xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_tx.c (revision fd6d41c5025e9fb45a115fc82d86e9983d1e9fd6)
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_remap_l1indirect(dmu_tx_t *tx, uint64_t object)
305 {
306 	dmu_tx_hold_t *txh;
307 
308 	ASSERT(tx->tx_txg == 0);
309 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
310 	    object, THT_WRITE, 0, 0);
311 	if (txh == NULL)
312 		return;
313 
314 	dnode_t *dn = txh->txh_dnode;
315 	(void) refcount_add_many(&txh->txh_space_towrite,
316 	    1ULL << dn->dn_indblkshift, FTAG);
317 	dmu_tx_count_dnode(txh);
318 }
319 
320 void
321 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
322 {
323 	dmu_tx_hold_t *txh;
324 
325 	ASSERT0(tx->tx_txg);
326 	ASSERT3U(len, <=, DMU_MAX_ACCESS);
327 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
328 
329 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
330 	if (txh != NULL) {
331 		dmu_tx_count_write(txh, off, len);
332 		dmu_tx_count_dnode(txh);
333 	}
334 }
335 
336 /*
337  * This function marks the transaction as being a "net free".  The end
338  * result is that refquotas will be disabled for this transaction, and
339  * this transaction will be able to use half of the pool space overhead
340  * (see dsl_pool_adjustedsize()).  Therefore this function should only
341  * be called for transactions that we expect will not cause a net increase
342  * in the amount of space used (but it's OK if that is occasionally not true).
343  */
344 void
345 dmu_tx_mark_netfree(dmu_tx_t *tx)
346 {
347 	tx->tx_netfree = B_TRUE;
348 }
349 
350 static void
351 dmu_tx_hold_free_impl(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
352 {
353 	dmu_tx_t *tx;
354 	dnode_t *dn;
355 	int err;
356 
357 	tx = txh->txh_tx;
358 	ASSERT(tx->tx_txg == 0);
359 
360 	dn = txh->txh_dnode;
361 	dmu_tx_count_dnode(txh);
362 
363 	if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
364 		return;
365 	if (len == DMU_OBJECT_END)
366 		len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
367 
368 	/*
369 	 * For i/o error checking, we read the first and last level-0
370 	 * blocks if they are not aligned, and all the level-1 blocks.
371 	 *
372 	 * Note:  dbuf_free_range() assumes that we have not instantiated
373 	 * any level-0 dbufs that will be completely freed.  Therefore we must
374 	 * exercise care to not read or count the first and last blocks
375 	 * if they are blocksize-aligned.
376 	 */
377 	if (dn->dn_datablkshift == 0) {
378 		if (off != 0 || len < dn->dn_datablksz)
379 			dmu_tx_count_write(txh, 0, dn->dn_datablksz);
380 	} else {
381 		/* first block will be modified if it is not aligned */
382 		if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
383 			dmu_tx_count_write(txh, off, 1);
384 		/* last block will be modified if it is not aligned */
385 		if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
386 			dmu_tx_count_write(txh, off + len, 1);
387 	}
388 
389 	/*
390 	 * Check level-1 blocks.
391 	 */
392 	if (dn->dn_nlevels > 1) {
393 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
394 		    SPA_BLKPTRSHIFT;
395 		uint64_t start = off >> shift;
396 		uint64_t end = (off + len) >> shift;
397 
398 		ASSERT(dn->dn_indblkshift != 0);
399 
400 		/*
401 		 * dnode_reallocate() can result in an object with indirect
402 		 * blocks having an odd data block size.  In this case,
403 		 * just check the single block.
