xref: /titanic_52/usr/src/uts/common/fs/zfs/dmu.c (revision ea46d7619be99679c4c99ed47508abe31d5e0979)
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, int flags)
101 {
102 	dnode_t *dn;
103 	uint64_t blkid;
104 	dmu_buf_impl_t *db;
105 	int err;
106 	int db_flags = DB_RF_CANFAIL;
107 
108 	if (flags & DMU_READ_NO_PREFETCH)
109 		db_flags |= DB_RF_NOPREFETCH;
110 
111 	err = dnode_hold(os, object, FTAG, &dn);
112 	if (err)
113 		return (err);
114 	blkid = dbuf_whichblock(dn, offset);
115 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
116 	db = dbuf_hold(dn, blkid, tag);
117 	rw_exit(&dn->dn_struct_rwlock);
118 	if (db == NULL) {
119 		err = EIO;
120 	} else {
121 		err = dbuf_read(db, NULL, db_flags);
122 		if (err) {
123 			dbuf_rele(db, tag);
124 			db = NULL;
125 		}
126 	}
127 
128 	dnode_rele(dn, FTAG);
129 	*dbp = &db->db;
130 	return (err);
131 }
132 
133 int
134 dmu_bonus_max(void)
135 {
136 	return (DN_MAX_BONUSLEN);
137 }
138 
139 int
140 dmu_set_bonus(dmu_buf_t *db, int newsize, dmu_tx_t *tx)
141 {
142 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
143 
144 	if (dn->dn_bonus != (dmu_buf_impl_t *)db)
145 		return (EINVAL);
146 	if (newsize < 0 || newsize > db->db_size)
147 		return (EINVAL);
148 	dnode_setbonuslen(dn, newsize, tx);
149 	return (0);
150 }
151 
152 int
153 dmu_set_bonustype(dmu_buf_t *db, dmu_object_type_t type, dmu_tx_t *tx)
154 {
155 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
156 
157 	if (type > DMU_OT_NUMTYPES)
158 		return (EINVAL);
159 
160 	if (dn->dn_bonus != (dmu_buf_impl_t *)db)
161 		return (EINVAL);
162 
163 	dnode_setbonus_type(dn, type, tx);
164 	return (0);
165 }
166 
167 int
168 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
169 {
170 	dnode_t *dn;
171 	int error;
172 
173 	error = dnode_hold(os, object, FTAG, &dn);
174 	dbuf_rm_spill(dn, tx);
175 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
176 	dnode_rm_spill(dn, tx);
177 	rw_exit(&dn->dn_struct_rwlock);
178 	dnode_rele(dn, FTAG);
179 	return (error);
180 }
181 
182 /*
183  * returns ENOENT, EIO, or 0.
184  */
185 int
186 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
187 {
188 	dnode_t *dn;
189 	dmu_buf_impl_t *db;
190 	int error;
191 
192 	error = dnode_hold(os, object, FTAG, &dn);
193 	if (error)
194 		return (error);
195 
196 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
197 	if (dn->dn_bonus == NULL) {
198 		rw_exit(&dn->dn_struct_rwlock);
199 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
200 		if (dn->dn_bonus == NULL)
201 			dbuf_create_bonus(dn);
202 	}
203 	db = dn->dn_bonus;
204 	rw_exit(&dn->dn_struct_rwlock);
205 
206 	/* as long as the bonus buf is held, the dnode will be held */
207 	if (refcount_add(&db->db_holds, tag) == 1)
208 		VERIFY(dnode_add_ref(dn, db));
209 
210 	dnode_rele(dn, FTAG);
211 
212 	VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
213 
214 	*dbp = &db->db;
215 	return (0);
216 }
217 
218 /*
219  * returns ENOENT, EIO, or 0.
220  *
221  * This interface will allocate a blank spill dbuf when a spill blk
222  * doesn't already exist on the dnode.
223  *
224  * if you only want to find an already existing spill db, then
225  * dmu_spill_hold_existing() should be used.
226  */
227 int
228 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
229 {
230 	dmu_buf_impl_t *db = NULL;
231 	int err;
232 
233 	if ((flags & DB_RF_HAVESTRUCT) == 0)
234 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
235 
236 	db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
237 
238 	if ((flags & DB_RF_HAVESTRUCT) == 0)
239 		rw_exit(&dn->dn_struct_rwlock);
240 
241 	ASSERT(db != NULL);
242 	err = dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | flags);
243 	*dbp = &db->db;
244 	return (err);
245 }
246 
247 int
248 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
249 {
250 	dnode_t *dn = ((dmu_buf_impl_t *)bonus)->db_dnode;
251 	int err;
252 
253 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA)
254 		return (EINVAL);
255 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
256 
257 	if (!dn->dn_have_spill) {
258 		rw_exit(&dn->dn_struct_rwlock);
259 		return (ENOENT);
260 	}
261 	err = dmu_spill_hold_by_dnode(dn, DB_RF_HAVESTRUCT, tag, dbp);
262 	rw_exit(&dn->dn_struct_rwlock);
263 	return (err);
264 }
265 
266 int
267 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
268 {
269 	return (dmu_spill_hold_by_dnode(((dmu_buf_impl_t *)bonus)->db_dnode,
270 	    0, tag, dbp));
271 }
272 
273 /*
274  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
275  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
276  * and can induce severe lock contention when writing to several files
277  * whose dnodes are in the same block.
