xref: /freebsd/sys/contrib/openzfs/module/zfs/dnode.c (revision 924226fba12cc9a228c73b956e1b7fa24c60b055)
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 static int
112 dnode_cons(void *arg, void *unused, int kmflag)
113 {
114 	(void) unused, (void) kmflag;
115 	dnode_t *dn = arg;
116 
117 	rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL);
118 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
119 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
120 	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
121 	cv_init(&dn->dn_nodnholds, NULL, CV_DEFAULT, NULL);
122 
123 	/*
124 	 * Every dbuf has a reference, and dropping a tracked reference is
125 	 * O(number of references), so don't track dn_holds.
126 	 */
127 	zfs_refcount_create_untracked(&dn->dn_holds);
128 	zfs_refcount_create(&dn->dn_tx_holds);
129 	list_link_init(&dn->dn_link);
130 
131 	memset(dn->dn_next_type, 0, sizeof (dn->dn_next_type));
132 	memset(dn->dn_next_nblkptr, 0, sizeof (dn->dn_next_nblkptr));
133 	memset(dn->dn_next_nlevels, 0, sizeof (dn->dn_next_nlevels));
134 	memset(dn->dn_next_indblkshift, 0, sizeof (dn->dn_next_indblkshift));
135 	memset(dn->dn_next_bonustype, 0, sizeof (dn->dn_next_bonustype));
136 	memset(dn->dn_rm_spillblk, 0, sizeof (dn->dn_rm_spillblk));
137 	memset(dn->dn_next_bonuslen, 0, sizeof (dn->dn_next_bonuslen));
138 	memset(dn->dn_next_blksz, 0, sizeof (dn->dn_next_blksz));
139 	memset(dn->dn_next_maxblkid, 0, sizeof (dn->dn_next_maxblkid));
140 
141 	for (int 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 static void
177 dnode_dest(void *arg, void *unused)
178 {
179 	(void) unused;
180 	dnode_t *dn = arg;
181 
182 	rw_destroy(&dn->dn_struct_rwlock);
183 	mutex_destroy(&dn->dn_mtx);
184 	mutex_destroy(&dn->dn_dbufs_mtx);
185 	cv_destroy(&dn->dn_notxholds);
186 	cv_destroy(&dn->dn_nodnholds);
187 	zfs_refcount_destroy(&dn->dn_holds);
188 	zfs_refcount_destroy(&dn->dn_tx_holds);
189 	ASSERT(!list_link_active(&dn->dn_link));
190 
191 	for (int i = 0; i < TXG_SIZE; i++) {
192 		ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
193 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
194 		list_destroy(&dn->dn_dirty_records[i]);
195 		ASSERT0(dn->dn_next_nblkptr[i]);
196 		ASSERT0(dn->dn_next_nlevels[i]);
197 		ASSERT0(dn->dn_next_indblkshift[i]);
198 		ASSERT0(dn->dn_next_bonustype[i]);
199 		ASSERT0(dn->dn_rm_spillblk[i]);
200 		ASSERT0(dn->dn_next_bonuslen[i]);
201 		ASSERT0(dn->dn_next_blksz[i]);
202 		ASSERT0(dn->dn_next_maxblkid[i]);
203 	}
204 
205 	ASSERT0(dn->dn_allocated_txg);
206 	ASSERT0(dn->dn_free_txg);
207 	ASSERT0(dn->dn_assigned_txg);
208 	ASSERT0(dn->dn_dirty_txg);
209 	ASSERT0(dn->dn_dirtyctx);
210 	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
211 	ASSERT3P(dn->dn_bonus, ==, NULL);
212 	ASSERT(!dn->dn_have_spill);
213 	ASSERT3P(dn->dn_zio, ==, NULL);
214 	ASSERT0(dn->dn_oldused);
215 	ASSERT0(dn->dn_oldflags);
216 	ASSERT0(dn->dn_olduid);
217 	ASSERT0(dn->dn_oldgid);
218 	ASSERT0(dn->dn_oldprojid);
219 	ASSERT0(dn->dn_newuid);
220 	ASSERT0(dn->dn_newgid);
221 	ASSERT0(dn->dn_newprojid);
222 	ASSERT0(dn->dn_id_flags);
223 
224 	ASSERT0(dn->dn_dbufs_count);
225 	avl_destroy(&dn->dn_dbufs);
226 }
227 
228 void
229 dnode_init(void)
230 {
231 	ASSERT(dnode_cache == NULL);
232 	dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
233 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
234 	kmem_cache_set_move(dnode_cache, dnode_move);
235 
236 	dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
237 	    KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
238 	    KSTAT_FLAG_VIRTUAL);
239 	if (dnode_ksp != NULL) {
240 		dnode_ksp->ks_data = &dnode_stats;
241 		kstat_install(dnode_ksp);
242 	}
243 }
244 
245 void
246 dnode_fini(void)
247 {
248 	if (dnode_ksp != NULL) {
249 		kstat_delete(dnode_ksp);
250 		dnode_ksp = NULL;
251 	}
252 
253 	kmem_cache_destroy(dnode_cache);
254 	dnode_cache = NULL;
255 }
256 
257 
258 #ifdef ZFS_DEBUG
259 void
260 dnode_verify(dnode_t *dn)
261 {
262 	int drop_struct_lock = FALSE;
263 
264 	ASSERT(dn->dn_phys);
265 	ASSERT(dn->dn_objset);
266 	ASSERT(dn->dn_handle->dnh_dnode == dn);
267 
268 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
269 
270 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
271 		return;
272 
273 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
274 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
275 		drop_struct_lock = TRUE;
276 	}
277 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
278 		int i;
279 		int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
280 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
281 		if (dn->dn_datablkshift) {
282 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
283 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
284 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
285 		}
286 		ASSERT3U(dn->dn_nlevels, <=, 30);
287 		ASSERT(DMU_OT_IS_VALID(dn->dn_type));
288 		ASSERT3U(dn->dn_nblkptr, >=, 1);
289 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
290 		ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
291 		ASSERT3U(dn->dn_datablksz, ==,
292 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
293 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
294 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
295 		    dn->dn_bonuslen, <=, max_bonuslen);
296 		for (i = 0; i < TXG_SIZE; i++) {
297 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
298 		}
299 	}
300 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
301 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
302 	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
303 	if (dn->dn_dbuf != NULL) {
304 		ASSERT3P(dn->dn_phys, ==,
305 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
306 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
307 	}
308 	if (drop_struct_lock)
309 		rw_exit(&dn->dn_struct_rwlock);
310 }
311 #endif
312 
313 void
314 dnode_byteswap(dnode_phys_t *dnp)
315 {
316 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
317 	int i;
318 
319 	if (dnp->dn_type == DMU_OT_NONE) {
320 		memset(dnp, 0, sizeof (dnode_phys_t));
321 		return;
322 	}
323 
324 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
325 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
326 	dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
327 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
328 	dnp->dn_used = BSWAP_64(dnp->dn_used);
329 
330 	/*
331 	 * dn_nblkptr is only one byte, so it's OK to read it in either
332 	 * byte order.  We can't read dn_bouslen.
333 	 */
334 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
335 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
336 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
337 		buf64[i] = BSWAP_64(buf64[i]);
338 
339 	/*
340 	 * OK to check dn_bonuslen for zero, because it won't matter if
341 	 * we have the wrong byte order.  This is necessary because the
342 	 * dnode dnode is smaller than a regular dnode.
343 	 */
344 	if (dnp->dn_bonuslen != 0) {
345 		/*
346 		 * Note that the bonus length calculated here may be
347 		 * longer than the actual bonus buffer.  This is because
348 		 * we always put the bonus buffer after the last block
349 		 * pointer (instead of packing it against the end of the
350 		 * dnode buffer).
351 		 */
352 		int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
353 		int slots = dnp->dn_extra_slots + 1;
354 		size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
355 		dmu_object_byteswap_t byteswap;
356 		ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
357 		byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
358 		dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
359 	}
360 
361 	/* Swap SPILL block if we have one */
362 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
363 		byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
364 }
365 
366 void
367 dnode_buf_byteswap(void *vbuf, size_t size)
368 {
369 	int i = 0;
370 
371 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
372 	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
373 
374 	while (i < size) {
375 		dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
376 		dnode_byteswap(dnp);
377 
378 		i += DNODE_MIN_SIZE;
379 		if (dnp->dn_type != DMU_OT_NONE)
380 			i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
381 	}
382 }
383 
384 void
385 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
386 {
387 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
388 
389 	dnode_setdirty(dn, tx);
390 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
391 	ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
392 	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
393 
394 	if (newsize < dn->dn_bonuslen) {
395 		/* clear any data after the end of the new size */
396 		size_t diff = dn->dn_bonuslen - newsize;
397 		char *data_end = ((char *)dn->dn_bonus->db.db_data) + newsize;
398 		memset(data_end, 0, diff);
399 	}
400 
401 	dn->dn_bonuslen = newsize;
402 	if (newsize == 0)
403 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
404 	else
405 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
406 	rw_exit(&dn->dn_struct_rwlock);
407 }
408 
409 void
410 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
411 {
412 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
413 	dnode_setdirty(dn, tx);
414 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
415 	dn->dn_bonustype = newtype;
416 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
417 	rw_exit(&dn->dn_struct_rwlock);
418 }
419 
420 void
421 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
422 {
423 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
424 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
425 	dnode_setdirty(dn, tx);
426 	dn->dn_rm_spillblk[tx->tx_txg & TXG_MASK] = DN_KILL_SPILLBLK;
427 	dn->dn_have_spill = B_FALSE;
428 }
429 
430 static void
431 dnode_setdblksz(dnode_t *dn, int size)
432 {
433 	ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
434 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
435 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
436 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
437 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
438 	dn->dn_datablksz = size;
439 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
440 	dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
441 }
442 
443 static dnode_t *
444 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
445     uint64_t object, dnode_handle_t *dnh)
446 {
447 	dnode_t *dn;
448 
449 	dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
450 	dn->dn_moved = 0;
451 
452 	/*
453 	 * Defer setting dn_objset until the dnode is ready to be a candidate
454 	 * for the dnode_move() callback.
