xref: /titanic_41/usr/src/uts/common/fs/zfs/dnode_sync.c (revision 1666a561bcf7351b8aa2b813412d0e84137a21b0)
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, 2015 by Delphix. All rights reserved.
25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
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_tx.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/spa.h>
36 #include <sys/range_tree.h>
37 #include <sys/zfeature.h>
38 
39 static void
dnode_increase_indirection(dnode_t * dn,dmu_tx_t * tx)40 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
41 {
42 	dmu_buf_impl_t *db;
43 	int txgoff = tx->tx_txg & TXG_MASK;
44 	int nblkptr = dn->dn_phys->dn_nblkptr;
45 	int old_toplvl = dn->dn_phys->dn_nlevels - 1;
46 	int new_level = dn->dn_next_nlevels[txgoff];
47 	int i;
48 
49 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
50 
51 	/* this dnode can't be paged out because it's dirty */
52 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
53 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
54 	ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
55 
56 	db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
57 	ASSERT(db != NULL);
58 
59 	dn->dn_phys->dn_nlevels = new_level;
60 	dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
61 	    dn->dn_object, dn->dn_phys->dn_nlevels);
62 
63 	/* check for existing blkptrs in the dnode */
64 	for (i = 0; i < nblkptr; i++)
65 		if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i]))
66 			break;
67 	if (i != nblkptr) {
68 		/* transfer dnode's block pointers to new indirect block */
69 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
70 		ASSERT(db->db.db_data);
71 		ASSERT(arc_released(db->db_buf));
72 		ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
73 		bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
74 		    sizeof (blkptr_t) * nblkptr);
75 		arc_buf_freeze(db->db_buf);
76 	}
77 
78 	/* set dbuf's parent pointers to new indirect buf */
79 	for (i = 0; i < nblkptr; i++) {
80 		dmu_buf_impl_t *child =
81 		    dbuf_find(dn->dn_objset, dn->dn_object, old_toplvl, i);
82 
83 		if (child == NULL)
84 			continue;
85 #ifdef	DEBUG
86 		DB_DNODE_ENTER(child);
87 		ASSERT3P(DB_DNODE(child), ==, dn);
88 		DB_DNODE_EXIT(child);
89 #endif	/* DEBUG */
90 		if (child->db_parent && child->db_parent != dn->dn_dbuf) {
91 			ASSERT(child->db_parent->db_level == db->db_level);
92 			ASSERT(child->db_blkptr !=
93 			    &dn->dn_phys->dn_blkptr[child->db_blkid]);
94 			mutex_exit(&child->db_mtx);
95 			continue;
96 		}
97 		ASSERT(child->db_parent == NULL ||
98 		    child->db_parent == dn->dn_dbuf);
99 
100 		child->db_parent = db;
101 		dbuf_add_ref(db, child);
102 		if (db->db.db_data)
103 			child->db_blkptr = (blkptr_t *)db->db.db_data + i;
104 		else
105 			child->db_blkptr = NULL;
106 		dprintf_dbuf_bp(child, child->db_blkptr,
107 		    "changed db_blkptr to new indirect %s", "");
108 
109 		mutex_exit(&child->db_mtx);
110 	}
111 
112 	bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
113 
114 	dbuf_rele(db, FTAG);
115 
116 	rw_exit(&dn->dn_struct_rwlock);
117 }
118 
119 static void
free_blocks(dnode_t * dn,blkptr_t * bp,int num,dmu_tx_t * tx)120 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
121 {
122 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
123 	uint64_t bytesfreed = 0;
124 
125 	dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
126 
127 	for (int i = 0; i < num; i++, bp++) {
128 		if (BP_IS_HOLE(bp))
129 			continue;
130 
131 		bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
132 		ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
133 
134 		/*
135 		 * Save some useful information on the holes being
136 		 * punched, including logical size, type, and indirection
137 		 * level. Retaining birth time enables detection of when
138 		 * holes are punched for reducing the number of free
139 		 * records transmitted during a zfs send.
