xref: /freebsd/sys/contrib/openzfs/module/zfs/dnode_sync.c (revision d8fbbd371ca11d9ad4b29b9d3a316885a5da0b15)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 
23 /*
24  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
25  * Copyright (c) 2012, 2020 by Delphix. All rights reserved.
26  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
27  * Copyright 2020 Oxide Computer Company
28  */
29 
30 #include <sys/zfs_context.h>
31 #include <sys/dbuf.h>
32 #include <sys/dnode.h>
33 #include <sys/dmu.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_recv.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/spa.h>
39 #include <sys/range_tree.h>
40 #include <sys/zfeature.h>
41 
42 static void
dnode_increase_indirection(dnode_t * dn,dmu_tx_t * tx)43 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
44 {
45 	dmu_buf_impl_t *db;
46 	int txgoff = tx->tx_txg & TXG_MASK;
47 	int nblkptr = dn->dn_phys->dn_nblkptr;
48 	int old_toplvl = dn->dn_phys->dn_nlevels - 1;
49 	int new_level = dn->dn_next_nlevels[txgoff];
50 	int i;
51 
52 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
53 
54 	/* this dnode can't be paged out because it's dirty */
55 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
56 	ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
57 
58 	db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
59 	ASSERT(db != NULL);
60 
61 	dn->dn_phys->dn_nlevels = new_level;
62 	dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
63 	    (u_longlong_t)dn->dn_object, dn->dn_phys->dn_nlevels);
64 
65 	/*
66 	 * Lock ordering requires that we hold the children's db_mutexes (by
67 	 * calling dbuf_find()) before holding the parent's db_rwlock.  The lock
68 	 * order is imposed by dbuf_read's steps of "grab the lock to protect
69 	 * db_parent, get db_parent, hold db_parent's db_rwlock".
70 	 */
71 	dmu_buf_impl_t *children[DN_MAX_NBLKPTR];
72 	ASSERT3U(nblkptr, <=, DN_MAX_NBLKPTR);
73 	for (i = 0; i < nblkptr; i++) {
74 		children[i] = dbuf_find(dn->dn_objset, dn->dn_object,
75 		    old_toplvl, i, NULL);
76 	}
77 
78 	/* transfer dnode's block pointers to new indirect block */
79 	(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
80 	if (dn->dn_dbuf != NULL)
81 		rw_enter(&dn->dn_dbuf->db_rwlock, RW_WRITER);
82 	rw_enter(&db->db_rwlock, RW_WRITER);
83 	ASSERT(db->db.db_data);
84 	ASSERT(arc_released(db->db_buf));
85 	ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
86 	memcpy(db->db.db_data, dn->dn_phys->dn_blkptr,
87 	    sizeof (blkptr_t) * nblkptr);
88 	arc_buf_freeze(db->db_buf);
89 
90 	/* set dbuf's parent pointers to new indirect buf */
91 	for (i = 0; i < nblkptr; i++) {
92 		dmu_buf_impl_t *child = children[i];
93 
94 		if (child == NULL)
95 			continue;
96 #ifdef	ZFS_DEBUG
97 		DB_DNODE_ENTER(child);
98 		ASSERT3P(DB_DNODE(child), ==, dn);
99 		DB_DNODE_EXIT(child);
100 #endif	/* DEBUG */
101 		if (child->db_parent && child->db_parent != dn->dn_dbuf) {
102 			ASSERT(child->db_parent->db_level == db->db_level);
103 			ASSERT(child->db_blkptr !=
104 			    &dn->dn_phys->dn_blkptr[child->db_blkid]);
105 			mutex_exit(&child->db_mtx);
106 			continue;
107 		}
108 		ASSERT(child->db_parent == NULL ||
109 		    child->db_parent == dn->dn_dbuf);
110 
111 		child->db_parent = db;
112 		dbuf_add_ref(db, child);
113 		if (db->db.db_data)
114 			child->db_blkptr = (blkptr_t *)db->db.db_data + i;
115 		else
116 			child->db_blkptr = NULL;
117 		dprintf_dbuf_bp(child, child->db_blkptr,
118 		    "changed db_blkptr to new indirect %s", "");
119 
120 		mutex_exit(&child->db_mtx);
121 	}
122 
123 	memset(dn->dn_phys->dn_blkptr, 0, sizeof (blkptr_t) * nblkptr);
124 
125 	rw_exit(&db->db_rwlock);
126 	if (dn->dn_dbuf != NULL)
127 		rw_exit(&dn->dn_dbuf->db_rwlock);
128 
129 	dbuf_rele(db, FTAG);
130 
131 	rw_exit(&dn->dn_struct_rwlock);
132 }
133 
134 static void
free_blocks(dnode_t * dn,blkptr_t * bp,int num,dmu_tx_t * tx)135 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
136 {
137 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
138 	uint64_t bytesfreed = 0;
139 
140 	dprintf("ds=%p obj=%llx num=%d\n", ds, (u_longlong_t)dn->dn_object,
141 	    num);
142 
143 	for (int i = 0; i < num; i++, bp++) {
144 		if (BP_IS_HOLE(bp))
145 			continue;
146 
147 		bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
148 		ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
149 
150 		/*
151 		 * Save some useful information on the holes being
152 		 * punched, including logical size, type, and indirection
153 		 * level. Retaining birth time enables detection of when
154 		 * holes are punched for reducing the number of free
155 		 * records transmitted during a zfs send.
