xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu.c (revision 1d8ccc7bfd8ab0f27a23c5253d95c22ceb9c27f4)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 #include <sys/dmu.h>
26 #include <sys/dmu_impl.h>
27 #include <sys/dmu_tx.h>
28 #include <sys/dbuf.h>
29 #include <sys/dnode.h>
30 #include <sys/zfs_context.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dmu_traverse.h>
33 #include <sys/dsl_dataset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_pool.h>
36 #include <sys/dsl_synctask.h>
37 #include <sys/dsl_prop.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/zfs_ioctl.h>
40 #include <sys/zap.h>
41 #include <sys/zio_checksum.h>
42 #include <sys/sa.h>
43 #ifdef _KERNEL
44 #include <sys/vmsystm.h>
45 #include <sys/zfs_znode.h>
46 #endif
47 
48 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
49 	{	byteswap_uint8_array,	TRUE,	"unallocated"		},
50 	{	zap_byteswap,		TRUE,	"object directory"	},
51 	{	byteswap_uint64_array,	TRUE,	"object array"		},
52 	{	byteswap_uint8_array,	TRUE,	"packed nvlist"		},
53 	{	byteswap_uint64_array,	TRUE,	"packed nvlist size"	},
54 	{	byteswap_uint64_array,	TRUE,	"bpobj"			},
55 	{	byteswap_uint64_array,	TRUE,	"bpobj 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,	"scan work queue"	},
88 	{	zap_byteswap,		TRUE,	"ZFS user/group used"	},
89 	{	zap_byteswap,		TRUE,	"ZFS user/group quota"	},
90 	{	zap_byteswap,		TRUE,	"snapshot refcount tags"},
91 	{	zap_byteswap,		TRUE,	"DDT ZAP algorithm"	},
92 	{	zap_byteswap,		TRUE,	"DDT statistics"	},
93 	{	byteswap_uint8_array,	TRUE,	"System attributes"	},
94 	{	zap_byteswap,		TRUE,	"SA master node"	},
95 	{	zap_byteswap,		TRUE,	"SA attr registration"	},
96 	{	zap_byteswap,		TRUE,	"SA attr layouts"	},
97 	{	zap_byteswap,		TRUE,	"scan translations"	},
98 	{	byteswap_uint8_array,	FALSE,	"deduplicated block"	},
99 	{	zap_byteswap,		TRUE,	"DSL deadlist map"	},
100 	{	byteswap_uint64_array,	TRUE,	"DSL deadlist map hdr"	},
101 	{	zap_byteswap,		TRUE,	"DSL dir clones"	},
102 	{	byteswap_uint64_array,	TRUE,	"bpobj subobj"		},
103 };
104 
105 int
106 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
107     void *tag, dmu_buf_t **dbp, int flags)
108 {
109 	dnode_t *dn;
110 	uint64_t blkid;
111 	dmu_buf_impl_t *db;
112 	int err;
113 	int db_flags = DB_RF_CANFAIL;
114 
115 	if (flags & DMU_READ_NO_PREFETCH)
116 		db_flags |= DB_RF_NOPREFETCH;
117 
118 	err = dnode_hold(os, object, FTAG, &dn);
119 	if (err)
120 		return (err);
121 	blkid = dbuf_whichblock(dn, offset);
122 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
123 	db = dbuf_hold(dn, blkid, tag);
124 	rw_exit(&dn->dn_struct_rwlock);
125 	if (db == NULL) {
126 		err = EIO;
127 	} else {
128 		err = dbuf_read(db, NULL, db_flags);
129 		if (err) {
130 			dbuf_rele(db, tag);
131 			db = NULL;
132 		}
133 	}
134 
135 	dnode_rele(dn, FTAG);
136 	*dbp = &db->db;
137 	return (err);
138 }
139 
140 int
141 dmu_bonus_max(void)
142 {
143 	return (DN_MAX_BONUSLEN);
144 }
145 
146 int
147 dmu_set_bonus(dmu_buf_t *db, int newsize, dmu_tx_t *tx)
148 {
149 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
150 
151 	if (dn->dn_bonus != (dmu_buf_impl_t *)db)
152 		return (EINVAL);
153 	if (newsize < 0 || newsize > db->db_size)
154 		return (EINVAL);
155 	dnode_setbonuslen(dn, newsize, tx);
156 	return (0);
157 }
158 
159 int
160 dmu_set_bonustype(dmu_buf_t *db, dmu_object_type_t type, dmu_tx_t *tx)
161 {
162 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
163 
164 	if (type > DMU_OT_NUMTYPES)
165 		return (EINVAL);
166 
167 	if (dn->dn_bonus != (dmu_buf_impl_t *)db)
168 		return (EINVAL);
169 
170 	dnode_setbonus_type(dn, type, tx);
171 	return (0);
172 }
173 
174 int
175 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
176 {
177 	dnode_t *dn;
178 	int error;
179 
180 	error = dnode_hold(os, object, FTAG, &dn);
181 	dbuf_rm_spill(dn, tx);
182 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
183 	dnode_rm_spill(dn, tx);
184 	rw_exit(&dn->dn_struct_rwlock);
185 	dnode_rele(dn, FTAG);
186 	return (error);
187 }
188 
189 /*
190  * returns ENOENT, EIO, or 0.
191  */
192 int
193 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
194 {
195 	dnode_t *dn;
196 	dmu_buf_impl_t *db;
197 	int error;
198 
199 	error = dnode_hold(os, object, FTAG, &dn);
200 	if (error)
201 		return (error);
202 
203 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
204 	if (dn->dn_bonus == NULL) {
205 		rw_exit(&dn->dn_struct_rwlock);
206 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
207 		if (dn->dn_bonus == NULL)
208 			dbuf_create_bonus(dn);
209 	}
210 	db = dn->dn_bonus;
211 	rw_exit(&dn->dn_struct_rwlock);
212 
213 	/* as long as the bonus buf is held, the dnode will be held */
214 	if (refcount_add(&db->db_holds, tag) == 1)
215 		VERIFY(dnode_add_ref(dn, db));
216 
217 	dnode_rele(dn, FTAG);
218 
219 	VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
220 
221 	*dbp = &db->db;
222 	return (0);
223 }
224 
225 /*
226  * returns ENOENT, EIO, or 0.
