xref: /titanic_52/usr/src/uts/common/fs/zfs/dmu.c (revision ff0e937b36dcde1a47ff7b00aa76a491c0dc07a8)
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,
186     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
187 {
188 	dsl_pool_t *dp = NULL;
189 	dmu_buf_t **dbp;
190 	uint64_t blkid, nblks, i;
191 	uint32_t flags;
192 	int err;
193 	zio_t *zio;
194 	hrtime_t start;
195 
196 	ASSERT(length <= DMU_MAX_ACCESS);
197 
198 	flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT;
199 	if (length > zfetch_array_rd_sz)
200 		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, 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);
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);
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 static int
380 get_next_chunk(dnode_t *dn, uint64_t *offset, uint64_t limit)
381 {
382 	uint64_t len = *offset - limit;
383 	uint64_t chunk_len = dn->dn_datablksz * DMU_MAX_DELETEBLKCNT;
384 	uint64_t subchunk =
385 	    dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
386 
387 	ASSERT(limit <= *offset);
388 
389 	if (len <= chunk_len) {
390 		*offset = limit;
391 		return (0);
392 	}
393 
394 	ASSERT(ISP2(subchunk));
395 
396 	while (*offset > limit) {
397 		uint64_t initial_offset = P2ROUNDUP(*offset, subchunk);
398 		uint64_t delta;
399 		int err;
400 
401 		/* skip over allocated data */
402 		err = dnode_next_offset(dn,
403 		    DNODE_FIND_HOLE|DNODE_FIND_BACKWARDS, offset, 1, 1, 0);
404 		if (err == ESRCH)
405 			*offset = limit;
406 		else if (err)
407 			return (err);
408 
409 		ASSERT3U(*offset, <=, initial_offset);
410 		*offset = P2ALIGN(*offset, subchunk);
411 		delta = initial_offset - *offset;
412 		if (delta >= chunk_len) {
413 			*offset += delta - chunk_len;
414 			return (0);
415 		}
416 		chunk_len -= delta;
417 
418 		/* skip over unallocated data */
419 		err = dnode_next_offset(dn,
420 		    DNODE_FIND_BACKWARDS, offset, 1, 1, 0);
421 		if (err == ESRCH)
422 			*offset = limit;
423 		else if (err)
424 			return (err);
425 
426 		if (*offset < limit)
427 			*offset = limit;
428 		ASSERT3U(*offset, <, initial_offset);
429 	}
430 	return (0);
431 }
432 
433 static int
434 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
435     uint64_t length, boolean_t free_dnode)
436 {
437 	dmu_tx_t *tx;
438 	uint64_t object_size, start, end, len;
439 	boolean_t trunc = (length == DMU_OBJECT_END);
440 	int align, err;
441 
442 	align = 1 << dn->dn_datablkshift;
443 	ASSERT(align > 0);
444 	object_size = align == 1 ? dn->dn_datablksz :
445 	    (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
446 
447 	end = offset + length;
448 	if (trunc || end > object_size)
449 		end = object_size;
450 	if (end <= offset)
451 		return (0);
452 	length = end - offset;
453 
454 	while (length) {
455 		start = end;
456 		/* assert(offset <= start) */
457 		err = get_next_chunk(dn, &start, offset);
458 		if (err)
459 			return (err);
460 		len = trunc ? DMU_OBJECT_END : end - start;
461 
462 		tx = dmu_tx_create(os);
463 		dmu_tx_hold_free(tx, dn->dn_object, start, len);
464 		err = dmu_tx_assign(tx, TXG_WAIT);
465 		if (err) {
466 			dmu_tx_abort(tx);
467 			return (err);
468 		}
469 
470 		dnode_free_range(dn, start, trunc ? -1 : len, tx);
471 
472 		if (start == 0 && free_dnode) {
473 			ASSERT(trunc);
474 			dnode_free(dn, tx);
475 		}
476 
477 		length -= end - start;
478 
479 		dmu_tx_commit(tx);
480 		end = start;
481 	}
482 	return (0);
483 }
484 
485 int
486 dmu_free_long_range(objset_t *os, uint64_t object,
487     uint64_t offset, uint64_t length)
488 {
489 	dnode_t *dn;
490 	int err;
491 
492 	err = dnode_hold(os->os, object, FTAG, &dn);
493 	if (err != 0)
494 		return (err);
495 	err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
496 	dnode_rele(dn, FTAG);
497 	return (err);
498 }
499 
500 int
501 dmu_free_object(objset_t *os, uint64_t object)
502 {
503 	dnode_t *dn;
504 	dmu_tx_t *tx;
505 	int err;
506 
507 	err = dnode_hold_impl(os->os, object, DNODE_MUST_BE_ALLOCATED,
508 	    FTAG, &dn);
509 	if (err != 0)
510 		return (err);
511 	if (dn->dn_nlevels == 1) {