455 	 */
456 	dn->dn_object = object;
457 	dn->dn_dbuf = db;
458 	dn->dn_handle = dnh;
459 	dn->dn_phys = dnp;
460 
461 	if (dnp->dn_datablkszsec) {
462 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
463 	} else {
464 		dn->dn_datablksz = 0;
465 		dn->dn_datablkszsec = 0;
466 		dn->dn_datablkshift = 0;
467 	}
468 	dn->dn_indblkshift = dnp->dn_indblkshift;
469 	dn->dn_nlevels = dnp->dn_nlevels;
470 	dn->dn_type = dnp->dn_type;
471 	dn->dn_nblkptr = dnp->dn_nblkptr;
472 	dn->dn_checksum = dnp->dn_checksum;
473 	dn->dn_compress = dnp->dn_compress;
474 	dn->dn_bonustype = dnp->dn_bonustype;
475 	dn->dn_bonuslen = dnp->dn_bonuslen;
476 	dn->dn_num_slots = dnp->dn_extra_slots + 1;
477 	dn->dn_maxblkid = dnp->dn_maxblkid;
478 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
479 	dn->dn_id_flags = 0;
480 
481 	dmu_zfetch_init(&dn->dn_zfetch, dn);
482 
483 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
484 	ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
485 	ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
486 
487 	mutex_enter(&os->os_lock);
488 
489 	/*
490 	 * Exclude special dnodes from os_dnodes so an empty os_dnodes
491 	 * signifies that the special dnodes have no references from
492 	 * their children (the entries in os_dnodes).  This allows
493 	 * dnode_destroy() to easily determine if the last child has
494 	 * been removed and then complete eviction of the objset.
495 	 */
496 	if (!DMU_OBJECT_IS_SPECIAL(object))
497 		list_insert_head(&os->os_dnodes, dn);
498 	membar_producer();
499 
500 	/*
501 	 * Everything else must be valid before assigning dn_objset
502 	 * makes the dnode eligible for dnode_move().
503 	 */
504 	dn->dn_objset = os;
505 
506 	dnh->dnh_dnode = dn;
507 	mutex_exit(&os->os_lock);
508 
509 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
510 
511 	return (dn);
512 }
513 
514 /*
515  * Caller must be holding the dnode handle, which is released upon return.
516  */
517 static void
518 dnode_destroy(dnode_t *dn)
519 {
520 	objset_t *os = dn->dn_objset;
521 	boolean_t complete_os_eviction = B_FALSE;
522 
523 	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
524 
525 	mutex_enter(&os->os_lock);
526 	POINTER_INVALIDATE(&dn->dn_objset);
527 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
528 		list_remove(&os->os_dnodes, dn);
529 		complete_os_eviction =
530 		    list_is_empty(&os->os_dnodes) &&
531 		    list_link_active(&os->os_evicting_node);
532 	}
533 	mutex_exit(&os->os_lock);
534 
535 	/* the dnode can no longer move, so we can release the handle */
536 	if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
537 		zrl_remove(&dn->dn_handle->dnh_zrlock);
538 
539 	dn->dn_allocated_txg = 0;
540 	dn->dn_free_txg = 0;
541 	dn->dn_assigned_txg = 0;
542 	dn->dn_dirty_txg = 0;
543 
544 	dn->dn_dirtyctx = 0;
545 	dn->dn_dirtyctx_firstset = NULL;
546 	if (dn->dn_bonus != NULL) {
547 		mutex_enter(&dn->dn_bonus->db_mtx);
548 		dbuf_destroy(dn->dn_bonus);
549 		dn->dn_bonus = NULL;
550 	}
551 	dn->dn_zio = NULL;
552 
553 	dn->dn_have_spill = B_FALSE;
554 	dn->dn_oldused = 0;
555 	dn->dn_oldflags = 0;
556 	dn->dn_olduid = 0;
557 	dn->dn_oldgid = 0;
558 	dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
559 	dn->dn_newuid = 0;
560 	dn->dn_newgid = 0;
561 	dn->dn_newprojid = ZFS_DEFAULT_PROJID;
562 	dn->dn_id_flags = 0;
563 
564 	dmu_zfetch_fini(&dn->dn_zfetch);
565 	kmem_cache_free(dnode_cache, dn);
566 	arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
567 
568 	if (complete_os_eviction)
569 		dmu_objset_evict_done(os);
570 }
571 
572 void
573 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
574     dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
575 {
576 	int i;
577 
578 	ASSERT3U(dn_slots, >, 0);
579 	ASSERT3U(dn_slots << DNODE_SHIFT, <=,
580 	    spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
581 	ASSERT3U(blocksize, <=,
582 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
583 	if (blocksize == 0)
584 		blocksize = 1 << zfs_default_bs;
585 	else
586 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
587 
588 	if (ibs == 0)
589 		ibs = zfs_default_ibs;
590 
591 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
592 
593 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
594 	    dn->dn_objset, (u_longlong_t)dn->dn_object,
595 	    (u_longlong_t)tx->tx_txg, blocksize, ibs, dn_slots);
596 	DNODE_STAT_BUMP(dnode_allocate);
597 
598 	ASSERT(dn->dn_type == DMU_OT_NONE);
599 	ASSERT0(memcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)));
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 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
753 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
754 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
755 
756 	/* Copy fields. */
757 	ndn->dn_objset = odn->dn_objset;
758 	ndn->dn_object = odn->dn_object;
759 	ndn->dn_dbuf = odn->dn_dbuf;
760 	ndn->dn_handle = odn->dn_handle;
761 	ndn->dn_phys = odn->dn_phys;
762 	ndn->dn_type = odn->dn_type;
763 	ndn->dn_bonuslen = odn->dn_bonuslen;
764 	ndn->dn_bonustype = odn->dn_bonustype;
765 	ndn->dn_nblkptr = odn->dn_nblkptr;
766 	ndn->dn_checksum = odn->dn_checksum;
767 	ndn->dn_compress = odn->dn_compress;
768 	ndn->dn_nlevels = odn->dn_nlevels;
769 	ndn->dn_indblkshift = odn->dn_indblkshift;
770 	ndn->dn_datablkshift = odn->dn_datablkshift;
771 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
772 	ndn->dn_datablksz = odn->dn_datablksz;
773 	ndn->dn_maxblkid = odn->dn_maxblkid;
774 	ndn->dn_num_slots = odn->dn_num_slots;
775 	memcpy(ndn->dn_next_type, odn->dn_next_type,
776 	    sizeof (odn->dn_next_type));
777 	memcpy(ndn->dn_next_nblkptr, odn->dn_next_nblkptr,
778 	    sizeof (odn->dn_next_nblkptr));
779 	memcpy(ndn->dn_next_nlevels, odn->dn_next_nlevels,
780 	    sizeof (odn->dn_next_nlevels));
781 	memcpy(ndn->dn_next_indblkshift, odn->dn_next_indblkshift,
782 	    sizeof (odn->dn_next_indblkshift));
783 	memcpy(ndn->dn_next_bonustype, odn->dn_next_bonustype,
784 	    sizeof (odn->dn_next_bonustype));
785 	memcpy(ndn->dn_rm_spillblk, odn->dn_rm_spillblk,
786 	    sizeof (odn->dn_rm_spillblk));
787 	memcpy(ndn->dn_next_bonuslen, odn->dn_next_bonuslen,
788 	    sizeof (odn->dn_next_bonuslen));
789 	memcpy(ndn->dn_next_blksz, odn->dn_next_blksz,
790 	    sizeof (odn->dn_next_blksz));
791 	memcpy(ndn->dn_next_maxblkid, odn->dn_next_maxblkid,
792 	    sizeof (odn->dn_next_maxblkid));
793 	for (int i = 0; i < TXG_SIZE; i++) {
794 		list_move_tail(&ndn->dn_dirty_records[i],
795 		    &odn->dn_dirty_records[i]);
796 	}
797 	memcpy(ndn->dn_free_ranges, odn->dn_free_ranges,
798 	    sizeof (odn->dn_free_ranges));
799 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
800 	ndn->dn_free_txg = odn->dn_free_txg;
801 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
802 	ndn->dn_dirty_txg = odn->dn_dirty_txg;
803 	ndn->dn_dirtyctx = odn->dn_dirtyctx;
804 	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
805 	ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0);
806 	zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
807 	ASSERT(avl_is_empty(&ndn->dn_dbufs));
808 	avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
809 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
810 	ndn->dn_bonus = odn->dn_bonus;
811 	ndn->dn_have_spill = odn->dn_have_spill;
812 	ndn->dn_zio = odn->dn_zio;
813 	ndn->dn_oldused = odn->dn_oldused;
814 	ndn->dn_oldflags = odn->dn_oldflags;
815 	ndn->dn_olduid = odn->dn_olduid;
816 	ndn->dn_oldgid = odn->dn_oldgid;
817 	ndn->dn_oldprojid = odn->dn_oldprojid;
818 	ndn->dn_newuid = odn->dn_newuid;
819 	ndn->dn_newgid = odn->dn_newgid;
820 	ndn->dn_newprojid = odn->dn_newprojid;
821 	ndn->dn_id_flags = odn->dn_id_flags;
822 	dmu_zfetch_init(&ndn->dn_zfetch, ndn);
823 
824 	/*
825 	 * Update back pointers. Updating the handle fixes the back pointer of
826 	 * every descendant dbuf as well as the bonus dbuf.
827 	 */
828 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
829 	ndn->dn_handle->dnh_dnode = ndn;
830 
831 	/*
832 	 * Invalidate the original dnode by clearing all of its back pointers.
833 	 */
834 	odn->dn_dbuf = NULL;
835 	odn->dn_handle = NULL;
836 	avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
837 	    offsetof(dmu_buf_impl_t, db_link));
838 	odn->dn_dbufs_count = 0;
839 	odn->dn_bonus = NULL;
840 	dmu_zfetch_fini(&odn->dn_zfetch);
841 
842 	/*
843 	 * Set the low bit of the objset pointer to ensure that dnode_move()
844 	 * recognizes the dnode as invalid in any subsequent callback.