227  *
228  * This interface will allocate a blank spill dbuf when a spill blk
229  * doesn't already exist on the dnode.
230  *
231  * if you only want to find an already existing spill db, then
232  * dmu_spill_hold_existing() should be used.
233  */
234 int
235 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
236 {
237 	dmu_buf_impl_t *db = NULL;
238 	int err;
239 
240 	if ((flags & DB_RF_HAVESTRUCT) == 0)
241 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
242 
243 	db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
244 
245 	if ((flags & DB_RF_HAVESTRUCT) == 0)
246 		rw_exit(&dn->dn_struct_rwlock);
247 
248 	ASSERT(db != NULL);
249 	err = dbuf_read(db, NULL, flags);
250 	if (err == 0)
251 		*dbp = &db->db;
252 	else
253 		dbuf_rele(db, tag);
254 	return (err);
255 }
256 
257 int
258 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
259 {
260 	dnode_t *dn = ((dmu_buf_impl_t *)bonus)->db_dnode;
261 	int err;
262 
263 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA)
264 		return (EINVAL);
265 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
266 
267 	if (!dn->dn_have_spill) {
268 		rw_exit(&dn->dn_struct_rwlock);
269 		return (ENOENT);
270 	}
271 	err = dmu_spill_hold_by_dnode(dn,
272 	    DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
273 	rw_exit(&dn->dn_struct_rwlock);
274 	return (err);
275 }
276 
277 int
278 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
279 {
280 	return (dmu_spill_hold_by_dnode(((dmu_buf_impl_t *)bonus)->db_dnode,
281 	    DB_RF_CANFAIL, tag, dbp));
282 }
283 
284 /*
285  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
286  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
287  * and can induce severe lock contention when writing to several files
288  * whose dnodes are in the same block.
289  */
290 static int
291 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
292     int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
293 {
294 	dsl_pool_t *dp = NULL;
295 	dmu_buf_t **dbp;
296 	uint64_t blkid, nblks, i;
297 	uint32_t dbuf_flags;
298 	int err;
299 	zio_t *zio;
300 	hrtime_t start;
301 
302 	ASSERT(length <= DMU_MAX_ACCESS);
303 
304 	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
305 	if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
306 		dbuf_flags |= DB_RF_NOPREFETCH;
307 
308 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
309 	if (dn->dn_datablkshift) {
310 		int blkshift = dn->dn_datablkshift;
311 		nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
312 		    P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
313 	} else {
314 		if (offset + length > dn->dn_datablksz) {
315 			zfs_panic_recover("zfs: accessing past end of object "
316 			    "%llx/%llx (size=%u access=%llu+%llu)",
317 			    (longlong_t)dn->dn_objset->
318 			    os_dsl_dataset->ds_object,
319 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
320 			    (longlong_t)offset, (longlong_t)length);
321 			rw_exit(&dn->dn_struct_rwlock);
322 			return (EIO);
323 		}
324 		nblks = 1;
325 	}
326 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
327 
328 	if (dn->dn_objset->os_dsl_dataset)
329 		dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool;
330 	if (dp && dsl_pool_sync_context(dp))
331 		start = gethrtime();
332 	zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
333 	blkid = dbuf_whichblock(dn, offset);
334 	for (i = 0; i < nblks; i++) {
335 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
336 		if (db == NULL) {
337 			rw_exit(&dn->dn_struct_rwlock);
338 			dmu_buf_rele_array(dbp, nblks, tag);
339 			zio_nowait(zio);
340 			return (EIO);
341 		}
342 		/* initiate async i/o */
343 		if (read) {
344 			(void) dbuf_read(db, zio, dbuf_flags);
345 		}
346 		dbp[i] = &db->db;
347 	}
348 	rw_exit(&dn->dn_struct_rwlock);
349 
350 	/* wait for async i/o */
351 	err = zio_wait(zio);
352 	/* track read overhead when we are in sync context */
353 	if (dp && dsl_pool_sync_context(dp))
354 		dp->dp_read_overhead += gethrtime() - start;
355 	if (err) {
356 		dmu_buf_rele_array(dbp, nblks, tag);
357 		return (err);
358 	}
359 
360 	/* wait for other io to complete */
361 	if (read) {
362 		for (i = 0; i < nblks; i++) {
363 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
364 			mutex_enter(&db->db_mtx);
365 			while (db->db_state == DB_READ ||
366 			    db->db_state == DB_FILL)
367 				cv_wait(&db->db_changed, &db->db_mtx);
368 			if (db->db_state == DB_UNCACHED)
369 				err = EIO;
370 			mutex_exit(&db->db_mtx);
371 			if (err) {
372 				dmu_buf_rele_array(dbp, nblks, tag);
373 				return (err);
374 			}
375 		}
376 	}
377 
378 	*numbufsp = nblks;
379 	*dbpp = dbp;
380 	return (0);
381 }
382 
383 static int
384 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
385     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
386 {
387 	dnode_t *dn;
388 	int err;
389 
390 	err = dnode_hold(os, object, FTAG, &dn);
391 	if (err)
392 		return (err);
393 
394 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
395 	    numbufsp, dbpp, DMU_READ_PREFETCH);
396 
397 	dnode_rele(dn, FTAG);
398 
399 	return (err);
400 }
401 
402 int
403 dmu_buf_hold_array_by_bonus(dmu_buf_t *db, uint64_t offset,
404     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
405 {
406 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
407 	int err;
408 
409 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
410 	    numbufsp, dbpp, DMU_READ_PREFETCH);
411 
412 	return (err);
413 }
414 
415 void
416 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
417 {
418 	int i;
419 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
420 
421 	if (numbufs == 0)
422 		return;
423 
424 	for (i = 0; i < numbufs; i++) {
425 		if (dbp[i])
426 			dbuf_rele(dbp[i], tag);
427 	}
428 
429 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
430 }
431 
432 void
433 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
434 {
435 	dnode_t *dn;
436 	uint64_t blkid;
437 	int nblks, i, err;
438 
439 	if (zfs_prefetch_disable)
440 		return;
441 
442 	if (len == 0) {  /* they're interested in the bonus buffer */
443 		dn = os->os_meta_dnode;
444 
445 		if (object == 0 || object >= DN_MAX_OBJECT)
446 			return;
447 
448 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
449 		blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
450 		dbuf_prefetch(dn, blkid);
451 		rw_exit(&dn->dn_struct_rwlock);
452 		return;
453 	}
454 
455 	/*
456 	 * XXX - Note, if the dnode for the requested object is not
457 	 * already cached, we will do a *synchronous* read in the
458 	 * dnode_hold() call.  The same is true for any indirects.
