xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_tx.c (revision 9fa718d2f477620f14e3f2948dd03e3470add804)
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  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  * Copyright (c) 2012 by Delphix. All rights reserved.
25  */
26 
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dbuf.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
33 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
34 #include <sys/dsl_pool.h>
35 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 #include <sys/spa.h>
37 #include <sys/sa.h>
38 #include <sys/sa_impl.h>
39 #include <sys/zfs_context.h>
40 #include <sys/varargs.h>
41 
42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43     uint64_t arg1, uint64_t arg2);
44 
45 
46 dmu_tx_t *
47 dmu_tx_create_dd(dsl_dir_t *dd)
48 {
49 	dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
50 	tx->tx_dir = dd;
51 	if (dd != NULL)
52 		tx->tx_pool = dd->dd_pool;
53 	list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
54 	    offsetof(dmu_tx_hold_t, txh_node));
55 	list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
56 	    offsetof(dmu_tx_callback_t, dcb_node));
57 #ifdef ZFS_DEBUG
58 	refcount_create(&tx->tx_space_written);
59 	refcount_create(&tx->tx_space_freed);
60 #endif
61 	return (tx);
62 }
63 
64 dmu_tx_t *
65 dmu_tx_create(objset_t *os)
66 {
67 	dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
68 	tx->tx_objset = os;
69 	tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
70 	return (tx);
71 }
72 
73 dmu_tx_t *
74 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
75 {
76 	dmu_tx_t *tx = dmu_tx_create_dd(NULL);
77 
78 	ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
79 	tx->tx_pool = dp;
80 	tx->tx_txg = txg;
81 	tx->tx_anyobj = TRUE;
82 
83 	return (tx);
84 }
85 
86 int
87 dmu_tx_is_syncing(dmu_tx_t *tx)
88 {
89 	return (tx->tx_anyobj);
90 }
91 
92 int
93 dmu_tx_private_ok(dmu_tx_t *tx)
94 {
95 	return (tx->tx_anyobj);
96 }
97 
98 static dmu_tx_hold_t *
99 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
100     enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
101 {
102 	dmu_tx_hold_t *txh;
103 	dnode_t *dn = NULL;
104 	int err;
105 
106 	if (object != DMU_NEW_OBJECT) {
107 		err = dnode_hold(os, object, tx, &dn);
108 		if (err) {
109 			tx->tx_err = err;
110 			return (NULL);
111 		}
112 
113 		if (err == 0 && tx->tx_txg != 0) {
114 			mutex_enter(&dn->dn_mtx);
115 			/*
116 			 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
117 			 * problem, but there's no way for it to happen (for
118 			 * now, at least).
119 			 */
120 			ASSERT(dn->dn_assigned_txg == 0);
121 			dn->dn_assigned_txg = tx->tx_txg;
122 			(void) refcount_add(&dn->dn_tx_holds, tx);
123 			mutex_exit(&dn->dn_mtx);
124 		}
125 	}
126 
127 	txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
128 	txh->txh_tx = tx;
129 	txh->txh_dnode = dn;
130 #ifdef ZFS_DEBUG
131 	txh->txh_type = type;
132 	txh->txh_arg1 = arg1;
133 	txh->txh_arg2 = arg2;
134 #endif
135 	list_insert_tail(&tx->tx_holds, txh);
136 
137 	return (txh);
138 }
139 
140 void
141 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
142 {
143 	/*
144 	 * If we're syncing, they can manipulate any object anyhow, and
145 	 * the hold on the dnode_t can cause problems.
146 	 */
147 	if (!dmu_tx_is_syncing(tx)) {
148 		(void) dmu_tx_hold_object_impl(tx, os,
149 		    object, THT_NEWOBJECT, 0, 0);
150 	}
151 }
152 
153 static int
154 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
155 {
156 	int err;
157 	dmu_buf_impl_t *db;
158 
159 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
160 	db = dbuf_hold_level(dn, level, blkid, FTAG);
161 	rw_exit(&dn->dn_struct_rwlock);
162 	if (db == NULL)
163 		return (EIO);
164 	err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
165 	dbuf_rele(db, FTAG);
166 	return (err);
167 }
168 
169 static void
170 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
171     int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
172 {
173 	objset_t *os = dn->dn_objset;
174 	dsl_dataset_t *ds = os->os_dsl_dataset;
175 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
176 	dmu_buf_impl_t *parent = NULL;
177 	blkptr_t *bp = NULL;
178 	uint64_t space;
179 
180 	if (level >= dn->dn_nlevels || history[level] == blkid)
181 		return;
182 
183 	history[level] = blkid;
184 
185 	space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
186 
187 	if (db == NULL || db == dn->dn_dbuf) {
188 		ASSERT(level != 0);
189 		db = NULL;
190 	} else {
191 		ASSERT(DB_DNODE(db) == dn);
192 		ASSERT(db->db_level == level);
193 		ASSERT(db->db.db_size == space);
194 		ASSERT(db->db_blkid == blkid);
195 		bp = db->db_blkptr;
196 		parent = db->db_parent;
197 	}
198 
199 	freeable = (bp && (freeable ||
200 	    dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
201 
202 	if (freeable)
203 		txh->txh_space_tooverwrite += space;
204 	else
205 		txh->txh_space_towrite += space;
206 	if (bp)
207 		txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
208 
209 	dmu_tx_count_twig(txh, dn, parent, level + 1,
210 	    blkid >> epbs, freeable, history);
211 }
212 
213 /* ARGSUSED */
214 static void
215 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
216 {
217 	dnode_t *dn = txh->txh_dnode;
218 	uint64_t start, end, i;
219 	int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
220 	int err = 0;
221 
222 	if (len == 0)
223 		return;
224 
225 	min_bs = SPA_MINBLOCKSHIFT;
226 	max_bs = SPA_MAXBLOCKSHIFT;
227 	min_ibs = DN_MIN_INDBLKSHIFT;
228 	max_ibs = DN_MAX_INDBLKSHIFT;
229 
230 	if (dn) {
231 		uint64_t history[DN_MAX_LEVELS];
232 		int nlvls = dn->dn_nlevels;
233 		int delta;
234 
235 		/*
236 		 * For i/o error checking, read the first and last level-0
237 		 * blocks (if they are not aligned), and all the level-1 blocks.
