xref: /freebsd/sys/contrib/openzfs/module/zfs/dmu.c (revision f5ce3f4ef562ea9fc4d8f9c13c268f48a5bacba7)
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, 2023, 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 	if (newsize < 0 || newsize > db_fake->db_size)
280 		return (SET_ERROR(EINVAL));
281 
282 	DB_DNODE_ENTER(db);
283 	dn = DB_DNODE(db);
284 
285 	if (dn->dn_bonus != db) {
286 		error = SET_ERROR(EINVAL);
287 	} else {
288 		dnode_setbonuslen(dn, newsize, tx);
289 		error = 0;
290 	}
291 
292 	DB_DNODE_EXIT(db);
293 	return (error);
294 }
295 
296 int
297 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
298 {
299 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
300 	dnode_t *dn;
301 	int error;
302 
303 	if (!DMU_OT_IS_VALID(type))
304 		return (SET_ERROR(EINVAL));
305 
306 	DB_DNODE_ENTER(db);
307 	dn = DB_DNODE(db);
308 
309 	if (dn->dn_bonus != db) {
310 		error = SET_ERROR(EINVAL);
311 	} else {
312 		dnode_setbonus_type(dn, type, tx);
313 		error = 0;
314 	}
315 
316 	DB_DNODE_EXIT(db);
317 	return (error);
318 }
319 
320 dmu_object_type_t
321 dmu_get_bonustype(dmu_buf_t *db_fake)
322 {
323 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
324 	dmu_object_type_t type;
325 
326 	DB_DNODE_ENTER(db);
327 	type = DB_DNODE(db)->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 	int err;
490 	uint32_t db_flags = DB_RF_CANFAIL;
491 
492 	if (flags & DMU_READ_NO_DECRYPT)
493 		db_flags |= DB_RF_NO_DECRYPT;
494 
495 	DB_DNODE_ENTER(db);
496 	err = dmu_spill_hold_by_dnode(DB_DNODE(db), db_flags, tag, dbp);
497 	DB_DNODE_EXIT(db);
498 
499 	return (err);
500 }
501 
502 /*
503  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
504  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
505  * and can induce severe lock contention when writing to several files
506  * whose dnodes are in the same block.
507  */
508 int
509 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
510     boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp,
511     uint32_t flags)
512 {
513 	dmu_buf_t **dbp;
514 	zstream_t *zs = NULL;
515 	uint64_t blkid, nblks, i;
516 	uint32_t dbuf_flags;
517 	int err;
518 	zio_t *zio = NULL;
519 	boolean_t missed = B_FALSE;
520 
521 	ASSERT(!read || length <= DMU_MAX_ACCESS);
522 
523 	/*
524 	 * Note: We directly notify the prefetch code of this read, so that
525 	 * we can tell it about the multi-block read.  dbuf_read() only knows
526 	 * about the one block it is accessing.
527 	 */
528 	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
529 	    DB_RF_NOPREFETCH;
530 
531 	if ((flags & DMU_READ_NO_DECRYPT) != 0)
532 		dbuf_flags |= DB_RF_NO_DECRYPT;
533 
534 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
535 	if (dn->dn_datablkshift) {
536 		int blkshift = dn->dn_datablkshift;
537 		nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
538 		    P2ALIGN_TYPED(offset, 1ULL << blkshift, uint64_t))
539 		    >> blkshift;
540 	} else {
541 		if (offset + length > dn->dn_datablksz) {
542 			zfs_panic_recover("zfs: accessing past end of object "
543 			    "%llx/%llx (size=%u access=%llu+%llu)",
544 			    (longlong_t)dn->dn_objset->
545 			    os_dsl_dataset->ds_object,
546 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
547 			    (longlong_t)offset, (longlong_t)length);
548 			rw_exit(&dn->dn_struct_rwlock);
549 			return (SET_ERROR(EIO));
550 		}
551 		nblks = 1;
552 	}
553 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
554 
555 	if (read)
556 		zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
557 		    ZIO_FLAG_CANFAIL);
558 	blkid = dbuf_whichblock(dn, 0, offset);
559 	if ((flags & DMU_READ_NO_PREFETCH) == 0) {
560 		/*
561 		 * Prepare the zfetch before initiating the demand reads, so
562 		 * that if multiple threads block on same indirect block, we
563 		 * base predictions on the original less racy request order.
564 		 */
565 		zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read,
566 		    B_TRUE);
567 	}
568 	for (i = 0; i < nblks; i++) {
569 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
570 		if (db == NULL) {
571 			if (zs) {
572 				dmu_zfetch_run(&dn->dn_zfetch, zs, missed,
573 				    B_TRUE);
574 			}
575 			rw_exit(&dn->dn_struct_rwlock);
576 			dmu_buf_rele_array(dbp, nblks, tag);
577 			if (read)
578 				zio_nowait(zio);
579 			return (SET_ERROR(EIO));
580 		}
581 
582 		/*
583 		 * Initiate async demand data read.
584 		 * We check the db_state after calling dbuf_read() because
585 		 * (1) dbuf_read() may change the state to CACHED due to a
586 		 * hit in the ARC, and (2) on a cache miss, a child will
587 		 * have been added to "zio" but not yet completed, so the
588 		 * state will not yet be CACHED.
589 		 */
590 		if (read) {
591 			if (i == nblks - 1 && blkid + i < dn->dn_maxblkid &&
592 			    offset + length < db->db.db_offset +
593 			    db->db.db_size) {
594 				if (offset <= db->db.db_offset)
595 					dbuf_flags |= DB_RF_PARTIAL_FIRST;
596 				else
597 					dbuf_flags |= DB_RF_PARTIAL_MORE;
598 			}
599 			(void) dbuf_read(db, zio, dbuf_flags);
600 			if (db->db_state != DB_CACHED)
601 				missed = B_TRUE;
602 		}
603 		dbp[i] = &db->db;
604 	}
605 
606 	if (!read)
607 		zfs_racct_write(length, nblks);
608 
609 	if (zs)
610 		dmu_zfetch_run(&dn->dn_zfetch, zs, missed, B_TRUE);
611 	rw_exit(&dn->dn_struct_rwlock);
612 
613 	if (read) {
614 		/* wait for async read i/o */
615 		err = zio_wait(zio);
616 		if (err) {
617 			dmu_buf_rele_array(dbp, nblks, tag);
618 			return (err);
619 		}
620 
621 		/* wait for other io to complete */
622 		for (i = 0; i < nblks; i++) {
623 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
624 			mutex_enter(&db->db_mtx);
625 			while (db->db_state == DB_READ ||
626 			    db->db_state == DB_FILL)
627 				cv_wait(&db->db_changed, &db->db_mtx);
628 			if (db->db_state == DB_UNCACHED)
629 				err = SET_ERROR(EIO);
630 			mutex_exit(&db->db_mtx);
631 			if (err) {
632 				dmu_buf_rele_array(dbp, nblks, tag);
633 				return (err);
634 			}
635 		}
636 	}
637 
638 	*numbufsp = nblks;
639 	*dbpp = dbp;
640 	return (0);
641 }
642 
643 int
644 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
645     uint64_t length, int read, const void *tag, int *numbufsp,
646     dmu_buf_t ***dbpp)
647 {
648 	dnode_t *dn;
649 	int err;
650 
651 	err = dnode_hold(os, object, FTAG, &dn);
652 	if (err)
653 		return (err);
654 
655 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
656 	    numbufsp, dbpp, DMU_READ_PREFETCH);
657 
658 	dnode_rele(dn, FTAG);
659 
660 	return (err);
661 }
662 
663 int
664 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
665     uint64_t length, boolean_t read, const void *tag, int *numbufsp,
666     dmu_buf_t ***dbpp)
667 {
668 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
669 	int err;
670 
671 	DB_DNODE_ENTER(db);
672 	err = dmu_buf_hold_array_by_dnode(DB_DNODE(db), offset, length, read,
673 	    tag, numbufsp, dbpp, DMU_READ_PREFETCH);
674 	DB_DNODE_EXIT(db);
675 
676 	return (err);
677 }
678 
679 void
680 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag)
681 {
682 	int i;
683 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
684 
685 	if (numbufs == 0)
686 		return;
687 
688 	for (i = 0; i < numbufs; i++) {
689 		if (dbp[i])
690 			dbuf_rele(dbp[i], tag);
691 	}
692 
693 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
694 }
695 
696 /*
697  * Issue prefetch I/Os for the given blocks.  If level is greater than 0, the
698  * indirect blocks prefetched will be those that point to the blocks containing
699  * the data starting at offset, and continuing to offset + len.  If the range
700  * is too long, prefetch the first dmu_prefetch_max bytes as requested, while
701  * for the rest only a higher level, also fitting within dmu_prefetch_max.  It
702  * should primarily help random reads, since for long sequential reads there is
703  * a speculative prefetcher.
704  *
705  * Note that if the indirect blocks above the blocks being prefetched are not
706  * in cache, they will be asynchronously read in.  Dnode read by dnode_hold()
707  * is currently synchronous.
708  */
709 void
710 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
711     uint64_t len, zio_priority_t pri)
712 {
713 	dnode_t *dn;
714 
715 	if (dmu_prefetch_max == 0 || len == 0) {
716 		dmu_prefetch_dnode(os, object, pri);
717 		return;
718 	}
719 
720 	if (dnode_hold(os, object, FTAG, &dn) != 0)
721 		return;
722 
723 	dmu_prefetch_by_dnode(dn, level, offset, len, pri);
724 
725 	dnode_rele(dn, FTAG);
726 }
727 
728 void
729 dmu_prefetch_by_dnode(dnode_t *dn, int64_t level, uint64_t offset,
730     uint64_t len, zio_priority_t pri)
731 {
732 	int64_t level2 = level;
733 	uint64_t start, end, start2, end2;
734 
735 	/*
736 	 * Depending on len we may do two prefetches: blocks [start, end) at
737 	 * level, and following blocks [start2, end2) at higher level2.