845 	 */
846 	POINTER_INVALIDATE(&odn->dn_objset);
847 
848 	/*
849 	 * Satisfy the destructor.
850 	 */
851 	for (int i = 0; i < TXG_SIZE; i++) {
852 		list_create(&odn->dn_dirty_records[i],
853 		    sizeof (dbuf_dirty_record_t),
854 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
855 		odn->dn_free_ranges[i] = NULL;
856 		odn->dn_next_nlevels[i] = 0;
857 		odn->dn_next_indblkshift[i] = 0;
858 		odn->dn_next_bonustype[i] = 0;
859 		odn->dn_rm_spillblk[i] = 0;
860 		odn->dn_next_bonuslen[i] = 0;
861 		odn->dn_next_blksz[i] = 0;
862 	}
863 	odn->dn_allocated_txg = 0;
864 	odn->dn_free_txg = 0;
865 	odn->dn_assigned_txg = 0;
866 	odn->dn_dirty_txg = 0;
867 	odn->dn_dirtyctx = 0;
868 	odn->dn_dirtyctx_firstset = NULL;
869 	odn->dn_have_spill = B_FALSE;
870 	odn->dn_zio = NULL;
871 	odn->dn_oldused = 0;
872 	odn->dn_oldflags = 0;
873 	odn->dn_olduid = 0;
874 	odn->dn_oldgid = 0;
875 	odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
876 	odn->dn_newuid = 0;
877 	odn->dn_newgid = 0;
878 	odn->dn_newprojid = ZFS_DEFAULT_PROJID;
879 	odn->dn_id_flags = 0;
880 
881 	/*
882 	 * Mark the dnode.
883 	 */
884 	ndn->dn_moved = 1;
885 	odn->dn_moved = (uint8_t)-1;
886 }
887 
888 static kmem_cbrc_t
889 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
890 {
891 	dnode_t *odn = buf, *ndn = newbuf;
892 	objset_t *os;
893 	int64_t refcount;
894 	uint32_t dbufs;
895 
896 	/*
897 	 * The dnode is on the objset's list of known dnodes if the objset
898 	 * pointer is valid. We set the low bit of the objset pointer when
899 	 * freeing the dnode to invalidate it, and the memory patterns written
900 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
901 	 * A newly created dnode sets the objset pointer last of all to indicate
902 	 * that the dnode is known and in a valid state to be moved by this
903 	 * function.
904 	 */
905 	os = odn->dn_objset;
906 	if (!POINTER_IS_VALID(os)) {
907 		DNODE_STAT_BUMP(dnode_move_invalid);
908 		return (KMEM_CBRC_DONT_KNOW);
909 	}
910 
911 	/*
912 	 * Ensure that the objset does not go away during the move.
913 	 */
914 	rw_enter(&os_lock, RW_WRITER);
915 	if (os != odn->dn_objset) {
916 		rw_exit(&os_lock);
917 		DNODE_STAT_BUMP(dnode_move_recheck1);
918 		return (KMEM_CBRC_DONT_KNOW);
919 	}
920 
921 	/*
922 	 * If the dnode is still valid, then so is the objset. We know that no
923 	 * valid objset can be freed while we hold os_lock, so we can safely
924 	 * ensure that the objset remains in use.
925 	 */
926 	mutex_enter(&os->os_lock);
927 
928 	/*
929 	 * Recheck the objset pointer in case the dnode was removed just before
930 	 * acquiring the lock.
931 	 */
932 	if (os != odn->dn_objset) {
933 		mutex_exit(&os->os_lock);
934 		rw_exit(&os_lock);
935 		DNODE_STAT_BUMP(dnode_move_recheck2);
936 		return (KMEM_CBRC_DONT_KNOW);
937 	}
938 
939 	/*
940 	 * At this point we know that as long as we hold os->os_lock, the dnode
941 	 * cannot be freed and fields within the dnode can be safely accessed.
942 	 * The objset listing this dnode cannot go away as long as this dnode is
943 	 * on its list.
944 	 */
945 	rw_exit(&os_lock);
946 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
947 		mutex_exit(&os->os_lock);
948 		DNODE_STAT_BUMP(dnode_move_special);
949 		return (KMEM_CBRC_NO);
950 	}
951 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
952 
953 	/*
954 	 * Lock the dnode handle to prevent the dnode from obtaining any new
955 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
956 	 * from accessing the dnode, so that we can discount their holds. The
957 	 * handle is safe to access because we know that while the dnode cannot
958 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
959 	 * safely move any dnode referenced only by dbufs.
960 	 */
961 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
962 		mutex_exit(&os->os_lock);
963 		DNODE_STAT_BUMP(dnode_move_handle);
964 		return (KMEM_CBRC_LATER);
965 	}
966 
967 	/*
968 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
969 	 * We need to guarantee that there is a hold for every dbuf in order to
970 	 * determine whether the dnode is actively referenced. Falsely matching
971 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
972 	 * that a thread already having an active dnode hold is about to add a
973 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
974 	 * progress.
975 	 */
976 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
977 		zrl_exit(&odn->dn_handle->dnh_zrlock);
978 		mutex_exit(&os->os_lock);
979 		DNODE_STAT_BUMP(dnode_move_rwlock);
980 		return (KMEM_CBRC_LATER);
981 	}
982 
983 	/*
984 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
985 	 * case, the dbuf count is decremented under the handle lock before the
986 	 * dbuf's hold is released. This order ensures that if we count the hold
987 	 * after the dbuf is removed but before its hold is released, we will
988 	 * treat the unmatched hold as active and exit safely. If we count the
989 	 * hold before the dbuf is removed, the hold is discounted, and the
990 	 * removal is blocked until the move completes.
991 	 */
992 	refcount = zfs_refcount_count(&odn->dn_holds);
993 	ASSERT(refcount >= 0);
994 	dbufs = DN_DBUFS_COUNT(odn);
995 
996 	/* We can't have more dbufs than dnode holds. */
997 	ASSERT3U(dbufs, <=, refcount);
998 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
999 	    uint32_t, dbufs);
1000 
1001 	if (refcount > dbufs) {
1002 		rw_exit(&odn->dn_struct_rwlock);
1003 		zrl_exit(&odn->dn_handle->dnh_zrlock);
1004 		mutex_exit(&os->os_lock);
1005 		DNODE_STAT_BUMP(dnode_move_active);
1006 		return (KMEM_CBRC_LATER);
1007 	}
1008 
1009 	rw_exit(&odn->dn_struct_rwlock);
1010 
1011 	/*
1012 	 * At this point we know that anyone with a hold on the dnode is not
1013 	 * actively referencing it. The dnode is known and in a valid state to
1014 	 * move. We're holding the locks needed to execute the critical section.
1015 	 */
1016 	dnode_move_impl(odn, ndn);
1017 
1018 	list_link_replace(&odn->dn_link, &ndn->dn_link);
1019 	/* If the dnode was safe to move, the refcount cannot have changed. */
1020 	ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1021 	ASSERT(dbufs == DN_DBUFS_COUNT(ndn));
1022 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1023 	mutex_exit(&os->os_lock);
1024 
1025 	return (KMEM_CBRC_YES);
1026 }
1027 #endif	/* _KERNEL */
1028 
1029 static void
1030 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1031 {
1032 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1033 
1034 	for (int i = idx; i < idx + slots; i++) {
1035 		dnode_handle_t *dnh = &children->dnc_children[i];
1036 		zrl_add(&dnh->dnh_zrlock);
1037 	}
1038 }
1039 
1040 static void
1041 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1042 {
1043 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1044 
1045 	for (int i = idx; i < idx + slots; i++) {
1046 		dnode_handle_t *dnh = &children->dnc_children[i];
1047 
1048 		if (zrl_is_locked(&dnh->dnh_zrlock))
1049 			zrl_exit(&dnh->dnh_zrlock);
1050 		else
1051 			zrl_remove(&dnh->dnh_zrlock);
1052 	}
1053 }
1054 
1055 static int
1056 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1057 {
1058 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1059 
1060 	for (int i = idx; i < idx + slots; i++) {
1061 		dnode_handle_t *dnh = &children->dnc_children[i];
1062 
1063 		if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1064 			for (int j = idx; j < i; j++) {
1065 				dnh = &children->dnc_children[j];
1066 				zrl_exit(&dnh->dnh_zrlock);
1067 			}
1068 
1069 			return (0);
1070 		}
1071 	}
1072 
1073 	return (1);
1074 }
1075 
1076 static void
1077 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1078 {
1079 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1080 
1081 	for (int i = idx; i < idx + slots; i++) {
1082 		dnode_handle_t *dnh = &children->dnc_children[i];
1083 		dnh->dnh_dnode = ptr;
1084 	}
1085 }
1086 
1087 static boolean_t
1088 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1089 {
1090 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1091 
1092 	/*
1093 	 * If all dnode slots are either already free or
1094 	 * evictable return B_TRUE.
