xref: /illumos-gate/usr/src/uts/common/fs/zfs/dnode_sync.c (revision 0f6d88aded0d165f5954688a9b13bac76c38da84)
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 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/dbuf.h>
29 #include <sys/dnode.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dataset.h>
34 #include <sys/spa.h>
35 #include <sys/range_tree.h>
36 #include <sys/zfeature.h>
37 
38 static void
39 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
40 {
41 	dmu_buf_impl_t *db;
42 	int txgoff = tx->tx_txg & TXG_MASK;
43 	int nblkptr = dn->dn_phys->dn_nblkptr;
44 	int old_toplvl = dn->dn_phys->dn_nlevels - 1;
45 	int new_level = dn->dn_next_nlevels[txgoff];
46 	int i;
47 
48 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
49 
50 	/* this dnode can't be paged out because it's dirty */
51 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
52 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
53 	ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
54 
55 	db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
56 	ASSERT(db != NULL);
57 
58 	dn->dn_phys->dn_nlevels = new_level;
59 	dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
60 	    dn->dn_object, dn->dn_phys->dn_nlevels);
61 
62 	/* check for existing blkptrs in the dnode */
63 	for (i = 0; i < nblkptr; i++)
64 		if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i]))
65 			break;
66 	if (i != nblkptr) {
67 		/* transfer dnode's block pointers to new indirect block */
68 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
69 		ASSERT(db->db.db_data);
70 		ASSERT(arc_released(db->db_buf));
71 		ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
72 		bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
73 		    sizeof (blkptr_t) * nblkptr);
74 		arc_buf_freeze(db->db_buf);
75 	}
76 
77 	/* set dbuf's parent pointers to new indirect buf */
78 	for (i = 0; i < nblkptr; i++) {
79 		dmu_buf_impl_t *child = dbuf_find(dn, old_toplvl, i);
80 
81 		if (child == NULL)
82 			continue;
83 #ifdef	DEBUG
84 		DB_DNODE_ENTER(child);
85 		ASSERT3P(DB_DNODE(child), ==, dn);
86 		DB_DNODE_EXIT(child);
87 #endif	/* DEBUG */
88 		if (child->db_parent && child->db_parent != dn->dn_dbuf) {
89 			ASSERT(child->db_parent->db_level == db->db_level);
90 			ASSERT(child->db_blkptr !=
91 			    &dn->dn_phys->dn_blkptr[child->db_blkid]);
92 			mutex_exit(&child->db_mtx);
93 			continue;
94 		}
95 		ASSERT(child->db_parent == NULL ||
96 		    child->db_parent == dn->dn_dbuf);
97 
98 		child->db_parent = db;
99 		dbuf_add_ref(db, child);
100 		if (db->db.db_data)
101 			child->db_blkptr = (blkptr_t *)db->db.db_data + i;
102 		else
103 			child->db_blkptr = NULL;
104 		dprintf_dbuf_bp(child, child->db_blkptr,
105 		    "changed db_blkptr to new indirect %s", "");
106 
107 		mutex_exit(&child->db_mtx);
108 	}
109 
110 	bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
111 
112 	dbuf_rele(db, FTAG);
113 
114 	rw_exit(&dn->dn_struct_rwlock);
115 }
116 
117 static void
118 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
119 {
120 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
121 	uint64_t bytesfreed = 0;
122 
123 	dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
124 
125 	for (int i = 0; i < num; i++, bp++) {
126 		if (BP_IS_HOLE(bp))
127 			continue;
128 
129 		bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
130 		ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
131 
132 		/*
133 		 * Save some useful information on the holes being
134 		 * punched, including logical size, type, and indirection
135 		 * level. Retaining birth time enables detection of when
136 		 * holes are punched for reducing the number of free
137 		 * records transmitted during a zfs send.
