xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_tx.c (revision 39e9b01c2242a38babd9d9a1d736c4388ac7a972)
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 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/dmu.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dbuf.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
34 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
35 #include <sys/dsl_pool.h>
36 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
37 #include <sys/spa.h>
38 #include <sys/zfs_context.h>
39 
40 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
41     uint64_t arg1, uint64_t arg2);
42 
43 
44 dmu_tx_t *
45 dmu_tx_create_dd(dsl_dir_t *dd)
46 {
47 	dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
48 	tx->tx_dir = dd;
49 	if (dd)
50 		tx->tx_pool = dd->dd_pool;
51 	list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
52 	    offsetof(dmu_tx_hold_t, txh_node));
53 #ifdef ZFS_DEBUG
54 	refcount_create(&tx->tx_space_written);
55 	refcount_create(&tx->tx_space_freed);
56 #endif
57 	return (tx);
58 }
59 
60 dmu_tx_t *
61 dmu_tx_create(objset_t *os)
62 {
63 	dmu_tx_t *tx = dmu_tx_create_dd(os->os->os_dsl_dataset->ds_dir);
64 	tx->tx_objset = os;
65 	tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os->os_dsl_dataset);
66 	return (tx);
67 }
68 
69 dmu_tx_t *
70 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
71 {
72 	dmu_tx_t *tx = dmu_tx_create_dd(NULL);
73 
74 	ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
75 	tx->tx_pool = dp;
76 	tx->tx_txg = txg;
77 	tx->tx_anyobj = TRUE;
78 
79 	return (tx);
80 }
81 
82 int
83 dmu_tx_is_syncing(dmu_tx_t *tx)
84 {
85 	return (tx->tx_anyobj);
86 }
87 
88 int
89 dmu_tx_private_ok(dmu_tx_t *tx)
90 {
91 	return (tx->tx_anyobj);
92 }
93 
94 static dmu_tx_hold_t *
95 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
96     enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
97 {
98 	dmu_tx_hold_t *txh;
99 	dnode_t *dn = NULL;
100 	int err;
101 
102 	if (object != DMU_NEW_OBJECT) {
103 		err = dnode_hold(os->os, object, tx, &dn);
104 		if (err) {
105 			tx->tx_err = err;
106 			return (NULL);
107 		}
108 
109 		if (err == 0 && tx->tx_txg != 0) {
110 			mutex_enter(&dn->dn_mtx);
111 			/*
112 			 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
113 			 * problem, but there's no way for it to happen (for
114 			 * now, at least).
115 			 */
116 			ASSERT(dn->dn_assigned_txg == 0);
117 			dn->dn_assigned_txg = tx->tx_txg;
118 			(void) refcount_add(&dn->dn_tx_holds, tx);
119 			mutex_exit(&dn->dn_mtx);
120 		}
121 	}
122 
123 	txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
124 	txh->txh_tx = tx;
125 	txh->txh_dnode = dn;
126 #ifdef ZFS_DEBUG
127 	txh->txh_type = type;
128 	txh->txh_arg1 = arg1;
129 	txh->txh_arg2 = arg2;
130 #endif
131 	list_insert_tail(&tx->tx_holds, txh);
132 
133 	return (txh);
134 }
135 
136 void
137 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
138 {
139 	/*
140 	 * If we're syncing, they can manipulate any object anyhow, and
141 	 * the hold on the dnode_t can cause problems.
142 	 */
143 	if (!dmu_tx_is_syncing(tx)) {
144 		(void) dmu_tx_hold_object_impl(tx, os,
145 		    object, THT_NEWOBJECT, 0, 0);
146 	}
147 }
148 
149 static int
150 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
151 {
152 	int err;
153 	dmu_buf_impl_t *db;
154 
155 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
156 	db = dbuf_hold_level(dn, level, blkid, FTAG);
157 	rw_exit(&dn->dn_struct_rwlock);
158 	if (db == NULL)
159 		return (EIO);
160 	err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
161 	dbuf_rele(db, FTAG);
162 	return (err);
163 }
164 
165 /* ARGSUSED */
166 static void
167 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
168 {
169 	dnode_t *dn = txh->txh_dnode;
170 	uint64_t start, end, i;
171 	int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
172 	int err = 0;
173 
174 	if (len == 0)
175 		return;
176 
177 	min_bs = SPA_MINBLOCKSHIFT;
178 	max_bs = SPA_MAXBLOCKSHIFT;
179 	min_ibs = DN_MIN_INDBLKSHIFT;
180 	max_ibs = DN_MAX_INDBLKSHIFT;
181 
182 
183 	/*
184 	 * For i/o error checking, read the first and last level-0
185 	 * blocks (if they are not aligned), and all the level-1 blocks.