278  */
279 static int
280 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
281     int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
282 {
283 	dsl_pool_t *dp = NULL;
284 	dmu_buf_t **dbp;
285 	uint64_t blkid, nblks, i;
286 	uint32_t dbuf_flags;
287 	int err;
288 	zio_t *zio;
289 	hrtime_t start;
290 
291 	ASSERT(length <= DMU_MAX_ACCESS);
292 
293 	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
294 	if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
295 		dbuf_flags |= DB_RF_NOPREFETCH;
296 
297 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
298 	if (dn->dn_datablkshift) {
299 		int blkshift = dn->dn_datablkshift;
300 		nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
301 		    P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
302 	} else {
303 		if (offset + length > dn->dn_datablksz) {
304 			zfs_panic_recover("zfs: accessing past end of object "
305 			    "%llx/%llx (size=%u access=%llu+%llu)",
306 			    (longlong_t)dn->dn_objset->
307 			    os_dsl_dataset->ds_object,
308 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
309 			    (longlong_t)offset, (longlong_t)length);
310 			rw_exit(&dn->dn_struct_rwlock);
311 			return (EIO);
312 		}
313 		nblks = 1;
314 	}
315 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
316 
317 	if (dn->dn_objset->os_dsl_dataset)
318 		dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool;
319 	if (dp && dsl_pool_sync_context(dp))
320 		start = gethrtime();
321 	zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
322 	blkid = dbuf_whichblock(dn, offset);
323 	for (i = 0; i < nblks; i++) {
324 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
325 		if (db == NULL) {
326 			rw_exit(&dn->dn_struct_rwlock);
327 			dmu_buf_rele_array(dbp, nblks, tag);
328 			zio_nowait(zio);
329 			return (EIO);
330 		}
331 		/* initiate async i/o */
332 		if (read) {
333 			(void) dbuf_read(db, zio, dbuf_flags);
334 		}
335 		dbp[i] = &db->db;
336 	}
337 	rw_exit(&dn->dn_struct_rwlock);
338 
339 	/* wait for async i/o */
340 	err = zio_wait(zio);
341 	/* track read overhead when we are in sync context */
342 	if (dp && dsl_pool_sync_context(dp))
343 		dp->dp_read_overhead += gethrtime() - start;
344 	if (err) {
345 		dmu_buf_rele_array(dbp, nblks, tag);
346 		return (err);
347 	}
348 
349 	/* wait for other io to complete */
350 	if (read) {
351 		for (i = 0; i < nblks; i++) {
352 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
353 			mutex_enter(&db->db_mtx);
354 			while (db->db_state == DB_READ ||
355 			    db->db_state == DB_FILL)
356 				cv_wait(&db->db_changed, &db->db_mtx);
357 			if (db->db_state == DB_UNCACHED)
358 				err = EIO;
359 			mutex_exit(&db->db_mtx);
360 			if (err) {
361 				dmu_buf_rele_array(dbp, nblks, tag);
362 				return (err);
363 			}
364 		}
365 	}
366 
367 	*numbufsp = nblks;
368 	*dbpp = dbp;
369 	return (0);
370 }
371 
372 static int
373 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
374     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
375 {
376 	dnode_t *dn;
377 	int err;
378 
379 	err = dnode_hold(os, object, FTAG, &dn);
380 	if (err)
381 		return (err);
382 
383 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
384 	    numbufsp, dbpp, DMU_READ_PREFETCH);
385 
386 	dnode_rele(dn, FTAG);
387 
388 	return (err);
389 }
390 
391 int
392 dmu_buf_hold_array_by_bonus(dmu_buf_t *db, uint64_t offset,
393     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
394 {
395 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
396 	int err;
397 
398 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
399 	    numbufsp, dbpp, DMU_READ_PREFETCH);
400 
401 	return (err);
402 }
403 
404 void
405 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
406 {
407 	int i;
408 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
409 
410 	if (numbufs == 0)
411 		return;
412 
413 	for (i = 0; i < numbufs; i++) {
414 		if (dbp[i])
415 			dbuf_rele(dbp[i], tag);
416 	}
417 
418 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
419 }
420 
421 void
422 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
423 {
424 	dnode_t *dn;
425 	uint64_t blkid;
426 	int nblks, i, err;
427 
428 	if (zfs_prefetch_disable)
429 		return;
430 
431 	if (len == 0) {  /* they're interested in the bonus buffer */
432 		dn = os->os_meta_dnode;
433 
434 		if (object == 0 || object >= DN_MAX_OBJECT)
435 			return;
436 
437 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
438 		blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
439 		dbuf_prefetch(dn, blkid);
440 		rw_exit(&dn->dn_struct_rwlock);
441 		return;
442 	}
443 
444 	/*
445 	 * XXX - Note, if the dnode for the requested object is not
446 	 * already cached, we will do a *synchronous* read in the
447 	 * dnode_hold() call.  The same is true for any indirects.
448 	 */
449 	err = dnode_hold(os, object, FTAG, &dn);
450 	if (err != 0)
451 		return;
452 
453 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
454 	if (dn->dn_datablkshift) {
455 		int blkshift = dn->dn_datablkshift;
456 		nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
457 		    P2ALIGN(offset, 1<<blkshift)) >> blkshift;
458 	} else {
459 		nblks = (offset < dn->dn_datablksz);
460 	}
461 
462 	if (nblks != 0) {
463 		blkid = dbuf_whichblock(dn, offset);
464 		for (i = 0; i < nblks; i++)
465 			dbuf_prefetch(dn, blkid+i);
466 	}
467 
468 	rw_exit(&dn->dn_struct_rwlock);
469 
470 	dnode_rele(dn, FTAG);
471 }
472 
473 /*
474  * Get the next "chunk" of file data to free.  We traverse the file from
475  * the end so that the file gets shorter over time (if we crashes in the
476  * middle, this will leave us in a better state).  We find allocated file
477  * data by simply searching the allocated level 1 indirects.