140 		 */
141 
142 		uint64_t lsize = BP_GET_LSIZE(bp);
143 		dmu_object_type_t type = BP_GET_TYPE(bp);
144 		uint64_t lvl = BP_GET_LEVEL(bp);
145 
146 		bzero(bp, sizeof (blkptr_t));
147 
148 		if (spa_feature_is_active(dn->dn_objset->os_spa,
149 		    SPA_FEATURE_HOLE_BIRTH)) {
150 			BP_SET_LSIZE(bp, lsize);
151 			BP_SET_TYPE(bp, type);
152 			BP_SET_LEVEL(bp, lvl);
153 			BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
154 		}
155 	}
156 	dnode_diduse_space(dn, -bytesfreed);
157 }
158 
159 #ifdef ZFS_DEBUG
160 static void
free_verify(dmu_buf_impl_t * db,uint64_t start,uint64_t end,dmu_tx_t * tx)161 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
162 {
163 	int off, num;
164 	int i, err, epbs;
165 	uint64_t txg = tx->tx_txg;
166 	dnode_t *dn;
167 
168 	DB_DNODE_ENTER(db);
169 	dn = DB_DNODE(db);
170 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
171 	off = start - (db->db_blkid * 1<<epbs);
172 	num = end - start + 1;
173 
174 	ASSERT3U(off, >=, 0);
175 	ASSERT3U(num, >=, 0);
176 	ASSERT3U(db->db_level, >, 0);
177 	ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
178 	ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
179 	ASSERT(db->db_blkptr != NULL);
180 
181 	for (i = off; i < off+num; i++) {
182 		uint64_t *buf;
183 		dmu_buf_impl_t *child;
184 		dbuf_dirty_record_t *dr;
185 		int j;
186 
187 		ASSERT(db->db_level == 1);
188 
189 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
190 		err = dbuf_hold_impl(dn, db->db_level-1,
191 		    (db->db_blkid << epbs) + i, TRUE, FALSE, FTAG, &child);
192 		rw_exit(&dn->dn_struct_rwlock);
193 		if (err == ENOENT)
194 			continue;
195 		ASSERT(err == 0);
196 		ASSERT(child->db_level == 0);
197 		dr = child->db_last_dirty;
198 		while (dr && dr->dr_txg > txg)
199 			dr = dr->dr_next;
200 		ASSERT(dr == NULL || dr->dr_txg == txg);
201 
202 		/* data_old better be zeroed */
203 		if (dr) {
204 			buf = dr->dt.dl.dr_data->b_data;
205 			for (j = 0; j < child->db.db_size >> 3; j++) {
206 				if (buf[j] != 0) {
207 					panic("freed data not zero: "
208 					    "child=%p i=%d off=%d num=%d\n",
209 					    (void *)child, i, off, num);
210 				}
211 			}
212 		}
213 
214 		/*
215 		 * db_data better be zeroed unless it's dirty in a
216 		 * future txg.
217 		 */
218 		mutex_enter(&child->db_mtx);
219 		buf = child->db.db_data;
220 		if (buf != NULL && child->db_state != DB_FILL &&
221 		    child->db_last_dirty == NULL) {
222 			for (j = 0; j < child->db.db_size >> 3; j++) {
223 				if (buf[j] != 0) {
224 					panic("freed data not zero: "
225 					    "child=%p i=%d off=%d num=%d\n",
226 					    (void *)child, i, off, num);
227 				}
228 			}
229 		}
230 		mutex_exit(&child->db_mtx);
231 
232 		dbuf_rele(child, FTAG);
233 	}
234 	DB_DNODE_EXIT(db);
235 }
236 #endif
237 
238 static void
free_children(dmu_buf_impl_t * db,uint64_t blkid,uint64_t nblks,dmu_tx_t * tx)239 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
240     dmu_tx_t *tx)
241 {
242 	dnode_t *dn;
243 	blkptr_t *bp;
244 	dmu_buf_impl_t *subdb;
245 	uint64_t start, end, dbstart, dbend, i;
246 	int epbs, shift;
247 
248 	/*
249 	 * There is a small possibility that this block will not be cached:
250 	 *   1 - if level > 1 and there are no children with level <= 1
251 	 *   2 - if this block was evicted since we read it from
252 	 *	 dmu_tx_hold_free().