156 		 */
157 
158 		uint64_t lsize = BP_GET_LSIZE(bp);
159 		dmu_object_type_t type = BP_GET_TYPE(bp);
160 		uint64_t lvl = BP_GET_LEVEL(bp);
161 
162 		memset(bp, 0, sizeof (blkptr_t));
163 
164 		if (spa_feature_is_active(dn->dn_objset->os_spa,
165 		    SPA_FEATURE_HOLE_BIRTH)) {
166 			BP_SET_LSIZE(bp, lsize);
167 			BP_SET_TYPE(bp, type);
168 			BP_SET_LEVEL(bp, lvl);
169 			BP_SET_BIRTH(bp, dmu_tx_get_txg(tx), 0);
170 		}
171 	}
172 	dnode_diduse_space(dn, -bytesfreed);
173 }
174 
175 #ifdef ZFS_DEBUG
176 static void
free_verify(dmu_buf_impl_t * db,uint64_t start,uint64_t end,dmu_tx_t * tx)177 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
178 {
179 	uint64_t off, num, i, j;
180 	unsigned int epbs;
181 	int err;
182 	uint64_t txg = tx->tx_txg;
183 	dnode_t *dn;
184 
185 	DB_DNODE_ENTER(db);
186 	dn = DB_DNODE(db);
187 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
188 	off = start - (db->db_blkid << epbs);
189 	num = end - start + 1;
190 
191 	ASSERT3U(dn->dn_phys->dn_indblkshift, >=, SPA_BLKPTRSHIFT);
192 	ASSERT3U(end + 1, >=, start);
193 	ASSERT3U(start, >=, (db->db_blkid << epbs));
194 	ASSERT3U(db->db_level, >, 0);
195 	ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
196 	ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
197 	ASSERT(db->db_blkptr != NULL);
198 
199 	for (i = off; i < off+num; i++) {
200 		uint64_t *buf;
201 		dmu_buf_impl_t *child;
202 		dbuf_dirty_record_t *dr;
203 
204 		ASSERT(db->db_level == 1);
205 
206 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
207 		err = dbuf_hold_impl(dn, db->db_level - 1,
208 		    (db->db_blkid << epbs) + i, TRUE, FALSE, FTAG, &child);
209 		rw_exit(&dn->dn_struct_rwlock);
210 		if (err == ENOENT)
211 			continue;
212 		ASSERT0(err);
213 		ASSERT0(child->db_level);
214 		dr = dbuf_find_dirty_eq(child, 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=%llu off=%llu "
223 					    "num=%llu\n",
224 					    (void *)child, (u_longlong_t)i,
225 					    (u_longlong_t)off,
226 					    (u_longlong_t)num);
227 				}
228 			}
229 		}
230 
231 		/*
232 		 * db_data better be zeroed unless it's dirty in a
233 		 * future txg.
234 		 */
235 		mutex_enter(&child->db_mtx);
236 		buf = child->db.db_data;
237 		if (buf != NULL && child->db_state != DB_FILL &&
238 		    list_is_empty(&child->db_dirty_records)) {
239 			for (j = 0; j < child->db.db_size >> 3; j++) {
240 				if (buf[j] != 0) {
241 					panic("freed data not zero: "
242 					    "child=%p i=%llu off=%llu "
243 					    "num=%llu\n",
244 					    (void *)child, (u_longlong_t)i,
245 					    (u_longlong_t)off,
246 					    (u_longlong_t)num);
247 				}
248 			}
249 		}
250 		mutex_exit(&child->db_mtx);
251 
252 		dbuf_rele(child, FTAG);
253 	}
254 	DB_DNODE_EXIT(db);
255 }
256 #endif
257 
258 /*
259  * We don't usually free the indirect blocks here.  If in one txg we have a
260  * free_range and a write to the same indirect block, it's important that we
261  * preserve the hole's birth times. Therefore, we don't free any any indirect
262  * blocks in free_children().  If an indirect block happens to turn into all
263  * holes, it will be freed by dbuf_write_children_ready, which happens at a
264  * point in the syncing process where we know for certain the contents of the
265  * indirect block.
