xref: /titanic_51/usr/src/uts/common/fs/zfs/dnode.c (revision e6546372a86065934a1b311176ecd1b1c543a222)
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, 2015 by Delphix. All rights reserved.
24  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25  * Copyright (c) 2014 Integros [integros.com]
26  */
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 #include <sys/range_tree.h>
41 
42 static kmem_cache_t *dnode_cache;
43 /*
44  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
45  * turned on when DEBUG is also defined.
46  */
47 #ifdef	DEBUG
48 #define	DNODE_STATS
49 #endif	/* DEBUG */
50 
51 #ifdef	DNODE_STATS
52 #define	DNODE_STAT_ADD(stat)			((stat)++)
53 #else
54 #define	DNODE_STAT_ADD(stat)			/* nothing */
55 #endif	/* DNODE_STATS */
56 
57 static dnode_phys_t dnode_phys_zero;
58 
59 int zfs_default_bs = SPA_MINBLOCKSHIFT;
60 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
61 
62 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
63 
64 static int
65 dbuf_compare(const void *x1, const void *x2)
66 {
67 	const dmu_buf_impl_t *d1 = x1;
68 	const dmu_buf_impl_t *d2 = x2;
69 
70 	if (d1->db_level < d2->db_level) {
71 		return (-1);
72 	}
73 	if (d1->db_level > d2->db_level) {
74 		return (1);
75 	}
76 
77 	if (d1->db_blkid < d2->db_blkid) {
78 		return (-1);
79 	}
80 	if (d1->db_blkid > d2->db_blkid) {
81 		return (1);
82 	}
83 
84 	if (d1->db_state == DB_SEARCH) {
85 		ASSERT3S(d2->db_state, !=, DB_SEARCH);
86 		return (-1);
87 	} else if (d2->db_state == DB_SEARCH) {
88 		ASSERT3S(d1->db_state, !=, DB_SEARCH);
89 		return (1);
90 	}
91 
92 	if ((uintptr_t)d1 < (uintptr_t)d2) {
93 		return (-1);
94 	}
95 	if ((uintptr_t)d1 > (uintptr_t)d2) {
96 		return (1);
97 	}
98 	return (0);
99 }
100 
101 /* ARGSUSED */
102 static int
103 dnode_cons(void *arg, void *unused, int kmflag)
104 {
105 	dnode_t *dn = arg;
106 	int i;
107 
108 	rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
109 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
110 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
111 	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
112 
113 	/*
114 	 * Every dbuf has a reference, and dropping a tracked reference is
115 	 * O(number of references), so don't track dn_holds.
116 	 */
117 	refcount_create_untracked(&dn->dn_holds);
118 	refcount_create(&dn->dn_tx_holds);
119 	list_link_init(&dn->dn_link);
120 
121 	bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
122 	bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
123 	bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
124 	bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
125 	bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
126 	bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
127 	bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
128 
129 	for (i = 0; i < TXG_SIZE; i++) {
130 		list_link_init(&dn->dn_dirty_link[i]);
131 		dn->dn_free_ranges[i] = NULL;
132 		list_create(&dn->dn_dirty_records[i],
133 		    sizeof (dbuf_dirty_record_t),
134 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
135 	}
136 
137 	dn->dn_allocated_txg = 0;
138 	dn->dn_free_txg = 0;
139 	dn->dn_assigned_txg = 0;
140 	dn->dn_dirtyctx = 0;
141 	dn->dn_dirtyctx_firstset = NULL;
142 	dn->dn_bonus = NULL;
143 	dn->dn_have_spill = B_FALSE;
144 	dn->dn_zio = NULL;
145 	dn->dn_oldused = 0;
146 	dn->dn_oldflags = 0;
147 	dn->dn_olduid = 0;
148 	dn->dn_oldgid = 0;
149 	dn->dn_newuid = 0;
150 	dn->dn_newgid = 0;
151 	dn->dn_id_flags = 0;
152 
153 	dn->dn_dbufs_count = 0;
154 	dn->dn_unlisted_l0_blkid = 0;
155 	avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
156 	    offsetof(dmu_buf_impl_t, db_link));
157 
158 	dn->dn_moved = 0;
159 	return (0);
160 }
161 
162 /* ARGSUSED */
163 static void
164 dnode_dest(void *arg, void *unused)
165 {
166 	int i;
167 	dnode_t *dn = arg;
168 
169 	rw_destroy(&dn->dn_struct_rwlock);
170 	mutex_destroy(&dn->dn_mtx);
171 	mutex_destroy(&dn->dn_dbufs_mtx);
172 	cv_destroy(&dn->dn_notxholds);
173 	refcount_destroy(&dn->dn_holds);
174 	refcount_destroy(&dn->dn_tx_holds);
175 	ASSERT(!list_link_active(&dn->dn_link));
176 
177 	for (i = 0; i < TXG_SIZE; i++) {
178 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
179 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
180 		list_destroy(&dn->dn_dirty_records[i]);
181 		ASSERT0(dn->dn_next_nblkptr[i]);
182 		ASSERT0(dn->dn_next_nlevels[i]);
183 		ASSERT0(dn->dn_next_indblkshift[i]);
184 		ASSERT0(dn->dn_next_bonustype[i]);
185 		ASSERT0(dn->dn_rm_spillblk[i]);
186 		ASSERT0(dn->dn_next_bonuslen[i]);
187 		ASSERT0(dn->dn_next_blksz[i]);
188 	}
189 
190 	ASSERT0(dn->dn_allocated_txg);
191 	ASSERT0(dn->dn_free_txg);
192 	ASSERT0(dn->dn_assigned_txg);
193 	ASSERT0(dn->dn_dirtyctx);
194 	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
195 	ASSERT3P(dn->dn_bonus, ==, NULL);
196 	ASSERT(!dn->dn_have_spill);
197 	ASSERT3P(dn->dn_zio, ==, NULL);
198 	ASSERT0(dn->dn_oldused);
199 	ASSERT0(dn->dn_oldflags);
200 	ASSERT0(dn->dn_olduid);
201 	ASSERT0(dn->dn_oldgid);
202 	ASSERT0(dn->dn_newuid);
203 	ASSERT0(dn->dn_newgid);
204 	ASSERT0(dn->dn_id_flags);
205 
206 	ASSERT0(dn->dn_dbufs_count);
207 	ASSERT0(dn->dn_unlisted_l0_blkid);
208 	avl_destroy(&dn->dn_dbufs);
209 }
210 
211 void
212 dnode_init(void)
213 {
214 	ASSERT(dnode_cache == NULL);
215 	dnode_cache = kmem_cache_create("dnode_t",
216 	    sizeof (dnode_t),
217 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
218 	kmem_cache_set_move(dnode_cache, dnode_move);
219 }
220 
221 void
222 dnode_fini(void)
223 {
224 	kmem_cache_destroy(dnode_cache);
225 	dnode_cache = NULL;
226 }
227 
228 
229 #ifdef ZFS_DEBUG
230 void
231 dnode_verify(dnode_t *dn)
232 {
233 	int drop_struct_lock = FALSE;
234 
235 	ASSERT(dn->dn_phys);
236 	ASSERT(dn->dn_objset);
237 	ASSERT(dn->dn_handle->dnh_dnode == dn);
238 
239 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
240 
241 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
242 		return;
243 
244 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
245 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
246 		drop_struct_lock = TRUE;
247 	}
248 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
249 		int i;
250 		ASSERT3U(dn->dn_indblkshift, >=, 0);
251 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
252 		if (dn->dn_datablkshift) {
253 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
254 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
255 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
256 		}
257 		ASSERT3U(dn->dn_nlevels, <=, 30);
258 		ASSERT(DMU_OT_IS_VALID(dn->dn_type));
259 		ASSERT3U(dn->dn_nblkptr, >=, 1);
260 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
261 		ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
262 		ASSERT3U(dn->dn_datablksz, ==,
263 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
264 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
265 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
266 		    dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
267 		for (i = 0; i < TXG_SIZE; i++) {
268 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
269 		}
270 	}
271 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
272 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
273 	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
274 	if (dn->dn_dbuf != NULL) {
275 		ASSERT3P(dn->dn_phys, ==,
276 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
277 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
278 	}
279 	if (drop_struct_lock)
280 		rw_exit(&dn->dn_struct_rwlock);
281 }
282 #endif
283 
284 void
285 dnode_byteswap(dnode_phys_t *dnp)
286 {
287 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
288 	int i;
289 
290 	if (dnp->dn_type == DMU_OT_NONE) {
291 		bzero(dnp, sizeof (dnode_phys_t));
292 		return;
293 	}
294 
295 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
296 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
297 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
298 	dnp->dn_used = BSWAP_64(dnp->dn_used);
299 
300 	/*
301 	 * dn_nblkptr is only one byte, so it's OK to read it in either
302 	 * byte order.  We can't read dn_bouslen.
