xref: /titanic_41/usr/src/uts/common/fs/zfs/dnode.c (revision 672986541be54a7a471bb088e60780c37e371d7e)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/zfs_context.h>
29 #include <sys/dbuf.h>
30 #include <sys/dnode.h>
31 #include <sys/dmu.h>
32 #include <sys/dmu_impl.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/spa.h>
38 #include <sys/zio.h>
39 #include <sys/dmu_zfetch.h>
40 
41 static int free_range_compar(const void *node1, const void *node2);
42 
43 static kmem_cache_t *dnode_cache;
44 
45 static dnode_phys_t dnode_phys_zero;
46 
47 int zfs_default_bs = SPA_MINBLOCKSHIFT;
48 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
49 
50 /* ARGSUSED */
51 static int
52 dnode_cons(void *arg, void *unused, int kmflag)
53 {
54 	int i;
55 	dnode_t *dn = arg;
56 	bzero(dn, sizeof (dnode_t));
57 
58 	rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
59 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
60 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
61 	refcount_create(&dn->dn_holds);
62 	refcount_create(&dn->dn_tx_holds);
63 
64 	for (i = 0; i < TXG_SIZE; i++) {
65 		avl_create(&dn->dn_ranges[i], free_range_compar,
66 		    sizeof (free_range_t),
67 		    offsetof(struct free_range, fr_node));
68 		list_create(&dn->dn_dirty_records[i],
69 		    sizeof (dbuf_dirty_record_t),
70 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
71 	}
72 
73 	list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
74 	    offsetof(dmu_buf_impl_t, db_link));
75 
76 	return (0);
77 }
78 
79 /* ARGSUSED */
80 static void
81 dnode_dest(void *arg, void *unused)
82 {
83 	int i;
84 	dnode_t *dn = arg;
85 
86 	rw_destroy(&dn->dn_struct_rwlock);
87 	mutex_destroy(&dn->dn_mtx);
88 	mutex_destroy(&dn->dn_dbufs_mtx);
89 	refcount_destroy(&dn->dn_holds);
90 	refcount_destroy(&dn->dn_tx_holds);
91 
92 	for (i = 0; i < TXG_SIZE; i++) {
93 		avl_destroy(&dn->dn_ranges[i]);
94 		list_destroy(&dn->dn_dirty_records[i]);
95 	}
96 
97 	list_destroy(&dn->dn_dbufs);
98 }
99 
100 void
101 dnode_init(void)
102 {
103 	dnode_cache = kmem_cache_create("dnode_t",
104 	    sizeof (dnode_t),
105 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
106 }
107 
108 void
109 dnode_fini(void)
110 {
111 	kmem_cache_destroy(dnode_cache);
112 }
113 
114 
115 #ifdef ZFS_DEBUG
116 void
117 dnode_verify(dnode_t *dn)
118 {
119 	int drop_struct_lock = FALSE;
120 
121 	ASSERT(dn->dn_phys);
122 	ASSERT(dn->dn_objset);
123 
124 	ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES);
125 
126 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
127 		return;
128 
129 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
130 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
131 		drop_struct_lock = TRUE;
132 	}
133 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
134 		int i;
135 		ASSERT3U(dn->dn_indblkshift, >=, 0);
136 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
137 		if (dn->dn_datablkshift) {
138 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
139 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
140 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
141 		}
142 		ASSERT3U(dn->dn_nlevels, <=, 30);
143 		ASSERT3U(dn->dn_type, <=, DMU_OT_NUMTYPES);
144 		ASSERT3U(dn->dn_nblkptr, >=, 1);
145 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
146 		ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
147 		ASSERT3U(dn->dn_datablksz, ==,
148 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
149 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
150 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
151 		    dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
152 		for (i = 0; i < TXG_SIZE; i++) {
153 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
154 		}
155 	}
156 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
157 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
158 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || dn->dn_dbuf != NULL);
159 	if (dn->dn_dbuf != NULL) {
160 		ASSERT3P(dn->dn_phys, ==,
161 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
162 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
163 	}
164 	if (drop_struct_lock)
165 		rw_exit(&dn->dn_struct_rwlock);
166 }
167 #endif
168 
169 void
170 dnode_byteswap(dnode_phys_t *dnp)
171 {
172 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
173 	int i;
174 
175 	if (dnp->dn_type == DMU_OT_NONE) {
176 		bzero(dnp, sizeof (dnode_phys_t));
177 		return;
178 	}
179 
180 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
181 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
182 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
183 	dnp->dn_used = BSWAP_64(dnp->dn_used);
184 
185 	/*
186 	 * dn_nblkptr is only one byte, so it's OK to read it in either
187 	 * byte order.  We can't read dn_bouslen.
188 	 */
189 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
190 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
191 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
192 		buf64[i] = BSWAP_64(buf64[i]);
193 
194 	/*
195 	 * OK to check dn_bonuslen for zero, because it won't matter if
196 	 * we have the wrong byte order.  This is necessary because the
197 	 * dnode dnode is smaller than a regular dnode.
198 	 */
199 	if (dnp->dn_bonuslen != 0) {
200 		/*
201 		 * Note that the bonus length calculated here may be
202 		 * longer than the actual bonus buffer.  This is because
203 		 * we always put the bonus buffer after the last block
204 		 * pointer (instead of packing it against the end of the
205 		 * dnode buffer).
