xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu.c (revision 0ba964ea)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/dmu.h>
27 #include <sys/dmu_impl.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/dbuf.h>
30 #include <sys/dnode.h>
31 #include <sys/zfs_context.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dmu_traverse.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_pool.h>
37 #include <sys/dsl_synctask.h>
38 #include <sys/dsl_prop.h>
39 #include <sys/dmu_zfetch.h>
40 #include <sys/zfs_ioctl.h>
41 #include <sys/zap.h>
42 #include <sys/zio_checksum.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 };
91 
92 int
93 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
94     void *tag, dmu_buf_t **dbp)
95 {
96 	dnode_t *dn;
97 	uint64_t blkid;
98 	dmu_buf_impl_t *db;
99 	int err;
100 
101 	err = dnode_hold(os->os, object, FTAG, &dn);
102 	if (err)
103 		return (err);
104 	blkid = dbuf_whichblock(dn, offset);
105 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
106 	db = dbuf_hold(dn, blkid, tag);
107 	rw_exit(&dn->dn_struct_rwlock);
108 	if (db == NULL) {
109 		err = EIO;
110 	} else {
111 		err = dbuf_read(db, NULL, DB_RF_CANFAIL);
112 		if (err) {
113 			dbuf_rele(db, tag);
114 			db = NULL;
115 		}
116 	}
117 
118 	dnode_rele(dn, FTAG);
119 	*dbp = &db->db;
120 	return (err);
121 }
122 
123 int
124 dmu_bonus_max(void)
125 {
126 	return (DN_MAX_BONUSLEN);
127 }
128 
129 int
130 dmu_set_bonus(dmu_buf_t *db, int newsize, dmu_tx_t *tx)
131 {
132 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
133 
134 	if (dn->dn_bonus != (dmu_buf_impl_t *)db)
135 		return (EINVAL);
136 	if (newsize < 0 || newsize > db->db_size)
137 		return (EINVAL);
138 	dnode_setbonuslen(dn, newsize, tx);
139 	return (0);
140 }
141 
142 /*
143  * returns ENOENT, EIO, or 0.
144  */
145 int
146 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
147 {
148 	dnode_t *dn;
149 	dmu_buf_impl_t *db;
150 	int error;
151 
152 	error = dnode_hold(os->os, object, FTAG, &dn);
153 	if (error)
154 		return (error);
155 
156 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
157 	if (dn->dn_bonus == NULL) {
158 		rw_exit(&dn->dn_struct_rwlock);
159 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
160 		if (dn->dn_bonus == NULL)
161 			dbuf_create_bonus(dn);
162 	}
163 	db = dn->dn_bonus;
164 	rw_exit(&dn->dn_struct_rwlock);
165 
166 	/* as long as the bonus buf is held, the dnode will be held */
167 	if (refcount_add(&db->db_holds, tag) == 1)
168 		VERIFY(dnode_add_ref(dn, db));
169 
170 	dnode_rele(dn, FTAG);
171 
172 	VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED));
173 
174 	*dbp = &db->db;
175 	return (0);
176 }
177 
178 /*
179  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
180  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
181  * and can induce severe lock contention when writing to several files
182  * whose dnodes are in the same block.
183  */
184 static int
185 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
186     int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
187 {
188 	dsl_pool_t *dp = NULL;
189 	dmu_buf_t **dbp;
190 	uint64_t blkid, nblks, i;
191 	uint32_t dbuf_flags;
192 	int err;
193 	zio_t *zio;
194 	hrtime_t start;
195 
196 	ASSERT(length <= DMU_MAX_ACCESS);
197 
198 	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT;
199 	if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
200 		dbuf_flags |= DB_RF_NOPREFETCH;
201 
202 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
203 	if (dn->dn_datablkshift) {
204 		int blkshift = dn->dn_datablkshift;
205 		nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
206 		    P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
207 	} else {
208 		if (offset + length > dn->dn_datablksz) {
209 			zfs_panic_recover("zfs: accessing past end of object "
210 			    "%llx/%llx (size=%u access=%llu+%llu)",
211 			    (longlong_t)dn->dn_objset->
212 			    os_dsl_dataset->ds_object,
213 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
214 			    (longlong_t)offset, (longlong_t)length);
215 			return (EIO);
216 		}
217 		nblks = 1;
218 	}
219 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
220 
221 	if (dn->dn_objset->os_dsl_dataset)
222 		dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool;
223 	if (dp && dsl_pool_sync_context(dp))
224 		start = gethrtime();
