xref: /titanic_50/usr/src/uts/common/fs/zfs/dnode_sync.c (revision b4756084ba1ef238a1e15b1585b853a7c1f85582)
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_tx.h>
30 #include <sys/dmu_objset.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/spa.h>
33 
34 static void
35 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
36 {
37 	dmu_buf_impl_t *db;
38 	int txgoff = tx->tx_txg & TXG_MASK;
39 	int nblkptr = dn->dn_phys->dn_nblkptr;
40 	int old_toplvl = dn->dn_phys->dn_nlevels - 1;
41 	int new_level = dn->dn_next_nlevels[txgoff];
42 	int i;
43 
44 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
45 
46 	/* this dnode can't be paged out because it's dirty */
47 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
48 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
49 	ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
50 
51 	db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
52 	ASSERT(db != NULL);
53 
54 	dn->dn_phys->dn_nlevels = new_level;
55 	dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
56 	    dn->dn_object, dn->dn_phys->dn_nlevels);
57 
58 	/* check for existing blkptrs in the dnode */
59 	for (i = 0; i < nblkptr; i++)
60 		if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i]))
61 			break;
62 	if (i != nblkptr) {
63 		/* transfer dnode's block pointers to new indirect block */
64 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
65 		ASSERT(db->db.db_data);
66 		ASSERT(arc_released(db->db_buf));
67 		ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
68 		bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
69 		    sizeof (blkptr_t) * nblkptr);
70 		arc_buf_freeze(db->db_buf);
71 	}
72 
73 	/* set dbuf's parent pointers to new indirect buf */
74 	for (i = 0; i < nblkptr; i++) {
75 		dmu_buf_impl_t *child = dbuf_find(dn, old_toplvl, i);
76 
77 		if (child == NULL)
78 			continue;
79 		ASSERT3P(child->db_dnode, ==, dn);
80 		if (child->db_parent && child->db_parent != dn->dn_dbuf) {
81 			ASSERT(child->db_parent->db_level == db->db_level);
82 			ASSERT(child->db_blkptr !=
83 			    &dn->dn_phys->dn_blkptr[child->db_blkid]);
84 			mutex_exit(&child->db_mtx);
85 			continue;
86 		}
87 		ASSERT(child->db_parent == NULL ||
88 		    child->db_parent == dn->dn_dbuf);
89 
90 		child->db_parent = db;
91 		dbuf_add_ref(db, child);
92 		if (db->db.db_data)
93 			child->db_blkptr = (blkptr_t *)db->db.db_data + i;
94 		else
95 			child->db_blkptr = NULL;
96 		dprintf_dbuf_bp(child, child->db_blkptr,
97 		    "changed db_blkptr to new indirect %s", "");
98 
99 		mutex_exit(&child->db_mtx);
100 	}
101 
102 	bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
103 
104 	dbuf_rele(db, FTAG);
105 
106 	rw_exit(&dn->dn_struct_rwlock);
107 }
108 
109 static int
110 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
111 {
112 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
113 	uint64_t bytesfreed = 0;
114 	int i, blocks_freed = 0;
115 
116 	dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
117 
118 	for (i = 0; i < num; i++, bp++) {
119 		if (BP_IS_HOLE(bp))
120 			continue;
121 