1095 	 */
1096 	for (int i = idx; i < idx + slots; i++) {
1097 		dnode_handle_t *dnh = &children->dnc_children[i];
1098 		dnode_t *dn = dnh->dnh_dnode;
1099 
1100 		if (dn == DN_SLOT_FREE) {
1101 			continue;
1102 		} else if (DN_SLOT_IS_PTR(dn)) {
1103 			mutex_enter(&dn->dn_mtx);
1104 			boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1105 			    zfs_refcount_is_zero(&dn->dn_holds) &&
1106 			    !DNODE_IS_DIRTY(dn));
1107 			mutex_exit(&dn->dn_mtx);
1108 
1109 			if (!can_free)
1110 				return (B_FALSE);
1111 			else
1112 				continue;
1113 		} else {
1114 			return (B_FALSE);
1115 		}
1116 	}
1117 
1118 	return (B_TRUE);
1119 }
1120 
1121 static void
1122 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1123 {
1124 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1125 
1126 	for (int i = idx; i < idx + slots; i++) {
1127 		dnode_handle_t *dnh = &children->dnc_children[i];
1128 
1129 		ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1130 
1131 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1132 			ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1133 			dnode_destroy(dnh->dnh_dnode);
1134 			dnh->dnh_dnode = DN_SLOT_FREE;
1135 		}
1136 	}
1137 }
1138 
1139 void
1140 dnode_free_interior_slots(dnode_t *dn)
1141 {
1142 	dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1143 	int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1144 	int idx = (dn->dn_object & (epb - 1)) + 1;
1145 	int slots = dn->dn_num_slots - 1;
1146 
1147 	if (slots == 0)
1148 		return;
1149 
1150 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1151 
1152 	while (!dnode_slots_tryenter(children, idx, slots)) {
1153 		DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1154 		cond_resched();
1155 	}
1156 
1157 	dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1158 	dnode_slots_rele(children, idx, slots);
1159 }
1160 
1161 void
1162 dnode_special_close(dnode_handle_t *dnh)
1163 {
1164 	dnode_t *dn = dnh->dnh_dnode;
1165 
1166 	/*
1167 	 * Ensure dnode_rele_and_unlock() has released dn_mtx, after final
1168 	 * zfs_refcount_remove()
1169 	 */
1170 	mutex_enter(&dn->dn_mtx);
1171 	if (zfs_refcount_count(&dn->dn_holds) > 0)
1172 		cv_wait(&dn->dn_nodnholds, &dn->dn_mtx);
1173 	mutex_exit(&dn->dn_mtx);
1174 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 0);
1175 
1176 	ASSERT(dn->dn_dbuf == NULL ||
1177 	    dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1178 	zrl_add(&dnh->dnh_zrlock);
1179 	dnode_destroy(dn); /* implicit zrl_remove() */
1180 	zrl_destroy(&dnh->dnh_zrlock);
1181 	dnh->dnh_dnode = NULL;
1182 }
1183 
1184 void
1185 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1186     dnode_handle_t *dnh)
1187 {
1188 	dnode_t *dn;
1189 
1190 	zrl_init(&dnh->dnh_zrlock);
1191 	VERIFY3U(1, ==, zrl_tryenter(&dnh->dnh_zrlock));
1192 
1193 	dn = dnode_create(os, dnp, NULL, object, dnh);
1194 	DNODE_VERIFY(dn);
1195 
1196 	zrl_exit(&dnh->dnh_zrlock);
1197 }
1198 
1199 static void
1200 dnode_buf_evict_async(void *dbu)
1201 {
1202 	dnode_children_t *dnc = dbu;
1203 
1204 	DNODE_STAT_BUMP(dnode_buf_evict);
1205 
1206 	for (int i = 0; i < dnc->dnc_count; i++) {
1207 		dnode_handle_t *dnh = &dnc->dnc_children[i];
1208 		dnode_t *dn;
1209 
1210 		/*
1211 		 * The dnode handle lock guards against the dnode moving to
1212 		 * another valid address, so there is no need here to guard
1213 		 * against changes to or from NULL.
1214 		 */
1215 		if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1216 			zrl_destroy(&dnh->dnh_zrlock);
1217 			dnh->dnh_dnode = DN_SLOT_UNINIT;
1218 			continue;
1219 		}
1220 
1221 		zrl_add(&dnh->dnh_zrlock);
1222 		dn = dnh->dnh_dnode;
1223 		/*
1224 		 * If there are holds on this dnode, then there should
1225 		 * be holds on the dnode's containing dbuf as well; thus
1226 		 * it wouldn't be eligible for eviction and this function
1227 		 * would not have been called.
1228 		 */
1229 		ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1230 		ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1231 
1232 		dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1233 		zrl_destroy(&dnh->dnh_zrlock);
1234 		dnh->dnh_dnode = DN_SLOT_UNINIT;
1235 	}
1236 	kmem_free(dnc, sizeof (dnode_children_t) +
1237 	    dnc->dnc_count * sizeof (dnode_handle_t));
1238 }
1239 
1240 /*
1241  * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1242  * to ensure the hole at the specified object offset is large enough to
1243  * hold the dnode being created. The slots parameter is also used to ensure
1244  * a dnode does not span multiple dnode blocks. In both of these cases, if
1245  * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1246  * are only possible when using DNODE_MUST_BE_FREE.
1247  *
1248  * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1249  * dnode_hold_impl() will check if the requested dnode is already consumed
1250  * as an extra dnode slot by an large dnode, in which case it returns
1251  * ENOENT.
1252  *
1253  * If the DNODE_DRY_RUN flag is set, we don't actually hold the dnode, just
1254  * return whether the hold would succeed or not. tag and dnp should set to
1255  * NULL in this case.
1256  *
1257  * errors:
1258  * EINVAL - Invalid object number or flags.
1259  * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1260  * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1261  *        - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1262  *        - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1263  * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1264  *        - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1265  * EIO    - I/O error when reading the meta dnode dbuf.
1266  *
1267  * succeeds even for free dnodes.
1268  */
1269 int
1270 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1271     void *tag, dnode_t **dnp)
1272 {
1273 	int epb, idx, err;
1274 	int drop_struct_lock = FALSE;
1275 	int type;
1276 	uint64_t blk;
1277 	dnode_t *mdn, *dn;
1278 	dmu_buf_impl_t *db;
1279 	dnode_children_t *dnc;
1280 	dnode_phys_t *dn_block;
1281 	dnode_handle_t *dnh;
1282 
1283 	ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1284 	ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1285 	IMPLY(flag & DNODE_DRY_RUN, (tag == NULL) && (dnp == NULL));
1286 
1287 	/*
1288 	 * If you are holding the spa config lock as writer, you shouldn't
1289 	 * be asking the DMU to do *anything* unless it's the root pool
1290 	 * which may require us to read from the root filesystem while
1291 	 * holding some (not all) of the locks as writer.
1292 	 */
1293 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1294 	    (spa_is_root(os->os_spa) &&
1295 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1296 
1297 	ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1298 
1299 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1300 	    object == DMU_PROJECTUSED_OBJECT) {
1301 		if (object == DMU_USERUSED_OBJECT)
1302 			dn = DMU_USERUSED_DNODE(os);
1303 		else if (object == DMU_GROUPUSED_OBJECT)
1304 			dn = DMU_GROUPUSED_DNODE(os);
1305 		else
1306 			dn = DMU_PROJECTUSED_DNODE(os);
1307 		if (dn == NULL)
1308 			return (SET_ERROR(ENOENT));
1309 		type = dn->dn_type;
1310 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1311 			return (SET_ERROR(ENOENT));
1312 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1313 			return (SET_ERROR(EEXIST));
1314 		DNODE_VERIFY(dn);
1315 		/* Don't actually hold if dry run, just return 0 */
1316 		if (!(flag & DNODE_DRY_RUN)) {
1317 			(void) zfs_refcount_add(&dn->dn_holds, tag);
1318 			*dnp = dn;
1319 		}
1320 		return (0);
1321 	}
1322 
1323 	if (object == 0 || object >= DN_MAX_OBJECT)
1324 		return (SET_ERROR(EINVAL));
1325 
1326 	mdn = DMU_META_DNODE(os);
1327 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1328 
1329 	DNODE_VERIFY(mdn);
1330 
1331 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1332 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1333 		drop_struct_lock = TRUE;
1334 	}
1335 
1336 	blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1337 	db = dbuf_hold(mdn, blk, FTAG);
1338 	if (drop_struct_lock)
1339 		rw_exit(&mdn->dn_struct_rwlock);
1340 	if (db == NULL) {
1341 		DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1342 		return (SET_ERROR(EIO));
1343 	}
1344 
1345 	/*
1346 	 * We do not need to decrypt to read the dnode so it doesn't matter
1347 	 * if we get the encrypted or decrypted version.
1348 	 */
1349 	err = dbuf_read(db, NULL, DB_RF_CANFAIL |
1350 	    DB_RF_NO_DECRYPT | DB_RF_NOPREFETCH);
1351 	if (err) {
1352 		DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1353 		dbuf_rele(db, FTAG);
1354 		return (err);
1355 	}
1356 
1357 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1358 	epb = db->db.db_size >> DNODE_SHIFT;
1359 
1360 	idx = object & (epb - 1);
1361 	dn_block = (dnode_phys_t *)db->db.db_data;
1362 
1363 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1364 	dnc = dmu_buf_get_user(&db->db);
1365 	dnh = NULL;
1366 	if (dnc == NULL) {
1367 		dnode_children_t *winner;
1368 		int skip = 0;
1369 
1370 		dnc = kmem_zalloc(sizeof (dnode_children_t) +
1371 		    epb * sizeof (dnode_handle_t), KM_SLEEP);
1372 		dnc->dnc_count = epb;
1373 		dnh = &dnc->dnc_children[0];
1374 
1375 		/* Initialize dnode slot status from dnode_phys_t */
1376 		for (int i = 0; i < epb; i++) {
1377 			zrl_init(&dnh[i].dnh_zrlock);
1378 
1379 			if (skip) {
1380 				skip--;
1381 				continue;
1382 			}
1383 
1384 			if (dn_block[i].dn_type != DMU_OT_NONE) {
1385 				int interior = dn_block[i].dn_extra_slots;
1386 
1387 				dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1388 				dnode_set_slots(dnc, i + 1, interior,
1389 				    DN_SLOT_INTERIOR);
1390 				skip = interior;
1391 			} else {
1392 				dnh[i].dnh_dnode = DN_SLOT_FREE;
1393 				skip = 0;
1394 			}
1395 		}
1396 
1397 		dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1398 		    dnode_buf_evict_async, NULL);
1399 		winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1400 		if (winner != NULL) {
1401 
1402 			for (int i = 0; i < epb; i++)
1403 				zrl_destroy(&dnh[i].dnh_zrlock);
1404 
1405 			kmem_free(dnc, sizeof (dnode_children_t) +
1406 			    epb * sizeof (dnode_handle_t));
1407 			dnc = winner;
1408 		}
1409 	}
1410 
1411 	ASSERT(dnc->dnc_count == epb);
1412 
1413 	if (flag & DNODE_MUST_BE_ALLOCATED) {
1414 		slots = 1;
1415 
1416 		dnode_slots_hold(dnc, idx, slots);
1417 		dnh = &dnc->dnc_children[idx];
1418 
1419 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1420 			dn = dnh->dnh_dnode;
1421 		} else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1422 			DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1423 			dnode_slots_rele(dnc, idx, slots);
1424 			dbuf_rele(db, FTAG);
1425 			return (SET_ERROR(EEXIST));
1426 		} else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1427 			DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1428 			dnode_slots_rele(dnc, idx, slots);
1429 			dbuf_rele(db, FTAG);
1430 			return (SET_ERROR(ENOENT));
1431 		} else {
1432 			dnode_slots_rele(dnc, idx, slots);
1433 			while (!dnode_slots_tryenter(dnc, idx, slots)) {
1434 				DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1435 				cond_resched();
1436 			}
1437 
1438 			/*
1439 			 * Someone else won the race and called dnode_create()
1440 			 * after we checked DN_SLOT_IS_PTR() above but before
1441 			 * we acquired the lock.