225 	zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
226 	blkid = dbuf_whichblock(dn, offset);
227 	for (i = 0; i < nblks; i++) {
228 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
229 		if (db == NULL) {
230 			rw_exit(&dn->dn_struct_rwlock);
231 			dmu_buf_rele_array(dbp, nblks, tag);
232 			zio_nowait(zio);
233 			return (EIO);
234 		}
235 		/* initiate async i/o */
236 		if (read) {
237 			rw_exit(&dn->dn_struct_rwlock);
238 			(void) dbuf_read(db, zio, dbuf_flags);
239 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
240 		}
241 		dbp[i] = &db->db;
242 	}
243 	rw_exit(&dn->dn_struct_rwlock);
244 
245 	/* wait for async i/o */
246 	err = zio_wait(zio);
247 	/* track read overhead when we are in sync context */
248 	if (dp && dsl_pool_sync_context(dp))
249 		dp->dp_read_overhead += gethrtime() - start;
250 	if (err) {
251 		dmu_buf_rele_array(dbp, nblks, tag);
252 		return (err);
253 	}
254 
255 	/* wait for other io to complete */
256 	if (read) {
257 		for (i = 0; i < nblks; i++) {
258 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
259 			mutex_enter(&db->db_mtx);
260 			while (db->db_state == DB_READ ||
261 			    db->db_state == DB_FILL)
262 				cv_wait(&db->db_changed, &db->db_mtx);
263 			if (db->db_state == DB_UNCACHED)
264 				err = EIO;
265 			mutex_exit(&db->db_mtx);
266 			if (err) {
267 				dmu_buf_rele_array(dbp, nblks, tag);
268 				return (err);
269 			}
270 		}
271 	}
272 
273 	*numbufsp = nblks;
274 	*dbpp = dbp;
275 	return (0);
276 }
277 
278 static int
279 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
280     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
281 {
282 	dnode_t *dn;
283 	int err;
284 
285 	err = dnode_hold(os->os, object, FTAG, &dn);
286 	if (err)
287 		return (err);
288 
289 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
290 	    numbufsp, dbpp, DMU_READ_PREFETCH);
291 
292 	dnode_rele(dn, FTAG);
293 
294 	return (err);
295 }
296 
297 int
298 dmu_buf_hold_array_by_bonus(dmu_buf_t *db, uint64_t offset,
299     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
300 {
301 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
302 	int err;
303 
304 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
305 	    numbufsp, dbpp, DMU_READ_PREFETCH);
306 
307 	return (err);
308 }
309 
310 void
311 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
312 {
313 	int i;
314 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
315 
316 	if (numbufs == 0)
317 		return;
318 
319 	for (i = 0; i < numbufs; i++) {
320 		if (dbp[i])
321 			dbuf_rele(dbp[i], tag);
322 	}
323 
324 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
325 }
326 
327 void
328 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
329 {
330 	dnode_t *dn;
331 	uint64_t blkid;
332 	int nblks, i, err;
333 
334 	if (zfs_prefetch_disable)
335 		return;
336 
337 	if (len == 0) {  /* they're interested in the bonus buffer */
338 		dn = os->os->os_meta_dnode;
339 
340 		if (object == 0 || object >= DN_MAX_OBJECT)
341 			return;
342 
343 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
344 		blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
345 		dbuf_prefetch(dn, blkid);
346 		rw_exit(&dn->dn_struct_rwlock);
347 		return;
348 	}
349 
350 	/*
351 	 * XXX - Note, if the dnode for the requested object is not
352 	 * already cached, we will do a *synchronous* read in the
353 	 * dnode_hold() call.  The same is true for any indirects.
354 	 */
355 	err = dnode_hold(os->os, object, FTAG, &dn);
356 	if (err != 0)
357 		return;
358 
359 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
360 	if (dn->dn_datablkshift) {
361 		int blkshift = dn->dn_datablkshift;
362 		nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
363 		    P2ALIGN(offset, 1<<blkshift)) >> blkshift;
364 	} else {
365 		nblks = (offset < dn->dn_datablksz);
366 	}
367 
368 	if (nblks != 0) {
369 		blkid = dbuf_whichblock(dn, offset);
370 		for (i = 0; i < nblks; i++)
371 			dbuf_prefetch(dn, blkid+i);
372 	}
373 
374 	rw_exit(&dn->dn_struct_rwlock);
375 
376 	dnode_rele(dn, FTAG);
377 }
378 
379 /*
380  * Get the next "chunk" of file data to free.  We traverse the file from
381  * the end so that the file gets shorter over time (if we crashes in the
382  * middle, this will leave us in a better state).  We find allocated file
383  * data by simply searching the allocated level 1 indirects.