738 	 */
739 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
740 	if (dn->dn_datablkshift != 0) {
741 		/*
742 		 * The object has multiple blocks.  Calculate the full range
743 		 * of blocks [start, end2) and then split it into two parts,
744 		 * so that the first [start, end) fits into dmu_prefetch_max.
745 		 */
746 		start = dbuf_whichblock(dn, level, offset);
747 		end2 = dbuf_whichblock(dn, level, offset + len - 1) + 1;
748 		uint8_t ibs = dn->dn_indblkshift;
749 		uint8_t bs = (level == 0) ? dn->dn_datablkshift : ibs;
750 		uint_t limit = P2ROUNDUP(dmu_prefetch_max, 1 << bs) >> bs;
751 		start2 = end = MIN(end2, start + limit);
752 
753 		/*
754 		 * Find level2 where [start2, end2) fits into dmu_prefetch_max.
755 		 */
756 		uint8_t ibps = ibs - SPA_BLKPTRSHIFT;
757 		limit = P2ROUNDUP(dmu_prefetch_max, 1 << ibs) >> ibs;
758 		do {
759 			level2++;
760 			start2 = P2ROUNDUP(start2, 1 << ibps) >> ibps;
761 			end2 = P2ROUNDUP(end2, 1 << ibps) >> ibps;
762 		} while (end2 - start2 > limit);
763 	} else {
764 		/* There is only one block.  Prefetch it or nothing. */
765 		start = start2 = end2 = 0;
766 		end = start + (level == 0 && offset < dn->dn_datablksz);
767 	}
768 
769 	for (uint64_t i = start; i < end; i++)
770 		dbuf_prefetch(dn, level, i, pri, 0);
771 	for (uint64_t i = start2; i < end2; i++)
772 		dbuf_prefetch(dn, level2, i, pri, 0);
773 	rw_exit(&dn->dn_struct_rwlock);
774 }
775 
776 typedef struct {
777 	kmutex_t	dpa_lock;
778 	kcondvar_t	dpa_cv;
779 	uint64_t	dpa_pending_io;
780 } dmu_prefetch_arg_t;
781 
782 static void
783 dmu_prefetch_done(void *arg, uint64_t level, uint64_t blkid, boolean_t issued)
784 {
785 	(void) level; (void) blkid; (void)issued;
786 	dmu_prefetch_arg_t *dpa = arg;
787 
788 	ASSERT0(level);
789 
790 	mutex_enter(&dpa->dpa_lock);
791 	ASSERT3U(dpa->dpa_pending_io, >, 0);
792 	if (--dpa->dpa_pending_io == 0)
793 		cv_broadcast(&dpa->dpa_cv);
794 	mutex_exit(&dpa->dpa_lock);
795 }
796 
797 static void
798 dmu_prefetch_wait_by_dnode(dnode_t *dn, uint64_t offset, uint64_t len)
799 {
800 	dmu_prefetch_arg_t dpa;
801 
802 	mutex_init(&dpa.dpa_lock, NULL, MUTEX_DEFAULT, NULL);
803 	cv_init(&dpa.dpa_cv, NULL, CV_DEFAULT, NULL);
804 
805 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
806 
807 	uint64_t start = dbuf_whichblock(dn, 0, offset);
808 	uint64_t end = dbuf_whichblock(dn, 0, offset + len - 1) + 1;
809 	dpa.dpa_pending_io = end - start;
810 
811 	for (uint64_t blk = start; blk < end; blk++) {
812 		(void) dbuf_prefetch_impl(dn, 0, blk, ZIO_PRIORITY_ASYNC_READ,
813 		    0, dmu_prefetch_done, &dpa);
814 	}
815 
816 	rw_exit(&dn->dn_struct_rwlock);
817 
818 	/* wait for prefetch L0 reads to finish */
819 	mutex_enter(&dpa.dpa_lock);
820 	while (dpa.dpa_pending_io > 0) {
821 		cv_wait(&dpa.dpa_cv, &dpa.dpa_lock);
822 
823 	}
824 	mutex_exit(&dpa.dpa_lock);
825 
826 	mutex_destroy(&dpa.dpa_lock);
827 	cv_destroy(&dpa.dpa_cv);
828 }
829 
830 /*
831  * Issue prefetch I/Os for the given L0 block range and wait for the I/O
832  * to complete. This does not enforce dmu_prefetch_max and will prefetch
833  * the entire range. The blocks are read from disk into the ARC but no
834  * decompression occurs (i.e., the dbuf cache is not required).
835  */
836 int
837 dmu_prefetch_wait(objset_t *os, uint64_t object, uint64_t offset, uint64_t size)
838 {
839 	dnode_t *dn;
840 	int err = 0;
841 
842 	err = dnode_hold(os, object, FTAG, &dn);
843 	if (err != 0)
844 		return (err);
845 
846 	/*
847 	 * Chunk the requests (16 indirects worth) so that we can be interrupted
848 	 */
849 	uint64_t chunksize;
850 	if (dn->dn_indblkshift) {
851 		uint64_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1);
852 		chunksize = (nbps * 16) << dn->dn_datablkshift;
853 	} else {
854 		chunksize = dn->dn_datablksz;
855 	}
856 
857 	while (size > 0) {
858 		uint64_t mylen = MIN(size, chunksize);
859 
860 		dmu_prefetch_wait_by_dnode(dn, offset, mylen);
861 
862 		offset += mylen;
863 		size -= mylen;
864 
865 		if (issig()) {
866 			err = SET_ERROR(EINTR);
867 			break;
868 		}
869 	}
870 
871 	dnode_rele(dn, FTAG);
872 
873 	return (err);
874 }
875 
876 /*
877  * Issue prefetch I/Os for the given object's dnode.
878  */
879 void
880 dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri)
881 {
882 	if (object == 0 || object >= DN_MAX_OBJECT)
883 		return;
884 
885 	dnode_t *dn = DMU_META_DNODE(os);
886 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
887 	uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t));
888 	dbuf_prefetch(dn, 0, blkid, pri, 0);
889 	rw_exit(&dn->dn_struct_rwlock);
890 }
891 
892 /*
893  * Get the next "chunk" of file data to free.  We traverse the file from
894  * the end so that the file gets shorter over time (if we crashes in the
895  * middle, this will leave us in a better state).  We find allocated file
896  * data by simply searching the allocated level 1 indirects.
897  *
898  * On input, *start should be the first offset that does not need to be
899  * freed (e.g. "offset + length").  On return, *start will be the first
900  * offset that should be freed and l1blks is set to the number of level 1
901  * indirect blocks found within the chunk.
902  */
903 static int
904 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
905 {
906 	uint64_t blks;
907 	uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
908 	/* bytes of data covered by a level-1 indirect block */
909 	uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
910 	    EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
911 
912 	ASSERT3U(minimum, <=, *start);
913 
914 	/* dn_nlevels == 1 means we don't have any L1 blocks */
915 	if (dn->dn_nlevels <= 1) {
916 		*l1blks = 0;
917 		*start = minimum;
918 		return (0);
919 	}
920 
921 	/*
922 	 * Check if we can free the entire range assuming that all of the
923 	 * L1 blocks in this range have data. If we can, we use this
924 	 * worst case value as an estimate so we can avoid having to look
925 	 * at the object's actual data.
926 	 */
927 	uint64_t total_l1blks =
928 	    (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
929 	    iblkrange;
930 	if (total_l1blks <= maxblks) {
931 		*l1blks = total_l1blks;
932 		*start = minimum;
933 		return (0);
934 	}
935 	ASSERT(ISP2(iblkrange));
936 
937 	for (blks = 0; *start > minimum && blks < maxblks; blks++) {
938 		int err;
939 
940 		/*
941 		 * dnode_next_offset(BACKWARDS) will find an allocated L1
942 		 * indirect block at or before the input offset.  We must
943 		 * decrement *start so that it is at the end of the region
944 		 * to search.
945 		 */
946 		(*start)--;
947 
948 		err = dnode_next_offset(dn,
949 		    DNODE_FIND_BACKWARDS, start, 2, 1, 0);
950 
951 		/* if there are no indirect blocks before start, we are done */
952 		if (err == ESRCH) {
953 			*start = minimum;
954 			break;
955 		} else if (err != 0) {
956 			*l1blks = blks;
957 			return (err);
958 		}
959 
960 		/* set start to the beginning of this L1 indirect */
961 		*start = P2ALIGN_TYPED(*start, iblkrange, uint64_t);
962 	}
963 	if (*start < minimum)
964 		*start = minimum;
965 	*l1blks = blks;
966 
967 	return (0);
968 }
969 
970 /*
971  * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
972  * otherwise return false.
973  * Used below in dmu_free_long_range_impl() to enable abort when unmounting
974  */
975 static boolean_t
976 dmu_objset_zfs_unmounting(objset_t *os)
977 {
978 #ifdef _KERNEL
979 	if (dmu_objset_type(os) == DMU_OST_ZFS)
980 		return (zfs_get_vfs_flag_unmounted(os));
981 #else
982 	(void) os;
983 #endif
984 	return (B_FALSE);
985 }
986 
987 static int
988 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
989     uint64_t length)
990 {
991 	uint64_t object_size;
992 	int err;
993 	uint64_t dirty_frees_threshold;
994 	dsl_pool_t *dp = dmu_objset_pool(os);
995 
996 	if (dn == NULL)
997 		return (SET_ERROR(EINVAL));
998 
999 	object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1000 	if (offset >= object_size)
1001 		return (0);
1002 
1003 	if (zfs_per_txg_dirty_frees_percent <= 100)
1004 		dirty_frees_threshold =
1005 		    zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
1006 	else
1007 		dirty_frees_threshold = zfs_dirty_data_max / 20;
1008 
1009 	if (length == DMU_OBJECT_END || offset + length > object_size)
1010 		length = object_size - offset;
1011 
1012 	while (length != 0) {
1013 		uint64_t chunk_end, chunk_begin, chunk_len;
1014 		uint64_t l1blks;
1015 		dmu_tx_t *tx;
1016 
1017 		if (dmu_objset_zfs_unmounting(dn->dn_objset))
1018 			return (SET_ERROR(EINTR));
1019 
1020 		chunk_end = chunk_begin = offset + length;
1021 
1022 		/* move chunk_begin backwards to the beginning of this chunk */
1023 		err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
1024 		if (err)
1025 			return (err);
1026 		ASSERT3U(chunk_begin, >=, offset);
1027 		ASSERT3U(chunk_begin, <=, chunk_end);
1028 
1029 		chunk_len = chunk_end - chunk_begin;
1030 
1031 		tx = dmu_tx_create(os);
1032 		dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
1033 
1034 		/*
1035 		 * Mark this transaction as typically resulting in a net
1036 		 * reduction in space used.
