,
2256 			    RANGE_SEG64, NULL, 0, 0);
2257 		}
2258 		range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2259 		range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2260 	}
2261 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2262 	    blkid, nblks, tx->tx_txg);
2263 	mutex_exit(&dn->dn_mtx);
2264 
2265 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2266 	dnode_setdirty(dn, tx);
2267 }
2268 
2269 static boolean_t
2270 dnode_spill_freed(dnode_t *dn)
2271 {
2272 	int i;
2273 
2274 	mutex_enter(&dn->dn_mtx);
2275 	for (i = 0; i < TXG_SIZE; i++) {
2276 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2277 			break;
2278 	}
2279 	mutex_exit(&dn->dn_mtx);
2280 	return (i < TXG_SIZE);
2281 }
2282 
2283 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2284 uint64_t
2285 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2286 {
2287 	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
2288 	int i;
2289 
2290 	if (blkid == DMU_BONUS_BLKID)
2291 		return (FALSE);
2292 
2293 	/*
2294 	 * If we're in the process of opening the pool, dp will not be
2295 	 * set yet, but there shouldn't be anything dirty.
2296 	 */
2297 	if (dp == NULL)
2298 		return (FALSE);
2299 
2300 	if (dn->dn_free_txg)
2301 		return (TRUE);
2302 
2303 	if (blkid == DMU_SPILL_BLKID)
2304 		return (dnode_spill_freed(dn));
2305 
2306 	mutex_enter(&dn->dn_mtx);
2307 	for (i = 0; i < TXG_SIZE; i++) {
2308 		if (dn->dn_free_ranges[i] != NULL &&
2309 		    range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2310 			break;
2311 	}
2312 	mutex_exit(&dn->dn_mtx);
2313 	return (i < TXG_SIZE);
2314 }
2315 
2316 /* call from syncing context when we actually write/free space for this dnode */
2317 void
2318 dnode_diduse_space(dnode_t *dn, int64_t delta)
2319 {
2320 	uint64_t space;
2321 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2322 	    dn, dn->dn_phys,
2323 	    (u_longlong_t)dn->dn_phys->dn_used,
2324 	    (longlong_t)delta);
2325 
2326 	mutex_enter(&dn->dn_mtx);
2327 	space = DN_USED_BYTES(dn->dn_phys);
2328 	if (delta > 0) {
2329 		ASSERT3U(space + delta, >=, space); /* no overflow */
2330 	} else {
2331 		ASSERT3U(space, >=, -delta); /* no underflow */
2332 	}
2333 	space += delta;
2334 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2335 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2336 		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2337 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2338 	} else {
2339 		dn->dn_phys->dn_used = space;
2340 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2341 	}
2342 	mutex_exit(&dn->dn_mtx);
2343 }
2344 
2345 /*
2346  * Scans a block at the indicated "level" looking for a hole or data,
2347  * depending on 'flags'.
2348  *
2349  * If level > 0, then we are scanning an indirect block looking at its
2350  * pointers.  If level == 0, then we are looking at a block of dnodes.
2351  *
2352  * If we don't find what we are looking for in the block, we return ESRCH.
2353  * Otherwise, return with *offset pointing to the beginning (if searching
2354  * forwards) or end (if searching backwards) of the range covered by the
2355  * block pointer we matched on (or dnode).
2356  *
2357  * The basic search algorithm used below by dnode_next_offset() is to
2358  * use this function to search up the block tree (widen the search) until
2359  * we find something (i.e., we don't return ESRCH) and then search back
2360  * down the tree (narrow the search) until we reach our original search
2361  * level.
2362  */
2363 static int
2364 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2365     int lvl, uint64_t blkfill, uint64_t txg)
2366 {
2367 	dmu_buf_impl_t *db = NULL;
2368 	void *data = NULL;
2369 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2370 	uint64_t epb = 1ULL << epbs;
2371 	uint64_t minfill, maxfill;
2372 	boolean_t hole;
2373 	int i, inc, error, span;
2374 
2375 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2376 
2377 	hole = ((flags & DNODE_FIND_HOLE) != 0);
2378 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2379 	ASSERT(txg == 0 || !hole);
2380 
2381 	if (lvl == dn->dn_phys->dn_nlevels) {
2382 		error = 0;
2383 		epb = dn->dn_phys->dn_nblkptr;
2384 		data = dn->dn_phys->dn_blkptr;
2385 	} else {
2386 		uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2387 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2388 		if (error) {
2389 			if (error != ENOENT)
2390 				return (error);
2391 			if (hole)
2392 				return (0);
2393 			/*
2394 			 * This can only happen when we are searching up
2395 			 * the block tree for data.  We don't really need to
2396 			 * adjust the offset, as we will just end up looking
2397 			 * at the pointer to this block in its parent, and its
2398 			 * going to be unallocated, so we will skip over it.
2399 			 */
2400 			return (SET_ERROR(ESRCH));
2401 		}
2402 		error = dbuf_read(db, NULL,
2403 		    DB_RF_CANFAIL | DB_RF_HAVESTRUCT |
2404 		    DB_RF_NO_DECRYPT | DB_RF_NOPREFETCH);
2405 		if (error) {
2406 			dbuf_rele(db, FTAG);
2407 			return (error);
2408 		}
2409 		data = db->db.db_data;
2410 		rw_enter(&db->db_rwlock, RW_READER);
2411 	}
2412 
2413 	if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2414 	    db->db_blkptr->blk_birth <= txg ||
2415 	    BP_IS_HOLE(db->db_blkptr))) {
2416 		/*
2417 		 * This can only happen when we are searching up the tree
2418 		 * and these conditions mean that we need to keep climbing.
