xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_tx.c (revision e914ace2e9d9bf2dbf9a1f1ce81cb776022096f5)
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 *
dmu_tx_create_dd(dsl_dir_t * dd)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 *
dmu_tx_create(objset_t * os)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 *
dmu_tx_create_assigned(struct dsl_pool * dp,uint64_t txg)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
dmu_tx_is_syncing(dmu_tx_t * tx)84 dmu_tx_is_syncing(dmu_tx_t *tx)
85 {
86 	return (tx->tx_anyobj);
87 }
88 
89 int
dmu_tx_private_ok(dmu_tx_t * tx)90 dmu_tx_private_ok(dmu_tx_t *tx)
91 {
92 	return (tx->tx_anyobj);
93 }
94 
95 static dmu_tx_hold_t *
dmu_tx_hold_dnode_impl(dmu_tx_t * tx,dnode_t * dn,enum dmu_tx_hold_type type,uint64_t arg1,uint64_t arg2)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) zfs_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) zfs_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 	zfs_refcount_create(&txh->txh_space_towrite);
121 	zfs_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 *
dmu_tx_hold_object_impl(dmu_tx_t * tx,objset_t * os,uint64_t object,enum dmu_tx_hold_type type,uint64_t arg1,uint64_t arg2)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
dmu_tx_add_new_object(dmu_tx_t * tx,dnode_t * dn)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
dmu_tx_check_ioerr(zio_t * zio,dnode_t * dn,int level,uint64_t blkid)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
dmu_tx_count_write(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)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) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
217 
218 	if (zfs_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
dmu_tx_count_dnode(dmu_tx_hold_t * txh)281 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
282 {
283 	(void) zfs_refcount_add_many(&txh->txh_space_towrite, DNODE_MIN_SIZE,
284 	    FTAG);
285 }
286 
287 void
dmu_tx_hold_write(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)288 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
289 {
290 	dmu_tx_hold_t *txh;
291 
292 	ASSERT0(tx->tx_txg);
293 	ASSERT3U(len, <=, DMU_MAX_ACCESS);
294 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
295 
296 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
297 	    object, THT_WRITE, off, len);
298 	if (txh != NULL) {
299 		dmu_tx_count_write(txh, off, len);
300 		dmu_tx_count_dnode(txh);
301 	}
302 }
303 
304 void
dmu_tx_hold_remap_l1indirect(dmu_tx_t * tx,uint64_t object)305 dmu_tx_hold_remap_l1indirect(dmu_tx_t *tx, uint64_t object)
306 {
307 	dmu_tx_hold_t *txh;
308 
309 	ASSERT(tx->tx_txg == 0);
310 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
311 	    object, THT_WRITE, 0, 0);
312 	if (txh == NULL)
313 		return;
314 
315 	dnode_t *dn = txh->txh_dnode;
316 	(void) zfs_refcount_add_many(&txh->txh_space_towrite,
317 	    1ULL << dn->dn_indblkshift, FTAG);
318 	dmu_tx_count_dnode(txh);
319 }
320 
321 void
dmu_tx_hold_write_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)322 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
323 {
324 	dmu_tx_hold_t *txh;
325 
326 	ASSERT0(tx->tx_txg);
327 	ASSERT3U(len, <=, DMU_MAX_ACCESS);
328 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
329 
330 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
331 	if (txh != NULL) {
332 		dmu_tx_count_write(txh, off, len);
333 		dmu_tx_count_dnode(txh);
334 	}
335 }
336 
337 /*
338  * This function marks the transaction as being a "net free".  The end
339  * result is that refquotas will be disabled for this transaction, and
340  * this transaction will be able to use half of the pool space overhead
341  * (see dsl_pool_adjustedsize()).  Therefore this function should only
342  * be called for transactions that we expect will not cause a net increase
343  * in the amount of space used (but it's OK if that is occasionally not true).
344  */
345 void
dmu_tx_mark_netfree(dmu_tx_t * tx)346 dmu_tx_mark_netfree(dmu_tx_t *tx)
347 {
348 	tx->tx_netfree = B_TRUE;
349 }
350 
351 static void
dmu_tx_hold_free_impl(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)352 dmu_tx_hold_free_impl(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
353 {
354 	dmu_tx_t *tx;
355 	dnode_t *dn;
356 	int err;
357 
358 	tx = txh->txh_tx;
359 	ASSERT(tx->tx_txg == 0);
360 
361 	dn = txh->txh_dnode;
362 	dmu_tx_count_dnode(txh);
363 
364 	if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
365 		return;
366 	if (len == DMU_OBJECT_END)
367 		len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
368 
369 	/*
370 	 * For i/o error checking, we read the first and last level-0
371 	 * blocks if they are not aligned, and all the level-1 blocks.
372 	 *
373 	 * Note:  dbuf_free_range() assumes that we have not instantiated
374 	 * any level-0 dbufs that will be completely freed.  Therefore we must
375 	 * exercise care to not read or count the first and last blocks
376 	 * if they are blocksize-aligned.
