xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_tx.c (revision 94ddd0900a8838f62bba15e270649a42f4ef9f81)
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, 2017 by Delphix. All rights reserved.
25  * Copyright (c) 2014 Integros [integros.com]
26  */
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>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_pool.h>
36 #include <sys/zap_impl.h>
37 #include <sys/spa.h>
38 #include <sys/sa.h>
39 #include <sys/sa_impl.h>
40 #include <sys/zfs_context.h>
41 #include <sys/varargs.h>
42 
43 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
44     uint64_t arg1, uint64_t arg2);
45 
46 
47 dmu_tx_t *
48 dmu_tx_create_dd(dsl_dir_t *dd)
49 {
50 	dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
51 	tx->tx_dir = dd;
52 	if (dd != NULL)
53 		tx->tx_pool = dd->dd_pool;
54 	list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
55 	    offsetof(dmu_tx_hold_t, txh_node));
56 	list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
57 	    offsetof(dmu_tx_callback_t, dcb_node));
58 	tx->tx_start = gethrtime();
59 	return (tx);
60 }
61 
62 dmu_tx_t *
63 dmu_tx_create(objset_t *os)
64 {
65 	dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
66 	tx->tx_objset = os;
67 	return (tx);
68 }
69 
70 dmu_tx_t *
71 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
72 {
73 	dmu_tx_t *tx = dmu_tx_create_dd(NULL);
74 
75 	txg_verify(dp->dp_spa, txg);
76 	tx->tx_pool = dp;
77 	tx->tx_txg = txg;
78 	tx->tx_anyobj = TRUE;
79 
80 	return (tx);
81 }
82 
83 int
84 dmu_tx_is_syncing(dmu_tx_t *tx)
85 {
86 	return (tx->tx_anyobj);
87 }
88 
89 int
90 dmu_tx_private_ok(dmu_tx_t *tx)
91 {
92 	return (tx->tx_anyobj);
93 }
94 
95 static dmu_tx_hold_t *
96 dmu_tx_hold_dnode_impl(dmu_tx_t *tx, dnode_t *dn, enum dmu_tx_hold_type type,
97     uint64_t arg1, uint64_t arg2)
98 {
99 	dmu_tx_hold_t *txh;
100 
101 	if (dn != NULL) {
102 		(void) refcount_add(&dn->dn_holds, tx);
103 		if (tx->tx_txg != 0) {
104 			mutex_enter(&dn->dn_mtx);
105 			/*
106 			 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
107 			 * problem, but there's no way for it to happen (for
108 			 * now, at least).
109 			 */
110 			ASSERT(dn->dn_assigned_txg == 0);
111 			dn->dn_assigned_txg = tx->tx_txg;
112 			(void) refcount_add(&dn->dn_tx_holds, tx);
113 			mutex_exit(&dn->dn_mtx);
114 		}
115 	}
116 
117 	txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
118 	txh->txh_tx = tx;
119 	txh->txh_dnode = dn;
120 	refcount_create(&txh->txh_space_towrite);
121 	refcount_create(&txh->txh_memory_tohold);
122 	txh->txh_type = type;
123 	txh->txh_arg1 = arg1;
124 	txh->txh_arg2 = arg2;
125 	list_insert_tail(&tx->tx_holds, txh);
126 
127 	return (txh);
128 }
129 
130 static dmu_tx_hold_t *
131 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
132     enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
133 {
134 	dnode_t *dn = NULL;
135 	dmu_tx_hold_t *txh;
136 	int err;
137 
138 	if (object != DMU_NEW_OBJECT) {
139 		err = dnode_hold(os, object, FTAG, &dn);
140 		if (err != 0) {
141 			tx->tx_err = err;
142 			return (NULL);
143 		}
144 	}
145 	txh = dmu_tx_hold_dnode_impl(tx, dn, type, arg1, arg2);
146 	if (dn != NULL)
147 		dnode_rele(dn, FTAG);
148 	return (txh);
149 }
150 
151 void
152 dmu_tx_add_new_object(dmu_tx_t *tx, dnode_t *dn)
153 {
154 	/*
155 	 * If we're syncing, they can manipulate any object anyhow, and
156 	 * the hold on the dnode_t can cause problems.
157 	 */
158 	if (!dmu_tx_is_syncing(tx))
159 		(void) dmu_tx_hold_dnode_impl(tx, dn, THT_NEWOBJECT, 0, 0);
160 }
161 
162 /*
163  * This function reads specified data from disk.  The specified data will
164  * be needed to perform the transaction -- i.e, it will be read after
165  * we do dmu_tx_assign().  There are two reasons that we read the data now
166  * (before dmu_tx_assign()):
167  *
168  * 1. Reading it now has potentially better performance.  The transaction
169  * has not yet been assigned, so the TXG is not held open, and also the
170  * caller typically has less locks held when calling dmu_tx_hold_*() than
171  * after the transaction has been assigned.  This reduces the lock (and txg)
172  * hold times, thus reducing lock contention.
173  *
174  * 2. It is easier for callers (primarily the ZPL) to handle i/o errors
175  * that are detected before they start making changes to the DMU state
176  * (i.e. now).  Once the transaction has been assigned, and some DMU
177  * state has been changed, it can be difficult to recover from an i/o
178  * error (e.g. to undo the changes already made in memory at the DMU
179  * layer).  Typically code to do so does not exist in the caller -- it
180  * assumes that the data has already been cached and thus i/o errors are
181  * not possible.