1037 		 */
1038 		dmu_tx_mark_netfree(tx);
1039 		err = dmu_tx_assign(tx, TXG_WAIT);
1040 		if (err) {
1041 			dmu_tx_abort(tx);
1042 			return (err);
1043 		}
1044 
1045 		uint64_t txg = dmu_tx_get_txg(tx);
1046 
1047 		mutex_enter(&dp->dp_lock);
1048 		uint64_t long_free_dirty =
1049 		    dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
1050 		mutex_exit(&dp->dp_lock);
1051 
1052 		/*
1053 		 * To avoid filling up a TXG with just frees, wait for
1054 		 * the next TXG to open before freeing more chunks if
1055 		 * we have reached the threshold of frees.
1056 		 */
1057 		if (dirty_frees_threshold != 0 &&
1058 		    long_free_dirty >= dirty_frees_threshold) {
1059 			DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
1060 			dmu_tx_commit(tx);
1061 			txg_wait_open(dp, 0, B_TRUE);
1062 			continue;
1063 		}
1064 
1065 		/*
1066 		 * In order to prevent unnecessary write throttling, for each
1067 		 * TXG, we track the cumulative size of L1 blocks being dirtied
1068 		 * in dnode_free_range() below. We compare this number to a
1069 		 * tunable threshold, past which we prevent new L1 dirty freeing
1070 		 * blocks from being added into the open TXG. See
1071 		 * dmu_free_long_range_impl() for details. The threshold
1072 		 * prevents write throttle activation due to dirty freeing L1
1073 		 * blocks taking up a large percentage of zfs_dirty_data_max.
1074 		 */
1075 		mutex_enter(&dp->dp_lock);
1076 		dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
1077 		    l1blks << dn->dn_indblkshift;
1078 		mutex_exit(&dp->dp_lock);
1079 		DTRACE_PROBE3(free__long__range,
1080 		    uint64_t, long_free_dirty, uint64_t, chunk_len,
1081 		    uint64_t, txg);
1082 		dnode_free_range(dn, chunk_begin, chunk_len, tx);
1083 
1084 		dmu_tx_commit(tx);
1085 
1086 		length -= chunk_len;
1087 	}
1088 	return (0);
1089 }
1090 
1091 int
1092 dmu_free_long_range(objset_t *os, uint64_t object,
1093     uint64_t offset, uint64_t length)
1094 {
1095 	dnode_t *dn;
1096 	int err;
1097 
1098 	err = dnode_hold(os, object, FTAG, &dn);
1099 	if (err != 0)
1100 		return (err);
1101 	err = dmu_free_long_range_impl(os, dn, offset, length);
1102 
1103 	/*
1104 	 * It is important to zero out the maxblkid when freeing the entire
1105 	 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
1106 	 * will take the fast path, and (b) dnode_reallocate() can verify
1107 	 * that the entire file has been freed.
1108 	 */
1109 	if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
1110 		dn->dn_maxblkid = 0;
1111 
1112 	dnode_rele(dn, FTAG);
1113 	return (err);
1114 }
1115 
1116 int
1117 dmu_free_long_object(objset_t *os, uint64_t object)
1118 {
1119 	dmu_tx_t *tx;
1120 	int err;
1121 
1122 	err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
1123 	if (err != 0)
1124 		return (err);
1125 
1126 	tx = dmu_tx_create(os);
1127 	dmu_tx_hold_bonus(tx, object);
1128 	dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
1129 	dmu_tx_mark_netfree(tx);
1130 	err = dmu_tx_assign(tx, TXG_WAIT);
1131 	if (err == 0) {
1132 		err = dmu_object_free(os, object, tx);
1133 		dmu_tx_commit(tx);
1134 	} else {
1135 		dmu_tx_abort(tx);
1136 	}
1137 
1138 	return (err);
1139 }
1140 
1141 int
1142 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
1143     uint64_t size, dmu_tx_t *tx)
1144 {
1145 	dnode_t *dn;
1146 	int err = dnode_hold(os, object, FTAG, &dn);
1147 	if (err)
1148 		return (err);
1149 	ASSERT(offset < UINT64_MAX);
1150 	ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
1151 	dnode_free_range(dn, offset, size, tx);
1152 	dnode_rele(dn, FTAG);
1153 	return (0);
1154 }
1155 
1156 static int
1157 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
1158     void *buf, uint32_t flags)
1159 {
1160 	dmu_buf_t **dbp;
1161 	int numbufs, err = 0;
1162 
1163 	/*
1164 	 * Deal with odd block sizes, where there can't be data past the first
1165 	 * block.  If we ever do the tail block optimization, we will need to
1166 	 * handle that here as well.
1167 	 */
1168 	if (dn->dn_maxblkid == 0) {
1169 		uint64_t newsz = offset > dn->dn_datablksz ? 0 :
1170 		    MIN(size, dn->dn_datablksz - offset);
1171 		memset((char *)buf + newsz, 0, size - newsz);
1172 		size = newsz;
1173 	}
1174 
1175 	while (size > 0) {
1176 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1177 		int i;
1178 
1179 		/*
1180 		 * NB: we could do this block-at-a-time, but it's nice
1181 		 * to be reading in parallel.
1182 		 */
1183 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1184 		    TRUE, FTAG, &numbufs, &dbp, flags);
1185 		if (err)
1186 			break;
1187 
1188 		for (i = 0; i < numbufs; i++) {
1189 			uint64_t tocpy;
1190 			int64_t bufoff;
1191 			dmu_buf_t *db = dbp[i];
1192 
1193 			ASSERT(size > 0);
1194 
1195 			bufoff = offset - db->db_offset;
1196 			tocpy = MIN(db->db_size - bufoff, size);
1197 
1198 			(void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1199 
1200 			offset += tocpy;
1201 			size -= tocpy;
1202 			buf = (char *)buf + tocpy;
1203 		}
1204 		dmu_buf_rele_array(dbp, numbufs, FTAG);
1205 	}
1206 	return (err);
1207 }
1208 
1209 int
1210 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1211     void *buf, uint32_t flags)
1212 {
1213 	dnode_t *dn;
1214 	int err;
1215 
1216 	err = dnode_hold(os, object, FTAG, &dn);
1217 	if (err != 0)
1218 		return (err);
1219 
1220 	err = dmu_read_impl(dn, offset, size, buf, flags);
1221 	dnode_rele(dn, FTAG);
1222 	return (err);
1223 }
1224 
1225 int
1226 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1227     uint32_t flags)
1228 {
1229 	return (dmu_read_impl(dn, offset, size, buf, flags));
1230 }
1231 
1232 static void
1233 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1234     const void *buf, dmu_tx_t *tx)
1235 {
1236 	int i;
1237 
1238 	for (i = 0; i < numbufs; i++) {
1239 		uint64_t tocpy;
1240 		int64_t bufoff;
1241 		dmu_buf_t *db = dbp[i];
1242 
1243 		ASSERT(size > 0);
1244 
1245 		bufoff = offset - db->db_offset;
1246 		tocpy = MIN(db->db_size - bufoff, size);
1247 
1248 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1249 
1250 		if (tocpy == db->db_size)
1251 			dmu_buf_will_fill(db, tx, B_FALSE);
1252 		else
1253 			dmu_buf_will_dirty(db, tx);
1254 
1255 		(void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1256 
1257 		if (tocpy == db->db_size)
1258 			dmu_buf_fill_done(db, tx, B_FALSE);
1259 
1260 		offset += tocpy;
1261 		size -= tocpy;
1262 		buf = (char *)buf + tocpy;
1263 	}
1264 }
1265 
1266 void
1267 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1268     const void *buf, dmu_tx_t *tx)
1269 {
1270 	dmu_buf_t **dbp;
1271 	int numbufs;
1272 
1273 	if (size == 0)
1274 		return;
1275 
1276 	VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1277 	    FALSE, FTAG, &numbufs, &dbp));
1278 	dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1279 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1280 }
1281 
1282 /*
1283  * Note: Lustre is an external consumer of this interface.
1284  */
1285 void
1286 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1287     const void *buf, dmu_tx_t *tx)
1288 {
1289 	dmu_buf_t **dbp;
1290 	int numbufs;
1291 
1292 	if (size == 0)
1293 		return;
1294 
1295 	VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1296 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1297 	dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1298 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1299 }
1300 
1301 void
1302 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1303     dmu_tx_t *tx)
1304 {
1305 	dmu_buf_t **dbp;
1306 	int numbufs, i;
1307 
1308 	if (size == 0)
1309 		return;
1310 
1311 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1312 	    FALSE, FTAG, &numbufs, &dbp));
1313 
1314 	for (i = 0; i < numbufs; i++) {
1315 		dmu_buf_t *db = dbp[i];
1316 
1317 		dmu_buf_will_not_fill(db, tx);
1318 	}
1319 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1320 }
1321 
1322 void
1323 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1324     void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1325     int compressed_size, int byteorder, dmu_tx_t *tx)
1326 {
1327 	dmu_buf_t *db;
1328 
1329 	ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1330 	ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1331 	VERIFY0(dmu_buf_hold_noread(os, object, offset,
1332 	    FTAG, &db));
1333 
1334 	dmu_buf_write_embedded(db,
1335 	    data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1336 	    uncompressed_size, compressed_size, byteorder, tx);
1337 
1338 	dmu_buf_rele(db, FTAG);
1339 }
1340 
1341 void
1342 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1343     dmu_tx_t *tx)
1344 {
1345 	int numbufs, i;
1346 	dmu_buf_t **dbp;
1347 
1348 	VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1349 	    &numbufs, &dbp));
1350 	for (i = 0; i < numbufs; i++)
1351 		dmu_buf_redact(dbp[i], tx);
1352 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1353 }
1354 
1355 #ifdef _KERNEL
1356 int
1357 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size)
1358 {
1359 	dmu_buf_t **dbp;
1360 	int numbufs, i, err;
1361 
1362 	/*
1363 	 * NB: we could do this block-at-a-time, but it's nice
1364 	 * to be reading in parallel.
