xref: /titanic_44/usr/src/uts/common/fs/zfs/dmu.c (revision 98d1cbfec254295273b6a761bc1861c0374bdf02)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 #include <sys/dmu.h>
26 #include <sys/dmu_impl.h>
27 #include <sys/dmu_tx.h>
28 #include <sys/dbuf.h>
29 #include <sys/dnode.h>
30 #include <sys/zfs_context.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dmu_traverse.h>
33 #include <sys/dsl_dataset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_pool.h>
36 #include <sys/dsl_synctask.h>
37 #include <sys/dsl_prop.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/zfs_ioctl.h>
40 #include <sys/zap.h>
41 #include <sys/zio_checksum.h>
42 #include <sys/sa.h>
43 #ifdef _KERNEL
44 #include <sys/vmsystm.h>
45 #include <sys/zfs_znode.h>
46 #endif
47 
48 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
49 	{	byteswap_uint8_array,	TRUE,	"unallocated"		},
50 	{	zap_byteswap,		TRUE,	"object directory"	},
51 	{	byteswap_uint64_array,	TRUE,	"object array"		},
52 	{	byteswap_uint8_array,	TRUE,	"packed nvlist"		},
53 	{	byteswap_uint64_array,	TRUE,	"packed nvlist size"	},
54 	{	byteswap_uint64_array,	TRUE,	"bplist"		},
55 	{	byteswap_uint64_array,	TRUE,	"bplist header"		},
56 	{	byteswap_uint64_array,	TRUE,	"SPA space map header"	},
57 	{	byteswap_uint64_array,	TRUE,	"SPA space map"		},
58 	{	byteswap_uint64_array,	TRUE,	"ZIL intent log"	},
59 	{	dnode_buf_byteswap,	TRUE,	"DMU dnode"		},
60 	{	dmu_objset_byteswap,	TRUE,	"DMU objset"		},
61 	{	byteswap_uint64_array,	TRUE,	"DSL directory"		},
62 	{	zap_byteswap,		TRUE,	"DSL directory child map"},
63 	{	zap_byteswap,		TRUE,	"DSL dataset snap map"	},
64 	{	zap_byteswap,		TRUE,	"DSL props"		},
65 	{	byteswap_uint64_array,	TRUE,	"DSL dataset"		},
66 	{	zfs_znode_byteswap,	TRUE,	"ZFS znode"		},
67 	{	zfs_oldacl_byteswap,	TRUE,	"ZFS V0 ACL"		},
68 	{	byteswap_uint8_array,	FALSE,	"ZFS plain file"	},
69 	{	zap_byteswap,		TRUE,	"ZFS directory"		},
70 	{	zap_byteswap,		TRUE,	"ZFS master node"	},
71 	{	zap_byteswap,		TRUE,	"ZFS delete queue"	},
72 	{	byteswap_uint8_array,	FALSE,	"zvol object"		},
73 	{	zap_byteswap,		TRUE,	"zvol prop"		},
74 	{	byteswap_uint8_array,	FALSE,	"other uint8[]"		},
75 	{	byteswap_uint64_array,	FALSE,	"other uint64[]"	},
76 	{	zap_byteswap,		TRUE,	"other ZAP"		},
77 	{	zap_byteswap,		TRUE,	"persistent error log"	},
78 	{	byteswap_uint8_array,	TRUE,	"SPA history"		},
79 	{	byteswap_uint64_array,	TRUE,	"SPA history offsets"	},
80 	{	zap_byteswap,		TRUE,	"Pool properties"	},
81 	{	zap_byteswap,		TRUE,	"DSL permissions"	},
82 	{	zfs_acl_byteswap,	TRUE,	"ZFS ACL"		},
83 	{	byteswap_uint8_array,	TRUE,	"ZFS SYSACL"		},
84 	{	byteswap_uint8_array,	TRUE,	"FUID table"		},
85 	{	byteswap_uint64_array,	TRUE,	"FUID table size"	},
86 	{	zap_byteswap,		TRUE,	"DSL dataset next clones"},
87 	{	zap_byteswap,		TRUE,	"scrub work queue"	},
88 	{	zap_byteswap,		TRUE,	"ZFS user/group used"	},
89 	{	zap_byteswap,		TRUE,	"ZFS user/group quota"	},
90 	{	zap_byteswap,		TRUE,	"snapshot refcount tags"},
91 	{	zap_byteswap,		TRUE,	"DDT ZAP algorithm"	},
92 	{	zap_byteswap,		TRUE,	"DDT statistics"	},
93 	{	byteswap_uint8_array,	TRUE,	"System attributes"	},
94 	{	zap_byteswap,		TRUE,	"SA master node"	},
95 	{	zap_byteswap,		TRUE,	"SA attr registration"	},
96 	{	zap_byteswap,		TRUE,	"SA attr layouts"	}, };
97 
98 int
99 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
100     void *tag, dmu_buf_t **dbp)
101 {
102 	dnode_t *dn;
103 	uint64_t blkid;
104 	dmu_buf_impl_t *db;
105 	int err;
106 
107 	err = dnode_hold(os, object, FTAG, &dn);
108 	if (err)
109 		return (err);
110 	blkid = dbuf_whichblock(dn, offset);
111 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
112 	db = dbuf_hold(dn, blkid, tag);
113 	rw_exit(&dn->dn_struct_rwlock);
114 	if (db == NULL) {
115 		err = EIO;
116 	} else {
117 		err = dbuf_read(db, NULL, DB_RF_CANFAIL);
118 		if (err) {
119 			dbuf_rele(db, tag);
120 			db = NULL;
121 		}
122 	}
123 
124 	dnode_rele(dn, FTAG);
125 	*dbp = &db->db;
126 	return (err);
127 }
128 
129 int
130 dmu_bonus_max(void)
131 {
132 	return (DN_MAX_BONUSLEN);
133 }
134 
135 int
136 dmu_set_bonus(dmu_buf_t *db, int newsize, dmu_tx_t *tx)
137 {
138 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
139 
140 	if (dn->dn_bonus != (dmu_buf_impl_t *)db)
141 		return (EINVAL);
142 	if (newsize < 0 || newsize > db->db_size)
143 		return (EINVAL);
144 	dnode_setbonuslen(dn, newsize, tx);
145 	return (0);
146 }
147 
148 int
149 dmu_set_bonustype(dmu_buf_t *db, dmu_object_type_t type, dmu_tx_t *tx)
150 {
151 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
152 
153 	if (type > DMU_OT_NUMTYPES)
154 		return (EINVAL);
155 
156 	if (dn->dn_bonus != (dmu_buf_impl_t *)db)
157 		return (EINVAL);
158 
159 	dnode_setbonus_type(dn, type, tx);
160 	return (0);
161 }
162 
163 int
164 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
165 {
166 	dnode_t *dn;
167 	int error;
168 
169 	error = dnode_hold(os, object, FTAG, &dn);
170 	dbuf_rm_spill(dn, tx);
171 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
172 	dnode_rm_spill(dn, tx);
173 	rw_exit(&dn->dn_struct_rwlock);
174 	dnode_rele(dn, FTAG);
175 	return (error);
176 }
177 
178 /*
179  * returns ENOENT, EIO, or 0.