182  *
183  * It has been observed that the i/o initiated here can be a performance
184  * problem, and it appears to be optional, because we don't look at the
185  * data which is read.  However, removing this read would only serve to
186  * move the work elsewhere (after the dmu_tx_assign()), where it may
187  * have a greater impact on performance (in addition to the impact on
188  * fault tolerance noted above).
189  */
190 static int
191 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
192 {
193 	int err;
194 	dmu_buf_impl_t *db;
195 
196 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
197 	db = dbuf_hold_level(dn, level, blkid, FTAG);
198 	rw_exit(&dn->dn_struct_rwlock);
199 	if (db == NULL)
200 		return (SET_ERROR(EIO));
201 	err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
202 	dbuf_rele(db, FTAG);
203 	return (err);
204 }
205 
206 /* ARGSUSED */
207 static void
208 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
209 {
210 	dnode_t *dn = txh->txh_dnode;
211 	int err = 0;
212 
213 	if (len == 0)
214 		return;
215 
216 	(void) refcount_add_many(&txh->txh_space_towrite, len, FTAG);
217 
218 	if (refcount_count(&txh->txh_space_towrite) > 2 * DMU_MAX_ACCESS)
219 		err = SET_ERROR(EFBIG);
220 
221 	if (dn == NULL)
222 		return;
223 
224 	/*
225 	 * For i/o error checking, read the blocks that will be needed
226 	 * to perform the write: the first and last level-0 blocks (if
227 	 * they are not aligned, i.e. if they are partial-block writes),
228 	 * and all the level-1 blocks.
229 	 */
230 	if (dn->dn_maxblkid == 0) {
231 		if (off < dn->dn_datablksz &&
232 		    (off > 0 || len < dn->dn_datablksz)) {
233 			err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
234 			if (err != 0) {
235 				txh->txh_tx->tx_err = err;
236 			}
237 		}
238 	} else {
239 		zio_t *zio = zio_root(dn->dn_objset->os_spa,
240 		    NULL, NULL, ZIO_FLAG_CANFAIL);
241 
242 		/* first level-0 block */
243 		uint64_t start = off >> dn->dn_datablkshift;
244 		if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
245 			err = dmu_tx_check_ioerr(zio, dn, 0, start);
246 			if (err != 0) {
247 				txh->txh_tx->tx_err = err;
248 			}
249 		}
250 
251 		/* last level-0 block */
252 		uint64_t end = (off + len - 1) >> dn->dn_datablkshift;
253 		if (end != start && end <= dn->dn_maxblkid &&
254 		    P2PHASE(off + len, dn->dn_datablksz)) {
255 			err = dmu_tx_check_ioerr(zio, dn, 0, end);
256 			if (err != 0) {
257 				txh->txh_tx->tx_err = err;
258 			}
259 		}
260 
261 		/* level-1 blocks */
262 		if (dn->dn_nlevels > 1) {
263 			int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
264 			for (uint64_t i = (start >> shft) + 1;
265 			    i < end >> shft; i++) {
266 				err = dmu_tx_check_ioerr(zio, dn, 1, i);
267 				if (err != 0) {
268 					txh->txh_tx->tx_err = err;
269 				}
270 			}
271 		}
272 
273 		err = zio_wait(zio);
274 		if (err != 0) {
275 			txh->txh_tx->tx_err = err;
276 		}
277 	}
278 }
279 
280 static void
281 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
282 {
283 	(void) refcount_add_many(&txh->txh_space_towrite, DNODE_SIZE, FTAG);
284 }
285 
286 void
287 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
288 {
289 	dmu_tx_hold_t *txh;
290 
291 	ASSERT0(tx->tx_txg);
292 	ASSERT3U(len, <=, DMU_MAX_ACCESS);
293 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
294 
295 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
296 	    object, THT_WRITE, off, len);
297 	if (txh != NULL) {
298 		dmu_tx_count_write(txh, off, len);
299 		dmu_tx_count_dnode(txh);
300 	}
301 }
302 
303 void
304 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
305 {
306 	dmu_tx_hold_t *txh;
307 
308 	ASSERT0(tx->tx_txg);
309 	ASSERT3U(len, <=, DMU_MAX_ACCESS);
310 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
311 
312 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
313 	if (txh != NULL) {
314 		dmu_tx_count_write(txh, off, len);
315 		dmu_tx_count_dnode(txh);
316 	}
317 }
318 
319 /*
320  * This function marks the transaction as being a "net free".  The end
321  * result is that refquotas will be disabled for this transaction, and
322  * this transaction will be able to use half of the pool space overhead
323  * (see dsl_pool_adjustedsize()).  Therefore this function should only
324  * be called for transactions that we expect will not cause a net increase
325  * in the amount of space used (but it's OK if that is occasionally not true).
326  */
327 void
328 dmu_tx_mark_netfree(dmu_tx_t *tx)
329 {
330 	tx->tx_netfree = B_TRUE;
331 }
332 
333 static void
334 dmu_tx_hold_free_impl(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
335 {
336 	dmu_tx_t *tx;
337 	dnode_t *dn;
338 	int err;
339 
340 	tx = txh->txh_tx;
341 	ASSERT(tx->tx_txg == 0);
342 
343 	dn = txh->txh_dnode;
344 	dmu_tx_count_dnode(txh);
345 
346 	if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
347 		return;
348 	if (len == DMU_OBJECT_END)
349 		len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
350 
351 	/*
352 	 * For i/o error checking, we read the first and last level-0
353 	 * blocks if they are not aligned, and all the level-1 blocks.