1365 	 */
1366 	err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1367 	    TRUE, FTAG, &numbufs, &dbp, 0);
1368 	if (err)
1369 		return (err);
1370 
1371 	for (i = 0; i < numbufs; i++) {
1372 		uint64_t tocpy;
1373 		int64_t bufoff;
1374 		dmu_buf_t *db = dbp[i];
1375 
1376 		ASSERT(size > 0);
1377 
1378 		bufoff = zfs_uio_offset(uio) - db->db_offset;
1379 		tocpy = MIN(db->db_size - bufoff, size);
1380 
1381 		err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1382 		    UIO_READ, uio);
1383 
1384 		if (err)
1385 			break;
1386 
1387 		size -= tocpy;
1388 	}
1389 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1390 
1391 	return (err);
1392 }
1393 
1394 /*
1395  * Read 'size' bytes into the uio buffer.
1396  * From object zdb->db_object.
1397  * Starting at zfs_uio_offset(uio).
1398  *
1399  * If the caller already has a dbuf in the target object
1400  * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1401  * because we don't have to find the dnode_t for the object.
1402  */
1403 int
1404 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size)
1405 {
1406 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1407 	int err;
1408 
1409 	if (size == 0)
1410 		return (0);
1411 
1412 	DB_DNODE_ENTER(db);
1413 	err = dmu_read_uio_dnode(DB_DNODE(db), uio, size);
1414 	DB_DNODE_EXIT(db);
1415 
1416 	return (err);
1417 }
1418 
1419 /*
1420  * Read 'size' bytes into the uio buffer.
1421  * From the specified object
1422  * Starting at offset zfs_uio_offset(uio).
1423  */
1424 int
1425 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size)
1426 {
1427 	dnode_t *dn;
1428 	int err;
1429 
1430 	if (size == 0)
1431 		return (0);
1432 
1433 	err = dnode_hold(os, object, FTAG, &dn);
1434 	if (err)
1435 		return (err);
1436 
1437 	err = dmu_read_uio_dnode(dn, uio, size);
1438 
1439 	dnode_rele(dn, FTAG);
1440 
1441 	return (err);
1442 }
1443 
1444 int
1445 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx)
1446 {
1447 	dmu_buf_t **dbp;
1448 	int numbufs;
1449 	int err = 0;
1450 	int i;
1451 
1452 	err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1453 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1454 	if (err)
1455 		return (err);
1456 
1457 	for (i = 0; i < numbufs; i++) {
1458 		uint64_t tocpy;
1459 		int64_t bufoff;
1460 		dmu_buf_t *db = dbp[i];
1461 
1462 		ASSERT(size > 0);
1463 
1464 		offset_t off = zfs_uio_offset(uio);
1465 		bufoff = off - db->db_offset;
1466 		tocpy = MIN(db->db_size - bufoff, size);
1467 
1468 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1469 
1470 		if (tocpy == db->db_size)
1471 			dmu_buf_will_fill(db, tx, B_TRUE);
1472 		else
1473 			dmu_buf_will_dirty(db, tx);
1474 
1475 		err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1476 		    tocpy, UIO_WRITE, uio);
1477 
1478 		if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) {
1479 			/* The fill was reverted.  Undo any uio progress. */
1480 			zfs_uio_advance(uio, off - zfs_uio_offset(uio));
1481 		}
1482 
1483 		if (err)
1484 			break;
1485 
1486 		size -= tocpy;
1487 	}
1488 
1489 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1490 	return (err);
1491 }
1492 
1493 /*
1494  * Write 'size' bytes from the uio buffer.
1495  * To object zdb->db_object.
1496  * Starting at offset zfs_uio_offset(uio).
1497  *
1498  * If the caller already has a dbuf in the target object
1499  * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1500  * because we don't have to find the dnode_t for the object.
1501  */
1502 int
1503 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1504     dmu_tx_t *tx)
1505 {
1506 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1507 	int err;
1508 
1509 	if (size == 0)
1510 		return (0);
1511 
1512 	DB_DNODE_ENTER(db);
1513 	err = dmu_write_uio_dnode(DB_DNODE(db), uio, size, tx);
1514 	DB_DNODE_EXIT(db);
1515 
1516 	return (err);
1517 }
1518 
1519 /*
1520  * Write 'size' bytes from the uio buffer.
1521  * To the specified object.
1522  * Starting at offset zfs_uio_offset(uio).
1523  */
1524 int
1525 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1526     dmu_tx_t *tx)
1527 {
1528 	dnode_t *dn;
1529 	int err;
1530 
1531 	if (size == 0)
1532 		return (0);
1533 
1534 	err = dnode_hold(os, object, FTAG, &dn);
1535 	if (err)
1536 		return (err);
1537 
1538 	err = dmu_write_uio_dnode(dn, uio, size, tx);
1539 
1540 	dnode_rele(dn, FTAG);
1541 
1542 	return (err);
1543 }
1544 #endif /* _KERNEL */
1545 
1546 static void
1547 dmu_cached_bps(spa_t *spa, blkptr_t *bps, uint_t nbps,
1548     uint64_t *l1sz, uint64_t *l2sz)
1549 {
1550 	int cached_flags;
1551 
1552 	if (bps == NULL)
1553 		return;
1554 
1555 	for (size_t blk_off = 0; blk_off < nbps; blk_off++) {
1556 		blkptr_t *bp = &bps[blk_off];
1557 
1558 		if (BP_IS_HOLE(bp))
1559 			continue;
1560 
1561 		cached_flags = arc_cached(spa, bp);
1562 		if (cached_flags == 0)
1563 			continue;
1564 
1565 		if ((cached_flags & (ARC_CACHED_IN_L1 | ARC_CACHED_IN_L2)) ==
1566 		    ARC_CACHED_IN_L2)
1567 			*l2sz += BP_GET_LSIZE(bp);
1568 		else
1569 			*l1sz += BP_GET_LSIZE(bp);
1570 	}
1571 }
1572 
1573 /*
1574  * Estimate DMU object cached size.
1575  */
1576 int
1577 dmu_object_cached_size(objset_t *os, uint64_t object,
1578     uint64_t *l1sz, uint64_t *l2sz)
1579 {
1580 	dnode_t *dn;
1581 	dmu_object_info_t doi;
1582 	int err = 0;
1583 
1584 	*l1sz = *l2sz = 0;
1585 
1586 	if (dnode_hold(os, object, FTAG, &dn) != 0)
1587 		return (0);
1588 
1589 	if (dn->dn_nlevels < 2) {
1590 		dnode_rele(dn, FTAG);
1591 		return (0);
1592 	}
1593 
1594 	dmu_object_info_from_dnode(dn, &doi);
1595 
1596 	for (uint64_t off = 0; off < doi.doi_max_offset;
1597 	    off += dmu_prefetch_max) {
1598 		/* dbuf_read doesn't prefetch L1 blocks. */
1599 		dmu_prefetch_by_dnode(dn, 1, off,
1600 		    dmu_prefetch_max, ZIO_PRIORITY_SYNC_READ);
1601 	}
1602 
1603 	/*
1604 	 * Hold all valid L1 blocks, asking ARC the status of each BP
1605 	 * contained in each such L1 block.
1606 	 */
1607 	uint_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1);
1608 	uint64_t l1blks = 1 + (dn->dn_maxblkid / nbps);
1609 
1610 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1611 	for (uint64_t blk = 0; blk < l1blks; blk++) {
1612 		dmu_buf_impl_t *db = NULL;
1613 
1614 		if (issig()) {
1615 			/*
1616 			 * On interrupt, get out, and bubble up EINTR
1617 			 */
1618 			err = EINTR;
1619 			break;
1620 		}
1621 
1622 		/*
1623 		 * If we get an i/o error here, the L1 can't be read,
1624 		 * and nothing under it could be cached, so we just
1625 		 * continue. Ignoring the error from dbuf_hold_impl
1626 		 * or from dbuf_read is then a reasonable choice.
1627 		 */
1628 		err = dbuf_hold_impl(dn, 1, blk, B_TRUE, B_FALSE, FTAG, &db);
1629 		if (err != 0) {
1630 			/*
1631 			 * ignore error and continue
1632 			 */
1633 			err = 0;
1634 			continue;
1635 		}
1636 
1637 		err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1638 		if (err == 0) {
1639 			dmu_cached_bps(dmu_objset_spa(os), db->db.db_data,
1640 			    nbps, l1sz, l2sz);
1641 		}
1642 		/*
1643 		 * error may be ignored, and we continue
1644 		 */
1645 		err = 0;
1646 		dbuf_rele(db, FTAG);
1647 	}
1648 	rw_exit(&dn->dn_struct_rwlock);
1649 
1650 	dnode_rele(dn, FTAG);
1651 	return (err);
1652 }
1653 
1654 /*
1655  * Allocate a loaned anonymous arc buffer.
1656  */
1657 arc_buf_t *
1658 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1659 {
1660 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1661 
1662 	return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1663 }
1664 
1665 /*
1666  * Free a loaned arc buffer.
1667  */
1668 void
1669 dmu_return_arcbuf(arc_buf_t *buf)
1670 {
1671 	arc_return_buf(buf, FTAG);
1672 	arc_buf_destroy(buf, FTAG);
1673 }
1674 
1675 /*
1676  * A "lightweight" write is faster than a regular write (e.g.
1677  * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1678  * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t.  However, the
1679  * data can not be read or overwritten until the transaction's txg has been
1680  * synced.  This makes it appropriate for workloads that are known to be
1681  * (temporarily) write-only, like "zfs receive".
