xref: /freebsd/sys/contrib/openzfs/module/zfs/dnode_sync.c (revision 19fae0f66023a97a9b464b3beeeabb2081f575b3)
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 https://opensource.org/licenses/CDDL-1.0.
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, 2020 by Delphix. All rights reserved.
25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26  * Copyright 2020 Oxide Computer Company
27  */
28 
29 #include <sys/zfs_context.h>
30 #include <sys/dbuf.h>
31 #include <sys/dnode.h>
32 #include <sys/dmu.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dmu_recv.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/spa.h>
38 #include <sys/range_tree.h>
39 #include <sys/zfeature.h>
40 
41 static void
42 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
43 {
44 	dmu_buf_impl_t *db;
45 	int txgoff = tx->tx_txg & TXG_MASK;
46 	int nblkptr = dn->dn_phys->dn_nblkptr;
47 	int old_toplvl = dn->dn_phys->dn_nlevels - 1;
48 	int new_level = dn->dn_next_nlevels[txgoff];
49 	int i;
50 
51 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
52 
53 	/* this dnode can't be paged out because it's dirty */
54 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
55 	ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
56 
57 	db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
58 	ASSERT(db != NULL);
59 
60 	dn->dn_phys->dn_nlevels = new_level;
61 	dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
62 	    (u_longlong_t)dn->dn_object, dn->dn_phys->dn_nlevels);
63 
64 	/*
65 	 * Lock ordering requires that we hold the children's db_mutexes (by
66 	 * calling dbuf_find()) before holding the parent's db_rwlock.  The lock
67 	 * order is imposed by dbuf_read's steps of "grab the lock to protect
68 	 * db_parent, get db_parent, hold db_parent's db_rwlock".
69 	 */
70 	dmu_buf_impl_t *children[DN_MAX_NBLKPTR];
71 	ASSERT3U(nblkptr, <=, DN_MAX_NBLKPTR);
72 	for (i = 0; i < nblkptr; i++) {
73 		children[i] = dbuf_find(dn->dn_objset, dn->dn_object,
74 		    old_toplvl, i, NULL);
75 	}
76 
77 	/* transfer dnode's block pointers to new indirect block */
78 	(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
79 	if (dn->dn_dbuf != NULL)
80 		rw_enter(&dn->dn_dbuf->db_rwlock, RW_WRITER);
81 	rw_enter(&db->db_rwlock, RW_WRITER);
82 	ASSERT(db->db.db_data);
83 	ASSERT(arc_released(db->db_buf));
84 	ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
85 	memcpy(db->db.db_data, dn->dn_phys->dn_blkptr,
86 	    sizeof (blkptr_t) * nblkptr);
87 	arc_buf_freeze(db->db_buf);
88 
89 	/* set dbuf's parent pointers to new indirect buf */
90 	for (i = 0; i < nblkptr; i++) {
91 		dmu_buf_impl_t *child = children[i];
92 
93 		if (child == NULL)
94 			continue;
95 #ifdef	ZFS_DEBUG
96 		DB_DNODE_ENTER(child);
97 		ASSERT3P(DB_DNODE(child), ==, dn);
98 		DB_DNODE_EXIT(child);
99 #endif	/* DEBUG */
100 		if (child->db_parent && child->db_parent != dn->dn_dbuf) {
101 			ASSERT(child->db_parent->db_level == db->db_level);
102 			ASSERT(child->db_blkptr !=
103 			    &dn->dn_phys->dn_blkptr[child->db_blkid]);
104 			mutex_exit(&child->db_mtx);
105 			continue;
106 		}
107 		ASSERT(child->db_parent == NULL ||
108 		    child->db_parent == dn->dn_dbuf);
109 
110 		child->db_parent = db;
111 		dbuf_add_ref(db, child);
112 		if (db->db.db_data)
113 			child->db_blkptr = (blkptr_t *)db->db.db_data + i;
114 		else
115 			child->db_blkptr = NULL;
116 		dprintf_dbuf_bp(child, child->db_blkptr,
117 		    "changed db_blkptr to new indirect %s", "");
118 
119 		mutex_exit(&child->db_mtx);
120 	}
121 
122 	memset(dn->dn_phys->dn_blkptr, 0, sizeof (blkptr_t) * nblkptr);
123 
124 	rw_exit(&db->db_rwlock);
125 	if (dn->dn_dbuf != NULL)
126 		rw_exit(&dn->dn_dbuf->db_rwlock);
127 
128 	dbuf_rele(db, FTAG);
129 
130 	rw_exit(&dn->dn_struct_rwlock);
131 }
132 
133 static void
134 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
135 {
136 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
137 	uint64_t bytesfreed = 0;
138 
139 	dprintf("ds=%p obj=%llx num=%d\n", ds, (u_longlong_t)dn->dn_object,
140 	    num);
141 
142 	for (int i = 0; i < num; i++, bp++) {
143 		if (BP_IS_HOLE(bp))
144 			continue;
145 
146 		bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
147 		ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
148 
149 		/*
150 		 * Save some useful information on the holes being
151 		 * punched, including logical size, type, and indirection
152 		 * level. Retaining birth time enables detection of when
153 		 * holes are punched for reducing the number of free
154 		 * records transmitted during a zfs send.
