xref: /illumos-gate/usr/src/uts/common/fs/zfs/dnode_sync.c (revision e4f7ccbe1e17940fdbde623116bcc8cd761d9ffc)
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, 2018 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	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 = child->db_last_dirty;
212 		while (dr && dr->dr_txg > txg)
213 			dr = dr->dr_next;
214 		ASSERT(dr == NULL || dr->dr_txg == txg);
215 
216 		/* data_old better be zeroed */
217 		if (dr) {
218 			buf = dr->dt.dl.dr_data->b_data;
219 			for (j = 0; j < child->db.db_size >> 3; j++) {
220 				if (buf[j] != 0) {
221 					panic("freed data not zero: "
222 					    "child=%p i=%d off=%d num=%d\n",
223 					    (void *)child, i, off, num);
224 				}
225 			}
226 		}
227 
228 		/*
229 		 * db_data better be zeroed unless it's dirty in a
230 		 * future txg.
231 		 */
232 		mutex_enter(&child->db_mtx);
233 		buf = child->db.db_data;
234 		if (buf != NULL && child->db_state != DB_FILL &&
235 		    child->db_last_dirty == NULL) {
236 			for (j = 0; j < child->db.db_size >> 3; j++) {
237 				if (buf[j] != 0) {
238 					panic("freed data not zero: "
239 					    "child=%p i=%d off=%d num=%d\n",
240 					    (void *)child, i, off, num);
241 				}
242 			}
243 		}
244 		mutex_exit(&child->db_mtx);
245 
246 		dbuf_rele(child, FTAG);
247 	}
248 	DB_DNODE_EXIT(db);
249 }
250 #endif
251 
252 /*
253  * We don't usually free the indirect blocks here.  If in one txg we have a
254  * free_range and a write to the same indirect block, it's important that we
255  * preserve the hole's birth times. Therefore, we don't free any any indirect
256  * blocks in free_children().  If an indirect block happens to turn into all
257  * holes, it will be freed by dbuf_write_children_ready, which happens at a
258  * point in the syncing process where we know for certain the contents of the
259  * indirect block.
260  *
261  * However, if we're freeing a dnode, its space accounting must go to zero
262  * before we actually try to free the dnode, or we will trip an assertion. In
263  * addition, we know the case described above cannot occur, because the dnode is
264  * being freed.  Therefore, we free the indirect blocks immediately in that
265  * case.
266  */
267 static void
268 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
269     boolean_t free_indirects, dmu_tx_t *tx)
270 {
271 	dnode_t *dn;
272 	blkptr_t *bp;
273 	dmu_buf_impl_t *subdb;
274 	uint64_t start, end, dbstart, dbend;
275 	unsigned int epbs, shift, i;
276 
277 	/*
278 	 * There is a small possibility that this block will not be cached:
279 	 *   1 - if level > 1 and there are no children with level <= 1
280 	 *   2 - if this block was evicted since we read it from
281 	 *	 dmu_tx_hold_free().
282 	 */
283 	if (db->db_state != DB_CACHED)
284 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
285 
286 	/*
287 	 * If we modify this indirect block, and we are not freeing the
288 	 * dnode (!free_indirects), then this indirect block needs to get
289 	 * written to disk by dbuf_write().  If it is dirty, we know it will
290 	 * be written (otherwise, we would have incorrect on-disk state
291 	 * because the space would be freed but still referenced by the BP
292 	 * in this indirect block).  Therefore we VERIFY that it is
293 	 * dirty.
294 	 *
295 	 * Our VERIFY covers some cases that do not actually have to be
296 	 * dirty, but the open-context code happens to dirty.  E.g. if the
297 	 * blocks we are freeing are all holes, because in that case, we
298 	 * are only freeing part of this indirect block, so it is an
299 	 * ancestor of the first or last block to be freed.  The first and
300 	 * last L1 indirect blocks are always dirtied by dnode_free_range().