354 	 *
355 	 * Note:  dbuf_free_range() assumes that we have not instantiated
356 	 * any level-0 dbufs that will be completely freed.  Therefore we must
357 	 * exercise care to not read or count the first and last blocks
358 	 * if they are blocksize-aligned.
359 	 */
360 	if (dn->dn_datablkshift == 0) {
361 		if (off != 0 || len < dn->dn_datablksz)
362 			dmu_tx_count_write(txh, 0, dn->dn_datablksz);
363 	} else {
364 		/* first block will be modified if it is not aligned */
365 		if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
366 			dmu_tx_count_write(txh, off, 1);
367 		/* last block will be modified if it is not aligned */
368 		if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
369 			dmu_tx_count_write(txh, off + len, 1);
370 	}
371 
372 	/*
373 	 * Check level-1 blocks.
374 	 */
375 	if (dn->dn_nlevels > 1) {
376 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
377 		    SPA_BLKPTRSHIFT;
378 		uint64_t start = off >> shift;
379 		uint64_t end = (off + len) >> shift;
380 
381 		ASSERT(dn->dn_indblkshift != 0);
382 
383 		/*
384 		 * dnode_reallocate() can result in an object with indirect
385 		 * blocks having an odd data block size.  In this case,
386 		 * just check the single block.
387 		 */
388 		if (dn->dn_datablkshift == 0)
389 			start = end = 0;
390 
391 		zio_t *zio = zio_root(tx->tx_pool->dp_spa,
392 		    NULL, NULL, ZIO_FLAG_CANFAIL);
393 		for (uint64_t i = start; i <= end; i++) {
394 			uint64_t ibyte = i << shift;
395 			err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
396 			i = ibyte >> shift;
397 			if (err == ESRCH || i > end)
398 				break;
399 			if (err != 0) {
400 				tx->tx_err = err;
401 				(void) zio_wait(zio);
402 				return;
403 			}
404 
405 			(void) refcount_add_many(&txh->txh_memory_tohold,
406 			    1 << dn->dn_indblkshift, FTAG);
407 
408 			err = dmu_tx_check_ioerr(zio, dn, 1, i);
409 			if (err != 0) {
410 				tx->tx_err = err;
411 				(void) zio_wait(zio);
412 				return;
413 			}
414 		}
415 		err = zio_wait(zio);
416 		if (err != 0) {
417 			tx->tx_err = err;
418 			return;
419 		}
420 	}
421 }
422 
423 void
424 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
425 {
426 	dmu_tx_hold_t *txh;
427 
428 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
429 	    object, THT_FREE, off, len);
430 	if (txh != NULL)
431 		(void) dmu_tx_hold_free_impl(txh, off, len);
432 }
433 
434 void
435 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
436 {
437 	dmu_tx_hold_t *txh;
438 
439 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
440 	if (txh != NULL)
441 		(void) dmu_tx_hold_free_impl(txh, off, len);
442 }
443 
444 static void
445 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
446 {
447 	dmu_tx_t *tx = txh->txh_tx;
448 	dnode_t *dn;
449 	int err;
450 
451 	ASSERT(tx->tx_txg == 0);
452 
453 	dn = txh->txh_dnode;
454 
455 	dmu_tx_count_dnode(txh);
456 
457 	/*
458 	 * Modifying a almost-full microzap is around the worst case (128KB)
459 	 *
460 	 * If it is a fat zap, the worst case would be 7*16KB=112KB:
461 	 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
462 	 * - 4 new blocks written if adding:
463 	 *    - 2 blocks for possibly split leaves,
464 	 *    - 2 grown ptrtbl blocks
465 	 */
466 	(void) refcount_add_many(&txh->txh_space_towrite,
467 	    MZAP_MAX_BLKSZ, FTAG);
468 
469 	if (dn == NULL)
470 		return;
471 
472 	ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
473 
474 	if (dn->dn_maxblkid == 0 || name == NULL) {
475 		/*
476 		 * This is a microzap (only one block), or we don't know
477 		 * the name.  Check the first block for i/o errors.
478 		 */
479 		err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
480 		if (err != 0) {
481 			tx->tx_err = err;
482 		}
483 	} else {
484 		/*
485 		 * Access the name so that we'll check for i/o errors to
486 		 * the leaf blocks, etc.  We ignore ENOENT, as this name
487 		 * may not yet exist.