377 	 */
378 	if (dn->dn_datablkshift == 0) {
379 		if (off != 0 || len < dn->dn_datablksz)
380 			dmu_tx_count_write(txh, 0, dn->dn_datablksz);
381 	} else {
382 		/* first block will be modified if it is not aligned */
383 		if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
384 			dmu_tx_count_write(txh, off, 1);
385 		/* last block will be modified if it is not aligned */
386 		if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
387 			dmu_tx_count_write(txh, off + len, 1);
388 	}
389 
390 	/*
391 	 * Check level-1 blocks.
392 	 */
393 	if (dn->dn_nlevels > 1) {
394 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
395 		    SPA_BLKPTRSHIFT;
396 		uint64_t start = off >> shift;
397 		uint64_t end = (off + len) >> shift;
398 
399 		ASSERT(dn->dn_indblkshift != 0);
400 
401 		/*
402 		 * dnode_reallocate() can result in an object with indirect
403 		 * blocks having an odd data block size.  In this case,
404 		 * just check the single block.
405 		 */
406 		if (dn->dn_datablkshift == 0)
407 			start = end = 0;
408 
409 		zio_t *zio = zio_root(tx->tx_pool->dp_spa,
410 		    NULL, NULL, ZIO_FLAG_CANFAIL);
411 		for (uint64_t i = start; i <= end; i++) {
412 			uint64_t ibyte = i << shift;
413 			err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
414 			i = ibyte >> shift;
415 			if (err == ESRCH || i > end)
416 				break;
417 			if (err != 0) {
418 				tx->tx_err = err;
419 				(void) zio_wait(zio);
420 				return;
421 			}
422 
423 			(void) zfs_refcount_add_many(&txh->txh_memory_tohold,
424 			    1 << dn->dn_indblkshift, FTAG);
425 
426 			err = dmu_tx_check_ioerr(zio, dn, 1, i);
427 			if (err != 0) {
428 				tx->tx_err = err;
429 				(void) zio_wait(zio);
430 				return;
431 			}
432 		}
433 		err = zio_wait(zio);
434 		if (err != 0) {
435 			tx->tx_err = err;
436 			return;
437 		}
438 	}
439 }
440 
441 void
dmu_tx_hold_free(dmu_tx_t * tx,uint64_t object,uint64_t off,uint64_t len)442 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
443 {
444 	dmu_tx_hold_t *txh;
445 
446 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
447 	    object, THT_FREE, off, len);
448 	if (txh != NULL)
449 		(void) dmu_tx_hold_free_impl(txh, off, len);
450 }
451 
452 void
dmu_tx_hold_free_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,uint64_t len)453 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
454 {
455 	dmu_tx_hold_t *txh;
456 
457 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
458 	if (txh != NULL)
459 		(void) dmu_tx_hold_free_impl(txh, off, len);
460 }
461 
462 static void
dmu_tx_hold_zap_impl(dmu_tx_hold_t * txh,const char * name)463 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
464 {
465 	dmu_tx_t *tx = txh->txh_tx;
466 	dnode_t *dn;
467 	int err;
468 
469 	ASSERT(tx->tx_txg == 0);
470 
471 	dn = txh->txh_dnode;
472 
473 	dmu_tx_count_dnode(txh);
474 
475 	/*
476 	 * Modifying a almost-full microzap is around the worst case (128KB)
477 	 *
478 	 * If it is a fat zap, the worst case would be 7*16KB=112KB:
479 	 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
480 	 * - 4 new blocks written if adding:
481 	 *    - 2 blocks for possibly split leaves,
482 	 *    - 2 grown ptrtbl blocks
483 	 */
484 	(void) zfs_refcount_add_many(&txh->txh_space_towrite,
485 	    MZAP_MAX_BLKSZ, FTAG);
486 
487 	if (dn == NULL)
488 		return;
489 
490 	ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
491 
492 	if (dn->dn_maxblkid == 0 || name == NULL) {
493 		/*
494 		 * This is a microzap (only one block), or we don't know
495 		 * the name.  Check the first block for i/o errors.
496 		 */
497 		err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
498 		if (err != 0) {
499 			tx->tx_err = err;
500 		}
501 	} else {
502 		/*
503 		 * Access the name so that we'll check for i/o errors to
504 		 * the leaf blocks, etc.  We ignore ENOENT, as this name
505 		 * may not yet exist.
