xref: /titanic_41/usr/src/uts/common/fs/zfs/dnode.c (revision 5203bc321053fb87d7073c7640548fab73634793)
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  */
24 
25 #include <sys/zfs_context.h>
26 #include <sys/dbuf.h>
27 #include <sys/dnode.h>
28 #include <sys/dmu.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dsl_dataset.h>
34 #include <sys/spa.h>
35 #include <sys/zio.h>
36 #include <sys/dmu_zfetch.h>
37 
38 static int free_range_compar(const void *node1, const void *node2);
39 
40 static kmem_cache_t *dnode_cache;
41 /*
42  * Define DNODE_STATS to turn on statistic gathering. By default, it is only
43  * turned on when DEBUG is also defined.
44  */
45 #ifdef	DEBUG
46 #define	DNODE_STATS
47 #endif	/* DEBUG */
48 
49 #ifdef	DNODE_STATS
50 #define	DNODE_STAT_ADD(stat)			((stat)++)
51 #else
52 #define	DNODE_STAT_ADD(stat)			/* nothing */
53 #endif	/* DNODE_STATS */
54 
55 static dnode_phys_t dnode_phys_zero;
56 
57 int zfs_default_bs = SPA_MINBLOCKSHIFT;
58 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
59 
60 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
61 
62 /* ARGSUSED */
63 static int
64 dnode_cons(void *arg, void *unused, int kmflag)
65 {
66 	dnode_t *dn = arg;
67 	int i;
68 
69 	rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
70 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
71 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
72 	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
73 
74 	refcount_create(&dn->dn_holds);
75 	refcount_create(&dn->dn_tx_holds);
76 	list_link_init(&dn->dn_link);
77 
78 	bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
79 	bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
80 	bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
81 	bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
82 	bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
83 	bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
84 	bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
85 
86 	for (i = 0; i < TXG_SIZE; i++) {
87 		list_link_init(&dn->dn_dirty_link[i]);
88 		avl_create(&dn->dn_ranges[i], free_range_compar,
89 		    sizeof (free_range_t),
90 		    offsetof(struct free_range, fr_node));
91 		list_create(&dn->dn_dirty_records[i],
92 		    sizeof (dbuf_dirty_record_t),
93 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
94 	}
95 
96 	dn->dn_allocated_txg = 0;
97 	dn->dn_free_txg = 0;
98 	dn->dn_assigned_txg = 0;
99 	dn->dn_dirtyctx = 0;
100 	dn->dn_dirtyctx_firstset = NULL;
101 	dn->dn_bonus = NULL;
102 	dn->dn_have_spill = B_FALSE;
103 	dn->dn_zio = NULL;
104 	dn->dn_oldused = 0;
105 	dn->dn_oldflags = 0;
106 	dn->dn_olduid = 0;
107 	dn->dn_oldgid = 0;
108 	dn->dn_newuid = 0;
109 	dn->dn_newgid = 0;
110 	dn->dn_id_flags = 0;
111 
112 	dn->dn_dbufs_count = 0;
113 	list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
114 	    offsetof(dmu_buf_impl_t, db_link));
115 
116 	dn->dn_moved = 0;
117 	return (0);
118 }
119 
120 /* ARGSUSED */
121 static void
122 dnode_dest(void *arg, void *unused)
123 {
124 	int i;
125 	dnode_t *dn = arg;
126 
127 	rw_destroy(&dn->dn_struct_rwlock);
128 	mutex_destroy(&dn->dn_mtx);
129 	mutex_destroy(&dn->dn_dbufs_mtx);
130 	cv_destroy(&dn->dn_notxholds);
131 	refcount_destroy(&dn->dn_holds);
132 	refcount_destroy(&dn->dn_tx_holds);
133 	ASSERT(!list_link_active(&dn->dn_link));
134 
135 	for (i = 0; i < TXG_SIZE; i++) {
136 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
137 		avl_destroy(&dn->dn_ranges[i]);
138 		list_destroy(&dn->dn_dirty_records[i]);
139 		ASSERT3U(dn->dn_next_nblkptr[i], ==, 0);
140 		ASSERT3U(dn->dn_next_nlevels[i], ==, 0);
141 		ASSERT3U(dn->dn_next_indblkshift[i], ==, 0);
142 		ASSERT3U(dn->dn_next_bonustype[i], ==, 0);
143 		ASSERT3U(dn->dn_rm_spillblk[i], ==, 0);
144 		ASSERT3U(dn->dn_next_bonuslen[i], ==, 0);
145 		ASSERT3U(dn->dn_next_blksz[i], ==, 0);
146 	}
147 
148 	ASSERT3U(dn->dn_allocated_txg, ==, 0);
149 	ASSERT3U(dn->dn_free_txg, ==, 0);
150 	ASSERT3U(dn->dn_assigned_txg, ==, 0);
151 	ASSERT3U(dn->dn_dirtyctx, ==, 0);
152 	ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
153 	ASSERT3P(dn->dn_bonus, ==, NULL);
154 	ASSERT(!dn->dn_have_spill);
155 	ASSERT3P(dn->dn_zio, ==, NULL);
156 	ASSERT3U(dn->dn_oldused, ==, 0);
157 	ASSERT3U(dn->dn_oldflags, ==, 0);
158 	ASSERT3U(dn->dn_olduid, ==, 0);
159 	ASSERT3U(dn->dn_oldgid, ==, 0);
160 	ASSERT3U(dn->dn_newuid, ==, 0);
161 	ASSERT3U(dn->dn_newgid, ==, 0);
162 	ASSERT3U(dn->dn_id_flags, ==, 0);
163 
164 	ASSERT3U(dn->dn_dbufs_count, ==, 0);
165 	list_destroy(&dn->dn_dbufs);
166 }
167 
168 void
169 dnode_init(void)
170 {
171 	ASSERT(dnode_cache == NULL);
172 	dnode_cache = kmem_cache_create("dnode_t",
173 	    sizeof (dnode_t),
174 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
175 	kmem_cache_set_move(dnode_cache, dnode_move);
176 }
177 
178 void
179 dnode_fini(void)
180 {
181 	kmem_cache_destroy(dnode_cache);
182 	dnode_cache = NULL;
183 }
184 
185 
186 #ifdef ZFS_DEBUG
187 void
188 dnode_verify(dnode_t *dn)
189 {
190 	int drop_struct_lock = FALSE;
191 
192 	ASSERT(dn->dn_phys);
193 	ASSERT(dn->dn_objset);
194 	ASSERT(dn->dn_handle->dnh_dnode == dn);
195 
196 	ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES);
197 
198 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
199 		return;
200 
201 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
202 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
203 		drop_struct_lock = TRUE;
204 	}
205 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
206 		int i;
207 		ASSERT3U(dn->dn_indblkshift, >=, 0);
208 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
209 		if (dn->dn_datablkshift) {
210 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
211 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
212 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
213 		}
214 		ASSERT3U(dn->dn_nlevels, <=, 30);
215 		ASSERT3U(dn->dn_type, <=, DMU_OT_NUMTYPES);
216 		ASSERT3U(dn->dn_nblkptr, >=, 1);
217 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
218 		ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
219 		ASSERT3U(dn->dn_datablksz, ==,
220 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
221 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
222 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
223 		    dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
224 		for (i = 0; i < TXG_SIZE; i++) {
225 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
226 		}
227 	}
228 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
229 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
230 	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
231 	if (dn->dn_dbuf != NULL) {
232 		ASSERT3P(dn->dn_phys, ==,
233 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
234 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
235 	}
236 	if (drop_struct_lock)
237 		rw_exit(&dn->dn_struct_rwlock);
238 }
239 #endif
240 
241 void
242 dnode_byteswap(dnode_phys_t *dnp)
243 {
244 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
245 	int i;
246 
247 	if (dnp->dn_type == DMU_OT_NONE) {
248 		bzero(dnp, sizeof (dnode_phys_t));
249 		return;
250 	}
251 
252 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
253 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
254 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
255 	dnp->dn_used = BSWAP_64(dnp->dn_used);
256 
257 	/*
258 	 * dn_nblkptr is only one byte, so it's OK to read it in either
259 	 * byte order.  We can't read dn_bouslen.
260 	 */
261 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
262 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
263 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
264 		buf64[i] = BSWAP_64(buf64[i]);
265 
266 	/*
267 	 * OK to check dn_bonuslen for zero, because it won't matter if
268 	 * we have the wrong byte order.  This is necessary because the
269 	 * dnode dnode is smaller than a regular dnode.
