xref: /titanic_41/usr/src/uts/common/fs/zfs/dnode_sync.c (revision bb25c06cca41ca78e5fb87fbb8e81d55beb18c95)
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 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
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/zio.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 i;
43 	uint64_t txg = tx->tx_txg;
44 
45 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
46 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
47 	/* this dnode can't be paged out because it's dirty */
48 
49 	db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
50 	ASSERT(db != NULL);
51 	for (i = 0; i < dn->dn_phys->dn_nblkptr; i++)
52 		if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i]))
53 			break;
54 	if (i != dn->dn_phys->dn_nblkptr) {
55 		ASSERT(list_link_active(&db->db_dirty_node[txg&TXG_MASK]));
56 
57 		(void) dbuf_read(db, NULL,
58 		    DB_RF_HAVESTRUCT | DB_RF_MUST_SUCCEED);
59 		arc_release(db->db_buf, db);
60 		/* copy dnode's block pointers to new indirect block */
61 		ASSERT3U(sizeof (blkptr_t) * dn->dn_phys->dn_nblkptr, <=,
62 		    db->db.db_size);
63 		bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
64 		    sizeof (blkptr_t) * dn->dn_phys->dn_nblkptr);
65 	}
66 
67 	dn->dn_phys->dn_nlevels += 1;
68 	dprintf("os=%p obj=%llu, increase to %d\n",
69 		dn->dn_objset, dn->dn_object,
70 		dn->dn_phys->dn_nlevels);
71 
72 	/* set dbuf's parent pointers to new indirect buf */
73 	for (i = 0; i < dn->dn_phys->dn_nblkptr; i++) {
74 		dmu_buf_impl_t *child =
75 		    dbuf_find(dn, dn->dn_phys->dn_nlevels-2, i);
76 		if (child == NULL)
77 			continue;
78 		if (child->db_dnode == NULL) {
79 			mutex_exit(&child->db_mtx);
80 			continue;
81 		}
82 
83 		if (child->db_parent == NULL ||
84 		    child->db_parent == dn->dn_dbuf) {
85 			dprintf_dbuf_bp(child, child->db_blkptr,
86 			    "changing db_blkptr to new indirect %s", "");
87 			child->db_parent = db;
88 			dbuf_add_ref(db, child);
89 			if (db->db.db_data) {
90 				child->db_blkptr =
91 				    (blkptr_t *)db->db.db_data + i;
92 			} else {
93 				child->db_blkptr = NULL;
94 			}
95 			dprintf_dbuf_bp(child, child->db_blkptr,
96 			    "changed db_blkptr to new indirect %s", "");
97 		}
98 		ASSERT3P(child->db_parent, ==, db);
99 
100 		mutex_exit(&child->db_mtx);
101 	}
102 
103 	bzero(dn->dn_phys->dn_blkptr,
104 		sizeof (blkptr_t) * dn->dn_phys->dn_nblkptr);
105 
106 	dbuf_rele(db, FTAG);
107 }
108 
109 static void
110 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
111 {
112 	objset_impl_t *os = dn->dn_objset;
113 	uint64_t bytesfreed = 0;
114 	int i;
115 
116 	dprintf("os=%p obj=%llx num=%d\n", os, dn->dn_object, num);
117 
118 	for (i = 0; i < num; i++, bp++) {
119 		if (BP_IS_HOLE(bp))
120 			continue;
121 
122 		bytesfreed += bp_get_dasize(os->os_spa, bp);
123 		ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
124 		dsl_dataset_block_kill(os->os_dsl_dataset, bp, tx);
125 	}
126 	dnode_diduse_space(dn, -bytesfreed);
127 }
128 
129 #ifdef ZFS_DEBUG
130 static void
131 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
132 {
133 	int off, num;
134 	int i, err, epbs;
135 	uint64_t txg = tx->tx_txg;
136 
137 	epbs = db->db_dnode->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