384  */
385 static int
386 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit)
387 {
388 	uint64_t len = *start - limit;
389 	uint64_t blkcnt = 0;
390 	uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1));
391 	uint64_t iblkrange =
392 	    dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
393 
394 	ASSERT(limit <= *start);
395 
396 	if (len <= iblkrange * maxblks) {
397 		*start = limit;
398 		return (0);
399 	}
400 	ASSERT(ISP2(iblkrange));
401 
402 	while (*start > limit && blkcnt < maxblks) {
403 		int err;
404 
405 		/* find next allocated L1 indirect */
406 		err = dnode_next_offset(dn,
407 		    DNODE_FIND_BACKWARDS, start, 2, 1, 0);
408 
409 		/* if there are no more, then we are done */
410 		if (err == ESRCH) {
411 			*start = limit;
412 			return (0);
413 		} else if (err) {
414 			return (err);
415 		}
416 		blkcnt += 1;
417 
418 		/* reset offset to end of "next" block back */
419 		*start = P2ALIGN(*start, iblkrange);
420 		if (*start <= limit)
421 			*start = limit;
422 		else
423 			*start -= 1;
424 	}
425 	return (0);
426 }
427 
428 static int
429 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
430     uint64_t length, boolean_t free_dnode)
431 {
432 	dmu_tx_t *tx;
433 	uint64_t object_size, start, end, len;
434 	boolean_t trunc = (length == DMU_OBJECT_END);
435 	int align, err;
436 
437 	align = 1 << dn->dn_datablkshift;
438 	ASSERT(align > 0);
439 	object_size = align == 1 ? dn->dn_datablksz :
440 	    (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
441 
442 	end = offset + length;
443 	if (trunc || end > object_size)
444 		end = object_size;
445 	if (end <= offset)
446 		return (0);
447 	length = end - offset;
448 
449 	while (length) {
450 		start = end;
451 		/* assert(offset <= start) */
452 		err = get_next_chunk(dn, &start, offset);
453 		if (err)
454 			return (err);
455 		len = trunc ? DMU_OBJECT_END : end - start;
456 
457 		tx = dmu_tx_create(os);
458 		dmu_tx_hold_free(tx, dn->dn_object, start, len);
459 		err = dmu_tx_assign(tx, TXG_WAIT);
460 		if (err) {
461 			dmu_tx_abort(tx);
462 			return (err);
463 		}
464 
465 		dnode_free_range(dn, start, trunc ? -1 : len, tx);
466 
467 		if (start == 0 && free_dnode) {
468 			ASSERT(trunc);
469 			dnode_free(dn, tx);
470 		}
471 
472 		length -= end - start;
473 
474 		dmu_tx_commit(tx);
475 		end = start;
476 	}
477 	return (0);
478 }
479 
480 int
481 dmu_free_long_range(objset_t *os, uint64_t object,
482     uint64_t offset, uint64_t length)
483 {
484 	dnode_t *dn;
485 	int err;
486 
487 	err = dnode_hold(os->os, object, FTAG, &dn);
488 	if (err != 0)
489 		return (err);
490 	err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
491 	dnode_rele(dn, FTAG);
492 	return (err);
493 }
494 
495 int
496 dmu_free_object(objset_t *os, uint64_t object)
497 {
498 	dnode_t *dn;
499 	dmu_tx_t *tx;
500 	int err;
501 
502 	err = dnode_hold_impl(os->os, object, DNODE_MUST_BE_ALLOCATED,
503 	    FTAG, &dn);
504 	if (err != 0)
505 		return (err);
506 	if (dn->dn_nlevels == 1) {
507 		tx = dmu_tx_create(os);
508 		dmu_tx_hold_bonus(tx, object);
509 		dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
510 		err = dmu_tx_assign(tx, TXG_WAIT);
511 		if (err == 0) {
512 			dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
513 			dnode_free(dn, tx);
514 			dmu_tx_commit(tx);
515 		} else {
516 			dmu_tx_abort(tx);
517 		}
518 	} else {
519 		err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
520 	}
521 	dnode_rele(dn, FTAG);
522 	return (err);
523 }
524 
525 int
526 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
527     uint64_t size, dmu_tx_t *tx)
528 {
529 	dnode_t *dn;
530 	int err = dnode_hold(os->os, object, FTAG, &dn);
531 	if (err)
532 		return (err);
533 	ASSERT(offset < UINT64_MAX);
534 	ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
535 	dnode_free_range(dn, offset, size, tx);
536 	dnode_rele(dn, FTAG);
537 	return (0);
538 }
539 
540 int
541 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
542     void *buf, uint32_t flags)
543 {
544 	dnode_t *dn;
545 	dmu_buf_t **dbp;
546 	int numbufs, i, err;
547 
548 	err = dnode_hold(os->os, object, FTAG, &dn);
549 	if (err)
550 		return (err);
551 
552 	/*
553 	 * Deal with odd block sizes, where there can't be data past the first
554 	 * block.  If we ever do the tail block optimization, we will need to
555 	 * handle that here as well.