266  *
267  * However, if we're freeing a dnode, its space accounting must go to zero
268  * before we actually try to free the dnode, or we will trip an assertion. In
269  * addition, we know the case described above cannot occur, because the dnode is
270  * being freed.  Therefore, we free the indirect blocks immediately in that
271  * case.
272  */
273 static void
free_children(dmu_buf_impl_t * db,uint64_t blkid,uint64_t nblks,boolean_t free_indirects,dmu_tx_t * tx)274 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
275     boolean_t free_indirects, dmu_tx_t *tx)
276 {
277 	dnode_t *dn;
278 	blkptr_t *bp;
279 	dmu_buf_impl_t *subdb;
280 	uint64_t start, end, dbstart, dbend;
281 	unsigned int epbs, shift, i;
282 
283 	/*
284 	 * There is a small possibility that this block will not be cached:
285 	 *   1 - if level > 1 and there are no children with level <= 1
286 	 *   2 - if this block was evicted since we read it from
287 	 *	 dmu_tx_hold_free().
288 	 */
289 	if (db->db_state != DB_CACHED)
290 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
291 
292 	/*
293 	 * If we modify this indirect block, and we are not freeing the
294 	 * dnode (!free_indirects), then this indirect block needs to get
295 	 * written to disk by dbuf_write().  If it is dirty, we know it will
296 	 * be written (otherwise, we would have incorrect on-disk state
297 	 * because the space would be freed but still referenced by the BP
298 	 * in this indirect block).  Therefore we VERIFY that it is
299 	 * dirty.
300 	 *
301 	 * Our VERIFY covers some cases that do not actually have to be
302 	 * dirty, but the open-context code happens to dirty.  E.g. if the
303 	 * blocks we are freeing are all holes, because in that case, we
304 	 * are only freeing part of this indirect block, so it is an
305 	 * ancestor of the first or last block to be freed.  The first and
306 	 * last L1 indirect blocks are always dirtied by dnode_free_range().
307 	 */
308 	db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
309 	VERIFY(BP_GET_FILL(db->db_blkptr) == 0 || db->db_dirtycnt > 0);
310 	dmu_buf_unlock_parent(db, dblt, FTAG);
311 
312 	dbuf_release_bp(db);
313 	bp = db->db.db_data;
314 
315 	DB_DNODE_ENTER(db);
316 	dn = DB_DNODE(db);
317 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
318 	ASSERT3U(epbs, <, 31);
319 	shift = (db->db_level - 1) * epbs;
320 	dbstart = db->db_blkid << epbs;
321 	start = blkid >> shift;
322 	if (dbstart < start) {
323 		bp += start - dbstart;
324 	} else {
325 		start = dbstart;
326 	}
327 	dbend = ((db->db_blkid + 1) << epbs) - 1;
328 	end = (blkid + nblks - 1) >> shift;
329 	if (dbend <= end)
330 		end = dbend;
331 
332 	ASSERT3U(start, <=, end);
333 
334 	if (db->db_level == 1) {
335 		FREE_VERIFY(db, start, end, tx);
336 		rw_enter(&db->db_rwlock, RW_WRITER);
337 		free_blocks(dn, bp, end - start + 1, tx);
338 		rw_exit(&db->db_rwlock);
339 	} else {
340 		for (uint64_t id = start; id <= end; id++, bp++) {
341 			if (BP_IS_HOLE(bp))
342 				continue;
343 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
344 			VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
345 			    id, TRUE, FALSE, FTAG, &subdb));
346 			rw_exit(&dn->dn_struct_rwlock);
347 			ASSERT3P(bp, ==, subdb->db_blkptr);
348 
349 			free_children(subdb, blkid, nblks, free_indirects, tx);
350 			dbuf_rele(subdb, FTAG);
351 		}
352 	}
353 
354 	if (free_indirects) {
355 		rw_enter(&db->db_rwlock, RW_WRITER);
356 		for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++)
357 			ASSERT(BP_IS_HOLE(bp));
358 		memset(db->db.db_data, 0, db->db.db_size);
359 		free_blocks(dn, db->db_blkptr, 1, tx);
360 		rw_exit(&db->db_rwlock);
361 	}
362 
363 	DB_DNODE_EXIT(db);
364 	arc_buf_freeze(db->db_buf);
365 }
366 
367 /*
368  * Traverse the indicated range of the provided file
369  * and "free" all the blocks contained there.