506 		 */
507 		err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
508 		if (err == EIO || err == ECKSUM || err == ENXIO) {
509 			tx->tx_err = err;
510 		}
511 	}
512 }
513 
514 void
dmu_tx_hold_zap(dmu_tx_t * tx,uint64_t object,int add,const char * name)515 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
516 {
517 	dmu_tx_hold_t *txh;
518 
519 	ASSERT0(tx->tx_txg);
520 
521 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
522 	    object, THT_ZAP, add, (uintptr_t)name);
523 	if (txh != NULL)
524 		dmu_tx_hold_zap_impl(txh, name);
525 }
526 
527 void
dmu_tx_hold_zap_by_dnode(dmu_tx_t * tx,dnode_t * dn,int add,const char * name)528 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
529 {
530 	dmu_tx_hold_t *txh;
531 
532 	ASSERT0(tx->tx_txg);
533 	ASSERT(dn != NULL);
534 
535 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
536 	if (txh != NULL)
537 		dmu_tx_hold_zap_impl(txh, name);
538 }
539 
540 void
dmu_tx_hold_bonus(dmu_tx_t * tx,uint64_t object)541 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
542 {
543 	dmu_tx_hold_t *txh;
544 
545 	ASSERT(tx->tx_txg == 0);
546 
547 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
548 	    object, THT_BONUS, 0, 0);
549 	if (txh)
550 		dmu_tx_count_dnode(txh);
551 }
552 
553 void
dmu_tx_hold_bonus_by_dnode(dmu_tx_t * tx,dnode_t * dn)554 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
555 {
556 	dmu_tx_hold_t *txh;
557 
558 	ASSERT0(tx->tx_txg);
559 
560 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
561 	if (txh)
562 		dmu_tx_count_dnode(txh);
563 }
564 
565 void
dmu_tx_hold_space(dmu_tx_t * tx,uint64_t space)566 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
567 {
568 	dmu_tx_hold_t *txh;
569 	ASSERT(tx->tx_txg == 0);
570 
571 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
572 	    DMU_NEW_OBJECT, THT_SPACE, space, 0);
573 
574 	(void) zfs_refcount_add_many(&txh->txh_space_towrite, space, FTAG);
575 }
576 
577 #ifdef ZFS_DEBUG
578 void
dmu_tx_dirty_buf(dmu_tx_t * tx,dmu_buf_impl_t * db)579 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
580 {
581 	boolean_t match_object = B_FALSE;
582 	boolean_t match_offset = B_FALSE;
583 
584 	DB_DNODE_ENTER(db);
585 	dnode_t *dn = DB_DNODE(db);
586 	ASSERT(tx->tx_txg != 0);
587 	ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
588 	ASSERT3U(dn->dn_object, ==, db->db.db_object);
589 
590 	if (tx->tx_anyobj) {
591 		DB_DNODE_EXIT(db);
592 		return;
593 	}
594 
595 	/* XXX No checking on the meta dnode for now */
596 	if (db->db.db_object == DMU_META_DNODE_OBJECT) {
597 		DB_DNODE_EXIT(db);
598 		return;
599 	}
600 
601 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
602 	    txh = list_next(&tx->tx_holds, txh)) {
603 		ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
604 		if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
605 			match_object = TRUE;
606 		if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
607 			int datablkshift = dn->dn_datablkshift ?
608 			    dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
609 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
610 			int shift = datablkshift + epbs * db->db_level;
611 			uint64_t beginblk = shift >= 64 ? 0 :
612 			    (txh->txh_arg1 >> shift);
613 			uint64_t endblk = shift >= 64 ? 0 :
614 			    ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
615 			uint64_t blkid = db->db_blkid;
616 
617 			/* XXX txh_arg2 better not be zero... */
618 
619 			dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
620 			    txh->txh_type, beginblk, endblk);
621 
622 			switch (txh->txh_type) {
623 			case THT_WRITE:
624 				if (blkid >= beginblk && blkid <= endblk)
625 					match_offset = TRUE;
626 				/*
627 				 * We will let this hold work for the bonus
628 				 * or spill buffer so that we don't need to
629 				 * hold it when creating a new object.
630 				 */
631 				if (blkid == DMU_BONUS_BLKID ||
632 				    blkid == DMU_SPILL_BLKID)
633 					match_offset = TRUE;
634 				/*
635 				 * They might have to increase nlevels,
636 				 * thus dirtying the new TLIBs.  Or the
637 				 * might have to change the block size,
638 				 * thus dirying the new lvl=0 blk=0.
639 				 */
640 				if (blkid == 0)
641 					match_offset = TRUE;
642 				break;
643 			case THT_FREE:
644 				/*
645 				 * We will dirty all the level 1 blocks in
646 				 * the free range and perhaps the first and
647 				 * last level 0 block.