404 		 */
405 		if (dn->dn_datablkshift == 0)
406 			start = end = 0;
407 
408 		zio_t *zio = zio_root(tx->tx_pool->dp_spa,
409 		    NULL, NULL, ZIO_FLAG_CANFAIL);
410 		for (uint64_t i = start; i <= end; i++) {
411 			uint64_t ibyte = i << shift;
412 			err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
413 			i = ibyte >> shift;
414 			if (err == ESRCH || i > end)
415 				break;
416 			if (err != 0) {
417 				tx->tx_err = err;
418 				(void) zio_wait(zio);
419 				return;
420 			}
421 
422 			(void) refcount_add_many(&txh->txh_memory_tohold,
423 			    1 << dn->dn_indblkshift, FTAG);
424 
425 			err = dmu_tx_check_ioerr(zio, dn, 1, i);
426 			if (err != 0) {
427 				tx->tx_err = err;
428 				(void) zio_wait(zio);
429 				return;
430 			}
431 		}
432 		err = zio_wait(zio);
433 		if (err != 0) {
434 			tx->tx_err = err;
435 			return;
436 		}
437 	}
438 }
439 
440 void
441 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
442 {
443 	dmu_tx_hold_t *txh;
444 
445 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
446 	    object, THT_FREE, off, len);
447 	if (txh != NULL)
448 		(void) dmu_tx_hold_free_impl(txh, off, len);
449 }
450 
451 void
452 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
453 {
454 	dmu_tx_hold_t *txh;
455 
456 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
457 	if (txh != NULL)
458 		(void) dmu_tx_hold_free_impl(txh, off, len);
459 }
460 
461 static void
462 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
463 {
464 	dmu_tx_t *tx = txh->txh_tx;
465 	dnode_t *dn;
466 	int err;
467 
468 	ASSERT(tx->tx_txg == 0);
469 
470 	dn = txh->txh_dnode;
471 
472 	dmu_tx_count_dnode(txh);
473 
474 	/*
475 	 * Modifying a almost-full microzap is around the worst case (128KB)
476 	 *
477 	 * If it is a fat zap, the worst case would be 7*16KB=112KB:
478 	 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
479 	 * - 4 new blocks written if adding:
480 	 *    - 2 blocks for possibly split leaves,
481 	 *    - 2 grown ptrtbl blocks
482 	 */
483 	(void) refcount_add_many(&txh->txh_space_towrite,
484 	    MZAP_MAX_BLKSZ, FTAG);
485 
486 	if (dn == NULL)
487 		return;
488 
489 	ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
490 
491 	if (dn->dn_maxblkid == 0 || name == NULL) {
492 		/*
493 		 * This is a microzap (only one block), or we don't know
494 		 * the name.  Check the first block for i/o errors.
495 		 */
496 		err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
497 		if (err != 0) {
498 			tx->tx_err = err;
499 		}
500 	} else {
501 		/*
502 		 * Access the name so that we'll check for i/o errors to
503 		 * the leaf blocks, etc.  We ignore ENOENT, as this name
504 		 * may not yet exist.
505 		 */
506 		err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
507 		if (err == EIO || err == ECKSUM || err == ENXIO) {
508 			tx->tx_err = err;
509 		}
510 	}
511 }
512 
513 void
514 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
515 {
516 	dmu_tx_hold_t *txh;
517 
518 	ASSERT0(tx->tx_txg);
519 
520 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
521 	    object, THT_ZAP, add, (uintptr_t)name);
522 	if (txh != NULL)
523 		dmu_tx_hold_zap_impl(txh, name);
524 }
525 
526 void
527 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
528 {
529 	dmu_tx_hold_t *txh;
530 
531 	ASSERT0(tx->tx_txg);
532 	ASSERT(dn != NULL);
533 
534 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
535 	if (txh != NULL)
536 		dmu_tx_hold_zap_impl(txh, name);
537 }
538 
539 void
540 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
541 {
542 	dmu_tx_hold_t *txh;
543 
544 	ASSERT(tx->tx_txg == 0);
545 
546 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
547 	    object, THT_BONUS, 0, 0);
548 	if (txh)
549 		dmu_tx_count_dnode(txh);
550 }
551 
552 void
553 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
554 {
555 	dmu_tx_hold_t *txh;
556 
557 	ASSERT0(tx->tx_txg);
558 
559 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
560 	if (txh)
561 		dmu_tx_count_dnode(txh);
562 }
563 
564 void
565 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
566 {
567 	dmu_tx_hold_t *txh;
568 	ASSERT(tx->tx_txg == 0);
569 
570 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
571 	    DMU_NEW_OBJECT, THT_SPACE, space, 0);
572 
573 	(void) refcount_add_many(&txh->txh_space_towrite, space, FTAG);
574 }
575 
576 #ifdef ZFS_DEBUG
577 void
578 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
579 {
580 	boolean_t match_object = B_FALSE;
581 	boolean_t match_offset = B_FALSE;
582 
583 	DB_DNODE_ENTER(db);
584 	dnode_t *dn = DB_DNODE(db);
585 	ASSERT(tx->tx_txg != 0);
586 	ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
587 	ASSERT3U(dn->dn_object, ==, db->db.db_object);
588 
589 	if (tx->tx_anyobj) {
590 		DB_DNODE_EXIT(db);
591 		return;
592 	}
593 
594 	/* XXX No checking on the meta dnode for now */
595 	if (db->db.db_object == DMU_META_DNODE_OBJECT) {
596 		DB_DNODE_EXIT(db);
597 		return;
598 	}
599 
600 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
601 	    txh = list_next(&tx->tx_holds, txh)) {
602 		ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
603 		if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
604 			match_object = TRUE;
605 		if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
606 			int datablkshift = dn->dn_datablkshift ?