303 	 */
304 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
305 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
306 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
307 		buf64[i] = BSWAP_64(buf64[i]);
308 
309 	/*
310 	 * OK to check dn_bonuslen for zero, because it won't matter if
311 	 * we have the wrong byte order.  This is necessary because the
312 	 * dnode dnode is smaller than a regular dnode.
313 	 */
314 	if (dnp->dn_bonuslen != 0) {
315 		/*
316 		 * Note that the bonus length calculated here may be
317 		 * longer than the actual bonus buffer.  This is because
318 		 * we always put the bonus buffer after the last block
319 		 * pointer (instead of packing it against the end of the
320 		 * dnode buffer).
321 		 */
322 		int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
323 		size_t len = DN_MAX_BONUSLEN - off;
324 		ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
325 		dmu_object_byteswap_t byteswap =
326 		    DMU_OT_BYTESWAP(dnp->dn_bonustype);
327 		dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
328 	}
329 
330 	/* Swap SPILL block if we have one */
331 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
332 		byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
333 
334 }
335 
336 void
337 dnode_buf_byteswap(void *vbuf, size_t size)
338 {
339 	dnode_phys_t *buf = vbuf;
340 	int i;
341 
342 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
343 	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
344 
345 	size >>= DNODE_SHIFT;
346 	for (i = 0; i < size; i++) {
347 		dnode_byteswap(buf);
348 		buf++;
349 	}
350 }
351 
352 void
353 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
354 {
355 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
356 
357 	dnode_setdirty(dn, tx);
358 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
359 	ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
360 	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
361 	dn->dn_bonuslen = newsize;
362 	if (newsize == 0)
363 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
364 	else
365 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
366 	rw_exit(&dn->dn_struct_rwlock);
367 }
368 
369 void
370 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
371 {
372 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
373 	dnode_setdirty(dn, tx);
374 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
375 	dn->dn_bonustype = newtype;
376 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
377 	rw_exit(&dn->dn_struct_rwlock);
378 }
379 
380 void
381 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
382 {
383 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
384 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
385 	dnode_setdirty(dn, tx);
386 	dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
387 	dn->dn_have_spill = B_FALSE;
388 }
389 
390 static void
391 dnode_setdblksz(dnode_t *dn, int size)
392 {
393 	ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
394 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
395 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
396 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
397 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
398 	dn->dn_datablksz = size;
399 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
400 	dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
401 }
402 
403 static dnode_t *
404 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
405     uint64_t object, dnode_handle_t *dnh)
406 {
407 	dnode_t *dn;
408 
409 	dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
410 	ASSERT(!POINTER_IS_VALID(dn->dn_objset));
411 	dn->dn_moved = 0;
412 
413 	/*
414 	 * Defer setting dn_objset until the dnode is ready to be a candidate
415 	 * for the dnode_move() callback.
416 	 */
417 	dn->dn_object = object;
418 	dn->dn_dbuf = db;
419 	dn->dn_handle = dnh;
420 	dn->dn_phys = dnp;
421 
422 	if (dnp->dn_datablkszsec) {
423 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
424 	} else {
425 		dn->dn_datablksz = 0;
426 		dn->dn_datablkszsec = 0;
427 		dn->dn_datablkshift = 0;
428 	}
429 	dn->dn_indblkshift = dnp->dn_indblkshift;
430 	dn->dn_nlevels = dnp->dn_nlevels;
431 	dn->dn_type = dnp->dn_type;
432 	dn->dn_nblkptr = dnp->dn_nblkptr;
433 	dn->dn_checksum = dnp->dn_checksum;
434 	dn->dn_compress = dnp->dn_compress;
435 	dn->dn_bonustype = dnp->dn_bonustype;
436 	dn->dn_bonuslen = dnp->dn_bonuslen;
437 	dn->dn_maxblkid = dnp->dn_maxblkid;
438 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
439 	dn->dn_id_flags = 0;
440 
441 	dmu_zfetch_init(&dn->dn_zfetch, dn);
442 
443 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
444 
445 	mutex_enter(&os->os_lock);
446 	if (dnh->dnh_dnode != NULL) {
447 		/* Lost the allocation race. */
448 		mutex_exit(&os->os_lock);
449 		kmem_cache_free(dnode_cache, dn);
450 		return (dnh->dnh_dnode);
451 	}
452 
453 	/*
454 	 * Exclude special dnodes from os_dnodes so an empty os_dnodes
455 	 * signifies that the special dnodes have no references from
456 	 * their children (the entries in os_dnodes).  This allows
457 	 * dnode_destroy() to easily determine if the last child has
458 	 * been removed and then complete eviction of the objset.
459 	 */
460 	if (!DMU_OBJECT_IS_SPECIAL(object))
461 		list_insert_head(&os->os_dnodes, dn);
462 	membar_producer();
463 
464 	/*
465 	 * Everything else must be valid before assigning dn_objset
466 	 * makes the dnode eligible for dnode_move().
467 	 */
468 	dn->dn_objset = os;
469 
470 	dnh->dnh_dnode = dn;
471 	mutex_exit(&os->os_lock);
472 
473 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
474 	return (dn);
475 }
476 
477 /*
478  * Caller must be holding the dnode handle, which is released upon return.
479  */
480 static void
481 dnode_destroy(dnode_t *dn)
482 {
483 	objset_t *os = dn->dn_objset;
484 	boolean_t complete_os_eviction = B_FALSE;
485 
486 	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
487 
488 	mutex_enter(&os->os_lock);
489 	POINTER_INVALIDATE(&dn->dn_objset);
490 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
491 		list_remove(&os->os_dnodes, dn);
492 		complete_os_eviction =
493 		    list_is_empty(&os->os_dnodes) &&
494 		    list_link_active(&os->os_evicting_node);
495 	}
496 	mutex_exit(&os->os_lock);
497 
498 	/* the dnode can no longer move, so we can release the handle */
499 	zrl_remove(&dn->dn_handle->dnh_zrlock);
500 
501 	dn->dn_allocated_txg = 0;
502 	dn->dn_free_txg = 0;
503 	dn->dn_assigned_txg = 0;
504 
505 	dn->dn_dirtyctx = 0;
506 	if (dn->dn_dirtyctx_firstset != NULL) {
507 		kmem_free(dn->dn_dirtyctx_firstset, 1);
508 		dn->dn_dirtyctx_firstset = NULL;
509 	}
510 	if (dn->dn_bonus != NULL) {
511 		mutex_enter(&dn->dn_bonus->db_mtx);
512 		dbuf_evict(dn->dn_bonus);
513 		dn->dn_bonus = NULL;
514 	}
515 	dn->dn_zio = NULL;
516 
517 	dn->dn_have_spill = B_FALSE;
518 	dn->dn_oldused = 0;
519 	dn->dn_oldflags = 0;
520 	dn->dn_olduid = 0;
521 	dn->dn_oldgid = 0;
522 	dn->dn_newuid = 0;
523 	dn->dn_newgid = 0;
524 	dn->dn_id_flags = 0;
525 	dn->dn_unlisted_l0_blkid = 0;
526 
527 	dmu_zfetch_fini(&dn->dn_zfetch);
528 	kmem_cache_free(dnode_cache, dn);
529 	arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
530 
531 	if (complete_os_eviction)
532 		dmu_objset_evict_done(os);
533 }
534 
535 void
536 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
537     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
538 {
539 	int i;
540 
541 	ASSERT3U(blocksize, <=,
542 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
543 	if (blocksize == 0)
544 		blocksize = 1 << zfs_default_bs;
545 	else
546 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
547 
548 	if (ibs == 0)
549 		ibs = zfs_default_ibs;
550 
551 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
552 
553 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
554 	    dn->dn_object, tx->tx_txg, blocksize, ibs);
555 
556 	ASSERT(dn->dn_type == DMU_OT_NONE);
557 	ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
558 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
559 	ASSERT(ot != DMU_OT_NONE);
560 	ASSERT(DMU_OT_IS_VALID(ot));
561 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
562 	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
563 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
564 	ASSERT(DMU_OT_IS_VALID(bonustype));
565 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
566 	ASSERT(dn->dn_type == DMU_OT_NONE);
567 	ASSERT0(dn->dn_maxblkid);
568 	ASSERT0(dn->dn_allocated_txg);
569 	ASSERT0(dn->dn_assigned_txg);
570 	ASSERT(refcount_is_zero(&dn->dn_tx_holds));
