xref: /titanic_44/usr/src/uts/common/fs/zfs/dnode.c (revision 9fa718d2f477620f14e3f2948dd03e3470add804)
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 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/dbuf.h>
28 #include <sys/dnode.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/spa.h>
36 #include <sys/zio.h>
37 #include <sys/dmu_zfetch.h>
38 
39 static int free_range_compar(const void *node1, const void *node2);
40 
41 static kmem_cache_t *dnode_cache;
42 /*
43  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
44  * turned on when DEBUG is also defined.
45  */
46 #ifdef	DEBUG
47 #define	DNODE_STATS
48 #endif	/* DEBUG */
49 
50 #ifdef	DNODE_STATS
51 #define	DNODE_STAT_ADD(stat)			((stat)++)
52 #else
53 #define	DNODE_STAT_ADD(stat)			/* nothing */
54 #endif	/* DNODE_STATS */
55 
56 static dnode_phys_t dnode_phys_zero;
57 
58 int zfs_default_bs = SPA_MINBLOCKSHIFT;
59 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
60 
61 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
62 
63 /* ARGSUSED */
64 static int
65 dnode_cons(void *arg, void *unused, int kmflag)
66 {
67 	dnode_t *dn = arg;
68 	int i;
69 
70 	rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
71 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
72 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
73 	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
74 
75 	refcount_create(&dn->dn_holds);
76 	refcount_create(&dn->dn_tx_holds);
77 	list_link_init(&dn->dn_link);
78 
79 	bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
80 	bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
81 	bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
82 	bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
83 	bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
84 	bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
85 	bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
86 
87 	for (i = 0; i < TXG_SIZE; i++) {
88 		list_link_init(&dn->dn_dirty_link[i]);
89 		avl_create(&dn->dn_ranges[i], free_range_compar,
90 		    sizeof (free_range_t),
91 		    offsetof(struct free_range, fr_node));
92 		list_create(&dn->dn_dirty_records[i],
93 		    sizeof (dbuf_dirty_record_t),
94 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
95 	}
96 
97 	dn->dn_allocated_txg = 0;
98 	dn->dn_free_txg = 0;
99 	dn->dn_assigned_txg = 0;
100 	dn->dn_dirtyctx = 0;
101 	dn->dn_dirtyctx_firstset = NULL;
102 	dn->dn_bonus = NULL;
103 	dn->dn_have_spill = B_FALSE;
104 	dn->dn_zio = NULL;
105 	dn->dn_oldused = 0;
106 	dn->dn_oldflags = 0;
107 	dn->dn_olduid = 0;
108 	dn->dn_oldgid = 0;
109 	dn->dn_newuid = 0;
110 	dn->dn_newgid = 0;
111 	dn->dn_id_flags = 0;
112 
113 	dn->dn_dbufs_count = 0;
114 	list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
115 	    offsetof(dmu_buf_impl_t, db_link));
116 
117 	dn->dn_moved = 0;
118 	return (0);
119 }
120 
121 /* ARGSUSED */
122 static void
123 dnode_dest(void *arg, void *unused)
124 {
125 	int i;
126 	dnode_t *dn = arg;
127 
128 	rw_destroy(&dn->dn_struct_rwlock);
129 	mutex_destroy(&dn->dn_mtx);
130 	mutex_destroy(&dn->dn_dbufs_mtx);
131 	cv_destroy(&dn->dn_notxholds);
132 	refcount_destroy(&dn->dn_holds);
133 	refcount_destroy(&dn->dn_tx_holds);
134 	ASSERT(!list_link_active(&dn->dn_link));
135 
136 	for (i = 0; i < TXG_SIZE; i++) {
137 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
138 		avl_destroy(&dn->dn_ranges[i]);
139 		list_destroy(&dn->dn_dirty_records[i]);
140 		ASSERT0(dn->dn_next_nblkptr[i]);
141 		ASSERT0(dn->dn_next_nlevels[i]);
142 		ASSERT0(dn->dn_next_indblkshift[i]);
143 		ASSERT0(dn->dn_next_bonustype[i]);
144 		ASSERT0(dn->dn_rm_spillblk[i]);
145 		ASSERT0(dn->dn_next_bonuslen[i]);
146 		ASSERT0(dn->dn_next_blksz[i]);
147 	}
148 
149 	ASSERT0(dn->dn_allocated_txg);
150 	ASSERT0(dn->dn_free_txg);
151 	ASSERT0(dn->dn_assigned_txg);
152 	ASSERT0(dn->dn_dirtyctx);
153 	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
154 	ASSERT3P(dn->dn_bonus, ==, NULL);
155 	ASSERT(!dn->dn_have_spill);
156 	ASSERT3P(dn->dn_zio, ==, NULL);
157 	ASSERT0(dn->dn_oldused);
158 	ASSERT0(dn->dn_oldflags);
159 	ASSERT0(dn->dn_olduid);
160 	ASSERT0(dn->dn_oldgid);
161 	ASSERT0(dn->dn_newuid);
162 	ASSERT0(dn->dn_newgid);
163 	ASSERT0(dn->dn_id_flags);
164 
165 	ASSERT0(dn->dn_dbufs_count);
166 	list_destroy(&dn->dn_dbufs);
167 }
168 
169 void
170 dnode_init(void)
171 {
172 	ASSERT(dnode_cache == NULL);
173 	dnode_cache = kmem_cache_create("dnode_t",
174 	    sizeof (dnode_t),
175 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
176 	kmem_cache_set_move(dnode_cache, dnode_move);
177 }
178 
179 void
180 dnode_fini(void)
181 {
182 	kmem_cache_destroy(dnode_cache);
183 	dnode_cache = NULL;
184 }
185 
186 
187 #ifdef ZFS_DEBUG
188 void
189 dnode_verify(dnode_t *dn)
190 {
191 	int drop_struct_lock = FALSE;
192 
193 	ASSERT(dn->dn_phys);
194 	ASSERT(dn->dn_objset);
195 	ASSERT(dn->dn_handle->dnh_dnode == dn);
196 
197 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
198 
199 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
200 		return;
201 
202 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
203 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
204 		drop_struct_lock = TRUE;
205 	}
206 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
207 		int i;
208 		ASSERT3U(dn->dn_indblkshift, >=, 0);
209 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
210 		if (dn->dn_datablkshift) {
211 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
212 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
213 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
214 		}
215 		ASSERT3U(dn->dn_nlevels, <=, 30);
216 		ASSERT(DMU_OT_IS_VALID(dn->dn_type));
217 		ASSERT3U(dn->dn_nblkptr, >=, 1);
218 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
219 		ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
220 		ASSERT3U(dn->dn_datablksz, ==,
221 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
222 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
223 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
224 		    dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
225 		for (i = 0; i < TXG_SIZE; i++) {
226 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
227 		}
228 	}
229 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
230 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
231 	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
232 	if (dn->dn_dbuf != NULL) {
233 		ASSERT3P(dn->dn_phys, ==,
234 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
235 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
236 	}
237 	if (drop_struct_lock)
238 		rw_exit(&dn->dn_struct_rwlock);
239 }
240 #endif
241 
242 void
243 dnode_byteswap(dnode_phys_t *dnp)
244 {
245 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
246 	int i;
247 
248 	if (dnp->dn_type == DMU_OT_NONE) {
249 		bzero(dnp, sizeof (dnode_phys_t));
250 		return;
251 	}
252 
253 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
254 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
255 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
256 	dnp->dn_used = BSWAP_64(dnp->dn_used);
257 
258 	/*
259 	 * dn_nblkptr is only one byte, so it's OK to read it in either
260 	 * byte order.  We can't read dn_bouslen.
261 	 */
262 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
263 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
264 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
265 		buf64[i] = BSWAP_64(buf64[i]);
266 
267 	/*
268 	 * OK to check dn_bonuslen for zero, because it won't matter if
269 	 * we have the wrong byte order.  This is necessary because the
270 	 * dnode dnode is smaller than a regular dnode.
271 	 */
272 	if (dnp->dn_bonuslen != 0) {
273 		/*
274 		 * Note that the bonus length calculated here may be
275 		 * longer than the actual bonus buffer.  This is because
276 		 * we always put the bonus buffer after the last block
277 		 * pointer (instead of packing it against the end of the
278 		 * dnode buffer).