238 		 */
239 		if (dn->dn_maxblkid == 0) {
240 			delta = dn->dn_datablksz;
241 			start = (off < dn->dn_datablksz) ? 0 : 1;
242 			end = (off+len <= dn->dn_datablksz) ? 0 : 1;
243 			if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
244 				err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
245 				if (err)
246 					goto out;
247 				delta -= off;
248 			}
249 		} else {
250 			zio_t *zio = zio_root(dn->dn_objset->os_spa,
251 			    NULL, NULL, ZIO_FLAG_CANFAIL);
252 
253 			/* first level-0 block */
254 			start = off >> dn->dn_datablkshift;
255 			if (P2PHASE(off, dn->dn_datablksz) ||
256 			    len < dn->dn_datablksz) {
257 				err = dmu_tx_check_ioerr(zio, dn, 0, start);
258 				if (err)
259 					goto out;
260 			}
261 
262 			/* last level-0 block */
263 			end = (off+len-1) >> dn->dn_datablkshift;
264 			if (end != start && end <= dn->dn_maxblkid &&
265 			    P2PHASE(off+len, dn->dn_datablksz)) {
266 				err = dmu_tx_check_ioerr(zio, dn, 0, end);
267 				if (err)
268 					goto out;
269 			}
270 
271 			/* level-1 blocks */
272 			if (nlvls > 1) {
273 				int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
274 				for (i = (start>>shft)+1; i < end>>shft; i++) {
275 					err = dmu_tx_check_ioerr(zio, dn, 1, i);
276 					if (err)
277 						goto out;
278 				}
279 			}
280 
281 			err = zio_wait(zio);
282 			if (err)
283 				goto out;
284 			delta = P2NPHASE(off, dn->dn_datablksz);
285 		}
286 
287 		if (dn->dn_maxblkid > 0) {
288 			/*
289 			 * The blocksize can't change,
290 			 * so we can make a more precise estimate.
291 			 */
292 			ASSERT(dn->dn_datablkshift != 0);
293 			min_bs = max_bs = dn->dn_datablkshift;
294 			min_ibs = max_ibs = dn->dn_indblkshift;
295 		} else if (dn->dn_indblkshift > max_ibs) {
296 			/*
297 			 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
298 			 * the code will still work correctly on older pools.
299 			 */
300 			min_ibs = max_ibs = dn->dn_indblkshift;
301 		}
302 
303 		/*
304 		 * If this write is not off the end of the file
305 		 * we need to account for overwrites/unref.
306 		 */
307 		if (start <= dn->dn_maxblkid) {
308 			for (int l = 0; l < DN_MAX_LEVELS; l++)
309 				history[l] = -1ULL;
310 		}
311 		while (start <= dn->dn_maxblkid) {
312 			dmu_buf_impl_t *db;
313 
314 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
315 			err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
316 			rw_exit(&dn->dn_struct_rwlock);
317 
318 			if (err) {
319 				txh->txh_tx->tx_err = err;
320 				return;
321 			}
322 
323 			dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
324 			    history);
325 			dbuf_rele(db, FTAG);
326 			if (++start > end) {
327 				/*
328 				 * Account for new indirects appearing
329 				 * before this IO gets assigned into a txg.
330 				 */
331 				bits = 64 - min_bs;
332 				epbs = min_ibs - SPA_BLKPTRSHIFT;
333 				for (bits -= epbs * (nlvls - 1);
334 				    bits >= 0; bits -= epbs)
335 					txh->txh_fudge += 1ULL << max_ibs;
336 				goto out;
337 			}
338 			off += delta;
339 			if (len >= delta)
340 				len -= delta;
341 			delta = dn->dn_datablksz;
342 		}
343 	}
344 
345 	/*
346 	 * 'end' is the last thing we will access, not one past.
347 	 * This way we won't overflow when accessing the last byte.
348 	 */
349 	start = P2ALIGN(off, 1ULL << max_bs);
350 	end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
351 	txh->txh_space_towrite += end - start + 1;
352 
353 	start >>= min_bs;
354 	end >>= min_bs;
355 
356 	epbs = min_ibs - SPA_BLKPTRSHIFT;
357 
358 	/*
359 	 * The object contains at most 2^(64 - min_bs) blocks,
360 	 * and each indirect level maps 2^epbs.
