xref: /freebsd/sys/contrib/openzfs/module/zfs/dmu.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
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  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2011, 2020 by Delphix. All rights reserved.
24  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
26  * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved.
27  * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
28  * Copyright (c) 2019 Datto Inc.
29  * Copyright (c) 2019, Klara Inc.
30  * Copyright (c) 2019, Allan Jude
31  * Copyright (c) 2022 Hewlett Packard Enterprise Development LP.
32  * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
33  */
34 
35 #include <sys/dmu.h>
36 #include <sys/dmu_impl.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/dbuf.h>
39 #include <sys/dnode.h>
40 #include <sys/zfs_context.h>
41 #include <sys/dmu_objset.h>
42 #include <sys/dmu_traverse.h>
43 #include <sys/dsl_dataset.h>
44 #include <sys/dsl_dir.h>
45 #include <sys/dsl_pool.h>
46 #include <sys/dsl_synctask.h>
47 #include <sys/dsl_prop.h>
48 #include <sys/dmu_zfetch.h>
49 #include <sys/zfs_ioctl.h>
50 #include <sys/zap.h>
51 #include <sys/zio_checksum.h>
52 #include <sys/zio_compress.h>
53 #include <sys/sa.h>
54 #include <sys/zfeature.h>
55 #include <sys/abd.h>
56 #include <sys/brt.h>
57 #include <sys/trace_zfs.h>
58 #include <sys/zfs_racct.h>
59 #include <sys/zfs_rlock.h>
60 #ifdef _KERNEL
61 #include <sys/vmsystm.h>
62 #include <sys/zfs_znode.h>
63 #endif
64 
65 /*
66  * Enable/disable nopwrite feature.
67  */
68 static int zfs_nopwrite_enabled = 1;
69 
70 /*
71  * Tunable to control percentage of dirtied L1 blocks from frees allowed into
72  * one TXG. After this threshold is crossed, additional dirty blocks from frees
73  * will wait until the next TXG.
74  * A value of zero will disable this throttle.
75  */
76 static uint_t zfs_per_txg_dirty_frees_percent = 30;
77 
78 /*
79  * Enable/disable forcing txg sync when dirty checking for holes with lseek().
80  * By default this is enabled to ensure accurate hole reporting, it can result
81  * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads.
82  * Disabling this option will result in holes never being reported in dirty
83  * files which is always safe.
84  */
85 static int zfs_dmu_offset_next_sync = 1;
86 
87 /*
88  * Limit the amount we can prefetch with one call to this amount.  This
89  * helps to limit the amount of memory that can be used by prefetching.
90  * Larger objects should be prefetched a bit at a time.
91  */
92 #ifdef _ILP32
93 uint_t dmu_prefetch_max = 8 * 1024 * 1024;
94 #else
95 uint_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE;
96 #endif
97 
98 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
99 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "unallocated"		},
100 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "object directory"	},
101 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "object array"		},
102 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "packed nvlist"		},
103 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "packed nvlist size"	},
104 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "bpobj"			},
105 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "bpobj header"		},
106 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "SPA space map header"	},
107 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "SPA space map"		},
108 	{DMU_BSWAP_UINT64, TRUE,  FALSE, TRUE,  "ZIL intent log"	},
109 	{DMU_BSWAP_DNODE,  TRUE,  FALSE, TRUE,  "DMU dnode"		},
110 	{DMU_BSWAP_OBJSET, TRUE,  TRUE,  FALSE, "DMU objset"		},
111 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "DSL directory"		},
112 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL directory child map"},
113 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL dataset snap map"	},
114 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL props"		},
115 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "DSL dataset"		},
116 	{DMU_BSWAP_ZNODE,  TRUE,  FALSE, FALSE, "ZFS znode"		},
117 	{DMU_BSWAP_OLDACL, TRUE,  FALSE, TRUE,  "ZFS V0 ACL"		},
118 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "ZFS plain file"	},
119 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS directory"		},
120 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "ZFS master node"	},
121 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS delete queue"	},
122 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "zvol object"		},
123 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "zvol prop"		},
124 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "other uint8[]"		},
125 	{DMU_BSWAP_UINT64, FALSE, FALSE, TRUE,  "other uint64[]"	},
126 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "other ZAP"		},
127 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "persistent error log"	},
128 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "SPA history"		},
129 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "SPA history offsets"	},
130 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "Pool properties"	},
131 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL permissions"	},
132 	{DMU_BSWAP_ACL,    TRUE,  FALSE, TRUE,  "ZFS ACL"		},
133 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, TRUE,  "ZFS SYSACL"		},
134 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, TRUE,  "FUID table"		},
135 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "FUID table size"	},
136 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL dataset next clones"},
137 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "scan work queue"	},
138 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS user/group/project used" },
139 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS user/group/project quota"},
140 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "snapshot refcount tags"},
141 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "DDT ZAP algorithm"	},
142 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "DDT statistics"	},
143 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, TRUE,	"System attributes"	},
144 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,	"SA master node"	},
145 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,	"SA attr registration"	},
146 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,	"SA attr layouts"	},
147 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "scan translations"	},
148 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "deduplicated block"	},
149 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL deadlist map"	},
150 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "DSL deadlist map hdr"	},
151 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL dir clones"	},
152 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "bpobj subobj"		}
153 };
154 
155 dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
156 	{	byteswap_uint8_array,	"uint8"		},
157 	{	byteswap_uint16_array,	"uint16"	},
158 	{	byteswap_uint32_array,	"uint32"	},
159 	{	byteswap_uint64_array,	"uint64"	},
160 	{	zap_byteswap,		"zap"		},
161 	{	dnode_buf_byteswap,	"dnode"		},
162 	{	dmu_objset_byteswap,	"objset"	},
163 	{	zfs_znode_byteswap,	"znode"		},
164 	{	zfs_oldacl_byteswap,	"oldacl"	},
165 	{	zfs_acl_byteswap,	"acl"		}
166 };
167 
168 int
169 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
170     const void *tag, dmu_buf_t **dbp)
171 {
172 	uint64_t blkid;
173 	dmu_buf_impl_t *db;
174 
175 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
176 	blkid = dbuf_whichblock(dn, 0, offset);
177 	db = dbuf_hold(dn, blkid, tag);
178 	rw_exit(&dn->dn_struct_rwlock);
179 
180 	if (db == NULL) {
181 		*dbp = NULL;
182 		return (SET_ERROR(EIO));
183 	}
184 
185 	*dbp = &db->db;
186 	return (0);
187 }
188 
189 int
190 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
191     const void *tag, dmu_buf_t **dbp)
192 {
193 	dnode_t *dn;
194 	uint64_t blkid;
195 	dmu_buf_impl_t *db;
196 	int err;
197 
198 	err = dnode_hold(os, object, FTAG, &dn);
199 	if (err)
200 		return (err);
201 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
202 	blkid = dbuf_whichblock(dn, 0, offset);
203 	db = dbuf_hold(dn, blkid, tag);
204 	rw_exit(&dn->dn_struct_rwlock);
205 	dnode_rele(dn, FTAG);
206 
207 	if (db == NULL) {
208 		*dbp = NULL;
209 		return (SET_ERROR(EIO));
210 	}
211 
212 	*dbp = &db->db;
213 	return (err);
214 }
215 
216 int
217 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
218     const void *tag, dmu_buf_t **dbp, int flags)
219 {
220 	int err;
221 	int db_flags = DB_RF_CANFAIL;
222 
223 	if (flags & DMU_READ_NO_PREFETCH)
224 		db_flags |= DB_RF_NOPREFETCH;
225 	if (flags & DMU_READ_NO_DECRYPT)
226 		db_flags |= DB_RF_NO_DECRYPT;
227 
228 	err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
229 	if (err == 0) {
230 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
231 		err = dbuf_read(db, NULL, db_flags);
232 		if (err != 0) {
233 			dbuf_rele(db, tag);
234 			*dbp = NULL;
235 		}
236 	}
237 
238 	return (err);
239 }
240 
241 int
242 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
243     const void *tag, dmu_buf_t **dbp, int flags)
244 {
245 	int err;
246 	int db_flags = DB_RF_CANFAIL;
247 
248 	if (flags & DMU_READ_NO_PREFETCH)
249 		db_flags |= DB_RF_NOPREFETCH;
250 	if (flags & DMU_READ_NO_DECRYPT)
251 		db_flags |= DB_RF_NO_DECRYPT;
252 
253 	err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
254 	if (err == 0) {
255 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
256 		err = dbuf_read(db, NULL, db_flags);
257 		if (err != 0) {
258 			dbuf_rele(db, tag);
259 			*dbp = NULL;
260 		}
261 	}
262 
263 	return (err);
264 }
265 
266 int
267 dmu_bonus_max(void)
268 {
269 	return (DN_OLD_MAX_BONUSLEN);
270 }
271 
272 int
273 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
274 {
275 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
276 	dnode_t *dn;
277 	int error;
278 
279 	DB_DNODE_ENTER(db);
280 	dn = DB_DNODE(db);
281 
282 	if (dn->dn_bonus != db) {
283 		error = SET_ERROR(EINVAL);
284 	} else if (newsize < 0 || newsize > db_fake->db_size) {
285 		error = SET_ERROR(EINVAL);
286 	} else {
287 		dnode_setbonuslen(dn, newsize, tx);
288 		error = 0;
289 	}
290 
291 	DB_DNODE_EXIT(db);
292 	return (error);
293 }
294 
295 int
296 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
297 {
298 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
299 	dnode_t *dn;
300 	int error;
301 
302 	DB_DNODE_ENTER(db);
303 	dn = DB_DNODE(db);
304 
305 	if (!DMU_OT_IS_VALID(type)) {
306 		error = SET_ERROR(EINVAL);
307 	} else if (dn->dn_bonus != db) {
308 		error = SET_ERROR(EINVAL);
309 	} else {
310 		dnode_setbonus_type(dn, type, tx);
311 		error = 0;
312 	}
313 
314 	DB_DNODE_EXIT(db);
315 	return (error);
316 }
317 
318 dmu_object_type_t
319 dmu_get_bonustype(dmu_buf_t *db_fake)
320 {
321 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
322 	dnode_t *dn;
323 	dmu_object_type_t type;
324 
325 	DB_DNODE_ENTER(db);
326 	dn = DB_DNODE(db);
327 	type = dn->dn_bonustype;
328 	DB_DNODE_EXIT(db);
329 
330 	return (type);
331 }
332 
333 int
334 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
335 {
336 	dnode_t *dn;
337 	int error;
338 
339 	error = dnode_hold(os, object, FTAG, &dn);
340 	dbuf_rm_spill(dn, tx);
341 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
342 	dnode_rm_spill(dn, tx);
343 	rw_exit(&dn->dn_struct_rwlock);
344 	dnode_rele(dn, FTAG);
345 	return (error);
346 }
347 
348 /*
349  * Lookup and hold the bonus buffer for the provided dnode.  If the dnode
350  * has not yet been allocated a new bonus dbuf a will be allocated.