512 		tx = dmu_tx_create(os);
513 		dmu_tx_hold_bonus(tx, object);
514 		dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
515 		err = dmu_tx_assign(tx, TXG_WAIT);
516 		if (err == 0) {
517 			dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
518 			dnode_free(dn, tx);
519 			dmu_tx_commit(tx);
520 		} else {
521 			dmu_tx_abort(tx);
522 		}
523 	} else {
524 		err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
525 	}
526 	dnode_rele(dn, FTAG);
527 	return (err);
528 }
529 
530 int
531 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
532     uint64_t size, dmu_tx_t *tx)
533 {
534 	dnode_t *dn;
535 	int err = dnode_hold(os->os, object, FTAG, &dn);
536 	if (err)
537 		return (err);
538 	ASSERT(offset < UINT64_MAX);
539 	ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
540 	dnode_free_range(dn, offset, size, tx);
541 	dnode_rele(dn, FTAG);
542 	return (0);
543 }
544 
545 int
546 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
547     void *buf)
548 {
549 	dnode_t *dn;
550 	dmu_buf_t **dbp;
551 	int numbufs, i, err;
552 
553 	err = dnode_hold(os->os, object, FTAG, &dn);
554 	if (err)
555 		return (err);
556 
557 	/*
558 	 * Deal with odd block sizes, where there can't be data past the first
559 	 * block.  If we ever do the tail block optimization, we will need to
560 	 * handle that here as well.
561 	 */
562 	if (dn->dn_datablkshift == 0) {
563 		int newsz = offset > dn->dn_datablksz ? 0 :
564 		    MIN(size, dn->dn_datablksz - offset);
565 		bzero((char *)buf + newsz, size - newsz);
566 		size = newsz;
567 	}
568 
569 	while (size > 0) {
570 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
571 
572 		/*
573 		 * NB: we could do this block-at-a-time, but it's nice
574 		 * to be reading in parallel.
575 		 */
576 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
577 		    TRUE, FTAG, &numbufs, &dbp);
578 		if (err)
579 			break;
580 
581 		for (i = 0; i < numbufs; i++) {
582 			int tocpy;
583 			int bufoff;
584 			dmu_buf_t *db = dbp[i];
585 
586 			ASSERT(size > 0);
587 
588 			bufoff = offset - db->db_offset;
589 			tocpy = (int)MIN(db->db_size - bufoff, size);
590 
591 			bcopy((char *)db->db_data + bufoff, buf, tocpy);
592 
593 			offset += tocpy;
594 			size -= tocpy;
595 			buf = (char *)buf + tocpy;
596 		}
597 		dmu_buf_rele_array(dbp, numbufs, FTAG);
598 	}
599 	dnode_rele(dn, FTAG);
600 	return (err);
601 }
602 
603 void
604 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
605     const void *buf, dmu_tx_t *tx)
606 {
607 	dmu_buf_t **dbp;
608 	int numbufs, i;
609 
610 	if (size == 0)
611 		return;
612 
613 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
614 	    FALSE, FTAG, &numbufs, &dbp));
615 
616 	for (i = 0; i < numbufs; i++) {
617 		int tocpy;
618 		int bufoff;
619 		dmu_buf_t *db = dbp[i];
620 
621 		ASSERT(size > 0);
622 
623 		bufoff = offset - db->db_offset;
624 		tocpy = (int)MIN(db->db_size - bufoff, size);
625 
626 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
627 
628 		if (tocpy == db->db_size)
629 			dmu_buf_will_fill(db, tx);
630 		else
631 			dmu_buf_will_dirty(db, tx);
632 
633 		bcopy(buf, (char *)db->db_data + bufoff, tocpy);
634 
635 		if (tocpy == db->db_size)
636 			dmu_buf_fill_done(db, tx);
637 
638 		offset += tocpy;
639 		size -= tocpy;
640 		buf = (char *)buf + tocpy;
641 	}
642 	dmu_buf_rele_array(dbp, numbufs, FTAG);
643 }
644 
645 void
646 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
647     dmu_tx_t *tx)
648 {
649 	dmu_buf_t **dbp;
650 	int numbufs, i;
651 
652 	if (size == 0)
653 		return;
654 
655 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
656 	    FALSE, FTAG, &numbufs, &dbp));
657 
658 	for (i = 0; i < numbufs; i++) {
659 		dmu_buf_t *db = dbp[i];
660 
661 		dmu_buf_will_not_fill(db, tx);
662 	}
663 	dmu_buf_rele_array(dbp, numbufs, FTAG);
664 }
665 
666 #ifdef _KERNEL
667 int
668 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
669 {
670 	dmu_buf_t **dbp;
671 	int numbufs, i, err;
672 
673 	/*
674 	 * NB: we could do this block-at-a-time, but it's nice
675 	 * to be reading in parallel.