301 	 */
302 	db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
303 	VERIFY(BP_GET_FILL(db->db_blkptr) == 0 || db->db_dirtycnt > 0);
304 	dmu_buf_unlock_parent(db, dblt, FTAG);
305 
306 	dbuf_release_bp(db);
307 	bp = db->db.db_data;
308 
309 	DB_DNODE_ENTER(db);
310 	dn = DB_DNODE(db);
311 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
312 	ASSERT3U(epbs, <, 31);
313 	shift = (db->db_level - 1) * epbs;
314 	dbstart = db->db_blkid << epbs;
315 	start = blkid >> shift;
316 	if (dbstart < start) {
317 		bp += start - dbstart;
318 	} else {
319 		start = dbstart;
320 	}
321 	dbend = ((db->db_blkid + 1) << epbs) - 1;
322 	end = (blkid + nblks - 1) >> shift;
323 	if (dbend <= end)
324 		end = dbend;
325 
326 	ASSERT3U(start, <=, end);
327 
328 	if (db->db_level == 1) {
329 		FREE_VERIFY(db, start, end, tx);
330 		rw_enter(&db->db_rwlock, RW_WRITER);
331 		free_blocks(dn, bp, end - start + 1, tx);
332 		rw_exit(&db->db_rwlock);
333 	} else {
334 		for (uint64_t id = start; id <= end; id++, bp++) {
335 			if (BP_IS_HOLE(bp))
336 				continue;
337 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
338 			VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
339 			    id, TRUE, FALSE, FTAG, &subdb));
340 			rw_exit(&dn->dn_struct_rwlock);
341 			ASSERT3P(bp, ==, subdb->db_blkptr);
342 
343 			free_children(subdb, blkid, nblks, free_indirects, tx);
344 			dbuf_rele(subdb, FTAG);
345 		}
346 	}
347 
348 	if (free_indirects) {
349 		rw_enter(&db->db_rwlock, RW_WRITER);
350 		for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++)
351 			ASSERT(BP_IS_HOLE(bp));
352 		bzero(db->db.db_data, db->db.db_size);
353 		free_blocks(dn, db->db_blkptr, 1, tx);
354 		rw_exit(&db->db_rwlock);
355 	}
356 
357 	DB_DNODE_EXIT(db);
358 	arc_buf_freeze(db->db_buf);
359 }
360 
361 /*
362  * Traverse the indicated range of the provided file
363  * and "free" all the blocks contained there.
364  */
365 static void
366 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
367     boolean_t free_indirects, dmu_tx_t *tx)
368 {
369 	blkptr_t *bp = dn->dn_phys->dn_blkptr;
370 	int dnlevel = dn->dn_phys->dn_nlevels;
371 	boolean_t trunc = B_FALSE;
372 
373 	if (blkid > dn->dn_phys->dn_maxblkid)
374 		return;
375 
376 	ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
377 	if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
378 		nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
379 		trunc = B_TRUE;
380 	}
381 
382 	/* There are no indirect blocks in the object */
383 	if (dnlevel == 1) {
384 		if (blkid >= dn->dn_phys->dn_nblkptr) {
385 			/* this range was never made persistent */
386 			return;
387 		}
388 		ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
389 		free_blocks(dn, bp + blkid, nblks, tx);
390 	} else {
391 		int shift = (dnlevel - 1) *
392 		    (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
393 		int start = blkid >> shift;
394 		int end = (blkid + nblks - 1) >> shift;
395 		dmu_buf_impl_t *db;
396 
397 		ASSERT(start < dn->dn_phys->dn_nblkptr);
398 		bp += start;
399 		for (int i = start; i <= end; i++, bp++) {
400 			if (BP_IS_HOLE(bp))
401 				continue;
402 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
403 			VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
404 			    TRUE, FALSE, FTAG, &db));
405 			rw_exit(&dn->dn_struct_rwlock);
406 			free_children(db, blkid, nblks, free_indirects, tx);
407 			dbuf_rele(db, FTAG);
408 		}
409 	}
410 
411 	/*
412 	 * Do not truncate the maxblkid if we are performing a raw
413 	 * receive. The raw receive sets the maxblkid manually and
414 	 * must not be overridden. Usually, the last DRR_FREE record
415 	 * will be at the maxblkid, because the source system sets
416 	 * the maxblkid when truncating. However, if the last block
417 	 * was freed by overwriting with zeros and being compressed
418 	 * away to a hole, the source system will generate a DRR_FREE
419 	 * record while leaving the maxblkid after the end of that
420 	 * record. In this case we need to leave the maxblkid as
421 	 * indicated in the DRR_OBJECT record, so that it matches the
422 	 * source system, ensuring that the cryptographic hashes will
423 	 * match.