361 	 */
362 	for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
363 		start >>= epbs;
364 		end >>= epbs;
365 		ASSERT3U(end, >=, start);
366 		txh->txh_space_towrite += (end - start + 1) << max_ibs;
367 		if (start != 0) {
368 			/*
369 			 * We also need a new blkid=0 indirect block
370 			 * to reference any existing file data.
371 			 */
372 			txh->txh_space_towrite += 1ULL << max_ibs;
373 		}
374 	}
375 
376 out:
377 	if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
378 	    2 * DMU_MAX_ACCESS)
379 		err = EFBIG;
380 
381 	if (err)
382 		txh->txh_tx->tx_err = err;
383 }
384 
385 static void
386 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
387 {
388 	dnode_t *dn = txh->txh_dnode;
389 	dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
390 	uint64_t space = mdn->dn_datablksz +
391 	    ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
392 
393 	if (dn && dn->dn_dbuf->db_blkptr &&
394 	    dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
395 	    dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
396 		txh->txh_space_tooverwrite += space;
397 		txh->txh_space_tounref += space;
398 	} else {
399 		txh->txh_space_towrite += space;
400 		if (dn && dn->dn_dbuf->db_blkptr)
401 			txh->txh_space_tounref += space;
402 	}
403 }
404 
405 void
406 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
407 {
408 	dmu_tx_hold_t *txh;
409 
410 	ASSERT(tx->tx_txg == 0);
411 	ASSERT(len < DMU_MAX_ACCESS);
412 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
413 
414 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
415 	    object, THT_WRITE, off, len);
416 	if (txh == NULL)
417 		return;
418 
419 	dmu_tx_count_write(txh, off, len);
420 	dmu_tx_count_dnode(txh);
421 }
422 
423 static void
424 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
425 {
426 	uint64_t blkid, nblks, lastblk;
427 	uint64_t space = 0, unref = 0, skipped = 0;
428 	dnode_t *dn = txh->txh_dnode;
429 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
430 	spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
431 	int epbs;
432 
433 	if (dn->dn_nlevels == 0)
434 		return;
435 
436 	/*
437 	 * The struct_rwlock protects us against dn_nlevels
438 	 * changing, in case (against all odds) we manage to dirty &
439 	 * sync out the changes after we check for being dirty.
440 	 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
441 	 */
442 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
443 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
444 	if (dn->dn_maxblkid == 0) {
445 		if (off == 0 && len >= dn->dn_datablksz) {
446 			blkid = 0;
447 			nblks = 1;
448 		} else {
449 			rw_exit(&dn->dn_struct_rwlock);
450 			return;
451 		}
452 	} else {
453 		blkid = off >> dn->dn_datablkshift;
454 		nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
455 
456 		if (blkid >= dn->dn_maxblkid) {
457 			rw_exit(&dn->dn_struct_rwlock);
458 			return;
459 		}
460 		if (blkid + nblks > dn->dn_maxblkid)
461 			nblks = dn->dn_maxblkid - blkid;
462 
463 	}
464 	if (dn->dn_nlevels == 1) {
465 		int i;
466 		for (i = 0; i < nblks; i++) {
467 			blkptr_t *bp = dn->dn_phys->dn_blkptr;
468 			ASSERT3U(blkid + i, <, dn->dn_nblkptr);
469 			bp += blkid + i;
470 			if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
471 				dprintf_bp(bp, "can free old%s", "");
472 				space += bp_get_dsize(spa, bp);
473 			}
474 			unref += BP_GET_ASIZE(bp);
475 		}
476 		nblks = 0;
477 	}
478 
479 	/*
480 	 * Add in memory requirements of higher-level indirects.
481 	 * This assumes a worst-possible scenario for dn_nlevels.
482 	 */
483 	{
484 		uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
485 		int level = (dn->dn_nlevels > 1) ? 2 : 1;
486 
487 		while (level++ < DN_MAX_LEVELS) {
488 			txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
489 			blkcnt = 1 + (blkcnt >> epbs);
490 		}
491 		ASSERT(blkcnt <= dn->dn_nblkptr);
492 	}
493 
494 	lastblk = blkid + nblks - 1;
495 	while (nblks) {
496 		dmu_buf_impl_t *dbuf;
497 		uint64_t ibyte, new_blkid;
498 		int epb = 1 << epbs;
499 		int err, i, blkoff, tochk;
500 		blkptr_t *bp;
501 
502 		ibyte = blkid << dn->dn_datablkshift;
503 		err = dnode_next_offset(dn,
504 		    DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
505 		new_blkid = ibyte >> dn->dn_datablkshift;
506 		if (err == ESRCH) {
507 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
508 			break;
509 		}
510 		if (err) {
511 			txh->txh_tx->tx_err = err;
512 			break;
513 		}
514 		if (new_blkid > lastblk) {
515 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
516 			break;
517 		}
518 
519 		if (new_blkid > blkid) {
520 			ASSERT((new_blkid >> epbs) > (blkid >> epbs));
521 			skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
522 			nblks -= new_blkid - blkid;
523 			blkid = new_blkid;
524 		}
525 		blkoff = P2PHASE(blkid, epb);
526 		tochk = MIN(epb - blkoff, nblks);
527 
528 		err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
529 		if (err) {
530 			txh->txh_tx->tx_err = err;
531 			break;
532 		}
533 
534 		txh->txh_memory_tohold += dbuf->db.db_size;
535 
536 		/*
537 		 * We don't check memory_tohold against DMU_MAX_ACCESS because
538 		 * memory_tohold is an over-estimation (especially the >L1
539 		 * indirect blocks), so it could fail.  Callers should have
540 		 * already verified that they will not be holding too much
541 		 * memory.