155 		 */
156 
157 		uint64_t lsize = BP_GET_LSIZE(bp);
158 		dmu_object_type_t type = BP_GET_TYPE(bp);
159 		uint64_t lvl = BP_GET_LEVEL(bp);
160 
161 		memset(bp, 0, sizeof (blkptr_t));
162 
163 		if (spa_feature_is_active(dn->dn_objset->os_spa,
164 		    SPA_FEATURE_HOLE_BIRTH)) {
165 			BP_SET_LSIZE(bp, lsize);
166 			BP_SET_TYPE(bp, type);
167 			BP_SET_LEVEL(bp, lvl);
168 			BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
169 		}
170 	}
171 	dnode_diduse_space(dn, -bytesfreed);
172 }
173 
174 #ifdef ZFS_DEBUG
175 static void
176 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
177 {
178 	uint64_t off, num, i, j;
179 	unsigned int epbs;
180 	int err;
181 	uint64_t txg = tx->tx_txg;
182 	dnode_t *dn;
183 
184 	DB_DNODE_ENTER(db);
185 	dn = DB_DNODE(db);
186 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
187 	off = start - (db->db_blkid << epbs);
188 	num = end - start + 1;
189 
190 	ASSERT3U(dn->dn_phys->dn_indblkshift, >=, SPA_BLKPTRSHIFT);
191 	ASSERT3U(end + 1, >=, start);
192 	ASSERT3U(start, >=, (db->db_blkid << epbs));
193 	ASSERT3U(db->db_level, >, 0);
194 	ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
195 	ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
196 	ASSERT(db->db_blkptr != NULL);
197 
198 	for (i = off; i < off+num; i++) {
199 		uint64_t *buf;
200 		dmu_buf_impl_t *child;
201 		dbuf_dirty_record_t *dr;
202 
203 		ASSERT(db->db_level == 1);
204 
205 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
206 		err = dbuf_hold_impl(dn, db->db_level - 1,
207 		    (db->db_blkid << epbs) + i, TRUE, FALSE, FTAG, &child);
208 		rw_exit(&dn->dn_struct_rwlock);
209 		if (err == ENOENT)
210 			continue;
211 		ASSERT(err == 0);
212 		ASSERT(child->db_level == 0);
213 		dr = dbuf_find_dirty_eq(child, txg);
214 
215 		/* data_old better be zeroed */
216 		if (dr) {
217 			buf = dr->dt.dl.dr_data->b_data;
218 			for (j = 0; j < child->db.db_size >> 3; j++) {
219 				if (buf[j] != 0) {
220 					panic("freed data not zero: "
221 					    "child=%p i=%llu off=%llu "
222 					    "num=%llu\n",
223 					    (void *)child, (u_longlong_t)i,
224 					    (u_longlong_t)off,
225 					    (u_longlong_t)num);
226 				}
227 			}
228 		}
229 
230 		/*
231 		 * db_data better be zeroed unless it's dirty in a
232 		 * future txg.
