xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_tx.c (revision 4a7f2a75ca428b8c3910f76ab6cea0dfadba8914)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/dmu.h>
27 #include <sys/dmu_impl.h>
28 #include <sys/dbuf.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/dmu_objset.h>
31 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
32 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
33 #include <sys/dsl_pool.h>
34 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
35 #include <sys/spa.h>
36 #include <sys/zfs_context.h>
37 
38 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
39     uint64_t arg1, uint64_t arg2);
40 
41 
42 dmu_tx_t *
43 dmu_tx_create_dd(dsl_dir_t *dd)
44 {
45 	dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
46 	tx->tx_dir = dd;
47 	if (dd)
48 		tx->tx_pool = dd->dd_pool;
49 	list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
50 	    offsetof(dmu_tx_hold_t, txh_node));
51 #ifdef ZFS_DEBUG
52 	refcount_create(&tx->tx_space_written);
53 	refcount_create(&tx->tx_space_freed);
54 #endif
55 	return (tx);
56 }
57 
58 dmu_tx_t *
59 dmu_tx_create(objset_t *os)
60 {
61 	dmu_tx_t *tx = dmu_tx_create_dd(os->os->os_dsl_dataset->ds_dir);
62 	tx->tx_objset = os;
63 	tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os->os_dsl_dataset);
64 	return (tx);
65 }
66 
67 dmu_tx_t *
68 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
69 {
70 	dmu_tx_t *tx = dmu_tx_create_dd(NULL);
71 
72 	ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
73 	tx->tx_pool = dp;
74 	tx->tx_txg = txg;
75 	tx->tx_anyobj = TRUE;
76 
77 	return (tx);
78 }
79 
80 int
81 dmu_tx_is_syncing(dmu_tx_t *tx)
82 {
83 	return (tx->tx_anyobj);
84 }
85 
86 int
87 dmu_tx_private_ok(dmu_tx_t *tx)
88 {
89 	return (tx->tx_anyobj);
90 }
91 
92 static dmu_tx_hold_t *
93 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
94     enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
95 {
96 	dmu_tx_hold_t *txh;
97 	dnode_t *dn = NULL;
98 	int err;
99 
100 	if (object != DMU_NEW_OBJECT) {
101 		err = dnode_hold(os->os, object, tx, &dn);
102 		if (err) {
103 			tx->tx_err = err;
104 			return (NULL);
105 		}
106 
107 		if (err == 0 && tx->tx_txg != 0) {
108 			mutex_enter(&dn->dn_mtx);
109 			/*
110 			 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
111 			 * problem, but there's no way for it to happen (for
112 			 * now, at least).
113 			 */
114 			ASSERT(dn->dn_assigned_txg == 0);
115 			dn->dn_assigned_txg = tx->tx_txg;
116 			(void) refcount_add(&dn->dn_tx_holds, tx);
117 			mutex_exit(&dn->dn_mtx);
118 		}
119 	}
120 
121 	txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
122 	txh->txh_tx = tx;
123 	txh->txh_dnode = dn;
124 #ifdef ZFS_DEBUG
125 	txh->txh_type = type;
126 	txh->txh_arg1 = arg1;
127 	txh->txh_arg2 = arg2;
128 #endif
129 	list_insert_tail(&tx->tx_holds, txh);
130 
131 	return (txh);
132 }
133 
134 void
135 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
136 {
137 	/*
138 	 * If we're syncing, they can manipulate any object anyhow, and
139 	 * the hold on the dnode_t can cause problems.