138 	off = start - (db->db_blkid * 1<<epbs);
139 	num = end - start + 1;
140 
141 	ASSERT3U(off, >=, 0);
142 	ASSERT3U(num, >=, 0);
143 	ASSERT3U(db->db_level, >, 0);
144 	ASSERT3U(db->db.db_size, ==, 1<<db->db_dnode->dn_phys->dn_indblkshift);
145 	ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
146 	ASSERT(db->db_blkptr != NULL);
147 
148 	for (i = off; i < off+num; i++) {
149 		uint64_t *buf;
150 		int j;
151 		dmu_buf_impl_t *child;
152 
153 		ASSERT(db->db_level == 1);
154 
155 		rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER);
156 		err = dbuf_hold_impl(db->db_dnode, db->db_level-1,
157 			(db->db_blkid << epbs) + i, TRUE, FTAG, &child);
158 		rw_exit(&db->db_dnode->dn_struct_rwlock);
159 		if (err == ENOENT)
160 			continue;
161 		ASSERT(err == 0);
162 		ASSERT(child->db_level == 0);
163 		ASSERT(!list_link_active(&child->db_dirty_node[txg&TXG_MASK]));
164 
165 		/* db_data_old better be zeroed */
166 		if (child->db_d.db_data_old[txg & TXG_MASK]) {
167 			buf = ((arc_buf_t *)child->db_d.db_data_old
168 			    [txg & TXG_MASK])->b_data;
169 			for (j = 0; j < child->db.db_size >> 3; j++) {
170 				if (buf[j] != 0) {
171 					panic("freed data not zero: "
172 					    "child=%p i=%d off=%d num=%d\n",
173 					    child, i, off, num);
174 				}
175 			}
176 		}
177 
178 		/*
179 		 * db_data better be zeroed unless it's dirty in a
180 		 * future txg.
181 		 */
182 		mutex_enter(&child->db_mtx);
183 		buf = child->db.db_data;
184 		if (buf != NULL && child->db_state != DB_FILL &&
185 		    !list_link_active(&child->db_dirty_node
186 			[(txg+1) & TXG_MASK]) &&
187 		    !list_link_active(&child->db_dirty_node
188 			[(txg+2) & TXG_MASK])) {
189 			for (j = 0; j < child->db.db_size >> 3; j++) {
190 				if (buf[j] != 0) {
191 					panic("freed data not zero: "
192 					    "child=%p i=%d off=%d num=%d\n",
193 					    child, i, off, num);
194 				}
195 			}
196 		}
197 		mutex_exit(&child->db_mtx);
198 
199 		dbuf_rele(child, FTAG);
200 	}
201 }
202 #endif
203 
204 static int
205 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks, int trunc,
206     dmu_tx_t *tx)
207 {
208 	dnode_t *dn = db->db_dnode;
209 	blkptr_t *bp;
210 	dmu_buf_impl_t *subdb;
211 	uint64_t start, end, dbstart, dbend, i;
212 	int epbs, shift, err;
213 	int txgoff = tx->tx_txg & TXG_MASK;
214 	int all = TRUE;
215 
216 	(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
217 	arc_release(db->db_buf, db);
218 	bp = (blkptr_t *)db->db.db_data;
219 
220 	epbs = db->db_dnode->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
221 	shift = (db->db_level - 1) * epbs;
222 	dbstart = db->db_blkid << epbs;
223 	start = blkid >> shift;
224 	if (dbstart < start) {
225 		bp += start - dbstart;
226 		all = FALSE;
227 	} else {
228 		start = dbstart;
229 	}
230 	dbend = ((db->db_blkid + 1) << epbs) - 1;
231 	end = (blkid + nblks - 1) >> shift;
232 	if (dbend <= end)
233 		end = dbend;
234 	else if (all)
235 		all = trunc;
236 	ASSERT3U(start, <=, end);
237 
238 	if (db->db_level == 1) {
239 		FREE_VERIFY(db, start, end, tx);
240 		free_blocks(dn, bp, end-start+1, tx);
241 		ASSERT(all || list_link_active(&db->db_dirty_node[txgoff]));
242 		return (all);
243 	}
244 
245 	for (i = start; i <= end; i++, bp++) {
246 		if (BP_IS_HOLE(bp))