571 	ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
572 	ASSERT(avl_is_empty(&dn->dn_dbufs));
573 
574 	for (i = 0; i < TXG_SIZE; i++) {
575 		ASSERT0(dn->dn_next_nblkptr[i]);
576 		ASSERT0(dn->dn_next_nlevels[i]);
577 		ASSERT0(dn->dn_next_indblkshift[i]);
578 		ASSERT0(dn->dn_next_bonuslen[i]);
579 		ASSERT0(dn->dn_next_bonustype[i]);
580 		ASSERT0(dn->dn_rm_spillblk[i]);
581 		ASSERT0(dn->dn_next_blksz[i]);
582 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
583 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
584 		ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
585 	}
586 
587 	dn->dn_type = ot;
588 	dnode_setdblksz(dn, blocksize);
589 	dn->dn_indblkshift = ibs;
590 	dn->dn_nlevels = 1;
591 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
592 		dn->dn_nblkptr = 1;
593 	else
594 		dn->dn_nblkptr = 1 +
595 		    ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
596 	dn->dn_bonustype = bonustype;
597 	dn->dn_bonuslen = bonuslen;
598 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
599 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
600 	dn->dn_dirtyctx = 0;
601 
602 	dn->dn_free_txg = 0;
603 	if (dn->dn_dirtyctx_firstset) {
604 		kmem_free(dn->dn_dirtyctx_firstset, 1);
605 		dn->dn_dirtyctx_firstset = NULL;
606 	}
607 
608 	dn->dn_allocated_txg = tx->tx_txg;
609 	dn->dn_id_flags = 0;
610 
611 	dnode_setdirty(dn, tx);
612 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
613 	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
614 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
615 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
616 }
617 
618 void
619 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
620     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
621 {
622 	int nblkptr;
623 
624 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
625 	ASSERT3U(blocksize, <=,
626 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
627 	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
628 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
629 	ASSERT(tx->tx_txg != 0);
630 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
631 	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
632 	    (bonustype == DMU_OT_SA && bonuslen == 0));
633 	ASSERT(DMU_OT_IS_VALID(bonustype));
634 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
635 
636 	/* clean up any unreferenced dbufs */
637 	dnode_evict_dbufs(dn);
638 
639 	dn->dn_id_flags = 0;
640 
641 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
642 	dnode_setdirty(dn, tx);
643 	if (dn->dn_datablksz != blocksize) {
644 		/* change blocksize */
645 		ASSERT(dn->dn_maxblkid == 0 &&
646 		    (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
647 		    dnode_block_freed(dn, 0)));
648 		dnode_setdblksz(dn, blocksize);
649 		dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
650 	}
651 	if (dn->dn_bonuslen != bonuslen)
652 		dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
653 
654 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
655 		nblkptr = 1;
656 	else
657 		nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
658 	if (dn->dn_bonustype != bonustype)
659 		dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
660 	if (dn->dn_nblkptr != nblkptr)
661 		dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
662 	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
663 		dbuf_rm_spill(dn, tx);
664 		dnode_rm_spill(dn, tx);
665 	}
666 	rw_exit(&dn->dn_struct_rwlock);
667 
668 	/* change type */
669 	dn->dn_type = ot;
670 
671 	/* change bonus size and type */
672 	mutex_enter(&dn->dn_mtx);
673 	dn->dn_bonustype = bonustype;
674 	dn->dn_bonuslen = bonuslen;
675 	dn->dn_nblkptr = nblkptr;
676 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
677 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
678 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
679 
680 	/* fix up the bonus db_size */
681 	if (dn->dn_bonus) {
682 		dn->dn_bonus->db.db_size =
683 		    DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
684 		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
685 	}
686 
687 	dn->dn_allocated_txg = tx->tx_txg;
688 	mutex_exit(&dn->dn_mtx);
689 }
690 
691 #ifdef	DNODE_STATS
692 static struct {
693 	uint64_t dms_dnode_invalid;
694 	uint64_t dms_dnode_recheck1;
695 	uint64_t dms_dnode_recheck2;
696 	uint64_t dms_dnode_special;
697 	uint64_t dms_dnode_handle;
698 	uint64_t dms_dnode_rwlock;
699 	uint64_t dms_dnode_active;
700 } dnode_move_stats;
701 #endif	/* DNODE_STATS */
702 
703 static void
704 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
705 {
706 	int i;
707 
708 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
709 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
710 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
711 	ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
712 
713 	/* Copy fields. */
714 	ndn->dn_objset = odn->dn_objset;
715 	ndn->dn_object = odn->dn_object;
716 	ndn->dn_dbuf = odn->dn_dbuf;
717 	ndn->dn_handle = odn->dn_handle;
718 	ndn->dn_phys = odn->dn_phys;
719 	ndn->dn_type = odn->dn_type;
720 	ndn->dn_bonuslen = odn->dn_bonuslen;
721 	ndn->dn_bonustype = odn->dn_bonustype;
722 	ndn->dn_nblkptr = odn->dn_nblkptr;
723 	ndn->dn_checksum = odn->dn_checksum;
724 	ndn->dn_compress = odn->dn_compress;
725 	ndn->dn_nlevels = odn->dn_nlevels;
726 	ndn->dn_indblkshift = odn->dn_indblkshift;
727 	ndn->dn_datablkshift = odn->dn_datablkshift;
728 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
729 	ndn->dn_datablksz = odn->dn_datablksz;
730 	ndn->dn_maxblkid = odn->dn_maxblkid;
731 	bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
732 	    sizeof (odn->dn_next_nblkptr));
733 	bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
734 	    sizeof (odn->dn_next_nlevels));
735 	bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
736 	    sizeof (odn->dn_next_indblkshift));
737 	bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
738 	    sizeof (odn->dn_next_bonustype));
739 	bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
740 	    sizeof (odn->dn_rm_spillblk));
741 	bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
742 	    sizeof (odn->dn_next_bonuslen));
743 	bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
744 	    sizeof (odn->dn_next_blksz));
745 	for (i = 0; i < TXG_SIZE; i++) {
746 		list_move_tail(&ndn->dn_dirty_records[i],
747 		    &odn->dn_dirty_records[i]);
748 	}
749 	bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
750 	    sizeof (odn->dn_free_ranges));
751 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
752 	ndn->dn_free_txg = odn->dn_free_txg;
753 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
754 	ndn->dn_dirtyctx = odn->dn_dirtyctx;
755 	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
756 	ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
757 	refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
758 	ASSERT(avl_is_empty(&ndn->dn_dbufs));
759 	avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
760 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
761 	ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
762 	ndn->dn_bonus = odn->dn_bonus;
763 	ndn->dn_have_spill = odn->dn_have_spill;
764 	ndn->dn_zio = odn->dn_zio;
765 	ndn->dn_oldused = odn->dn_oldused;
766 	ndn->dn_oldflags = odn->dn_oldflags;
767 	ndn->dn_olduid = odn->dn_olduid;
768 	ndn->dn_oldgid = odn->dn_oldgid;
769 	ndn->dn_newuid = odn->dn_newuid;
770 	ndn->dn_newgid = odn->dn_newgid;
771 	ndn->dn_id_flags = odn->dn_id_flags;
772 	dmu_zfetch_init(&ndn->dn_zfetch, NULL);
773 	list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
774 	ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
775 
776 	/*
777 	 * Update back pointers. Updating the handle fixes the back pointer of
778 	 * every descendant dbuf as well as the bonus dbuf.
779 	 */
780 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
781 	ndn->dn_handle->dnh_dnode = ndn;
782 	if (ndn->dn_zfetch.zf_dnode == odn) {
783 		ndn->dn_zfetch.zf_dnode = ndn;
784 	}
785 
786 	/*
787 	 * Invalidate the original dnode by clearing all of its back pointers.
788 	 */
789 	odn->dn_dbuf = NULL;
790 	odn->dn_handle = NULL;
791 	avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
792 	    offsetof(dmu_buf_impl_t, db_link));
793 	odn->dn_dbufs_count = 0;
794 	odn->dn_unlisted_l0_blkid = 0;
795 	odn->dn_bonus = NULL;
796 	odn->dn_zfetch.zf_dnode = NULL;
797 
798 	/*
799 	 * Set the low bit of the objset pointer to ensure that dnode_move()
800 	 * recognizes the dnode as invalid in any subsequent callback.