648 				 */
649 				if (blkid >= beginblk && (blkid <= endblk ||
650 				    txh->txh_arg2 == DMU_OBJECT_END))
651 					match_offset = TRUE;
652 				break;
653 			case THT_SPILL:
654 				if (blkid == DMU_SPILL_BLKID)
655 					match_offset = TRUE;
656 				break;
657 			case THT_BONUS:
658 				if (blkid == DMU_BONUS_BLKID)
659 					match_offset = TRUE;
660 				break;
661 			case THT_ZAP:
662 				match_offset = TRUE;
663 				break;
664 			case THT_NEWOBJECT:
665 				match_object = TRUE;
666 				break;
667 			default:
668 				ASSERT(!"bad txh_type");
669 			}
670 		}
671 		if (match_object && match_offset) {
672 			DB_DNODE_EXIT(db);
673 			return;
674 		}
675 	}
676 	DB_DNODE_EXIT(db);
677 	panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
678 	    (u_longlong_t)db->db.db_object, db->db_level,
679 	    (u_longlong_t)db->db_blkid);
680 }
681 #endif
682 
683 /*
684  * If we can't do 10 iops, something is wrong.  Let us go ahead
685  * and hit zfs_dirty_data_max.
686  */
687 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
688 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
689 
690 /*
691  * We delay transactions when we've determined that the backend storage
692  * isn't able to accommodate the rate of incoming writes.
693  *
694  * If there is already a transaction waiting, we delay relative to when
695  * that transaction finishes waiting.  This way the calculated min_time
696  * is independent of the number of threads concurrently executing
697  * transactions.
698  *
699  * If we are the only waiter, wait relative to when the transaction
700  * started, rather than the current time.  This credits the transaction for
701  * "time already served", e.g. reading indirect blocks.
702  *
703  * The minimum time for a transaction to take is calculated as:
704  *     min_time = scale * (dirty - min) / (max - dirty)
705  *     min_time is then capped at zfs_delay_max_ns.
706  *
707  * The delay has two degrees of freedom that can be adjusted via tunables.
708  * The percentage of dirty data at which we start to delay is defined by
709  * zfs_delay_min_dirty_percent. This should typically be at or above
710  * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
711  * delay after writing at full speed has failed to keep up with the incoming
712  * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
713  * speaking, this variable determines the amount of delay at the midpoint of
714  * the curve.
715  *
716  * delay
717  *  10ms +-------------------------------------------------------------*+
718  *       |                                                             *|
719  *   9ms +                                                             *+
720  *       |                                                             *|
721  *   8ms +                                                             *+
722  *       |                                                            * |
723  *   7ms +                                                            * +
724  *       |                                                            * |
725  *   6ms +                                                            * +
726  *       |                                                            * |
727  *   5ms +                                                           *  +
728  *       |                                                           *  |
729  *   4ms +                                                           *  +
730  *       |                                                           *  |
731  *   3ms +                                                          *   +
732  *       |                                                          *   |
733  *   2ms +                                              (midpoint) *    +
734  *       |                                                  |    **     |
735  *   1ms +                                                  v ***       +
736  *       |             zfs_delay_scale ---------->     ********         |
737  *     0 +-------------------------------------*********----------------+
738  *       0%                    <- zfs_dirty_data_max ->               100%
739  *
740  * Note that since the delay is added to the outstanding time remaining on the
741  * most recent transaction, the delay is effectively the inverse of IOPS.
742  * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
743  * was chosen such that small changes in the amount of accumulated dirty data
744  * in the first 3/4 of the curve yield relatively small differences in the
745  * amount of delay.
746  *
747  * The effects can be easier to understand when the amount of delay is
748  * represented on a log scale:
749  *
750  * delay
751  * 100ms +-------------------------------------------------------------++
752  *       +                                                              +
753  *       |                                                              |
754  *       +                                                             *+
755  *  10ms +                                                             *+
756  *       +                                                           ** +
757  *       |                                              (midpoint)  **  |
758  *       +                                                  |     **    +
759  *   1ms +                                                  v ****      +
760  *       +             zfs_delay_scale ---------->        *****         +
761  *       |                                             ****             |
762  *       +                                          ****                +
763  * 100us +                                        **                    +
764  *       +                                       *                      +
765  *       |                                      *                       |
766  *       +                                     *                        +
767  *  10us +                                     *                        +
768  *       +                                                              +
769  *       |                                                              |
770  *       +                                                              +
771  *       +--------------------------------------------------------------+
772  *       0%                    <- zfs_dirty_data_max ->               100%
773  *
774  * Note here that only as the amount of dirty data approaches its limit does
775  * the delay start to increase rapidly. The goal of a properly tuned system
776  * should be to keep the amount of dirty data out of that range by first
777  * ensuring that the appropriate limits are set for the I/O scheduler to reach
778  * optimal throughput on the backend storage, and then by changing the value
779  * of zfs_delay_scale to increase the steepness of the curve.
780  */
781 static void
dmu_tx_delay(dmu_tx_t * tx,uint64_t dirty)782 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
783 {
784 	dsl_pool_t *dp = tx->tx_pool;
785 	uint64_t delay_min_bytes =
786 	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
787 	hrtime_t wakeup, min_tx_time, now;
788 
789 	if (dirty <= delay_min_bytes)
790 		return;
791 
792 	/*
793 	 * The caller has already waited until we are under the max.