206 		 */
207 		int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
208 		size_t len = DN_MAX_BONUSLEN - off;
209 		ASSERT3U(dnp->dn_bonustype, <, DMU_OT_NUMTYPES);
210 		dmu_ot[dnp->dn_bonustype].ot_byteswap(dnp->dn_bonus + off, len);
211 	}
212 }
213 
214 void
215 dnode_buf_byteswap(void *vbuf, size_t size)
216 {
217 	dnode_phys_t *buf = vbuf;
218 	int i;
219 
220 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
221 	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
222 
223 	size >>= DNODE_SHIFT;
224 	for (i = 0; i < size; i++) {
225 		dnode_byteswap(buf);
226 		buf++;
227 	}
228 }
229 
230 static int
231 free_range_compar(const void *node1, const void *node2)
232 {
233 	const free_range_t *rp1 = node1;
234 	const free_range_t *rp2 = node2;
235 
236 	if (rp1->fr_blkid < rp2->fr_blkid)
237 		return (-1);
238 	else if (rp1->fr_blkid > rp2->fr_blkid)
239 		return (1);
240 	else return (0);
241 }
242 
243 static void
244 dnode_setdblksz(dnode_t *dn, int size)
245 {
246 	ASSERT3U(P2PHASE(size, SPA_MINBLOCKSIZE), ==, 0);
247 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
248 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
249 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
250 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
251 	dn->dn_datablksz = size;
252 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
253 	dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0;
254 }
255 
256 static dnode_t *
257 dnode_create(objset_impl_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
258     uint64_t object)
259 {
260 	dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
261 	(void) dnode_cons(dn, NULL, 0); /* XXX */
262 
263 	dn->dn_objset = os;
264 	dn->dn_object = object;
265 	dn->dn_dbuf = db;
266 	dn->dn_phys = dnp;
267 
268 	if (dnp->dn_datablkszsec)
269 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
270 	dn->dn_indblkshift = dnp->dn_indblkshift;
271 	dn->dn_nlevels = dnp->dn_nlevels;
272 	dn->dn_type = dnp->dn_type;
273 	dn->dn_nblkptr = dnp->dn_nblkptr;
274 	dn->dn_checksum = dnp->dn_checksum;
275 	dn->dn_compress = dnp->dn_compress;
276 	dn->dn_bonustype = dnp->dn_bonustype;
277 	dn->dn_bonuslen = dnp->dn_bonuslen;
278 	dn->dn_maxblkid = dnp->dn_maxblkid;
279 
280 	dmu_zfetch_init(&dn->dn_zfetch, dn);
281 
282 	ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES);
283 	mutex_enter(&os->os_lock);
284 	list_insert_head(&os->os_dnodes, dn);
285 	mutex_exit(&os->os_lock);
286 
287 	arc_space_consume(sizeof (dnode_t));
288 	return (dn);
289 }
290 
291 static void
292 dnode_destroy(dnode_t *dn)
293 {
294 	objset_impl_t *os = dn->dn_objset;
295 
296 #ifdef ZFS_DEBUG
297 	int i;
298 
299 	for (i = 0; i < TXG_SIZE; i++) {
300 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
301 		ASSERT(NULL == list_head(&dn->dn_dirty_records[i]));
302 		ASSERT(0 == avl_numnodes(&dn->dn_ranges[i]));
303 	}
304 	ASSERT(NULL == list_head(&dn->dn_dbufs));
305 #endif
306 
307 	mutex_enter(&os->os_lock);
308 	list_remove(&os->os_dnodes, dn);
309 	mutex_exit(&os->os_lock);
310 
311 	if (dn->dn_dirtyctx_firstset) {
312 		kmem_free(dn->dn_dirtyctx_firstset, 1);
313 		dn->dn_dirtyctx_firstset = NULL;
314 	}
315 	dmu_zfetch_rele(&dn->dn_zfetch);
316 	if (dn->dn_bonus) {
317 		mutex_enter(&dn->dn_bonus->db_mtx);
318 		dbuf_evict(dn->dn_bonus);
319 		dn->dn_bonus = NULL;
320 	}
321 	kmem_cache_free(dnode_cache, dn);
322 	arc_space_return(sizeof (dnode_t));
323 }
324 
325 void
326 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
327     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
328 {
329 	int i;
330 
331 	if (blocksize == 0)
332 		blocksize = 1 << zfs_default_bs;
333 	else if (blocksize > SPA_MAXBLOCKSIZE)
334 		blocksize = SPA_MAXBLOCKSIZE;
335 	else
336 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
337 
338 	if (ibs == 0)
339 		ibs = zfs_default_ibs;
340 
341 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
342 
343 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
344 	    dn->dn_object, tx->tx_txg, blocksize, ibs);
345 
346 	ASSERT(dn->dn_type == DMU_OT_NONE);
347 	ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
348 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
349 	ASSERT(ot != DMU_OT_NONE);
350 	ASSERT3U(ot, <, DMU_OT_NUMTYPES);
351 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
352 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
353 	ASSERT3U(bonustype, <, DMU_OT_NUMTYPES);
354 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
355 	ASSERT(dn->dn_type == DMU_OT_NONE);
356 	ASSERT3U(dn->dn_maxblkid, ==, 0);
357 	ASSERT3U(dn->dn_allocated_txg, ==, 0);
358 	ASSERT3U(dn->dn_assigned_txg, ==, 0);
359 	ASSERT(refcount_is_zero(&dn->dn_tx_holds));
360 	ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
361 	ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
362 
363 	for (i = 0; i < TXG_SIZE; i++) {
364 		ASSERT3U(dn->dn_next_nlevels[i], ==, 0);
365 		ASSERT3U(dn->dn_next_indblkshift[i], ==, 0);
366 		ASSERT3U(dn->dn_next_blksz[i], ==, 0);
367 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
368 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
369 		ASSERT3U(avl_numnodes(&dn->dn_ranges[i]), ==, 0);
370 	}
371 
372 	dn->dn_type = ot;
373 	dnode_setdblksz(dn, blocksize);
374 	dn->dn_indblkshift = ibs;
375 	dn->dn_nlevels = 1;
376 	dn->dn_nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
377 	dn->dn_bonustype = bonustype;
378 	dn->dn_bonuslen = bonuslen;
379 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
380 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
381 	dn->dn_dirtyctx = 0;
382 
383 	dn->dn_free_txg = 0;
384 	if (dn->dn_dirtyctx_firstset) {
385 		kmem_free(dn->dn_dirtyctx_firstset, 1);
386 		dn->dn_dirtyctx_firstset = NULL;
387 	}
388 
389 	dn->dn_allocated_txg = tx->tx_txg;
390 
391 	dnode_setdirty(dn, tx);
392 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
393 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
394 }
395 
396 void
397 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
398     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
399 {
400 	int i;
401 	dmu_buf_impl_t *db = NULL;
402 
403 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
404 	ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
405 	ASSERT3U(blocksize % SPA_MINBLOCKSIZE, ==, 0);
406 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
407 	ASSERT(tx->tx_txg != 0);
408 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
409 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
410 	ASSERT3U(bonustype, <, DMU_OT_NUMTYPES);
411 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
412 
413 	for (i = 0; i < TXG_SIZE; i++)
414 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
415 
416 	/* clean up any unreferenced dbufs */
417 	(void) dnode_evict_dbufs(dn, 0);
418 	ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
419 
420 	/*
421 	 * XXX I should really have a generation number to tell if we
422 	 * need to do this...