801 	 */
802 	POINTER_INVALIDATE(&odn->dn_objset);
803 
804 	/*
805 	 * Satisfy the destructor.
806 	 */
807 	for (i = 0; i < TXG_SIZE; i++) {
808 		list_create(&odn->dn_dirty_records[i],
809 		    sizeof (dbuf_dirty_record_t),
810 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
811 		odn->dn_free_ranges[i] = NULL;
812 		odn->dn_next_nlevels[i] = 0;
813 		odn->dn_next_indblkshift[i] = 0;
814 		odn->dn_next_bonustype[i] = 0;
815 		odn->dn_rm_spillblk[i] = 0;
816 		odn->dn_next_bonuslen[i] = 0;
817 		odn->dn_next_blksz[i] = 0;
818 	}
819 	odn->dn_allocated_txg = 0;
820 	odn->dn_free_txg = 0;
821 	odn->dn_assigned_txg = 0;
822 	odn->dn_dirtyctx = 0;
823 	odn->dn_dirtyctx_firstset = NULL;
824 	odn->dn_have_spill = B_FALSE;
825 	odn->dn_zio = NULL;
826 	odn->dn_oldused = 0;
827 	odn->dn_oldflags = 0;
828 	odn->dn_olduid = 0;
829 	odn->dn_oldgid = 0;
830 	odn->dn_newuid = 0;
831 	odn->dn_newgid = 0;
832 	odn->dn_id_flags = 0;
833 
834 	/*
835 	 * Mark the dnode.
836 	 */
837 	ndn->dn_moved = 1;
838 	odn->dn_moved = (uint8_t)-1;
839 }
840 
841 #ifdef	_KERNEL
842 /*ARGSUSED*/
843 static kmem_cbrc_t
844 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
845 {
846 	dnode_t *odn = buf, *ndn = newbuf;
847 	objset_t *os;
848 	int64_t refcount;
849 	uint32_t dbufs;
850 
851 	/*
852 	 * The dnode is on the objset's list of known dnodes if the objset
853 	 * pointer is valid. We set the low bit of the objset pointer when
854 	 * freeing the dnode to invalidate it, and the memory patterns written
855 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
856 	 * A newly created dnode sets the objset pointer last of all to indicate
857 	 * that the dnode is known and in a valid state to be moved by this
858 	 * function.
859 	 */
860 	os = odn->dn_objset;
861 	if (!POINTER_IS_VALID(os)) {
862 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
863 		return (KMEM_CBRC_DONT_KNOW);
864 	}
865 
866 	/*
867 	 * Ensure that the objset does not go away during the move.
868 	 */
869 	rw_enter(&os_lock, RW_WRITER);
870 	if (os != odn->dn_objset) {
871 		rw_exit(&os_lock);
872 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
873 		return (KMEM_CBRC_DONT_KNOW);
874 	}
875 
876 	/*
877 	 * If the dnode is still valid, then so is the objset. We know that no
878 	 * valid objset can be freed while we hold os_lock, so we can safely
879 	 * ensure that the objset remains in use.
880 	 */
881 	mutex_enter(&os->os_lock);
882 
883 	/*
884 	 * Recheck the objset pointer in case the dnode was removed just before
885 	 * acquiring the lock.
886 	 */
887 	if (os != odn->dn_objset) {
888 		mutex_exit(&os->os_lock);
889 		rw_exit(&os_lock);
890 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
891 		return (KMEM_CBRC_DONT_KNOW);
892 	}
893 
894 	/*
895 	 * At this point we know that as long as we hold os->os_lock, the dnode
896 	 * cannot be freed and fields within the dnode can be safely accessed.
897 	 * The objset listing this dnode cannot go away as long as this dnode is
898 	 * on its list.
899 	 */
900 	rw_exit(&os_lock);
901 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
902 		mutex_exit(&os->os_lock);
903 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
904 		return (KMEM_CBRC_NO);
905 	}
906 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
907 
908 	/*
909 	 * Lock the dnode handle to prevent the dnode from obtaining any new
910 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
911 	 * from accessing the dnode, so that we can discount their holds. The
912 	 * handle is safe to access because we know that while the dnode cannot
913 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
914 	 * safely move any dnode referenced only by dbufs.
915 	 */
916 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
917 		mutex_exit(&os->os_lock);
918 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
919 		return (KMEM_CBRC_LATER);
920 	}
921 
922 	/*
923 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
924 	 * We need to guarantee that there is a hold for every dbuf in order to
925 	 * determine whether the dnode is actively referenced. Falsely matching
926 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
927 	 * that a thread already having an active dnode hold is about to add a
928 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
929 	 * progress.
930 	 */
931 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
932 		zrl_exit(&odn->dn_handle->dnh_zrlock);
933 		mutex_exit(&os->os_lock);
934 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
935 		return (KMEM_CBRC_LATER);
936 	}
937 
938 	/*
939 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
940 	 * case, the dbuf count is decremented under the handle lock before the
941 	 * dbuf's hold is released. This order ensures that if we count the hold
942 	 * after the dbuf is removed but before its hold is released, we will
943 	 * treat the unmatched hold as active and exit safely. If we count the
944 	 * hold before the dbuf is removed, the hold is discounted, and the
945 	 * removal is blocked until the move completes.
946 	 */
947 	refcount = refcount_count(&odn->dn_holds);
948 	ASSERT(refcount >= 0);
949 	dbufs = odn->dn_dbufs_count;
950 
951 	/* We can't have more dbufs than dnode holds. */
952 	ASSERT3U(dbufs, <=, refcount);
953 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
954 	    uint32_t, dbufs);
955 
956 	if (refcount > dbufs) {
957 		rw_exit(&odn->dn_struct_rwlock);
958 		zrl_exit(&odn->dn_handle->dnh_zrlock);
959 		mutex_exit(&os->os_lock);
960 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
961 		return (KMEM_CBRC_LATER);
962 	}
963 
964 	rw_exit(&odn->dn_struct_rwlock);
965 
966 	/*
967 	 * At this point we know that anyone with a hold on the dnode is not
968 	 * actively referencing it. The dnode is known and in a valid state to
969 	 * move. We're holding the locks needed to execute the critical section.
970 	 */
971 	dnode_move_impl(odn, ndn);
972 
973 	list_link_replace(&odn->dn_link, &ndn->dn_link);
974 	/* If the dnode was safe to move, the refcount cannot have changed. */
975 	ASSERT(refcount == refcount_count(&ndn->dn_holds));
976 	ASSERT(dbufs == ndn->dn_dbufs_count);
977 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
978 	mutex_exit(&os->os_lock);
979 
980 	return (KMEM_CBRC_YES);
981 }
982 #endif	/* _KERNEL */
983 
984 void
985 dnode_special_close(dnode_handle_t *dnh)
986 {
987 	dnode_t *dn = dnh->dnh_dnode;
988 
989 	/*
990 	 * Wait for final references to the dnode to clear.  This can
991 	 * only happen if the arc is asyncronously evicting state that
992 	 * has a hold on this dnode while we are trying to evict this
993 	 * dnode.
994 	 */
995 	while (refcount_count(&dn->dn_holds) > 0)
996 		delay(1);
997 	ASSERT(dn->dn_dbuf == NULL ||
998 	    dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
999 	zrl_add(&dnh->dnh_zrlock);
1000 	dnode_destroy(dn); /* implicit zrl_remove() */
1001 	zrl_destroy(&dnh->dnh_zrlock);
1002 	dnh->dnh_dnode = NULL;
1003 }
1004 
1005 void
1006 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1007     dnode_handle_t *dnh)
1008 {
1009 	dnode_t *dn;
1010 
1011 	dn = dnode_create(os, dnp, NULL, object, dnh);
1012 	zrl_init(&dnh->dnh_zrlock);
1013 	DNODE_VERIFY(dn);
1014 }
1015 
1016 static void
1017 dnode_buf_evict_async(void *dbu)
1018 {
1019 	dnode_children_t *children_dnodes = dbu;
1020 	int i;
1021 
1022 	for (i = 0; i < children_dnodes->dnc_count; i++) {
1023 		dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1024 		dnode_t *dn;
1025 
1026 		/*
1027 		 * The dnode handle lock guards against the dnode moving to
1028 		 * another valid address, so there is no need here to guard
1029 		 * against changes to or from NULL.