140 	 */
141 	if (!dmu_tx_is_syncing(tx)) {
142 		(void) dmu_tx_hold_object_impl(tx, os,
143 		    object, THT_NEWOBJECT, 0, 0);
144 	}
145 }
146 
147 static int
148 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
149 {
150 	int err;
151 	dmu_buf_impl_t *db;
152 
153 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
154 	db = dbuf_hold_level(dn, level, blkid, FTAG);
155 	rw_exit(&dn->dn_struct_rwlock);
156 	if (db == NULL)
157 		return (EIO);
158 	err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
159 	dbuf_rele(db, FTAG);
160 	return (err);
161 }
162 
163 static void
164 dmu_tx_count_indirects(dmu_tx_hold_t *txh, dmu_buf_impl_t *db,
165     boolean_t freeable, dmu_buf_impl_t **history)
166 {
167 	int i = db->db_level + 1;
168 	dnode_t *dn = db->db_dnode;
169 
170 	if (i >= dn->dn_nlevels)
171 		return;
172 
173 	db = db->db_parent;
174 	if (db == NULL) {
175 		uint64_t lvls = dn->dn_nlevels - i;
176 
177 		txh->txh_space_towrite += lvls << dn->dn_indblkshift;
178 		return;
179 	}
180 
181 	if (db != history[i]) {
182 		dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
183 		uint64_t space = 1ULL << dn->dn_indblkshift;
184 
185 		freeable = (db->db_blkptr && (freeable ||
186 		    dsl_dataset_block_freeable(ds, db->db_blkptr->blk_birth)));
187 		if (freeable)
188 			txh->txh_space_tooverwrite += space;
189 		else
190 			txh->txh_space_towrite += space;
191 		if (db->db_blkptr)
192 			txh->txh_space_tounref += space;
193 		history[i] = db;
194 		dmu_tx_count_indirects(txh, db, freeable, history);
195 	}
196 }
197 
198 /* ARGSUSED */
199 static void
200 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
201 {
202 	dnode_t *dn = txh->txh_dnode;
203 	uint64_t start, end, i;
204 	int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
205 	int err = 0;
206 
207 	if (len == 0)
208 		return;
209 
210 	min_bs = SPA_MINBLOCKSHIFT;
211 	max_bs = SPA_MAXBLOCKSHIFT;
212 	min_ibs = DN_MIN_INDBLKSHIFT;
213 	max_ibs = DN_MAX_INDBLKSHIFT;
214 
215 	if (dn) {
216 		dmu_buf_impl_t *last[DN_MAX_LEVELS];
217 		int nlvls = dn->dn_nlevels;
218 		int delta;
219 
220 		/*
221 		 * For i/o error checking, read the first and last level-0
222 		 * blocks (if they are not aligned), and all the level-1 blocks.
223 		 */
224 		if (dn->dn_maxblkid == 0) {
225 			delta = dn->dn_datablksz;
226 			start = (off < dn->dn_datablksz) ? 0 : 1;
227 			end = (off+len <= dn->dn_datablksz) ? 0 : 1;
228 			if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
229 				err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
230 				if (err)
231 					goto out;
232 				delta -= off;
233 			}
234 		} else {
235 			zio_t *zio = zio_root(dn->dn_objset->os_spa,
236 			    NULL, NULL, ZIO_FLAG_CANFAIL);
237 
238 			/* first level-0 block */
239 			start = off >> dn->dn_datablkshift;
240 			if (P2PHASE(off, dn->dn_datablksz) ||
241 			    len < dn->dn_datablksz) {
242 				err = dmu_tx_check_ioerr(zio, dn, 0, start);
243 				if (err)
244 					goto out;
245 			}
246 
247 			/* last level-0 block */
248 			end = (off+len-1) >> dn->dn_datablkshift;
249 			if (end != start && end <= dn->dn_maxblkid &&
250 			    P2PHASE(off+len, dn->dn_datablksz)) {
251 				err = dmu_tx_check_ioerr(zio, dn, 0, end);
252 				if (err)
253 					goto out;
254 			}
255 
256 			/* level-1 blocks */
257 			if (nlvls > 1) {
258 				int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
259 				for (i = (start>>shft)+1; i < end>>shft; i++) {
260 					err = dmu_tx_check_ioerr(zio, dn, 1, i);
261 					if (err)
262 						goto out;
263 				}
264 			}
265 
266 			err = zio_wait(zio);
267 			if (err)
268 				goto out;
269 			delta = P2NPHASE(off, dn->dn_datablksz);
270 		}
271 
272 		if (dn->dn_maxblkid > 0) {
273 			/*
274 			 * The blocksize can't change,
275 			 * so we can make a more precise estimate.
276 			 */
277 			ASSERT(dn->dn_datablkshift != 0);
278 			min_bs = max_bs = dn->dn_datablkshift;
279 			min_ibs = max_ibs = dn->dn_indblkshift;
280 		} else if (dn->dn_indblkshift > max_ibs) {
281 			/*
282 			 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
283 			 * the code will still work correctly on older pools.
284 			 */
285 			min_ibs = max_ibs = dn->dn_indblkshift;
286 		}
287 
288 		/*
289 		 * If this write is not off the end of the file
290 		 * we need to account for overwrites/unref.
291 		 */
292 		if (start <= dn->dn_maxblkid)
293 			bzero(last, sizeof (dmu_buf_impl_t *) * DN_MAX_LEVELS);
294 		while (start <= dn->dn_maxblkid) {
295 			spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
296 			dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
297 			dmu_buf_impl_t *db;
298 
299 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
300 			db = dbuf_hold_level(dn, 0, start, FTAG);
301 			rw_exit(&dn->dn_struct_rwlock);
302 			if (db->db_blkptr && dsl_dataset_block_freeable(ds,
303 			    db->db_blkptr->blk_birth)) {
304 				dprintf_bp(db->db_blkptr, "can free old%s", "");
305 				txh->txh_space_tooverwrite += dn->dn_datablksz;
306 				txh->txh_space_tounref += dn->dn_datablksz;
307 				dmu_tx_count_indirects(txh, db, TRUE, last);
308 			} else {
309 				txh->txh_space_towrite += dn->dn_datablksz;
310 				if (db->db_blkptr)
311 					txh->txh_space_tounref +=
312 					    bp_get_dasize(spa, db->db_blkptr);
313 				dmu_tx_count_indirects(txh, db, FALSE, last);
314 			}
315 			dbuf_rele(db, FTAG);
316 			if (++start > end) {
317 				/*
318 				 * Account for new indirects appearing
319 				 * before this IO gets assigned into a txg.
