xref: /freebsd/sys/contrib/openzfs/module/zfs/dnode_sync.c (revision 3078531de10dcae44b253a35125c949ff4235284)
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] =
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 	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 	int off, num;
179 	int i, err, epbs;
180 	uint64_t txg = tx->tx_txg;
181 	dnode_t *dn;
182 
183 	DB_DNODE_ENTER(db);
184 	dn = DB_DNODE(db);
185 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
186 	off = start - (db->db_blkid * 1<<epbs);
187 	num = end - start + 1;
188 
189 	ASSERT3U(off, >=, 0);
190 	ASSERT3U(num, >=, 0);
191 	ASSERT3U(db->db_level, >, 0);
192 	ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
193 	ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
194 	ASSERT(db->db_blkptr != NULL);
195 
196 	for (i = off; i < off+num; i++) {
197 		uint64_t *buf;
198 		dmu_buf_impl_t *child;
199 		dbuf_dirty_record_t *dr;
200 		int j;
201 
202 		ASSERT(db->db_level == 1);
203 
204 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
205 		err = dbuf_hold_impl(dn, db->db_level - 1,
206 		    (db->db_blkid << epbs) + i, TRUE, FALSE, FTAG, &child);
207 		rw_exit(&dn->dn_struct_rwlock);
208 		if (err == ENOENT)
209 			continue;
210 		ASSERT(err == 0);
211 		ASSERT(child->db_level == 0);
212 		dr = dbuf_find_dirty_eq(child, txg);
213 
214 		/* data_old better be zeroed */
215 		if (dr) {
216 			buf = dr->dt.dl.dr_data->b_data;
217 			for (j = 0; j < child->db.db_size >> 3; j++) {
218 				if (buf[j] != 0) {
219 					panic("freed data not zero: "
220 					    "child=%p i=%d off=%d num=%d\n",
221 					    (void *)child, i, off, num);
222 				}
223 			}
224 		}
225 
226 		/*
227 		 * db_data better be zeroed unless it's dirty in a
228 		 * future txg.
229 		 */
230 		mutex_enter(&child->db_mtx);
231 		buf = child->db.db_data;
232 		if (buf != NULL && child->db_state != DB_FILL &&
233 		    list_is_empty(&child->db_dirty_records)) {
234 			for (j = 0; j < child->db.db_size >> 3; j++) {
235 				if (buf[j] != 0) {
236 					panic("freed data not zero: "
237 					    "child=%p i=%d off=%d num=%d\n",
238 					    (void *)child, i, off, num);
239 				}
240 			}
241 		}
242 		mutex_exit(&child->db_mtx);
243 
244 		dbuf_rele(child, FTAG);
245 	}
246 	DB_DNODE_EXIT(db);
247 }
248 #endif
249 
250 /*
251  * We don't usually free the indirect blocks here.  If in one txg we have a
252  * free_range and a write to the same indirect block, it's important that we
253  * preserve the hole's birth times. Therefore, we don't free any any indirect
254  * blocks in free_children().  If an indirect block happens to turn into all
255  * holes, it will be freed by dbuf_write_children_ready, which happens at a
256  * point in the syncing process where we know for certain the contents of the
257  * indirect block.
258  *
259  * However, if we're freeing a dnode, its space accounting must go to zero
260  * before we actually try to free the dnode, or we will trip an assertion. In
261  * addition, we know the case described above cannot occur, because the dnode is
262  * being freed.  Therefore, we free the indirect blocks immediately in that
263  * case.
264  */
265 static void
266 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
267     boolean_t free_indirects, dmu_tx_t *tx)
268 {
269 	dnode_t *dn;
270 	blkptr_t *bp;
271 	dmu_buf_impl_t *subdb;
272 	uint64_t start, end, dbstart, dbend;
273 	unsigned int epbs, shift, i;
274 
275 	/*
276 	 * There is a small possibility that this block will not be cached:
277 	 *   1 - if level > 1 and there are no children with level <= 1
278 	 *   2 - if this block was evicted since we read it from
279 	 *	 dmu_tx_hold_free().