351  * Returns ENOENT, EIO, or 0.
352  */
353 int dmu_bonus_hold_by_dnode(dnode_t *dn, const void *tag, dmu_buf_t **dbp,
354     uint32_t flags)
355 {
356 	dmu_buf_impl_t *db;
357 	int error;
358 	uint32_t db_flags = DB_RF_MUST_SUCCEED;
359 
360 	if (flags & DMU_READ_NO_PREFETCH)
361 		db_flags |= DB_RF_NOPREFETCH;
362 	if (flags & DMU_READ_NO_DECRYPT)
363 		db_flags |= DB_RF_NO_DECRYPT;
364 
365 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
366 	if (dn->dn_bonus == NULL) {
367 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
368 			rw_exit(&dn->dn_struct_rwlock);
369 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
370 		}
371 		if (dn->dn_bonus == NULL)
372 			dbuf_create_bonus(dn);
373 	}
374 	db = dn->dn_bonus;
375 
376 	/* as long as the bonus buf is held, the dnode will be held */
377 	if (zfs_refcount_add(&db->db_holds, tag) == 1) {
378 		VERIFY(dnode_add_ref(dn, db));
379 		atomic_inc_32(&dn->dn_dbufs_count);
380 	}
381 
382 	/*
383 	 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
384 	 * hold and incrementing the dbuf count to ensure that dnode_move() sees
385 	 * a dnode hold for every dbuf.
386 	 */
387 	rw_exit(&dn->dn_struct_rwlock);
388 
389 	error = dbuf_read(db, NULL, db_flags);
390 	if (error) {
391 		dnode_evict_bonus(dn);
392 		dbuf_rele(db, tag);
393 		*dbp = NULL;
394 		return (error);
395 	}
396 
397 	*dbp = &db->db;
398 	return (0);
399 }
400 
401 int
402 dmu_bonus_hold(objset_t *os, uint64_t object, const void *tag, dmu_buf_t **dbp)
403 {
404 	dnode_t *dn;
405 	int error;
406 
407 	error = dnode_hold(os, object, FTAG, &dn);
408 	if (error)
409 		return (error);
410 
411 	error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH);
412 	dnode_rele(dn, FTAG);
413 
414 	return (error);
415 }
416 
417 /*
418  * returns ENOENT, EIO, or 0.
419  *
420  * This interface will allocate a blank spill dbuf when a spill blk
421  * doesn't already exist on the dnode.
422  *
423  * if you only want to find an already existing spill db, then
424  * dmu_spill_hold_existing() should be used.
425  */
426 int
427 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, const void *tag,
428     dmu_buf_t **dbp)
429 {
430 	dmu_buf_impl_t *db = NULL;
431 	int err;
432 
433 	if ((flags & DB_RF_HAVESTRUCT) == 0)
434 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
435 
436 	db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
437 
438 	if ((flags & DB_RF_HAVESTRUCT) == 0)
439 		rw_exit(&dn->dn_struct_rwlock);
440 
441 	if (db == NULL) {
442 		*dbp = NULL;
443 		return (SET_ERROR(EIO));
444 	}
445 	err = dbuf_read(db, NULL, flags);
446 	if (err == 0)
447 		*dbp = &db->db;
448 	else {
449 		dbuf_rele(db, tag);
450 		*dbp = NULL;
451 	}
452 	return (err);
453 }
454 
455 int
456 dmu_spill_hold_existing(dmu_buf_t *bonus, const void *tag, dmu_buf_t **dbp)
457 {
458 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
459 	dnode_t *dn;
460 	int err;
461 
462 	DB_DNODE_ENTER(db);
463 	dn = DB_DNODE(db);
464 
465 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
466 		err = SET_ERROR(EINVAL);
467 	} else {
468 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
469 
470 		if (!dn->dn_have_spill) {
471 			err = SET_ERROR(ENOENT);
472 		} else {
473 			err = dmu_spill_hold_by_dnode(dn,
474 			    DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
475 		}
476 
477 		rw_exit(&dn->dn_struct_rwlock);
478 	}
479 
480 	DB_DNODE_EXIT(db);
481 	return (err);
482 }
483 
484 int
485 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, const void *tag,
486     dmu_buf_t **dbp)
487 {
488 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
489 	dnode_t *dn;
490 	int err;
491 	uint32_t db_flags = DB_RF_CANFAIL;
492 
493 	if (flags & DMU_READ_NO_DECRYPT)
494 		db_flags |= DB_RF_NO_DECRYPT;
495 
496 	DB_DNODE_ENTER(db);
497 	dn = DB_DNODE(db);
498 	err = dmu_spill_hold_by_dnode(dn, db_flags, tag, dbp);
499 	DB_DNODE_EXIT(db);
500 
501 	return (err);
502 }
503 
504 /*
505  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
506  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
507  * and can induce severe lock contention when writing to several files
508  * whose dnodes are in the same block.
509  */
510 int
511 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
512     boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp,
513     uint32_t flags)
514 {
515 	dmu_buf_t **dbp;
516 	zstream_t *zs = NULL;
517 	uint64_t blkid, nblks, i;
518 	uint32_t dbuf_flags;
519 	int err;
520 	zio_t *zio = NULL;
521 	boolean_t missed = B_FALSE;
522 
523 	ASSERT(!read || length <= DMU_MAX_ACCESS);
524 
525 	/*
526 	 * Note: We directly notify the prefetch code of this read, so that
527 	 * we can tell it about the multi-block read.  dbuf_read() only knows
528 	 * about the one block it is accessing.
529 	 */
530 	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
531 	    DB_RF_NOPREFETCH;
532 
533 	if ((flags & DMU_READ_NO_DECRYPT) != 0)
534 		dbuf_flags |= DB_RF_NO_DECRYPT;
535 
536 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
537 	if (dn->dn_datablkshift) {
538 		int blkshift = dn->dn_datablkshift;
539 		nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
540 		    P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
541 	} else {
542 		if (offset + length > dn->dn_datablksz) {
543 			zfs_panic_recover("zfs: accessing past end of object "
544 			    "%llx/%llx (size=%u access=%llu+%llu)",
545 			    (longlong_t)dn->dn_objset->
546 			    os_dsl_dataset->ds_object,
547 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
548 			    (longlong_t)offset, (longlong_t)length);
549 			rw_exit(&dn->dn_struct_rwlock);
550 			return (SET_ERROR(EIO));
551 		}
552 		nblks = 1;
553 	}
554 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
555 
556 	if (read)
557 		zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
558 		    ZIO_FLAG_CANFAIL);
559 	blkid = dbuf_whichblock(dn, 0, offset);
560 	if ((flags & DMU_READ_NO_PREFETCH) == 0) {
561 		/*
562 		 * Prepare the zfetch before initiating the demand reads, so
563 		 * that if multiple threads block on same indirect block, we
564 		 * base predictions on the original less racy request order.
565 		 */
566 		zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read,
567 		    B_TRUE);
568 	}
569 	for (i = 0; i < nblks; i++) {
570 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
571 		if (db == NULL) {
572 			if (zs)
573 				dmu_zfetch_run(zs, missed, B_TRUE);
574 			rw_exit(&dn->dn_struct_rwlock);
575 			dmu_buf_rele_array(dbp, nblks, tag);
576 			if (read)
577 				zio_nowait(zio);
578 			return (SET_ERROR(EIO));
579 		}
580 
581 		/*
582 		 * Initiate async demand data read.
583 		 * We check the db_state after calling dbuf_read() because
584 		 * (1) dbuf_read() may change the state to CACHED due to a
585 		 * hit in the ARC, and (2) on a cache miss, a child will
586 		 * have been added to "zio" but not yet completed, so the
587 		 * state will not yet be CACHED.
588 		 */
589 		if (read) {
590 			if (i == nblks - 1 && blkid + i < dn->dn_maxblkid &&
591 			    offset + length < db->db.db_offset +
592 			    db->db.db_size) {
593 				if (offset <= db->db.db_offset)
594 					dbuf_flags |= DB_RF_PARTIAL_FIRST;
595 				else
596 					dbuf_flags |= DB_RF_PARTIAL_MORE;
597 			}
598 			(void) dbuf_read(db, zio, dbuf_flags);
599 			if (db->db_state != DB_CACHED)
600 				missed = B_TRUE;
601 		}
602 		dbp[i] = &db->db;
603 	}
604 
605 	if (!read)
606 		zfs_racct_write(length, nblks);
607 
608 	if (zs)
609 		dmu_zfetch_run(zs, missed, B_TRUE);
610 	rw_exit(&dn->dn_struct_rwlock);
611 
612 	if (read) {
613 		/* wait for async read i/o */
614 		err = zio_wait(zio);
615 		if (err) {
616 			dmu_buf_rele_array(dbp, nblks, tag);
617 			return (err);
618 		}
619 
620 		/* wait for other io to complete */
621 		for (i = 0; i < nblks; i++) {
622 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
623 			mutex_enter(&db->db_mtx);
624 			while (db->db_state == DB_READ ||
625 			    db->db_state == DB_FILL)
626 				cv_wait(&db->db_changed, &db->db_mtx);
627 			if (db->db_state == DB_UNCACHED)
628 				err = SET_ERROR(EIO);
629 			mutex_exit(&db->db_mtx);
630 			if (err) {
631 				dmu_buf_rele_array(dbp, nblks, tag);
632 				return (err);
633 			}
634 		}
635 	}
636 
637 	*numbufsp = nblks;
638 	*dbpp = dbp;
639 	return (0);
640 }
641 
642 int
643 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
644     uint64_t length, int read, const void *tag, int *numbufsp,
645     dmu_buf_t ***dbpp)
646 {
647 	dnode_t *dn;
648 	int err;
649 
650 	err = dnode_hold(os, object, FTAG, &dn);
651 	if (err)
652 		return (err);
653 
654 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
655 	    numbufsp, dbpp, DMU_READ_PREFETCH);
656 
657 	dnode_rele(dn, FTAG);
658 
659 	return (err);
660 }
661 
662 int
663 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
664     uint64_t length, boolean_t read, const void *tag, int *numbufsp,
665     dmu_buf_t ***dbpp)
666 {
667 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
668 	dnode_t *dn;
669 	int err;
670 
671 	DB_DNODE_ENTER(db);
672 	dn = DB_DNODE(db);
673 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
674 	    numbufsp, dbpp, DMU_READ_PREFETCH);
675 	DB_DNODE_EXIT(db);
676 
677 	return (err);
678 }
679 
680 void
681 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag)
682 {
683 	int i;
684 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
685 
686 	if (numbufs == 0)
687 		return;
688 
689 	for (i = 0; i < numbufs; i++) {
690 		if (dbp[i])
691 			dbuf_rele(dbp[i], tag);
692 	}
693 
694 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
695 }
696 
697 /*
698  * Issue prefetch I/Os for the given blocks.  If level is greater than 0, the
699  * indirect blocks prefetched will be those that point to the blocks containing
700  * the data starting at offset, and continuing to offset + len.  If the range
701  * it too long, prefetch the first dmu_prefetch_max bytes as requested, while
702  * for the rest only a higher level, also fitting within dmu_prefetch_max.  It
703  * should primarily help random reads, since for long sequential reads there is
704  * a speculative prefetcher.