478  */
479 static int
480 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit)
481 {
482 	uint64_t len = *start - limit;
483 	uint64_t blkcnt = 0;
484 	uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1));
485 	uint64_t iblkrange =
486 	    dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
487 
488 	ASSERT(limit <= *start);
489 
490 	if (len <= iblkrange * maxblks) {
491 		*start = limit;
492 		return (0);
493 	}
494 	ASSERT(ISP2(iblkrange));
495 
496 	while (*start > limit && blkcnt < maxblks) {
497 		int err;
498 
499 		/* find next allocated L1 indirect */
500 		err = dnode_next_offset(dn,
501 		    DNODE_FIND_BACKWARDS, start, 2, 1, 0);
502 
503 		/* if there are no more, then we are done */
504 		if (err == ESRCH) {
505 			*start = limit;
506 			return (0);
507 		} else if (err) {
508 			return (err);
509 		}
510 		blkcnt += 1;
511 
512 		/* reset offset to end of "next" block back */
513 		*start = P2ALIGN(*start, iblkrange);
514 		if (*start <= limit)
515 			*start = limit;
516 		else
517 			*start -= 1;
518 	}
519 	return (0);
520 }
521 
522 static int
523 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
524     uint64_t length, boolean_t free_dnode)
525 {
526 	dmu_tx_t *tx;
527 	uint64_t object_size, start, end, len;
528 	boolean_t trunc = (length == DMU_OBJECT_END);
529 	int align, err;
530 
531 	align = 1 << dn->dn_datablkshift;
532 	ASSERT(align > 0);
533 	object_size = align == 1 ? dn->dn_datablksz :
534 	    (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
535 
536 	end = offset + length;
537 	if (trunc || end > object_size)
538 		end = object_size;
539 	if (end <= offset)
540 		return (0);
541 	length = end - offset;
542 
543 	while (length) {
544 		start = end;
545 		/* assert(offset <= start) */
546 		err = get_next_chunk(dn, &start, offset);
547 		if (err)
548 			return (err);
549 		len = trunc ? DMU_OBJECT_END : end - start;
550 
551 		tx = dmu_tx_create(os);
552 		dmu_tx_hold_free(tx, dn->dn_object, start, len);
553 		err = dmu_tx_assign(tx, TXG_WAIT);
554 		if (err) {
555 			dmu_tx_abort(tx);
556 			return (err);
557 		}
558 
559 		dnode_free_range(dn, start, trunc ? -1 : len, tx);
560 
561 		if (start == 0 && free_dnode) {
562 			ASSERT(trunc);
563 			dnode_free(dn, tx);
564 		}
565 
566 		length -= end - start;
567 
568 		dmu_tx_commit(tx);
569 		end = start;
570 	}
571 	return (0);
572 }
573 
574 int
575 dmu_free_long_range(objset_t *os, uint64_t object,
576     uint64_t offset, uint64_t length)
577 {
578 	dnode_t *dn;
579 	int err;
580 
581 	err = dnode_hold(os, object, FTAG, &dn);
582 	if (err != 0)
583 		return (err);
584 	err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
585 	dnode_rele(dn, FTAG);
586 	return (err);
587 }
588 
589 int
590 dmu_free_object(objset_t *os, uint64_t object)
591 {
592 	dnode_t *dn;
593 	dmu_tx_t *tx;
594 	int err;
595 
596 	err = dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED,
597 	    FTAG, &dn);
598 	if (err != 0)
599 		return (err);
600 	if (dn->dn_nlevels == 1) {
601 		tx = dmu_tx_create(os);
602 		dmu_tx_hold_bonus(tx, object);
603 		dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
604 		err = dmu_tx_assign(tx, TXG_WAIT);
605 		if (err == 0) {
606 			dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
607 			dnode_free(dn, tx);
608 			dmu_tx_commit(tx);
609 		} else {
610 			dmu_tx_abort(tx);
611 		}
612 	} else {
613 		err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
614 	}
615 	dnode_rele(dn, FTAG);
616 	return (err);
617 }
618 
619 int
620 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
621     uint64_t size, dmu_tx_t *tx)
622 {
623 	dnode_t *dn;
624 	int err = dnode_hold(os, object, FTAG, &dn);
625 	if (err)
626 		return (err);
627 	ASSERT(offset < UINT64_MAX);
628 	ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
629 	dnode_free_range(dn, offset, size, tx);
630 	dnode_rele(dn, FTAG);
631 	return (0);
632 }
633 
634 int
635 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
636     void *buf, uint32_t flags)
637 {
638 	dnode_t *dn;
639 	dmu_buf_t **dbp;
640 	int numbufs, err;
641 
642 	err = dnode_hold(os, object, FTAG, &dn);
643 	if (err)
644 		return (err);
645 
646 	/*
647 	 * Deal with odd block sizes, where there can't be data past the first
648 	 * block.  If we ever do the tail block optimization, we will need to
649 	 * handle that here as well.
650 	 */
651 	if (dn->dn_maxblkid == 0) {
652 		int newsz = offset > dn->dn_datablksz ? 0 :
653 		    MIN(size, dn->dn_datablksz - offset);
654 		bzero((char *)buf + newsz, size - newsz);
655 		size = newsz;
656 	}
657 
658 	while (size > 0) {
659 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
660 		int i;
661 
662 		/*
663 		 * NB: we could do this block-at-a-time, but it's nice
664 		 * to be reading in parallel.