280 	 */
281 	if (db->db_state != DB_CACHED)
282 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
283 
284 	/*
285 	 * If we modify this indirect block, and we are not freeing the
286 	 * dnode (!free_indirects), then this indirect block needs to get
287 	 * written to disk by dbuf_write().  If it is dirty, we know it will
288 	 * be written (otherwise, we would have incorrect on-disk state
289 	 * because the space would be freed but still referenced by the BP
290 	 * in this indirect block).  Therefore we VERIFY that it is
291 	 * dirty.
292 	 *
293 	 * Our VERIFY covers some cases that do not actually have to be
294 	 * dirty, but the open-context code happens to dirty.  E.g. if the
295 	 * blocks we are freeing are all holes, because in that case, we
296 	 * are only freeing part of this indirect block, so it is an
297 	 * ancestor of the first or last block to be freed.  The first and
298 	 * last L1 indirect blocks are always dirtied by dnode_free_range().
299 	 */
300 	db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
301 	VERIFY(BP_GET_FILL(db->db_blkptr) == 0 || db->db_dirtycnt > 0);
302 	dmu_buf_unlock_parent(db, dblt, FTAG);
303 
304 	dbuf_release_bp(db);
305 	bp = db->db.db_data;
306 
307 	DB_DNODE_ENTER(db);
308 	dn = DB_DNODE(db);
309 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
310 	ASSERT3U(epbs, <, 31);
311 	shift = (db->db_level - 1) * epbs;
312 	dbstart = db->db_blkid << epbs;
313 	start = blkid >> shift;
314 	if (dbstart < start) {
315 		bp += start - dbstart;
316 	} else {
317 		start = dbstart;
318 	}
319 	dbend = ((db->db_blkid + 1) << epbs) - 1;
320 	end = (blkid + nblks - 1) >> shift;
321 	if (dbend <= end)
322 		end = dbend;
323 
324 	ASSERT3U(start, <=, end);
325 
326 	if (db->db_level == 1) {
327 		FREE_VERIFY(db, start, end, tx);
328 		rw_enter(&db->db_rwlock, RW_WRITER);
329 		free_blocks(dn, bp, end - start + 1, tx);
330 		rw_exit(&db->db_rwlock);
331 	} else {
332 		for (uint64_t id = start; id <= end; id++, bp++) {
333 			if (BP_IS_HOLE(bp))
334 				continue;
335 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
336 			VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
337 			    id, TRUE, FALSE, FTAG, &subdb));
338 			rw_exit(&dn->dn_struct_rwlock);
339 			ASSERT3P(bp, ==, subdb->db_blkptr);
340 
341 			free_children(subdb, blkid, nblks, free_indirects, tx);
342 			dbuf_rele(subdb, FTAG);
343 		}
344 	}
345 
346 	if (free_indirects) {
347 		rw_enter(&db->db_rwlock, RW_WRITER);
348 		for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++)
349 			ASSERT(BP_IS_HOLE(bp));
350 		memset(db->db.db_data, 0, db->db.db_size);
351 		free_blocks(dn, db->db_blkptr, 1, tx);
352 		rw_exit(&db->db_rwlock);
353 	}
354 
355 	DB_DNODE_EXIT(db);
356 	arc_buf_freeze(db->db_buf);
357 }
358 
359 /*
360  * Traverse the indicated range of the provided file
361  * and "free" all the blocks contained there.