676 	 */
677 	err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
678 	    &numbufs, &dbp);
679 	if (err)
680 		return (err);
681 
682 	for (i = 0; i < numbufs; i++) {
683 		int tocpy;
684 		int bufoff;
685 		dmu_buf_t *db = dbp[i];
686 
687 		ASSERT(size > 0);
688 
689 		bufoff = uio->uio_loffset - db->db_offset;
690 		tocpy = (int)MIN(db->db_size - bufoff, size);
691 
692 		err = uiomove((char *)db->db_data + bufoff, tocpy,
693 		    UIO_READ, uio);
694 		if (err)
695 			break;
696 
697 		size -= tocpy;
698 	}
699 	dmu_buf_rele_array(dbp, numbufs, FTAG);
700 
701 	return (err);
702 }
703 
704 int
705 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
706     dmu_tx_t *tx)
707 {
708 	dmu_buf_t **dbp;
709 	int numbufs, i;
710 	int err = 0;
711 
712 	if (size == 0)
713 		return (0);
714 
715 	err = dmu_buf_hold_array(os, object, uio->uio_loffset, size,
716 	    FALSE, FTAG, &numbufs, &dbp);
717 	if (err)
718 		return (err);
719 
720 	for (i = 0; i < numbufs; i++) {
721 		int tocpy;
722 		int bufoff;
723 		dmu_buf_t *db = dbp[i];
724 
725 		ASSERT(size > 0);
726 
727 		bufoff = uio->uio_loffset - db->db_offset;
728 		tocpy = (int)MIN(db->db_size - bufoff, size);
729 
730 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
731 
732 		if (tocpy == db->db_size)
733 			dmu_buf_will_fill(db, tx);
734 		else
735 			dmu_buf_will_dirty(db, tx);
736 
737 		/*
738 		 * XXX uiomove could block forever (eg. nfs-backed
739 		 * pages).  There needs to be a uiolockdown() function
740 		 * to lock the pages in memory, so that uiomove won't
741 		 * block.
742 		 */
743 		err = uiomove((char *)db->db_data + bufoff, tocpy,
744 		    UIO_WRITE, uio);
745 
746 		if (tocpy == db->db_size)
747 			dmu_buf_fill_done(db, tx);
748 
749 		if (err)
750 			break;
751 
752 		size -= tocpy;
753 	}
754 	dmu_buf_rele_array(dbp, numbufs, FTAG);
755 	return (err);
756 }
757 
758 int
759 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
760     page_t *pp, dmu_tx_t *tx)
761 {
762 	dmu_buf_t **dbp;
763 	int numbufs, i;
764 	int err;
765 
766 	if (size == 0)
767 		return (0);
768 
769 	err = dmu_buf_hold_array(os, object, offset, size,
770 	    FALSE, FTAG, &numbufs, &dbp);
771 	if (err)
772 		return (err);
773 
774 	for (i = 0; i < numbufs; i++) {
775 		int tocpy, copied, thiscpy;
776 		int bufoff;
777 		dmu_buf_t *db = dbp[i];
778 		caddr_t va;
779 
780 		ASSERT(size > 0);
781 		ASSERT3U(db->db_size, >=, PAGESIZE);
782 
783 		bufoff = offset - db->db_offset;
784 		tocpy = (int)MIN(db->db_size - bufoff, size);
785 
786 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
787 
788 		if (tocpy == db->db_size)
789 			dmu_buf_will_fill(db, tx);
790 		else
791 			dmu_buf_will_dirty(db, tx);
792 
793 		for (copied = 0; copied < tocpy; copied += PAGESIZE) {
794 			ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
795 			thiscpy = MIN(PAGESIZE, tocpy - copied);
796 			va = zfs_map_page(pp, S_READ);
797 			bcopy(va, (char *)db->db_data + bufoff, thiscpy);
798 			zfs_unmap_page(pp, va);
799 			pp = pp->p_next;
800 			bufoff += PAGESIZE;
801 		}
802 
803 		if (tocpy == db->db_size)
804 			dmu_buf_fill_done(db, tx);
805 
806 		if (err)
807 			break;
808 
809 		offset += tocpy;
810 		size -= tocpy;
811 	}
812 	dmu_buf_rele_array(dbp, numbufs, FTAG);
813 	return (err);
814 }
815 #endif
816 
817 /*
818  * Allocate a loaned anonymous arc buffer.