488 		 */
489 		err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
490 		if (err == EIO || err == ECKSUM || err == ENXIO) {
491 			tx->tx_err = err;
492 		}
493 	}
494 }
495 
496 void
497 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
498 {
499 	dmu_tx_hold_t *txh;
500 
501 	ASSERT0(tx->tx_txg);
502 
503 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
504 	    object, THT_ZAP, add, (uintptr_t)name);
505 	if (txh != NULL)
506 		dmu_tx_hold_zap_impl(txh, name);
507 }
508 
509 void
510 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
511 {
512 	dmu_tx_hold_t *txh;
513 
514 	ASSERT0(tx->tx_txg);
515 	ASSERT(dn != NULL);
516 
517 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
518 	if (txh != NULL)
519 		dmu_tx_hold_zap_impl(txh, name);
520 }
521 
522 void
523 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
524 {
525 	dmu_tx_hold_t *txh;
526 
527 	ASSERT(tx->tx_txg == 0);
528 
529 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
530 	    object, THT_BONUS, 0, 0);
531 	if (txh)
532 		dmu_tx_count_dnode(txh);
533 }
534 
535 void
536 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
537 {
538 	dmu_tx_hold_t *txh;
539 
540 	ASSERT0(tx->tx_txg);
541 
542 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
543 	if (txh)
544 		dmu_tx_count_dnode(txh);
545 }
546 
547 void
548 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
549 {
550 	dmu_tx_hold_t *txh;
551 	ASSERT(tx->tx_txg == 0);
552 
553 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
554 	    DMU_NEW_OBJECT, THT_SPACE, space, 0);
555 
556 	(void) refcount_add_many(&txh->txh_space_towrite, space, FTAG);
557 }
558 
559 #ifdef ZFS_DEBUG
560 void
561 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
562 {
563 	boolean_t match_object = B_FALSE;
564 	boolean_t match_offset = B_FALSE;
565 
566 	DB_DNODE_ENTER(db);
567 	dnode_t *dn = DB_DNODE(db);
568 	ASSERT(tx->tx_txg != 0);
569 	ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
570 	ASSERT3U(dn->dn_object, ==, db->db.db_object);
571 
572 	if (tx->tx_anyobj) {
573 		DB_DNODE_EXIT(db);
574 		return;
575 	}
576 
577 	/* XXX No checking on the meta dnode for now */
578 	if (db->db.db_object == DMU_META_DNODE_OBJECT) {
579 		DB_DNODE_EXIT(db);
580 		return;
581 	}
582 
583 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
584 	    txh = list_next(&tx->tx_holds, txh)) {
585 		ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
586 		if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
587 			match_object = TRUE;
588 		if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
589 			int datablkshift = dn->dn_datablkshift ?
590 			    dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
591 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
592 			int shift = datablkshift + epbs * db->db_level;
593 			uint64_t beginblk = shift >= 64 ? 0 :
594 			    (txh->txh_arg1 >> shift);
595 			uint64_t endblk = shift >= 64 ? 0 :
596 			    ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
597 			uint64_t blkid = db->db_blkid;
598 
599 			/* XXX txh_arg2 better not be zero... */
600 
601 			dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
602 			    txh->txh_type, beginblk, endblk);
603 
604 			switch (txh->txh_type) {
605 			case THT_WRITE:
606 				if (blkid >= beginblk && blkid <= endblk)
607 					match_offset = TRUE;
608 				/*
609 				 * We will let this hold work for the bonus
610 				 * or spill buffer so that we don't need to
611 				 * hold it when creating a new object.
612 				 */
613 				if (blkid == DMU_BONUS_BLKID ||
614 				    blkid == DMU_SPILL_BLKID)
615 					match_offset = TRUE;
616 				/*
617 				 * They might have to increase nlevels,
618 				 * thus dirtying the new TLIBs.  Or the
619 				 * might have to change the block size,
620 				 * thus dirying the new lvl=0 blk=0.
621 				 */
622 				if (blkid == 0)
623 					match_offset = TRUE;
624 				break;
625 			case THT_FREE:
626 				/*
627 				 * We will dirty all the level 1 blocks in
628 				 * the free range and perhaps the first and
629 				 * last level 0 block.
630 				 */
631 				if (blkid >= beginblk && (blkid <= endblk ||
632 				    txh->txh_arg2 == DMU_OBJECT_END))
633 					match_offset = TRUE;
634 				break;
635 			case THT_SPILL:
636 				if (blkid == DMU_SPILL_BLKID)
637 					match_offset = TRUE;
638 				break;
639 			case THT_BONUS:
640 				if (blkid == DMU_BONUS_BLKID)
641 					match_offset = TRUE;
642 				break;
643 			case THT_ZAP:
644 				match_offset = TRUE;
645 				break;
646 			case THT_NEWOBJECT:
647 				match_object = TRUE;
648 				break;
649 			default:
650 				ASSERT(!"bad txh_type");
651 			}
652 		}
653 		if (match_object && match_offset) {
654 			DB_DNODE_EXIT(db);
655 			return;
656 		}
657 	}
658 	DB_DNODE_EXIT(db);
659 	panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
660 	    (u_longlong_t)db->db.db_object, db->db_level,
661 	    (u_longlong_t)db->db_blkid);
662 }
663 #endif
664 
665 /*
666  * If we can't do 10 iops, something is wrong.  Let us go ahead
667  * and hit zfs_dirty_data_max.
668  */
669 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
670 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
671 
672 /*
673  * We delay transactions when we've determined that the backend storage
674  * isn't able to accommodate the rate of incoming writes.
675  *
676  * If there is already a transaction waiting, we delay relative to when
677  * that transaction finishes waiting.  This way the calculated min_time
678  * is independent of the number of threads concurrently executing
679  * transactions.
680  *
681  * If we are the only waiter, wait relative to when the transaction
682  * started, rather than the current time.  This credits the transaction for
683  * "time already served", e.g. reading indirect blocks.