424 	 */
425 	if (trunc && !dn->dn_objset->os_raw_receive) {
426 		dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
427 
428 		uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
429 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
430 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
431 		    dn->dn_phys->dn_maxblkid == 0 ||
432 		    dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
433 	}
434 }
435 
436 typedef struct dnode_sync_free_range_arg {
437 	dnode_t *dsfra_dnode;
438 	dmu_tx_t *dsfra_tx;
439 	boolean_t dsfra_free_indirects;
440 } dnode_sync_free_range_arg_t;
441 
442 static void
443 dnode_sync_free_range(void *arg, uint64_t blkid, uint64_t nblks)
444 {
445 	dnode_sync_free_range_arg_t *dsfra = arg;
446 	dnode_t *dn = dsfra->dsfra_dnode;
447 
448 	mutex_exit(&dn->dn_mtx);
449 	dnode_sync_free_range_impl(dn, blkid, nblks,
450 	    dsfra->dsfra_free_indirects, dsfra->dsfra_tx);
451 	mutex_enter(&dn->dn_mtx);
452 }
453 
454 /*
455  * Try to kick all the dnode's dbufs out of the cache...
456  */
457 void
458 dnode_evict_dbufs(dnode_t *dn)
459 {
460 	dmu_buf_impl_t db_marker;
461 	dmu_buf_impl_t *db, *db_next;
462 
463 	mutex_enter(&dn->dn_dbufs_mtx);
464 	for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
465 
466 #ifdef	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 	dnode_evict_bonus(dn);
507 }
508 
509 void
510 dnode_evict_bonus(dnode_t *dn)
511 {
512 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
513 	if (dn->dn_bonus != NULL) {
514 		if (zfs_refcount_is_zero(&dn->dn_bonus->db_holds)) {
515 			mutex_enter(&dn->dn_bonus->db_mtx);
516 			dbuf_destroy(dn->dn_bonus);
517 			dn->dn_bonus = NULL;
518 		} else {
519 			dn->dn_bonus->db_pending_evict = TRUE;
520 		}
521 	}
522 	rw_exit(&dn->dn_struct_rwlock);
523 }
524 
525 static void
526 dnode_undirty_dbufs(list_t *list)
527 {
528 	dbuf_dirty_record_t *dr;
529 
530 	while (dr = list_head(list)) {
531 		dmu_buf_impl_t *db = dr->dr_dbuf;
532 		uint64_t txg = dr->dr_txg;
533 
534 		if (db->db_level != 0)
535 			dnode_undirty_dbufs(&dr->dt.di.dr_children);
536 
537 		mutex_enter(&db->db_mtx);
538 		/* XXX - use dbuf_undirty()? */
539 		list_remove(list, dr);
540 		ASSERT(db->db_last_dirty == dr);
541 		db->db_last_dirty = NULL;
542 		db->db_dirtycnt -= 1;
543 		if (db->db_level == 0) {
544 			ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
545 			    dr->dt.dl.dr_data == db->db_buf);
546 			dbuf_unoverride(dr);
547 		} else {
548 			mutex_destroy(&dr->dt.di.dr_mtx);
549 			list_destroy(&dr->dt.di.dr_children);
550 		}
551 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
552 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
553 	}
554 }
555 
556 static void
557 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
558 {
559 	int txgoff = tx->tx_txg & TXG_MASK;
560 
561 	ASSERT(dmu_tx_is_syncing(tx));
562 
563 	/*
564 	 * Our contents should have been freed in dnode_sync() by the
565 	 * free range record inserted by the caller of dnode_free().