819  */
820 arc_buf_t *
821 dmu_request_arcbuf(dmu_buf_t *handle, int size)
822 {
823 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
824 
825 	return (arc_loan_buf(dn->dn_objset->os_spa, size));
826 }
827 
828 /*
829  * Free a loaned arc buffer.
830  */
831 void
832 dmu_return_arcbuf(arc_buf_t *buf)
833 {
834 	arc_return_buf(buf, FTAG);
835 	VERIFY(arc_buf_remove_ref(buf, FTAG) == 1);
836 }
837 
838 /*
839  * When possible directly assign passed loaned arc buffer to a dbuf.
840  * If this is not possible copy the contents of passed arc buf via
841  * dmu_write().
842  */
843 void
844 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
845     dmu_tx_t *tx)
846 {
847 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
848 	dmu_buf_impl_t *db;
849 	uint32_t blksz = (uint32_t)arc_buf_size(buf);
850 	uint64_t blkid;
851 
852 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
853 	blkid = dbuf_whichblock(dn, offset);
854 	VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
855 	rw_exit(&dn->dn_struct_rwlock);
856 
857 	if (offset == db->db.db_offset && blksz == db->db.db_size) {
858 		dbuf_assign_arcbuf(db, buf, tx);
859 		dbuf_rele(db, FTAG);
860 	} else {
861 		dbuf_rele(db, FTAG);
862 		ASSERT(dn->dn_objset->os.os == dn->dn_objset);
863 		dmu_write(&dn->dn_objset->os, dn->dn_object, offset, blksz,
864 		    buf->b_data, tx);
865 		dmu_return_arcbuf(buf);
866 	}
867 }
868 
869 typedef struct {
870 	dbuf_dirty_record_t	*dr;
871 	dmu_sync_cb_t		*done;
872 	void			*arg;
873 } dmu_sync_arg_t;
874 
875 /* ARGSUSED */
876 static void
877 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
878 {
879 	blkptr_t *bp = zio->io_bp;
880 
881 	if (!BP_IS_HOLE(bp)) {
882 		dmu_sync_arg_t *in = varg;
883 		dbuf_dirty_record_t *dr = in->dr;
884 		dmu_buf_impl_t *db = dr->dr_dbuf;
885 		ASSERT(BP_GET_TYPE(bp) == db->db_dnode->dn_type);
886 		ASSERT(BP_GET_LEVEL(bp) == 0);
887 		bp->blk_fill = 1;
888 	}
889 }
890 
891 /* ARGSUSED */
892 static void
893 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
894 {
895 	dmu_sync_arg_t *in = varg;
896 	dbuf_dirty_record_t *dr = in->dr;
897 	dmu_buf_impl_t *db = dr->dr_dbuf;
898 	dmu_sync_cb_t *done = in->done;
899 
900 	mutex_enter(&db->db_mtx);
901 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
902 	dr->dt.dl.dr_overridden_by = *zio->io_bp; /* structure assignment */
903 	dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
904 	cv_broadcast(&db->db_changed);
905 	mutex_exit(&db->db_mtx);
906 
907 	if (done)
908 		done(&(db->db), in->arg);
909 
910 	kmem_free(in, sizeof (dmu_sync_arg_t));
911 }
912 
913 /*
914  * Intent log support: sync the block associated with db to disk.