320 				 */
321 				bits = 64 - min_bs;
322 				epbs = min_ibs - SPA_BLKPTRSHIFT;
323 				for (bits -= epbs * (nlvls - 1);
324 				    bits >= 0; bits -= epbs)
325 					txh->txh_fudge += 1ULL << max_ibs;
326 				goto out;
327 			}
328 			off += delta;
329 			if (len >= delta)
330 				len -= delta;
331 			delta = dn->dn_datablksz;
332 		}
333 	}
334 
335 	/*
336 	 * 'end' is the last thing we will access, not one past.
337 	 * This way we won't overflow when accessing the last byte.
338 	 */
339 	start = P2ALIGN(off, 1ULL << max_bs);
340 	end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
341 	txh->txh_space_towrite += end - start + 1;
342 
343 	start >>= min_bs;
344 	end >>= min_bs;
345 
346 	epbs = min_ibs - SPA_BLKPTRSHIFT;
347 
348 	/*
349 	 * The object contains at most 2^(64 - min_bs) blocks,
350 	 * and each indirect level maps 2^epbs.
351 	 */
352 	for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
353 		start >>= epbs;
354 		end >>= epbs;
355 		ASSERT3U(end, >=, start);
356 		txh->txh_space_towrite += (end - start + 1) << max_ibs;
357 		if (start != 0) {
358 			/*
359 			 * We also need a new blkid=0 indirect block
360 			 * to reference any existing file data.
361 			 */
362 			txh->txh_space_towrite += 1ULL << max_ibs;
363 		}
364 	}
365 
366 out:
367 	if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
368 	    2 * DMU_MAX_ACCESS)
369 		err = EFBIG;
370 
371 	if (err)
372 		txh->txh_tx->tx_err = err;
373 }
374 
375 static void
376 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
377 {
378 	dnode_t *dn = txh->txh_dnode;
379 	dnode_t *mdn = txh->txh_tx->tx_objset->os->os_meta_dnode;
380 	uint64_t space = mdn->dn_datablksz +
381 	    ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
382 
383 	if (dn && dn->dn_dbuf->db_blkptr &&
384 	    dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
385 	    dn->dn_dbuf->db_blkptr->blk_birth)) {
386 		txh->txh_space_tooverwrite += space;
387 		txh->txh_space_tounref += space;
388 	} else {
389 		txh->txh_space_towrite += space;
390 		if (dn && dn->dn_dbuf->db_blkptr)
391 			txh->txh_space_tounref += space;
392 	}
393 }
394 
395 void
396 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
397 {
398 	dmu_tx_hold_t *txh;
399 
400 	ASSERT(tx->tx_txg == 0);
401 	ASSERT(len < DMU_MAX_ACCESS);
402 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
403 
404 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
405 	    object, THT_WRITE, off, len);
406 	if (txh == NULL)
407 		return;
408 
409 	dmu_tx_count_write(txh, off, len);
410 	dmu_tx_count_dnode(txh);
411 }
412 
413 static void
414 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
415 {
416 	uint64_t blkid, nblks, lastblk;
417 	uint64_t space = 0, unref = 0, skipped = 0;
418 	dnode_t *dn = txh->txh_dnode;
419 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
420 	spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
421 	int epbs;
422 
423 	if (dn->dn_nlevels == 0)
424 		return;
425 
426 	/*
427 	 * The struct_rwlock protects us against dn_nlevels
428 	 * changing, in case (against all odds) we manage to dirty &
429 	 * sync out the changes after we check for being dirty.
430 	 * Also, dbuf_hold_level() wants us to have the struct_rwlock.