1030 		 */
1031 		if (dnh->dnh_dnode == NULL) {
1032 			zrl_destroy(&dnh->dnh_zrlock);
1033 			continue;
1034 		}
1035 
1036 		zrl_add(&dnh->dnh_zrlock);
1037 		dn = dnh->dnh_dnode;
1038 		/*
1039 		 * If there are holds on this dnode, then there should
1040 		 * be holds on the dnode's containing dbuf as well; thus
1041 		 * it wouldn't be eligible for eviction and this function
1042 		 * would not have been called.
1043 		 */
1044 		ASSERT(refcount_is_zero(&dn->dn_holds));
1045 		ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1046 
1047 		dnode_destroy(dn); /* implicit zrl_remove() */
1048 		zrl_destroy(&dnh->dnh_zrlock);
1049 		dnh->dnh_dnode = NULL;
1050 	}
1051 	kmem_free(children_dnodes, sizeof (dnode_children_t) +
1052 	    children_dnodes->dnc_count * sizeof (dnode_handle_t));
1053 }
1054 
1055 /*
1056  * errors:
1057  * EINVAL - invalid object number.
1058  * EIO - i/o error.
1059  * succeeds even for free dnodes.
1060  */
1061 int
1062 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1063     void *tag, dnode_t **dnp)
1064 {
1065 	int epb, idx, err;
1066 	int drop_struct_lock = FALSE;
1067 	int type;
1068 	uint64_t blk;
1069 	dnode_t *mdn, *dn;
1070 	dmu_buf_impl_t *db;
1071 	dnode_children_t *children_dnodes;
1072 	dnode_handle_t *dnh;
1073 
1074 	/*
1075 	 * If you are holding the spa config lock as writer, you shouldn't
1076 	 * be asking the DMU to do *anything* unless it's the root pool
1077 	 * which may require us to read from the root filesystem while
1078 	 * holding some (not all) of the locks as writer.
1079 	 */
1080 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1081 	    (spa_is_root(os->os_spa) &&
1082 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1083 
1084 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1085 		dn = (object == DMU_USERUSED_OBJECT) ?
1086 		    DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1087 		if (dn == NULL)
1088 			return (SET_ERROR(ENOENT));
1089 		type = dn->dn_type;
1090 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1091 			return (SET_ERROR(ENOENT));
1092 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1093 			return (SET_ERROR(EEXIST));
1094 		DNODE_VERIFY(dn);
1095 		(void) refcount_add(&dn->dn_holds, tag);
1096 		*dnp = dn;
1097 		return (0);
1098 	}
1099 
1100 	if (object == 0 || object >= DN_MAX_OBJECT)
1101 		return (SET_ERROR(EINVAL));
1102 
1103 	mdn = DMU_META_DNODE(os);
1104 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1105 
1106 	DNODE_VERIFY(mdn);
1107 
1108 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1109 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1110 		drop_struct_lock = TRUE;
1111 	}
1112 
1113 	blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1114 
1115 	db = dbuf_hold(mdn, blk, FTAG);
1116 	if (drop_struct_lock)
1117 		rw_exit(&mdn->dn_struct_rwlock);
1118 	if (db == NULL)
1119 		return (SET_ERROR(EIO));
1120 	err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1121 	if (err) {
1122 		dbuf_rele(db, FTAG);
1123 		return (err);
1124 	}
1125 
1126 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1127 	epb = db->db.db_size >> DNODE_SHIFT;
1128 
1129 	idx = object & (epb-1);
1130 
1131 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1132 	children_dnodes = dmu_buf_get_user(&db->db);
1133 	if (children_dnodes == NULL) {
1134 		int i;
1135 		dnode_children_t *winner;
1136 		children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1137 		    epb * sizeof (dnode_handle_t), KM_SLEEP);
1138 		children_dnodes->dnc_count = epb;
1139 		dnh = &children_dnodes->dnc_children[0];
1140 		for (i = 0; i < epb; i++) {
1141 			zrl_init(&dnh[i].dnh_zrlock);
1142 		}
1143 		dmu_buf_init_user(&children_dnodes->dnc_dbu, NULL,
1144 		    dnode_buf_evict_async, NULL);
1145 		winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1146 		if (winner != NULL) {
1147 
1148 			for (i = 0; i < epb; i++) {
1149 				zrl_destroy(&dnh[i].dnh_zrlock);
1150 			}
1151 
1152 			kmem_free(children_dnodes, sizeof (dnode_children_t) +
1153 			    epb * sizeof (dnode_handle_t));
1154 			children_dnodes = winner;
1155 		}
1156 	}
1157 	ASSERT(children_dnodes->dnc_count == epb);
1158 
1159 	dnh = &children_dnodes->dnc_children[idx];
1160 	zrl_add(&dnh->dnh_zrlock);
1161 	dn = dnh->dnh_dnode;
1162 	if (dn == NULL) {
1163 		dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1164 
1165 		dn = dnode_create(os, phys, db, object, dnh);
1166 	}
1167 
1168 	mutex_enter(&dn->dn_mtx);
1169 	type = dn->dn_type;
1170 	if (dn->dn_free_txg ||
1171 	    ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1172 	    ((flag & DNODE_MUST_BE_FREE) &&
1173 	    (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1174 		mutex_exit(&dn->dn_mtx);
1175 		zrl_remove(&dnh->dnh_zrlock);
1176 		dbuf_rele(db, FTAG);
1177 		return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1178 	}
1179 	if (refcount_add(&dn->dn_holds, tag) == 1)
1180 		dbuf_add_ref(db, dnh);
1181 	mutex_exit(&dn->dn_mtx);
1182 
1183 	/* Now we can rely on the hold to prevent the dnode from moving. */
1184 	zrl_remove(&dnh->dnh_zrlock);
1185 
1186 	DNODE_VERIFY(dn);
1187 	ASSERT3P(dn->dn_dbuf, ==, db);
1188 	ASSERT3U(dn->dn_object, ==, object);
1189 	dbuf_rele(db, FTAG);
1190 
1191 	*dnp = dn;
1192 	return (0);
1193 }
1194 
1195 /*
1196  * Return held dnode if the object is allocated, NULL if not.
1197  */
1198 int
1199 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1200 {
1201 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1202 }
1203 
1204 /*
1205  * Can only add a reference if there is already at least one
1206  * reference on the dnode.  Returns FALSE if unable to add a
1207  * new reference.
1208  */
1209 boolean_t
1210 dnode_add_ref(dnode_t *dn, void *tag)
1211 {
1212 	mutex_enter(&dn->dn_mtx);
1213 	if (refcount_is_zero(&dn->dn_holds)) {
1214 		mutex_exit(&dn->dn_mtx);
1215 		return (FALSE);
1216 	}
1217 	VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1218 	mutex_exit(&dn->dn_mtx);
1219 	return (TRUE);
1220 }
1221 
1222 void
1223 dnode_rele(dnode_t *dn, void *tag)
1224 {
1225 	mutex_enter(&dn->dn_mtx);
1226 	dnode_rele_and_unlock(dn, tag);
1227 }
1228 
1229 void
1230 dnode_rele_and_unlock(dnode_t *dn, void *tag)
1231 {
1232 	uint64_t refs;
1233 	/* Get while the hold prevents the dnode from moving. */
1234 	dmu_buf_impl_t *db = dn->dn_dbuf;
1235 	dnode_handle_t *dnh = dn->dn_handle;
1236 
1237 	refs = refcount_remove(&dn->dn_holds, tag);
1238 	mutex_exit(&dn->dn_mtx);
1239 
1240 	/*
1241 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1242 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1243 	 * prevent the dnode from moving, since releasing the last hold could
1244 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1245 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1246 	 * other direct or indirect hold on the dnode must first drop the dnode
1247 	 * handle.
1248 	 */
1249 	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1250 
1251 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1252 	if (refs == 0 && db != NULL) {
1253 		/*
1254 		 * Another thread could add a hold to the dnode handle in
1255 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1256 		 * hold on the parent dbuf prevents the handle from being
1257 		 * destroyed, the hold on the handle is OK. We can't yet assert
1258 		 * that the handle has zero references, but that will be
1259 		 * asserted anyway when the handle gets destroyed.