423 	 */
424 	if (blocksize != dn->dn_datablksz ||
425 	    dn->dn_bonustype != bonustype || dn->dn_bonuslen != bonuslen) {
426 		/* free all old data */
427 		dnode_free_range(dn, 0, -1ULL, tx);
428 	}
429 
430 	/* change blocksize */
431 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
432 	if (blocksize != dn->dn_datablksz &&
433 	    (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
434 	    list_head(&dn->dn_dbufs) != NULL)) {
435 		db = dbuf_hold(dn, 0, FTAG);
436 		dbuf_new_size(db, blocksize, tx);
437 	}
438 	dnode_setdblksz(dn, blocksize);
439 	dnode_setdirty(dn, tx);
440 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
441 	rw_exit(&dn->dn_struct_rwlock);
442 	if (db) {
443 		dbuf_rele(db, FTAG);
444 		db = NULL;
445 	}
446 
447 	/* change type */
448 	dn->dn_type = ot;
449 
450 	if (dn->dn_bonuslen != bonuslen) {
451 		/* change bonus size */
452 		if (bonuslen == 0)
453 			bonuslen = 1; /* XXX */
454 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
455 		if (dn->dn_bonus == NULL)
456 			dn->dn_bonus = dbuf_create_bonus(dn);
457 		db = dn->dn_bonus;
458 		rw_exit(&dn->dn_struct_rwlock);
459 		if (refcount_add(&db->db_holds, FTAG) == 1)
460 			dnode_add_ref(dn, db);
461 		VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED));
462 		mutex_enter(&db->db_mtx);
463 		ASSERT3U(db->db.db_size, ==, dn->dn_bonuslen);
464 		ASSERT(db->db.db_data != NULL);
465 		db->db.db_size = bonuslen;
466 		mutex_exit(&db->db_mtx);
467 		(void) dbuf_dirty(db, tx);
468 	}
469 
470 	/* change bonus size and type */
471 	mutex_enter(&dn->dn_mtx);
472 	dn->dn_bonustype = bonustype;
473 	dn->dn_bonuslen = bonuslen;
474 	dn->dn_nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
475 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
476 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
477 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
478 
479 	/*
480 	 * NB: we have to do the dbuf_rele after we've changed the
481 	 * dn_bonuslen, for the sake of dbuf_verify().
482 	 */
483 	if (db)
484 		dbuf_rele(db, FTAG);
485 
486 	dn->dn_allocated_txg = tx->tx_txg;
487 	mutex_exit(&dn->dn_mtx);
488 }
489 
490 void
491 dnode_special_close(dnode_t *dn)
492 {
493 	/*
494 	 * Wait for final references to the dnode to clear.  This can
495 	 * only happen if the arc is asyncronously evicting state that
496 	 * has a hold on this dnode while we are trying to evict this
497 	 * dnode.
498 	 */
499 	while (refcount_count(&dn->dn_holds) > 0)
500 		delay(1);
501 	dnode_destroy(dn);
502 }
503 
504 dnode_t *
505 dnode_special_open(objset_impl_t *os, dnode_phys_t *dnp, uint64_t object)
506 {
507 	dnode_t *dn = dnode_create(os, dnp, NULL, object);
508 	DNODE_VERIFY(dn);
509 	return (dn);
510 }
511 
512 static void
513 dnode_buf_pageout(dmu_buf_t *db, void *arg)
514 {
515 	dnode_t **children_dnodes = arg;
516 	int i;
517 	int epb = db->db_size >> DNODE_SHIFT;
518 
519 	for (i = 0; i < epb; i++) {
520 		dnode_t *dn = children_dnodes[i];
521 		int n;
522 
523 		if (dn == NULL)
524 			continue;
525 #ifdef ZFS_DEBUG
526 		/*
527 		 * If there are holds on this dnode, then there should
528 		 * be holds on the dnode's containing dbuf as well; thus
529 		 * it wouldn't be eligable for eviction and this function
530 		 * would not have been called.
531 		 */
532 		ASSERT(refcount_is_zero(&dn->dn_holds));
533 		ASSERT(list_head(&dn->dn_dbufs) == NULL);
534 		ASSERT(refcount_is_zero(&dn->dn_tx_holds));
535 
536 		for (n = 0; n < TXG_SIZE; n++)
537 			ASSERT(!list_link_active(&dn->dn_dirty_link[n]));
538 #endif
539 		children_dnodes[i] = NULL;
540 		dnode_destroy(dn);
541 	}
542 	kmem_free(children_dnodes, epb * sizeof (dnode_t *));
543 }
544 
545 /*
546  * errors:
547  * EINVAL - invalid object number.