705  *
706  * Note that if the indirect blocks above the blocks being prefetched are not
707  * in cache, they will be asynchronously read in.  Dnode read by dnode_hold()
708  * is currently synchronous.
709  */
710 void
711 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
712     uint64_t len, zio_priority_t pri)
713 {
714 	dnode_t *dn;
715 	int64_t level2 = level;
716 	uint64_t start, end, start2, end2;
717 
718 	if (dmu_prefetch_max == 0 || len == 0) {
719 		dmu_prefetch_dnode(os, object, pri);
720 		return;
721 	}
722 
723 	if (dnode_hold(os, object, FTAG, &dn) != 0)
724 		return;
725 
726 	/*
727 	 * Depending on len we may do two prefetches: blocks [start, end) at
728 	 * level, and following blocks [start2, end2) at higher level2.
729 	 */
730 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
731 	if (dn->dn_datablkshift != 0) {
732 		/*
733 		 * The object has multiple blocks.  Calculate the full range
734 		 * of blocks [start, end2) and then split it into two parts,
735 		 * so that the first [start, end) fits into dmu_prefetch_max.
736 		 */
737 		start = dbuf_whichblock(dn, level, offset);
738 		end2 = dbuf_whichblock(dn, level, offset + len - 1) + 1;
739 		uint8_t ibs = dn->dn_indblkshift;
740 		uint8_t bs = (level == 0) ? dn->dn_datablkshift : ibs;
741 		uint_t limit = P2ROUNDUP(dmu_prefetch_max, 1 << bs) >> bs;
742 		start2 = end = MIN(end2, start + limit);
743 
744 		/*
745 		 * Find level2 where [start2, end2) fits into dmu_prefetch_max.
746 		 */
747 		uint8_t ibps = ibs - SPA_BLKPTRSHIFT;
748 		limit = P2ROUNDUP(dmu_prefetch_max, 1 << ibs) >> ibs;
749 		do {
750 			level2++;
751 			start2 = P2ROUNDUP(start2, 1 << ibps) >> ibps;
752 			end2 = P2ROUNDUP(end2, 1 << ibps) >> ibps;
753 		} while (end2 - start2 > limit);
754 	} else {
755 		/* There is only one block.  Prefetch it or nothing. */
756 		start = start2 = end2 = 0;
757 		end = start + (level == 0 && offset < dn->dn_datablksz);
758 	}
759 
760 	for (uint64_t i = start; i < end; i++)
761 		dbuf_prefetch(dn, level, i, pri, 0);
762 	for (uint64_t i = start2; i < end2; i++)
763 		dbuf_prefetch(dn, level2, i, pri, 0);
764 	rw_exit(&dn->dn_struct_rwlock);
765 
766 	dnode_rele(dn, FTAG);
767 }
768 
769 /*
770  * Issue prefetch I/Os for the given object's dnode.
771  */
772 void
773 dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri)
774 {
775 	if (object == 0 || object >= DN_MAX_OBJECT)
776 		return;
777 
778 	dnode_t *dn = DMU_META_DNODE(os);
779 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
780 	uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t));
781 	dbuf_prefetch(dn, 0, blkid, pri, 0);
782 	rw_exit(&dn->dn_struct_rwlock);
783 }
784 
785 /*
786  * Get the next "chunk" of file data to free.  We traverse the file from
787  * the end so that the file gets shorter over time (if we crashes in the
788  * middle, this will leave us in a better state).  We find allocated file
789  * data by simply searching the allocated level 1 indirects.
790  *
791  * On input, *start should be the first offset that does not need to be
792  * freed (e.g. "offset + length").  On return, *start will be the first
793  * offset that should be freed and l1blks is set to the number of level 1
794  * indirect blocks found within the chunk.
795  */
796 static int
797 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
798 {
799 	uint64_t blks;
800 	uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
801 	/* bytes of data covered by a level-1 indirect block */
802 	uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
803 	    EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
804 
805 	ASSERT3U(minimum, <=, *start);
806 
807 	/*
808 	 * Check if we can free the entire range assuming that all of the
809 	 * L1 blocks in this range have data. If we can, we use this
810 	 * worst case value as an estimate so we can avoid having to look
811 	 * at the object's actual data.
812 	 */
813 	uint64_t total_l1blks =
814 	    (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
815 	    iblkrange;
816 	if (total_l1blks <= maxblks) {
817 		*l1blks = total_l1blks;
818 		*start = minimum;
819 		return (0);
820 	}
821 	ASSERT(ISP2(iblkrange));
822 
823 	for (blks = 0; *start > minimum && blks < maxblks; blks++) {
824 		int err;
825 
826 		/*
827 		 * dnode_next_offset(BACKWARDS) will find an allocated L1
828 		 * indirect block at or before the input offset.  We must
829 		 * decrement *start so that it is at the end of the region
830 		 * to search.
831 		 */
832 		(*start)--;
833 
834 		err = dnode_next_offset(dn,
835 		    DNODE_FIND_BACKWARDS, start, 2, 1, 0);
836 
837 		/* if there are no indirect blocks before start, we are done */
838 		if (err == ESRCH) {
839 			*start = minimum;
840 			break;
841 		} else if (err != 0) {
842 			*l1blks = blks;
843 			return (err);
844 		}
845 
846 		/* set start to the beginning of this L1 indirect */
847 		*start = P2ALIGN(*start, iblkrange);
848 	}
849 	if (*start < minimum)
850 		*start = minimum;
851 	*l1blks = blks;
852 
853 	return (0);
854 }
855 
856 /*
857  * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
858  * otherwise return false.
859  * Used below in dmu_free_long_range_impl() to enable abort when unmounting
860  */
861 static boolean_t
862 dmu_objset_zfs_unmounting(objset_t *os)
863 {
864 #ifdef _KERNEL
865 	if (dmu_objset_type(os) == DMU_OST_ZFS)
866 		return (zfs_get_vfs_flag_unmounted(os));
867 #else
868 	(void) os;
869 #endif
870 	return (B_FALSE);
871 }
872 
873 static int
874 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
875     uint64_t length)
876 {
877 	uint64_t object_size;
878 	int err;
879 	uint64_t dirty_frees_threshold;
880 	dsl_pool_t *dp = dmu_objset_pool(os);
881 
882 	if (dn == NULL)
883 		return (SET_ERROR(EINVAL));
884 
885 	object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
886 	if (offset >= object_size)
887 		return (0);
888 
889 	if (zfs_per_txg_dirty_frees_percent <= 100)
890 		dirty_frees_threshold =
891 		    zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
892 	else
893 		dirty_frees_threshold = zfs_dirty_data_max / 20;
894 
895 	if (length == DMU_OBJECT_END || offset + length > object_size)
896 		length = object_size - offset;
897 
898 	while (length != 0) {
899 		uint64_t chunk_end, chunk_begin, chunk_len;
900 		uint64_t l1blks;
901 		dmu_tx_t *tx;
902 
903 		if (dmu_objset_zfs_unmounting(dn->dn_objset))
904 			return (SET_ERROR(EINTR));
905 
906 		chunk_end = chunk_begin = offset + length;
907 
908 		/* move chunk_begin backwards to the beginning of this chunk */
909 		err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
910 		if (err)
911 			return (err);
912 		ASSERT3U(chunk_begin, >=, offset);
913 		ASSERT3U(chunk_begin, <=, chunk_end);
914 
915 		chunk_len = chunk_end - chunk_begin;
916 
917 		tx = dmu_tx_create(os);
918 		dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
919 
920 		/*
921 		 * Mark this transaction as typically resulting in a net
922 		 * reduction in space used.
923 		 */
924 		dmu_tx_mark_netfree(tx);
925 		err = dmu_tx_assign(tx, TXG_WAIT);
926 		if (err) {
927 			dmu_tx_abort(tx);
928 			return (err);
929 		}
930 
931 		uint64_t txg = dmu_tx_get_txg(tx);
932 
933 		mutex_enter(&dp->dp_lock);
934 		uint64_t long_free_dirty =
935 		    dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
936 		mutex_exit(&dp->dp_lock);
937 
938 		/*
939 		 * To avoid filling up a TXG with just frees, wait for
940 		 * the next TXG to open before freeing more chunks if
941 		 * we have reached the threshold of frees.
942 		 */
943 		if (dirty_frees_threshold != 0 &&
944 		    long_free_dirty >= dirty_frees_threshold) {
945 			DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
946 			dmu_tx_commit(tx);
947 			txg_wait_open(dp, 0, B_TRUE);
948 			continue;
949 		}
950 
951 		/*
952 		 * In order to prevent unnecessary write throttling, for each
953 		 * TXG, we track the cumulative size of L1 blocks being dirtied
954 		 * in dnode_free_range() below. We compare this number to a
955 		 * tunable threshold, past which we prevent new L1 dirty freeing
956 		 * blocks from being added into the open TXG. See
957 		 * dmu_free_long_range_impl() for details. The threshold
958 		 * prevents write throttle activation due to dirty freeing L1
959 		 * blocks taking up a large percentage of zfs_dirty_data_max.
960 		 */
961 		mutex_enter(&dp->dp_lock);
962 		dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
963 		    l1blks << dn->dn_indblkshift;
964 		mutex_exit(&dp->dp_lock);
965 		DTRACE_PROBE3(free__long__range,
966 		    uint64_t, long_free_dirty, uint64_t, chunk_len,
967 		    uint64_t, txg);
968 		dnode_free_range(dn, chunk_begin, chunk_len, tx);
969 
970 		dmu_tx_commit(tx);
971 
972 		length -= chunk_len;
973 	}
974 	return (0);
975 }
976 
977 int
978 dmu_free_long_range(objset_t *os, uint64_t object,
979     uint64_t offset, uint64_t length)
980 {
981 	dnode_t *dn;
982 	int err;
983 
984 	err = dnode_hold(os, object, FTAG, &dn);
985 	if (err != 0)
986 		return (err);
987 	err = dmu_free_long_range_impl(os, dn, offset, length);
988 
989 	/*
990 	 * It is important to zero out the maxblkid when freeing the entire
991 	 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
992 	 * will take the fast path, and (b) dnode_reallocate() can verify
993 	 * that the entire file has been freed.