459 	 */
460 	err = dnode_hold(os, object, FTAG, &dn);
461 	if (err != 0)
462 		return;
463 
464 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
465 	if (dn->dn_datablkshift) {
466 		int blkshift = dn->dn_datablkshift;
467 		nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
468 		    P2ALIGN(offset, 1<<blkshift)) >> blkshift;
469 	} else {
470 		nblks = (offset < dn->dn_datablksz);
471 	}
472 
473 	if (nblks != 0) {
474 		blkid = dbuf_whichblock(dn, offset);
475 		for (i = 0; i < nblks; i++)
476 			dbuf_prefetch(dn, blkid+i);
477 	}
478 
479 	rw_exit(&dn->dn_struct_rwlock);
480 
481 	dnode_rele(dn, FTAG);
482 }
483 
484 /*
485  * Get the next "chunk" of file data to free.  We traverse the file from
486  * the end so that the file gets shorter over time (if we crashes in the
487  * middle, this will leave us in a better state).  We find allocated file
488  * data by simply searching the allocated level 1 indirects.
489  */
490 static int
491 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit)
492 {
493 	uint64_t len = *start - limit;
494 	uint64_t blkcnt = 0;
495 	uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1));
496 	uint64_t iblkrange =
497 	    dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
498 
499 	ASSERT(limit <= *start);
500 
501 	if (len <= iblkrange * maxblks) {
502 		*start = limit;
503 		return (0);
504 	}
505 	ASSERT(ISP2(iblkrange));
506 
507 	while (*start > limit && blkcnt < maxblks) {
508 		int err;
509 
510 		/* find next allocated L1 indirect */
511 		err = dnode_next_offset(dn,
512 		    DNODE_FIND_BACKWARDS, start, 2, 1, 0);
513 
514 		/* if there are no more, then we are done */
515 		if (err == ESRCH) {
516 			*start = limit;
517 			return (0);
518 		} else if (err) {
519 			return (err);
520 		}
521 		blkcnt += 1;
522 
523 		/* reset offset to end of "next" block back */
524 		*start = P2ALIGN(*start, iblkrange);
525 		if (*start <= limit)
526 			*start = limit;
527 		else
528 			*start -= 1;
529 	}
530 	return (0);
531 }
532 
533 static int
534 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
535     uint64_t length, boolean_t free_dnode)
536 {
537 	dmu_tx_t *tx;
538 	uint64_t object_size, start, end, len;
539 	boolean_t trunc = (length == DMU_OBJECT_END);
540 	int align, err;
541 
542 	align = 1 << dn->dn_datablkshift;
543 	ASSERT(align > 0);
544 	object_size = align == 1 ? dn->dn_datablksz :
545 	    (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
546 
547 	end = offset + length;
548 	if (trunc || end > object_size)
549 		end = object_size;
550 	if (end <= offset)
551 		return (0);
552 	length = end - offset;
553 
554 	while (length) {
555 		start = end;
556 		/* assert(offset <= start) */
557 		err = get_next_chunk(dn, &start, offset);
558 		if (err)
559 			return (err);
560 		len = trunc ? DMU_OBJECT_END : end - start;
561 
562 		tx = dmu_tx_create(os);
563 		dmu_tx_hold_free(tx, dn->dn_object, start, len);
564 		err = dmu_tx_assign(tx, TXG_WAIT);
565 		if (err) {
566 			dmu_tx_abort(tx);
567 			return (err);
568 		}
569 
570 		dnode_free_range(dn, start, trunc ? -1 : len, tx);
571 
572 		if (start == 0 && free_dnode) {
573 			ASSERT(trunc);
574 			dnode_free(dn, tx);
575 		}
576 
577 		length -= end - start;
578 
579 		dmu_tx_commit(tx);
580 		end = start;
581 	}
582 	return (0);
583 }
584 
585 int
586 dmu_free_long_range(objset_t *os, uint64_t object,
587     uint64_t offset, uint64_t length)
588 {
589 	dnode_t *dn;
590 	int err;
591 
592 	err = dnode_hold(os, object, FTAG, &dn);
593 	if (err != 0)
594 		return (err);
595 	err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
596 	dnode_rele(dn, FTAG);
597 	return (err);
598 }
599 
600 int
601 dmu_free_object(objset_t *os, uint64_t object)
602 {
603 	dnode_t *dn;
604 	dmu_tx_t *tx;
605 	int err;
606 
607 	err = dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED,
608 	    FTAG, &dn);
609 	if (err != 0)
610 		return (err);
611 	if (dn->dn_nlevels == 1) {
612 		tx = dmu_tx_create(os);
613 		dmu_tx_hold_bonus(tx, object);
614 		dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
615 		err = dmu_tx_assign(tx, TXG_WAIT);
616 		if (err == 0) {
617 			dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
618 			dnode_free(dn, tx);
619 			dmu_tx_commit(tx);
620 		} else {
621 			dmu_tx_abort(tx);
622 		}
623 	} else {
624 		err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
625 	}
626 	dnode_rele(dn, FTAG);
627 	return (err);
628 }
629 
630 int
631 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
632     uint64_t size, dmu_tx_t *tx)
633 {
634 	dnode_t *dn;
635 	int err = dnode_hold(os, object, FTAG, &dn);
636 	if (err)
637 		return (err);
638 	ASSERT(offset < UINT64_MAX);
639 	ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
640 	dnode_free_range(dn, offset, size, tx);
641 	dnode_rele(dn, FTAG);
642 	return (0);
643 }
644 
645 int
646 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
647     void *buf, uint32_t flags)
648 {
649 	dnode_t *dn;
650 	dmu_buf_t **dbp;
651 	int numbufs, err;
652 
653 	err = dnode_hold(os, object, FTAG, &dn);
654 	if (err)
655 		return (err);
656 
657 	/*
658 	 * Deal with odd block sizes, where there can't be data past the first
659 	 * block.  If we ever do the tail block optimization, we will need to
660 	 * handle that here as well.