362  */
363 static void
364 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
365     boolean_t free_indirects, dmu_tx_t *tx)
366 {
367 	blkptr_t *bp = dn->dn_phys->dn_blkptr;
368 	int dnlevel = dn->dn_phys->dn_nlevels;
369 	boolean_t trunc = B_FALSE;
370 
371 	if (blkid > dn->dn_phys->dn_maxblkid)
372 		return;
373 
374 	ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
375 	if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
376 		nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
377 		trunc = B_TRUE;
378 	}
379 
380 	/* There are no indirect blocks in the object */
381 	if (dnlevel == 1) {
382 		if (blkid >= dn->dn_phys->dn_nblkptr) {
383 			/* this range was never made persistent */
384 			return;
385 		}
386 		ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
387 		free_blocks(dn, bp + blkid, nblks, tx);
388 	} else {
389 		int shift = (dnlevel - 1) *
390 		    (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
391 		int start = blkid >> shift;
392 		int end = (blkid + nblks - 1) >> shift;
393 		dmu_buf_impl_t *db;
394 
395 		ASSERT(start < dn->dn_phys->dn_nblkptr);
396 		bp += start;
397 		for (int i = start; i <= end; i++, bp++) {
398 			if (BP_IS_HOLE(bp))
399 				continue;
400 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
401 			VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
402 			    TRUE, FALSE, FTAG, &db));
403 			rw_exit(&dn->dn_struct_rwlock);
404 			free_children(db, blkid, nblks, free_indirects, tx);
405 			dbuf_rele(db, FTAG);
406 		}
407 	}
408 
409 	/*
410 	 * Do not truncate the maxblkid if we are performing a raw
411 	 * receive. The raw receive sets the maxblkid manually and
412 	 * must not be overridden. Usually, the last DRR_FREE record
413 	 * will be at the maxblkid, because the source system sets
414 	 * the maxblkid when truncating. However, if the last block
415 	 * was freed by overwriting with zeros and being compressed
416 	 * away to a hole, the source system will generate a DRR_FREE
417 	 * record while leaving the maxblkid after the end of that
418 	 * record. In this case we need to leave the maxblkid as
419 	 * indicated in the DRR_OBJECT record, so that it matches the
420 	 * source system, ensuring that the cryptographic hashes will
421 	 * match.
422 	 */
423 	if (trunc && !dn->dn_objset->os_raw_receive) {
424 		uint64_t off __maybe_unused;
425 		dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
426 
427 		off = (dn->dn_phys->dn_maxblkid + 1) *
428 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
429 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
430 		    dn->dn_phys->dn_maxblkid == 0 ||
431 		    dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
432 	}
433 }
434 
435 typedef struct dnode_sync_free_range_arg {
436 	dnode_t *dsfra_dnode;
437 	dmu_tx_t *dsfra_tx;
438 	boolean_t dsfra_free_indirects;
439 } dnode_sync_free_range_arg_t;
440 
441 static void
442 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
443 {
444 	dnode_sync_free_range_arg_t *dsfra = arg;
445 	dnode_t *dn = dsfra->dsfra_dnode;
446 
447 	mutex_exit(&dn->dn_mtx);
448 	dnode_sync_free_range_impl(dn, blkid, nblks,
449 	    dsfra->dsfra_free_indirects, dsfra->dsfra_tx);
450 	mutex_enter(&dn->dn_mtx);
451 }
452 
453 /*
454  * Try to kick all the dnode's dbufs out of the cache...
455  */
456 void
457 dnode_evict_dbufs(dnode_t *dn)
458 {
459 	dmu_buf_impl_t *db_marker;
460 	dmu_buf_impl_t *db, *db_next;
461 
462 	db_marker = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
463 
464 	mutex_enter(&dn->dn_dbufs_mtx);
465 	for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
466 
467 #ifdef	ZFS_DEBUG
468 		DB_DNODE_ENTER(db);
469 		ASSERT3P(DB_DNODE(db), ==, dn);
470 		DB_DNODE_EXIT(db);
471 #endif	/* DEBUG */
472 
473 		mutex_enter(&db->db_mtx);
474 		if (db->db_state != DB_EVICTING &&
475 		    zfs_refcount_is_zero(&db->db_holds)) {
476 			db_marker->db_level = db->db_level;
477 			db_marker->db_blkid = db->db_blkid;
478 			db_marker->db_state = DB_SEARCH;
479 			avl_insert_here(&dn->dn_dbufs, db_marker, db,
480 			    AVL_BEFORE);
481 
482 			/*
483 			 * We need to use the "marker" dbuf rather than
484 			 * simply getting the next dbuf, because
485 			 * dbuf_destroy() may actually remove multiple dbufs.