180  */
181 int
182 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
183 {
184 	dnode_t *dn;
185 	dmu_buf_impl_t *db;
186 	int error;
187 
188 	error = dnode_hold(os, object, FTAG, &dn);
189 	if (error)
190 		return (error);
191 
192 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
193 	if (dn->dn_bonus == NULL) {
194 		rw_exit(&dn->dn_struct_rwlock);
195 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
196 		if (dn->dn_bonus == NULL)
197 			dbuf_create_bonus(dn);
198 	}
199 	db = dn->dn_bonus;
200 	rw_exit(&dn->dn_struct_rwlock);
201 
202 	/* as long as the bonus buf is held, the dnode will be held */
203 	if (refcount_add(&db->db_holds, tag) == 1)
204 		VERIFY(dnode_add_ref(dn, db));
205 
206 	dnode_rele(dn, FTAG);
207 
208 	VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED));
209 
210 	*dbp = &db->db;
211 	return (0);
212 }
213 
214 /*
215  * returns ENOENT, EIO, or 0.
216  *
217  * This interface will allocate a blank spill dbuf when a spill blk
218  * doesn't already exist on the dnode.
219  *
220  * if you only want to find an already existing spill db, then
221  * dmu_spill_hold_existing() should be used.
222  */
223 int
224 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
225 {
226 	dmu_buf_impl_t *db = NULL;
227 	int err;
228 
229 	if ((flags & DB_RF_HAVESTRUCT) == 0)
230 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
231 
232 	db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
233 
234 	if ((flags & DB_RF_HAVESTRUCT) == 0)
235 		rw_exit(&dn->dn_struct_rwlock);
236 
237 	ASSERT(db != NULL);
238 	err = dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | flags);
239 	*dbp = &db->db;
240 	return (err);
241 }
242 
243 int
244 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
245 {
246 	dnode_t *dn = ((dmu_buf_impl_t *)bonus)->db_dnode;
247 	int err;
248 
249 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA)
250 		return (EINVAL);
251 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
252 
253 	if (!dn->dn_have_spill) {
254 		rw_exit(&dn->dn_struct_rwlock);
255 		return (ENOENT);
256 	}
257 	err = dmu_spill_hold_by_dnode(dn, DB_RF_HAVESTRUCT, tag, dbp);
258 	rw_exit(&dn->dn_struct_rwlock);
259 	return (err);
260 }
261 
262 int
263 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
264 {
265 	return (dmu_spill_hold_by_dnode(((dmu_buf_impl_t *)bonus)->db_dnode,
266 	    0, tag, dbp));
267 }
268 
269 /*
270  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
271  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
272  * and can induce severe lock contention when writing to several files
273  * whose dnodes are in the same block.
274  */
275 static int
276 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
277     int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
278 {
279 	dsl_pool_t *dp = NULL;
280 	dmu_buf_t **dbp;
281 	uint64_t blkid, nblks, i;
282 	uint32_t dbuf_flags;
283 	int err;
284 	zio_t *zio;
285 	hrtime_t start;
286 
287 	ASSERT(length <= DMU_MAX_ACCESS);
288 
289 	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
290 	if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
291 		dbuf_flags |= DB_RF_NOPREFETCH;
292 
293 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
294 	if (dn->dn_datablkshift) {
295 		int blkshift = dn->dn_datablkshift;
296 		nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
297 		    P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
298 	} else {
299 		if (offset + length > dn->dn_datablksz) {
300 			zfs_panic_recover("zfs: accessing past end of object "
301 			    "%llx/%llx (size=%u access=%llu+%llu)",
302 			    (longlong_t)dn->dn_objset->
303 			    os_dsl_dataset->ds_object,
304 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
305 			    (longlong_t)offset, (longlong_t)length);
306 			rw_exit(&dn->dn_struct_rwlock);
307 			return (EIO);
308 		}
309 		nblks = 1;
310 	}
311 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
312 
313 	if (dn->dn_objset->os_dsl_dataset)
314 		dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool;
315 	if (dp && dsl_pool_sync_context(dp))
316 		start = gethrtime();
317 	zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
318 	blkid = dbuf_whichblock(dn, offset);
319 	for (i = 0; i < nblks; i++) {
320 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
321 		if (db == NULL) {
322 			rw_exit(&dn->dn_struct_rwlock);
323 			dmu_buf_rele_array(dbp, nblks, tag);
324 			zio_nowait(zio);
325 			return (EIO);
326 		}
327 		/* initiate async i/o */
328 		if (read) {
329 			(void) dbuf_read(db, zio, dbuf_flags);
330 		}
331 		dbp[i] = &db->db;
332 	}
333 	rw_exit(&dn->dn_struct_rwlock);
334 
335 	/* wait for async i/o */
336 	err = zio_wait(zio);
337 	/* track read overhead when we are in sync context */
338 	if (dp && dsl_pool_sync_context(dp))
339 		dp->dp_read_overhead += gethrtime() - start;
340 	if (err) {
341 		dmu_buf_rele_array(dbp, nblks, tag);
342 		return (err);
343 	}
344 
345 	/* wait for other io to complete */
346 	if (read) {
347 		for (i = 0; i < nblks; i++) {
348 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
349 			mutex_enter(&db->db_mtx);
350 			while (db->db_state == DB_READ ||
351 			    db->db_state == DB_FILL)
352 				cv_wait(&db->db_changed, &db->db_mtx);
353 			if (db->db_state == DB_UNCACHED)
354 				err = EIO;
355 			mutex_exit(&db->db_mtx);
356 			if (err) {
357 				dmu_buf_rele_array(dbp, nblks, tag);
358 				return (err);
359 			}
360 		}
361 	}
362 
363 	*numbufsp = nblks;
364 	*dbpp = dbp;
365 	return (0);
366 }
367 
368 static int
369 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
370     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
371 {
372 	dnode_t *dn;
373 	int err;
374 
375 	err = dnode_hold(os, object, FTAG, &dn);
376 	if (err)
377 		return (err);
378 
379 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
380 	    numbufsp, dbpp, DMU_READ_PREFETCH);
381 
382 	dnode_rele(dn, FTAG);
383 
384 	return (err);
385 }
386 
387 int
388 dmu_buf_hold_array_by_bonus(dmu_buf_t *db, uint64_t offset,
389     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
390 {
391 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
392 	int err;
393 
394 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
395 	    numbufsp, dbpp, DMU_READ_PREFETCH);
396 
397 	return (err);
398 }
399 
400 void
401 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
402 {
403 	int i;
404 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
405 
406 	if (numbufs == 0)
407 		return;
408 
409 	for (i = 0; i < numbufs; i++) {
410 		if (dbp[i])
411 			dbuf_rele(dbp[i], tag);
412 	}
413 
414 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
415 }
416 
417 void
418 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
419 {
420 	dnode_t *dn;
421 	uint64_t blkid;
422 	int nblks, i, err;
423 
424 	if (zfs_prefetch_disable)
425 		return;
426 
427 	if (len == 0) {  /* they're interested in the bonus buffer */
428 		dn = os->os_meta_dnode;
429 
430 		if (object == 0 || object >= DN_MAX_OBJECT)
431 			return;
432 
433 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
434 		blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
435 		dbuf_prefetch(dn, blkid);
436 		rw_exit(&dn->dn_struct_rwlock);
437 		return;
438 	}
439 
440 	/*
441 	 * XXX - Note, if the dnode for the requested object is not
442 	 * already cached, we will do a *synchronous* read in the
443 	 * dnode_hold() call.  The same is true for any indirects.
