xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu.c (revision 802b83c445ef5ffc9777155491dfe4fcd9793946)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/dmu.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dbuf.h>
32 #include <sys/dnode.h>
33 #include <sys/zfs_context.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dmu_traverse.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dsl_pool.h>
39 #include <sys/dsl_synctask.h>
40 #include <sys/dsl_prop.h>
41 #include <sys/dmu_zfetch.h>
42 #include <sys/zfs_ioctl.h>
43 #include <sys/zap.h>
44 #include <sys/zio_checksum.h>
45 #ifdef _KERNEL
46 #include <sys/vmsystm.h>
47 #endif
48 
49 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
50 	{	byteswap_uint8_array,	TRUE,	"unallocated"		},
51 	{	zap_byteswap,		TRUE,	"object directory"	},
52 	{	byteswap_uint64_array,	TRUE,	"object array"		},
53 	{	byteswap_uint8_array,	TRUE,	"packed nvlist"		},
54 	{	byteswap_uint64_array,	TRUE,	"packed nvlist size"	},
55 	{	byteswap_uint64_array,	TRUE,	"bplist"		},
56 	{	byteswap_uint64_array,	TRUE,	"bplist header"		},
57 	{	byteswap_uint64_array,	TRUE,	"SPA space map header"	},
58 	{	byteswap_uint64_array,	TRUE,	"SPA space map"		},
59 	{	byteswap_uint64_array,	TRUE,	"ZIL intent log"	},
60 	{	dnode_buf_byteswap,	TRUE,	"DMU dnode"		},
61 	{	dmu_objset_byteswap,	TRUE,	"DMU objset"		},
62 	{	byteswap_uint64_array,	TRUE,	"DSL directory"		},
63 	{	zap_byteswap,		TRUE,	"DSL directory child map"},
64 	{	zap_byteswap,		TRUE,	"DSL dataset snap map"	},
65 	{	zap_byteswap,		TRUE,	"DSL props"		},
66 	{	byteswap_uint64_array,	TRUE,	"DSL dataset"		},
67 	{	zfs_znode_byteswap,	TRUE,	"ZFS znode"		},
68 	{	zfs_acl_byteswap,	TRUE,	"ZFS ACL"		},
69 	{	byteswap_uint8_array,	FALSE,	"ZFS plain file"	},
70 	{	zap_byteswap,		TRUE,	"ZFS directory"		},
71 	{	zap_byteswap,		TRUE,	"ZFS master node"	},
72 	{	zap_byteswap,		TRUE,	"ZFS delete queue"	},
73 	{	byteswap_uint8_array,	FALSE,	"zvol object"		},
74 	{	zap_byteswap,		TRUE,	"zvol prop"		},
75 	{	byteswap_uint8_array,	FALSE,	"other uint8[]"		},
76 	{	byteswap_uint64_array,	FALSE,	"other uint64[]"	},
77 	{	zap_byteswap,		TRUE,	"other ZAP"		},
78 	{	zap_byteswap,		TRUE,	"persistent error log"	},
79 	{	byteswap_uint8_array,	TRUE,	"SPA history"		},
80 	{	byteswap_uint64_array,	TRUE,	"SPA history offsets"	},
81 	{	zap_byteswap,		TRUE,	"Pool properties"	},
82 	{	zap_byteswap,		TRUE,	"DSL permissions"	}
83 };
84 
85 int
86 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
87     void *tag, dmu_buf_t **dbp)
88 {
89 	dnode_t *dn;
90 	uint64_t blkid;
91 	dmu_buf_impl_t *db;
92 	int err;
93 
94 	err = dnode_hold(os->os, object, FTAG, &dn);
95 	if (err)
96 		return (err);
97 	blkid = dbuf_whichblock(dn, offset);
98 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
99 	db = dbuf_hold(dn, blkid, tag);
100 	rw_exit(&dn->dn_struct_rwlock);
101 	if (db == NULL) {
102 		err = EIO;
103 	} else {
104 		err = dbuf_read(db, NULL, DB_RF_CANFAIL);
105 		if (err) {
106 			dbuf_rele(db, tag);
107 			db = NULL;
108 		}
109 	}
110 
111 	dnode_rele(dn, FTAG);
112 	*dbp = &db->db;
113 	return (err);
114 }
115 
116 int
117 dmu_bonus_max(void)
118 {
119 	return (DN_MAX_BONUSLEN);
120 }
121 
122 int
123 dmu_set_bonus(dmu_buf_t *db, int newsize, dmu_tx_t *tx)
124 {
125 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
126 
127 	if (dn->dn_bonus != (dmu_buf_impl_t *)db)
128 		return (EINVAL);
129 	if (newsize < 0 || newsize > db->db_size)
130 		return (EINVAL);
131 	dnode_setbonuslen(dn, newsize, tx);
132 	return (0);
133 }
134 
135 /*
136  * returns ENOENT, EIO, or 0.