486 			 * It can call itself recursively on the parent dbuf,
487 			 * which may also be removed from dn_dbufs.  The code
488 			 * flow would look like:
489 			 *
490 			 * dbuf_destroy():
491 			 *   dnode_rele_and_unlock(parent_dbuf, evicting=TRUE):
492 			 *	if (!cacheable || pending_evict)
493 			 *	  dbuf_destroy()
494 			 */
495 			dbuf_destroy(db);
496 
497 			db_next = AVL_NEXT(&dn->dn_dbufs, db_marker);
498 			avl_remove(&dn->dn_dbufs, db_marker);
499 		} else {
500 			db->db_pending_evict = TRUE;
501 			mutex_exit(&db->db_mtx);
502 			db_next = AVL_NEXT(&dn->dn_dbufs, db);
503 		}
504 	}
505 	mutex_exit(&dn->dn_dbufs_mtx);
506 
507 	kmem_free(db_marker, sizeof (dmu_buf_impl_t));
508 
509 	dnode_evict_bonus(dn);
510 }
511 
512 void
513 dnode_evict_bonus(dnode_t *dn)
514 {
515 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
516 	if (dn->dn_bonus != NULL) {
517 		if (zfs_refcount_is_zero(&dn->dn_bonus->db_holds)) {
518 			mutex_enter(&dn->dn_bonus->db_mtx);
519 			dbuf_destroy(dn->dn_bonus);
520 			dn->dn_bonus = NULL;
521 		} else {
522 			dn->dn_bonus->db_pending_evict = TRUE;
523 		}
524 	}
525 	rw_exit(&dn->dn_struct_rwlock);
526 }
527 
528 static void
529 dnode_undirty_dbufs(list_t *list)
530 {
531 	dbuf_dirty_record_t *dr;
532 
533 	while ((dr = list_head(list))) {
534 		dmu_buf_impl_t *db = dr->dr_dbuf;
535 		uint64_t txg = dr->dr_txg;
536 
537 		if (db->db_level != 0)
538 			dnode_undirty_dbufs(&dr->dt.di.dr_children);
539 
540 		mutex_enter(&db->db_mtx);
541 		/* XXX - use dbuf_undirty()? */
542 		list_remove(list, dr);
543 		ASSERT(list_head(&db->db_dirty_records) == dr);
544 		list_remove_head(&db->db_dirty_records);
545 		ASSERT(list_is_empty(&db->db_dirty_records));
546 		db->db_dirtycnt -= 1;
547 		if (db->db_level == 0) {
548 			ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
549 			    dr->dt.dl.dr_data == db->db_buf);
550 			dbuf_unoverride(dr);
551 		} else {
552 			mutex_destroy(&dr->dt.di.dr_mtx);
553 			list_destroy(&dr->dt.di.dr_children);
554 		}
555 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
556 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
557 	}
558 }
559 
560 static void
561 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
562 {
563 	int txgoff = tx->tx_txg & TXG_MASK;
564 
565 	ASSERT(dmu_tx_is_syncing(tx));
566 
567 	/*
568 	 * Our contents should have been freed in dnode_sync() by the
569 	 * free range record inserted by the caller of dnode_free().
570 	 */
571 	ASSERT0(DN_USED_BYTES(dn->dn_phys));
572 	ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
573 
574 	dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
575 	dnode_evict_dbufs(dn);
576 
577 	/*
578 	 * XXX - It would be nice to assert this, but we may still
579 	 * have residual holds from async evictions from the arc...
580 	 *
581 	 * zfs_obj_to_path() also depends on this being
582 	 * commented out.
