xref: /freebsd/sys/contrib/openzfs/module/zfs/dnode_sync.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
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, 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 	    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] =
74 		    dbuf_find(dn->dn_objset, dn->dn_object, old_toplvl, i);
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 	bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
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 	bzero(dn->dn_phys->dn_blkptr, 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, dn->dn_object, num);
140 
141 	for (int i = 0; i < num; i++, bp++) {
142 		if (BP_IS_HOLE(bp))
143 			continue;
144 
145 		bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
146 		ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
147 
148 		/*
149 		 * Save some useful information on the holes being
150 		 * punched, including logical size, type, and indirection
151 		 * level. Retaining birth time enables detection of when
152 		 * holes are punched for reducing the number of free
153 		 * records transmitted during a zfs send.
154 		 */
155 
156 		uint64_t lsize = BP_GET_LSIZE(bp);
157 		dmu_object_type_t type = BP_GET_TYPE(bp);
158 		uint64_t lvl = BP_GET_LEVEL(bp);
159 
160 		bzero(bp, sizeof (blkptr_t));
161 
162 		if (spa_feature_is_active(dn->dn_objset->os_spa,
163 		    SPA_FEATURE_HOLE_BIRTH)) {
164 			BP_SET_LSIZE(bp, lsize);
165 			BP_SET_TYPE(bp, type);
166 			BP_SET_LEVEL(bp, lvl);
167 			BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
168 		}
169 	}
170 	dnode_diduse_space(dn, -bytesfreed);
171 }
172 
173 #ifdef ZFS_DEBUG
174 static void
175 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
176 {
177 	int off, num;
178 	int i, err, epbs;
179 	uint64_t txg = tx->tx_txg;
180 	dnode_t *dn;
181 
182 	DB_DNODE_ENTER(db);
183 	dn = DB_DNODE(db);
184 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
185 	off = start - (db->db_blkid * 1<<epbs);
186 	num = end - start + 1;
187 
188 	ASSERT3U(off, >=, 0);
189 	ASSERT3U(num, >=, 0);
190 	ASSERT3U(db->db_level, >, 0);
191 	ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
192 	ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
193 	ASSERT(db->db_blkptr != NULL);
194 
195 	for (i = off; i < off+num; i++) {
196 		uint64_t *buf;
197 		dmu_buf_impl_t *child;
198 		dbuf_dirty_record_t *dr;
199 		int j;
200 
201 		ASSERT(db->db_level == 1);
202 
203 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
204 		err = dbuf_hold_impl(dn, db->db_level - 1,
205 		    (db->db_blkid << epbs) + i, TRUE, FALSE, FTAG, &child);
206 		rw_exit(&dn->dn_struct_rwlock);
207 		if (err == ENOENT)
208 			continue;
209 		ASSERT(err == 0);
210 		ASSERT(child->db_level == 0);
211 		dr = dbuf_find_dirty_eq(child, txg);
212 
213 		/* data_old better be zeroed */
214 		if (dr) {
215 			buf = dr->dt.dl.dr_data->b_data;
216 			for (j = 0; j < child->db.db_size >> 3; j++) {
217 				if (buf[j] != 0) {
218 					panic("freed data not zero: "
219 					    "child=%p i=%d off=%d num=%d\n",
220 					    (void *)child, i, off, num);
221 				}
222 			}
223 		}
224 
225 		/*
226 		 * db_data better be zeroed unless it's dirty in a
227 		 * future txg.
228 		 */
229 		mutex_enter(&child->db_mtx);
230 		buf = child->db.db_data;
231 		if (buf != NULL && child->db_state != DB_FILL &&
232 		    list_is_empty(&child->db_dirty_records)) {
233 			for (j = 0; j < child->db.db_size >> 3; j++) {
234 				if (buf[j] != 0) {
235 					panic("freed data not zero: "
236 					    "child=%p i=%d off=%d num=%d\n",
237 					    (void *)child, i, off, num);
238 				}
239 			}
240 		}
241 		mutex_exit(&child->db_mtx);
242 
243 		dbuf_rele(child, FTAG);
244 	}
245 	DB_DNODE_EXIT(db);
246 }
247 #endif
248 
249 /*
250  * We don't usually free the indirect blocks here.  If in one txg we have a
251  * free_range and a write to the same indirect block, it's important that we
252  * preserve the hole's birth times. Therefore, we don't free any any indirect
253  * blocks in free_children().  If an indirect block happens to turn into all
254  * holes, it will be freed by dbuf_write_children_ready, which happens at a
255  * point in the syncing process where we know for certain the contents of the
256  * indirect block.