431 	 */
432 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
433 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
434 	if (dn->dn_maxblkid == 0) {
435 		if (off == 0 && len >= dn->dn_datablksz) {
436 			blkid = 0;
437 			nblks = 1;
438 		} else {
439 			rw_exit(&dn->dn_struct_rwlock);
440 			return;
441 		}
442 	} else {
443 		blkid = off >> dn->dn_datablkshift;
444 		nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
445 
446 		if (blkid >= dn->dn_maxblkid) {
447 			rw_exit(&dn->dn_struct_rwlock);
448 			return;
449 		}
450 		if (blkid + nblks > dn->dn_maxblkid)
451 			nblks = dn->dn_maxblkid - blkid;
452 
453 	}
454 	if (dn->dn_nlevels == 1) {
455 		int i;
456 		for (i = 0; i < nblks; i++) {
457 			blkptr_t *bp = dn->dn_phys->dn_blkptr;
458 			ASSERT3U(blkid + i, <, dn->dn_nblkptr);
459 			bp += blkid + i;
460 			if (dsl_dataset_block_freeable(ds, bp->blk_birth)) {
461 				dprintf_bp(bp, "can free old%s", "");
462 				space += bp_get_dasize(spa, bp);
463 			}
464 			unref += BP_GET_ASIZE(bp);
465 		}
466 		nblks = 0;
467 	}
468 
469 	/*
470 	 * Add in memory requirements of higher-level indirects.
471 	 * This assumes a worst-possible scenario for dn_nlevels.
472 	 */
473 	{
474 		uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
475 		int level = (dn->dn_nlevels > 1) ? 2 : 1;
476 
477 		while (level++ < DN_MAX_LEVELS) {
478 			txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
479 			blkcnt = 1 + (blkcnt >> epbs);
480 		}
481 		ASSERT(blkcnt <= dn->dn_nblkptr);
482 	}
483 
484 	lastblk = blkid + nblks - 1;
485 	while (nblks) {
486 		dmu_buf_impl_t *dbuf;
487 		uint64_t ibyte, new_blkid;
488 		int epb = 1 << epbs;
489 		int err, i, blkoff, tochk;
490 		blkptr_t *bp;
491 
492 		ibyte = blkid << dn->dn_datablkshift;
493 		err = dnode_next_offset(dn,
494 		    DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
495 		new_blkid = ibyte >> dn->dn_datablkshift;
496 		if (err == ESRCH) {
497 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
498 			break;
499 		}
500 		if (err) {
501 			txh->txh_tx->tx_err = err;
502 			break;
503 		}
504 		if (new_blkid > lastblk) {
505 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
506 			break;
507 		}
508 
509 		if (new_blkid > blkid) {
510 			ASSERT((new_blkid >> epbs) > (blkid >> epbs));
511 			skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
512 			nblks -= new_blkid - blkid;
513 			blkid = new_blkid;
514 		}
515 		blkoff = P2PHASE(blkid, epb);
516 		tochk = MIN(epb - blkoff, nblks);
517 
518 		dbuf = dbuf_hold_level(dn, 1, blkid >> epbs, FTAG);
519 
520 		txh->txh_memory_tohold += dbuf->db.db_size;
521 		if (txh->txh_memory_tohold > DMU_MAX_ACCESS) {
522 			txh->txh_tx->tx_err = E2BIG;
523 			dbuf_rele(dbuf, FTAG);
524 			break;
525 		}
526 		err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
527 		if (err != 0) {
528 			txh->txh_tx->tx_err = err;
529 			dbuf_rele(dbuf, FTAG);
530 			break;
531 		}
532 
533 		bp = dbuf->db.db_data;
534 		bp += blkoff;
535 
536 		for (i = 0; i < tochk; i++) {
537 			if (dsl_dataset_block_freeable(ds, bp[i].blk_birth)) {
538 				dprintf_bp(&bp[i], "can free old%s", "");
539 				space += bp_get_dasize(spa, &bp[i]);
540 			}
541 			unref += BP_GET_ASIZE(bp);
542 		}
543 		dbuf_rele(dbuf, FTAG);
544 
545 		blkid += tochk;
546 		nblks -= tochk;
547 	}
548 	rw_exit(&dn->dn_struct_rwlock);
549 
550 	/* account for new level 1 indirect blocks that might show up */
551 	if (skipped > 0) {
552 		txh->txh_fudge += skipped << dn->dn_indblkshift;
553 		skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
554 		txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
555 	}
556 	txh->txh_space_tofree += space;
557 	txh->txh_space_tounref += unref;
558 }
559 
560 void
561 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
562 {
563 	dmu_tx_hold_t *txh;
564 	dnode_t *dn;
565 	uint64_t start, end, i;
566 	int err, shift;
567 	zio_t *zio;
568 
569 	ASSERT(tx->tx_txg == 0);
570 
571 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
572 	    object, THT_FREE, off, len);
573 	if (txh == NULL)
574 		return;
575 	dn = txh->txh_dnode;
576 
577 	/* first block */
578 	if (off != 0)
579 		dmu_tx_count_write(txh, off, 1);
580 	/* last block */
581 	if (len != DMU_OBJECT_END)
582 		dmu_tx_count_write(txh, off+len, 1);
583 
584 	if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
585 		return;
586 	if (len == DMU_OBJECT_END)
587 		len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
588 
589 	/*
590 	 * For i/o error checking, read the first and last level-0
591 	 * blocks, and all the level-1 blocks.  The above count_write's
592 	 * have already taken care of the level-0 blocks.