1260 		 */
1261 		dbuf_rele(db, dnh);
1262 	}
1263 }
1264 
1265 void
1266 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1267 {
1268 	objset_t *os = dn->dn_objset;
1269 	uint64_t txg = tx->tx_txg;
1270 
1271 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1272 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1273 		return;
1274 	}
1275 
1276 	DNODE_VERIFY(dn);
1277 
1278 #ifdef ZFS_DEBUG
1279 	mutex_enter(&dn->dn_mtx);
1280 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1281 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1282 	mutex_exit(&dn->dn_mtx);
1283 #endif
1284 
1285 	/*
1286 	 * Determine old uid/gid when necessary
1287 	 */
1288 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1289 
1290 	mutex_enter(&os->os_lock);
1291 
1292 	/*
1293 	 * If we are already marked dirty, we're done.
1294 	 */
1295 	if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1296 		mutex_exit(&os->os_lock);
1297 		return;
1298 	}
1299 
1300 	ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1301 	    !avl_is_empty(&dn->dn_dbufs));
1302 	ASSERT(dn->dn_datablksz != 0);
1303 	ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1304 	ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1305 	ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1306 
1307 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1308 	    dn->dn_object, txg);
1309 
1310 	if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1311 		list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1312 	} else {
1313 		list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1314 	}
1315 
1316 	mutex_exit(&os->os_lock);
1317 
1318 	/*
1319 	 * The dnode maintains a hold on its containing dbuf as
1320 	 * long as there are holds on it.  Each instantiated child
1321 	 * dbuf maintains a hold on the dnode.  When the last child
1322 	 * drops its hold, the dnode will drop its hold on the
1323 	 * containing dbuf. We add a "dirty hold" here so that the
1324 	 * dnode will hang around after we finish processing its
1325 	 * children.
1326 	 */
1327 	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1328 
1329 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1330 
1331 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1332 }
1333 
1334 void
1335 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1336 {
1337 	int txgoff = tx->tx_txg & TXG_MASK;
1338 
1339 	dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1340 
1341 	/* we should be the only holder... hopefully */
1342 	/* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1343 
1344 	mutex_enter(&dn->dn_mtx);
1345 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1346 		mutex_exit(&dn->dn_mtx);
1347 		return;
1348 	}
1349 	dn->dn_free_txg = tx->tx_txg;
1350 	mutex_exit(&dn->dn_mtx);
1351 
1352 	/*
1353 	 * If the dnode is already dirty, it needs to be moved from
1354 	 * the dirty list to the free list.
1355 	 */
1356 	mutex_enter(&dn->dn_objset->os_lock);
1357 	if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1358 		list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1359 		list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1360 		mutex_exit(&dn->dn_objset->os_lock);
1361 	} else {
1362 		mutex_exit(&dn->dn_objset->os_lock);
1363 		dnode_setdirty(dn, tx);
1364 	}
1365 }
1366 
1367 /*
1368  * Try to change the block size for the indicated dnode.  This can only
1369  * succeed if there are no blocks allocated or dirty beyond first block
1370  */
1371 int
1372 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1373 {
1374 	dmu_buf_impl_t *db;
1375 	int err;
1376 
1377 	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1378 	if (size == 0)
1379 		size = SPA_MINBLOCKSIZE;
1380 	else
1381 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1382 
1383 	if (ibs == dn->dn_indblkshift)
1384 		ibs = 0;
1385 
1386 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1387 		return (0);
1388 
1389 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1390 
1391 	/* Check for any allocated blocks beyond the first */
1392 	if (dn->dn_maxblkid != 0)
1393 		goto fail;
1394 
1395 	mutex_enter(&dn->dn_dbufs_mtx);
1396 	for (db = avl_first(&dn->dn_dbufs); db != NULL;
1397 	    db = AVL_NEXT(&dn->dn_dbufs, db)) {
1398 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1399 		    db->db_blkid != DMU_SPILL_BLKID) {
1400 			mutex_exit(&dn->dn_dbufs_mtx);
1401 			goto fail;
1402 		}
1403 	}
1404 	mutex_exit(&dn->dn_dbufs_mtx);
1405 
1406 	if (ibs && dn->dn_nlevels != 1)
1407 		goto fail;
1408 
1409 	/* resize the old block */
1410 	err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1411 	if (err == 0)
1412 		dbuf_new_size(db, size, tx);
1413 	else if (err != ENOENT)
1414 		goto fail;
1415 
1416 	dnode_setdblksz(dn, size);
1417 	dnode_setdirty(dn, tx);
1418 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1419 	if (ibs) {
1420 		dn->dn_indblkshift = ibs;
1421 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1422 	}
1423 	/* rele after we have fixed the blocksize in the dnode */
1424 	if (db)
1425 		dbuf_rele(db, FTAG);
1426 
1427 	rw_exit(&dn->dn_struct_rwlock);
1428 	return (0);
1429 
1430 fail:
1431 	rw_exit(&dn->dn_struct_rwlock);
1432 	return (SET_ERROR(ENOTSUP));
1433 }
1434 
1435 /* read-holding callers must not rely on the lock being continuously held */
1436 void
1437 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1438 {
1439 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1440 	int epbs, new_nlevels;
1441 	uint64_t sz;
1442 
1443 	ASSERT(blkid != DMU_BONUS_BLKID);
1444 
1445 	ASSERT(have_read ?
1446 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
1447 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
1448 
1449 	/*
1450 	 * if we have a read-lock, check to see if we need to do any work
1451 	 * before upgrading to a write-lock.
1452 	 */
1453 	if (have_read) {
1454 		if (blkid <= dn->dn_maxblkid)
1455 			return;
1456 
1457 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1458 			rw_exit(&dn->dn_struct_rwlock);
1459 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1460 		}
1461 	}
1462 
1463 	if (blkid <= dn->dn_maxblkid)
1464 		goto out;
1465 
1466 	dn->dn_maxblkid = blkid;
1467 
1468 	/*
1469 	 * Compute the number of levels necessary to support the new maxblkid.
1470 	 */
1471 	new_nlevels = 1;
1472 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1473 	for (sz = dn->dn_nblkptr;
1474 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1475 		new_nlevels++;
1476 
1477 	if (new_nlevels > dn->dn_nlevels) {
1478 		int old_nlevels = dn->dn_nlevels;
1479 		dmu_buf_impl_t *db;
1480 		list_t *list;
1481 		dbuf_dirty_record_t *new, *dr, *dr_next;
1482 
1483 		dn->dn_nlevels = new_nlevels;
1484 
1485 		ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1486 		dn->dn_next_nlevels[txgoff] = new_nlevels;
1487 
1488 		/* dirty the left indirects */
1489 		db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1490 		ASSERT(db != NULL);
1491 		new = dbuf_dirty(db, tx);
1492 		dbuf_rele(db, FTAG);
1493 
1494 		/* transfer the dirty records to the new indirect */
1495 		mutex_enter(&dn->dn_mtx);
1496 		mutex_enter(&new->dt.di.dr_mtx);
1497 		list = &dn->dn_dirty_records[txgoff];
1498 		for (dr = list_head(list); dr; dr = dr_next) {
1499 			dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1500 			if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1501 			    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1502 			    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1503 				ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1504 				list_remove(&dn->dn_dirty_records[txgoff], dr);
1505 				list_insert_tail(&new->dt.di.dr_children, dr);
1506 				dr->dr_parent = new;
1507 			}
1508 		}
1509 		mutex_exit(&new->dt.di.dr_mtx);
1510 		mutex_exit(&dn->dn_mtx);
1511 	}
1512 
1513 out:
1514 	if (have_read)
1515 		rw_downgrade(&dn->dn_struct_rwlock);
1516 }
1517 
1518 static void
1519 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1520 {
1521 	dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1522 	if (db != NULL) {
1523 		dmu_buf_will_dirty(&db->db, tx);
1524 		dbuf_rele(db, FTAG);
1525 	}
1526 }
1527 
1528 void
1529 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1530 {
1531 	dmu_buf_impl_t *db;
1532 	uint64_t blkoff, blkid, nblks;
1533 	int blksz, blkshift, head, tail;
1534 	int trunc = FALSE;
1535 	int epbs;
1536 
1537 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1538 	blksz = dn->dn_datablksz;
1539 	blkshift = dn->dn_datablkshift;
1540 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1541 
1542 	if (len == DMU_OBJECT_END) {
1543 		len = UINT64_MAX - off;
1544 		trunc = TRUE;
1545 	}
1546 
1547 	/*
1548 	 * First, block align the region to free:
1549 	 */
1550 	if (ISP2(blksz)) {
1551 		head = P2NPHASE(off, blksz);
1552 		blkoff = P2PHASE(off, blksz);
1553 		if ((off >> blkshift) > dn->dn_maxblkid)
1554 			goto out;
1555 	} else {
1556 		ASSERT(dn->dn_maxblkid == 0);
1557 		if (off == 0 && len >= blksz) {
1558 			/*
1559 			 * Freeing the whole block; fast-track this request.
