xref: /freebsd/sys/contrib/openzfs/module/zfs/dmu.c (revision efa8679e7f69c9cc225613827d9f75644cca5b3b)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, 2020 by Delphix. All rights reserved.
25  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27  * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved.
28  * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
29  * Copyright (c) 2019 Datto Inc.
30  * Copyright (c) 2019, 2023, Klara Inc.
31  * Copyright (c) 2019, Allan Jude
32  * Copyright (c) 2022 Hewlett Packard Enterprise Development LP.
33  * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
34  */
35 
36 #include <sys/dmu.h>
37 #include <sys/dmu_impl.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/dbuf.h>
40 #include <sys/dnode.h>
41 #include <sys/zfs_context.h>
42 #include <sys/dmu_objset.h>
43 #include <sys/dmu_traverse.h>
44 #include <sys/dsl_dataset.h>
45 #include <sys/dsl_dir.h>
46 #include <sys/dsl_pool.h>
47 #include <sys/dsl_synctask.h>
48 #include <sys/dsl_prop.h>
49 #include <sys/dmu_zfetch.h>
50 #include <sys/zfs_ioctl.h>
51 #include <sys/zap.h>
52 #include <sys/zio_checksum.h>
53 #include <sys/zio_compress.h>
54 #include <sys/sa.h>
55 #include <sys/zfeature.h>
56 #include <sys/abd.h>
57 #include <sys/brt.h>
58 #include <sys/trace_zfs.h>
59 #include <sys/zfs_racct.h>
60 #include <sys/zfs_rlock.h>
61 #ifdef _KERNEL
62 #include <sys/vmsystm.h>
63 #include <sys/zfs_znode.h>
64 #endif
65 
66 /*
67  * Enable/disable nopwrite feature.
68  */
69 static int zfs_nopwrite_enabled = 1;
70 
71 /*
72  * Tunable to control percentage of dirtied L1 blocks from frees allowed into
73  * one TXG. After this threshold is crossed, additional dirty blocks from frees
74  * will wait until the next TXG.
75  * A value of zero will disable this throttle.
76  */
77 static uint_t zfs_per_txg_dirty_frees_percent = 30;
78 
79 /*
80  * Enable/disable forcing txg sync when dirty checking for holes with lseek().
81  * By default this is enabled to ensure accurate hole reporting, it can result
82  * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads.
83  * Disabling this option will result in holes never being reported in dirty
84  * files which is always safe.
85  */
86 static int zfs_dmu_offset_next_sync = 1;
87 
88 /*
89  * Limit the amount we can prefetch with one call to this amount.  This
90  * helps to limit the amount of memory that can be used by prefetching.
91  * Larger objects should be prefetched a bit at a time.
92  */
93 #ifdef _ILP32
94 uint_t dmu_prefetch_max = 8 * 1024 * 1024;
95 #else
96 uint_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE;
97 #endif
98 
99 /*
100  * Override copies= for dedup state objects. 0 means the traditional behaviour
101  * (ie the default for the containing objset ie 3 for the MOS).
102  */
103 uint_t dmu_ddt_copies = 0;
104 
105 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
106 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "unallocated"		},
107 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "object directory"	},
108 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "object array"		},
109 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "packed nvlist"		},
110 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "packed nvlist size"	},
111 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "bpobj"			},
112 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "bpobj header"		},
113 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "SPA space map header"	},
114 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "SPA space map"		},
115 	{DMU_BSWAP_UINT64, TRUE,  FALSE, TRUE,  "ZIL intent log"	},
116 	{DMU_BSWAP_DNODE,  TRUE,  FALSE, TRUE,  "DMU dnode"		},
117 	{DMU_BSWAP_OBJSET, TRUE,  TRUE,  FALSE, "DMU objset"		},
118 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "DSL directory"		},
119 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL directory child map"},
120 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL dataset snap map"	},
121 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL props"		},
122 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "DSL dataset"		},
123 	{DMU_BSWAP_ZNODE,  TRUE,  FALSE, FALSE, "ZFS znode"		},
124 	{DMU_BSWAP_OLDACL, TRUE,  FALSE, TRUE,  "ZFS V0 ACL"		},
125 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "ZFS plain file"	},
126 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS directory"		},
127 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "ZFS master node"	},
128 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS delete queue"	},
129 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "zvol object"		},
130 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "zvol prop"		},
131 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "other uint8[]"		},
132 	{DMU_BSWAP_UINT64, FALSE, FALSE, TRUE,  "other uint64[]"	},
133 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "other ZAP"		},
134 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "persistent error log"	},
135 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "SPA history"		},
136 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "SPA history offsets"	},
137 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "Pool properties"	},
138 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL permissions"	},
139 	{DMU_BSWAP_ACL,    TRUE,  FALSE, TRUE,  "ZFS ACL"		},
140 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, TRUE,  "ZFS SYSACL"		},
141 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, TRUE,  "FUID table"		},
142 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "FUID table size"	},
143 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL dataset next clones"},
144 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "scan work queue"	},
145 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS user/group/project used" },
146 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS user/group/project quota"},
147 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "snapshot refcount tags"},
148 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "DDT ZAP algorithm"	},
149 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "DDT statistics"	},
150 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, TRUE,	"System attributes"	},
151 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,	"SA master node"	},
152 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,	"SA attr registration"	},
153 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,	"SA attr layouts"	},
154 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "scan translations"	},
155 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "deduplicated block"	},
156 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL deadlist map"	},
157 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "DSL deadlist map hdr"	},
158 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL dir clones"	},
159 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "bpobj subobj"		}
160 };
161 
162 dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
163 	{	byteswap_uint8_array,	"uint8"		},
164 	{	byteswap_uint16_array,	"uint16"	},
165 	{	byteswap_uint32_array,	"uint32"	},
166 	{	byteswap_uint64_array,	"uint64"	},
167 	{	zap_byteswap,		"zap"		},
168 	{	dnode_buf_byteswap,	"dnode"		},
169 	{	dmu_objset_byteswap,	"objset"	},
170 	{	zfs_znode_byteswap,	"znode"		},
171 	{	zfs_oldacl_byteswap,	"oldacl"	},
172 	{	zfs_acl_byteswap,	"acl"		}
173 };
174 
175 int
dmu_buf_hold_noread_by_dnode(dnode_t * dn,uint64_t offset,const void * tag,dmu_buf_t ** dbp)176 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
177     const void *tag, dmu_buf_t **dbp)
178 {
179 	uint64_t blkid;
180 	dmu_buf_impl_t *db;
181 
182 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
183 	blkid = dbuf_whichblock(dn, 0, offset);
184 	db = dbuf_hold(dn, blkid, tag);
185 	rw_exit(&dn->dn_struct_rwlock);
186 
187 	if (db == NULL) {
188 		*dbp = NULL;
189 		return (SET_ERROR(EIO));
190 	}
191 
192 	*dbp = &db->db;
193 	return (0);
194 }
195 
196 int
dmu_buf_hold_noread(objset_t * os,uint64_t object,uint64_t offset,const void * tag,dmu_buf_t ** dbp)197 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
198     const void *tag, dmu_buf_t **dbp)
199 {
200 	dnode_t *dn;
201 	uint64_t blkid;
202 	dmu_buf_impl_t *db;
203 	int err;
204 
205 	err = dnode_hold(os, object, FTAG, &dn);
206 	if (err)
207 		return (err);
208 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
209 	blkid = dbuf_whichblock(dn, 0, offset);
210 	db = dbuf_hold(dn, blkid, tag);
211 	rw_exit(&dn->dn_struct_rwlock);
212 	dnode_rele(dn, FTAG);
213 
214 	if (db == NULL) {
215 		*dbp = NULL;
216 		return (SET_ERROR(EIO));
217 	}
218 
219 	*dbp = &db->db;
220 	return (err);
221 }
222 
223 int
dmu_buf_hold_by_dnode(dnode_t * dn,uint64_t offset,const void * tag,dmu_buf_t ** dbp,dmu_flags_t flags)224 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
225     const void *tag, dmu_buf_t **dbp, dmu_flags_t flags)
226 {
227 	int err;
228 
229 	err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
230 	if (err == 0) {
231 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
232 		err = dbuf_read(db, NULL, flags | DB_RF_CANFAIL);
233 		if (err != 0) {
234 			dbuf_rele(db, tag);
235 			*dbp = NULL;
236 		}
237 	}
238 
239 	return (err);
240 }
241 
242 int
dmu_buf_hold(objset_t * os,uint64_t object,uint64_t offset,const void * tag,dmu_buf_t ** dbp,dmu_flags_t flags)243 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
244     const void *tag, dmu_buf_t **dbp, dmu_flags_t flags)
245 {
246 	int err;
247 
248 	err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
249 	if (err == 0) {
250 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
251 		err = dbuf_read(db, NULL, flags | DB_RF_CANFAIL);
252 		if (err != 0) {
253 			dbuf_rele(db, tag);
254 			*dbp = NULL;
255 		}
256 	}
257 
258 	return (err);
259 }
260 
261 int
dmu_bonus_max(void)262 dmu_bonus_max(void)
263 {
264 	return (DN_OLD_MAX_BONUSLEN);
265 }
266 
267 int
dmu_set_bonus(dmu_buf_t * db_fake,int newsize,dmu_tx_t * tx)268 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
269 {
270 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
271 	dnode_t *dn;
272 	int error;
273 
274 	if (newsize < 0 || newsize > db_fake->db_size)
275 		return (SET_ERROR(EINVAL));
276 
277 	DB_DNODE_ENTER(db);
278 	dn = DB_DNODE(db);
279 
280 	if (dn->dn_bonus != db) {
281 		error = SET_ERROR(EINVAL);
282 	} else {
283 		dnode_setbonuslen(dn, newsize, tx);
284 		error = 0;
285 	}
286 
287 	DB_DNODE_EXIT(db);
288 	return (error);
289 }
290 
291 int
dmu_set_bonustype(dmu_buf_t * db_fake,dmu_object_type_t type,dmu_tx_t * tx)292 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
293 {
294 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
295 	dnode_t *dn;
296 	int error;
297 
298 	if (!DMU_OT_IS_VALID(type))
299 		return (SET_ERROR(EINVAL));
300 
301 	DB_DNODE_ENTER(db);
302 	dn = DB_DNODE(db);
303 
304 	if (dn->dn_bonus != db) {
305 		error = SET_ERROR(EINVAL);
306 	} else {
307 		dnode_setbonus_type(dn, type, tx);
308 		error = 0;
309 	}
310 
311 	DB_DNODE_EXIT(db);
312 	return (error);
313 }
314 
315 dmu_object_type_t
dmu_get_bonustype(dmu_buf_t * db_fake)316 dmu_get_bonustype(dmu_buf_t *db_fake)
317 {
318 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
319 	dmu_object_type_t type;
320 
321 	DB_DNODE_ENTER(db);
322 	type = DB_DNODE(db)->dn_bonustype;
323 	DB_DNODE_EXIT(db);
324 
325 	return (type);
326 }
327 
328 int
dmu_rm_spill(objset_t * os,uint64_t object,dmu_tx_t * tx)329 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
330 {
331 	dnode_t *dn;
332 	int error;
333 
334 	error = dnode_hold(os, object, FTAG, &dn);
335 	dbuf_rm_spill(dn, tx);
336 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
337 	dnode_rm_spill(dn, tx);
338 	rw_exit(&dn->dn_struct_rwlock);
339 	dnode_rele(dn, FTAG);
340 	return (error);
341 }
342 
343 /*
344  * Lookup and hold the bonus buffer for the provided dnode.  If the dnode
345  * has not yet been allocated a new bonus dbuf a will be allocated.
346  * Returns ENOENT, EIO, or 0.
347  */
dmu_bonus_hold_by_dnode(dnode_t * dn,const void * tag,dmu_buf_t ** dbp,dmu_flags_t flags)348 int dmu_bonus_hold_by_dnode(dnode_t *dn, const void *tag, dmu_buf_t **dbp,
349     dmu_flags_t flags)
350 {
351 	dmu_buf_impl_t *db;
352 	int error;
353 
354 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
355 	if (dn->dn_bonus == NULL) {
356 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
357 			rw_exit(&dn->dn_struct_rwlock);
358 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
359 		}
360 		if (dn->dn_bonus == NULL)
361 			dbuf_create_bonus(dn);
362 	}
363 	db = dn->dn_bonus;
364 
365 	/* as long as the bonus buf is held, the dnode will be held */
366 	if (zfs_refcount_add(&db->db_holds, tag) == 1) {
367 		VERIFY(dnode_add_ref(dn, db));
368 		atomic_inc_32(&dn->dn_dbufs_count);
369 	}
370 
371 	/*
372 	 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
373 	 * hold and incrementing the dbuf count to ensure that dnode_move() sees
374 	 * a dnode hold for every dbuf.
375 	 */
376 	rw_exit(&dn->dn_struct_rwlock);
377 
378 	error = dbuf_read(db, NULL, flags | DB_RF_CANFAIL);
379 	if (error) {
380 		dnode_evict_bonus(dn);
381 		dbuf_rele(db, tag);
382 		*dbp = NULL;
383 		return (error);
384 	}
385 
386 	*dbp = &db->db;
387 	return (0);
388 }
389 
390 int
dmu_bonus_hold(objset_t * os,uint64_t object,const void * tag,dmu_buf_t ** dbp)391 dmu_bonus_hold(objset_t *os, uint64_t object, const void *tag, dmu_buf_t **dbp)
392 {
393 	dnode_t *dn;
394 	int error;
395 
396 	error = dnode_hold(os, object, FTAG, &dn);
397 	if (error)
398 		return (error);
399 
400 	error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH);
401 	dnode_rele(dn, FTAG);
402 
403 	return (error);
404 }
405 
406 /*
407  * returns ENOENT, EIO, or 0.
408  *
409  * This interface will allocate a blank spill dbuf when a spill blk
410  * doesn't already exist on the dnode.
411  *
412  * if you only want to find an already existing spill db, then
413  * dmu_spill_hold_existing() should be used.