665 		 */
666 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
667 		    TRUE, FTAG, &numbufs, &dbp, flags);
668 		if (err)
669 			break;
670 
671 		for (i = 0; i < numbufs; i++) {
672 			int tocpy;
673 			int bufoff;
674 			dmu_buf_t *db = dbp[i];
675 
676 			ASSERT(size > 0);
677 
678 			bufoff = offset - db->db_offset;
679 			tocpy = (int)MIN(db->db_size - bufoff, size);
680 
681 			bcopy((char *)db->db_data + bufoff, buf, tocpy);
682 
683 			offset += tocpy;
684 			size -= tocpy;
685 			buf = (char *)buf + tocpy;
686 		}
687 		dmu_buf_rele_array(dbp, numbufs, FTAG);
688 	}
689 	dnode_rele(dn, FTAG);
690 	return (err);
691 }
692 
693 void
694 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
695     const void *buf, dmu_tx_t *tx)
696 {
697 	dmu_buf_t **dbp;
698 	int numbufs, i;
699 
700 	if (size == 0)
701 		return;
702 
703 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
704 	    FALSE, FTAG, &numbufs, &dbp));
705 
706 	for (i = 0; i < numbufs; i++) {
707 		int tocpy;
708 		int bufoff;
709 		dmu_buf_t *db = dbp[i];
710 
711 		ASSERT(size > 0);
712 
713 		bufoff = offset - db->db_offset;
714 		tocpy = (int)MIN(db->db_size - bufoff, size);
715 
716 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
717 
718 		if (tocpy == db->db_size)
719 			dmu_buf_will_fill(db, tx);
720 		else
721 			dmu_buf_will_dirty(db, tx);
722 
723 		bcopy(buf, (char *)db->db_data + bufoff, tocpy);
724 
725 		if (tocpy == db->db_size)
726 			dmu_buf_fill_done(db, tx);
727 
728 		offset += tocpy;
729 		size -= tocpy;
730 		buf = (char *)buf + tocpy;
731 	}
732 	dmu_buf_rele_array(dbp, numbufs, FTAG);
733 }
734 
735 void
736 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
737     dmu_tx_t *tx)
738 {
739 	dmu_buf_t **dbp;
740 	int numbufs, i;
741 
742 	if (size == 0)
743 		return;
744 
745 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
746 	    FALSE, FTAG, &numbufs, &dbp));
747 
748 	for (i = 0; i < numbufs; i++) {
749 		dmu_buf_t *db = dbp[i];
750 
751 		dmu_buf_will_not_fill(db, tx);
752 	}
753 	dmu_buf_rele_array(dbp, numbufs, FTAG);
754 }
755 
756 /*
757  * DMU support for xuio
758  */
759 kstat_t *xuio_ksp = NULL;
760 
761 int
762 dmu_xuio_init(xuio_t *xuio, int nblk)
763 {
764 	dmu_xuio_t *priv;
765 	uio_t *uio = &xuio->xu_uio;
766 
767 	uio->uio_iovcnt = nblk;
768 	uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
769 
770 	priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
771 	priv->cnt = nblk;
772 	priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
773 	priv->iovp = uio->uio_iov;
774 	XUIO_XUZC_PRIV(xuio) = priv;
775 
776 	if (XUIO_XUZC_RW(xuio) == UIO_READ)
777 		XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
778 	else
779 		XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
780 
781 	return (0);
782 }
783 
784 void
785 dmu_xuio_fini(xuio_t *xuio)
786 {
787 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
788 	int nblk = priv->cnt;
789 
790 	kmem_free(priv->iovp, nblk * sizeof (iovec_t));
791 	kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
792 	kmem_free(priv, sizeof (dmu_xuio_t));
793 
794 	if (XUIO_XUZC_RW(xuio) == UIO_READ)
795 		XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
796 	else
797 		XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
798 }
799 
800 /*
801  * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
802  * and increase priv->next by 1.
803  */
804 int
805 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
806 {
807 	struct iovec *iov;
808 	uio_t *uio = &xuio->xu_uio;
809 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
810 	int i = priv->next++;
811 
812 	ASSERT(i < priv->cnt);
813 	ASSERT(off + n <= arc_buf_size(abuf));
814 	iov = uio->uio_iov + i;
815 	iov->iov_base = (char *)abuf->b_data + off;
816 	iov->iov_len = n;
817 	priv->bufs[i] = abuf;
818 	return (0);
819 }
820 
821 int
822 dmu_xuio_cnt(xuio_t *xuio)
823 {
824 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
825 	return (priv->cnt);
826 }
827 
828 arc_buf_t *
829 dmu_xuio_arcbuf(xuio_t *xuio, int i)
830 {
831 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
832 
833 	ASSERT(i < priv->cnt);
834 	return (priv->bufs[i]);
835 }
836 
837 void
838 dmu_xuio_clear(xuio_t *xuio, int i)
839 {
840 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
841 
842 	ASSERT(i < priv->cnt);
843 	priv->bufs[i] = NULL;
844 }
845 
846 static void
847 xuio_stat_init(void)
848 {
849 	xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
850 	    KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
851 	    KSTAT_FLAG_VIRTUAL);
852 	if (xuio_ksp != NULL) {
853 		xuio_ksp->ks_data = &xuio_stats;
854 		kstat_install(xuio_ksp);
855 	}
856 }
857 
858 static void
859 xuio_stat_fini(void)
860 {
861 	if (xuio_ksp != NULL) {
862 		kstat_delete(xuio_ksp);
863 		xuio_ksp = NULL;
864 	}
865 }
866 
867 void
868 xuio_stat_wbuf_copied()
869 {
870 	XUIOSTAT_BUMP(xuiostat_wbuf_copied);
871 }
872 
873 void
874 xuio_stat_wbuf_nocopy()
875 {
876 	XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
877 }
878 
879 #ifdef _KERNEL
880 int
881 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
882 {
883 	dmu_buf_t **dbp;
884 	int numbufs, i, err;
885 	xuio_t *xuio = NULL;
886 
887 	/*
888 	 * NB: we could do this block-at-a-time, but it's nice
889 	 * to be reading in parallel.