444 	 */
445 	err = dnode_hold(os, object, FTAG, &dn);
446 	if (err != 0)
447 		return;
448 
449 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
450 	if (dn->dn_datablkshift) {
451 		int blkshift = dn->dn_datablkshift;
452 		nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
453 		    P2ALIGN(offset, 1<<blkshift)) >> blkshift;
454 	} else {
455 		nblks = (offset < dn->dn_datablksz);
456 	}
457 
458 	if (nblks != 0) {
459 		blkid = dbuf_whichblock(dn, offset);
460 		for (i = 0; i < nblks; i++)
461 			dbuf_prefetch(dn, blkid+i);
462 	}
463 
464 	rw_exit(&dn->dn_struct_rwlock);
465 
466 	dnode_rele(dn, FTAG);
467 }
468 
469 /*
470  * Get the next "chunk" of file data to free.  We traverse the file from
471  * the end so that the file gets shorter over time (if we crashes in the
472  * middle, this will leave us in a better state).  We find allocated file
473  * data by simply searching the allocated level 1 indirects.
474  */
475 static int
476 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit)
477 {
478 	uint64_t len = *start - limit;
479 	uint64_t blkcnt = 0;
480 	uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1));
481 	uint64_t iblkrange =
482 	    dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
483 
484 	ASSERT(limit <= *start);
485 
486 	if (len <= iblkrange * maxblks) {
487 		*start = limit;
488 		return (0);
489 	}
490 	ASSERT(ISP2(iblkrange));
491 
492 	while (*start > limit && blkcnt < maxblks) {
493 		int err;
494 
495 		/* find next allocated L1 indirect */
496 		err = dnode_next_offset(dn,
497 		    DNODE_FIND_BACKWARDS, start, 2, 1, 0);
498 
499 		/* if there are no more, then we are done */
500 		if (err == ESRCH) {
501 			*start = limit;
502 			return (0);
503 		} else if (err) {
504 			return (err);
505 		}
506 		blkcnt += 1;
507 
508 		/* reset offset to end of "next" block back */
509 		*start = P2ALIGN(*start, iblkrange);
510 		if (*start <= limit)
511 			*start = limit;
512 		else
513 			*start -= 1;
514 	}
515 	return (0);
516 }
517 
518 static int
519 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
520     uint64_t length, boolean_t free_dnode)
521 {
522 	dmu_tx_t *tx;
523 	uint64_t object_size, start, end, len;
524 	boolean_t trunc = (length == DMU_OBJECT_END);
525 	int align, err;
526 
527 	align = 1 << dn->dn_datablkshift;
528 	ASSERT(align > 0);
529 	object_size = align == 1 ? dn->dn_datablksz :
530 	    (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
531 
532 	end = offset + length;
533 	if (trunc || end > object_size)
534 		end = object_size;
535 	if (end <= offset)
536 		return (0);
537 	length = end - offset;
538 
539 	while (length) {
540 		start = end;
541 		/* assert(offset <= start) */
542 		err = get_next_chunk(dn, &start, offset);
543 		if (err)
544 			return (err);
545 		len = trunc ? DMU_OBJECT_END : end - start;
546 
547 		tx = dmu_tx_create(os);
548 		dmu_tx_hold_free(tx, dn->dn_object, start, len);
549 		err = dmu_tx_assign(tx, TXG_WAIT);
550 		if (err) {
551 			dmu_tx_abort(tx);
552 			return (err);
553 		}
554 
555 		dnode_free_range(dn, start, trunc ? -1 : len, tx);
556 
557 		if (start == 0 && free_dnode) {
558 			ASSERT(trunc);
559 			dnode_free(dn, tx);
560 		}
561 
562 		length -= end - start;
563 
564 		dmu_tx_commit(tx);
565 		end = start;
566 	}
567 	return (0);
568 }
569 
570 int
571 dmu_free_long_range(objset_t *os, uint64_t object,
572     uint64_t offset, uint64_t length)
573 {
574 	dnode_t *dn;
575 	int err;
576 
577 	err = dnode_hold(os, object, FTAG, &dn);
578 	if (err != 0)
579 		return (err);
580 	err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
581 	dnode_rele(dn, FTAG);
582 	return (err);
583 }
584 
585 int
586 dmu_free_object(objset_t *os, uint64_t object)
587 {
588 	dnode_t *dn;
589 	dmu_tx_t *tx;
590 	int err;
591 
592 	err = dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED,
593 	    FTAG, &dn);
594 	if (err != 0)
595 		return (err);
596 	if (dn->dn_nlevels == 1) {
597 		tx = dmu_tx_create(os);
598 		dmu_tx_hold_bonus(tx, object);
599 		dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
600 		err = dmu_tx_assign(tx, TXG_WAIT);
601 		if (err == 0) {
602 			dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
603 			dnode_free(dn, tx);
604 			dmu_tx_commit(tx);
605 		} else {
606 			dmu_tx_abort(tx);
607 		}
608 	} else {
609 		err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
610 	}
611 	dnode_rele(dn, FTAG);
612 	return (err);
613 }
614 
615 int
616 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
617     uint64_t size, dmu_tx_t *tx)
618 {
619 	dnode_t *dn;
620 	int err = dnode_hold(os, object, FTAG, &dn);
621 	if (err)
622 		return (err);
623 	ASSERT(offset < UINT64_MAX);
624 	ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
625 	dnode_free_range(dn, offset, size, tx);
626 	dnode_rele(dn, FTAG);
627 	return (0);
628 }
629 
630 int
631 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
632     void *buf, uint32_t flags)
633 {
634 	dnode_t *dn;
635 	dmu_buf_t **dbp;
636 	int numbufs, err;
637 
638 	err = dnode_hold(os, object, FTAG, &dn);
639 	if (err)
640 		return (err);
641 
642 	/*
643 	 * Deal with odd block sizes, where there can't be data past the first
644 	 * block.  If we ever do the tail block optimization, we will need to
645 	 * handle that here as well.