661 	 */
662 	if (dn->dn_maxblkid == 0) {
663 		int newsz = offset > dn->dn_datablksz ? 0 :
664 		    MIN(size, dn->dn_datablksz - offset);
665 		bzero((char *)buf + newsz, size - newsz);
666 		size = newsz;
667 	}
668 
669 	while (size > 0) {
670 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
671 		int i;
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_by_dnode(dn, offset, mylen,
678 		    TRUE, FTAG, &numbufs, &dbp, flags);
679 		if (err)
680 			break;
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 = offset - db->db_offset;
690 			tocpy = (int)MIN(db->db_size - bufoff, size);
691 
692 			bcopy((char *)db->db_data + bufoff, buf, tocpy);
693 
694 			offset += tocpy;
695 			size -= tocpy;
696 			buf = (char *)buf + tocpy;
697 		}
698 		dmu_buf_rele_array(dbp, numbufs, FTAG);
699 	}
700 	dnode_rele(dn, FTAG);
701 	return (err);
702 }
703 
704 void
705 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
706     const void *buf, dmu_tx_t *tx)
707 {
708 	dmu_buf_t **dbp;
709 	int numbufs, i;
710 
711 	if (size == 0)
712 		return;
713 
714 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
715 	    FALSE, FTAG, &numbufs, &dbp));
716 
717 	for (i = 0; i < numbufs; i++) {
718 		int tocpy;
719 		int bufoff;
720 		dmu_buf_t *db = dbp[i];
721 
722 		ASSERT(size > 0);
723 
724 		bufoff = offset - db->db_offset;
725 		tocpy = (int)MIN(db->db_size - bufoff, size);
726 
727 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
728 
729 		if (tocpy == db->db_size)
730 			dmu_buf_will_fill(db, tx);
731 		else
732 			dmu_buf_will_dirty(db, tx);
733 
734 		bcopy(buf, (char *)db->db_data + bufoff, tocpy);
735 
736 		if (tocpy == db->db_size)
737 			dmu_buf_fill_done(db, tx);
738 
739 		offset += tocpy;
740 		size -= tocpy;
741 		buf = (char *)buf + tocpy;
742 	}
743 	dmu_buf_rele_array(dbp, numbufs, FTAG);
744 }
745 
746 void
747 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
748     dmu_tx_t *tx)
749 {
750 	dmu_buf_t **dbp;
751 	int numbufs, i;
752 
753 	if (size == 0)
754 		return;
755 
756 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
757 	    FALSE, FTAG, &numbufs, &dbp));
758 
759 	for (i = 0; i < numbufs; i++) {
760 		dmu_buf_t *db = dbp[i];
761 
762 		dmu_buf_will_not_fill(db, tx);
763 	}
764 	dmu_buf_rele_array(dbp, numbufs, FTAG);
765 }
766 
767 /*
768  * DMU support for xuio
769  */
770 kstat_t *xuio_ksp = NULL;
771 
772 int
773 dmu_xuio_init(xuio_t *xuio, int nblk)
774 {
775 	dmu_xuio_t *priv;
776 	uio_t *uio = &xuio->xu_uio;
777 
778 	uio->uio_iovcnt = nblk;
779 	uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
780 
781 	priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
782 	priv->cnt = nblk;
783 	priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
784 	priv->iovp = uio->uio_iov;
785 	XUIO_XUZC_PRIV(xuio) = priv;
786 
787 	if (XUIO_XUZC_RW(xuio) == UIO_READ)
788 		XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
789 	else
790 		XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
791 
792 	return (0);
793 }
794 
795 void
796 dmu_xuio_fini(xuio_t *xuio)
797 {
798 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
799 	int nblk = priv->cnt;
800 
801 	kmem_free(priv->iovp, nblk * sizeof (iovec_t));
802 	kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
803 	kmem_free(priv, sizeof (dmu_xuio_t));
804 
805 	if (XUIO_XUZC_RW(xuio) == UIO_READ)
806 		XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
807 	else
808 		XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
809 }
810 
811 /*
812  * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
813  * and increase priv->next by 1.
814  */
815 int
816 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
817 {
818 	struct iovec *iov;
819 	uio_t *uio = &xuio->xu_uio;
820 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
821 	int i = priv->next++;
822 
823 	ASSERT(i < priv->cnt);
824 	ASSERT(off + n <= arc_buf_size(abuf));
825 	iov = uio->uio_iov + i;
826 	iov->iov_base = (char *)abuf->b_data + off;
827 	iov->iov_len = n;
828 	priv->bufs[i] = abuf;
829 	return (0);
830 }
831 
832 int
833 dmu_xuio_cnt(xuio_t *xuio)
834 {
835 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
836 	return (priv->cnt);
837 }
838 
839 arc_buf_t *
840 dmu_xuio_arcbuf(xuio_t *xuio, int i)
841 {
842 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
843 
844 	ASSERT(i < priv->cnt);
845 	return (priv->bufs[i]);
846 }
847 
848 void
849 dmu_xuio_clear(xuio_t *xuio, int i)
850 {
851 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
852 
853 	ASSERT(i < priv->cnt);
854 	priv->bufs[i] = NULL;
855 }
856 
857 static void
858 xuio_stat_init(void)
859 {
860 	xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
861 	    KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
862 	    KSTAT_FLAG_VIRTUAL);
863 	if (xuio_ksp != NULL) {
864 		xuio_ksp->ks_data = &xuio_stats;
865 		kstat_install(xuio_ksp);
866 	}
867 }
868 
869 static void
870 xuio_stat_fini(void)
871 {
872 	if (xuio_ksp != NULL) {
873 		kstat_delete(xuio_ksp);
874 		xuio_ksp = NULL;
875 	}
876 }
877 
878 void
879 xuio_stat_wbuf_copied()
880 {
881 	XUIOSTAT_BUMP(xuiostat_wbuf_copied);
882 }
883 
884 void
885 xuio_stat_wbuf_nocopy()
886 {
887 	XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
888 }
889 
890 #ifdef _KERNEL
891 int
892 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
893 {
894 	dmu_buf_t **dbp;
895 	int numbufs, i, err;
896 	xuio_t *xuio = NULL;
897 
898 	/*
899 	 * NB: we could do this block-at-a-time, but it's nice
900 	 * to be reading in parallel.