794 	 * We make them pass us the amount of dirty data so we don't
795 	 * have to handle the case of it being >= the max, which could
796 	 * cause a divide-by-zero if it's == the max.
797 	 */
798 	ASSERT3U(dirty, <, zfs_dirty_data_max);
799 
800 	now = gethrtime();
801 	min_tx_time = zfs_delay_scale *
802 	    (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
803 	if (now > tx->tx_start + min_tx_time)
804 		return;
805 
806 	min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
807 
808 	DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
809 	    uint64_t, min_tx_time);
810 
811 	mutex_enter(&dp->dp_lock);
812 	wakeup = MAX(tx->tx_start + min_tx_time,
813 	    dp->dp_last_wakeup + min_tx_time);
814 	dp->dp_last_wakeup = wakeup;
815 	mutex_exit(&dp->dp_lock);
816 
817 #ifdef _KERNEL
818 	mutex_enter(&curthread->t_delay_lock);
819 	while (cv_timedwait_hires(&curthread->t_delay_cv,
820 	    &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
821 	    CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
822 		continue;
823 	mutex_exit(&curthread->t_delay_lock);
824 #else
825 	hrtime_t delta = wakeup - gethrtime();
826 	struct timespec ts;
827 	ts.tv_sec = delta / NANOSEC;
828 	ts.tv_nsec = delta % NANOSEC;
829 	(void) nanosleep(&ts, NULL);
830 #endif
831 }
832 
833 /*
834  * This routine attempts to assign the transaction to a transaction group.
835  * To do so, we must determine if there is sufficient free space on disk.
836  *
837  * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
838  * on it), then it is assumed that there is sufficient free space,
839  * unless there's insufficient slop space in the pool (see the comment
840  * above spa_slop_shift in spa_misc.c).
841  *
842  * If it is not a "netfree" transaction, then if the data already on disk
843  * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
844  * ENOSPC.  Otherwise, if the current rough estimate of pending changes,
845  * plus the rough estimate of this transaction's changes, may exceed the
846  * allowed usage, then this will fail with ERESTART, which will cause the
847  * caller to wait for the pending changes to be written to disk (by waiting
848  * for the next TXG to open), and then check the space usage again.
849  *
850  * The rough estimate of pending changes is comprised of the sum of:
851  *
852  *  - this transaction's holds' txh_space_towrite
853  *
854  *  - dd_tempreserved[], which is the sum of in-flight transactions'
855  *    holds' txh_space_towrite (i.e. those transactions that have called
856  *    dmu_tx_assign() but not yet called dmu_tx_commit()).
857  *
858  *  - dd_space_towrite[], which is the amount of dirtied dbufs.
859  *
860  * Note that all of these values are inflated by spa_get_worst_case_asize(),
861  * which means that we may get ERESTART well before we are actually in danger
862  * of running out of space, but this also mitigates any small inaccuracies
863  * in the rough estimate (e.g. txh_space_towrite doesn't take into account
864  * indirect blocks, and dd_space_towrite[] doesn't take into account changes
865  * to the MOS).
866  *
867  * Note that due to this algorithm, it is possible to exceed the allowed
868  * usage by one transaction.  Also, as we approach the allowed usage,
869  * we will allow a very limited amount of changes into each TXG, thus
870  * decreasing performance.
871  */
872 static int
dmu_tx_try_assign(dmu_tx_t * tx,uint64_t txg_how)873 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
874 {
875 	spa_t *spa = tx->tx_pool->dp_spa;
876 
877 	ASSERT0(tx->tx_txg);
878 
879 	if (tx->tx_err)
880 		return (tx->tx_err);
881 
882 	if (spa_suspended(spa)) {
883 		/*
884 		 * If the user has indicated a blocking failure mode
885 		 * then return ERESTART which will block in dmu_tx_wait().
886 		 * Otherwise, return EIO so that an error can get
887 		 * propagated back to the VOP calls.
888 		 *
889 		 * Note that we always honor the txg_how flag regardless
890 		 * of the failuremode setting.
891 		 */
892 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
893 		    !(txg_how & TXG_WAIT))
894 			return (SET_ERROR(EIO));
895 
896 		return (SET_ERROR(ERESTART));
897 	}
898 
899 	if (!tx->tx_dirty_delayed &&
900 	    dsl_pool_need_dirty_delay(tx->tx_pool)) {
901 		tx->tx_wait_dirty = B_TRUE;
902 		return (SET_ERROR(ERESTART));
903 	}
904 
905 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
906 	tx->tx_needassign_txh = NULL;
907 
908 	/*
909 	 * NB: No error returns are allowed after txg_hold_open, but
910 	 * before processing the dnode holds, due to the
911 	 * dmu_tx_unassign() logic.