593 	 */
594 	if (dn->dn_nlevels > 1) {
595 		shift = dn->dn_datablkshift + dn->dn_indblkshift -
596 		    SPA_BLKPTRSHIFT;
597 		start = off >> shift;
598 		end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
599 
600 		zio = zio_root(tx->tx_pool->dp_spa,
601 		    NULL, NULL, ZIO_FLAG_CANFAIL);
602 		for (i = start; i <= end; i++) {
603 			uint64_t ibyte = i << shift;
604 			err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
605 			i = ibyte >> shift;
606 			if (err == ESRCH)
607 				break;
608 			if (err) {
609 				tx->tx_err = err;
610 				return;
611 			}
612 
613 			err = dmu_tx_check_ioerr(zio, dn, 1, i);
614 			if (err) {
615 				tx->tx_err = err;
616 				return;
617 			}
618 		}
619 		err = zio_wait(zio);
620 		if (err) {
621 			tx->tx_err = err;
622 			return;
623 		}
624 	}
625 
626 	dmu_tx_count_dnode(txh);
627 	dmu_tx_count_free(txh, off, len);
628 }
629 
630 void
631 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, char *name)
632 {
633 	dmu_tx_hold_t *txh;
634 	dnode_t *dn;
635 	uint64_t nblocks;
636 	int epbs, err;
637 
638 	ASSERT(tx->tx_txg == 0);
639 
640 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
641 	    object, THT_ZAP, add, (uintptr_t)name);
642 	if (txh == NULL)
643 		return;
644 	dn = txh->txh_dnode;
645 
646 	dmu_tx_count_dnode(txh);
647 
648 	if (dn == NULL) {
649 		/*
650 		 * We will be able to fit a new object's entries into one leaf
651 		 * block.  So there will be at most 2 blocks total,
652 		 * including the header block.
653 		 */
654 		dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
655 		return;
656 	}
657 
658 	ASSERT3P(dmu_ot[dn->dn_type].ot_byteswap, ==, zap_byteswap);
659 
660 	if (dn->dn_maxblkid == 0 && !add) {
661 		/*
662 		 * If there is only one block  (i.e. this is a micro-zap)
663 		 * and we are not adding anything, the accounting is simple.
664 		 */
665 		err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
666 		if (err) {
667 			tx->tx_err = err;
668 			return;
669 		}
670 
671 		/*
672 		 * Use max block size here, since we don't know how much
673 		 * the size will change between now and the dbuf dirty call.
674 		 */
675 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
676 		    dn->dn_phys->dn_blkptr[0].blk_birth)) {
677 			txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
678 		} else {
679 			txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
680 			txh->txh_space_tounref +=
681 			    BP_GET_ASIZE(dn->dn_phys->dn_blkptr);
682 		}
683 		return;
684 	}
685 
686 	if (dn->dn_maxblkid > 0 && name) {
687 		/*
688 		 * access the name in this fat-zap so that we'll check
689 		 * for i/o errors to the leaf blocks, etc.
690 		 */
691 		err = zap_lookup(&dn->dn_objset->os, dn->dn_object, name,
692 		    8, 0, NULL);
693 		if (err == EIO) {
694 			tx->tx_err = err;
695 			return;
696 		}
697 	}
698 
699 	/*
700 	 * 3 blocks overwritten: target leaf, ptrtbl block, header block
701 	 * 3 new blocks written if adding: new split leaf, 2 grown ptrtbl blocks
702 	 */
703 	dmu_tx_count_write(txh, dn->dn_maxblkid * dn->dn_datablksz,
704 	    (3 + (add ? 3 : 0)) * dn->dn_datablksz);
705 
706 	/*
707 	 * If the modified blocks are scattered to the four winds,
708 	 * we'll have to modify an indirect twig for each.
709 	 */
710 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
711 	for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
712 		txh->txh_space_towrite += 3 << dn->dn_indblkshift;
713 }
714 
715 void
716 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
717 {
718 	dmu_tx_hold_t *txh;
719 
720 	ASSERT(tx->tx_txg == 0);
721 
722 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
723 	    object, THT_BONUS, 0, 0);
724 	if (txh)
725 		dmu_tx_count_dnode(txh);
726 }
727 
728 void
729 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
730 {
731 	dmu_tx_hold_t *txh;
732 	ASSERT(tx->tx_txg == 0);
733 
734 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
735 	    DMU_NEW_OBJECT, THT_SPACE, space, 0);
736 
737 	txh->txh_space_towrite += space;
738 }
739 
740 int
741 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
742 {
743 	dmu_tx_hold_t *txh;
744 	int holds = 0;
745 
746 	/*
747 	 * By asserting that the tx is assigned, we're counting the
748 	 * number of dn_tx_holds, which is the same as the number of
749 	 * dn_holds.  Otherwise, we'd be counting dn_holds, but
750 	 * dn_tx_holds could be 0.