994 	 */
995 	if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
996 		dn->dn_maxblkid = 0;
997 
998 	dnode_rele(dn, FTAG);
999 	return (err);
1000 }
1001 
1002 int
1003 dmu_free_long_object(objset_t *os, uint64_t object)
1004 {
1005 	dmu_tx_t *tx;
1006 	int err;
1007 
1008 	err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
1009 	if (err != 0)
1010 		return (err);
1011 
1012 	tx = dmu_tx_create(os);
1013 	dmu_tx_hold_bonus(tx, object);
1014 	dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
1015 	dmu_tx_mark_netfree(tx);
1016 	err = dmu_tx_assign(tx, TXG_WAIT);
1017 	if (err == 0) {
1018 		err = dmu_object_free(os, object, tx);
1019 		dmu_tx_commit(tx);
1020 	} else {
1021 		dmu_tx_abort(tx);
1022 	}
1023 
1024 	return (err);
1025 }
1026 
1027 int
1028 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
1029     uint64_t size, dmu_tx_t *tx)
1030 {
1031 	dnode_t *dn;
1032 	int err = dnode_hold(os, object, FTAG, &dn);
1033 	if (err)
1034 		return (err);
1035 	ASSERT(offset < UINT64_MAX);
1036 	ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
1037 	dnode_free_range(dn, offset, size, tx);
1038 	dnode_rele(dn, FTAG);
1039 	return (0);
1040 }
1041 
1042 static int
1043 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
1044     void *buf, uint32_t flags)
1045 {
1046 	dmu_buf_t **dbp;
1047 	int numbufs, err = 0;
1048 
1049 	/*
1050 	 * Deal with odd block sizes, where there can't be data past the first
1051 	 * block.  If we ever do the tail block optimization, we will need to
1052 	 * handle that here as well.
1053 	 */
1054 	if (dn->dn_maxblkid == 0) {
1055 		uint64_t newsz = offset > dn->dn_datablksz ? 0 :
1056 		    MIN(size, dn->dn_datablksz - offset);
1057 		memset((char *)buf + newsz, 0, size - newsz);
1058 		size = newsz;
1059 	}
1060 
1061 	while (size > 0) {
1062 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1063 		int i;
1064 
1065 		/*
1066 		 * NB: we could do this block-at-a-time, but it's nice
1067 		 * to be reading in parallel.
1068 		 */
1069 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1070 		    TRUE, FTAG, &numbufs, &dbp, flags);
1071 		if (err)
1072 			break;
1073 
1074 		for (i = 0; i < numbufs; i++) {
1075 			uint64_t tocpy;
1076 			int64_t bufoff;
1077 			dmu_buf_t *db = dbp[i];
1078 
1079 			ASSERT(size > 0);
1080 
1081 			bufoff = offset - db->db_offset;
1082 			tocpy = MIN(db->db_size - bufoff, size);
1083 
1084 			(void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1085 
1086 			offset += tocpy;
1087 			size -= tocpy;
1088 			buf = (char *)buf + tocpy;
1089 		}
1090 		dmu_buf_rele_array(dbp, numbufs, FTAG);
1091 	}
1092 	return (err);
1093 }
1094 
1095 int
1096 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1097     void *buf, uint32_t flags)
1098 {
1099 	dnode_t *dn;
1100 	int err;
1101 
1102 	err = dnode_hold(os, object, FTAG, &dn);
1103 	if (err != 0)
1104 		return (err);
1105 
1106 	err = dmu_read_impl(dn, offset, size, buf, flags);
1107 	dnode_rele(dn, FTAG);
1108 	return (err);
1109 }
1110 
1111 int
1112 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1113     uint32_t flags)
1114 {
1115 	return (dmu_read_impl(dn, offset, size, buf, flags));
1116 }
1117 
1118 static void
1119 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1120     const void *buf, dmu_tx_t *tx)
1121 {
1122 	int i;
1123 
1124 	for (i = 0; i < numbufs; i++) {
1125 		uint64_t tocpy;
1126 		int64_t bufoff;
1127 		dmu_buf_t *db = dbp[i];
1128 
1129 		ASSERT(size > 0);
1130 
1131 		bufoff = offset - db->db_offset;
1132 		tocpy = MIN(db->db_size - bufoff, size);
1133 
1134 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1135 
1136 		if (tocpy == db->db_size)
1137 			dmu_buf_will_fill(db, tx, B_FALSE);
1138 		else
1139 			dmu_buf_will_dirty(db, tx);
1140 
1141 		(void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1142 
1143 		if (tocpy == db->db_size)
1144 			dmu_buf_fill_done(db, tx, B_FALSE);
1145 
1146 		offset += tocpy;
1147 		size -= tocpy;
1148 		buf = (char *)buf + tocpy;
1149 	}
1150 }
1151 
1152 void
1153 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1154     const void *buf, dmu_tx_t *tx)
1155 {
1156 	dmu_buf_t **dbp;
1157 	int numbufs;
1158 
1159 	if (size == 0)
1160 		return;
1161 
1162 	VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1163 	    FALSE, FTAG, &numbufs, &dbp));
1164 	dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1165 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1166 }
1167 
1168 /*
1169  * Note: Lustre is an external consumer of this interface.
1170  */
1171 void
1172 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1173     const void *buf, dmu_tx_t *tx)
1174 {
1175 	dmu_buf_t **dbp;
1176 	int numbufs;
1177 
1178 	if (size == 0)
1179 		return;
1180 
1181 	VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1182 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1183 	dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1184 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1185 }
1186 
1187 void
1188 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1189     dmu_tx_t *tx)
1190 {
1191 	dmu_buf_t **dbp;
1192 	int numbufs, i;
1193 
1194 	if (size == 0)
1195 		return;
1196 
1197 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1198 	    FALSE, FTAG, &numbufs, &dbp));
1199 
1200 	for (i = 0; i < numbufs; i++) {
1201 		dmu_buf_t *db = dbp[i];
1202 
1203 		dmu_buf_will_not_fill(db, tx);
1204 	}
1205 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1206 }
1207 
1208 void
1209 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1210     void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1211     int compressed_size, int byteorder, dmu_tx_t *tx)
1212 {
1213 	dmu_buf_t *db;
1214 
1215 	ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1216 	ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1217 	VERIFY0(dmu_buf_hold_noread(os, object, offset,
1218 	    FTAG, &db));
1219 
1220 	dmu_buf_write_embedded(db,
1221 	    data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1222 	    uncompressed_size, compressed_size, byteorder, tx);
1223 
1224 	dmu_buf_rele(db, FTAG);
1225 }
1226 
1227 void
1228 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1229     dmu_tx_t *tx)
1230 {
1231 	int numbufs, i;
1232 	dmu_buf_t **dbp;
1233 
1234 	VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1235 	    &numbufs, &dbp));
1236 	for (i = 0; i < numbufs; i++)
1237 		dmu_buf_redact(dbp[i], tx);
1238 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1239 }
1240 
1241 #ifdef _KERNEL
1242 int
1243 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size)
1244 {
1245 	dmu_buf_t **dbp;
1246 	int numbufs, i, err;
1247 
1248 	/*
1249 	 * NB: we could do this block-at-a-time, but it's nice
1250 	 * to be reading in parallel.
1251 	 */
1252 	err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1253 	    TRUE, FTAG, &numbufs, &dbp, 0);
1254 	if (err)
1255 		return (err);
1256 
1257 	for (i = 0; i < numbufs; i++) {
1258 		uint64_t tocpy;
1259 		int64_t bufoff;
1260 		dmu_buf_t *db = dbp[i];
1261 
1262 		ASSERT(size > 0);
1263 
1264 		bufoff = zfs_uio_offset(uio) - db->db_offset;
1265 		tocpy = MIN(db->db_size - bufoff, size);
1266 
1267 		err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1268 		    UIO_READ, uio);
1269 
1270 		if (err)
1271 			break;
1272 
1273 		size -= tocpy;
1274 	}
1275 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1276 
1277 	return (err);
1278 }
1279 
1280 /*
1281  * Read 'size' bytes into the uio buffer.
1282  * From object zdb->db_object.
1283  * Starting at zfs_uio_offset(uio).
1284  *
1285  * If the caller already has a dbuf in the target object
1286  * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1287  * because we don't have to find the dnode_t for the object.
1288  */
1289 int
1290 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size)
1291 {
1292 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1293 	dnode_t *dn;
1294 	int err;
1295 
1296 	if (size == 0)
1297 		return (0);
1298 
1299 	DB_DNODE_ENTER(db);
1300 	dn = DB_DNODE(db);
1301 	err = dmu_read_uio_dnode(dn, uio, size);
1302 	DB_DNODE_EXIT(db);
1303 
1304 	return (err);
1305 }
1306 
1307 /*
1308  * Read 'size' bytes into the uio buffer.
1309  * From the specified object
1310  * Starting at offset zfs_uio_offset(uio).
1311  */
1312 int
1313 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size)
1314 {
1315 	dnode_t *dn;
1316 	int err;
1317 
1318 	if (size == 0)
1319 		return (0);
1320 
1321 	err = dnode_hold(os, object, FTAG, &dn);
1322 	if (err)
1323 		return (err);
1324 
1325 	err = dmu_read_uio_dnode(dn, uio, size);
1326 
1327 	dnode_rele(dn, FTAG);
1328 
1329 	return (err);
1330 }
1331 
1332 int
1333 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx)
1334 {
1335 	dmu_buf_t **dbp;
1336 	int numbufs;
1337 	int err = 0;
1338 	int i;
1339 
1340 	err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1341 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1342 	if (err)
1343 		return (err);
1344 
1345 	for (i = 0; i < numbufs; i++) {
1346 		uint64_t tocpy;
1347 		int64_t bufoff;
1348 		dmu_buf_t *db = dbp[i];
1349 
1350 		ASSERT(size > 0);
1351 
1352 		offset_t off = zfs_uio_offset(uio);
1353 		bufoff = off - db->db_offset;
1354 		tocpy = MIN(db->db_size - bufoff, size);
1355 
1356 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1357 
1358 		if (tocpy == db->db_size)
1359 			dmu_buf_will_fill(db, tx, B_TRUE);
1360 		else
1361 			dmu_buf_will_dirty(db, tx);
1362 
1363 		err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1364 		    tocpy, UIO_WRITE, uio);
1365 
1366 		if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) {
1367 			/* The fill was reverted.  Undo any uio progress. */
1368 			zfs_uio_advance(uio, off - zfs_uio_offset(uio));
1369 		}
1370 
1371 		if (err)
1372 			break;
1373 
1374 		size -= tocpy;
1375 	}
1376 
1377 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1378 	return (err);
1379 }
1380 
1381 /*
1382  * Write 'size' bytes from the uio buffer.