607 			    dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
608 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
609 			int shift = datablkshift + epbs * db->db_level;
610 			uint64_t beginblk = shift >= 64 ? 0 :
611 			    (txh->txh_arg1 >> shift);
612 			uint64_t endblk = shift >= 64 ? 0 :
613 			    ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
614 			uint64_t blkid = db->db_blkid;
615 
616 			/* XXX txh_arg2 better not be zero... */
617 
618 			dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
619 			    txh->txh_type, beginblk, endblk);
620 
621 			switch (txh->txh_type) {
622 			case THT_WRITE:
623 				if (blkid >= beginblk && blkid <= endblk)
624 					match_offset = TRUE;
625 				/*
626 				 * We will let this hold work for the bonus
627 				 * or spill buffer so that we don't need to
628 				 * hold it when creating a new object.
629 				 */
630 				if (blkid == DMU_BONUS_BLKID ||
631 				    blkid == DMU_SPILL_BLKID)
632 					match_offset = TRUE;
633 				/*
634 				 * They might have to increase nlevels,
635 				 * thus dirtying the new TLIBs.  Or the
636 				 * might have to change the block size,
637 				 * thus dirying the new lvl=0 blk=0.
638 				 */
639 				if (blkid == 0)
640 					match_offset = TRUE;
641 				break;
642 			case THT_FREE:
643 				/*
644 				 * We will dirty all the level 1 blocks in
645 				 * the free range and perhaps the first and
646 				 * last level 0 block.
647 				 */
648 				if (blkid >= beginblk && (blkid <= endblk ||
649 				    txh->txh_arg2 == DMU_OBJECT_END))
650 					match_offset = TRUE;
651 				break;
652 			case THT_SPILL:
653 				if (blkid == DMU_SPILL_BLKID)
654 					match_offset = TRUE;
655 				break;
656 			case THT_BONUS:
657 				if (blkid == DMU_BONUS_BLKID)
658 					match_offset = TRUE;
659 				break;
660 			case THT_ZAP:
661 				match_offset = TRUE;
662 				break;
663 			case THT_NEWOBJECT:
664 				match_object = TRUE;
665 				break;
666 			default:
667 				ASSERT(!"bad txh_type");
668 			}
669 		}
670 		if (match_object && match_offset) {
671 			DB_DNODE_EXIT(db);
672 			return;
673 		}
674 	}
675 	DB_DNODE_EXIT(db);
676 	panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
677 	    (u_longlong_t)db->db.db_object, db->db_level,
678 	    (u_longlong_t)db->db_blkid);
679 }
680 #endif
681 
682 /*
683  * If we can't do 10 iops, something is wrong.  Let us go ahead
684  * and hit zfs_dirty_data_max.
685  */
686 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
687 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
688 
689 /*
690  * We delay transactions when we've determined that the backend storage
691  * isn't able to accommodate the rate of incoming writes.
692  *
693  * If there is already a transaction waiting, we delay relative to when
694  * that transaction finishes waiting.  This way the calculated min_time
695  * is independent of the number of threads concurrently executing
696  * transactions.
697  *
698  * If we are the only waiter, wait relative to when the transaction
699  * started, rather than the current time.  This credits the transaction for
700  * "time already served", e.g. reading indirect blocks.
701  *
702  * The minimum time for a transaction to take is calculated as:
703  *     min_time = scale * (dirty - min) / (max - dirty)
704  *     min_time is then capped at zfs_delay_max_ns.
705  *
706  * The delay has two degrees of freedom that can be adjusted via tunables.