186 	 */
187 
188 	if (dn) {
189 		if (dn->dn_maxblkid == 0) {
190 			err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
191 			if (err)
192 				goto out;
193 		} else {
194 			zio_t *zio = zio_root(dn->dn_objset->os_spa,
195 			    NULL, NULL, ZIO_FLAG_CANFAIL);
196 
197 			/* first level-0 block */
198 			start = off >> dn->dn_datablkshift;
199 			if (P2PHASE(off, dn->dn_datablksz) ||
200 			    len < dn->dn_datablksz) {
201 				err = dmu_tx_check_ioerr(zio, dn, 0, start);
202 				if (err)
203 					goto out;
204 			}
205 
206 			/* last level-0 block */
207 			end = (off+len-1) >> dn->dn_datablkshift;
208 			if (end != start &&
209 			    P2PHASE(off+len, dn->dn_datablksz)) {
210 				err = dmu_tx_check_ioerr(zio, dn, 0, end);
211 				if (err)
212 					goto out;
213 			}
214 
215 			/* level-1 blocks */
216 			if (dn->dn_nlevels > 1) {
217 				start >>= dn->dn_indblkshift - SPA_BLKPTRSHIFT;
218 				end >>= dn->dn_indblkshift - SPA_BLKPTRSHIFT;
219 				for (i = start+1; i < end; i++) {
220 					err = dmu_tx_check_ioerr(zio, dn, 1, i);
221 					if (err)
222 						goto out;
223 				}
224 			}
225 
226 			err = zio_wait(zio);
227 			if (err)
228 				goto out;
229 		}
230 	}
231 
232 	/*
233 	 * If there's more than one block, the blocksize can't change,
234 	 * so we can make a more precise estimate.  Alternatively,
235 	 * if the dnode's ibs is larger than max_ibs, always use that.
236 	 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
237 	 * the code will still work correctly on existing pools.
238 	 */
239 	if (dn && (dn->dn_maxblkid != 0 || dn->dn_indblkshift > max_ibs)) {
240 		min_ibs = max_ibs = dn->dn_indblkshift;
241 		if (dn->dn_datablkshift != 0)
242 			min_bs = max_bs = dn->dn_datablkshift;
243 	}
244 
245 	/*
246 	 * 'end' is the last thing we will access, not one past.
247 	 * This way we won't overflow when accessing the last byte.
248 	 */
249 	start = P2ALIGN(off, 1ULL << max_bs);
250 	end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
251 	txh->txh_space_towrite += end - start + 1;
252 
253 	start >>= min_bs;
254 	end >>= min_bs;
255 
256 	epbs = min_ibs - SPA_BLKPTRSHIFT;
257 
258 	/*
259 	 * The object contains at most 2^(64 - min_bs) blocks,
260 	 * and each indirect level maps 2^epbs.
261 	 */
262 	for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
263 		start >>= epbs;
264 		end >>= epbs;
265 		/*
266 		 * If we increase the number of levels of indirection,
267 		 * we'll need new blkid=0 indirect blocks.  If start == 0,
268 		 * we're already accounting for that blocks; and if end == 0,
269 		 * we can't increase the number of levels beyond that.