279 		 */
280 		int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
281 		size_t len = DN_MAX_BONUSLEN - off;
282 		ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
283 		dmu_object_byteswap_t byteswap =
284 		    DMU_OT_BYTESWAP(dnp->dn_bonustype);
285 		dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
286 	}
287 
288 	/* Swap SPILL block if we have one */
289 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
290 		byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
291 
292 }
293 
294 void
295 dnode_buf_byteswap(void *vbuf, size_t size)
296 {
297 	dnode_phys_t *buf = vbuf;
298 	int i;
299 
300 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
301 	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
302 
303 	size >>= DNODE_SHIFT;
304 	for (i = 0; i < size; i++) {
305 		dnode_byteswap(buf);
306 		buf++;
307 	}
308 }
309 
310 static int
311 free_range_compar(const void *node1, const void *node2)
312 {
313 	const free_range_t *rp1 = node1;
314 	const free_range_t *rp2 = node2;
315 
316 	if (rp1->fr_blkid < rp2->fr_blkid)
317 		return (-1);
318 	else if (rp1->fr_blkid > rp2->fr_blkid)
319 		return (1);
320 	else return (0);
321 }
322 
323 void
324 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
325 {
326 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
327 
328 	dnode_setdirty(dn, tx);
329 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
330 	ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
331 	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
332 	dn->dn_bonuslen = newsize;
333 	if (newsize == 0)
334 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
335 	else
336 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
337 	rw_exit(&dn->dn_struct_rwlock);
338 }
339 
340 void
341 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
342 {
343 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
344 	dnode_setdirty(dn, tx);
345 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
346 	dn->dn_bonustype = newtype;
347 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
348 	rw_exit(&dn->dn_struct_rwlock);
349 }
350 
351 void
352 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
353 {
354 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
355 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
356 	dnode_setdirty(dn, tx);
357 	dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
358 	dn->dn_have_spill = B_FALSE;
359 }
360 
361 static void
362 dnode_setdblksz(dnode_t *dn, int size)
363 {
364 	ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
365 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
366 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
367 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
368 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
369 	dn->dn_datablksz = size;
370 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
371 	dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0;
372 }
373 
374 static dnode_t *
375 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
376     uint64_t object, dnode_handle_t *dnh)
377 {
378 	dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
379 
380 	ASSERT(!POINTER_IS_VALID(dn->dn_objset));
381 	dn->dn_moved = 0;
382 
383 	/*
384 	 * Defer setting dn_objset until the dnode is ready to be a candidate
385 	 * for the dnode_move() callback.
386 	 */
387 	dn->dn_object = object;
388 	dn->dn_dbuf = db;
389 	dn->dn_handle = dnh;
390 	dn->dn_phys = dnp;
391 
392 	if (dnp->dn_datablkszsec) {
393 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
394 	} else {
395 		dn->dn_datablksz = 0;
396 		dn->dn_datablkszsec = 0;
397 		dn->dn_datablkshift = 0;
398 	}
399 	dn->dn_indblkshift = dnp->dn_indblkshift;
400 	dn->dn_nlevels = dnp->dn_nlevels;
401 	dn->dn_type = dnp->dn_type;
402 	dn->dn_nblkptr = dnp->dn_nblkptr;
403 	dn->dn_checksum = dnp->dn_checksum;
404 	dn->dn_compress = dnp->dn_compress;
405 	dn->dn_bonustype = dnp->dn_bonustype;
406 	dn->dn_bonuslen = dnp->dn_bonuslen;
407 	dn->dn_maxblkid = dnp->dn_maxblkid;
408 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
409 	dn->dn_id_flags = 0;
410 
411 	dmu_zfetch_init(&dn->dn_zfetch, dn);
412 
413 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
414 
415 	mutex_enter(&os->os_lock);
416 	list_insert_head(&os->os_dnodes, dn);
417 	membar_producer();
418 	/*
419 	 * Everything else must be valid before assigning dn_objset makes the
420 	 * dnode eligible for dnode_move().
421 	 */
422 	dn->dn_objset = os;
423 	mutex_exit(&os->os_lock);
424 
425 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
426 	return (dn);
427 }
428 
429 /*
430  * Caller must be holding the dnode handle, which is released upon return.
431  */
432 static void
433 dnode_destroy(dnode_t *dn)
434 {
435 	objset_t *os = dn->dn_objset;
436 
437 	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
438 
439 	mutex_enter(&os->os_lock);
440 	POINTER_INVALIDATE(&dn->dn_objset);
441 	list_remove(&os->os_dnodes, dn);
442 	mutex_exit(&os->os_lock);
443 
444 	/* the dnode can no longer move, so we can release the handle */
445 	zrl_remove(&dn->dn_handle->dnh_zrlock);
446 
447 	dn->dn_allocated_txg = 0;
448 	dn->dn_free_txg = 0;
449 	dn->dn_assigned_txg = 0;
450 
451 	dn->dn_dirtyctx = 0;
452 	if (dn->dn_dirtyctx_firstset != NULL) {
453 		kmem_free(dn->dn_dirtyctx_firstset, 1);
454 		dn->dn_dirtyctx_firstset = NULL;
455 	}
456 	if (dn->dn_bonus != NULL) {
457 		mutex_enter(&dn->dn_bonus->db_mtx);
458 		dbuf_evict(dn->dn_bonus);
459 		dn->dn_bonus = NULL;
460 	}
461 	dn->dn_zio = NULL;
462 
463 	dn->dn_have_spill = B_FALSE;
464 	dn->dn_oldused = 0;
465 	dn->dn_oldflags = 0;
466 	dn->dn_olduid = 0;
467 	dn->dn_oldgid = 0;
468 	dn->dn_newuid = 0;
469 	dn->dn_newgid = 0;
470 	dn->dn_id_flags = 0;
471 
472 	dmu_zfetch_rele(&dn->dn_zfetch);
473 	kmem_cache_free(dnode_cache, dn);
474 	arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
475 }
476 
477 void
478 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
479     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
480 {
481 	int i;
482 
483 	if (blocksize == 0)
484 		blocksize = 1 << zfs_default_bs;
485 	else if (blocksize > SPA_MAXBLOCKSIZE)
486 		blocksize = SPA_MAXBLOCKSIZE;
487 	else
488 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
489 
490 	if (ibs == 0)
491 		ibs = zfs_default_ibs;
492 
493 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
494 
495 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
496 	    dn->dn_object, tx->tx_txg, blocksize, ibs);
497 
498 	ASSERT(dn->dn_type == DMU_OT_NONE);
499 	ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
500 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
501 	ASSERT(ot != DMU_OT_NONE);
502 	ASSERT(DMU_OT_IS_VALID(ot));
503 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
504 	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
505 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
506 	ASSERT(DMU_OT_IS_VALID(bonustype));
507 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
508 	ASSERT(dn->dn_type == DMU_OT_NONE);
509 	ASSERT0(dn->dn_maxblkid);
510 	ASSERT0(dn->dn_allocated_txg);
511 	ASSERT0(dn->dn_assigned_txg);
512 	ASSERT(refcount_is_zero(&dn->dn_tx_holds));
513 	ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
514 	ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
515 
516 	for (i = 0; i < TXG_SIZE; i++) {
517 		ASSERT0(dn->dn_next_nblkptr[i]);
518 		ASSERT0(dn->dn_next_nlevels[i]);
519 		ASSERT0(dn->dn_next_indblkshift[i]);
520 		ASSERT0(dn->dn_next_bonuslen[i]);
521 		ASSERT0(dn->dn_next_bonustype[i]);
522 		ASSERT0(dn->dn_rm_spillblk[i]);
523 		ASSERT0(dn->dn_next_blksz[i]);
524 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
525 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
526 		ASSERT0(avl_numnodes(&dn->dn_ranges[i]));
527 	}
528 
529 	dn->dn_type = ot;
530 	dnode_setdblksz(dn, blocksize);
531 	dn->dn_indblkshift = ibs;
532 	dn->dn_nlevels = 1;
533 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
534 		dn->dn_nblkptr = 1;
535 	else
536 		dn->dn_nblkptr = 1 +
537 		    ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
538 	dn->dn_bonustype = bonustype;
539 	dn->dn_bonuslen = bonuslen;
540 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
541 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
542 	dn->dn_dirtyctx = 0;
543 
544 	dn->dn_free_txg = 0;
545 	if (dn->dn_dirtyctx_firstset) {