137  */
138 int
139 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
140 {
141 	dnode_t *dn;
142 	dmu_buf_impl_t *db;
143 	int error;
144 
145 	error = dnode_hold(os->os, object, FTAG, &dn);
146 	if (error)
147 		return (error);
148 
149 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
150 	if (dn->dn_bonus == NULL) {
151 		rw_exit(&dn->dn_struct_rwlock);
152 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
153 		if (dn->dn_bonus == NULL)
154 			dbuf_create_bonus(dn);
155 	}
156 	db = dn->dn_bonus;
157 	rw_exit(&dn->dn_struct_rwlock);
158 
159 	/* as long as the bonus buf is held, the dnode will be held */
160 	if (refcount_add(&db->db_holds, tag) == 1)
161 		VERIFY(dnode_add_ref(dn, db));
162 
163 	dnode_rele(dn, FTAG);
164 
165 	VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED));
166 
167 	*dbp = &db->db;
168 	return (0);
169 }
170 
171 /*
172  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
173  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
174  * and can induce severe lock contention when writing to several files
175  * whose dnodes are in the same block.
176  */
177 static int
178 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset,
179     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
180 {
181 	dmu_buf_t **dbp;
182 	uint64_t blkid, nblks, i;
183 	uint32_t flags;
184 	int err;
185 	zio_t *zio;
186 
187 	ASSERT(length <= DMU_MAX_ACCESS);
188 
189 	flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT;
190 	if (length > zfetch_array_rd_sz)
191 		flags |= DB_RF_NOPREFETCH;
192 
193 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
194 	if (dn->dn_datablkshift) {
195 		int blkshift = dn->dn_datablkshift;
196 		nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
197 		    P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
198 	} else {
199 		if (offset + length > dn->dn_datablksz) {
200 			zfs_panic_recover("zfs: accessing past end of object "
201 			    "%llx/%llx (size=%u access=%llu+%llu)",
202 			    (longlong_t)dn->dn_objset->
203 			    os_dsl_dataset->ds_object,
204 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
205 			    (longlong_t)offset, (longlong_t)length);
206 			return (EIO);
207 		}
208 		nblks = 1;
209 	}
210 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
211 
212 	zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, TRUE);
213 	blkid = dbuf_whichblock(dn, offset);
214 	for (i = 0; i < nblks; i++) {
215 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
216 		if (db == NULL) {
217 			rw_exit(&dn->dn_struct_rwlock);
218 			dmu_buf_rele_array(dbp, nblks, tag);
219 			zio_nowait(zio);
220 			return (EIO);
221 		}
222 		/* initiate async i/o */
223 		if (read) {
224 			rw_exit(&dn->dn_struct_rwlock);
225 			(void) dbuf_read(db, zio, flags);
226 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
227 		}
228 		dbp[i] = &db->db;
229 	}
230 	rw_exit(&dn->dn_struct_rwlock);
231 
232 	/* wait for async i/o */
233 	err = zio_wait(zio);
234 	if (err) {
235 		dmu_buf_rele_array(dbp, nblks, tag);
236 		return (err);
237 	}
238 
239 	/* wait for other io to complete */
240 	if (read) {
241 		for (i = 0; i < nblks; i++) {
242 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
243 			mutex_enter(&db->db_mtx);
244 			while (db->db_state == DB_READ ||
245 			    db->db_state == DB_FILL)
246 				cv_wait(&db->db_changed, &db->db_mtx);
247 			if (db->db_state == DB_UNCACHED)
248 				err = EIO;
249 			mutex_exit(&db->db_mtx);
250 			if (err) {
251 				dmu_buf_rele_array(dbp, nblks, tag);
252 				return (err);
253 			}
254 		}
255 	}
256 
257 	*numbufsp = nblks;
258 	*dbpp = dbp;
259 	return (0);
260 }
261 
262 static int
263 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