270 		 */
271 		if (start != 0 && end != 0)
272 			txh->txh_space_towrite += 1ULL << max_ibs;
273 		txh->txh_space_towrite += (end - start + 1) << max_ibs;
274 	}
275 
276 	ASSERT(txh->txh_space_towrite < 2 * DMU_MAX_ACCESS);
277 
278 out:
279 	if (err)
280 		txh->txh_tx->tx_err = err;
281 }
282 
283 static void
284 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
285 {
286 	dnode_t *dn = txh->txh_dnode;
287 	dnode_t *mdn = txh->txh_tx->tx_objset->os->os_meta_dnode;
288 	uint64_t space = mdn->dn_datablksz +
289 	    ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
290 
291 	if (dn && dn->dn_dbuf->db_blkptr &&
292 	    dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
293 	    dn->dn_dbuf->db_blkptr->blk_birth)) {
294 		txh->txh_space_tooverwrite += space;
295 	} else {
296 		txh->txh_space_towrite += space;
297 		if (dn && dn->dn_dbuf->db_blkptr)
298 			txh->txh_space_tounref += space;
299 	}
300 }
301 
302 void
303 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
304 {
305 	dmu_tx_hold_t *txh;
306 
307 	ASSERT(tx->tx_txg == 0);
308 	ASSERT(len < DMU_MAX_ACCESS);
309 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
310 
311 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
312 	    object, THT_WRITE, off, len);
313 	if (txh == NULL)
314 		return;
315 
316 	dmu_tx_count_write(txh, off, len);
317 	dmu_tx_count_dnode(txh);
318 }
319 
320 static void
321 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
322 {
323 	uint64_t blkid, nblks, lastblk;
324 	uint64_t space = 0, unref = 0, skipped = 0;
325 	dnode_t *dn = txh->txh_dnode;
326 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
327 	spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
328 	int epbs;
329 
330 	if (dn->dn_nlevels == 0)
331 		return;
332 
333 	/*
334 	 * The struct_rwlock protects us against dn_nlevels
335 	 * changing, in case (against all odds) we manage to dirty &
336 	 * sync out the changes after we check for being dirty.
337 	 * Also, dbuf_hold_level() wants us to have the struct_rwlock.
338 	 */
339 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
340 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
341 	if (dn->dn_maxblkid == 0) {
342 		if (off == 0 && len >= dn->dn_datablksz) {
343 			blkid = 0;
344 			nblks = 1;
345 		} else {
346 			rw_exit(&dn->dn_struct_rwlock);
347 			return;
348 		}
349 	} else {
350 		blkid = off >> dn->dn_datablkshift;
351 		nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
352 
353 		if (blkid >= dn->dn_maxblkid) {
354 			rw_exit(&dn->dn_struct_rwlock);
355 			return;
356 		}
357 		if (blkid + nblks > dn->dn_maxblkid)
358 			nblks = dn->dn_maxblkid - blkid;
359 
360 	}
361 	if (dn->dn_nlevels == 1) {
362 		int i;
363 		for (i = 0; i < nblks; i++) {
364 			blkptr_t *bp = dn->dn_phys->dn_blkptr;
365 			ASSERT3U(blkid + i, <, dn->dn_nblkptr);
366 			bp += blkid + i;
367 			if (dsl_dataset_block_freeable(ds, bp->blk_birth)) {
368 				dprintf_bp(bp, "can free old%s", "");
369 				space += bp_get_dasize(spa, bp);
370 			}
371 			unref += BP_GET_ASIZE(bp);
372 		}
373 		nblks = 0;
374 	}
375 
376 	/*
377 	 * Add in memory requirements of higher-level indirects.
378 	 * This assumes a worst-possible scenario for dn_nlevels.
379 	 */
380 	{
381 		uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
382 		int level = (dn->dn_nlevels > 1) ? 2 : 1;
383 
384 		while (level++ < DN_MAX_LEVELS) {
385 			txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
386 			blkcnt = 1 + (blkcnt >> epbs);
387 		}
388 		ASSERT(blkcnt <= dn->dn_nblkptr);
389 	}
390 
391 	lastblk = blkid + nblks - 1;
392 	while (nblks) {
393 		dmu_buf_impl_t *dbuf;
394 		uint64_t ibyte, new_blkid;
395 		int epb = 1 << epbs;
396 		int err, i, blkoff, tochk;
397 		blkptr_t *bp;
398 
399 		ibyte = blkid << dn->dn_datablkshift;
400 		err = dnode_next_offset(dn,
401 		    DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
402 		new_blkid = ibyte >> dn->dn_datablkshift;
403 		if (err == ESRCH) {
404 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
405 			break;
406 		}
407 		if (err) {
408 			txh->txh_tx->tx_err = err;