566 	 */
567 	ASSERT0(DN_USED_BYTES(dn->dn_phys));
568 	ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
569 
570 	dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
571 	dnode_evict_dbufs(dn);
572 
573 	/*
574 	 * XXX - It would be nice to assert this, but we may still
575 	 * have residual holds from async evictions from the arc...
576 	 *
577 	 * zfs_obj_to_path() also depends on this being
578 	 * commented out.
579 	 *
580 	 * ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 1);
581 	 */
582 
583 	/* Undirty next bits */
584 	dn->dn_next_nlevels[txgoff] = 0;
585 	dn->dn_next_indblkshift[txgoff] = 0;
586 	dn->dn_next_blksz[txgoff] = 0;
587 	dn->dn_next_maxblkid[txgoff] = 0;
588 
589 	/* ASSERT(blkptrs are zero); */
590 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
591 	ASSERT(dn->dn_type != DMU_OT_NONE);
592 
593 	ASSERT(dn->dn_free_txg > 0);
594 	if (dn->dn_allocated_txg != dn->dn_free_txg)
595 		dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
596 	bzero(dn->dn_phys, sizeof (dnode_phys_t) * dn->dn_num_slots);
597 	dnode_free_interior_slots(dn);
598 
599 	mutex_enter(&dn->dn_mtx);
600 	dn->dn_type = DMU_OT_NONE;
601 	dn->dn_maxblkid = 0;
602 	dn->dn_allocated_txg = 0;
603 	dn->dn_free_txg = 0;
604 	dn->dn_have_spill = B_FALSE;
605 	dn->dn_num_slots = 1;
606 	mutex_exit(&dn->dn_mtx);
607 
608 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
609 
610 	dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
611 	/*
612 	 * Now that we've released our hold, the dnode may
613 	 * be evicted, so we mustn't access it.
614 	 */
615 }
616 
617 /*
618  * Write out the dnode's dirty buffers.
619  */
620 void
621 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
622 {
623 	objset_t *os = dn->dn_objset;
624 	dnode_phys_t *dnp = dn->dn_phys;
625 	int txgoff = tx->tx_txg & TXG_MASK;
626 	list_t *list = &dn->dn_dirty_records[txgoff];
627 	static const dnode_phys_t zerodn = { 0 };
628 	boolean_t kill_spill = B_FALSE;
629 
630 	ASSERT(dmu_tx_is_syncing(tx));
631 	ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
632 	ASSERT(dnp->dn_type != DMU_OT_NONE ||
633 	    bcmp(dnp, &zerodn, DNODE_MIN_SIZE) == 0);
634 	DNODE_VERIFY(dn);
635 
636 	ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
637 
638 	/*
639 	 * Do user accounting if it is enabled and this is not
640 	 * an encrypted receive.