414  */
415 int
dmu_spill_hold_by_dnode(dnode_t * dn,dmu_flags_t flags,const void * tag,dmu_buf_t ** dbp)416 dmu_spill_hold_by_dnode(dnode_t *dn, dmu_flags_t flags, const void *tag,
417     dmu_buf_t **dbp)
418 {
419 	dmu_buf_impl_t *db = NULL;
420 	int err;
421 
422 	if ((flags & DB_RF_HAVESTRUCT) == 0)
423 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
424 
425 	db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
426 
427 	if ((flags & DB_RF_HAVESTRUCT) == 0)
428 		rw_exit(&dn->dn_struct_rwlock);
429 
430 	if (db == NULL) {
431 		*dbp = NULL;
432 		return (SET_ERROR(EIO));
433 	}
434 	err = dbuf_read(db, NULL, flags);
435 	if (err == 0)
436 		*dbp = &db->db;
437 	else {
438 		dbuf_rele(db, tag);
439 		*dbp = NULL;
440 	}
441 	return (err);
442 }
443 
444 int
dmu_spill_hold_existing(dmu_buf_t * bonus,const void * tag,dmu_buf_t ** dbp)445 dmu_spill_hold_existing(dmu_buf_t *bonus, const void *tag, dmu_buf_t **dbp)
446 {
447 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
448 	dnode_t *dn;
449 	int err;
450 
451 	DB_DNODE_ENTER(db);
452 	dn = DB_DNODE(db);
453 
454 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
455 		err = SET_ERROR(EINVAL);
456 	} else {
457 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
458 
459 		if (!dn->dn_have_spill) {
460 			err = SET_ERROR(ENOENT);
461 		} else {
462 			err = dmu_spill_hold_by_dnode(dn,
463 			    DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
464 		}
465 
466 		rw_exit(&dn->dn_struct_rwlock);
467 	}
468 
469 	DB_DNODE_EXIT(db);
470 	return (err);
471 }
472 
473 int
dmu_spill_hold_by_bonus(dmu_buf_t * bonus,dmu_flags_t flags,const void * tag,dmu_buf_t ** dbp)474 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, dmu_flags_t flags, const void *tag,
475     dmu_buf_t **dbp)
476 {
477 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
478 	int err;
479 
480 	DB_DNODE_ENTER(db);
481 	err = dmu_spill_hold_by_dnode(DB_DNODE(db), flags, tag, dbp);
482 	DB_DNODE_EXIT(db);
483 
484 	return (err);
485 }
486 
487 /*
488  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
489  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
490  * and can induce severe lock contention when writing to several files
491  * whose dnodes are in the same block.
492  */
493 int
dmu_buf_hold_array_by_dnode(dnode_t * dn,uint64_t offset,uint64_t length,boolean_t read,const void * tag,int * numbufsp,dmu_buf_t *** dbpp,dmu_flags_t flags)494 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
495     boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp,
496     dmu_flags_t flags)
497 {
498 	dmu_buf_t **dbp;
499 	zstream_t *zs = NULL;
500 	uint64_t blkid, nblks, i;
501 	dmu_flags_t dbuf_flags;
502 	int err;
503 	zio_t *zio = NULL;
504 	boolean_t missed = B_FALSE;
505 
506 	ASSERT(!read || length <= DMU_MAX_ACCESS);
507 
508 	/*
509 	 * Note: We directly notify the prefetch code of this read, so that
510 	 * we can tell it about the multi-block read.  dbuf_read() only knows
511 	 * about the one block it is accessing.
512 	 */
513 	dbuf_flags = (flags & ~DMU_READ_PREFETCH) | DMU_READ_NO_PREFETCH |
514 	    DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
515 
516 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
517 	if (dn->dn_datablkshift) {
518 		int blkshift = dn->dn_datablkshift;
519 		nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
520 		    P2ALIGN_TYPED(offset, 1ULL << blkshift, uint64_t))
521 		    >> blkshift;
522 	} else {
523 		if (offset + length > dn->dn_datablksz) {
524 			zfs_panic_recover("zfs: accessing past end of object "
525 			    "%llx/%llx (size=%u access=%llu+%llu)",
526 			    (longlong_t)dn->dn_objset->
527 			    os_dsl_dataset->ds_object,
528 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
529 			    (longlong_t)offset, (longlong_t)length);
530 			rw_exit(&dn->dn_struct_rwlock);
531 			return (SET_ERROR(EIO));
532 		}
533 		nblks = 1;
534 	}
535 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
536 
537 	if (read)
538 		zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
539 		    ZIO_FLAG_CANFAIL);
540 	blkid = dbuf_whichblock(dn, 0, offset);
541 	if ((flags & DMU_READ_NO_PREFETCH) == 0) {
542 		/*
543 		 * Prepare the zfetch before initiating the demand reads, so
544 		 * that if multiple threads block on same indirect block, we
545 		 * base predictions on the original less racy request order.
546 		 */
547 		zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks,
548 		    read && !(flags & DMU_DIRECTIO), B_TRUE);
549 	}
550 	for (i = 0; i < nblks; i++) {
551 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
552 		if (db == NULL) {
553 			if (zs) {
554 				dmu_zfetch_run(&dn->dn_zfetch, zs, missed,
555 				    B_TRUE, (flags & DMU_UNCACHEDIO));
556 			}
557 			rw_exit(&dn->dn_struct_rwlock);
558 			dmu_buf_rele_array(dbp, nblks, tag);
559 			if (read)
560 				zio_nowait(zio);
561 			return (SET_ERROR(EIO));
562 		}
563 
564 		/*
565 		 * Initiate async demand data read.
566 		 * We check the db_state after calling dbuf_read() because
567 		 * (1) dbuf_read() may change the state to CACHED due to a
568 		 * hit in the ARC, and (2) on a cache miss, a child will
569 		 * have been added to "zio" but not yet completed, so the
570 		 * state will not yet be CACHED.
571 		 */
572 		if (read) {
573 			if (i == nblks - 1 && blkid + i < dn->dn_maxblkid &&
574 			    offset + length < db->db.db_offset +
575 			    db->db.db_size) {
576 				if (offset <= db->db.db_offset)
577 					dbuf_flags |= DMU_PARTIAL_FIRST;
578 				else
579 					dbuf_flags |= DMU_PARTIAL_MORE;
580 			}
581 			(void) dbuf_read(db, zio, dbuf_flags);
582 			if (db->db_state != DB_CACHED)
583 				missed = B_TRUE;
584 		}
585 		dbp[i] = &db->db;
586 	}
587 
588 	/*
589 	 * If we are doing O_DIRECT we still hold the dbufs, even for reads,
590 	 * but we do not issue any reads here. We do not want to account for
591 	 * writes in this case.
592 	 *
593 	 * O_DIRECT write/read accounting takes place in
594 	 * dmu_{write/read}_abd().
595 	 */
596 	if (!read && ((flags & DMU_DIRECTIO) == 0))
597 		zfs_racct_write(dn->dn_objset->os_spa, length, nblks, flags);
598 
599 	if (zs) {
600 		dmu_zfetch_run(&dn->dn_zfetch, zs, missed, B_TRUE,
601 		    (flags & DMU_UNCACHEDIO));
602 	}
603 	rw_exit(&dn->dn_struct_rwlock);
604 
605 	if (read) {
606 		/* wait for async read i/o */
607 		err = zio_wait(zio);
608 		if (err) {
609 			dmu_buf_rele_array(dbp, nblks, tag);
610 			return (err);
611 		}
612 
613 		/* wait for other io to complete */
614 		for (i = 0; i < nblks; i++) {
615 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
616 			mutex_enter(&db->db_mtx);
617 			while (db->db_state == DB_READ ||
618 			    db->db_state == DB_FILL)
619 				cv_wait(&db->db_changed, &db->db_mtx);
620 			if (db->db_state == DB_UNCACHED)
621 				err = SET_ERROR(EIO);
622 			mutex_exit(&db->db_mtx);
623 			if (err) {
624 				dmu_buf_rele_array(dbp, nblks, tag);
625 				return (err);
626 			}
627 		}
628 	}
629 
630 	*numbufsp = nblks;
631 	*dbpp = dbp;
632 	return (0);
633 }
634 
635 int
dmu_buf_hold_array(objset_t * os,uint64_t object,uint64_t offset,uint64_t length,int read,const void * tag,int * numbufsp,dmu_buf_t *** dbpp,dmu_flags_t flags)636 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
637     uint64_t length, int read, const void *tag, int *numbufsp,
638     dmu_buf_t ***dbpp, dmu_flags_t flags)
639 {
640 	dnode_t *dn;
641 	int err;
642 
643 	err = dnode_hold(os, object, FTAG, &dn);
644 	if (err)
645 		return (err);
646 
647 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
648 	    numbufsp, dbpp, flags);
649 
650 	dnode_rele(dn, FTAG);
651 
652 	return (err);
653 }
654 
655 int
dmu_buf_hold_array_by_bonus(dmu_buf_t * db_fake,uint64_t offset,uint64_t length,boolean_t read,const void * tag,int * numbufsp,dmu_buf_t *** dbpp,dmu_flags_t flags)656 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
657     uint64_t length, boolean_t read, const void *tag, int *numbufsp,
658     dmu_buf_t ***dbpp, dmu_flags_t flags)
659 {
660 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
661 	int err;
662 
663 	DB_DNODE_ENTER(db);
664 	err = dmu_buf_hold_array_by_dnode(DB_DNODE(db), offset, length, read,
665 	    tag, numbufsp, dbpp, flags);
666 	DB_DNODE_EXIT(db);
667 
668 	return (err);
669 }
670 
671 void
dmu_buf_rele_array(dmu_buf_t ** dbp_fake,int numbufs,const void * tag)672 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag)
673 {
674 	int i;
675 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
676 
677 	if (numbufs == 0)
678 		return;
679 
680 	for (i = 0; i < numbufs; i++) {
681 		if (dbp[i])
682 			dbuf_rele(dbp[i], tag);
683 	}
684 
685 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
686 }
687 
688 /*
689  * Issue prefetch I/Os for the given blocks.  If level is greater than 0, the
690  * indirect blocks prefetched will be those that point to the blocks containing
691  * the data starting at offset, and continuing to offset + len.  If the range
692  * is too long, prefetch the first dmu_prefetch_max bytes as requested, while
693  * for the rest only a higher level, also fitting within dmu_prefetch_max.  It
694  * should primarily help random reads, since for long sequential reads there is
695  * a speculative prefetcher.
696  *
697  * Note that if the indirect blocks above the blocks being prefetched are not
698  * in cache, they will be asynchronously read in.  Dnode read by dnode_hold()
699  * is currently synchronous.
700  */
701 void
dmu_prefetch(objset_t * os,uint64_t object,int64_t level,uint64_t offset,uint64_t len,zio_priority_t pri)702 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
703     uint64_t len, zio_priority_t pri)
704 {
705 	dnode_t *dn;
706 
707 	if (dmu_prefetch_max == 0 || len == 0) {
708 		dmu_prefetch_dnode(os, object, pri);
709 		return;
710 	}
711 
712 	if (dnode_hold(os, object, FTAG, &dn) != 0)
713 		return;
714 
715 	dmu_prefetch_by_dnode(dn, level, offset, len, pri);
716 
717 	dnode_rele(dn, FTAG);
718 }
719 
720 void
dmu_prefetch_by_dnode(dnode_t * dn,int64_t level,uint64_t offset,uint64_t len,zio_priority_t pri)721 dmu_prefetch_by_dnode(dnode_t *dn, int64_t level, uint64_t offset,
722     uint64_t len, zio_priority_t pri)
723 {
724 	int64_t level2 = level;
725 	uint64_t start, end, start2, end2;
726 
727 	/*
728 	 * Depending on len we may do two prefetches: blocks [start, end) at
729 	 * level, and following blocks [start2, end2) at higher level2.
730 	 */
731 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
732 	if (dn->dn_datablkshift != 0) {
733 
734 		/*
735 		 * Limit prefetch to present blocks.
736 		 */
737 		uint64_t size = (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
738 		if (offset >= size) {
739 			rw_exit(&dn->dn_struct_rwlock);
740 			return;
741 		}
742 		if (offset + len < offset || offset + len > size)
743 			len = size - offset;
744 
745 		/*
746 		 * The object has multiple blocks.  Calculate the full range
747 		 * of blocks [start, end2) and then split it into two parts,
748 		 * so that the first [start, end) fits into dmu_prefetch_max.
749 		 */
750 		start = dbuf_whichblock(dn, level, offset);
751 		end2 = dbuf_whichblock(dn, level, offset + len - 1) + 1;
752 		uint8_t ibs = dn->dn_indblkshift;
753 		uint8_t bs = (level == 0) ? dn->dn_datablkshift : ibs;
754 		uint_t limit = P2ROUNDUP(dmu_prefetch_max, 1 << bs) >> bs;
755 		start2 = end = MIN(end2, start + limit);
756 
757 		/*
758 		 * Find level2 where [start2, end2) fits into dmu_prefetch_max.
759 		 */
760 		uint8_t ibps = ibs - SPA_BLKPTRSHIFT;
761 		limit = P2ROUNDUP(dmu_prefetch_max, 1 << ibs) >> ibs;
762 		if (limit == 0)
763 			end2 = start2;
764 		do {
765 			level2++;
766 			start2 = P2ROUNDUP(start2, 1 << ibps) >> ibps;
767 			end2 = P2ROUNDUP(end2, 1 << ibps) >> ibps;
768 		} while (end2 - start2 > limit);
769 	} else {
770 		/* There is only one block.  Prefetch it or nothing. */
771 		start = start2 = end2 = 0;
772 		end = start + (level == 0 && offset < dn->dn_datablksz);
773 	}
774 
775 	for (uint64_t i = start; i < end; i++)
776 		dbuf_prefetch(dn, level, i, pri, 0);
777 	for (uint64_t i = start2; i < end2; i++)
778 		dbuf_prefetch(dn, level2, i, pri, 0);
779 	rw_exit(&dn->dn_struct_rwlock);
780 }
781 
782 /*
783  * Prime a prefetch for sequential accesses from offset for at least len bytes.