1442 			 */
1443 			if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1444 				DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1445 				dn = dnh->dnh_dnode;
1446 			} else {
1447 				dn = dnode_create(os, dn_block + idx, db,
1448 				    object, dnh);
1449 			}
1450 		}
1451 
1452 		mutex_enter(&dn->dn_mtx);
1453 		if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1454 			DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1455 			mutex_exit(&dn->dn_mtx);
1456 			dnode_slots_rele(dnc, idx, slots);
1457 			dbuf_rele(db, FTAG);
1458 			return (SET_ERROR(ENOENT));
1459 		}
1460 
1461 		/* Don't actually hold if dry run, just return 0 */
1462 		if (flag & DNODE_DRY_RUN) {
1463 			mutex_exit(&dn->dn_mtx);
1464 			dnode_slots_rele(dnc, idx, slots);
1465 			dbuf_rele(db, FTAG);
1466 			return (0);
1467 		}
1468 
1469 		DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1470 	} else if (flag & DNODE_MUST_BE_FREE) {
1471 
1472 		if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1473 			DNODE_STAT_BUMP(dnode_hold_free_overflow);
1474 			dbuf_rele(db, FTAG);
1475 			return (SET_ERROR(ENOSPC));
1476 		}
1477 
1478 		dnode_slots_hold(dnc, idx, slots);
1479 
1480 		if (!dnode_check_slots_free(dnc, idx, slots)) {
1481 			DNODE_STAT_BUMP(dnode_hold_free_misses);
1482 			dnode_slots_rele(dnc, idx, slots);
1483 			dbuf_rele(db, FTAG);
1484 			return (SET_ERROR(ENOSPC));
1485 		}
1486 
1487 		dnode_slots_rele(dnc, idx, slots);
1488 		while (!dnode_slots_tryenter(dnc, idx, slots)) {
1489 			DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1490 			cond_resched();
1491 		}
1492 
1493 		if (!dnode_check_slots_free(dnc, idx, slots)) {
1494 			DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1495 			dnode_slots_rele(dnc, idx, slots);
1496 			dbuf_rele(db, FTAG);
1497 			return (SET_ERROR(ENOSPC));
1498 		}
1499 
1500 		/*
1501 		 * Allocated but otherwise free dnodes which would
1502 		 * be in the interior of a multi-slot dnodes need
1503 		 * to be freed.  Single slot dnodes can be safely
1504 		 * re-purposed as a performance optimization.
1505 		 */
1506 		if (slots > 1)
1507 			dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1508 
1509 		dnh = &dnc->dnc_children[idx];
1510 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1511 			dn = dnh->dnh_dnode;
1512 		} else {
1513 			dn = dnode_create(os, dn_block + idx, db,
1514 			    object, dnh);
1515 		}
1516 
1517 		mutex_enter(&dn->dn_mtx);
1518 		if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1519 			DNODE_STAT_BUMP(dnode_hold_free_refcount);
1520 			mutex_exit(&dn->dn_mtx);
1521 			dnode_slots_rele(dnc, idx, slots);
1522 			dbuf_rele(db, FTAG);
1523 			return (SET_ERROR(EEXIST));
1524 		}
1525 
1526 		/* Don't actually hold if dry run, just return 0 */
1527 		if (flag & DNODE_DRY_RUN) {
1528 			mutex_exit(&dn->dn_mtx);
1529 			dnode_slots_rele(dnc, idx, slots);
1530 			dbuf_rele(db, FTAG);
1531 			return (0);
1532 		}
1533 
1534 		dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1535 		DNODE_STAT_BUMP(dnode_hold_free_hits);
1536 	} else {
1537 		dbuf_rele(db, FTAG);
1538 		return (SET_ERROR(EINVAL));
1539 	}
1540 
1541 	ASSERT0(dn->dn_free_txg);
1542 
1543 	if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1544 		dbuf_add_ref(db, dnh);
1545 
1546 	mutex_exit(&dn->dn_mtx);
1547 
1548 	/* Now we can rely on the hold to prevent the dnode from moving. */
1549 	dnode_slots_rele(dnc, idx, slots);
1550 
1551 	DNODE_VERIFY(dn);
1552 	ASSERT3P(dnp, !=, NULL);
1553 	ASSERT3P(dn->dn_dbuf, ==, db);
1554 	ASSERT3U(dn->dn_object, ==, object);
1555 	dbuf_rele(db, FTAG);
1556 
1557 	*dnp = dn;
1558 	return (0);
1559 }
1560 
1561 /*
1562  * Return held dnode if the object is allocated, NULL if not.
1563  */
1564 int
1565 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1566 {
1567 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1568 	    dnp));
1569 }
1570 
1571 /*
1572  * Can only add a reference if there is already at least one
1573  * reference on the dnode.  Returns FALSE if unable to add a
1574  * new reference.
1575  */
1576 boolean_t
1577 dnode_add_ref(dnode_t *dn, void *tag)
1578 {
1579 	mutex_enter(&dn->dn_mtx);
1580 	if (zfs_refcount_is_zero(&dn->dn_holds)) {
1581 		mutex_exit(&dn->dn_mtx);
1582 		return (FALSE);
1583 	}
1584 	VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1585 	mutex_exit(&dn->dn_mtx);
1586 	return (TRUE);
1587 }
1588 
1589 void
1590 dnode_rele(dnode_t *dn, void *tag)
1591 {
1592 	mutex_enter(&dn->dn_mtx);
1593 	dnode_rele_and_unlock(dn, tag, B_FALSE);
1594 }
1595 
1596 void
1597 dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1598 {
1599 	uint64_t refs;
1600 	/* Get while the hold prevents the dnode from moving. */
1601 	dmu_buf_impl_t *db = dn->dn_dbuf;
1602 	dnode_handle_t *dnh = dn->dn_handle;
1603 
1604 	refs = zfs_refcount_remove(&dn->dn_holds, tag);
1605 	if (refs == 0)
1606 		cv_broadcast(&dn->dn_nodnholds);
1607 	mutex_exit(&dn->dn_mtx);
1608 	/* dnode could get destroyed at this point, so don't use it anymore */
1609 
1610 	/*
1611 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1612 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1613 	 * prevent the dnode from moving, since releasing the last hold could
1614 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1615 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1616 	 * other direct or indirect hold on the dnode must first drop the dnode
1617 	 * handle.
1618 	 */
1619 #ifdef ZFS_DEBUG
1620 	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1621 #endif
1622 
1623 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1624 	if (refs == 0 && db != NULL) {
1625 		/*
1626 		 * Another thread could add a hold to the dnode handle in
1627 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1628 		 * hold on the parent dbuf prevents the handle from being
1629 		 * destroyed, the hold on the handle is OK. We can't yet assert
1630 		 * that the handle has zero references, but that will be
1631 		 * asserted anyway when the handle gets destroyed.
1632 		 */
1633 		mutex_enter(&db->db_mtx);
1634 		dbuf_rele_and_unlock(db, dnh, evicting);
1635 	}
1636 }
1637 
1638 /*
1639  * Test whether we can create a dnode at the specified location.
1640  */
1641 int
1642 dnode_try_claim(objset_t *os, uint64_t object, int slots)
1643 {
1644 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_FREE | DNODE_DRY_RUN,
1645 	    slots, NULL, NULL));
1646 }
1647 
1648 /*
1649  * Checks if the dnode contains any uncommitted dirty records.
1650  */
1651 boolean_t
1652 dnode_is_dirty(dnode_t *dn)
1653 {
1654 	mutex_enter(&dn->dn_mtx);
1655 
1656 	for (int i = 0; i < TXG_SIZE; i++) {
1657 		if (multilist_link_active(&dn->dn_dirty_link[i])) {
1658 			mutex_exit(&dn->dn_mtx);
1659 			return (B_TRUE);
1660 		}
1661 	}
1662 
1663 	mutex_exit(&dn->dn_mtx);
1664 
1665 	return (B_FALSE);
1666 }
1667 
1668 void
1669 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1670 {
1671 	objset_t *os = dn->dn_objset;
1672 	uint64_t txg = tx->tx_txg;
1673 
1674 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1675 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1676 		return;
1677 	}
1678 
1679 	DNODE_VERIFY(dn);
1680 
1681 #ifdef ZFS_DEBUG
1682 	mutex_enter(&dn->dn_mtx);
1683 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1684 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1685 	mutex_exit(&dn->dn_mtx);
1686 #endif
1687 
1688 	/*
1689 	 * Determine old uid/gid when necessary
1690 	 */
1691 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1692 
1693 	multilist_t *dirtylist = &os->os_dirty_dnodes[txg & TXG_MASK];
1694 	multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1695 
1696 	/*
1697 	 * If we are already marked dirty, we're done.