684  *
685  * The minimum time for a transaction to take is calculated as:
686  *     min_time = scale * (dirty - min) / (max - dirty)
687  *     min_time is then capped at zfs_delay_max_ns.
688  *
689  * The delay has two degrees of freedom that can be adjusted via tunables.
690  * The percentage of dirty data at which we start to delay is defined by
691  * zfs_delay_min_dirty_percent. This should typically be at or above
692  * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
693  * delay after writing at full speed has failed to keep up with the incoming
694  * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
695  * speaking, this variable determines the amount of delay at the midpoint of
696  * the curve.
697  *
698  * delay
699  *  10ms +-------------------------------------------------------------*+
700  *       |                                                             *|
701  *   9ms +                                                             *+
702  *       |                                                             *|
703  *   8ms +                                                             *+
704  *       |                                                            * |
705  *   7ms +                                                            * +
706  *       |                                                            * |
707  *   6ms +                                                            * +
708  *       |                                                            * |
709  *   5ms +                                                           *  +
710  *       |                                                           *  |
711  *   4ms +                                                           *  +
712  *       |                                                           *  |
713  *   3ms +                                                          *   +
714  *       |                                                          *   |
715  *   2ms +                                              (midpoint) *    +
716  *       |                                                  |    **     |
717  *   1ms +                                                  v ***       +
718  *       |             zfs_delay_scale ---------->     ********         |
719  *     0 +-------------------------------------*********----------------+
720  *       0%                    <- zfs_dirty_data_max ->               100%
721  *
722  * Note that since the delay is added to the outstanding time remaining on the
723  * most recent transaction, the delay is effectively the inverse of IOPS.
724  * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
725  * was chosen such that small changes in the amount of accumulated dirty data
726  * in the first 3/4 of the curve yield relatively small differences in the
727  * amount of delay.
728  *
729  * The effects can be easier to understand when the amount of delay is
730  * represented on a log scale:
731  *
732  * delay
733  * 100ms +-------------------------------------------------------------++
734  *       +                                                              +
735  *       |                                                              |
736  *       +                                                             *+
737  *  10ms +                                                             *+
738  *       +                                                           ** +
739  *       |                                              (midpoint)  **  |
740  *       +                                                  |     **    +
741  *   1ms +                                                  v ****      +
742  *       +             zfs_delay_scale ---------->        *****         +
743  *       |                                             ****             |
744  *       +                                          ****                +
745  * 100us +                                        **                    +
746  *       +                                       *                      +
747  *       |                                      *                       |
748  *       +                                     *                        +
749  *  10us +                                     *                        +
750  *       +                                                              +
751  *       |                                                              |
752  *       +                                                              +
753  *       +--------------------------------------------------------------+
754  *       0%                    <- zfs_dirty_data_max ->               100%
755  *
756  * Note here that only as the amount of dirty data approaches its limit does
757  * the delay start to increase rapidly. The goal of a properly tuned system
758  * should be to keep the amount of dirty data out of that range by first
759  * ensuring that the appropriate limits are set for the I/O scheduler to reach
760  * optimal throughput on the backend storage, and then by changing the value
761  * of zfs_delay_scale to increase the steepness of the curve.
762  */
763 static void
764 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
765 {
766 	dsl_pool_t *dp = tx->tx_pool;
767 	uint64_t delay_min_bytes =
768 	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
769 	hrtime_t wakeup, min_tx_time, now;
770 
771 	if (dirty <= delay_min_bytes)
772 		return;
773 
774 	/*
775 	 * The caller has already waited until we are under the max.
776 	 * We make them pass us the amount of dirty data so we don't
777 	 * have to handle the case of it being >= the max, which could
778 	 * cause a divide-by-zero if it's == the max.
779 	 */
780 	ASSERT3U(dirty, <, zfs_dirty_data_max);
781 
782 	now = gethrtime();
783 	min_tx_time = zfs_delay_scale *
784 	    (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
785 	if (now > tx->tx_start + min_tx_time)
786 		return;
787 
788 	min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
789 
790 	DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
791 	    uint64_t, min_tx_time);
792 
793 	mutex_enter(&dp->dp_lock);
794 	wakeup = MAX(tx->tx_start + min_tx_time,
795 	    dp->dp_last_wakeup + min_tx_time);
796 	dp->dp_last_wakeup = wakeup;
797 	mutex_exit(&dp->dp_lock);
798 
799 #ifdef _KERNEL
800 	mutex_enter(&curthread->t_delay_lock);
801 	while (cv_timedwait_hires(&curthread->t_delay_cv,
802 	    &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
803 	    CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
804 		continue;
805 	mutex_exit(&curthread->t_delay_lock);
806 #else
807 	hrtime_t delta = wakeup - gethrtime();
808 	struct timespec ts;
809 	ts.tv_sec = delta / NANOSEC;
810 	ts.tv_nsec = delta % NANOSEC;
811 	(void) nanosleep(&ts, NULL);
812 #endif
813 }
814 
815 /*
816  * This routine attempts to assign the transaction to a transaction group.
817  * To do so, we must determine if there is sufficient free space on disk.
818  *
819  * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
820  * on it), then it is assumed that there is sufficient free space,
821  * unless there's insufficient slop space in the pool (see the comment
822  * above spa_slop_shift in spa_misc.c).