646 	 */
647 	if (dn->dn_maxblkid == 0) {
648 		int newsz = offset > dn->dn_datablksz ? 0 :
649 		    MIN(size, dn->dn_datablksz - offset);
650 		bzero((char *)buf + newsz, size - newsz);
651 		size = newsz;
652 	}
653 
654 	while (size > 0) {
655 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
656 		int i;
657 
658 		/*
659 		 * NB: we could do this block-at-a-time, but it's nice
660 		 * to be reading in parallel.
661 		 */
662 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
663 		    TRUE, FTAG, &numbufs, &dbp, flags);
664 		if (err)
665 			break;
666 
667 		for (i = 0; i < numbufs; i++) {
668 			int tocpy;
669 			int bufoff;
670 			dmu_buf_t *db = dbp[i];
671 
672 			ASSERT(size > 0);
673 
674 			bufoff = offset - db->db_offset;
675 			tocpy = (int)MIN(db->db_size - bufoff, size);
676 
677 			bcopy((char *)db->db_data + bufoff, buf, tocpy);
678 
679 			offset += tocpy;
680 			size -= tocpy;
681 			buf = (char *)buf + tocpy;
682 		}
683 		dmu_buf_rele_array(dbp, numbufs, FTAG);
684 	}
685 	dnode_rele(dn, FTAG);
686 	return (err);
687 }
688 
689 void
690 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
691     const void *buf, dmu_tx_t *tx)
692 {
693 	dmu_buf_t **dbp;
694 	int numbufs, i;
695 
696 	if (size == 0)
697 		return;
698 
699 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
700 	    FALSE, FTAG, &numbufs, &dbp));
701 
702 	for (i = 0; i < numbufs; i++) {
703 		int tocpy;
704 		int bufoff;
705 		dmu_buf_t *db = dbp[i];
706 
707 		ASSERT(size > 0);
708 
709 		bufoff = offset - db->db_offset;
710 		tocpy = (int)MIN(db->db_size - bufoff, size);
711 
712 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
713 
714 		if (tocpy == db->db_size)
715 			dmu_buf_will_fill(db, tx);
716 		else
717 			dmu_buf_will_dirty(db, tx);
718 
719 		bcopy(buf, (char *)db->db_data + bufoff, tocpy);
720 
721 		if (tocpy == db->db_size)
722 			dmu_buf_fill_done(db, tx);
723 
724 		offset += tocpy;
725 		size -= tocpy;
726 		buf = (char *)buf + tocpy;
727 	}
728 	dmu_buf_rele_array(dbp, numbufs, FTAG);
729 }
730 
731 void
732 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
733     dmu_tx_t *tx)
734 {
735 	dmu_buf_t **dbp;
736 	int numbufs, i;
737 
738 	if (size == 0)
739 		return;
740 
741 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
742 	    FALSE, FTAG, &numbufs, &dbp));
743 
744 	for (i = 0; i < numbufs; i++) {
745 		dmu_buf_t *db = dbp[i];
746 
747 		dmu_buf_will_not_fill(db, tx);
748 	}
749 	dmu_buf_rele_array(dbp, numbufs, FTAG);
750 }
751 
752 /*
753  * DMU support for xuio
754  */
755 kstat_t *xuio_ksp = NULL;
756 
757 int
758 dmu_xuio_init(xuio_t *xuio, int nblk)
759 {
760 	dmu_xuio_t *priv;
761 	uio_t *uio = &xuio->xu_uio;
762 
763 	uio->uio_iovcnt = nblk;
764 	uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
765 
766 	priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
767 	priv->cnt = nblk;
768 	priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
769 	priv->iovp = uio->uio_iov;
770 	XUIO_XUZC_PRIV(xuio) = priv;
771 
772 	if (XUIO_XUZC_RW(xuio) == UIO_READ)
773 		XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
774 	else
775 		XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
776 
777 	return (0);
778 }
779 
780 void
781 dmu_xuio_fini(xuio_t *xuio)
782 {
783 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
784 	int nblk = priv->cnt;
785 
786 	kmem_free(priv->iovp, nblk * sizeof (iovec_t));
787 	kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
788 	kmem_free(priv, sizeof (dmu_xuio_t));
789 
790 	if (XUIO_XUZC_RW(xuio) == UIO_READ)
791 		XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
792 	else
793 		XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
794 }
795 
796 /*
797  * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
798  * and increase priv->next by 1.