583 	 *
584 	 * ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 1);
585 	 */
586 
587 	/* Undirty next bits */
588 	dn->dn_next_nlevels[txgoff] = 0;
589 	dn->dn_next_indblkshift[txgoff] = 0;
590 	dn->dn_next_blksz[txgoff] = 0;
591 	dn->dn_next_maxblkid[txgoff] = 0;
592 
593 	/* ASSERT(blkptrs are zero); */
594 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
595 	ASSERT(dn->dn_type != DMU_OT_NONE);
596 
597 	ASSERT(dn->dn_free_txg > 0);
598 	if (dn->dn_allocated_txg != dn->dn_free_txg)
599 		dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
600 	memset(dn->dn_phys, 0, sizeof (dnode_phys_t) * dn->dn_num_slots);
601 	dnode_free_interior_slots(dn);
602 
603 	mutex_enter(&dn->dn_mtx);
604 	dn->dn_type = DMU_OT_NONE;
605 	dn->dn_maxblkid = 0;
606 	dn->dn_allocated_txg = 0;
607 	dn->dn_free_txg = 0;
608 	dn->dn_have_spill = B_FALSE;
609 	dn->dn_num_slots = 1;
610 	mutex_exit(&dn->dn_mtx);
611 
612 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
613 
614 	dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
615 	/*
616 	 * Now that we've released our hold, the dnode may
617 	 * be evicted, so we mustn't access it.
618 	 */
619 }
620 
621 /*
622  * Write out the dnode's dirty buffers.
623  */
624 void
625 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
626 {
627 	objset_t *os = dn->dn_objset;
628 	dnode_phys_t *dnp = dn->dn_phys;
629 	int txgoff = tx->tx_txg & TXG_MASK;
630 	list_t *list = &dn->dn_dirty_records[txgoff];
631 	static const dnode_phys_t zerodn __maybe_unused = { 0 };
632 	boolean_t kill_spill = B_FALSE;
633 
634 	ASSERT(dmu_tx_is_syncing(tx));
635 	ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
636 	ASSERT(dnp->dn_type != DMU_OT_NONE ||
637 	    memcmp(dnp, &zerodn, DNODE_MIN_SIZE) == 0);
638 	DNODE_VERIFY(dn);
639 
640 	ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
641 
642 	/*
643 	 * Do user accounting if it is enabled and this is not
644 	 * an encrypted receive.
645 	 */
646 	if (dmu_objset_userused_enabled(os) &&
647 	    !DMU_OBJECT_IS_SPECIAL(dn->dn_object) &&
648 	    (!os->os_encrypted || !dmu_objset_is_receiving(os))) {
649 		mutex_enter(&dn->dn_mtx);
650 		dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
651 		dn->dn_oldflags = dn->dn_phys->dn_flags;
652 		dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
653 		if (dmu_objset_userobjused_enabled(dn->dn_objset))
654 			dn->dn_phys->dn_flags |=
655 			    DNODE_FLAG_USEROBJUSED_ACCOUNTED;
656 		mutex_exit(&dn->dn_mtx);
657 		dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
658 	} else if (!(os->os_encrypted && dmu_objset_is_receiving(os))) {
659 		/*
660 		 * Once we account for it, we should always account for it,
661 		 * except for the case of a raw receive. We will not be able
662 		 * to account for it until the receiving dataset has been
663 		 * mounted.