257  *
258  * However, if we're freeing a dnode, its space accounting must go to zero
259  * before we actually try to free the dnode, or we will trip an assertion. In
260  * addition, we know the case described above cannot occur, because the dnode is
261  * being freed.  Therefore, we free the indirect blocks immediately in that
262  * case.
263  */
264 static void
265 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
266     boolean_t free_indirects, dmu_tx_t *tx)
267 {
268 	dnode_t *dn;
269 	blkptr_t *bp;
270 	dmu_buf_impl_t *subdb;
271 	uint64_t start, end, dbstart, dbend;
272 	unsigned int epbs, shift, i;
273 
274 	/*
275 	 * There is a small possibility that this block will not be cached:
276 	 *   1 - if level > 1 and there are no children with level <= 1
277 	 *   2 - if this block was evicted since we read it from
278 	 *	 dmu_tx_hold_free().
279 	 */
280 	if (db->db_state != DB_CACHED)
281 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
282 
283 	/*
284 	 * If we modify this indirect block, and we are not freeing the
285 	 * dnode (!free_indirects), then this indirect block needs to get
286 	 * written to disk by dbuf_write().  If it is dirty, we know it will
287 	 * be written (otherwise, we would have incorrect on-disk state
288 	 * because the space would be freed but still referenced by the BP
289 	 * in this indirect block).  Therefore we VERIFY that it is
290 	 * dirty.
291 	 *
292 	 * Our VERIFY covers some cases that do not actually have to be
293 	 * dirty, but the open-context code happens to dirty.  E.g. if the
294 	 * blocks we are freeing are all holes, because in that case, we
295 	 * are only freeing part of this indirect block, so it is an
296 	 * ancestor of the first or last block to be freed.  The first and
297 	 * last L1 indirect blocks are always dirtied by dnode_free_range().
298 	 */
299 	db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
300 	VERIFY(BP_GET_FILL(db->db_blkptr) == 0 || db->db_dirtycnt > 0);
301 	dmu_buf_unlock_parent(db, dblt, FTAG);
302 
303 	dbuf_release_bp(db);
304 	bp = db->db.db_data;
305 
306 	DB_DNODE_ENTER(db);
307 	dn = DB_DNODE(db);
308 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
309 	ASSERT3U(epbs, <, 31);
310 	shift = (db->db_level - 1) * epbs;
311 	dbstart = db->db_blkid << epbs;
312 	start = blkid >> shift;
313 	if (dbstart < start) {
314 		bp += start - dbstart;
315 	} else {
316 		start = dbstart;
317 	}
318 	dbend = ((db->db_blkid + 1) << epbs) - 1;
319 	end = (blkid + nblks - 1) >> shift;
320 	if (dbend <= end)
321 		end = dbend;
322 
323 	ASSERT3U(start, <=, end);
324 
325 	if (db->db_level == 1) {
326 		FREE_VERIFY(db, start, end, tx);
327 		rw_enter(&db->db_rwlock, RW_WRITER);
328 		free_blocks(dn, bp, end - start + 1, tx);
329 		rw_exit(&db->db_rwlock);
330 	} else {
331 		for (uint64_t id = start; id <= end; id++, bp++) {
332 			if (BP_IS_HOLE(bp))
333 				continue;
334 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
335 			VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
336 			    id, TRUE, FALSE, FTAG, &subdb));
337 			rw_exit(&dn->dn_struct_rwlock);
338 			ASSERT3P(bp, ==, subdb->db_blkptr);
339 
340 			free_children(subdb, blkid, nblks, free_indirects, tx);
341 			dbuf_rele(subdb, FTAG);
342 		}
343 	}
344 
345 	if (free_indirects) {
346 		rw_enter(&db->db_rwlock, RW_WRITER);
347 		for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++)
348 			ASSERT(BP_IS_HOLE(bp));
349 		bzero(db->db.db_data, db->db.db_size);
350 		free_blocks(dn, db->db_blkptr, 1, tx);
351 		rw_exit(&db->db_rwlock);
352 	}
353 
354 	DB_DNODE_EXIT(db);
355 	arc_buf_freeze(db->db_buf);
356 }
357 
358 /*
359  * Traverse the indicated range of the provided file
360  * and "free" all the blocks contained there.