707  * The percentage of dirty data at which we start to delay is defined by
708  * zfs_delay_min_dirty_percent. This should typically be at or above
709  * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
710  * delay after writing at full speed has failed to keep up with the incoming
711  * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
712  * speaking, this variable determines the amount of delay at the midpoint of
713  * the curve.
714  *
715  * delay
716  *  10ms +-------------------------------------------------------------*+
717  *       |                                                             *|
718  *   9ms +                                                             *+
719  *       |                                                             *|
720  *   8ms +                                                             *+
721  *       |                                                            * |
722  *   7ms +                                                            * +
723  *       |                                                            * |
724  *   6ms +                                                            * +
725  *       |                                                            * |
726  *   5ms +                                                           *  +
727  *       |                                                           *  |
728  *   4ms +                                                           *  +
729  *       |                                                           *  |
730  *   3ms +                                                          *   +
731  *       |                                                          *   |
732  *   2ms +                                              (midpoint) *    +
733  *       |                                                  |    **     |
734  *   1ms +                                                  v ***       +
735  *       |             zfs_delay_scale ---------->     ********         |
736  *     0 +-------------------------------------*********----------------+
737  *       0%                    <- zfs_dirty_data_max ->               100%
738  *
739  * Note that since the delay is added to the outstanding time remaining on the
740  * most recent transaction, the delay is effectively the inverse of IOPS.
741  * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
742  * was chosen such that small changes in the amount of accumulated dirty data
743  * in the first 3/4 of the curve yield relatively small differences in the
744  * amount of delay.
745  *
746  * The effects can be easier to understand when the amount of delay is
747  * represented on a log scale:
748  *
749  * delay
750  * 100ms +-------------------------------------------------------------++
751  *       +                                                              +
752  *       |                                                              |
753  *       +                                                             *+
754  *  10ms +                                                             *+
755  *       +                                                           ** +
756  *       |                                              (midpoint)  **  |
757  *       +                                                  |     **    +
758  *   1ms +                                                  v ****      +
759  *       +             zfs_delay_scale ---------->        *****         +
760  *       |                                             ****             |
761  *       +                                          ****                +
762  * 100us +                                        **                    +
763  *       +                                       *                      +
764  *       |                                      *                       |
765  *       +                                     *                        +
766  *  10us +                                     *                        +
767  *       +                                                              +
768  *       |                                                              |
769  *       +                                                              +
770  *       +--------------------------------------------------------------+
771  *       0%                    <- zfs_dirty_data_max ->               100%
772  *
773  * Note here that only as the amount of dirty data approaches its limit does
774  * the delay start to increase rapidly. The goal of a properly tuned system
775  * should be to keep the amount of dirty data out of that range by first
776  * ensuring that the appropriate limits are set for the I/O scheduler to reach
777  * optimal throughput on the backend storage, and then by changing the value
778  * of zfs_delay_scale to increase the steepness of the curve.
779  */
780 static void
781 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
782 {
783 	dsl_pool_t *dp = tx->tx_pool;
784 	uint64_t delay_min_bytes =
785 	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
786 	hrtime_t wakeup, min_tx_time, now;
787 
788 	if (dirty <= delay_min_bytes)
789 		return;
790 
791 	/*
792 	 * The caller has already waited until we are under the max.
793 	 * We make them pass us the amount of dirty data so we don't
794 	 * have to handle the case of it being >= the max, which could
795 	 * cause a divide-by-zero if it's == the max.
796 	 */
797 	ASSERT3U(dirty, <, zfs_dirty_data_max);
798 
799 	now = gethrtime();
800 	min_tx_time = zfs_delay_scale *
801 	    (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
802 	if (now > tx->tx_start + min_tx_time)
803 		return;
804 
805 	min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
806 
807 	DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
808 	    uint64_t, min_tx_time);
809 
810 	mutex_enter(&dp->dp_lock);
811 	wakeup = MAX(tx->tx_start + min_tx_time,
812 	    dp->dp_last_wakeup + min_tx_time);
813 	dp->dp_last_wakeup = wakeup;
814 	mutex_exit(&dp->dp_lock);
815 
816 #ifdef _KERNEL
817 	mutex_enter(&curthread->t_delay_lock);
818 	while (cv_timedwait_hires(&curthread->t_delay_cv,
819 	    &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
820 	    CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
821 		continue;
822 	mutex_exit(&curthread->t_delay_lock);
823 #else
824 	hrtime_t delta = wakeup - gethrtime();
825 	struct timespec ts;
826 	ts.tv_sec = delta / NANOSEC;
827 	ts.tv_nsec = delta % NANOSEC;
828 	(void) nanosleep(&ts, NULL);
829 #endif
830 }
831 
832 /*
833  * This routine attempts to assign the transaction to a transaction group.