556 	 */
557 	if (dn->dn_datablkshift == 0) {
558 		int newsz = offset > dn->dn_datablksz ? 0 :
559 		    MIN(size, dn->dn_datablksz - offset);
560 		bzero((char *)buf + newsz, size - newsz);
561 		size = newsz;
562 	}
563 
564 	while (size > 0) {
565 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
566 
567 		/*
568 		 * NB: we could do this block-at-a-time, but it's nice
569 		 * to be reading in parallel.
570 		 */
571 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
572 		    TRUE, FTAG, &numbufs, &dbp, flags);
573 		if (err)
574 			break;
575 
576 		for (i = 0; i < numbufs; i++) {
577 			int tocpy;
578 			int bufoff;
579 			dmu_buf_t *db = dbp[i];
580 
581 			ASSERT(size > 0);
582 
583 			bufoff = offset - db->db_offset;
584 			tocpy = (int)MIN(db->db_size - bufoff, size);
585 
586 			bcopy((char *)db->db_data + bufoff, buf, tocpy);
587 
588 			offset += tocpy;
589 			size -= tocpy;
590 			buf = (char *)buf + tocpy;
591 		}
592 		dmu_buf_rele_array(dbp, numbufs, FTAG);
593 	}
594 	dnode_rele(dn, FTAG);
595 	return (err);
596 }
597 
598 void
599 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
600     const void *buf, dmu_tx_t *tx)
601 {
602 	dmu_buf_t **dbp;
603 	int numbufs, i;
604 
605 	if (size == 0)
606 		return;
607 
608 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
609 	    FALSE, FTAG, &numbufs, &dbp));
610 
611 	for (i = 0; i < numbufs; i++) {
612 		int tocpy;
613 		int bufoff;
614 		dmu_buf_t *db = dbp[i];
615 
616 		ASSERT(size > 0);
617 
618 		bufoff = offset - db->db_offset;
619 		tocpy = (int)MIN(db->db_size - bufoff, size);
620 
621 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
622 
623 		if (tocpy == db->db_size)
624 			dmu_buf_will_fill(db, tx);
625 		else
626 			dmu_buf_will_dirty(db, tx);
627 
628 		bcopy(buf, (char *)db->db_data + bufoff, tocpy);
629 
630 		if (tocpy == db->db_size)
631 			dmu_buf_fill_done(db, tx);
632 
633 		offset += tocpy;
634 		size -= tocpy;
635 		buf = (char *)buf + tocpy;
636 	}
637 	dmu_buf_rele_array(dbp, numbufs, FTAG);
638 }
639 
640 void
641 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
642     dmu_tx_t *tx)
643 {
644 	dmu_buf_t **dbp;
645 	int numbufs, i;
646 
647 	if (size == 0)
648 		return;
649 
650 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
651 	    FALSE, FTAG, &numbufs, &dbp));
652 
653 	for (i = 0; i < numbufs; i++) {
654 		dmu_buf_t *db = dbp[i];
655 
656 		dmu_buf_will_not_fill(db, tx);
657 	}
658 	dmu_buf_rele_array(dbp, numbufs, FTAG);
659 }
660 
661 #ifdef _KERNEL
662 int
663 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
664 {
665 	dmu_buf_t **dbp;
666 	int numbufs, i, err;
667 
668 	/*
669 	 * NB: we could do this block-at-a-time, but it's nice
670 	 * to be reading in parallel.
671 	 */
672 	err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
673 	    &numbufs, &dbp);
674 	if (err)
675 		return (err);
676 
677 	for (i = 0; i < numbufs; i++) {
678 		int tocpy;
679 		int bufoff;
680 		dmu_buf_t *db = dbp[i];
681 
682 		ASSERT(size > 0);
683 
684 		bufoff = uio->uio_loffset - db->db_offset;
685 		tocpy = (int)MIN(db->db_size - bufoff, size);
686 
687 		err = uiomove((char *)db->db_data + bufoff, tocpy,
688 		    UIO_READ, uio);
689 		if (err)
690 			break;
691 
692 		size -= tocpy;
693 	}
694 	dmu_buf_rele_array(dbp, numbufs, FTAG);
695 
696 	return (err);
697 }
698 
699 int
700 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
701     dmu_tx_t *tx)
702 {
703 	dmu_buf_t **dbp;
704 	int numbufs, i;
705 	int err = 0;
706 
707 	if (size == 0)
708 		return (0);
709 
710 	err = dmu_buf_hold_array(os, object, uio->uio_loffset, size,
711 	    FALSE, FTAG, &numbufs, &dbp);
712 	if (err)
713 		return (err);
714 
715 	for (i = 0; i < numbufs; i++) {
716 		int tocpy;
717 		int bufoff;
718 		dmu_buf_t *db = dbp[i];
719 
720 		ASSERT(size > 0);
721 
722 		bufoff = uio->uio_loffset - db->db_offset;
723 		tocpy = (int)MIN(db->db_size - bufoff, size);
724 
725 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
726 
727 		if (tocpy == db->db_size)
728 			dmu_buf_will_fill(db, tx);
729 		else
730 			dmu_buf_will_dirty(db, tx);
731 
732 		/*
733 		 * XXX uiomove could block forever (eg. nfs-backed
734 		 * pages).  There needs to be a uiolockdown() function
735 		 * to lock the pages in memory, so that uiomove won't
736 		 * block.