409 			break;
410 		}
411 		if (new_blkid > lastblk) {
412 			skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
413 			break;
414 		}
415 
416 		if (new_blkid > blkid) {
417 			ASSERT((new_blkid >> epbs) > (blkid >> epbs));
418 			skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
419 			nblks -= new_blkid - blkid;
420 			blkid = new_blkid;
421 		}
422 		blkoff = P2PHASE(blkid, epb);
423 		tochk = MIN(epb - blkoff, nblks);
424 
425 		dbuf = dbuf_hold_level(dn, 1, blkid >> epbs, FTAG);
426 
427 		txh->txh_memory_tohold += dbuf->db.db_size;
428 		if (txh->txh_memory_tohold > DMU_MAX_ACCESS) {
429 			txh->txh_tx->tx_err = E2BIG;
430 			dbuf_rele(dbuf, FTAG);
431 			break;
432 		}
433 		err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
434 		if (err != 0) {
435 			txh->txh_tx->tx_err = err;
436 			dbuf_rele(dbuf, FTAG);
437 			break;
438 		}
439 
440 		bp = dbuf->db.db_data;
441 		bp += blkoff;
442 
443 		for (i = 0; i < tochk; i++) {
444 			if (dsl_dataset_block_freeable(ds, bp[i].blk_birth)) {
445 				dprintf_bp(&bp[i], "can free old%s", "");
446 				space += bp_get_dasize(spa, &bp[i]);
447 			}
448 			unref += BP_GET_ASIZE(bp);
449 		}
450 		dbuf_rele(dbuf, FTAG);
451 
452 		blkid += tochk;
453 		nblks -= tochk;
454 	}
455 	rw_exit(&dn->dn_struct_rwlock);
456 
457 	/* account for new level 1 indirect blocks that might show up */
458 	if (skipped > 0) {
459 		txh->txh_fudge += skipped << dn->dn_indblkshift;
460 		skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
461 		txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
462 	}
463 	txh->txh_space_tofree += space;
464 	txh->txh_space_tounref += unref;
465 }
466 
467 void
468 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
469 {
470 	dmu_tx_hold_t *txh;
471 	dnode_t *dn;
472 	uint64_t start, end, i;
473 	int err, shift;
474 	zio_t *zio;
475 
476 	ASSERT(tx->tx_txg == 0);
477 
478 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
479 	    object, THT_FREE, off, len);
480 	if (txh == NULL)
481 		return;
482 	dn = txh->txh_dnode;
483 
484 	/* first block */
485 	if (off != 0)
486 		dmu_tx_count_write(txh, off, 1);
487 	/* last block */
488 	if (len != DMU_OBJECT_END)
489 		dmu_tx_count_write(txh, off+len, 1);
490 
491 	if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
492 		return;
493 	if (len == DMU_OBJECT_END)
494 		len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
495 
496 	/*
497 	 * For i/o error checking, read the first and last level-0
498 	 * blocks, and all the level-1 blocks.  The above count_write's
499 	 * have already taken care of the level-0 blocks.
500 	 */
501 	if (dn->dn_nlevels > 1) {
502 		shift = dn->dn_datablkshift + dn->dn_indblkshift -
503 		    SPA_BLKPTRSHIFT;
504 		start = off >> shift;
505 		end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
506 
507 		zio = zio_root(tx->tx_pool->dp_spa,
508 		    NULL, NULL, ZIO_FLAG_CANFAIL);
509 		for (i = start; i <= end; i++) {
510 			uint64_t ibyte = i << shift;
511 			err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
512 			i = ibyte >> shift;
513 			if (err == ESRCH)
514 				break;
515 			if (err) {
516 				tx->tx_err = err;
517 				return;
518 			}
519 
520 			err = dmu_tx_check_ioerr(zio, dn, 1, i);
521 			if (err) {
522 				tx->tx_err = err;
523 				return;
524 			}
525 		}
526 		err = zio_wait(zio);
527 		if (err) {
528 			tx->tx_err = err;
529 			return;
530 		}
531 	}
532 
533 	dmu_tx_count_dnode(txh);
534 	dmu_tx_count_free(txh, off, len);
535 }
536 
537 void
538 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, char *name)
539 {
540 	dmu_tx_hold_t *txh;
541 	dnode_t *dn;
542 	uint64_t nblocks;
543 	int epbs, err;
544 
545 	ASSERT(tx->tx_txg == 0);
546 
547 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
548 	    object, THT_ZAP, add, (uintptr_t)name);
549 	if (txh == NULL)
550 		return;
551 	dn = txh->txh_dnode;
552 
553 	dmu_tx_count_dnode(txh);
554 
555 	if (dn == NULL) {
556 		/*
557 		 * We will be able to fit a new object's entries into one leaf
558 		 * block.  So there will be at most 2 blocks total,
559 		 * including the header block.