1383  * To object zdb->db_object.
1384  * Starting at offset zfs_uio_offset(uio).
1385  *
1386  * If the caller already has a dbuf in the target object
1387  * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1388  * because we don't have to find the dnode_t for the object.
1389  */
1390 int
1391 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1392     dmu_tx_t *tx)
1393 {
1394 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1395 	dnode_t *dn;
1396 	int err;
1397 
1398 	if (size == 0)
1399 		return (0);
1400 
1401 	DB_DNODE_ENTER(db);
1402 	dn = DB_DNODE(db);
1403 	err = dmu_write_uio_dnode(dn, uio, size, tx);
1404 	DB_DNODE_EXIT(db);
1405 
1406 	return (err);
1407 }
1408 
1409 /*
1410  * Write 'size' bytes from the uio buffer.
1411  * To the specified object.
1412  * Starting at offset zfs_uio_offset(uio).
1413  */
1414 int
1415 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1416     dmu_tx_t *tx)
1417 {
1418 	dnode_t *dn;
1419 	int err;
1420 
1421 	if (size == 0)
1422 		return (0);
1423 
1424 	err = dnode_hold(os, object, FTAG, &dn);
1425 	if (err)
1426 		return (err);
1427 
1428 	err = dmu_write_uio_dnode(dn, uio, size, tx);
1429 
1430 	dnode_rele(dn, FTAG);
1431 
1432 	return (err);
1433 }
1434 #endif /* _KERNEL */
1435 
1436 /*
1437  * Allocate a loaned anonymous arc buffer.
1438  */
1439 arc_buf_t *
1440 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1441 {
1442 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1443 
1444 	return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1445 }
1446 
1447 /*
1448  * Free a loaned arc buffer.
1449  */
1450 void
1451 dmu_return_arcbuf(arc_buf_t *buf)
1452 {
1453 	arc_return_buf(buf, FTAG);
1454 	arc_buf_destroy(buf, FTAG);
1455 }
1456 
1457 /*
1458  * A "lightweight" write is faster than a regular write (e.g.
1459  * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1460  * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t.  However, the
1461  * data can not be read or overwritten until the transaction's txg has been
1462  * synced.  This makes it appropriate for workloads that are known to be
1463  * (temporarily) write-only, like "zfs receive".
1464  *
1465  * A single block is written, starting at the specified offset in bytes.  If
1466  * the call is successful, it returns 0 and the provided abd has been
1467  * consumed (the caller should not free it).
1468  */
1469 int
1470 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1471     const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx)
1472 {
1473 	dbuf_dirty_record_t *dr =
1474 	    dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
1475 	if (dr == NULL)
1476 		return (SET_ERROR(EIO));
1477 	dr->dt.dll.dr_abd = abd;
1478 	dr->dt.dll.dr_props = *zp;
1479 	dr->dt.dll.dr_flags = flags;
1480 	return (0);
1481 }
1482 
1483 /*
1484  * When possible directly assign passed loaned arc buffer to a dbuf.
1485  * If this is not possible copy the contents of passed arc buf via
1486  * dmu_write().
1487  */
1488 int
1489 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1490     dmu_tx_t *tx)
1491 {
1492 	dmu_buf_impl_t *db;
1493 	objset_t *os = dn->dn_objset;
1494 	uint64_t object = dn->dn_object;
1495 	uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1496 	uint64_t blkid;
1497 
1498 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1499 	blkid = dbuf_whichblock(dn, 0, offset);
1500 	db = dbuf_hold(dn, blkid, FTAG);
1501 	rw_exit(&dn->dn_struct_rwlock);
1502 	if (db == NULL)
1503 		return (SET_ERROR(EIO));
1504 
1505 	/*
1506 	 * We can only assign if the offset is aligned and the arc buf is the
1507 	 * same size as the dbuf.
1508 	 */
1509 	if (offset == db->db.db_offset && blksz == db->db.db_size) {
1510 		zfs_racct_write(blksz, 1);
1511 		dbuf_assign_arcbuf(db, buf, tx);
1512 		dbuf_rele(db, FTAG);
1513 	} else {
1514 		/* compressed bufs must always be assignable to their dbuf */
1515 		ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1516 		ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1517 
1518 		dbuf_rele(db, FTAG);
1519 		dmu_write(os, object, offset, blksz, buf->b_data, tx);
1520 		dmu_return_arcbuf(buf);
1521 	}
1522 
1523 	return (0);
1524 }
1525 
1526 int
1527 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1528     dmu_tx_t *tx)
1529 {
1530 	int err;
1531 	dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1532 
1533 	DB_DNODE_ENTER(dbuf);
1534 	err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx);
1535 	DB_DNODE_EXIT(dbuf);
1536 
1537 	return (err);
1538 }
1539 
1540 typedef struct {
1541 	dbuf_dirty_record_t	*dsa_dr;
1542 	dmu_sync_cb_t		*dsa_done;
1543 	zgd_t			*dsa_zgd;
1544 	dmu_tx_t		*dsa_tx;
1545 } dmu_sync_arg_t;
1546 
1547 static void
1548 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1549 {
1550 	(void) buf;
1551 	dmu_sync_arg_t *dsa = varg;
1552 	dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1553 	blkptr_t *bp = zio->io_bp;
1554 
1555 	if (zio->io_error == 0) {
1556 		if (BP_IS_HOLE(bp)) {
1557 			/*
1558 			 * A block of zeros may compress to a hole, but the
1559 			 * block size still needs to be known for replay.
1560 			 */
1561 			BP_SET_LSIZE(bp, db->db_size);
1562 		} else if (!BP_IS_EMBEDDED(bp)) {
1563 			ASSERT(BP_GET_LEVEL(bp) == 0);
1564 			BP_SET_FILL(bp, 1);
1565 		}
1566 	}
1567 }
1568 
1569 static void
1570 dmu_sync_late_arrival_ready(zio_t *zio)
1571 {
1572 	dmu_sync_ready(zio, NULL, zio->io_private);
1573 }
1574 
1575 static void
1576 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1577 {
1578 	(void) buf;
1579 	dmu_sync_arg_t *dsa = varg;
1580 	dbuf_dirty_record_t *dr = dsa->dsa_dr;
1581 	dmu_buf_impl_t *db = dr->dr_dbuf;
1582 	zgd_t *zgd = dsa->dsa_zgd;
1583 
1584 	/*
1585 	 * Record the vdev(s) backing this blkptr so they can be flushed after
1586 	 * the writes for the lwb have completed.
1587 	 */
1588 	if (zio->io_error == 0) {
1589 		zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1590 	}
1591 
1592 	mutex_enter(&db->db_mtx);
1593 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1594 	if (zio->io_error == 0) {
1595 		dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1596 		if (dr->dt.dl.dr_nopwrite) {
1597 			blkptr_t *bp = zio->io_bp;
1598 			blkptr_t *bp_orig = &zio->io_bp_orig;
1599 			uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1600 
1601 			ASSERT(BP_EQUAL(bp, bp_orig));
1602 			VERIFY(BP_EQUAL(bp, db->db_blkptr));
1603 			ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1604 			VERIFY(zio_checksum_table[chksum].ci_flags &
1605 			    ZCHECKSUM_FLAG_NOPWRITE);
1606 		}
1607 		dr->dt.dl.dr_overridden_by = *zio->io_bp;
1608 		dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1609 		dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1610 
1611 		/*
1612 		 * Old style holes are filled with all zeros, whereas
1613 		 * new-style holes maintain their lsize, type, level,
1614 		 * and birth time (see zio_write_compress). While we
1615 		 * need to reset the BP_SET_LSIZE() call that happened
1616 		 * in dmu_sync_ready for old style holes, we do *not*
1617 		 * want to wipe out the information contained in new
1618 		 * style holes. Thus, only zero out the block pointer if
1619 		 * it's an old style hole.
1620 		 */
1621 		if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1622 		    dr->dt.dl.dr_overridden_by.blk_birth == 0)
1623 			BP_ZERO(&dr->dt.dl.dr_overridden_by);
1624 	} else {
1625 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1626 	}
1627 	cv_broadcast(&db->db_changed);
1628 	mutex_exit(&db->db_mtx);
1629 
1630 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1631 
1632 	kmem_free(dsa, sizeof (*dsa));
1633 }
1634 
1635 static void
1636 dmu_sync_late_arrival_done(zio_t *zio)
1637 {
1638 	blkptr_t *bp = zio->io_bp;
1639 	dmu_sync_arg_t *dsa = zio->io_private;
1640 	zgd_t *zgd = dsa->dsa_zgd;
1641 
1642 	if (zio->io_error == 0) {
1643 		/*
1644 		 * Record the vdev(s) backing this blkptr so they can be
1645 		 * flushed after the writes for the lwb have completed.
1646 		 */
1647 		zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1648 
1649 		if (!BP_IS_HOLE(bp)) {
1650 			blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
1651 			ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1652 			ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1653 			ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1654 			ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1655 			zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1656 		}
1657 	}
1658 
1659 	dmu_tx_commit(dsa->dsa_tx);
1660 
1661 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1662 
1663 	abd_free(zio->io_abd);
1664 	kmem_free(dsa, sizeof (*dsa));
1665 }
1666 
1667 static int
1668 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1669     zio_prop_t *zp, zbookmark_phys_t *zb)
1670 {
1671 	dmu_sync_arg_t *dsa;
1672 	dmu_tx_t *tx;
1673 	int error;
1674 
1675 	error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL,
1676 	    DB_RF_CANFAIL | DB_RF_NOPREFETCH);
1677 	if (error != 0)
1678 		return (error);
1679 
1680 	tx = dmu_tx_create(os);
1681 	dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1682 	/*
1683 	 * This transaction does not produce any dirty data or log blocks, so
1684 	 * it should not be throttled.  All other cases wait for TXG sync, by
1685 	 * which time the log block we are writing will be obsolete, so we can
1686 	 * skip waiting and just return error here instead.