737 		 */
738 		err = uiomove((char *)db->db_data + bufoff, tocpy,
739 		    UIO_WRITE, uio);
740 
741 		if (tocpy == db->db_size)
742 			dmu_buf_fill_done(db, tx);
743 
744 		if (err)
745 			break;
746 
747 		size -= tocpy;
748 	}
749 	dmu_buf_rele_array(dbp, numbufs, FTAG);
750 	return (err);
751 }
752 
753 int
754 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
755     page_t *pp, dmu_tx_t *tx)
756 {
757 	dmu_buf_t **dbp;
758 	int numbufs, i;
759 	int err;
760 
761 	if (size == 0)
762 		return (0);
763 
764 	err = dmu_buf_hold_array(os, object, offset, size,
765 	    FALSE, FTAG, &numbufs, &dbp);
766 	if (err)
767 		return (err);
768 
769 	for (i = 0; i < numbufs; i++) {
770 		int tocpy, copied, thiscpy;
771 		int bufoff;
772 		dmu_buf_t *db = dbp[i];
773 		caddr_t va;
774 
775 		ASSERT(size > 0);
776 		ASSERT3U(db->db_size, >=, PAGESIZE);
777 
778 		bufoff = offset - db->db_offset;
779 		tocpy = (int)MIN(db->db_size - bufoff, size);
780 
781 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
782 
783 		if (tocpy == db->db_size)
784 			dmu_buf_will_fill(db, tx);
785 		else
786 			dmu_buf_will_dirty(db, tx);
787 
788 		for (copied = 0; copied < tocpy; copied += PAGESIZE) {
789 			ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
790 			thiscpy = MIN(PAGESIZE, tocpy - copied);
791 			va = zfs_map_page(pp, S_READ);
792 			bcopy(va, (char *)db->db_data + bufoff, thiscpy);
793 			zfs_unmap_page(pp, va);
794 			pp = pp->p_next;
795 			bufoff += PAGESIZE;
796 		}
797 
798 		if (tocpy == db->db_size)
799 			dmu_buf_fill_done(db, tx);
800 
801 		offset += tocpy;
802 		size -= tocpy;
803 	}
804 	dmu_buf_rele_array(dbp, numbufs, FTAG);
805 	return (err);
806 }
807 #endif
808 
809 /*
810  * Allocate a loaned anonymous arc buffer.
811  */
812 arc_buf_t *
813 dmu_request_arcbuf(dmu_buf_t *handle, int size)
814 {
815 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
816 
817 	return (arc_loan_buf(dn->dn_objset->os_spa, size));
818 }
819 
820 /*
821  * Free a loaned arc buffer.
822  */
823 void
824 dmu_return_arcbuf(arc_buf_t *buf)
825 {
826 	arc_return_buf(buf, FTAG);
827 	VERIFY(arc_buf_remove_ref(buf, FTAG) == 1);
828 }
829 
830 /*
831  * When possible directly assign passed loaned arc buffer to a dbuf.
832  * If this is not possible copy the contents of passed arc buf via
833  * dmu_write().