270 	 */
271 	if (dnp->dn_bonuslen != 0) {
272 		/*
273 		 * Note that the bonus length calculated here may be
274 		 * longer than the actual bonus buffer.  This is because
275 		 * we always put the bonus buffer after the last block
276 		 * pointer (instead of packing it against the end of the
277 		 * dnode buffer).
278 		 */
279 		int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
280 		size_t len = DN_MAX_BONUSLEN - off;
281 		ASSERT3U(dnp->dn_bonustype, <, DMU_OT_NUMTYPES);
282 		dmu_ot[dnp->dn_bonustype].ot_byteswap(dnp->dn_bonus + off, len);
283 	}
284 
285 	/* Swap SPILL block if we have one */
286 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
287 		byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
288 
289 }
290 
291 void
292 dnode_buf_byteswap(void *vbuf, size_t size)
293 {
294 	dnode_phys_t *buf = vbuf;
295 	int i;
296 
297 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
298 	ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
299 
300 	size >>= DNODE_SHIFT;
301 	for (i = 0; i < size; i++) {
302 		dnode_byteswap(buf);
303 		buf++;
304 	}
305 }
306 
307 static int
308 free_range_compar(const void *node1, const void *node2)
309 {
310 	const free_range_t *rp1 = node1;
311 	const free_range_t *rp2 = node2;
312 
313 	if (rp1->fr_blkid < rp2->fr_blkid)
314 		return (-1);
315 	else if (rp1->fr_blkid > rp2->fr_blkid)
316 		return (1);
317 	else return (0);
318 }
319 
320 void
321 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
322 {
323 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
324 
325 	dnode_setdirty(dn, tx);
326 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
327 	ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
328 	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
329 	dn->dn_bonuslen = newsize;
330 	if (newsize == 0)
331 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
332 	else
333 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
334 	rw_exit(&dn->dn_struct_rwlock);
335 }
336 
337 void
338 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
339 {
340 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
341 	dnode_setdirty(dn, tx);
342 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
343 	dn->dn_bonustype = newtype;
344 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
345 	rw_exit(&dn->dn_struct_rwlock);
346 }
347 
348 void
349 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
350 {
351 	ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
352 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
353 	dnode_setdirty(dn, tx);
354 	dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
355 	dn->dn_have_spill = B_FALSE;
356 }
357 
358 static void
359 dnode_setdblksz(dnode_t *dn, int size)
360 {
361 	ASSERT3U(P2PHASE(size, SPA_MINBLOCKSIZE), ==, 0);
362 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
363 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
364 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
365 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
366 	dn->dn_datablksz = size;
367 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
368 	dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0;
369 }
370 
371 static dnode_t *
372 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
373     uint64_t object, dnode_handle_t *dnh)
374 {
375 	dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
376 
377 	ASSERT(!POINTER_IS_VALID(dn->dn_objset));
378 	dn->dn_moved = 0;
379 
380 	/*
381 	 * Defer setting dn_objset until the dnode is ready to be a candidate
382 	 * for the dnode_move() callback.
383 	 */
384 	dn->dn_object = object;
385 	dn->dn_dbuf = db;
386 	dn->dn_handle = dnh;
387 	dn->dn_phys = dnp;
388 
389 	if (dnp->dn_datablkszsec) {
390 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
391 	} else {
392 		dn->dn_datablksz = 0;
393 		dn->dn_datablkszsec = 0;
394 		dn->dn_datablkshift = 0;
395 	}
396 	dn->dn_indblkshift = dnp->dn_indblkshift;
397 	dn->dn_nlevels = dnp->dn_nlevels;
398 	dn->dn_type = dnp->dn_type;
399 	dn->dn_nblkptr = dnp->dn_nblkptr;
400 	dn->dn_checksum = dnp->dn_checksum;
401 	dn->dn_compress = dnp->dn_compress;
402 	dn->dn_bonustype = dnp->dn_bonustype;
403 	dn->dn_bonuslen = dnp->dn_bonuslen;
404 	dn->dn_maxblkid = dnp->dn_maxblkid;
405 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
406 	dn->dn_id_flags = 0;
407 
408 	dmu_zfetch_init(&dn->dn_zfetch, dn);
409 
410 	ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES);
411 
412 	mutex_enter(&os->os_lock);
413 	list_insert_head(&os->os_dnodes, dn);
414 	membar_producer();
415 	/*
416 	 * Everything else must be valid before assigning dn_objset makes the
417 	 * dnode eligible for dnode_move().
418 	 */
419 	dn->dn_objset = os;
420 	mutex_exit(&os->os_lock);
421 
422 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
423 	return (dn);
424 }
425 
426 /*
427  * Caller must be holding the dnode handle, which is released upon return.
428  */
429 static void
430 dnode_destroy(dnode_t *dn)
431 {
432 	objset_t *os = dn->dn_objset;
433 
434 	ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
435 
436 	mutex_enter(&os->os_lock);
437 	POINTER_INVALIDATE(&dn->dn_objset);
438 	list_remove(&os->os_dnodes, dn);
439 	mutex_exit(&os->os_lock);
440 
441 	/* the dnode can no longer move, so we can release the handle */
442 	zrl_remove(&dn->dn_handle->dnh_zrlock);
443 
444 	dn->dn_allocated_txg = 0;
445 	dn->dn_free_txg = 0;
446 	dn->dn_assigned_txg = 0;
447 
448 	dn->dn_dirtyctx = 0;
449 	if (dn->dn_dirtyctx_firstset != NULL) {
450 		kmem_free(dn->dn_dirtyctx_firstset, 1);
451 		dn->dn_dirtyctx_firstset = NULL;
452 	}
453 	if (dn->dn_bonus != NULL) {
454 		mutex_enter(&dn->dn_bonus->db_mtx);
455 		dbuf_evict(dn->dn_bonus);
456 		dn->dn_bonus = NULL;
457 	}
458 	dn->dn_zio = NULL;
459 
460 	dn->dn_have_spill = B_FALSE;
461 	dn->dn_oldused = 0;
462 	dn->dn_oldflags = 0;
463 	dn->dn_olduid = 0;
464 	dn->dn_oldgid = 0;
465 	dn->dn_newuid = 0;
466 	dn->dn_newgid = 0;
467 	dn->dn_id_flags = 0;
468 
469 	dmu_zfetch_rele(&dn->dn_zfetch);
470 	kmem_cache_free(dnode_cache, dn);
471 	arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
472 }
473 
474 void
475 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
476     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
477 {
478 	int i;
479 
480 	if (blocksize == 0)
481 		blocksize = 1 << zfs_default_bs;
482 	else if (blocksize > SPA_MAXBLOCKSIZE)
483 		blocksize = SPA_MAXBLOCKSIZE;
484 	else
485 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
486 
487 	if (ibs == 0)
488 		ibs = zfs_default_ibs;
489 
490 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
491 
492 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
493 	    dn->dn_object, tx->tx_txg, blocksize, ibs);
494 
495 	ASSERT(dn->dn_type == DMU_OT_NONE);
496 	ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
497 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
498 	ASSERT(ot != DMU_OT_NONE);
499 	ASSERT3U(ot, <, DMU_OT_NUMTYPES);
500 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
501 	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
502 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
503 	ASSERT3U(bonustype, <, DMU_OT_NUMTYPES);
504 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
505 	ASSERT(dn->dn_type == DMU_OT_NONE);
506 	ASSERT3U(dn->dn_maxblkid, ==, 0);
507 	ASSERT3U(dn->dn_allocated_txg, ==, 0);
508 	ASSERT3U(dn->dn_assigned_txg, ==, 0);
509 	ASSERT(refcount_is_zero(&dn->dn_tx_holds));
510 	ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
511 	ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
512 
513 	for (i = 0; i < TXG_SIZE; i++) {
514 		ASSERT3U(dn->dn_next_nblkptr[i], ==, 0);
515 		ASSERT3U(dn->dn_next_nlevels[i], ==, 0);
516 		ASSERT3U(dn->dn_next_indblkshift[i], ==, 0);
517 		ASSERT3U(dn->dn_next_bonuslen[i], ==, 0);
518 		ASSERT3U(dn->dn_next_bonustype[i], ==, 0);
519 		ASSERT3U(dn->dn_rm_spillblk[i], ==, 0);
520 		ASSERT3U(dn->dn_next_blksz[i], ==, 0);
521 		ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
522 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
523 		ASSERT3U(avl_numnodes(&dn->dn_ranges[i]), ==, 0);
524 	}
525 
526 	dn->dn_type = ot;
527 	dnode_setdblksz(dn, blocksize);
528 	dn->dn_indblkshift = ibs;
529 	dn->dn_nlevels = 1;
530 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
531 		dn->dn_nblkptr = 1;
532 	else
533 		dn->dn_nblkptr = 1 +
534 		    ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
535 	dn->dn_bonustype = bonustype;
536 	dn->dn_bonuslen = bonuslen;
537 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