138 		 */
139 
140 		uint64_t lsize = BP_GET_LSIZE(bp);
141 		dmu_object_type_t type = BP_GET_TYPE(bp);
142 		uint64_t lvl = BP_GET_LEVEL(bp);
143 
144 		bzero(bp, sizeof (blkptr_t));
145 
146 		if (spa_feature_is_active(dn->dn_objset->os_spa,
147 		    SPA_FEATURE_HOLE_BIRTH)) {
148 			BP_SET_LSIZE(bp, lsize);
149 			BP_SET_TYPE(bp, type);
150 			BP_SET_LEVEL(bp, lvl);
151 			BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
152 		}
153 	}
154 	dnode_diduse_space(dn, -bytesfreed);
155 }
156 
157 #ifdef ZFS_DEBUG
158 static void
159 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
160 {
161 	int off, num;
162 	int i, err, epbs;
163 	uint64_t txg = tx->tx_txg;
164 	dnode_t *dn;
165 
166 	DB_DNODE_ENTER(db);
167 	dn = DB_DNODE(db);
168 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
169 	off = start - (db->db_blkid * 1<<epbs);
170 	num = end - start + 1;
171 
172 	ASSERT3U(off, >=, 0);
173 	ASSERT3U(num, >=, 0);
174 	ASSERT3U(db->db_level, >, 0);
175 	ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
176 	ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
177 	ASSERT(db->db_blkptr != NULL);
178 
179 	for (i = off; i < off+num; i++) {
180 		uint64_t *buf;
181 		dmu_buf_impl_t *child;
182 		dbuf_dirty_record_t *dr;
183 		int j;
184 
185 		ASSERT(db->db_level == 1);
186 
187 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
188 		err = dbuf_hold_impl(dn, db->db_level-1,
189 		    (db->db_blkid << epbs) + i, TRUE, FTAG, &child);
190 		rw_exit(&dn->dn_struct_rwlock);
191 		if (err == ENOENT)
192 			continue;
193 		ASSERT(err == 0);
194 		ASSERT(child->db_level == 0);
195 		dr = child->db_last_dirty;
196 		while (dr && dr->dr_txg > txg)
197 			dr = dr->dr_next;
198 		ASSERT(dr == NULL || dr->dr_txg == txg);
199 
200 		/* data_old better be zeroed */
201 		if (dr) {
202 			buf = dr->dt.dl.dr_data->b_data;
203 			for (j = 0; j < child->db.db_size >> 3; j++) {
204 				if (buf[j] != 0) {
205 					panic("freed data not zero: "
206 					    "child=%p i=%d off=%d num=%d\n",
207 					    (void *)child, i, off, num);
208 				}
209 			}
210 		}
211 
212 		/*
213 		 * db_data better be zeroed unless it's dirty in a
214 		 * future txg.
215 		 */
216 		mutex_enter(&child->db_mtx);
217 		buf = child->db.db_data;
218 		if (buf != NULL && child->db_state != DB_FILL &&
219 		    child->db_last_dirty == NULL) {
220 			for (j = 0; j < child->db.db_size >> 3; j++) {
221 				if (buf[j] != 0) {
222 					panic("freed data not zero: "
223 					    "child=%p i=%d off=%d num=%d\n",
224 					    (void *)child, i, off, num);
225 				}
226 			}
227 		}
228 		mutex_exit(&child->db_mtx);
229 
230 		dbuf_rele(child, FTAG);
231 	}
232 	DB_DNODE_EXIT(db);
233 }
234 #endif
235 
236 static void
237 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
238     dmu_tx_t *tx)
239 {
240 	dnode_t *dn;
241 	blkptr_t *bp;
242 	dmu_buf_impl_t *subdb;
243 	uint64_t start, end, dbstart, dbend, i;
244 	int epbs, shift;
245 
246 	/*
247 	 * There is a small possibility that this block will not be cached:
248 	 *   1 - if level > 1 and there are no children with level <= 1
249 	 *   2 - if this block was evicted since we read it from
250 	 *	 dmu_tx_hold_free().