264     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
265 {
266 	dnode_t *dn;
267 	int err;
268 
269 	err = dnode_hold(os->os, object, FTAG, &dn);
270 	if (err)
271 		return (err);
272 
273 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
274 	    numbufsp, dbpp);
275 
276 	dnode_rele(dn, FTAG);
277 
278 	return (err);
279 }
280 
281 int
282 dmu_buf_hold_array_by_bonus(dmu_buf_t *db, uint64_t offset,
283     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
284 {
285 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
286 	int err;
287 
288 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
289 	    numbufsp, dbpp);
290 
291 	return (err);
292 }
293 
294 void
295 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
296 {
297 	int i;
298 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
299 
300 	if (numbufs == 0)
301 		return;
302 
303 	for (i = 0; i < numbufs; i++) {
304 		if (dbp[i])
305 			dbuf_rele(dbp[i], tag);
306 	}
307 
308 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
309 }
310 
311 void
312 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
313 {
314 	dnode_t *dn;
315 	uint64_t blkid;
316 	int nblks, i, err;
317 
318 	if (zfs_prefetch_disable)
319 		return;
320 
321 	if (len == 0) {  /* they're interested in the bonus buffer */
322 		dn = os->os->os_meta_dnode;
323 
324 		if (object == 0 || object >= DN_MAX_OBJECT)
325 			return;
326 
327 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
328 		blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
329 		dbuf_prefetch(dn, blkid);
330 		rw_exit(&dn->dn_struct_rwlock);
331 		return;
332 	}
333 
334 	/*
335 	 * XXX - Note, if the dnode for the requested object is not
336 	 * already cached, we will do a *synchronous* read in the
337 	 * dnode_hold() call.  The same is true for any indirects.
338 	 */
339 	err = dnode_hold(os->os, object, FTAG, &dn);
340 	if (err != 0)
341 		return;
342 
343 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
344 	if (dn->dn_datablkshift) {
345 		int blkshift = dn->dn_datablkshift;
346 		nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
347 		    P2ALIGN(offset, 1<<blkshift)) >> blkshift;
348 	} else {
349 		nblks = (offset < dn->dn_datablksz);
350 	}
351 
352 	if (nblks != 0) {
353 		blkid = dbuf_whichblock(dn, offset);
354 		for (i = 0; i < nblks; i++)
355 			dbuf_prefetch(dn, blkid+i);
356 	}
357 
358 	rw_exit(&dn->dn_struct_rwlock);
359 
360 	dnode_rele(dn, FTAG);
361 }
362 
363 int
364 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
365     uint64_t size, dmu_tx_t *tx)
366 {
367 	dnode_t *dn;
368 	int err = dnode_hold(os->os, object, FTAG, &dn);
369 	if (err)
370 		return (err);
371 	ASSERT(offset < UINT64_MAX);
372 	ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
373 	dnode_free_range(dn, offset, size, tx);
374 	dnode_rele(dn, FTAG);
375 	return (0);
376 }
377 
378 int
379 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
380     void *buf)
381 {
382 	dnode_t *dn;
383 	dmu_buf_t **dbp;
384 	int numbufs, i, err;
385 
386 	err = dnode_hold(os->os, object, FTAG, &dn);
387 	if (err)
388 		return (err);
389 
390 	/*
391 	 * Deal with odd block sizes, where there can't be data past the first
392 	 * block.  If we ever do the tail block optimization, we will need to
393 	 * handle that here as well.
394 	 */
395 	if (dn->dn_datablkshift == 0) {
396 		int newsz = offset > dn->dn_datablksz ? 0 :
397 		    MIN(size, dn->dn_datablksz - offset);
398 		bzero((char *)buf + newsz, size - newsz);
399 		size = newsz;
400 	}
401 
402 	while (size > 0) {
403 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
404 
405 		/*
406 		 * NB: we could do this block-at-a-time, but it's nice
407 		 * to be reading in parallel.