1682  *
1683  * A single block is written, starting at the specified offset in bytes.  If
1684  * the call is successful, it returns 0 and the provided abd has been
1685  * consumed (the caller should not free it).
1686  */
1687 int
1688 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1689     const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx)
1690 {
1691 	dbuf_dirty_record_t *dr =
1692 	    dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
1693 	if (dr == NULL)
1694 		return (SET_ERROR(EIO));
1695 	dr->dt.dll.dr_abd = abd;
1696 	dr->dt.dll.dr_props = *zp;
1697 	dr->dt.dll.dr_flags = flags;
1698 	return (0);
1699 }
1700 
1701 /*
1702  * When possible directly assign passed loaned arc buffer to a dbuf.
1703  * If this is not possible copy the contents of passed arc buf via
1704  * dmu_write().
1705  */
1706 int
1707 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1708     dmu_tx_t *tx)
1709 {
1710 	dmu_buf_impl_t *db;
1711 	objset_t *os = dn->dn_objset;
1712 	uint64_t object = dn->dn_object;
1713 	uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1714 	uint64_t blkid;
1715 
1716 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1717 	blkid = dbuf_whichblock(dn, 0, offset);
1718 	db = dbuf_hold(dn, blkid, FTAG);
1719 	rw_exit(&dn->dn_struct_rwlock);
1720 	if (db == NULL)
1721 		return (SET_ERROR(EIO));
1722 
1723 	/*
1724 	 * We can only assign if the offset is aligned and the arc buf is the
1725 	 * same size as the dbuf.
1726 	 */
1727 	if (offset == db->db.db_offset && blksz == db->db.db_size) {
1728 		zfs_racct_write(blksz, 1);
1729 		dbuf_assign_arcbuf(db, buf, tx);
1730 		dbuf_rele(db, FTAG);
1731 	} else {
1732 		/* compressed bufs must always be assignable to their dbuf */
1733 		ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1734 		ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1735 
1736 		dbuf_rele(db, FTAG);
1737 		dmu_write(os, object, offset, blksz, buf->b_data, tx);
1738 		dmu_return_arcbuf(buf);
1739 	}
1740 
1741 	return (0);
1742 }
1743 
1744 int
1745 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1746     dmu_tx_t *tx)
1747 {
1748 	int err;
1749 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1750 
1751 	DB_DNODE_ENTER(db);
1752 	err = dmu_assign_arcbuf_by_dnode(DB_DNODE(db), offset, buf, tx);
1753 	DB_DNODE_EXIT(db);
1754 
1755 	return (err);
1756 }
1757 
1758 typedef struct {
1759 	dbuf_dirty_record_t	*dsa_dr;
1760 	dmu_sync_cb_t		*dsa_done;
1761 	zgd_t			*dsa_zgd;
1762 	dmu_tx_t		*dsa_tx;
1763 } dmu_sync_arg_t;
1764 
1765 static void
1766 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1767 {
1768 	(void) buf;
1769 	dmu_sync_arg_t *dsa = varg;
1770 	dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1771 	blkptr_t *bp = zio->io_bp;
1772 
1773 	if (zio->io_error == 0) {
1774 		if (BP_IS_HOLE(bp)) {
1775 			/*
1776 			 * A block of zeros may compress to a hole, but the
1777 			 * block size still needs to be known for replay.
1778 			 */
1779 			BP_SET_LSIZE(bp, db->db_size);
1780 		} else if (!BP_IS_EMBEDDED(bp)) {
1781 			ASSERT(BP_GET_LEVEL(bp) == 0);
1782 			BP_SET_FILL(bp, 1);
1783 		}
1784 	}
1785 }
1786 
1787 static void
1788 dmu_sync_late_arrival_ready(zio_t *zio)
1789 {
1790 	dmu_sync_ready(zio, NULL, zio->io_private);
1791 }
1792 
1793 static void
1794 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1795 {
1796 	(void) buf;
1797 	dmu_sync_arg_t *dsa = varg;
1798 	dbuf_dirty_record_t *dr = dsa->dsa_dr;
1799 	dmu_buf_impl_t *db = dr->dr_dbuf;
1800 	zgd_t *zgd = dsa->dsa_zgd;
1801 
1802 	/*
1803 	 * Record the vdev(s) backing this blkptr so they can be flushed after
1804 	 * the writes for the lwb have completed.
1805 	 */
1806 	if (zio->io_error == 0) {
1807 		zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1808 	}
1809 
1810 	mutex_enter(&db->db_mtx);
1811 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1812 	if (zio->io_error == 0) {
1813 		dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1814 		if (dr->dt.dl.dr_nopwrite) {
1815 			blkptr_t *bp = zio->io_bp;
1816 			blkptr_t *bp_orig = &zio->io_bp_orig;
1817 			uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1818 
1819 			ASSERT(BP_EQUAL(bp, bp_orig));
1820 			VERIFY(BP_EQUAL(bp, db->db_blkptr));
1821 			ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1822 			VERIFY(zio_checksum_table[chksum].ci_flags &
1823 			    ZCHECKSUM_FLAG_NOPWRITE);
1824 		}
1825 		dr->dt.dl.dr_overridden_by = *zio->io_bp;
1826 		dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1827 		dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1828 
1829 		/*
1830 		 * Old style holes are filled with all zeros, whereas
1831 		 * new-style holes maintain their lsize, type, level,
1832 		 * and birth time (see zio_write_compress). While we
1833 		 * need to reset the BP_SET_LSIZE() call that happened
1834 		 * in dmu_sync_ready for old style holes, we do *not*
1835 		 * want to wipe out the information contained in new
1836 		 * style holes. Thus, only zero out the block pointer if
1837 		 * it's an old style hole.
1838 		 */
1839 		if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1840 		    BP_GET_LOGICAL_BIRTH(&dr->dt.dl.dr_overridden_by) == 0)
1841 			BP_ZERO(&dr->dt.dl.dr_overridden_by);
1842 	} else {
1843 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1844 	}
1845 	cv_broadcast(&db->db_changed);
1846 	mutex_exit(&db->db_mtx);
1847 
1848 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1849 
1850 	kmem_free(dsa, sizeof (*dsa));
1851 }
1852 
1853 static void
1854 dmu_sync_late_arrival_done(zio_t *zio)
1855 {
1856 	blkptr_t *bp = zio->io_bp;
1857 	dmu_sync_arg_t *dsa = zio->io_private;
1858 	zgd_t *zgd = dsa->dsa_zgd;
1859 
1860 	if (zio->io_error == 0) {
1861 		/*
1862 		 * Record the vdev(s) backing this blkptr so they can be
1863 		 * flushed after the writes for the lwb have completed.
1864 		 */
1865 		zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1866 
1867 		if (!BP_IS_HOLE(bp)) {
1868 			blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
1869 			ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1870 			ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1871 			ASSERT(BP_GET_LOGICAL_BIRTH(zio->io_bp) == zio->io_txg);
1872 			ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1873 			zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1874 		}
1875 	}
1876 
1877 	dmu_tx_commit(dsa->dsa_tx);
1878 
1879 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1880 
1881 	abd_free(zio->io_abd);
1882 	kmem_free(dsa, sizeof (*dsa));
1883 }
1884 
1885 static int
1886 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1887     zio_prop_t *zp, zbookmark_phys_t *zb)
1888 {
1889 	dmu_sync_arg_t *dsa;
1890 	dmu_tx_t *tx;
1891 	int error;
1892 
1893 	error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL,
1894 	    DB_RF_CANFAIL | DB_RF_NOPREFETCH);
1895 	if (error != 0)
1896 		return (error);
1897 
1898 	tx = dmu_tx_create(os);
1899 	dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1900 	/*
1901 	 * This transaction does not produce any dirty data or log blocks, so
1902 	 * it should not be throttled.  All other cases wait for TXG sync, by
1903 	 * which time the log block we are writing will be obsolete, so we can
1904 	 * skip waiting and just return error here instead.
1905 	 */
1906 	if (dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE) != 0) {
1907 		dmu_tx_abort(tx);
1908 		/* Make zl_get_data do txg_waited_synced() */
1909 		return (SET_ERROR(EIO));
1910 	}
1911 
1912 	/*
1913 	 * In order to prevent the zgd's lwb from being free'd prior to
1914 	 * dmu_sync_late_arrival_done() being called, we have to ensure
1915 	 * the lwb's "max txg" takes this tx's txg into account.
1916 	 */
1917 	zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1918 
1919 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1920 	dsa->dsa_dr = NULL;
1921 	dsa->dsa_done = done;
1922 	dsa->dsa_zgd = zgd;
1923 	dsa->dsa_tx = tx;
1924 
1925 	/*
1926 	 * Since we are currently syncing this txg, it's nontrivial to
1927 	 * determine what BP to nopwrite against, so we disable nopwrite.
1928 	 *
1929 	 * When syncing, the db_blkptr is initially the BP of the previous
1930 	 * txg.  We can not nopwrite against it because it will be changed
1931 	 * (this is similar to the non-late-arrival case where the dbuf is
1932 	 * dirty in a future txg).
1933 	 *
1934 	 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1935 	 * We can not nopwrite against it because although the BP will not
1936 	 * (typically) be changed, the data has not yet been persisted to this
1937 	 * location.
1938 	 *
1939 	 * Finally, when dbuf_write_done() is called, it is theoretically
1940 	 * possible to always nopwrite, because the data that was written in
1941 	 * this txg is the same data that we are trying to write.  However we
1942 	 * would need to check that this dbuf is not dirty in any future
1943 	 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1944 	 * don't nopwrite in this case.
1945 	 */
1946 	zp->zp_nopwrite = B_FALSE;
1947 
1948 	zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1949 	    abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1950 	    zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1951 	    dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done,
1952 	    dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1953 
1954 	return (0);
1955 }
1956 
1957 /*
1958  * Intent log support: sync the block associated with db to disk.
1959  * N.B. and XXX: the caller is responsible for making sure that the
1960  * data isn't changing while dmu_sync() is writing it.
1961  *
1962  * Return values:
1963  *
1964  *	EEXIST: this txg has already been synced, so there's nothing to do.