253 	 */
254 	if (db->db_state != DB_CACHED)
255 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
256 
257 	dbuf_release_bp(db);
258 	bp = db->db.db_data;
259 
260 	DB_DNODE_ENTER(db);
261 	dn = DB_DNODE(db);
262 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
263 	shift = (db->db_level - 1) * epbs;
264 	dbstart = db->db_blkid << epbs;
265 	start = blkid >> shift;
266 	if (dbstart < start) {
267 		bp += start - dbstart;
268 	} else {
269 		start = dbstart;
270 	}
271 	dbend = ((db->db_blkid + 1) << epbs) - 1;
272 	end = (blkid + nblks - 1) >> shift;
273 	if (dbend <= end)
274 		end = dbend;
275 
276 	ASSERT3U(start, <=, end);
277 
278 	if (db->db_level == 1) {
279 		FREE_VERIFY(db, start, end, tx);
280 		free_blocks(dn, bp, end-start+1, tx);
281 	} else {
282 		for (i = start; i <= end; i++, bp++) {
283 			if (BP_IS_HOLE(bp))
284 				continue;
285 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
286 			VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
287 			    i, TRUE, FALSE, FTAG, &subdb));
288 			rw_exit(&dn->dn_struct_rwlock);
289 			ASSERT3P(bp, ==, subdb->db_blkptr);
290 
291 			free_children(subdb, blkid, nblks, tx);
292 			dbuf_rele(subdb, FTAG);
293 		}
294 	}
295 
296 	/* If this whole block is free, free ourself too. */
297 	for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
298 		if (!BP_IS_HOLE(bp))
299 			break;
300 	}
301 	if (i == 1 << epbs) {
302 		/* didn't find any non-holes */
303 		bzero(db->db.db_data, db->db.db_size);
304 		free_blocks(dn, db->db_blkptr, 1, tx);
305 	} else {
306 		/*
307 		 * Partial block free; must be marked dirty so that it
308 		 * will be written out.
309 		 */
310 		ASSERT(db->db_dirtycnt > 0);
311 	}
312 
313 	DB_DNODE_EXIT(db);
314 	arc_buf_freeze(db->db_buf);
315 }
316 
317 /*
318  * Traverse the indicated range of the provided file
319  * and "free" all the blocks contained there.
320  */
321 static void
dnode_sync_free_range_impl(dnode_t * dn,uint64_t blkid,uint64_t nblks,dmu_tx_t * tx)322 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
323     dmu_tx_t *tx)
324 {
325 	blkptr_t *bp = dn->dn_phys->dn_blkptr;
326 	int dnlevel = dn->dn_phys->dn_nlevels;
327 	boolean_t trunc = B_FALSE;
328 
329 	if (blkid > dn->dn_phys->dn_maxblkid)
330 		return;
331 
332 	ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
333 	if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
334 		nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
335 		trunc = B_TRUE;
336 	}
337 
338 	/* There are no indirect blocks in the object */
339 	if (dnlevel == 1) {
340 		if (blkid >= dn->dn_phys->dn_nblkptr) {
341 			/* this range was never made persistent */
342 			return;
343 		}
344 		ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
345 		free_blocks(dn, bp + blkid, nblks, tx);
346 	} else {
347 		int shift = (dnlevel - 1) *
348 		    (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
349 		int start = blkid >> shift;
350 		int end = (blkid + nblks - 1) >> shift;
351 		dmu_buf_impl_t *db;
352 
353 		ASSERT(start < dn->dn_phys->dn_nblkptr);
354 		bp += start;
355 		for (int i = start; i <= end; i++, bp++) {
356 			if (BP_IS_HOLE(bp))
357 				continue;
358 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
359 			VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
360 			    TRUE, FALSE, FTAG, &db));
361 			rw_exit(&dn->dn_struct_rwlock);
362 
363 			free_children(db, blkid, nblks, tx);
364 			dbuf_rele(db, FTAG);
365 		}
366 	}
367 
368 	if (trunc) {
369 		dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
370 
371 		uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
372 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
373 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
374 		    dn->dn_phys->dn_maxblkid == 0 ||
375 		    dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
376 	}
377 }
378 
379 typedef struct dnode_sync_free_range_arg {
380 	dnode_t *dsfra_dnode;
381 	dmu_tx_t *dsfra_tx;
382 } dnode_sync_free_range_arg_t;
383 
384 static void
dnode_sync_free_range(void * arg,uint64_t blkid,uint64_t nblks)385 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
386 {
387 	dnode_sync_free_range_arg_t *dsfra = arg;
388 	dnode_t *dn = dsfra->dsfra_dnode;
389 
390 	mutex_exit(&dn->dn_mtx);
391 	dnode_sync_free_range_impl(dn, blkid, nblks, dsfra->dsfra_tx);
392 	mutex_enter(&dn->dn_mtx);
393 }
394 
395 /*
396  * Try to kick all the dnode's dbufs out of the cache...