408 		 */
409 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
410 		    TRUE, FTAG, &numbufs, &dbp);
411 		if (err)
412 			break;
413 
414 		for (i = 0; i < numbufs; i++) {
415 			int tocpy;
416 			int bufoff;
417 			dmu_buf_t *db = dbp[i];
418 
419 			ASSERT(size > 0);
420 
421 			bufoff = offset - db->db_offset;
422 			tocpy = (int)MIN(db->db_size - bufoff, size);
423 
424 			bcopy((char *)db->db_data + bufoff, buf, tocpy);
425 
426 			offset += tocpy;
427 			size -= tocpy;
428 			buf = (char *)buf + tocpy;
429 		}
430 		dmu_buf_rele_array(dbp, numbufs, FTAG);
431 	}
432 	dnode_rele(dn, FTAG);
433 	return (err);
434 }
435 
436 void
437 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
438     const void *buf, dmu_tx_t *tx)
439 {
440 	dmu_buf_t **dbp;
441 	int numbufs, i;
442 
443 	if (size == 0)
444 		return;
445 
446 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
447 	    FALSE, FTAG, &numbufs, &dbp));
448 
449 	for (i = 0; i < numbufs; i++) {
450 		int tocpy;
451 		int bufoff;
452 		dmu_buf_t *db = dbp[i];
453 
454 		ASSERT(size > 0);
455 
456 		bufoff = offset - db->db_offset;
457 		tocpy = (int)MIN(db->db_size - bufoff, size);
458 
459 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
460 
461 		if (tocpy == db->db_size)
462 			dmu_buf_will_fill(db, tx);
463 		else
464 			dmu_buf_will_dirty(db, tx);
465 
466 		bcopy(buf, (char *)db->db_data + bufoff, tocpy);
467 
468 		if (tocpy == db->db_size)
469 			dmu_buf_fill_done(db, tx);
470 
471 		offset += tocpy;
472 		size -= tocpy;
473 		buf = (char *)buf + tocpy;
474 	}
475 	dmu_buf_rele_array(dbp, numbufs, FTAG);
476 }
477 
478 #ifdef _KERNEL
479 int
480 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
481 {
482 	dmu_buf_t **dbp;
483 	int numbufs, i, err;
484 
485 	/*
486 	 * NB: we could do this block-at-a-time, but it's nice
487 	 * to be reading in parallel.
488 	 */
489 	err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
490 	    &numbufs, &dbp);
491 	if (err)
492 		return (err);
493 
494 	for (i = 0; i < numbufs; i++) {
495 		int tocpy;
496 		int bufoff;
497 		dmu_buf_t *db = dbp[i];
498 
499 		ASSERT(size > 0);
500 
501 		bufoff = uio->uio_loffset - db->db_offset;
502 		tocpy = (int)MIN(db->db_size - bufoff, size);
503 
504 		err = uiomove((char *)db->db_data + bufoff, tocpy,
505 		    UIO_READ, uio);
506 		if (err)
507 			break;
508 
509 		size -= tocpy;
510 	}
511 	dmu_buf_rele_array(dbp, numbufs, FTAG);
512 
513 	return (err);
514 }
515 
516 int
517 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
518     dmu_tx_t *tx)
519 {
520 	dmu_buf_t **dbp;
521 	int numbufs, i;
522 	int err = 0;
523 
524 	if (size == 0)
525 		return (0);
526 
527 	err = dmu_buf_hold_array(os, object, uio->uio_loffset, size,
528 	    FALSE, FTAG, &numbufs, &dbp);
529 	if (err)
530 		return (err);
531 
532 	for (i = 0; i < numbufs; i++) {
533 		int tocpy;
534 		int bufoff;
535 		dmu_buf_t *db = dbp[i];
536 
537 		ASSERT(size > 0);
538 
539 		bufoff = uio->uio_loffset - db->db_offset;
540 		tocpy = (int)MIN(db->db_size - bufoff, size);
541 
542 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
543 
544 		if (tocpy == db->db_size)
545 			dmu_buf_will_fill(db, tx);
546 		else
547 			dmu_buf_will_dirty(db, tx);
548 
549 		/*
550 		 * XXX uiomove could block forever (eg. nfs-backed
551 		 * pages).  There needs to be a uiolockdown() function
552 		 * to lock the pages in memory, so that uiomove won't
553 		 * block.