1687 	 */
1688 	if (dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE) != 0) {
1689 		dmu_tx_abort(tx);
1690 		/* Make zl_get_data do txg_waited_synced() */
1691 		return (SET_ERROR(EIO));
1692 	}
1693 
1694 	/*
1695 	 * In order to prevent the zgd's lwb from being free'd prior to
1696 	 * dmu_sync_late_arrival_done() being called, we have to ensure
1697 	 * the lwb's "max txg" takes this tx's txg into account.
1698 	 */
1699 	zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1700 
1701 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1702 	dsa->dsa_dr = NULL;
1703 	dsa->dsa_done = done;
1704 	dsa->dsa_zgd = zgd;
1705 	dsa->dsa_tx = tx;
1706 
1707 	/*
1708 	 * Since we are currently syncing this txg, it's nontrivial to
1709 	 * determine what BP to nopwrite against, so we disable nopwrite.
1710 	 *
1711 	 * When syncing, the db_blkptr is initially the BP of the previous
1712 	 * txg.  We can not nopwrite against it because it will be changed
1713 	 * (this is similar to the non-late-arrival case where the dbuf is
1714 	 * dirty in a future txg).
1715 	 *
1716 	 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1717 	 * We can not nopwrite against it because although the BP will not
1718 	 * (typically) be changed, the data has not yet been persisted to this
1719 	 * location.
1720 	 *
1721 	 * Finally, when dbuf_write_done() is called, it is theoretically
1722 	 * possible to always nopwrite, because the data that was written in
1723 	 * this txg is the same data that we are trying to write.  However we
1724 	 * would need to check that this dbuf is not dirty in any future
1725 	 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1726 	 * don't nopwrite in this case.
1727 	 */
1728 	zp->zp_nopwrite = B_FALSE;
1729 
1730 	zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1731 	    abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1732 	    zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1733 	    dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done,
1734 	    dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1735 
1736 	return (0);
1737 }
1738 
1739 /*
1740  * Intent log support: sync the block associated with db to disk.
1741  * N.B. and XXX: the caller is responsible for making sure that the
1742  * data isn't changing while dmu_sync() is writing it.
1743  *
1744  * Return values:
1745  *
1746  *	EEXIST: this txg has already been synced, so there's nothing to do.
1747  *		The caller should not log the write.
1748  *
1749  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1750  *		The caller should not log the write.
1751  *
1752  *	EALREADY: this block is already in the process of being synced.
1753  *		The caller should track its progress (somehow).
1754  *
1755  *	EIO: could not do the I/O.
1756  *		The caller should do a txg_wait_synced().
1757  *
1758  *	0: the I/O has been initiated.
1759  *		The caller should log this blkptr in the done callback.
1760  *		It is possible that the I/O will fail, in which case
1761  *		the error will be reported to the done callback and
1762  *		propagated to pio from zio_done().
1763  */
1764 int
1765 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1766 {
1767 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1768 	objset_t *os = db->db_objset;
1769 	dsl_dataset_t *ds = os->os_dsl_dataset;
1770 	dbuf_dirty_record_t *dr, *dr_next;
1771 	dmu_sync_arg_t *dsa;
1772 	zbookmark_phys_t zb;
1773 	zio_prop_t zp;
1774 	dnode_t *dn;
1775 
1776 	ASSERT(pio != NULL);
1777 	ASSERT(txg != 0);
1778 
1779 	SET_BOOKMARK(&zb, ds->ds_object,
1780 	    db->db.db_object, db->db_level, db->db_blkid);
1781 
1782 	DB_DNODE_ENTER(db);
1783 	dn = DB_DNODE(db);
1784 	dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1785 	DB_DNODE_EXIT(db);
1786 
1787 	/*
1788 	 * If we're frozen (running ziltest), we always need to generate a bp.
1789 	 */
1790 	if (txg > spa_freeze_txg(os->os_spa))
1791 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1792 
1793 	/*
1794 	 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1795 	 * and us.  If we determine that this txg is not yet syncing,
1796 	 * but it begins to sync a moment later, that's OK because the
1797 	 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1798 	 */
1799 	mutex_enter(&db->db_mtx);
1800 
1801 	if (txg <= spa_last_synced_txg(os->os_spa)) {
1802 		/*
1803 		 * This txg has already synced.  There's nothing to do.
1804 		 */
1805 		mutex_exit(&db->db_mtx);
1806 		return (SET_ERROR(EEXIST));
1807 	}
1808 
1809 	if (txg <= spa_syncing_txg(os->os_spa)) {
1810 		/*
1811 		 * This txg is currently syncing, so we can't mess with
1812 		 * the dirty record anymore; just write a new log block.
1813 		 */
1814 		mutex_exit(&db->db_mtx);
1815 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1816 	}
1817 
1818 	dr = dbuf_find_dirty_eq(db, txg);
1819 
1820 	if (dr == NULL) {
1821 		/*
1822 		 * There's no dr for this dbuf, so it must have been freed.
1823 		 * There's no need to log writes to freed blocks, so we're done.
1824 		 */
1825 		mutex_exit(&db->db_mtx);
1826 		return (SET_ERROR(ENOENT));
1827 	}
1828 
1829 	dr_next = list_next(&db->db_dirty_records, dr);
1830 	ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
1831 
1832 	if (db->db_blkptr != NULL) {
1833 		/*
1834 		 * We need to fill in zgd_bp with the current blkptr so that
1835 		 * the nopwrite code can check if we're writing the same
1836 		 * data that's already on disk.  We can only nopwrite if we
1837 		 * are sure that after making the copy, db_blkptr will not
1838 		 * change until our i/o completes.  We ensure this by
1839 		 * holding the db_mtx, and only allowing nopwrite if the
1840 		 * block is not already dirty (see below).  This is verified
1841 		 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1842 		 * not changed.
1843 		 */
1844 		*zgd->zgd_bp = *db->db_blkptr;
1845 	}
1846 
1847 	/*
1848 	 * Assume the on-disk data is X, the current syncing data (in
1849 	 * txg - 1) is Y, and the current in-memory data is Z (currently
1850 	 * in dmu_sync).
1851 	 *
1852 	 * We usually want to perform a nopwrite if X and Z are the
1853 	 * same.  However, if Y is different (i.e. the BP is going to
1854 	 * change before this write takes effect), then a nopwrite will
1855 	 * be incorrect - we would override with X, which could have
1856 	 * been freed when Y was written.
1857 	 *
1858 	 * (Note that this is not a concern when we are nop-writing from
1859 	 * syncing context, because X and Y must be identical, because
1860 	 * all previous txgs have been synced.)
1861 	 *
1862 	 * Therefore, we disable nopwrite if the current BP could change
1863 	 * before this TXG.  There are two ways it could change: by
1864 	 * being dirty (dr_next is non-NULL), or by being freed
1865 	 * (dnode_block_freed()).  This behavior is verified by
1866 	 * zio_done(), which VERIFYs that the override BP is identical
1867 	 * to the on-disk BP.
1868 	 */
1869 	DB_DNODE_ENTER(db);
1870 	dn = DB_DNODE(db);
1871 	if (dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1872 		zp.zp_nopwrite = B_FALSE;
1873 	DB_DNODE_EXIT(db);
1874 
1875 	ASSERT(dr->dr_txg == txg);
1876 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1877 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1878 		/*
1879 		 * We have already issued a sync write for this buffer,
1880 		 * or this buffer has already been synced.  It could not
1881 		 * have been dirtied since, or we would have cleared the state.
1882 		 */
1883 		mutex_exit(&db->db_mtx);
1884 		return (SET_ERROR(EALREADY));
1885 	}
1886 
1887 	ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1888 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1889 	mutex_exit(&db->db_mtx);
1890 
1891 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1892 	dsa->dsa_dr = dr;
1893 	dsa->dsa_done = done;
1894 	dsa->dsa_zgd = zgd;
1895 	dsa->dsa_tx = NULL;
1896 
1897 	zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp,
1898 	    dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), dbuf_is_l2cacheable(db),
1899 	    &zp, dmu_sync_ready, NULL, dmu_sync_done, dsa,
1900 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1901 
1902 	return (0);
1903 }
1904 
1905 int
1906 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
1907 {
1908 	dnode_t *dn;
1909 	int err;
1910 
1911 	err = dnode_hold(os, object, FTAG, &dn);
1912 	if (err)
1913 		return (err);
1914 	err = dnode_set_nlevels(dn, nlevels, tx);
1915 	dnode_rele(dn, FTAG);
1916 	return (err);
1917 }
1918 
1919 int
1920 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1921     dmu_tx_t *tx)
1922 {
1923 	dnode_t *dn;
1924 	int err;
1925 
1926 	err = dnode_hold(os, object, FTAG, &dn);
1927 	if (err)
1928 		return (err);
1929 	err = dnode_set_blksz(dn, size, ibs, tx);
1930 	dnode_rele(dn, FTAG);
1931 	return (err);
1932 }
1933 
1934 int
1935 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
1936     dmu_tx_t *tx)
1937 {
1938 	dnode_t *dn;
1939 	int err;
1940 
1941 	err = dnode_hold(os, object, FTAG, &dn);
1942 	if (err)
1943 		return (err);
1944 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1945 	dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE);
1946 	rw_exit(&dn->dn_struct_rwlock);
1947 	dnode_rele(dn, FTAG);
1948 	return (0);
1949 }
1950 
1951 void
1952 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1953     dmu_tx_t *tx)
1954 {
1955 	dnode_t *dn;
1956 
1957 	/*
1958 	 * Send streams include each object's checksum function.  This
1959 	 * check ensures that the receiving system can understand the
1960 	 * checksum function transmitted.
1961 	 */
1962 	ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1963 
1964 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
1965 	ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1966 	dn->dn_checksum = checksum;
1967 	dnode_setdirty(dn, tx);
1968 	dnode_rele(dn, FTAG);
1969 }
1970 
1971 void
1972 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1973     dmu_tx_t *tx)
1974 {
1975 	dnode_t *dn;
1976 
1977 	/*
1978 	 * Send streams include each object's compression function.  This
1979 	 * check ensures that the receiving system can understand the
1980 	 * compression function transmitted.
1981 	 */
1982 	ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1983 
1984 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
1985 	dn->dn_compress = compress;
1986 	dnode_setdirty(dn, tx);
1987 	dnode_rele(dn, FTAG);
1988 }
1989 
1990 /*
1991  * When the "redundant_metadata" property is set to "most", only indirect
1992  * blocks of this level and higher will have an additional ditto block.