542 		 */
543 
544 		err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
545 		if (err != 0) {
546 			txh->txh_tx->tx_err = err;
547 			dbuf_rele(dbuf, FTAG);
548 			break;
549 		}
550 
551 		bp = dbuf->db.db_data;
552 		bp += blkoff;
553 
554 		for (i = 0; i < tochk; i++) {
555 			if (dsl_dataset_block_freeable(ds, &bp[i],
556 			    bp[i].blk_birth)) {
557 				dprintf_bp(&bp[i], "can free old%s", "");
558 				space += bp_get_dsize(spa, &bp[i]);
559 			}
560 			unref += BP_GET_ASIZE(bp);
561 		}
562 		dbuf_rele(dbuf, FTAG);
563 
564 		blkid += tochk;
565 		nblks -= tochk;
566 	}
567 	rw_exit(&dn->dn_struct_rwlock);
568 
569 	/* account for new level 1 indirect blocks that might show up */
570 	if (skipped > 0) {
571 		txh->txh_fudge += skipped << dn->dn_indblkshift;
572 		skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
573 		txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
574 	}
575 	txh->txh_space_tofree += space;
576 	txh->txh_space_tounref += unref;
577 }
578 
579 void
580 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
581 {
582 	dmu_tx_hold_t *txh;
583 	dnode_t *dn;
584 	uint64_t start, end, i;
585 	int err, shift;
586 	zio_t *zio;
587 
588 	ASSERT(tx->tx_txg == 0);
589 
590 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
591 	    object, THT_FREE, off, len);
592 	if (txh == NULL)
593 		return;
594 	dn = txh->txh_dnode;
595 
596 	/* first block */
597 	if (off != 0)
598 		dmu_tx_count_write(txh, off, 1);
599 	/* last block */
600 	if (len != DMU_OBJECT_END)
601 		dmu_tx_count_write(txh, off+len, 1);
602 
603 	dmu_tx_count_dnode(txh);
604 
605 	if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
606 		return;
607 	if (len == DMU_OBJECT_END)
608 		len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
609 
610 	/*
611 	 * For i/o error checking, read the first and last level-0
612 	 * blocks, and all the level-1 blocks.  The above count_write's
613 	 * have already taken care of the level-0 blocks.
614 	 */
615 	if (dn->dn_nlevels > 1) {
616 		shift = dn->dn_datablkshift + dn->dn_indblkshift -
617 		    SPA_BLKPTRSHIFT;
618 		start = off >> shift;
619 		end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
620 
621 		zio = zio_root(tx->tx_pool->dp_spa,
622 		    NULL, NULL, ZIO_FLAG_CANFAIL);
623 		for (i = start; i <= end; i++) {
624 			uint64_t ibyte = i << shift;
625 			err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
626 			i = ibyte >> shift;
627 			if (err == ESRCH)
628 				break;
629 			if (err) {
630 				tx->tx_err = err;
631 				return;
632 			}
633 
634 			err = dmu_tx_check_ioerr(zio, dn, 1, i);
635 			if (err) {
636 				tx->tx_err = err;
637 				return;
638 			}
639 		}
640 		err = zio_wait(zio);
641 		if (err) {
642 			tx->tx_err = err;
643 			return;
644 		}
645 	}
646 
647 	dmu_tx_count_free(txh, off, len);
648 }
649 
650 void
651 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
652 {
653 	dmu_tx_hold_t *txh;
654 	dnode_t *dn;
655 	uint64_t nblocks;
656 	int epbs, err;
657 
658 	ASSERT(tx->tx_txg == 0);
659 
660 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
661 	    object, THT_ZAP, add, (uintptr_t)name);
662 	if (txh == NULL)
663 		return;
664 	dn = txh->txh_dnode;
665 
666 	dmu_tx_count_dnode(txh);
667 
668 	if (dn == NULL) {
669 		/*
670 		 * We will be able to fit a new object's entries into one leaf
671 		 * block.  So there will be at most 2 blocks total,
672 		 * including the header block.
673 		 */
674 		dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
675 		return;
676 	}
677 
678 	ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
679 
680 	if (dn->dn_maxblkid == 0 && !add) {
681 		blkptr_t *bp;
682 
683 		/*
684 		 * If there is only one block  (i.e. this is a micro-zap)
685 		 * and we are not adding anything, the accounting is simple.
686 		 */
687 		err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
688 		if (err) {
689 			tx->tx_err = err;
690 			return;
691 		}
692 
693 		/*
694 		 * Use max block size here, since we don't know how much
695 		 * the size will change between now and the dbuf dirty call.