251 	 */
252 	if (db->db_state != DB_CACHED)
253 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
254 
255 	dbuf_release_bp(db);
256 	bp = db->db.db_data;
257 
258 	DB_DNODE_ENTER(db);
259 	dn = DB_DNODE(db);
260 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
261 	shift = (db->db_level - 1) * epbs;
262 	dbstart = db->db_blkid << epbs;
263 	start = blkid >> shift;
264 	if (dbstart < start) {
265 		bp += start - dbstart;
266 	} else {
267 		start = dbstart;
268 	}
269 	dbend = ((db->db_blkid + 1) << epbs) - 1;
270 	end = (blkid + nblks - 1) >> shift;
271 	if (dbend <= end)
272 		end = dbend;
273 
274 	ASSERT3U(start, <=, end);
275 
276 	if (db->db_level == 1) {
277 		FREE_VERIFY(db, start, end, tx);
278 		free_blocks(dn, bp, end-start+1, tx);
279 	} else {
280 		for (i = start; i <= end; i++, bp++) {
281 			if (BP_IS_HOLE(bp))
282 				continue;
283 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
284 			VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
285 			    i, B_TRUE, FTAG, &subdb));
286 			rw_exit(&dn->dn_struct_rwlock);
287 			ASSERT3P(bp, ==, subdb->db_blkptr);
288 
289 			free_children(subdb, blkid, nblks, tx);
290 			dbuf_rele(subdb, FTAG);
291 		}
292 	}
293 
294 	/* If this whole block is free, free ourself too. */
295 	for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
296 		if (!BP_IS_HOLE(bp))
297 			break;
298 	}
299 	if (i == 1 << epbs) {
300 		/* didn't find any non-holes */
301 		bzero(db->db.db_data, db->db.db_size);
302 		free_blocks(dn, db->db_blkptr, 1, tx);
303 	} else {
304 		/*
305 		 * Partial block free; must be marked dirty so that it
306 		 * will be written out.
307 		 */
308 		ASSERT(db->db_dirtycnt > 0);
309 	}
310 
311 	DB_DNODE_EXIT(db);
312 	arc_buf_freeze(db->db_buf);
313 }
314 
315 /*
316  * Traverse the indicated range of the provided file
317  * and "free" all the blocks contained there.
318  */
319 static void
320 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
321     dmu_tx_t *tx)
322 {
323 	blkptr_t *bp = dn->dn_phys->dn_blkptr;
324 	int dnlevel = dn->dn_phys->dn_nlevels;
325 	boolean_t trunc = B_FALSE;
326 
327 	if (blkid > dn->dn_phys->dn_maxblkid)
328 		return;
329 
330 	ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
331 	if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
332 		nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
333 		trunc = B_TRUE;
334 	}
335 
336 	/* There are no indirect blocks in the object */
337 	if (dnlevel == 1) {
338 		if (blkid >= dn->dn_phys->dn_nblkptr) {
339 			/* this range was never made persistent */
340 			return;
341 		}
342 		ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
343 		free_blocks(dn, bp + blkid, nblks, tx);
344 	} else {
345 		int shift = (dnlevel - 1) *
346 		    (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
347 		int start = blkid >> shift;
348 		int end = (blkid + nblks - 1) >> shift;
349 		dmu_buf_impl_t *db;
350 
351 		ASSERT(start < dn->dn_phys->dn_nblkptr);
352 		bp += start;
353 		for (int i = start; i <= end; i++, bp++) {
354 			if (BP_IS_HOLE(bp))
355 				continue;
356 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
357 			VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
358 			    TRUE, FTAG, &db));
359 			rw_exit(&dn->dn_struct_rwlock);
360 
361 			free_children(db, blkid, nblks, tx);
362 			dbuf_rele(db, FTAG);
363 		}
364 	}
365 
366 	if (trunc) {
367 		dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
368 
369 		uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
370 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
371 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
372 		    dn->dn_phys->dn_maxblkid == 0 ||
373 		    dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
374 	}
375 }
376 
377 typedef struct dnode_sync_free_range_arg {
378 	dnode_t *dsfra_dnode;
379 	dmu_tx_t *dsfra_tx;
380 } dnode_sync_free_range_arg_t;
381 
382 static void
383 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
384 {
385 	dnode_sync_free_range_arg_t *dsfra = arg;
386 	dnode_t *dn = dsfra->dsfra_dnode;
387 
388 	mutex_exit(&dn->dn_mtx);
389 	dnode_sync_free_range_impl(dn, blkid, nblks, dsfra->dsfra_tx);
390 	mutex_enter(&dn->dn_mtx);
391 }
392 
393 /*
394  * Try to kick all the dnode's dbufs out of the cache...