915  * N.B. and XXX: the caller is responsible for making sure that the
916  * data isn't changing while dmu_sync() is writing it.
917  *
918  * Return values:
919  *
920  *	EEXIST: this txg has already been synced, so there's nothing to to.
921  *		The caller should not log the write.
922  *
923  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
924  *		The caller should not log the write.
925  *
926  *	EALREADY: this block is already in the process of being synced.
927  *		The caller should track its progress (somehow).
928  *
929  *	EINPROGRESS: the IO has been initiated.
930  *		The caller should log this blkptr in the callback.
931  *
932  *	0: completed.  Sets *bp to the blkptr just written.
933  *		The caller should log this blkptr immediately.
934  */
935 int
936 dmu_sync(zio_t *pio, dmu_buf_t *db_fake,
937     blkptr_t *bp, uint64_t txg, dmu_sync_cb_t *done, void *arg)
938 {
939 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
940 	objset_impl_t *os = db->db_objset;
941 	dsl_pool_t *dp = os->os_dsl_dataset->ds_dir->dd_pool;
942 	tx_state_t *tx = &dp->dp_tx;
943 	dbuf_dirty_record_t *dr;
944 	dmu_sync_arg_t *in;
945 	zbookmark_t zb;
946 	writeprops_t wp = { 0 };
947 	zio_t *zio;
948 	int err;
949 
950 	ASSERT(BP_IS_HOLE(bp));
951 	ASSERT(txg != 0);
952 
953 	dprintf("dmu_sync txg=%llu, s,o,q %llu %llu %llu\n",
954 	    txg, tx->tx_synced_txg, tx->tx_open_txg, tx->tx_quiesced_txg);
955 
956 	/*
957 	 * XXX - would be nice if we could do this without suspending...
958 	 */
959 	txg_suspend(dp);
960 
961 	/*
962 	 * If this txg already synced, there's nothing to do.
963 	 */
964 	if (txg <= tx->tx_synced_txg) {
965 		txg_resume(dp);
966 		/*
967 		 * If we're running ziltest, we need the blkptr regardless.
968 		 */
969 		if (txg > spa_freeze_txg(dp->dp_spa)) {
970 			/* if db_blkptr == NULL, this was an empty write */
971 			if (db->db_blkptr)
972 				*bp = *db->db_blkptr; /* structure assignment */
973 			return (0);
974 		}
975 		return (EEXIST);
976 	}
977 
978 	mutex_enter(&db->db_mtx);
979 
980 	if (txg == tx->tx_syncing_txg) {
981 		while (db->db_data_pending) {
982 			/*
983 			 * IO is in-progress.  Wait for it to finish.
984 			 * XXX - would be nice to be able to somehow "attach"
985 			 * this zio to the parent zio passed in.
986 			 */
987 			cv_wait(&db->db_changed, &db->db_mtx);
988 			if (!db->db_data_pending &&
989 			    db->db_blkptr && BP_IS_HOLE(db->db_blkptr)) {
990 				/*
991 				 * IO was compressed away
992 				 */
993 				*bp = *db->db_blkptr; /* structure assignment */
994 				mutex_exit(&db->db_mtx);
995 				txg_resume(dp);
996 				return (0);
997 			}
998 			ASSERT(db->db_data_pending ||
999 			    (db->db_blkptr && db->db_blkptr->blk_birth == txg));
1000 		}
1001 
1002 		if (db->db_blkptr && db->db_blkptr->blk_birth == txg) {
1003 			/*
1004 			 * IO is already completed.
1005 			 */
1006 			*bp = *db->db_blkptr; /* structure assignment */
1007 			mutex_exit(&db->db_mtx);
1008 			txg_resume(dp);
1009 			return (0);
1010 		}
1011 	}
1012 
1013 	dr = db->db_last_dirty;
1014 	while (dr && dr->dr_txg > txg)
1015 		dr = dr->dr_next;
1016 	if (dr == NULL || dr->dr_txg < txg) {
1017 		/*
1018 		 * This dbuf isn't dirty, must have been free_range'd.
1019 		 * There's no need to log writes to freed blocks, so we're done.
1020 		 */
1021 		mutex_exit(&db->db_mtx);
1022 		txg_resume(dp);
1023 		return (ENOENT);
1024 	}
1025 
1026 	ASSERT(dr->dr_txg == txg);
1027 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
1028 		/*
1029 		 * We have already issued a sync write for this buffer.