560 		 */
561 		dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
562 		return;
563 	}
564 
565 	ASSERT3P(dmu_ot[dn->dn_type].ot_byteswap, ==, zap_byteswap);
566 
567 	if (dn->dn_maxblkid == 0 && !add) {
568 		/*
569 		 * If there is only one block  (i.e. this is a micro-zap)
570 		 * and we are not adding anything, the accounting is simple.
571 		 */
572 		err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
573 		if (err) {
574 			tx->tx_err = err;
575 			return;
576 		}
577 
578 		/*
579 		 * Use max block size here, since we don't know how much
580 		 * the size will change between now and the dbuf dirty call.
581 		 */
582 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
583 		    dn->dn_phys->dn_blkptr[0].blk_birth)) {
584 			txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
585 		} else {
586 			txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
587 			txh->txh_space_tounref +=
588 			    BP_GET_ASIZE(dn->dn_phys->dn_blkptr);
589 		}
590 		return;
591 	}
592 
593 	if (dn->dn_maxblkid > 0 && name) {
594 		/*
595 		 * access the name in this fat-zap so that we'll check
596 		 * for i/o errors to the leaf blocks, etc.
597 		 */
598 		err = zap_lookup(&dn->dn_objset->os, dn->dn_object, name,
599 		    8, 0, NULL);
600 		if (err == EIO) {
601 			tx->tx_err = err;
602 			return;
603 		}
604 	}
605 
606 	/*
607 	 * 3 blocks overwritten: target leaf, ptrtbl block, header block
608 	 * 3 new blocks written if adding: new split leaf, 2 grown ptrtbl blocks
609 	 */
610 	dmu_tx_count_write(txh, dn->dn_maxblkid * dn->dn_datablksz,
611 	    (3 + (add ? 3 : 0)) << dn->dn_datablkshift);
612 
613 	/*
614 	 * If the modified blocks are scattered to the four winds,
615 	 * we'll have to modify an indirect twig for each.
616 	 */
617 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
618 	for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
619 		txh->txh_space_towrite += 3 << dn->dn_indblkshift;
620 }
621 
622 void
623 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
624 {
625 	dmu_tx_hold_t *txh;
626 
627 	ASSERT(tx->tx_txg == 0);
628 
629 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
630 	    object, THT_BONUS, 0, 0);
631 	if (txh)
632 		dmu_tx_count_dnode(txh);
633 }
634 
635 void
636 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
637 {
638 	dmu_tx_hold_t *txh;
639 	ASSERT(tx->tx_txg == 0);
640 
641 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
642 	    DMU_NEW_OBJECT, THT_SPACE, space, 0);
643 
644 	txh->txh_space_towrite += space;
645 }
646 
647 int
648 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
649 {
650 	dmu_tx_hold_t *txh;
651 	int holds = 0;
652 
653 	/*
654 	 * By asserting that the tx is assigned, we're counting the
655 	 * number of dn_tx_holds, which is the same as the number of
656 	 * dn_holds.  Otherwise, we'd be counting dn_holds, but
657 	 * dn_tx_holds could be 0.
658 	 */
659 	ASSERT(tx->tx_txg != 0);
660 
661 	/* if (tx->tx_anyobj == TRUE) */
662 		/* return (0); */
663 
664 	for (txh = list_head(&tx->tx_holds); txh;
665 	    txh = list_next(&tx->tx_holds, txh)) {
666 		if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
667 			holds++;
668 	}
669 
670 	return (holds);
671 }
672 
673 #ifdef ZFS_DEBUG
674 void
675 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
676 {
677 	dmu_tx_hold_t *txh;
678 	int match_object = FALSE, match_offset = FALSE;
679 	dnode_t *dn = db->db_dnode;
680 
681 	ASSERT(tx->tx_txg != 0);
682 	ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset->os);
683 	ASSERT3U(dn->dn_object, ==, db->db.db_object);
684 
685 	if (tx->tx_anyobj)
686 		return;
687 
688 	/* XXX No checking on the meta dnode for now */
689 	if (db->db.db_object == DMU_META_DNODE_OBJECT)
690 		return;
691 
692 	for (txh = list_head(&tx->tx_holds); txh;
693 	    txh = list_next(&tx->tx_holds, txh)) {
694 		ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
695 		if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
696 			match_object = TRUE;
697 		if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
698 			int datablkshift = dn->dn_datablkshift ?