912 	 */
913 
914 	uint64_t towrite = 0;
915 	uint64_t tohold = 0;
916 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
917 	    txh = list_next(&tx->tx_holds, txh)) {
918 		dnode_t *dn = txh->txh_dnode;
919 		if (dn != NULL) {
920 			mutex_enter(&dn->dn_mtx);
921 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
922 				mutex_exit(&dn->dn_mtx);
923 				tx->tx_needassign_txh = txh;
924 				return (SET_ERROR(ERESTART));
925 			}
926 			if (dn->dn_assigned_txg == 0)
927 				dn->dn_assigned_txg = tx->tx_txg;
928 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
929 			(void) zfs_refcount_add(&dn->dn_tx_holds, tx);
930 			mutex_exit(&dn->dn_mtx);
931 		}
932 		towrite += zfs_refcount_count(&txh->txh_space_towrite);
933 		tohold += zfs_refcount_count(&txh->txh_memory_tohold);
934 	}
935 
936 	/* needed allocation: worst-case estimate of write space */
937 	uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite);
938 	/* calculate memory footprint estimate */
939 	uint64_t memory = towrite + tohold;
940 
941 	if (tx->tx_dir != NULL && asize != 0) {
942 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
943 		    asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx);
944 		if (err != 0)
945 			return (err);
946 	}
947 
948 	return (0);
949 }
950 
951 static void
dmu_tx_unassign(dmu_tx_t * tx)952 dmu_tx_unassign(dmu_tx_t *tx)
953 {
954 	if (tx->tx_txg == 0)
955 		return;
956 
957 	txg_rele_to_quiesce(&tx->tx_txgh);
958 
959 	/*
960 	 * Walk the transaction's hold list, removing the hold on the
961 	 * associated dnode, and notifying waiters if the refcount drops to 0.
962 	 */
963 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
964 	    txh != tx->tx_needassign_txh;
965 	    txh = list_next(&tx->tx_holds, txh)) {
966 		dnode_t *dn = txh->txh_dnode;
967 
968 		if (dn == NULL)
969 			continue;
970 		mutex_enter(&dn->dn_mtx);
971 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
972 
973 		if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
974 			dn->dn_assigned_txg = 0;
975 			cv_broadcast(&dn->dn_notxholds);
976 		}
977 		mutex_exit(&dn->dn_mtx);
978 	}
979 
980 	txg_rele_to_sync(&tx->tx_txgh);
981 
982 	tx->tx_lasttried_txg = tx->tx_txg;
983 	tx->tx_txg = 0;
984 }
985 
986 /*
987  * Assign tx to a transaction group; txg_how is a bitmask:
988  *
989  * If TXG_WAIT is set and the currently open txg is full, this function
990  * will wait until there's a new txg. This should be used when no locks
991  * are being held. With this bit set, this function will only fail if
992  * we're truly out of space (or over quota).
993  *
994  * If TXG_WAIT is *not* set and we can't assign into the currently open
995  * txg without blocking, this function will return immediately with
996  * ERESTART. This should be used whenever locks are being held.  On an
997  * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
998  * and try again.
999  *
1000  * If TXG_NOTHROTTLE is set, this indicates that this tx should not be
1001  * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
1002  * details on the throttle). This is used by the VFS operations, after
1003  * they have already called dmu_tx_wait() (though most likely on a
1004  * different tx).
1005  */
1006 int
dmu_tx_assign(dmu_tx_t * tx,uint64_t txg_how)1007 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1008 {
1009 	int err;
1010 
1011 	ASSERT(tx->tx_txg == 0);
1012 	ASSERT0(txg_how & ~(TXG_WAIT | TXG_NOTHROTTLE));
1013 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1014 
1015 	/* If we might wait, we must not hold the config lock. */
1016 	IMPLY((txg_how & TXG_WAIT), !dsl_pool_config_held(tx->tx_pool));
1017 
1018 	if ((txg_how & TXG_NOTHROTTLE))
1019 		tx->tx_dirty_delayed = B_TRUE;
1020 
1021 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1022 		dmu_tx_unassign(tx);
1023 
1024 		if (err != ERESTART || !(txg_how & TXG_WAIT))
1025 			return (err);
1026 
1027 		dmu_tx_wait(tx);
1028 	}
1029 
1030 	txg_rele_to_quiesce(&tx->tx_txgh);
1031 
1032 	return (0);
1033 }
1034 
1035 void
dmu_tx_wait(dmu_tx_t * tx)1036 dmu_tx_wait(dmu_tx_t *tx)
1037 {
1038 	spa_t *spa = tx->tx_pool->dp_spa;
1039 	dsl_pool_t *dp = tx->tx_pool;
1040 
1041 	ASSERT(tx->tx_txg == 0);
1042 	ASSERT(!dsl_pool_config_held(tx->tx_pool));
1043 
1044 	if (tx->tx_wait_dirty) {
1045 		/*
1046 		 * dmu_tx_try_assign() has determined that we need to wait
1047 		 * because we've consumed much or all of the dirty buffer
1048 		 * space.