546 		kmem_free(dn->dn_dirtyctx_firstset, 1);
547 		dn->dn_dirtyctx_firstset = NULL;
548 	}
549 
550 	dn->dn_allocated_txg = tx->tx_txg;
551 	dn->dn_id_flags = 0;
552 
553 	dnode_setdirty(dn, tx);
554 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
555 	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
556 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
557 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
558 }
559 
560 void
561 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
562     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
563 {
564 	int nblkptr;
565 
566 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
567 	ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
568 	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
569 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
570 	ASSERT(tx->tx_txg != 0);
571 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
572 	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
573 	    (bonustype == DMU_OT_SA && bonuslen == 0));
574 	ASSERT(DMU_OT_IS_VALID(bonustype));
575 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
576 
577 	/* clean up any unreferenced dbufs */
578 	dnode_evict_dbufs(dn);
579 
580 	dn->dn_id_flags = 0;
581 
582 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
583 	dnode_setdirty(dn, tx);
584 	if (dn->dn_datablksz != blocksize) {
585 		/* change blocksize */
586 		ASSERT(dn->dn_maxblkid == 0 &&
587 		    (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
588 		    dnode_block_freed(dn, 0)));
589 		dnode_setdblksz(dn, blocksize);
590 		dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
591 	}
592 	if (dn->dn_bonuslen != bonuslen)
593 		dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
594 
595 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
596 		nblkptr = 1;
597 	else
598 		nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
599 	if (dn->dn_bonustype != bonustype)
600 		dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
601 	if (dn->dn_nblkptr != nblkptr)
602 		dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
603 	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
604 		dbuf_rm_spill(dn, tx);
605 		dnode_rm_spill(dn, tx);
606 	}
607 	rw_exit(&dn->dn_struct_rwlock);
608 
609 	/* change type */
610 	dn->dn_type = ot;
611 
612 	/* change bonus size and type */
613 	mutex_enter(&dn->dn_mtx);
614 	dn->dn_bonustype = bonustype;
615 	dn->dn_bonuslen = bonuslen;
616 	dn->dn_nblkptr = nblkptr;
617 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
618 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
619 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
620 
621 	/* fix up the bonus db_size */
622 	if (dn->dn_bonus) {
623 		dn->dn_bonus->db.db_size =
624 		    DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
625 		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
626 	}
627 
628 	dn->dn_allocated_txg = tx->tx_txg;
629 	mutex_exit(&dn->dn_mtx);
630 }
631 
632 #ifdef	DNODE_STATS
633 static struct {
634 	uint64_t dms_dnode_invalid;
635 	uint64_t dms_dnode_recheck1;
636 	uint64_t dms_dnode_recheck2;
637 	uint64_t dms_dnode_special;
638 	uint64_t dms_dnode_handle;
639 	uint64_t dms_dnode_rwlock;
640 	uint64_t dms_dnode_active;
641 } dnode_move_stats;
642 #endif	/* DNODE_STATS */
643 
644 static void
645 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
646 {
647 	int i;
648 
649 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
650 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
651 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
652 	ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
653 
654 	/* Copy fields. */
655 	ndn->dn_objset = odn->dn_objset;
656 	ndn->dn_object = odn->dn_object;
657 	ndn->dn_dbuf = odn->dn_dbuf;
658 	ndn->dn_handle = odn->dn_handle;
659 	ndn->dn_phys = odn->dn_phys;
660 	ndn->dn_type = odn->dn_type;
661 	ndn->dn_bonuslen = odn->dn_bonuslen;
662 	ndn->dn_bonustype = odn->dn_bonustype;
663 	ndn->dn_nblkptr = odn->dn_nblkptr;
664 	ndn->dn_checksum = odn->dn_checksum;
665 	ndn->dn_compress = odn->dn_compress;
666 	ndn->dn_nlevels = odn->dn_nlevels;
667 	ndn->dn_indblkshift = odn->dn_indblkshift;
668 	ndn->dn_datablkshift = odn->dn_datablkshift;
669 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
670 	ndn->dn_datablksz = odn->dn_datablksz;
671 	ndn->dn_maxblkid = odn->dn_maxblkid;
672 	bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
673 	    sizeof (odn->dn_next_nblkptr));
674 	bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
675 	    sizeof (odn->dn_next_nlevels));
676 	bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
677 	    sizeof (odn->dn_next_indblkshift));
678 	bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
679 	    sizeof (odn->dn_next_bonustype));
680 	bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
681 	    sizeof (odn->dn_rm_spillblk));
682 	bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
683 	    sizeof (odn->dn_next_bonuslen));
684 	bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
685 	    sizeof (odn->dn_next_blksz));
686 	for (i = 0; i < TXG_SIZE; i++) {
687 		list_move_tail(&ndn->dn_dirty_records[i],
688 		    &odn->dn_dirty_records[i]);
689 	}
690 	bcopy(&odn->dn_ranges[0], &ndn->dn_ranges[0], sizeof (odn->dn_ranges));
691 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
692 	ndn->dn_free_txg = odn->dn_free_txg;
693 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
694 	ndn->dn_dirtyctx = odn->dn_dirtyctx;
695 	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
696 	ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
697 	refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
698 	ASSERT(list_is_empty(&ndn->dn_dbufs));
699 	list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs);
700 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
701 	ndn->dn_bonus = odn->dn_bonus;
702 	ndn->dn_have_spill = odn->dn_have_spill;
703 	ndn->dn_zio = odn->dn_zio;
704 	ndn->dn_oldused = odn->dn_oldused;
705 	ndn->dn_oldflags = odn->dn_oldflags;
706 	ndn->dn_olduid = odn->dn_olduid;
707 	ndn->dn_oldgid = odn->dn_oldgid;
708 	ndn->dn_newuid = odn->dn_newuid;
709 	ndn->dn_newgid = odn->dn_newgid;
710 	ndn->dn_id_flags = odn->dn_id_flags;
711 	dmu_zfetch_init(&ndn->dn_zfetch, NULL);
712 	list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
713 	ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
714 	ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
715 	ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
716 
717 	/*
718 	 * Update back pointers. Updating the handle fixes the back pointer of
719 	 * every descendant dbuf as well as the bonus dbuf.
720 	 */
721 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
722 	ndn->dn_handle->dnh_dnode = ndn;
723 	if (ndn->dn_zfetch.zf_dnode == odn) {
724 		ndn->dn_zfetch.zf_dnode = ndn;
725 	}
726 
727 	/*
728 	 * Invalidate the original dnode by clearing all of its back pointers.
729 	 */
730 	odn->dn_dbuf = NULL;
731 	odn->dn_handle = NULL;
732 	list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t),
733 	    offsetof(dmu_buf_impl_t, db_link));
734 	odn->dn_dbufs_count = 0;
735 	odn->dn_bonus = NULL;
736 	odn->dn_zfetch.zf_dnode = NULL;
737 
738 	/*
739 	 * Set the low bit of the objset pointer to ensure that dnode_move()
740 	 * recognizes the dnode as invalid in any subsequent callback.
741 	 */
742 	POINTER_INVALIDATE(&odn->dn_objset);
743 
744 	/*
745 	 * Satisfy the destructor.
746 	 */
747 	for (i = 0; i < TXG_SIZE; i++) {
748 		list_create(&odn->dn_dirty_records[i],
749 		    sizeof (dbuf_dirty_record_t),
750 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
751 		odn->dn_ranges[i].avl_root = NULL;
752 		odn->dn_ranges[i].avl_numnodes = 0;
753 		odn->dn_next_nlevels[i] = 0;
754 		odn->dn_next_indblkshift[i] = 0;
755 		odn->dn_next_bonustype[i] = 0;
756 		odn->dn_rm_spillblk[i] = 0;
757 		odn->dn_next_bonuslen[i] = 0;
758 		odn->dn_next_blksz[i] = 0;
759 	}
760 	odn->dn_allocated_txg = 0;
761 	odn->dn_free_txg = 0;
762 	odn->dn_assigned_txg = 0;
763 	odn->dn_dirtyctx = 0;
764 	odn->dn_dirtyctx_firstset = NULL;
765 	odn->dn_have_spill = B_FALSE;
766 	odn->dn_zio = NULL;
767 	odn->dn_oldused = 0;
768 	odn->dn_oldflags = 0;
769 	odn->dn_olduid = 0;
770 	odn->dn_oldgid = 0;
771 	odn->dn_newuid = 0;
772 	odn->dn_newgid = 0;
773 	odn->dn_id_flags = 0;
774 
775 	/*
776 	 * Mark the dnode.