1698 	 */
1699 	if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1700 		multilist_sublist_unlock(mls);
1701 		return;
1702 	}
1703 
1704 	ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1705 	    !avl_is_empty(&dn->dn_dbufs));
1706 	ASSERT(dn->dn_datablksz != 0);
1707 	ASSERT0(dn->dn_next_bonuslen[txg & TXG_MASK]);
1708 	ASSERT0(dn->dn_next_blksz[txg & TXG_MASK]);
1709 	ASSERT0(dn->dn_next_bonustype[txg & TXG_MASK]);
1710 
1711 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1712 	    (u_longlong_t)dn->dn_object, (u_longlong_t)txg);
1713 
1714 	multilist_sublist_insert_head(mls, dn);
1715 
1716 	multilist_sublist_unlock(mls);
1717 
1718 	/*
1719 	 * The dnode maintains a hold on its containing dbuf as
1720 	 * long as there are holds on it.  Each instantiated child
1721 	 * dbuf maintains a hold on the dnode.  When the last child
1722 	 * drops its hold, the dnode will drop its hold on the
1723 	 * containing dbuf. We add a "dirty hold" here so that the
1724 	 * dnode will hang around after we finish processing its
1725 	 * children.
1726 	 */
1727 	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1728 
1729 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1730 
1731 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1732 }
1733 
1734 void
1735 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1736 {
1737 	mutex_enter(&dn->dn_mtx);
1738 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1739 		mutex_exit(&dn->dn_mtx);
1740 		return;
1741 	}
1742 	dn->dn_free_txg = tx->tx_txg;
1743 	mutex_exit(&dn->dn_mtx);
1744 
1745 	dnode_setdirty(dn, tx);
1746 }
1747 
1748 /*
1749  * Try to change the block size for the indicated dnode.  This can only
1750  * succeed if there are no blocks allocated or dirty beyond first block
1751  */
1752 int
1753 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1754 {
1755 	dmu_buf_impl_t *db;
1756 	int err;
1757 
1758 	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1759 	if (size == 0)
1760 		size = SPA_MINBLOCKSIZE;
1761 	else
1762 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1763 
1764 	if (ibs == dn->dn_indblkshift)
1765 		ibs = 0;
1766 
1767 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1768 		return (0);
1769 
1770 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1771 
1772 	/* Check for any allocated blocks beyond the first */
1773 	if (dn->dn_maxblkid != 0)
1774 		goto fail;
1775 
1776 	mutex_enter(&dn->dn_dbufs_mtx);
1777 	for (db = avl_first(&dn->dn_dbufs); db != NULL;
1778 	    db = AVL_NEXT(&dn->dn_dbufs, db)) {
1779 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1780 		    db->db_blkid != DMU_SPILL_BLKID) {
1781 			mutex_exit(&dn->dn_dbufs_mtx);
1782 			goto fail;
1783 		}
1784 	}
1785 	mutex_exit(&dn->dn_dbufs_mtx);
1786 
1787 	if (ibs && dn->dn_nlevels != 1)
1788 		goto fail;
1789 
1790 	/* resize the old block */
1791 	err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1792 	if (err == 0) {
1793 		dbuf_new_size(db, size, tx);
1794 	} else if (err != ENOENT) {
1795 		goto fail;
1796 	}
1797 
1798 	dnode_setdblksz(dn, size);
1799 	dnode_setdirty(dn, tx);
1800 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1801 	if (ibs) {
1802 		dn->dn_indblkshift = ibs;
1803 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1804 	}
1805 	/* release after we have fixed the blocksize in the dnode */
1806 	if (db)
1807 		dbuf_rele(db, FTAG);
1808 
1809 	rw_exit(&dn->dn_struct_rwlock);
1810 	return (0);
1811 
1812 fail:
1813 	rw_exit(&dn->dn_struct_rwlock);
1814 	return (SET_ERROR(ENOTSUP));
1815 }
1816 
1817 static void
1818 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1819 {
1820 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1821 	int old_nlevels = dn->dn_nlevels;
1822 	dmu_buf_impl_t *db;
1823 	list_t *list;
1824 	dbuf_dirty_record_t *new, *dr, *dr_next;
1825 
1826 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1827 
1828 	ASSERT3U(new_nlevels, >, dn->dn_nlevels);
1829 	dn->dn_nlevels = new_nlevels;
1830 
1831 	ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1832 	dn->dn_next_nlevels[txgoff] = new_nlevels;
1833 
1834 	/* dirty the left indirects */
1835 	db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1836 	ASSERT(db != NULL);
1837 	new = dbuf_dirty(db, tx);
1838 	dbuf_rele(db, FTAG);
1839 
1840 	/* transfer the dirty records to the new indirect */
1841 	mutex_enter(&dn->dn_mtx);
1842 	mutex_enter(&new->dt.di.dr_mtx);
1843 	list = &dn->dn_dirty_records[txgoff];
1844 	for (dr = list_head(list); dr; dr = dr_next) {
1845 		dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1846 
1847 		IMPLY(dr->dr_dbuf == NULL, old_nlevels == 1);
1848 		if (dr->dr_dbuf == NULL ||
1849 		    (dr->dr_dbuf->db_level == old_nlevels - 1 &&
1850 		    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1851 		    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID)) {
1852 			list_remove(&dn->dn_dirty_records[txgoff], dr);
1853 			list_insert_tail(&new->dt.di.dr_children, dr);
1854 			dr->dr_parent = new;
1855 		}
1856 	}
1857 	mutex_exit(&new->dt.di.dr_mtx);
1858 	mutex_exit(&dn->dn_mtx);
1859 }
1860 
1861 int
1862 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
1863 {
1864 	int ret = 0;
1865 
1866 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1867 
1868 	if (dn->dn_nlevels == nlevels) {
1869 		ret = 0;
1870 		goto out;
1871 	} else if (nlevels < dn->dn_nlevels) {
1872 		ret = SET_ERROR(EINVAL);
1873 		goto out;
1874 	}
1875 
1876 	dnode_set_nlevels_impl(dn, nlevels, tx);
1877 
1878 out:
1879 	rw_exit(&dn->dn_struct_rwlock);
1880 	return (ret);
1881 }
1882 
1883 /* read-holding callers must not rely on the lock being continuously held */
1884 void
1885 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read,
1886     boolean_t force)
1887 {
1888 	int epbs, new_nlevels;
1889 	uint64_t sz;
1890 
1891 	ASSERT(blkid != DMU_BONUS_BLKID);
1892 
1893 	ASSERT(have_read ?
1894 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
1895 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
1896 
1897 	/*
1898 	 * if we have a read-lock, check to see if we need to do any work
1899 	 * before upgrading to a write-lock.
1900 	 */
1901 	if (have_read) {
1902 		if (blkid <= dn->dn_maxblkid)
1903 			return;
1904 
1905 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1906 			rw_exit(&dn->dn_struct_rwlock);
1907 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1908 		}
1909 	}
1910 
1911 	/*
1912 	 * Raw sends (indicated by the force flag) require that we take the
1913 	 * given blkid even if the value is lower than the current value.
1914 	 */
1915 	if (!force && blkid <= dn->dn_maxblkid)
1916 		goto out;
1917 
1918 	/*
1919 	 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff]
1920 	 * to indicate that this field is set. This allows us to set the
1921 	 * maxblkid to 0 on an existing object in dnode_sync().
1922 	 */
1923 	dn->dn_maxblkid = blkid;
1924 	dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] =
1925 	    blkid | DMU_NEXT_MAXBLKID_SET;
1926 
1927 	/*
1928 	 * Compute the number of levels necessary to support the new maxblkid.
1929 	 * Raw sends will ensure nlevels is set correctly for us.
1930 	 */
1931 	new_nlevels = 1;
1932 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1933 	for (sz = dn->dn_nblkptr;
1934 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1935 		new_nlevels++;
1936 
1937 	ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
1938 
1939 	if (!force) {
1940 		if (new_nlevels > dn->dn_nlevels)
1941 			dnode_set_nlevels_impl(dn, new_nlevels, tx);
1942 	} else {
1943 		ASSERT3U(dn->dn_nlevels, >=, new_nlevels);
1944 	}
1945 
1946 out:
1947 	if (have_read)
1948 		rw_downgrade(&dn->dn_struct_rwlock);
1949 }
1950 
1951 static void
1952 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1953 {
1954 	dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1955 	if (db != NULL) {
1956 		dmu_buf_will_dirty(&db->db, tx);
1957 		dbuf_rele(db, FTAG);
1958 	}
1959 }
1960 
1961 /*
1962  * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1963  * and end_blkid.
1964  */
1965 static void
1966 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1967     dmu_tx_t *tx)
1968 {
1969 	dmu_buf_impl_t *db_search;
1970 	dmu_buf_impl_t *db;
1971 	avl_index_t where;
1972 
1973 	db_search = kmem_zalloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1974 
1975 	mutex_enter(&dn->dn_dbufs_mtx);
1976 
1977 	db_search->db_level = 1;
1978 	db_search->db_blkid = start_blkid + 1;
1979 	db_search->db_state = DB_SEARCH;
1980 	for (;;) {
1981 
1982 		db = avl_find(&dn->dn_dbufs, db_search, &where);
1983 		if (db == NULL)
1984 			db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1985 
1986 		if (db == NULL || db->db_level != 1 ||
1987 		    db->db_blkid >= end_blkid) {
1988 			break;
1989 		}
1990 
1991 		/*
1992 		 * Setup the next blkid we want to search for.
1993 		 */
1994 		db_search->db_blkid = db->db_blkid + 1;
1995 		ASSERT3U(db->db_blkid, >=, start_blkid);
1996 
1997 		/*
1998 		 * If the dbuf transitions to DB_EVICTING while we're trying
1999 		 * to dirty it, then we will be unable to discover it in
2000 		 * the dbuf hash table. This will result in a call to
2001 		 * dbuf_create() which needs to acquire the dn_dbufs_mtx
2002 		 * lock. To avoid a deadlock, we drop the lock before
2003 		 * dirtying the level-1 dbuf.