834  * To do so, we must determine if there is sufficient free space on disk.
835  *
836  * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
837  * on it), then it is assumed that there is sufficient free space,
838  * unless there's insufficient slop space in the pool (see the comment
839  * above spa_slop_shift in spa_misc.c).
840  *
841  * If it is not a "netfree" transaction, then if the data already on disk
842  * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
843  * ENOSPC.  Otherwise, if the current rough estimate of pending changes,
844  * plus the rough estimate of this transaction's changes, may exceed the
845  * allowed usage, then this will fail with ERESTART, which will cause the
846  * caller to wait for the pending changes to be written to disk (by waiting
847  * for the next TXG to open), and then check the space usage again.
848  *
849  * The rough estimate of pending changes is comprised of the sum of:
850  *
851  *  - this transaction's holds' txh_space_towrite
852  *
853  *  - dd_tempreserved[], which is the sum of in-flight transactions'
854  *    holds' txh_space_towrite (i.e. those transactions that have called
855  *    dmu_tx_assign() but not yet called dmu_tx_commit()).
856  *
857  *  - dd_space_towrite[], which is the amount of dirtied dbufs.
858  *
859  * Note that all of these values are inflated by spa_get_worst_case_asize(),
860  * which means that we may get ERESTART well before we are actually in danger
861  * of running out of space, but this also mitigates any small inaccuracies
862  * in the rough estimate (e.g. txh_space_towrite doesn't take into account
863  * indirect blocks, and dd_space_towrite[] doesn't take into account changes
864  * to the MOS).
865  *
866  * Note that due to this algorithm, it is possible to exceed the allowed
867  * usage by one transaction.  Also, as we approach the allowed usage,
868  * we will allow a very limited amount of changes into each TXG, thus
869  * decreasing performance.
870  */
871 static int
872 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
873 {
874 	spa_t *spa = tx->tx_pool->dp_spa;
875 
876 	ASSERT0(tx->tx_txg);
877 
878 	if (tx->tx_err)
879 		return (tx->tx_err);
880 
881 	if (spa_suspended(spa)) {
882 		/*
883 		 * If the user has indicated a blocking failure mode
884 		 * then return ERESTART which will block in dmu_tx_wait().
885 		 * Otherwise, return EIO so that an error can get
886 		 * propagated back to the VOP calls.
887 		 *
888 		 * Note that we always honor the txg_how flag regardless
889 		 * of the failuremode setting.
890 		 */
891 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
892 		    !(txg_how & TXG_WAIT))
893 			return (SET_ERROR(EIO));
894 
895 		return (SET_ERROR(ERESTART));
896 	}
897 
898 	if (!tx->tx_dirty_delayed &&
899 	    dsl_pool_need_dirty_delay(tx->tx_pool)) {
900 		tx->tx_wait_dirty = B_TRUE;
901 		return (SET_ERROR(ERESTART));
902 	}
903 
904 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
905 	tx->tx_needassign_txh = NULL;
906 
907 	/*
908 	 * NB: No error returns are allowed after txg_hold_open, but
909 	 * before processing the dnode holds, due to the
910 	 * dmu_tx_unassign() logic.