538 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
539 	dn->dn_dirtyctx = 0;
540 
541 	dn->dn_free_txg = 0;
542 	if (dn->dn_dirtyctx_firstset) {
543 		kmem_free(dn->dn_dirtyctx_firstset, 1);
544 		dn->dn_dirtyctx_firstset = NULL;
545 	}
546 
547 	dn->dn_allocated_txg = tx->tx_txg;
548 	dn->dn_id_flags = 0;
549 
550 	dnode_setdirty(dn, tx);
551 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
552 	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
553 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
554 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
555 }
556 
557 void
558 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
559     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
560 {
561 	int nblkptr;
562 
563 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
564 	ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
565 	ASSERT3U(blocksize % SPA_MINBLOCKSIZE, ==, 0);
566 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
567 	ASSERT(tx->tx_txg != 0);
568 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
569 	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
570 	    (bonustype == DMU_OT_SA && bonuslen == 0));
571 	ASSERT3U(bonustype, <, DMU_OT_NUMTYPES);
572 	ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
573 
574 	/* clean up any unreferenced dbufs */
575 	dnode_evict_dbufs(dn);
576 
577 	dn->dn_id_flags = 0;
578 
579 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
580 	dnode_setdirty(dn, tx);
581 	if (dn->dn_datablksz != blocksize) {
582 		/* change blocksize */
583 		ASSERT(dn->dn_maxblkid == 0 &&
584 		    (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
585 		    dnode_block_freed(dn, 0)));
586 		dnode_setdblksz(dn, blocksize);
587 		dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
588 	}
589 	if (dn->dn_bonuslen != bonuslen)
590 		dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
591 
592 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
593 		nblkptr = 1;
594 	else
595 		nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
596 	if (dn->dn_bonustype != bonustype)
597 		dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
598 	if (dn->dn_nblkptr != nblkptr)
599 		dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
600 	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
601 		dbuf_rm_spill(dn, tx);
602 		dnode_rm_spill(dn, tx);
603 	}
604 	rw_exit(&dn->dn_struct_rwlock);
605 
606 	/* change type */
607 	dn->dn_type = ot;
608 
609 	/* change bonus size and type */
610 	mutex_enter(&dn->dn_mtx);
611 	dn->dn_bonustype = bonustype;
612 	dn->dn_bonuslen = bonuslen;
613 	dn->dn_nblkptr = nblkptr;
614 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
615 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
616 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
617 
618 	/* fix up the bonus db_size */
619 	if (dn->dn_bonus) {
620 		dn->dn_bonus->db.db_size =
621 		    DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
622 		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
623 	}
624 
625 	dn->dn_allocated_txg = tx->tx_txg;
626 	mutex_exit(&dn->dn_mtx);
627 }
628 
629 #ifdef	DNODE_STATS
630 static struct {
631 	uint64_t dms_dnode_invalid;
632 	uint64_t dms_dnode_recheck1;
633 	uint64_t dms_dnode_recheck2;
634 	uint64_t dms_dnode_special;
635 	uint64_t dms_dnode_handle;
636 	uint64_t dms_dnode_rwlock;
637 	uint64_t dms_dnode_active;
638 } dnode_move_stats;
639 #endif	/* DNODE_STATS */
640 
641 static void
642 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
643 {
644 	int i;
645 
646 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
647 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
648 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
649 	ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
650 
651 	/* Copy fields. */
652 	ndn->dn_objset = odn->dn_objset;
653 	ndn->dn_object = odn->dn_object;
654 	ndn->dn_dbuf = odn->dn_dbuf;
655 	ndn->dn_handle = odn->dn_handle;
656 	ndn->dn_phys = odn->dn_phys;
657 	ndn->dn_type = odn->dn_type;
658 	ndn->dn_bonuslen = odn->dn_bonuslen;
659 	ndn->dn_bonustype = odn->dn_bonustype;
660 	ndn->dn_nblkptr = odn->dn_nblkptr;
661 	ndn->dn_checksum = odn->dn_checksum;
662 	ndn->dn_compress = odn->dn_compress;
663 	ndn->dn_nlevels = odn->dn_nlevels;
664 	ndn->dn_indblkshift = odn->dn_indblkshift;
665 	ndn->dn_datablkshift = odn->dn_datablkshift;
666 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
667 	ndn->dn_datablksz = odn->dn_datablksz;
668 	ndn->dn_maxblkid = odn->dn_maxblkid;
669 	bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
670 	    sizeof (odn->dn_next_nblkptr));
671 	bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
672 	    sizeof (odn->dn_next_nlevels));
673 	bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
674 	    sizeof (odn->dn_next_indblkshift));
675 	bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
676 	    sizeof (odn->dn_next_bonustype));
677 	bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
678 	    sizeof (odn->dn_rm_spillblk));
679 	bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
680 	    sizeof (odn->dn_next_bonuslen));
681 	bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
682 	    sizeof (odn->dn_next_blksz));
683 	for (i = 0; i < TXG_SIZE; i++) {
684 		list_move_tail(&ndn->dn_dirty_records[i],
685 		    &odn->dn_dirty_records[i]);
686 	}
687 	bcopy(&odn->dn_ranges[0], &ndn->dn_ranges[0], sizeof (odn->dn_ranges));
688 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
689 	ndn->dn_free_txg = odn->dn_free_txg;
690 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
691 	ndn->dn_dirtyctx = odn->dn_dirtyctx;
692 	ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
693 	ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
694 	refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
695 	ASSERT(list_is_empty(&ndn->dn_dbufs));
696 	list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs);
697 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
698 	ndn->dn_bonus = odn->dn_bonus;
699 	ndn->dn_have_spill = odn->dn_have_spill;
700 	ndn->dn_zio = odn->dn_zio;
701 	ndn->dn_oldused = odn->dn_oldused;
702 	ndn->dn_oldflags = odn->dn_oldflags;
703 	ndn->dn_olduid = odn->dn_olduid;
704 	ndn->dn_oldgid = odn->dn_oldgid;
705 	ndn->dn_newuid = odn->dn_newuid;
706 	ndn->dn_newgid = odn->dn_newgid;
707 	ndn->dn_id_flags = odn->dn_id_flags;
708 	dmu_zfetch_init(&ndn->dn_zfetch, NULL);
709 	list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
710 	ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
711 	ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
712 	ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
713 
714 	/*
715 	 * Update back pointers. Updating the handle fixes the back pointer of
716 	 * every descendant dbuf as well as the bonus dbuf.
717 	 */
718 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
719 	ndn->dn_handle->dnh_dnode = ndn;
720 	if (ndn->dn_zfetch.zf_dnode == odn) {
721 		ndn->dn_zfetch.zf_dnode = ndn;
722 	}
723 
724 	/*
725 	 * Invalidate the original dnode by clearing all of its back pointers.
726 	 */
727 	odn->dn_dbuf = NULL;
728 	odn->dn_handle = NULL;
729 	list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t),
730 	    offsetof(dmu_buf_impl_t, db_link));
731 	odn->dn_dbufs_count = 0;
732 	odn->dn_bonus = NULL;
733 	odn->dn_zfetch.zf_dnode = NULL;
734 
735 	/*
736 	 * Set the low bit of the objset pointer to ensure that dnode_move()
737 	 * recognizes the dnode as invalid in any subsequent callback.
738 	 */
739 	POINTER_INVALIDATE(&odn->dn_objset);
740 
741 	/*
742 	 * Satisfy the destructor.
743 	 */
744 	for (i = 0; i < TXG_SIZE; i++) {
745 		list_create(&odn->dn_dirty_records[i],
746 		    sizeof (dbuf_dirty_record_t),
747 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
748 		odn->dn_ranges[i].avl_root = NULL;
749 		odn->dn_ranges[i].avl_numnodes = 0;
750 		odn->dn_next_nlevels[i] = 0;
751 		odn->dn_next_indblkshift[i] = 0;
752 		odn->dn_next_bonustype[i] = 0;
753 		odn->dn_rm_spillblk[i] = 0;
754 		odn->dn_next_bonuslen[i] = 0;
755 		odn->dn_next_blksz[i] = 0;
756 	}
757 	odn->dn_allocated_txg = 0;
758 	odn->dn_free_txg = 0;
759 	odn->dn_assigned_txg = 0;
760 	odn->dn_dirtyctx = 0;
761 	odn->dn_dirtyctx_firstset = NULL;
762 	odn->dn_have_spill = B_FALSE;
763 	odn->dn_zio = NULL;
764 	odn->dn_oldused = 0;
765 	odn->dn_oldflags = 0;
766 	odn->dn_olduid = 0;
767 	odn->dn_oldgid = 0;
768 	odn->dn_newuid = 0;
769 	odn->dn_newgid = 0;
770 	odn->dn_id_flags = 0;
771 
772 	/*
773 	 * Mark the dnode.