122 		bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
123 		ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
124 		bzero(bp, sizeof (blkptr_t));
125 		blocks_freed += 1;
126 	}
127 	dnode_diduse_space(dn, -bytesfreed);
128 	return (blocks_freed);
129 }
130 
131 #ifdef ZFS_DEBUG
132 static void
133 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
134 {
135 	int off, num;
136 	int i, err, epbs;
137 	uint64_t txg = tx->tx_txg;
138 
139 	epbs = db->db_dnode->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
140 	off = start - (db->db_blkid * 1<<epbs);
141 	num = end - start + 1;
142 
143 	ASSERT3U(off, >=, 0);
144 	ASSERT3U(num, >=, 0);
145 	ASSERT3U(db->db_level, >, 0);
146 	ASSERT3U(db->db.db_size, ==, 1<<db->db_dnode->dn_phys->dn_indblkshift);
147 	ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
148 	ASSERT(db->db_blkptr != NULL);
149 
150 	for (i = off; i < off+num; i++) {
151 		uint64_t *buf;
152 		dmu_buf_impl_t *child;
153 		dbuf_dirty_record_t *dr;
154 		int j;
155 
156 		ASSERT(db->db_level == 1);
157 
158 		rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER);
159 		err = dbuf_hold_impl(db->db_dnode, db->db_level-1,
160 		    (db->db_blkid << epbs) + i, TRUE, FTAG, &child);
161 		rw_exit(&db->db_dnode->dn_struct_rwlock);
162 		if (err == ENOENT)
163 			continue;
164 		ASSERT(err == 0);
165 		ASSERT(child->db_level == 0);
166 		dr = child->db_last_dirty;
167 		while (dr && dr->dr_txg > txg)
168 			dr = dr->dr_next;
169 		ASSERT(dr == NULL || dr->dr_txg == txg);
170 
171 		/* data_old better be zeroed */
172 		if (dr) {
173 			buf = dr->dt.dl.dr_data->b_data;
174 			for (j = 0; j < child->db.db_size >> 3; j++) {
175 				if (buf[j] != 0) {
176 					panic("freed data not zero: "
177 					    "child=%p i=%d off=%d num=%d\n",
178 					    (void *)child, i, off, num);
179 				}
180 			}
181 		}
182 
183 		/*
184 		 * db_data better be zeroed unless it's dirty in a
185 		 * future txg.
186 		 */
187 		mutex_enter(&child->db_mtx);
188 		buf = child->db.db_data;
189 		if (buf != NULL && child->db_state != DB_FILL &&
190 		    child->db_last_dirty == NULL) {
191 			for (j = 0; j < child->db.db_size >> 3; j++) {
192 				if (buf[j] != 0) {
193 					panic("freed data not zero: "
194 					    "child=%p i=%d off=%d num=%d\n",
195 					    (void *)child, i, off, num);
196 				}
197 			}
198 		}
199 		mutex_exit(&child->db_mtx);
200 
201 		dbuf_rele(child, FTAG);
202 	}
203 }
204 #endif
205 
206 #define	ALL -1
207 
208 static int
209 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks, int trunc,
210     dmu_tx_t *tx)
211 {
212 	dnode_t *dn = db->db_dnode;
213 	blkptr_t *bp;
214 	dmu_buf_impl_t *subdb;
215 	uint64_t start, end, dbstart, dbend, i;
216 	int epbs, shift, err;
217 	int all = TRUE;
218 	int blocks_freed = 0;
219 
220 	/*
221 	 * There is a small possibility that this block will not be cached:
222 	 *   1 - if level > 1 and there are no children with level <= 1
223 	 *   2 - if we didn't get a dirty hold (because this block had just
224 	 *	 finished being written -- and so had no holds), and then this
225 	 *	 block got evicted before we got here.