548  * EIO - i/o error.
549  * succeeds even for free dnodes.
550  */
551 int
552 dnode_hold_impl(objset_impl_t *os, uint64_t object, int flag,
553     void *tag, dnode_t **dnp)
554 {
555 	int epb, idx, err;
556 	int drop_struct_lock = FALSE;
557 	int type;
558 	uint64_t blk;
559 	dnode_t *mdn, *dn;
560 	dmu_buf_impl_t *db;
561 	dnode_t **children_dnodes;
562 
563 	if (object == 0 || object >= DN_MAX_OBJECT)
564 		return (EINVAL);
565 
566 	mdn = os->os_meta_dnode;
567 
568 	DNODE_VERIFY(mdn);
569 
570 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
571 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
572 		drop_struct_lock = TRUE;
573 	}
574 
575 	blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
576 
577 	db = dbuf_hold(mdn, blk, FTAG);
578 	if (drop_struct_lock)
579 		rw_exit(&mdn->dn_struct_rwlock);
580 	if (db == NULL)
581 		return (EIO);
582 	err = dbuf_read(db, NULL, DB_RF_CANFAIL);
583 	if (err) {
584 		dbuf_rele(db, FTAG);
585 		return (err);
586 	}
587 
588 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
589 	epb = db->db.db_size >> DNODE_SHIFT;
590 
591 	idx = object & (epb-1);
592 
593 	children_dnodes = dmu_buf_get_user(&db->db);
594 	if (children_dnodes == NULL) {
595 		dnode_t **winner;
596 		children_dnodes = kmem_zalloc(epb * sizeof (dnode_t *),
597 		    KM_SLEEP);
598 		if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
599 		    dnode_buf_pageout)) {
600 			kmem_free(children_dnodes, epb * sizeof (dnode_t *));
601 			children_dnodes = winner;
602 		}
603 	}
604 
605 	if ((dn = children_dnodes[idx]) == NULL) {
606 		dnode_phys_t *dnp = (dnode_phys_t *)db->db.db_data+idx;
607 		dnode_t *winner;
608 
609 		dn = dnode_create(os, dnp, db, object);
610 		winner = atomic_cas_ptr(&children_dnodes[idx], NULL, dn);
611 		if (winner != NULL) {
612 			dnode_destroy(dn);
613 			dn = winner;
614 		}
615 	}
616 
617 	mutex_enter(&dn->dn_mtx);
618 	type = dn->dn_type;
619 	if (dn->dn_free_txg ||
620 	    ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
621 	    ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)) {
622 		mutex_exit(&dn->dn_mtx);
623 		dbuf_rele(db, FTAG);
624 		return (type == DMU_OT_NONE ? ENOENT : EEXIST);
625 	}
626 	mutex_exit(&dn->dn_mtx);
627 
628 	if (refcount_add(&dn->dn_holds, tag) == 1)
629 		dbuf_add_ref(db, dn);
630 
631 	DNODE_VERIFY(dn);
632 	ASSERT3P(dn->dn_dbuf, ==, db);
633 	ASSERT3U(dn->dn_object, ==, object);
634 	dbuf_rele(db, FTAG);
635 
636 	*dnp = dn;
637 	return (0);
638 }
639 
640 /*
641  * Return held dnode if the object is allocated, NULL if not.
642  */
643 int
644 dnode_hold(objset_impl_t *os, uint64_t object, void *tag, dnode_t **dnp)
645 {
646 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
647 }
648 
649 void
650 dnode_add_ref(dnode_t *dn, void *tag)
651 {
652 	ASSERT(refcount_count(&dn->dn_holds) > 0);
653 	(void) refcount_add(&dn->dn_holds, tag);
654 }
655 
656 void
657 dnode_rele(dnode_t *dn, void *tag)
658 {
659 	uint64_t refs;
660 
661 	refs = refcount_remove(&dn->dn_holds, tag);
662 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
663 	if (refs == 0 && dn->dn_dbuf)
664 		dbuf_rele(dn->dn_dbuf, dn);
665 }
666 
667 void
668 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
669 {
670 	objset_impl_t *os = dn->dn_objset;
671 	uint64_t txg = tx->tx_txg;
672 
673 	if (dn->dn_object == DMU_META_DNODE_OBJECT)
674 		return;
675 
676 	DNODE_VERIFY(dn);
677 
678 #ifdef ZFS_DEBUG
679 	mutex_enter(&dn->dn_mtx);
680 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
681 	/* ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg); */
682 	mutex_exit(&dn->dn_mtx);
683 #endif
684 
685 	mutex_enter(&os->os_lock);
686 
687 	/*
688 	 * If we are already marked dirty, we're done.
689 	 */
690 	if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
691 		mutex_exit(&os->os_lock);
692 		return;
693 	}
694 
695 	ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
696 	ASSERT(dn->dn_datablksz != 0);
697 	ASSERT3U(dn->dn_next_blksz[txg&TXG_MASK], ==, 0);
698 
699 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
700 	    dn->dn_object, txg);
701 
702 	if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
703 		list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
704 	} else {
705 		list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
706 	}
707 
708 	mutex_exit(&os->os_lock);
709 
710 	/*
711 	 * The dnode maintains a hold on its containing dbuf as
712 	 * long as there are holds on it.  Each instantiated child
713 	 * dbuf maintaines a hold on the dnode.  When the last child
714 	 * drops its hold, the dnode will drop its hold on the
715 	 * containing dbuf. We add a "dirty hold" here so that the
716 	 * dnode will hang around after we finish processing its
717 	 * children.