696 		 */
697 		bp = &dn->dn_phys->dn_blkptr[0];
698 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
699 		    bp, bp->blk_birth))
700 			txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
701 		else
702 			txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
703 		if (!BP_IS_HOLE(bp))
704 			txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
705 		return;
706 	}
707 
708 	if (dn->dn_maxblkid > 0 && name) {
709 		/*
710 		 * access the name in this fat-zap so that we'll check
711 		 * for i/o errors to the leaf blocks, etc.
712 		 */
713 		err = zap_lookup(dn->dn_objset, dn->dn_object, name,
714 		    8, 0, NULL);
715 		if (err == EIO) {
716 			tx->tx_err = err;
717 			return;
718 		}
719 	}
720 
721 	err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
722 	    &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
723 
724 	/*
725 	 * If the modified blocks are scattered to the four winds,
726 	 * we'll have to modify an indirect twig for each.
727 	 */
728 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
729 	for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
730 		if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
731 			txh->txh_space_towrite += 3 << dn->dn_indblkshift;
732 		else
733 			txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
734 }
735 
736 void
737 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
738 {
739 	dmu_tx_hold_t *txh;
740 
741 	ASSERT(tx->tx_txg == 0);
742 
743 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
744 	    object, THT_BONUS, 0, 0);
745 	if (txh)
746 		dmu_tx_count_dnode(txh);
747 }
748 
749 void
750 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
751 {
752 	dmu_tx_hold_t *txh;
753 	ASSERT(tx->tx_txg == 0);
754 
755 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
756 	    DMU_NEW_OBJECT, THT_SPACE, space, 0);
757 
758 	txh->txh_space_towrite += space;
759 }
760 
761 int
762 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
763 {
764 	dmu_tx_hold_t *txh;
765 	int holds = 0;
766 
767 	/*
768 	 * By asserting that the tx is assigned, we're counting the
769 	 * number of dn_tx_holds, which is the same as the number of
770 	 * dn_holds.  Otherwise, we'd be counting dn_holds, but
771 	 * dn_tx_holds could be 0.
772 	 */
773 	ASSERT(tx->tx_txg != 0);
774 
775 	/* if (tx->tx_anyobj == TRUE) */
776 		/* return (0); */
777 
778 	for (txh = list_head(&tx->tx_holds); txh;
779 	    txh = list_next(&tx->tx_holds, txh)) {
780 		if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
781 			holds++;
782 	}
783 
784 	return (holds);
785 }
786 
787 #ifdef ZFS_DEBUG
788 void
789 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
790 {
791 	dmu_tx_hold_t *txh;
792 	int match_object = FALSE, match_offset = FALSE;
793 	dnode_t *dn;
794 
795 	DB_DNODE_ENTER(db);
796 	dn = DB_DNODE(db);
797 	ASSERT(tx->tx_txg != 0);
798 	ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
799 	ASSERT3U(dn->dn_object, ==, db->db.db_object);
800 
801 	if (tx->tx_anyobj) {
802 		DB_DNODE_EXIT(db);
803 		return;
804 	}
805 
806 	/* XXX No checking on the meta dnode for now */
807 	if (db->db.db_object == DMU_META_DNODE_OBJECT) {
808 		DB_DNODE_EXIT(db);
809 		return;
810 	}
811 
812 	for (txh = list_head(&tx->tx_holds); txh;
813 	    txh = list_next(&tx->tx_holds, txh)) {
814 		ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
815 		if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
816 			match_object = TRUE;
817 		if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
818 			int datablkshift = dn->dn_datablkshift ?
819 			    dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
820 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
821 			int shift = datablkshift + epbs * db->db_level;
822 			uint64_t beginblk = shift >= 64 ? 0 :
823 			    (txh->txh_arg1 >> shift);
824 			uint64_t endblk = shift >= 64 ? 0 :
825 			    ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
826 			uint64_t blkid = db->db_blkid;
827 
828 			/* XXX txh_arg2 better not be zero... */
829 
830 			dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
831 			    txh->txh_type, beginblk, endblk);
832 
833 			switch (txh->txh_type) {
834 			case THT_WRITE:
835 				if (blkid >= beginblk && blkid <= endblk)
836 					match_offset = TRUE;
837 				/*
838 				 * We will let this hold work for the bonus
839 				 * or spill buffer so that we don't need to
840 				 * hold it when creating a new object.
841 				 */
842 				if (blkid == DMU_BONUS_BLKID ||
843 				    blkid == DMU_SPILL_BLKID)
844 					match_offset = TRUE;
845 				/*
846 				 * They might have to increase nlevels,
847 				 * thus dirtying the new TLIBs.  Or the
848 				 * might have to change the block size,
849 				 * thus dirying the new lvl=0 blk=0.
850 				 */
851 				if (blkid == 0)
852 					match_offset = TRUE;
853 				break;
854 			case THT_FREE:
855 				/*
856 				 * We will dirty all the level 1 blocks in
857 				 * the free range and perhaps the first and
858 				 * last level 0 block.