774 	 */
775 	ndn->dn_moved = 1;
776 	odn->dn_moved = (uint8_t)-1;
777 }
778 
779 #ifdef	_KERNEL
780 /*ARGSUSED*/
781 static kmem_cbrc_t
782 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
783 {
784 	dnode_t *odn = buf, *ndn = newbuf;
785 	objset_t *os;
786 	int64_t refcount;
787 	uint32_t dbufs;
788 
789 	/*
790 	 * The dnode is on the objset's list of known dnodes if the objset
791 	 * pointer is valid. We set the low bit of the objset pointer when
792 	 * freeing the dnode to invalidate it, and the memory patterns written
793 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
794 	 * A newly created dnode sets the objset pointer last of all to indicate
795 	 * that the dnode is known and in a valid state to be moved by this
796 	 * function.
797 	 */
798 	os = odn->dn_objset;
799 	if (!POINTER_IS_VALID(os)) {
800 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
801 		return (KMEM_CBRC_DONT_KNOW);
802 	}
803 
804 	/*
805 	 * Ensure that the objset does not go away during the move.
806 	 */
807 	rw_enter(&os_lock, RW_WRITER);
808 	if (os != odn->dn_objset) {
809 		rw_exit(&os_lock);
810 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
811 		return (KMEM_CBRC_DONT_KNOW);
812 	}
813 
814 	/*
815 	 * If the dnode is still valid, then so is the objset. We know that no
816 	 * valid objset can be freed while we hold os_lock, so we can safely
817 	 * ensure that the objset remains in use.
818 	 */
819 	mutex_enter(&os->os_lock);
820 
821 	/*
822 	 * Recheck the objset pointer in case the dnode was removed just before
823 	 * acquiring the lock.
824 	 */
825 	if (os != odn->dn_objset) {
826 		mutex_exit(&os->os_lock);
827 		rw_exit(&os_lock);
828 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
829 		return (KMEM_CBRC_DONT_KNOW);
830 	}
831 
832 	/*
833 	 * At this point we know that as long as we hold os->os_lock, the dnode
834 	 * cannot be freed and fields within the dnode can be safely accessed.
835 	 * The objset listing this dnode cannot go away as long as this dnode is
836 	 * on its list.
837 	 */
838 	rw_exit(&os_lock);
839 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
840 		mutex_exit(&os->os_lock);
841 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
842 		return (KMEM_CBRC_NO);
843 	}
844 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
845 
846 	/*
847 	 * Lock the dnode handle to prevent the dnode from obtaining any new
848 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
849 	 * from accessing the dnode, so that we can discount their holds. The
850 	 * handle is safe to access because we know that while the dnode cannot
851 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
852 	 * safely move any dnode referenced only by dbufs.
853 	 */
854 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
855 		mutex_exit(&os->os_lock);
856 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
857 		return (KMEM_CBRC_LATER);
858 	}
859 
860 	/*
861 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
862 	 * We need to guarantee that there is a hold for every dbuf in order to
863 	 * determine whether the dnode is actively referenced. Falsely matching
864 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
865 	 * that a thread already having an active dnode hold is about to add a
866 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
867 	 * progress.
868 	 */
869 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
870 		zrl_exit(&odn->dn_handle->dnh_zrlock);
871 		mutex_exit(&os->os_lock);
872 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
873 		return (KMEM_CBRC_LATER);
874 	}
875 
876 	/*
877 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
878 	 * case, the dbuf count is decremented under the handle lock before the
879 	 * dbuf's hold is released. This order ensures that if we count the hold
880 	 * after the dbuf is removed but before its hold is released, we will
881 	 * treat the unmatched hold as active and exit safely. If we count the
882 	 * hold before the dbuf is removed, the hold is discounted, and the
883 	 * removal is blocked until the move completes.
884 	 */
885 	refcount = refcount_count(&odn->dn_holds);
886 	ASSERT(refcount >= 0);
887 	dbufs = odn->dn_dbufs_count;
888 
889 	/* We can't have more dbufs than dnode holds. */
890 	ASSERT3U(dbufs, <=, refcount);
891 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
892 	    uint32_t, dbufs);
893 
894 	if (refcount > dbufs) {
895 		rw_exit(&odn->dn_struct_rwlock);
896 		zrl_exit(&odn->dn_handle->dnh_zrlock);
897 		mutex_exit(&os->os_lock);
898 		DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
899 		return (KMEM_CBRC_LATER);
900 	}
901 
902 	rw_exit(&odn->dn_struct_rwlock);
903 
904 	/*
905 	 * At this point we know that anyone with a hold on the dnode is not
906 	 * actively referencing it. The dnode is known and in a valid state to
907 	 * move. We're holding the locks needed to execute the critical section.
908 	 */
909 	dnode_move_impl(odn, ndn);
910 
911 	list_link_replace(&odn->dn_link, &ndn->dn_link);
912 	/* If the dnode was safe to move, the refcount cannot have changed. */
913 	ASSERT(refcount == refcount_count(&ndn->dn_holds));
914 	ASSERT(dbufs == ndn->dn_dbufs_count);
915 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
916 	mutex_exit(&os->os_lock);
917 
918 	return (KMEM_CBRC_YES);
919 }
920 #endif	/* _KERNEL */
921 
922 void
923 dnode_special_close(dnode_handle_t *dnh)
924 {
925 	dnode_t *dn = dnh->dnh_dnode;
926 
927 	/*
928 	 * Wait for final references to the dnode to clear.  This can
929 	 * only happen if the arc is asyncronously evicting state that
930 	 * has a hold on this dnode while we are trying to evict this
931 	 * dnode.
932 	 */
933 	while (refcount_count(&dn->dn_holds) > 0)
934 		delay(1);
935 	zrl_add(&dnh->dnh_zrlock);
936 	dnode_destroy(dn); /* implicit zrl_remove() */
937 	zrl_destroy(&dnh->dnh_zrlock);
938 	dnh->dnh_dnode = NULL;
939 }
940 
941 dnode_t *
942 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
943     dnode_handle_t *dnh)
944 {
945 	dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
946 	dnh->dnh_dnode = dn;
947 	zrl_init(&dnh->dnh_zrlock);
948 	DNODE_VERIFY(dn);
949 	return (dn);
950 }
951 
952 static void
953 dnode_buf_pageout(dmu_buf_t *db, void *arg)
954 {
955 	dnode_children_t *children_dnodes = arg;
956 	int i;
957 	int epb = db->db_size >> DNODE_SHIFT;
958 
959 	ASSERT(epb == children_dnodes->dnc_count);
960 
961 	for (i = 0; i < epb; i++) {
962 		dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
963 		dnode_t *dn;
964 
965 		/*
966 		 * The dnode handle lock guards against the dnode moving to
967 		 * another valid address, so there is no need here to guard
968 		 * against changes to or from NULL.
969 		 */
970 		if (dnh->dnh_dnode == NULL) {
971 			zrl_destroy(&dnh->dnh_zrlock);
972 			continue;
973 		}
974 
975 		zrl_add(&dnh->dnh_zrlock);
976 		dn = dnh->dnh_dnode;
977 		/*
978 		 * If there are holds on this dnode, then there should
979 		 * be holds on the dnode's containing dbuf as well; thus
980 		 * it wouldn't be eligible for eviction and this function
981 		 * would not have been called.
982 		 */
983 		ASSERT(refcount_is_zero(&dn->dn_holds));
984 		ASSERT(refcount_is_zero(&dn->dn_tx_holds));
985 
986 		dnode_destroy(dn); /* implicit zrl_remove() */
987 		zrl_destroy(&dnh->dnh_zrlock);
988 		dnh->dnh_dnode = NULL;
989 	}
990 	kmem_free(children_dnodes, sizeof (dnode_children_t) +
991 	    (epb - 1) * sizeof (dnode_handle_t));
992 }
993 
994 /*
995  * errors:
996  * EINVAL - invalid object number.
997  * EIO - i/o error.
998  * succeeds even for free dnodes.