699 			    dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
700 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
701 			int shift = datablkshift + epbs * db->db_level;
702 			uint64_t beginblk = shift >= 64 ? 0 :
703 			    (txh->txh_arg1 >> shift);
704 			uint64_t endblk = shift >= 64 ? 0 :
705 			    ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
706 			uint64_t blkid = db->db_blkid;
707 
708 			/* XXX txh_arg2 better not be zero... */
709 
710 			dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
711 			    txh->txh_type, beginblk, endblk);
712 
713 			switch (txh->txh_type) {
714 			case THT_WRITE:
715 				if (blkid >= beginblk && blkid <= endblk)
716 					match_offset = TRUE;
717 				/*
718 				 * We will let this hold work for the bonus
719 				 * buffer so that we don't need to hold it
720 				 * when creating a new object.
721 				 */
722 				if (blkid == DB_BONUS_BLKID)
723 					match_offset = TRUE;
724 				/*
725 				 * They might have to increase nlevels,
726 				 * thus dirtying the new TLIBs.  Or the
727 				 * might have to change the block size,
728 				 * thus dirying the new lvl=0 blk=0.
729 				 */
730 				if (blkid == 0)
731 					match_offset = TRUE;
732 				break;
733 			case THT_FREE:
734 				/*
735 				 * We will dirty all the level 1 blocks in
736 				 * the free range and perhaps the first and
737 				 * last level 0 block.
738 				 */
739 				if (blkid >= beginblk && (blkid <= endblk ||
740 				    txh->txh_arg2 == DMU_OBJECT_END))
741 					match_offset = TRUE;
742 				break;
743 			case THT_BONUS:
744 				if (blkid == DB_BONUS_BLKID)
745 					match_offset = TRUE;
746 				break;
747 			case THT_ZAP:
748 				match_offset = TRUE;
749 				break;
750 			case THT_NEWOBJECT:
751 				match_object = TRUE;
752 				break;
753 			default:
754 				ASSERT(!"bad txh_type");
755 			}
756 		}
757 		if (match_object && match_offset)
758 			return;
759 	}
760 	panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
761 	    (u_longlong_t)db->db.db_object, db->db_level,
762 	    (u_longlong_t)db->db_blkid);
763 }
764 #endif
765 
766 static int
767 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
768 {
769 	dmu_tx_hold_t *txh;
770 	spa_t *spa = tx->tx_pool->dp_spa;
771 	uint64_t memory, asize, fsize, usize;
772 	uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
773 
774 	ASSERT3U(tx->tx_txg, ==, 0);
775 
776 	if (tx->tx_err)
777 		return (tx->tx_err);
778 
779 	if (spa_state(spa) == POOL_STATE_IO_FAILURE) {
780 		/*
781 		 * If the user has indicated a blocking failure mode
782 		 * then return ERESTART which will block in dmu_tx_wait().
783 		 * Otherwise, return EIO so that an error can get
784 		 * propagated back to the VOP calls.
785 		 *
786 		 * Note that we always honor the txg_how flag regardless
787 		 * of the failuremode setting.
788 		 */
789 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
790 		    txg_how != TXG_WAIT)
791 			return (EIO);
792 
793 		return (ERESTART);
794 	}
795 
796 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
797 	tx->tx_needassign_txh = NULL;
798 
799 	/*
800 	 * NB: No error returns are allowed after txg_hold_open, but
801 	 * before processing the dnode holds, due to the
802 	 * dmu_tx_unassign() logic.