784  */
785 void
dmu_prefetch_stream(objset_t * os,uint64_t object,uint64_t offset,uint64_t len,boolean_t start_now)786 dmu_prefetch_stream(objset_t *os, uint64_t object, uint64_t offset,
787     uint64_t len, boolean_t start_now)
788 {
789 	dnode_t *dn;
790 
791 	if (dnode_hold(os, object, FTAG, &dn) != 0)
792 		return;
793 	dmu_prefetch_stream_by_dnode(dn, offset, len, start_now);
794 	dnode_rele(dn, FTAG);
795 }
796 
797 void
dmu_prefetch_stream_by_dnode(dnode_t * dn,uint64_t offset,uint64_t len,boolean_t start_now)798 dmu_prefetch_stream_by_dnode(dnode_t *dn, uint64_t offset, uint64_t len,
799     boolean_t start_now)
800 {
801 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
802 	if (dn->dn_datablkshift != 0) {
803 		uint64_t start = dbuf_whichblock(dn, 0, offset);
804 		if (len == 0) {
805 			if (dmu_zfetch_prime(&dn->dn_zfetch, start, start) &&
806 			    start_now) {
807 				dmu_zfetch(&dn->dn_zfetch, start, 0, B_TRUE,
808 				    B_TRUE, B_TRUE, B_FALSE);
809 			}
810 		} else {
811 			uint64_t end = dbuf_whichblock(dn, 0, offset + len - 1);
812 			if (start == end) {
813 				if (start_now) {
814 					dbuf_prefetch(dn, 0, start,
815 					    ZIO_PRIORITY_ASYNC_READ, 0);
816 				}
817 			} else if (
818 			    dmu_zfetch_prime(&dn->dn_zfetch, start, end + 1) &&
819 			    start_now) {
820 				dmu_zfetch(&dn->dn_zfetch, start, 0, B_TRUE,
821 				    B_TRUE, B_TRUE, B_FALSE);
822 			}
823 		}
824 	} else if (offset < dn->dn_datablksz && start_now) {
825 		dbuf_prefetch(dn, 0, 0, ZIO_PRIORITY_ASYNC_READ, 0);
826 	}
827 	rw_exit(&dn->dn_struct_rwlock);
828 }
829 
830 typedef struct {
831 	kmutex_t	dpa_lock;
832 	kcondvar_t	dpa_cv;
833 	uint64_t	dpa_pending_io;
834 } dmu_prefetch_arg_t;
835 
836 static void
dmu_prefetch_done(void * arg,uint64_t level,uint64_t blkid,boolean_t issued)837 dmu_prefetch_done(void *arg, uint64_t level, uint64_t blkid, boolean_t issued)
838 {
839 	(void) level; (void) blkid; (void)issued;
840 	dmu_prefetch_arg_t *dpa = arg;
841 
842 	ASSERT0(level);
843 
844 	mutex_enter(&dpa->dpa_lock);
845 	ASSERT3U(dpa->dpa_pending_io, >, 0);
846 	if (--dpa->dpa_pending_io == 0)
847 		cv_broadcast(&dpa->dpa_cv);
848 	mutex_exit(&dpa->dpa_lock);
849 }
850 
851 static void
dmu_prefetch_wait_by_dnode(dnode_t * dn,uint64_t offset,uint64_t len)852 dmu_prefetch_wait_by_dnode(dnode_t *dn, uint64_t offset, uint64_t len)
853 {
854 	dmu_prefetch_arg_t dpa;
855 
856 	mutex_init(&dpa.dpa_lock, NULL, MUTEX_DEFAULT, NULL);
857 	cv_init(&dpa.dpa_cv, NULL, CV_DEFAULT, NULL);
858 
859 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
860 
861 	uint64_t start = dbuf_whichblock(dn, 0, offset);
862 	uint64_t end = dbuf_whichblock(dn, 0, offset + len - 1) + 1;
863 	dpa.dpa_pending_io = end - start;
864 
865 	for (uint64_t blk = start; blk < end; blk++) {
866 		(void) dbuf_prefetch_impl(dn, 0, blk, ZIO_PRIORITY_ASYNC_READ,
867 		    0, dmu_prefetch_done, &dpa);
868 	}
869 
870 	rw_exit(&dn->dn_struct_rwlock);
871 
872 	/* wait for prefetch L0 reads to finish */
873 	mutex_enter(&dpa.dpa_lock);
874 	while (dpa.dpa_pending_io > 0) {
875 		cv_wait(&dpa.dpa_cv, &dpa.dpa_lock);
876 
877 	}
878 	mutex_exit(&dpa.dpa_lock);
879 
880 	mutex_destroy(&dpa.dpa_lock);
881 	cv_destroy(&dpa.dpa_cv);
882 }
883 
884 /*
885  * Issue prefetch I/Os for the given L0 block range and wait for the I/O
886  * to complete. This does not enforce dmu_prefetch_max and will prefetch
887  * the entire range. The blocks are read from disk into the ARC but no
888  * decompression occurs (i.e., the dbuf cache is not required).
889  */
890 int
dmu_prefetch_wait(objset_t * os,uint64_t object,uint64_t offset,uint64_t size)891 dmu_prefetch_wait(objset_t *os, uint64_t object, uint64_t offset, uint64_t size)
892 {
893 	dnode_t *dn;
894 	int err = 0;
895 
896 	err = dnode_hold(os, object, FTAG, &dn);
897 	if (err != 0)
898 		return (err);
899 
900 	/*
901 	 * Chunk the requests (16 indirects worth) so that we can be
902 	 * interrupted.  Prefetch at least SPA_MAXBLOCKSIZE at a time
903 	 * to better utilize pools with smaller block sizes.
904 	 */
905 	uint64_t chunksize;
906 	if (dn->dn_indblkshift) {
907 		uint64_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1);
908 		chunksize = (nbps * 16) << dn->dn_datablkshift;
909 		chunksize = MAX(chunksize, SPA_MAXBLOCKSIZE);
910 	} else {
911 		chunksize = dn->dn_datablksz;
912 	}
913 
914 	while (size > 0) {
915 		uint64_t mylen = MIN(size, chunksize);
916 
917 		dmu_prefetch_wait_by_dnode(dn, offset, mylen);
918 
919 		offset += mylen;
920 		size -= mylen;
921 
922 		if (issig()) {
923 			err = SET_ERROR(EINTR);
924 			break;
925 		}
926 	}
927 
928 	dnode_rele(dn, FTAG);
929 
930 	return (err);
931 }
932 
933 /*
934  * Issue prefetch I/Os for the given object's dnode.
935  */
936 void
dmu_prefetch_dnode(objset_t * os,uint64_t object,zio_priority_t pri)937 dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri)
938 {
939 	if (object == 0 || object >= DN_MAX_OBJECT)
940 		return;
941 
942 	dnode_t *dn = DMU_META_DNODE(os);
943 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
944 	uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t));
945 	dbuf_prefetch(dn, 0, blkid, pri, 0);
946 	rw_exit(&dn->dn_struct_rwlock);
947 }
948 
949 /*
950  * Advisory cache eviction for a byte range of an object.
951  */
952 void
dmu_evict_range(objset_t * os,uint64_t object,uint64_t offset,uint64_t len)953 dmu_evict_range(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
954 {
955 	dnode_t *dn;
956 
957 	if (len == 0)
958 		return;
959 	if (dnode_hold(os, object, FTAG, &dn) != 0)
960 		return;
961 
962 	/*
963 	 * Exclude the last block if the range end is not block-aligned:
964 	 * a sequential access may continue into that block.  The first
965 	 * block is included even when partially covered since backwards
966 	 * access patterns are rare.
967 	 */
968 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
969 	uint64_t start, end;
970 	if (dn->dn_datablkshift != 0) {
971 		start = dbuf_whichblock(dn, 0, offset);
972 		end = dbuf_whichblock(dn, 0, offset + len);
973 	} else {
974 		start = (offset >= dn->dn_datablksz);
975 		end = (offset + len >= dn->dn_datablksz);
976 	}
977 	if (end > start)
978 		dbuf_evict_range(dn, start, end - 1);
979 	rw_exit(&dn->dn_struct_rwlock);
980 
981 	dnode_rele(dn, FTAG);
982 }
983 
984 /*
985  * Get the next "chunk" of file data to free.  We traverse the file from
986  * the end so that the file gets shorter over time (if we crash in the
987  * middle, this will leave us in a better state).  We find allocated file
988  * data by simply searching the allocated level 1 indirects.
989  *
990  * On input, *start should be the first offset that does not need to be
991  * freed (e.g. "offset + length").  On return, *start will be the first
992  * offset that should be freed and l1blks is set to the number of level 1
993  * indirect blocks found within the chunk.
994  */
995 static int
get_next_chunk(dnode_t * dn,uint64_t * start,uint64_t minimum,uint64_t * l1blks)996 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
997 {
998 	uint64_t blks;
999 	uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
1000 	/* bytes of data covered by a level-1 indirect block */
1001 	uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
1002 	    EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
1003 
1004 	ASSERT3U(minimum, <=, *start);
1005 
1006 	/* dn_nlevels == 1 means we don't have any L1 blocks */
1007 	if (dn->dn_nlevels <= 1) {
1008 		*l1blks = 0;
1009 		*start = minimum;
1010 		return (0);
1011 	}
1012 
1013 	/*
1014 	 * Check if we can free the entire range assuming that all of the
1015 	 * L1 blocks in this range have data. If we can, we use this
1016 	 * worst case value as an estimate so we can avoid having to look
1017 	 * at the object's actual data.
1018 	 */
1019 	uint64_t total_l1blks =
1020 	    (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
1021 	    iblkrange;
1022 	if (total_l1blks <= maxblks) {
1023 		*l1blks = total_l1blks;
1024 		*start = minimum;
1025 		return (0);
1026 	}
1027 	ASSERT(ISP2(iblkrange));
1028 
1029 	for (blks = 0; *start > minimum && blks < maxblks; blks++) {
1030 		int err;
1031 
1032 		/*
1033 		 * dnode_next_offset(BACKWARDS) will find an allocated L1
1034 		 * indirect block at or before the input offset.  We must
1035 		 * decrement *start so that it is at the end of the region
1036 		 * to search.
1037 		 */
1038 		(*start)--;
1039 
1040 		err = dnode_next_offset(dn,
1041 		    DNODE_FIND_BACKWARDS, start, 2, 1, 0);
1042 
1043 		/* if there are no indirect blocks before start, we are done */
1044 		if (err == ESRCH) {
1045 			*start = minimum;
1046 			break;
1047 		} else if (err != 0) {
1048 			*l1blks = blks;
1049 			return (err);
1050 		}
1051 
1052 		/* set start to the beginning of this L1 indirect */
1053 		*start = P2ALIGN_TYPED(*start, iblkrange, uint64_t);
1054 	}
1055 	if (*start < minimum)
1056 		*start = minimum;
1057 	*l1blks = blks;
1058 
1059 	return (0);
1060 }
1061 
1062 /*
1063  * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
1064  * otherwise return false.
1065  * Used below in dmu_free_long_range_impl() to enable abort when unmounting
1066  */
1067 static boolean_t
dmu_objset_zfs_unmounting(objset_t * os)1068 dmu_objset_zfs_unmounting(objset_t *os)
1069 {
1070 #ifdef _KERNEL
1071 	if (dmu_objset_type(os) == DMU_OST_ZFS)
1072 		return (zfs_get_vfs_flag_unmounted(os));
1073 #else
1074 	(void) os;
1075 #endif
1076 	return (B_FALSE);
1077 }
1078 
1079 static int
dmu_free_long_range_impl(objset_t * os,dnode_t * dn,uint64_t offset,uint64_t length)1080 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
1081     uint64_t length)
1082 {
1083 	uint64_t object_size;
1084 	int err;
1085 	uint64_t dirty_frees_threshold;
1086 	dsl_pool_t *dp = dmu_objset_pool(os);
1087 
1088 	if (dn == NULL)
1089 		return (SET_ERROR(EINVAL));
1090 
1091 	object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1092 	if (offset >= object_size)
1093 		return (0);
1094 
1095 	if (zfs_per_txg_dirty_frees_percent <= 100)
1096 		dirty_frees_threshold =
1097 		    zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
1098 	else
1099 		dirty_frees_threshold = zfs_dirty_data_max / 20;
1100 
1101 	if (length == DMU_OBJECT_END || offset + length > object_size)
1102 		length = object_size - offset;
1103 
1104 	while (length != 0) {
1105 		uint64_t chunk_end, chunk_begin, chunk_len;
1106 		uint64_t l1blks;
1107 		dmu_tx_t *tx;
1108 
1109 		if (dmu_objset_zfs_unmounting(dn->dn_objset))
1110 			return (SET_ERROR(EINTR));
1111 
1112 		chunk_end = chunk_begin = offset + length;
1113 
1114 		/* move chunk_begin backwards to the beginning of this chunk */
1115 		err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
1116 		if (err)
1117 			return (err);
1118 		ASSERT3U(chunk_begin, >=, offset);
1119 		ASSERT3U(chunk_begin, <=, chunk_end);
1120 
1121 		chunk_len = chunk_end - chunk_begin;
1122 
1123 		tx = dmu_tx_create(os);
1124 		dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
1125 
1126 		/*
1127 		 * Mark this transaction as typically resulting in a net
1128 		 * reduction in space used.
1129 		 */
1130 		dmu_tx_mark_netfree(tx);
1131 		err = dmu_tx_assign(tx, DMU_TX_WAIT);
1132 		if (err) {
1133 			dmu_tx_abort(tx);
1134 			return (err);
1135 		}
1136 
1137 		uint64_t txg = dmu_tx_get_txg(tx);
1138 
1139 		mutex_enter(&dp->dp_lock);
1140 		uint64_t long_free_dirty =
1141 		    dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
1142 		mutex_exit(&dp->dp_lock);
1143 
1144 		/*
1145 		 * To avoid filling up a TXG with just frees, wait for
1146 		 * the next TXG to open before freeing more chunks if
1147 		 * we have reached the threshold of frees.
1148 		 */
1149 		if (dirty_frees_threshold != 0 &&
1150 		    long_free_dirty >= dirty_frees_threshold) {
1151 			DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
1152 			dmu_tx_commit(tx);
1153 			txg_wait_open(dp, 0, B_TRUE);
1154 			continue;
1155 		}
1156 
1157 		/*
1158 		 * In order to prevent unnecessary write throttling, for each
1159 		 * TXG, we track the cumulative size of L1 blocks being dirtied
1160 		 * in dnode_free_range() below. We compare this number to a
1161 		 * tunable threshold, past which we prevent new L1 dirty freeing
1162 		 * blocks from being added into the open TXG. See
1163 		 * dmu_free_long_range_impl() for details. The threshold
1164 		 * prevents write throttle activation due to dirty freeing L1
1165 		 * blocks taking up a large percentage of zfs_dirty_data_max.