554 		 */
555 		err = uiomove((char *)db->db_data + bufoff, tocpy,
556 		    UIO_WRITE, uio);
557 
558 		if (tocpy == db->db_size)
559 			dmu_buf_fill_done(db, tx);
560 
561 		if (err)
562 			break;
563 
564 		size -= tocpy;
565 	}
566 	dmu_buf_rele_array(dbp, numbufs, FTAG);
567 	return (err);
568 }
569 
570 int
571 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
572     page_t *pp, dmu_tx_t *tx)
573 {
574 	dmu_buf_t **dbp;
575 	int numbufs, i;
576 	int err;
577 
578 	if (size == 0)
579 		return (0);
580 
581 	err = dmu_buf_hold_array(os, object, offset, size,
582 	    FALSE, FTAG, &numbufs, &dbp);
583 	if (err)
584 		return (err);
585 
586 	for (i = 0; i < numbufs; i++) {
587 		int tocpy, copied, thiscpy;
588 		int bufoff;
589 		dmu_buf_t *db = dbp[i];
590 		caddr_t va;
591 
592 		ASSERT(size > 0);
593 		ASSERT3U(db->db_size, >=, PAGESIZE);
594 
595 		bufoff = offset - db->db_offset;
596 		tocpy = (int)MIN(db->db_size - bufoff, size);
597 
598 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
599 
600 		if (tocpy == db->db_size)
601 			dmu_buf_will_fill(db, tx);
602 		else
603 			dmu_buf_will_dirty(db, tx);
604 
605 		for (copied = 0; copied < tocpy; copied += PAGESIZE) {
606 			ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
607 			thiscpy = MIN(PAGESIZE, tocpy - copied);
608 			va = ppmapin(pp, PROT_READ, (caddr_t)-1);
609 			bcopy(va, (char *)db->db_data + bufoff, thiscpy);
610 			ppmapout(va);
611 			pp = pp->p_next;
612 			bufoff += PAGESIZE;
613 		}
614 
615 		if (tocpy == db->db_size)
616 			dmu_buf_fill_done(db, tx);
617 
618 		if (err)
619 			break;
620 
621 		offset += tocpy;
622 		size -= tocpy;
623 	}
624 	dmu_buf_rele_array(dbp, numbufs, FTAG);
625 	return (err);
626 }
627 #endif
628 
629 typedef struct {
630 	dbuf_dirty_record_t	*dr;
631 	dmu_sync_cb_t		*done;
632 	void			*arg;
633 } dmu_sync_arg_t;
634 
635 /* ARGSUSED */
636 static void
637 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
638 {
639 	dmu_sync_arg_t *in = varg;
640 	dbuf_dirty_record_t *dr = in->dr;
641 	dmu_buf_impl_t *db = dr->dr_dbuf;
642 	dmu_sync_cb_t *done = in->done;
643 
644 	if (!BP_IS_HOLE(zio->io_bp)) {
645 		zio->io_bp->blk_fill = 1;
646 		BP_SET_TYPE(zio->io_bp, db->db_dnode->dn_type);
647 		BP_SET_LEVEL(zio->io_bp, 0);
648 	}
649 
650 	mutex_enter(&db->db_mtx);
651 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
652 	dr->dt.dl.dr_overridden_by = *zio->io_bp; /* structure assignment */
653 	dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
654 	cv_broadcast(&db->db_changed);
655 	mutex_exit(&db->db_mtx);
656 
657 	if (done)
658 		done(&(db->db), in->arg);
659 
660 	kmem_free(in, sizeof (dmu_sync_arg_t));
661 }
662 
663 /*
664  * Intent log support: sync the block associated with db to disk.
665  * N.B. and XXX: the caller is responsible for making sure that the
666  * data isn't changing while dmu_sync() is writing it.
667  *
668  * Return values:
669  *
670  *	EEXIST: this txg has already been synced, so there's nothing to to.
671  *		The caller should not log the write.
672  *
673  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
674  *		The caller should not log the write.
675  *
676  *	EALREADY: this block is already in the process of being synced.