247 			continue;
248 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
249 		err = dbuf_hold_impl(dn, db->db_level-1, i, TRUE, FTAG, &subdb);
250 		ASSERT3U(err, ==, 0);
251 		rw_exit(&dn->dn_struct_rwlock);
252 
253 		if (free_children(subdb, blkid, nblks, trunc, tx)) {
254 			ASSERT3P(subdb->db_blkptr, ==, bp);
255 			free_blocks(dn, bp, 1, tx);
256 		} else {
257 			all = FALSE;
258 		}
259 		dbuf_rele(subdb, FTAG);
260 	}
261 #ifdef ZFS_DEBUG
262 	bp -= (end-start)+1;
263 	for (i = start; i <= end; i++, bp++) {
264 		if (i == start && blkid != 0)
265 			continue;
266 		else if (i == end && !trunc)
267 			continue;
268 		ASSERT3U(bp->blk_birth, ==, 0);
269 	}
270 #endif
271 	ASSERT(all || list_link_active(&db->db_dirty_node[txgoff]));
272 	return (all);
273 }
274 
275 /*
276  * free_range: Traverse the indicated range of the provided file
277  * and "free" all the blocks contained there.
278  */
279 static void
280 dnode_sync_free_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
281 {
282 	blkptr_t *bp = dn->dn_phys->dn_blkptr;
283 	dmu_buf_impl_t *db;
284 	int trunc, start, end, shift, i, err;
285 	int dnlevel = dn->dn_phys->dn_nlevels;
286 
287 	if (blkid > dn->dn_phys->dn_maxblkid)
288 		return;
289 
290 	ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
291 	trunc = blkid + nblks > dn->dn_phys->dn_maxblkid;
292 	if (trunc)
293 		nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
294 
295 	/* There are no indirect blocks in the object */
296 	if (dnlevel == 1) {
297 		if (blkid >= dn->dn_phys->dn_nblkptr) {
298 			/* this range was never made persistent */
299 			return;
300 		}
301 		ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
302 		free_blocks(dn, bp + blkid, nblks, tx);
303 		if (trunc) {
304 			uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
305 			    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
306 			dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
307 			ASSERT(off < dn->dn_phys->dn_maxblkid ||
308 			    dn->dn_phys->dn_maxblkid == 0 ||
309 			    dnode_next_offset(dn, FALSE, &off,
310 			    1, 1, 0) != 0);
311 		}
312 		return;
313 	}
314 
315 	shift = (dnlevel - 1) * (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
316 	start = blkid >> shift;
317 	ASSERT(start < dn->dn_phys->dn_nblkptr);
318 	end = (blkid + nblks - 1) >> shift;
319 	bp += start;
320 	for (i = start; i <= end; i++, bp++) {
321 		if (BP_IS_HOLE(bp))
322 			continue;
323 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
324 		err = dbuf_hold_impl(dn, dnlevel-1, i, TRUE, FTAG, &db);
325 		ASSERT3U(err, ==, 0);
326 		rw_exit(&dn->dn_struct_rwlock);
327 
328 		if (free_children(db, blkid, nblks, trunc, tx)) {
329 			ASSERT3P(db->db_blkptr, ==, bp);
330 			free_blocks(dn, bp, 1, tx);
331 		}
332 		dbuf_rele(db, FTAG);
333 	}
334 	if (trunc) {
335 		uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
336 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
337 		dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
338 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
339 		    dn->dn_phys->dn_maxblkid == 0 ||
340 		    dnode_next_offset(dn, FALSE, &off, 1, 1, 0) != 0);
341 	}
342 }
343 
344 /*
345  * Try to kick all the dnodes dbufs out of the cache...