799  */
800 int
801 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
802 {
803 	struct iovec *iov;
804 	uio_t *uio = &xuio->xu_uio;
805 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
806 	int i = priv->next++;
807 
808 	ASSERT(i < priv->cnt);
809 	ASSERT(off + n <= arc_buf_size(abuf));
810 	iov = uio->uio_iov + i;
811 	iov->iov_base = (char *)abuf->b_data + off;
812 	iov->iov_len = n;
813 	priv->bufs[i] = abuf;
814 	return (0);
815 }
816 
817 int
818 dmu_xuio_cnt(xuio_t *xuio)
819 {
820 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
821 	return (priv->cnt);
822 }
823 
824 arc_buf_t *
825 dmu_xuio_arcbuf(xuio_t *xuio, int i)
826 {
827 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
828 
829 	ASSERT(i < priv->cnt);
830 	return (priv->bufs[i]);
831 }
832 
833 void
834 dmu_xuio_clear(xuio_t *xuio, int i)
835 {
836 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
837 
838 	ASSERT(i < priv->cnt);
839 	priv->bufs[i] = NULL;
840 }
841 
842 static void
843 xuio_stat_init(void)
844 {
845 	xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
846 	    KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
847 	    KSTAT_FLAG_VIRTUAL);
848 	if (xuio_ksp != NULL) {
849 		xuio_ksp->ks_data = &xuio_stats;
850 		kstat_install(xuio_ksp);
851 	}
852 }
853 
854 static void
855 xuio_stat_fini(void)
856 {
857 	if (xuio_ksp != NULL) {
858 		kstat_delete(xuio_ksp);
859 		xuio_ksp = NULL;
860 	}
861 }
862 
863 void
864 xuio_stat_wbuf_copied()
865 {
866 	XUIOSTAT_BUMP(xuiostat_wbuf_copied);
867 }
868 
869 void
870 xuio_stat_wbuf_nocopy()
871 {
872 	XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
873 }
874 
875 #ifdef _KERNEL
876 int
877 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
878 {
879 	dmu_buf_t **dbp;
880 	int numbufs, i, err;
881 	xuio_t *xuio = NULL;
882 
883 	/*
884 	 * NB: we could do this block-at-a-time, but it's nice
885 	 * to be reading in parallel.
886 	 */
887 	err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
888 	    &numbufs, &dbp);
889 	if (err)
890 		return (err);
891 
892 	if (uio->uio_extflg == UIO_XUIO)
893 		xuio = (xuio_t *)uio;
894 
895 	for (i = 0; i < numbufs; i++) {
896 		int tocpy;
897 		int bufoff;
898 		dmu_buf_t *db = dbp[i];
899 
900 		ASSERT(size > 0);
901 
902 		bufoff = uio->uio_loffset - db->db_offset;
903 		tocpy = (int)MIN(db->db_size - bufoff, size);
904 
905 		if (xuio) {
906 			dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
907 			arc_buf_t *dbuf_abuf = dbi->db_buf;
908 			arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
909 			err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
910 			if (!err) {
911 				uio->uio_resid -= tocpy;
912 				uio->uio_loffset += tocpy;
913 			}
914 
915 			if (abuf == dbuf_abuf)
916 				XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
917 			else
918 				XUIOSTAT_BUMP(xuiostat_rbuf_copied);
919 		} else {
920 			err = uiomove((char *)db->db_data + bufoff, tocpy,
921 			    UIO_READ, uio);
922 		}
923 		if (err)
924 			break;
925 
926 		size -= tocpy;
927 	}
928 	dmu_buf_rele_array(dbp, numbufs, FTAG);
929 
930 	return (err);
931 }
932 
933 static int
934 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
935 {
936 	dmu_buf_t **dbp;
937 	int numbufs;
938 	int err = 0;
939 	int i;
940 
941 	err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
942 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
943 	if (err)
944 		return (err);
945 
946 	for (i = 0; i < numbufs; i++) {
947 		int tocpy;
948 		int bufoff;
949 		dmu_buf_t *db = dbp[i];
950 
951 		ASSERT(size > 0);
952 
953 		bufoff = uio->uio_loffset - db->db_offset;
954 		tocpy = (int)MIN(db->db_size - bufoff, size);
955 
956 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
957 
958 		if (tocpy == db->db_size)
959 			dmu_buf_will_fill(db, tx);
960 		else
961 			dmu_buf_will_dirty(db, tx);
962 
963 		/*
964 		 * XXX uiomove could block forever (eg. nfs-backed
965 		 * pages).  There needs to be a uiolockdown() function
966 		 * to lock the pages in memory, so that uiomove won't
967 		 * block.
968 		 */
969 		err = uiomove((char *)db->db_data + bufoff, tocpy,
970 		    UIO_WRITE, uio);
971 
972 		if (tocpy == db->db_size)
973 			dmu_buf_fill_done(db, tx);
974 
975 		if (err)
976 			break;
977 
978 		size -= tocpy;
979 	}
980 
981 	dmu_buf_rele_array(dbp, numbufs, FTAG);
982 	return (err);
983 }
984 
985 int
986 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
987     dmu_tx_t *tx)
988 {
989 	if (size == 0)
990 		return (0);
991 
992 	return (dmu_write_uio_dnode(((dmu_buf_impl_t *)zdb)->db_dnode,
993 	    uio, size, tx));
994 }
995 
996 int
997 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
998     dmu_tx_t *tx)
999 {
1000 	dnode_t *dn;
1001 	int err;
1002 
1003 	if (size == 0)
1004 		return (0);
1005 
1006 	err = dnode_hold(os, object, FTAG, &dn);
1007 	if (err)
1008 		return (err);
1009 
1010 	err = dmu_write_uio_dnode(dn, uio, size, tx);
1011 
1012 	dnode_rele(dn, FTAG);
1013 
1014 	return (err);
1015 }
1016 
1017 int
1018 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1019     page_t *pp, dmu_tx_t *tx)
1020 {
1021 	dmu_buf_t **dbp;
1022 	int numbufs, i;
1023 	int err;
1024 
1025 	if (size == 0)
1026 		return (0);
1027 
1028 	err = dmu_buf_hold_array(os, object, offset, size,
1029 	    FALSE, FTAG, &numbufs, &dbp);
1030 	if (err)
1031 		return (err);
1032 
1033 	for (i = 0; i < numbufs; i++) {
1034 		int tocpy, copied, thiscpy;
1035 		int bufoff;
1036 		dmu_buf_t *db = dbp[i];
1037 		caddr_t va;
1038 
1039 		ASSERT(size > 0);
1040 		ASSERT3U(db->db_size, >=, PAGESIZE);
1041 
1042 		bufoff = offset - db->db_offset;
1043 		tocpy = (int)MIN(db->db_size - bufoff, size);
1044 
1045 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1046 
1047 		if (tocpy == db->db_size)
1048 			dmu_buf_will_fill(db, tx);
1049 		else
1050 			dmu_buf_will_dirty(db, tx);
1051 
1052 		for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1053 			ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1054 			thiscpy = MIN(PAGESIZE, tocpy - copied);
1055 			va = zfs_map_page(pp, S_READ);
1056 			bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1057 			zfs_unmap_page(pp, va);
1058 			pp = pp->p_next;
1059 			bufoff += PAGESIZE;
1060 		}
1061 
1062 		if (tocpy == db->db_size)
1063 			dmu_buf_fill_done(db, tx);
1064 
1065 		offset += tocpy;
1066 		size -= tocpy;
1067 	}
1068 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1069 	return (err);
1070 }
1071 #endif
1072 
1073 /*
1074  * Allocate a loaned anonymous arc buffer.