890 	 */
891 	err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
892 	    &numbufs, &dbp);
893 	if (err)
894 		return (err);
895 
896 	if (uio->uio_extflg == UIO_XUIO)
897 		xuio = (xuio_t *)uio;
898 
899 	for (i = 0; i < numbufs; i++) {
900 		int tocpy;
901 		int bufoff;
902 		dmu_buf_t *db = dbp[i];
903 
904 		ASSERT(size > 0);
905 
906 		bufoff = uio->uio_loffset - db->db_offset;
907 		tocpy = (int)MIN(db->db_size - bufoff, size);
908 
909 		if (xuio) {
910 			dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
911 			arc_buf_t *dbuf_abuf = dbi->db_buf;
912 			arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
913 			err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
914 			if (!err) {
915 				uio->uio_resid -= tocpy;
916 				uio->uio_loffset += tocpy;
917 			}
918 
919 			if (abuf == dbuf_abuf)
920 				XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
921 			else
922 				XUIOSTAT_BUMP(xuiostat_rbuf_copied);
923 		} else {
924 			err = uiomove((char *)db->db_data + bufoff, tocpy,
925 			    UIO_READ, uio);
926 		}
927 		if (err)
928 			break;
929 
930 		size -= tocpy;
931 	}
932 	dmu_buf_rele_array(dbp, numbufs, FTAG);
933 
934 	return (err);
935 }
936 
937 static int
938 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
939 {
940 	dmu_buf_t **dbp;
941 	int numbufs;
942 	int err = 0;
943 	int i;
944 
945 	err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
946 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
947 	if (err)
948 		return (err);
949 
950 	for (i = 0; i < numbufs; i++) {
951 		int tocpy;
952 		int bufoff;
953 		dmu_buf_t *db = dbp[i];
954 
955 		ASSERT(size > 0);
956 
957 		bufoff = uio->uio_loffset - db->db_offset;
958 		tocpy = (int)MIN(db->db_size - bufoff, size);
959 
960 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
961 
962 		if (tocpy == db->db_size)
963 			dmu_buf_will_fill(db, tx);
964 		else
965 			dmu_buf_will_dirty(db, tx);
966 
967 		/*
968 		 * XXX uiomove could block forever (eg. nfs-backed
969 		 * pages).  There needs to be a uiolockdown() function
970 		 * to lock the pages in memory, so that uiomove won't
971 		 * block.
972 		 */
973 		err = uiomove((char *)db->db_data + bufoff, tocpy,
974 		    UIO_WRITE, uio);
975 
976 		if (tocpy == db->db_size)
977 			dmu_buf_fill_done(db, tx);
978 
979 		if (err)
980 			break;
981 
982 		size -= tocpy;
983 	}
984 
985 	dmu_buf_rele_array(dbp, numbufs, FTAG);
986 	return (err);
987 }
988 
989 int
990 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
991     dmu_tx_t *tx)
992 {
993 	if (size == 0)
994 		return (0);
995 
996 	return (dmu_write_uio_dnode(((dmu_buf_impl_t *)zdb)->db_dnode,
997 	    uio, size, tx));
998 }
999 
1000 int
1001 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1002     dmu_tx_t *tx)
1003 {
1004 	dnode_t *dn;
1005 	int err;
1006 
1007 	if (size == 0)
1008 		return (0);
1009 
1010 	err = dnode_hold(os, object, FTAG, &dn);
1011 	if (err)
1012 		return (err);
1013 
1014 	err = dmu_write_uio_dnode(dn, uio, size, tx);
1015 
1016 	dnode_rele(dn, FTAG);
1017 
1018 	return (err);
1019 }
1020 
1021 int
1022 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1023     page_t *pp, dmu_tx_t *tx)
1024 {
1025 	dmu_buf_t **dbp;
1026 	int numbufs, i;
1027 	int err;
1028 
1029 	if (size == 0)
1030 		return (0);
1031 
1032 	err = dmu_buf_hold_array(os, object, offset, size,
1033 	    FALSE, FTAG, &numbufs, &dbp);
1034 	if (err)
1035 		return (err);
1036 
1037 	for (i = 0; i < numbufs; i++) {
1038 		int tocpy, copied, thiscpy;
1039 		int bufoff;
1040 		dmu_buf_t *db = dbp[i];
1041 		caddr_t va;
1042 
1043 		ASSERT(size > 0);
1044 		ASSERT3U(db->db_size, >=, PAGESIZE);
1045 
1046 		bufoff = offset - db->db_offset;
1047 		tocpy = (int)MIN(db->db_size - bufoff, size);
1048 
1049 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1050 
1051 		if (tocpy == db->db_size)
1052 			dmu_buf_will_fill(db, tx);
1053 		else
1054 			dmu_buf_will_dirty(db, tx);
1055 
1056 		for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1057 			ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1058 			thiscpy = MIN(PAGESIZE, tocpy - copied);
1059 			va = zfs_map_page(pp, S_READ);
1060 			bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1061 			zfs_unmap_page(pp, va);
1062 			pp = pp->p_next;
1063 			bufoff += PAGESIZE;
1064 		}
1065 
1066 		if (tocpy == db->db_size)
1067 			dmu_buf_fill_done(db, tx);
1068 
1069 		offset += tocpy;
1070 		size -= tocpy;
1071 	}
1072 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1073 	return (err);
1074 }
1075 #endif
1076 
1077 /*
1078  * Allocate a loaned anonymous arc buffer.
1079  */
1080 arc_buf_t *
1081 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1082 {
1083 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
1084 
1085 	return (arc_loan_buf(dn->dn_objset->os_spa, size));
1086 }
1087 
1088 /*
1089  * Free a loaned arc buffer.
1090  */
1091 void
1092 dmu_return_arcbuf(arc_buf_t *buf)
1093 {
1094 	arc_return_buf(buf, FTAG);
1095 	VERIFY(arc_buf_remove_ref(buf, FTAG) == 1);
1096 }
1097 
1098 /*
1099  * When possible directly assign passed loaned arc buffer to a dbuf.
1100  * If this is not possible copy the contents of passed arc buf via
1101  * dmu_write().