346  */
347 int
348 dnode_evict_dbufs(dnode_t *dn, int try)
349 {
350 	int progress;
351 	int pass = 0;
352 
353 	do {
354 		dmu_buf_impl_t *db, marker;
355 		int evicting = FALSE;
356 
357 		progress = FALSE;
358 		mutex_enter(&dn->dn_dbufs_mtx);
359 		list_insert_tail(&dn->dn_dbufs, &marker);
360 		db = list_head(&dn->dn_dbufs);
361 		for (; db != &marker; db = list_head(&dn->dn_dbufs)) {
362 			list_remove(&dn->dn_dbufs, db);
363 			list_insert_tail(&dn->dn_dbufs, db);
364 
365 			mutex_enter(&db->db_mtx);
366 			if (db->db_state == DB_EVICTING) {
367 				progress = TRUE;
368 				evicting = TRUE;
369 				mutex_exit(&db->db_mtx);
370 			} else if (refcount_is_zero(&db->db_holds)) {
371 				progress = TRUE;
372 				ASSERT(!arc_released(db->db_buf));
373 				dbuf_clear(db); /* exits db_mtx for us */
374 			} else {
375 				mutex_exit(&db->db_mtx);
376 			}
377 
378 		}
379 		list_remove(&dn->dn_dbufs, &marker);
380 		/*
381 		 * NB: we need to drop dn_dbufs_mtx between passes so
382 		 * that any DB_EVICTING dbufs can make progress.
383 		 * Ideally, we would have some cv we could wait on, but
384 		 * since we don't, just wait a bit to give the other
385 		 * thread a chance to run.
386 		 */
387 		mutex_exit(&dn->dn_dbufs_mtx);
388 		if (evicting)
389 			delay(1);
390 		pass++;
391 		ASSERT(pass < 100); /* sanity check */
392 	} while (progress);
393 
394 	/*
395 	 * This function works fine even if it can't evict everything.
396 	 * If were only asked to try to evict everything then
397 	 * return an error if we can't. Otherwise panic as the caller
398 	 * expects total eviction.
399 	 */
400 	if (list_head(&dn->dn_dbufs) != NULL) {
401 		if (try) {
402 			return (1);
403 		} else {
404 			panic("dangling dbufs (dn=%p, dbuf=%p)\n",
405 			    dn, list_head(&dn->dn_dbufs));
406 		}
407 	}
408 
409 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
410 	if (dn->dn_bonus && refcount_is_zero(&dn->dn_bonus->db_holds)) {
411 		mutex_enter(&dn->dn_bonus->db_mtx);
412 		dbuf_evict(dn->dn_bonus);
413 		dn->dn_bonus = NULL;
414 	}
415 	rw_exit(&dn->dn_struct_rwlock);
416 	return (0);
417 }
418 
419 static int
420 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
421 {
422 	dmu_buf_impl_t *db;
423 	int txgoff = tx->tx_txg & TXG_MASK;
424 
425 	ASSERT(dmu_tx_is_syncing(tx));
426 
427 	/* Undirty all buffers */
428 	while (db = list_head(&dn->dn_dirty_dbufs[txgoff])) {
429 		mutex_enter(&db->db_mtx);
430 		/* XXX - use dbuf_undirty()? */
431 		list_remove(&dn->dn_dirty_dbufs[txgoff], db);
432 		if (db->db_level == 0) {
433 			ASSERT(db->db_blkid == DB_BONUS_BLKID ||
434 			    db->db_d.db_data_old[txgoff] == db->db_buf);
435 			if (db->db_d.db_overridden_by[txgoff])
436 				dbuf_unoverride(db, tx->tx_txg);
437 			db->db_d.db_data_old[txgoff] = NULL;
438 		}
439 		db->db_dirtycnt -= 1;
440 		mutex_exit(&db->db_mtx);
441 		dbuf_rele(db, (void *)(uintptr_t)tx->tx_txg);
442 	}
443 
444 	(void) dnode_evict_dbufs(dn, 0);
445 	ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
446 
447 	/*
448 	 * XXX - It would be nice to assert this, but we may still
449 	 * have residual holds from async evictions from the arc...