233 		 */
234 		mutex_enter(&child->db_mtx);
235 		buf = child->db.db_data;
236 		if (buf != NULL && child->db_state != DB_FILL &&
237 		    list_is_empty(&child->db_dirty_records)) {
238 			for (j = 0; j < child->db.db_size >> 3; j++) {
239 				if (buf[j] != 0) {
240 					panic("freed data not zero: "
241 					    "child=%p i=%llu off=%llu "
242 					    "num=%llu\n",
243 					    (void *)child, (u_longlong_t)i,
244 					    (u_longlong_t)off,
245 					    (u_longlong_t)num);
246 				}
247 			}
248 		}
249 		mutex_exit(&child->db_mtx);
250 
251 		dbuf_rele(child, FTAG);
252 	}
253 	DB_DNODE_EXIT(db);
254 }
255 #endif
256 
257 /*
258  * We don't usually free the indirect blocks here.  If in one txg we have a
259  * free_range and a write to the same indirect block, it's important that we
260  * preserve the hole's birth times. Therefore, we don't free any any indirect
261  * blocks in free_children().  If an indirect block happens to turn into all
262  * holes, it will be freed by dbuf_write_children_ready, which happens at a
263  * point in the syncing process where we know for certain the contents of the
264  * indirect block.
265  *
266  * However, if we're freeing a dnode, its space accounting must go to zero
267  * before we actually try to free the dnode, or we will trip an assertion. In
268  * addition, we know the case described above cannot occur, because the dnode is
269  * being freed.  Therefore, we free the indirect blocks immediately in that
270  * case.
271  */
272 static void
273 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
274     boolean_t free_indirects, dmu_tx_t *tx)
275 {
276 	dnode_t *dn;
277 	blkptr_t *bp;
278 	dmu_buf_impl_t *subdb;
279 	uint64_t start, end, dbstart, dbend;
280 	unsigned int epbs, shift, i;
281 
282 	/*
283 	 * There is a small possibility that this block will not be cached:
284 	 *   1 - if level > 1 and there are no children with level <= 1
285 	 *   2 - if this block was evicted since we read it from
286 	 *	 dmu_tx_hold_free().
287 	 */
288 	if (db->db_state != DB_CACHED)
289 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
290 
291 	/*
292 	 * If we modify this indirect block, and we are not freeing the
293 	 * dnode (!free_indirects), then this indirect block needs to get
294 	 * written to disk by dbuf_write().  If it is dirty, we know it will
295 	 * be written (otherwise, we would have incorrect on-disk state
296 	 * because the space would be freed but still referenced by the BP
297 	 * in this indirect block).  Therefore we VERIFY that it is
298 	 * dirty.
299 	 *
300 	 * Our VERIFY covers some cases that do not actually have to be
301 	 * dirty, but the open-context code happens to dirty.  E.g. if the
302 	 * blocks we are freeing are all holes, because in that case, we
303 	 * are only freeing part of this indirect block, so it is an
304 	 * ancestor of the first or last block to be freed.  The first and
305 	 * last L1 indirect blocks are always dirtied by dnode_free_range().
306 	 */
307 	db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
308 	VERIFY(BP_GET_FILL(db->db_blkptr) == 0 || db->db_dirtycnt > 0);
309 	dmu_buf_unlock_parent(db, dblt, FTAG);
310 
311 	dbuf_release_bp(db);
312 	bp = db->db.db_data;
313 
314 	DB_DNODE_ENTER(db);
315 	dn = DB_DNODE(db);
316 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
317 	ASSERT3U(epbs, <, 31);
318 	shift = (db->db_level - 1) * epbs;
319 	dbstart = db->db_blkid << epbs;
320 	start = blkid >> shift;
321 	if (dbstart < start) {
322 		bp += start - dbstart;
323 	} else {
324 		start = dbstart;
325 	}
326 	dbend = ((db->db_blkid + 1) << epbs) - 1;
327 	end = (blkid + nblks - 1) >> shift;
328 	if (dbend <= end)
329 		end = dbend;
330 
331 	ASSERT3U(start, <=, end);
332 
333 	if (db->db_level == 1) {
334 		FREE_VERIFY(db, start, end, tx);
335 		rw_enter(&db->db_rwlock, RW_WRITER);
336 		free_blocks(dn, bp, end - start + 1, tx);
337 		rw_exit(&db->db_rwlock);
338 	} else {
339 		for (uint64_t id = start; id <= end; id++, bp++) {
340 			if (BP_IS_HOLE(bp))
341 				continue;
342 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
343 			VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
344 			    id, TRUE, FALSE, FTAG, &subdb));
345 			rw_exit(&dn->dn_struct_rwlock);
346 			ASSERT3P(bp, ==, subdb->db_blkptr);
347 
348 			free_children(subdb, blkid, nblks, free_indirects, tx);
349 			dbuf_rele(subdb, FTAG);
350 		}
351 	}
352 
353 	if (free_indirects) {
354 		rw_enter(&db->db_rwlock, RW_WRITER);
355 		for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++)
356 			ASSERT(BP_IS_HOLE(bp));
357 		memset(db->db.db_data, 0, db->db.db_size);
358 		free_blocks(dn, db->db_blkptr, 1, tx);
359 		rw_exit(&db->db_rwlock);
360 	}
361 
362 	DB_DNODE_EXIT(db);
363 	arc_buf_freeze(db->db_buf);
364 }
365 
366 /*
367  * Traverse the indicated range of the provided file
368  * and "free" all the blocks contained there.