2004 		 */
2005 		mutex_exit(&dn->dn_dbufs_mtx);
2006 		dnode_dirty_l1(dn, db->db_blkid, tx);
2007 		mutex_enter(&dn->dn_dbufs_mtx);
2008 	}
2009 
2010 #ifdef ZFS_DEBUG
2011 	/*
2012 	 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
2013 	 */
2014 	db_search->db_level = 1;
2015 	db_search->db_blkid = start_blkid + 1;
2016 	db_search->db_state = DB_SEARCH;
2017 	db = avl_find(&dn->dn_dbufs, db_search, &where);
2018 	if (db == NULL)
2019 		db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
2020 	for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
2021 		if (db->db_level != 1 || db->db_blkid >= end_blkid)
2022 			break;
2023 		if (db->db_state != DB_EVICTING)
2024 			ASSERT(db->db_dirtycnt > 0);
2025 	}
2026 #endif
2027 	kmem_free(db_search, sizeof (dmu_buf_impl_t));
2028 	mutex_exit(&dn->dn_dbufs_mtx);
2029 }
2030 
2031 void
2032 dnode_set_dirtyctx(dnode_t *dn, dmu_tx_t *tx, void *tag)
2033 {
2034 	/*
2035 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
2036 	 * initialize the objset.
2037 	 */
2038 	if (dn->dn_dirtyctx == DN_UNDIRTIED) {
2039 		dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2040 
2041 		if (ds != NULL) {
2042 			rrw_enter(&ds->ds_bp_rwlock, RW_READER, tag);
2043 		}
2044 		if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
2045 			if (dmu_tx_is_syncing(tx))
2046 				dn->dn_dirtyctx = DN_DIRTY_SYNC;
2047 			else
2048 				dn->dn_dirtyctx = DN_DIRTY_OPEN;
2049 			dn->dn_dirtyctx_firstset = tag;
2050 		}
2051 		if (ds != NULL) {
2052 			rrw_exit(&ds->ds_bp_rwlock, tag);
2053 		}
2054 	}
2055 }
2056 
2057 static void
2058 dnode_partial_zero(dnode_t *dn, uint64_t off, uint64_t blkoff, uint64_t len,
2059     dmu_tx_t *tx)
2060 {
2061 	dmu_buf_impl_t *db;
2062 	int res;
2063 
2064 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2065 	res = dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off), TRUE, FALSE,
2066 	    FTAG, &db);
2067 	rw_exit(&dn->dn_struct_rwlock);
2068 	if (res == 0) {
2069 		db_lock_type_t dblt;
2070 		boolean_t dirty;
2071 
2072 		dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
2073 		/* don't dirty if not on disk and not dirty */
2074 		dirty = !list_is_empty(&db->db_dirty_records) ||
2075 		    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr));
2076 		dmu_buf_unlock_parent(db, dblt, FTAG);
2077 		if (dirty) {
2078 			caddr_t data;
2079 
2080 			dmu_buf_will_dirty(&db->db, tx);
2081 			data = db->db.db_data;
2082 			memset(data + blkoff, 0, len);
2083 		}
2084 		dbuf_rele(db, FTAG);
2085 	}
2086 }
2087 
2088 void
2089 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
2090 {
2091 	uint64_t blkoff, blkid, nblks;
2092 	int blksz, blkshift, head, tail;
2093 	int trunc = FALSE;
2094 	int epbs;
2095 
2096 	blksz = dn->dn_datablksz;
2097 	blkshift = dn->dn_datablkshift;
2098 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2099 
2100 	if (len == DMU_OBJECT_END) {
2101 		len = UINT64_MAX - off;
2102 		trunc = TRUE;
2103 	}
2104 
2105 	/*
2106 	 * First, block align the region to free:
2107 	 */
2108 	if (ISP2(blksz)) {
2109 		head = P2NPHASE(off, blksz);
2110 		blkoff = P2PHASE(off, blksz);
2111 		if ((off >> blkshift) > dn->dn_maxblkid)
2112 			return;
2113 	} else {
2114 		ASSERT(dn->dn_maxblkid == 0);
2115 		if (off == 0 && len >= blksz) {
2116 			/*
2117 			 * Freeing the whole block; fast-track this request.
2118 			 */
2119 			blkid = 0;
2120 			nblks = 1;
2121 			if (dn->dn_nlevels > 1) {
2122 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2123 				dnode_dirty_l1(dn, 0, tx);
2124 				rw_exit(&dn->dn_struct_rwlock);
2125 			}
2126 			goto done;
2127 		} else if (off >= blksz) {
2128 			/* Freeing past end-of-data */
2129 			return;
2130 		} else {
2131 			/* Freeing part of the block. */
2132 			head = blksz - off;
2133 			ASSERT3U(head, >, 0);
2134 		}
2135 		blkoff = off;
2136 	}
2137 	/* zero out any partial block data at the start of the range */
2138 	if (head) {
2139 		ASSERT3U(blkoff + head, ==, blksz);
2140 		if (len < head)
2141 			head = len;
2142 		dnode_partial_zero(dn, off, blkoff, head, tx);
2143 		off += head;
2144 		len -= head;
2145 	}
2146 
2147 	/* If the range was less than one block, we're done */
2148 	if (len == 0)
2149 		return;
2150 
2151 	/* If the remaining range is past end of file, we're done */
2152 	if ((off >> blkshift) > dn->dn_maxblkid)
2153 		return;
2154 
2155 	ASSERT(ISP2(blksz));
2156 	if (trunc)
2157 		tail = 0;
2158 	else
2159 		tail = P2PHASE(len, blksz);
2160 
2161 	ASSERT0(P2PHASE(off, blksz));
2162 	/* zero out any partial block data at the end of the range */
2163 	if (tail) {
2164 		if (len < tail)
2165 			tail = len;
2166 		dnode_partial_zero(dn, off + len, 0, tail, tx);
2167 		len -= tail;
2168 	}
2169 
2170 	/* If the range did not include a full block, we are done */
2171 	if (len == 0)
2172 		return;
2173 
2174 	ASSERT(IS_P2ALIGNED(off, blksz));
2175 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2176 	blkid = off >> blkshift;
2177 	nblks = len >> blkshift;
2178 	if (trunc)
2179 		nblks += 1;
2180 
2181 	/*
2182 	 * Dirty all the indirect blocks in this range.  Note that only
2183 	 * the first and last indirect blocks can actually be written
2184 	 * (if they were partially freed) -- they must be dirtied, even if
2185 	 * they do not exist on disk yet.  The interior blocks will
2186 	 * be freed by free_children(), so they will not actually be written.
2187 	 * Even though these interior blocks will not be written, we
2188 	 * dirty them for two reasons:
2189 	 *
2190 	 *  - It ensures that the indirect blocks remain in memory until
2191 	 *    syncing context.  (They have already been prefetched by
2192 	 *    dmu_tx_hold_free(), so we don't have to worry about reading
2193 	 *    them serially here.)
2194 	 *
2195 	 *  - The dirty space accounting will put pressure on the txg sync
2196 	 *    mechanism to begin syncing, and to delay transactions if there
2197 	 *    is a large amount of freeing.  Even though these indirect
2198 	 *    blocks will not be written, we could need to write the same
2199 	 *    amount of space if we copy the freed BPs into deadlists.
2200 	 */
2201 	if (dn->dn_nlevels > 1) {
2202 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2203 		uint64_t first, last;
2204 
2205 		first = blkid >> epbs;
2206 		dnode_dirty_l1(dn, first, tx);
2207 		if (trunc)
2208 			last = dn->dn_maxblkid >> epbs;
2209 		else
2210 			last = (blkid + nblks - 1) >> epbs;
2211 		if (last != first)
2212 			dnode_dirty_l1(dn, last, tx);
2213 
2214 		dnode_dirty_l1range(dn, first, last, tx);
2215 
2216 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2217 		    SPA_BLKPTRSHIFT;
2218 		for (uint64_t i = first + 1; i < last; i++) {
2219 			/*
2220 			 * Set i to the blockid of the next non-hole
2221 			 * level-1 indirect block at or after i.  Note
2222 			 * that dnode_next_offset() operates in terms of
2223 			 * level-0-equivalent bytes.
2224 			 */
2225 			uint64_t ibyte = i << shift;
2226 			int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2227 			    &ibyte, 2, 1, 0);
2228 			i = ibyte >> shift;
2229 			if (i >= last)
2230 				break;
2231 
2232 			/*
2233 			 * Normally we should not see an error, either
2234 			 * from dnode_next_offset() or dbuf_hold_level()
2235 			 * (except for ESRCH from dnode_next_offset).
2236 			 * If there is an i/o error, then when we read
2237 			 * this block in syncing context, it will use
2238 			 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2239 			 * to the "failmode" property.  dnode_next_offset()
2240 			 * doesn't have a flag to indicate MUSTSUCCEED.
2241 			 */
2242 			if (err != 0)
2243 				break;
2244 
2245 			dnode_dirty_l1(dn, i, tx);
2246 		}
2247 		rw_exit(&dn->dn_struct_rwlock);
2248 	}
2249 
2250 done:
2251 	/*
2252 	 * Add this range to the dnode range list.
2253 	 * We will finish up this free operation in the syncing phase.
2254 	 */
2255 	mutex_enter(&dn->dn_mtx);
2256 	{
2257 		int txgoff = tx->tx_txg & TXG_MASK;
2258 		if (dn->dn_free_ranges[txgoff] == NULL) {
2259 			dn->dn_free_ranges[txgoff] = range_tree_create(NULL,
2260 			    RANGE_SEG64, NULL, 0, 0);
2261 		}
2262 		range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2263 		range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2264 	}
2265 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2266 	    (u_longlong_t)blkid, (u_longlong_t)nblks,
2267 	    (u_longlong_t)tx->tx_txg);
2268 	mutex_exit(&dn->dn_mtx);
2269 
2270 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2271 	dnode_setdirty(dn, tx);
2272 }
2273 
2274 static boolean_t
2275 dnode_spill_freed(dnode_t *dn)
2276 {
2277 	int i;
2278 
2279 	mutex_enter(&dn->dn_mtx);
2280 	for (i = 0; i < TXG_SIZE; i++) {
2281 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2282 			break;
2283 	}
2284 	mutex_exit(&dn->dn_mtx);
2285 	return (i < TXG_SIZE);
2286 }
2287 
2288 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2289 uint64_t
2290 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2291 {
2292 	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
2293 	int i;
2294 
2295 	if (blkid == DMU_BONUS_BLKID)
2296 		return (FALSE);
2297 
2298 	/*
2299 	 * If we're in the process of opening the pool, dp will not be
2300 	 * set yet, but there shouldn't be anything dirty.