450 	 *
451 	 * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
452 	 */
453 
454 	/* Undirty next bits */
455 	dn->dn_next_nlevels[txgoff] = 0;
456 	dn->dn_next_indblkshift[txgoff] = 0;
457 	dn->dn_next_blksz[txgoff] = 0;
458 
459 	/* free up all the blocks in the file. */
460 	dnode_sync_free_range(dn, 0, dn->dn_phys->dn_maxblkid+1, tx);
461 	ASSERT3U(DN_USED_BYTES(dn->dn_phys), ==, 0);
462 
463 	/* ASSERT(blkptrs are zero); */
464 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
465 	ASSERT(dn->dn_type != DMU_OT_NONE);
466 
467 	ASSERT(dn->dn_free_txg > 0);
468 	if (dn->dn_allocated_txg != dn->dn_free_txg)
469 		dbuf_will_dirty(dn->dn_dbuf, tx);
470 	bzero(dn->dn_phys, sizeof (dnode_phys_t));
471 
472 	mutex_enter(&dn->dn_mtx);
473 	dn->dn_type = DMU_OT_NONE;
474 	dn->dn_maxblkid = 0;
475 	dn->dn_allocated_txg = 0;
476 	mutex_exit(&dn->dn_mtx);
477 
478 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
479 
480 	dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
481 	/*
482 	 * Now that we've released our hold, the dnode may
483 	 * be evicted, so we musn't access it.
484 	 */
485 	return (1);
486 }
487 
488 /*
489  * Write out the dnode's dirty buffers at the specified level.
490  * This may create more dirty buffers at the next level up.
491  *
492  * NOTE: The dnode is kept in memory by being dirty.  Once the
493  * dirty bit is cleared, it may be evicted.  Beware of this!
494  */
495 int
496 dnode_sync(dnode_t *dn, int level, zio_t *zio, dmu_tx_t *tx)
497 {
498 	free_range_t *rp;
499 	int txgoff = tx->tx_txg & TXG_MASK;
500 	dnode_phys_t *dnp = dn->dn_phys;
501 
502 	ASSERT(dmu_tx_is_syncing(tx));
503 	ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
504 	DNODE_VERIFY(dn);
505 
506 	/*
507 	 * Make sure the dbuf for the dn_phys is released before we modify it.
508 	 */
509 	if (dn->dn_dbuf)
510 		arc_release(dn->dn_dbuf->db_buf, dn->dn_dbuf);
511 
512 	mutex_enter(&dn->dn_mtx);
513 	if (dn->dn_allocated_txg == tx->tx_txg) {
514 		/* The dnode is newly allocated or reallocated */
515 		if (dnp->dn_type == DMU_OT_NONE) {
516 			/* this is a first alloc, not a realloc */
517 			/* XXX shouldn't the phys already be zeroed? */
518 			bzero(dnp, DNODE_CORE_SIZE);
519 			dnp->dn_nlevels = 1;
520 		}
521 
522 		if (dn->dn_nblkptr > dnp->dn_nblkptr) {
523 			/* zero the new blkptrs we are gaining */
524 			bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
525 			    sizeof (blkptr_t) *
526 			    (dn->dn_nblkptr - dnp->dn_nblkptr));
527 		}
528 		dnp->dn_type = dn->dn_type;
529 		dnp->dn_bonustype = dn->dn_bonustype;
530 		dnp->dn_bonuslen = dn->dn_bonuslen;
531 		dnp->dn_nblkptr = dn->dn_nblkptr;
532 	}
533 
534 	ASSERT(level != 0 || dnp->dn_nlevels > 1 ||
535 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
536 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
537 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
538 
539 	if (dn->dn_next_blksz[txgoff]) {
540 		ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
541 		    SPA_MINBLOCKSIZE) == 0);
542 		ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
543 		    list_head(&dn->dn_dirty_dbufs[txgoff]) != NULL ||
544 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
545 		    dnp->dn_datablkszsec);
546 		dnp->dn_datablkszsec =
547 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
548 		dn->dn_next_blksz[txgoff] = 0;
549 	}
550 
551 	if (dn->dn_next_indblkshift[txgoff]) {
552 		ASSERT(dnp->dn_nlevels == 1);
553 		dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
554 		dn->dn_next_indblkshift[txgoff] = 0;
555 	}
556 
557 	/*
558 	 * Just take the live (open-context) values for checksum and compress.