859 				 */
860 				if (blkid >= beginblk && (blkid <= endblk ||
861 				    txh->txh_arg2 == DMU_OBJECT_END))
862 					match_offset = TRUE;
863 				break;
864 			case THT_SPILL:
865 				if (blkid == DMU_SPILL_BLKID)
866 					match_offset = TRUE;
867 				break;
868 			case THT_BONUS:
869 				if (blkid == DMU_BONUS_BLKID)
870 					match_offset = TRUE;
871 				break;
872 			case THT_ZAP:
873 				match_offset = TRUE;
874 				break;
875 			case THT_NEWOBJECT:
876 				match_object = TRUE;
877 				break;
878 			default:
879 				ASSERT(!"bad txh_type");
880 			}
881 		}
882 		if (match_object && match_offset) {
883 			DB_DNODE_EXIT(db);
884 			return;
885 		}
886 	}
887 	DB_DNODE_EXIT(db);
888 	panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
889 	    (u_longlong_t)db->db.db_object, db->db_level,
890 	    (u_longlong_t)db->db_blkid);
891 }
892 #endif
893 
894 static int
895 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
896 {
897 	dmu_tx_hold_t *txh;
898 	spa_t *spa = tx->tx_pool->dp_spa;
899 	uint64_t memory, asize, fsize, usize;
900 	uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
901 
902 	ASSERT0(tx->tx_txg);
903 
904 	if (tx->tx_err)
905 		return (tx->tx_err);
906 
907 	if (spa_suspended(spa)) {
908 		/*
909 		 * If the user has indicated a blocking failure mode
910 		 * then return ERESTART which will block in dmu_tx_wait().
911 		 * Otherwise, return EIO so that an error can get
912 		 * propagated back to the VOP calls.
913 		 *
914 		 * Note that we always honor the txg_how flag regardless
915 		 * of the failuremode setting.
916 		 */
917 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
918 		    txg_how != TXG_WAIT)
919 			return (EIO);
920 
921 		return (ERESTART);
922 	}
923 
924 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
925 	tx->tx_needassign_txh = NULL;
926 
927 	/*
928 	 * NB: No error returns are allowed after txg_hold_open, but
929 	 * before processing the dnode holds, due to the
930 	 * dmu_tx_unassign() logic.
931 	 */
932 
933 	towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
934 	for (txh = list_head(&tx->tx_holds); txh;
935 	    txh = list_next(&tx->tx_holds, txh)) {
936 		dnode_t *dn = txh->txh_dnode;
937 		if (dn != NULL) {
938 			mutex_enter(&dn->dn_mtx);
939 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
940 				mutex_exit(&dn->dn_mtx);
941 				tx->tx_needassign_txh = txh;
942 				return (ERESTART);
943 			}
944 			if (dn->dn_assigned_txg == 0)
945 				dn->dn_assigned_txg = tx->tx_txg;
946 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
947 			(void) refcount_add(&dn->dn_tx_holds, tx);
948 			mutex_exit(&dn->dn_mtx);
949 		}
950 		towrite += txh->txh_space_towrite;
951 		tofree += txh->txh_space_tofree;
952 		tooverwrite += txh->txh_space_tooverwrite;
953 		tounref += txh->txh_space_tounref;
954 		tohold += txh->txh_memory_tohold;
955 		fudge += txh->txh_fudge;
956 	}
957 
958 	/*
959 	 * NB: This check must be after we've held the dnodes, so that
960 	 * the dmu_tx_unassign() logic will work properly
961 	 */
962 	if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
963 		return (ERESTART);
964 
965 	/*
966 	 * If a snapshot has been taken since we made our estimates,
967 	 * assume that we won't be able to free or overwrite anything.
968 	 */
969 	if (tx->tx_objset &&
970 	    dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
971 	    tx->tx_lastsnap_txg) {
972 		towrite += tooverwrite;
973 		tooverwrite = tofree = 0;
974 	}
975 
976 	/* needed allocation: worst-case estimate of write space */
977 	asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
978 	/* freed space estimate: worst-case overwrite + free estimate */
979 	fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
980 	/* convert unrefd space to worst-case estimate */
981 	usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
982 	/* calculate memory footprint estimate */
983 	memory = towrite + tooverwrite + tohold;
984 
985 #ifdef ZFS_DEBUG
986 	/*
987 	 * Add in 'tohold' to account for our dirty holds on this memory
988 	 * XXX - the "fudge" factor is to account for skipped blocks that
989 	 * we missed because dnode_next_offset() misses in-core-only blocks.
990 	 */
991 	tx->tx_space_towrite = asize +
992 	    spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
993 	tx->tx_space_tofree = tofree;
994 	tx->tx_space_tooverwrite = tooverwrite;
995 	tx->tx_space_tounref = tounref;
996 #endif
997 
998 	if (tx->tx_dir && asize != 0) {
999 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1000 		    asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1001 		if (err)
1002 			return (err);
1003 	}
1004 
1005 	return (0);
1006 }
1007 
1008 static void
1009 dmu_tx_unassign(dmu_tx_t *tx)
1010 {
1011 	dmu_tx_hold_t *txh;
1012 
1013 	if (tx->tx_txg == 0)
1014 		return;
1015 
1016 	txg_rele_to_quiesce(&tx->tx_txgh);
1017 
1018 	for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1019 	    txh = list_next(&tx->tx_holds, txh)) {
1020 		dnode_t *dn = txh->txh_dnode;
1021 
1022 		if (dn == NULL)
1023 			continue;
1024 		mutex_enter(&dn->dn_mtx);
1025 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1026 
1027 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1028 			dn->dn_assigned_txg = 0;
1029 			cv_broadcast(&dn->dn_notxholds);
1030 		}
1031 		mutex_exit(&dn->dn_mtx);
1032 	}
1033 
1034 	txg_rele_to_sync(&tx->tx_txgh);
1035 
1036 	tx->tx_lasttried_txg = tx->tx_txg;
1037 	tx->tx_txg = 0;
1038 }
1039 
1040 /*
1041  * Assign tx to a transaction group.  txg_how can be one of:
1042  *
1043  * (1)	TXG_WAIT.  If the current open txg is full, waits until there's
1044  *	a new one.  This should be used when you're not holding locks.