1102  */
1103 void
1104 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1105     dmu_tx_t *tx)
1106 {
1107 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
1108 	dmu_buf_impl_t *db;
1109 	uint32_t blksz = (uint32_t)arc_buf_size(buf);
1110 	uint64_t blkid;
1111 
1112 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1113 	blkid = dbuf_whichblock(dn, offset);
1114 	VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1115 	rw_exit(&dn->dn_struct_rwlock);
1116 
1117 	if (offset == db->db.db_offset && blksz == db->db.db_size) {
1118 		dbuf_assign_arcbuf(db, buf, tx);
1119 		dbuf_rele(db, FTAG);
1120 	} else {
1121 		dbuf_rele(db, FTAG);
1122 		dmu_write(dn->dn_objset, dn->dn_object, offset, blksz,
1123 		    buf->b_data, tx);
1124 		dmu_return_arcbuf(buf);
1125 		XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1126 	}
1127 }
1128 
1129 typedef struct {
1130 	dbuf_dirty_record_t	*dsa_dr;
1131 	dmu_sync_cb_t		*dsa_done;
1132 	zgd_t			*dsa_zgd;
1133 	dmu_tx_t		*dsa_tx;
1134 } dmu_sync_arg_t;
1135 
1136 /* ARGSUSED */
1137 static void
1138 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1139 {
1140 	dmu_sync_arg_t *dsa = varg;
1141 	dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1142 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1143 	blkptr_t *bp = zio->io_bp;
1144 
1145 	if (zio->io_error == 0) {
1146 		if (BP_IS_HOLE(bp)) {
1147 			/*
1148 			 * A block of zeros may compress to a hole, but the
1149 			 * block size still needs to be known for replay.
1150 			 */
1151 			BP_SET_LSIZE(bp, db->db_size);
1152 		} else {
1153 			ASSERT(BP_GET_TYPE(bp) == dn->dn_type);
1154 			ASSERT(BP_GET_LEVEL(bp) == 0);
1155 			bp->blk_fill = 1;
1156 		}
1157 	}
1158 }
1159 
1160 static void
1161 dmu_sync_late_arrival_ready(zio_t *zio)
1162 {
1163 	dmu_sync_ready(zio, NULL, zio->io_private);
1164 }
1165 
1166 /* ARGSUSED */
1167 static void
1168 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1169 {
1170 	dmu_sync_arg_t *dsa = varg;
1171 	dbuf_dirty_record_t *dr = dsa->dsa_dr;
1172 	dmu_buf_impl_t *db = dr->dr_dbuf;
1173 
1174 	mutex_enter(&db->db_mtx);
1175 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1176 	if (zio->io_error == 0) {
1177 		dr->dt.dl.dr_overridden_by = *zio->io_bp;
1178 		dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1179 		dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1180 		if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1181 			BP_ZERO(&dr->dt.dl.dr_overridden_by);
1182 	} else {
1183 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1184 	}
1185 	cv_broadcast(&db->db_changed);
1186 	mutex_exit(&db->db_mtx);
1187 
1188 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1189 
1190 	kmem_free(dsa, sizeof (*dsa));
1191 }
1192 
1193 static void
1194 dmu_sync_late_arrival_done(zio_t *zio)
1195 {
1196 	blkptr_t *bp = zio->io_bp;
1197 	dmu_sync_arg_t *dsa = zio->io_private;
1198 
1199 	if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1200 		ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1201 		ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1202 		zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1203 	}
1204 
1205 	dmu_tx_commit(dsa->dsa_tx);
1206 
1207 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1208 
1209 	kmem_free(dsa, sizeof (*dsa));
1210 }
1211 
1212 static int
1213 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1214     zio_prop_t *zp, zbookmark_t *zb)
1215 {
1216 	dmu_sync_arg_t *dsa;
1217 	dmu_tx_t *tx;
1218 
1219 	tx = dmu_tx_create(os);
1220 	dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1221 	if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1222 		dmu_tx_abort(tx);
1223 		return (EIO);	/* Make zl_get_data do txg_waited_synced() */
1224 	}
1225 
1226 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1227 	dsa->dsa_dr = NULL;
1228 	dsa->dsa_done = done;
1229 	dsa->dsa_zgd = zgd;
1230 	dsa->dsa_tx = tx;
1231 
1232 	zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1233 	    zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1234 	    dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1235 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1236 
1237 	return (0);
1238 }
1239 
1240 /*
1241  * Intent log support: sync the block associated with db to disk.
1242  * N.B. and XXX: the caller is responsible for making sure that the
1243  * data isn't changing while dmu_sync() is writing it.
1244  *
1245  * Return values:
1246  *
1247  *	EEXIST: this txg has already been synced, so there's nothing to to.
1248  *		The caller should not log the write.
1249  *
1250  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1251  *		The caller should not log the write.
1252  *
1253  *	EALREADY: this block is already in the process of being synced.
1254  *		The caller should track its progress (somehow).
1255  *
1256  *	EIO: could not do the I/O.
1257  *		The caller should do a txg_wait_synced().
1258  *
1259  *	0: the I/O has been initiated.
1260  *		The caller should log this blkptr in the done callback.
1261  *		It is possible that the I/O will fail, in which case
1262  *		the error will be reported to the done callback and
1263  *		propagated to pio from zio_done().
1264  */
1265 int
1266 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1267 {
1268 	blkptr_t *bp = zgd->zgd_bp;
1269 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1270 	objset_t *os = db->db_objset;
1271 	dsl_dataset_t *ds = os->os_dsl_dataset;
1272 	dbuf_dirty_record_t *dr;
1273 	dmu_sync_arg_t *dsa;
1274 	zbookmark_t zb;
1275 	zio_prop_t zp;
1276 
1277 	ASSERT(pio != NULL);
1278 	ASSERT(BP_IS_HOLE(bp));
1279 	ASSERT(txg != 0);
1280 
1281 	SET_BOOKMARK(&zb, ds->ds_object,
1282 	    db->db.db_object, db->db_level, db->db_blkid);
1283 
1284 	dmu_write_policy(os, db->db_dnode, db->db_level, WP_DMU_SYNC, &zp);
1285 
1286 	/*
1287 	 * If we're frozen (running ziltest), we always need to generate a bp.
1288 	 */
1289 	if (txg > spa_freeze_txg(os->os_spa))
1290 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1291 
1292 	/*
1293 	 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1294 	 * and us.  If we determine that this txg is not yet syncing,
1295 	 * but it begins to sync a moment later, that's OK because the
1296 	 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1297 	 */
1298 	mutex_enter(&db->db_mtx);
1299 
1300 	if (txg <= spa_last_synced_txg(os->os_spa)) {
1301 		/*
1302 		 * This txg has already synced.  There's nothing to do.