369  */
370 static void
371 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
372     boolean_t free_indirects, dmu_tx_t *tx)
373 {
374 	blkptr_t *bp = dn->dn_phys->dn_blkptr;
375 	int dnlevel = dn->dn_phys->dn_nlevels;
376 	boolean_t trunc = B_FALSE;
377 
378 	if (blkid > dn->dn_phys->dn_maxblkid)
379 		return;
380 
381 	ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
382 	if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
383 		nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
384 		trunc = B_TRUE;
385 	}
386 
387 	/* There are no indirect blocks in the object */
388 	if (dnlevel == 1) {
389 		if (blkid >= dn->dn_phys->dn_nblkptr) {
390 			/* this range was never made persistent */
391 			return;
392 		}
393 		ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
394 		free_blocks(dn, bp + blkid, nblks, tx);
395 	} else {
396 		int shift = (dnlevel - 1) *
397 		    (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
398 		int start = blkid >> shift;
399 		int end = (blkid + nblks - 1) >> shift;
400 		dmu_buf_impl_t *db;
401 
402 		ASSERT(start < dn->dn_phys->dn_nblkptr);
403 		bp += start;
404 		for (int i = start; i <= end; i++, bp++) {
405 			if (BP_IS_HOLE(bp))
406 				continue;
407 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
408 			VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
409 			    TRUE, FALSE, FTAG, &db));
410 			rw_exit(&dn->dn_struct_rwlock);
411 			free_children(db, blkid, nblks, free_indirects, tx);
412 			dbuf_rele(db, FTAG);
413 		}
414 	}
415 
416 	/*
417 	 * Do not truncate the maxblkid if we are performing a raw
418 	 * receive. The raw receive sets the maxblkid manually and
419 	 * must not be overridden. Usually, the last DRR_FREE record
420 	 * will be at the maxblkid, because the source system sets
421 	 * the maxblkid when truncating. However, if the last block
422 	 * was freed by overwriting with zeros and being compressed
423 	 * away to a hole, the source system will generate a DRR_FREE
424 	 * record while leaving the maxblkid after the end of that
425 	 * record. In this case we need to leave the maxblkid as
426 	 * indicated in the DRR_OBJECT record, so that it matches the
427 	 * source system, ensuring that the cryptographic hashes will
428 	 * match.
429 	 */
430 	if (trunc && !dn->dn_objset->os_raw_receive) {
431 		uint64_t off __maybe_unused;
432 		dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
433 
434 		off = (dn->dn_phys->dn_maxblkid + 1) *
435 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
436 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
437 		    dn->dn_phys->dn_maxblkid == 0 ||
438 		    dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
439 	}
440 }
441 
442 typedef struct dnode_sync_free_range_arg {
443 	dnode_t *dsfra_dnode;
444 	dmu_tx_t *dsfra_tx;
445 	boolean_t dsfra_free_indirects;
446 } dnode_sync_free_range_arg_t;
447 
448 static void
449 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
450 {
451 	dnode_sync_free_range_arg_t *dsfra = arg;
452 	dnode_t *dn = dsfra->dsfra_dnode;
453 
454 	mutex_exit(&dn->dn_mtx);
455 	dnode_sync_free_range_impl(dn, blkid, nblks,
456 	    dsfra->dsfra_free_indirects, dsfra->dsfra_tx);
457 	mutex_enter(&dn->dn_mtx);
458 }
459 
460 /*
461  * Try to kick all the dnode's dbufs out of the cache...