999  */
1000 int
1001 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1002     void *tag, dnode_t **dnp)
1003 {
1004 	int epb, idx, err;
1005 	int drop_struct_lock = FALSE;
1006 	int type;
1007 	uint64_t blk;
1008 	dnode_t *mdn, *dn;
1009 	dmu_buf_impl_t *db;
1010 	dnode_children_t *children_dnodes;
1011 	dnode_handle_t *dnh;
1012 
1013 	/*
1014 	 * If you are holding the spa config lock as writer, you shouldn't
1015 	 * be asking the DMU to do *anything* unless it's the root pool
1016 	 * which may require us to read from the root filesystem while
1017 	 * holding some (not all) of the locks as writer.
1018 	 */
1019 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1020 	    (spa_is_root(os->os_spa) &&
1021 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1022 
1023 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1024 		dn = (object == DMU_USERUSED_OBJECT) ?
1025 		    DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1026 		if (dn == NULL)
1027 			return (ENOENT);
1028 		type = dn->dn_type;
1029 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1030 			return (ENOENT);
1031 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1032 			return (EEXIST);
1033 		DNODE_VERIFY(dn);
1034 		(void) refcount_add(&dn->dn_holds, tag);
1035 		*dnp = dn;
1036 		return (0);
1037 	}
1038 
1039 	if (object == 0 || object >= DN_MAX_OBJECT)
1040 		return (EINVAL);
1041 
1042 	mdn = DMU_META_DNODE(os);
1043 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1044 
1045 	DNODE_VERIFY(mdn);
1046 
1047 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1048 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1049 		drop_struct_lock = TRUE;
1050 	}
1051 
1052 	blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1053 
1054 	db = dbuf_hold(mdn, blk, FTAG);
1055 	if (drop_struct_lock)
1056 		rw_exit(&mdn->dn_struct_rwlock);
1057 	if (db == NULL)
1058 		return (EIO);
1059 	err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1060 	if (err) {
1061 		dbuf_rele(db, FTAG);
1062 		return (err);
1063 	}
1064 
1065 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1066 	epb = db->db.db_size >> DNODE_SHIFT;
1067 
1068 	idx = object & (epb-1);
1069 
1070 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1071 	children_dnodes = dmu_buf_get_user(&db->db);
1072 	if (children_dnodes == NULL) {
1073 		int i;
1074 		dnode_children_t *winner;
1075 		children_dnodes = kmem_alloc(sizeof (dnode_children_t) +
1076 		    (epb - 1) * sizeof (dnode_handle_t), KM_SLEEP);
1077 		children_dnodes->dnc_count = epb;
1078 		dnh = &children_dnodes->dnc_children[0];
1079 		for (i = 0; i < epb; i++) {
1080 			zrl_init(&dnh[i].dnh_zrlock);
1081 			dnh[i].dnh_dnode = NULL;
1082 		}
1083 		if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1084 		    dnode_buf_pageout)) {
1085 			kmem_free(children_dnodes, sizeof (dnode_children_t) +
1086 			    (epb - 1) * sizeof (dnode_handle_t));
1087 			children_dnodes = winner;
1088 		}
1089 	}
1090 	ASSERT(children_dnodes->dnc_count == epb);
1091 
1092 	dnh = &children_dnodes->dnc_children[idx];
1093 	zrl_add(&dnh->dnh_zrlock);
1094 	if ((dn = dnh->dnh_dnode) == NULL) {
1095 		dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1096 		dnode_t *winner;
1097 
1098 		dn = dnode_create(os, phys, db, object, dnh);
1099 		winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1100 		if (winner != NULL) {
1101 			zrl_add(&dnh->dnh_zrlock);
1102 			dnode_destroy(dn); /* implicit zrl_remove() */
1103 			dn = winner;
1104 		}
1105 	}
1106 
1107 	mutex_enter(&dn->dn_mtx);
1108 	type = dn->dn_type;
1109 	if (dn->dn_free_txg ||
1110 	    ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1111 	    ((flag & DNODE_MUST_BE_FREE) &&
1112 	    (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1113 		mutex_exit(&dn->dn_mtx);
1114 		zrl_remove(&dnh->dnh_zrlock);
1115 		dbuf_rele(db, FTAG);
1116 		return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1117 	}
1118 	mutex_exit(&dn->dn_mtx);
1119 
1120 	if (refcount_add(&dn->dn_holds, tag) == 1)
1121 		dbuf_add_ref(db, dnh);
1122 	/* Now we can rely on the hold to prevent the dnode from moving. */
1123 	zrl_remove(&dnh->dnh_zrlock);
1124 
1125 	DNODE_VERIFY(dn);
1126 	ASSERT3P(dn->dn_dbuf, ==, db);
1127 	ASSERT3U(dn->dn_object, ==, object);
1128 	dbuf_rele(db, FTAG);
1129 
1130 	*dnp = dn;
1131 	return (0);
1132 }
1133 
1134 /*
1135  * Return held dnode if the object is allocated, NULL if not.
1136  */
1137 int
1138 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1139 {
1140 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1141 }
1142 
1143 /*
1144  * Can only add a reference if there is already at least one
1145  * reference on the dnode.  Returns FALSE if unable to add a
1146  * new reference.
1147  */
1148 boolean_t
1149 dnode_add_ref(dnode_t *dn, void *tag)
1150 {
1151 	mutex_enter(&dn->dn_mtx);
1152 	if (refcount_is_zero(&dn->dn_holds)) {
1153 		mutex_exit(&dn->dn_mtx);
1154 		return (FALSE);
1155 	}
1156 	VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1157 	mutex_exit(&dn->dn_mtx);
1158 	return (TRUE);
1159 }
1160 
1161 void
1162 dnode_rele(dnode_t *dn, void *tag)
1163 {
1164 	uint64_t refs;
1165 	/* Get while the hold prevents the dnode from moving. */
1166 	dmu_buf_impl_t *db = dn->dn_dbuf;
1167 	dnode_handle_t *dnh = dn->dn_handle;
1168 
1169 	mutex_enter(&dn->dn_mtx);
1170 	refs = refcount_remove(&dn->dn_holds, tag);
1171 	mutex_exit(&dn->dn_mtx);
1172 
1173 	/*
1174 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1175 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1176 	 * prevent the dnode from moving, since releasing the last hold could
1177 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1178 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1179 	 * other direct or indirect hold on the dnode must first drop the dnode
1180 	 * handle.
1181 	 */
1182 	ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1183 
1184 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1185 	if (refs == 0 && db != NULL) {
1186 		/*
1187 		 * Another thread could add a hold to the dnode handle in
1188 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1189 		 * hold on the parent dbuf prevents the handle from being
1190 		 * destroyed, the hold on the handle is OK. We can't yet assert
1191 		 * that the handle has zero references, but that will be
1192 		 * asserted anyway when the handle gets destroyed.
1193 		 */
1194 		dbuf_rele(db, dnh);
1195 	}
1196 }
1197 
1198 void
1199 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1200 {
1201 	objset_t *os = dn->dn_objset;
1202 	uint64_t txg = tx->tx_txg;
1203 
1204 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1205 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1206 		return;
1207 	}
1208 
1209 	DNODE_VERIFY(dn);
1210 
1211 #ifdef ZFS_DEBUG
1212 	mutex_enter(&dn->dn_mtx);
1213 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1214 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1215 	mutex_exit(&dn->dn_mtx);
1216 #endif
1217 
1218 	/*
1219 	 * Determine old uid/gid when necessary
1220 	 */
1221 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1222 
1223 	mutex_enter(&os->os_lock);
1224 
1225 	/*
1226 	 * If we are already marked dirty, we're done.
1227 	 */
1228 	if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1229 		mutex_exit(&os->os_lock);
1230 		return;
1231 	}
1232 
1233 	ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
1234 	ASSERT(dn->dn_datablksz != 0);
1235 	ASSERT3U(dn->dn_next_bonuslen[txg&TXG_MASK], ==, 0);
1236 	ASSERT3U(dn->dn_next_blksz[txg&TXG_MASK], ==, 0);
1237 	ASSERT3U(dn->dn_next_bonustype[txg&TXG_MASK], ==, 0);
1238 
1239 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1240 	    dn->dn_object, txg);
1241 
1242 	if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1243 		list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1244 	} else {
1245 		list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1246 	}
1247 
1248 	mutex_exit(&os->os_lock);
1249 
1250 	/*
1251 	 * The dnode maintains a hold on its containing dbuf as
1252 	 * long as there are holds on it.  Each instantiated child
1253 	 * dbuf maintains a hold on the dnode.  When the last child
1254 	 * drops its hold, the dnode will drop its hold on the
1255 	 * containing dbuf. We add a "dirty hold" here so that the
1256 	 * dnode will hang around after we finish processing its
1257 	 * children.