1303 		 */
1304 		mutex_exit(&db->db_mtx);
1305 		return (EEXIST);
1306 	}
1307 
1308 	if (txg <= spa_syncing_txg(os->os_spa)) {
1309 		/*
1310 		 * This txg is currently syncing, so we can't mess with
1311 		 * the dirty record anymore; just write a new log block.
1312 		 */
1313 		mutex_exit(&db->db_mtx);
1314 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1315 	}
1316 
1317 	dr = db->db_last_dirty;
1318 	while (dr && dr->dr_txg != txg)
1319 		dr = dr->dr_next;
1320 
1321 	if (dr == NULL) {
1322 		/*
1323 		 * There's no dr for this dbuf, so it must have been freed.
1324 		 * There's no need to log writes to freed blocks, so we're done.
1325 		 */
1326 		mutex_exit(&db->db_mtx);
1327 		return (ENOENT);
1328 	}
1329 
1330 	ASSERT(dr->dr_txg == txg);
1331 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1332 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1333 		/*
1334 		 * We have already issued a sync write for this buffer,
1335 		 * or this buffer has already been synced.  It could not
1336 		 * have been dirtied since, or we would have cleared the state.
1337 		 */
1338 		mutex_exit(&db->db_mtx);
1339 		return (EALREADY);
1340 	}
1341 
1342 	ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1343 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1344 	mutex_exit(&db->db_mtx);
1345 
1346 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1347 	dsa->dsa_dr = dr;
1348 	dsa->dsa_done = done;
1349 	dsa->dsa_zgd = zgd;
1350 	dsa->dsa_tx = NULL;
1351 
1352 	zio_nowait(arc_write(pio, os->os_spa, txg,
1353 	    bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp,
1354 	    dmu_sync_ready, dmu_sync_done, dsa,
1355 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1356 
1357 	return (0);
1358 }
1359 
1360 int
1361 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1362 	dmu_tx_t *tx)
1363 {
1364 	dnode_t *dn;
1365 	int err;
1366 
1367 	err = dnode_hold(os, object, FTAG, &dn);
1368 	if (err)
1369 		return (err);
1370 	err = dnode_set_blksz(dn, size, ibs, tx);
1371 	dnode_rele(dn, FTAG);
1372 	return (err);
1373 }
1374 
1375 void
1376 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1377 	dmu_tx_t *tx)
1378 {
1379 	dnode_t *dn;
1380 
1381 	/* XXX assumes dnode_hold will not get an i/o error */
1382 	(void) dnode_hold(os, object, FTAG, &dn);
1383 	ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1384 	dn->dn_checksum = checksum;
1385 	dnode_setdirty(dn, tx);
1386 	dnode_rele(dn, FTAG);
1387 }
1388 
1389 void
1390 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1391 	dmu_tx_t *tx)
1392 {
1393 	dnode_t *dn;
1394 
1395 	/* XXX assumes dnode_hold will not get an i/o error */
1396 	(void) dnode_hold(os, object, FTAG, &dn);
1397 	ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1398 	dn->dn_compress = compress;
1399 	dnode_setdirty(dn, tx);
1400 	dnode_rele(dn, FTAG);
1401 }
1402 
1403 int zfs_mdcomp_disable = 0;
1404 
1405 void
1406 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1407 {
1408 	dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1409 	boolean_t ismd = (level > 0 || dmu_ot[type].ot_metadata);
1410 	enum zio_checksum checksum = os->os_checksum;
1411 	enum zio_compress compress = os->os_compress;
1412 	enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1413 	boolean_t dedup;
1414 	boolean_t dedup_verify = os->os_dedup_verify;
1415 	int copies = os->os_copies;
1416 
1417 	/*
1418 	 * Determine checksum setting.
1419 	 */
1420 	if (ismd) {
1421 		/*
1422 		 * Metadata always gets checksummed.  If the data
1423 		 * checksum is multi-bit correctable, and it's not a
1424 		 * ZBT-style checksum, then it's suitable for metadata
1425 		 * as well.  Otherwise, the metadata checksum defaults
1426 		 * to fletcher4.
1427 		 */
1428 		if (zio_checksum_table[checksum].ci_correctable < 1 ||
1429 		    zio_checksum_table[checksum].ci_eck)
1430 			checksum = ZIO_CHECKSUM_FLETCHER_4;
1431 	} else {
1432 		checksum = zio_checksum_select(dn->dn_checksum, checksum);
1433 	}
1434 
1435 	/*
1436 	 * Determine compression setting.
1437 	 */
1438 	if (ismd) {
1439 		/*
1440 		 * XXX -- we should design a compression algorithm
1441 		 * that specializes in arrays of bps.
1442 		 */
1443 		compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1444 		    ZIO_COMPRESS_LZJB;
1445 	} else {
1446 		compress = zio_compress_select(dn->dn_compress, compress);
1447 	}
1448 
1449 	/*
1450 	 * Determine dedup setting.  If we are in dmu_sync(), we won't
1451 	 * actually dedup now because that's all done in syncing context;
1452 	 * but we do want to use the dedup checkum.  If the checksum is not
1453 	 * strong enough to ensure unique signatures, force dedup_verify.