462  */
463 void
464 dnode_evict_dbufs(dnode_t *dn)
465 {
466 	dmu_buf_impl_t *db_marker;
467 	dmu_buf_impl_t *db, *db_next;
468 
469 	db_marker = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
470 
471 	mutex_enter(&dn->dn_dbufs_mtx);
472 	for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
473 
474 #ifdef	ZFS_DEBUG
475 		DB_DNODE_ENTER(db);
476 		ASSERT3P(DB_DNODE(db), ==, dn);
477 		DB_DNODE_EXIT(db);
478 #endif	/* DEBUG */
479 
480 		mutex_enter(&db->db_mtx);
481 		if (db->db_state != DB_EVICTING &&
482 		    zfs_refcount_is_zero(&db->db_holds)) {
483 			db_marker->db_level = db->db_level;
484 			db_marker->db_blkid = db->db_blkid;
485 			db_marker->db_state = DB_SEARCH;
486 			avl_insert_here(&dn->dn_dbufs, db_marker, db,
487 			    AVL_BEFORE);
488 
489 			/*
490 			 * We need to use the "marker" dbuf rather than
491 			 * simply getting the next dbuf, because
492 			 * dbuf_destroy() may actually remove multiple dbufs.
493 			 * It can call itself recursively on the parent dbuf,
494 			 * which may also be removed from dn_dbufs.  The code
495 			 * flow would look like:
496 			 *
497 			 * dbuf_destroy():
498 			 *   dnode_rele_and_unlock(parent_dbuf, evicting=TRUE):
499 			 *	if (!cacheable || pending_evict)
500 			 *	  dbuf_destroy()
501 			 */
502 			dbuf_destroy(db);
503 
504 			db_next = AVL_NEXT(&dn->dn_dbufs, db_marker);
505 			avl_remove(&dn->dn_dbufs, db_marker);
506 		} else {
507 			db->db_pending_evict = TRUE;
508 			mutex_exit(&db->db_mtx);
509 			db_next = AVL_NEXT(&dn->dn_dbufs, db);
510 		}
511 	}
512 	mutex_exit(&dn->dn_dbufs_mtx);
513 
514 	kmem_free(db_marker, sizeof (dmu_buf_impl_t));
515 
516 	dnode_evict_bonus(dn);
517 }
518 
519 void
520 dnode_evict_bonus(dnode_t *dn)
521 {
522 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
523 	if (dn->dn_bonus != NULL) {
524 		if (zfs_refcount_is_zero(&dn->dn_bonus->db_holds)) {
525 			mutex_enter(&dn->dn_bonus->db_mtx);
526 			dbuf_destroy(dn->dn_bonus);
527 			dn->dn_bonus = NULL;
528 		} else {
529 			dn->dn_bonus->db_pending_evict = TRUE;
530 		}
531 	}
532 	rw_exit(&dn->dn_struct_rwlock);
533 }
534 
535 static void
536 dnode_undirty_dbufs(list_t *list)
537 {
538 	dbuf_dirty_record_t *dr;
539 
540 	while ((dr = list_head(list))) {
541 		dmu_buf_impl_t *db = dr->dr_dbuf;
542 		uint64_t txg = dr->dr_txg;
543 
544 		if (db->db_level != 0)
545 			dnode_undirty_dbufs(&dr->dt.di.dr_children);
546 
547 		mutex_enter(&db->db_mtx);
548 		/* XXX - use dbuf_undirty()? */
549 		list_remove(list, dr);
550 		ASSERT(list_head(&db->db_dirty_records) == dr);
551 		list_remove_head(&db->db_dirty_records);
552 		ASSERT(list_is_empty(&db->db_dirty_records));
553 		db->db_dirtycnt -= 1;
554 		if (db->db_level == 0) {
555 			ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
556 			    dr->dt.dl.dr_data == db->db_buf);
557 			dbuf_unoverride(dr);
558 		} else {
559 			mutex_destroy(&dr->dt.di.dr_mtx);
560 			list_destroy(&dr->dt.di.dr_children);
561 		}
562 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
563 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
564 	}
565 }
566 
567 static void
568 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
569 {
570 	int txgoff = tx->tx_txg & TXG_MASK;
571 
572 	ASSERT(dmu_tx_is_syncing(tx));
573 
574 	/*
575 	 * Our contents should have been freed in dnode_sync() by the
576 	 * free range record inserted by the caller of dnode_free().