226 	 */
227 	if (db->db_state != DB_CACHED)
228 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
229 
230 	arc_release(db->db_buf, db);
231 	bp = (blkptr_t *)db->db.db_data;
232 
233 	epbs = db->db_dnode->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
234 	shift = (db->db_level - 1) * epbs;
235 	dbstart = db->db_blkid << epbs;
236 	start = blkid >> shift;
237 	if (dbstart < start) {
238 		bp += start - dbstart;
239 		all = FALSE;
240 	} else {
241 		start = dbstart;
242 	}
243 	dbend = ((db->db_blkid + 1) << epbs) - 1;
244 	end = (blkid + nblks - 1) >> shift;
245 	if (dbend <= end)
246 		end = dbend;
247 	else if (all)
248 		all = trunc;
249 	ASSERT3U(start, <=, end);
250 
251 	if (db->db_level == 1) {
252 		FREE_VERIFY(db, start, end, tx);
253 		blocks_freed = free_blocks(dn, bp, end-start+1, tx);
254 		arc_buf_freeze(db->db_buf);
255 		ASSERT(all || blocks_freed == 0 || db->db_last_dirty);
256 		return (all ? ALL : blocks_freed);
257 	}
258 
259 	for (i = start; i <= end; i++, bp++) {
260 		if (BP_IS_HOLE(bp))
261 			continue;
262 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
263 		err = dbuf_hold_impl(dn, db->db_level-1, i, TRUE, FTAG, &subdb);
264 		ASSERT3U(err, ==, 0);
265 		rw_exit(&dn->dn_struct_rwlock);
266 
267 		if (free_children(subdb, blkid, nblks, trunc, tx) == ALL) {
268 			ASSERT3P(subdb->db_blkptr, ==, bp);
269 			blocks_freed += free_blocks(dn, bp, 1, tx);
270 		} else {
271 			all = FALSE;
272 		}
273 		dbuf_rele(subdb, FTAG);
274 	}
275 	arc_buf_freeze(db->db_buf);
276 #ifdef ZFS_DEBUG
277 	bp -= (end-start)+1;
278 	for (i = start; i <= end; i++, bp++) {
279 		if (i == start && blkid != 0)
280 			continue;
281 		else if (i == end && !trunc)
282 			continue;
283 		ASSERT3U(bp->blk_birth, ==, 0);
284 	}
285 #endif
286 	ASSERT(all || blocks_freed == 0 || db->db_last_dirty);
287 	return (all ? ALL : blocks_freed);
288 }
289 
290 /*
291  * free_range: Traverse the indicated range of the provided file
292  * and "free" all the blocks contained there.
293  */
294 static void
295 dnode_sync_free_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
296 {
297 	blkptr_t *bp = dn->dn_phys->dn_blkptr;
298 	dmu_buf_impl_t *db;
299 	int trunc, start, end, shift, i, err;
300 	int dnlevel = dn->dn_phys->dn_nlevels;
301 
302 	if (blkid > dn->dn_phys->dn_maxblkid)
303 		return;
304 
305 	ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
306 	trunc = blkid + nblks > dn->dn_phys->dn_maxblkid;
307 	if (trunc)
308 		nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
309 
310 	/* There are no indirect blocks in the object */
311 	if (dnlevel == 1) {
312 		if (blkid >= dn->dn_phys->dn_nblkptr) {
313 			/* this range was never made persistent */
314 			return;
315 		}
316 		ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
317 		(void) free_blocks(dn, bp + blkid, nblks, tx);
318 		if (trunc) {
319 			uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
320 			    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
321 			dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
322 			ASSERT(off < dn->dn_phys->dn_maxblkid ||
323 			    dn->dn_phys->dn_maxblkid == 0 ||
324 			    dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
325 		}
326 		return;
327 	}
328 
329 	shift = (dnlevel - 1) * (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
330 	start = blkid >> shift;
331 	ASSERT(start < dn->dn_phys->dn_nblkptr);
332 	end = (blkid + nblks - 1) >> shift;
333 	bp += start;
334 	for (i = start; i <= end; i++, bp++) {
335 		if (BP_IS_HOLE(bp))
336 			continue;
337 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
338 		err = dbuf_hold_impl(dn, dnlevel-1, i, TRUE, FTAG, &db);
339 		ASSERT3U(err, ==, 0);
340 		rw_exit(&dn->dn_struct_rwlock);
341 
342 		if (free_children(db, blkid, nblks, trunc, tx) == ALL) {
343 			ASSERT3P(db->db_blkptr, ==, bp);
344 			(void) free_blocks(dn, bp, 1, tx);
345 		}
346 		dbuf_rele(db, FTAG);
347 	}
348 	if (trunc) {
349 		uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
350 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
351 		dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
352 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
353 		    dn->dn_phys->dn_maxblkid == 0 ||
354 		    dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
355 	}
356 }
357 
358 /*
359  * Try to kick all the dnodes dbufs out of the cache...