2419 		 */
2420 		error = SET_ERROR(ESRCH);
2421 	} else if (lvl == 0) {
2422 		dnode_phys_t *dnp = data;
2423 
2424 		ASSERT(dn->dn_type == DMU_OT_DNODE);
2425 		ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2426 
2427 		for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2428 		    i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2429 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2430 				break;
2431 		}
2432 
2433 		if (i == blkfill)
2434 			error = SET_ERROR(ESRCH);
2435 
2436 		*offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2437 		    (i << DNODE_SHIFT);
2438 	} else {
2439 		blkptr_t *bp = data;
2440 		uint64_t start = *offset;
2441 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
2442 		minfill = 0;
2443 		maxfill = blkfill << ((lvl - 1) * epbs);
2444 
2445 		if (hole)
2446 			maxfill--;
2447 		else
2448 			minfill++;
2449 
2450 		if (span >= 8 * sizeof (*offset)) {
2451 			/* This only happens on the highest indirection level */
2452 			ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1);
2453 			*offset = 0;
2454 		} else {
2455 			*offset = *offset >> span;
2456 		}
2457 
2458 		for (i = BF64_GET(*offset, 0, epbs);
2459 		    i >= 0 && i < epb; i += inc) {
2460 			if (BP_GET_FILL(&bp[i]) >= minfill &&
2461 			    BP_GET_FILL(&bp[i]) <= maxfill &&
2462 			    (hole || bp[i].blk_birth > txg))
2463 				break;
2464 			if (inc > 0 || *offset > 0)
2465 				*offset += inc;
2466 		}
2467 
2468 		if (span >= 8 * sizeof (*offset)) {
2469 			*offset = start;
2470 		} else {
2471 			*offset = *offset << span;
2472 		}
2473 
2474 		if (inc < 0) {
2475 			/* traversing backwards; position offset at the end */
2476 			ASSERT3U(*offset, <=, start);
2477 			*offset = MIN(*offset + (1ULL << span) - 1, start);
2478 		} else if (*offset < start) {
2479 			*offset = start;
2480 		}
2481 		if (i < 0 || i >= epb)
2482 			error = SET_ERROR(ESRCH);
2483 	}
2484 
2485 	if (db != NULL) {
2486 		rw_exit(&db->db_rwlock);
2487 		dbuf_rele(db, FTAG);
2488 	}
2489 
2490 	return (error);
2491 }
2492 
2493 /*
2494  * Find the next hole, data, or sparse region at or after *offset.
2495  * The value 'blkfill' tells us how many items we expect to find
2496  * in an L0 data block; this value is 1 for normal objects,
2497  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2498  * DNODES_PER_BLOCK when searching for sparse regions thereof.
2499  *
2500  * Examples:
2501  *
2502  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2503  *	Finds the next/previous hole/data in a file.
2504  *	Used in dmu_offset_next().
2505  *
2506  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2507  *	Finds the next free/allocated dnode an objset's meta-dnode.
2508  *	Only finds objects that have new contents since txg (ie.
2509  *	bonus buffer changes and content removal are ignored).
2510  *	Used in dmu_object_next().
2511  *
2512  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2513  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
2514  *	Used in dmu_object_alloc().
2515  */
2516 int
2517 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2518     int minlvl, uint64_t blkfill, uint64_t txg)
2519 {
2520 	uint64_t initial_offset = *offset;
2521 	int lvl, maxlvl;
2522 	int error = 0;
2523 
2524 	if (!(flags & DNODE_FIND_HAVELOCK))
2525 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
2526 
2527 	if (dn->dn_phys->dn_nlevels == 0) {
2528 		error = SET_ERROR(ESRCH);
2529 		goto out;
2530 	}
2531 
2532 	if (dn->dn_datablkshift == 0) {
2533 		if (*offset < dn->dn_datablksz) {
2534 			if (flags & DNODE_FIND_HOLE)
2535 				*offset = dn->dn_datablksz;
2536 		} else {
2537 			error = SET_ERROR(ESRCH);
2538 		}
2539 		goto out;
2540 	}
2541 
2542 	maxlvl = dn->dn_phys->dn_nlevels;
2543 
2544 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2545 		error = dnode_next_offset_level(dn,
2546 		    flags, offset, lvl, blkfill, txg);
2547 		if (error != ESRCH)
2548 			break;
2549 	}
2550 
2551 	while (error == 0 && --lvl >= minlvl) {
2552 		error = dnode_next_offset_level(dn,
2553 		    flags, offset, lvl, blkfill, txg);
2554 	}
2555 
2556 	/*
2557 	 * There's always a "virtual hole" at the end of the object, even
2558 	 * if all BP's which physically exist are non-holes.
2559 	 */
2560 	if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2561 	    minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2562 		error = 0;
2563 	}
2564 
2565 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2566 	    initial_offset < *offset : initial_offset > *offset))
2567 		error = SET_ERROR(ESRCH);
2568 out:
2569 	if (!(flags & DNODE_FIND_HAVELOCK))
2570 		rw_exit(&dn->dn_struct_rwlock);
2571 
2572 	return (error);
2573 }
2574 
2575 #if defined(_KERNEL)
2576 EXPORT_SYMBOL(dnode_hold);
2577 EXPORT_SYMBOL(dnode_rele);
2578 EXPORT_SYMBOL(dnode_set_nlevels);
2579 EXPORT_SYMBOL(dnode_set_blksz);
2580 EXPORT_SYMBOL(dnode_free_range);
2581 EXPORT_SYMBOL(dnode_evict_dbufs);
2582 EXPORT_SYMBOL(dnode_evict_bonus);
2583 #endif
2584