559 	 * Strictly speaking it's a future leak, but nothing bad happens if we
560 	 * start using the new checksum or compress algorithm a little early.
561 	 */
562 	dnp->dn_checksum = dn->dn_checksum;
563 	dnp->dn_compress = dn->dn_compress;
564 
565 	mutex_exit(&dn->dn_mtx);
566 
567 	/* process all the "freed" ranges in the file */
568 	if (dn->dn_free_txg == 0 || dn->dn_free_txg > tx->tx_txg) {
569 		for (rp = avl_last(&dn->dn_ranges[txgoff]); rp != NULL;
570 		    rp = AVL_PREV(&dn->dn_ranges[txgoff], rp))
571 			dnode_sync_free_range(dn,
572 			    rp->fr_blkid, rp->fr_nblks, tx);
573 	}
574 	mutex_enter(&dn->dn_mtx);
575 	for (rp = avl_first(&dn->dn_ranges[txgoff]); rp; ) {
576 		free_range_t *last = rp;
577 		rp = AVL_NEXT(&dn->dn_ranges[txgoff], rp);
578 		avl_remove(&dn->dn_ranges[txgoff], last);
579 		kmem_free(last, sizeof (free_range_t));
580 	}
581 	mutex_exit(&dn->dn_mtx);
582 
583 	if (dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg) {
584 		ASSERT3U(level, ==, 0);
585 		return (dnode_sync_free(dn, tx));
586 	}
587 
588 	if (dn->dn_next_nlevels[txgoff]) {
589 		int new_lvl = dn->dn_next_nlevels[txgoff];
590 
591 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
592 		while (new_lvl > dnp->dn_nlevels)
593 			dnode_increase_indirection(dn, tx);
594 		rw_exit(&dn->dn_struct_rwlock);
595 		dn->dn_next_nlevels[txgoff] = 0;
596 	}
597 
598 	if (level == dnp->dn_nlevels) {
599 		uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
600 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
601 
602 		/* we've already synced out all data and indirect blocks */
603 		/* there are no more dirty dbufs under this dnode */
604 		ASSERT3P(list_head(&dn->dn_dirty_dbufs[txgoff]), ==, NULL);
605 		ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= tx->tx_txg);
606 
607 		/* NB: the "off < maxblkid" is to catch overflow */
608 		/*
609 		 * NB: if blocksize is changing, we could get confused,
610 		 * so only bother if there are multiple blocks and thus
611 		 * it can't be changing.
612 		 */
613 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
614 		    dn->dn_phys->dn_maxblkid == 0 ||
615 		    dnode_next_offset(dn, FALSE, &off, 1, 1, 0) != 0);
616 
617 		ASSERT(dnp->dn_nlevels > 1 ||
618 		    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
619 		    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
620 		    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
621 
622 		if (dn->dn_object != DMU_META_DNODE_OBJECT) {
623 			dbuf_will_dirty(dn->dn_dbuf, tx);
624 			dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
625 		}
626 
627 		/*
628 		 * Now that we've dropped the reference, the dnode may
629 		 * be evicted, so we musn't access it.
630 		 */
631 		return (1);
632 	} else {
633 		dmu_buf_impl_t *db, *db_next;
634 		list_t *list = &dn->dn_dirty_dbufs[txgoff];
635 		/*
636 		 * Iterate over the list, removing and sync'ing dbufs
637 		 * which are on the level we want, and leaving others.
638 		 */
639 		for (db = list_head(list); db; db = db_next) {
640 			db_next = list_next(list, db);
641 			if (db->db_level == level) {
642 				list_remove(list, db);
643 				dbuf_sync(db, zio, tx);
644 			}
645 		}
646 		return (0);
647 	}
648 }
649