751 	 */
752 	ASSERT(tx->tx_txg != 0);
753 
754 	/* if (tx->tx_anyobj == TRUE) */
755 		/* return (0); */
756 
757 	for (txh = list_head(&tx->tx_holds); txh;
758 	    txh = list_next(&tx->tx_holds, txh)) {
759 		if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
760 			holds++;
761 	}
762 
763 	return (holds);
764 }
765 
766 #ifdef ZFS_DEBUG
767 void
768 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
769 {
770 	dmu_tx_hold_t *txh;
771 	int match_object = FALSE, match_offset = FALSE;
772 	dnode_t *dn = db->db_dnode;
773 
774 	ASSERT(tx->tx_txg != 0);
775 	ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset->os);
776 	ASSERT3U(dn->dn_object, ==, db->db.db_object);
777 
778 	if (tx->tx_anyobj)
779 		return;
780 
781 	/* XXX No checking on the meta dnode for now */
782 	if (db->db.db_object == DMU_META_DNODE_OBJECT)
783 		return;
784 
785 	for (txh = list_head(&tx->tx_holds); txh;
786 	    txh = list_next(&tx->tx_holds, txh)) {
787 		ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
788 		if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
789 			match_object = TRUE;
790 		if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
791 			int datablkshift = dn->dn_datablkshift ?
792 			    dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
793 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
794 			int shift = datablkshift + epbs * db->db_level;
795 			uint64_t beginblk = shift >= 64 ? 0 :
796 			    (txh->txh_arg1 >> shift);
797 			uint64_t endblk = shift >= 64 ? 0 :
798 			    ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
799 			uint64_t blkid = db->db_blkid;
800 
801 			/* XXX txh_arg2 better not be zero... */
802 
803 			dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
804 			    txh->txh_type, beginblk, endblk);
805 
806 			switch (txh->txh_type) {
807 			case THT_WRITE:
808 				if (blkid >= beginblk && blkid <= endblk)
809 					match_offset = TRUE;
810 				/*
811 				 * We will let this hold work for the bonus
812 				 * buffer so that we don't need to hold it
813 				 * when creating a new object.
814 				 */
815 				if (blkid == DB_BONUS_BLKID)
816 					match_offset = TRUE;
817 				/*
818 				 * They might have to increase nlevels,
819 				 * thus dirtying the new TLIBs.  Or the
820 				 * might have to change the block size,
821 				 * thus dirying the new lvl=0 blk=0.
822 				 */
823 				if (blkid == 0)
824 					match_offset = TRUE;
825 				break;
826 			case THT_FREE:
827 				/*
828 				 * We will dirty all the level 1 blocks in
829 				 * the free range and perhaps the first and
830 				 * last level 0 block.
831 				 */
832 				if (blkid >= beginblk && (blkid <= endblk ||
833 				    txh->txh_arg2 == DMU_OBJECT_END))
834 					match_offset = TRUE;
835 				break;
836 			case THT_BONUS:
837 				if (blkid == DB_BONUS_BLKID)
838 					match_offset = TRUE;
839 				break;
840 			case THT_ZAP:
841 				match_offset = TRUE;
842 				break;
843 			case THT_NEWOBJECT:
844 				match_object = TRUE;
845 				break;
846 			default:
847 				ASSERT(!"bad txh_type");
848 			}
849 		}
850 		if (match_object && match_offset)
851 			return;
852 	}
853 	panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
854 	    (u_longlong_t)db->db.db_object, db->db_level,
855 	    (u_longlong_t)db->db_blkid);
856 }
857 #endif
858 
859 static int
860 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
861 {
862 	dmu_tx_hold_t *txh;
863 	spa_t *spa = tx->tx_pool->dp_spa;
864 	uint64_t memory, asize, fsize, usize;
865 	uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
866 
867 	ASSERT3U(tx->tx_txg, ==, 0);
868 
869 	if (tx->tx_err)
870 		return (tx->tx_err);
871 
872 	if (spa_suspended(spa)) {
873 		/*
874 		 * If the user has indicated a blocking failure mode
875 		 * then return ERESTART which will block in dmu_tx_wait().
876 		 * Otherwise, return EIO so that an error can get
877 		 * propagated back to the VOP calls.
878 		 *
879 		 * Note that we always honor the txg_how flag regardless
880 		 * of the failuremode setting.