1965  *		The caller should not log the write.
1966  *
1967  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1968  *		The caller should not log the write.
1969  *
1970  *	EALREADY: this block is already in the process of being synced.
1971  *		The caller should track its progress (somehow).
1972  *
1973  *	EIO: could not do the I/O.
1974  *		The caller should do a txg_wait_synced().
1975  *
1976  *	0: the I/O has been initiated.
1977  *		The caller should log this blkptr in the done callback.
1978  *		It is possible that the I/O will fail, in which case
1979  *		the error will be reported to the done callback and
1980  *		propagated to pio from zio_done().
1981  */
1982 int
1983 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1984 {
1985 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1986 	objset_t *os = db->db_objset;
1987 	dsl_dataset_t *ds = os->os_dsl_dataset;
1988 	dbuf_dirty_record_t *dr, *dr_next;
1989 	dmu_sync_arg_t *dsa;
1990 	zbookmark_phys_t zb;
1991 	zio_prop_t zp;
1992 
1993 	ASSERT(pio != NULL);
1994 	ASSERT(txg != 0);
1995 
1996 	SET_BOOKMARK(&zb, ds->ds_object,
1997 	    db->db.db_object, db->db_level, db->db_blkid);
1998 
1999 	DB_DNODE_ENTER(db);
2000 	dmu_write_policy(os, DB_DNODE(db), db->db_level, WP_DMU_SYNC, &zp);
2001 	DB_DNODE_EXIT(db);
2002 
2003 	/*
2004 	 * If we're frozen (running ziltest), we always need to generate a bp.
2005 	 */
2006 	if (txg > spa_freeze_txg(os->os_spa))
2007 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2008 
2009 	/*
2010 	 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
2011 	 * and us.  If we determine that this txg is not yet syncing,
2012 	 * but it begins to sync a moment later, that's OK because the
2013 	 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
2014 	 */
2015 	mutex_enter(&db->db_mtx);
2016 
2017 	if (txg <= spa_last_synced_txg(os->os_spa)) {
2018 		/*
2019 		 * This txg has already synced.  There's nothing to do.
2020 		 */
2021 		mutex_exit(&db->db_mtx);
2022 		return (SET_ERROR(EEXIST));
2023 	}
2024 
2025 	if (txg <= spa_syncing_txg(os->os_spa)) {
2026 		/*
2027 		 * This txg is currently syncing, so we can't mess with
2028 		 * the dirty record anymore; just write a new log block.
2029 		 */
2030 		mutex_exit(&db->db_mtx);
2031 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2032 	}
2033 
2034 	dr = dbuf_find_dirty_eq(db, txg);
2035 
2036 	if (dr == NULL) {
2037 		/*
2038 		 * There's no dr for this dbuf, so it must have been freed.
2039 		 * There's no need to log writes to freed blocks, so we're done.
2040 		 */
2041 		mutex_exit(&db->db_mtx);
2042 		return (SET_ERROR(ENOENT));
2043 	}
2044 
2045 	dr_next = list_next(&db->db_dirty_records, dr);
2046 	ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
2047 
2048 	if (db->db_blkptr != NULL) {
2049 		/*
2050 		 * We need to fill in zgd_bp with the current blkptr so that
2051 		 * the nopwrite code can check if we're writing the same
2052 		 * data that's already on disk.  We can only nopwrite if we
2053 		 * are sure that after making the copy, db_blkptr will not
2054 		 * change until our i/o completes.  We ensure this by
2055 		 * holding the db_mtx, and only allowing nopwrite if the
2056 		 * block is not already dirty (see below).  This is verified
2057 		 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2058 		 * not changed.
2059 		 */
2060 		*zgd->zgd_bp = *db->db_blkptr;
2061 	}
2062 
2063 	/*
2064 	 * Assume the on-disk data is X, the current syncing data (in
2065 	 * txg - 1) is Y, and the current in-memory data is Z (currently
2066 	 * in dmu_sync).
2067 	 *
2068 	 * We usually want to perform a nopwrite if X and Z are the
2069 	 * same.  However, if Y is different (i.e. the BP is going to
2070 	 * change before this write takes effect), then a nopwrite will
2071 	 * be incorrect - we would override with X, which could have
2072 	 * been freed when Y was written.
2073 	 *
2074 	 * (Note that this is not a concern when we are nop-writing from
2075 	 * syncing context, because X and Y must be identical, because
2076 	 * all previous txgs have been synced.)
2077 	 *
2078 	 * Therefore, we disable nopwrite if the current BP could change
2079 	 * before this TXG.  There are two ways it could change: by
2080 	 * being dirty (dr_next is non-NULL), or by being freed
2081 	 * (dnode_block_freed()).  This behavior is verified by
2082 	 * zio_done(), which VERIFYs that the override BP is identical
2083 	 * to the on-disk BP.
2084 	 */
2085 	if (dr_next != NULL) {
2086 		zp.zp_nopwrite = B_FALSE;
2087 	} else {
2088 		DB_DNODE_ENTER(db);
2089 		if (dnode_block_freed(DB_DNODE(db), db->db_blkid))
2090 			zp.zp_nopwrite = B_FALSE;
2091 		DB_DNODE_EXIT(db);
2092 	}
2093 
2094 	ASSERT(dr->dr_txg == txg);
2095 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
2096 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2097 		/*
2098 		 * We have already issued a sync write for this buffer,
2099 		 * or this buffer has already been synced.  It could not
2100 		 * have been dirtied since, or we would have cleared the state.
2101 		 */
2102 		mutex_exit(&db->db_mtx);
2103 		return (SET_ERROR(EALREADY));
2104 	}
2105 
2106 	ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2107 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
2108 	mutex_exit(&db->db_mtx);
2109 
2110 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2111 	dsa->dsa_dr = dr;
2112 	dsa->dsa_done = done;
2113 	dsa->dsa_zgd = zgd;
2114 	dsa->dsa_tx = NULL;
2115 
2116 	zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp,
2117 	    dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), dbuf_is_l2cacheable(db),
2118 	    &zp, dmu_sync_ready, NULL, dmu_sync_done, dsa,
2119 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
2120 
2121 	return (0);
2122 }
2123 
2124 int
2125 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
2126 {
2127 	dnode_t *dn;
2128 	int err;
2129 
2130 	err = dnode_hold(os, object, FTAG, &dn);
2131 	if (err)
2132 		return (err);
2133 	err = dnode_set_nlevels(dn, nlevels, tx);
2134 	dnode_rele(dn, FTAG);
2135 	return (err);
2136 }
2137 
2138 int
2139 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
2140     dmu_tx_t *tx)
2141 {
2142 	dnode_t *dn;
2143 	int err;
2144 
2145 	err = dnode_hold(os, object, FTAG, &dn);
2146 	if (err)
2147 		return (err);
2148 	err = dnode_set_blksz(dn, size, ibs, tx);
2149 	dnode_rele(dn, FTAG);
2150 	return (err);
2151 }
2152 
2153 int
2154 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
2155     dmu_tx_t *tx)
2156 {
2157 	dnode_t *dn;
2158 	int err;
2159 
2160 	err = dnode_hold(os, object, FTAG, &dn);
2161 	if (err)
2162 		return (err);
2163 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2164 	dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE);
2165 	rw_exit(&dn->dn_struct_rwlock);
2166 	dnode_rele(dn, FTAG);
2167 	return (0);
2168 }
2169 
2170 void
2171 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2172     dmu_tx_t *tx)
2173 {
2174 	dnode_t *dn;
2175 
2176 	/*
2177 	 * Send streams include each object's checksum function.  This
2178 	 * check ensures that the receiving system can understand the
2179 	 * checksum function transmitted.
2180 	 */
2181 	ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2182 
2183 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
2184 	ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2185 	dn->dn_checksum = checksum;
2186 	dnode_setdirty(dn, tx);
2187 	dnode_rele(dn, FTAG);
2188 }
2189 
2190 void
2191 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2192     dmu_tx_t *tx)
2193 {
2194 	dnode_t *dn;
2195 
2196 	/*
2197 	 * Send streams include each object's compression function.  This
2198 	 * check ensures that the receiving system can understand the
2199 	 * compression function transmitted.
2200 	 */
2201 	ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2202 
2203 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
2204 	dn->dn_compress = compress;
2205 	dnode_setdirty(dn, tx);
2206 	dnode_rele(dn, FTAG);
2207 }
2208 
2209 /*
2210  * When the "redundant_metadata" property is set to "most", only indirect
2211  * blocks of this level and higher will have an additional ditto block.
2212  */
2213 static const int zfs_redundant_metadata_most_ditto_level = 2;
2214 
2215 void
2216 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2217 {
2218 	dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2219 	boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2220 	    (wp & WP_SPILL));
2221 	enum zio_checksum checksum = os->os_checksum;
2222 	enum zio_compress compress = os->os_compress;
2223 	uint8_t complevel = os->os_complevel;
2224 	enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2225 	boolean_t dedup = B_FALSE;
2226 	boolean_t nopwrite = B_FALSE;
2227 	boolean_t dedup_verify = os->os_dedup_verify;
2228 	boolean_t encrypt = B_FALSE;
2229 	int copies = os->os_copies;
2230 
2231 	/*
2232 	 * We maintain different write policies for each of the following
2233 	 * types of data:
2234 	 *	 1. metadata
2235 	 *	 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2236 	 *	 3. all other level 0 blocks
2237 	 */
2238 	if (ismd) {
2239 		/*
2240 		 * XXX -- we should design a compression algorithm
2241 		 * that specializes in arrays of bps.
2242 		 */
2243 		compress = zio_compress_select(os->os_spa,
2244 		    ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2245 
2246 		/*
2247 		 * Metadata always gets checksummed.  If the data
2248 		 * checksum is multi-bit correctable, and it's not a
2249 		 * ZBT-style checksum, then it's suitable for metadata
2250 		 * as well.  Otherwise, the metadata checksum defaults
2251 		 * to fletcher4.