1258 	 */
1259 	VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1260 
1261 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1262 
1263 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1264 }
1265 
1266 void
1267 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1268 {
1269 	int txgoff = tx->tx_txg & TXG_MASK;
1270 
1271 	dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1272 
1273 	/* we should be the only holder... hopefully */
1274 	/* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1275 
1276 	mutex_enter(&dn->dn_mtx);
1277 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1278 		mutex_exit(&dn->dn_mtx);
1279 		return;
1280 	}
1281 	dn->dn_free_txg = tx->tx_txg;
1282 	mutex_exit(&dn->dn_mtx);
1283 
1284 	/*
1285 	 * If the dnode is already dirty, it needs to be moved from
1286 	 * the dirty list to the free list.
1287 	 */
1288 	mutex_enter(&dn->dn_objset->os_lock);
1289 	if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1290 		list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1291 		list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1292 		mutex_exit(&dn->dn_objset->os_lock);
1293 	} else {
1294 		mutex_exit(&dn->dn_objset->os_lock);
1295 		dnode_setdirty(dn, tx);
1296 	}
1297 }
1298 
1299 /*
1300  * Try to change the block size for the indicated dnode.  This can only
1301  * succeed if there are no blocks allocated or dirty beyond first block
1302  */
1303 int
1304 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1305 {
1306 	dmu_buf_impl_t *db, *db_next;
1307 	int err;
1308 
1309 	if (size == 0)
1310 		size = SPA_MINBLOCKSIZE;
1311 	if (size > SPA_MAXBLOCKSIZE)
1312 		size = SPA_MAXBLOCKSIZE;
1313 	else
1314 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1315 
1316 	if (ibs == dn->dn_indblkshift)
1317 		ibs = 0;
1318 
1319 	if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1320 		return (0);
1321 
1322 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1323 
1324 	/* Check for any allocated blocks beyond the first */
1325 	if (dn->dn_phys->dn_maxblkid != 0)
1326 		goto fail;
1327 
1328 	mutex_enter(&dn->dn_dbufs_mtx);
1329 	for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
1330 		db_next = list_next(&dn->dn_dbufs, db);
1331 
1332 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1333 		    db->db_blkid != DMU_SPILL_BLKID) {
1334 			mutex_exit(&dn->dn_dbufs_mtx);
1335 			goto fail;
1336 		}
1337 	}
1338 	mutex_exit(&dn->dn_dbufs_mtx);
1339 
1340 	if (ibs && dn->dn_nlevels != 1)
1341 		goto fail;
1342 
1343 	/* resize the old block */
1344 	err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1345 	if (err == 0)
1346 		dbuf_new_size(db, size, tx);
1347 	else if (err != ENOENT)
1348 		goto fail;
1349 
1350 	dnode_setdblksz(dn, size);
1351 	dnode_setdirty(dn, tx);
1352 	dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1353 	if (ibs) {
1354 		dn->dn_indblkshift = ibs;
1355 		dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1356 	}
1357 	/* rele after we have fixed the blocksize in the dnode */
1358 	if (db)
1359 		dbuf_rele(db, FTAG);
1360 
1361 	rw_exit(&dn->dn_struct_rwlock);
1362 	return (0);
1363 
1364 fail:
1365 	rw_exit(&dn->dn_struct_rwlock);
1366 	return (ENOTSUP);
1367 }
1368 
1369 /* read-holding callers must not rely on the lock being continuously held */
1370 void
1371 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1372 {
1373 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
1374 	int epbs, new_nlevels;
1375 	uint64_t sz;
1376 
1377 	ASSERT(blkid != DMU_BONUS_BLKID);
1378 
1379 	ASSERT(have_read ?
1380 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
1381 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
1382 
1383 	/*
1384 	 * if we have a read-lock, check to see if we need to do any work
1385 	 * before upgrading to a write-lock.
1386 	 */
1387 	if (have_read) {
1388 		if (blkid <= dn->dn_maxblkid)
1389 			return;
1390 
1391 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1392 			rw_exit(&dn->dn_struct_rwlock);
1393 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1394 		}
1395 	}
1396 
1397 	if (blkid <= dn->dn_maxblkid)
1398 		goto out;
1399 
1400 	dn->dn_maxblkid = blkid;
1401 
1402 	/*
1403 	 * Compute the number of levels necessary to support the new maxblkid.
1404 	 */
1405 	new_nlevels = 1;
1406 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1407 	for (sz = dn->dn_nblkptr;
1408 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1409 		new_nlevels++;
1410 
1411 	if (new_nlevels > dn->dn_nlevels) {
1412 		int old_nlevels = dn->dn_nlevels;
1413 		dmu_buf_impl_t *db;
1414 		list_t *list;
1415 		dbuf_dirty_record_t *new, *dr, *dr_next;
1416 
1417 		dn->dn_nlevels = new_nlevels;
1418 
1419 		ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1420 		dn->dn_next_nlevels[txgoff] = new_nlevels;
1421 
1422 		/* dirty the left indirects */
1423 		db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1424 		ASSERT(db != NULL);
1425 		new = dbuf_dirty(db, tx);
1426 		dbuf_rele(db, FTAG);
1427 
1428 		/* transfer the dirty records to the new indirect */
1429 		mutex_enter(&dn->dn_mtx);
1430 		mutex_enter(&new->dt.di.dr_mtx);
1431 		list = &dn->dn_dirty_records[txgoff];
1432 		for (dr = list_head(list); dr; dr = dr_next) {
1433 			dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1434 			if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1435 			    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1436 			    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1437 				ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1438 				list_remove(&dn->dn_dirty_records[txgoff], dr);
1439 				list_insert_tail(&new->dt.di.dr_children, dr);
1440 				dr->dr_parent = new;
1441 			}
1442 		}
1443 		mutex_exit(&new->dt.di.dr_mtx);
1444 		mutex_exit(&dn->dn_mtx);
1445 	}
1446 
1447 out:
1448 	if (have_read)
1449 		rw_downgrade(&dn->dn_struct_rwlock);
1450 }
1451 
1452 void
1453 dnode_clear_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
1454 {
1455 	avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1456 	avl_index_t where;
1457 	free_range_t *rp;
1458 	free_range_t rp_tofind;
1459 	uint64_t endblk = blkid + nblks;
1460 
1461 	ASSERT(MUTEX_HELD(&dn->dn_mtx));
1462 	ASSERT(nblks <= UINT64_MAX - blkid); /* no overflow */
1463 
1464 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1465 	    blkid, nblks, tx->tx_txg);
1466 	rp_tofind.fr_blkid = blkid;
1467 	rp = avl_find(tree, &rp_tofind, &where);
1468 	if (rp == NULL)
1469 		rp = avl_nearest(tree, where, AVL_BEFORE);
1470 	if (rp == NULL)
1471 		rp = avl_nearest(tree, where, AVL_AFTER);
1472 
1473 	while (rp && (rp->fr_blkid <= blkid + nblks)) {
1474 		uint64_t fr_endblk = rp->fr_blkid + rp->fr_nblks;
1475 		free_range_t *nrp = AVL_NEXT(tree, rp);
1476 
1477 		if (blkid <= rp->fr_blkid && endblk >= fr_endblk) {
1478 			/* clear this entire range */
1479 			avl_remove(tree, rp);
1480 			kmem_free(rp, sizeof (free_range_t));
1481 		} else if (blkid <= rp->fr_blkid &&
1482 		    endblk > rp->fr_blkid && endblk < fr_endblk) {
1483 			/* clear the beginning of this range */
1484 			rp->fr_blkid = endblk;
1485 			rp->fr_nblks = fr_endblk - endblk;
1486 		} else if (blkid > rp->fr_blkid && blkid < fr_endblk &&
1487 		    endblk >= fr_endblk) {
1488 			/* clear the end of this range */
1489 			rp->fr_nblks = blkid - rp->fr_blkid;
1490 		} else if (blkid > rp->fr_blkid && endblk < fr_endblk) {
1491 			/* clear a chunk out of this range */
1492 			free_range_t *new_rp =
1493 			    kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1494 
1495 			new_rp->fr_blkid = endblk;
1496 			new_rp->fr_nblks = fr_endblk - endblk;
1497 			avl_insert_here(tree, new_rp, rp, AVL_AFTER);
1498 			rp->fr_nblks = blkid - rp->fr_blkid;
1499 		}
1500 		/* there may be no overlap */
1501 		rp = nrp;
1502 	}
1503 }
1504 
1505 void
1506 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1507 {
1508 	dmu_buf_impl_t *db;
1509 	uint64_t blkoff, blkid, nblks;
1510 	int blksz, blkshift, head, tail;