1454 	 */
1455 	dedup = (!ismd && dedup_checksum != ZIO_CHECKSUM_OFF);
1456 	if (dedup) {
1457 		checksum = dedup_checksum;
1458 		if (!zio_checksum_table[checksum].ci_dedup)
1459 			dedup_verify = 1;
1460 	}
1461 
1462 	if (wp & WP_DMU_SYNC)
1463 		dedup = 0;
1464 
1465 	if (wp & WP_NOFILL) {
1466 		ASSERT(!ismd && level == 0);
1467 		checksum = ZIO_CHECKSUM_OFF;
1468 		compress = ZIO_COMPRESS_OFF;
1469 		dedup = B_FALSE;
1470 	}
1471 
1472 	zp->zp_checksum = checksum;
1473 	zp->zp_compress = compress;
1474 	zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1475 	zp->zp_level = level;
1476 	zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1477 	zp->zp_dedup = dedup;
1478 	zp->zp_dedup_verify = dedup && dedup_verify;
1479 }
1480 
1481 int
1482 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1483 {
1484 	dnode_t *dn;
1485 	int i, err;
1486 
1487 	err = dnode_hold(os, object, FTAG, &dn);
1488 	if (err)
1489 		return (err);
1490 	/*
1491 	 * Sync any current changes before
1492 	 * we go trundling through the block pointers.
1493 	 */
1494 	for (i = 0; i < TXG_SIZE; i++) {
1495 		if (list_link_active(&dn->dn_dirty_link[i]))
1496 			break;
1497 	}
1498 	if (i != TXG_SIZE) {
1499 		dnode_rele(dn, FTAG);
1500 		txg_wait_synced(dmu_objset_pool(os), 0);
1501 		err = dnode_hold(os, object, FTAG, &dn);
1502 		if (err)
1503 			return (err);
1504 	}
1505 
1506 	err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1507 	dnode_rele(dn, FTAG);
1508 
1509 	return (err);
1510 }
1511 
1512 void
1513 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1514 {
1515 	dnode_phys_t *dnp;
1516 
1517 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1518 	mutex_enter(&dn->dn_mtx);
1519 
1520 	dnp = dn->dn_phys;
1521 
1522 	doi->doi_data_block_size = dn->dn_datablksz;
1523 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
1524 	    1ULL << dn->dn_indblkshift : 0;
1525 	doi->doi_type = dn->dn_type;
1526 	doi->doi_bonus_type = dn->dn_bonustype;
1527 	doi->doi_bonus_size = dn->dn_bonuslen;
1528 	doi->doi_indirection = dn->dn_nlevels;
1529 	doi->doi_checksum = dn->dn_checksum;
1530 	doi->doi_compress = dn->dn_compress;
1531 	doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1532 	doi->doi_max_offset = (dnp->dn_maxblkid + 1) * dn->dn_datablksz;
1533 	doi->doi_fill_count = 0;
1534 	for (int i = 0; i < dnp->dn_nblkptr; i++)
1535 		doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1536 
1537 	mutex_exit(&dn->dn_mtx);
1538 	rw_exit(&dn->dn_struct_rwlock);
1539 }
1540 
1541 /*
1542  * Get information on a DMU object.
1543  * If doi is NULL, just indicates whether the object exists.
1544  */
1545 int
1546 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1547 {
1548 	dnode_t *dn;
1549 	int err = dnode_hold(os, object, FTAG, &dn);
1550 
1551 	if (err)
1552 		return (err);
1553 
1554 	if (doi != NULL)
1555 		dmu_object_info_from_dnode(dn, doi);
1556 
1557 	dnode_rele(dn, FTAG);
1558 	return (0);
1559 }
1560 
1561 /*
1562  * As above, but faster; can be used when you have a held dbuf in hand.
1563  */
1564 void
1565 dmu_object_info_from_db(dmu_buf_t *db, dmu_object_info_t *doi)
1566 {
1567 	dmu_object_info_from_dnode(((dmu_buf_impl_t *)db)->db_dnode, doi);
1568 }
1569 
1570 /*
1571  * Faster still when you only care about the size.
1572  * This is specifically optimized for zfs_getattr().
1573  */
1574 void
1575 dmu_object_size_from_db(dmu_buf_t *db, uint32_t *blksize, u_longlong_t *nblk512)
1576 {
1577 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1578 
1579 	*blksize = dn->dn_datablksz;
1580 	/* add 1 for dnode space */
1581 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1582 	    SPA_MINBLOCKSHIFT) + 1;
1583 }
1584 
1585 void
1586 byteswap_uint64_array(void *vbuf, size_t size)
1587 {
1588 	uint64_t *buf = vbuf;
1589 	size_t count = size >> 3;
1590 	int i;
1591 
1592 	ASSERT((size & 7) == 0);
1593 
1594 	for (i = 0; i < count; i++)
1595 		buf[i] = BSWAP_64(buf[i]);
1596 }
1597 
1598 void
1599 byteswap_uint32_array(void *vbuf, size_t size)
1600 {
1601 	uint32_t *buf = vbuf;
1602 	size_t count = size >> 2;
1603 	int i;
1604 
1605 	ASSERT((size & 3) == 0);
1606 
1607 	for (i = 0; i < count; i++)
1608 		buf[i] = BSWAP_32(buf[i]);
1609 }
1610 
1611 void
1612 byteswap_uint16_array(void *vbuf, size_t size)
1613 {
1614 	uint16_t *buf = vbuf;
1615 	size_t count = size >> 1;
1616 	int i;
1617 
1618 	ASSERT((size & 1) == 0);
1619 
1620 	for (i = 0; i < count; i++)
1621 		buf[i] = BSWAP_16(buf[i]);
1622 }
1623 
1624 /* ARGSUSED */
1625 void
1626 byteswap_uint8_array(void *vbuf, size_t size)
1627 {
1628 }
1629 
1630 void
1631 dmu_init(void)
1632 {
1633 	dbuf_init();
1634 	dnode_init();
1635 	zfetch_init();
1636 	arc_init();
1637 	l2arc_init();
1638 	xuio_stat_init();
1639 	sa_cache_init();
1640 }
1641 
1642 void
1643 dmu_fini(void)
1644 {
1645 	arc_fini();
1646 	zfetch_fini();
1647 	dnode_fini();
1648 	dbuf_fini();
1649 	l2arc_fini();
1650 	xuio_stat_fini();
1651 	sa_cache_fini();
1652 }
1653