901 	 */
902 	err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
903 	    &numbufs, &dbp);
904 	if (err)
905 		return (err);
906 
907 	if (uio->uio_extflg == UIO_XUIO)
908 		xuio = (xuio_t *)uio;
909 
910 	for (i = 0; i < numbufs; i++) {
911 		int tocpy;
912 		int bufoff;
913 		dmu_buf_t *db = dbp[i];
914 
915 		ASSERT(size > 0);
916 
917 		bufoff = uio->uio_loffset - db->db_offset;
918 		tocpy = (int)MIN(db->db_size - bufoff, size);
919 
920 		if (xuio) {
921 			dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
922 			arc_buf_t *dbuf_abuf = dbi->db_buf;
923 			arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
924 			err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
925 			if (!err) {
926 				uio->uio_resid -= tocpy;
927 				uio->uio_loffset += tocpy;
928 			}
929 
930 			if (abuf == dbuf_abuf)
931 				XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
932 			else
933 				XUIOSTAT_BUMP(xuiostat_rbuf_copied);
934 		} else {
935 			err = uiomove((char *)db->db_data + bufoff, tocpy,
936 			    UIO_READ, uio);
937 		}
938 		if (err)
939 			break;
940 
941 		size -= tocpy;
942 	}
943 	dmu_buf_rele_array(dbp, numbufs, FTAG);
944 
945 	return (err);
946 }
947 
948 static int
949 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
950 {
951 	dmu_buf_t **dbp;
952 	int numbufs;
953 	int err = 0;
954 	int i;
955 
956 	err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
957 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
958 	if (err)
959 		return (err);
960 
961 	for (i = 0; i < numbufs; i++) {
962 		int tocpy;
963 		int bufoff;
964 		dmu_buf_t *db = dbp[i];
965 
966 		ASSERT(size > 0);
967 
968 		bufoff = uio->uio_loffset - db->db_offset;
969 		tocpy = (int)MIN(db->db_size - bufoff, size);
970 
971 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
972 
973 		if (tocpy == db->db_size)
974 			dmu_buf_will_fill(db, tx);
975 		else
976 			dmu_buf_will_dirty(db, tx);
977 
978 		/*
979 		 * XXX uiomove could block forever (eg. nfs-backed
980 		 * pages).  There needs to be a uiolockdown() function
981 		 * to lock the pages in memory, so that uiomove won't
982 		 * block.
983 		 */
984 		err = uiomove((char *)db->db_data + bufoff, tocpy,
985 		    UIO_WRITE, uio);
986 
987 		if (tocpy == db->db_size)
988 			dmu_buf_fill_done(db, tx);
989 
990 		if (err)
991 			break;
992 
993 		size -= tocpy;
994 	}
995 
996 	dmu_buf_rele_array(dbp, numbufs, FTAG);
997 	return (err);
998 }
999 
1000 int
1001 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1002     dmu_tx_t *tx)
1003 {
1004 	if (size == 0)
1005 		return (0);
1006 
1007 	return (dmu_write_uio_dnode(((dmu_buf_impl_t *)zdb)->db_dnode,
1008 	    uio, size, tx));
1009 }
1010 
1011 int
1012 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1013     dmu_tx_t *tx)
1014 {
1015 	dnode_t *dn;
1016 	int err;
1017 
1018 	if (size == 0)
1019 		return (0);
1020 
1021 	err = dnode_hold(os, object, FTAG, &dn);
1022 	if (err)
1023 		return (err);
1024 
1025 	err = dmu_write_uio_dnode(dn, uio, size, tx);
1026 
1027 	dnode_rele(dn, FTAG);
1028 
1029 	return (err);
1030 }
1031 
1032 int
1033 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1034     page_t *pp, dmu_tx_t *tx)
1035 {
1036 	dmu_buf_t **dbp;
1037 	int numbufs, i;
1038 	int err;
1039 
1040 	if (size == 0)
1041 		return (0);
1042 
1043 	err = dmu_buf_hold_array(os, object, offset, size,
1044 	    FALSE, FTAG, &numbufs, &dbp);
1045 	if (err)
1046 		return (err);
1047 
1048 	for (i = 0; i < numbufs; i++) {
1049 		int tocpy, copied, thiscpy;
1050 		int bufoff;
1051 		dmu_buf_t *db = dbp[i];
1052 		caddr_t va;
1053 
1054 		ASSERT(size > 0);
1055 		ASSERT3U(db->db_size, >=, PAGESIZE);
1056 
1057 		bufoff = offset - db->db_offset;
1058 		tocpy = (int)MIN(db->db_size - bufoff, size);
1059 
1060 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1061 
1062 		if (tocpy == db->db_size)
1063 			dmu_buf_will_fill(db, tx);
1064 		else
1065 			dmu_buf_will_dirty(db, tx);
1066 
1067 		for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1068 			ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1069 			thiscpy = MIN(PAGESIZE, tocpy - copied);
1070 			va = zfs_map_page(pp, S_READ);
1071 			bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1072 			zfs_unmap_page(pp, va);
1073 			pp = pp->p_next;
1074 			bufoff += PAGESIZE;
1075 		}
1076 
1077 		if (tocpy == db->db_size)
1078 			dmu_buf_fill_done(db, tx);
1079 
1080 		offset += tocpy;
1081 		size -= tocpy;
1082 	}
1083 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1084 	return (err);
1085 }
1086 #endif
1087 
1088 /*
1089  * Allocate a loaned anonymous arc buffer.
1090  */
1091 arc_buf_t *
1092 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1093 {
1094 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
1095 
1096 	return (arc_loan_buf(dn->dn_objset->os_spa, size));
1097 }
1098 
1099 /*
1100  * Free a loaned arc buffer.
1101  */
1102 void
1103 dmu_return_arcbuf(arc_buf_t *buf)
1104 {
1105 	arc_return_buf(buf, FTAG);
1106 	VERIFY(arc_buf_remove_ref(buf, FTAG) == 1);
1107 }
1108 
1109 /*
1110  * When possible directly assign passed loaned arc buffer to a dbuf.