397  */
398 void
dnode_evict_dbufs(dnode_t * dn)399 dnode_evict_dbufs(dnode_t *dn)
400 {
401 	dmu_buf_impl_t db_marker;
402 	dmu_buf_impl_t *db, *db_next;
403 
404 	mutex_enter(&dn->dn_dbufs_mtx);
405 	for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
406 
407 #ifdef	DEBUG
408 		DB_DNODE_ENTER(db);
409 		ASSERT3P(DB_DNODE(db), ==, dn);
410 		DB_DNODE_EXIT(db);
411 #endif	/* DEBUG */
412 
413 		mutex_enter(&db->db_mtx);
414 		if (db->db_state != DB_EVICTING &&
415 		    refcount_is_zero(&db->db_holds)) {
416 			db_marker.db_level = db->db_level;
417 			db_marker.db_blkid = db->db_blkid;
418 			db_marker.db_state = DB_SEARCH;
419 			avl_insert_here(&dn->dn_dbufs, &db_marker, db,
420 			    AVL_BEFORE);
421 
422 			dbuf_clear(db);
423 
424 			db_next = AVL_NEXT(&dn->dn_dbufs, &db_marker);
425 			avl_remove(&dn->dn_dbufs, &db_marker);
426 		} else {
427 			db->db_pending_evict = TRUE;
428 			mutex_exit(&db->db_mtx);
429 			db_next = AVL_NEXT(&dn->dn_dbufs, db);
430 		}
431 	}
432 	mutex_exit(&dn->dn_dbufs_mtx);
433 
434 	dnode_evict_bonus(dn);
435 }
436 
437 void
dnode_evict_bonus(dnode_t * dn)438 dnode_evict_bonus(dnode_t *dn)
439 {
440 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
441 	if (dn->dn_bonus != NULL) {
442 		if (refcount_is_zero(&dn->dn_bonus->db_holds)) {
443 			mutex_enter(&dn->dn_bonus->db_mtx);
444 			dbuf_evict(dn->dn_bonus);
445 			dn->dn_bonus = NULL;
446 		} else {
447 			dn->dn_bonus->db_pending_evict = TRUE;
448 		}
449 	}
450 	rw_exit(&dn->dn_struct_rwlock);
451 }
452 
453 static void
dnode_undirty_dbufs(list_t * list)454 dnode_undirty_dbufs(list_t *list)
455 {
456 	dbuf_dirty_record_t *dr;
457 
458 	while (dr = list_head(list)) {
459 		dmu_buf_impl_t *db = dr->dr_dbuf;
460 		uint64_t txg = dr->dr_txg;
461 
462 		if (db->db_level != 0)
463 			dnode_undirty_dbufs(&dr->dt.di.dr_children);
464 
465 		mutex_enter(&db->db_mtx);
466 		/* XXX - use dbuf_undirty()? */
467 		list_remove(list, dr);
468 		ASSERT(db->db_last_dirty == dr);
469 		db->db_last_dirty = NULL;
470 		db->db_dirtycnt -= 1;
471 		if (db->db_level == 0) {
472 			ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
473 			    dr->dt.dl.dr_data == db->db_buf);
474 			dbuf_unoverride(dr);
475 		} else {
476 			mutex_destroy(&dr->dt.di.dr_mtx);
477 			list_destroy(&dr->dt.di.dr_children);
478 		}
479 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
480 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
481 	}
482 }
483 
484 static void
dnode_sync_free(dnode_t * dn,dmu_tx_t * tx)485 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
486 {
487 	int txgoff = tx->tx_txg & TXG_MASK;
488 
489 	ASSERT(dmu_tx_is_syncing(tx));
490 
491 	/*
492 	 * Our contents should have been freed in dnode_sync() by the
493 	 * free range record inserted by the caller of dnode_free().