1049 		 */
1050 		mutex_enter(&dp->dp_lock);
1051 		while (dp->dp_dirty_total >= zfs_dirty_data_max)
1052 			cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1053 		uint64_t dirty = dp->dp_dirty_total;
1054 		mutex_exit(&dp->dp_lock);
1055 
1056 		dmu_tx_delay(tx, dirty);
1057 
1058 		tx->tx_wait_dirty = B_FALSE;
1059 
1060 		/*
1061 		 * Note: setting tx_dirty_delayed only has effect if the
1062 		 * caller used TX_WAIT.  Otherwise they are going to
1063 		 * destroy this tx and try again.  The common case,
1064 		 * zfs_write(), uses TX_WAIT.
1065 		 */
1066 		tx->tx_dirty_delayed = B_TRUE;
1067 	} else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1068 		/*
1069 		 * If the pool is suspended we need to wait until it
1070 		 * is resumed.  Note that it's possible that the pool
1071 		 * has become active after this thread has tried to
1072 		 * obtain a tx.  If that's the case then tx_lasttried_txg
1073 		 * would not have been set.
1074 		 */
1075 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1076 	} else if (tx->tx_needassign_txh) {
1077 		/*
1078 		 * A dnode is assigned to the quiescing txg.  Wait for its
1079 		 * transaction to complete.
1080 		 */
1081 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1082 
1083 		mutex_enter(&dn->dn_mtx);
1084 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1085 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1086 		mutex_exit(&dn->dn_mtx);
1087 		tx->tx_needassign_txh = NULL;
1088 	} else {
1089 		/*
1090 		 * If we have a lot of dirty data just wait until we sync
1091 		 * out a TXG at which point we'll hopefully have synced
1092 		 * a portion of the changes.
1093 		 */
1094 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1095 	}
1096 }
1097 
1098 static void
dmu_tx_destroy(dmu_tx_t * tx)1099 dmu_tx_destroy(dmu_tx_t *tx)
1100 {
1101 	dmu_tx_hold_t *txh;
1102 
1103 	while ((txh = list_head(&tx->tx_holds)) != NULL) {
1104 		dnode_t *dn = txh->txh_dnode;
1105 
1106 		list_remove(&tx->tx_holds, txh);
1107 		zfs_refcount_destroy_many(&txh->txh_space_towrite,
1108 		    zfs_refcount_count(&txh->txh_space_towrite));
1109 		zfs_refcount_destroy_many(&txh->txh_memory_tohold,
1110 		    zfs_refcount_count(&txh->txh_memory_tohold));
1111 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1112 		if (dn != NULL)
1113 			dnode_rele(dn, tx);
1114 	}
1115 
1116 	list_destroy(&tx->tx_callbacks);
1117 	list_destroy(&tx->tx_holds);
1118 	kmem_free(tx, sizeof (dmu_tx_t));
1119 }
1120 
1121 void
dmu_tx_commit(dmu_tx_t * tx)1122 dmu_tx_commit(dmu_tx_t *tx)
1123 {
1124 	ASSERT(tx->tx_txg != 0);
1125 
1126 	/*
1127 	 * Go through the transaction's hold list and remove holds on
1128 	 * associated dnodes, notifying waiters if no holds remain.
1129 	 */
1130 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1131 	    txh = list_next(&tx->tx_holds, txh)) {
1132 		dnode_t *dn = txh->txh_dnode;
1133 
1134 		if (dn == NULL)
1135 			continue;
1136 
1137 		mutex_enter(&dn->dn_mtx);
1138 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1139 
1140 		if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1141 			dn->dn_assigned_txg = 0;
1142 			cv_broadcast(&dn->dn_notxholds);
1143 		}
1144 		mutex_exit(&dn->dn_mtx);
1145 	}
1146 
1147 	if (tx->tx_tempreserve_cookie)
1148 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1149 
1150 	if (!list_is_empty(&tx->tx_callbacks))
1151 		txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1152 
1153 	if (tx->tx_anyobj == FALSE)
1154 		txg_rele_to_sync(&tx->tx_txgh);
1155 
1156 	dmu_tx_destroy(tx);
1157 }
1158 
1159 void
dmu_tx_abort(dmu_tx_t * tx)1160 dmu_tx_abort(dmu_tx_t *tx)
1161 {
1162 	ASSERT(tx->tx_txg == 0);
1163 
1164 	/*
1165 	 * Call any registered callbacks with an error code.