718 	 */
719 	dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg);
720 
721 	(void) dbuf_dirty(dn->dn_dbuf, tx);
722 
723 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
724 }
725 
726 void
727 dnode_free(dnode_t *dn, dmu_tx_t *tx)
728 {
729 	int txgoff = tx->tx_txg & TXG_MASK;
730 
731 	dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
732 
733 	/* we should be the only holder... hopefully */
734 	/* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
735 
736 	mutex_enter(&dn->dn_mtx);
737 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
738 		mutex_exit(&dn->dn_mtx);
739 		return;
740 	}
741 	dn->dn_free_txg = tx->tx_txg;
742 	mutex_exit(&dn->dn_mtx);
743 
744 	/*
745 	 * If the dnode is already dirty, it needs to be moved from
746 	 * the dirty list to the free list.
747 	 */
748 	mutex_enter(&dn->dn_objset->os_lock);
749 	if (list_link_active(&dn->dn_dirty_link[txgoff])) {
750 		list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
751 		list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
752 		mutex_exit(&dn->dn_objset->os_lock);
753 	} else {
754 		mutex_exit(&dn->dn_objset->os_lock);
755 		dnode_setdirty(dn, tx);
756 	}
757 }
758 
759 /*
760  * Try to change the block size for the indicated dnode.  This can only
761  * succeed if there are no blocks allocated or dirty beyond first block
762  */
763 int
764 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
765 {
766 	dmu_buf_impl_t *db, *db_next;
767 	int have_db0 = FALSE;
768 
769 	if (size == 0)
770 		size = SPA_MINBLOCKSIZE;
771 	if (size > SPA_MAXBLOCKSIZE)
772 		size = SPA_MAXBLOCKSIZE;
773 	else
774 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
775 
776 	if (ibs == dn->dn_indblkshift)
777 		ibs = 0;
778 
779 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
780 		return (0);
781 
782 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
783 
784 	/* Check for any allocated blocks beyond the first */
785 	if (dn->dn_phys->dn_maxblkid != 0)
786 		goto fail;
787 
788 	mutex_enter(&dn->dn_dbufs_mtx);
789 	for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
790 		db_next = list_next(&dn->dn_dbufs, db);
791 
792 		if (db->db_blkid == 0) {
793 			have_db0 = TRUE;
794 		} else if (db->db_blkid != DB_BONUS_BLKID) {
795 			mutex_exit(&dn->dn_dbufs_mtx);
796 			goto fail;
797 		}
798 	}
799 	mutex_exit(&dn->dn_dbufs_mtx);
800 
801 	if (ibs && dn->dn_nlevels != 1)
802 		goto fail;
803 
804 	db = NULL;
805 	if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) || have_db0) {
806 		/* obtain the old block */
807 		db = dbuf_hold(dn, 0, FTAG);
808 		dbuf_new_size(db, size, tx);
809 	}
810 
811 	dnode_setdblksz(dn, size);
812 	dnode_setdirty(dn, tx);
813 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
814 	if (ibs) {
815 		dn->dn_indblkshift = ibs;
816 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
817 	}
818 
819 	if (db)
820 		dbuf_rele(db, FTAG);
821 
822 	rw_exit(&dn->dn_struct_rwlock);
823 	return (0);
824 
825 fail:
826 	rw_exit(&dn->dn_struct_rwlock);
827 	return (ENOTSUP);
828 }
829 
830 void
831 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx)
832 {
833 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
834 	int drop_struct_lock = FALSE;
835 	int epbs, new_nlevels;
836 	uint64_t sz;
837 
838 	ASSERT(blkid != DB_BONUS_BLKID);
839 
840 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
841 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
842 		drop_struct_lock = TRUE;
843 	}
844 
845 	if (blkid <= dn->dn_maxblkid)
846 		goto out;
847 
848 	dn->dn_maxblkid = blkid;
849 
850 	/*
851 	 * Compute the number of levels necessary to support the new maxblkid.
852 	 */
853 	new_nlevels = 1;
854 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
855 	for (sz = dn->dn_nblkptr;
856 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
857 		new_nlevels++;
858 
859 	if (new_nlevels > dn->dn_nlevels) {
860 		int old_nlevels = dn->dn_nlevels;
861 		dmu_buf_impl_t *db;
862 		list_t *list;
863 		dbuf_dirty_record_t *new, *dr, *dr_next;
864 
865 		dn->dn_nlevels = new_nlevels;
866 
867 		ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
868 		dn->dn_next_nlevels[txgoff] = new_nlevels;
869 
870 		/* dirty the left indirects */
871 		db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
872 		new = dbuf_dirty(db, tx);
873 		dbuf_rele(db, FTAG);
874 
875 		/* transfer the dirty records to the new indirect */
876 		mutex_enter(&dn->dn_mtx);
877 		mutex_enter(&new->dt.di.dr_mtx);
878 		list = &dn->dn_dirty_records[txgoff];
879 		for (dr = list_head(list); dr; dr = dr_next) {
880 			dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
881 			if (dr->dr_dbuf->db_level != new_nlevels-1 &&
882 			    dr->dr_dbuf->db_blkid != DB_BONUS_BLKID) {
883 				ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
884 				list_remove(&dn->dn_dirty_records[txgoff], dr);
885 				list_insert_tail(&new->dt.di.dr_children, dr);
886 				dr->dr_parent = new;
887 			}
888 		}
889 		mutex_exit(&new->dt.di.dr_mtx);
890 		mutex_exit(&dn->dn_mtx);
891 	}
892 
893 out:
894 	if (drop_struct_lock)
895 		rw_exit(&dn->dn_struct_rwlock);
896 }
897 
898 void
899 dnode_clear_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
900 {
901 	avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
902 	avl_index_t where;
903 	free_range_t *rp;
904 	free_range_t rp_tofind;
905 	uint64_t endblk = blkid + nblks;
906 
907 	ASSERT(MUTEX_HELD(&dn->dn_mtx));
908 	ASSERT(nblks <= UINT64_MAX - blkid); /* no overflow */