370  */
371 static void
dnode_sync_free_range_impl(dnode_t * dn,uint64_t blkid,uint64_t nblks,boolean_t free_indirects,dmu_tx_t * tx)372 dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
373     boolean_t free_indirects, dmu_tx_t *tx)
374 {
375 	blkptr_t *bp = dn->dn_phys->dn_blkptr;
376 	int dnlevel = dn->dn_phys->dn_nlevels;
377 	boolean_t trunc = B_FALSE;
378 
379 	if (blkid > dn->dn_phys->dn_maxblkid)
380 		return;
381 
382 	ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
383 	if (blkid + nblks > dn->dn_phys->dn_maxblkid) {
384 		nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
385 		trunc = B_TRUE;
386 	}
387 
388 	/* There are no indirect blocks in the object */
389 	if (dnlevel == 1) {
390 		if (blkid >= dn->dn_phys->dn_nblkptr) {
391 			/* this range was never made persistent */
392 			return;
393 		}
394 		ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
395 		free_blocks(dn, bp + blkid, nblks, tx);
396 	} else {
397 		int shift = (dnlevel - 1) *
398 		    (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
399 		int start = blkid >> shift;
400 		int end = (blkid + nblks - 1) >> shift;
401 		dmu_buf_impl_t *db;
402 
403 		ASSERT(start < dn->dn_phys->dn_nblkptr);
404 		bp += start;
405 		for (int i = start; i <= end; i++, bp++) {
406 			if (BP_IS_HOLE(bp))
407 				continue;
408 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
409 			VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
410 			    TRUE, FALSE, FTAG, &db));
411 			rw_exit(&dn->dn_struct_rwlock);
412 			free_children(db, blkid, nblks, free_indirects, tx);
413 			dbuf_rele(db, FTAG);
414 		}
415 	}
416 
417 	/*
418 	 * Do not truncate the maxblkid if we are performing a raw
419 	 * receive. The raw receive sets the maxblkid manually and
420 	 * must not be overridden. Usually, the last DRR_FREE record
421 	 * will be at the maxblkid, because the source system sets
422 	 * the maxblkid when truncating. However, if the last block
423 	 * was freed by overwriting with zeros and being compressed
424 	 * away to a hole, the source system will generate a DRR_FREE
425 	 * record while leaving the maxblkid after the end of that
426 	 * record. In this case we need to leave the maxblkid as
427 	 * indicated in the DRR_OBJECT record, so that it matches the
428 	 * source system, ensuring that the cryptographic hashes will
429 	 * match.
430 	 */
431 	if (trunc && !dn->dn_objset->os_raw_receive) {
432 		uint64_t off __maybe_unused;
433 		dn->dn_phys->dn_maxblkid = blkid == 0 ? 0 : blkid - 1;
434 
435 		off = (dn->dn_phys->dn_maxblkid + 1) *
436 		    (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
437 		ASSERT(off < dn->dn_phys->dn_maxblkid ||
438 		    dn->dn_phys->dn_maxblkid == 0 ||
439 		    dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
440 	}
441 }
442 
443 /*
444  * Try to kick all the dnode's dbufs out of the cache...
445  */
446 void
dnode_evict_dbufs(dnode_t * dn)447 dnode_evict_dbufs(dnode_t *dn)
448 {
449 	dmu_buf_impl_t *db_marker;
450 	dmu_buf_impl_t *db, *db_next;
451 
452 	db_marker = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
453 
454 	mutex_enter(&dn->dn_dbufs_mtx);
455 	for (db = avl_first(&dn->dn_dbufs); db != NULL; db = db_next) {
456 
457 #ifdef	ZFS_DEBUG
458 		DB_DNODE_ENTER(db);
459 		ASSERT3P(DB_DNODE(db), ==, dn);
460 		DB_DNODE_EXIT(db);
461 #endif	/* DEBUG */
462 
463 		mutex_enter(&db->db_mtx);
464 		if (db->db_state != DB_EVICTING &&
465 		    zfs_refcount_is_zero(&db->db_holds)) {
466 			db_marker->db_level = db->db_level;
467 			db_marker->db_blkid = db->db_blkid;
468 			/*
469 			 * Insert a MARKER node with the same level and blkid.