1166 		 */
1167 		mutex_enter(&dp->dp_lock);
1168 		dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
1169 		    l1blks << dn->dn_indblkshift;
1170 		mutex_exit(&dp->dp_lock);
1171 		DTRACE_PROBE3(free__long__range,
1172 		    uint64_t, long_free_dirty, uint64_t, chunk_len,
1173 		    uint64_t, txg);
1174 		dnode_free_range(dn, chunk_begin, chunk_len, tx);
1175 
1176 		dmu_tx_commit(tx);
1177 
1178 		length -= chunk_len;
1179 	}
1180 	return (0);
1181 }
1182 
1183 int
dmu_free_long_range(objset_t * os,uint64_t object,uint64_t offset,uint64_t length)1184 dmu_free_long_range(objset_t *os, uint64_t object,
1185     uint64_t offset, uint64_t length)
1186 {
1187 	dnode_t *dn;
1188 	int err;
1189 
1190 	err = dnode_hold(os, object, FTAG, &dn);
1191 	if (err != 0)
1192 		return (err);
1193 	err = dmu_free_long_range_impl(os, dn, offset, length);
1194 
1195 	/*
1196 	 * It is important to zero out the maxblkid when freeing the entire
1197 	 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
1198 	 * will take the fast path, and (b) dnode_reallocate() can verify
1199 	 * that the entire file has been freed.
1200 	 */
1201 	if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
1202 		dn->dn_maxblkid = 0;
1203 
1204 	dnode_rele(dn, FTAG);
1205 	return (err);
1206 }
1207 
1208 int
dmu_free_long_object(objset_t * os,uint64_t object)1209 dmu_free_long_object(objset_t *os, uint64_t object)
1210 {
1211 	dmu_tx_t *tx;
1212 	int err;
1213 
1214 	err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
1215 	if (err != 0)
1216 		return (err);
1217 
1218 	tx = dmu_tx_create(os);
1219 	dmu_tx_hold_bonus(tx, object);
1220 	dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
1221 	dmu_tx_mark_netfree(tx);
1222 	err = dmu_tx_assign(tx, DMU_TX_WAIT);
1223 	if (err == 0) {
1224 		err = dmu_object_free(os, object, tx);
1225 		dmu_tx_commit(tx);
1226 	} else {
1227 		dmu_tx_abort(tx);
1228 	}
1229 
1230 	return (err);
1231 }
1232 
1233 int
dmu_free_range(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,dmu_tx_t * tx)1234 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
1235     uint64_t size, dmu_tx_t *tx)
1236 {
1237 	dnode_t *dn;
1238 	int err = dnode_hold(os, object, FTAG, &dn);
1239 	if (err)
1240 		return (err);
1241 	ASSERT(offset < UINT64_MAX);
1242 	ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
1243 	dnode_free_range(dn, offset, size, tx);
1244 	dnode_rele(dn, FTAG);
1245 	return (0);
1246 }
1247 
1248 static int
dmu_read_impl(dnode_t * dn,uint64_t offset,uint64_t size,void * buf,dmu_flags_t flags)1249 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
1250     void *buf, dmu_flags_t flags)
1251 {
1252 	dmu_buf_t **dbp;
1253 	int numbufs, err = 0;
1254 
1255 	/*
1256 	 * Deal with odd block sizes, where there can't be data past the first
1257 	 * block. If we ever do the tail block optimization, we will need to
1258 	 * handle that here as well.
1259 	 */
1260 	if (dn->dn_maxblkid == 0) {
1261 		uint64_t newsz = offset > dn->dn_datablksz ? 0 :
1262 		    MIN(size, dn->dn_datablksz - offset);
1263 		memset((char *)buf + newsz, 0, size - newsz);
1264 		size = newsz;
1265 	}
1266 
1267 	if (size == 0)
1268 		return (0);
1269 
1270 	/* Allow Direct I/O when requested and properly aligned */
1271 	if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned(buf) &&
1272 	    zfs_dio_aligned(offset, size, PAGESIZE)) {
1273 		abd_t *data = abd_get_from_buf(buf, size);
1274 		err = dmu_read_abd(dn, offset, size, data, flags);
1275 		abd_free(data);
1276 		return (err);
1277 	}
1278 	flags &= ~DMU_DIRECTIO;
1279 
1280 	while (size > 0) {
1281 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1282 		int i;
1283 
1284 		/*
1285 		 * NB: we could do this block-at-a-time, but it's nice
1286 		 * to be reading in parallel.
1287 		 */
1288 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1289 		    TRUE, FTAG, &numbufs, &dbp, flags);
1290 		if (err)
1291 			break;
1292 
1293 		for (i = 0; i < numbufs; i++) {
1294 			uint64_t tocpy;
1295 			int64_t bufoff;
1296 			dmu_buf_t *db = dbp[i];
1297 
1298 			ASSERT(size > 0);
1299 
1300 			bufoff = offset - db->db_offset;
1301 			tocpy = MIN(db->db_size - bufoff, size);
1302 
1303 			ASSERT(db->db_data != NULL);
1304 			(void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1305 
1306 			offset += tocpy;
1307 			size -= tocpy;
1308 			buf = (char *)buf + tocpy;
1309 		}
1310 		dmu_buf_rele_array(dbp, numbufs, FTAG);
1311 	}
1312 	return (err);
1313 }
1314 
1315 int
dmu_read(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,void * buf,dmu_flags_t flags)1316 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1317     void *buf, dmu_flags_t flags)
1318 {
1319 	dnode_t *dn;
1320 	int err;
1321 
1322 	err = dnode_hold(os, object, FTAG, &dn);
1323 	if (err != 0)
1324 		return (err);
1325 
1326 	err = dmu_read_impl(dn, offset, size, buf, flags);
1327 	dnode_rele(dn, FTAG);
1328 	return (err);
1329 }
1330 
1331 int
dmu_read_by_dnode(dnode_t * dn,uint64_t offset,uint64_t size,void * buf,dmu_flags_t flags)1332 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1333     dmu_flags_t flags)
1334 {
1335 	return (dmu_read_impl(dn, offset, size, buf, flags));
1336 }
1337 
1338 static void
dmu_write_impl(dmu_buf_t ** dbp,int numbufs,uint64_t offset,uint64_t size,const void * buf,dmu_tx_t * tx,dmu_flags_t flags)1339 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1340     const void *buf, dmu_tx_t *tx, dmu_flags_t flags)
1341 {
1342 	int i;
1343 
1344 	for (i = 0; i < numbufs; i++) {
1345 		uint64_t tocpy;
1346 		int64_t bufoff;
1347 		dmu_buf_t *db = dbp[i];
1348 
1349 		ASSERT(size > 0);
1350 
1351 		bufoff = offset - db->db_offset;
1352 		tocpy = MIN(db->db_size - bufoff, size);
1353 
1354 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1355 
1356 		if (tocpy == db->db_size) {
1357 			dmu_buf_will_fill_flags(db, tx, B_FALSE, flags);
1358 		} else {
1359 			if (i == numbufs - 1 && bufoff + tocpy < db->db_size) {
1360 				if (bufoff == 0)
1361 					flags |= DMU_PARTIAL_FIRST;
1362 				else
1363 					flags |= DMU_PARTIAL_MORE;
1364 			}
1365 			dmu_buf_will_dirty_flags(db, tx, flags);
1366 		}
1367 
1368 		ASSERT(db->db_data != NULL);
1369 		(void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1370 
1371 		if (tocpy == db->db_size)
1372 			dmu_buf_fill_done(db, tx, B_FALSE);
1373 
1374 		offset += tocpy;
1375 		size -= tocpy;
1376 		buf = (char *)buf + tocpy;
1377 	}
1378 }
1379 
1380 void
dmu_write(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,const void * buf,dmu_tx_t * tx,dmu_flags_t flags)1381 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1382     const void *buf, dmu_tx_t *tx, dmu_flags_t flags)
1383 {
1384 	dmu_buf_t **dbp;
1385 	int numbufs;
1386 
1387 	if (size == 0)
1388 		return;
1389 
1390 	VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1391 	    FALSE, FTAG, &numbufs, &dbp, flags));
1392 	dmu_write_impl(dbp, numbufs, offset, size, buf, tx, flags);
1393 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1394 }
1395 
1396 int
dmu_write_by_dnode(dnode_t * dn,uint64_t offset,uint64_t size,const void * buf,dmu_tx_t * tx,dmu_flags_t flags)1397 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1398     const void *buf, dmu_tx_t *tx, dmu_flags_t flags)
1399 {
1400 	dmu_buf_t **dbp;
1401 	int numbufs;
1402 	int error;
1403 
1404 	if (size == 0)
1405 		return (0);
1406 
1407 	/* Allow Direct I/O when requested and properly aligned */
1408 	if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned((void *)buf) &&
1409 	    zfs_dio_aligned(offset, size, dn->dn_datablksz)) {
1410 		abd_t *data = abd_get_from_buf((void *)buf, size);
1411 		error = dmu_write_abd(dn, offset, size, data, flags, tx);
1412 		abd_free(data);
1413 		return (error);
1414 	}
1415 	flags &= ~DMU_DIRECTIO;
1416 
1417 	VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1418 	    FALSE, FTAG, &numbufs, &dbp, flags));
1419 	dmu_write_impl(dbp, numbufs, offset, size, buf, tx, flags);
1420 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1421 	return (0);
1422 }
1423 
1424 void
dmu_prealloc(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,dmu_tx_t * tx)1425 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1426     dmu_tx_t *tx)
1427 {
1428 	dmu_buf_t **dbp;
1429 	int numbufs, i;
1430 
1431 	if (size == 0)
1432 		return;
1433 
1434 	VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1435 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1436 
1437 	for (i = 0; i < numbufs; i++) {
1438 		dmu_buf_t *db = dbp[i];
1439 
1440 		dmu_buf_will_not_fill(db, tx);
1441 	}
1442 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1443 }
1444 
1445 void
dmu_write_embedded(objset_t * os,uint64_t object,uint64_t offset,void * data,uint8_t etype,uint8_t comp,int uncompressed_size,int compressed_size,int byteorder,dmu_tx_t * tx)1446 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1447     void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1448     int compressed_size, int byteorder, dmu_tx_t *tx)
1449 {
1450 	dmu_buf_t *db;
1451 
1452 	ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1453 	ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1454 	VERIFY0(dmu_buf_hold_noread(os, object, offset,
1455 	    FTAG, &db));
1456 
1457 	dmu_buf_write_embedded(db,
1458 	    data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1459 	    uncompressed_size, compressed_size, byteorder, tx);
1460 
1461 	dmu_buf_rele(db, FTAG);
1462 }
1463 
1464 void
dmu_redact(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,dmu_tx_t * tx)1465 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1466     dmu_tx_t *tx)
1467 {
1468 	int numbufs, i;
1469 	dmu_buf_t **dbp;
1470 
1471 	VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1472 	    &numbufs, &dbp, DMU_READ_PREFETCH));
1473 	for (i = 0; i < numbufs; i++)
1474 		dmu_buf_redact(dbp[i], tx);
1475 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1476 }
1477 
1478 #ifdef _KERNEL
1479 int
dmu_read_uio_dnode(dnode_t * dn,zfs_uio_t * uio,uint64_t size,dmu_flags_t flags)1480 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size,
1481     dmu_flags_t flags)
1482 {
1483 	dmu_buf_t **dbp;
1484 	int numbufs, i, err;
1485 
1486 	if ((flags & DMU_DIRECTIO) && (uio->uio_extflg & UIO_DIRECT))
1487 		return (dmu_read_uio_direct(dn, uio, size, flags));
1488 	flags &= ~DMU_DIRECTIO;
1489 
1490 	/*
1491 	 * NB: we could do this block-at-a-time, but it's nice
1492 	 * to be reading in parallel.
1493 	 */
1494 	err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1495 	    TRUE, FTAG, &numbufs, &dbp, flags);
1496 	if (err)
1497 		return (err);
1498 
1499 	for (i = 0; i < numbufs; i++) {
1500 		uint64_t tocpy;
1501 		int64_t bufoff;
1502 		dmu_buf_t *db = dbp[i];
1503 
1504 		ASSERT(size > 0);
1505 
1506 		bufoff = zfs_uio_offset(uio) - db->db_offset;
1507 		tocpy = MIN(db->db_size - bufoff, size);
1508 
1509 		ASSERT(db->db_data != NULL);
1510 		err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1511 		    UIO_READ, uio);
1512 
1513 		if (err)
1514 			break;
1515 
1516 		size -= tocpy;
1517 	}
1518 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1519 
1520 	return (err);
1521 }
1522 
1523 /*
1524  * Read 'size' bytes into the uio buffer.
1525  * From object zdb->db_object.
1526  * Starting at zfs_uio_offset(uio).
1527  *
1528  * If the caller already has a dbuf in the target object
1529  * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1530  * because we don't have to find the dnode_t for the object.
1531  */
1532 int
dmu_read_uio_dbuf(dmu_buf_t * zdb,zfs_uio_t * uio,uint64_t size,dmu_flags_t flags)1533 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1534     dmu_flags_t flags)
1535 {
1536 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1537 	int err;
1538 
1539 	if (size == 0)
1540 		return (0);
1541 
1542 	DB_DNODE_ENTER(db);
1543 	err = dmu_read_uio_dnode(DB_DNODE(db), uio, size, flags);
1544 	DB_DNODE_EXIT(db);
1545 
1546 	return (err);
1547 }
1548 
1549 /*
1550  * Read 'size' bytes into the uio buffer.
1551  * From the specified object
1552  * Starting at offset zfs_uio_offset(uio).
1553  */
1554 int
dmu_read_uio(objset_t * os,uint64_t object,zfs_uio_t * uio,uint64_t size,dmu_flags_t flags)1555 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1556     dmu_flags_t flags)
1557 {
1558 	dnode_t *dn;
1559 	int err;
1560 
1561 	if (size == 0)
1562 		return (0);
1563 
1564 	err = dnode_hold(os, object, FTAG, &dn);
1565 	if (err)
1566 		return (err);
1567 
1568 	err = dmu_read_uio_dnode(dn, uio, size, flags);
1569 
1570 	dnode_rele(dn, FTAG);
1571 
1572 	return (err);
1573 }
1574 
1575 int
dmu_write_uio_dnode(dnode_t * dn,zfs_uio_t * uio,uint64_t size,dmu_tx_t * tx,dmu_flags_t flags)1576 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx,
1577     dmu_flags_t flags)
1578 {
1579 	dmu_buf_t **dbp;
1580 	int numbufs;
1581 	int err = 0;
1582 	uint64_t write_size;
1583 	dmu_flags_t oflags = flags;
1584 
1585 top:
1586 	write_size = size;
1587 
1588 	/*
1589 	 * We only allow Direct I/O writes to happen if we are block
1590 	 * sized aligned. Otherwise, we pass the write off to the ARC.