360  */
361 void
362 dnode_evict_dbufs(dnode_t *dn)
363 {
364 	int progress;
365 	int pass = 0;
366 
367 	do {
368 		dmu_buf_impl_t *db, marker;
369 		int evicting = FALSE;
370 
371 		progress = FALSE;
372 		mutex_enter(&dn->dn_dbufs_mtx);
373 		list_insert_tail(&dn->dn_dbufs, &marker);
374 		db = list_head(&dn->dn_dbufs);
375 		for (; db != &marker; db = list_head(&dn->dn_dbufs)) {
376 			list_remove(&dn->dn_dbufs, db);
377 			list_insert_tail(&dn->dn_dbufs, db);
378 			ASSERT3P(db->db_dnode, ==, dn);
379 
380 			mutex_enter(&db->db_mtx);
381 			if (db->db_state == DB_EVICTING) {
382 				progress = TRUE;
383 				evicting = TRUE;
384 				mutex_exit(&db->db_mtx);
385 			} else if (refcount_is_zero(&db->db_holds)) {
386 				progress = TRUE;
387 				dbuf_clear(db); /* exits db_mtx for us */
388 			} else {
389 				mutex_exit(&db->db_mtx);
390 			}
391 
392 		}
393 		list_remove(&dn->dn_dbufs, &marker);
394 		/*
395 		 * NB: we need to drop dn_dbufs_mtx between passes so
396 		 * that any DB_EVICTING dbufs can make progress.
397 		 * Ideally, we would have some cv we could wait on, but
398 		 * since we don't, just wait a bit to give the other
399 		 * thread a chance to run.
400 		 */
401 		mutex_exit(&dn->dn_dbufs_mtx);
402 		if (evicting)
403 			delay(1);
404 		pass++;
405 		ASSERT(pass < 100); /* sanity check */
406 	} while (progress);
407 
408 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
409 	if (dn->dn_bonus && refcount_is_zero(&dn->dn_bonus->db_holds)) {
410 		mutex_enter(&dn->dn_bonus->db_mtx);
411 		dbuf_evict(dn->dn_bonus);
412 		dn->dn_bonus = NULL;
413 	}
414 	rw_exit(&dn->dn_struct_rwlock);
415 }
416 
417 static void
418 dnode_undirty_dbufs(list_t *list)
419 {
420 	dbuf_dirty_record_t *dr;
421 
422 	while (dr = list_head(list)) {
423 		dmu_buf_impl_t *db = dr->dr_dbuf;
424 		uint64_t txg = dr->dr_txg;
425 
426 		if (db->db_level != 0)
427 			dnode_undirty_dbufs(&dr->dt.di.dr_children);
428 
429 		mutex_enter(&db->db_mtx);
430 		/* XXX - use dbuf_undirty()? */
431 		list_remove(list, dr);
432 		ASSERT(db->db_last_dirty == dr);
433 		db->db_last_dirty = NULL;
434 		db->db_dirtycnt -= 1;
435 		if (db->db_level == 0) {
436 			ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
437 			    dr->dt.dl.dr_data == db->db_buf);
438 			dbuf_unoverride(dr);
439 		}
440 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
441 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
442 	}
443 }
444 
445 static void
446 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
447 {
448 	int txgoff = tx->tx_txg & TXG_MASK;
449 
450 	ASSERT(dmu_tx_is_syncing(tx));
451 
452 	/*
453 	 * Our contents should have been freed in dnode_sync() by the
454 	 * free range record inserted by the caller of dnode_free().
455 	 */
456 	ASSERT3U(DN_USED_BYTES(dn->dn_phys), ==, 0);
457 	ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
458 
459 	dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
460 	dnode_evict_dbufs(dn);
461 	ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
462 
463 	/*
464 	 * XXX - It would be nice to assert this, but we may still
465 	 * have residual holds from async evictions from the arc...