470 			 * And to resolve any ties in dbuf_compare() use the
471 			 * pointer of the dbuf that we are evicting. Pass the
472 			 * address in db_parent.
473 			 */
474 			db_marker->db_state = DB_MARKER;
475 			db_marker->db_parent = (void *)((uintptr_t)db - 1);
476 			avl_insert_here(&dn->dn_dbufs, db_marker, db,
477 			    AVL_BEFORE);
478 
479 			/*
480 			 * We need to use the "marker" dbuf rather than
481 			 * simply getting the next dbuf, because
482 			 * dbuf_destroy() may actually remove multiple dbufs.
483 			 * It can call itself recursively on the parent dbuf,
484 			 * which may also be removed from dn_dbufs.  The code
485 			 * flow would look like:
486 			 *
487 			 * dbuf_destroy():
488 			 *   dnode_rele_and_unlock(parent_dbuf, evicting=TRUE):
489 			 *	if (!cacheable || pending_evict)
490 			 *	  dbuf_destroy()
491 			 */
492 			dbuf_destroy(db);
493 
494 			db_next = AVL_NEXT(&dn->dn_dbufs, db_marker);
495 			avl_remove(&dn->dn_dbufs, db_marker);
496 		} else {
497 			db->db_pending_evict = TRUE;
498 			db->db_partial_read = FALSE;
499 			mutex_exit(&db->db_mtx);
500 			db_next = AVL_NEXT(&dn->dn_dbufs, db);
501 		}
502 	}
503 	mutex_exit(&dn->dn_dbufs_mtx);
504 
505 	kmem_free(db_marker, sizeof (dmu_buf_impl_t));
506 
507 	dnode_evict_bonus(dn);
508 }
509 
510 void
dnode_evict_bonus(dnode_t * dn)511 dnode_evict_bonus(dnode_t *dn)
512 {
513 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
514 	if (dn->dn_bonus != NULL) {
515 		if (zfs_refcount_is_zero(&dn->dn_bonus->db_holds)) {
516 			mutex_enter(&dn->dn_bonus->db_mtx);
517 			dbuf_destroy(dn->dn_bonus);
518 			dn->dn_bonus = NULL;
519 		} else {
520 			dn->dn_bonus->db_pending_evict = TRUE;
521 		}
522 	}
523 	rw_exit(&dn->dn_struct_rwlock);
524 }
525 
526 static void
dnode_undirty_dbufs(list_t * list)527 dnode_undirty_dbufs(list_t *list)
528 {
529 	dbuf_dirty_record_t *dr;
530 
531 	while ((dr = list_head(list))) {
532 		dmu_buf_impl_t *db = dr->dr_dbuf;
533 		uint64_t txg = dr->dr_txg;
534 
535 		if (db->db_level != 0)
536 			dnode_undirty_dbufs(&dr->dt.di.dr_children);
537 
538 		mutex_enter(&db->db_mtx);
539 		/* XXX - use dbuf_undirty()? */
540 		list_remove(list, dr);
541 		ASSERT(list_head(&db->db_dirty_records) == dr);
542 		list_remove_head(&db->db_dirty_records);
543 		ASSERT(list_is_empty(&db->db_dirty_records));
544 		db->db_dirtycnt -= 1;
545 		if (db->db_level == 0) {
546 			ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
547 			    dr->dt.dl.dr_data == db->db_buf);
548 			dbuf_unoverride(dr);
549 		} else {
550 			mutex_destroy(&dr->dt.di.dr_mtx);
551 			list_destroy(&dr->dt.di.dr_children);
552 		}
553 		kmem_cache_free(dbuf_dirty_kmem_cache, dr);
554 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
555 	}
556 }
557 
558 static void
dnode_sync_free(dnode_t * dn,dmu_tx_t * tx)559 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
560 {
561 	int txgoff = tx->tx_txg & TXG_MASK;
562 
563 	ASSERT(dmu_tx_is_syncing(tx));
564 
565 	/*
566 	 * Our contents should have been freed in dnode_sync() by the
567 	 * free range record inserted by the caller of dnode_free().
568 	 */
569 	ASSERT0(DN_USED_BYTES(dn->dn_phys));
570 	ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
571 
572 	dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
573 	dnode_evict_dbufs(dn);
574 
575 	/*
576 	 * XXX - It would be nice to assert this, but we may still
577 	 * have residual holds from async evictions from the arc...
578 	 *
579 	 * zfs_obj_to_path() also depends on this being
580 	 * commented out.