1591 	 */
1592 	if ((flags & DMU_DIRECTIO) && (uio->uio_extflg & UIO_DIRECT) &&
1593 	    (write_size >= dn->dn_datablksz)) {
1594 		if (zfs_dio_aligned(zfs_uio_offset(uio), write_size,
1595 		    dn->dn_datablksz)) {
1596 			return (dmu_write_uio_direct(dn, uio, size, flags, tx));
1597 		} else if (write_size > dn->dn_datablksz &&
1598 		    zfs_dio_offset_aligned(zfs_uio_offset(uio),
1599 		    dn->dn_datablksz)) {
1600 			write_size =
1601 			    dn->dn_datablksz * (write_size / dn->dn_datablksz);
1602 			err = dmu_write_uio_direct(dn, uio, write_size, flags,
1603 			    tx);
1604 			if (err == 0) {
1605 				size -= write_size;
1606 				goto top;
1607 			} else {
1608 				return (err);
1609 			}
1610 		} else {
1611 			write_size =
1612 			    P2PHASE(zfs_uio_offset(uio), dn->dn_datablksz);
1613 		}
1614 	}
1615 	flags &= ~DMU_DIRECTIO;
1616 
1617 	err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), write_size,
1618 	    FALSE, FTAG, &numbufs, &dbp, flags);
1619 	if (err)
1620 		return (err);
1621 
1622 	for (int i = 0; i < numbufs; i++) {
1623 		uint64_t tocpy;
1624 		int64_t bufoff;
1625 		dmu_buf_t *db = dbp[i];
1626 
1627 		ASSERT(write_size > 0);
1628 
1629 		offset_t off = zfs_uio_offset(uio);
1630 		bufoff = off - db->db_offset;
1631 		tocpy = MIN(db->db_size - bufoff, write_size);
1632 
1633 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1634 
1635 		if (tocpy == db->db_size) {
1636 			dmu_buf_will_fill_flags(db, tx, B_TRUE, flags);
1637 		} else {
1638 			if (i == numbufs - 1 && bufoff + tocpy < db->db_size) {
1639 				if (bufoff == 0)
1640 					flags |= DMU_PARTIAL_FIRST;
1641 				else
1642 					flags |= DMU_PARTIAL_MORE;
1643 			}
1644 			dmu_buf_will_dirty_flags(db, tx, flags);
1645 		}
1646 
1647 		ASSERT(db->db_data != NULL);
1648 		err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1649 		    tocpy, UIO_WRITE, uio);
1650 
1651 		if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) {
1652 			/* The fill was reverted.  Undo any uio progress. */
1653 			zfs_uio_advance(uio, off - zfs_uio_offset(uio));
1654 		}
1655 
1656 		if (err)
1657 			break;
1658 
1659 		write_size -= tocpy;
1660 		size -= tocpy;
1661 	}
1662 
1663 	IMPLY(err == 0, write_size == 0);
1664 
1665 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1666 
1667 	if ((oflags & DMU_DIRECTIO) && (uio->uio_extflg & UIO_DIRECT) &&
1668 	    err == 0 && size > 0) {
1669 		flags = oflags;
1670 		goto top;
1671 	}
1672 	IMPLY(err == 0, size == 0);
1673 
1674 	return (err);
1675 }
1676 
1677 /*
1678  * Write 'size' bytes from the uio buffer.
1679  * To object zdb->db_object.
1680  * Starting at offset zfs_uio_offset(uio).
1681  *
1682  * If the caller already has a dbuf in the target object
1683  * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1684  * because we don't have to find the dnode_t for the object.
1685  */
1686 int
dmu_write_uio_dbuf(dmu_buf_t * zdb,zfs_uio_t * uio,uint64_t size,dmu_tx_t * tx,dmu_flags_t flags)1687 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1688     dmu_tx_t *tx, dmu_flags_t flags)
1689 {
1690 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1691 	int err;
1692 
1693 	if (size == 0)
1694 		return (0);
1695 
1696 	DB_DNODE_ENTER(db);
1697 	err = dmu_write_uio_dnode(DB_DNODE(db), uio, size, tx, flags);
1698 	DB_DNODE_EXIT(db);
1699 
1700 	return (err);
1701 }
1702 
1703 /*
1704  * Write 'size' bytes from the uio buffer.
1705  * To the specified object.
1706  * Starting at offset zfs_uio_offset(uio).
1707  */
1708 int
dmu_write_uio(objset_t * os,uint64_t object,zfs_uio_t * uio,uint64_t size,dmu_tx_t * tx,dmu_flags_t flags)1709 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1710     dmu_tx_t *tx, dmu_flags_t flags)
1711 {
1712 	dnode_t *dn;
1713 	int err;
1714 
1715 	if (size == 0)
1716 		return (0);
1717 
1718 	err = dnode_hold(os, object, FTAG, &dn);
1719 	if (err)
1720 		return (err);
1721 
1722 	err = dmu_write_uio_dnode(dn, uio, size, tx, flags);
1723 
1724 	dnode_rele(dn, FTAG);
1725 
1726 	return (err);
1727 }
1728 #endif /* _KERNEL */
1729 
1730 static void
dmu_cached_bps(spa_t * spa,blkptr_t * bps,uint_t nbps,uint64_t * l1sz,uint64_t * l2sz)1731 dmu_cached_bps(spa_t *spa, blkptr_t *bps, uint_t nbps,
1732     uint64_t *l1sz, uint64_t *l2sz)
1733 {
1734 	int cached_flags;
1735 
1736 	if (bps == NULL)
1737 		return;
1738 
1739 	for (size_t blk_off = 0; blk_off < nbps; blk_off++) {
1740 		blkptr_t *bp = &bps[blk_off];
1741 
1742 		if (BP_IS_HOLE(bp))
1743 			continue;
1744 
1745 		cached_flags = arc_cached(spa, bp);
1746 		if (cached_flags == 0)
1747 			continue;
1748 
1749 		if ((cached_flags & (ARC_CACHED_IN_L1 | ARC_CACHED_IN_L2)) ==
1750 		    ARC_CACHED_IN_L2)
1751 			*l2sz += BP_GET_LSIZE(bp);
1752 		else
1753 			*l1sz += BP_GET_LSIZE(bp);
1754 	}
1755 }
1756 
1757 /*
1758  * Estimate DMU object cached size.
1759  */
1760 int
dmu_object_cached_size(objset_t * os,uint64_t object,uint64_t * l1sz,uint64_t * l2sz)1761 dmu_object_cached_size(objset_t *os, uint64_t object,
1762     uint64_t *l1sz, uint64_t *l2sz)
1763 {
1764 	dnode_t *dn;
1765 	dmu_object_info_t doi;
1766 	int err = 0;
1767 
1768 	*l1sz = *l2sz = 0;
1769 
1770 	if (dnode_hold(os, object, FTAG, &dn) != 0)
1771 		return (0);
1772 
1773 	if (dn->dn_nlevels < 2) {
1774 		dnode_rele(dn, FTAG);
1775 		return (0);
1776 	}
1777 
1778 	dmu_object_info_from_dnode(dn, &doi);
1779 
1780 	for (uint64_t off = 0; off < doi.doi_max_offset &&
1781 	    dmu_prefetch_max > 0; off += dmu_prefetch_max) {
1782 		/* dbuf_read doesn't prefetch L1 blocks. */
1783 		dmu_prefetch_by_dnode(dn, 1, off,
1784 		    dmu_prefetch_max, ZIO_PRIORITY_SYNC_READ);
1785 	}
1786 
1787 	/*
1788 	 * Hold all valid L1 blocks, asking ARC the status of each BP
1789 	 * contained in each such L1 block.
1790 	 */
1791 	uint_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1);
1792 	uint64_t l1blks = 1 + (dn->dn_maxblkid / nbps);
1793 
1794 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1795 	for (uint64_t blk = 0; blk < l1blks; blk++) {
1796 		dmu_buf_impl_t *db = NULL;
1797 
1798 		if (issig()) {
1799 			/*
1800 			 * On interrupt, get out, and bubble up EINTR
1801 			 */
1802 			err = EINTR;
1803 			break;
1804 		}
1805 
1806 		/*
1807 		 * If we get an i/o error here, the L1 can't be read,
1808 		 * and nothing under it could be cached, so we just
1809 		 * continue. Ignoring the error from dbuf_hold_impl
1810 		 * or from dbuf_read is then a reasonable choice.
1811 		 */
1812 		err = dbuf_hold_impl(dn, 1, blk, B_TRUE, B_FALSE, FTAG, &db);
1813 		if (err != 0) {
1814 			/*
1815 			 * ignore error and continue
1816 			 */
1817 			err = 0;
1818 			continue;
1819 		}
1820 
1821 		err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1822 		if (err == 0) {
1823 			dmu_cached_bps(dmu_objset_spa(os), db->db.db_data,
1824 			    nbps, l1sz, l2sz);
1825 		}
1826 		/*
1827 		 * error may be ignored, and we continue
1828 		 */
1829 		err = 0;
1830 		dbuf_rele(db, FTAG);
1831 	}
1832 	rw_exit(&dn->dn_struct_rwlock);
1833 
1834 	dnode_rele(dn, FTAG);
1835 	return (err);
1836 }
1837 
1838 /*
1839  * Allocate a loaned anonymous arc buffer.
1840  */
1841 arc_buf_t *
dmu_request_arcbuf(dmu_buf_t * handle,int size)1842 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1843 {
1844 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1845 
1846 	return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1847 }
1848 
1849 /*
1850  * Free a loaned arc buffer.
1851  */
1852 void
dmu_return_arcbuf(arc_buf_t * buf)1853 dmu_return_arcbuf(arc_buf_t *buf)
1854 {
1855 	arc_return_buf(buf, FTAG);
1856 	arc_buf_destroy(buf, FTAG);
1857 }
1858 
1859 /*
1860  * A "lightweight" write is faster than a regular write (e.g.
1861  * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1862  * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t.  However, the
1863  * data can not be read or overwritten until the transaction's txg has been
1864  * synced.  This makes it appropriate for workloads that are known to be
1865  * (temporarily) write-only, like "zfs receive".
1866  *
1867  * A single block is written, starting at the specified offset in bytes.  If
1868  * the call is successful, it returns 0 and the provided abd has been
1869  * consumed (the caller should not free it).
1870  */
1871 int
dmu_lightweight_write_by_dnode(dnode_t * dn,uint64_t offset,abd_t * abd,const zio_prop_t * zp,zio_flag_t flags,dmu_tx_t * tx)1872 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1873     const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx)
1874 {
1875 	dbuf_dirty_record_t *dr =
1876 	    dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
1877 	if (dr == NULL)
1878 		return (SET_ERROR(EIO));
1879 	dr->dt.dll.dr_abd = abd;
1880 	dr->dt.dll.dr_props = *zp;
1881 	dr->dt.dll.dr_flags = flags;
1882 	return (0);
1883 }
1884 
1885 /*
1886  * When possible directly assign passed loaned arc buffer to a dbuf.
1887  * If this is not possible copy the contents of passed arc buf via
1888  * dmu_write().
1889  */
1890 int
dmu_assign_arcbuf_by_dnode(dnode_t * dn,uint64_t offset,arc_buf_t * buf,dmu_tx_t * tx,dmu_flags_t flags)1891 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1892     dmu_tx_t *tx, dmu_flags_t flags)
1893 {
1894 	dmu_buf_impl_t *db;
1895 	objset_t *os = dn->dn_objset;
1896 	uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1897 	uint64_t blkid;
1898 
1899 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1900 	blkid = dbuf_whichblock(dn, 0, offset);
1901 	db = dbuf_hold(dn, blkid, FTAG);
1902 	rw_exit(&dn->dn_struct_rwlock);
1903 	if (db == NULL)
1904 		return (SET_ERROR(EIO));
1905 
1906 	/*
1907 	 * We can only assign if the offset is aligned and the arc buf is the
1908 	 * same size as the dbuf.
1909 	 */
1910 	if (offset == db->db.db_offset && blksz == db->db.db_size) {
1911 		zfs_racct_write(os->os_spa, blksz, 1, flags);
1912 		dbuf_assign_arcbuf(db, buf, tx, flags);
1913 		dbuf_rele(db, FTAG);
1914 	} else {
1915 		/* compressed bufs must always be assignable to their dbuf */
1916 		ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1917 		ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1918 
1919 		dbuf_rele(db, FTAG);
1920 		dmu_write_by_dnode(dn, offset, blksz, buf->b_data, tx, flags);
1921 		dmu_return_arcbuf(buf);
1922 	}
1923 
1924 	return (0);
1925 }
1926 
1927 int
dmu_assign_arcbuf_by_dbuf(dmu_buf_t * handle,uint64_t offset,arc_buf_t * buf,dmu_tx_t * tx,dmu_flags_t flags)1928 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1929     dmu_tx_t *tx, dmu_flags_t flags)
1930 {
1931 	int err;
1932 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1933 
1934 	DB_DNODE_ENTER(db);
1935 	err = dmu_assign_arcbuf_by_dnode(DB_DNODE(db), offset, buf, tx, flags);
1936 	DB_DNODE_EXIT(db);
1937 
1938 	return (err);
1939 }
1940 
1941 void
dmu_sync_ready(zio_t * zio,arc_buf_t * buf,void * varg)1942 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1943 {
1944 	(void) buf;
1945 	dmu_sync_arg_t *dsa = varg;
1946 
1947 	if (zio->io_error == 0) {
1948 		dbuf_dirty_record_t *dr = dsa->dsa_dr;
1949 		blkptr_t *bp = zio->io_bp;
1950 
1951 		if (BP_IS_HOLE(bp)) {
1952 			dmu_buf_t *db = NULL;
1953 			if (dr)
1954 				db = &(dr->dr_dbuf->db);
1955 			else
1956 				db = dsa->dsa_zgd->zgd_db;
1957 			/*
1958 			 * A block of zeros may compress to a hole, but the
1959 			 * block size still needs to be known for replay.