2252 		 */
2253 		if (!(zio_checksum_table[checksum].ci_flags &
2254 		    ZCHECKSUM_FLAG_METADATA) ||
2255 		    (zio_checksum_table[checksum].ci_flags &
2256 		    ZCHECKSUM_FLAG_EMBEDDED))
2257 			checksum = ZIO_CHECKSUM_FLETCHER_4;
2258 
2259 		switch (os->os_redundant_metadata) {
2260 		case ZFS_REDUNDANT_METADATA_ALL:
2261 			copies++;
2262 			break;
2263 		case ZFS_REDUNDANT_METADATA_MOST:
2264 			if (level >= zfs_redundant_metadata_most_ditto_level ||
2265 			    DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))
2266 				copies++;
2267 			break;
2268 		case ZFS_REDUNDANT_METADATA_SOME:
2269 			if (DMU_OT_IS_CRITICAL(type))
2270 				copies++;
2271 			break;
2272 		case ZFS_REDUNDANT_METADATA_NONE:
2273 			break;
2274 		}
2275 	} else if (wp & WP_NOFILL) {
2276 		ASSERT(level == 0);
2277 
2278 		/*
2279 		 * If we're writing preallocated blocks, we aren't actually
2280 		 * writing them so don't set any policy properties.  These
2281 		 * blocks are currently only used by an external subsystem
2282 		 * outside of zfs (i.e. dump) and not written by the zio
2283 		 * pipeline.
2284 		 */
2285 		compress = ZIO_COMPRESS_OFF;
2286 		checksum = ZIO_CHECKSUM_OFF;
2287 	} else {
2288 		compress = zio_compress_select(os->os_spa, dn->dn_compress,
2289 		    compress);
2290 		complevel = zio_complevel_select(os->os_spa, compress,
2291 		    complevel, complevel);
2292 
2293 		checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2294 		    zio_checksum_select(dn->dn_checksum, checksum) :
2295 		    dedup_checksum;
2296 
2297 		/*
2298 		 * Determine dedup setting.  If we are in dmu_sync(),
2299 		 * we won't actually dedup now because that's all
2300 		 * done in syncing context; but we do want to use the
2301 		 * dedup checksum.  If the checksum is not strong
2302 		 * enough to ensure unique signatures, force
2303 		 * dedup_verify.
2304 		 */
2305 		if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2306 			dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2307 			if (!(zio_checksum_table[checksum].ci_flags &
2308 			    ZCHECKSUM_FLAG_DEDUP))
2309 				dedup_verify = B_TRUE;
2310 		}
2311 
2312 		/*
2313 		 * Enable nopwrite if we have secure enough checksum
2314 		 * algorithm (see comment in zio_nop_write) and
2315 		 * compression is enabled.  We don't enable nopwrite if
2316 		 * dedup is enabled as the two features are mutually
2317 		 * exclusive.
2318 		 */
2319 		nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2320 		    ZCHECKSUM_FLAG_NOPWRITE) &&
2321 		    compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2322 	}
2323 
2324 	/*
2325 	 * All objects in an encrypted objset are protected from modification
2326 	 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2327 	 * in the bp, so we cannot use all copies. Encrypted objects are also
2328 	 * not subject to nopwrite since writing the same data will still
2329 	 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2330 	 * to avoid ambiguity in the dedup code since the DDT does not store
2331 	 * object types.
2332 	 */
2333 	if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2334 		encrypt = B_TRUE;
2335 
2336 		if (DMU_OT_IS_ENCRYPTED(type)) {
2337 			copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2338 			nopwrite = B_FALSE;
2339 		} else {
2340 			dedup = B_FALSE;
2341 		}
2342 
2343 		if (level <= 0 &&
2344 		    (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2345 			compress = ZIO_COMPRESS_EMPTY;
2346 		}
2347 	}
2348 
2349 	zp->zp_compress = compress;
2350 	zp->zp_complevel = complevel;
2351 	zp->zp_checksum = checksum;
2352 	zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2353 	zp->zp_level = level;
2354 	zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2355 	zp->zp_dedup = dedup;
2356 	zp->zp_dedup_verify = dedup && dedup_verify;
2357 	zp->zp_nopwrite = nopwrite;
2358 	zp->zp_encrypt = encrypt;
2359 	zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2360 	memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN);
2361 	memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN);
2362 	memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN);
2363 	zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ?
2364 	    os->os_zpl_special_smallblock : 0;
2365 	zp->zp_storage_type = dn ? dn->dn_storage_type : DMU_OT_NONE;
2366 
2367 	ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2368 }
2369 
2370 /*
2371  * Reports the location of data and holes in an object.  In order to
2372  * accurately report holes all dirty data must be synced to disk.  This
2373  * causes extremely poor performance when seeking for holes in a dirty file.
2374  * As a compromise, only provide hole data when the dnode is clean.  When
2375  * a dnode is dirty report the dnode as having no holes by returning EBUSY
2376  * which is always safe to do.
2377  */
2378 int
2379 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2380 {
2381 	dnode_t *dn;
2382 	int restarted = 0, err;
2383 
2384 restart:
2385 	err = dnode_hold(os, object, FTAG, &dn);
2386 	if (err)
2387 		return (err);
2388 
2389 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2390 
2391 	if (dnode_is_dirty(dn)) {
2392 		/*
2393 		 * If the zfs_dmu_offset_next_sync module option is enabled
2394 		 * then hole reporting has been requested.  Dirty dnodes
2395 		 * must be synced to disk to accurately report holes.
2396 		 *
2397 		 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is
2398 		 * held by the caller only a single restart will be required.
2399 		 * We tolerate callers which do not hold the rangelock by
2400 		 * returning EBUSY and not reporting holes after one restart.
2401 		 */
2402 		if (zfs_dmu_offset_next_sync) {
2403 			rw_exit(&dn->dn_struct_rwlock);
2404 			dnode_rele(dn, FTAG);
2405 
2406 			if (restarted)
2407 				return (SET_ERROR(EBUSY));
2408 
2409 			txg_wait_synced(dmu_objset_pool(os), 0);
2410 			restarted = 1;
2411 			goto restart;
2412 		}
2413 
2414 		err = SET_ERROR(EBUSY);
2415 	} else {
2416 		err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK |
2417 		    (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2418 	}
2419 
2420 	rw_exit(&dn->dn_struct_rwlock);
2421 	dnode_rele(dn, FTAG);
2422 
2423 	return (err);
2424 }
2425 
2426 int
2427 dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2428     blkptr_t *bps, size_t *nbpsp)
2429 {
2430 	dmu_buf_t **dbp, *dbuf;
2431 	dmu_buf_impl_t *db;
2432 	blkptr_t *bp;
2433 	int error, numbufs;
2434 
2435 	error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2436 	    &numbufs, &dbp);
2437 	if (error != 0) {
2438 		if (error == ESRCH) {
2439 			error = SET_ERROR(ENXIO);
2440 		}
2441 		return (error);
2442 	}
2443 
2444 	ASSERT3U(numbufs, <=, *nbpsp);
2445 
2446 	for (int i = 0; i < numbufs; i++) {
2447 		dbuf = dbp[i];
2448 		db = (dmu_buf_impl_t *)dbuf;
2449 
2450 		mutex_enter(&db->db_mtx);
2451 
2452 		if (!list_is_empty(&db->db_dirty_records)) {
2453 			dbuf_dirty_record_t *dr;
2454 
2455 			dr = list_head(&db->db_dirty_records);
2456 			if (dr->dt.dl.dr_brtwrite) {
2457 				/*
2458 				 * This is very special case where we clone a
2459 				 * block and in the same transaction group we
2460 				 * read its BP (most likely to clone the clone).
2461 				 */
2462 				bp = &dr->dt.dl.dr_overridden_by;
2463 			} else {
2464 				/*
2465 				 * The block was modified in the same
2466 				 * transaction group.
2467 				 */
2468 				mutex_exit(&db->db_mtx);
2469 				error = SET_ERROR(EAGAIN);
2470 				goto out;
2471 			}
2472 		} else {
2473 			bp = db->db_blkptr;
2474 		}
2475 
2476 		mutex_exit(&db->db_mtx);
2477 
2478 		if (bp == NULL) {
2479 			/*
2480 			 * The file size was increased, but the block was never
2481 			 * written, otherwise we would either have the block
2482 			 * pointer or the dirty record and would not get here.
2483 			 * It is effectively a hole, so report it as such.
2484 			 */
2485 			BP_ZERO(&bps[i]);
2486 			continue;
2487 		}
2488 		/*
2489 		 * Make sure we clone only data blocks.
2490 		 */
2491 		if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) {
2492 			error = SET_ERROR(EINVAL);
2493 			goto out;
2494 		}
2495 
2496 		/*
2497 		 * If the block was allocated in transaction group that is not
2498 		 * yet synced, we could clone it, but we couldn't write this
2499 		 * operation into ZIL, or it may be impossible to replay, since
2500 		 * the block may appear not yet allocated at that point.
2501 		 */
2502 		if (BP_GET_BIRTH(bp) > spa_freeze_txg(os->os_spa)) {
2503 			error = SET_ERROR(EINVAL);
2504 			goto out;
2505 		}
2506 		if (BP_GET_BIRTH(bp) > spa_last_synced_txg(os->os_spa)) {
2507 			error = SET_ERROR(EAGAIN);
2508 			goto out;
2509 		}
2510 
2511 		bps[i] = *bp;
2512 	}
2513 
2514 	*nbpsp = numbufs;
2515 out:
2516 	dmu_buf_rele_array(dbp, numbufs, FTAG);
2517 
2518 	return (error);
2519 }
2520 
2521 int
2522 dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2523     dmu_tx_t *tx, const blkptr_t *bps, size_t nbps)
2524 {
2525 	spa_t *spa;
2526 	dmu_buf_t **dbp, *dbuf;
2527 	dmu_buf_impl_t *db;
2528 	struct dirty_leaf *dl;
2529 	dbuf_dirty_record_t *dr;
2530 	const blkptr_t *bp;
2531 	int error = 0, i, numbufs;
2532 
2533 	spa = os->os_spa;
2534 
2535 	VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2536 	    &numbufs, &dbp));
2537 	ASSERT3U(nbps, ==, numbufs);
2538 
2539 	/*
2540 	 * Before we start cloning make sure that the dbufs sizes match new BPs
2541 	 * sizes. If they don't, that's a no-go, as we are not able to shrink
2542 	 * dbufs.