834  */
835 void
836 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
837     dmu_tx_t *tx)
838 {
839 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
840 	dmu_buf_impl_t *db;
841 	uint32_t blksz = (uint32_t)arc_buf_size(buf);
842 	uint64_t blkid;
843 
844 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
845 	blkid = dbuf_whichblock(dn, offset);
846 	VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
847 	rw_exit(&dn->dn_struct_rwlock);
848 
849 	if (offset == db->db.db_offset && blksz == db->db.db_size) {
850 		dbuf_assign_arcbuf(db, buf, tx);
851 		dbuf_rele(db, FTAG);
852 	} else {
853 		dbuf_rele(db, FTAG);
854 		ASSERT(dn->dn_objset->os.os == dn->dn_objset);
855 		dmu_write(&dn->dn_objset->os, dn->dn_object, offset, blksz,
856 		    buf->b_data, tx);
857 		dmu_return_arcbuf(buf);
858 	}
859 }
860 
861 typedef struct {
862 	dbuf_dirty_record_t	*dr;
863 	dmu_sync_cb_t		*done;
864 	void			*arg;
865 } dmu_sync_arg_t;
866 
867 /* ARGSUSED */
868 static void
869 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
870 {
871 	blkptr_t *bp = zio->io_bp;
872 
873 	if (!BP_IS_HOLE(bp)) {
874 		dmu_sync_arg_t *in = varg;
875 		dbuf_dirty_record_t *dr = in->dr;
876 		dmu_buf_impl_t *db = dr->dr_dbuf;
877 		ASSERT(BP_GET_TYPE(bp) == db->db_dnode->dn_type);
878 		ASSERT(BP_GET_LEVEL(bp) == 0);
879 		bp->blk_fill = 1;
880 	}
881 }
882 
883 /* ARGSUSED */
884 static void
885 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
886 {
887 	dmu_sync_arg_t *in = varg;
888 	dbuf_dirty_record_t *dr = in->dr;
889 	dmu_buf_impl_t *db = dr->dr_dbuf;
890 	dmu_sync_cb_t *done = in->done;
891 
892 	mutex_enter(&db->db_mtx);
893 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
894 	dr->dt.dl.dr_overridden_by = *zio->io_bp; /* structure assignment */
895 	dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
896 	cv_broadcast(&db->db_changed);
897 	mutex_exit(&db->db_mtx);
898 
899 	if (done)
900 		done(&(db->db), in->arg);
901 
902 	kmem_free(in, sizeof (dmu_sync_arg_t));
903 }
904 
905 /*
906  * Intent log support: sync the block associated with db to disk.
907  * N.B. and XXX: the caller is responsible for making sure that the
908  * data isn't changing while dmu_sync() is writing it.
909  *
910  * Return values:
911  *
912  *	EEXIST: this txg has already been synced, so there's nothing to to.
913  *		The caller should not log the write.
914  *
915  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
916  *		The caller should not log the write.
917  *
918  *	EALREADY: this block is already in the process of being synced.
919  *		The caller should track its progress (somehow).
920  *
921  *	EINPROGRESS: the IO has been initiated.
922  *		The caller should log this blkptr in the callback.
923  *
924  *	0: completed.  Sets *bp to the blkptr just written.
925  *		The caller should log this blkptr immediately.
926  */
927 int
928 dmu_sync(zio_t *pio, dmu_buf_t *db_fake,
929     blkptr_t *bp, uint64_t txg, dmu_sync_cb_t *done, void *arg)
930 {
931 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
932 	objset_impl_t *os = db->db_objset;
933 	dsl_pool_t *dp = os->os_dsl_dataset->ds_dir->dd_pool;
934 	tx_state_t *tx = &dp->dp_tx;
935 	dbuf_dirty_record_t *dr;
936 	dmu_sync_arg_t *in;
937 	zbookmark_t zb;
938 	writeprops_t wp = { 0 };
939 	zio_t *zio;
940 	int err;
941 
942 	ASSERT(BP_IS_HOLE(bp));
943 	ASSERT(txg != 0);
944 
945 	dprintf("dmu_sync txg=%llu, s,o,q %llu %llu %llu\n",
946 	    txg, tx->tx_synced_txg, tx->tx_open_txg, tx->tx_quiesced_txg);
947 
948 	/*
949 	 * XXX - would be nice if we could do this without suspending...
950 	 */
951 	txg_suspend(dp);
952 
953 	/*
954 	 * If this txg already synced, there's nothing to do.
955 	 */
956 	if (txg <= tx->tx_synced_txg) {
957 		txg_resume(dp);
958 		/*
959 		 * If we're running ziltest, we need the blkptr regardless.
960 		 */
961 		if (txg > spa_freeze_txg(dp->dp_spa)) {
962 			/* if db_blkptr == NULL, this was an empty write */
963 			if (db->db_blkptr)
964 				*bp = *db->db_blkptr; /* structure assignment */
965 			return (0);
966 		}
967 		return (EEXIST);
968 	}
969 
970 	mutex_enter(&db->db_mtx);
971 
972 	if (txg == tx->tx_syncing_txg) {
973 		while (db->db_data_pending) {
974 			/*
975 			 * IO is in-progress.  Wait for it to finish.
976 			 * XXX - would be nice to be able to somehow "attach"
977 			 * this zio to the parent zio passed in.
978 			 */
979 			cv_wait(&db->db_changed, &db->db_mtx);
980 			if (!db->db_data_pending &&
981 			    db->db_blkptr && BP_IS_HOLE(db->db_blkptr)) {
982 				/*
983 				 * IO was compressed away
984 				 */
985 				*bp = *db->db_blkptr; /* structure assignment */
986 				mutex_exit(&db->db_mtx);
987 				txg_resume(dp);
988 				return (0);
989 			}
990 			ASSERT(db->db_data_pending ||
991 			    (db->db_blkptr && db->db_blkptr->blk_birth == txg));
992 		}
993 
994 		if (db->db_blkptr && db->db_blkptr->blk_birth == txg) {
995 			/*
996 			 * IO is already completed.