1960 			 */
1961 			BP_SET_LSIZE(bp, db->db_size);
1962 		} else if (!BP_IS_EMBEDDED(bp)) {
1963 			ASSERT0(BP_GET_LEVEL(bp));
1964 			BP_SET_FILL(bp, 1);
1965 		}
1966 	}
1967 }
1968 
1969 static void
dmu_sync_late_arrival_ready(zio_t * zio)1970 dmu_sync_late_arrival_ready(zio_t *zio)
1971 {
1972 	dmu_sync_ready(zio, NULL, zio->io_private);
1973 }
1974 
1975 void
dmu_sync_done(zio_t * zio,arc_buf_t * buf,void * varg)1976 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1977 {
1978 	(void) buf;
1979 	dmu_sync_arg_t *dsa = varg;
1980 	dbuf_dirty_record_t *dr = dsa->dsa_dr;
1981 	dmu_buf_impl_t *db = dr->dr_dbuf;
1982 	zgd_t *zgd = dsa->dsa_zgd;
1983 
1984 	/*
1985 	 * Record the vdev(s) backing this blkptr so they can be flushed after
1986 	 * the writes for the lwb have completed.
1987 	 */
1988 	if (zgd && zio->io_error == 0) {
1989 		zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1990 	}
1991 
1992 	mutex_enter(&db->db_mtx);
1993 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1994 	if (zio->io_error == 0) {
1995 		ASSERT0(dr->dt.dl.dr_has_raw_params);
1996 		dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1997 		if (dr->dt.dl.dr_nopwrite) {
1998 			blkptr_t *bp = zio->io_bp;
1999 			blkptr_t *bp_orig = &zio->io_bp_orig;
2000 			uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
2001 
2002 			ASSERT(BP_EQUAL(bp, bp_orig));
2003 			VERIFY(BP_EQUAL(bp, db->db_blkptr));
2004 			ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
2005 			VERIFY(zio_checksum_table[chksum].ci_flags &
2006 			    ZCHECKSUM_FLAG_NOPWRITE);
2007 		}
2008 		dr->dt.dl.dr_overridden_by = *zio->io_bp;
2009 		dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
2010 		dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
2011 		dr->dt.dl.dr_gang_copies = zio->io_prop.zp_gang_copies;
2012 
2013 		/*
2014 		 * Old style holes are filled with all zeros, whereas
2015 		 * new-style holes maintain their lsize, type, level,
2016 		 * and birth time (see zio_write_compress). While we
2017 		 * need to reset the BP_SET_LSIZE() call that happened
2018 		 * in dmu_sync_ready for old style holes, we do *not*
2019 		 * want to wipe out the information contained in new
2020 		 * style holes. Thus, only zero out the block pointer if
2021 		 * it's an old style hole.
2022 		 */
2023 		if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
2024 		    BP_GET_LOGICAL_BIRTH(&dr->dt.dl.dr_overridden_by) == 0)
2025 			BP_ZERO(&dr->dt.dl.dr_overridden_by);
2026 	} else {
2027 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2028 	}
2029 
2030 	cv_broadcast(&db->db_changed);
2031 	mutex_exit(&db->db_mtx);
2032 
2033 	if (dsa->dsa_done)
2034 		dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
2035 
2036 	kmem_free(dsa, sizeof (*dsa));
2037 }
2038 
2039 static void
dmu_sync_late_arrival_done(zio_t * zio)2040 dmu_sync_late_arrival_done(zio_t *zio)
2041 {
2042 	blkptr_t *bp = zio->io_bp;
2043 	dmu_sync_arg_t *dsa = zio->io_private;
2044 	zgd_t *zgd = dsa->dsa_zgd;
2045 
2046 	if (zio->io_error == 0) {
2047 		/*
2048 		 * Record the vdev(s) backing this blkptr so they can be
2049 		 * flushed after the writes for the lwb have completed.
2050 		 */
2051 		zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
2052 
2053 		if (!BP_IS_HOLE(bp)) {
2054 			blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
2055 			ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
2056 			ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
2057 			ASSERT(BP_GET_BIRTH(zio->io_bp) == zio->io_txg);
2058 			ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
2059 			zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
2060 		}
2061 	}
2062 
2063 	dmu_tx_commit(dsa->dsa_tx);
2064 
2065 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
2066 
2067 	abd_free(zio->io_abd);
2068 	kmem_free(dsa, sizeof (*dsa));
2069 }
2070 
2071 static int
dmu_sync_late_arrival(zio_t * pio,objset_t * os,dmu_sync_cb_t * done,zgd_t * zgd,zio_prop_t * zp,zbookmark_phys_t * zb)2072 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
2073     zio_prop_t *zp, zbookmark_phys_t *zb)
2074 {
2075 	dmu_sync_arg_t *dsa;
2076 	dmu_tx_t *tx;
2077 	int error;
2078 
2079 	error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL,
2080 	    DB_RF_CANFAIL | DMU_READ_NO_PREFETCH | DMU_KEEP_CACHING);
2081 	if (error != 0)
2082 		return (error);
2083 
2084 	tx = dmu_tx_create(os);
2085 	dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
2086 	/*
2087 	 * This transaction does not produce any dirty data or log blocks, so
2088 	 * it should not be throttled.  All other cases wait for TXG sync, by
2089 	 * which time the log block we are writing will be obsolete, so we can
2090 	 * skip waiting and just return error here instead.
2091 	 */
2092 	if (dmu_tx_assign(tx, DMU_TX_NOWAIT | DMU_TX_NOTHROTTLE) != 0) {
2093 		dmu_tx_abort(tx);
2094 		/* Make zl_get_data do txg_waited_synced() */
2095 		return (SET_ERROR(EIO));
2096 	}
2097 
2098 	/*
2099 	 * In order to prevent the zgd's lwb from being free'd prior to
2100 	 * dmu_sync_late_arrival_done() being called, we have to ensure
2101 	 * the lwb's "max txg" takes this tx's txg into account.
2102 	 */
2103 	zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
2104 
2105 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2106 	dsa->dsa_dr = NULL;
2107 	dsa->dsa_done = done;
2108 	dsa->dsa_zgd = zgd;
2109 	dsa->dsa_tx = tx;
2110 
2111 	/*
2112 	 * Since we are currently syncing this txg, it's nontrivial to
2113 	 * determine what BP to nopwrite against, so we disable nopwrite.
2114 	 *
2115 	 * When syncing, the db_blkptr is initially the BP of the previous
2116 	 * txg.  We can not nopwrite against it because it will be changed
2117 	 * (this is similar to the non-late-arrival case where the dbuf is
2118 	 * dirty in a future txg).
2119 	 *
2120 	 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
2121 	 * We can not nopwrite against it because although the BP will not
2122 	 * (typically) be changed, the data has not yet been persisted to this
2123 	 * location.
2124 	 *
2125 	 * Finally, when dbuf_write_done() is called, it is theoretically
2126 	 * possible to always nopwrite, because the data that was written in
2127 	 * this txg is the same data that we are trying to write.  However we
2128 	 * would need to check that this dbuf is not dirty in any future
2129 	 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
2130 	 * don't nopwrite in this case.
2131 	 */
2132 	zp->zp_nopwrite = B_FALSE;
2133 
2134 	zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
2135 	    abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
2136 	    zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
2137 	    dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done,
2138 	    dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
2139 
2140 	return (0);
2141 }
2142 
2143 /*
2144  * Intent log support: sync the block associated with db to disk.
2145  * N.B. and XXX: the caller is responsible for making sure that the
2146  * data isn't changing while dmu_sync() is writing it.
2147  *
2148  * Return values:
2149  *
2150  *	EEXIST: this txg has already been synced, so there's nothing to do.
2151  *		The caller should not log the write.
2152  *
2153  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
2154  *		The caller should not log the write.
2155  *
2156  *	EALREADY: this block is already in the process of being synced.
2157  *		The caller should track its progress (somehow).
2158  *
2159  *	EIO: could not do the I/O.
2160  *		The caller should do a txg_wait_synced().
2161  *
2162  *	0: the I/O has been initiated.
2163  *		The caller should log this blkptr in the done callback.
2164  *		It is possible that the I/O will fail, in which case
2165  *		the error will be reported to the done callback and
2166  *		propagated to pio from zio_done().
2167  */
2168 int
dmu_sync(zio_t * pio,uint64_t txg,dmu_sync_cb_t * done,zgd_t * zgd)2169 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
2170 {
2171 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
2172 	objset_t *os = db->db_objset;
2173 	dsl_dataset_t *ds = os->os_dsl_dataset;
2174 	dbuf_dirty_record_t *dr, *dr_next;
2175 	dmu_sync_arg_t *dsa;
2176 	zbookmark_phys_t zb;
2177 	zio_prop_t zp;
2178 
2179 	ASSERT(pio != NULL);
2180 	ASSERT(txg != 0);
2181 
2182 	SET_BOOKMARK(&zb, ds->ds_object,
2183 	    db->db.db_object, db->db_level, db->db_blkid);
2184 
2185 	DB_DNODE_ENTER(db);
2186 	dmu_write_policy(os, DB_DNODE(db), db->db_level, WP_DMU_SYNC, &zp);
2187 	DB_DNODE_EXIT(db);
2188 
2189 	/*
2190 	 * If we're frozen (running ziltest), we always need to generate a bp.
2191 	 */
2192 	if (txg > spa_freeze_txg(os->os_spa))
2193 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2194 
2195 	/*
2196 	 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
2197 	 * and us.  If we determine that this txg is not yet syncing,
2198 	 * but it begins to sync a moment later, that's OK because the
2199 	 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
2200 	 */
2201 	mutex_enter(&db->db_mtx);
2202 
2203 	if (txg <= spa_last_synced_txg(os->os_spa)) {
2204 		/*
2205 		 * This txg has already synced.  There's nothing to do.
2206 		 */
2207 		mutex_exit(&db->db_mtx);
2208 		return (SET_ERROR(EEXIST));
2209 	}
2210 
2211 	if (txg <= spa_syncing_txg(os->os_spa)) {
2212 		/*
2213 		 * This txg is currently syncing, so we can't mess with
2214 		 * the dirty record anymore; just write a new log block.
2215 		 */
2216 		mutex_exit(&db->db_mtx);
2217 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2218 	}
2219 
2220 	dr = dbuf_find_dirty_eq(db, txg);
2221 
2222 	if (dr == NULL) {
2223 		/*
2224 		 * There's no dr for this dbuf, so it must have been freed.
2225 		 * There's no need to log writes to freed blocks, so we're done.
2226 		 */
2227 		mutex_exit(&db->db_mtx);
2228 		return (SET_ERROR(ENOENT));
2229 	}
2230 
2231 	dr_next = list_next(&db->db_dirty_records, dr);
2232 	ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
2233 
2234 	if (db->db_blkptr != NULL) {
2235 		/*
2236 		 * We need to fill in zgd_bp with the current blkptr so that
2237 		 * the nopwrite code can check if we're writing the same
2238 		 * data that's already on disk.  We can only nopwrite if we
2239 		 * are sure that after making the copy, db_blkptr will not
2240 		 * change until our i/o completes.  We ensure this by
2241 		 * holding the db_mtx, and only allowing nopwrite if the
2242 		 * block is not already dirty (see below).  This is verified
2243 		 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2244 		 * not changed.
2245 		 */
2246 		*zgd->zgd_bp = *db->db_blkptr;
2247 	}
2248 
2249 	/*
2250 	 * Assume the on-disk data is X, the current syncing data (in
2251 	 * txg - 1) is Y, and the current in-memory data is Z (currently
2252 	 * in dmu_sync).
2253 	 *
2254 	 * We usually want to perform a nopwrite if X and Z are the
2255 	 * same.  However, if Y is different (i.e. the BP is going to
2256 	 * change before this write takes effect), then a nopwrite will
2257 	 * be incorrect - we would override with X, which could have
2258 	 * been freed when Y was written.
2259 	 *
2260 	 * (Note that this is not a concern when we are nop-writing from
2261 	 * syncing context, because X and Y must be identical, because
2262 	 * all previous txgs have been synced.)
2263 	 *
2264 	 * Therefore, we disable nopwrite if the current BP could change
2265 	 * before this TXG.  There are two ways it could change: by
2266 	 * being dirty (dr_next is non-NULL), or by being freed
2267 	 * (dnode_block_freed()).  This behavior is verified by
2268 	 * zio_done(), which VERIFYs that the override BP is identical
2269 	 * to the on-disk BP.
2270 	 */
2271 	if (dr_next != NULL) {
2272 		zp.zp_nopwrite = B_FALSE;
2273 	} else {
2274 		DB_DNODE_ENTER(db);
2275 		if (dnode_block_freed(DB_DNODE(db), db->db_blkid))
2276 			zp.zp_nopwrite = B_FALSE;
2277 		DB_DNODE_EXIT(db);
2278 	}
2279 
2280 	ASSERT(dr->dr_txg == txg);
2281 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
2282 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2283 		/*
2284 		 * We have already issued a sync write for this buffer,
2285 		 * or this buffer has already been synced.  It could not
2286 		 * have been dirtied since, or we would have cleared the state.
2287 		 */
2288 		mutex_exit(&db->db_mtx);
2289 		return (SET_ERROR(EALREADY));
2290 	}
2291 
2292 	ASSERT0(dr->dt.dl.dr_has_raw_params);
2293 	ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2294 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
2295 	mutex_exit(&db->db_mtx);
2296 
2297 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2298 	dsa->dsa_dr = dr;
2299 	dsa->dsa_done = done;
2300 	dsa->dsa_zgd = zgd;
2301 	dsa->dsa_tx = NULL;
2302 
2303 	zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp,
2304 	    dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db),
2305 	    dbuf_is_l2cacheable(db, NULL), &zp, dmu_sync_ready, NULL,
2306 	    dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL,
2307 	    &zb));
2308 
2309 	return (0);
2310 }
2311 
2312 int
dmu_object_set_nlevels(objset_t * os,uint64_t object,int nlevels,dmu_tx_t * tx)2313 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
2314 {
2315 	dnode_t *dn;
2316 	int err;
2317 
2318 	err = dnode_hold(os, object, FTAG, &dn);
2319 	if (err)
2320 		return (err);
2321 	err = dnode_set_nlevels(dn, nlevels, tx);
2322 	dnode_rele(dn, FTAG);
2323 	return (err);
2324 }
2325 
2326 int
dmu_object_set_blocksize(objset_t * os,uint64_t object,uint64_t size,int ibs,dmu_tx_t * tx)2327 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
2328     dmu_tx_t *tx)
2329 {
2330 	dnode_t *dn;
2331 	int err;
2332 
2333 	err = dnode_hold(os, object, FTAG, &dn);
2334 	if (err)
2335 		return (err);
2336 	err = dnode_set_blksz(dn, size, ibs, tx);
2337 	dnode_rele(dn, FTAG);
2338 	return (err);
2339 }
2340 
2341 int
dmu_object_set_maxblkid(objset_t * os,uint64_t object,uint64_t maxblkid,dmu_tx_t * tx)2342 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
2343     dmu_tx_t *tx)
2344 {
2345 	dnode_t *dn;
2346 	int err;
2347 
2348 	err = dnode_hold(os, object, FTAG, &dn);
2349 	if (err)
2350 		return (err);
2351 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2352 	dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE);
2353 	rw_exit(&dn->dn_struct_rwlock);
2354 	dnode_rele(dn, FTAG);
2355 	return (0);
2356 }
2357 
2358 void
dmu_object_set_checksum(objset_t * os,uint64_t object,uint8_t checksum,dmu_tx_t * tx)2359 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2360     dmu_tx_t *tx)
2361 {
2362 	dnode_t *dn;
2363 
2364 	/*
2365 	 * Send streams include each object's checksum function.  This
2366 	 * check ensures that the receiving system can understand the
2367 	 * checksum function transmitted.