1045  *	If will only fail if we're truly out of space (or over quota).
1046  *
1047  * (2)	TXG_NOWAIT.  If we can't assign into the current open txg without
1048  *	blocking, returns immediately with ERESTART.  This should be used
1049  *	whenever you're holding locks.  On an ERESTART error, the caller
1050  *	should drop locks, do a dmu_tx_wait(tx), and try again.
1051  *
1052  * (3)	A specific txg.  Use this if you need to ensure that multiple
1053  *	transactions all sync in the same txg.  Like TXG_NOWAIT, it
1054  *	returns ERESTART if it can't assign you into the requested txg.
1055  */
1056 int
1057 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1058 {
1059 	int err;
1060 
1061 	ASSERT(tx->tx_txg == 0);
1062 	ASSERT(txg_how != 0);
1063 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1064 
1065 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1066 		dmu_tx_unassign(tx);
1067 
1068 		if (err != ERESTART || txg_how != TXG_WAIT)
1069 			return (err);
1070 
1071 		dmu_tx_wait(tx);
1072 	}
1073 
1074 	txg_rele_to_quiesce(&tx->tx_txgh);
1075 
1076 	return (0);
1077 }
1078 
1079 void
1080 dmu_tx_wait(dmu_tx_t *tx)
1081 {
1082 	spa_t *spa = tx->tx_pool->dp_spa;
1083 
1084 	ASSERT(tx->tx_txg == 0);
1085 
1086 	/*
1087 	 * It's possible that the pool has become active after this thread
1088 	 * has tried to obtain a tx. If that's the case then his
1089 	 * tx_lasttried_txg would not have been assigned.
1090 	 */
1091 	if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1092 		txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1093 	} else if (tx->tx_needassign_txh) {
1094 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1095 
1096 		mutex_enter(&dn->dn_mtx);
1097 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1098 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1099 		mutex_exit(&dn->dn_mtx);
1100 		tx->tx_needassign_txh = NULL;
1101 	} else {
1102 		txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1103 	}
1104 }
1105 
1106 void
1107 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1108 {
1109 #ifdef ZFS_DEBUG
1110 	if (tx->tx_dir == NULL || delta == 0)
1111 		return;
1112 
1113 	if (delta > 0) {
1114 		ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1115 		    tx->tx_space_towrite);
1116 		(void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1117 	} else {
1118 		(void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1119 	}
1120 #endif
1121 }
1122 
1123 void
1124 dmu_tx_commit(dmu_tx_t *tx)
1125 {
1126 	dmu_tx_hold_t *txh;
1127 
1128 	ASSERT(tx->tx_txg != 0);
1129 
1130 	while (txh = list_head(&tx->tx_holds)) {
1131 		dnode_t *dn = txh->txh_dnode;
1132 
1133 		list_remove(&tx->tx_holds, txh);
1134 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1135 		if (dn == NULL)
1136 			continue;
1137 		mutex_enter(&dn->dn_mtx);
1138 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1139 
1140 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1141 			dn->dn_assigned_txg = 0;
1142 			cv_broadcast(&dn->dn_notxholds);
1143 		}
1144 		mutex_exit(&dn->dn_mtx);
1145 		dnode_rele(dn, tx);
1146 	}
1147 
1148 	if (tx->tx_tempreserve_cookie)
1149 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1150 
1151 	if (!list_is_empty(&tx->tx_callbacks))
1152 		txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1153 
1154 	if (tx->tx_anyobj == FALSE)
1155 		txg_rele_to_sync(&tx->tx_txgh);
1156 
1157 	list_destroy(&tx->tx_callbacks);
1158 	list_destroy(&tx->tx_holds);
1159 #ifdef ZFS_DEBUG
1160 	dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1161 	    tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1162 	    tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1163 	refcount_destroy_many(&tx->tx_space_written,
1164 	    refcount_count(&tx->tx_space_written));
1165 	refcount_destroy_many(&tx->tx_space_freed,
1166 	    refcount_count(&tx->tx_space_freed));
1167 #endif
1168 	kmem_free(tx, sizeof (dmu_tx_t));
1169 }
1170 
1171 void
1172 dmu_tx_abort(dmu_tx_t *tx)
1173 {
1174 	dmu_tx_hold_t *txh;
1175 
1176 	ASSERT(tx->tx_txg == 0);
1177 
1178 	while (txh = list_head(&tx->tx_holds)) {
1179 		dnode_t *dn = txh->txh_dnode;
1180 
1181 		list_remove(&tx->tx_holds, txh);
1182 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1183 		if (dn != NULL)
1184 			dnode_rele(dn, tx);
1185 	}
1186 
1187 	/*
1188 	 * Call any registered callbacks with an error code.