997 			 */
998 			*bp = *db->db_blkptr; /* structure assignment */
999 			mutex_exit(&db->db_mtx);
1000 			txg_resume(dp);
1001 			return (0);
1002 		}
1003 	}
1004 
1005 	dr = db->db_last_dirty;
1006 	while (dr && dr->dr_txg > txg)
1007 		dr = dr->dr_next;
1008 	if (dr == NULL || dr->dr_txg < txg) {
1009 		/*
1010 		 * This dbuf isn't dirty, must have been free_range'd.
1011 		 * There's no need to log writes to freed blocks, so we're done.
1012 		 */
1013 		mutex_exit(&db->db_mtx);
1014 		txg_resume(dp);
1015 		return (ENOENT);
1016 	}
1017 
1018 	ASSERT(dr->dr_txg == txg);
1019 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
1020 		/*
1021 		 * We have already issued a sync write for this buffer.
1022 		 */
1023 		mutex_exit(&db->db_mtx);
1024 		txg_resume(dp);
1025 		return (EALREADY);
1026 	} else if (dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1027 		/*
1028 		 * This buffer has already been synced.  It could not
1029 		 * have been dirtied since, or we would have cleared the state.
1030 		 */
1031 		*bp = dr->dt.dl.dr_overridden_by; /* structure assignment */
1032 		mutex_exit(&db->db_mtx);
1033 		txg_resume(dp);
1034 		return (0);
1035 	}
1036 
1037 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1038 	in = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1039 	in->dr = dr;
1040 	in->done = done;
1041 	in->arg = arg;
1042 	mutex_exit(&db->db_mtx);
1043 	txg_resume(dp);
1044 
1045 	zb.zb_objset = os->os_dsl_dataset->ds_object;
1046 	zb.zb_object = db->db.db_object;
1047 	zb.zb_level = db->db_level;
1048 	zb.zb_blkid = db->db_blkid;
1049 
1050 	wp.wp_type = db->db_dnode->dn_type;
1051 	wp.wp_level = db->db_level;
1052 	wp.wp_copies = os->os_copies;
1053 	wp.wp_dnchecksum = db->db_dnode->dn_checksum;
1054 	wp.wp_oschecksum = os->os_checksum;
1055 	wp.wp_dncompress = db->db_dnode->dn_compress;
1056 	wp.wp_oscompress = os->os_compress;
1057 
1058 	ASSERT(BP_IS_HOLE(bp));
1059 
1060 	zio = arc_write(pio, os->os_spa, &wp, DBUF_IS_L2CACHEABLE(db),
1061 	    txg, bp, dr->dt.dl.dr_data, dmu_sync_ready, dmu_sync_done, in,
1062 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
1063 	if (pio) {
1064 		zio_nowait(zio);
1065 		err = EINPROGRESS;
1066 	} else {
1067 		err = zio_wait(zio);
1068 		ASSERT(err == 0);
1069 	}
1070 	return (err);
1071 }
1072 
1073 int
1074 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1075 	dmu_tx_t *tx)
1076 {
1077 	dnode_t *dn;
1078 	int err;
1079 
1080 	err = dnode_hold(os->os, object, FTAG, &dn);
1081 	if (err)
1082 		return (err);
1083 	err = dnode_set_blksz(dn, size, ibs, tx);
1084 	dnode_rele(dn, FTAG);
1085 	return (err);
1086 }
1087 
1088 void
1089 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1090 	dmu_tx_t *tx)
1091 {
1092 	dnode_t *dn;
1093 
1094 	/* XXX assumes dnode_hold will not get an i/o error */
1095 	(void) dnode_hold(os->os, object, FTAG, &dn);
1096 	ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1097 	dn->dn_checksum = checksum;
1098 	dnode_setdirty(dn, tx);
1099 	dnode_rele(dn, FTAG);
1100 }
1101 
1102 void
1103 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1104 	dmu_tx_t *tx)
1105 {
1106 	dnode_t *dn;
1107 
1108 	/* XXX assumes dnode_hold will not get an i/o error */
1109 	(void) dnode_hold(os->os, object, FTAG, &dn);
1110 	ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1111 	dn->dn_compress = compress;
1112 	dnode_setdirty(dn, tx);
1113 	dnode_rele(dn, FTAG);
1114 }
1115 
1116 int
1117 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1118 {
1119 	dnode_t *dn;
1120 	int i, err;
1121 
1122 	err = dnode_hold(os->os, object, FTAG, &dn);
1123 	if (err)
1124 		return (err);
1125 	/*
1126 	 * Sync any current changes before
1127 	 * we go trundling through the block pointers.