777 	 */
778 	ndn->dn_moved = 1;
779 	odn->dn_moved = (uint8_t)-1;
780 }
781 
782 #ifdef	_KERNEL
783 /*ARGSUSED*/
784 static kmem_cbrc_t
785 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
786 {
787 	dnode_t *odn = buf, *ndn = newbuf;
788 	objset_t *os;
789 	int64_t refcount;
790 	uint32_t dbufs;
791 
792 	/*
793 	 * The dnode is on the objset's list of known dnodes if the objset
794 	 * pointer is valid. We set the low bit of the objset pointer when
795 	 * freeing the dnode to invalidate it, and the memory patterns written
796 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
797 	 * A newly created dnode sets the objset pointer last of all to indicate
798 	 * that the dnode is known and in a valid state to be moved by this
799 	 * function.
800 	 */
801 	os = odn->dn_objset;
802 	if (!POINTER_IS_VALID(os)) {
803 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
804 		return (KMEM_CBRC_DONT_KNOW);
805 	}
806 
807 	/*
808 	 * Ensure that the objset does not go away during the move.
809 	 */
810 	rw_enter(&os_lock, RW_WRITER);
811 	if (os != odn->dn_objset) {
812 		rw_exit(&os_lock);
813 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
814 		return (KMEM_CBRC_DONT_KNOW);
815 	}
816 
817 	/*
818 	 * If the dnode is still valid, then so is the objset. We know that no
819 	 * valid objset can be freed while we hold os_lock, so we can safely
820 	 * ensure that the objset remains in use.
821 	 */
822 	mutex_enter(&os->os_lock);
823 
824 	/*
825 	 * Recheck the objset pointer in case the dnode was removed just before
826 	 * acquiring the lock.
827 	 */
828 	if (os != odn->dn_objset) {
829 		mutex_exit(&os->os_lock);
830 		rw_exit(&os_lock);
831 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
832 		return (KMEM_CBRC_DONT_KNOW);
833 	}
834 
835 	/*
836 	 * At this point we know that as long as we hold os->os_lock, the dnode
837 	 * cannot be freed and fields within the dnode can be safely accessed.
838 	 * The objset listing this dnode cannot go away as long as this dnode is
839 	 * on its list.
840 	 */
841 	rw_exit(&os_lock);
842 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
843 		mutex_exit(&os->os_lock);
844 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
845 		return (KMEM_CBRC_NO);
846 	}
847 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
848 
849 	/*
850 	 * Lock the dnode handle to prevent the dnode from obtaining any new
851 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
852 	 * from accessing the dnode, so that we can discount their holds. The
853 	 * handle is safe to access because we know that while the dnode cannot
854 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
855 	 * safely move any dnode referenced only by dbufs.
856 	 */
857 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
858 		mutex_exit(&os->os_lock);
859 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
860 		return (KMEM_CBRC_LATER);
861 	}
862 
863 	/*
864 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
865 	 * We need to guarantee that there is a hold for every dbuf in order to
866 	 * determine whether the dnode is actively referenced. Falsely matching
867 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
868 	 * that a thread already having an active dnode hold is about to add a
869 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
870 	 * progress.
871 	 */
872 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
873 		zrl_exit(&odn->dn_handle->dnh_zrlock);
874 		mutex_exit(&os->os_lock);
875 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
876 		return (KMEM_CBRC_LATER);
877 	}
878 
879 	/*
880 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
881 	 * case, the dbuf count is decremented under the handle lock before the
882 	 * dbuf's hold is released. This order ensures that if we count the hold
883 	 * after the dbuf is removed but before its hold is released, we will
884 	 * treat the unmatched hold as active and exit safely. If we count the
885 	 * hold before the dbuf is removed, the hold is discounted, and the
886 	 * removal is blocked until the move completes.
887 	 */
888 	refcount = refcount_count(&odn->dn_holds);
889 	ASSERT(refcount >= 0);
890 	dbufs = odn->dn_dbufs_count;
891 
892 	/* We can't have more dbufs than dnode holds. */
893 	ASSERT3U(dbufs, <=, refcount);
894 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
895 	    uint32_t, dbufs);
896 
897 	if (refcount > dbufs) {
898 		rw_exit(&odn->dn_struct_rwlock);
899 		zrl_exit(&odn->dn_handle->dnh_zrlock);
900 		mutex_exit(&os->os_lock);
901 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
902 		return (KMEM_CBRC_LATER);
903 	}
904 
905 	rw_exit(&odn->dn_struct_rwlock);
906 
907 	/*
908 	 * At this point we know that anyone with a hold on the dnode is not
909 	 * actively referencing it. The dnode is known and in a valid state to
910 	 * move. We're holding the locks needed to execute the critical section.
911 	 */
912 	dnode_move_impl(odn, ndn);
913 
914 	list_link_replace(&odn->dn_link, &ndn->dn_link);
915 	/* If the dnode was safe to move, the refcount cannot have changed. */
916 	ASSERT(refcount == refcount_count(&ndn->dn_holds));
917 	ASSERT(dbufs == ndn->dn_dbufs_count);
918 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
919 	mutex_exit(&os->os_lock);
920 
921 	return (KMEM_CBRC_YES);
922 }
923 #endif	/* _KERNEL */
924 
925 void
926 dnode_special_close(dnode_handle_t *dnh)
927 {
928 	dnode_t *dn = dnh->dnh_dnode;
929 
930 	/*
931 	 * Wait for final references to the dnode to clear.  This can
932 	 * only happen if the arc is asyncronously evicting state that
933 	 * has a hold on this dnode while we are trying to evict this
934 	 * dnode.
935 	 */
936 	while (refcount_count(&dn->dn_holds) > 0)
937 		delay(1);
938 	zrl_add(&dnh->dnh_zrlock);
939 	dnode_destroy(dn); /* implicit zrl_remove() */
940 	zrl_destroy(&dnh->dnh_zrlock);
941 	dnh->dnh_dnode = NULL;
942 }
943 
944 dnode_t *
945 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
946     dnode_handle_t *dnh)
947 {
948 	dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
949 	dnh->dnh_dnode = dn;
950 	zrl_init(&dnh->dnh_zrlock);
951 	DNODE_VERIFY(dn);
952 	return (dn);
953 }
954 
955 static void
956 dnode_buf_pageout(dmu_buf_t *db, void *arg)
957 {
958 	dnode_children_t *children_dnodes = arg;
959 	int i;
960 	int epb = db->db_size >> DNODE_SHIFT;
961 
962 	ASSERT(epb == children_dnodes->dnc_count);
963 
964 	for (i = 0; i < epb; i++) {
965 		dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
966 		dnode_t *dn;
967 
968 		/*
969 		 * The dnode handle lock guards against the dnode moving to
970 		 * another valid address, so there is no need here to guard
971 		 * against changes to or from NULL.
972 		 */
973 		if (dnh->dnh_dnode == NULL) {
974 			zrl_destroy(&dnh->dnh_zrlock);
975 			continue;
976 		}
977 
978 		zrl_add(&dnh->dnh_zrlock);
979 		dn = dnh->dnh_dnode;
980 		/*
981 		 * If there are holds on this dnode, then there should
982 		 * be holds on the dnode's containing dbuf as well; thus
983 		 * it wouldn't be eligible for eviction and this function
984 		 * would not have been called.
985 		 */
986 		ASSERT(refcount_is_zero(&dn->dn_holds));
987 		ASSERT(refcount_is_zero(&dn->dn_tx_holds));
988 
989 		dnode_destroy(dn); /* implicit zrl_remove() */
990 		zrl_destroy(&dnh->dnh_zrlock);
991 		dnh->dnh_dnode = NULL;
992 	}
993 	kmem_free(children_dnodes, sizeof (dnode_children_t) +
994 	    (epb - 1) * sizeof (dnode_handle_t));
995 }
996 
997 /*
998  * errors:
999  * EINVAL - invalid object number.
1000  * EIO - i/o error.
1001  * succeeds even for free dnodes.