803 	 */
804 
805 	towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
806 	for (txh = list_head(&tx->tx_holds); txh;
807 	    txh = list_next(&tx->tx_holds, txh)) {
808 		dnode_t *dn = txh->txh_dnode;
809 		if (dn != NULL) {
810 			mutex_enter(&dn->dn_mtx);
811 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
812 				mutex_exit(&dn->dn_mtx);
813 				tx->tx_needassign_txh = txh;
814 				return (ERESTART);
815 			}
816 			if (dn->dn_assigned_txg == 0)
817 				dn->dn_assigned_txg = tx->tx_txg;
818 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
819 			(void) refcount_add(&dn->dn_tx_holds, tx);
820 			mutex_exit(&dn->dn_mtx);
821 		}
822 		towrite += txh->txh_space_towrite;
823 		tofree += txh->txh_space_tofree;
824 		tooverwrite += txh->txh_space_tooverwrite;
825 		tounref += txh->txh_space_tounref;
826 		tohold += txh->txh_memory_tohold;
827 		fudge += txh->txh_fudge;
828 	}
829 
830 	/*
831 	 * NB: This check must be after we've held the dnodes, so that
832 	 * the dmu_tx_unassign() logic will work properly
833 	 */
834 	if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
835 		return (ERESTART);
836 
837 	/*
838 	 * If a snapshot has been taken since we made our estimates,
839 	 * assume that we won't be able to free or overwrite anything.
840 	 */
841 	if (tx->tx_objset &&
842 	    dsl_dataset_prev_snap_txg(tx->tx_objset->os->os_dsl_dataset) >
843 	    tx->tx_lastsnap_txg) {
844 		towrite += tooverwrite;
845 		tooverwrite = tofree = 0;
846 	}
847 
848 	/* needed allocation: worst-case estimate of write space */
849 	asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
850 	/* freed space estimate: worst-case overwrite + free estimate */
851 	fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
852 	/* convert unrefd space to worst-case estimate */
853 	usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
854 	/* calculate memory footprint estimate */
855 	memory = towrite + tooverwrite + tohold;
856 
857 #ifdef ZFS_DEBUG
858 	/*
859 	 * Add in 'tohold' to account for our dirty holds on this memory
860 	 * XXX - the "fudge" factor is to account for skipped blocks that
861 	 * we missed because dnode_next_offset() misses in-core-only blocks.
862 	 */
863 	tx->tx_space_towrite = asize +
864 	    spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
865 	tx->tx_space_tofree = tofree;
866 	tx->tx_space_tooverwrite = tooverwrite;
867 	tx->tx_space_tounref = tounref;
868 #endif
869 
870 	if (tx->tx_dir && asize != 0) {
871 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
872 		    asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
873 		if (err)
874 			return (err);
875 	}
876 
877 	return (0);
878 }
879 
880 static void
881 dmu_tx_unassign(dmu_tx_t *tx)
882 {
883 	dmu_tx_hold_t *txh;
884 
885 	if (tx->tx_txg == 0)
886 		return;
887 
888 	txg_rele_to_quiesce(&tx->tx_txgh);
889 
890 	for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
891 	    txh = list_next(&tx->tx_holds, txh)) {
892 		dnode_t *dn = txh->txh_dnode;
893 
894 		if (dn == NULL)
895 			continue;
896 		mutex_enter(&dn->dn_mtx);
897 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
898 
899 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
900 			dn->dn_assigned_txg = 0;
901 			cv_broadcast(&dn->dn_notxholds);
902 		}
903 		mutex_exit(&dn->dn_mtx);
904 	}
905 
906 	txg_rele_to_sync(&tx->tx_txgh);
907 
908 	tx->tx_lasttried_txg = tx->tx_txg;
909 	tx->tx_txg = 0;
910 }
911 
912 /*
913  * Assign tx to a transaction group.  txg_how can be one of:
914  *
915  * (1)	TXG_WAIT.  If the current open txg is full, waits until there's
916  *	a new one.  This should be used when you're not holding locks.
917  *	If will only fail if we're truly out of space (or over quota).
918  *
919  * (2)	TXG_NOWAIT.  If we can't assign into the current open txg without
920  *	blocking, returns immediately with ERESTART.  This should be used
921  *	whenever you're holding locks.  On an ERESTART error, the caller
922  *	should drop locks, do a dmu_tx_wait(tx), and try again.
923  *
924  * (3)	A specific txg.  Use this if you need to ensure that multiple
925  *	transactions all sync in the same txg.  Like TXG_NOWAIT, it
926  *	returns ERESTART if it can't assign you into the requested txg.