1128 	 */
1129 	for (i = 0; i < TXG_SIZE; i++) {
1130 		if (list_link_active(&dn->dn_dirty_link[i]))
1131 			break;
1132 	}
1133 	if (i != TXG_SIZE) {
1134 		dnode_rele(dn, FTAG);
1135 		txg_wait_synced(dmu_objset_pool(os), 0);
1136 		err = dnode_hold(os->os, object, FTAG, &dn);
1137 		if (err)
1138 			return (err);
1139 	}
1140 
1141 	err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1142 	dnode_rele(dn, FTAG);
1143 
1144 	return (err);
1145 }
1146 
1147 void
1148 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1149 {
1150 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1151 	mutex_enter(&dn->dn_mtx);
1152 
1153 	doi->doi_data_block_size = dn->dn_datablksz;
1154 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
1155 	    1ULL << dn->dn_indblkshift : 0;
1156 	doi->doi_indirection = dn->dn_nlevels;
1157 	doi->doi_checksum = dn->dn_checksum;
1158 	doi->doi_compress = dn->dn_compress;
1159 	doi->doi_physical_blks = (DN_USED_BYTES(dn->dn_phys) +
1160 	    SPA_MINBLOCKSIZE/2) >> SPA_MINBLOCKSHIFT;
1161 	doi->doi_max_block_offset = dn->dn_phys->dn_maxblkid;
1162 	doi->doi_type = dn->dn_type;
1163 	doi->doi_bonus_size = dn->dn_bonuslen;
1164 	doi->doi_bonus_type = dn->dn_bonustype;
1165 
1166 	mutex_exit(&dn->dn_mtx);
1167 	rw_exit(&dn->dn_struct_rwlock);
1168 }
1169 
1170 /*
1171  * Get information on a DMU object.
1172  * If doi is NULL, just indicates whether the object exists.
1173  */
1174 int
1175 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1176 {
1177 	dnode_t *dn;
1178 	int err = dnode_hold(os->os, object, FTAG, &dn);
1179 
1180 	if (err)
1181 		return (err);
1182 
1183 	if (doi != NULL)
1184 		dmu_object_info_from_dnode(dn, doi);
1185 
1186 	dnode_rele(dn, FTAG);
1187 	return (0);
1188 }
1189 
1190 /*
1191  * As above, but faster; can be used when you have a held dbuf in hand.
1192  */
1193 void
1194 dmu_object_info_from_db(dmu_buf_t *db, dmu_object_info_t *doi)
1195 {
1196 	dmu_object_info_from_dnode(((dmu_buf_impl_t *)db)->db_dnode, doi);
1197 }
1198 
1199 /*
1200  * Faster still when you only care about the size.
1201  * This is specifically optimized for zfs_getattr().
1202  */
1203 void
1204 dmu_object_size_from_db(dmu_buf_t *db, uint32_t *blksize, u_longlong_t *nblk512)
1205 {
1206 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1207 
1208 	*blksize = dn->dn_datablksz;
1209 	/* add 1 for dnode space */
1210 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1211 	    SPA_MINBLOCKSHIFT) + 1;
1212 }
1213 
1214 void
1215 byteswap_uint64_array(void *vbuf, size_t size)
1216 {
1217 	uint64_t *buf = vbuf;
1218 	size_t count = size >> 3;
1219 	int i;
1220 
1221 	ASSERT((size & 7) == 0);
1222 
1223 	for (i = 0; i < count; i++)
1224 		buf[i] = BSWAP_64(buf[i]);
1225 }
1226 
1227 void
1228 byteswap_uint32_array(void *vbuf, size_t size)
1229 {
1230 	uint32_t *buf = vbuf;
1231 	size_t count = size >> 2;
1232 	int i;
1233 
1234 	ASSERT((size & 3) == 0);
1235 
1236 	for (i = 0; i < count; i++)
1237 		buf[i] = BSWAP_32(buf[i]);
1238 }
1239 
1240 void
1241 byteswap_uint16_array(void *vbuf, size_t size)
1242 {
1243 	uint16_t *buf = vbuf;
1244 	size_t count = size >> 1;
1245 	int i;
1246 
1247 	ASSERT((size & 1) == 0);
1248 
1249 	for (i = 0; i < count; i++)
1250 		buf[i] = BSWAP_16(buf[i]);
1251 }
1252 
1253 /* ARGSUSED */
1254 void
1255 byteswap_uint8_array(void *vbuf, size_t size)
1256 {
1257 }
1258 
1259 void
1260 dmu_init(void)
1261 {
1262 	dbuf_init();
1263 	dnode_init();
1264 	arc_init();
1265 	l2arc_init();
1266 }
1267 
1268 void
1269 dmu_fini(void)
1270 {
1271 	arc_fini();
1272 	dnode_fini();
1273 	dbuf_fini();
1274 	l2arc_fini();
1275 }
1276