577 	 */
578 	ASSERT0(DN_USED_BYTES(dn->dn_phys));
579 	ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
580 
581 	dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
582 	dnode_evict_dbufs(dn);
583 
584 	/*
585 	 * XXX - It would be nice to assert this, but we may still
586 	 * have residual holds from async evictions from the arc...
587 	 *
588 	 * zfs_obj_to_path() also depends on this being
589 	 * commented out.
590 	 *
591 	 * ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 1);
592 	 */
593 
594 	/* Undirty next bits */
595 	dn->dn_next_nlevels[txgoff] = 0;
596 	dn->dn_next_indblkshift[txgoff] = 0;
597 	dn->dn_next_blksz[txgoff] = 0;
598 	dn->dn_next_maxblkid[txgoff] = 0;
599 
600 	/* ASSERT(blkptrs are zero); */
601 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
602 	ASSERT(dn->dn_type != DMU_OT_NONE);
603 
604 	ASSERT(dn->dn_free_txg > 0);
605 	if (dn->dn_allocated_txg != dn->dn_free_txg)
606 		dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
607 	memset(dn->dn_phys, 0, sizeof (dnode_phys_t) * dn->dn_num_slots);
608 	dnode_free_interior_slots(dn);
609 
610 	mutex_enter(&dn->dn_mtx);
611 	dn->dn_type = DMU_OT_NONE;
612 	dn->dn_maxblkid = 0;
613 	dn->dn_allocated_txg = 0;
614 	dn->dn_free_txg = 0;
615 	dn->dn_have_spill = B_FALSE;
616 	dn->dn_num_slots = 1;
617 	mutex_exit(&dn->dn_mtx);
618 
619 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
620 
621 	dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
622 	/*
623 	 * Now that we've released our hold, the dnode may
624 	 * be evicted, so we mustn't access it.
625 	 */
626 }
627 
628 /*
629  * Write out the dnode's dirty buffers.
630  */
631 void
632 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
633 {
634 	objset_t *os = dn->dn_objset;
635 	dnode_phys_t *dnp = dn->dn_phys;
636 	int txgoff = tx->tx_txg & TXG_MASK;
637 	list_t *list = &dn->dn_dirty_records[txgoff];
638 	static const dnode_phys_t zerodn __maybe_unused = { 0 };
639 	boolean_t kill_spill = B_FALSE;
640 
641 	ASSERT(dmu_tx_is_syncing(tx));
642 	ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
643 	ASSERT(dnp->dn_type != DMU_OT_NONE ||
644 	    memcmp(dnp, &zerodn, DNODE_MIN_SIZE) == 0);
645 	DNODE_VERIFY(dn);
646 
647 	ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
648 
649 	/*
650 	 * Do user accounting if it is enabled and this is not
651 	 * an encrypted receive.
652 	 */
653 	if (dmu_objset_userused_enabled(os) &&
654 	    !DMU_OBJECT_IS_SPECIAL(dn->dn_object) &&
655 	    (!os->os_encrypted || !dmu_objset_is_receiving(os))) {
656 		mutex_enter(&dn->dn_mtx);
657 		dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
658 		dn->dn_oldflags = dn->dn_phys->dn_flags;
659 		dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
660 		if (dmu_objset_userobjused_enabled(dn->dn_objset))
661 			dn->dn_phys->dn_flags |=
662 			    DNODE_FLAG_USEROBJUSED_ACCOUNTED;
663 		mutex_exit(&dn->dn_mtx);
664 		dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
665 	} else if (!(os->os_encrypted && dmu_objset_is_receiving(os))) {
666 		/*
667 		 * Once we account for it, we should always account for it,
668 		 * except for the case of a raw receive. We will not be able
669 		 * to account for it until the receiving dataset has been
670 		 * mounted.