1002  */
1003 int
1004 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1005     void *tag, dnode_t **dnp)
1006 {
1007 	int epb, idx, err;
1008 	int drop_struct_lock = FALSE;
1009 	int type;
1010 	uint64_t blk;
1011 	dnode_t *mdn, *dn;
1012 	dmu_buf_impl_t *db;
1013 	dnode_children_t *children_dnodes;
1014 	dnode_handle_t *dnh;
1015 
1016 	/*
1017 	 * If you are holding the spa config lock as writer, you shouldn't
1018 	 * be asking the DMU to do *anything* unless it's the root pool
1019 	 * which may require us to read from the root filesystem while
1020 	 * holding some (not all) of the locks as writer.
1021 	 */
1022 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1023 	    (spa_is_root(os->os_spa) &&
1024 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1025 
1026 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1027 		dn = (object == DMU_USERUSED_OBJECT) ?
1028 		    DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1029 		if (dn == NULL)
1030 			return (ENOENT);
1031 		type = dn->dn_type;
1032 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1033 			return (ENOENT);
1034 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1035 			return (EEXIST);
1036 		DNODE_VERIFY(dn);
1037 		(void) refcount_add(&dn->dn_holds, tag);
1038 		*dnp = dn;
1039 		return (0);
1040 	}
1041 
1042 	if (object == 0 || object >= DN_MAX_OBJECT)
1043 		return (EINVAL);
1044 
1045 	mdn = DMU_META_DNODE(os);
1046 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1047 
1048 	DNODE_VERIFY(mdn);
1049 
1050 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1051 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1052 		drop_struct_lock = TRUE;
1053 	}
1054 
1055 	blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1056 
1057 	db = dbuf_hold(mdn, blk, FTAG);
1058 	if (drop_struct_lock)
1059 		rw_exit(&mdn->dn_struct_rwlock);
1060 	if (db == NULL)
1061 		return (EIO);
1062 	err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1063 	if (err) {
1064 		dbuf_rele(db, FTAG);
1065 		return (err);
1066 	}
1067 
1068 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1069 	epb = db->db.db_size >> DNODE_SHIFT;
1070 
1071 	idx = object & (epb-1);
1072 
1073 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1074 	children_dnodes = dmu_buf_get_user(&db->db);
1075 	if (children_dnodes == NULL) {
1076 		int i;
1077 		dnode_children_t *winner;
1078 		children_dnodes = kmem_alloc(sizeof (dnode_children_t) +
1079 		    (epb - 1) * sizeof (dnode_handle_t), KM_SLEEP);
1080 		children_dnodes->dnc_count = epb;
1081 		dnh = &children_dnodes->dnc_children[0];
1082 		for (i = 0; i < epb; i++) {
1083 			zrl_init(&dnh[i].dnh_zrlock);
1084 			dnh[i].dnh_dnode = NULL;
1085 		}
1086 		if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1087 		    dnode_buf_pageout)) {
1088 			kmem_free(children_dnodes, sizeof (dnode_children_t) +
1089 			    (epb - 1) * sizeof (dnode_handle_t));
1090 			children_dnodes = winner;
1091 		}
1092 	}
1093 	ASSERT(children_dnodes->dnc_count == epb);
1094 
1095 	dnh = &children_dnodes->dnc_children[idx];
1096 	zrl_add(&dnh->dnh_zrlock);
1097 	if ((dn = dnh->dnh_dnode) == NULL) {
1098 		dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1099 		dnode_t *winner;
1100 
1101 		dn = dnode_create(os, phys, db, object, dnh);
1102 		winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1103 		if (winner != NULL) {
1104 			zrl_add(&dnh->dnh_zrlock);
1105 			dnode_destroy(dn); /* implicit zrl_remove() */
1106 			dn = winner;
1107 		}
1108 	}
1109 
1110 	mutex_enter(&dn->dn_mtx);
1111 	type = dn->dn_type;
1112 	if (dn->dn_free_txg ||
1113 	    ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1114 	    ((flag & DNODE_MUST_BE_FREE) &&
1115 	    (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1116 		mutex_exit(&dn->dn_mtx);
1117 		zrl_remove(&dnh->dnh_zrlock);
1118 		dbuf_rele(db, FTAG);
1119 		return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1120 	}
1121 	mutex_exit(&dn->dn_mtx);
1122 
1123 	if (refcount_add(&dn->dn_holds, tag) == 1)
1124 		dbuf_add_ref(db, dnh);
1125 	/* Now we can rely on the hold to prevent the dnode from moving. */
1126 	zrl_remove(&dnh->dnh_zrlock);
1127 
1128 	DNODE_VERIFY(dn);
1129 	ASSERT3P(dn->dn_dbuf, ==, db);
1130 	ASSERT3U(dn->dn_object, ==, object);
1131 	dbuf_rele(db, FTAG);
1132 
1133 	*dnp = dn;
1134 	return (0);
1135 }
1136 
1137 /*
1138  * Return held dnode if the object is allocated, NULL if not.
1139  */
1140 int
1141 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1142 {
1143 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1144 }
1145 
1146 /*
1147  * Can only add a reference if there is already at least one
1148  * reference on the dnode.  Returns FALSE if unable to add a
1149  * new reference.
1150  */
1151 boolean_t
1152 dnode_add_ref(dnode_t *dn, void *tag)
1153 {
1154 	mutex_enter(&dn->dn_mtx);
1155 	if (refcount_is_zero(&dn->dn_holds)) {
1156 		mutex_exit(&dn->dn_mtx);
1157 		return (FALSE);
1158 	}
1159 	VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1160 	mutex_exit(&dn->dn_mtx);
1161 	return (TRUE);
1162 }
1163 
1164 void
1165 dnode_rele(dnode_t *dn, void *tag)
1166 {
1167 	uint64_t refs;
1168 	/* Get while the hold prevents the dnode from moving. */
1169 	dmu_buf_impl_t *db = dn->dn_dbuf;
1170 	dnode_handle_t *dnh = dn->dn_handle;
1171 
1172 	mutex_enter(&dn->dn_mtx);
1173 	refs = refcount_remove(&dn->dn_holds, tag);
1174 	mutex_exit(&dn->dn_mtx);
1175 
1176 	/*
1177 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1178 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1179 	 * prevent the dnode from moving, since releasing the last hold could
1180 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1181 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1182 	 * other direct or indirect hold on the dnode must first drop the dnode
1183 	 * handle.
1184 	 */
1185 	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1186 
1187 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1188 	if (refs == 0 && db != NULL) {
1189 		/*
1190 		 * Another thread could add a hold to the dnode handle in
1191 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1192 		 * hold on the parent dbuf prevents the handle from being
1193 		 * destroyed, the hold on the handle is OK. We can't yet assert
1194 		 * that the handle has zero references, but that will be
1195 		 * asserted anyway when the handle gets destroyed.
1196 		 */
1197 		dbuf_rele(db, dnh);
1198 	}
1199 }
1200 
1201 void
1202 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1203 {
1204 	objset_t *os = dn->dn_objset;
1205 	uint64_t txg = tx->tx_txg;
1206 
1207 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1208 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1209 		return;
1210 	}
1211 
1212 	DNODE_VERIFY(dn);
1213 
1214 #ifdef ZFS_DEBUG
1215 	mutex_enter(&dn->dn_mtx);
1216 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1217 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1218 	mutex_exit(&dn->dn_mtx);
1219 #endif
1220 
1221 	/*
1222 	 * Determine old uid/gid when necessary
1223 	 */
1224 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1225 
1226 	mutex_enter(&os->os_lock);
1227 
1228 	/*
1229 	 * If we are already marked dirty, we're done.
1230 	 */
1231 	if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1232 		mutex_exit(&os->os_lock);
1233 		return;
1234 	}
1235 
1236 	ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
1237 	ASSERT(dn->dn_datablksz != 0);
1238 	ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1239 	ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1240 	ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1241 
1242 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1243 	    dn->dn_object, txg);
1244 
1245 	if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1246 		list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1247 	} else {
1248 		list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1249 	}
1250 
1251 	mutex_exit(&os->os_lock);
1252 
1253 	/*
1254 	 * The dnode maintains a hold on its containing dbuf as
1255 	 * long as there are holds on it.  Each instantiated child
1256 	 * dbuf maintains a hold on the dnode.  When the last child
1257 	 * drops its hold, the dnode will drop its hold on the
1258 	 * containing dbuf. We add a "dirty hold" here so that the
1259 	 * dnode will hang around after we finish processing its
1260 	 * children.