1993  */
1994 static const int zfs_redundant_metadata_most_ditto_level = 2;
1995 
1996 void
1997 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1998 {
1999 	dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2000 	boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2001 	    (wp & WP_SPILL));
2002 	enum zio_checksum checksum = os->os_checksum;
2003 	enum zio_compress compress = os->os_compress;
2004 	uint8_t complevel = os->os_complevel;
2005 	enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2006 	boolean_t dedup = B_FALSE;
2007 	boolean_t nopwrite = B_FALSE;
2008 	boolean_t dedup_verify = os->os_dedup_verify;
2009 	boolean_t encrypt = B_FALSE;
2010 	int copies = os->os_copies;
2011 
2012 	/*
2013 	 * We maintain different write policies for each of the following
2014 	 * types of data:
2015 	 *	 1. metadata
2016 	 *	 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2017 	 *	 3. all other level 0 blocks
2018 	 */
2019 	if (ismd) {
2020 		/*
2021 		 * XXX -- we should design a compression algorithm
2022 		 * that specializes in arrays of bps.
2023 		 */
2024 		compress = zio_compress_select(os->os_spa,
2025 		    ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2026 
2027 		/*
2028 		 * Metadata always gets checksummed.  If the data
2029 		 * checksum is multi-bit correctable, and it's not a
2030 		 * ZBT-style checksum, then it's suitable for metadata
2031 		 * as well.  Otherwise, the metadata checksum defaults
2032 		 * to fletcher4.
2033 		 */
2034 		if (!(zio_checksum_table[checksum].ci_flags &
2035 		    ZCHECKSUM_FLAG_METADATA) ||
2036 		    (zio_checksum_table[checksum].ci_flags &
2037 		    ZCHECKSUM_FLAG_EMBEDDED))
2038 			checksum = ZIO_CHECKSUM_FLETCHER_4;
2039 
2040 		switch (os->os_redundant_metadata) {
2041 		case ZFS_REDUNDANT_METADATA_ALL:
2042 			copies++;
2043 			break;
2044 		case ZFS_REDUNDANT_METADATA_MOST:
2045 			if (level >= zfs_redundant_metadata_most_ditto_level ||
2046 			    DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))
2047 				copies++;
2048 			break;
2049 		case ZFS_REDUNDANT_METADATA_SOME:
2050 			if (DMU_OT_IS_CRITICAL(type))
2051 				copies++;
2052 			break;
2053 		case ZFS_REDUNDANT_METADATA_NONE:
2054 			break;
2055 		}
2056 	} else if (wp & WP_NOFILL) {
2057 		ASSERT(level == 0);
2058 
2059 		/*
2060 		 * If we're writing preallocated blocks, we aren't actually
2061 		 * writing them so don't set any policy properties.  These
2062 		 * blocks are currently only used by an external subsystem
2063 		 * outside of zfs (i.e. dump) and not written by the zio
2064 		 * pipeline.
2065 		 */
2066 		compress = ZIO_COMPRESS_OFF;
2067 		checksum = ZIO_CHECKSUM_OFF;
2068 	} else {
2069 		compress = zio_compress_select(os->os_spa, dn->dn_compress,
2070 		    compress);
2071 		complevel = zio_complevel_select(os->os_spa, compress,
2072 		    complevel, complevel);
2073 
2074 		checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2075 		    zio_checksum_select(dn->dn_checksum, checksum) :
2076 		    dedup_checksum;
2077 
2078 		/*
2079 		 * Determine dedup setting.  If we are in dmu_sync(),
2080 		 * we won't actually dedup now because that's all
2081 		 * done in syncing context; but we do want to use the
2082 		 * dedup checksum.  If the checksum is not strong
2083 		 * enough to ensure unique signatures, force
2084 		 * dedup_verify.
2085 		 */
2086 		if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2087 			dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2088 			if (!(zio_checksum_table[checksum].ci_flags &
2089 			    ZCHECKSUM_FLAG_DEDUP))
2090 				dedup_verify = B_TRUE;
2091 		}
2092 
2093 		/*
2094 		 * Enable nopwrite if we have secure enough checksum
2095 		 * algorithm (see comment in zio_nop_write) and
2096 		 * compression is enabled.  We don't enable nopwrite if
2097 		 * dedup is enabled as the two features are mutually
2098 		 * exclusive.
2099 		 */
2100 		nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2101 		    ZCHECKSUM_FLAG_NOPWRITE) &&
2102 		    compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2103 	}
2104 
2105 	/*
2106 	 * All objects in an encrypted objset are protected from modification
2107 	 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2108 	 * in the bp, so we cannot use all copies. Encrypted objects are also
2109 	 * not subject to nopwrite since writing the same data will still
2110 	 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2111 	 * to avoid ambiguity in the dedup code since the DDT does not store
2112 	 * object types.
2113 	 */
2114 	if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2115 		encrypt = B_TRUE;
2116 
2117 		if (DMU_OT_IS_ENCRYPTED(type)) {
2118 			copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2119 			nopwrite = B_FALSE;
2120 		} else {
2121 			dedup = B_FALSE;
2122 		}
2123 
2124 		if (level <= 0 &&
2125 		    (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2126 			compress = ZIO_COMPRESS_EMPTY;
2127 		}
2128 	}
2129 
2130 	zp->zp_compress = compress;
2131 	zp->zp_complevel = complevel;
2132 	zp->zp_checksum = checksum;
2133 	zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2134 	zp->zp_level = level;
2135 	zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2136 	zp->zp_dedup = dedup;
2137 	zp->zp_dedup_verify = dedup && dedup_verify;
2138 	zp->zp_nopwrite = nopwrite;
2139 	zp->zp_encrypt = encrypt;
2140 	zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2141 	memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN);
2142 	memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN);
2143 	memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN);
2144 	zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ?
2145 	    os->os_zpl_special_smallblock : 0;
2146 
2147 	ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2148 }
2149 
2150 /*
2151  * Reports the location of data and holes in an object.  In order to
2152  * accurately report holes all dirty data must be synced to disk.  This
2153  * causes extremely poor performance when seeking for holes in a dirty file.
2154  * As a compromise, only provide hole data when the dnode is clean.  When
2155  * a dnode is dirty report the dnode as having no holes by returning EBUSY
2156  * which is always safe to do.
2157  */
2158 int
2159 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2160 {
2161 	dnode_t *dn;
2162 	int restarted = 0, err;
2163 
2164 restart:
2165 	err = dnode_hold(os, object, FTAG, &dn);
2166 	if (err)
2167 		return (err);
2168 
2169 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2170 
2171 	if (dnode_is_dirty(dn)) {
2172 		/*
2173 		 * If the zfs_dmu_offset_next_sync module option is enabled
2174 		 * then hole reporting has been requested.  Dirty dnodes
2175 		 * must be synced to disk to accurately report holes.
2176 		 *
2177 		 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is
2178 		 * held by the caller only a single restart will be required.
2179 		 * We tolerate callers which do not hold the rangelock by
2180 		 * returning EBUSY and not reporting holes after one restart.
2181 		 */
2182 		if (zfs_dmu_offset_next_sync) {
2183 			rw_exit(&dn->dn_struct_rwlock);
2184 			dnode_rele(dn, FTAG);
2185 
2186 			if (restarted)
2187 				return (SET_ERROR(EBUSY));
2188 
2189 			txg_wait_synced(dmu_objset_pool(os), 0);
2190 			restarted = 1;
2191 			goto restart;
2192 		}
2193 
2194 		err = SET_ERROR(EBUSY);
2195 	} else {
2196 		err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK |
2197 		    (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2198 	}
2199 
2200 	rw_exit(&dn->dn_struct_rwlock);
2201 	dnode_rele(dn, FTAG);
2202 
2203 	return (err);
2204 }
2205 
2206 int
2207 dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2208     blkptr_t *bps, size_t *nbpsp)
2209 {
2210 	dmu_buf_t **dbp, *dbuf;
2211 	dmu_buf_impl_t *db;
2212 	blkptr_t *bp;
2213 	int error, numbufs;
2214 
2215 	error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2216 	    &numbufs, &dbp);
2217 	if (error != 0) {
2218 		if (error == ESRCH) {
2219 			error = SET_ERROR(ENXIO);
2220 		}
2221 		return (error);
2222 	}
2223 
2224 	ASSERT3U(numbufs, <=, *nbpsp);
2225 
2226 	for (int i = 0; i < numbufs; i++) {
2227 		dbuf = dbp[i];
2228 		db = (dmu_buf_impl_t *)dbuf;
2229 
2230 		mutex_enter(&db->db_mtx);
2231 
2232 		if (!list_is_empty(&db->db_dirty_records)) {
2233 			dbuf_dirty_record_t *dr;
2234 
2235 			dr = list_head(&db->db_dirty_records);
2236 			if (dr->dt.dl.dr_brtwrite) {
2237 				/*
2238 				 * This is very special case where we clone a
2239 				 * block and in the same transaction group we
2240 				 * read its BP (most likely to clone the clone).
2241 				 */
2242 				bp = &dr->dt.dl.dr_overridden_by;
2243 			} else {
2244 				/*
2245 				 * The block was modified in the same
2246 				 * transaction group.
2247 				 */
2248 				mutex_exit(&db->db_mtx);
2249 				error = SET_ERROR(EAGAIN);
2250 				goto out;
2251 			}
2252 		} else {
2253 			bp = db->db_blkptr;
2254 		}
2255 
2256 		mutex_exit(&db->db_mtx);
2257 
2258 		if (bp == NULL) {
2259 			/*
2260 			 * The block was created in this transaction group,
2261 			 * so it has no BP yet.
2262 			 */
2263 			error = SET_ERROR(EAGAIN);
2264 			goto out;
2265 		}
2266 		/*
2267 		 * Make sure we clone only data blocks.
2268 		 */
2269 		if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) {
2270 			error = SET_ERROR(EINVAL);
2271 			goto out;
2272 		}
2273 
2274 		/*
2275 		 * If the block was allocated in transaction group that is not
2276 		 * yet synced, we could clone it, but we couldn't write this
2277 		 * operation into ZIL, or it may be impossible to replay, since
2278 		 * the block may appear not yet allocated at that point.