361  */
362 static void
363 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
364     boolean_t free_indirects, dmu_tx_t *tx)
365 {
366 	blkptr_t *bp = dn->dn_phys->dn_blkptr;
367 	int dnlevel = dn->dn_phys->dn_nlevels;
368 	boolean_t trunc = B_FALSE;
369 
370 	if (blkid > dn->dn_phys->dn_maxblkid)
371 		return;
372 
373 	ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
374 	if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
375 		nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
376 		trunc = B_TRUE;
377 	}
378 
379 	/* There are no indirect blocks in the object */
380 	if (dnlevel == 1) {
381 		if (blkid >= dn->dn_phys->dn_nblkptr) {
382 			/* this range was never made persistent */
383 			return;
384 		}
385 		ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
386 		free_blocks(dn, bp + blkid, nblks, tx);
387 	} else {
388 		int shift = (dnlevel - 1) *
389 		    (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
390 		int start = blkid >> shift;
391 		int end = (blkid + nblks - 1) >> shift;
392 		dmu_buf_impl_t *db;
393 
394 		ASSERT(start < dn->dn_phys->dn_nblkptr);
395 		bp += start;
396 		for (int i = start; i <= end; i++, bp++) {
397 			if (BP_IS_HOLE(bp))
398 				continue;
399 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
400 			VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
401 			    TRUE, FALSE, FTAG, &db));
402 			rw_exit(&dn->dn_struct_rwlock);
403 			free_children(db, blkid, nblks, free_indirects, tx);
404 			dbuf_rele(db, FTAG);
405 		}
406 	}
407 
408 	/*
409 	 * Do not truncate the maxblkid if we are performing a raw
410 	 * receive. The raw receive sets the maxblkid manually and
411 	 * must not be overridden. Usually, the last DRR_FREE record
412 	 * will be at the maxblkid, because the source system sets
413 	 * the maxblkid when truncating. However, if the last block
414 	 * was freed by overwriting with zeros and being compressed
415 	 * away to a hole, the source system will generate a DRR_FREE
416 	 * record while leaving the maxblkid after the end of that
417 	 * record. In this case we need to leave the maxblkid as
418 	 * indicated in the DRR_OBJECT record, so that it matches the
419 	 * source system, ensuring that the cryptographic hashes will
420 	 * match.
421 	 */
422 	if (trunc && !dn->dn_objset->os_raw_receive) {
423 		uint64_t off __maybe_unused;
424 		dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
425 
426 		off = (dn->dn_phys->dn_maxblkid + 1) *
427 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
428 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
429 		    dn->dn_phys->dn_maxblkid == 0 ||
430 		    dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
431 	}
432 }
433 
434 typedef struct dnode_sync_free_range_arg {
435 	dnode_t *dsfra_dnode;
436 	dmu_tx_t *dsfra_tx;
437 	boolean_t dsfra_free_indirects;
438 } dnode_sync_free_range_arg_t;
439 
440 static void
441 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
442 {
443 	dnode_sync_free_range_arg_t *dsfra = arg;
444 	dnode_t *dn = dsfra->dsfra_dnode;
445 
446 	mutex_exit(&dn->dn_mtx);
447 	dnode_sync_free_range_impl(dn, blkid, nblks,
448 	    dsfra->dsfra_free_indirects, dsfra->dsfra_tx);
449 	mutex_enter(&dn->dn_mtx);
450 }
451 
452 /*
453  * Try to kick all the dnode's dbufs out of the cache...