1261 	 */
1262 	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1263 
1264 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1265 
1266 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1267 }
1268 
1269 void
1270 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1271 {
1272 	int txgoff = tx->tx_txg & TXG_MASK;
1273 
1274 	dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1275 
1276 	/* we should be the only holder... hopefully */
1277 	/* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1278 
1279 	mutex_enter(&dn->dn_mtx);
1280 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1281 		mutex_exit(&dn->dn_mtx);
1282 		return;
1283 	}
1284 	dn->dn_free_txg = tx->tx_txg;
1285 	mutex_exit(&dn->dn_mtx);
1286 
1287 	/*
1288 	 * If the dnode is already dirty, it needs to be moved from
1289 	 * the dirty list to the free list.
1290 	 */
1291 	mutex_enter(&dn->dn_objset->os_lock);
1292 	if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1293 		list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1294 		list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1295 		mutex_exit(&dn->dn_objset->os_lock);
1296 	} else {
1297 		mutex_exit(&dn->dn_objset->os_lock);
1298 		dnode_setdirty(dn, tx);
1299 	}
1300 }
1301 
1302 /*
1303  * Try to change the block size for the indicated dnode.  This can only
1304  * succeed if there are no blocks allocated or dirty beyond first block
1305  */
1306 int
1307 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1308 {
1309 	dmu_buf_impl_t *db, *db_next;
1310 	int err;
1311 
1312 	if (size == 0)
1313 		size = SPA_MINBLOCKSIZE;
1314 	if (size > SPA_MAXBLOCKSIZE)
1315 		size = SPA_MAXBLOCKSIZE;
1316 	else
1317 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1318 
1319 	if (ibs == dn->dn_indblkshift)
1320 		ibs = 0;
1321 
1322 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1323 		return (0);
1324 
1325 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1326 
1327 	/* Check for any allocated blocks beyond the first */
1328 	if (dn->dn_phys->dn_maxblkid != 0)
1329 		goto fail;
1330 
1331 	mutex_enter(&dn->dn_dbufs_mtx);
1332 	for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
1333 		db_next = list_next(&dn->dn_dbufs, db);
1334 
1335 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1336 		    db->db_blkid != DMU_SPILL_BLKID) {
1337 			mutex_exit(&dn->dn_dbufs_mtx);
1338 			goto fail;
1339 		}
1340 	}
1341 	mutex_exit(&dn->dn_dbufs_mtx);
1342 
1343 	if (ibs && dn->dn_nlevels != 1)
1344 		goto fail;
1345 
1346 	/* resize the old block */
1347 	err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1348 	if (err == 0)
1349 		dbuf_new_size(db, size, tx);
1350 	else if (err != ENOENT)
1351 		goto fail;
1352 
1353 	dnode_setdblksz(dn, size);
1354 	dnode_setdirty(dn, tx);
1355 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1356 	if (ibs) {
1357 		dn->dn_indblkshift = ibs;
1358 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1359 	}
1360 	/* rele after we have fixed the blocksize in the dnode */
1361 	if (db)
1362 		dbuf_rele(db, FTAG);
1363 
1364 	rw_exit(&dn->dn_struct_rwlock);
1365 	return (0);
1366 
1367 fail:
1368 	rw_exit(&dn->dn_struct_rwlock);
1369 	return (ENOTSUP);
1370 }
1371 
1372 /* read-holding callers must not rely on the lock being continuously held */
1373 void
1374 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1375 {
1376 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1377 	int epbs, new_nlevels;
1378 	uint64_t sz;
1379 
1380 	ASSERT(blkid != DMU_BONUS_BLKID);
1381 
1382 	ASSERT(have_read ?
1383 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
1384 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
1385 
1386 	/*
1387 	 * if we have a read-lock, check to see if we need to do any work
1388 	 * before upgrading to a write-lock.
1389 	 */
1390 	if (have_read) {
1391 		if (blkid <= dn->dn_maxblkid)
1392 			return;
1393 
1394 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1395 			rw_exit(&dn->dn_struct_rwlock);
1396 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1397 		}
1398 	}
1399 
1400 	if (blkid <= dn->dn_maxblkid)
1401 		goto out;
1402 
1403 	dn->dn_maxblkid = blkid;
1404 
1405 	/*
1406 	 * Compute the number of levels necessary to support the new maxblkid.
1407 	 */
1408 	new_nlevels = 1;
1409 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1410 	for (sz = dn->dn_nblkptr;
1411 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1412 		new_nlevels++;
1413 
1414 	if (new_nlevels > dn->dn_nlevels) {
1415 		int old_nlevels = dn->dn_nlevels;
1416 		dmu_buf_impl_t *db;
1417 		list_t *list;
1418 		dbuf_dirty_record_t *new, *dr, *dr_next;
1419 
1420 		dn->dn_nlevels = new_nlevels;
1421 
1422 		ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1423 		dn->dn_next_nlevels[txgoff] = new_nlevels;
1424 
1425 		/* dirty the left indirects */
1426 		db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1427 		ASSERT(db != NULL);
1428 		new = dbuf_dirty(db, tx);
1429 		dbuf_rele(db, FTAG);
1430 
1431 		/* transfer the dirty records to the new indirect */
1432 		mutex_enter(&dn->dn_mtx);
1433 		mutex_enter(&new->dt.di.dr_mtx);
1434 		list = &dn->dn_dirty_records[txgoff];
1435 		for (dr = list_head(list); dr; dr = dr_next) {
1436 			dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1437 			if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1438 			    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1439 			    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1440 				ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1441 				list_remove(&dn->dn_dirty_records[txgoff], dr);
1442 				list_insert_tail(&new->dt.di.dr_children, dr);
1443 				dr->dr_parent = new;
1444 			}
1445 		}
1446 		mutex_exit(&new->dt.di.dr_mtx);
1447 		mutex_exit(&dn->dn_mtx);
1448 	}
1449 
1450 out:
1451 	if (have_read)
1452 		rw_downgrade(&dn->dn_struct_rwlock);
1453 }
1454 
1455 void
1456 dnode_clear_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
1457 {
1458 	avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1459 	avl_index_t where;
1460 	free_range_t *rp;
1461 	free_range_t rp_tofind;
1462 	uint64_t endblk = blkid + nblks;
1463 
1464 	ASSERT(MUTEX_HELD(&dn->dn_mtx));
1465 	ASSERT(nblks <= UINT64_MAX - blkid); /* no overflow */
1466 
1467 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1468 	    blkid, nblks, tx->tx_txg);
1469 	rp_tofind.fr_blkid = blkid;
1470 	rp = avl_find(tree, &rp_tofind, &where);
1471 	if (rp == NULL)
1472 		rp = avl_nearest(tree, where, AVL_BEFORE);
1473 	if (rp == NULL)
1474 		rp = avl_nearest(tree, where, AVL_AFTER);
1475 
1476 	while (rp && (rp->fr_blkid <= blkid + nblks)) {
1477 		uint64_t fr_endblk = rp->fr_blkid + rp->fr_nblks;
1478 		free_range_t *nrp = AVL_NEXT(tree, rp);
1479 
1480 		if (blkid <= rp->fr_blkid && endblk >= fr_endblk) {
1481 			/* clear this entire range */
1482 			avl_remove(tree, rp);
1483 			kmem_free(rp, sizeof (free_range_t));
1484 		} else if (blkid <= rp->fr_blkid &&
1485 		    endblk > rp->fr_blkid && endblk < fr_endblk) {
1486 			/* clear the beginning of this range */
1487 			rp->fr_blkid = endblk;
1488 			rp->fr_nblks = fr_endblk - endblk;
1489 		} else if (blkid > rp->fr_blkid && blkid < fr_endblk &&
1490 		    endblk >= fr_endblk) {
1491 			/* clear the end of this range */
1492 			rp->fr_nblks = blkid - rp->fr_blkid;
1493 		} else if (blkid > rp->fr_blkid && endblk < fr_endblk) {
1494 			/* clear a chunk out of this range */
1495 			free_range_t *new_rp =
1496 			    kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1497 
1498 			new_rp->fr_blkid = endblk;
1499 			new_rp->fr_nblks = fr_endblk - endblk;
1500 			avl_insert_here(tree, new_rp, rp, AVL_AFTER);
1501 			rp->fr_nblks = blkid - rp->fr_blkid;
1502 		}
1503 		/* there may be no overlap */
1504 		rp = nrp;
1505 	}
1506 }
1507 
1508 void
1509 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1510 {
1511 	dmu_buf_impl_t *db;
1512 	uint64_t blkoff, blkid, nblks;
1513 	int blksz, blkshift, head, tail;