641 	 */
642 	if (dmu_objset_userused_enabled(os) &&
643 	    !DMU_OBJECT_IS_SPECIAL(dn->dn_object) &&
644 	    (!os->os_encrypted || !dmu_objset_is_receiving(os))) {
645 		mutex_enter(&dn->dn_mtx);
646 		dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
647 		dn->dn_oldflags = dn->dn_phys->dn_flags;
648 		dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
649 		if (dmu_objset_userobjused_enabled(dn->dn_objset))
650 			dn->dn_phys->dn_flags |=
651 			    DNODE_FLAG_USEROBJUSED_ACCOUNTED;
652 		mutex_exit(&dn->dn_mtx);
653 		dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
654 	} else {
655 		/* Once we account for it, we should always account for it */
656 		ASSERT(!(dn->dn_phys->dn_flags &
657 		    DNODE_FLAG_USERUSED_ACCOUNTED));
658 		ASSERT(!(dn->dn_phys->dn_flags &
659 		    DNODE_FLAG_USEROBJUSED_ACCOUNTED));
660 	}
661 
662 	mutex_enter(&dn->dn_mtx);
663 	if (dn->dn_allocated_txg == tx->tx_txg) {
664 		/* The dnode is newly allocated or reallocated */
665 		if (dnp->dn_type == DMU_OT_NONE) {
666 			/* this is a first alloc, not a realloc */
667 			dnp->dn_nlevels = 1;
668 			dnp->dn_nblkptr = dn->dn_nblkptr;
669 		}
670 
671 		dnp->dn_type = dn->dn_type;
672 		dnp->dn_bonustype = dn->dn_bonustype;
673 		dnp->dn_bonuslen = dn->dn_bonuslen;
674 	}
675 
676 	dnp->dn_extra_slots = dn->dn_num_slots - 1;
677 
678 	ASSERT(dnp->dn_nlevels > 1 ||
679 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
680 	    BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
681 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
682 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
683 	ASSERT(dnp->dn_nlevels < 2 ||
684 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
685 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
686 
687 	if (dn->dn_next_type[txgoff] != 0) {
688 		dnp->dn_type = dn->dn_type;
689 		dn->dn_next_type[txgoff] = 0;
690 	}
691 
692 	if (dn->dn_next_blksz[txgoff] != 0) {
693 		ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
694 		    SPA_MINBLOCKSIZE) == 0);
695 		ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
696 		    dn->dn_maxblkid == 0 || list_head(list) != NULL ||
697 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
698 		    dnp->dn_datablkszsec ||
699 		    !range_tree_is_empty(dn->dn_free_ranges[txgoff]));
700 		dnp->dn_datablkszsec =
701 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
702 		dn->dn_next_blksz[txgoff] = 0;
703 	}
704 
705 	if (dn->dn_next_bonuslen[txgoff] != 0) {
706 		if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
707 			dnp->dn_bonuslen = 0;
708 		else
709 			dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
710 		ASSERT(dnp->dn_bonuslen <=
711 		    DN_SLOTS_TO_BONUSLEN(dnp->dn_extra_slots + 1));
712 		dn->dn_next_bonuslen[txgoff] = 0;
713 	}
714 
715 	if (dn->dn_next_bonustype[txgoff] != 0) {
716 		ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
717 		dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
718 		dn->dn_next_bonustype[txgoff] = 0;
719 	}
720 
721 	boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
722 	    dn->dn_free_txg <= tx->tx_txg;
723 
724 	/*
725 	 * Remove the spill block if we have been explicitly asked to
726 	 * remove it, or if the object is being removed.
727 	 */
728 	if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
729 		if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
730 			kill_spill = B_TRUE;
731 		dn->dn_rm_spillblk[txgoff] = 0;
732 	}
733 
734 	if (dn->dn_next_indblkshift[txgoff] != 0) {
735 		ASSERT(dnp->dn_nlevels == 1);
736 		dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
737 		dn->dn_next_indblkshift[txgoff] = 0;
738 	}
739 
740 	/*
741 	 * Just take the live (open-context) values for checksum and compress.
742 	 * Strictly speaking it's a future leak, but nothing bad happens if we
743 	 * start using the new checksum or compress algorithm a little early.