581 	 *
582 	 * ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 1);
583 	 */
584 
585 	/* Undirty next bits */
586 	dn->dn_next_nlevels[txgoff] = 0;
587 	dn->dn_next_indblkshift[txgoff] = 0;
588 	dn->dn_next_blksz[txgoff] = 0;
589 	dn->dn_next_maxblkid[txgoff] = 0;
590 
591 	/* ASSERT(blkptrs are zero); */
592 	ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
593 	ASSERT(dn->dn_type != DMU_OT_NONE);
594 
595 	ASSERT(dn->dn_free_txg > 0);
596 	if (dn->dn_allocated_txg != dn->dn_free_txg)
597 		dmu_buf_will_dirty(&dn->dn_dbuf->db, tx);
598 	memset(dn->dn_phys, 0, sizeof (dnode_phys_t) * dn->dn_num_slots);
599 	dnode_free_interior_slots(dn);
600 
601 	mutex_enter(&dn->dn_mtx);
602 	dn->dn_type = DMU_OT_NONE;
603 	dn->dn_maxblkid = 0;
604 	dn->dn_allocated_txg = 0;
605 	dn->dn_free_txg = 0;
606 	dn->dn_have_spill = B_FALSE;
607 	dn->dn_num_slots = 1;
608 	mutex_exit(&dn->dn_mtx);
609 
610 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
611 
612 	dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
613 	/*
614 	 * Now that we've released our hold, the dnode may
615 	 * be evicted, so we mustn't access it.
616 	 */
617 }
618 
619 /*
620  * We cannot simply detach the range tree (set dn_free_ranges to NULL)
621  * before processing it because dnode_block_freed() relies on it to
622  * correctly identify blocks that have been freed in the current TXG
623  * (for dbuf_read() calls on holes). If we detached it early, a concurrent
624  * reader might see the block as valid on disk and return stale data
625  * instead of zeros.
626  *
627  * We also can't use zfs_range_tree_walk() nor zfs_range_tree_vacate()
628  * with a callback that drops dn_mtx (dnode_sync_free_range()). This is
629  * unsafe because another thread (spa_sync_deferred_frees() ->
630  * dnode_free_range()) could acquire dn_mtx and modify the tree while the
631  * walk or vacate was in progress. This leads to tree corruption or panic
632  * when we resume.
633  *
634  * To fix the race while maintaining visibility, we process the tree
635  * incrementally. We pick a segment, drop the lock to sync it, and
636  * re-acquire the lock to remove it. By always restarting from the head
637  * of the tree, we ensure we are never using an invalid iterator.
638  * We use zfs_range_tree_clear() instead of ..._remove() because the range
639  * might have already been removed while the lock was dropped (specifically
640  * in the dbuf_dirty path mentioned above). ..._clear() handles this
641  * gracefully, while ..._remove() would panic on a missing segment.
642  */
643 static void
dnode_sync_free_ranges(dnode_t * dn,dmu_tx_t * tx)644 dnode_sync_free_ranges(dnode_t *dn, dmu_tx_t *tx)
645 {
646 	int txgoff = tx->tx_txg & TXG_MASK;
647 
648 	mutex_enter(&dn->dn_mtx);
649 	zfs_range_tree_t *rt = dn->dn_free_ranges[txgoff];
650 	if (rt != NULL) {
651 		boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
652 		    dn->dn_free_txg <= tx->tx_txg;
653 		zfs_range_seg_t *rs;
654 
655 		if (freeing_dnode) {
656 			ASSERT(zfs_range_tree_contains(rt, 0,
657 			    dn->dn_maxblkid + 1));
658 		}
659 
660 		while ((rs = zfs_range_tree_first(rt)) != NULL) {
661 			uint64_t start = zfs_rs_get_start(rs, rt);
662 			uint64_t size = zfs_rs_get_end(rs, rt) - start;
663 
664 			mutex_exit(&dn->dn_mtx);
665 			dnode_sync_free_range_impl(dn, start, size,
666 			    freeing_dnode, tx);
667 			mutex_enter(&dn->dn_mtx);
668 
669 			zfs_range_tree_clear(rt, start, size);
670 		}
671 		zfs_range_tree_destroy(rt);
672 		dn->dn_free_ranges[txgoff] = NULL;
673 	}
674 	mutex_exit(&dn->dn_mtx);
675 }
676 
677 /*
678  * Write out the dnode's dirty buffers.
679  * Does not wait for zio completions.