911 	 */
912 
913 	uint64_t towrite = 0;
914 	uint64_t tohold = 0;
915 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
916 	    txh = list_next(&tx->tx_holds, txh)) {
917 		dnode_t *dn = txh->txh_dnode;
918 		if (dn != NULL) {
919 			mutex_enter(&dn->dn_mtx);
920 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
921 				mutex_exit(&dn->dn_mtx);
922 				tx->tx_needassign_txh = txh;
923 				return (SET_ERROR(ERESTART));
924 			}
925 			if (dn->dn_assigned_txg == 0)
926 				dn->dn_assigned_txg = tx->tx_txg;
927 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
928 			(void) refcount_add(&dn->dn_tx_holds, tx);
929 			mutex_exit(&dn->dn_mtx);
930 		}
931 		towrite += refcount_count(&txh->txh_space_towrite);
932 		tohold += refcount_count(&txh->txh_memory_tohold);
933 	}
934 
935 	/* needed allocation: worst-case estimate of write space */
936 	uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite);
937 	/* calculate memory footprint estimate */
938 	uint64_t memory = towrite + tohold;
939 
940 	if (tx->tx_dir != NULL && asize != 0) {
941 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
942 		    asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx);
943 		if (err != 0)
944 			return (err);
945 	}
946 
947 	return (0);
948 }
949 
950 static void
951 dmu_tx_unassign(dmu_tx_t *tx)
952 {
953 	if (tx->tx_txg == 0)
954 		return;
955 
956 	txg_rele_to_quiesce(&tx->tx_txgh);
957 
958 	/*
959 	 * Walk the transaction's hold list, removing the hold on the
960 	 * associated dnode, and notifying waiters if the refcount drops to 0.
961 	 */
962 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
963 	    txh != tx->tx_needassign_txh;
964 	    txh = list_next(&tx->tx_holds, txh)) {
965 		dnode_t *dn = txh->txh_dnode;
966 
967 		if (dn == NULL)
968 			continue;
969 		mutex_enter(&dn->dn_mtx);
970 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
971 
972 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
973 			dn->dn_assigned_txg = 0;
974 			cv_broadcast(&dn->dn_notxholds);
975 		}
976 		mutex_exit(&dn->dn_mtx);
977 	}
978 
979 	txg_rele_to_sync(&tx->tx_txgh);
980 
981 	tx->tx_lasttried_txg = tx->tx_txg;
982 	tx->tx_txg = 0;
983 }
984 
985 /*
986  * Assign tx to a transaction group; txg_how is a bitmask:
987  *
988  * If TXG_WAIT is set and the currently open txg is full, this function
989  * will wait until there's a new txg. This should be used when no locks
990  * are being held. With this bit set, this function will only fail if
991  * we're truly out of space (or over quota).
992  *
993  * If TXG_WAIT is *not* set and we can't assign into the currently open
994  * txg without blocking, this function will return immediately with
995  * ERESTART. This should be used whenever locks are being held.  On an
996  * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
997  * and try again.
998  *
999  * If TXG_NOTHROTTLE is set, this indicates that this tx should not be
1000  * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
1001  * details on the throttle). This is used by the VFS operations, after
1002  * they have already called dmu_tx_wait() (though most likely on a
1003  * different tx).
1004  */
1005 int
1006 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1007 {
1008 	int err;
1009 
1010 	ASSERT(tx->tx_txg == 0);
1011 	ASSERT0(txg_how & ~(TXG_WAIT | TXG_NOTHROTTLE));
1012 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1013 
1014 	/* If we might wait, we must not hold the config lock. */
1015 	IMPLY((txg_how & TXG_WAIT), !dsl_pool_config_held(tx->tx_pool));
1016 
1017 	if ((txg_how & TXG_NOTHROTTLE))
1018 		tx->tx_dirty_delayed = B_TRUE;
1019 
1020 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1021 		dmu_tx_unassign(tx);
1022 
1023 		if (err != ERESTART || !(txg_how & TXG_WAIT))
1024 			return (err);
1025 
1026 		dmu_tx_wait(tx);
1027 	}
1028 
1029 	txg_rele_to_quiesce(&tx->tx_txgh);
1030 
1031 	return (0);
1032 }
1033 
1034 void
1035 dmu_tx_wait(dmu_tx_t *tx)
1036 {
1037 	spa_t *spa = tx->tx_pool->dp_spa;
1038 	dsl_pool_t *dp = tx->tx_pool;
1039 
1040 	ASSERT(tx->tx_txg == 0);
1041 	ASSERT(!dsl_pool_config_held(tx->tx_pool));
1042 
1043 	if (tx->tx_wait_dirty) {
1044 		/*
1045 		 * dmu_tx_try_assign() has determined that we need to wait
1046 		 * because we've consumed much or all of the dirty buffer
1047 		 * space.