395  */
396 void
397 dnode_evict_dbufs(dnode_t *dn)
398 {
399 	int progress;
400 	int pass = 0;
401 
402 	do {
403 		dmu_buf_impl_t *db, *db_next;
404 		int evicting = FALSE;
405 
406 		progress = FALSE;
407 		mutex_enter(&dn->dn_dbufs_mtx);
408 		for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
409 			db_next = AVL_NEXT(&dn->dn_dbufs, db);
410 #ifdef	DEBUG
411 			DB_DNODE_ENTER(db);
412 			ASSERT3P(DB_DNODE(db), ==, dn);
413 			DB_DNODE_EXIT(db);
414 #endif	/* DEBUG */
415 
416 			mutex_enter(&db->db_mtx);
417 			if (db->db_state == DB_EVICTING) {
418 				progress = TRUE;
419 				evicting = TRUE;
420 				mutex_exit(&db->db_mtx);
421 			} else if (refcount_is_zero(&db->db_holds)) {
422 				progress = TRUE;
423 				dbuf_clear(db); /* exits db_mtx for us */
424 			} else {
425 				mutex_exit(&db->db_mtx);
426 			}
427 
428 		}
429 		/*
430 		 * NB: we need to drop dn_dbufs_mtx between passes so
431 		 * that any DB_EVICTING dbufs can make progress.
432 		 * Ideally, we would have some cv we could wait on, but
433 		 * since we don't, just wait a bit to give the other
434 		 * thread a chance to run.
435 		 */
436 		mutex_exit(&dn->dn_dbufs_mtx);
437 		if (evicting)
438 			delay(1);
439 		pass++;
440 		ASSERT(pass < 100); /* sanity check */
441 	} while (progress);
442 
443 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
444 	if (dn->dn_bonus && refcount_is_zero(&dn->dn_bonus->db_holds)) {
445 		mutex_enter(&dn->dn_bonus->db_mtx);
446 		dbuf_evict(dn->dn_bonus);
447 		dn->dn_bonus = NULL;
448 	}
449 	rw_exit(&dn->dn_struct_rwlock);
450 }
451 
452 static void
453 dnode_undirty_dbufs(list_t *list)
454 {
455 	dbuf_dirty_record_t *dr;
456 
457 	while (dr = list_head(list)) {
458 		dmu_buf_impl_t *db = dr->dr_dbuf;
459 		uint64_t txg = dr->dr_txg;
460 
461 		if (db->db_level != 0)
462 			dnode_undirty_dbufs(&dr->dt.di.dr_children);
463 
464 		mutex_enter(&db->db_mtx);
465 		/* XXX - use dbuf_undirty()? */
466 		list_remove(list, dr);
467 		ASSERT(db->db_last_dirty == dr);
468 		db->db_last_dirty = NULL;
469 		db->db_dirtycnt -= 1;
470 		if (db->db_level == 0) {
471 			ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
472 			    dr->dt.dl.dr_data == db->db_buf);
473 			dbuf_unoverride(dr);
474 		} else {
475 			mutex_destroy(&dr->dt.di.dr_mtx);
476 			list_destroy(&dr->dt.di.dr_children);
477 		}
478 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
479 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
480 	}
481 }
482 
483 static void
484 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
485 {
486 	int txgoff = tx->tx_txg & TXG_MASK;
487 
488 	ASSERT(dmu_tx_is_syncing(tx));
489 
490 	/*
491 	 * Our contents should have been freed in dnode_sync() by the
492 	 * free range record inserted by the caller of dnode_free().
493 	 */
494 	ASSERT0(DN_USED_BYTES(dn->dn_phys));
495 	ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
496 
497 	dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
498 	dnode_evict_dbufs(dn);
499 	ASSERT(avl_is_empty(&dn->dn_dbufs));
500 	ASSERT3P(dn->dn_bonus, ==, NULL);
501 
502 	/*
503 	 * XXX - It would be nice to assert this, but we may still
504 	 * have residual holds from async evictions from the arc...