1166 	 */
1167 	if (!list_is_empty(&tx->tx_callbacks))
1168 		dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1169 
1170 	dmu_tx_destroy(tx);
1171 }
1172 
1173 uint64_t
dmu_tx_get_txg(dmu_tx_t * tx)1174 dmu_tx_get_txg(dmu_tx_t *tx)
1175 {
1176 	ASSERT(tx->tx_txg != 0);
1177 	return (tx->tx_txg);
1178 }
1179 
1180 dsl_pool_t *
dmu_tx_pool(dmu_tx_t * tx)1181 dmu_tx_pool(dmu_tx_t *tx)
1182 {
1183 	ASSERT(tx->tx_pool != NULL);
1184 	return (tx->tx_pool);
1185 }
1186 
1187 void
dmu_tx_callback_register(dmu_tx_t * tx,dmu_tx_callback_func_t * func,void * data)1188 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1189 {
1190 	dmu_tx_callback_t *dcb;
1191 
1192 	dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1193 
1194 	dcb->dcb_func = func;
1195 	dcb->dcb_data = data;
1196 
1197 	list_insert_tail(&tx->tx_callbacks, dcb);
1198 }
1199 
1200 /*
1201  * Call all the commit callbacks on a list, with a given error code.
1202  */
1203 void
dmu_tx_do_callbacks(list_t * cb_list,int error)1204 dmu_tx_do_callbacks(list_t *cb_list, int error)
1205 {
1206 	dmu_tx_callback_t *dcb;
1207 
1208 	while ((dcb = list_head(cb_list)) != NULL) {
1209 		list_remove(cb_list, dcb);
1210 		dcb->dcb_func(dcb->dcb_data, error);
1211 		kmem_free(dcb, sizeof (dmu_tx_callback_t));
1212 	}
1213 }
1214 
1215 /*
1216  * Interface to hold a bunch of attributes.
1217  * used for creating new files.
1218  * attrsize is the total size of all attributes
1219  * to be added during object creation
1220  *
1221  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1222  */
1223 
1224 /*
1225  * hold necessary attribute name for attribute registration.
1226  * should be a very rare case where this is needed.  If it does
1227  * happen it would only happen on the first write to the file system.
1228  */
1229 static void
dmu_tx_sa_registration_hold(sa_os_t * sa,dmu_tx_t * tx)1230 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1231 {
1232 	if (!sa->sa_need_attr_registration)
1233 		return;
1234 
1235 	for (int i = 0; i != sa->sa_num_attrs; i++) {
1236 		if (!sa->sa_attr_table[i].sa_registered) {
1237 			if (sa->sa_reg_attr_obj)
1238 				dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1239 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1240 			else
1241 				dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1242 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1243 		}
1244 	}
1245 }
1246 
1247 void
dmu_tx_hold_spill(dmu_tx_t * tx,uint64_t object)1248 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1249 {
1250 	dmu_tx_hold_t *txh;
1251 
1252 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1253 	    THT_SPILL, 0, 0);
1254 	if (txh != NULL)
1255 		(void) zfs_refcount_add_many(&txh->txh_space_towrite,
1256 		    SPA_OLD_MAXBLOCKSIZE, FTAG);
1257 }
1258 
1259 void
dmu_tx_hold_sa_create(dmu_tx_t * tx,int attrsize)1260 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1261 {
1262 	sa_os_t *sa = tx->tx_objset->os_sa;
1263 
1264 	dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1265 
1266 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1267 		return;
1268 
1269 	if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1270 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1271 	} else {
1272 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1273 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1274 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1275 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1276 	}
1277 
1278 	dmu_tx_sa_registration_hold(sa, tx);
1279 
1280 	if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill)
1281 		return;
1282 
1283 	(void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1284 	    THT_SPILL, 0, 0);
1285 }
1286 
1287 /*
1288  * Hold SA attribute
1289  *
1290  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1291  *
1292  * variable_size is the total size of all variable sized attributes
1293  * passed to this function.  It is not the total size of all
1294  * variable size attributes that *may* exist on this object.
1295  */
1296 void
dmu_tx_hold_sa(dmu_tx_t * tx,sa_handle_t * hdl,boolean_t may_grow)1297 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1298 {
1299 	uint64_t object;
1300 	sa_os_t *sa = tx->tx_objset->os_sa;
1301 
1302 	ASSERT(hdl != NULL);
1303 
1304 	object = sa_handle_object(hdl);
1305 
1306 	dmu_tx_hold_bonus(tx, object);
1307 
1308 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1309 		return;
1310 
1311 	if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1312 	    tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1313 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1314 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1315 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1316 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1317 	}
1318 
1319 	dmu_tx_sa_registration_hold(sa, tx);
1320 
1321 	if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1322 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1323 
1324 	if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1325 		ASSERT(tx->tx_txg == 0);
1326 		dmu_tx_hold_spill(tx, object);
1327 	} else {
1328 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1329 		dnode_t *dn;
1330 
1331 		DB_DNODE_ENTER(db);
1332 		dn = DB_DNODE(db);
1333 		if (dn->dn_have_spill) {
1334 			ASSERT(tx->tx_txg == 0);
1335 			dmu_tx_hold_spill(tx, object);
1336 		}
1337 		DB_DNODE_EXIT(db);
1338 	}
1339 }
1340