454  */
455 void
456 dnode_evict_dbufs(dnode_t *dn)
457 {
458 	dmu_buf_impl_t *db_marker;
459 	dmu_buf_impl_t *db, *db_next;
460 
461 	db_marker = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
462 
463 	mutex_enter(&dn->dn_dbufs_mtx);
464 	for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
465 
466 #ifdef	ZFS_DEBUG
467 		DB_DNODE_ENTER(db);
468 		ASSERT3P(DB_DNODE(db), ==, dn);
469 		DB_DNODE_EXIT(db);
470 #endif	/* DEBUG */
471 
472 		mutex_enter(&db->db_mtx);
473 		if (db->db_state != DB_EVICTING &&
474 		    zfs_refcount_is_zero(&db->db_holds)) {
475 			db_marker->db_level = db->db_level;
476 			db_marker->db_blkid = db->db_blkid;
477 			db_marker->db_state = DB_SEARCH;
478 			avl_insert_here(&dn->dn_dbufs, db_marker, db,
479 			    AVL_BEFORE);
480 
481 			/*
482 			 * We need to use the "marker" dbuf rather than
483 			 * simply getting the next dbuf, because
484 			 * dbuf_destroy() may actually remove multiple dbufs.
485 			 * It can call itself recursively on the parent dbuf,
486 			 * which may also be removed from dn_dbufs.  The code
487 			 * flow would look like:
488 			 *
489 			 * dbuf_destroy():
490 			 *   dnode_rele_and_unlock(parent_dbuf, evicting=TRUE):
491 			 *	if (!cacheable || pending_evict)
492 			 *	  dbuf_destroy()
493 			 */
494 			dbuf_destroy(db);
495 
496 			db_next = AVL_NEXT(&dn->dn_dbufs, db_marker);
497 			avl_remove(&dn->dn_dbufs, db_marker);
498 		} else {
499 			db->db_pending_evict = TRUE;
500 			mutex_exit(&db->db_mtx);
501 			db_next = AVL_NEXT(&dn->dn_dbufs, db);
502 		}
503 	}
504 	mutex_exit(&dn->dn_dbufs_mtx);
505 
506 	kmem_free(db_marker, sizeof (dmu_buf_impl_t));
507 
508 	dnode_evict_bonus(dn);
509 }
510 
511 void
512 dnode_evict_bonus(dnode_t *dn)
513 {
514 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
515 	if (dn->dn_bonus != NULL) {
516 		if (zfs_refcount_is_zero(&dn->dn_bonus->db_holds)) {
517 			mutex_enter(&dn->dn_bonus->db_mtx);
518 			dbuf_destroy(dn->dn_bonus);
519 			dn->dn_bonus = NULL;
520 		} else {
521 			dn->dn_bonus->db_pending_evict = TRUE;
522 		}
523 	}
524 	rw_exit(&dn->dn_struct_rwlock);
525 }
526 
527 static void
528 dnode_undirty_dbufs(list_t *list)
529 {
530 	dbuf_dirty_record_t *dr;
531 
532 	while ((dr = list_head(list))) {
533 		dmu_buf_impl_t *db = dr->dr_dbuf;
534 		uint64_t txg = dr->dr_txg;
535 
536 		if (db->db_level != 0)
537 			dnode_undirty_dbufs(&dr->dt.di.dr_children);
538 
539 		mutex_enter(&db->db_mtx);
540 		/* XXX - use dbuf_undirty()? */
541 		list_remove(list, dr);
542 		ASSERT(list_head(&db->db_dirty_records) == dr);
543 		list_remove_head(&db->db_dirty_records);
544 		ASSERT(list_is_empty(&db->db_dirty_records));
545 		db->db_dirtycnt -= 1;
546 		if (db->db_level == 0) {
547 			ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
548 			    dr->dt.dl.dr_data == db->db_buf);
549 			dbuf_unoverride(dr);
550 		} else {
551 			mutex_destroy(&dr->dt.di.dr_mtx);
552 			list_destroy(&dr->dt.di.dr_children);
553 		}
554 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
555 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
556 	}
557 }
558 
559 static void
560 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
561 {
562 	int txgoff = tx->tx_txg & TXG_MASK;
563 
564 	ASSERT(dmu_tx_is_syncing(tx));
565 
566 	/*
567 	 * Our contents should have been freed in dnode_sync() by the
568 	 * free range record inserted by the caller of dnode_free().