2368 	 */
2369 	ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2370 
2371 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
2372 	ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2373 	dn->dn_checksum = checksum;
2374 	dnode_setdirty(dn, tx);
2375 	dnode_rele(dn, FTAG);
2376 }
2377 
2378 void
dmu_object_set_compress(objset_t * os,uint64_t object,uint8_t compress,dmu_tx_t * tx)2379 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2380     dmu_tx_t *tx)
2381 {
2382 	dnode_t *dn;
2383 
2384 	/*
2385 	 * Send streams include each object's compression function.  This
2386 	 * check ensures that the receiving system can understand the
2387 	 * compression function transmitted.
2388 	 */
2389 	ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2390 
2391 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
2392 	dn->dn_compress = compress;
2393 	dnode_setdirty(dn, tx);
2394 	dnode_rele(dn, FTAG);
2395 }
2396 
2397 /*
2398  * When the "redundant_metadata" property is set to "most", only indirect
2399  * blocks of this level and higher will have an additional ditto block.
2400  */
2401 static const int zfs_redundant_metadata_most_ditto_level = 2;
2402 
2403 void
dmu_write_policy(objset_t * os,dnode_t * dn,int level,int wp,zio_prop_t * zp)2404 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2405 {
2406 	dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2407 	boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2408 	    (wp & WP_SPILL));
2409 	enum zio_checksum checksum = os->os_checksum;
2410 	enum zio_compress compress = os->os_compress;
2411 	uint8_t complevel = os->os_complevel;
2412 	enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2413 	boolean_t dedup = B_FALSE;
2414 	boolean_t nopwrite = B_FALSE;
2415 	boolean_t dedup_verify = os->os_dedup_verify;
2416 	boolean_t encrypt = B_FALSE;
2417 	int copies = os->os_copies;
2418 	int gang_copies = os->os_copies;
2419 
2420 	/*
2421 	 * We maintain different write policies for each of the following
2422 	 * types of data:
2423 	 *	 1. metadata
2424 	 *	 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2425 	 *	 3. all other level 0 blocks
2426 	 */
2427 	if (ismd) {
2428 		/*
2429 		 * XXX -- we should design a compression algorithm
2430 		 * that specializes in arrays of bps.
2431 		 */
2432 		compress = zio_compress_select(os->os_spa,
2433 		    ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2434 
2435 		/*
2436 		 * Metadata always gets checksummed.  If the data
2437 		 * checksum is multi-bit correctable, and it's not a
2438 		 * ZBT-style checksum, then it's suitable for metadata
2439 		 * as well.  Otherwise, the metadata checksum defaults
2440 		 * to fletcher4.
2441 		 */
2442 		if (!(zio_checksum_table[checksum].ci_flags &
2443 		    ZCHECKSUM_FLAG_METADATA) ||
2444 		    (zio_checksum_table[checksum].ci_flags &
2445 		    ZCHECKSUM_FLAG_EMBEDDED))
2446 			checksum = ZIO_CHECKSUM_FLETCHER_4;
2447 
2448 		switch (os->os_redundant_metadata) {
2449 		case ZFS_REDUNDANT_METADATA_ALL:
2450 			copies++;
2451 			gang_copies++;
2452 			break;
2453 		case ZFS_REDUNDANT_METADATA_MOST:
2454 			if (level >= zfs_redundant_metadata_most_ditto_level ||
2455 			    DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))
2456 				copies++;
2457 			if (level + 1 >=
2458 			    zfs_redundant_metadata_most_ditto_level ||
2459 			    DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))
2460 				gang_copies++;
2461 			break;
2462 		case ZFS_REDUNDANT_METADATA_SOME:
2463 			if (DMU_OT_IS_CRITICAL(type, level)) {
2464 				copies++;
2465 				gang_copies++;
2466 			} else if (DMU_OT_IS_METADATA(type)) {
2467 				gang_copies++;
2468 			}
2469 			break;
2470 		case ZFS_REDUNDANT_METADATA_NONE:
2471 			break;
2472 		}
2473 
2474 		if (dmu_ddt_copies > 0) {
2475 			/*
2476 			 * If this tunable is set, and this is a write for a
2477 			 * dedup entry store (zap or log), then we treat it
2478 			 * something like ZFS_REDUNDANT_METADATA_MOST on a
2479 			 * regular dataset: this many copies, and one more for
2480 			 * "higher" indirect blocks. This specific exception is
2481 			 * necessary because dedup objects are stored in the
2482 			 * MOS, which always has the highest possible copies.
2483 			 */
2484 			dmu_object_type_t stype =
2485 			    dn ? dn->dn_storage_type : DMU_OT_NONE;
2486 			if (stype == DMU_OT_NONE)
2487 				stype = type;
2488 			if (stype == DMU_OT_DDT_ZAP) {
2489 				copies = dmu_ddt_copies;
2490 				if (level >=
2491 				    zfs_redundant_metadata_most_ditto_level)
2492 					copies++;
2493 			}
2494 		}
2495 	} else if (wp & WP_NOFILL) {
2496 		ASSERT0(level);
2497 
2498 		/*
2499 		 * If we're writing preallocated blocks, we aren't actually
2500 		 * writing them so don't set any policy properties.  These
2501 		 * blocks are currently only used by an external subsystem
2502 		 * outside of zfs (i.e. dump) and not written by the zio
2503 		 * pipeline.
2504 		 */
2505 		compress = ZIO_COMPRESS_OFF;
2506 		checksum = ZIO_CHECKSUM_OFF;
2507 	} else {
2508 		compress = zio_compress_select(os->os_spa, dn->dn_compress,
2509 		    compress);
2510 		complevel = zio_complevel_select(os->os_spa, compress,
2511 		    complevel, complevel);
2512 
2513 		/*
2514 		 * Storing many references to an all zeros block in the dedup
2515 		 * table would be expensive.  Instead, if dedup is enabled,
2516 		 * store them as holes even if compression is not enabled.
2517 		 */
2518 		if (compress == ZIO_COMPRESS_OFF &&
2519 		    dedup_checksum != ZIO_CHECKSUM_OFF)
2520 			compress = ZIO_COMPRESS_EMPTY;
2521 
2522 		checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2523 		    zio_checksum_select(dn->dn_checksum, checksum) :
2524 		    dedup_checksum;
2525 
2526 		/*
2527 		 * Determine dedup setting.  If we are in dmu_sync(),
2528 		 * we won't actually dedup now because that's all
2529 		 * done in syncing context; but we do want to use the
2530 		 * dedup checksum.  If the checksum is not strong
2531 		 * enough to ensure unique signatures, force
2532 		 * dedup_verify.
2533 		 */
2534 		if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2535 			dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2536 			if (!(zio_checksum_table[checksum].ci_flags &
2537 			    ZCHECKSUM_FLAG_DEDUP))
2538 				dedup_verify = B_TRUE;
2539 		}
2540 
2541 		/*
2542 		 * Enable nopwrite if we have secure enough checksum
2543 		 * algorithm (see comment in zio_nop_write) and
2544 		 * compression is enabled.  We don't enable nopwrite if
2545 		 * dedup is enabled as the two features are mutually
2546 		 * exclusive.
2547 		 */
2548 		nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2549 		    ZCHECKSUM_FLAG_NOPWRITE) &&
2550 		    compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2551 
2552 		if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2553 		    (os->os_redundant_metadata ==
2554 		    ZFS_REDUNDANT_METADATA_MOST &&
2555 		    zfs_redundant_metadata_most_ditto_level <= 1))
2556 			gang_copies++;
2557 	}
2558 
2559 	/*
2560 	 * All objects in an encrypted objset are protected from modification
2561 	 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2562 	 * in the bp, so we cannot use all copies. Encrypted objects are also
2563 	 * not subject to nopwrite since writing the same data will still
2564 	 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2565 	 * to avoid ambiguity in the dedup code since the DDT does not store
2566 	 * object types.
2567 	 */
2568 	if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2569 		encrypt = B_TRUE;
2570 
2571 		if (DMU_OT_IS_ENCRYPTED(type)) {
2572 			copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2573 			gang_copies = MIN(gang_copies, SPA_DVAS_PER_BP - 1);
2574 			nopwrite = B_FALSE;
2575 		} else {
2576 			dedup = B_FALSE;
2577 		}
2578 
2579 		if (level <= 0 &&
2580 		    (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2581 			compress = ZIO_COMPRESS_EMPTY;
2582 		}
2583 	}
2584 
2585 	zp->zp_compress = compress;
2586 	zp->zp_complevel = complevel;
2587 	zp->zp_checksum = checksum;
2588 	zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2589 	zp->zp_level = level;
2590 	zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2591 	zp->zp_gang_copies = MIN(MAX(gang_copies, copies),
2592 	    spa_max_replication(os->os_spa));
2593 	zp->zp_dedup = dedup;
2594 	zp->zp_dedup_verify = dedup && dedup_verify;
2595 	zp->zp_nopwrite = nopwrite;
2596 	zp->zp_encrypt = encrypt;
2597 	zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2598 	zp->zp_direct_write = (wp & WP_DIRECT_WR) ? B_TRUE : B_FALSE;
2599 	zp->zp_rewrite = B_FALSE;
2600 	memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN);
2601 	memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN);
2602 	memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN);
2603 	zp->zp_zpl_smallblk = os->os_zpl_special_smallblock;
2604 	zp->zp_storage_type = dn ? dn->dn_storage_type : DMU_OT_NONE;
2605 
2606 	ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2607 }
2608 
2609 /*
2610  * Reports the location of data and holes in an object.  In order to
2611  * accurately report holes all dirty data must be synced to disk.  This
2612  * causes extremely poor performance when seeking for holes in a dirty file.
2613  * As a compromise, only provide hole data when the dnode is clean.  When
2614  * a dnode is dirty report the dnode as having no holes by returning EBUSY
2615  * which is always safe to do.
2616  */
2617 int
dmu_offset_next(objset_t * os,uint64_t object,boolean_t hole,uint64_t * off)2618 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2619 {
2620 	dnode_t *dn;
2621 	uint64_t txg, maxtxg = 0;
2622 	int err;
2623 
2624 restart:
2625 	err = dnode_hold(os, object, FTAG, &dn);
2626 	if (err)
2627 		return (err);
2628 
2629 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2630 
2631 	if (dnode_is_dirty(dn)) {
2632 		/*
2633 		 * If the zfs_dmu_offset_next_sync module option is enabled
2634 		 * then hole reporting has been requested.  Dirty dnodes
2635 		 * must be synced to disk to accurately report holes.
2636 		 *
2637 		 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is
2638 		 * held by the caller only limited restarts will be required.
2639 		 * We tolerate callers which do not hold the rangelock by
2640 		 * returning EBUSY and not reporting holes after at most
2641 		 * TXG_CONCURRENT_STATES (3) restarts.
2642 		 */
2643 		if (zfs_dmu_offset_next_sync) {
2644 			rw_exit(&dn->dn_struct_rwlock);
2645 			dnode_rele(dn, FTAG);
2646 
2647 			if (maxtxg == 0) {
2648 				txg = spa_last_synced_txg(dmu_objset_spa(os));
2649 				maxtxg = txg + TXG_CONCURRENT_STATES;
2650 			} else if (txg >= maxtxg)
2651 				return (SET_ERROR(EBUSY));
2652 
2653 			txg_wait_synced(dmu_objset_pool(os), ++txg);
2654 			goto restart;
2655 		}
2656 
2657 		err = SET_ERROR(EBUSY);
2658 	} else {
2659 		err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK |
2660 		    (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2661 	}
2662 
2663 	rw_exit(&dn->dn_struct_rwlock);
2664 	dnode_rele(dn, FTAG);
2665 
2666 	return (err);
2667 }
2668 
2669 int
dmu_read_l0_bps(objset_t * os,uint64_t object,uint64_t offset,uint64_t length,blkptr_t * bps,size_t * nbpsp)2670 dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2671     blkptr_t *bps, size_t *nbpsp)
2672 {
2673 	dmu_buf_t **dbp, *dbuf;
2674 	dmu_buf_impl_t *db;
2675 	blkptr_t *bp;
2676 	int error, numbufs;
2677 
2678 	error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2679 	    &numbufs, &dbp, DMU_READ_PREFETCH);
2680 	if (error != 0) {
2681 		if (error == ESRCH) {
2682 			error = SET_ERROR(ENXIO);
2683 		}
2684 		return (error);
2685 	}
2686 
2687 	ASSERT3U(numbufs, <=, *nbpsp);
2688 
2689 	for (int i = 0; i < numbufs; i++) {
2690 		dbuf = dbp[i];
2691 		db = (dmu_buf_impl_t *)dbuf;
2692 
2693 		mutex_enter(&db->db_mtx);
2694 
2695 		if (!list_is_empty(&db->db_dirty_records)) {
2696 			dbuf_dirty_record_t *dr;
2697 
2698 			dr = list_head(&db->db_dirty_records);
2699 			if (dr->dt.dl.dr_brtwrite) {
2700 				/*
2701 				 * This is very special case where we clone a
2702 				 * block and in the same transaction group we
2703 				 * read its BP (most likely to clone the clone).