1075  */
1076 arc_buf_t *
1077 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1078 {
1079 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
1080 
1081 	return (arc_loan_buf(dn->dn_objset->os_spa, size));
1082 }
1083 
1084 /*
1085  * Free a loaned arc buffer.
1086  */
1087 void
1088 dmu_return_arcbuf(arc_buf_t *buf)
1089 {
1090 	arc_return_buf(buf, FTAG);
1091 	VERIFY(arc_buf_remove_ref(buf, FTAG) == 1);
1092 }
1093 
1094 /*
1095  * When possible directly assign passed loaned arc buffer to a dbuf.
1096  * If this is not possible copy the contents of passed arc buf via
1097  * dmu_write().
1098  */
1099 void
1100 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1101     dmu_tx_t *tx)
1102 {
1103 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
1104 	dmu_buf_impl_t *db;
1105 	uint32_t blksz = (uint32_t)arc_buf_size(buf);
1106 	uint64_t blkid;
1107 
1108 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1109 	blkid = dbuf_whichblock(dn, offset);
1110 	VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1111 	rw_exit(&dn->dn_struct_rwlock);
1112 
1113 	if (offset == db->db.db_offset && blksz == db->db.db_size) {
1114 		dbuf_assign_arcbuf(db, buf, tx);
1115 		dbuf_rele(db, FTAG);
1116 	} else {
1117 		dbuf_rele(db, FTAG);
1118 		dmu_write(dn->dn_objset, dn->dn_object, offset, blksz,
1119 		    buf->b_data, tx);
1120 		dmu_return_arcbuf(buf);
1121 		XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1122 	}
1123 }
1124 
1125 typedef struct {
1126 	dbuf_dirty_record_t	*dsa_dr;
1127 	dmu_sync_cb_t		*dsa_done;
1128 	zgd_t			*dsa_zgd;
1129 	dmu_tx_t		*dsa_tx;
1130 } dmu_sync_arg_t;
1131 
1132 /* ARGSUSED */
1133 static void
1134 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1135 {
1136 	dmu_sync_arg_t *dsa = varg;
1137 	dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1138 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1139 	blkptr_t *bp = zio->io_bp;
1140 
1141 	if (zio->io_error == 0) {
1142 		if (BP_IS_HOLE(bp)) {
1143 			/*
1144 			 * A block of zeros may compress to a hole, but the
1145 			 * block size still needs to be known for replay.
1146 			 */
1147 			BP_SET_LSIZE(bp, db->db_size);
1148 		} else {
1149 			ASSERT(BP_GET_TYPE(bp) == dn->dn_type);
1150 			ASSERT(BP_GET_LEVEL(bp) == 0);
1151 			bp->blk_fill = 1;
1152 		}
1153 	}
1154 }
1155 
1156 static void
1157 dmu_sync_late_arrival_ready(zio_t *zio)
1158 {
1159 	dmu_sync_ready(zio, NULL, zio->io_private);
1160 }
1161 
1162 /* ARGSUSED */
1163 static void
1164 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1165 {
1166 	dmu_sync_arg_t *dsa = varg;
1167 	dbuf_dirty_record_t *dr = dsa->dsa_dr;
1168 	dmu_buf_impl_t *db = dr->dr_dbuf;
1169 
1170 	mutex_enter(&db->db_mtx);
1171 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1172 	if (zio->io_error == 0) {
1173 		dr->dt.dl.dr_overridden_by = *zio->io_bp;
1174 		dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1175 		dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1176 		if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1177 			BP_ZERO(&dr->dt.dl.dr_overridden_by);
1178 	} else {
1179 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1180 	}
1181 	cv_broadcast(&db->db_changed);
1182 	mutex_exit(&db->db_mtx);
1183 
1184 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1185 
1186 	kmem_free(dsa, sizeof (*dsa));
1187 }
1188 
1189 static void
1190 dmu_sync_late_arrival_done(zio_t *zio)
1191 {
1192 	blkptr_t *bp = zio->io_bp;
1193 	dmu_sync_arg_t *dsa = zio->io_private;
1194 
1195 	if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1196 		ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1197 		ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1198 		zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1199 	}
1200 
1201 	dmu_tx_commit(dsa->dsa_tx);
1202 
1203 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1204 
1205 	kmem_free(dsa, sizeof (*dsa));
1206 }
1207 
1208 static int
1209 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1210     zio_prop_t *zp, zbookmark_t *zb)
1211 {
1212 	dmu_sync_arg_t *dsa;
1213 	dmu_tx_t *tx;
1214 
1215 	tx = dmu_tx_create(os);
1216 	dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1217 	if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1218 		dmu_tx_abort(tx);
1219 		return (EIO);	/* Make zl_get_data do txg_waited_synced() */
1220 	}
1221 
1222 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1223 	dsa->dsa_dr = NULL;
1224 	dsa->dsa_done = done;
1225 	dsa->dsa_zgd = zgd;
1226 	dsa->dsa_tx = tx;
1227 
1228 	zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1229 	    zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1230 	    dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1231 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1232 
1233 	return (0);
1234 }
1235 
1236 /*
1237  * Intent log support: sync the block associated with db to disk.
1238  * N.B. and XXX: the caller is responsible for making sure that the
1239  * data isn't changing while dmu_sync() is writing it.
1240  *
1241  * Return values:
1242  *
1243  *	EEXIST: this txg has already been synced, so there's nothing to to.
1244  *		The caller should not log the write.
1245  *
1246  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1247  *		The caller should not log the write.
1248  *
1249  *	EALREADY: this block is already in the process of being synced.
1250  *		The caller should track its progress (somehow).
1251  *
1252  *	EIO: could not do the I/O.
1253  *		The caller should do a txg_wait_synced().
1254  *
1255  *	0: the I/O has been initiated.
1256  *		The caller should log this blkptr in the done callback.
1257  *		It is possible that the I/O will fail, in which case
1258  *		the error will be reported to the done callback and
1259  *		propagated to pio from zio_done().