505 	 *
506 	 * zfs_obj_to_path() also depends on this being
507 	 * commented out.
508 	 *
509 	 * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
510 	 */
511 
512 	/* Undirty next bits */
513 	dn->dn_next_nlevels[txgoff] = 0;
514 	dn->dn_next_indblkshift[txgoff] = 0;
515 	dn->dn_next_blksz[txgoff] = 0;
516 
517 	/* ASSERT(blkptrs are zero); */
518 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
519 	ASSERT(dn->dn_type != DMU_OT_NONE);
520 
521 	ASSERT(dn->dn_free_txg > 0);
522 	if (dn->dn_allocated_txg != dn->dn_free_txg)
523 		dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
524 	bzero(dn->dn_phys, sizeof (dnode_phys_t));
525 
526 	mutex_enter(&dn->dn_mtx);
527 	dn->dn_type = DMU_OT_NONE;
528 	dn->dn_maxblkid = 0;
529 	dn->dn_allocated_txg = 0;
530 	dn->dn_free_txg = 0;
531 	dn->dn_have_spill = B_FALSE;
532 	mutex_exit(&dn->dn_mtx);
533 
534 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
535 
536 	dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
537 	/*
538 	 * Now that we've released our hold, the dnode may
539 	 * be evicted, so we musn't access it.
540 	 */
541 }
542 
543 /*
544  * Write out the dnode's dirty buffers.
545  */
546 void
547 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
548 {
549 	dnode_phys_t *dnp = dn->dn_phys;
550 	int txgoff = tx->tx_txg & TXG_MASK;
551 	list_t *list = &dn->dn_dirty_records[txgoff];
552 	static const dnode_phys_t zerodn = { 0 };
553 	boolean_t kill_spill = B_FALSE;
554 
555 	ASSERT(dmu_tx_is_syncing(tx));
556 	ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
557 	ASSERT(dnp->dn_type != DMU_OT_NONE ||
558 	    bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
559 	DNODE_VERIFY(dn);
560 
561 	ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
562 
563 	if (dmu_objset_userused_enabled(dn->dn_objset) &&
564 	    !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
565 		mutex_enter(&dn->dn_mtx);
566 		dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
567 		dn->dn_oldflags = dn->dn_phys->dn_flags;
568 		dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
569 		mutex_exit(&dn->dn_mtx);
570 		dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
571 	} else {
572 		/* Once we account for it, we should always account for it. */
573 		ASSERT(!(dn->dn_phys->dn_flags &
574 		    DNODE_FLAG_USERUSED_ACCOUNTED));
575 	}
576 
577 	mutex_enter(&dn->dn_mtx);
578 	if (dn->dn_allocated_txg == tx->tx_txg) {
579 		/* The dnode is newly allocated or reallocated */
580 		if (dnp->dn_type == DMU_OT_NONE) {
581 			/* this is a first alloc, not a realloc */
582 			dnp->dn_nlevels = 1;
583 			dnp->dn_nblkptr = dn->dn_nblkptr;
584 		}
585 
586 		dnp->dn_type = dn->dn_type;
587 		dnp->dn_bonustype = dn->dn_bonustype;
588 		dnp->dn_bonuslen = dn->dn_bonuslen;
589 	}
590 	ASSERT(dnp->dn_nlevels > 1 ||
591 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
592 	    BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
593 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
594 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
595 	ASSERT(dnp->dn_nlevels < 2 ||
596 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
597 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
598 
599 	if (dn->dn_next_type[txgoff] != 0) {
600 		dnp->dn_type = dn->dn_type;
601 		dn->dn_next_type[txgoff] = 0;
602 	}
603 
604 	if (dn->dn_next_blksz[txgoff] != 0) {
605 		ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
606 		    SPA_MINBLOCKSIZE) == 0);
607 		ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
608 		    dn->dn_maxblkid == 0 || list_head(list) != NULL ||
609 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
610 		    dnp->dn_datablkszsec ||
611 		    range_tree_space(dn->dn_free_ranges[txgoff]) != 0);
612 		dnp->dn_datablkszsec =
613 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
614 		dn->dn_next_blksz[txgoff] = 0;
615 	}
616 
617 	if (dn->dn_next_bonuslen[txgoff] != 0) {
618 		if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
619 			dnp->dn_bonuslen = 0;
620 		else
621 			dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
622 		ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
623 		dn->dn_next_bonuslen[txgoff] = 0;
624 	}
625 
626 	if (dn->dn_next_bonustype[txgoff] != 0) {
627 		ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
628 		dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
629 		dn->dn_next_bonustype[txgoff] = 0;
630 	}
631 
632 	boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
633 	    dn->dn_free_txg <= tx->tx_txg;
634 
635 	/*
636 	 * We will either remove a spill block when a file is being removed
637 	 * or we have been asked to remove it.