1511 	int trunc = FALSE;
1512 	int epbs;
1513 
1514 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1515 	blksz = dn->dn_datablksz;
1516 	blkshift = dn->dn_datablkshift;
1517 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1518 
1519 	if (len == -1ULL) {
1520 		len = UINT64_MAX - off;
1521 		trunc = TRUE;
1522 	}
1523 
1524 	/*
1525 	 * First, block align the region to free:
1526 	 */
1527 	if (ISP2(blksz)) {
1528 		head = P2NPHASE(off, blksz);
1529 		blkoff = P2PHASE(off, blksz);
1530 		if ((off >> blkshift) > dn->dn_maxblkid)
1531 			goto out;
1532 	} else {
1533 		ASSERT(dn->dn_maxblkid == 0);
1534 		if (off == 0 && len >= blksz) {
1535 			/* Freeing the whole block; fast-track this request */
1536 			blkid = 0;
1537 			nblks = 1;
1538 			goto done;
1539 		} else if (off >= blksz) {
1540 			/* Freeing past end-of-data */
1541 			goto out;
1542 		} else {
1543 			/* Freeing part of the block. */
1544 			head = blksz - off;
1545 			ASSERT3U(head, >, 0);
1546 		}
1547 		blkoff = off;
1548 	}
1549 	/* zero out any partial block data at the start of the range */
1550 	if (head) {
1551 		ASSERT3U(blkoff + head, ==, blksz);
1552 		if (len < head)
1553 			head = len;
1554 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1555 		    FTAG, &db) == 0) {
1556 			caddr_t data;
1557 
1558 			/* don't dirty if it isn't on disk and isn't dirty */
1559 			if (db->db_last_dirty ||
1560 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1561 				rw_exit(&dn->dn_struct_rwlock);
1562 				dbuf_will_dirty(db, tx);
1563 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1564 				data = db->db.db_data;
1565 				bzero(data + blkoff, head);
1566 			}
1567 			dbuf_rele(db, FTAG);
1568 		}
1569 		off += head;
1570 		len -= head;
1571 	}
1572 
1573 	/* If the range was less than one block, we're done */
1574 	if (len == 0)
1575 		goto out;
1576 
1577 	/* If the remaining range is past end of file, we're done */
1578 	if ((off >> blkshift) > dn->dn_maxblkid)
1579 		goto out;
1580 
1581 	ASSERT(ISP2(blksz));
1582 	if (trunc)
1583 		tail = 0;
1584 	else
1585 		tail = P2PHASE(len, blksz);
1586 
1587 	ASSERT3U(P2PHASE(off, blksz), ==, 0);
1588 	/* zero out any partial block data at the end of the range */
1589 	if (tail) {
1590 		if (len < tail)
1591 			tail = len;
1592 		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1593 		    TRUE, FTAG, &db) == 0) {
1594 			/* don't dirty if not on disk and not dirty */
1595 			if (db->db_last_dirty ||
1596 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1597 				rw_exit(&dn->dn_struct_rwlock);
1598 				dbuf_will_dirty(db, tx);
1599 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1600 				bzero(db->db.db_data, tail);
1601 			}
1602 			dbuf_rele(db, FTAG);
1603 		}
1604 		len -= tail;
1605 	}
1606 
1607 	/* If the range did not include a full block, we are done */
1608 	if (len == 0)
1609 		goto out;
1610 
1611 	ASSERT(IS_P2ALIGNED(off, blksz));
1612 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1613 	blkid = off >> blkshift;
1614 	nblks = len >> blkshift;
1615 	if (trunc)
1616 		nblks += 1;
1617 
1618 	/*
1619 	 * Read in and mark all the level-1 indirects dirty,
1620 	 * so that they will stay in memory until syncing phase.
1621 	 * Always dirty the first and last indirect to make sure
1622 	 * we dirty all the partial indirects.
1623 	 */
1624 	if (dn->dn_nlevels > 1) {
1625 		uint64_t i, first, last;
1626 		int shift = epbs + dn->dn_datablkshift;
1627 
1628 		first = blkid >> epbs;
1629 		if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1630 			dbuf_will_dirty(db, tx);
1631 			dbuf_rele(db, FTAG);
1632 		}
1633 		if (trunc)
1634 			last = dn->dn_maxblkid >> epbs;
1635 		else
1636 			last = (blkid + nblks - 1) >> epbs;
1637 		if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1638 			dbuf_will_dirty(db, tx);
1639 			dbuf_rele(db, FTAG);
1640 		}
1641 		for (i = first + 1; i < last; i++) {
1642 			uint64_t ibyte = i << shift;
1643 			int err;
1644 
1645 			err = dnode_next_offset(dn,
1646 			    DNODE_FIND_HAVELOCK, &ibyte, 1, 1, 0);
1647 			i = ibyte >> shift;
1648 			if (err == ESRCH || i >= last)
1649 				break;
1650 			ASSERT(err == 0);
1651 			db = dbuf_hold_level(dn, 1, i, FTAG);
1652 			if (db) {
1653 				dbuf_will_dirty(db, tx);
1654 				dbuf_rele(db, FTAG);
1655 			}
1656 		}
1657 	}
1658 done:
1659 	/*
1660 	 * Add this range to the dnode range list.
1661 	 * We will finish up this free operation in the syncing phase.
1662 	 */
1663 	mutex_enter(&dn->dn_mtx);
1664 	dnode_clear_range(dn, blkid, nblks, tx);
1665 	{
1666 		free_range_t *rp, *found;
1667 		avl_index_t where;
1668 		avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1669 
1670 		/* Add new range to dn_ranges */
1671 		rp = kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1672 		rp->fr_blkid = blkid;
1673 		rp->fr_nblks = nblks;
1674 		found = avl_find(tree, rp, &where);
1675 		ASSERT(found == NULL);
1676 		avl_insert(tree, rp, where);
1677 		dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1678 		    blkid, nblks, tx->tx_txg);
1679 	}
1680 	mutex_exit(&dn->dn_mtx);
1681 
1682 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1683 	dnode_setdirty(dn, tx);
1684 out:
1685 	if (trunc && dn->dn_maxblkid >= (off >> blkshift))
1686 		dn->dn_maxblkid = (off >> blkshift ? (off >> blkshift) - 1 : 0);
1687 
1688 	rw_exit(&dn->dn_struct_rwlock);
1689 }
1690 
1691 static boolean_t
1692 dnode_spill_freed(dnode_t *dn)
1693 {
1694 	int i;
1695 
1696 	mutex_enter(&dn->dn_mtx);
1697 	for (i = 0; i < TXG_SIZE; i++) {
1698 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1699 			break;
1700 	}
1701 	mutex_exit(&dn->dn_mtx);
1702 	return (i < TXG_SIZE);
1703 }
1704 
1705 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1706 uint64_t
1707 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1708 {
1709 	free_range_t range_tofind;
1710 	void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1711 	int i;
1712 
1713 	if (blkid == DMU_BONUS_BLKID)
1714 		return (FALSE);
1715 
1716 	/*
1717 	 * If we're in the process of opening the pool, dp will not be
1718 	 * set yet, but there shouldn't be anything dirty.
1719 	 */
1720 	if (dp == NULL)
1721 		return (FALSE);
1722 
1723 	if (dn->dn_free_txg)
1724 		return (TRUE);
1725 
1726 	if (blkid == DMU_SPILL_BLKID)
1727 		return (dnode_spill_freed(dn));
1728 
1729 	range_tofind.fr_blkid = blkid;
1730 	mutex_enter(&dn->dn_mtx);
1731 	for (i = 0; i < TXG_SIZE; i++) {
1732 		free_range_t *range_found;
1733 		avl_index_t idx;
1734 
1735 		range_found = avl_find(&dn->dn_ranges[i], &range_tofind, &idx);
1736 		if (range_found) {
1737 			ASSERT(range_found->fr_nblks > 0);
1738 			break;
1739 		}
1740 		range_found = avl_nearest(&dn->dn_ranges[i], idx, AVL_BEFORE);
1741 		if (range_found &&
1742 		    range_found->fr_blkid + range_found->fr_nblks > blkid)
1743 			break;
1744 	}
1745 	mutex_exit(&dn->dn_mtx);
1746 	return (i < TXG_SIZE);
1747 }
1748 
1749 /* call from syncing context when we actually write/free space for this dnode */
1750 void
1751 dnode_diduse_space(dnode_t *dn, int64_t delta)
1752 {
1753 	uint64_t space;
1754 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1755 	    dn, dn->dn_phys,
1756 	    (u_longlong_t)dn->dn_phys->dn_used,
1757 	    (longlong_t)delta);
1758 
1759 	mutex_enter(&dn->dn_mtx);
1760 	space = DN_USED_BYTES(dn->dn_phys);
1761 	if (delta > 0) {
1762 		ASSERT3U(space + delta, >=, space); /* no overflow */
1763 	} else {
1764 		ASSERT3U(space, >=, -delta); /* no underflow */
1765 	}
1766 	space += delta;
1767 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1768 		ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1769 		ASSERT3U(P2PHASE(space, 1<<DEV_BSHIFT), ==, 0);
1770 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1771 	} else {
1772 		dn->dn_phys->dn_used = space;
1773 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1774 	}
1775 	mutex_exit(&dn->dn_mtx);
1776 }
1777 
1778 /*
1779  * Call when we think we're going to write/free space in open context.