1048 		 */
1049 		mutex_enter(&dp->dp_lock);
1050 		while (dp->dp_dirty_total >= zfs_dirty_data_max)
1051 			cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1052 		uint64_t dirty = dp->dp_dirty_total;
1053 		mutex_exit(&dp->dp_lock);
1054 
1055 		dmu_tx_delay(tx, dirty);
1056 
1057 		tx->tx_wait_dirty = B_FALSE;
1058 
1059 		/*
1060 		 * Note: setting tx_dirty_delayed only has effect if the
1061 		 * caller used TX_WAIT.  Otherwise they are going to
1062 		 * destroy this tx and try again.  The common case,
1063 		 * zfs_write(), uses TX_WAIT.
1064 		 */
1065 		tx->tx_dirty_delayed = B_TRUE;
1066 	} else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1067 		/*
1068 		 * If the pool is suspended we need to wait until it
1069 		 * is resumed.  Note that it's possible that the pool
1070 		 * has become active after this thread has tried to
1071 		 * obtain a tx.  If that's the case then tx_lasttried_txg
1072 		 * would not have been set.
1073 		 */
1074 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1075 	} else if (tx->tx_needassign_txh) {
1076 		/*
1077 		 * A dnode is assigned to the quiescing txg.  Wait for its
1078 		 * transaction to complete.
1079 		 */
1080 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1081 
1082 		mutex_enter(&dn->dn_mtx);
1083 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1084 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1085 		mutex_exit(&dn->dn_mtx);
1086 		tx->tx_needassign_txh = NULL;
1087 	} else {
1088 		/*
1089 		 * If we have a lot of dirty data just wait until we sync
1090 		 * out a TXG at which point we'll hopefully have synced
1091 		 * a portion of the changes.
1092 		 */
1093 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1094 	}
1095 }
1096 
1097 static void
1098 dmu_tx_destroy(dmu_tx_t *tx)
1099 {
1100 	dmu_tx_hold_t *txh;
1101 
1102 	while ((txh = list_head(&tx->tx_holds)) != NULL) {
1103 		dnode_t *dn = txh->txh_dnode;
1104 
1105 		list_remove(&tx->tx_holds, txh);
1106 		refcount_destroy_many(&txh->txh_space_towrite,
1107 		    refcount_count(&txh->txh_space_towrite));
1108 		refcount_destroy_many(&txh->txh_memory_tohold,
1109 		    refcount_count(&txh->txh_memory_tohold));
1110 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1111 		if (dn != NULL)
1112 			dnode_rele(dn, tx);
1113 	}
1114 
1115 	list_destroy(&tx->tx_callbacks);
1116 	list_destroy(&tx->tx_holds);
1117 	kmem_free(tx, sizeof (dmu_tx_t));
1118 }
1119 
1120 void
1121 dmu_tx_commit(dmu_tx_t *tx)
1122 {
1123 	ASSERT(tx->tx_txg != 0);
1124 
1125 	/*
1126 	 * Go through the transaction's hold list and remove holds on
1127 	 * associated dnodes, notifying waiters if no holds remain.
1128 	 */
1129 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1130 	    txh = list_next(&tx->tx_holds, txh)) {
1131 		dnode_t *dn = txh->txh_dnode;
1132 
1133 		if (dn == NULL)
1134 			continue;
1135 
1136 		mutex_enter(&dn->dn_mtx);
1137 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1138 
1139 		if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1140 			dn->dn_assigned_txg = 0;
1141 			cv_broadcast(&dn->dn_notxholds);
1142 		}
1143 		mutex_exit(&dn->dn_mtx);
1144 	}
1145 
1146 	if (tx->tx_tempreserve_cookie)
1147 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1148 
1149 	if (!list_is_empty(&tx->tx_callbacks))
1150 		txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1151 
1152 	if (tx->tx_anyobj == FALSE)
1153 		txg_rele_to_sync(&tx->tx_txgh);
1154 
1155 	dmu_tx_destroy(tx);
1156 }
1157 
1158 void
1159 dmu_tx_abort(dmu_tx_t *tx)
1160 {
1161 	ASSERT(tx->tx_txg == 0);
1162 
1163 	/*
1164 	 * Call any registered callbacks with an error code.