909 
910 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
911 	    blkid, nblks, tx->tx_txg);
912 	rp_tofind.fr_blkid = blkid;
913 	rp = avl_find(tree, &rp_tofind, &where);
914 	if (rp == NULL)
915 		rp = avl_nearest(tree, where, AVL_BEFORE);
916 	if (rp == NULL)
917 		rp = avl_nearest(tree, where, AVL_AFTER);
918 
919 	while (rp && (rp->fr_blkid <= blkid + nblks)) {
920 		uint64_t fr_endblk = rp->fr_blkid + rp->fr_nblks;
921 		free_range_t *nrp = AVL_NEXT(tree, rp);
922 
923 		if (blkid <= rp->fr_blkid && endblk >= fr_endblk) {
924 			/* clear this entire range */
925 			avl_remove(tree, rp);
926 			kmem_free(rp, sizeof (free_range_t));
927 		} else if (blkid <= rp->fr_blkid &&
928 		    endblk > rp->fr_blkid && endblk < fr_endblk) {
929 			/* clear the beginning of this range */
930 			rp->fr_blkid = endblk;
931 			rp->fr_nblks = fr_endblk - endblk;
932 		} else if (blkid > rp->fr_blkid && blkid < fr_endblk &&
933 		    endblk >= fr_endblk) {
934 			/* clear the end of this range */
935 			rp->fr_nblks = blkid - rp->fr_blkid;
936 		} else if (blkid > rp->fr_blkid && endblk < fr_endblk) {
937 			/* clear a chunk out of this range */
938 			free_range_t *new_rp =
939 			    kmem_alloc(sizeof (free_range_t), KM_SLEEP);
940 
941 			new_rp->fr_blkid = endblk;
942 			new_rp->fr_nblks = fr_endblk - endblk;
943 			avl_insert_here(tree, new_rp, rp, AVL_AFTER);
944 			rp->fr_nblks = blkid - rp->fr_blkid;
945 		}
946 		/* there may be no overlap */
947 		rp = nrp;
948 	}
949 }
950 
951 void
952 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
953 {
954 	dmu_buf_impl_t *db;
955 	uint64_t blkoff, blkid, nblks;
956 	int blksz, head;
957 	int trunc = FALSE;
958 
959 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
960 	blksz = dn->dn_datablksz;
961 
962 	/* If the range is past the end of the file, this is a no-op */
963 	if (off >= blksz * (dn->dn_maxblkid+1))
964 		goto out;
965 	if (len == -1ULL) {
966 		len = UINT64_MAX - off;
967 		trunc = TRUE;
968 	}
969 
970 	/*
971 	 * First, block align the region to free:
972 	 */
973 	if (ISP2(blksz)) {
974 		head = P2NPHASE(off, blksz);
975 		blkoff = P2PHASE(off, blksz);
976 	} else {
977 		ASSERT(dn->dn_maxblkid == 0);
978 		if (off == 0 && len >= blksz) {
979 			/* Freeing the whole block; don't do any head. */
980 			head = 0;
981 		} else {
982 			/* Freeing part of the block. */
983 			head = blksz - off;
984 			ASSERT3U(head, >, 0);
985 		}
986 		blkoff = off;
987 	}
988 	/* zero out any partial block data at the start of the range */
989 	if (head) {
990 		ASSERT3U(blkoff + head, ==, blksz);
991 		if (len < head)
992 			head = len;
993 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
994 		    FTAG, &db) == 0) {
995 			caddr_t data;
996 
997 			/* don't dirty if it isn't on disk and isn't dirty */
998 			if (db->db_last_dirty ||
999 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1000 				rw_exit(&dn->dn_struct_rwlock);
1001 				dbuf_will_dirty(db, tx);
1002 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1003 				data = db->db.db_data;
1004 				bzero(data + blkoff, head);
1005 			}
1006 			dbuf_rele(db, FTAG);
1007 		}
1008 		off += head;
1009 		len -= head;
1010 	}
1011 
1012 	/* If the range was less than one block, we're done */
1013 	if (len == 0 || off >= blksz * (dn->dn_maxblkid+1))
1014 		goto out;
1015 
1016 	if (!ISP2(blksz)) {
1017 		/*
1018 		 * They are freeing the whole block of a
1019 		 * non-power-of-two blocksize file.  Skip all the messy
1020 		 * math.
1021 		 */
1022 		ASSERT3U(off, ==, 0);
1023 		ASSERT3U(len, >=, blksz);
1024 		blkid = 0;
1025 		nblks = 1;
1026 	} else {
1027 		int tail;
1028 		int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1029 		int blkshift = dn->dn_datablkshift;
1030 
1031 		/* If the remaining range is past end of file, we're done */
1032 		if (off > dn->dn_maxblkid << blkshift)
1033 			goto out;
1034 
1035 		if (off + len == UINT64_MAX)
1036 			tail = 0;
1037 		else
1038 			tail = P2PHASE(len, blksz);
1039 
1040 		ASSERT3U(P2PHASE(off, blksz), ==, 0);
1041 		/* zero out any partial block data at the end of the range */
1042 		if (tail) {
1043 			if (len < tail)
1044 				tail = len;
1045 			if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1046 			    TRUE, FTAG, &db) == 0) {
1047 				/* don't dirty if not on disk and not dirty */
1048 				if (db->db_last_dirty ||
1049 				    (db->db_blkptr &&
1050 				    !BP_IS_HOLE(db->db_blkptr))) {
1051 					rw_exit(&dn->dn_struct_rwlock);
1052 					dbuf_will_dirty(db, tx);
1053 					rw_enter(&dn->dn_struct_rwlock,
1054 					    RW_WRITER);
1055 					bzero(db->db.db_data, tail);
1056 				}
1057 				dbuf_rele(db, FTAG);
1058 			}
1059 			len -= tail;
1060 		}
1061 		/* If the range did not include a full block, we are done */
1062 		if (len == 0)
1063 			goto out;
1064 
1065 		/* dirty the left indirects */
1066 		if (dn->dn_nlevels > 1 && off != 0) {
1067 			db = dbuf_hold_level(dn, 1,
1068 			    (off - head) >> (blkshift + epbs), FTAG);
1069 			dbuf_will_dirty(db, tx);
1070 			dbuf_rele(db, FTAG);
1071 		}
1072 
1073 		/* dirty the right indirects */
1074 		if (dn->dn_nlevels > 1 && !trunc) {
1075 			db = dbuf_hold_level(dn, 1,
1076 			    (off + len + tail - 1) >> (blkshift + epbs), FTAG);
1077 			dbuf_will_dirty(db, tx);
1078 			dbuf_rele(db, FTAG);
1079 		}
1080 
1081 		/*
1082 		 * Finally, add this range to the dnode range list, we
1083 		 * will finish up this free operation in the syncing phase.