2543 	 */
2544 	for (i = 0; i < numbufs; i++) {
2545 		dbuf = dbp[i];
2546 		db = (dmu_buf_impl_t *)dbuf;
2547 		bp = &bps[i];
2548 
2549 		ASSERT0(db->db_level);
2550 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2551 		ASSERT(db->db_blkid != DMU_SPILL_BLKID);
2552 
2553 		if (!BP_IS_HOLE(bp) && BP_GET_LSIZE(bp) != dbuf->db_size) {
2554 			error = SET_ERROR(EXDEV);
2555 			goto out;
2556 		}
2557 	}
2558 
2559 	for (i = 0; i < numbufs; i++) {
2560 		dbuf = dbp[i];
2561 		db = (dmu_buf_impl_t *)dbuf;
2562 		bp = &bps[i];
2563 
2564 		ASSERT0(db->db_level);
2565 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2566 		ASSERT(db->db_blkid != DMU_SPILL_BLKID);
2567 		ASSERT(BP_IS_HOLE(bp) || dbuf->db_size == BP_GET_LSIZE(bp));
2568 
2569 		dmu_buf_will_clone(dbuf, tx);
2570 
2571 		mutex_enter(&db->db_mtx);
2572 
2573 		dr = list_head(&db->db_dirty_records);
2574 		VERIFY(dr != NULL);
2575 		ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2576 		dl = &dr->dt.dl;
2577 		dl->dr_overridden_by = *bp;
2578 		if (!BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) != 0) {
2579 			if (!BP_IS_EMBEDDED(bp)) {
2580 				BP_SET_BIRTH(&dl->dr_overridden_by, dr->dr_txg,
2581 				    BP_GET_BIRTH(bp));
2582 			} else {
2583 				BP_SET_LOGICAL_BIRTH(&dl->dr_overridden_by,
2584 				    dr->dr_txg);
2585 			}
2586 		}
2587 		dl->dr_brtwrite = B_TRUE;
2588 		dl->dr_override_state = DR_OVERRIDDEN;
2589 
2590 		mutex_exit(&db->db_mtx);
2591 
2592 		/*
2593 		 * When data in embedded into BP there is no need to create
2594 		 * BRT entry as there is no data block. Just copy the BP as
2595 		 * it contains the data.
2596 		 */
2597 		if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
2598 			brt_pending_add(spa, bp, tx);
2599 		}
2600 	}
2601 out:
2602 	dmu_buf_rele_array(dbp, numbufs, FTAG);
2603 
2604 	return (error);
2605 }
2606 
2607 void
2608 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2609 {
2610 	dnode_phys_t *dnp = dn->dn_phys;
2611 
2612 	doi->doi_data_block_size = dn->dn_datablksz;
2613 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
2614 	    1ULL << dn->dn_indblkshift : 0;
2615 	doi->doi_type = dn->dn_type;
2616 	doi->doi_bonus_type = dn->dn_bonustype;
2617 	doi->doi_bonus_size = dn->dn_bonuslen;
2618 	doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2619 	doi->doi_indirection = dn->dn_nlevels;
2620 	doi->doi_checksum = dn->dn_checksum;
2621 	doi->doi_compress = dn->dn_compress;
2622 	doi->doi_nblkptr = dn->dn_nblkptr;
2623 	doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2624 	doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2625 	doi->doi_fill_count = 0;
2626 	for (int i = 0; i < dnp->dn_nblkptr; i++)
2627 		doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2628 }
2629 
2630 void
2631 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2632 {
2633 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2634 	mutex_enter(&dn->dn_mtx);
2635 
2636 	__dmu_object_info_from_dnode(dn, doi);
2637 
2638 	mutex_exit(&dn->dn_mtx);
2639 	rw_exit(&dn->dn_struct_rwlock);
2640 }
2641 
2642 /*
2643  * Get information on a DMU object.
2644  * If doi is NULL, just indicates whether the object exists.
2645  */
2646 int
2647 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2648 {
2649 	dnode_t *dn;
2650 	int err = dnode_hold(os, object, FTAG, &dn);
2651 
2652 	if (err)
2653 		return (err);
2654 
2655 	if (doi != NULL)
2656 		dmu_object_info_from_dnode(dn, doi);
2657 
2658 	dnode_rele(dn, FTAG);
2659 	return (0);
2660 }
2661 
2662 /*
2663  * As above, but faster; can be used when you have a held dbuf in hand.
2664  */
2665 void
2666 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2667 {
2668 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2669 
2670 	DB_DNODE_ENTER(db);
2671 	dmu_object_info_from_dnode(DB_DNODE(db), doi);
2672 	DB_DNODE_EXIT(db);
2673 }
2674 
2675 /*
2676  * Faster still when you only care about the size.
2677  */
2678 void
2679 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2680     u_longlong_t *nblk512)
2681 {
2682 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2683 	dnode_t *dn;
2684 
2685 	DB_DNODE_ENTER(db);
2686 	dn = DB_DNODE(db);
2687 
2688 	*blksize = dn->dn_datablksz;
2689 	/* add in number of slots used for the dnode itself */
2690 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2691 	    SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2692 	DB_DNODE_EXIT(db);
2693 }
2694 
2695 void
2696 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2697 {
2698 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2699 
2700 	DB_DNODE_ENTER(db);
2701 	*dnsize = DB_DNODE(db)->dn_num_slots << DNODE_SHIFT;
2702 	DB_DNODE_EXIT(db);
2703 }
2704 
2705 void
2706 byteswap_uint64_array(void *vbuf, size_t size)
2707 {
2708 	uint64_t *buf = vbuf;
2709 	size_t count = size >> 3;
2710 	int i;
2711 
2712 	ASSERT((size & 7) == 0);
2713 
2714 	for (i = 0; i < count; i++)
2715 		buf[i] = BSWAP_64(buf[i]);
2716 }
2717 
2718 void
2719 byteswap_uint32_array(void *vbuf, size_t size)
2720 {
2721 	uint32_t *buf = vbuf;
2722 	size_t count = size >> 2;
2723 	int i;
2724 
2725 	ASSERT((size & 3) == 0);
2726 
2727 	for (i = 0; i < count; i++)
2728 		buf[i] = BSWAP_32(buf[i]);
2729 }
2730 
2731 void
2732 byteswap_uint16_array(void *vbuf, size_t size)
2733 {
2734 	uint16_t *buf = vbuf;
2735 	size_t count = size >> 1;
2736 	int i;
2737 
2738 	ASSERT((size & 1) == 0);
2739 
2740 	for (i = 0; i < count; i++)
2741 		buf[i] = BSWAP_16(buf[i]);
2742 }
2743 
2744 void
2745 byteswap_uint8_array(void *vbuf, size_t size)
2746 {
2747 	(void) vbuf, (void) size;
2748 }
2749 
2750 void
2751 dmu_init(void)
2752 {
2753 	abd_init();
2754 	zfs_dbgmsg_init();
2755 	sa_cache_init();
2756 	dmu_objset_init();
2757 	dnode_init();
2758 	zfetch_init();
2759 	dmu_tx_init();
2760 	l2arc_init();
2761 	arc_init();
2762 	dbuf_init();
2763 }
2764 
2765 void
2766 dmu_fini(void)
2767 {
2768 	arc_fini(); /* arc depends on l2arc, so arc must go first */
2769 	l2arc_fini();
2770 	dmu_tx_fini();
2771 	zfetch_fini();
2772 	dbuf_fini();
2773 	dnode_fini();
2774 	dmu_objset_fini();
2775 	sa_cache_fini();
2776 	zfs_dbgmsg_fini();
2777 	abd_fini();
2778 }
2779 
2780 EXPORT_SYMBOL(dmu_bonus_hold);
2781 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2782 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2783 EXPORT_SYMBOL(dmu_buf_rele_array);
2784 EXPORT_SYMBOL(dmu_prefetch);
2785 EXPORT_SYMBOL(dmu_prefetch_by_dnode);
2786 EXPORT_SYMBOL(dmu_prefetch_dnode);
2787 EXPORT_SYMBOL(dmu_free_range);
2788 EXPORT_SYMBOL(dmu_free_long_range);
2789 EXPORT_SYMBOL(dmu_free_long_object);
2790 EXPORT_SYMBOL(dmu_read);
2791 EXPORT_SYMBOL(dmu_read_by_dnode);
2792 EXPORT_SYMBOL(dmu_write);
2793 EXPORT_SYMBOL(dmu_write_by_dnode);
2794 EXPORT_SYMBOL(dmu_prealloc);
2795 EXPORT_SYMBOL(dmu_object_info);
2796 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2797 EXPORT_SYMBOL(dmu_object_info_from_db);
2798 EXPORT_SYMBOL(dmu_object_size_from_db);
2799 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2800 EXPORT_SYMBOL(dmu_object_set_nlevels);
2801 EXPORT_SYMBOL(dmu_object_set_blocksize);
2802 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2803 EXPORT_SYMBOL(dmu_object_set_checksum);
2804 EXPORT_SYMBOL(dmu_object_set_compress);
2805 EXPORT_SYMBOL(dmu_offset_next);
2806 EXPORT_SYMBOL(dmu_write_policy);
2807 EXPORT_SYMBOL(dmu_sync);
2808 EXPORT_SYMBOL(dmu_request_arcbuf);
2809 EXPORT_SYMBOL(dmu_return_arcbuf);
2810 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2811 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2812 EXPORT_SYMBOL(dmu_buf_hold);
2813 EXPORT_SYMBOL(dmu_ot);
2814 
2815 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
2816 	"Enable NOP writes");
2817 
2818 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW,
2819 	"Percentage of dirtied blocks from frees in one TXG");
2820 
2821 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
2822 	"Enable forcing txg sync to find holes");
2823 
2824 /* CSTYLED */
2825 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW,
2826 	"Limit one prefetch call to this size");
2827