1260  */
1261 int
1262 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1263 {
1264 	blkptr_t *bp = zgd->zgd_bp;
1265 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1266 	objset_t *os = db->db_objset;
1267 	dsl_dataset_t *ds = os->os_dsl_dataset;
1268 	dbuf_dirty_record_t *dr;
1269 	dmu_sync_arg_t *dsa;
1270 	zbookmark_t zb;
1271 	zio_prop_t zp;
1272 
1273 	ASSERT(pio != NULL);
1274 	ASSERT(BP_IS_HOLE(bp));
1275 	ASSERT(txg != 0);
1276 
1277 	SET_BOOKMARK(&zb, ds->ds_object,
1278 	    db->db.db_object, db->db_level, db->db_blkid);
1279 
1280 	dmu_write_policy(os, db->db_dnode, db->db_level, WP_DMU_SYNC, &zp);
1281 
1282 	/*
1283 	 * If we're frozen (running ziltest), we always need to generate a bp.
1284 	 */
1285 	if (txg > spa_freeze_txg(os->os_spa))
1286 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1287 
1288 	/*
1289 	 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1290 	 * and us.  If we determine that this txg is not yet syncing,
1291 	 * but it begins to sync a moment later, that's OK because the
1292 	 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1293 	 */
1294 	mutex_enter(&db->db_mtx);
1295 
1296 	if (txg <= spa_last_synced_txg(os->os_spa)) {
1297 		/*
1298 		 * This txg has already synced.  There's nothing to do.
1299 		 */
1300 		mutex_exit(&db->db_mtx);
1301 		return (EEXIST);
1302 	}
1303 
1304 	if (txg <= spa_syncing_txg(os->os_spa)) {
1305 		/*
1306 		 * This txg is currently syncing, so we can't mess with
1307 		 * the dirty record anymore; just write a new log block.
1308 		 */
1309 		mutex_exit(&db->db_mtx);
1310 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1311 	}
1312 
1313 	dr = db->db_last_dirty;
1314 	while (dr && dr->dr_txg != txg)
1315 		dr = dr->dr_next;
1316 
1317 	if (dr == NULL) {
1318 		/*
1319 		 * There's no dr for this dbuf, so it must have been freed.
1320 		 * There's no need to log writes to freed blocks, so we're done.
1321 		 */
1322 		mutex_exit(&db->db_mtx);
1323 		return (ENOENT);
1324 	}
1325 
1326 	ASSERT(dr->dr_txg == txg);
1327 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1328 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1329 		/*
1330 		 * We have already issued a sync write for this buffer,
1331 		 * or this buffer has already been synced.  It could not
1332 		 * have been dirtied since, or we would have cleared the state.
1333 		 */
1334 		mutex_exit(&db->db_mtx);
1335 		return (EALREADY);
1336 	}
1337 
1338 	ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1339 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1340 	mutex_exit(&db->db_mtx);
1341 
1342 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1343 	dsa->dsa_dr = dr;
1344 	dsa->dsa_done = done;
1345 	dsa->dsa_zgd = zgd;
1346 	dsa->dsa_tx = NULL;
1347 
1348 	zio_nowait(arc_write(pio, os->os_spa, txg,
1349 	    bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp,
1350 	    dmu_sync_ready, dmu_sync_done, dsa,
1351 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1352 
1353 	return (0);
1354 }
1355 
1356 int
1357 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1358 	dmu_tx_t *tx)
1359 {
1360 	dnode_t *dn;
1361 	int err;
1362 
1363 	err = dnode_hold(os, object, FTAG, &dn);
1364 	if (err)
1365 		return (err);
1366 	err = dnode_set_blksz(dn, size, ibs, tx);
1367 	dnode_rele(dn, FTAG);
1368 	return (err);
1369 }
1370 
1371 void
1372 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1373 	dmu_tx_t *tx)
1374 {
1375 	dnode_t *dn;
1376 
1377 	/* XXX assumes dnode_hold will not get an i/o error */
1378 	(void) dnode_hold(os, object, FTAG, &dn);
1379 	ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1380 	dn->dn_checksum = checksum;
1381 	dnode_setdirty(dn, tx);
1382 	dnode_rele(dn, FTAG);
1383 }
1384 
1385 void
1386 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1387 	dmu_tx_t *tx)
1388 {
1389 	dnode_t *dn;
1390 
1391 	/* XXX assumes dnode_hold will not get an i/o error */
1392 	(void) dnode_hold(os, object, FTAG, &dn);
1393 	ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1394 	dn->dn_compress = compress;
1395 	dnode_setdirty(dn, tx);
1396 	dnode_rele(dn, FTAG);
1397 }
1398 
1399 int zfs_mdcomp_disable = 0;
1400 
1401 void
1402 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1403 {
1404 	dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1405 	boolean_t ismd = (level > 0 || dmu_ot[type].ot_metadata);
1406 	enum zio_checksum checksum = os->os_checksum;
1407 	enum zio_compress compress = os->os_compress;
1408 	enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1409 	boolean_t dedup;
1410 	boolean_t dedup_verify = os->os_dedup_verify;
1411 	int copies = os->os_copies;
1412 
1413 	/*
1414 	 * Determine checksum setting.
1415 	 */
1416 	if (ismd) {
1417 		/*
1418 		 * Metadata always gets checksummed.  If the data
1419 		 * checksum is multi-bit correctable, and it's not a
1420 		 * ZBT-style checksum, then it's suitable for metadata
1421 		 * as well.  Otherwise, the metadata checksum defaults
1422 		 * to fletcher4.
1423 		 */
1424 		if (zio_checksum_table[checksum].ci_correctable < 1 ||
1425 		    zio_checksum_table[checksum].ci_eck)
1426 			checksum = ZIO_CHECKSUM_FLETCHER_4;
1427 	} else {
1428 		checksum = zio_checksum_select(dn->dn_checksum, checksum);
1429 	}
1430 
1431 	/*
1432 	 * Determine compression setting.
1433 	 */
1434 	if (ismd) {
1435 		/*
1436 		 * XXX -- we should design a compression algorithm
1437 		 * that specializes in arrays of bps.
1438 		 */
1439 		compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1440 		    ZIO_COMPRESS_LZJB;
1441 	} else {
1442 		compress = zio_compress_select(dn->dn_compress, compress);
1443 	}
1444 
1445 	/*
1446 	 * Determine dedup setting.  If we are in dmu_sync(), we won't
1447 	 * actually dedup now because that's all done in syncing context;
1448 	 * but we do want to use the dedup checkum.  If the checksum is not
1449 	 * strong enough to ensure unique signatures, force dedup_verify.