1780  * Be conservative (ie. OK to write less than this or free more than
1781  * this, but don't write more or free less).
1782  */
1783 void
1784 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1785 {
1786 	objset_t *os = dn->dn_objset;
1787 	dsl_dataset_t *ds = os->os_dsl_dataset;
1788 
1789 	if (space > 0)
1790 		space = spa_get_asize(os->os_spa, space);
1791 
1792 	if (ds)
1793 		dsl_dir_willuse_space(ds->ds_dir, space, tx);
1794 
1795 	dmu_tx_willuse_space(tx, space);
1796 }
1797 
1798 /*
1799  * This function scans a block at the indicated "level" looking for
1800  * a hole or data (depending on 'flags').  If level > 0, then we are
1801  * scanning an indirect block looking at its pointers.  If level == 0,
1802  * then we are looking at a block of dnodes.  If we don't find what we
1803  * are looking for in the block, we return ESRCH.  Otherwise, return
1804  * with *offset pointing to the beginning (if searching forwards) or
1805  * end (if searching backwards) of the range covered by the block
1806  * pointer we matched on (or dnode).
1807  *
1808  * The basic search algorithm used below by dnode_next_offset() is to
1809  * use this function to search up the block tree (widen the search) until
1810  * we find something (i.e., we don't return ESRCH) and then search back
1811  * down the tree (narrow the search) until we reach our original search
1812  * level.
1813  */
1814 static int
1815 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1816 	int lvl, uint64_t blkfill, uint64_t txg)
1817 {
1818 	dmu_buf_impl_t *db = NULL;
1819 	void *data = NULL;
1820 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1821 	uint64_t epb = 1ULL << epbs;
1822 	uint64_t minfill, maxfill;
1823 	boolean_t hole;
1824 	int i, inc, error, span;
1825 
1826 	dprintf("probing object %llu offset %llx level %d of %u\n",
1827 	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1828 
1829 	hole = ((flags & DNODE_FIND_HOLE) != 0);
1830 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1831 	ASSERT(txg == 0 || !hole);
1832 
1833 	if (lvl == dn->dn_phys->dn_nlevels) {
1834 		error = 0;
1835 		epb = dn->dn_phys->dn_nblkptr;
1836 		data = dn->dn_phys->dn_blkptr;
1837 	} else {
1838 		uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1839 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1840 		if (error) {
1841 			if (error != ENOENT)
1842 				return (error);
1843 			if (hole)
1844 				return (0);
1845 			/*
1846 			 * This can only happen when we are searching up
1847 			 * the block tree for data.  We don't really need to
1848 			 * adjust the offset, as we will just end up looking
1849 			 * at the pointer to this block in its parent, and its
1850 			 * going to be unallocated, so we will skip over it.
1851 			 */
1852 			return (ESRCH);
1853 		}
1854 		error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1855 		if (error) {
1856 			dbuf_rele(db, FTAG);
1857 			return (error);
1858 		}
1859 		data = db->db.db_data;
1860 	}
1861 
1862 	if (db && txg &&
1863 	    (db->db_blkptr == NULL || db->db_blkptr->blk_birth <= txg)) {
1864 		/*
1865 		 * This can only happen when we are searching up the tree
1866 		 * and these conditions mean that we need to keep climbing.
1867 		 */
1868 		error = ESRCH;
1869 	} else if (lvl == 0) {
1870 		dnode_phys_t *dnp = data;
1871 		span = DNODE_SHIFT;
1872 		ASSERT(dn->dn_type == DMU_OT_DNODE);
1873 
1874 		for (i = (*offset >> span) & (blkfill - 1);
1875 		    i >= 0 && i < blkfill; i += inc) {
1876 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1877 				break;
1878 			*offset += (1ULL << span) * inc;
1879 		}
1880 		if (i < 0 || i == blkfill)
1881 			error = ESRCH;
1882 	} else {
1883 		blkptr_t *bp = data;
1884 		uint64_t start = *offset;
1885 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
1886 		minfill = 0;
1887 		maxfill = blkfill << ((lvl - 1) * epbs);
1888 
1889 		if (hole)
1890 			maxfill--;
1891 		else
1892 			minfill++;
1893 
1894 		*offset = *offset >> span;
1895 		for (i = BF64_GET(*offset, 0, epbs);
1896 		    i >= 0 && i < epb; i += inc) {
1897 			if (bp[i].blk_fill >= minfill &&
1898 			    bp[i].blk_fill <= maxfill &&
1899 			    (hole || bp[i].blk_birth > txg))
1900 				break;
1901 			if (inc > 0 || *offset > 0)
1902 				*offset += inc;
1903 		}
1904 		*offset = *offset << span;
1905 		if (inc < 0) {
1906 			/* traversing backwards; position offset at the end */
1907 			ASSERT3U(*offset, <=, start);
1908 			*offset = MIN(*offset + (1ULL << span) - 1, start);
1909 		} else if (*offset < start) {
1910 			*offset = start;
1911 		}
1912 		if (i < 0 || i >= epb)
1913 			error = ESRCH;
1914 	}
1915 
1916 	if (db)
1917 		dbuf_rele(db, FTAG);
1918 
1919 	return (error);
1920 }
1921 
1922 /*
1923  * Find the next hole, data, or sparse region at or after *offset.
1924  * The value 'blkfill' tells us how many items we expect to find
1925  * in an L0 data block; this value is 1 for normal objects,
1926  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1927  * DNODES_PER_BLOCK when searching for sparse regions thereof.
1928  *
1929  * Examples:
1930  *
1931  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1932  *	Finds the next/previous hole/data in a file.
1933  *	Used in dmu_offset_next().
1934  *
1935  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1936  *	Finds the next free/allocated dnode an objset's meta-dnode.
1937  *	Only finds objects that have new contents since txg (ie.
1938  *	bonus buffer changes and content removal are ignored).
1939  *	Used in dmu_object_next().
1940  *
1941  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1942  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
1943  *	Used in dmu_object_alloc().
1944  */
1945 int
1946 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1947     int minlvl, uint64_t blkfill, uint64_t txg)
1948 {
1949 	uint64_t initial_offset = *offset;
1950 	int lvl, maxlvl;
1951 	int error = 0;
1952 
1953 	if (!(flags & DNODE_FIND_HAVELOCK))
1954 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1955 
1956 	if (dn->dn_phys->dn_nlevels == 0) {
1957 		error = ESRCH;
1958 		goto out;
1959 	}
1960 
1961 	if (dn->dn_datablkshift == 0) {
1962 		if (*offset < dn->dn_datablksz) {
1963 			if (flags & DNODE_FIND_HOLE)
1964 				*offset = dn->dn_datablksz;
1965 		} else {
1966 			error = ESRCH;
1967 		}
1968 		goto out;
1969 	}
1970 
1971 	maxlvl = dn->dn_phys->dn_nlevels;
1972 
1973 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1974 		error = dnode_next_offset_level(dn,
1975 		    flags, offset, lvl, blkfill, txg);
1976 		if (error != ESRCH)
1977 			break;
1978 	}
1979 
1980 	while (error == 0 && --lvl >= minlvl) {
1981 		error = dnode_next_offset_level(dn,
1982 		    flags, offset, lvl, blkfill, txg);
1983 	}
1984 
1985 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1986 	    initial_offset < *offset : initial_offset > *offset))
1987 		error = ESRCH;
1988 out:
1989 	if (!(flags & DNODE_FIND_HAVELOCK))
1990 		rw_exit(&dn->dn_struct_rwlock);
1991 
1992 	return (error);
1993 }
1994