1111  * If this is not possible copy the contents of passed arc buf via
1112  * dmu_write().
1113  */
1114 void
1115 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1116     dmu_tx_t *tx)
1117 {
1118 	dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
1119 	dmu_buf_impl_t *db;
1120 	uint32_t blksz = (uint32_t)arc_buf_size(buf);
1121 	uint64_t blkid;
1122 
1123 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1124 	blkid = dbuf_whichblock(dn, offset);
1125 	VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1126 	rw_exit(&dn->dn_struct_rwlock);
1127 
1128 	if (offset == db->db.db_offset && blksz == db->db.db_size) {
1129 		dbuf_assign_arcbuf(db, buf, tx);
1130 		dbuf_rele(db, FTAG);
1131 	} else {
1132 		dbuf_rele(db, FTAG);
1133 		dmu_write(dn->dn_objset, dn->dn_object, offset, blksz,
1134 		    buf->b_data, tx);
1135 		dmu_return_arcbuf(buf);
1136 		XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1137 	}
1138 }
1139 
1140 typedef struct {
1141 	dbuf_dirty_record_t	*dsa_dr;
1142 	dmu_sync_cb_t		*dsa_done;
1143 	zgd_t			*dsa_zgd;
1144 	dmu_tx_t		*dsa_tx;
1145 } dmu_sync_arg_t;
1146 
1147 /* ARGSUSED */
1148 static void
1149 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1150 {
1151 	dmu_sync_arg_t *dsa = varg;
1152 	dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1153 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1154 	blkptr_t *bp = zio->io_bp;
1155 
1156 	if (zio->io_error == 0) {
1157 		if (BP_IS_HOLE(bp)) {
1158 			/*
1159 			 * A block of zeros may compress to a hole, but the
1160 			 * block size still needs to be known for replay.
1161 			 */
1162 			BP_SET_LSIZE(bp, db->db_size);
1163 		} else {
1164 			ASSERT(BP_GET_TYPE(bp) == dn->dn_type);
1165 			ASSERT(BP_GET_LEVEL(bp) == 0);
1166 			bp->blk_fill = 1;
1167 		}
1168 	}
1169 }
1170 
1171 static void
1172 dmu_sync_late_arrival_ready(zio_t *zio)
1173 {
1174 	dmu_sync_ready(zio, NULL, zio->io_private);
1175 }
1176 
1177 /* ARGSUSED */
1178 static void
1179 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1180 {
1181 	dmu_sync_arg_t *dsa = varg;
1182 	dbuf_dirty_record_t *dr = dsa->dsa_dr;
1183 	dmu_buf_impl_t *db = dr->dr_dbuf;
1184 
1185 	mutex_enter(&db->db_mtx);
1186 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1187 	if (zio->io_error == 0) {
1188 		dr->dt.dl.dr_overridden_by = *zio->io_bp;
1189 		dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1190 		dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1191 		if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1192 			BP_ZERO(&dr->dt.dl.dr_overridden_by);
1193 	} else {
1194 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1195 	}
1196 	cv_broadcast(&db->db_changed);
1197 	mutex_exit(&db->db_mtx);
1198 
1199 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1200 
1201 	kmem_free(dsa, sizeof (*dsa));
1202 }
1203 
1204 static void
1205 dmu_sync_late_arrival_done(zio_t *zio)
1206 {
1207 	blkptr_t *bp = zio->io_bp;
1208 	dmu_sync_arg_t *dsa = zio->io_private;
1209 
1210 	if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1211 		ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1212 		ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1213 		zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1214 	}
1215 
1216 	dmu_tx_commit(dsa->dsa_tx);
1217 
1218 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1219 
1220 	kmem_free(dsa, sizeof (*dsa));
1221 }
1222 
1223 static int
1224 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1225     zio_prop_t *zp, zbookmark_t *zb)
1226 {
1227 	dmu_sync_arg_t *dsa;
1228 	dmu_tx_t *tx;
1229 
1230 	tx = dmu_tx_create(os);
1231 	dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1232 	if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1233 		dmu_tx_abort(tx);
1234 		return (EIO);	/* Make zl_get_data do txg_waited_synced() */
1235 	}
1236 
1237 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1238 	dsa->dsa_dr = NULL;
1239 	dsa->dsa_done = done;
1240 	dsa->dsa_zgd = zgd;
1241 	dsa->dsa_tx = tx;
1242 
1243 	zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1244 	    zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1245 	    dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1246 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1247 
1248 	return (0);
1249 }
1250 
1251 /*
1252  * Intent log support: sync the block associated with db to disk.
1253  * N.B. and XXX: the caller is responsible for making sure that the
1254  * data isn't changing while dmu_sync() is writing it.
1255  *
1256  * Return values:
1257  *
1258  *	EEXIST: this txg has already been synced, so there's nothing to to.
1259  *		The caller should not log the write.
1260  *
1261  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1262  *		The caller should not log the write.
1263  *
1264  *	EALREADY: this block is already in the process of being synced.
1265  *		The caller should track its progress (somehow).
1266  *
1267  *	EIO: could not do the I/O.
1268  *		The caller should do a txg_wait_synced().
1269  *
1270  *	0: the I/O has been initiated.
1271  *		The caller should log this blkptr in the done callback.
1272  *		It is possible that the I/O will fail, in which case
1273  *		the error will be reported to the done callback and
1274  *		propagated to pio from zio_done().
1275  */
1276 int
1277 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1278 {
1279 	blkptr_t *bp = zgd->zgd_bp;
1280 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1281 	objset_t *os = db->db_objset;
1282 	dsl_dataset_t *ds = os->os_dsl_dataset;
1283 	dbuf_dirty_record_t *dr;
1284 	dmu_sync_arg_t *dsa;
1285 	zbookmark_t zb;
1286 	zio_prop_t zp;
1287 
1288 	ASSERT(pio != NULL);
1289 	ASSERT(BP_IS_HOLE(bp));
1290 	ASSERT(txg != 0);
1291 
1292 	SET_BOOKMARK(&zb, ds->ds_object,
1293 	    db->db.db_object, db->db_level, db->db_blkid);
1294 
1295 	dmu_write_policy(os, db->db_dnode, db->db_level, WP_DMU_SYNC, &zp);
1296 
1297 	/*
1298 	 * If we're frozen (running ziltest), we always need to generate a bp.