1084 		 */
1085 		ASSERT(IS_P2ALIGNED(off, 1<<blkshift));
1086 		ASSERT(off + len == UINT64_MAX ||
1087 		    IS_P2ALIGNED(len, 1<<blkshift));
1088 		blkid = off >> blkshift;
1089 		nblks = len >> blkshift;
1090 
1091 		if (trunc)
1092 			dn->dn_maxblkid = (blkid ? blkid - 1 : 0);
1093 	}
1094 
1095 	mutex_enter(&dn->dn_mtx);
1096 	dnode_clear_range(dn, blkid, nblks, tx);
1097 	{
1098 		free_range_t *rp, *found;
1099 		avl_index_t where;
1100 		avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1101 
1102 		/* Add new range to dn_ranges */
1103 		rp = kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1104 		rp->fr_blkid = blkid;
1105 		rp->fr_nblks = nblks;
1106 		found = avl_find(tree, rp, &where);
1107 		ASSERT(found == NULL);
1108 		avl_insert(tree, rp, where);
1109 		dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1110 		    blkid, nblks, tx->tx_txg);
1111 	}
1112 	mutex_exit(&dn->dn_mtx);
1113 
1114 	dbuf_free_range(dn, blkid, nblks, tx);
1115 	dnode_setdirty(dn, tx);
1116 out:
1117 	rw_exit(&dn->dn_struct_rwlock);
1118 }
1119 
1120 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1121 uint64_t
1122 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1123 {
1124 	free_range_t range_tofind;
1125 	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1126 	int i;
1127 
1128 	if (blkid == DB_BONUS_BLKID)
1129 		return (FALSE);
1130 
1131 	/*
1132 	 * If we're in the process of opening the pool, dp will not be
1133 	 * set yet, but there shouldn't be anything dirty.
1134 	 */
1135 	if (dp == NULL)
1136 		return (FALSE);
1137 
1138 	if (dn->dn_free_txg)
1139 		return (TRUE);
1140 
1141 	/*
1142 	 * If dn_datablkshift is not set, then there's only a single
1143 	 * block, in which case there will never be a free range so it
1144 	 * won't matter.
1145 	 */
1146 	range_tofind.fr_blkid = blkid;
1147 	mutex_enter(&dn->dn_mtx);
1148 	for (i = 0; i < TXG_SIZE; i++) {
1149 		free_range_t *range_found;
1150 		avl_index_t idx;
1151 
1152 		range_found = avl_find(&dn->dn_ranges[i], &range_tofind, &idx);
1153 		if (range_found) {
1154 			ASSERT(range_found->fr_nblks > 0);
1155 			break;
1156 		}
1157 		range_found = avl_nearest(&dn->dn_ranges[i], idx, AVL_BEFORE);
1158 		if (range_found &&
1159 		    range_found->fr_blkid + range_found->fr_nblks > blkid)
1160 			break;
1161 	}
1162 	mutex_exit(&dn->dn_mtx);
1163 	return (i < TXG_SIZE);
1164 }
1165 
1166 /* call from syncing context when we actually write/free space for this dnode */
1167 void
1168 dnode_diduse_space(dnode_t *dn, int64_t delta)
1169 {
1170 	uint64_t space;
1171 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1172 	    dn, dn->dn_phys,
1173 	    (u_longlong_t)dn->dn_phys->dn_used,
1174 	    (longlong_t)delta);
1175 
1176 	mutex_enter(&dn->dn_mtx);
1177 	space = DN_USED_BYTES(dn->dn_phys);
1178 	if (delta > 0) {
1179 		ASSERT3U(space + delta, >=, space); /* no overflow */
1180 	} else {
1181 		ASSERT3U(space, >=, -delta); /* no underflow */
1182 	}
1183 	space += delta;
1184 	if (spa_version(dn->dn_objset->os_spa) < ZFS_VERSION_DNODE_BYTES) {
1185 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1186 		ASSERT3U(P2PHASE(space, 1<<DEV_BSHIFT), ==, 0);
1187 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1188 	} else {
1189 		dn->dn_phys->dn_used = space;
1190 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1191 	}
1192 	mutex_exit(&dn->dn_mtx);
1193 }
1194 
1195 /*
1196  * Call when we think we're going to write/free space in open context.
1197  * Be conservative (ie. OK to write less than this or free more than
1198  * this, but don't write more or free less).