664 		 */
665 		ASSERT(!(dn->dn_phys->dn_flags &
666 		    DNODE_FLAG_USERUSED_ACCOUNTED));
667 		ASSERT(!(dn->dn_phys->dn_flags &
668 		    DNODE_FLAG_USEROBJUSED_ACCOUNTED));
669 	}
670 
671 	mutex_enter(&dn->dn_mtx);
672 	if (dn->dn_allocated_txg == tx->tx_txg) {
673 		/* The dnode is newly allocated or reallocated */
674 		if (dnp->dn_type == DMU_OT_NONE) {
675 			/* this is a first alloc, not a realloc */
676 			dnp->dn_nlevels = 1;
677 			dnp->dn_nblkptr = dn->dn_nblkptr;
678 		}
679 
680 		dnp->dn_type = dn->dn_type;
681 		dnp->dn_bonustype = dn->dn_bonustype;
682 		dnp->dn_bonuslen = dn->dn_bonuslen;
683 	}
684 
685 	dnp->dn_extra_slots = dn->dn_num_slots - 1;
686 
687 	ASSERT(dnp->dn_nlevels > 1 ||
688 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
689 	    BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
690 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
691 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
692 	ASSERT(dnp->dn_nlevels < 2 ||
693 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
694 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
695 
696 	if (dn->dn_next_type[txgoff] != 0) {
697 		dnp->dn_type = dn->dn_type;
698 		dn->dn_next_type[txgoff] = 0;
699 	}
700 
701 	if (dn->dn_next_blksz[txgoff] != 0) {
702 		ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
703 		    SPA_MINBLOCKSIZE) == 0);
704 		ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
705 		    dn->dn_maxblkid == 0 || list_head(list) != NULL ||
706 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
707 		    dnp->dn_datablkszsec ||
708 		    !range_tree_is_empty(dn->dn_free_ranges[txgoff]));
709 		dnp->dn_datablkszsec =
710 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
711 		dn->dn_next_blksz[txgoff] = 0;
712 	}
713 
714 	if (dn->dn_next_bonuslen[txgoff] != 0) {
715 		if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
716 			dnp->dn_bonuslen = 0;
717 		else
718 			dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
719 		ASSERT(dnp->dn_bonuslen <=
720 		    DN_SLOTS_TO_BONUSLEN(dnp->dn_extra_slots + 1));
721 		dn->dn_next_bonuslen[txgoff] = 0;
722 	}
723 
724 	if (dn->dn_next_bonustype[txgoff] != 0) {
725 		ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
726 		dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
727 		dn->dn_next_bonustype[txgoff] = 0;
728 	}
729 
730 	boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
731 	    dn->dn_free_txg <= tx->tx_txg;
732 
733 	/*
734 	 * Remove the spill block if we have been explicitly asked to
735 	 * remove it, or if the object is being removed.
736 	 */
737 	if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
738 		if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
739 			kill_spill = B_TRUE;
740 		dn->dn_rm_spillblk[txgoff] = 0;
741 	}
742 
743 	if (dn->dn_next_indblkshift[txgoff] != 0) {
744 		ASSERT(dnp->dn_nlevels == 1);
745 		dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
746 		dn->dn_next_indblkshift[txgoff] = 0;
747 	}
748 
749 	/*
750 	 * Just take the live (open-context) values for checksum and compress.
751 	 * Strictly speaking it's a future leak, but nothing bad happens if we
752 	 * start using the new checksum or compress algorithm a little early.
753 	 */
754 	dnp->dn_checksum = dn->dn_checksum;
755 	dnp->dn_compress = dn->dn_compress;
756 
757 	mutex_exit(&dn->dn_mtx);
758 
759 	if (kill_spill) {
760 		free_blocks(dn, DN_SPILL_BLKPTR(dn->dn_phys), 1, tx);
761 		mutex_enter(&dn->dn_mtx);
762 		dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
763 		mutex_exit(&dn->dn_mtx);
764 	}
765 
766 	/* process all the "freed" ranges in the file */
767 	if (dn->dn_free_ranges[txgoff] != NULL) {
768 		dnode_sync_free_range_arg_t dsfra;
769 		dsfra.dsfra_dnode = dn;
770 		dsfra.dsfra_tx = tx;
771 		dsfra.dsfra_free_indirects = freeing_dnode;
772 		mutex_enter(&dn->dn_mtx);
773 		if (freeing_dnode) {
774 			ASSERT(range_tree_contains(dn->dn_free_ranges[txgoff],
775 			    0, dn->dn_maxblkid + 1));
776 		}
777 		/*
778 		 * Because dnode_sync_free_range() must drop dn_mtx during its
779 		 * processing, using it as a callback to range_tree_vacate() is
780 		 * not safe.  No other operations (besides destroy) are allowed
781 		 * once range_tree_vacate() has begun, and dropping dn_mtx
782 		 * would leave a window open for another thread to observe that
783 		 * invalid (and unsafe) state.