1299 	 */
1300 	if (txg > spa_freeze_txg(os->os_spa))
1301 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1302 
1303 	/*
1304 	 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1305 	 * and us.  If we determine that this txg is not yet syncing,
1306 	 * but it begins to sync a moment later, that's OK because the
1307 	 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1308 	 */
1309 	mutex_enter(&db->db_mtx);
1310 
1311 	if (txg <= spa_last_synced_txg(os->os_spa)) {
1312 		/*
1313 		 * This txg has already synced.  There's nothing to do.
1314 		 */
1315 		mutex_exit(&db->db_mtx);
1316 		return (EEXIST);
1317 	}
1318 
1319 	if (txg <= spa_syncing_txg(os->os_spa)) {
1320 		/*
1321 		 * This txg is currently syncing, so we can't mess with
1322 		 * the dirty record anymore; just write a new log block.
1323 		 */
1324 		mutex_exit(&db->db_mtx);
1325 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1326 	}
1327 
1328 	dr = db->db_last_dirty;
1329 	while (dr && dr->dr_txg != txg)
1330 		dr = dr->dr_next;
1331 
1332 	if (dr == NULL) {
1333 		/*
1334 		 * There's no dr for this dbuf, so it must have been freed.
1335 		 * There's no need to log writes to freed blocks, so we're done.
1336 		 */
1337 		mutex_exit(&db->db_mtx);
1338 		return (ENOENT);
1339 	}
1340 
1341 	ASSERT(dr->dr_txg == txg);
1342 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1343 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1344 		/*
1345 		 * We have already issued a sync write for this buffer,
1346 		 * or this buffer has already been synced.  It could not
1347 		 * have been dirtied since, or we would have cleared the state.
1348 		 */
1349 		mutex_exit(&db->db_mtx);
1350 		return (EALREADY);
1351 	}
1352 
1353 	ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1354 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1355 	mutex_exit(&db->db_mtx);
1356 
1357 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1358 	dsa->dsa_dr = dr;
1359 	dsa->dsa_done = done;
1360 	dsa->dsa_zgd = zgd;
1361 	dsa->dsa_tx = NULL;
1362 
1363 	zio_nowait(arc_write(pio, os->os_spa, txg,
1364 	    bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp,
1365 	    dmu_sync_ready, dmu_sync_done, dsa,
1366 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1367 
1368 	return (0);
1369 }
1370 
1371 int
1372 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1373 	dmu_tx_t *tx)
1374 {
1375 	dnode_t *dn;
1376 	int err;
1377 
1378 	err = dnode_hold(os, object, FTAG, &dn);
1379 	if (err)
1380 		return (err);
1381 	err = dnode_set_blksz(dn, size, ibs, tx);
1382 	dnode_rele(dn, FTAG);
1383 	return (err);
1384 }
1385 
1386 void
1387 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1388 	dmu_tx_t *tx)
1389 {
1390 	dnode_t *dn;
1391 
1392 	/* XXX assumes dnode_hold will not get an i/o error */
1393 	(void) dnode_hold(os, object, FTAG, &dn);
1394 	ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1395 	dn->dn_checksum = checksum;
1396 	dnode_setdirty(dn, tx);
1397 	dnode_rele(dn, FTAG);
1398 }
1399 
1400 void
1401 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1402 	dmu_tx_t *tx)
1403 {
1404 	dnode_t *dn;
1405 
1406 	/* XXX assumes dnode_hold will not get an i/o error */
1407 	(void) dnode_hold(os, object, FTAG, &dn);
1408 	ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1409 	dn->dn_compress = compress;
1410 	dnode_setdirty(dn, tx);
1411 	dnode_rele(dn, FTAG);
1412 }
1413 
1414 int zfs_mdcomp_disable = 0;
1415 
1416 void
1417 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1418 {
1419 	dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1420 	boolean_t ismd = (level > 0 || dmu_ot[type].ot_metadata ||
1421 	    (wp & WP_SPILL));
1422 	enum zio_checksum checksum = os->os_checksum;
1423 	enum zio_compress compress = os->os_compress;
1424 	enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1425 	boolean_t dedup;
1426 	boolean_t dedup_verify = os->os_dedup_verify;
1427 	int copies = os->os_copies;
1428 
1429 	/*
1430 	 * Determine checksum setting.
1431 	 */
1432 	if (ismd) {
1433 		/*
1434 		 * Metadata always gets checksummed.  If the data
1435 		 * checksum is multi-bit correctable, and it's not a
1436 		 * ZBT-style checksum, then it's suitable for metadata
1437 		 * as well.  Otherwise, the metadata checksum defaults
1438 		 * to fletcher4.
1439 		 */
1440 		if (zio_checksum_table[checksum].ci_correctable < 1 ||
1441 		    zio_checksum_table[checksum].ci_eck)
1442 			checksum = ZIO_CHECKSUM_FLETCHER_4;
1443 	} else {
1444 		checksum = zio_checksum_select(dn->dn_checksum, checksum);
1445 	}
1446 
1447 	/*
1448 	 * Determine compression setting.
1449 	 */
1450 	if (ismd) {
1451 		/*
1452 		 * XXX -- we should design a compression algorithm
1453 		 * that specializes in arrays of bps.
1454 		 */
1455 		compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1456 		    ZIO_COMPRESS_LZJB;
1457 	} else {
1458 		compress = zio_compress_select(dn->dn_compress, compress);
1459 	}
1460 
1461 	/*
1462 	 * Determine dedup setting.  If we are in dmu_sync(), we won't
1463 	 * actually dedup now because that's all done in syncing context;
1464 	 * but we do want to use the dedup checkum.  If the checksum is not
1465 	 * strong enough to ensure unique signatures, force dedup_verify.