2704 				 */
2705 				bp = &dr->dt.dl.dr_overridden_by;
2706 			} else {
2707 				/*
2708 				 * The block was modified in the same
2709 				 * transaction group.
2710 				 */
2711 				mutex_exit(&db->db_mtx);
2712 				error = SET_ERROR(EAGAIN);
2713 				goto out;
2714 			}
2715 		} else {
2716 			bp = db->db_blkptr;
2717 		}
2718 
2719 		mutex_exit(&db->db_mtx);
2720 
2721 		if (bp == NULL) {
2722 			/*
2723 			 * The file size was increased, but the block was never
2724 			 * written, otherwise we would either have the block
2725 			 * pointer or the dirty record and would not get here.
2726 			 * It is effectively a hole, so report it as such.
2727 			 */
2728 			BP_ZERO(&bps[i]);
2729 			continue;
2730 		}
2731 		/*
2732 		 * Make sure we clone only data blocks.
2733 		 */
2734 		if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) {
2735 			error = SET_ERROR(EINVAL);
2736 			goto out;
2737 		}
2738 
2739 		/*
2740 		 * If the block was allocated in transaction group that is not
2741 		 * yet synced, we could clone it, but we couldn't write this
2742 		 * operation into ZIL, or it may be impossible to replay, since
2743 		 * the block may appear not yet allocated at that point.
2744 		 */
2745 		if (BP_GET_PHYSICAL_BIRTH(bp) > spa_freeze_txg(os->os_spa)) {
2746 			error = SET_ERROR(EINVAL);
2747 			goto out;
2748 		}
2749 		if (BP_GET_PHYSICAL_BIRTH(bp) >
2750 		    spa_last_synced_txg(os->os_spa)) {
2751 			error = SET_ERROR(EAGAIN);
2752 			goto out;
2753 		}
2754 
2755 		bps[i] = *bp;
2756 	}
2757 
2758 	*nbpsp = numbufs;
2759 out:
2760 	dmu_buf_rele_array(dbp, numbufs, FTAG);
2761 
2762 	return (error);
2763 }
2764 
2765 int
dmu_brt_clone(objset_t * os,uint64_t object,uint64_t offset,uint64_t length,dmu_tx_t * tx,const blkptr_t * bps,size_t nbps)2766 dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2767     dmu_tx_t *tx, const blkptr_t *bps, size_t nbps)
2768 {
2769 	spa_t *spa;
2770 	dmu_buf_t **dbp, *dbuf;
2771 	dmu_buf_impl_t *db;
2772 	struct dirty_leaf *dl;
2773 	dbuf_dirty_record_t *dr;
2774 	const blkptr_t *bp;
2775 	int error = 0, i, numbufs;
2776 
2777 	spa = os->os_spa;
2778 
2779 	VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2780 	    &numbufs, &dbp, DMU_READ_PREFETCH));
2781 	ASSERT3U(nbps, ==, numbufs);
2782 
2783 	/*
2784 	 * Before we start cloning make sure that the dbufs sizes match new BPs
2785 	 * sizes. If they don't, that's a no-go, as we are not able to shrink
2786 	 * dbufs.
2787 	 */
2788 	for (i = 0; i < numbufs; i++) {
2789 		dbuf = dbp[i];
2790 		db = (dmu_buf_impl_t *)dbuf;
2791 		bp = &bps[i];
2792 
2793 		ASSERT3U(db->db.db_object, !=, DMU_META_DNODE_OBJECT);
2794 		ASSERT0(db->db_level);
2795 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2796 		ASSERT(db->db_blkid != DMU_SPILL_BLKID);
2797 
2798 		if (!BP_IS_HOLE(bp) && BP_GET_LSIZE(bp) != dbuf->db_size) {
2799 			error = SET_ERROR(EXDEV);
2800 			goto out;
2801 		}
2802 	}
2803 
2804 	for (i = 0; i < numbufs; i++) {
2805 		dbuf = dbp[i];
2806 		db = (dmu_buf_impl_t *)dbuf;
2807 		bp = &bps[i];
2808 
2809 		dmu_buf_will_clone_or_dio(dbuf, tx);
2810 
2811 		mutex_enter(&db->db_mtx);
2812 
2813 		dr = list_head(&db->db_dirty_records);
2814 		VERIFY(dr != NULL);
2815 		ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2816 		dl = &dr->dt.dl;
2817 		ASSERT0(dl->dr_has_raw_params);
2818 		dl->dr_overridden_by = *bp;
2819 		if (!BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) != 0) {
2820 			if (!BP_IS_EMBEDDED(bp)) {
2821 				BP_SET_BIRTH(&dl->dr_overridden_by, dr->dr_txg,
2822 				    BP_GET_PHYSICAL_BIRTH(bp));
2823 				BP_SET_REWRITE(&dl->dr_overridden_by, 0);
2824 			} else {
2825 				BP_SET_LOGICAL_BIRTH(&dl->dr_overridden_by,
2826 				    dr->dr_txg);
2827 			}
2828 		}
2829 		dl->dr_brtwrite = B_TRUE;
2830 		dl->dr_override_state = DR_OVERRIDDEN;
2831 
2832 		mutex_exit(&db->db_mtx);
2833 
2834 		/*
2835 		 * When data in embedded into BP there is no need to create
2836 		 * BRT entry as there is no data block. Just copy the BP as
2837 		 * it contains the data.
2838 		 */
2839 		if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
2840 			brt_pending_add(spa, bp, tx);
2841 		}
2842 	}
2843 out:
2844 	dmu_buf_rele_array(dbp, numbufs, FTAG);
2845 
2846 	return (error);
2847 }
2848 
2849 void
__dmu_object_info_from_dnode(dnode_t * dn,dmu_object_info_t * doi)2850 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2851 {
2852 	dnode_phys_t *dnp = dn->dn_phys;
2853 
2854 	doi->doi_data_block_size = dn->dn_datablksz;
2855 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
2856 	    1ULL << dn->dn_indblkshift : 0;
2857 	doi->doi_type = dn->dn_type;
2858 	doi->doi_bonus_type = dn->dn_bonustype;
2859 	doi->doi_bonus_size = dn->dn_bonuslen;
2860 	doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2861 	doi->doi_indirection = dn->dn_nlevels;
2862 	doi->doi_checksum = dn->dn_checksum;
2863 	doi->doi_compress = dn->dn_compress;
2864 	doi->doi_nblkptr = dn->dn_nblkptr;
2865 	doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2866 	doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2867 	doi->doi_fill_count = 0;
2868 	for (int i = 0; i < dnp->dn_nblkptr; i++)
2869 		doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2870 }
2871 
2872 void
dmu_object_info_from_dnode(dnode_t * dn,dmu_object_info_t * doi)2873 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2874 {
2875 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2876 	mutex_enter(&dn->dn_mtx);
2877 
2878 	__dmu_object_info_from_dnode(dn, doi);
2879 
2880 	mutex_exit(&dn->dn_mtx);
2881 	rw_exit(&dn->dn_struct_rwlock);
2882 }
2883 
2884 /*
2885  * Get information on a DMU object.
2886  * If doi is NULL, just indicates whether the object exists.
2887  */
2888 int
dmu_object_info(objset_t * os,uint64_t object,dmu_object_info_t * doi)2889 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2890 {
2891 	dnode_t *dn;
2892 	int err = dnode_hold(os, object, FTAG, &dn);
2893 
2894 	if (err)
2895 		return (err);
2896 
2897 	if (doi != NULL)
2898 		dmu_object_info_from_dnode(dn, doi);
2899 
2900 	dnode_rele(dn, FTAG);
2901 	return (0);
2902 }
2903 
2904 /*
2905  * As above, but faster; can be used when you have a held dbuf in hand.
2906  */
2907 void
dmu_object_info_from_db(dmu_buf_t * db_fake,dmu_object_info_t * doi)2908 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2909 {
2910 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2911 
2912 	DB_DNODE_ENTER(db);
2913 	dmu_object_info_from_dnode(DB_DNODE(db), doi);
2914 	DB_DNODE_EXIT(db);
2915 }
2916 
2917 /*
2918  * Faster still when you only care about the size.
2919  */
2920 void
dmu_object_size_from_db(dmu_buf_t * db_fake,uint32_t * blksize,u_longlong_t * nblk512)2921 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2922     u_longlong_t *nblk512)
2923 {
2924 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2925 	dnode_t *dn;
2926 
2927 	DB_DNODE_ENTER(db);
2928 	dn = DB_DNODE(db);
2929 
2930 	*blksize = dn->dn_datablksz;
2931 	/* add in number of slots used for the dnode itself */
2932 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2933 	    SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2934 	DB_DNODE_EXIT(db);
2935 }
2936 
2937 void
dmu_object_dnsize_from_db(dmu_buf_t * db_fake,int * dnsize)2938 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2939 {
2940 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2941 
2942 	DB_DNODE_ENTER(db);
2943 	*dnsize = DB_DNODE(db)->dn_num_slots << DNODE_SHIFT;
2944 	DB_DNODE_EXIT(db);
2945 }
2946 
2947 void
byteswap_uint64_array(void * vbuf,size_t size)2948 byteswap_uint64_array(void *vbuf, size_t size)
2949 {
2950 	uint64_t *buf = vbuf;
2951 	size_t count = size >> 3;
2952 	int i;
2953 
2954 	ASSERT0((size & 7));
2955 
2956 	for (i = 0; i < count; i++)
2957 		buf[i] = BSWAP_64(buf[i]);
2958 }
2959 
2960 void
byteswap_uint32_array(void * vbuf,size_t size)2961 byteswap_uint32_array(void *vbuf, size_t size)
2962 {
2963 	uint32_t *buf = vbuf;
2964 	size_t count = size >> 2;
2965 	int i;
2966 
2967 	ASSERT0((size & 3));
2968 
2969 	for (i = 0; i < count; i++)
2970 		buf[i] = BSWAP_32(buf[i]);
2971 }
2972 
2973 void
byteswap_uint16_array(void * vbuf,size_t size)2974 byteswap_uint16_array(void *vbuf, size_t size)
2975 {
2976 	uint16_t *buf = vbuf;
2977 	size_t count = size >> 1;
2978 	int i;
2979 
2980 	ASSERT0((size & 1));
2981 
2982 	for (i = 0; i < count; i++)
2983 		buf[i] = BSWAP_16(buf[i]);
2984 }
2985 
2986 void
byteswap_uint8_array(void * vbuf,size_t size)2987 byteswap_uint8_array(void *vbuf, size_t size)
2988 {
2989 	(void) vbuf, (void) size;
2990 }
2991 
2992 void
dmu_init(void)2993 dmu_init(void)
2994 {
2995 	abd_init();
2996 	zfs_dbgmsg_init();
2997 	sa_cache_init();
2998 	dmu_objset_init();
2999 	dnode_init();
3000 	zfetch_init();
3001 	dmu_tx_init();
3002 	l2arc_init();
3003 	arc_init();
3004 	dbuf_init();
3005 }
3006 
3007 void
dmu_fini(void)3008 dmu_fini(void)
3009 {
3010 	arc_fini(); /* arc depends on l2arc, so arc must go first */
3011 	l2arc_fini();
3012 	dmu_tx_fini();
3013 	zfetch_fini();
3014 	dbuf_fini();
3015 	dnode_fini();
3016 	dmu_objset_fini();
3017 	sa_cache_fini();
3018 	zfs_dbgmsg_fini();
3019 	abd_fini();
3020 }
3021 
3022 EXPORT_SYMBOL(dmu_bonus_hold);
3023 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
3024 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
3025 EXPORT_SYMBOL(dmu_buf_rele_array);
3026 EXPORT_SYMBOL(dmu_prefetch);
3027 EXPORT_SYMBOL(dmu_prefetch_by_dnode);
3028 EXPORT_SYMBOL(dmu_prefetch_dnode);
3029 EXPORT_SYMBOL(dmu_prefetch_stream);
3030 EXPORT_SYMBOL(dmu_prefetch_stream_by_dnode);
3031 EXPORT_SYMBOL(dmu_free_range);
3032 EXPORT_SYMBOL(dmu_free_long_range);
3033 EXPORT_SYMBOL(dmu_free_long_object);
3034 EXPORT_SYMBOL(dmu_read);
3035 EXPORT_SYMBOL(dmu_read_by_dnode);
3036 EXPORT_SYMBOL(dmu_read_uio);
3037 EXPORT_SYMBOL(dmu_read_uio_dbuf);
3038 EXPORT_SYMBOL(dmu_read_uio_dnode);
3039 EXPORT_SYMBOL(dmu_write);
3040 EXPORT_SYMBOL(dmu_write_by_dnode);
3041 EXPORT_SYMBOL(dmu_write_uio);
3042 EXPORT_SYMBOL(dmu_write_uio_dbuf);
3043 EXPORT_SYMBOL(dmu_write_uio_dnode);
3044 EXPORT_SYMBOL(dmu_prealloc);
3045 EXPORT_SYMBOL(dmu_object_info);
3046 EXPORT_SYMBOL(dmu_object_info_from_dnode);
3047 EXPORT_SYMBOL(dmu_object_info_from_db);
3048 EXPORT_SYMBOL(dmu_object_size_from_db);
3049 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
3050 EXPORT_SYMBOL(dmu_object_set_nlevels);
3051 EXPORT_SYMBOL(dmu_object_set_blocksize);
3052 EXPORT_SYMBOL(dmu_object_set_maxblkid);
3053 EXPORT_SYMBOL(dmu_object_set_checksum);
3054 EXPORT_SYMBOL(dmu_object_set_compress);
3055 EXPORT_SYMBOL(dmu_offset_next);
3056 EXPORT_SYMBOL(dmu_write_policy);
3057 EXPORT_SYMBOL(dmu_sync);
3058 EXPORT_SYMBOL(dmu_request_arcbuf);
3059 EXPORT_SYMBOL(dmu_return_arcbuf);
3060 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
3061 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
3062 EXPORT_SYMBOL(dmu_buf_hold);
3063 EXPORT_SYMBOL(dmu_ot);
3064 
3065 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
3066 	"Enable NOP writes");
3067 
3068 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW,
3069 	"Percentage of dirtied blocks from frees in one TXG");
3070 
3071 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
3072 	"Enable forcing txg sync to find holes");
3073 
3074 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW,
3075 	"Limit one prefetch call to this size");
3076 
3077 ZFS_MODULE_PARAM(zfs, , dmu_ddt_copies, UINT, ZMOD_RW,
3078 	"Override copies= for dedup objects");
3079