671 		 */
672 		ASSERT(!(dn->dn_phys->dn_flags &
673 		    DNODE_FLAG_USERUSED_ACCOUNTED));
674 		ASSERT(!(dn->dn_phys->dn_flags &
675 		    DNODE_FLAG_USEROBJUSED_ACCOUNTED));
676 	}
677 
678 	mutex_enter(&dn->dn_mtx);
679 	if (dn->dn_allocated_txg == tx->tx_txg) {
680 		/* The dnode is newly allocated or reallocated */
681 		if (dnp->dn_type == DMU_OT_NONE) {
682 			/* this is a first alloc, not a realloc */
683 			dnp->dn_nlevels = 1;
684 			dnp->dn_nblkptr = dn->dn_nblkptr;
685 		}
686 
687 		dnp->dn_type = dn->dn_type;
688 		dnp->dn_bonustype = dn->dn_bonustype;
689 		dnp->dn_bonuslen = dn->dn_bonuslen;
690 	}
691 
692 	dnp->dn_extra_slots = dn->dn_num_slots - 1;
693 
694 	ASSERT(dnp->dn_nlevels > 1 ||
695 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
696 	    BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
697 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
698 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
699 	ASSERT(dnp->dn_nlevels < 2 ||
700 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
701 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
702 
703 	if (dn->dn_next_type[txgoff] != 0) {
704 		dnp->dn_type = dn->dn_type;
705 		dn->dn_next_type[txgoff] = 0;
706 	}
707 
708 	if (dn->dn_next_blksz[txgoff] != 0) {
709 		ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
710 		    SPA_MINBLOCKSIZE) == 0);
711 		ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
712 		    dn->dn_maxblkid == 0 || list_head(list) != NULL ||
713 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
714 		    dnp->dn_datablkszsec ||
715 		    !range_tree_is_empty(dn->dn_free_ranges[txgoff]));
716 		dnp->dn_datablkszsec =
717 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
718 		dn->dn_next_blksz[txgoff] = 0;
719 	}
720 
721 	if (dn->dn_next_bonuslen[txgoff] != 0) {
722 		if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
723 			dnp->dn_bonuslen = 0;
724 		else
725 			dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
726 		ASSERT(dnp->dn_bonuslen <=
727 		    DN_SLOTS_TO_BONUSLEN(dnp->dn_extra_slots + 1));
728 		dn->dn_next_bonuslen[txgoff] = 0;
729 	}
730 
731 	if (dn->dn_next_bonustype[txgoff] != 0) {
732 		ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
733 		dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
734 		dn->dn_next_bonustype[txgoff] = 0;
735 	}
736 
737 	boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
738 	    dn->dn_free_txg <= tx->tx_txg;
739 
740 	/*
741 	 * Remove the spill block if we have been explicitly asked to
742 	 * remove it, or if the object is being removed.
743 	 */
744 	if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
745 		if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
746 			kill_spill = B_TRUE;
747 		dn->dn_rm_spillblk[txgoff] = 0;
748 	}
749 
750 	if (dn->dn_next_indblkshift[txgoff] != 0) {
751 		ASSERT(dnp->dn_nlevels == 1);
752 		dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
753 		dn->dn_next_indblkshift[txgoff] = 0;
754 	}
755 
756 	/*
757 	 * Just take the live (open-context) values for checksum and compress.
758 	 * Strictly speaking it's a future leak, but nothing bad happens if we
759 	 * start using the new checksum or compress algorithm a little early.