569 	 */
570 	ASSERT0(DN_USED_BYTES(dn->dn_phys));
571 	ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
572 
573 	dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
574 	dnode_evict_dbufs(dn);
575 
576 	/*
577 	 * XXX - It would be nice to assert this, but we may still
578 	 * have residual holds from async evictions from the arc...
579 	 *
580 	 * zfs_obj_to_path() also depends on this being
581 	 * commented out.
582 	 *
583 	 * ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 1);
584 	 */
585 
586 	/* Undirty next bits */
587 	dn->dn_next_nlevels[txgoff] = 0;
588 	dn->dn_next_indblkshift[txgoff] = 0;
589 	dn->dn_next_blksz[txgoff] = 0;
590 	dn->dn_next_maxblkid[txgoff] = 0;
591 
592 	/* ASSERT(blkptrs are zero); */
593 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
594 	ASSERT(dn->dn_type != DMU_OT_NONE);
595 
596 	ASSERT(dn->dn_free_txg > 0);
597 	if (dn->dn_allocated_txg != dn->dn_free_txg)
598 		dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
599 	bzero(dn->dn_phys, sizeof (dnode_phys_t) * dn->dn_num_slots);
600 	dnode_free_interior_slots(dn);
601 
602 	mutex_enter(&dn->dn_mtx);
603 	dn->dn_type = DMU_OT_NONE;
604 	dn->dn_maxblkid = 0;
605 	dn->dn_allocated_txg = 0;
606 	dn->dn_free_txg = 0;
607 	dn->dn_have_spill = B_FALSE;
608 	dn->dn_num_slots = 1;
609 	mutex_exit(&dn->dn_mtx);
610 
611 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
612 
613 	dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
614 	/*
615 	 * Now that we've released our hold, the dnode may
616 	 * be evicted, so we mustn't access it.
617 	 */
618 }
619 
620 /*
621  * Write out the dnode's dirty buffers.
622  */
623 void
624 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
625 {
626 	objset_t *os = dn->dn_objset;
627 	dnode_phys_t *dnp = dn->dn_phys;
628 	int txgoff = tx->tx_txg & TXG_MASK;
629 	list_t *list = &dn->dn_dirty_records[txgoff];
630 	static const dnode_phys_t zerodn __maybe_unused = { 0 };
631 	boolean_t kill_spill = B_FALSE;
632 
633 	ASSERT(dmu_tx_is_syncing(tx));
634 	ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
635 	ASSERT(dnp->dn_type != DMU_OT_NONE ||
636 	    bcmp(dnp, &zerodn, DNODE_MIN_SIZE) == 0);
637 	DNODE_VERIFY(dn);
638 
639 	ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
640 
641 	/*
642 	 * Do user accounting if it is enabled and this is not
643 	 * an encrypted receive.