494 	 */
495 	ASSERT0(DN_USED_BYTES(dn->dn_phys));
496 	ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
497 
498 	dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
499 	dnode_evict_dbufs(dn);
500 
501 	/*
502 	 * XXX - It would be nice to assert this, but we may still
503 	 * have residual holds from async evictions from the arc...
504 	 *
505 	 * zfs_obj_to_path() also depends on this being
506 	 * commented out.
507 	 *
508 	 * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
509 	 */
510 
511 	/* Undirty next bits */
512 	dn->dn_next_nlevels[txgoff] = 0;
513 	dn->dn_next_indblkshift[txgoff] = 0;
514 	dn->dn_next_blksz[txgoff] = 0;
515 
516 	/* ASSERT(blkptrs are zero); */
517 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
518 	ASSERT(dn->dn_type != DMU_OT_NONE);
519 
520 	ASSERT(dn->dn_free_txg > 0);
521 	if (dn->dn_allocated_txg != dn->dn_free_txg)
522 		dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
523 	bzero(dn->dn_phys, sizeof (dnode_phys_t));
524 
525 	mutex_enter(&dn->dn_mtx);
526 	dn->dn_type = DMU_OT_NONE;
527 	dn->dn_maxblkid = 0;
528 	dn->dn_allocated_txg = 0;
529 	dn->dn_free_txg = 0;
530 	dn->dn_have_spill = B_FALSE;
531 	mutex_exit(&dn->dn_mtx);
532 
533 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
534 
535 	dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
536 	/*
537 	 * Now that we've released our hold, the dnode may
538 	 * be evicted, so we musn't access it.
539 	 */
540 }
541 
542 /*
543  * Write out the dnode's dirty buffers.
544  */
545 void
dnode_sync(dnode_t * dn,dmu_tx_t * tx)546 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
547 {
548 	dnode_phys_t *dnp = dn->dn_phys;
549 	int txgoff = tx->tx_txg & TXG_MASK;
550 	list_t *list = &dn->dn_dirty_records[txgoff];
551 	static const dnode_phys_t zerodn = { 0 };
552 	boolean_t kill_spill = B_FALSE;
553 
554 	ASSERT(dmu_tx_is_syncing(tx));
555 	ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
556 	ASSERT(dnp->dn_type != DMU_OT_NONE ||
557 	    bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
558 	DNODE_VERIFY(dn);
559 
560 	ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
561 
562 	if (dmu_objset_userused_enabled(dn->dn_objset) &&
563 	    !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
564 		mutex_enter(&dn->dn_mtx);
565 		dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
566 		dn->dn_oldflags = dn->dn_phys->dn_flags;
567 		dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
568 		mutex_exit(&dn->dn_mtx);
569 		dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
570 	} else {
571 		/* Once we account for it, we should always account for it. */
572 		ASSERT(!(dn->dn_phys->dn_flags &
573 		    DNODE_FLAG_USERUSED_ACCOUNTED));
574 	}
575 
576 	mutex_enter(&dn->dn_mtx);
577 	if (dn->dn_allocated_txg == tx->tx_txg) {
578 		/* The dnode is newly allocated or reallocated */
579 		if (dnp->dn_type == DMU_OT_NONE) {
580 			/* this is a first alloc, not a realloc */
581 			dnp->dn_nlevels = 1;
582 			dnp->dn_nblkptr = dn->dn_nblkptr;
583 		}
584 
585 		dnp->dn_type = dn->dn_type;
586 		dnp->dn_bonustype = dn->dn_bonustype;
587 		dnp->dn_bonuslen = dn->dn_bonuslen;
588 	}
589 	ASSERT(dnp->dn_nlevels > 1 ||
590 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
591 	    BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
592 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
593 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
594 	ASSERT(dnp->dn_nlevels < 2 ||
595 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
596 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
597 
598 	if (dn->dn_next_type[txgoff] != 0) {
599 		dnp->dn_type = dn->dn_type;
600 		dn->dn_next_type[txgoff] = 0;
601 	}
602 
603 	if (dn->dn_next_blksz[txgoff] != 0) {
604 		ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
605 		    SPA_MINBLOCKSIZE) == 0);
606 		ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
607 		    dn->dn_maxblkid == 0 || list_head(list) != NULL ||
608 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
609 		    dnp->dn_datablkszsec ||
610 		    range_tree_space(dn->dn_free_ranges[txgoff]) != 0);
611 		dnp->dn_datablkszsec =
612 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
613 		dn->dn_next_blksz[txgoff] = 0;
614 	}
615 
616 	if (dn->dn_next_bonuslen[txgoff] != 0) {
617 		if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
618 			dnp->dn_bonuslen = 0;
619 		else
620 			dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
621 		ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
622 		dn->dn_next_bonuslen[txgoff] = 0;
623 	}
624 
625 	if (dn->dn_next_bonustype[txgoff] != 0) {
626 		ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
627 		dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
628 		dn->dn_next_bonustype[txgoff] = 0;
629 	}
630 
631 	boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
632 	    dn->dn_free_txg <= tx->tx_txg;
633 
634 	/*
635 	 * Remove the spill block if we have been explicitly asked to
636 	 * remove it, or if the object is being removed.