680  */
681 void
dnode_sync(dnode_t * dn,dmu_tx_t * tx)682 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
683 {
684 	objset_t *os = dn->dn_objset;
685 	dnode_phys_t *dnp = dn->dn_phys;
686 	int txgoff = tx->tx_txg & TXG_MASK;
687 	list_t *list = &dn->dn_dirty_records[txgoff];
688 	static const dnode_phys_t zerodn __maybe_unused = { 0 };
689 	boolean_t kill_spill = B_FALSE;
690 
691 	ASSERT(dmu_tx_is_syncing(tx));
692 	ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
693 	ASSERT(dnp->dn_type != DMU_OT_NONE ||
694 	    memcmp(dnp, &zerodn, DNODE_MIN_SIZE) == 0);
695 	DNODE_VERIFY(dn);
696 
697 	ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
698 
699 	/*
700 	 * Do user accounting if it is enabled and this is not
701 	 * an encrypted receive.
702 	 */
703 	if (dmu_objset_userused_enabled(os) &&
704 	    !DMU_OBJECT_IS_SPECIAL(dn->dn_object) &&
705 	    (!os->os_encrypted || !dmu_objset_is_receiving(os))) {
706 		mutex_enter(&dn->dn_mtx);
707 		dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
708 		dn->dn_oldflags = dn->dn_phys->dn_flags;
709 		dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
710 		if (dmu_objset_userobjused_enabled(dn->dn_objset))
711 			dn->dn_phys->dn_flags |=
712 			    DNODE_FLAG_USEROBJUSED_ACCOUNTED;
713 		mutex_exit(&dn->dn_mtx);
714 		dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
715 	} else if (!(os->os_encrypted && dmu_objset_is_receiving(os))) {
716 		/*
717 		 * Once we account for it, we should always account for it,
718 		 * except for the case of a raw receive. We will not be able
719 		 * to account for it until the receiving dataset has been
720 		 * mounted.
721 		 */
722 		ASSERT(!(dn->dn_phys->dn_flags &
723 		    DNODE_FLAG_USERUSED_ACCOUNTED));
724 		ASSERT(!(dn->dn_phys->dn_flags &
725 		    DNODE_FLAG_USEROBJUSED_ACCOUNTED));
726 	}
727 
728 	mutex_enter(&dn->dn_mtx);
729 	if (dn->dn_allocated_txg == tx->tx_txg) {
730 		/* The dnode is newly allocated or reallocated */
731 		if (dnp->dn_type == DMU_OT_NONE) {
732 			/* this is a first alloc, not a realloc */
733 			dnp->dn_nlevels = 1;
734 			dnp->dn_nblkptr = dn->dn_nblkptr;
735 		}
736 
737 		dnp->dn_type = dn->dn_type;
738 		dnp->dn_bonustype = dn->dn_bonustype;
739 		dnp->dn_bonuslen = dn->dn_bonuslen;
740 	}
741 
742 	dnp->dn_extra_slots = dn->dn_num_slots - 1;
743 
744 	ASSERT(dnp->dn_nlevels > 1 ||
745 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
746 	    BP_IS_EMBEDDED(&dnp->dn_blkptr[0]) ||
747 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
748 	    dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
749 	ASSERT(dnp->dn_nlevels < 2 ||
750 	    BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
751 	    BP_GET_LSIZE(&dnp->dn_blkptr[0]) == 1 << dnp->dn_indblkshift);
752 
753 	if (dn->dn_next_type[txgoff] != 0) {
754 		dnp->dn_type = dn->dn_type;
755 		dn->dn_next_type[txgoff] = 0;
756 	}
757 
758 	if (dn->dn_next_blksz[txgoff] != 0) {
759 		ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
760 		    SPA_MINBLOCKSIZE) == 0);
761 		ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
762 		    dn->dn_maxblkid == 0 || list_head(list) != NULL ||
763 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
764 		    dnp->dn_datablkszsec ||
765 		    !zfs_range_tree_is_empty(dn->dn_free_ranges[txgoff]));
766 		dnp->dn_datablkszsec =
767 		    dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
768 		dn->dn_next_blksz[txgoff] = 0;
769 	}
770 
771 	if (dn->dn_next_bonuslen[txgoff] != 0) {
772 		if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
773 			dnp->dn_bonuslen = 0;
774 		else
775 			dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
776 		ASSERT(dnp->dn_bonuslen <=
777 		    DN_SLOTS_TO_BONUSLEN(dnp->dn_extra_slots + 1));
778 		dn->dn_next_bonuslen[txgoff] = 0;
779 	}
780 
781 	if (dn->dn_next_bonustype[txgoff] != 0) {
782 		ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
783 		dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
784 		dn->dn_next_bonustype[txgoff] = 0;
785 	}
786 
787 	boolean_t freeing_dnode = dn->dn_free_txg > 0 &&
788 	    dn->dn_free_txg <= tx->tx_txg;
789 
790 	/*
791 	 * Remove the spill block if we have been explicitly asked to
792 	 * remove it, or if the object is being removed.