638 	 */
639 	if (dn->dn_rm_spillblk[txgoff] ||
640 	    ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) && freeing_dnode)) {
641 		if ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
642 			kill_spill = B_TRUE;
643 		dn->dn_rm_spillblk[txgoff] = 0;
644 	}
645 
646 	if (dn->dn_next_indblkshift[txgoff] != 0) {
647 		ASSERT(dnp->dn_nlevels == 1);
648 		dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
649 		dn->dn_next_indblkshift[txgoff] = 0;
650 	}
651 
652 	/*
653 	 * Just take the live (open-context) values for checksum and compress.
654 	 * Strictly speaking it's a future leak, but nothing bad happens if we
655 	 * start using the new checksum or compress algorithm a little early.
656 	 */
657 	dnp->dn_checksum = dn->dn_checksum;
658 	dnp->dn_compress = dn->dn_compress;
659 
660 	mutex_exit(&dn->dn_mtx);
661 
662 	if (kill_spill) {
663 		free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
664 		mutex_enter(&dn->dn_mtx);
665 		dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
666 		mutex_exit(&dn->dn_mtx);
667 	}
668 
669 	/* process all the "freed" ranges in the file */
670 	if (dn->dn_free_ranges[txgoff] != NULL) {
671 		dnode_sync_free_range_arg_t dsfra;
672 		dsfra.dsfra_dnode = dn;
673 		dsfra.dsfra_tx = tx;
674 		mutex_enter(&dn->dn_mtx);
675 		range_tree_vacate(dn->dn_free_ranges[txgoff],
676 		    dnode_sync_free_range, &dsfra);
677 		range_tree_destroy(dn->dn_free_ranges[txgoff]);
678 		dn->dn_free_ranges[txgoff] = NULL;
679 		mutex_exit(&dn->dn_mtx);
680 	}
681 
682 	if (freeing_dnode) {
683 		dnode_sync_free(dn, tx);
684 		return;
685 	}
686 
687 	if (dn->dn_next_nblkptr[txgoff]) {
688 		/* this should only happen on a realloc */
689 		ASSERT(dn->dn_allocated_txg == tx->tx_txg);
690 		if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
691 			/* zero the new blkptrs we are gaining */
692 			bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
693 			    sizeof (blkptr_t) *
694 			    (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
695 #ifdef ZFS_DEBUG
696 		} else {
697 			int i;
698 			ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
699 			/* the blkptrs we are losing better be unallocated */
700 			for (i = dn->dn_next_nblkptr[txgoff];
701 			    i < dnp->dn_nblkptr; i++)
702 				ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
703 #endif
704 		}
705 		mutex_enter(&dn->dn_mtx);
706 		dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
707 		dn->dn_next_nblkptr[txgoff] = 0;
708 		mutex_exit(&dn->dn_mtx);
709 	}
710 
711 	if (dn->dn_next_nlevels[txgoff]) {
712 		dnode_increase_indirection(dn, tx);
713 		dn->dn_next_nlevels[txgoff] = 0;
714 	}
715 
716 	dbuf_sync_list(list, tx);
717 
718 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
719 		ASSERT3P(list_head(list), ==, NULL);
720 		dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
721 	}
722 
723 	/*
724 	 * Although we have dropped our reference to the dnode, it
725 	 * can't be evicted until its written, and we haven't yet
726 	 * initiated the IO for the dnode's dbuf.
727 	 */
728 }
729