823  *
824  * If it is not a "netfree" transaction, then if the data already on disk
825  * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
826  * ENOSPC.  Otherwise, if the current rough estimate of pending changes,
827  * plus the rough estimate of this transaction's changes, may exceed the
828  * allowed usage, then this will fail with ERESTART, which will cause the
829  * caller to wait for the pending changes to be written to disk (by waiting
830  * for the next TXG to open), and then check the space usage again.
831  *
832  * The rough estimate of pending changes is comprised of the sum of:
833  *
834  *  - this transaction's holds' txh_space_towrite
835  *
836  *  - dd_tempreserved[], which is the sum of in-flight transactions'
837  *    holds' txh_space_towrite (i.e. those transactions that have called
838  *    dmu_tx_assign() but not yet called dmu_tx_commit()).
839  *
840  *  - dd_space_towrite[], which is the amount of dirtied dbufs.
841  *
842  * Note that all of these values are inflated by spa_get_worst_case_asize(),
843  * which means that we may get ERESTART well before we are actually in danger
844  * of running out of space, but this also mitigates any small inaccuracies
845  * in the rough estimate (e.g. txh_space_towrite doesn't take into account
846  * indirect blocks, and dd_space_towrite[] doesn't take into account changes
847  * to the MOS).
848  *
849  * Note that due to this algorithm, it is possible to exceed the allowed
850  * usage by one transaction.  Also, as we approach the allowed usage,
851  * we will allow a very limited amount of changes into each TXG, thus
852  * decreasing performance.
853  */
854 static int
855 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
856 {
857 	spa_t *spa = tx->tx_pool->dp_spa;
858 
859 	ASSERT0(tx->tx_txg);
860 
861 	if (tx->tx_err)
862 		return (tx->tx_err);
863 
864 	if (spa_suspended(spa)) {
865 		/*
866 		 * If the user has indicated a blocking failure mode
867 		 * then return ERESTART which will block in dmu_tx_wait().
868 		 * Otherwise, return EIO so that an error can get
869 		 * propagated back to the VOP calls.
870 		 *
871 		 * Note that we always honor the txg_how flag regardless
872 		 * of the failuremode setting.
873 		 */
874 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
875 		    txg_how != TXG_WAIT)
876 			return (SET_ERROR(EIO));
877 
878 		return (SET_ERROR(ERESTART));
879 	}
880 
881 	if (!tx->tx_waited &&
882 	    dsl_pool_need_dirty_delay(tx->tx_pool)) {
883 		tx->tx_wait_dirty = B_TRUE;
884 		return (SET_ERROR(ERESTART));
885 	}
886 
887 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
888 	tx->tx_needassign_txh = NULL;
889 
890 	/*
891 	 * NB: No error returns are allowed after txg_hold_open, but
892 	 * before processing the dnode holds, due to the
893 	 * dmu_tx_unassign() logic.
894 	 */
895 
896 	uint64_t towrite = 0;
897 	uint64_t tohold = 0;
898 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
899 	    txh = list_next(&tx->tx_holds, txh)) {
900 		dnode_t *dn = txh->txh_dnode;
901 		if (dn != NULL) {
902 			mutex_enter(&dn->dn_mtx);
903 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
904 				mutex_exit(&dn->dn_mtx);
905 				tx->tx_needassign_txh = txh;
906 				return (SET_ERROR(ERESTART));
907 			}
908 			if (dn->dn_assigned_txg == 0)
909 				dn->dn_assigned_txg = tx->tx_txg;
910 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
911 			(void) refcount_add(&dn->dn_tx_holds, tx);
912 			mutex_exit(&dn->dn_mtx);
913 		}
914 		towrite += refcount_count(&txh->txh_space_towrite);
915 		tohold += refcount_count(&txh->txh_memory_tohold);
916 	}
917 
918 	/* needed allocation: worst-case estimate of write space */
919 	uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite);
920 	/* calculate memory footprint estimate */
921 	uint64_t memory = towrite + tohold;
922 
923 	if (tx->tx_dir != NULL && asize != 0) {
924 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
925 		    asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx);
926 		if (err != 0)
927 			return (err);
928 	}
929 
930 	return (0);
931 }
932 
933 static void
934 dmu_tx_unassign(dmu_tx_t *tx)
935 {
936 	if (tx->tx_txg == 0)
937 		return;
938 
939 	txg_rele_to_quiesce(&tx->tx_txgh);
940 
941 	/*
942 	 * Walk the transaction's hold list, removing the hold on the
943 	 * associated dnode, and notifying waiters if the refcount drops to 0.
944 	 */
945 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
946 	    txh != tx->tx_needassign_txh;
947 	    txh = list_next(&tx->tx_holds, txh)) {
948 		dnode_t *dn = txh->txh_dnode;
949 
950 		if (dn == NULL)
951 			continue;
952 		mutex_enter(&dn->dn_mtx);
953 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
954 
955 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
956 			dn->dn_assigned_txg = 0;
957 			cv_broadcast(&dn->dn_notxholds);
958 		}
959 		mutex_exit(&dn->dn_mtx);
960 	}
961 
962 	txg_rele_to_sync(&tx->tx_txgh);
963 
964 	tx->tx_lasttried_txg = tx->tx_txg;
965 	tx->tx_txg = 0;
966 }
967 
968 /*
969  * Assign tx to a transaction group.  txg_how can be one of:
970  *
971  * (1)	TXG_WAIT.  If the current open txg is full, waits until there's
972  *	a new one.  This should be used when you're not holding locks.