881 		 */
882 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
883 		    txg_how != TXG_WAIT)
884 			return (EIO);
885 
886 		return (ERESTART);
887 	}
888 
889 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
890 	tx->tx_needassign_txh = NULL;
891 
892 	/*
893 	 * NB: No error returns are allowed after txg_hold_open, but
894 	 * before processing the dnode holds, due to the
895 	 * dmu_tx_unassign() logic.
896 	 */
897 
898 	towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
899 	for (txh = list_head(&tx->tx_holds); txh;
900 	    txh = list_next(&tx->tx_holds, txh)) {
901 		dnode_t *dn = txh->txh_dnode;
902 		if (dn != NULL) {
903 			mutex_enter(&dn->dn_mtx);
904 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
905 				mutex_exit(&dn->dn_mtx);
906 				tx->tx_needassign_txh = txh;
907 				return (ERESTART);
908 			}
909 			if (dn->dn_assigned_txg == 0)
910 				dn->dn_assigned_txg = tx->tx_txg;
911 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
912 			(void) refcount_add(&dn->dn_tx_holds, tx);
913 			mutex_exit(&dn->dn_mtx);
914 		}
915 		towrite += txh->txh_space_towrite;
916 		tofree += txh->txh_space_tofree;
917 		tooverwrite += txh->txh_space_tooverwrite;
918 		tounref += txh->txh_space_tounref;
919 		tohold += txh->txh_memory_tohold;
920 		fudge += txh->txh_fudge;
921 	}
922 
923 	/*
924 	 * NB: This check must be after we've held the dnodes, so that
925 	 * the dmu_tx_unassign() logic will work properly
926 	 */
927 	if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
928 		return (ERESTART);
929 
930 	/*
931 	 * If a snapshot has been taken since we made our estimates,
932 	 * assume that we won't be able to free or overwrite anything.
933 	 */
934 	if (tx->tx_objset &&
935 	    dsl_dataset_prev_snap_txg(tx->tx_objset->os->os_dsl_dataset) >
936 	    tx->tx_lastsnap_txg) {
937 		towrite += tooverwrite;
938 		tooverwrite = tofree = 0;
939 	}
940 
941 	/* needed allocation: worst-case estimate of write space */
942 	asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
943 	/* freed space estimate: worst-case overwrite + free estimate */
944 	fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
945 	/* convert unrefd space to worst-case estimate */
946 	usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
947 	/* calculate memory footprint estimate */
948 	memory = towrite + tooverwrite + tohold;
949 
950 #ifdef ZFS_DEBUG
951 	/*
952 	 * Add in 'tohold' to account for our dirty holds on this memory
953 	 * XXX - the "fudge" factor is to account for skipped blocks that
954 	 * we missed because dnode_next_offset() misses in-core-only blocks.
955 	 */
956 	tx->tx_space_towrite = asize +
957 	    spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
958 	tx->tx_space_tofree = tofree;
959 	tx->tx_space_tooverwrite = tooverwrite;
960 	tx->tx_space_tounref = tounref;
961 #endif
962 
963 	if (tx->tx_dir && asize != 0) {
964 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
965 		    asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
966 		if (err)
967 			return (err);
968 	}
969 
970 	return (0);
971 }
972 
973 static void
974 dmu_tx_unassign(dmu_tx_t *tx)
975 {
976 	dmu_tx_hold_t *txh;
977 
978 	if (tx->tx_txg == 0)
979 		return;
980 
981 	txg_rele_to_quiesce(&tx->tx_txgh);
982 
983 	for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
984 	    txh = list_next(&tx->tx_holds, txh)) {
985 		dnode_t *dn = txh->txh_dnode;
986 
987 		if (dn == NULL)
988 			continue;
989 		mutex_enter(&dn->dn_mtx);
990 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
991 
992 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
993 			dn->dn_assigned_txg = 0;
994 			cv_broadcast(&dn->dn_notxholds);
995 		}
996 		mutex_exit(&dn->dn_mtx);
997 	}
998 
999 	txg_rele_to_sync(&tx->tx_txgh);
1000 
1001 	tx->tx_lasttried_txg = tx->tx_txg;
1002 	tx->tx_txg = 0;
1003 }
1004 
1005 /*
1006  * Assign tx to a transaction group.  txg_how can be one of:
1007  *
1008  * (1)	TXG_WAIT.  If the current open txg is full, waits until there's
1009  *	a new one.  This should be used when you're not holding locks.
1010  *	If will only fail if we're truly out of space (or over quota).
1011  *
1012  * (2)	TXG_NOWAIT.  If we can't assign into the current open txg without
1013  *	blocking, returns immediately with ERESTART.  This should be used
1014  *	whenever you're holding locks.  On an ERESTART error, the caller
1015  *	should drop locks, do a dmu_tx_wait(tx), and try again.