644 	 */
645 	if (dmu_objset_userused_enabled(os) &&
646 	    !DMU_OBJECT_IS_SPECIAL(dn->dn_object) &&
647 	    (!os->os_encrypted || !dmu_objset_is_receiving(os))) {
648 		mutex_enter(&dn->dn_mtx);
649 		dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
650 		dn->dn_oldflags = dn->dn_phys->dn_flags;
651 		dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
652 		if (dmu_objset_userobjused_enabled(dn->dn_objset))
653 			dn->dn_phys->dn_flags |=
654 			    DNODE_FLAG_USEROBJUSED_ACCOUNTED;
655 		mutex_exit(&dn->dn_mtx);
656 		dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
657 	} else {
658 		/* Once we account for it, we should always account for it */
659 		ASSERT(!(dn->dn_phys->dn_flags &
660 		    DNODE_FLAG_USERUSED_ACCOUNTED));
661 		ASSERT(!(dn->dn_phys->dn_flags &
662 		    DNODE_FLAG_USEROBJUSED_ACCOUNTED));
663 	}
664 
665 	mutex_enter(&dn->dn_mtx);
666 	if (dn->dn_allocated_txg == tx->tx_txg) {
667 		/* The dnode is newly allocated or reallocated */
668 		if (dnp->dn_type == DMU_OT_NONE) {
669 			/* this is a first alloc, not a realloc */
670 			dnp->dn_nlevels = 1;
671 			dnp->dn_nblkptr = dn->dn_nblkptr;
672 		}
673 
674 		dnp->dn_type = dn->dn_type;
675 		dnp->dn_bonustype = dn->dn_bonustype;
676 		dnp->dn_bonuslen = dn->dn_bonuslen;
677 	}
678 
679 	dnp->dn_extra_slots = dn->dn_num_slots - 1;
680 
681 	ASSERT(dnp->dn_nlevels > 1 ||
682 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
683 	    BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
684 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
685 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
686 	ASSERT(dnp->dn_nlevels < 2 ||
687 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
688 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
689 
690 	if (dn->dn_next_type[txgoff] != 0) {
691 		dnp->dn_type = dn->dn_type;
692 		dn->dn_next_type[txgoff] = 0;
693 	}
694 
695 	if (dn->dn_next_blksz[txgoff] != 0) {
696 		ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
697 		    SPA_MINBLOCKSIZE) == 0);
698 		ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
699 		    dn->dn_maxblkid == 0 || list_head(list) != NULL ||
700 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
701 		    dnp->dn_datablkszsec ||
702 		    !range_tree_is_empty(dn->dn_free_ranges[txgoff]));
703 		dnp->dn_datablkszsec =
704 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
705 		dn->dn_next_blksz[txgoff] = 0;
706 	}
707 
708 	if (dn->dn_next_bonuslen[txgoff] != 0) {
709 		if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
710 			dnp->dn_bonuslen = 0;
711 		else
712 			dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
713 		ASSERT(dnp->dn_bonuslen <=
714 		    DN_SLOTS_TO_BONUSLEN(dnp->dn_extra_slots + 1));
715 		dn->dn_next_bonuslen[txgoff] = 0;
716 	}
717 
718 	if (dn->dn_next_bonustype[txgoff] != 0) {
719 		ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
720 		dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
721 		dn->dn_next_bonustype[txgoff] = 0;
722 	}
723 
724 	boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
725 	    dn->dn_free_txg <= tx->tx_txg;
726 
727 	/*
728 	 * Remove the spill block if we have been explicitly asked to
729 	 * remove it, or if the object is being removed.
730 	 */
731 	if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
732 		if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
733 			kill_spill = B_TRUE;
734 		dn->dn_rm_spillblk[txgoff] = 0;
735 	}
736 
737 	if (dn->dn_next_indblkshift[txgoff] != 0) {
738 		ASSERT(dnp->dn_nlevels == 1);
739 		dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
740 		dn->dn_next_indblkshift[txgoff] = 0;
741 	}
742 
743 	/*
744 	 * Just take the live (open-context) values for checksum and compress.
745 	 * Strictly speaking it's a future leak, but nothing bad happens if we
746 	 * start using the new checksum or compress algorithm a little early.