1466 	 */
1467 	dedup = (!ismd && dedup_checksum != ZIO_CHECKSUM_OFF);
1468 	if (dedup) {
1469 		checksum = dedup_checksum;
1470 		if (!zio_checksum_table[checksum].ci_dedup)
1471 			dedup_verify = 1;
1472 	}
1473 
1474 	if (wp & WP_DMU_SYNC)
1475 		dedup = 0;
1476 
1477 	if (wp & WP_NOFILL) {
1478 		ASSERT(!ismd && level == 0);
1479 		checksum = ZIO_CHECKSUM_OFF;
1480 		compress = ZIO_COMPRESS_OFF;
1481 		dedup = B_FALSE;
1482 	}
1483 
1484 	zp->zp_checksum = checksum;
1485 	zp->zp_compress = compress;
1486 	zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1487 	zp->zp_level = level;
1488 	zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1489 	zp->zp_dedup = dedup;
1490 	zp->zp_dedup_verify = dedup && dedup_verify;
1491 }
1492 
1493 int
1494 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1495 {
1496 	dnode_t *dn;
1497 	int i, err;
1498 
1499 	err = dnode_hold(os, object, FTAG, &dn);
1500 	if (err)
1501 		return (err);
1502 	/*
1503 	 * Sync any current changes before
1504 	 * we go trundling through the block pointers.
1505 	 */
1506 	for (i = 0; i < TXG_SIZE; i++) {
1507 		if (list_link_active(&dn->dn_dirty_link[i]))
1508 			break;
1509 	}
1510 	if (i != TXG_SIZE) {
1511 		dnode_rele(dn, FTAG);
1512 		txg_wait_synced(dmu_objset_pool(os), 0);
1513 		err = dnode_hold(os, object, FTAG, &dn);
1514 		if (err)
1515 			return (err);
1516 	}
1517 
1518 	err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1519 	dnode_rele(dn, FTAG);
1520 
1521 	return (err);
1522 }
1523 
1524 void
1525 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1526 {
1527 	dnode_phys_t *dnp;
1528 
1529 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1530 	mutex_enter(&dn->dn_mtx);
1531 
1532 	dnp = dn->dn_phys;
1533 
1534 	doi->doi_data_block_size = dn->dn_datablksz;
1535 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
1536 	    1ULL << dn->dn_indblkshift : 0;
1537 	doi->doi_type = dn->dn_type;
1538 	doi->doi_bonus_type = dn->dn_bonustype;
1539 	doi->doi_bonus_size = dn->dn_bonuslen;
1540 	doi->doi_indirection = dn->dn_nlevels;
1541 	doi->doi_checksum = dn->dn_checksum;
1542 	doi->doi_compress = dn->dn_compress;
1543 	doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1544 	doi->doi_max_offset = (dnp->dn_maxblkid + 1) * dn->dn_datablksz;
1545 	doi->doi_fill_count = 0;
1546 	for (int i = 0; i < dnp->dn_nblkptr; i++)
1547 		doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1548 
1549 	mutex_exit(&dn->dn_mtx);
1550 	rw_exit(&dn->dn_struct_rwlock);
1551 }
1552 
1553 /*
1554  * Get information on a DMU object.
1555  * If doi is NULL, just indicates whether the object exists.
1556  */
1557 int
1558 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1559 {
1560 	dnode_t *dn;
1561 	int err = dnode_hold(os, object, FTAG, &dn);
1562 
1563 	if (err)
1564 		return (err);
1565 
1566 	if (doi != NULL)
1567 		dmu_object_info_from_dnode(dn, doi);
1568 
1569 	dnode_rele(dn, FTAG);
1570 	return (0);
1571 }
1572 
1573 /*
1574  * As above, but faster; can be used when you have a held dbuf in hand.
1575  */
1576 void
1577 dmu_object_info_from_db(dmu_buf_t *db, dmu_object_info_t *doi)
1578 {
1579 	dmu_object_info_from_dnode(((dmu_buf_impl_t *)db)->db_dnode, doi);
1580 }
1581 
1582 /*
1583  * Faster still when you only care about the size.
1584  * This is specifically optimized for zfs_getattr().
1585  */
1586 void
1587 dmu_object_size_from_db(dmu_buf_t *db, uint32_t *blksize, u_longlong_t *nblk512)
1588 {
1589 	dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1590 
1591 	*blksize = dn->dn_datablksz;
1592 	/* add 1 for dnode space */
1593 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1594 	    SPA_MINBLOCKSHIFT) + 1;
1595 }
1596 
1597 void
1598 byteswap_uint64_array(void *vbuf, size_t size)
1599 {
1600 	uint64_t *buf = vbuf;
1601 	size_t count = size >> 3;
1602 	int i;
1603 
1604 	ASSERT((size & 7) == 0);
1605 
1606 	for (i = 0; i < count; i++)
1607 		buf[i] = BSWAP_64(buf[i]);
1608 }
1609 
1610 void
1611 byteswap_uint32_array(void *vbuf, size_t size)
1612 {
1613 	uint32_t *buf = vbuf;
1614 	size_t count = size >> 2;
1615 	int i;
1616 
1617 	ASSERT((size & 3) == 0);
1618 
1619 	for (i = 0; i < count; i++)
1620 		buf[i] = BSWAP_32(buf[i]);
1621 }
1622 
1623 void
1624 byteswap_uint16_array(void *vbuf, size_t size)
1625 {
1626 	uint16_t *buf = vbuf;
1627 	size_t count = size >> 1;
1628 	int i;
1629 
1630 	ASSERT((size & 1) == 0);
1631 
1632 	for (i = 0; i < count; i++)
1633 		buf[i] = BSWAP_16(buf[i]);
1634 }
1635 
1636 /* ARGSUSED */
1637 void
1638 byteswap_uint8_array(void *vbuf, size_t size)
1639 {
1640 }
1641 
1642 void
1643 dmu_init(void)
1644 {
1645 	zfs_dbgmsg_init();
1646 	dbuf_init();
1647 	dnode_init();
1648 	zfetch_init();
1649 	arc_init();
1650 	l2arc_init();
1651 	xuio_stat_init();
1652 	sa_cache_init();
1653 }
1654 
1655 void
1656 dmu_fini(void)
1657 {
1658 	arc_fini();
1659 	zfetch_fini();
1660 	dnode_fini();
1661 	dbuf_fini();
1662 	l2arc_fini();
1663 	xuio_stat_fini();
1664 	sa_cache_fini();
1665 	zfs_dbgmsg_fini();
1666 }
1667