1199  */
1200 void
1201 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1202 {
1203 	objset_impl_t *os = dn->dn_objset;
1204 	dsl_dataset_t *ds = os->os_dsl_dataset;
1205 
1206 	if (space > 0)
1207 		space = spa_get_asize(os->os_spa, space);
1208 
1209 	if (ds)
1210 		dsl_dir_willuse_space(ds->ds_dir, space, tx);
1211 
1212 	dmu_tx_willuse_space(tx, space);
1213 }
1214 
1215 static int
1216 dnode_next_offset_level(dnode_t *dn, boolean_t hole, uint64_t *offset,
1217 	int lvl, uint64_t blkfill, uint64_t txg)
1218 {
1219 	dmu_buf_impl_t *db = NULL;
1220 	void *data = NULL;
1221 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1222 	uint64_t epb = 1ULL << epbs;
1223 	uint64_t minfill, maxfill;
1224 	int i, error, span;
1225 
1226 	dprintf("probing object %llu offset %llx level %d of %u\n",
1227 	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1228 
1229 	if (lvl == dn->dn_phys->dn_nlevels) {
1230 		error = 0;
1231 		epb = dn->dn_phys->dn_nblkptr;
1232 		data = dn->dn_phys->dn_blkptr;
1233 	} else {
1234 		uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1235 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1236 		if (error) {
1237 			if (error == ENOENT)
1238 				return (hole ? 0 : ESRCH);
1239 			return (error);
1240 		}
1241 		error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1242 		if (error) {
1243 			dbuf_rele(db, FTAG);
1244 			return (error);
1245 		}
1246 		data = db->db.db_data;
1247 	}
1248 
1249 	if (db && txg &&
1250 	    (db->db_blkptr == NULL || db->db_blkptr->blk_birth <= txg)) {
1251 		error = ESRCH;
1252 	} else if (lvl == 0) {
1253 		dnode_phys_t *dnp = data;
1254 		span = DNODE_SHIFT;
1255 		ASSERT(dn->dn_type == DMU_OT_DNODE);
1256 
1257 		for (i = (*offset >> span) & (blkfill - 1); i < blkfill; i++) {
1258 			boolean_t newcontents = B_TRUE;
1259 			if (txg) {
1260 				int j;
1261 				newcontents = B_FALSE;
1262 				for (j = 0; j < dnp[i].dn_nblkptr; j++) {
1263 					if (dnp[i].dn_blkptr[j].blk_birth > txg)
1264 						newcontents = B_TRUE;
1265 				}
1266 			}
1267 			if (!dnp[i].dn_type == hole && newcontents)
1268 				break;
1269 			*offset += 1ULL << span;
1270 		}
1271 		if (i == blkfill)
1272 			error = ESRCH;
1273 	} else {
1274 		blkptr_t *bp = data;
1275 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
1276 		minfill = 0;
1277 		maxfill = blkfill << ((lvl - 1) * epbs);
1278 
1279 		if (hole)
1280 			maxfill--;
1281 		else
1282 			minfill++;
1283 
1284 		for (i = (*offset >> span) & ((1ULL << epbs) - 1);
1285 		    i < epb; i++) {
1286 			if (bp[i].blk_fill >= minfill &&
1287 			    bp[i].blk_fill <= maxfill &&
1288 			    bp[i].blk_birth > txg)
1289 				break;
1290 			*offset += 1ULL << span;
1291 		}
1292 		if (i >= epb)
1293 			error = ESRCH;
1294 	}
1295 
1296 	if (db)
1297 		dbuf_rele(db, FTAG);
1298 
1299 	return (error);
1300 }
1301 
1302 /*
1303  * Find the next hole, data, or sparse region at or after *offset.
1304  * The value 'blkfill' tells us how many items we expect to find
1305  * in an L0 data block; this value is 1 for normal objects,
1306  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1307  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1308  *
1309  * Examples:
1310  *
1311  * dnode_next_offset(dn, hole, offset, 1, 1, 0);
1312  *	Finds the next hole/data in a file.
1313  *	Used in dmu_offset_next().
1314  *
1315  * dnode_next_offset(mdn, hole, offset, 0, DNODES_PER_BLOCK, txg);
1316  *	Finds the next free/allocated dnode an objset's meta-dnode.
1317  *	Only finds objects that have new contents since txg (ie.
1318  *	bonus buffer changes and content removal are ignored).
1319  *	Used in dmu_object_next().
1320  *
1321  * dnode_next_offset(mdn, TRUE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1322  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
1323  *	Used in dmu_object_alloc().
1324  */
1325 int
1326 dnode_next_offset(dnode_t *dn, boolean_t hole, uint64_t *offset,
1327     int minlvl, uint64_t blkfill, uint64_t txg)
1328 {
1329 	int lvl, maxlvl;
1330 	int error = 0;
1331 	uint64_t initial_offset = *offset;
1332 
1333 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1334 
1335 	if (dn->dn_phys->dn_nlevels == 0) {
1336 		rw_exit(&dn->dn_struct_rwlock);
1337 		return (ESRCH);
1338 	}
1339 
1340 	if (dn->dn_datablkshift == 0) {
1341 		if (*offset < dn->dn_datablksz) {
1342 			if (hole)
1343 				*offset = dn->dn_datablksz;
1344 		} else {
1345 			error = ESRCH;
1346 		}
1347 		rw_exit(&dn->dn_struct_rwlock);
1348 		return (error);
1349 	}
1350 
1351 	maxlvl = dn->dn_phys->dn_nlevels;
1352 
1353 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1354 		error = dnode_next_offset_level(dn,
1355 		    hole, offset, lvl, blkfill, txg);
1356 		if (error != ESRCH)
1357 			break;
1358 	}
1359 
1360 	while (--lvl >= minlvl && error == 0) {
1361 		error = dnode_next_offset_level(dn,
1362 		    hole, offset, lvl, blkfill, txg);
1363 	}
1364 
1365 	rw_exit(&dn->dn_struct_rwlock);
1366 
1367 	if (error == 0 && initial_offset > *offset)
1368 		error = ESRCH;
1369 
1370 	return (error);
1371 }
1372