973  *	It will only fail if we're truly out of space (or over quota).
974  *
975  * (2)	TXG_NOWAIT.  If we can't assign into the current open txg without
976  *	blocking, returns immediately with ERESTART.  This should be used
977  *	whenever you're holding locks.  On an ERESTART error, the caller
978  *	should drop locks, do a dmu_tx_wait(tx), and try again.
979  *
980  * (3)  TXG_WAITED.  Like TXG_NOWAIT, but indicates that dmu_tx_wait()
981  *      has already been called on behalf of this operation (though
982  *      most likely on a different tx).
983  */
984 int
985 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
986 {
987 	int err;
988 
989 	ASSERT(tx->tx_txg == 0);
990 	ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
991 	    txg_how == TXG_WAITED);
992 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
993 
994 	/* If we might wait, we must not hold the config lock. */
995 	ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
996 
997 	if (txg_how == TXG_WAITED)
998 		tx->tx_waited = B_TRUE;
999 
1000 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1001 		dmu_tx_unassign(tx);
1002 
1003 		if (err != ERESTART || txg_how != TXG_WAIT)
1004 			return (err);
1005 
1006 		dmu_tx_wait(tx);
1007 	}
1008 
1009 	txg_rele_to_quiesce(&tx->tx_txgh);
1010 
1011 	return (0);
1012 }
1013 
1014 void
1015 dmu_tx_wait(dmu_tx_t *tx)
1016 {
1017 	spa_t *spa = tx->tx_pool->dp_spa;
1018 	dsl_pool_t *dp = tx->tx_pool;
1019 
1020 	ASSERT(tx->tx_txg == 0);
1021 	ASSERT(!dsl_pool_config_held(tx->tx_pool));
1022 
1023 	if (tx->tx_wait_dirty) {
1024 		/*
1025 		 * dmu_tx_try_assign() has determined that we need to wait
1026 		 * because we've consumed much or all of the dirty buffer
1027 		 * space.
1028 		 */
1029 		mutex_enter(&dp->dp_lock);
1030 		while (dp->dp_dirty_total >= zfs_dirty_data_max)
1031 			cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1032 		uint64_t dirty = dp->dp_dirty_total;
1033 		mutex_exit(&dp->dp_lock);
1034 
1035 		dmu_tx_delay(tx, dirty);
1036 
1037 		tx->tx_wait_dirty = B_FALSE;
1038 
1039 		/*
1040 		 * Note: setting tx_waited only has effect if the caller
1041 		 * used TX_WAIT.  Otherwise they are going to destroy
1042 		 * this tx and try again.  The common case, zfs_write(),
1043 		 * uses TX_WAIT.
1044 		 */
1045 		tx->tx_waited = B_TRUE;
1046 	} else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1047 		/*
1048 		 * If the pool is suspended we need to wait until it
1049 		 * is resumed.  Note that it's possible that the pool
1050 		 * has become active after this thread has tried to
1051 		 * obtain a tx.  If that's the case then tx_lasttried_txg
1052 		 * would not have been set.
1053 		 */
1054 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1055 	} else if (tx->tx_needassign_txh) {
1056 		/*
1057 		 * A dnode is assigned to the quiescing txg.  Wait for its
1058 		 * transaction to complete.
1059 		 */
1060 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1061 
1062 		mutex_enter(&dn->dn_mtx);
1063 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1064 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1065 		mutex_exit(&dn->dn_mtx);
1066 		tx->tx_needassign_txh = NULL;
1067 	} else {
1068 		txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1069 	}
1070 }
1071 
1072 static void
1073 dmu_tx_destroy(dmu_tx_t *tx)
1074 {
1075 	dmu_tx_hold_t *txh;
1076 
1077 	while ((txh = list_head(&tx->tx_holds)) != NULL) {
1078 		dnode_t *dn = txh->txh_dnode;
1079 
1080 		list_remove(&tx->tx_holds, txh);
1081 		refcount_destroy_many(&txh->txh_space_towrite,
1082 		    refcount_count(&txh->txh_space_towrite));
1083 		refcount_destroy_many(&txh->txh_memory_tohold,
1084 		    refcount_count(&txh->txh_memory_tohold));
1085 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1086 		if (dn != NULL)
1087 			dnode_rele(dn, tx);
1088 	}
1089 
1090 	list_destroy(&tx->tx_callbacks);
1091 	list_destroy(&tx->tx_holds);
1092 	kmem_free(tx, sizeof (dmu_tx_t));
1093 }
1094 
1095 void
1096 dmu_tx_commit(dmu_tx_t *tx)
1097 {
1098 	ASSERT(tx->tx_txg != 0);
1099 
1100 	/*
1101 	 * Go through the transaction's hold list and remove holds on
1102 	 * associated dnodes, notifying waiters if no holds remain.