677  *		The caller should track its progress (somehow).
678  *
679  *	EINPROGRESS: the IO has been initiated.
680  *		The caller should log this blkptr in the callback.
681  *
682  *	0: completed.  Sets *bp to the blkptr just written.
683  *		The caller should log this blkptr immediately.
684  */
685 int
686 dmu_sync(zio_t *pio, dmu_buf_t *db_fake,
687     blkptr_t *bp, uint64_t txg, dmu_sync_cb_t *done, void *arg)
688 {
689 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
690 	objset_impl_t *os = db->db_objset;
691 	dsl_pool_t *dp = os->os_dsl_dataset->ds_dir->dd_pool;
692 	tx_state_t *tx = &dp->dp_tx;
693 	dbuf_dirty_record_t *dr;
694 	dmu_sync_arg_t *in;
695 	zbookmark_t zb;
696 	zio_t *zio;
697 	int zio_flags;
698 	int err;
699 
700 	ASSERT(BP_IS_HOLE(bp));
701 	ASSERT(txg != 0);
702 
703 
704 	dprintf("dmu_sync txg=%llu, s,o,q %llu %llu %llu\n",
705 	    txg, tx->tx_synced_txg, tx->tx_open_txg, tx->tx_quiesced_txg);
706 
707 	/*
708 	 * XXX - would be nice if we could do this without suspending...
709 	 */
710 	txg_suspend(dp);
711 
712 	/*
713 	 * If this txg already synced, there's nothing to do.
714 	 */
715 	if (txg <= tx->tx_synced_txg) {
716 		txg_resume(dp);
717 		/*
718 		 * If we're running ziltest, we need the blkptr regardless.
719 		 */
720 		if (txg > spa_freeze_txg(dp->dp_spa)) {
721 			/* if db_blkptr == NULL, this was an empty write */
722 			if (db->db_blkptr)
723 				*bp = *db->db_blkptr; /* structure assignment */
724 			return (0);
725 		}
726 		return (EEXIST);
727 	}
728 
729 	mutex_enter(&db->db_mtx);
730 
731 	if (txg == tx->tx_syncing_txg) {
732 		while (db->db_data_pending) {
733 			/*
734 			 * IO is in-progress.  Wait for it to finish.
735 			 * XXX - would be nice to be able to somehow "attach"
736 			 * this zio to the parent zio passed in.
737 			 */
738 			cv_wait(&db->db_changed, &db->db_mtx);
739 			if (!db->db_data_pending &&
740 			    db->db_blkptr && BP_IS_HOLE(db->db_blkptr)) {
741 				/*
742 				 * IO was compressed away
743 				 */
744 				*bp = *db->db_blkptr; /* structure assignment */
745 				mutex_exit(&db->db_mtx);
746 				txg_resume(dp);
747 				return (0);
748 			}
749 			ASSERT(db->db_data_pending ||
750 			    (db->db_blkptr && db->db_blkptr->blk_birth == txg));
751 		}
752 
753 		if (db->db_blkptr && db->db_blkptr->blk_birth == txg) {
754 			/*
755 			 * IO is already completed.
756 			 */
757 			*bp = *db->db_blkptr; /* structure assignment */
758 			mutex_exit(&db->db_mtx);
759 			txg_resume(dp);
760 			return (0);
761 		}
762 	}
763 
764 	dr = db->db_last_dirty;
765 	while (dr && dr->dr_txg > txg)
766 		dr = dr->dr_next;
767 	if (dr == NULL || dr->dr_txg < txg) {
768 		/*
769 		 * This dbuf isn't dirty, must have been free_range'd.
770 		 * There's no need to log writes to freed blocks, so we're done.
771 		 */
772 		mutex_exit(&db->db_mtx);
773 		txg_resume(dp);
774 		return (ENOENT);
775 	}
776 
777 	ASSERT(dr->dr_txg == txg);
778 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
779 		/*
780 		 * We have already issued a sync write for this buffer.
781 		 */
782 		mutex_exit(&db->db_mtx);
783 		txg_resume(dp);
784 		return (EALREADY);
785 	} else if (dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
786 		/*
787 		 * This buffer has already been synced.  It could not
788 		 * have been dirtied since, or we would have cleared the state.