927  */
928 int
929 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
930 {
931 	int err;
932 
933 	ASSERT(tx->tx_txg == 0);
934 	ASSERT(txg_how != 0);
935 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
936 
937 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
938 		dmu_tx_unassign(tx);
939 
940 		if (err != ERESTART || txg_how != TXG_WAIT)
941 			return (err);
942 
943 		dmu_tx_wait(tx);
944 	}
945 
946 	txg_rele_to_quiesce(&tx->tx_txgh);
947 
948 	return (0);
949 }
950 
951 void
952 dmu_tx_wait(dmu_tx_t *tx)
953 {
954 	spa_t *spa = tx->tx_pool->dp_spa;
955 
956 	ASSERT(tx->tx_txg == 0);
957 
958 	/*
959 	 * It's possible that the pool has become active after this thread
960 	 * has tried to obtain a tx. If that's the case then his
961 	 * tx_lasttried_txg would not have been assigned.
962 	 */
963 	if (spa_state(spa) == POOL_STATE_IO_FAILURE ||
964 	    tx->tx_lasttried_txg == 0) {
965 		txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
966 	} else if (tx->tx_needassign_txh) {
967 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
968 
969 		mutex_enter(&dn->dn_mtx);
970 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
971 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
972 		mutex_exit(&dn->dn_mtx);
973 		tx->tx_needassign_txh = NULL;
974 	} else {
975 		txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
976 	}
977 }
978 
979 void
980 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
981 {
982 #ifdef ZFS_DEBUG
983 	if (tx->tx_dir == NULL || delta == 0)
984 		return;
985 
986 	if (delta > 0) {
987 		ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
988 		    tx->tx_space_towrite);
989 		(void) refcount_add_many(&tx->tx_space_written, delta, NULL);
990 	} else {
991 		(void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
992 	}
993 #endif
994 }
995 
996 void
997 dmu_tx_commit(dmu_tx_t *tx)
998 {
999 	dmu_tx_hold_t *txh;
1000 
1001 	ASSERT(tx->tx_txg != 0);
1002 
1003 	while (txh = list_head(&tx->tx_holds)) {
1004 		dnode_t *dn = txh->txh_dnode;
1005 
1006 		list_remove(&tx->tx_holds, txh);
1007 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1008 		if (dn == NULL)
1009 			continue;
1010 		mutex_enter(&dn->dn_mtx);
1011 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1012 
1013 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1014 			dn->dn_assigned_txg = 0;
1015 			cv_broadcast(&dn->dn_notxholds);
1016 		}
1017 		mutex_exit(&dn->dn_mtx);
1018 		dnode_rele(dn, tx);
1019 	}
1020 
1021 	if (tx->tx_tempreserve_cookie)
1022 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1023 
1024 	if (tx->tx_anyobj == FALSE)
1025 		txg_rele_to_sync(&tx->tx_txgh);
1026 	list_destroy(&tx->tx_holds);
1027 #ifdef ZFS_DEBUG
1028 	dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1029 	    tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1030 	    tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1031 	refcount_destroy_many(&tx->tx_space_written,
1032 	    refcount_count(&tx->tx_space_written));
1033 	refcount_destroy_many(&tx->tx_space_freed,
1034 	    refcount_count(&tx->tx_space_freed));
1035 #endif
1036 	kmem_free(tx, sizeof (dmu_tx_t));
1037 }
1038 
1039 void
1040 dmu_tx_abort(dmu_tx_t *tx)
1041 {
1042 	dmu_tx_hold_t *txh;
1043 
1044 	ASSERT(tx->tx_txg == 0);
1045 
1046 	while (txh = list_head(&tx->tx_holds)) {
1047 		dnode_t *dn = txh->txh_dnode;
1048 
1049 		list_remove(&tx->tx_holds, txh);
1050 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1051 		if (dn != NULL)
1052 			dnode_rele(dn, tx);
1053 	}
1054 	list_destroy(&tx->tx_holds);
1055 #ifdef ZFS_DEBUG
1056 	refcount_destroy_many(&tx->tx_space_written,
1057 	    refcount_count(&tx->tx_space_written));
1058 	refcount_destroy_many(&tx->tx_space_freed,
1059 	    refcount_count(&tx->tx_space_freed));
1060 #endif
1061 	kmem_free(tx, sizeof (dmu_tx_t));
1062 }
1063 
1064 uint64_t
1065 dmu_tx_get_txg(dmu_tx_t *tx)
1066 {
1067 	ASSERT(tx->tx_txg != 0);
1068 	return (tx->tx_txg);
1069 }
1070