1103 	 */
1104 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1105 	    txh = list_next(&tx->tx_holds, txh)) {
1106 		dnode_t *dn = txh->txh_dnode;
1107 
1108 		if (dn == NULL)
1109 			continue;
1110 
1111 		mutex_enter(&dn->dn_mtx);
1112 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1113 
1114 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1115 			dn->dn_assigned_txg = 0;
1116 			cv_broadcast(&dn->dn_notxholds);
1117 		}
1118 		mutex_exit(&dn->dn_mtx);
1119 	}
1120 
1121 	if (tx->tx_tempreserve_cookie)
1122 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1123 
1124 	if (!list_is_empty(&tx->tx_callbacks))
1125 		txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1126 
1127 	if (tx->tx_anyobj == FALSE)
1128 		txg_rele_to_sync(&tx->tx_txgh);
1129 
1130 	dmu_tx_destroy(tx);
1131 }
1132 
1133 void
1134 dmu_tx_abort(dmu_tx_t *tx)
1135 {
1136 	ASSERT(tx->tx_txg == 0);
1137 
1138 	/*
1139 	 * Call any registered callbacks with an error code.
1140 	 */
1141 	if (!list_is_empty(&tx->tx_callbacks))
1142 		dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1143 
1144 	dmu_tx_destroy(tx);
1145 }
1146 
1147 uint64_t
1148 dmu_tx_get_txg(dmu_tx_t *tx)
1149 {
1150 	ASSERT(tx->tx_txg != 0);
1151 	return (tx->tx_txg);
1152 }
1153 
1154 dsl_pool_t *
1155 dmu_tx_pool(dmu_tx_t *tx)
1156 {
1157 	ASSERT(tx->tx_pool != NULL);
1158 	return (tx->tx_pool);
1159 }
1160 
1161 void
1162 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1163 {
1164 	dmu_tx_callback_t *dcb;
1165 
1166 	dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1167 
1168 	dcb->dcb_func = func;
1169 	dcb->dcb_data = data;
1170 
1171 	list_insert_tail(&tx->tx_callbacks, dcb);
1172 }
1173 
1174 /*
1175  * Call all the commit callbacks on a list, with a given error code.
1176  */
1177 void
1178 dmu_tx_do_callbacks(list_t *cb_list, int error)
1179 {
1180 	dmu_tx_callback_t *dcb;
1181 
1182 	while ((dcb = list_head(cb_list)) != NULL) {
1183 		list_remove(cb_list, dcb);
1184 		dcb->dcb_func(dcb->dcb_data, error);
1185 		kmem_free(dcb, sizeof (dmu_tx_callback_t));
1186 	}
1187 }
1188 
1189 /*
1190  * Interface to hold a bunch of attributes.
1191  * used for creating new files.
1192  * attrsize is the total size of all attributes
1193  * to be added during object creation
1194  *
1195  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1196  */
1197 
1198 /*
1199  * hold necessary attribute name for attribute registration.
1200  * should be a very rare case where this is needed.  If it does
1201  * happen it would only happen on the first write to the file system.
1202  */
1203 static void
1204 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1205 {
1206 	if (!sa->sa_need_attr_registration)
1207 		return;
1208 
1209 	for (int i = 0; i != sa->sa_num_attrs; i++) {
1210 		if (!sa->sa_attr_table[i].sa_registered) {
1211 			if (sa->sa_reg_attr_obj)
1212 				dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1213 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1214 			else
1215 				dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1216 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1217 		}
1218 	}
1219 }
1220 
1221 void
1222 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1223 {
1224 	dmu_tx_hold_t *txh = dmu_tx_hold_object_impl(tx,
1225 	    tx->tx_objset, object, THT_SPILL, 0, 0);
1226 
1227 	(void) refcount_add_many(&txh->txh_space_towrite,
1228 	    SPA_OLD_MAXBLOCKSIZE, FTAG);
1229 }
1230 
1231 void
1232 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1233 {
1234 	sa_os_t *sa = tx->tx_objset->os_sa;
1235 
1236 	dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1237 
1238 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1239 		return;
1240 
1241 	if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1242 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1243 	} else {
1244 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1245 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1246 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1247 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1248 	}
1249 
1250 	dmu_tx_sa_registration_hold(sa, tx);
1251 
1252 	if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1253 		return;
1254 
1255 	(void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1256 	    THT_SPILL, 0, 0);
1257 }
1258 
1259 /*
1260  * Hold SA attribute
1261  *
1262  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1263  *
1264  * variable_size is the total size of all variable sized attributes
1265  * passed to this function.  It is not the total size of all
1266  * variable size attributes that *may* exist on this object.
1267  */
1268 void
1269 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1270 {
1271 	uint64_t object;
1272 	sa_os_t *sa = tx->tx_objset->os_sa;
1273 
1274 	ASSERT(hdl != NULL);
1275 
1276 	object = sa_handle_object(hdl);
1277 
1278 	dmu_tx_hold_bonus(tx, object);
1279 
1280 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1281 		return;
1282 
1283 	if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1284 	    tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1285 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1286 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1287 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1288 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1289 	}
1290 
1291 	dmu_tx_sa_registration_hold(sa, tx);
1292 
1293 	if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1294 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1295 
1296 	if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1297 		ASSERT(tx->tx_txg == 0);
1298 		dmu_tx_hold_spill(tx, object);
1299 	} else {
1300 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1301 		dnode_t *dn;
1302 
1303 		DB_DNODE_ENTER(db);
1304 		dn = DB_DNODE(db);
1305 		if (dn->dn_have_spill) {
1306 			ASSERT(tx->tx_txg == 0);
1307 			dmu_tx_hold_spill(tx, object);
1308 		}
1309 		DB_DNODE_EXIT(db);
1310 	}
1311 }
1312