793 	 */
794 	if (dn->dn_rm_spillblk[txgoff] || freeing_dnode) {
795 		if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
796 			kill_spill = B_TRUE;
797 		dn->dn_rm_spillblk[txgoff] = 0;
798 	}
799 
800 	if (dn->dn_next_indblkshift[txgoff] != 0) {
801 		ASSERT(dnp->dn_nlevels == 1);
802 		dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
803 		dn->dn_next_indblkshift[txgoff] = 0;
804 	}
805 
806 	/*
807 	 * Just take the live (open-context) values for checksum and compress.
808 	 * Strictly speaking it's a future leak, but nothing bad happens if we
809 	 * start using the new checksum or compress algorithm a little early.
810 	 */
811 	dnp->dn_checksum = dn->dn_checksum;
812 	dnp->dn_compress = dn->dn_compress;
813 
814 	mutex_exit(&dn->dn_mtx);
815 
816 	if (kill_spill) {
817 		free_blocks(dn, DN_SPILL_BLKPTR(dn->dn_phys), 1, tx);
818 		mutex_enter(&dn->dn_mtx);
819 		dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
820 		mutex_exit(&dn->dn_mtx);
821 	}
822 
823 	/* process all the "freed" ranges in the file */
824 	dnode_sync_free_ranges(dn, tx);
825 
826 	if (freeing_dnode) {
827 		dn->dn_objset->os_freed_dnodes++;
828 		dnode_sync_free(dn, tx);
829 		return;
830 	}
831 
832 	if (dn->dn_num_slots > DNODE_MIN_SLOTS) {
833 		dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
834 		mutex_enter(&ds->ds_lock);
835 		ds->ds_feature_activation[SPA_FEATURE_LARGE_DNODE] =
836 		    (void *)B_TRUE;
837 		mutex_exit(&ds->ds_lock);
838 	}
839 
840 	if (dn->dn_next_nlevels[txgoff]) {
841 		dnode_increase_indirection(dn, tx);
842 		dn->dn_next_nlevels[txgoff] = 0;
843 	}
844 
845 	/*
846 	 * This must be done after dnode_sync_free_ranges()
847 	 * and dnode_increase_indirection(). See dnode_new_blkid()
848 	 * for an explanation of the high bit being set.
849 	 */
850 	if (dn->dn_next_maxblkid[txgoff]) {
851 		mutex_enter(&dn->dn_mtx);
852 		dnp->dn_maxblkid =
853 		    dn->dn_next_maxblkid[txgoff] & ~DMU_NEXT_MAXBLKID_SET;
854 		dn->dn_next_maxblkid[txgoff] = 0;
855 		mutex_exit(&dn->dn_mtx);
856 	}
857 
858 	if (dn->dn_next_nblkptr[txgoff]) {
859 		/* this should only happen on a realloc */
860 		ASSERT(dn->dn_allocated_txg == tx->tx_txg);
861 		if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
862 			/* zero the new blkptrs we are gaining */
863 			memset(dnp->dn_blkptr + dnp->dn_nblkptr, 0,
864 			    sizeof (blkptr_t) *
865 			    (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
866 #ifdef ZFS_DEBUG
867 		} else {
868 			int i;
869 			ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
870 			/* the blkptrs we are losing better be unallocated */
871 			for (i = 0; i < dnp->dn_nblkptr; i++) {
872 				if (i >= dn->dn_next_nblkptr[txgoff])
873 					ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
874 			}
875 #endif
876 		}
877 		mutex_enter(&dn->dn_mtx);
878 		dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
879 		dn->dn_next_nblkptr[txgoff] = 0;
880 		mutex_exit(&dn->dn_mtx);
881 	}
882 
883 	dbuf_sync_list(list, dn->dn_phys->dn_nlevels - 1, tx);
884 
885 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
886 		ASSERT0P(list_head(list));
887 		dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
888 	}
889 
890 	ASSERT3U(dnp->dn_bonuslen, <=, DN_MAX_BONUS_LEN(dnp));
891 
892 	/*
893 	 * Although we have dropped our reference to the dnode, it
894 	 * can't be evicted until its written, and we haven't yet
895 	 * initiated the IO for the dnode's dbuf.  Additionally, the caller
896 	 * has already added a reference to the dnode because it's on the
897 	 * os_synced_dnodes list.
898 	 */
899 }
900