2301 	 */
2302 	if (dp == NULL)
2303 		return (FALSE);
2304 
2305 	if (dn->dn_free_txg)
2306 		return (TRUE);
2307 
2308 	if (blkid == DMU_SPILL_BLKID)
2309 		return (dnode_spill_freed(dn));
2310 
2311 	mutex_enter(&dn->dn_mtx);
2312 	for (i = 0; i < TXG_SIZE; i++) {
2313 		if (dn->dn_free_ranges[i] != NULL &&
2314 		    range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2315 			break;
2316 	}
2317 	mutex_exit(&dn->dn_mtx);
2318 	return (i < TXG_SIZE);
2319 }
2320 
2321 /* call from syncing context when we actually write/free space for this dnode */
2322 void
2323 dnode_diduse_space(dnode_t *dn, int64_t delta)
2324 {
2325 	uint64_t space;
2326 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2327 	    dn, dn->dn_phys,
2328 	    (u_longlong_t)dn->dn_phys->dn_used,
2329 	    (longlong_t)delta);
2330 
2331 	mutex_enter(&dn->dn_mtx);
2332 	space = DN_USED_BYTES(dn->dn_phys);
2333 	if (delta > 0) {
2334 		ASSERT3U(space + delta, >=, space); /* no overflow */
2335 	} else {
2336 		ASSERT3U(space, >=, -delta); /* no underflow */
2337 	}
2338 	space += delta;
2339 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2340 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2341 		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2342 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2343 	} else {
2344 		dn->dn_phys->dn_used = space;
2345 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2346 	}
2347 	mutex_exit(&dn->dn_mtx);
2348 }
2349 
2350 /*
2351  * Scans a block at the indicated "level" looking for a hole or data,
2352  * depending on 'flags'.
2353  *
2354  * If level > 0, then we are scanning an indirect block looking at its
2355  * pointers.  If level == 0, then we are looking at a block of dnodes.
2356  *
2357  * If we don't find what we are looking for in the block, we return ESRCH.
2358  * Otherwise, return with *offset pointing to the beginning (if searching
2359  * forwards) or end (if searching backwards) of the range covered by the
2360  * block pointer we matched on (or dnode).
2361  *
2362  * The basic search algorithm used below by dnode_next_offset() is to
2363  * use this function to search up the block tree (widen the search) until
2364  * we find something (i.e., we don't return ESRCH) and then search back
2365  * down the tree (narrow the search) until we reach our original search
2366  * level.
2367  */
2368 static int
2369 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2370     int lvl, uint64_t blkfill, uint64_t txg)
2371 {
2372 	dmu_buf_impl_t *db = NULL;
2373 	void *data = NULL;
2374 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2375 	uint64_t epb = 1ULL << epbs;
2376 	uint64_t minfill, maxfill;
2377 	boolean_t hole;
2378 	int i, inc, error, span;
2379 
2380 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2381 
2382 	hole = ((flags & DNODE_FIND_HOLE) != 0);
2383 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2384 	ASSERT(txg == 0 || !hole);
2385 
2386 	if (lvl == dn->dn_phys->dn_nlevels) {
2387 		error = 0;
2388 		epb = dn->dn_phys->dn_nblkptr;
2389 		data = dn->dn_phys->dn_blkptr;
2390 	} else {
2391 		uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2392 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2393 		if (error) {
2394 			if (error != ENOENT)
2395 				return (error);
2396 			if (hole)
2397 				return (0);
2398 			/*
2399 			 * This can only happen when we are searching up
2400 			 * the block tree for data.  We don't really need to
2401 			 * adjust the offset, as we will just end up looking
2402 			 * at the pointer to this block in its parent, and its
2403 			 * going to be unallocated, so we will skip over it.
2404 			 */
2405 			return (SET_ERROR(ESRCH));
2406 		}
2407 		error = dbuf_read(db, NULL,
2408 		    DB_RF_CANFAIL | DB_RF_HAVESTRUCT |
2409 		    DB_RF_NO_DECRYPT | DB_RF_NOPREFETCH);
2410 		if (error) {
2411 			dbuf_rele(db, FTAG);
2412 			return (error);
2413 		}
2414 		data = db->db.db_data;
2415 		rw_enter(&db->db_rwlock, RW_READER);
2416 	}
2417 
2418 	if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2419 	    db->db_blkptr->blk_birth <= txg ||
2420 	    BP_IS_HOLE(db->db_blkptr))) {
2421 		/*
2422 		 * This can only happen when we are searching up the tree
2423 		 * and these conditions mean that we need to keep climbing.
2424 		 */
2425 		error = SET_ERROR(ESRCH);
2426 	} else if (lvl == 0) {
2427 		dnode_phys_t *dnp = data;
2428 
2429 		ASSERT(dn->dn_type == DMU_OT_DNODE);
2430 		ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2431 
2432 		for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2433 		    i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2434 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2435 				break;
2436 		}
2437 
2438 		if (i == blkfill)
2439 			error = SET_ERROR(ESRCH);
2440 
2441 		*offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2442 		    (i << DNODE_SHIFT);
2443 	} else {
2444 		blkptr_t *bp = data;
2445 		uint64_t start = *offset;
2446 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
2447 		minfill = 0;
2448 		maxfill = blkfill << ((lvl - 1) * epbs);
2449 
2450 		if (hole)
2451 			maxfill--;
2452 		else
2453 			minfill++;
2454 
2455 		if (span >= 8 * sizeof (*offset)) {
2456 			/* This only happens on the highest indirection level */
2457 			ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1);
2458 			*offset = 0;
2459 		} else {
2460 			*offset = *offset >> span;
2461 		}
2462 
2463 		for (i = BF64_GET(*offset, 0, epbs);
2464 		    i >= 0 && i < epb; i += inc) {
2465 			if (BP_GET_FILL(&bp[i]) >= minfill &&
2466 			    BP_GET_FILL(&bp[i]) <= maxfill &&
2467 			    (hole || bp[i].blk_birth > txg))
2468 				break;
2469 			if (inc > 0 || *offset > 0)
2470 				*offset += inc;
2471 		}
2472 
2473 		if (span >= 8 * sizeof (*offset)) {
2474 			*offset = start;
2475 		} else {
2476 			*offset = *offset << span;
2477 		}
2478 
2479 		if (inc < 0) {
2480 			/* traversing backwards; position offset at the end */
2481 			ASSERT3U(*offset, <=, start);
2482 			*offset = MIN(*offset + (1ULL << span) - 1, start);
2483 		} else if (*offset < start) {
2484 			*offset = start;
2485 		}
2486 		if (i < 0 || i >= epb)
2487 			error = SET_ERROR(ESRCH);
2488 	}
2489 
2490 	if (db != NULL) {
2491 		rw_exit(&db->db_rwlock);
2492 		dbuf_rele(db, FTAG);
2493 	}
2494 
2495 	return (error);
2496 }
2497 
2498 /*
2499  * Find the next hole, data, or sparse region at or after *offset.
2500  * The value 'blkfill' tells us how many items we expect to find
2501  * in an L0 data block; this value is 1 for normal objects,
2502  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2503  * DNODES_PER_BLOCK when searching for sparse regions thereof.
2504  *
2505  * Examples:
2506  *
2507  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2508  *	Finds the next/previous hole/data in a file.
2509  *	Used in dmu_offset_next().
2510  *
2511  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2512  *	Finds the next free/allocated dnode an objset's meta-dnode.
2513  *	Only finds objects that have new contents since txg (ie.
2514  *	bonus buffer changes and content removal are ignored).
2515  *	Used in dmu_object_next().
2516  *
2517  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2518  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
2519  *	Used in dmu_object_alloc().
2520  */
2521 int
2522 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2523     int minlvl, uint64_t blkfill, uint64_t txg)
2524 {
2525 	uint64_t initial_offset = *offset;
2526 	int lvl, maxlvl;
2527 	int error = 0;
2528 
2529 	if (!(flags & DNODE_FIND_HAVELOCK))
2530 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
2531 
2532 	if (dn->dn_phys->dn_nlevels == 0) {
2533 		error = SET_ERROR(ESRCH);
2534 		goto out;
2535 	}
2536 
2537 	if (dn->dn_datablkshift == 0) {
2538 		if (*offset < dn->dn_datablksz) {
2539 			if (flags & DNODE_FIND_HOLE)
2540 				*offset = dn->dn_datablksz;
2541 		} else {
2542 			error = SET_ERROR(ESRCH);
2543 		}
2544 		goto out;
2545 	}
2546 
2547 	maxlvl = dn->dn_phys->dn_nlevels;
2548 
2549 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2550 		error = dnode_next_offset_level(dn,
2551 		    flags, offset, lvl, blkfill, txg);
2552 		if (error != ESRCH)
2553 			break;
2554 	}
2555 
2556 	while (error == 0 && --lvl >= minlvl) {
2557 		error = dnode_next_offset_level(dn,
2558 		    flags, offset, lvl, blkfill, txg);
2559 	}
2560 
2561 	/*
2562 	 * There's always a "virtual hole" at the end of the object, even
2563 	 * if all BP's which physically exist are non-holes.
2564 	 */
2565 	if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2566 	    minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2567 		error = 0;
2568 	}
2569 
2570 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2571 	    initial_offset < *offset : initial_offset > *offset))
2572 		error = SET_ERROR(ESRCH);
2573 out:
2574 	if (!(flags & DNODE_FIND_HAVELOCK))
2575 		rw_exit(&dn->dn_struct_rwlock);
2576 
2577 	return (error);
2578 }
2579 
2580 #if defined(_KERNEL)
2581 EXPORT_SYMBOL(dnode_hold);
2582 EXPORT_SYMBOL(dnode_rele);
2583 EXPORT_SYMBOL(dnode_set_nlevels);
2584 EXPORT_SYMBOL(dnode_set_blksz);
2585 EXPORT_SYMBOL(dnode_free_range);
2586 EXPORT_SYMBOL(dnode_evict_dbufs);
2587 EXPORT_SYMBOL(dnode_evict_bonus);
2588 #endif
2589