1450 	 */
1451 	dedup = (!ismd && dedup_checksum != ZIO_CHECKSUM_OFF);
1452 	if (dedup) {
1453 		checksum = dedup_checksum;
1454 		if (!zio_checksum_table[checksum].ci_dedup)
1455 			dedup_verify = 1;
1456 	}
1457 
1458 	if (wp & WP_DMU_SYNC)
1459 		dedup = 0;
1460 
1461 	if (wp & WP_NOFILL) {
1462 		ASSERT(!ismd && level == 0);
1463 		checksum = ZIO_CHECKSUM_OFF;
1464 		compress = ZIO_COMPRESS_OFF;
1465 		dedup = B_FALSE;
1466 	}
1467 
1468 	zp->zp_checksum = checksum;
1469 	zp->zp_compress = compress;
1470 	zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1471 	zp->zp_level = level;
1472 	zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1473 	zp->zp_dedup = dedup;
1474 	zp->zp_dedup_verify = dedup && dedup_verify;
1475 }
1476 
1477 int
1478 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1479 {
1480 	dnode_t *dn;
1481 	int i, err;
1482 
1483 	err = dnode_hold(os, object, FTAG, &dn);
1484 	if (err)
1485 		return (err);
1486 	/*
1487 	 * Sync any current changes before
1488 	 * we go trundling through the block pointers.
1489 	 */
1490 	for (i = 0; i < TXG_SIZE; i++) {
1491 		if (list_link_active(&dn->dn_dirty_link[i]))
1492 			break;
1493 	}
1494 	if (i != TXG_SIZE) {
1495 		dnode_rele(dn, FTAG);
1496 		txg_wait_synced(dmu_objset_pool(os), 0);
1497 		err = dnode_hold(os, object, FTAG, &dn);
1498 		if (err)
1499 			return (err);
1500 	}
1501 
1502 	err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1503 	dnode_rele(dn, FTAG);
1504 
1505 	return (err);
1506 }
1507 
1508 void
1509 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1510 {
1511 	dnode_phys_t *dnp;
1512 
1513 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1514 	mutex_enter(&dn->dn_mtx);
1515 
1516 	dnp = dn->dn_phys;
1517 
1518 	doi->doi_data_block_size = dn->dn_datablksz;
1519 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
1520 	    1ULL << dn->dn_indblkshift : 0;
1521 	doi->doi_type = dn->dn_type;
1522 	doi->doi_bonus_type = dn->dn_bonustype;
1523 	doi->doi_bonus_size = dn->dn_bonuslen;
1524 	doi->doi_indirection = dn->dn_nlevels;
1525 	doi->doi_checksum = dn->dn_checksum;
1526 	doi->doi_compress = dn->dn_compress;
1527 	doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1528 	doi->doi_max_offset = (dnp->dn_maxblkid + 1) * dn->dn_datablksz;
1529 	doi->doi_fill_count = 0;
1530 	for (int i = 0; i < dnp->dn_nblkptr; i++)
1531 		doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1532 
1533 	mutex_exit(&dn->dn_mtx);
1534 	rw_exit(&dn->dn_struct_rwlock);
1535 }
1536 
1537 /*
1538  * Get information on a DMU object.
1539  * If doi is NULL, just indicates whether the object exists.
1540  */
1541 int
1542 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1543 {
1544 	dnode_t *dn;
1545 	int err = dnode_hold(os, object, FTAG, &dn);
1546 
1547 	if (err)
1548 		return (err);
1549 
1550 	if (doi != NULL)
1551 		dmu_object_info_from_dnode(dn, doi);
1552 
1553 	dnode_rele(dn, FTAG);
1554 	return (0);
1555 }
1556 
1557 /*
1558  * As above, but faster; can be used when you have a held dbuf in hand.
1559  */
1560 void
1561 dmu_object_info_from_db(dmu_buf_t *db, dmu_object_info_t *doi)
1562 {
1563 	dmu_object_info_from_dnode(((dmu_buf_impl_t *)db)->db_dnode, doi);
1564 }
1565 
1566 /*
1567  * Faster still when you only care about the size.
1568  * This is specifically optimized for zfs_getattr().
1569  */
1570 void
1571 dmu_object_size_from_db(dmu_buf_t *db, uint32_t *blksize, u_longlong_t *nblk512)
1572 {
1573 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1574 
1575 	*blksize = dn->dn_datablksz;
1576 	/* add 1 for dnode space */
1577 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1578 	    SPA_MINBLOCKSHIFT) + 1;
1579 }
1580 
1581 void
1582 byteswap_uint64_array(void *vbuf, size_t size)
1583 {
1584 	uint64_t *buf = vbuf;
1585 	size_t count = size >> 3;
1586 	int i;
1587 
1588 	ASSERT((size & 7) == 0);
1589 
1590 	for (i = 0; i < count; i++)
1591 		buf[i] = BSWAP_64(buf[i]);
1592 }
1593 
1594 void
1595 byteswap_uint32_array(void *vbuf, size_t size)
1596 {
1597 	uint32_t *buf = vbuf;
1598 	size_t count = size >> 2;
1599 	int i;
1600 
1601 	ASSERT((size & 3) == 0);
1602 
1603 	for (i = 0; i < count; i++)
1604 		buf[i] = BSWAP_32(buf[i]);
1605 }
1606 
1607 void
1608 byteswap_uint16_array(void *vbuf, size_t size)
1609 {
1610 	uint16_t *buf = vbuf;
1611 	size_t count = size >> 1;
1612 	int i;
1613 
1614 	ASSERT((size & 1) == 0);
1615 
1616 	for (i = 0; i < count; i++)
1617 		buf[i] = BSWAP_16(buf[i]);
1618 }
1619 
1620 /* ARGSUSED */
1621 void
1622 byteswap_uint8_array(void *vbuf, size_t size)
1623 {
1624 }
1625 
1626 void
1627 dmu_init(void)
1628 {
1629 	dbuf_init();
1630 	dnode_init();
1631 	zfetch_init();
1632 	arc_init();
1633 	l2arc_init();
1634 	xuio_stat_init();
1635 	sa_cache_init();
1636 }
1637 
1638 void
1639 dmu_fini(void)
1640 {
1641 	arc_fini();
1642 	zfetch_fini();
1643 	dnode_fini();
1644 	dbuf_fini();
1645 	l2arc_fini();
1646 	xuio_stat_fini();
1647 	sa_cache_fini();
1648 }
1649