1189 	 */
1190 	if (!list_is_empty(&tx->tx_callbacks))
1191 		dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1192 
1193 	list_destroy(&tx->tx_callbacks);
1194 	list_destroy(&tx->tx_holds);
1195 #ifdef ZFS_DEBUG
1196 	refcount_destroy_many(&tx->tx_space_written,
1197 	    refcount_count(&tx->tx_space_written));
1198 	refcount_destroy_many(&tx->tx_space_freed,
1199 	    refcount_count(&tx->tx_space_freed));
1200 #endif
1201 	kmem_free(tx, sizeof (dmu_tx_t));
1202 }
1203 
1204 uint64_t
1205 dmu_tx_get_txg(dmu_tx_t *tx)
1206 {
1207 	ASSERT(tx->tx_txg != 0);
1208 	return (tx->tx_txg);
1209 }
1210 
1211 void
1212 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1213 {
1214 	dmu_tx_callback_t *dcb;
1215 
1216 	dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1217 
1218 	dcb->dcb_func = func;
1219 	dcb->dcb_data = data;
1220 
1221 	list_insert_tail(&tx->tx_callbacks, dcb);
1222 }
1223 
1224 /*
1225  * Call all the commit callbacks on a list, with a given error code.
1226  */
1227 void
1228 dmu_tx_do_callbacks(list_t *cb_list, int error)
1229 {
1230 	dmu_tx_callback_t *dcb;
1231 
1232 	while (dcb = list_head(cb_list)) {
1233 		list_remove(cb_list, dcb);
1234 		dcb->dcb_func(dcb->dcb_data, error);
1235 		kmem_free(dcb, sizeof (dmu_tx_callback_t));
1236 	}
1237 }
1238 
1239 /*
1240  * Interface to hold a bunch of attributes.
1241  * used for creating new files.
1242  * attrsize is the total size of all attributes
1243  * to be added during object creation
1244  *
1245  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1246  */
1247 
1248 /*
1249  * hold necessary attribute name for attribute registration.
1250  * should be a very rare case where this is needed.  If it does
1251  * happen it would only happen on the first write to the file system.
1252  */
1253 static void
1254 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1255 {
1256 	int i;
1257 
1258 	if (!sa->sa_need_attr_registration)
1259 		return;
1260 
1261 	for (i = 0; i != sa->sa_num_attrs; i++) {
1262 		if (!sa->sa_attr_table[i].sa_registered) {
1263 			if (sa->sa_reg_attr_obj)
1264 				dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1265 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1266 			else
1267 				dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1268 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1269 		}
1270 	}
1271 }
1272 
1273 
1274 void
1275 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1276 {
1277 	dnode_t *dn;
1278 	dmu_tx_hold_t *txh;
1279 
1280 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1281 	    THT_SPILL, 0, 0);
1282 
1283 	dn = txh->txh_dnode;
1284 
1285 	if (dn == NULL)
1286 		return;
1287 
1288 	/* If blkptr doesn't exist then add space to towrite */
1289 	if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1290 		txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1291 	} else {
1292 		blkptr_t *bp;
1293 
1294 		bp = &dn->dn_phys->dn_spill;
1295 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1296 		    bp, bp->blk_birth))
1297 			txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1298 		else
1299 			txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1300 		if (!BP_IS_HOLE(bp))
1301 			txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1302 	}
1303 }
1304 
1305 void
1306 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1307 {
1308 	sa_os_t *sa = tx->tx_objset->os_sa;
1309 
1310 	dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1311 
1312 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1313 		return;
1314 
1315 	if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1316 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1317 	else {
1318 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1319 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1320 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1321 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1322 	}
1323 
1324 	dmu_tx_sa_registration_hold(sa, tx);
1325 
1326 	if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1327 		return;
1328 
1329 	(void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1330 	    THT_SPILL, 0, 0);
1331 }
1332 
1333 /*
1334  * Hold SA attribute
1335  *
1336  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1337  *
1338  * variable_size is the total size of all variable sized attributes
1339  * passed to this function.  It is not the total size of all
1340  * variable size attributes that *may* exist on this object.
1341  */
1342 void
1343 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1344 {
1345 	uint64_t object;
1346 	sa_os_t *sa = tx->tx_objset->os_sa;
1347 
1348 	ASSERT(hdl != NULL);
1349 
1350 	object = sa_handle_object(hdl);
1351 
1352 	dmu_tx_hold_bonus(tx, object);
1353 
1354 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1355 		return;
1356 
1357 	if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1358 	    tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1359 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1360 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1361 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1362 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1363 	}
1364 
1365 	dmu_tx_sa_registration_hold(sa, tx);
1366 
1367 	if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1368 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1369 
1370 	if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1371 		ASSERT(tx->tx_txg == 0);
1372 		dmu_tx_hold_spill(tx, object);
1373 	} else {
1374 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1375 		dnode_t *dn;
1376 
1377 		DB_DNODE_ENTER(db);
1378 		dn = DB_DNODE(db);
1379 		if (dn->dn_have_spill) {
1380 			ASSERT(tx->tx_txg == 0);
1381 			dmu_tx_hold_spill(tx, object);
1382 		}
1383 		DB_DNODE_EXIT(db);
1384 	}
1385 }
1386