1030 		 */
1031 		mutex_exit(&db->db_mtx);
1032 		txg_resume(dp);
1033 		return (EALREADY);
1034 	} else if (dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1035 		/*
1036 		 * This buffer has already been synced.  It could not
1037 		 * have been dirtied since, or we would have cleared the state.
1038 		 */
1039 		*bp = dr->dt.dl.dr_overridden_by; /* structure assignment */
1040 		mutex_exit(&db->db_mtx);
1041 		txg_resume(dp);
1042 		return (0);
1043 	}
1044 
1045 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1046 	in = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1047 	in->dr = dr;
1048 	in->done = done;
1049 	in->arg = arg;
1050 	mutex_exit(&db->db_mtx);
1051 	txg_resume(dp);
1052 
1053 	zb.zb_objset = os->os_dsl_dataset->ds_object;
1054 	zb.zb_object = db->db.db_object;
1055 	zb.zb_level = db->db_level;
1056 	zb.zb_blkid = db->db_blkid;
1057 
1058 	wp.wp_type = db->db_dnode->dn_type;
1059 	wp.wp_level = db->db_level;
1060 	wp.wp_copies = os->os_copies;
1061 	wp.wp_dnchecksum = db->db_dnode->dn_checksum;
1062 	wp.wp_oschecksum = os->os_checksum;
1063 	wp.wp_dncompress = db->db_dnode->dn_compress;
1064 	wp.wp_oscompress = os->os_compress;
1065 
1066 	ASSERT(BP_IS_HOLE(bp));
1067 
1068 	zio = arc_write(pio, os->os_spa, &wp, DBUF_IS_L2CACHEABLE(db),
1069 	    txg, bp, dr->dt.dl.dr_data, dmu_sync_ready, dmu_sync_done, in,
1070 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
1071 	if (pio) {
1072 		zio_nowait(zio);
1073 		err = EINPROGRESS;
1074 	} else {
1075 		err = zio_wait(zio);
1076 		ASSERT(err == 0);
1077 	}
1078 	return (err);
1079 }
1080 
1081 int
1082 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1083 	dmu_tx_t *tx)
1084 {
1085 	dnode_t *dn;
1086 	int err;
1087 
1088 	err = dnode_hold(os->os, object, FTAG, &dn);
1089 	if (err)
1090 		return (err);
1091 	err = dnode_set_blksz(dn, size, ibs, tx);
1092 	dnode_rele(dn, FTAG);
1093 	return (err);
1094 }
1095 
1096 void
1097 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1098 	dmu_tx_t *tx)
1099 {
1100 	dnode_t *dn;
1101 
1102 	/* XXX assumes dnode_hold will not get an i/o error */
1103 	(void) dnode_hold(os->os, object, FTAG, &dn);
1104 	ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1105 	dn->dn_checksum = checksum;
1106 	dnode_setdirty(dn, tx);
1107 	dnode_rele(dn, FTAG);
1108 }
1109 
1110 void
1111 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1112 	dmu_tx_t *tx)
1113 {
1114 	dnode_t *dn;
1115 
1116 	/* XXX assumes dnode_hold will not get an i/o error */
1117 	(void) dnode_hold(os->os, object, FTAG, &dn);
1118 	ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1119 	dn->dn_compress = compress;
1120 	dnode_setdirty(dn, tx);
1121 	dnode_rele(dn, FTAG);
1122 }
1123 
1124 int
1125 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1126 {
1127 	dnode_t *dn;
1128 	int i, err;
1129 
1130 	err = dnode_hold(os->os, object, FTAG, &dn);
1131 	if (err)
1132 		return (err);
1133 	/*
1134 	 * Sync any current changes before
1135 	 * we go trundling through the block pointers.