1560 			 * Note that we won't dirty any indirect blocks,
1561 			 * which is fine because we will be freeing the entire
1562 			 * file and thus all indirect blocks will be freed
1563 			 * by free_children().
1564 			 */
1565 			blkid = 0;
1566 			nblks = 1;
1567 			goto done;
1568 		} else if (off >= blksz) {
1569 			/* Freeing past end-of-data */
1570 			goto out;
1571 		} else {
1572 			/* Freeing part of the block. */
1573 			head = blksz - off;
1574 			ASSERT3U(head, >, 0);
1575 		}
1576 		blkoff = off;
1577 	}
1578 	/* zero out any partial block data at the start of the range */
1579 	if (head) {
1580 		ASSERT3U(blkoff + head, ==, blksz);
1581 		if (len < head)
1582 			head = len;
1583 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1584 		    TRUE, FALSE, FTAG, &db) == 0) {
1585 			caddr_t data;
1586 
1587 			/* don't dirty if it isn't on disk and isn't dirty */
1588 			if (db->db_last_dirty ||
1589 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1590 				rw_exit(&dn->dn_struct_rwlock);
1591 				dmu_buf_will_dirty(&db->db, tx);
1592 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1593 				data = db->db.db_data;
1594 				bzero(data + blkoff, head);
1595 			}
1596 			dbuf_rele(db, FTAG);
1597 		}
1598 		off += head;
1599 		len -= head;
1600 	}
1601 
1602 	/* If the range was less than one block, we're done */
1603 	if (len == 0)
1604 		goto out;
1605 
1606 	/* If the remaining range is past end of file, we're done */
1607 	if ((off >> blkshift) > dn->dn_maxblkid)
1608 		goto out;
1609 
1610 	ASSERT(ISP2(blksz));
1611 	if (trunc)
1612 		tail = 0;
1613 	else
1614 		tail = P2PHASE(len, blksz);
1615 
1616 	ASSERT0(P2PHASE(off, blksz));
1617 	/* zero out any partial block data at the end of the range */
1618 	if (tail) {
1619 		if (len < tail)
1620 			tail = len;
1621 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1622 		    TRUE, FALSE, FTAG, &db) == 0) {
1623 			/* don't dirty if not on disk and not dirty */
1624 			if (db->db_last_dirty ||
1625 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1626 				rw_exit(&dn->dn_struct_rwlock);
1627 				dmu_buf_will_dirty(&db->db, tx);
1628 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1629 				bzero(db->db.db_data, tail);
1630 			}
1631 			dbuf_rele(db, FTAG);
1632 		}
1633 		len -= tail;
1634 	}
1635 
1636 	/* If the range did not include a full block, we are done */
1637 	if (len == 0)
1638 		goto out;
1639 
1640 	ASSERT(IS_P2ALIGNED(off, blksz));
1641 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1642 	blkid = off >> blkshift;
1643 	nblks = len >> blkshift;
1644 	if (trunc)
1645 		nblks += 1;
1646 
1647 	/*
1648 	 * Dirty all the indirect blocks in this range.  Note that only
1649 	 * the first and last indirect blocks can actually be written
1650 	 * (if they were partially freed) -- they must be dirtied, even if
1651 	 * they do not exist on disk yet.  The interior blocks will
1652 	 * be freed by free_children(), so they will not actually be written.
1653 	 * Even though these interior blocks will not be written, we
1654 	 * dirty them for two reasons:
1655 	 *
1656 	 *  - It ensures that the indirect blocks remain in memory until
1657 	 *    syncing context.  (They have already been prefetched by
1658 	 *    dmu_tx_hold_free(), so we don't have to worry about reading
1659 	 *    them serially here.)
1660 	 *
1661 	 *  - The dirty space accounting will put pressure on the txg sync
1662 	 *    mechanism to begin syncing, and to delay transactions if there
1663 	 *    is a large amount of freeing.  Even though these indirect
1664 	 *    blocks will not be written, we could need to write the same
1665 	 *    amount of space if we copy the freed BPs into deadlists.
1666 	 */
1667 	if (dn->dn_nlevels > 1) {
1668 		uint64_t first, last;
1669 
1670 		first = blkid >> epbs;
1671 		dnode_dirty_l1(dn, first, tx);
1672 		if (trunc)
1673 			last = dn->dn_maxblkid >> epbs;
1674 		else
1675 			last = (blkid + nblks - 1) >> epbs;
1676 		if (last != first)
1677 			dnode_dirty_l1(dn, last, tx);
1678 
1679 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1680 		    SPA_BLKPTRSHIFT;
1681 		for (uint64_t i = first + 1; i < last; i++) {
1682 			/*
1683 			 * Set i to the blockid of the next non-hole
1684 			 * level-1 indirect block at or after i.  Note
1685 			 * that dnode_next_offset() operates in terms of
1686 			 * level-0-equivalent bytes.
1687 			 */
1688 			uint64_t ibyte = i << shift;
1689 			int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1690 			    &ibyte, 2, 1, 0);
1691 			i = ibyte >> shift;
1692 			if (i >= last)
1693 				break;
1694 
1695 			/*
1696 			 * Normally we should not see an error, either
1697 			 * from dnode_next_offset() or dbuf_hold_level()
1698 			 * (except for ESRCH from dnode_next_offset).
1699 			 * If there is an i/o error, then when we read
1700 			 * this block in syncing context, it will use
1701 			 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1702 			 * to the "failmode" property.  dnode_next_offset()
1703 			 * doesn't have a flag to indicate MUSTSUCCEED.
1704 			 */
1705 			if (err != 0)
1706 				break;
1707 
1708 			dnode_dirty_l1(dn, i, tx);
1709 		}
1710 	}
1711 
1712 done:
1713 	/*
1714 	 * Add this range to the dnode range list.
1715 	 * We will finish up this free operation in the syncing phase.
1716 	 */
1717 	mutex_enter(&dn->dn_mtx);
1718 	int txgoff = tx->tx_txg & TXG_MASK;
1719 	if (dn->dn_free_ranges[txgoff] == NULL) {
1720 		dn->dn_free_ranges[txgoff] =
1721 		    range_tree_create(NULL, NULL, &dn->dn_mtx);
1722 	}
1723 	range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1724 	range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1725 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1726 	    blkid, nblks, tx->tx_txg);
1727 	mutex_exit(&dn->dn_mtx);
1728 
1729 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1730 	dnode_setdirty(dn, tx);
1731 out:
1732 
1733 	rw_exit(&dn->dn_struct_rwlock);
1734 }
1735 
1736 static boolean_t
1737 dnode_spill_freed(dnode_t *dn)
1738 {
1739 	int i;
1740 
1741 	mutex_enter(&dn->dn_mtx);
1742 	for (i = 0; i < TXG_SIZE; i++) {
1743 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1744 			break;
1745 	}
1746 	mutex_exit(&dn->dn_mtx);
1747 	return (i < TXG_SIZE);
1748 }
1749 
1750 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1751 uint64_t
1752 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1753 {
1754 	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1755 	int i;
1756 
1757 	if (blkid == DMU_BONUS_BLKID)
1758 		return (FALSE);
1759 
1760 	/*
1761 	 * If we're in the process of opening the pool, dp will not be
1762 	 * set yet, but there shouldn't be anything dirty.
1763 	 */
1764 	if (dp == NULL)
1765 		return (FALSE);
1766 
1767 	if (dn->dn_free_txg)
1768 		return (TRUE);
1769 
1770 	if (blkid == DMU_SPILL_BLKID)
1771 		return (dnode_spill_freed(dn));
1772 
1773 	mutex_enter(&dn->dn_mtx);
1774 	for (i = 0; i < TXG_SIZE; i++) {
1775 		if (dn->dn_free_ranges[i] != NULL &&
1776 		    range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1777 			break;
1778 	}
1779 	mutex_exit(&dn->dn_mtx);
1780 	return (i < TXG_SIZE);
1781 }
1782 
1783 /* call from syncing context when we actually write/free space for this dnode */
1784 void
1785 dnode_diduse_space(dnode_t *dn, int64_t delta)
1786 {
1787 	uint64_t space;
1788 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1789 	    dn, dn->dn_phys,
1790 	    (u_longlong_t)dn->dn_phys->dn_used,
1791 	    (longlong_t)delta);
1792 
1793 	mutex_enter(&dn->dn_mtx);
1794 	space = DN_USED_BYTES(dn->dn_phys);
1795 	if (delta > 0) {
1796 		ASSERT3U(space + delta, >=, space); /* no overflow */
1797 	} else {
1798 		ASSERT3U(space, >=, -delta); /* no underflow */
1799 	}
1800 	space += delta;
1801 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1802 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1803 		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1804 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1805 	} else {
1806 		dn->dn_phys->dn_used = space;
1807 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1808 	}
1809 	mutex_exit(&dn->dn_mtx);
1810 }
1811 
1812 /*
1813  * Call when we think we're going to write/free space in open context to track
1814  * the amount of memory in use by the currently open txg.