637 	 */
638 	if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
639 		if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
640 			kill_spill = B_TRUE;
641 		dn->dn_rm_spillblk[txgoff] = 0;
642 	}
643 
644 	if (dn->dn_next_indblkshift[txgoff] != 0) {
645 		ASSERT(dnp->dn_nlevels == 1);
646 		dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
647 		dn->dn_next_indblkshift[txgoff] = 0;
648 	}
649 
650 	/*
651 	 * Just take the live (open-context) values for checksum and compress.
652 	 * Strictly speaking it's a future leak, but nothing bad happens if we
653 	 * start using the new checksum or compress algorithm a little early.
654 	 */
655 	dnp->dn_checksum = dn->dn_checksum;
656 	dnp->dn_compress = dn->dn_compress;
657 
658 	mutex_exit(&dn->dn_mtx);
659 
660 	if (kill_spill) {
661 		free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
662 		mutex_enter(&dn->dn_mtx);
663 		dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
664 		mutex_exit(&dn->dn_mtx);
665 	}
666 
667 	/* process all the "freed" ranges in the file */
668 	if (dn->dn_free_ranges[txgoff] != NULL) {
669 		dnode_sync_free_range_arg_t dsfra;
670 		dsfra.dsfra_dnode = dn;
671 		dsfra.dsfra_tx = tx;
672 		mutex_enter(&dn->dn_mtx);
673 		range_tree_vacate(dn->dn_free_ranges[txgoff],
674 		    dnode_sync_free_range, &dsfra);
675 		range_tree_destroy(dn->dn_free_ranges[txgoff]);
676 		dn->dn_free_ranges[txgoff] = NULL;
677 		mutex_exit(&dn->dn_mtx);
678 	}
679 
680 	if (freeing_dnode) {
681 		dnode_sync_free(dn, tx);
682 		return;
683 	}
684 
685 	if (dn->dn_next_nlevels[txgoff]) {
686 		dnode_increase_indirection(dn, tx);
687 		dn->dn_next_nlevels[txgoff] = 0;
688 	}
689 
690 	if (dn->dn_next_nblkptr[txgoff]) {
691 		/* this should only happen on a realloc */
692 		ASSERT(dn->dn_allocated_txg == tx->tx_txg);
693 		if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
694 			/* zero the new blkptrs we are gaining */
695 			bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
696 			    sizeof (blkptr_t) *
697 			    (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
698 #ifdef ZFS_DEBUG
699 		} else {
700 			int i;
701 			ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
702 			/* the blkptrs we are losing better be unallocated */
703 			for (i = dn->dn_next_nblkptr[txgoff];
704 			    i < dnp->dn_nblkptr; i++)
705 				ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
706 #endif
707 		}
708 		mutex_enter(&dn->dn_mtx);
709 		dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
710 		dn->dn_next_nblkptr[txgoff] = 0;
711 		mutex_exit(&dn->dn_mtx);
712 	}
713 
714 	dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
715 
716 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
717 		ASSERT3P(list_head(list), ==, NULL);
718 		dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
719 	}
720 
721 	/*
722 	 * Although we have dropped our reference to the dnode, it
723 	 * can't be evicted until its written, and we haven't yet
724 	 * initiated the IO for the dnode's dbuf.
725 	 */
726 }
727