744 	 */
745 	dnp->dn_checksum = dn->dn_checksum;
746 	dnp->dn_compress = dn->dn_compress;
747 
748 	mutex_exit(&dn->dn_mtx);
749 
750 	if (kill_spill) {
751 		free_blocks(dn, DN_SPILL_BLKPTR(dn->dn_phys), 1, tx);
752 		mutex_enter(&dn->dn_mtx);
753 		dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
754 		mutex_exit(&dn->dn_mtx);
755 	}
756 
757 	/* process all the "freed" ranges in the file */
758 	if (dn->dn_free_ranges[txgoff] != NULL) {
759 		dnode_sync_free_range_arg_t dsfra;
760 		dsfra.dsfra_dnode = dn;
761 		dsfra.dsfra_tx = tx;
762 		dsfra.dsfra_free_indirects = freeing_dnode;
763 		mutex_enter(&dn->dn_mtx);
764 		if (freeing_dnode) {
765 			ASSERT(range_tree_contains(dn->dn_free_ranges[txgoff],
766 			    0, dn->dn_maxblkid + 1));
767 		}
768 		/*
769 		 * Because dnode_sync_free_range() must drop dn_mtx during its
770 		 * processing, using it as a callback to range_tree_vacate() is
771 		 * not safe.  No other operations (besides destroy) are allowed
772 		 * once range_tree_vacate() has begun, and dropping dn_mtx
773 		 * would leave a window open for another thread to observe that
774 		 * invalid (and unsafe) state.
775 		 */
776 		range_tree_walk(dn->dn_free_ranges[txgoff],
777 		    dnode_sync_free_range, &dsfra);
778 		range_tree_vacate(dn->dn_free_ranges[txgoff], NULL, NULL);
779 		range_tree_destroy(dn->dn_free_ranges[txgoff]);
780 		dn->dn_free_ranges[txgoff] = NULL;
781 		mutex_exit(&dn->dn_mtx);
782 	}
783 
784 	if (freeing_dnode) {
785 		dn->dn_objset->os_freed_dnodes++;
786 		dnode_sync_free(dn, tx);
787 		return;
788 	}
789 
790 	if (dn->dn_num_slots > DNODE_MIN_SLOTS) {
791 		dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
792 		mutex_enter(&ds->ds_lock);
793 		ds->ds_feature_activation_needed[SPA_FEATURE_LARGE_DNODE] =
794 		    B_TRUE;
795 		mutex_exit(&ds->ds_lock);
796 	}
797 
798 	if (dn->dn_next_nlevels[txgoff]) {
799 		dnode_increase_indirection(dn, tx);
800 		dn->dn_next_nlevels[txgoff] = 0;
801 	}
802 
803 	/*
804 	 * This must be done after dnode_sync_free_range()
805 	 * and dnode_increase_indirection(). See dnode_new_blkid()
806 	 * for an explanation of the high bit being set.
807 	 */
808 	if (dn->dn_next_maxblkid[txgoff]) {
809 		mutex_enter(&dn->dn_mtx);
810 		dnp->dn_maxblkid =
811 		    dn->dn_next_maxblkid[txgoff] & ~DMU_NEXT_MAXBLKID_SET;
812 		dn->dn_next_maxblkid[txgoff] = 0;
813 		mutex_exit(&dn->dn_mtx);
814 	}
815 
816 	if (dn->dn_next_nblkptr[txgoff]) {
817 		/* this should only happen on a realloc */
818 		ASSERT(dn->dn_allocated_txg == tx->tx_txg);
819 		if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
820 			/* zero the new blkptrs we are gaining */
821 			bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
822 			    sizeof (blkptr_t) *
823 			    (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
824 #ifdef ZFS_DEBUG
825 		} else {
826 			int i;
827 			ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
828 			/* the blkptrs we are losing better be unallocated */
829 			for (i = dn->dn_next_nblkptr[txgoff];
830 			    i < dnp->dn_nblkptr; i++)
831 				ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
832 #endif
833 		}
834 		mutex_enter(&dn->dn_mtx);
835 		dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
836 		dn->dn_next_nblkptr[txgoff] = 0;
837 		mutex_exit(&dn->dn_mtx);
838 	}
839 
840 	dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
841 
842 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
843 		ASSERT3P(list_head(list), ==, NULL);
844 		dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
845 	}
846 
847 	/*
848 	 * Although we have dropped our reference to the dnode, it
849 	 * can't be evicted until its written, and we haven't yet
850 	 * initiated the IO for the dnode's dbuf.
851 	 */
852 }
853