760 	 */
761 	dnp->dn_checksum = dn->dn_checksum;
762 	dnp->dn_compress = dn->dn_compress;
763 
764 	mutex_exit(&dn->dn_mtx);
765 
766 	if (kill_spill) {
767 		free_blocks(dn, DN_SPILL_BLKPTR(dn->dn_phys), 1, tx);
768 		mutex_enter(&dn->dn_mtx);
769 		dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
770 		mutex_exit(&dn->dn_mtx);
771 	}
772 
773 	/* process all the "freed" ranges in the file */
774 	if (dn->dn_free_ranges[txgoff] != NULL) {
775 		dnode_sync_free_range_arg_t dsfra;
776 		dsfra.dsfra_dnode = dn;
777 		dsfra.dsfra_tx = tx;
778 		dsfra.dsfra_free_indirects = freeing_dnode;
779 		mutex_enter(&dn->dn_mtx);
780 		if (freeing_dnode) {
781 			ASSERT(range_tree_contains(dn->dn_free_ranges[txgoff],
782 			    0, dn->dn_maxblkid + 1));
783 		}
784 		/*
785 		 * Because dnode_sync_free_range() must drop dn_mtx during its
786 		 * processing, using it as a callback to range_tree_vacate() is
787 		 * not safe.  No other operations (besides destroy) are allowed
788 		 * once range_tree_vacate() has begun, and dropping dn_mtx
789 		 * would leave a window open for another thread to observe that
790 		 * invalid (and unsafe) state.
791 		 */
792 		range_tree_walk(dn->dn_free_ranges[txgoff],
793 		    dnode_sync_free_range, &dsfra);
794 		range_tree_vacate(dn->dn_free_ranges[txgoff], NULL, NULL);
795 		range_tree_destroy(dn->dn_free_ranges[txgoff]);
796 		dn->dn_free_ranges[txgoff] = NULL;
797 		mutex_exit(&dn->dn_mtx);
798 	}
799 
800 	if (freeing_dnode) {
801 		dn->dn_objset->os_freed_dnodes++;
802 		dnode_sync_free(dn, tx);
803 		return;
804 	}
805 
806 	if (dn->dn_num_slots > DNODE_MIN_SLOTS) {
807 		dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
808 		mutex_enter(&ds->ds_lock);
809 		ds->ds_feature_activation[SPA_FEATURE_LARGE_DNODE] =
810 		    (void *)B_TRUE;
811 		mutex_exit(&ds->ds_lock);
812 	}
813 
814 	if (dn->dn_next_nlevels[txgoff]) {
815 		dnode_increase_indirection(dn, tx);
816 		dn->dn_next_nlevels[txgoff] = 0;
817 	}
818 
819 	/*
820 	 * This must be done after dnode_sync_free_range()
821 	 * and dnode_increase_indirection(). See dnode_new_blkid()
822 	 * for an explanation of the high bit being set.
823 	 */
824 	if (dn->dn_next_maxblkid[txgoff]) {
825 		mutex_enter(&dn->dn_mtx);
826 		dnp->dn_maxblkid =
827 		    dn->dn_next_maxblkid[txgoff] & ~DMU_NEXT_MAXBLKID_SET;
828 		dn->dn_next_maxblkid[txgoff] = 0;
829 		mutex_exit(&dn->dn_mtx);
830 	}
831 
832 	if (dn->dn_next_nblkptr[txgoff]) {
833 		/* this should only happen on a realloc */
834 		ASSERT(dn->dn_allocated_txg == tx->tx_txg);
835 		if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
836 			/* zero the new blkptrs we are gaining */
837 			memset(dnp->dn_blkptr + dnp->dn_nblkptr, 0,
838 			    sizeof (blkptr_t) *
839 			    (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
840 #ifdef ZFS_DEBUG
841 		} else {
842 			int i;
843 			ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
844 			/* the blkptrs we are losing better be unallocated */
845 			for (i = 0; i < dnp->dn_nblkptr; i++) {
846 				if (i >= dn->dn_next_nblkptr[txgoff])
847 					ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
848 			}
849 #endif
850 		}
851 		mutex_enter(&dn->dn_mtx);
852 		dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
853 		dn->dn_next_nblkptr[txgoff] = 0;
854 		mutex_exit(&dn->dn_mtx);
855 	}
856 
857 	dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
858 
859 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
860 		ASSERT3P(list_head(list), ==, NULL);
861 		dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
862 	}
863 
864 	ASSERT3U(dnp->dn_bonuslen, <=, DN_MAX_BONUS_LEN(dnp));
865 
866 	/*
867 	 * Although we have dropped our reference to the dnode, it
868 	 * can't be evicted until its written, and we haven't yet
869 	 * initiated the IO for the dnode's dbuf.  Additionally, the caller
870 	 * has already added a reference to the dnode because it's on the
871 	 * os_synced_dnodes list.
872 	 */
873 }
874