1514 	int trunc = FALSE;
1515 	int epbs;
1516 
1517 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1518 	blksz = dn->dn_datablksz;
1519 	blkshift = dn->dn_datablkshift;
1520 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1521 
1522 	if (len == -1ULL) {
1523 		len = UINT64_MAX - off;
1524 		trunc = TRUE;
1525 	}
1526 
1527 	/*
1528 	 * First, block align the region to free:
1529 	 */
1530 	if (ISP2(blksz)) {
1531 		head = P2NPHASE(off, blksz);
1532 		blkoff = P2PHASE(off, blksz);
1533 		if ((off >> blkshift) > dn->dn_maxblkid)
1534 			goto out;
1535 	} else {
1536 		ASSERT(dn->dn_maxblkid == 0);
1537 		if (off == 0 && len >= blksz) {
1538 			/* Freeing the whole block; fast-track this request */
1539 			blkid = 0;
1540 			nblks = 1;
1541 			goto done;
1542 		} else if (off >= blksz) {
1543 			/* Freeing past end-of-data */
1544 			goto out;
1545 		} else {
1546 			/* Freeing part of the block. */
1547 			head = blksz - off;
1548 			ASSERT3U(head, >, 0);
1549 		}
1550 		blkoff = off;
1551 	}
1552 	/* zero out any partial block data at the start of the range */
1553 	if (head) {
1554 		ASSERT3U(blkoff + head, ==, blksz);
1555 		if (len < head)
1556 			head = len;
1557 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1558 		    FTAG, &db) == 0) {
1559 			caddr_t data;
1560 
1561 			/* don't dirty if it isn't on disk and isn't dirty */
1562 			if (db->db_last_dirty ||
1563 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1564 				rw_exit(&dn->dn_struct_rwlock);
1565 				dbuf_will_dirty(db, tx);
1566 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1567 				data = db->db.db_data;
1568 				bzero(data + blkoff, head);
1569 			}
1570 			dbuf_rele(db, FTAG);
1571 		}
1572 		off += head;
1573 		len -= head;
1574 	}
1575 
1576 	/* If the range was less than one block, we're done */
1577 	if (len == 0)
1578 		goto out;
1579 
1580 	/* If the remaining range is past end of file, we're done */
1581 	if ((off >> blkshift) > dn->dn_maxblkid)
1582 		goto out;
1583 
1584 	ASSERT(ISP2(blksz));
1585 	if (trunc)
1586 		tail = 0;
1587 	else
1588 		tail = P2PHASE(len, blksz);
1589 
1590 	ASSERT0(P2PHASE(off, blksz));
1591 	/* zero out any partial block data at the end of the range */
1592 	if (tail) {
1593 		if (len < tail)
1594 			tail = len;
1595 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1596 		    TRUE, FTAG, &db) == 0) {
1597 			/* don't dirty if not on disk and not dirty */
1598 			if (db->db_last_dirty ||
1599 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1600 				rw_exit(&dn->dn_struct_rwlock);
1601 				dbuf_will_dirty(db, tx);
1602 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1603 				bzero(db->db.db_data, tail);
1604 			}
1605 			dbuf_rele(db, FTAG);
1606 		}
1607 		len -= tail;
1608 	}
1609 
1610 	/* If the range did not include a full block, we are done */
1611 	if (len == 0)
1612 		goto out;
1613 
1614 	ASSERT(IS_P2ALIGNED(off, blksz));
1615 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1616 	blkid = off >> blkshift;
1617 	nblks = len >> blkshift;
1618 	if (trunc)
1619 		nblks += 1;
1620 
1621 	/*
1622 	 * Read in and mark all the level-1 indirects dirty,
1623 	 * so that they will stay in memory until syncing phase.
1624 	 * Always dirty the first and last indirect to make sure
1625 	 * we dirty all the partial indirects.
1626 	 */
1627 	if (dn->dn_nlevels > 1) {
1628 		uint64_t i, first, last;
1629 		int shift = epbs + dn->dn_datablkshift;
1630 
1631 		first = blkid >> epbs;
1632 		if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1633 			dbuf_will_dirty(db, tx);
1634 			dbuf_rele(db, FTAG);
1635 		}
1636 		if (trunc)
1637 			last = dn->dn_maxblkid >> epbs;
1638 		else
1639 			last = (blkid + nblks - 1) >> epbs;
1640 		if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1641 			dbuf_will_dirty(db, tx);
1642 			dbuf_rele(db, FTAG);
1643 		}
1644 		for (i = first + 1; i < last; i++) {
1645 			uint64_t ibyte = i << shift;
1646 			int err;
1647 
1648 			err = dnode_next_offset(dn,
1649 			    DNODE_FIND_HAVELOCK, &ibyte, 1, 1, 0);
1650 			i = ibyte >> shift;
1651 			if (err == ESRCH || i >= last)
1652 				break;
1653 			ASSERT(err == 0);
1654 			db = dbuf_hold_level(dn, 1, i, FTAG);
1655 			if (db) {
1656 				dbuf_will_dirty(db, tx);
1657 				dbuf_rele(db, FTAG);
1658 			}
1659 		}
1660 	}
1661 done:
1662 	/*
1663 	 * Add this range to the dnode range list.
1664 	 * We will finish up this free operation in the syncing phase.
1665 	 */
1666 	mutex_enter(&dn->dn_mtx);
1667 	dnode_clear_range(dn, blkid, nblks, tx);
1668 	{
1669 		free_range_t *rp, *found;
1670 		avl_index_t where;
1671 		avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1672 
1673 		/* Add new range to dn_ranges */
1674 		rp = kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1675 		rp->fr_blkid = blkid;
1676 		rp->fr_nblks = nblks;
1677 		found = avl_find(tree, rp, &where);
1678 		ASSERT(found == NULL);
1679 		avl_insert(tree, rp, where);
1680 		dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1681 		    blkid, nblks, tx->tx_txg);
1682 	}
1683 	mutex_exit(&dn->dn_mtx);
1684 
1685 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1686 	dnode_setdirty(dn, tx);
1687 out:
1688 	if (trunc && dn->dn_maxblkid >= (off >> blkshift))
1689 		dn->dn_maxblkid = (off >> blkshift ? (off >> blkshift) - 1 : 0);
1690 
1691 	rw_exit(&dn->dn_struct_rwlock);
1692 }
1693 
1694 static boolean_t
1695 dnode_spill_freed(dnode_t *dn)
1696 {
1697 	int i;
1698 
1699 	mutex_enter(&dn->dn_mtx);
1700 	for (i = 0; i < TXG_SIZE; i++) {
1701 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1702 			break;
1703 	}
1704 	mutex_exit(&dn->dn_mtx);
1705 	return (i < TXG_SIZE);
1706 }
1707 
1708 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1709 uint64_t
1710 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1711 {
1712 	free_range_t range_tofind;
1713 	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1714 	int i;
1715 
1716 	if (blkid == DMU_BONUS_BLKID)
1717 		return (FALSE);
1718 
1719 	/*
1720 	 * If we're in the process of opening the pool, dp will not be
1721 	 * set yet, but there shouldn't be anything dirty.
1722 	 */
1723 	if (dp == NULL)
1724 		return (FALSE);
1725 
1726 	if (dn->dn_free_txg)
1727 		return (TRUE);
1728 
1729 	if (blkid == DMU_SPILL_BLKID)
1730 		return (dnode_spill_freed(dn));
1731 
1732 	range_tofind.fr_blkid = blkid;
1733 	mutex_enter(&dn->dn_mtx);
1734 	for (i = 0; i < TXG_SIZE; i++) {
1735 		free_range_t *range_found;
1736 		avl_index_t idx;
1737 
1738 		range_found = avl_find(&dn->dn_ranges[i], &range_tofind, &idx);
1739 		if (range_found) {
1740 			ASSERT(range_found->fr_nblks > 0);
1741 			break;
1742 		}
1743 		range_found = avl_nearest(&dn->dn_ranges[i], idx, AVL_BEFORE);
1744 		if (range_found &&
1745 		    range_found->fr_blkid + range_found->fr_nblks > blkid)
1746 			break;
1747 	}
1748 	mutex_exit(&dn->dn_mtx);
1749 	return (i < TXG_SIZE);
1750 }
1751 
1752 /* call from syncing context when we actually write/free space for this dnode */
1753 void
1754 dnode_diduse_space(dnode_t *dn, int64_t delta)
1755 {
1756 	uint64_t space;
1757 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1758 	    dn, dn->dn_phys,
1759 	    (u_longlong_t)dn->dn_phys->dn_used,
1760 	    (longlong_t)delta);
1761 
1762 	mutex_enter(&dn->dn_mtx);
1763 	space = DN_USED_BYTES(dn->dn_phys);
1764 	if (delta > 0) {
1765 		ASSERT3U(space + delta, >=, space); /* no overflow */
1766 	} else {
1767 		ASSERT3U(space, >=, -delta); /* no underflow */
1768 	}
1769 	space += delta;
1770 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1771 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1772 		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1773 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1774 	} else {
1775 		dn->dn_phys->dn_used = space;
1776 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1777 	}
1778 	mutex_exit(&dn->dn_mtx);
1779 }
1780 
1781 /*
1782  * Call when we think we're going to write/free space in open context.