466 	 *
467 	 * zfs_obj_to_path() also depends on this being
468 	 * commented out.
469 	 *
470 	 * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
471 	 */
472 
473 	/* Undirty next bits */
474 	dn->dn_next_nlevels[txgoff] = 0;
475 	dn->dn_next_indblkshift[txgoff] = 0;
476 	dn->dn_next_blksz[txgoff] = 0;
477 
478 	/* ASSERT(blkptrs are zero); */
479 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
480 	ASSERT(dn->dn_type != DMU_OT_NONE);
481 
482 	ASSERT(dn->dn_free_txg > 0);
483 	if (dn->dn_allocated_txg != dn->dn_free_txg)
484 		dbuf_will_dirty(dn->dn_dbuf, tx);
485 	bzero(dn->dn_phys, sizeof (dnode_phys_t));
486 
487 	mutex_enter(&dn->dn_mtx);
488 	dn->dn_type = DMU_OT_NONE;
489 	dn->dn_maxblkid = 0;
490 	dn->dn_allocated_txg = 0;
491 	dn->dn_free_txg = 0;
492 	dn->dn_have_spill = B_FALSE;
493 	mutex_exit(&dn->dn_mtx);
494 
495 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
496 
497 	dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
498 	/*
499 	 * Now that we've released our hold, the dnode may
500 	 * be evicted, so we musn't access it.
501 	 */
502 }
503 
504 /*
505  * Write out the dnode's dirty buffers.
506  */
507 void
508 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
509 {
510 	free_range_t *rp;
511 	dnode_phys_t *dnp = dn->dn_phys;
512 	int txgoff = tx->tx_txg & TXG_MASK;
513 	list_t *list = &dn->dn_dirty_records[txgoff];
514 	static const dnode_phys_t zerodn = { 0 };
515 	boolean_t kill_spill = B_FALSE;
516 
517 	ASSERT(dmu_tx_is_syncing(tx));
518 	ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
519 	ASSERT(dnp->dn_type != DMU_OT_NONE ||
520 	    bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
521 	DNODE_VERIFY(dn);
522 
523 	ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
524 
525 	if (dmu_objset_userused_enabled(dn->dn_objset) &&
526 	    !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
527 		mutex_enter(&dn->dn_mtx);
528 		dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
529 		dn->dn_oldflags = dn->dn_phys->dn_flags;
530 		dn->dn_id_flags |= DN_ID_SYNC;
531 		dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
532 		mutex_exit(&dn->dn_mtx);
533 		dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
534 	} else {
535 		/* Once we account for it, we should always account for it. */
536 		ASSERT(!(dn->dn_phys->dn_flags &
537 		    DNODE_FLAG_USERUSED_ACCOUNTED));
538 	}
539 
540 	mutex_enter(&dn->dn_mtx);
541 	if (dn->dn_allocated_txg == tx->tx_txg) {
542 		/* The dnode is newly allocated or reallocated */
543 		if (dnp->dn_type == DMU_OT_NONE) {
544 			/* this is a first alloc, not a realloc */
545 			dnp->dn_nlevels = 1;
546 			dnp->dn_nblkptr = dn->dn_nblkptr;
547 		}
548 
549 		dnp->dn_type = dn->dn_type;
550 		dnp->dn_bonustype = dn->dn_bonustype;
551 		dnp->dn_bonuslen = dn->dn_bonuslen;
552 	}
553 
554 	ASSERT(dnp->dn_nlevels > 1 ||
555 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
556 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
557 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
558 
559 	if (dn->dn_next_blksz[txgoff]) {
560 		ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
561 		    SPA_MINBLOCKSIZE) == 0);
562 		ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
563 		    dn->dn_maxblkid == 0 || list_head(list) != NULL ||
564 		    avl_last(&dn->dn_ranges[txgoff]) ||
565 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
566 		    dnp->dn_datablkszsec);
567 		dnp->dn_datablkszsec =
568 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
569 		dn->dn_next_blksz[txgoff] = 0;
570 	}
571 
572 	if (dn->dn_next_bonuslen[txgoff]) {
573 		if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
574 			dnp->dn_bonuslen = 0;
575 		else
576 			dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
577 		ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
578 		dn->dn_next_bonuslen[txgoff] = 0;
579 	}
580 
581 	if (dn->dn_next_bonustype[txgoff]) {
582 		ASSERT(dn->dn_next_bonustype[txgoff] < DMU_OT_NUMTYPES);
583 		dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
584 		dn->dn_next_bonustype[txgoff] = 0;
585 	}
586 
587 	/*
588 	 * We will either remove a spill block when a file is being removed
589 	 * or we have been asked to remove it.