789 		 */
790 		*bp = dr->dt.dl.dr_overridden_by; /* structure assignment */
791 		mutex_exit(&db->db_mtx);
792 		txg_resume(dp);
793 		return (0);
794 	}
795 
796 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
797 	in = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
798 	in->dr = dr;
799 	in->done = done;
800 	in->arg = arg;
801 	mutex_exit(&db->db_mtx);
802 	txg_resume(dp);
803 
804 	zb.zb_objset = os->os_dsl_dataset->ds_object;
805 	zb.zb_object = db->db.db_object;
806 	zb.zb_level = db->db_level;
807 	zb.zb_blkid = db->db_blkid;
808 	zio_flags = ZIO_FLAG_MUSTSUCCEED;
809 	if (dmu_ot[db->db_dnode->dn_type].ot_metadata || zb.zb_level != 0)
810 		zio_flags |= ZIO_FLAG_METADATA;
811 	zio = arc_write(pio, os->os_spa,
812 	    zio_checksum_select(db->db_dnode->dn_checksum, os->os_checksum),
813 	    zio_compress_select(db->db_dnode->dn_compress, os->os_compress),
814 	    dmu_get_replication_level(os, &zb, db->db_dnode->dn_type),
815 	    txg, bp, dr->dt.dl.dr_data, NULL, dmu_sync_done, in,
816 	    ZIO_PRIORITY_SYNC_WRITE, zio_flags, &zb);
817 
818 	if (pio) {
819 		zio_nowait(zio);
820 		err = EINPROGRESS;
821 	} else {
822 		err = zio_wait(zio);
823 		ASSERT(err == 0);
824 	}
825 	return (err);
826 }
827 
828 int
829 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
830 	dmu_tx_t *tx)
831 {
832 	dnode_t *dn;
833 	int err;
834 
835 	err = dnode_hold(os->os, object, FTAG, &dn);
836 	if (err)
837 		return (err);
838 	err = dnode_set_blksz(dn, size, ibs, tx);
839 	dnode_rele(dn, FTAG);
840 	return (err);
841 }
842 
843 void
844 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
845 	dmu_tx_t *tx)
846 {
847 	dnode_t *dn;
848 
849 	/* XXX assumes dnode_hold will not get an i/o error */
850 	(void) dnode_hold(os->os, object, FTAG, &dn);
851 	ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
852 	dn->dn_checksum = checksum;
853 	dnode_setdirty(dn, tx);
854 	dnode_rele(dn, FTAG);
855 }
856 
857 void
858 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
859 	dmu_tx_t *tx)
860 {
861 	dnode_t *dn;
862 
863 	/* XXX assumes dnode_hold will not get an i/o error */
864 	(void) dnode_hold(os->os, object, FTAG, &dn);
865 	ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
866 	dn->dn_compress = compress;
867 	dnode_setdirty(dn, tx);
868 	dnode_rele(dn, FTAG);
869 }
870 
871 int
872 dmu_get_replication_level(objset_impl_t *os,
873     zbookmark_t *zb, dmu_object_type_t ot)
874 {
875 	int ncopies = os->os_copies;
876 
877 	/* If it's the mos, it should have max copies set. */
878 	ASSERT(zb->zb_objset != 0 ||
879 	    ncopies == spa_max_replication(os->os_spa));
880 
881 	if (dmu_ot[ot].ot_metadata || zb->zb_level != 0)
882 		ncopies++;
883 	return (MIN(ncopies, spa_max_replication(os->os_spa)));
884 }
885 
886 int
887 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
888 {
889 	dnode_t *dn;
890 	int i, err;
891 
892 	err = dnode_hold(os->os, object, FTAG, &dn);
893 	if (err)
894 		return (err);
895 	/*
896 	 * Sync any current changes before
897 	 * we go trundling through the block pointers.