1783  * Be conservative (ie. OK to write less than this or free more than
1784  * this, but don't write more or free less).
1785  */
1786 void
1787 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1788 {
1789 	objset_t *os = dn->dn_objset;
1790 	dsl_dataset_t *ds = os->os_dsl_dataset;
1791 
1792 	if (space > 0)
1793 		space = spa_get_asize(os->os_spa, space);
1794 
1795 	if (ds)
1796 		dsl_dir_willuse_space(ds->ds_dir, space, tx);
1797 
1798 	dmu_tx_willuse_space(tx, space);
1799 }
1800 
1801 /*
1802  * This function scans a block at the indicated "level" looking for
1803  * a hole or data (depending on 'flags').  If level > 0, then we are
1804  * scanning an indirect block looking at its pointers.  If level == 0,
1805  * then we are looking at a block of dnodes.  If we don't find what we
1806  * are looking for in the block, we return ESRCH.  Otherwise, return
1807  * with *offset pointing to the beginning (if searching forwards) or
1808  * end (if searching backwards) of the range covered by the block
1809  * pointer we matched on (or dnode).
1810  *
1811  * The basic search algorithm used below by dnode_next_offset() is to
1812  * use this function to search up the block tree (widen the search) until
1813  * we find something (i.e., we don't return ESRCH) and then search back
1814  * down the tree (narrow the search) until we reach our original search
1815  * level.
1816  */
1817 static int
1818 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1819 	int lvl, uint64_t blkfill, uint64_t txg)
1820 {
1821 	dmu_buf_impl_t *db = NULL;
1822 	void *data = NULL;
1823 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1824 	uint64_t epb = 1ULL << epbs;
1825 	uint64_t minfill, maxfill;
1826 	boolean_t hole;
1827 	int i, inc, error, span;
1828 
1829 	dprintf("probing object %llu offset %llx level %d of %u\n",
1830 	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1831 
1832 	hole = ((flags & DNODE_FIND_HOLE) != 0);
1833 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1834 	ASSERT(txg == 0 || !hole);
1835 
1836 	if (lvl == dn->dn_phys->dn_nlevels) {
1837 		error = 0;
1838 		epb = dn->dn_phys->dn_nblkptr;
1839 		data = dn->dn_phys->dn_blkptr;
1840 	} else {
1841 		uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1842 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1843 		if (error) {
1844 			if (error != ENOENT)
1845 				return (error);
1846 			if (hole)
1847 				return (0);
1848 			/*
1849 			 * This can only happen when we are searching up
1850 			 * the block tree for data.  We don't really need to
1851 			 * adjust the offset, as we will just end up looking
1852 			 * at the pointer to this block in its parent, and its
1853 			 * going to be unallocated, so we will skip over it.
1854 			 */
1855 			return (ESRCH);
1856 		}
1857 		error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1858 		if (error) {
1859 			dbuf_rele(db, FTAG);
1860 			return (error);
1861 		}
1862 		data = db->db.db_data;
1863 	}
1864 
1865 	if (db && txg &&
1866 	    (db->db_blkptr == NULL || db->db_blkptr->blk_birth <= txg)) {
1867 		/*
1868 		 * This can only happen when we are searching up the tree
1869 		 * and these conditions mean that we need to keep climbing.
1870 		 */
1871 		error = ESRCH;
1872 	} else if (lvl == 0) {
1873 		dnode_phys_t *dnp = data;
1874 		span = DNODE_SHIFT;
1875 		ASSERT(dn->dn_type == DMU_OT_DNODE);
1876 
1877 		for (i = (*offset >> span) & (blkfill - 1);
1878 		    i >= 0 && i < blkfill; i += inc) {
1879 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1880 				break;
1881 			*offset += (1ULL << span) * inc;
1882 		}
1883 		if (i < 0 || i == blkfill)
1884 			error = ESRCH;
1885 	} else {
1886 		blkptr_t *bp = data;
1887 		uint64_t start = *offset;
1888 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
1889 		minfill = 0;
1890 		maxfill = blkfill << ((lvl - 1) * epbs);
1891 
1892 		if (hole)
1893 			maxfill--;
1894 		else
1895 			minfill++;
1896 
1897 		*offset = *offset >> span;
1898 		for (i = BF64_GET(*offset, 0, epbs);
1899 		    i >= 0 && i < epb; i += inc) {
1900 			if (bp[i].blk_fill >= minfill &&
1901 			    bp[i].blk_fill <= maxfill &&
1902 			    (hole || bp[i].blk_birth > txg))
1903 				break;
1904 			if (inc > 0 || *offset > 0)
1905 				*offset += inc;
1906 		}
1907 		*offset = *offset << span;
1908 		if (inc < 0) {
1909 			/* traversing backwards; position offset at the end */
1910 			ASSERT3U(*offset, <=, start);
1911 			*offset = MIN(*offset + (1ULL << span) - 1, start);
1912 		} else if (*offset < start) {
1913 			*offset = start;
1914 		}
1915 		if (i < 0 || i >= epb)
1916 			error = ESRCH;
1917 	}
1918 
1919 	if (db)
1920 		dbuf_rele(db, FTAG);
1921 
1922 	return (error);
1923 }
1924 
1925 /*
1926  * Find the next hole, data, or sparse region at or after *offset.
1927  * The value 'blkfill' tells us how many items we expect to find
1928  * in an L0 data block; this value is 1 for normal objects,
1929  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1930  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1931  *
1932  * Examples:
1933  *
1934  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1935  *	Finds the next/previous hole/data in a file.
1936  *	Used in dmu_offset_next().
1937  *
1938  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1939  *	Finds the next free/allocated dnode an objset's meta-dnode.
1940  *	Only finds objects that have new contents since txg (ie.
1941  *	bonus buffer changes and content removal are ignored).
1942  *	Used in dmu_object_next().
1943  *
1944  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1945  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
1946  *	Used in dmu_object_alloc().
1947  */
1948 int
1949 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1950     int minlvl, uint64_t blkfill, uint64_t txg)
1951 {
1952 	uint64_t initial_offset = *offset;
1953 	int lvl, maxlvl;
1954 	int error = 0;
1955 
1956 	if (!(flags & DNODE_FIND_HAVELOCK))
1957 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1958 
1959 	if (dn->dn_phys->dn_nlevels == 0) {
1960 		error = ESRCH;
1961 		goto out;
1962 	}
1963 
1964 	if (dn->dn_datablkshift == 0) {
1965 		if (*offset < dn->dn_datablksz) {
1966 			if (flags & DNODE_FIND_HOLE)
1967 				*offset = dn->dn_datablksz;
1968 		} else {
1969 			error = ESRCH;
1970 		}
1971 		goto out;
1972 	}
1973 
1974 	maxlvl = dn->dn_phys->dn_nlevels;
1975 
1976 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1977 		error = dnode_next_offset_level(dn,
1978 		    flags, offset, lvl, blkfill, txg);
1979 		if (error != ESRCH)
1980 			break;
1981 	}
1982 
1983 	while (error == 0 && --lvl >= minlvl) {
1984 		error = dnode_next_offset_level(dn,
1985 		    flags, offset, lvl, blkfill, txg);
1986 	}
1987 
1988 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1989 	    initial_offset < *offset : initial_offset > *offset))
1990 		error = ESRCH;
1991 out:
1992 	if (!(flags & DNODE_FIND_HAVELOCK))
1993 		rw_exit(&dn->dn_struct_rwlock);
1994 
1995 	return (error);
1996 }
1997