784 		 */
785 		range_tree_walk(dn->dn_free_ranges[txgoff],
786 		    dnode_sync_free_range, &dsfra);
787 		range_tree_vacate(dn->dn_free_ranges[txgoff], NULL, NULL);
788 		range_tree_destroy(dn->dn_free_ranges[txgoff]);
789 		dn->dn_free_ranges[txgoff] = NULL;
790 		mutex_exit(&dn->dn_mtx);
791 	}
792 
793 	if (freeing_dnode) {
794 		dn->dn_objset->os_freed_dnodes++;
795 		dnode_sync_free(dn, tx);
796 		return;
797 	}
798 
799 	if (dn->dn_num_slots > DNODE_MIN_SLOTS) {
800 		dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
801 		mutex_enter(&ds->ds_lock);
802 		ds->ds_feature_activation[SPA_FEATURE_LARGE_DNODE] =
803 		    (void *)B_TRUE;
804 		mutex_exit(&ds->ds_lock);
805 	}
806 
807 	if (dn->dn_next_nlevels[txgoff]) {
808 		dnode_increase_indirection(dn, tx);
809 		dn->dn_next_nlevels[txgoff] = 0;
810 	}
811 
812 	/*
813 	 * This must be done after dnode_sync_free_range()
814 	 * and dnode_increase_indirection(). See dnode_new_blkid()
815 	 * for an explanation of the high bit being set.
816 	 */
817 	if (dn->dn_next_maxblkid[txgoff]) {
818 		mutex_enter(&dn->dn_mtx);
819 		dnp->dn_maxblkid =
820 		    dn->dn_next_maxblkid[txgoff] & ~DMU_NEXT_MAXBLKID_SET;
821 		dn->dn_next_maxblkid[txgoff] = 0;
822 		mutex_exit(&dn->dn_mtx);
823 	}
824 
825 	if (dn->dn_next_nblkptr[txgoff]) {
826 		/* this should only happen on a realloc */
827 		ASSERT(dn->dn_allocated_txg == tx->tx_txg);
828 		if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
829 			/* zero the new blkptrs we are gaining */
830 			memset(dnp->dn_blkptr + dnp->dn_nblkptr, 0,
831 			    sizeof (blkptr_t) *
832 			    (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
833 #ifdef ZFS_DEBUG
834 		} else {
835 			int i;
836 			ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
837 			/* the blkptrs we are losing better be unallocated */
838 			for (i = 0; i < dnp->dn_nblkptr; i++) {
839 				if (i >= dn->dn_next_nblkptr[txgoff])
840 					ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
841 			}
842 #endif
843 		}
844 		mutex_enter(&dn->dn_mtx);
845 		dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
846 		dn->dn_next_nblkptr[txgoff] = 0;
847 		mutex_exit(&dn->dn_mtx);
848 	}
849 
850 	dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
851 
852 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
853 		ASSERT3P(list_head(list), ==, NULL);
854 		dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
855 	}
856 
857 	ASSERT3U(dnp->dn_bonuslen, <=, DN_MAX_BONUS_LEN(dnp));
858 
859 	/*
860 	 * Although we have dropped our reference to the dnode, it
861 	 * can't be evicted until its written, and we haven't yet
862 	 * initiated the IO for the dnode's dbuf.  Additionally, the caller
863 	 * has already added a reference to the dnode because it's on the
864 	 * os_synced_dnodes list.
865 	 */
866 }
867