1815  */
1816 void
1817 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1818 {
1819 	objset_t *os = dn->dn_objset;
1820 	dsl_dataset_t *ds = os->os_dsl_dataset;
1821 	int64_t aspace = spa_get_asize(os->os_spa, space);
1822 
1823 	if (ds != NULL) {
1824 		dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1825 		dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1826 	}
1827 
1828 	dmu_tx_willuse_space(tx, aspace);
1829 }
1830 
1831 /*
1832  * Scans a block at the indicated "level" looking for a hole or data,
1833  * depending on 'flags'.
1834  *
1835  * If level > 0, then we are scanning an indirect block looking at its
1836  * pointers.  If level == 0, then we are looking at a block of dnodes.
1837  *
1838  * If we don't find what we are looking for in the block, we return ESRCH.
1839  * Otherwise, return with *offset pointing to the beginning (if searching
1840  * forwards) or end (if searching backwards) of the range covered by the
1841  * block pointer we matched on (or dnode).
1842  *
1843  * The basic search algorithm used below by dnode_next_offset() is to
1844  * use this function to search up the block tree (widen the search) until
1845  * we find something (i.e., we don't return ESRCH) and then search back
1846  * down the tree (narrow the search) until we reach our original search
1847  * level.
1848  */
1849 static int
1850 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1851     int lvl, uint64_t blkfill, uint64_t txg)
1852 {
1853 	dmu_buf_impl_t *db = NULL;
1854 	void *data = NULL;
1855 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1856 	uint64_t epb = 1ULL << epbs;
1857 	uint64_t minfill, maxfill;
1858 	boolean_t hole;
1859 	int i, inc, error, span;
1860 
1861 	dprintf("probing object %llu offset %llx level %d of %u\n",
1862 	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1863 
1864 	hole = ((flags & DNODE_FIND_HOLE) != 0);
1865 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1866 	ASSERT(txg == 0 || !hole);
1867 
1868 	if (lvl == dn->dn_phys->dn_nlevels) {
1869 		error = 0;
1870 		epb = dn->dn_phys->dn_nblkptr;
1871 		data = dn->dn_phys->dn_blkptr;
1872 	} else {
1873 		uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1874 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1875 		if (error) {
1876 			if (error != ENOENT)
1877 				return (error);
1878 			if (hole)
1879 				return (0);
1880 			/*
1881 			 * This can only happen when we are searching up
1882 			 * the block tree for data.  We don't really need to
1883 			 * adjust the offset, as we will just end up looking
1884 			 * at the pointer to this block in its parent, and its
1885 			 * going to be unallocated, so we will skip over it.
1886 			 */
1887 			return (SET_ERROR(ESRCH));
1888 		}
1889 		error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1890 		if (error) {
1891 			dbuf_rele(db, FTAG);
1892 			return (error);
1893 		}
1894 		data = db->db.db_data;
1895 	}
1896 
1897 
1898 	if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1899 	    db->db_blkptr->blk_birth <= txg ||
1900 	    BP_IS_HOLE(db->db_blkptr))) {
1901 		/*
1902 		 * This can only happen when we are searching up the tree
1903 		 * and these conditions mean that we need to keep climbing.
1904 		 */
1905 		error = SET_ERROR(ESRCH);
1906 	} else if (lvl == 0) {
1907 		dnode_phys_t *dnp = data;
1908 		span = DNODE_SHIFT;
1909 		ASSERT(dn->dn_type == DMU_OT_DNODE);
1910 
1911 		for (i = (*offset >> span) & (blkfill - 1);
1912 		    i >= 0 && i < blkfill; i += inc) {
1913 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1914 				break;
1915 			*offset += (1ULL << span) * inc;
1916 		}
1917 		if (i < 0 || i == blkfill)
1918 			error = SET_ERROR(ESRCH);
1919 	} else {
1920 		blkptr_t *bp = data;
1921 		uint64_t start = *offset;
1922 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
1923 		minfill = 0;
1924 		maxfill = blkfill << ((lvl - 1) * epbs);
1925 
1926 		if (hole)
1927 			maxfill--;
1928 		else
1929 			minfill++;
1930 
1931 		*offset = *offset >> span;
1932 		for (i = BF64_GET(*offset, 0, epbs);
1933 		    i >= 0 && i < epb; i += inc) {
1934 			if (BP_GET_FILL(&bp[i]) >= minfill &&
1935 			    BP_GET_FILL(&bp[i]) <= maxfill &&
1936 			    (hole || bp[i].blk_birth > txg))
1937 				break;
1938 			if (inc > 0 || *offset > 0)
1939 				*offset += inc;
1940 		}
1941 		*offset = *offset << span;
1942 		if (inc < 0) {
1943 			/* traversing backwards; position offset at the end */
1944 			ASSERT3U(*offset, <=, start);
1945 			*offset = MIN(*offset + (1ULL << span) - 1, start);
1946 		} else if (*offset < start) {
1947 			*offset = start;
1948 		}
1949 		if (i < 0 || i >= epb)
1950 			error = SET_ERROR(ESRCH);
1951 	}
1952 
1953 	if (db)
1954 		dbuf_rele(db, FTAG);
1955 
1956 	return (error);
1957 }
1958 
1959 /*
1960  * Find the next hole, data, or sparse region at or after *offset.
1961  * The value 'blkfill' tells us how many items we expect to find
1962  * in an L0 data block; this value is 1 for normal objects,
1963  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1964  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1965  *
1966  * Examples:
1967  *
1968  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1969  *	Finds the next/previous hole/data in a file.
1970  *	Used in dmu_offset_next().
1971  *
1972  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1973  *	Finds the next free/allocated dnode an objset's meta-dnode.
1974  *	Only finds objects that have new contents since txg (ie.
1975  *	bonus buffer changes and content removal are ignored).
1976  *	Used in dmu_object_next().
1977  *
1978  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1979  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
1980  *	Used in dmu_object_alloc().
1981  */
1982 int
1983 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1984     int minlvl, uint64_t blkfill, uint64_t txg)
1985 {
1986 	uint64_t initial_offset = *offset;
1987 	int lvl, maxlvl;
1988 	int error = 0;
1989 
1990 	if (!(flags & DNODE_FIND_HAVELOCK))
1991 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1992 
1993 	if (dn->dn_phys->dn_nlevels == 0) {
1994 		error = SET_ERROR(ESRCH);
1995 		goto out;
1996 	}
1997 
1998 	if (dn->dn_datablkshift == 0) {
1999 		if (*offset < dn->dn_datablksz) {
2000 			if (flags & DNODE_FIND_HOLE)
2001 				*offset = dn->dn_datablksz;
2002 		} else {
2003 			error = SET_ERROR(ESRCH);
2004 		}
2005 		goto out;
2006 	}
2007 
2008 	maxlvl = dn->dn_phys->dn_nlevels;
2009 
2010 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2011 		error = dnode_next_offset_level(dn,
2012 		    flags, offset, lvl, blkfill, txg);
2013 		if (error != ESRCH)
2014 			break;
2015 	}
2016 
2017 	while (error == 0 && --lvl >= minlvl) {
2018 		error = dnode_next_offset_level(dn,
2019 		    flags, offset, lvl, blkfill, txg);
2020 	}
2021 
2022 	/*
2023 	 * There's always a "virtual hole" at the end of the object, even
2024 	 * if all BP's which physically exist are non-holes.
2025 	 */
2026 	if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2027 	    minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2028 		error = 0;
2029 	}
2030 
2031 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2032 	    initial_offset < *offset : initial_offset > *offset))
2033 		error = SET_ERROR(ESRCH);
2034 out:
2035 	if (!(flags & DNODE_FIND_HAVELOCK))
2036 		rw_exit(&dn->dn_struct_rwlock);
2037 
2038 	return (error);
2039 }
2040