1165 	 */
1166 	if (!list_is_empty(&tx->tx_callbacks))
1167 		dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1168 
1169 	dmu_tx_destroy(tx);
1170 }
1171 
1172 uint64_t
1173 dmu_tx_get_txg(dmu_tx_t *tx)
1174 {
1175 	ASSERT(tx->tx_txg != 0);
1176 	return (tx->tx_txg);
1177 }
1178 
1179 dsl_pool_t *
1180 dmu_tx_pool(dmu_tx_t *tx)
1181 {
1182 	ASSERT(tx->tx_pool != NULL);
1183 	return (tx->tx_pool);
1184 }
1185 
1186 void
1187 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1188 {
1189 	dmu_tx_callback_t *dcb;
1190 
1191 	dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1192 
1193 	dcb->dcb_func = func;
1194 	dcb->dcb_data = data;
1195 
1196 	list_insert_tail(&tx->tx_callbacks, dcb);
1197 }
1198 
1199 /*
1200  * Call all the commit callbacks on a list, with a given error code.
1201  */
1202 void
1203 dmu_tx_do_callbacks(list_t *cb_list, int error)
1204 {
1205 	dmu_tx_callback_t *dcb;
1206 
1207 	while ((dcb = list_head(cb_list)) != NULL) {
1208 		list_remove(cb_list, dcb);
1209 		dcb->dcb_func(dcb->dcb_data, error);
1210 		kmem_free(dcb, sizeof (dmu_tx_callback_t));
1211 	}
1212 }
1213 
1214 /*
1215  * Interface to hold a bunch of attributes.
1216  * used for creating new files.
1217  * attrsize is the total size of all attributes
1218  * to be added during object creation
1219  *
1220  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1221  */
1222 
1223 /*
1224  * hold necessary attribute name for attribute registration.
1225  * should be a very rare case where this is needed.  If it does
1226  * happen it would only happen on the first write to the file system.
1227  */
1228 static void
1229 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1230 {
1231 	if (!sa->sa_need_attr_registration)
1232 		return;
1233 
1234 	for (int i = 0; i != sa->sa_num_attrs; i++) {
1235 		if (!sa->sa_attr_table[i].sa_registered) {
1236 			if (sa->sa_reg_attr_obj)
1237 				dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1238 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1239 			else
1240 				dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1241 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1242 		}
1243 	}
1244 }
1245 
1246 void
1247 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1248 {
1249 	dmu_tx_hold_t *txh = dmu_tx_hold_object_impl(tx,
1250 	    tx->tx_objset, object, THT_SPILL, 0, 0);
1251 
1252 	(void) refcount_add_many(&txh->txh_space_towrite,
1253 	    SPA_OLD_MAXBLOCKSIZE, FTAG);
1254 }
1255 
1256 void
1257 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1258 {
1259 	sa_os_t *sa = tx->tx_objset->os_sa;
1260 
1261 	dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1262 
1263 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1264 		return;
1265 
1266 	if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1267 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1268 	} else {
1269 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1270 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1271 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1272 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1273 	}
1274 
1275 	dmu_tx_sa_registration_hold(sa, tx);
1276 
1277 	if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1278 		return;
1279 
1280 	(void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1281 	    THT_SPILL, 0, 0);
1282 }
1283 
1284 /*
1285  * Hold SA attribute
1286  *
1287  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1288  *
1289  * variable_size is the total size of all variable sized attributes
1290  * passed to this function.  It is not the total size of all
1291  * variable size attributes that *may* exist on this object.
1292  */
1293 void
1294 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1295 {
1296 	uint64_t object;
1297 	sa_os_t *sa = tx->tx_objset->os_sa;
1298 
1299 	ASSERT(hdl != NULL);
1300 
1301 	object = sa_handle_object(hdl);
1302 
1303 	dmu_tx_hold_bonus(tx, object);
1304 
1305 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1306 		return;
1307 
1308 	if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1309 	    tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1310 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1311 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1312 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1313 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1314 	}
1315 
1316 	dmu_tx_sa_registration_hold(sa, tx);
1317 
1318 	if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1319 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1320 
1321 	if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1322 		ASSERT(tx->tx_txg == 0);
1323 		dmu_tx_hold_spill(tx, object);
1324 	} else {
1325 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1326 		dnode_t *dn;
1327 
1328 		DB_DNODE_ENTER(db);
1329 		dn = DB_DNODE(db);
1330 		if (dn->dn_have_spill) {
1331 			ASSERT(tx->tx_txg == 0);
1332 			dmu_tx_hold_spill(tx, object);
1333 		}
1334 		DB_DNODE_EXIT(db);
1335 	}
1336 }
1337