2279 		 */
2280 		if (BP_PHYSICAL_BIRTH(bp) > spa_freeze_txg(os->os_spa)) {
2281 			error = SET_ERROR(EINVAL);
2282 			goto out;
2283 		}
2284 		if (BP_PHYSICAL_BIRTH(bp) > spa_last_synced_txg(os->os_spa)) {
2285 			error = SET_ERROR(EAGAIN);
2286 			goto out;
2287 		}
2288 
2289 		bps[i] = *bp;
2290 	}
2291 
2292 	*nbpsp = numbufs;
2293 out:
2294 	dmu_buf_rele_array(dbp, numbufs, FTAG);
2295 
2296 	return (error);
2297 }
2298 
2299 int
2300 dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2301     dmu_tx_t *tx, const blkptr_t *bps, size_t nbps)
2302 {
2303 	spa_t *spa;
2304 	dmu_buf_t **dbp, *dbuf;
2305 	dmu_buf_impl_t *db;
2306 	struct dirty_leaf *dl;
2307 	dbuf_dirty_record_t *dr;
2308 	const blkptr_t *bp;
2309 	int error = 0, i, numbufs;
2310 
2311 	spa = os->os_spa;
2312 
2313 	VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2314 	    &numbufs, &dbp));
2315 	ASSERT3U(nbps, ==, numbufs);
2316 
2317 	/*
2318 	 * Before we start cloning make sure that the dbufs sizes match new BPs
2319 	 * sizes. If they don't, that's a no-go, as we are not able to shrink
2320 	 * dbufs.
2321 	 */
2322 	for (i = 0; i < numbufs; i++) {
2323 		dbuf = dbp[i];
2324 		db = (dmu_buf_impl_t *)dbuf;
2325 		bp = &bps[i];
2326 
2327 		ASSERT0(db->db_level);
2328 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2329 		ASSERT(db->db_blkid != DMU_SPILL_BLKID);
2330 
2331 		if (!BP_IS_HOLE(bp) && BP_GET_LSIZE(bp) != dbuf->db_size) {
2332 			error = SET_ERROR(EXDEV);
2333 			goto out;
2334 		}
2335 	}
2336 
2337 	for (i = 0; i < numbufs; i++) {
2338 		dbuf = dbp[i];
2339 		db = (dmu_buf_impl_t *)dbuf;
2340 		bp = &bps[i];
2341 
2342 		ASSERT0(db->db_level);
2343 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2344 		ASSERT(db->db_blkid != DMU_SPILL_BLKID);
2345 		ASSERT(BP_IS_HOLE(bp) || dbuf->db_size == BP_GET_LSIZE(bp));
2346 
2347 		dmu_buf_will_clone(dbuf, tx);
2348 
2349 		mutex_enter(&db->db_mtx);
2350 
2351 		dr = list_head(&db->db_dirty_records);
2352 		VERIFY(dr != NULL);
2353 		ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2354 		dl = &dr->dt.dl;
2355 		dl->dr_overridden_by = *bp;
2356 		dl->dr_brtwrite = B_TRUE;
2357 		dl->dr_override_state = DR_OVERRIDDEN;
2358 		if (BP_IS_HOLE(bp)) {
2359 			dl->dr_overridden_by.blk_birth = 0;
2360 			dl->dr_overridden_by.blk_phys_birth = 0;
2361 		} else {
2362 			dl->dr_overridden_by.blk_birth = dr->dr_txg;
2363 			if (!BP_IS_EMBEDDED(bp)) {
2364 				dl->dr_overridden_by.blk_phys_birth =
2365 				    BP_PHYSICAL_BIRTH(bp);
2366 			}
2367 		}
2368 
2369 		mutex_exit(&db->db_mtx);
2370 
2371 		/*
2372 		 * When data in embedded into BP there is no need to create
2373 		 * BRT entry as there is no data block. Just copy the BP as
2374 		 * it contains the data.
2375 		 */
2376 		if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
2377 			brt_pending_add(spa, bp, tx);
2378 		}
2379 	}
2380 out:
2381 	dmu_buf_rele_array(dbp, numbufs, FTAG);
2382 
2383 	return (error);
2384 }
2385 
2386 void
2387 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2388 {
2389 	dnode_phys_t *dnp = dn->dn_phys;
2390 
2391 	doi->doi_data_block_size = dn->dn_datablksz;
2392 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
2393 	    1ULL << dn->dn_indblkshift : 0;
2394 	doi->doi_type = dn->dn_type;
2395 	doi->doi_bonus_type = dn->dn_bonustype;
2396 	doi->doi_bonus_size = dn->dn_bonuslen;
2397 	doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2398 	doi->doi_indirection = dn->dn_nlevels;
2399 	doi->doi_checksum = dn->dn_checksum;
2400 	doi->doi_compress = dn->dn_compress;
2401 	doi->doi_nblkptr = dn->dn_nblkptr;
2402 	doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2403 	doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2404 	doi->doi_fill_count = 0;
2405 	for (int i = 0; i < dnp->dn_nblkptr; i++)
2406 		doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2407 }
2408 
2409 void
2410 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2411 {
2412 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2413 	mutex_enter(&dn->dn_mtx);
2414 
2415 	__dmu_object_info_from_dnode(dn, doi);
2416 
2417 	mutex_exit(&dn->dn_mtx);
2418 	rw_exit(&dn->dn_struct_rwlock);
2419 }
2420 
2421 /*
2422  * Get information on a DMU object.
2423  * If doi is NULL, just indicates whether the object exists.
2424  */
2425 int
2426 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2427 {
2428 	dnode_t *dn;
2429 	int err = dnode_hold(os, object, FTAG, &dn);
2430 
2431 	if (err)
2432 		return (err);
2433 
2434 	if (doi != NULL)
2435 		dmu_object_info_from_dnode(dn, doi);
2436 
2437 	dnode_rele(dn, FTAG);
2438 	return (0);
2439 }
2440 
2441 /*
2442  * As above, but faster; can be used when you have a held dbuf in hand.
2443  */
2444 void
2445 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2446 {
2447 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2448 
2449 	DB_DNODE_ENTER(db);
2450 	dmu_object_info_from_dnode(DB_DNODE(db), doi);
2451 	DB_DNODE_EXIT(db);
2452 }
2453 
2454 /*
2455  * Faster still when you only care about the size.
2456  */
2457 void
2458 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2459     u_longlong_t *nblk512)
2460 {
2461 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2462 	dnode_t *dn;
2463 
2464 	DB_DNODE_ENTER(db);
2465 	dn = DB_DNODE(db);
2466 
2467 	*blksize = dn->dn_datablksz;
2468 	/* add in number of slots used for the dnode itself */
2469 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2470 	    SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2471 	DB_DNODE_EXIT(db);
2472 }
2473 
2474 void
2475 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2476 {
2477 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2478 	dnode_t *dn;
2479 
2480 	DB_DNODE_ENTER(db);
2481 	dn = DB_DNODE(db);
2482 	*dnsize = dn->dn_num_slots << DNODE_SHIFT;
2483 	DB_DNODE_EXIT(db);
2484 }
2485 
2486 void
2487 byteswap_uint64_array(void *vbuf, size_t size)
2488 {
2489 	uint64_t *buf = vbuf;
2490 	size_t count = size >> 3;
2491 	int i;
2492 
2493 	ASSERT((size & 7) == 0);
2494 
2495 	for (i = 0; i < count; i++)
2496 		buf[i] = BSWAP_64(buf[i]);
2497 }
2498 
2499 void
2500 byteswap_uint32_array(void *vbuf, size_t size)
2501 {
2502 	uint32_t *buf = vbuf;
2503 	size_t count = size >> 2;
2504 	int i;
2505 
2506 	ASSERT((size & 3) == 0);
2507 
2508 	for (i = 0; i < count; i++)
2509 		buf[i] = BSWAP_32(buf[i]);
2510 }
2511 
2512 void
2513 byteswap_uint16_array(void *vbuf, size_t size)
2514 {
2515 	uint16_t *buf = vbuf;
2516 	size_t count = size >> 1;
2517 	int i;
2518 
2519 	ASSERT((size & 1) == 0);
2520 
2521 	for (i = 0; i < count; i++)
2522 		buf[i] = BSWAP_16(buf[i]);
2523 }
2524 
2525 void
2526 byteswap_uint8_array(void *vbuf, size_t size)
2527 {
2528 	(void) vbuf, (void) size;
2529 }
2530 
2531 void
2532 dmu_init(void)
2533 {
2534 	abd_init();
2535 	zfs_dbgmsg_init();
2536 	sa_cache_init();
2537 	dmu_objset_init();
2538 	dnode_init();
2539 	zfetch_init();
2540 	dmu_tx_init();
2541 	l2arc_init();
2542 	arc_init();
2543 	dbuf_init();
2544 }
2545 
2546 void
2547 dmu_fini(void)
2548 {
2549 	arc_fini(); /* arc depends on l2arc, so arc must go first */
2550 	l2arc_fini();
2551 	dmu_tx_fini();
2552 	zfetch_fini();
2553 	dbuf_fini();
2554 	dnode_fini();
2555 	dmu_objset_fini();
2556 	sa_cache_fini();
2557 	zfs_dbgmsg_fini();
2558 	abd_fini();
2559 }
2560 
2561 EXPORT_SYMBOL(dmu_bonus_hold);
2562 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2563 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2564 EXPORT_SYMBOL(dmu_buf_rele_array);
2565 EXPORT_SYMBOL(dmu_prefetch);
2566 EXPORT_SYMBOL(dmu_free_range);
2567 EXPORT_SYMBOL(dmu_free_long_range);
2568 EXPORT_SYMBOL(dmu_free_long_object);
2569 EXPORT_SYMBOL(dmu_read);
2570 EXPORT_SYMBOL(dmu_read_by_dnode);
2571 EXPORT_SYMBOL(dmu_write);
2572 EXPORT_SYMBOL(dmu_write_by_dnode);
2573 EXPORT_SYMBOL(dmu_prealloc);
2574 EXPORT_SYMBOL(dmu_object_info);
2575 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2576 EXPORT_SYMBOL(dmu_object_info_from_db);
2577 EXPORT_SYMBOL(dmu_object_size_from_db);
2578 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2579 EXPORT_SYMBOL(dmu_object_set_nlevels);
2580 EXPORT_SYMBOL(dmu_object_set_blocksize);
2581 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2582 EXPORT_SYMBOL(dmu_object_set_checksum);
2583 EXPORT_SYMBOL(dmu_object_set_compress);
2584 EXPORT_SYMBOL(dmu_offset_next);
2585 EXPORT_SYMBOL(dmu_write_policy);
2586 EXPORT_SYMBOL(dmu_sync);
2587 EXPORT_SYMBOL(dmu_request_arcbuf);
2588 EXPORT_SYMBOL(dmu_return_arcbuf);
2589 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2590 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2591 EXPORT_SYMBOL(dmu_buf_hold);
2592 EXPORT_SYMBOL(dmu_ot);
2593 
2594 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
2595 	"Enable NOP writes");
2596 
2597 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW,
2598 	"Percentage of dirtied blocks from frees in one TXG");
2599 
2600 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
2601 	"Enable forcing txg sync to find holes");
2602 
2603 /* CSTYLED */
2604 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW,
2605 	"Limit one prefetch call to this size");
2606