1016  *
1017  * (3)	A specific txg.  Use this if you need to ensure that multiple
1018  *	transactions all sync in the same txg.  Like TXG_NOWAIT, it
1019  *	returns ERESTART if it can't assign you into the requested txg.
1020  */
1021 int
1022 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1023 {
1024 	int err;
1025 
1026 	ASSERT(tx->tx_txg == 0);
1027 	ASSERT(txg_how != 0);
1028 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1029 
1030 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1031 		dmu_tx_unassign(tx);
1032 
1033 		if (err != ERESTART || txg_how != TXG_WAIT)
1034 			return (err);
1035 
1036 		dmu_tx_wait(tx);
1037 	}
1038 
1039 	txg_rele_to_quiesce(&tx->tx_txgh);
1040 
1041 	return (0);
1042 }
1043 
1044 void
1045 dmu_tx_wait(dmu_tx_t *tx)
1046 {
1047 	spa_t *spa = tx->tx_pool->dp_spa;
1048 
1049 	ASSERT(tx->tx_txg == 0);
1050 
1051 	/*
1052 	 * It's possible that the pool has become active after this thread
1053 	 * has tried to obtain a tx. If that's the case then his
1054 	 * tx_lasttried_txg would not have been assigned.
1055 	 */
1056 	if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1057 		txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1058 	} else if (tx->tx_needassign_txh) {
1059 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1060 
1061 		mutex_enter(&dn->dn_mtx);
1062 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1063 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1064 		mutex_exit(&dn->dn_mtx);
1065 		tx->tx_needassign_txh = NULL;
1066 	} else {
1067 		txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1068 	}
1069 }
1070 
1071 void
1072 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1073 {
1074 #ifdef ZFS_DEBUG
1075 	if (tx->tx_dir == NULL || delta == 0)
1076 		return;
1077 
1078 	if (delta > 0) {
1079 		ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1080 		    tx->tx_space_towrite);
1081 		(void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1082 	} else {
1083 		(void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1084 	}
1085 #endif
1086 }
1087 
1088 void
1089 dmu_tx_commit(dmu_tx_t *tx)
1090 {
1091 	dmu_tx_hold_t *txh;
1092 
1093 	ASSERT(tx->tx_txg != 0);
1094 
1095 	while (txh = list_head(&tx->tx_holds)) {
1096 		dnode_t *dn = txh->txh_dnode;
1097 
1098 		list_remove(&tx->tx_holds, txh);
1099 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1100 		if (dn == NULL)
1101 			continue;
1102 		mutex_enter(&dn->dn_mtx);
1103 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1104 
1105 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1106 			dn->dn_assigned_txg = 0;
1107 			cv_broadcast(&dn->dn_notxholds);
1108 		}
1109 		mutex_exit(&dn->dn_mtx);
1110 		dnode_rele(dn, tx);
1111 	}
1112 
1113 	if (tx->tx_tempreserve_cookie)
1114 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1115 
1116 	if (tx->tx_anyobj == FALSE)
1117 		txg_rele_to_sync(&tx->tx_txgh);
1118 	list_destroy(&tx->tx_holds);
1119 #ifdef ZFS_DEBUG
1120 	dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1121 	    tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1122 	    tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1123 	refcount_destroy_many(&tx->tx_space_written,
1124 	    refcount_count(&tx->tx_space_written));
1125 	refcount_destroy_many(&tx->tx_space_freed,
1126 	    refcount_count(&tx->tx_space_freed));
1127 #endif
1128 	kmem_free(tx, sizeof (dmu_tx_t));
1129 }
1130 
1131 void
1132 dmu_tx_abort(dmu_tx_t *tx)
1133 {
1134 	dmu_tx_hold_t *txh;
1135 
1136 	ASSERT(tx->tx_txg == 0);
1137 
1138 	while (txh = list_head(&tx->tx_holds)) {
1139 		dnode_t *dn = txh->txh_dnode;
1140 
1141 		list_remove(&tx->tx_holds, txh);
1142 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1143 		if (dn != NULL)
1144 			dnode_rele(dn, tx);
1145 	}
1146 	list_destroy(&tx->tx_holds);
1147 #ifdef ZFS_DEBUG
1148 	refcount_destroy_many(&tx->tx_space_written,
1149 	    refcount_count(&tx->tx_space_written));
1150 	refcount_destroy_many(&tx->tx_space_freed,
1151 	    refcount_count(&tx->tx_space_freed));
1152 #endif
1153 	kmem_free(tx, sizeof (dmu_tx_t));
1154 }
1155 
1156 uint64_t
1157 dmu_tx_get_txg(dmu_tx_t *tx)
1158 {
1159 	ASSERT(tx->tx_txg != 0);
1160 	return (tx->tx_txg);
1161 }
1162