1136 	 */
1137 	for (i = 0; i < TXG_SIZE; i++) {
1138 		if (list_link_active(&dn->dn_dirty_link[i]))
1139 			break;
1140 	}
1141 	if (i != TXG_SIZE) {
1142 		dnode_rele(dn, FTAG);
1143 		txg_wait_synced(dmu_objset_pool(os), 0);
1144 		err = dnode_hold(os->os, object, FTAG, &dn);
1145 		if (err)
1146 			return (err);
1147 	}
1148 
1149 	err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1150 	dnode_rele(dn, FTAG);
1151 
1152 	return (err);
1153 }
1154 
1155 void
1156 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1157 {
1158 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1159 	mutex_enter(&dn->dn_mtx);
1160 
1161 	doi->doi_data_block_size = dn->dn_datablksz;
1162 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
1163 	    1ULL << dn->dn_indblkshift : 0;
1164 	doi->doi_indirection = dn->dn_nlevels;
1165 	doi->doi_checksum = dn->dn_checksum;
1166 	doi->doi_compress = dn->dn_compress;
1167 	doi->doi_physical_blks = (DN_USED_BYTES(dn->dn_phys) +
1168 	    SPA_MINBLOCKSIZE/2) >> SPA_MINBLOCKSHIFT;
1169 	doi->doi_max_block_offset = dn->dn_phys->dn_maxblkid;
1170 	doi->doi_type = dn->dn_type;
1171 	doi->doi_bonus_size = dn->dn_bonuslen;
1172 	doi->doi_bonus_type = dn->dn_bonustype;
1173 
1174 	mutex_exit(&dn->dn_mtx);
1175 	rw_exit(&dn->dn_struct_rwlock);
1176 }
1177 
1178 /*
1179  * Get information on a DMU object.
1180  * If doi is NULL, just indicates whether the object exists.
1181  */
1182 int
1183 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1184 {
1185 	dnode_t *dn;
1186 	int err = dnode_hold(os->os, object, FTAG, &dn);
1187 
1188 	if (err)
1189 		return (err);
1190 
1191 	if (doi != NULL)
1192 		dmu_object_info_from_dnode(dn, doi);
1193 
1194 	dnode_rele(dn, FTAG);
1195 	return (0);
1196 }
1197 
1198 /*
1199  * As above, but faster; can be used when you have a held dbuf in hand.
1200  */
1201 void
1202 dmu_object_info_from_db(dmu_buf_t *db, dmu_object_info_t *doi)
1203 {
1204 	dmu_object_info_from_dnode(((dmu_buf_impl_t *)db)->db_dnode, doi);
1205 }
1206 
1207 /*
1208  * Faster still when you only care about the size.
1209  * This is specifically optimized for zfs_getattr().
1210  */
1211 void
1212 dmu_object_size_from_db(dmu_buf_t *db, uint32_t *blksize, u_longlong_t *nblk512)
1213 {
1214 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1215 
1216 	*blksize = dn->dn_datablksz;
1217 	/* add 1 for dnode space */
1218 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1219 	    SPA_MINBLOCKSHIFT) + 1;
1220 }
1221 
1222 void
1223 byteswap_uint64_array(void *vbuf, size_t size)
1224 {
1225 	uint64_t *buf = vbuf;
1226 	size_t count = size >> 3;
1227 	int i;
1228 
1229 	ASSERT((size & 7) == 0);
1230 
1231 	for (i = 0; i < count; i++)
1232 		buf[i] = BSWAP_64(buf[i]);
1233 }
1234 
1235 void
1236 byteswap_uint32_array(void *vbuf, size_t size)
1237 {
1238 	uint32_t *buf = vbuf;
1239 	size_t count = size >> 2;
1240 	int i;
1241 
1242 	ASSERT((size & 3) == 0);
1243 
1244 	for (i = 0; i < count; i++)
1245 		buf[i] = BSWAP_32(buf[i]);
1246 }
1247 
1248 void
1249 byteswap_uint16_array(void *vbuf, size_t size)
1250 {
1251 	uint16_t *buf = vbuf;
1252 	size_t count = size >> 1;
1253 	int i;
1254 
1255 	ASSERT((size & 1) == 0);
1256 
1257 	for (i = 0; i < count; i++)
1258 		buf[i] = BSWAP_16(buf[i]);
1259 }
1260 
1261 /* ARGSUSED */
1262 void
1263 byteswap_uint8_array(void *vbuf, size_t size)
1264 {
1265 }
1266 
1267 void
1268 dmu_init(void)
1269 {
1270 	dbuf_init();
1271 	dnode_init();
1272 	arc_init();
1273 	l2arc_init();
1274 }
1275 
1276 void
1277 dmu_fini(void)
1278 {
1279 	arc_fini();
1280 	dnode_fini();
1281 	dbuf_fini();
1282 	l2arc_fini();
1283 }
1284