898 	 */
899 	for (i = 0; i < TXG_SIZE; i++) {
900 		if (list_link_active(&dn->dn_dirty_link[i]))
901 			break;
902 	}
903 	if (i != TXG_SIZE) {
904 		dnode_rele(dn, FTAG);
905 		txg_wait_synced(dmu_objset_pool(os), 0);
906 		err = dnode_hold(os->os, object, FTAG, &dn);
907 		if (err)
908 			return (err);
909 	}
910 
911 	err = dnode_next_offset(dn, hole, off, 1, 1, 0);
912 	dnode_rele(dn, FTAG);
913 
914 	return (err);
915 }
916 
917 void
918 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
919 {
920 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
921 	mutex_enter(&dn->dn_mtx);
922 
923 	doi->doi_data_block_size = dn->dn_datablksz;
924 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
925 	    1ULL << dn->dn_indblkshift : 0;
926 	doi->doi_indirection = dn->dn_nlevels;
927 	doi->doi_checksum = dn->dn_checksum;
928 	doi->doi_compress = dn->dn_compress;
929 	doi->doi_physical_blks = (DN_USED_BYTES(dn->dn_phys) +
930 	    SPA_MINBLOCKSIZE/2) >> SPA_MINBLOCKSHIFT;
931 	doi->doi_max_block_offset = dn->dn_phys->dn_maxblkid;
932 	doi->doi_type = dn->dn_type;
933 	doi->doi_bonus_size = dn->dn_bonuslen;
934 	doi->doi_bonus_type = dn->dn_bonustype;
935 
936 	mutex_exit(&dn->dn_mtx);
937 	rw_exit(&dn->dn_struct_rwlock);
938 }
939 
940 /*
941  * Get information on a DMU object.
942  * If doi is NULL, just indicates whether the object exists.
943  */
944 int
945 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
946 {
947 	dnode_t *dn;
948 	int err = dnode_hold(os->os, object, FTAG, &dn);
949 
950 	if (err)
951 		return (err);
952 
953 	if (doi != NULL)
954 		dmu_object_info_from_dnode(dn, doi);
955 
956 	dnode_rele(dn, FTAG);
957 	return (0);
958 }
959 
960 /*
961  * As above, but faster; can be used when you have a held dbuf in hand.
962  */
963 void
964 dmu_object_info_from_db(dmu_buf_t *db, dmu_object_info_t *doi)
965 {
966 	dmu_object_info_from_dnode(((dmu_buf_impl_t *)db)->db_dnode, doi);
967 }
968 
969 /*
970  * Faster still when you only care about the size.
971  * This is specifically optimized for zfs_getattr().
972  */
973 void
974 dmu_object_size_from_db(dmu_buf_t *db, uint32_t *blksize, u_longlong_t *nblk512)
975 {
976 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
977 
978 	*blksize = dn->dn_datablksz;
979 	/* add 1 for dnode space */
980 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
981 	    SPA_MINBLOCKSHIFT) + 1;
982 }
983 
984 void
985 byteswap_uint64_array(void *vbuf, size_t size)
986 {
987 	uint64_t *buf = vbuf;
988 	size_t count = size >> 3;
989 	int i;
990 
991 	ASSERT((size & 7) == 0);
992 
993 	for (i = 0; i < count; i++)
994 		buf[i] = BSWAP_64(buf[i]);
995 }
996 
997 void
998 byteswap_uint32_array(void *vbuf, size_t size)
999 {
1000 	uint32_t *buf = vbuf;
1001 	size_t count = size >> 2;
1002 	int i;
1003 
1004 	ASSERT((size & 3) == 0);
1005 
1006 	for (i = 0; i < count; i++)
1007 		buf[i] = BSWAP_32(buf[i]);
1008 }
1009 
1010 void
1011 byteswap_uint16_array(void *vbuf, size_t size)
1012 {
1013 	uint16_t *buf = vbuf;
1014 	size_t count = size >> 1;
1015 	int i;
1016 
1017 	ASSERT((size & 1) == 0);
1018 
1019 	for (i = 0; i < count; i++)
1020 		buf[i] = BSWAP_16(buf[i]);
1021 }
1022 
1023 /* ARGSUSED */
1024 void
1025 byteswap_uint8_array(void *vbuf, size_t size)
1026 {
1027 }
1028 
1029 void
1030 dmu_init(void)
1031 {
1032 	dbuf_init();
1033 	dnode_init();
1034 	arc_init();
1035 }
1036 
1037 void
1038 dmu_fini(void)
1039 {
1040 	arc_fini();
1041 	dnode_fini();
1042 	dbuf_fini();
1043 }
1044