747 	 */
748 	dnp->dn_checksum = dn->dn_checksum;
749 	dnp->dn_compress = dn->dn_compress;
750 
751 	mutex_exit(&dn->dn_mtx);
752 
753 	if (kill_spill) {
754 		free_blocks(dn, DN_SPILL_BLKPTR(dn->dn_phys), 1, tx);
755 		mutex_enter(&dn->dn_mtx);
756 		dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
757 		mutex_exit(&dn->dn_mtx);
758 	}
759 
760 	/* process all the "freed" ranges in the file */
761 	if (dn->dn_free_ranges[txgoff] != NULL) {
762 		dnode_sync_free_range_arg_t dsfra;
763 		dsfra.dsfra_dnode = dn;
764 		dsfra.dsfra_tx = tx;
765 		dsfra.dsfra_free_indirects = freeing_dnode;
766 		mutex_enter(&dn->dn_mtx);
767 		if (freeing_dnode) {
768 			ASSERT(range_tree_contains(dn->dn_free_ranges[txgoff],
769 			    0, dn->dn_maxblkid + 1));
770 		}
771 		/*
772 		 * Because dnode_sync_free_range() must drop dn_mtx during its
773 		 * processing, using it as a callback to range_tree_vacate() is
774 		 * not safe.  No other operations (besides destroy) are allowed
775 		 * once range_tree_vacate() has begun, and dropping dn_mtx
776 		 * would leave a window open for another thread to observe that
777 		 * invalid (and unsafe) state.
778 		 */
779 		range_tree_walk(dn->dn_free_ranges[txgoff],
780 		    dnode_sync_free_range, &dsfra);
781 		range_tree_vacate(dn->dn_free_ranges[txgoff], NULL, NULL);
782 		range_tree_destroy(dn->dn_free_ranges[txgoff]);
783 		dn->dn_free_ranges[txgoff] = NULL;
784 		mutex_exit(&dn->dn_mtx);
785 	}
786 
787 	if (freeing_dnode) {
788 		dn->dn_objset->os_freed_dnodes++;
789 		dnode_sync_free(dn, tx);
790 		return;
791 	}
792 
793 	if (dn->dn_num_slots > DNODE_MIN_SLOTS) {
794 		dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
795 		mutex_enter(&ds->ds_lock);
796 		ds->ds_feature_activation[SPA_FEATURE_LARGE_DNODE] =
797 		    (void *)B_TRUE;
798 		mutex_exit(&ds->ds_lock);
799 	}
800 
801 	if (dn->dn_next_nlevels[txgoff]) {
802 		dnode_increase_indirection(dn, tx);
803 		dn->dn_next_nlevels[txgoff] = 0;
804 	}
805 
806 	/*
807 	 * This must be done after dnode_sync_free_range()
808 	 * and dnode_increase_indirection(). See dnode_new_blkid()
809 	 * for an explanation of the high bit being set.
810 	 */
811 	if (dn->dn_next_maxblkid[txgoff]) {
812 		mutex_enter(&dn->dn_mtx);
813 		dnp->dn_maxblkid =
814 		    dn->dn_next_maxblkid[txgoff] & ~DMU_NEXT_MAXBLKID_SET;
815 		dn->dn_next_maxblkid[txgoff] = 0;
816 		mutex_exit(&dn->dn_mtx);
817 	}
818 
819 	if (dn->dn_next_nblkptr[txgoff]) {
820 		/* this should only happen on a realloc */
821 		ASSERT(dn->dn_allocated_txg == tx->tx_txg);
822 		if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
823 			/* zero the new blkptrs we are gaining */
824 			bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
825 			    sizeof (blkptr_t) *
826 			    (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
827 #ifdef ZFS_DEBUG
828 		} else {
829 			int i;
830 			ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
831 			/* the blkptrs we are losing better be unallocated */
832 			for (i = 0; i < dnp->dn_nblkptr; i++) {
833 				if (i >= dn->dn_next_nblkptr[txgoff])
834 					ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
835 			}
836 #endif
837 		}
838 		mutex_enter(&dn->dn_mtx);
839 		dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
840 		dn->dn_next_nblkptr[txgoff] = 0;
841 		mutex_exit(&dn->dn_mtx);
842 	}
843 
844 	dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
845 
846 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
847 		ASSERT3P(list_head(list), ==, NULL);
848 		dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
849 	}
850 
851 	/*
852 	 * Although we have dropped our reference to the dnode, it
853 	 * can't be evicted until its written, and we haven't yet
854 	 * initiated the IO for the dnode's dbuf.  Additionally, the caller
855 	 * has already added a reference to the dnode because it's on the
856 	 * os_synced_dnodes list.
857 	 */
858 }
859