590 	 */
591 	if (dn->dn_rm_spillblk[txgoff] ||
592 	    ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) &&
593 	    dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg)) {
594 		if ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
595 			kill_spill = B_TRUE;
596 		dn->dn_rm_spillblk[txgoff] = 0;
597 	}
598 
599 	if (dn->dn_next_indblkshift[txgoff]) {
600 		ASSERT(dnp->dn_nlevels == 1);
601 		dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
602 		dn->dn_next_indblkshift[txgoff] = 0;
603 	}
604 
605 	/*
606 	 * Just take the live (open-context) values for checksum and compress.
607 	 * Strictly speaking it's a future leak, but nothing bad happens if we
608 	 * start using the new checksum or compress algorithm a little early.
609 	 */
610 	dnp->dn_checksum = dn->dn_checksum;
611 	dnp->dn_compress = dn->dn_compress;
612 
613 	mutex_exit(&dn->dn_mtx);
614 
615 	if (kill_spill) {
616 		(void) free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
617 		mutex_enter(&dn->dn_mtx);
618 		dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
619 		mutex_exit(&dn->dn_mtx);
620 	}
621 
622 	/* process all the "freed" ranges in the file */
623 	while (rp = avl_last(&dn->dn_ranges[txgoff])) {
624 		dnode_sync_free_range(dn, rp->fr_blkid, rp->fr_nblks, tx);
625 		/* grab the mutex so we don't race with dnode_block_freed() */
626 		mutex_enter(&dn->dn_mtx);
627 		avl_remove(&dn->dn_ranges[txgoff], rp);
628 		mutex_exit(&dn->dn_mtx);
629 		kmem_free(rp, sizeof (free_range_t));
630 	}
631 
632 	if (dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg) {
633 		dnode_sync_free(dn, tx);
634 		return;
635 	}
636 
637 	if (dn->dn_next_nblkptr[txgoff]) {
638 		/* this should only happen on a realloc */
639 		ASSERT(dn->dn_allocated_txg == tx->tx_txg);
640 		if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
641 			/* zero the new blkptrs we are gaining */
642 			bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
643 			    sizeof (blkptr_t) *
644 			    (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
645 #ifdef ZFS_DEBUG
646 		} else {
647 			int i;
648 			ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
649 			/* the blkptrs we are losing better be unallocated */
650 			for (i = dn->dn_next_nblkptr[txgoff];
651 			    i < dnp->dn_nblkptr; i++)
652 				ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
653 #endif
654 		}
655 		mutex_enter(&dn->dn_mtx);
656 		dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
657 		dn->dn_next_nblkptr[txgoff] = 0;
658 		mutex_exit(&dn->dn_mtx);
659 	}
660 
661 	if (dn->dn_next_nlevels[txgoff]) {
662 		dnode_increase_indirection(dn, tx);
663 		dn->dn_next_nlevels[txgoff] = 0;
664 	}
665 
666 	dbuf_sync_list(list, tx);
667 
668 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
669 		ASSERT3P(list_head(list), ==, NULL);
670 		dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
671 	}
672 
673 	/*
674 	 * Although we have dropped our reference to the dnode, it
675 	 * can't be evicted until its written, and we haven't yet
676 	 * initiated the IO for the dnode's dbuf.
677 	 */
678 }
679